{"repo_id":"Human3R","entity_id":"py:demo","uri":"program://Human3R/module/demo#L1-L685","kind":"module","name":"demo","path":"demo.py","language":"python","start_line":1,"end_line":685,"context_start_line":1,"context_end_line":685,"code":"#!/usr/bin/env python3\n\"\"\"\nModified from CUT3R: https://github.com/CUT3R/CUT3R\n\nOnline Human-Scene Reconstruction Inference and Visualization Script\n\nThis script performs inference using the ARCroco3DStereo model and visualizes the\nresulting 3D scene point clouds and SMPLX sequences with the SceneHumanViewer. \nUse the command-line arguments to adjust parameters \nsuch as the model checkpoint path, image sequence directory, image size, device, etc.\n\nExample:\n    python demo.py --model_path src/human3r.pth --size 512 \\\n        --seq_path examples/GoodMornin1.mp4 --subsample 1 --vis_threshold 2 \\\n        --downsample_factor 1 --use_ttt3r --reset_interval 100\n\"\"\"\n\nimport os\nimport numpy as np\nimport torch\nimport time\nimport glob\nimport random\nimport cv2\nimport argparse\nimport tempfile\nimport shutil\nfrom copy import deepcopy\nfrom add_ckpt_path import add_path_to_dust3r\nimport imageio.v2 as iio\nimport roma\n\n# Set random seed for reproducibility.\nrandom.seed(42)\n\n\ndef parse_args():\n    \"\"\"Parse command-line arguments.\"\"\"\n    parser = argparse.ArgumentParser(\n        description=\"Run 3D point cloud inference and visualization using ARCroco3DStereo.\"\n    )\n    parser.add_argument(\n        \"--model_path\",\n        type=str,\n        default=\"src/cut3r_512_dpt_4_64.pth\",\n        help=\"Path to the pretrained model checkpoint.\",\n    )\n    parser.add_argument(\n        \"--seq_path\",\n        type=str,\n        default=\"\",\n        help=\"Path to the directory containing the image sequence.\",\n    )\n    parser.add_argument(\n        \"--device\",\n        type=str,\n        default=\"cuda\",\n        help=\"Device to run inference on (e.g., 'cuda' or 'cpu').\",\n    )\n    parser.add_argument(\n        \"--size\",\n        type=int,\n        default=\"512\",\n        help=\"Shape that input images will be rescaled to; if using 224+linear model, choose 224 otherwise 512\",\n    )\n    parser.add_argument(\n        \"--vis_threshold\",\n        type=float,\n        default=1.5,\n        help=\"Visualization threshold for the viewer. Ranging from 1 to INF\",\n    )\n    parser.add_argument(\n        \"--msk_threshold\",\n        type=float,\n        default=0.1,\n        help=\"Mask threshold. Ranging from 0 to 1\",\n    )\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"./tmp\",\n        help=\"value for tempfile.tempdir\",\n    )\n    parser.add_argument(\n        \"--save_smpl\",\n        action=\"store_true\",\n        help=\"Save smpl results.\",\n    )\n    parser.add_argument(\n        \"--save_video\",\n        action=\"store_true\",\n        help=\"Save smpl video.\",\n    )\n    parser.add_argument(\n        \"--max_frames\",\n        type=int,\n        default=None,\n        help=\"Max frames to use. Default is None (use all images).\",\n    )\n    parser.add_argument(\n        \"--subsample\",\n        type=int,\n        default=1,\n        help=\"Subsample factor for input images. Default is 1 (use all images).\",\n    )\n    parser.add_argument(\n        \"--reset_interval\", \n        type=int, \n        default=10000000\n        )\n    parser.add_argument(\n        \"--use_ttt3r\",\n        action=\"store_true\",\n        help=\"Use TTT3R.\",\n        default=False\n    )\n    parser.add_argument(\n        \"--downsample_factor\",\n        type=int,\n        default=10,\n        help=\"Point cloud downsample factor for the viewer\",\n    )\n    parser.add_argument(\n        \"--smpl_downsample\",\n        type=int,\n        default=1,\n        help=\"SMPL sequence downsample factor for the viewer\",\n    )\n    parser.add_argument(\n        \"--camera_downsample\",\n        type=int,\n        default=1,\n        help=\"Camera motion downsample factor for the viewer\",\n    )\n    parser.add_argument(\n        \"--mask_morph\",\n        type=int,\n        default=10,\n        help=\"Mask morphology for the viewer\",\n    )\n    return parser.parse_args()\n\n\ndef prepare_input(\n    img_paths, \n    img_mask, \n    size, \n    raymaps=None, \n    raymap_mask=None, \n    revisit=1, \n    update=True, \n    img_res=None, \n    reset_interval=100\n):\n    \"\"\"\n    Prepare input views for inference from a list of image paths.\n\n    Args:\n        img_paths (list): List of image file paths.\n        img_mask (list of bool): Flags indicating valid images.\n        size (int): Target image size.\n        raymaps (list, optional): List of ray maps.\n        raymap_mask (list, optional): Flags indicating valid ray maps.\n        revisit (int): How many times to revisit each view.\n        update (bool): Whether to update the state on revisits.\n\n    Returns:\n        list: A list of view dictionaries.\n    \"\"\"\n    # Import image loader (delayed import needed after adding ckpt path).\n    from src.dust3r.utils.image import load_images, pad_image\n    from dust3r.utils.geometry import get_camera_parameters\n\n    images = load_images(img_paths, size=size)\n    if img_res is not None:\n        K_mhmr = get_camera_parameters(img_res, device=\"cpu\") # if use pseudo K\n\n    views = []\n    if raymaps is None and raymap_mask is None:\n        # Only images are provided.\n        for i in range(len(images)):\n            view = {\n                \"img\": images[i][\"img\"],\n                \"ray_map\": torch.full(\n                    (\n                        images[i][\"img\"].shape[0],\n                        6,\n                        images[i][\"img\"].shape[-2],\n                        images[i][\"img\"].shape[-1],\n                    ),\n                    torch.nan,\n                ),\n                \"true_shape\": torch.from_numpy(images[i][\"true_shape\"]),\n                \"idx\": i,\n                \"instance\": str(i),\n                \"camera_pose\": torch.from_numpy(\n                    np.eye(4, dtype=np.float32)\n                    ).unsqueeze(0),\n                \"img_mask\": torch.tensor(True).unsqueeze(0),\n                \"ray_mask\": torch.tensor(False).unsqueeze(0),\n                \"update\": torch.tensor(True).unsqueeze(0),\n                \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n            }\n            if img_res is not None:\n                view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                view[\"K_mhmr\"] = K_mhmr\n            views.append(view)\n            if (i+1) % reset_interval == 0:\n                overlap_view = deepcopy(view)\n                overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                views.append(overlap_view)\n    else:\n        # Combine images and raymaps.\n        num_views = len(images) + len(raymaps)\n        assert len(img_mask) == len(raymap_mask) == num_views\n        assert sum(img_mask) == len(images) and sum(raymap_mask) == len(raymaps)\n\n        j = 0\n        k = 0\n        for i in range(num_views):\n            view = {\n                \"img\": (\n                    images[j][\"img\"]\n                    if img_mask[i]\n                    else torch.full_like(images[0][\"img\"], torch.nan)\n                ),\n                \"ray_map\": (\n                    raymaps[k]\n                    if raymap_mask[i]\n                    else torch.full_like(raymaps[0], torch.nan)\n                ),\n                \"true_shape\": (\n                    torch.from_numpy(images[j][\"true_shape\"])\n                    if img_mask[i]\n                    else torch.from_numpy(np.int32([raymaps[k].shape[1:-1][::-1]]))\n                ),\n                \"idx\": i,\n                \"instance\": str(i),\n                \"camera_pose\": torch.from_numpy(\n                    np.eye(4, dtype=np.float32)\n                    ).unsqueeze(0),\n                \"img_mask\": torch.tensor(img_mask[i]).unsqueeze(0),\n                \"ray_mask\": torch.tensor(raymap_mask[i]).unsqueeze(0),\n                \"update\": torch.tensor(img_mask[i]).unsqueeze(0),\n                \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n            }\n            if img_res is not None:\n                view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                view[\"K_mhmr\"] = K_mhmr\n            if img_mask[i]:\n                j += 1\n            if raymap_mask[i]:\n                k += 1\n            views.append(view)\n            if (i+1) % reset_interval == 0:\n                overlap_view = deepcopy(view)\n                overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                views.append(overlap_view)\n        assert j == len(images) and k == len(raymaps)\n\n    if revisit > 1:\n        new_views = []\n        for r in range(revisit):\n            for i, view in enumerate(views):\n                new_view = deepcopy(view)\n                new_view[\"idx\"] = r * len(views) + i\n                new_view[\"instance\"] = str(r * len(views) + i)\n                if r > 0 and not update:\n                    new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                new_views.append(new_view)\n        return new_views\n\n    return views\n\ndef prepare_output(\n        outputs, outdir, revisit=1, use_pose=True, \n        save_smpl=False, save_video=False, img_res=None, subsample=1):\n    \"\"\"\n    Process inference outputs to generate point clouds and camera parameters for visualization.\n\n    Args:\n        outputs (dict): Inference outputs.\n        revisit (int): Number of revisits per view.\n        use_pose (bool): Whether to transform points using camera pose.\n        save_smpl (bool): Whether to save smpl results.\n        save_video (bool): Whether to save smpl video.\n\n    Returns:\n        tuple: (points, colors, confidence, camera parameters dictionary)\n    \"\"\"\n    from src.dust3r.utils.camera import pose_encoding_to_camera\n    from src.dust3r.post_process import estimate_focal_knowing_depth\n    from src.dust3r.utils.geometry import geotrf, matrix_cumprod\n    from src.dust3r.utils import SMPL_Layer, vis_heatmap, render_meshes\n    from src.dust3r.utils.image import unpad_image\n    from viser_utils import get_color\n\n    # Only keep the outputs corresponding to one full pass.\n    valid_length = len(outputs[\"pred\"]) // revisit\n    outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n    outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n    # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n    reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n    shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n    outputs[\"pred\"] = [\n        pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n    outputs[\"views\"] = [\n        view for view, mask in zip(outputs[\"views\"], shifted_reset_mask) if not mask]\n    reset_mask = reset_mask[~shifted_reset_mask]\n\n    pts3ds_self_ls = [output[\"pts3d_in_self_view\"] for output in outputs[\"pred\"]]\n    pts3ds_other = [output[\"pts3d_in_other_view\"] for output in outputs[\"pred\"]]\n    conf_self = [output[\"conf_self\"] for output in outputs[\"pred\"]]\n    conf_other = [output[\"conf\"] for output in outputs[\"pred\"]]\n    pts3ds_self = torch.cat(pts3ds_self_ls, 0)\n\n    # Recover camera poses.\n    pr_poses = [\n        pose_encoding_to_camera(pred[\"camera_pose\"].clone()).cpu()\n        for pred in outputs[\"pred\"]\n    ]\n\n    # reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n    if reset_mask.any():\n        pr_poses = torch.cat(pr_poses, 0)\n        identity = torch.eye(4, device=pr_poses.device)\n        reset_poses = torch.where(reset_mask.unsqueeze(-1).unsqueeze(-1), pr_poses, identity)\n        cumulative_bases = matrix_cumprod(reset_poses)\n        shifted_bases = torch.cat([identity.unsqueeze(0), cumulative_bases[:-1]], dim=0)\n        pr_poses = torch.einsum('bij,bjk->bik', shifted_bases, pr_poses)\n        # keeps only reset_mask=False pr_poses\n        pr_poses = list(pr_poses.unsqueeze(1).unbind(0))\n\n    R_c2w = torch.cat([pr_pose[:, :3, :3] for pr_pose in pr_poses], 0)\n    t_c2w = torch.cat([pr_pose[:, :3, 3] for pr_pose in pr_poses], 0)\n\n    if use_pose:\n        transformed_pts3ds_other = []\n        for pose, pself in zip(pr_poses, pts3ds_self):\n            transformed_pts3ds_other.append(geotrf(pose, pself.unsqueeze(0)))\n        pts3ds_other = transformed_pts3ds_other\n        conf_other = conf_self\n\n    # Estimate focal length based on depth.\n    B, H, W, _ = pts3ds_self.shape\n    pp = torch.tensor([W // 2, H // 2], device=pts3ds_self.device).float().repeat(B, 1)\n    focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n    colors = [\n        0.5 * (output[\"img\"].permute(0, 2, 3, 1) + 1.0) for output in outputs[\"views\"]\n    ]\n\n    cam_dict = {\n        \"focal\": focal.numpy(),\n        \"pp\": pp.numpy(),\n        \"R\": R_c2w.numpy(),\n        \"t\": t_c2w.numpy(),\n    }\n\n    pts3ds_self_tosave = pts3ds_self  # B, H, W, 3\n    depths_tosave = pts3ds_self_tosave[..., 2]\n    pts3ds_other_tosave = torch.cat(pts3ds_other)  # B, H, W, 3\n    conf_self_tosave = torch.cat(conf_self)  # B, H, W\n    conf_other_tosave = torch.cat(conf_other)  # B, H, W\n    colors_tosave = torch.cat(\n        [\n            0.5 * (output[\"img\"].permute(0, 2, 3, 1) + 1.0)\n            for output in outputs[\"views\"]\n        ]\n    )  # [B, H, W, 3]\n    cam2world_tosave = torch.cat(pr_poses)  # B, 4, 4\n    intrinsics_tosave = (\n        torch.eye(3).unsqueeze(0).repeat(cam2world_tosave.shape[0], 1, 1)\n    )  # B, 3, 3\n    intrinsics_tosave[:, 0, 0] = focal.detach()\n    intrinsics_tosave[:, 1, 1] = focal.detach()\n    intrinsics_tosave[:, 0, 2] = pp[:, 0]\n    intrinsics_tosave[:, 1, 2] = pp[:, 1]\n\n    # get SMPL parameters from outputs\n    smpl_shape = [output.get(\n        \"smpl_shape\", torch.empty(1,0,10))[0] for output in outputs[\"pred\"]]\n    smpl_rotvec = [roma.rotmat_to_rotvec(\n        output.get(\n            \"smpl_rotmat\", torch.empty(1,0,53,3,3))[0]) for output in outputs[\"pred\"]]\n    smpl_transl = [output.get(\n        \"smpl_transl\", torch.empty(1,0,3))[0] for output in outputs[\"pred\"]]\n    smpl_expression = [output.get(\n        \"smpl_expression\", [None])[0] for output in outputs[\"pred\"]]\n    smpl_id = [output.get(\n        \"smpl_id\", torch.empty(1,0))[0] for output in outputs[\"pred\"]]\n    # smpl_loc = [output.get(\n    #     \"smpl_loc\", torch.empty(1,0,2))[0] for output in outputs[\"pred\"]]\n    # K_mhmr = [output.get(\n    #     \"K_mhmr\", torch.empty(1,0,3))[0] for output in outputs[\"views\"]]\n        \n    if save_smpl:\n        smpl_scores = [\n            output[\"smpl_scores\"][...,0] for output in outputs[\"pred\"]]\n        if img_res is not None:\n            smpl_scores = [\n                unpad_image(s, [H, W])[0] for s in smpl_scores]\n\n    has_mask = \"msk\" in outputs[\"pred\"][0]\n    if has_mask:\n        msks = [output[\"msk\"][...,0] for output in outputs[\"pred\"]]\n        if img_res is not None:\n            msks = [unpad_image(m, [H, W]) for m in msks]\n    else:\n        msks = [torch.zeros(1, H, W) for _ in range(B)]\n\n    # SMPL layer\n    smpl_layer = SMPL_Layer(type='smplx', \n                            gender='neutral', \n                            num_betas=smpl_shape[0].shape[-1], \n                            kid=False, \n                            person_center='head')\n    smpl_faces = smpl_layer.bm_x.faces\n\n    # os.makedirs(os.path.join(outdir, \"depth\"), exist_ok=True)\n    # os.makedirs(os.path.join(outdir, \"conf\"), exist_ok=True)\n    # os.makedirs(os.path.join(outdir, \"color\"), exist_ok=True)\n    # os.makedirs(os.path.join(outdir, \"camera\"), exist_ok=True)\n\n    all_verts = []\n    for f_id in range(B):\n        n_humans_i = smpl_shape[f_id].shape[0]\n        \n        if n_humans_i > 0:\n            with torch.no_grad():\n                smpl_out = smpl_layer(\n                    smpl_rotvec[f_id], \n                    smpl_shape[f_id], \n                    smpl_transl[f_id], \n                    None, None, \n                    K=intrinsics_tosave[f_id].expand(n_humans_i, -1 , -1), \n                    expression=smpl_expression[f_id])\n        \n        depth = depths_tosave[f_id].numpy()\n        conf = conf_self_tosave[f_id].numpy()\n        color = colors_tosave[f_id].numpy()\n        c2w = cam2world_tosave[f_id].numpy()\n        intrins = intrinsics_tosave[f_id].numpy()\n\n        if n_humans_i > 0:\n            # transform smpl verts to world coordinates\n            all_verts.append(geotrf(pr_poses[f_id], smpl_out['smpl_v3d'].unsqueeze(0))[0])\n            pr_verts = [t.numpy() for t in smpl_out['smpl_v3d'].unbind(0)]\n            pr_faces = [smpl_faces] * n_humans_i\n        else:\n            pr_verts = []\n            pr_faces = []\n            all_verts.append(torch.empty(0))\n\n        if save_smpl:\n            hm = vis_heatmap(colors_tosave[f_id], smpl_scores[f_id]).numpy()\n            img_array_np = (color * 255).astype(np.uint8)\n            smpl_rend = render_meshes(img_array_np.copy(), pr_verts, pr_faces,\n                                        {'focal': intrins[[0,1],[0,1]], \n                                        'princpt': intrins[[0,1],[-1,-1]]},\n                                        color=[get_color(i)/255 for i in smpl_id[f_id]])\n            if has_mask:\n                msk_array_np = vis_heatmap(colors_tosave[f_id], msks[f_id][0]).numpy()\n                color_smpl = np.concatenate([\n                    img_array_np, \n                    (msk_array_np * 255).astype(np.uint8), \n                    (hm * 255).astype(np.uint8), \n                    smpl_rend], 1)\n            else:\n                color_smpl = np.concatenate([\n                    img_array_np, \n                    (hm * 255).astype(np.uint8), \n                    smpl_rend], 1)\n        \n        # np.save(os.path.join(outdir, \"depth\", f\"{f_id:06d}.npy\"), depth)\n        # np.save(os.path.join(outdir, \"conf\", f\"{f_id:06d}.npy\"), conf)\n        # iio.imwrite(\n        #     os.path.join(outdir, \"color\", f\"{f_id:06d}.png\"),\n        #     (color * 255).astype(np.uint8),\n        # )\n        # np.savez(\n        #     os.path.join(outdir, \"camera\", f\"{f_id:06d}.npz\"),\n        #     pose=c2w,\n        #     intrinsics=intrins,\n        # )\n\n        # Save smpl results\n        if save_smpl:\n            os.makedirs(os.path.join(outdir, \"color_smpl\"), exist_ok=True)\n            iio.imwrite(\n                os.path.join(outdir, \"color_smpl\", f\"{f_id:06d}.png\"),\n                color_smpl,\n            )\n            # os.makedirs(os.path.join(outdir, \"smpl\"), exist_ok=True)\n            # np.savez(\n            #     os.path.join(outdir, \"smpl\", f\"{f_id:06d}.npz\"),\n            #     scores=smpl_scores[f_id].numpy(),\n            #     msk=msks[f_id].numpy() if has_mask else None,\n            #     shape=smpl_shape[f_id].numpy(),\n            #     rotvec=smpl_rotvec[f_id].numpy(),\n            #     transl=smpl_transl[f_id].numpy(),\n            #     expression=smpl_expression[f_id].numpy() if smpl_expression[f_id] is not None else None\n            # )\n\n    if save_smpl and save_video:\n        frames_dir = os.path.join(outdir, \"color_smpl\")\n        video_path = os.path.join(outdir, \"output_video.mp4\")\n        output_fps = 30 // subsample\n        os.system(f'/usr/bin/ffmpeg -y -framerate {output_fps} -i \"{frames_dir}/%06d.png\" '\n                f'-vf \"scale=trunc(iw/2)*2:trunc(ih/2)*2\" '\n                f'-vcodec h264 -preset fast -profile:v baseline -pix_fmt yuv420p '\n                f'-movflags +faststart -b:v 5000k \"{video_path}\"')\n    \n    return (\n        pts3ds_other,\n        colors, \n        conf_other, \n        cam_dict, \n        all_verts, \n        smpl_faces,\n        smpl_id,\n        msks\n    )\n\ndef parse_seq_path(p):\n    if os.path.isdir(p):\n        img_paths = sorted(glob.glob(f\"{p}/*\"))\n        tmpdirname = None\n    else:\n        cap = cv2.VideoCapture(p)\n        if not cap.isOpened():\n            raise ValueError(f\"Error opening video file {p}\")\n        video_fps = cap.get(cv2.CAP_PROP_FPS)\n        total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))\n        if video_fps == 0:\n            cap.release()\n            raise ValueError(f\"Error: Video FPS is 0 for {p}\")\n        frame_interval = 1\n        frame_indices = list(range(0, total_frames, frame_interval))\n        print(\n            f\" - Video FPS: {video_fps}, Frame Interval: {frame_interval}, Total Frames to Read: {len(frame_indic\n# ... truncated ...","source_hash":"d7f1b65a79599b5081acf2f9a0b80bd6656913e66c74093a514651e32499b044","truncated":true}
{"repo_id":"Human3R","entity_id":"py:demo.parse_args","uri":"program://Human3R/function/demo.parse_args#L37-L141","kind":"function","name":"parse_args","path":"demo.py","language":"python","start_line":37,"end_line":141,"context_start_line":17,"context_end_line":161,"code":"\nimport os\nimport numpy as np\nimport torch\nimport time\nimport glob\nimport random\nimport cv2\nimport argparse\nimport tempfile\nimport shutil\nfrom copy import deepcopy\nfrom add_ckpt_path import add_path_to_dust3r\nimport imageio.v2 as iio\nimport roma\n\n# Set random seed for reproducibility.\nrandom.seed(42)\n\n\ndef parse_args():\n    \"\"\"Parse command-line arguments.\"\"\"\n    parser = argparse.ArgumentParser(\n        description=\"Run 3D point cloud inference and visualization using ARCroco3DStereo.\"\n    )\n    parser.add_argument(\n        \"--model_path\",\n        type=str,\n        default=\"src/cut3r_512_dpt_4_64.pth\",\n        help=\"Path to the pretrained model checkpoint.\",\n    )\n    parser.add_argument(\n        \"--seq_path\",\n        type=str,\n        default=\"\",\n        help=\"Path to the directory containing the image sequence.\",\n    )\n    parser.add_argument(\n        \"--device\",\n        type=str,\n        default=\"cuda\",\n        help=\"Device to run inference on (e.g., 'cuda' or 'cpu').\",\n    )\n    parser.add_argument(\n        \"--size\",\n        type=int,\n        default=\"512\",\n        help=\"Shape that input images will be rescaled to; if using 224+linear model, choose 224 otherwise 512\",\n    )\n    parser.add_argument(\n        \"--vis_threshold\",\n        type=float,\n        default=1.5,\n        help=\"Visualization threshold for the viewer. Ranging from 1 to INF\",\n    )\n    parser.add_argument(\n        \"--msk_threshold\",\n        type=float,\n        default=0.1,\n        help=\"Mask threshold. Ranging from 0 to 1\",\n    )\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"./tmp\",\n        help=\"value for tempfile.tempdir\",\n    )\n    parser.add_argument(\n        \"--save_smpl\",\n        action=\"store_true\",\n        help=\"Save smpl results.\",\n    )\n    parser.add_argument(\n        \"--save_video\",\n        action=\"store_true\",\n        help=\"Save smpl video.\",\n    )\n    parser.add_argument(\n        \"--max_frames\",\n        type=int,\n        default=None,\n        help=\"Max frames to use. Default is None (use all images).\",\n    )\n    parser.add_argument(\n        \"--subsample\",\n        type=int,\n        default=1,\n        help=\"Subsample factor for input images. Default is 1 (use all images).\",\n    )\n    parser.add_argument(\n        \"--reset_interval\", \n        type=int, \n        default=10000000\n        )\n    parser.add_argument(\n        \"--use_ttt3r\",\n        action=\"store_true\",\n        help=\"Use TTT3R.\",\n        default=False\n    )\n    parser.add_argument(\n        \"--downsample_factor\",\n        type=int,\n        default=10,\n        help=\"Point cloud downsample factor for the viewer\",\n    )\n    parser.add_argument(\n        \"--smpl_downsample\",\n        type=int,\n        default=1,\n        help=\"SMPL sequence downsample factor for the viewer\",\n    )\n    parser.add_argument(\n        \"--camera_downsample\",\n        type=int,\n        default=1,\n        help=\"Camera motion downsample factor for the viewer\",\n    )\n    parser.add_argument(\n        \"--mask_morph\",\n        type=int,\n        default=10,\n        help=\"Mask morphology for the viewer\",\n    )\n    return parser.parse_args()\n\n\ndef prepare_input(\n    img_paths, \n    img_mask, \n    size, \n    raymaps=None, \n    raymap_mask=None, \n    revisit=1, \n    update=True, \n    img_res=None, \n    reset_interval=100\n):\n    \"\"\"\n    Prepare input views for inference from a list of image paths.\n\n    Args:\n        img_paths (list): List of image file paths.\n        img_mask (list of bool): Flags indicating valid images.\n        size (int): Target image size.","source_hash":"d7f1b65a79599b5081acf2f9a0b80bd6656913e66c74093a514651e32499b044","truncated":false}
{"repo_id":"Human3R","entity_id":"py:demo.prepare_input","uri":"program://Human3R/function/demo.prepare_input#L144-L273","kind":"function","name":"prepare_input","path":"demo.py","language":"python","start_line":144,"end_line":273,"context_start_line":124,"context_end_line":293,"code":"        \"--smpl_downsample\",\n        type=int,\n        default=1,\n        help=\"SMPL sequence downsample factor for the viewer\",\n    )\n    parser.add_argument(\n        \"--camera_downsample\",\n        type=int,\n        default=1,\n        help=\"Camera motion downsample factor for the viewer\",\n    )\n    parser.add_argument(\n        \"--mask_morph\",\n        type=int,\n        default=10,\n        help=\"Mask morphology for the viewer\",\n    )\n    return parser.parse_args()\n\n\ndef prepare_input(\n    img_paths, \n    img_mask, \n    size, \n    raymaps=None, \n    raymap_mask=None, \n    revisit=1, \n    update=True, \n    img_res=None, \n    reset_interval=100\n):\n    \"\"\"\n    Prepare input views for inference from a list of image paths.\n\n    Args:\n        img_paths (list): List of image file paths.\n        img_mask (list of bool): Flags indicating valid images.\n        size (int): Target image size.\n        raymaps (list, optional): List of ray maps.\n        raymap_mask (list, optional): Flags indicating valid ray maps.\n        revisit (int): How many times to revisit each view.\n        update (bool): Whether to update the state on revisits.\n\n    Returns:\n        list: A list of view dictionaries.\n    \"\"\"\n    # Import image loader (delayed import needed after adding ckpt path).\n    from src.dust3r.utils.image import load_images, pad_image\n    from dust3r.utils.geometry import get_camera_parameters\n\n    images = load_images(img_paths, size=size)\n    if img_res is not None:\n        K_mhmr = get_camera_parameters(img_res, device=\"cpu\") # if use pseudo K\n\n    views = []\n    if raymaps is None and raymap_mask is None:\n        # Only images are provided.\n        for i in range(len(images)):\n            view = {\n                \"img\": images[i][\"img\"],\n                \"ray_map\": torch.full(\n                    (\n                        images[i][\"img\"].shape[0],\n                        6,\n                        images[i][\"img\"].shape[-2],\n                        images[i][\"img\"].shape[-1],\n                    ),\n                    torch.nan,\n                ),\n                \"true_shape\": torch.from_numpy(images[i][\"true_shape\"]),\n                \"idx\": i,\n                \"instance\": str(i),\n                \"camera_pose\": torch.from_numpy(\n                    np.eye(4, dtype=np.float32)\n                    ).unsqueeze(0),\n                \"img_mask\": torch.tensor(True).unsqueeze(0),\n                \"ray_mask\": torch.tensor(False).unsqueeze(0),\n                \"update\": torch.tensor(True).unsqueeze(0),\n                \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n            }\n            if img_res is not None:\n                view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                view[\"K_mhmr\"] = K_mhmr\n            views.append(view)\n            if (i+1) % reset_interval == 0:\n                overlap_view = deepcopy(view)\n                overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                views.append(overlap_view)\n    else:\n        # Combine images and raymaps.\n        num_views = len(images) + len(raymaps)\n        assert len(img_mask) == len(raymap_mask) == num_views\n        assert sum(img_mask) == len(images) and sum(raymap_mask) == len(raymaps)\n\n        j = 0\n        k = 0\n        for i in range(num_views):\n            view = {\n                \"img\": (\n                    images[j][\"img\"]\n                    if img_mask[i]\n                    else torch.full_like(images[0][\"img\"], torch.nan)\n                ),\n                \"ray_map\": (\n                    raymaps[k]\n                    if raymap_mask[i]\n                    else torch.full_like(raymaps[0], torch.nan)\n                ),\n                \"true_shape\": (\n                    torch.from_numpy(images[j][\"true_shape\"])\n                    if img_mask[i]\n                    else torch.from_numpy(np.int32([raymaps[k].shape[1:-1][::-1]]))\n                ),\n                \"idx\": i,\n                \"instance\": str(i),\n                \"camera_pose\": torch.from_numpy(\n                    np.eye(4, dtype=np.float32)\n                    ).unsqueeze(0),\n                \"img_mask\": torch.tensor(img_mask[i]).unsqueeze(0),\n                \"ray_mask\": torch.tensor(raymap_mask[i]).unsqueeze(0),\n                \"update\": torch.tensor(img_mask[i]).unsqueeze(0),\n                \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n            }\n            if img_res is not None:\n                view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                view[\"K_mhmr\"] = K_mhmr\n            if img_mask[i]:\n                j += 1\n            if raymap_mask[i]:\n                k += 1\n            views.append(view)\n            if (i+1) % reset_interval == 0:\n                overlap_view = deepcopy(view)\n                overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                views.append(overlap_view)\n        assert j == len(images) and k == len(raymaps)\n\n    if revisit > 1:\n        new_views = []\n        for r in range(revisit):\n            for i, view in enumerate(views):\n                new_view = deepcopy(view)\n                new_view[\"idx\"] = r * len(views) + i\n                new_view[\"instance\"] = str(r * len(views) + i)\n                if r > 0 and not update:\n                    new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                new_views.append(new_view)\n        return new_views\n\n    return views\n\ndef prepare_output(\n        outputs, outdir, revisit=1, use_pose=True, \n        save_smpl=False, save_video=False, img_res=None, subsample=1):\n    \"\"\"\n    Process inference outputs to generate point clouds and camera parameters for visualization.\n\n    Args:\n        outputs (dict): Inference outputs.\n        revisit (int): Number of revisits per view.\n        use_pose (bool): Whether to transform points using camera pose.\n        save_smpl (bool): Whether to save smpl results.\n        save_video (bool): Whether to save smpl video.\n\n    Returns:\n        tuple: (points, colors, confidence, camera parameters dictionary)\n    \"\"\"\n    from src.dust3r.utils.camera import pose_encoding_to_camera\n    from src.dust3r.post_process import estimate_focal_knowing_depth\n    from src.dust3r.utils.geometry import geotrf, matrix_cumprod","source_hash":"d7f1b65a79599b5081acf2f9a0b80bd6656913e66c74093a514651e32499b044","truncated":false}
{"repo_id":"Human3R","entity_id":"py:demo.prepare_output","uri":"program://Human3R/function/demo.prepare_output#L275-L524","kind":"function","name":"prepare_output","path":"demo.py","language":"python","start_line":275,"end_line":524,"context_start_line":255,"context_end_line":544,"code":"            if (i+1) % reset_interval == 0:\n                overlap_view = deepcopy(view)\n                overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                views.append(overlap_view)\n        assert j == len(images) and k == len(raymaps)\n\n    if revisit > 1:\n        new_views = []\n        for r in range(revisit):\n            for i, view in enumerate(views):\n                new_view = deepcopy(view)\n                new_view[\"idx\"] = r * len(views) + i\n                new_view[\"instance\"] = str(r * len(views) + i)\n                if r > 0 and not update:\n                    new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                new_views.append(new_view)\n        return new_views\n\n    return views\n\ndef prepare_output(\n        outputs, outdir, revisit=1, use_pose=True, \n        save_smpl=False, save_video=False, img_res=None, subsample=1):\n    \"\"\"\n    Process inference outputs to generate point clouds and camera parameters for visualization.\n\n    Args:\n        outputs (dict): Inference outputs.\n        revisit (int): Number of revisits per view.\n        use_pose (bool): Whether to transform points using camera pose.\n        save_smpl (bool): Whether to save smpl results.\n        save_video (bool): Whether to save smpl video.\n\n    Returns:\n        tuple: (points, colors, confidence, camera parameters dictionary)\n    \"\"\"\n    from src.dust3r.utils.camera import pose_encoding_to_camera\n    from src.dust3r.post_process import estimate_focal_knowing_depth\n    from src.dust3r.utils.geometry import geotrf, matrix_cumprod\n    from src.dust3r.utils import SMPL_Layer, vis_heatmap, render_meshes\n    from src.dust3r.utils.image import unpad_image\n    from viser_utils import get_color\n\n    # Only keep the outputs corresponding to one full pass.\n    valid_length = len(outputs[\"pred\"]) // revisit\n    outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n    outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n    # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n    reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n    shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n    outputs[\"pred\"] = [\n        pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n    outputs[\"views\"] = [\n        view for view, mask in zip(outputs[\"views\"], shifted_reset_mask) if not mask]\n    reset_mask = reset_mask[~shifted_reset_mask]\n\n    pts3ds_self_ls = [output[\"pts3d_in_self_view\"] for output in outputs[\"pred\"]]\n    pts3ds_other = [output[\"pts3d_in_other_view\"] for output in outputs[\"pred\"]]\n    conf_self = [output[\"conf_self\"] for output in outputs[\"pred\"]]\n    conf_other = [output[\"conf\"] for output in outputs[\"pred\"]]\n    pts3ds_self = torch.cat(pts3ds_self_ls, 0)\n\n    # Recover camera poses.\n    pr_poses = [\n        pose_encoding_to_camera(pred[\"camera_pose\"].clone()).cpu()\n        for pred in outputs[\"pred\"]\n    ]\n\n    # reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n    if reset_mask.any():\n        pr_poses = torch.cat(pr_poses, 0)\n        identity = torch.eye(4, device=pr_poses.device)\n        reset_poses = torch.where(reset_mask.unsqueeze(-1).unsqueeze(-1), pr_poses, identity)\n        cumulative_bases = matrix_cumprod(reset_poses)\n        shifted_bases = torch.cat([identity.unsqueeze(0), cumulative_bases[:-1]], dim=0)\n        pr_poses = torch.einsum('bij,bjk->bik', shifted_bases, pr_poses)\n        # keeps only reset_mask=False pr_poses\n        pr_poses = list(pr_poses.unsqueeze(1).unbind(0))\n\n    R_c2w = torch.cat([pr_pose[:, :3, :3] for pr_pose in pr_poses], 0)\n    t_c2w = torch.cat([pr_pose[:, :3, 3] for pr_pose in pr_poses], 0)\n\n    if use_pose:\n        transformed_pts3ds_other = []\n        for pose, pself in zip(pr_poses, pts3ds_self):\n            transformed_pts3ds_other.append(geotrf(pose, pself.unsqueeze(0)))\n        pts3ds_other = transformed_pts3ds_other\n        conf_other = conf_self\n\n    # Estimate focal length based on depth.\n    B, H, W, _ = pts3ds_self.shape\n    pp = torch.tensor([W // 2, H // 2], device=pts3ds_self.device).float().repeat(B, 1)\n    focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n    colors = [\n        0.5 * (output[\"img\"].permute(0, 2, 3, 1) + 1.0) for output in outputs[\"views\"]\n    ]\n\n    cam_dict = {\n        \"focal\": focal.numpy(),\n        \"pp\": pp.numpy(),\n        \"R\": R_c2w.numpy(),\n        \"t\": t_c2w.numpy(),\n    }\n\n    pts3ds_self_tosave = pts3ds_self  # B, H, W, 3\n    depths_tosave = pts3ds_self_tosave[..., 2]\n    pts3ds_other_tosave = torch.cat(pts3ds_other)  # B, H, W, 3\n    conf_self_tosave = torch.cat(conf_self)  # B, H, W\n    conf_other_tosave = torch.cat(conf_other)  # B, H, W\n    colors_tosave = torch.cat(\n        [\n            0.5 * (output[\"img\"].permute(0, 2, 3, 1) + 1.0)\n            for output in outputs[\"views\"]\n        ]\n    )  # [B, H, W, 3]\n    cam2world_tosave = torch.cat(pr_poses)  # B, 4, 4\n    intrinsics_tosave = (\n        torch.eye(3).unsqueeze(0).repeat(cam2world_tosave.shape[0], 1, 1)\n    )  # B, 3, 3\n    intrinsics_tosave[:, 0, 0] = focal.detach()\n    intrinsics_tosave[:, 1, 1] = focal.detach()\n    intrinsics_tosave[:, 0, 2] = pp[:, 0]\n    intrinsics_tosave[:, 1, 2] = pp[:, 1]\n\n    # get SMPL parameters from outputs\n    smpl_shape = [output.get(\n        \"smpl_shape\", torch.empty(1,0,10))[0] for output in outputs[\"pred\"]]\n    smpl_rotvec = [roma.rotmat_to_rotvec(\n        output.get(\n            \"smpl_rotmat\", torch.empty(1,0,53,3,3))[0]) for output in outputs[\"pred\"]]\n    smpl_transl = [output.get(\n        \"smpl_transl\", torch.empty(1,0,3))[0] for output in outputs[\"pred\"]]\n    smpl_expression = [output.get(\n        \"smpl_expression\", [None])[0] for output in outputs[\"pred\"]]\n    smpl_id = [output.get(\n        \"smpl_id\", torch.empty(1,0))[0] for output in outputs[\"pred\"]]\n    # smpl_loc = [output.get(\n    #     \"smpl_loc\", torch.empty(1,0,2))[0] for output in outputs[\"pred\"]]\n    # K_mhmr = [output.get(\n    #     \"K_mhmr\", torch.empty(1,0,3))[0] for output in outputs[\"views\"]]\n        \n    if save_smpl:\n        smpl_scores = [\n            output[\"smpl_scores\"][...,0] for output in outputs[\"pred\"]]\n        if img_res is not None:\n            smpl_scores = [\n                unpad_image(s, [H, W])[0] for s in smpl_scores]\n\n    has_mask = \"msk\" in outputs[\"pred\"][0]\n    if has_mask:\n        msks = [output[\"msk\"][...,0] for output in outputs[\"pred\"]]\n        if img_res is not None:\n            msks = [unpad_image(m, [H, W]) for m in msks]\n    else:\n        msks = [torch.zeros(1, H, W) for _ in range(B)]\n\n    # SMPL layer\n    smpl_layer = SMPL_Layer(type='smplx', \n                            gender='neutral', \n                            num_betas=smpl_shape[0].shape[-1], \n                            kid=False, \n                            person_center='head')\n    smpl_faces = smpl_layer.bm_x.faces\n\n    # os.makedirs(os.path.join(outdir, \"depth\"), exist_ok=True)\n    # os.makedirs(os.path.join(outdir, \"conf\"), exist_ok=True)\n    # os.makedirs(os.path.join(outdir, \"color\"), exist_ok=True)\n    # os.makedirs(os.path.join(outdir, \"camera\"), exist_ok=True)\n\n    all_verts = []\n    for f_id in range(B):\n        n_humans_i = smpl_shape[f_id].shape[0]\n        \n        if n_humans_i > 0:\n            with torch.no_grad():\n                smpl_out = smpl_layer(\n                    smpl_rotvec[f_id], \n                    smpl_shape[f_id], \n                    smpl_transl[f_id], \n                    None, None, \n                    K=intrinsics_tosave[f_id].expand(n_humans_i, -1 , -1), \n                    expression=smpl_expression[f_id])\n        \n        depth = depths_tosave[f_id].numpy()\n        conf = conf_self_tosave[f_id].numpy()\n        color = colors_tosave[f_id].numpy()\n        c2w = cam2world_tosave[f_id].numpy()\n        intrins = intrinsics_tosave[f_id].numpy()\n\n        if n_humans_i > 0:\n            # transform smpl verts to world coordinates\n            all_verts.append(geotrf(pr_poses[f_id], smpl_out['smpl_v3d'].unsqueeze(0))[0])\n            pr_verts = [t.numpy() for t in smpl_out['smpl_v3d'].unbind(0)]\n            pr_faces = [smpl_faces] * n_humans_i\n        else:\n            pr_verts = []\n            pr_faces = []\n            all_verts.append(torch.empty(0))\n\n        if save_smpl:\n            hm = vis_heatmap(colors_tosave[f_id], smpl_scores[f_id]).numpy()\n            img_array_np = (color * 255).astype(np.uint8)\n            smpl_rend = render_meshes(img_array_np.copy(), pr_verts, pr_faces,\n                                        {'focal': intrins[[0,1],[0,1]], \n                                        'princpt': intrins[[0,1],[-1,-1]]},\n                                        color=[get_color(i)/255 for i in smpl_id[f_id]])\n            if has_mask:\n                msk_array_np = vis_heatmap(colors_tosave[f_id], msks[f_id][0]).numpy()\n                color_smpl = np.concatenate([\n                    img_array_np, \n                    (msk_array_np * 255).astype(np.uint8), \n                    (hm * 255).astype(np.uint8), \n                    smpl_rend], 1)\n            else:\n                color_smpl = np.concatenate([\n                    img_array_np, \n                    (hm * 255).astype(np.uint8), \n                    smpl_rend], 1)\n        \n        # np.save(os.path.join(outdir, \"depth\", f\"{f_id:06d}.npy\"), depth)\n        # np.save(os.path.join(outdir, \"conf\", f\"{f_id:06d}.npy\"), conf)\n        # iio.imwrite(\n        #     os.path.join(outdir, \"color\", f\"{f_id:06d}.png\"),\n        #     (color * 255).astype(np.uint8),\n        # )\n        # np.savez(\n        #     os.path.join(outdir, \"camera\", f\"{f_id:06d}.npz\"),\n        #     pose=c2w,\n        #     intrinsics=intrins,\n        # )\n\n        # Save smpl results\n        if save_smpl:\n            os.makedirs(os.path.join(outdir, \"color_smpl\"), exist_ok=True)\n            iio.imwrite(\n                os.path.join(outdir, \"color_smpl\", f\"{f_id:06d}.png\"),\n                color_smpl,\n            )\n            # os.makedirs(os.path.join(outdir, \"smpl\"), exist_ok=True)\n            # np.savez(\n            #     os.path.join(outdir, \"smpl\", f\"{f_id:06d}.npz\"),\n            #     scores=smpl_scores[f_id].numpy(),\n            #     msk=msks[f_id].numpy() if has_mask else None,\n            #     shape=smpl_shape[f_id].numpy(),\n            #     rotvec=smpl_rotvec[f_id].numpy(),\n            #     transl=smpl_transl[f_id].numpy(),\n            #     expression=smpl_expression[f_id].numpy() if smpl_expression[f_id] is not None else None\n            # )\n\n    if save_smpl and save_video:\n        frames_dir = os.path.join(outdir, \"color_smpl\")\n        video_path = os.path.join(outdir, \"output_video.mp4\")\n        output_fps = 30 // subsample\n        os.system(f'/usr/bin/ffmpeg -y -framerate {output_fps} -i \"{frames_dir}/%06d.png\" '\n                f'-vf \"scale=trunc(iw/2)*2:trunc(ih/2)*2\" '\n                f'-vcodec h264 -preset fast -profile:v baseline -pix_fmt yuv420p '\n                f'-movflags +faststart -b:v 5000k \"{video_path}\"')\n    \n    return (\n        pts3ds_other,\n        colors, \n        conf_other, \n        cam_dict, \n        all_verts, \n        smpl_faces,\n        smpl_id,\n        msks\n    )\n\ndef parse_seq_path(p):\n    if os.path.isdir(p):\n        img_paths = sorted(glob.glob(f\"{p}/*\"))\n        tmpdirname = None\n    else:\n        cap = cv2.VideoCapture(p)\n        if not cap.isOpened():\n            raise ValueError(f\"Error opening video file {p}\")\n        video_fps = cap.get(cv2.CAP_PROP_FPS)\n        total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))\n        if video_fps == 0:\n            cap.release()\n            raise ValueError(f\"Error: Video FPS is 0 for {p}\")\n        frame_interval = 1\n        frame_indices = list(range(0, total_frames, frame_interval))\n        print(\n            f\" - Video FPS: {video_fps}, Frame Interval: {frame_interval}, Total Frames to Read: {len(frame_indices)}\"\n        )\n        img_paths = []","source_hash":"d7f1b65a79599b5081acf2f9a0b80bd6656913e66c74093a514651e32499b044","truncated":false}
{"repo_id":"Human3R","entity_id":"py:demo.parse_seq_path","uri":"program://Human3R/function/demo.parse_seq_path#L526-L555","kind":"function","name":"parse_seq_path","path":"demo.py","language":"python","start_line":526,"end_line":555,"context_start_line":506,"context_end_line":575,"code":"    if save_smpl and save_video:\n        frames_dir = os.path.join(outdir, \"color_smpl\")\n        video_path = os.path.join(outdir, \"output_video.mp4\")\n        output_fps = 30 // subsample\n        os.system(f'/usr/bin/ffmpeg -y -framerate {output_fps} -i \"{frames_dir}/%06d.png\" '\n                f'-vf \"scale=trunc(iw/2)*2:trunc(ih/2)*2\" '\n                f'-vcodec h264 -preset fast -profile:v baseline -pix_fmt yuv420p '\n                f'-movflags +faststart -b:v 5000k \"{video_path}\"')\n    \n    return (\n        pts3ds_other,\n        colors, \n        conf_other, \n        cam_dict, \n        all_verts, \n        smpl_faces,\n        smpl_id,\n        msks\n    )\n\ndef parse_seq_path(p):\n    if os.path.isdir(p):\n        img_paths = sorted(glob.glob(f\"{p}/*\"))\n        tmpdirname = None\n    else:\n        cap = cv2.VideoCapture(p)\n        if not cap.isOpened():\n            raise ValueError(f\"Error opening video file {p}\")\n        video_fps = cap.get(cv2.CAP_PROP_FPS)\n        total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))\n        if video_fps == 0:\n            cap.release()\n            raise ValueError(f\"Error: Video FPS is 0 for {p}\")\n        frame_interval = 1\n        frame_indices = list(range(0, total_frames, frame_interval))\n        print(\n            f\" - Video FPS: {video_fps}, Frame Interval: {frame_interval}, Total Frames to Read: {len(frame_indices)}\"\n        )\n        img_paths = []\n        tmpdirname = tempfile.mkdtemp()\n        for i in frame_indices:\n            cap.set(cv2.CAP_PROP_POS_FRAMES, i)\n            ret, frame = cap.read()\n            if not ret:\n                break\n            frame_path = os.path.join(tmpdirname, f\"frame_{i}.jpg\")\n            cv2.imwrite(frame_path, frame)\n            img_paths.append(frame_path)\n        cap.release()\n    return img_paths, tmpdirname\n\n\ndef run_inference(args):\n    \"\"\"\n    Execute the full inference and visualization pipeline.\n\n    Args:\n        args: Parsed command-line arguments.\n    \"\"\"\n    # Set up the computation device.\n    device = args.device\n    if device == \"cuda\" and not torch.cuda.is_available():\n        print(\"CUDA not available. Switching to CPU.\")\n        device = \"cpu\"\n\n    # Add the checkpoint path (required for model imports in the dust3r package).\n    add_path_to_dust3r(args.model_path)\n\n    # Import model and inference functions after adding the ckpt path.\n    from src.dust3r.inference import inference_recurrent_lighter","source_hash":"d7f1b65a79599b5081acf2f9a0b80bd6656913e66c74093a514651e32499b044","truncated":false}
{"repo_id":"Human3R","entity_id":"py:demo.run_inference","uri":"program://Human3R/function/demo.run_inference#L558-L670","kind":"function","name":"run_inference","path":"demo.py","language":"python","start_line":558,"end_line":670,"context_start_line":538,"context_end_line":685,"code":"            raise ValueError(f\"Error: Video FPS is 0 for {p}\")\n        frame_interval = 1\n        frame_indices = list(range(0, total_frames, frame_interval))\n        print(\n            f\" - Video FPS: {video_fps}, Frame Interval: {frame_interval}, Total Frames to Read: {len(frame_indices)}\"\n        )\n        img_paths = []\n        tmpdirname = tempfile.mkdtemp()\n        for i in frame_indices:\n            cap.set(cv2.CAP_PROP_POS_FRAMES, i)\n            ret, frame = cap.read()\n            if not ret:\n                break\n            frame_path = os.path.join(tmpdirname, f\"frame_{i}.jpg\")\n            cv2.imwrite(frame_path, frame)\n            img_paths.append(frame_path)\n        cap.release()\n    return img_paths, tmpdirname\n\n\ndef run_inference(args):\n    \"\"\"\n    Execute the full inference and visualization pipeline.\n\n    Args:\n        args: Parsed command-line arguments.\n    \"\"\"\n    # Set up the computation device.\n    device = args.device\n    if device == \"cuda\" and not torch.cuda.is_available():\n        print(\"CUDA not available. Switching to CPU.\")\n        device = \"cpu\"\n\n    # Add the checkpoint path (required for model imports in the dust3r package).\n    add_path_to_dust3r(args.model_path)\n\n    # Import model and inference functions after adding the ckpt path.\n    from src.dust3r.inference import inference_recurrent_lighter\n    from src.dust3r.model import ARCroco3DStereo\n    from viser_utils import SceneHumanViewer\n\n    # Prepare image file paths.\n    img_paths, tmpdirname = parse_seq_path(args.seq_path)\n    if not img_paths:\n        print(f\"No images found in {args.seq_path}. Please verify the path.\")\n        return\n    \n    if args.max_frames is not None:\n        img_paths = img_paths[:args.max_frames]\n    img_paths = img_paths[::args.subsample]\n\n    print(f\"Found {len(img_paths)} images in {args.seq_path}.\")\n    img_mask = [True] * len(img_paths)\n\n    # Load and prepare the model.\n    print(f\"Loading model from {args.model_path}...\")\n    model = ARCroco3DStereo.from_pretrained(args.model_path).to(device)\n    model.eval()\n\n    # Prepare input views.\n    print(\"Preparing input views...\")\n    img_res = getattr(model, 'mhmr_img_res', None)\n    views = prepare_input(\n        img_paths=img_paths,\n        img_mask=img_mask,\n        size=args.size,\n        revisit=1,\n        update=True,\n        img_res=img_res,\n        reset_interval=args.reset_interval\n    )\n\n    if tmpdirname is not None:\n        shutil.rmtree(tmpdirname)\n\n    # Run inference.\n    print(\"Running inference...\")\n    start_time = time.time()\n    outputs, _ = inference_recurrent_lighter(\n        views, model, device, use_ttt3r=args.use_ttt3r)\n    total_time = time.time() - start_time\n    per_frame_time = total_time / len(views)\n    print(\n        f\"Inference completed in {total_time:.2f} seconds (average {per_frame_time:.2f} s per frame).\"\n    )\n\n    # Process outputs for visualization.\n    print(\"Preparing output for visualization...\")\n    (\n        pts3ds_other, \n        colors, \n        conf, \n        cam_dict, \n        all_smpl_verts, \n        smpl_faces,\n        smpl_id,\n        msks,\n        ) = prepare_output(\n        outputs, args.output_dir, 1, True, \n        args.save_smpl, args.save_video, img_res, args.subsample\n    )\n\n    # Convert tensors to numpy arrays for visualization.\n    pts3ds_to_vis = [p.cpu().numpy() for p in pts3ds_other]\n    colors_to_vis = [c.cpu().numpy() for c in colors]\n    msks_to_vis = [m.cpu().numpy() for m in msks]\n    conf_to_vis = [c.cpu().numpy() for c in conf]\n    edge_colors = [None] * len(pts3ds_to_vis)\n    verts_to_vis = [p.cpu().numpy() for p in all_smpl_verts]\n\n    # Create and run the point cloud viewer.\n    print(\"Launching point cloud viewer...\")\n    viewer = SceneHumanViewer(\n        pts3ds_to_vis,\n        colors_to_vis,\n        conf_to_vis,\n        cam_dict,\n        verts_to_vis,\n        smpl_faces,\n        smpl_id,\n        msks_to_vis,\n        device=device,\n        edge_color_list=edge_colors,\n        show_camera=True,\n        vis_threshold=args.vis_threshold,\n        msk_threshold=args.msk_threshold,\n        mask_morph=args.mask_morph,\n        size = args.size,\n        downsample_factor=args.downsample_factor,\n        smpl_downsample_factor=args.smpl_downsample,\n        camera_downsample_factor=args.camera_downsample\n    )\n    viewer.run()\n\n\ndef main():\n    args = parse_args()\n    if not args.seq_path:\n        print(\n            \"No inputs found! Please use our gradio demo if you would like to iteractively upload inputs.\"\n        )\n        return\n    else:\n        run_inference(args)\n\n\nif __name__ == \"__main__\":\n    main()","source_hash":"d7f1b65a79599b5081acf2f9a0b80bd6656913e66c74093a514651e32499b044","truncated":false}
{"repo_id":"Human3R","entity_id":"py:demo.main","uri":"program://Human3R/function/demo.main#L673-L681","kind":"function","name":"main","path":"demo.py","language":"python","start_line":673,"end_line":681,"context_start_line":653,"context_end_line":685,"code":"        conf_to_vis,\n        cam_dict,\n        verts_to_vis,\n        smpl_faces,\n        smpl_id,\n        msks_to_vis,\n        device=device,\n        edge_color_list=edge_colors,\n        show_camera=True,\n        vis_threshold=args.vis_threshold,\n        msk_threshold=args.msk_threshold,\n        mask_morph=args.mask_morph,\n        size = args.size,\n        downsample_factor=args.downsample_factor,\n        smpl_downsample_factor=args.smpl_downsample,\n        camera_downsample_factor=args.camera_downsample\n    )\n    viewer.run()\n\n\ndef main():\n    args = parse_args()\n    if not args.seq_path:\n        print(\n            \"No inputs found! Please use our gradio demo if you would like to iteractively upload inputs.\"\n        )\n        return\n    else:\n        run_inference(args)\n\n\nif __name__ == \"__main__\":\n    main()","source_hash":"d7f1b65a79599b5081acf2f9a0b80bd6656913e66c74093a514651e32499b044","truncated":false}
{"repo_id":"Human3R","entity_id":"py:add_ckpt_path","uri":"program://Human3R/module/add_ckpt_path#L1-L9","kind":"module","name":"add_ckpt_path","path":"add_ckpt_path.py","language":"python","start_line":1,"end_line":9,"context_start_line":1,"context_end_line":9,"code":"import sys\nimport os\nimport os.path as path\n\n\ndef add_path_to_dust3r(ckpt):\n    HERE_PATH = os.path.dirname(os.path.abspath(ckpt))\n    # workaround for sibling import\n    sys.path.insert(0, HERE_PATH)","source_hash":"40c45aec68241a855e4a09e6a7c95aeadb4ff6fc5fa40568ee98b586b5dc6d5a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:add_ckpt_path.add_path_to_dust3r","uri":"program://Human3R/function/add_ckpt_path.add_path_to_dust3r#L6-L9","kind":"function","name":"add_path_to_dust3r","path":"add_ckpt_path.py","language":"python","start_line":6,"end_line":9,"context_start_line":1,"context_end_line":9,"code":"import sys\nimport os\nimport os.path as path\n\n\ndef add_path_to_dust3r(ckpt):\n    HERE_PATH = os.path.dirname(os.path.abspath(ckpt))\n    # workaround for sibling import\n    sys.path.insert(0, HERE_PATH)","source_hash":"40c45aec68241a855e4a09e6a7c95aeadb4ff6fc5fa40568ee98b586b5dc6d5a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils","uri":"program://Human3R/module/viser_utils#L1-L1392","kind":"module","name":"viser_utils","path":"viser_utils.py","language":"python","start_line":1,"end_line":1392,"context_start_line":1,"context_end_line":1392,"code":"import torch\nimport os\nfrom matplotlib.figure import Figure\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nimport matplotlib as mpl\nimport cv2\nimport numpy as np\nimport matplotlib.cm as cm\nimport viser\nimport viser.transforms as tf\nimport time\nimport trimesh\nimport dataclasses\nfrom scipy.spatial.transform import Rotation\nfrom skimage.morphology import binary_dilation, binary_erosion, disk\nfrom src.dust3r.viz import (\n    add_scene_cam,\n    CAM_COLORS,\n    OPENGL,\n    pts3d_to_trimesh,\n    cat_meshes,\n)\n\n\ndef todevice(batch, device, callback=None, non_blocking=False):\n    \"\"\"Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy).\n\n    batch: list, tuple, dict of tensors or other things\n    device: pytorch device or 'numpy'\n    callback: function that would be called on every sub-elements.\n    \"\"\"\n    if callback:\n        batch = callback(batch)\n\n    if isinstance(batch, dict):\n        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef segment_sky(image):\n    import cv2\n    from scipy import ndimage\n\n    # Convert to HSV\n    image = to_numpy(image)\n    if np.issubdtype(image.dtype, np.floating):\n        image = np.uint8(255 * image.clip(min=0, max=1))\n    hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\n\n    # Define range for blue color and create mask\n    lower_blue = np.array([0, 0, 100])\n    upper_blue = np.array([30, 255, 255])\n    mask = cv2.inRange(hsv, lower_blue, upper_blue).view(bool)\n\n    # add luminous gray\n    mask |= (hsv[:, :, 1] < 10) & (hsv[:, :, 2] > 150)\n    mask |= (hsv[:, :, 1] < 30) & (hsv[:, :, 2] > 180)\n    mask |= (hsv[:, :, 1] < 50) & (hsv[:, :, 2] > 220)\n\n    # Morphological operations\n    kernel = np.ones((5, 5), np.uint8)\n    mask2 = ndimage.binary_opening(mask, structure=kernel)\n\n    # keep only largest CC\n    _, labels, stats, _ = cv2.connectedComponentsWithStats(\n        mask2.view(np.uint8), connectivity=8\n    )\n    cc_sizes = stats[1:, cv2.CC_STAT_AREA]\n    order = cc_sizes.argsort()[::-1]  # bigger first\n    i = 0\n    selection = []\n    while i < len(order) and cc_sizes[order[i]] > cc_sizes[order[0]] / 2:\n        selection.append(1 + order[i])\n        i += 1\n    mask3 = np.in1d(labels, selection).reshape(labels.shape)\n\n    # Apply mask\n    return torch.from_numpy(mask3)\n\n\ndef convert_scene_output_to_glb(\n    outdir,\n    imgs,\n    pts3d,\n    mask,\n    focals,\n    cams2world,\n    cam_size=0.05,\n    show_cam=True,\n    cam_color=None,\n    as_pointcloud=False,\n    transparent_cams=False,\n    silent=False,\n    save_name=None,\n):\n    assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)\n    pts3d = to_numpy(pts3d)\n    imgs = to_numpy(imgs)\n    focals = to_numpy(focals)\n    cams2world = to_numpy(cams2world)\n\n    scene = trimesh.Scene()\n\n    # full pointcloud\n    if as_pointcloud:\n        pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])\n        col = np.concatenate([p[m] for p, m in zip(imgs, mask)])\n        pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3))\n        scene.add_geometry(pct)\n    else:\n        meshes = []\n        for i in range(len(imgs)):\n            meshes.append(pts3d_to_trimesh(imgs[i], pts3d[i], mask[i]))\n        mesh = trimesh.Trimesh(**cat_meshes(meshes))\n        scene.add_geometry(mesh)\n\n    # add each camera\n    if show_cam:\n        for i, pose_c2w in enumerate(cams2world):\n            if isinstance(cam_color, list):\n                camera_edge_color = cam_color[i]\n            else:\n                camera_edge_color = cam_color or CAM_COLORS[i % len(CAM_COLORS)]\n            add_scene_cam(\n                scene,\n                pose_c2w,\n                camera_edge_color,\n                None if transparent_cams else imgs[i],\n                focals[i],\n                imsize=imgs[i].shape[1::-1],\n                screen_width=cam_size,\n            )\n\n    rot = np.eye(4)\n    rot[:3, :3] = Rotation.from_euler(\"y\", np.deg2rad(180)).as_matrix()\n    scene.apply_transform(np.linalg.inv(cams2world[0] @ OPENGL @ rot))\n    if save_name is None:\n        save_name = \"scene\"\n    outfile = os.path.join(outdir, save_name + \".glb\")\n    if not silent:\n        print(\"(exporting 3D scene to\", outfile, \")\")\n    scene.export(file_obj=outfile)\n    return outfile\n\n\n@dataclasses.dataclass\nclass CameraState(object):\n    fov: float\n    aspect: float\n    c2w: np.ndarray\n\n    def get_K(self, img_wh):\n        W, H = img_wh\n        focal_length = H / 2.0 / np.tan(self.fov / 2.0)\n        K = np.array(\n            [\n                [focal_length, 0.0, W / 2.0],\n                [0.0, focal_length, H / 2.0],\n                [0.0, 0.0, 1.0],\n            ]\n        )\n        return K\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n\n    tick_cnt = 6\n    tick_loc = np.linspace(vmin, vmax, tick_cnt)\n    cb1 = mpl.colorbar.ColorbarBase(\n        ax, cmap=cmap, norm=norm, ticks=tick_loc, orientation=\"vertical\"\n    )\n\n    tick_label = [str(np.round(x, cbar_precision)) for x in tick_loc]\n    if cbar_precision == 0:\n        tick_label = [x[:-2] for x in tick_label]\n\n    cb1.set_ticklabels(tick_label)\n\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]\n    \"\"\"\n    if range is not None:\n        vmin, vmax = range\n    elif mask is not None:\n\n        vmin = np.min(x[mask][np.nonzero(x[mask])])\n        vmax = np.max(x[mask])\n\n        x[np.logical_not(mask)] = vmin\n\n    else:\n        vmin, vmax = np.percentile(x, (1, 100))\n        vmax += 1e-6\n\n    x = np.clip(x, vmin, vmax)\n    x = (x - vmin) / (vmax - vmin)\n\n    cmap = cm.get_cmap(cmap_name)\n    x_new = cmap(x)[:, :, :3]\n\n    if mask is not None:\n        mask = np.float32(mask[:, :, np.newaxis])\n        x_new = x_new * mask + np.ones_like(x_new) * (1.0 - mask)\n\n    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out\n\ndef get_color(idx):\n    root_dir = os.path.dirname(os.path.abspath(__file__))\n    colors_path = os.path.join(root_dir, \"src/models/smpl_colors.txt\")\n    colors = np.loadtxt(colors_path).astype(int)\n    return colors[idx % len(colors)]\n\nclass SceneHumanViewer:\n    def __init__(\n        self,\n        pc_list,\n        color_list,\n        conf_list,\n        cam_dict,\n        all_smpl_verts,\n        smpl_faces,\n        smpl_id,\n        msk_list,\n        gt_cam_dict=None,\n        gt_smpl_verts=None,\n        image_mask=None,\n        edge_color_list=None,\n        device=\"cpu\",\n        port=8080,\n        show_camera=True,\n        show_gt_camera=False,\n        show_gt_smpl=False,\n        vis_threshold=1,\n        msk_threshold=0.1,\n        mask_morph=0,\n        size=512,\n        downsample_factor=10,\n        smpl_downsample_factor=1,\n        camera_downsample_factor=1,\n    ):\n        self.size=size\n        self.server = viser.ViserServer(port=port)\n        self.server.set_up_direction(\"-y\")\n        self.device = device\n        self.conf_list = conf_list\n        self.msk_list = msk_list\n        self.vis_threshold = vis_threshold\n        self.msk_threshold = msk_threshold\n        self.mask_morph = mask_morph\n        self.show_background = True\n        self.show_foreground = False\n        self.tt = lambda x: torch.from_numpy(x).float().to(device)\n        self.pcs, self.all_steps = self.read_data(\n            pc_list, color_list, conf_list, msk_list, \n            all_smpl_verts, smpl_faces, smpl_id, edge_color_list, gt_smpl_verts\n        )\n        # Fast lookup from step id to its sequential index\n        self.step_to_index = {step: idx for idx, step in enumerate(self.all_steps)}\n        self.cam_dict = cam_dict\n        self.gt_cam_dict = gt_cam_dict\n        self.gt_smpl_verts = gt_smpl_verts\n        self.num_frames = len(self.all_steps)\n        self.image_mask = image_mask\n        self.show_camera = show_camera\n        self.show_gt_camera = show_gt_camera and gt_cam_dict is not None\n        self.show_gt_smpl = show_gt_smpl and gt_smpl_verts is not None\n        self.on_replay = False\n        self.vis_pts_list = []\n        self.traj_list = []\n        self.orig_img_list = [x[0] for x in color_list]\n        self.via_points = []\n        self._updating_point_clouds = False\n        \n        # Performance optimization for dynamic opacity\n        self._last_opacity_update_step = -1\n        self._opacity_update_throttle = 0  # Frames to skip between updates\n        self._opacity_frame_counter = 0\n\n        gui_reset_up = self.server.gui.add_button(\n            \"Reset up direction\",\n            hint=\"Set the camera control 'up' direction to the current camera's 'up'.\",\n        )\n\n        @gui_reset_up.on_click\n        def _(event: viser.GuiEvent) -> None:\n            client = event.client\n            assert client is not None\n            client.camera.up_direction = tf.SO3(client.camera.wxyz) @ np.array(\n                [0.0, -1.0, 0.0]\n            )\n\n        button3 = self.server.gui.add_button(\"4D (Only Show Current Frame)\")\n        button4 = self.server.gui.add_button(\"3D (Show All Frames)\")\n        button5 = self.server.gui.add_button(\"Hybrid (Current SMPL + All Points)\")\n        self.is_render = False\n        self.fourd = False\n        self.hybrid_mode = False\n\n        @button3.on_click\n        def _(event: viser.GuiEvent) -> None:\n            self.fourd = True\n            self.hybrid_mode = False\n\n        @button4.on_click\n        def _(event: viser.GuiEvent) -> None:\n            self.fourd = False\n            self.hybrid_mode = False\n\n        @button5.on_click\n        def _(event: viser.GuiEvent) -> None:\n            self.fourd = False\n            self.hybrid_mode = True\n\n        self.gui_show_background = self.server.add_gui_checkbox(\n            \"Show Background\", True)\n\n        @self.gui_show_background.on_update\n        def _(_) -> None:\n            self.show_background = self.gui_show_background.value\n            self._update_point_clouds()\n\n        self.gui_show_foreground = self.server.add_gui_checkbox(\n            \"Show Foreground\", False)\n\n        @self.gui_show_foreground.on_update\n        def _(_) -> None:\n            self.show_foreground = self.gui_show_foreground.value\n            self._update_point_clouds()\n\n        self.gui_show_smpl = self.server.add_gui_checkbox(\"Show SMPL\", True)\n\n        @self.gui_show_smpl.on_update\n        def _(_) -> None:\n            # Update SMPL visibility considering downsampling factor\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'frame_nodes'):\n                self._update_smpl_visibility()\n\n        self.focal_slider = self.server.add_gui_slider(\n            \"Focal Length\",\n            min=0.1,\n            max=99999,\n            step=1,\n            initial_value=533,\n        )\n\n        self.psize_slider = self.server.add_gui_slider(\n            \"Point Size\",\n            min=0.0001,\n            max=0.1,\n            step=0.0001,\n            initial_value=0.005,\n        )\n        self.camsize_slider = self.server.add_gui_slider(\n            \"Camera Size\",\n            min=0.01,\n            max=0.5,\n            step=0.01,\n            initial_value=0.1,\n        )\n\n        self.downsample_slider = self.server.add_gui_slider(\n            \"Downsample Factor\",\n            min=1,\n            max=1000,\n            step=1,\n            initial_value=downsample_factor,\n        )\n        self.vis_threshold_slider = self.server.add_gui_slider(\n            \"Visibility Threshold\",\n            min=0.1,\n            max=50.0,\n            step=0.1,\n            initial_value=self.vis_threshold,\n        )\n        \n        self.mask_morph_slider = self.server.add_gui_slider(\n            \"Mask Morphology\",\n            min=-20.0,\n            max=100.0,\n            step=0.1,\n            initial_value=self.mask_morph,\n        )\n\n        # SMPL downsampling controls\n        self.smpl_downsample_slider = self.server.add_gui_slider(\n            \"SMPL Downsample\",\n            min=1,\n            max=200,\n            step=1,\n            initial_value=smpl_downsample_factor,\n        )\n        \n        # Camera downsampling controls\n        self.camera_downsample_slider = self.server.add_gui_slider(\n            \"Camera Downsample\",\n            min=1,\n            max=200,\n            step=1,\n            initial_value=camera_downsample_factor,\n        )\n        \n        @self.camera_downsample_slider.on_update\n        def _(_) -> None:\n            # Apply camera downsampling changes immediately\n            if hasattr(self, 'cam_handles'):\n                self._update_camera_visibility()\n            if hasattr(self, 'gt_cam_handles'):\n                self._update_gt_camera_visibility()\n\n        # Mesh opacity by time controls\n        self.mesh_time_opacity_checkbox = self.server.add_gui_checkbox(\n            \"Mesh Opacity by Time\", False\n        )\n        self.min_mesh_opacity_slider = self.server.add_gui_slider(\n            \"Min Mesh Opacity\",\n            min=0.0,\n            max=1.0,\n            step=0.05,\n            initial_value=0.1,\n        )\n\n        @self.mesh_time_opacity_checkbox.on_update\n        def _(_) -> None:\n            # Apply opacity changes to existing meshes immediately\n            if hasattr(self, 'mesh_handles'):\n                # If dynamic opacity is enabled, prefer dynamic update\n                if self.dynamic_opacity_checkbox.value and hasattr(self, 'current_step_index'):\n                    self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                else:\n                    self._update_mesh_opacities()\n\n        @self.min_mesh_opacity_slider.on_update\n        def _(_) -> None:\n            # Apply opacity changes to existing meshes immediately\n            if hasattr(self, 'mesh_handles'):\n                # If dynamic opacity is enabled, prefer dynamic update so min opacity affects decay baseline\n                if self.dynamic_opacity_checkbox.value and hasattr(self, 'current_step_index'):\n                    self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                else:\n                    self._update_mesh_opacities()\n\n        # Dynamic opacity controls\n        self.dynamic_opacity_checkbox = self.server.add_gui_checkbox(\n            \"Dynamic Opacity\", False\n        )\n        self.opacity_decay_len_slider = self.server.add_gui_slider(\n            \"Decay Length\",\n            min=1,\n            max=max(10, len(self.all_steps)),\n            step=1,\n            initial_value=min(10, max(1, len(self.all_steps))),\n        )\n        \n        # Performance control for dynamic opacity\n        self.opacity_update_throttle_slider = self.server.add_gui_slider(\n            \"Opacity Update Throttle\",\n            min=0,\n            max=10,\n            step=1,\n            initial_value=0,\n            hint=\"Skip N frames between opacity updates (0=update every frame, higher=better performance)\"\n        )\n        \n        # Performance preset buttons\n        self.performance_preset_buttons = self.server.add_gui_button_group(\n            \"Performance Presets\", \n            (\"High Quality\", \"Balanced\")\n        )\n\n        @self.dynamic_opacity_checkbox.on_update\n        def _(_) -> None:\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'current_step_index'):\n                if self.dynamic_opacity_checkbox.value:\n                    # Turning on dynamic opacity: update immediately\n                    self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                else:\n                    # Turning off dynamic opacity: fall back to static-by-time if enabled,\n                    # otherwise set all to fully opaque\n                    if self.mesh_time_opacity_checkbox.value:\n                        self._update_mesh_opacities()\n                    else:\n                        for handle in self.mesh_handles:\n                            handle.opacity = 1.0\n                        for handle in self.gt_mesh_handles:\n                            handle.opacity = 1.0\n\n        @self.opacity_decay_len_slider.on_update\n        def _(_) -> None:\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'current_step_index'):\n                self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                \n        @self.opacity_update_throttle_slider.on_update\n        def _(_) -> None:\n            self._opacity_update_throttle = int(self.opacity_update_throttle_slider.value)\n            self._opacity_frame_counter = 0  # Reset counter when throttle changes\n            # Force update to apply changes immediately\n            if hasattr(self, 'current_step_index') and self.dynamic_opacity_checkbox.value:\n                self._update_dynamic_opacities(self.current_step_index, force_update=True)\n        \n        @self.performance_preset_buttons.on_click\n        def _(_) -> None:\n            preset = self.performance_\n# ... truncated ...","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":true}
{"repo_id":"Human3R","entity_id":"py:viser_utils.todevice","uri":"program://Human3R/function/viser_utils.todevice#L26-L51","kind":"function","name":"todevice","path":"viser_utils.py","language":"python","start_line":26,"end_line":51,"context_start_line":6,"context_end_line":71,"code":"import matplotlib as mpl\nimport cv2\nimport numpy as np\nimport matplotlib.cm as cm\nimport viser\nimport viser.transforms as tf\nimport time\nimport trimesh\nimport dataclasses\nfrom scipy.spatial.transform import Rotation\nfrom skimage.morphology import binary_dilation, binary_erosion, disk\nfrom src.dust3r.viz import (\n    add_scene_cam,\n    CAM_COLORS,\n    OPENGL,\n    pts3d_to_trimesh,\n    cat_meshes,\n)\n\n\ndef todevice(batch, device, callback=None, non_blocking=False):\n    \"\"\"Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy).\n\n    batch: list, tuple, dict of tensors or other things\n    device: pytorch device or 'numpy'\n    callback: function that would be called on every sub-elements.\n    \"\"\"\n    if callback:\n        batch = callback(batch)\n\n    if isinstance(batch, dict):\n        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef segment_sky(image):\n    import cv2\n    from scipy import ndimage\n\n    # Convert to HSV\n    image = to_numpy(image)\n    if np.issubdtype(image.dtype, np.floating):\n        image = np.uint8(255 * image.clip(min=0, max=1))\n    hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\n\n    # Define range for blue color and create mask","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.to_numpy","uri":"program://Human3R/function/viser_utils.to_numpy#L57-L58","kind":"function","name":"to_numpy","path":"viser_utils.py","language":"python","start_line":57,"end_line":58,"context_start_line":37,"context_end_line":78,"code":"        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef segment_sky(image):\n    import cv2\n    from scipy import ndimage\n\n    # Convert to HSV\n    image = to_numpy(image)\n    if np.issubdtype(image.dtype, np.floating):\n        image = np.uint8(255 * image.clip(min=0, max=1))\n    hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\n\n    # Define range for blue color and create mask\n    lower_blue = np.array([0, 0, 100])\n    upper_blue = np.array([30, 255, 255])\n    mask = cv2.inRange(hsv, lower_blue, upper_blue).view(bool)\n\n    # add luminous gray\n    mask |= (hsv[:, :, 1] < 10) & (hsv[:, :, 2] > 150)\n    mask |= (hsv[:, :, 1] < 30) & (hsv[:, :, 2] > 180)","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.segment_sky","uri":"program://Human3R/function/viser_utils.segment_sky#L61-L99","kind":"function","name":"segment_sky","path":"viser_utils.py","language":"python","start_line":61,"end_line":99,"context_start_line":41,"context_end_line":119,"code":"\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef segment_sky(image):\n    import cv2\n    from scipy import ndimage\n\n    # Convert to HSV\n    image = to_numpy(image)\n    if np.issubdtype(image.dtype, np.floating):\n        image = np.uint8(255 * image.clip(min=0, max=1))\n    hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\n\n    # Define range for blue color and create mask\n    lower_blue = np.array([0, 0, 100])\n    upper_blue = np.array([30, 255, 255])\n    mask = cv2.inRange(hsv, lower_blue, upper_blue).view(bool)\n\n    # add luminous gray\n    mask |= (hsv[:, :, 1] < 10) & (hsv[:, :, 2] > 150)\n    mask |= (hsv[:, :, 1] < 30) & (hsv[:, :, 2] > 180)\n    mask |= (hsv[:, :, 1] < 50) & (hsv[:, :, 2] > 220)\n\n    # Morphological operations\n    kernel = np.ones((5, 5), np.uint8)\n    mask2 = ndimage.binary_opening(mask, structure=kernel)\n\n    # keep only largest CC\n    _, labels, stats, _ = cv2.connectedComponentsWithStats(\n        mask2.view(np.uint8), connectivity=8\n    )\n    cc_sizes = stats[1:, cv2.CC_STAT_AREA]\n    order = cc_sizes.argsort()[::-1]  # bigger first\n    i = 0\n    selection = []\n    while i < len(order) and cc_sizes[order[i]] > cc_sizes[order[0]] / 2:\n        selection.append(1 + order[i])\n        i += 1\n    mask3 = np.in1d(labels, selection).reshape(labels.shape)\n\n    # Apply mask\n    return torch.from_numpy(mask3)\n\n\ndef convert_scene_output_to_glb(\n    outdir,\n    imgs,\n    pts3d,\n    mask,\n    focals,\n    cams2world,\n    cam_size=0.05,\n    show_cam=True,\n    cam_color=None,\n    as_pointcloud=False,\n    transparent_cams=False,\n    silent=False,\n    save_name=None,\n):\n    assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)\n    pts3d = to_numpy(pts3d)\n    imgs = to_numpy(imgs)","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.convert_scene_output_to_glb","uri":"program://Human3R/function/viser_utils.convert_scene_output_to_glb#L102-L164","kind":"function","name":"convert_scene_output_to_glb","path":"viser_utils.py","language":"python","start_line":102,"end_line":164,"context_start_line":82,"context_end_line":184,"code":"    kernel = np.ones((5, 5), np.uint8)\n    mask2 = ndimage.binary_opening(mask, structure=kernel)\n\n    # keep only largest CC\n    _, labels, stats, _ = cv2.connectedComponentsWithStats(\n        mask2.view(np.uint8), connectivity=8\n    )\n    cc_sizes = stats[1:, cv2.CC_STAT_AREA]\n    order = cc_sizes.argsort()[::-1]  # bigger first\n    i = 0\n    selection = []\n    while i < len(order) and cc_sizes[order[i]] > cc_sizes[order[0]] / 2:\n        selection.append(1 + order[i])\n        i += 1\n    mask3 = np.in1d(labels, selection).reshape(labels.shape)\n\n    # Apply mask\n    return torch.from_numpy(mask3)\n\n\ndef convert_scene_output_to_glb(\n    outdir,\n    imgs,\n    pts3d,\n    mask,\n    focals,\n    cams2world,\n    cam_size=0.05,\n    show_cam=True,\n    cam_color=None,\n    as_pointcloud=False,\n    transparent_cams=False,\n    silent=False,\n    save_name=None,\n):\n    assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)\n    pts3d = to_numpy(pts3d)\n    imgs = to_numpy(imgs)\n    focals = to_numpy(focals)\n    cams2world = to_numpy(cams2world)\n\n    scene = trimesh.Scene()\n\n    # full pointcloud\n    if as_pointcloud:\n        pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])\n        col = np.concatenate([p[m] for p, m in zip(imgs, mask)])\n        pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3))\n        scene.add_geometry(pct)\n    else:\n        meshes = []\n        for i in range(len(imgs)):\n            meshes.append(pts3d_to_trimesh(imgs[i], pts3d[i], mask[i]))\n        mesh = trimesh.Trimesh(**cat_meshes(meshes))\n        scene.add_geometry(mesh)\n\n    # add each camera\n    if show_cam:\n        for i, pose_c2w in enumerate(cams2world):\n            if isinstance(cam_color, list):\n                camera_edge_color = cam_color[i]\n            else:\n                camera_edge_color = cam_color or CAM_COLORS[i % len(CAM_COLORS)]\n            add_scene_cam(\n                scene,\n                pose_c2w,\n                camera_edge_color,\n                None if transparent_cams else imgs[i],\n                focals[i],\n                imsize=imgs[i].shape[1::-1],\n                screen_width=cam_size,\n            )\n\n    rot = np.eye(4)\n    rot[:3, :3] = Rotation.from_euler(\"y\", np.deg2rad(180)).as_matrix()\n    scene.apply_transform(np.linalg.inv(cams2world[0] @ OPENGL @ rot))\n    if save_name is None:\n        save_name = \"scene\"\n    outfile = os.path.join(outdir, save_name + \".glb\")\n    if not silent:\n        print(\"(exporting 3D scene to\", outfile, \")\")\n    scene.export(file_obj=outfile)\n    return outfile\n\n\n@dataclasses.dataclass\nclass CameraState(object):\n    fov: float\n    aspect: float\n    c2w: np.ndarray\n\n    def get_K(self, img_wh):\n        W, H = img_wh\n        focal_length = H / 2.0 / np.tan(self.fov / 2.0)\n        K = np.array(\n            [\n                [focal_length, 0.0, W / 2.0],\n                [0.0, focal_length, H / 2.0],\n                [0.0, 0.0, 1.0],\n            ]\n        )\n        return K\n","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.CameraState","uri":"program://Human3R/class/viser_utils.CameraState#L168-L183","kind":"class","name":"CameraState","path":"viser_utils.py","language":"python","start_line":168,"end_line":183,"context_start_line":148,"context_end_line":203,"code":"                camera_edge_color,\n                None if transparent_cams else imgs[i],\n                focals[i],\n                imsize=imgs[i].shape[1::-1],\n                screen_width=cam_size,\n            )\n\n    rot = np.eye(4)\n    rot[:3, :3] = Rotation.from_euler(\"y\", np.deg2rad(180)).as_matrix()\n    scene.apply_transform(np.linalg.inv(cams2world[0] @ OPENGL @ rot))\n    if save_name is None:\n        save_name = \"scene\"\n    outfile = os.path.join(outdir, save_name + \".glb\")\n    if not silent:\n        print(\"(exporting 3D scene to\", outfile, \")\")\n    scene.export(file_obj=outfile)\n    return outfile\n\n\n@dataclasses.dataclass\nclass CameraState(object):\n    fov: float\n    aspect: float\n    c2w: np.ndarray\n\n    def get_K(self, img_wh):\n        W, H = img_wh\n        focal_length = H / 2.0 / np.tan(self.fov / 2.0)\n        K = np.array(\n            [\n                [focal_length, 0.0, W / 2.0],\n                [0.0, focal_length, H / 2.0],\n                [0.0, 0.0, 1.0],\n            ]\n        )\n        return K\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.get_vertical_colorbar","uri":"program://Human3R/function/viser_utils.get_vertical_colorbar#L186-L230","kind":"function","name":"get_vertical_colorbar","path":"viser_utils.py","language":"python","start_line":186,"end_line":230,"context_start_line":166,"context_end_line":250,"code":"\n@dataclasses.dataclass\nclass CameraState(object):\n    fov: float\n    aspect: float\n    c2w: np.ndarray\n\n    def get_K(self, img_wh):\n        W, H = img_wh\n        focal_length = H / 2.0 / np.tan(self.fov / 2.0)\n        K = np.array(\n            [\n                [focal_length, 0.0, W / 2.0],\n                [0.0, focal_length, H / 2.0],\n                [0.0, 0.0, 1.0],\n            ]\n        )\n        return K\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n\n    tick_cnt = 6\n    tick_loc = np.linspace(vmin, vmax, tick_cnt)\n    cb1 = mpl.colorbar.ColorbarBase(\n        ax, cmap=cmap, norm=norm, ticks=tick_loc, orientation=\"vertical\"\n    )\n\n    tick_label = [str(np.round(x, cbar_precision)) for x in tick_loc]\n    if cbar_precision == 0:\n        tick_label = [x[:-2] for x in tick_label]\n\n    cb1.set_ticklabels(tick_label)\n\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.colorize_np","uri":"program://Human3R/function/viser_utils.colorize_np#L233-L292","kind":"function","name":"colorize_np","path":"viser_utils.py","language":"python","start_line":233,"end_line":292,"context_start_line":213,"context_end_line":312,"code":"\n    cb1.set_ticklabels(tick_label)\n\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]\n    \"\"\"\n    if range is not None:\n        vmin, vmax = range\n    elif mask is not None:\n\n        vmin = np.min(x[mask][np.nonzero(x[mask])])\n        vmax = np.max(x[mask])\n\n        x[np.logical_not(mask)] = vmin\n\n    else:\n        vmin, vmax = np.percentile(x, (1, 100))\n        vmax += 1e-6\n\n    x = np.clip(x, vmin, vmax)\n    x = (x - vmin) / (vmax - vmin)\n\n    cmap = cm.get_cmap(cmap_name)\n    x_new = cmap(x)[:, :, :3]\n\n    if mask is not None:\n        mask = np.float32(mask[:, :, np.newaxis])\n        x_new = x_new * mask + np.ones_like(x_new) * (1.0 - mask)\n\n    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.colorize","uri":"program://Human3R/function/viser_utils.colorize#L295-L323","kind":"function","name":"colorize","path":"viser_utils.py","language":"python","start_line":295,"end_line":323,"context_start_line":275,"context_end_line":343,"code":"    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out\n\ndef get_color(idx):\n    root_dir = os.path.dirname(os.path.abspath(__file__))\n    colors_path = os.path.join(root_dir, \"src/models/smpl_colors.txt\")\n    colors = np.loadtxt(colors_path).astype(int)\n    return colors[idx % len(colors)]\n\nclass SceneHumanViewer:\n    def __init__(\n        self,\n        pc_list,\n        color_list,\n        conf_list,\n        cam_dict,\n        all_smpl_verts,\n        smpl_faces,\n        smpl_id,\n        msk_list,\n        gt_cam_dict=None,\n        gt_smpl_verts=None,","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.get_color","uri":"program://Human3R/function/viser_utils.get_color#L325-L329","kind":"function","name":"get_color","path":"viser_utils.py","language":"python","start_line":325,"end_line":329,"context_start_line":305,"context_end_line":349,"code":"    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out\n\ndef get_color(idx):\n    root_dir = os.path.dirname(os.path.abspath(__file__))\n    colors_path = os.path.join(root_dir, \"src/models/smpl_colors.txt\")\n    colors = np.loadtxt(colors_path).astype(int)\n    return colors[idx % len(colors)]\n\nclass SceneHumanViewer:\n    def __init__(\n        self,\n        pc_list,\n        color_list,\n        conf_list,\n        cam_dict,\n        all_smpl_verts,\n        smpl_faces,\n        smpl_id,\n        msk_list,\n        gt_cam_dict=None,\n        gt_smpl_verts=None,\n        image_mask=None,\n        edge_color_list=None,\n        device=\"cpu\",\n        port=8080,\n        show_camera=True,\n        show_gt_camera=False,","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.SceneHumanViewer","uri":"program://Human3R/class/viser_utils.SceneHumanViewer#L331-L1392","kind":"class","name":"SceneHumanViewer","path":"viser_utils.py","language":"python","start_line":331,"end_line":1392,"context_start_line":311,"context_end_line":1392,"code":"        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out\n\ndef get_color(idx):\n    root_dir = os.path.dirname(os.path.abspath(__file__))\n    colors_path = os.path.join(root_dir, \"src/models/smpl_colors.txt\")\n    colors = np.loadtxt(colors_path).astype(int)\n    return colors[idx % len(colors)]\n\nclass SceneHumanViewer:\n    def __init__(\n        self,\n        pc_list,\n        color_list,\n        conf_list,\n        cam_dict,\n        all_smpl_verts,\n        smpl_faces,\n        smpl_id,\n        msk_list,\n        gt_cam_dict=None,\n        gt_smpl_verts=None,\n        image_mask=None,\n        edge_color_list=None,\n        device=\"cpu\",\n        port=8080,\n        show_camera=True,\n        show_gt_camera=False,\n        show_gt_smpl=False,\n        vis_threshold=1,\n        msk_threshold=0.1,\n        mask_morph=0,\n        size=512,\n        downsample_factor=10,\n        smpl_downsample_factor=1,\n        camera_downsample_factor=1,\n    ):\n        self.size=size\n        self.server = viser.ViserServer(port=port)\n        self.server.set_up_direction(\"-y\")\n        self.device = device\n        self.conf_list = conf_list\n        self.msk_list = msk_list\n        self.vis_threshold = vis_threshold\n        self.msk_threshold = msk_threshold\n        self.mask_morph = mask_morph\n        self.show_background = True\n        self.show_foreground = False\n        self.tt = lambda x: torch.from_numpy(x).float().to(device)\n        self.pcs, self.all_steps = self.read_data(\n            pc_list, color_list, conf_list, msk_list, \n            all_smpl_verts, smpl_faces, smpl_id, edge_color_list, gt_smpl_verts\n        )\n        # Fast lookup from step id to its sequential index\n        self.step_to_index = {step: idx for idx, step in enumerate(self.all_steps)}\n        self.cam_dict = cam_dict\n        self.gt_cam_dict = gt_cam_dict\n        self.gt_smpl_verts = gt_smpl_verts\n        self.num_frames = len(self.all_steps)\n        self.image_mask = image_mask\n        self.show_camera = show_camera\n        self.show_gt_camera = show_gt_camera and gt_cam_dict is not None\n        self.show_gt_smpl = show_gt_smpl and gt_smpl_verts is not None\n        self.on_replay = False\n        self.vis_pts_list = []\n        self.traj_list = []\n        self.orig_img_list = [x[0] for x in color_list]\n        self.via_points = []\n        self._updating_point_clouds = False\n        \n        # Performance optimization for dynamic opacity\n        self._last_opacity_update_step = -1\n        self._opacity_update_throttle = 0  # Frames to skip between updates\n        self._opacity_frame_counter = 0\n\n        gui_reset_up = self.server.gui.add_button(\n            \"Reset up direction\",\n            hint=\"Set the camera control 'up' direction to the current camera's 'up'.\",\n        )\n\n        @gui_reset_up.on_click\n        def _(event: viser.GuiEvent) -> None:\n            client = event.client\n            assert client is not None\n            client.camera.up_direction = tf.SO3(client.camera.wxyz) @ np.array(\n                [0.0, -1.0, 0.0]\n            )\n\n        button3 = self.server.gui.add_button(\"4D (Only Show Current Frame)\")\n        button4 = self.server.gui.add_button(\"3D (Show All Frames)\")\n        button5 = self.server.gui.add_button(\"Hybrid (Current SMPL + All Points)\")\n        self.is_render = False\n        self.fourd = False\n        self.hybrid_mode = False\n\n        @button3.on_click\n        def _(event: viser.GuiEvent) -> None:\n            self.fourd = True\n            self.hybrid_mode = False\n\n        @button4.on_click\n        def _(event: viser.GuiEvent) -> None:\n            self.fourd = False\n            self.hybrid_mode = False\n\n        @button5.on_click\n        def _(event: viser.GuiEvent) -> None:\n            self.fourd = False\n            self.hybrid_mode = True\n\n        self.gui_show_background = self.server.add_gui_checkbox(\n            \"Show Background\", True)\n\n        @self.gui_show_background.on_update\n        def _(_) -> None:\n            self.show_background = self.gui_show_background.value\n            self._update_point_clouds()\n\n        self.gui_show_foreground = self.server.add_gui_checkbox(\n            \"Show Foreground\", False)\n\n        @self.gui_show_foreground.on_update\n        def _(_) -> None:\n            self.show_foreground = self.gui_show_foreground.value\n            self._update_point_clouds()\n\n        self.gui_show_smpl = self.server.add_gui_checkbox(\"Show SMPL\", True)\n\n        @self.gui_show_smpl.on_update\n        def _(_) -> None:\n            # Update SMPL visibility considering downsampling factor\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'frame_nodes'):\n                self._update_smpl_visibility()\n\n        self.focal_slider = self.server.add_gui_slider(\n            \"Focal Length\",\n            min=0.1,\n            max=99999,\n            step=1,\n            initial_value=533,\n        )\n\n        self.psize_slider = self.server.add_gui_slider(\n            \"Point Size\",\n            min=0.0001,\n            max=0.1,\n            step=0.0001,\n            initial_value=0.005,\n        )\n        self.camsize_slider = self.server.add_gui_slider(\n            \"Camera Size\",\n            min=0.01,\n            max=0.5,\n            step=0.01,\n            initial_value=0.1,\n        )\n\n        self.downsample_slider = self.server.add_gui_slider(\n            \"Downsample Factor\",\n            min=1,\n            max=1000,\n            step=1,\n            initial_value=downsample_factor,\n        )\n        self.vis_threshold_slider = self.server.add_gui_slider(\n            \"Visibility Threshold\",\n            min=0.1,\n            max=50.0,\n            step=0.1,\n            initial_value=self.vis_threshold,\n        )\n        \n        self.mask_morph_slider = self.server.add_gui_slider(\n            \"Mask Morphology\",\n            min=-20.0,\n            max=100.0,\n            step=0.1,\n            initial_value=self.mask_morph,\n        )\n\n        # SMPL downsampling controls\n        self.smpl_downsample_slider = self.server.add_gui_slider(\n            \"SMPL Downsample\",\n            min=1,\n            max=200,\n            step=1,\n            initial_value=smpl_downsample_factor,\n        )\n        \n        # Camera downsampling controls\n        self.camera_downsample_slider = self.server.add_gui_slider(\n            \"Camera Downsample\",\n            min=1,\n            max=200,\n            step=1,\n            initial_value=camera_downsample_factor,\n        )\n        \n        @self.camera_downsample_slider.on_update\n        def _(_) -> None:\n            # Apply camera downsampling changes immediately\n            if hasattr(self, 'cam_handles'):\n                self._update_camera_visibility()\n            if hasattr(self, 'gt_cam_handles'):\n                self._update_gt_camera_visibility()\n\n        # Mesh opacity by time controls\n        self.mesh_time_opacity_checkbox = self.server.add_gui_checkbox(\n            \"Mesh Opacity by Time\", False\n        )\n        self.min_mesh_opacity_slider = self.server.add_gui_slider(\n            \"Min Mesh Opacity\",\n            min=0.0,\n            max=1.0,\n            step=0.05,\n            initial_value=0.1,\n        )\n\n        @self.mesh_time_opacity_checkbox.on_update\n        def _(_) -> None:\n            # Apply opacity changes to existing meshes immediately\n            if hasattr(self, 'mesh_handles'):\n                # If dynamic opacity is enabled, prefer dynamic update\n                if self.dynamic_opacity_checkbox.value and hasattr(self, 'current_step_index'):\n                    self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                else:\n                    self._update_mesh_opacities()\n\n        @self.min_mesh_opacity_slider.on_update\n        def _(_) -> None:\n            # Apply opacity changes to existing meshes immediately\n            if hasattr(self, 'mesh_handles'):\n                # If dynamic opacity is enabled, prefer dynamic update so min opacity affects decay baseline\n                if self.dynamic_opacity_checkbox.value and hasattr(self, 'current_step_index'):\n                    self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                else:\n                    self._update_mesh_opacities()\n\n        # Dynamic opacity controls\n        self.dynamic_opacity_checkbox = self.server.add_gui_checkbox(\n            \"Dynamic Opacity\", False\n        )\n        self.opacity_decay_len_slider = self.server.add_gui_slider(\n            \"Decay Length\",\n            min=1,\n            max=max(10, len(self.all_steps)),\n            step=1,\n            initial_value=min(10, max(1, len(self.all_steps))),\n        )\n        \n        # Performance control for dynamic opacity\n        self.opacity_update_throttle_slider = self.server.add_gui_slider(\n            \"Opacity Update Throttle\",\n            min=0,\n            max=10,\n            step=1,\n            initial_value=0,\n            hint=\"Skip N frames between opacity updates (0=update every frame, higher=better performance)\"\n        )\n        \n        # Performance preset buttons\n        self.performance_preset_buttons = self.server.add_gui_button_group(\n            \"Performance Presets\", \n            (\"High Quality\", \"Balanced\")\n        )\n\n        @self.dynamic_opacity_checkbox.on_update\n        def _(_) -> None:\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'current_step_index'):\n                if self.dynamic_opacity_checkbox.value:\n                    # Turning on dynamic opacity: update immediately\n                    self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                else:\n                    # Turning off dynamic opacity: fall back to static-by-time if enabled,\n                    # otherwise set all to fully opaque\n                    if self.mesh_time_opacity_checkbox.value:\n                        self._update_mesh_opacities()\n                    else:\n                        for handle in self.mesh_handles:\n                            handle.opacity = 1.0\n                        for handle in self.gt_mesh_handles:\n                            handle.opacity = 1.0\n\n        @self.opacity_decay_len_slider.on_update\n        def _(_) -> None:\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'current_step_index'):\n                self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                \n        @self.opacity_update_throttle_slider.on_update\n        def _(_) -> None:\n            self._opacity_update_throttle = int(self.opacity_update_throttle_slider.value)\n            self._opacity_frame_counter = 0  # Reset counter when throttle changes\n            # Force update to apply changes immediately\n            if hasattr(self, 'current_step_index') and self.dynamic_opacity_checkbox.value:\n                self._update_dynamic_opacities(self.current_step_index, force_update=True)\n        \n        @self.performance_preset_buttons.on_click\n        def _(_) -> None:\n            preset = self.performance_preset_buttons.value\n            if preset == \"High Quality\":\n                # Best visual quality, may be slower\n                self.opacity_update_throttle_slider.value = 0\n                self._opacity_update_throttle = 0\n            elif preset == \"Balanced\":\n                # Good balance of quality and performance\n                self.opacity_update_throttle_slider.value = 2\n                self._opacity_update_throttle = 2\n                \n            self._opacity_frame_counter = 0\n            # Apply preset immediately\n            if hasattr(self, 'current_step_index') and self.dynamic_opacity_checkbox.value:\n                self._update_dynamic_opacities(self.current_step_index, force_update=True)\n\n        self.show_camera_checkbox = self.server.add_gui_checkbox(\n            \"Show Camera\", \n            initial_value=self.show_camera\n        )\n\n        self.show_gt_camera_checkbox = self.server.add_gui_checkbox(\n            \"Show GT Camera\", \n            initial_value=self.show_gt_camera\n        )\n\n        self.show_gt_smpl_checkbox = self.server.add_gui_checkbox(\n            \"Show GT SMPL\", \n            initial_value=self.show_gt_smpl\n        )\n   \n        self.pc_handles = []\n        self.cam_handles = []\n        self.gt_cam_handles = []\n        self.mesh_handles = []\n        self.mesh_step_mapping = []\n        self.gt_mesh_handles = []\n        self.gt_mesh_step_mapping = []\n\n        @self.psize_slider.on_update\n        def _(_) -> None:\n            for handle in self.pc_handles:\n                handle.point_size = self.psize_slider.value\n\n        @self.camsize_slider.on_update\n        def _(_) -> None:\n            for handle in self.cam_handles:\n                handle.scale = self.camsize_slider.value\n                handle.line_thickness = 0.03 * handle.scale\n            for handle in self.gt_cam_handles:\n                handle.scale = self.camsize_slider.value\n                handle.line_thickness = 0.03 * handle.scale\n\n        @self.downsample_slider.on_update\n        def _(_) -> None:\n            # when downsampling factor changes, regenerate all point clouds\n            if hasattr(self, 'frame_nodes'):\n                self._update_point_clouds()\n\n        @self.show_camera_checkbox.on_update\n        def _(_) -> None:\n            # update internal state\n            self.show_camera = self.show_camera_checkbox.value\n            \n            if self.show_camera:\n                # if camera display is enabled, ensure all cameras are visible with downsampling\n                self._update_camera_visibility()\n                    \n                # check if any cameras are missing\n                if hasattr(self, 'frame_nodes') and len(self.cam_handles) < len(self.frame_nodes):\n                    for i in range(len(self.cam_handles), len(self.frame_nodes)):\n                        if i < len(self.all_steps):\n                            step = self.all_steps[i]\n                            self.add_camera(step)\n                    # Apply downsampling to newly added cameras\n                    self._update_camera_visibility()\n            else:\n                # if camera display is disabled, hide all cameras\n                for handle in self.cam_handles:\n                    handle.visible = False\n\n        @self.show_gt_camera_checkbox.on_update\n        def _(_) -> None:\n            # update internal state\n            self.show_gt_camera = self.show_gt_camera_checkbox.value\n            \n            if self.show_gt_camera and self.gt_cam_dict is not None:\n                # if GT camera display is enabled, ensure all GT cameras are visible with downsampling\n                self._update_gt_camera_visibility()\n                    \n                # check if any GT cameras are missing\n                if hasattr(self, 'frame_nodes') and len(self.gt_cam_handles) < len(self.frame_nodes):\n                    for i in range(len(self.gt_cam_handles), len(self.frame_nodes)):\n                        if i < len(self.all_steps):\n                            step = self.all_steps[i]\n                            self.add_gt_camera(step)\n                    # Apply downsampling to newly added GT cameras\n                    self._update_gt_camera_visibility()\n            else:\n                # if GT camera display is disabled, hide all GT cameras\n                for handle in self.gt_cam_handles:\n                    handle.visible = False\n\n        @self.show_gt_smpl_checkbox.on_update\n        def _(_) -> None:\n            # update internal state\n            self.show_gt_smpl = self.show_gt_smpl_checkbox.value\n            \n            if self.show_gt_smpl and self.gt_smpl_verts is not None:\n                # if GT SMPL display is enabled, ensure all GT meshes are visible\n                for handle in self.gt_mesh_handles:\n                    handle.visible = True\n                    \n                # check if any GT meshes are missing\n                if hasattr(self, 'frame_nodes') and len(self.gt_mesh_handles) < len(self.frame_nodes):\n                    for i in range(len(self.gt_mesh_handles), len(self.frame_nodes)):\n                        if i < len(self.all_steps):\n                            step = self.all_steps[i]\n                            self.add_gt_smpl(step)\n            else:\n                # if GT SMPL display is disabled, hide all GT meshes\n                for handle in self.gt_mesh_handles:\n                    handle.visible = False\n\n        @self.vis_threshold_slider.on_update\n        def _(_) -> None:\n            # when visibility threshold changes, update threshold and regenerate point clouds\n            self.vis_threshold = self.vis_threshold_slider.value\n            if hasattr(self, 'frame_nodes'):\n                self._update_point_clouds()\n        \n        @self.mask_morph_slider.on_update\n        def _(_) -> None:\n            # when mask morphology changes, update morphology and regenerate point clouds\n            self.mask_morph = self.mask_morph_slider.value\n            if hasattr(self, 'frame_nodes'):\n                self._update_point_clouds()\n\n        @self.smpl_downsample_slider.on_update\n        def _(_) -> None:\n            # when SMPL downsampling factor changes, update SMPL visibility\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'frame_nodes'):\n                self._update_smpl_visibility()\n\n        self.server.on_client_connect(self._connect_client)\n\n    def get_camera_state(self, client: viser.ClientHandle) -> CameraState:\n        camera = client.camera\n        c2w = np.concatenate(\n            [\n                np.concatenate(\n                    [tf.SO3(camera.wxyz).as_matrix(), camera.position[:, None]], 1\n                ),\n                [[0, 0, 0, 1]],\n            ],\n            0,\n        )\n        return CameraState(\n            fov=camera.fov,\n            aspect=camera.aspect,\n            c2w=c2w,\n        )\n\n    @staticmethod\n    def generate_pseudo_intrinsics(h, w):\n        focal = (h**2 + w**2) ** 0.5\n        return np.array([[focal, 0, w // 2], [0, focal, h // 2], [0, 0, 1]]).astype(\n            np.float32\n        )\n\n    def get_ray_map(self, c2w, h, w, intrinsics=None):\n        if intrinsics is None:\n            intrinsics = self.generate_pseudo_intrinsics(h, w)\n        i, j = np.meshgrid(np.arange(w), np.arange(h), indexing=\"xy\")\n        grid = np.stack([i, j, np.ones_like(i)], axis=-1)\n        ro = c2w[:3, 3]\n        rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n        rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n        rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n        ro = np.broadcast_to(ro, (h, w, 3))\n        ray_map = np.concatenate([ro, rd], axis=-1)\n        return ray_map\n\n    def set_camera_loc(camera, pose, K):\n        \"\"\"\n        pose: 4x4 matrix\n        K: 3x3 matrix\n        \"\"\"\n        fx, fy = K[0, 0], K[1, 1]\n        cx, cy = K[0, 2], K[1, 2]\n        aspect = float(cx) / float(cy)\n        fov = 2 * np.arctan(2 * cx / fx)\n        wxyz_xyz = tf.SE3.from_matrix(pose).wxyz_xyz\n        wxyz = wxyz_xyz[:4]\n        xyz = wxyz_xyz[4:]\n        camera.wxyz = wxyz\n        camera.position = xyz\n        camera.fov = fov\n        camera.aspect = aspect\n\n    def _connect_client(self, client: viser.ClientHandle):\n        from src.dust3r.inference import inference_step\n        from src.dust3r.utils.geometry import geotrf\n\n        wxyz_panel = client.gui.add_text(\"wxyz:\", f\"{client.camera.wxyz}\")\n        position_panel = client.gui.add_text(\"position:\", f\"{client.camera.position}\")\n        fov_panel = client.gui.add_text(\n            \"fov:\", f\"{2\n# ... truncated ...","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":true}
{"repo_id":"Human3R","entity_id":"py:viser_utils.get_K","uri":"program://Human3R/function/viser_utils.get_K#L173-L183","kind":"function","name":"get_K","path":"viser_utils.py","language":"python","start_line":173,"end_line":183,"context_start_line":153,"context_end_line":203,"code":"            )\n\n    rot = np.eye(4)\n    rot[:3, :3] = Rotation.from_euler(\"y\", np.deg2rad(180)).as_matrix()\n    scene.apply_transform(np.linalg.inv(cams2world[0] @ OPENGL @ rot))\n    if save_name is None:\n        save_name = \"scene\"\n    outfile = os.path.join(outdir, save_name + \".glb\")\n    if not silent:\n        print(\"(exporting 3D scene to\", outfile, \")\")\n    scene.export(file_obj=outfile)\n    return outfile\n\n\n@dataclasses.dataclass\nclass CameraState(object):\n    fov: float\n    aspect: float\n    c2w: np.ndarray\n\n    def get_K(self, img_wh):\n        W, H = img_wh\n        focal_length = H / 2.0 / np.tan(self.fov / 2.0)\n        K = np.array(\n            [\n                [focal_length, 0.0, W / 2.0],\n                [0.0, focal_length, H / 2.0],\n                [0.0, 0.0, 1.0],\n            ]\n        )\n        return K\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.__init__","uri":"program://Human3R/function/viser_utils.__init__#L332-L763","kind":"function","name":"__init__","path":"viser_utils.py","language":"python","start_line":332,"end_line":763,"context_start_line":312,"context_end_line":783,"code":"        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out\n\ndef get_color(idx):\n    root_dir = os.path.dirname(os.path.abspath(__file__))\n    colors_path = os.path.join(root_dir, \"src/models/smpl_colors.txt\")\n    colors = np.loadtxt(colors_path).astype(int)\n    return colors[idx % len(colors)]\n\nclass SceneHumanViewer:\n    def __init__(\n        self,\n        pc_list,\n        color_list,\n        conf_list,\n        cam_dict,\n        all_smpl_verts,\n        smpl_faces,\n        smpl_id,\n        msk_list,\n        gt_cam_dict=None,\n        gt_smpl_verts=None,\n        image_mask=None,\n        edge_color_list=None,\n        device=\"cpu\",\n        port=8080,\n        show_camera=True,\n        show_gt_camera=False,\n        show_gt_smpl=False,\n        vis_threshold=1,\n        msk_threshold=0.1,\n        mask_morph=0,\n        size=512,\n        downsample_factor=10,\n        smpl_downsample_factor=1,\n        camera_downsample_factor=1,\n    ):\n        self.size=size\n        self.server = viser.ViserServer(port=port)\n        self.server.set_up_direction(\"-y\")\n        self.device = device\n        self.conf_list = conf_list\n        self.msk_list = msk_list\n        self.vis_threshold = vis_threshold\n        self.msk_threshold = msk_threshold\n        self.mask_morph = mask_morph\n        self.show_background = True\n        self.show_foreground = False\n        self.tt = lambda x: torch.from_numpy(x).float().to(device)\n        self.pcs, self.all_steps = self.read_data(\n            pc_list, color_list, conf_list, msk_list, \n            all_smpl_verts, smpl_faces, smpl_id, edge_color_list, gt_smpl_verts\n        )\n        # Fast lookup from step id to its sequential index\n        self.step_to_index = {step: idx for idx, step in enumerate(self.all_steps)}\n        self.cam_dict = cam_dict\n        self.gt_cam_dict = gt_cam_dict\n        self.gt_smpl_verts = gt_smpl_verts\n        self.num_frames = len(self.all_steps)\n        self.image_mask = image_mask\n        self.show_camera = show_camera\n        self.show_gt_camera = show_gt_camera and gt_cam_dict is not None\n        self.show_gt_smpl = show_gt_smpl and gt_smpl_verts is not None\n        self.on_replay = False\n        self.vis_pts_list = []\n        self.traj_list = []\n        self.orig_img_list = [x[0] for x in color_list]\n        self.via_points = []\n        self._updating_point_clouds = False\n        \n        # Performance optimization for dynamic opacity\n        self._last_opacity_update_step = -1\n        self._opacity_update_throttle = 0  # Frames to skip between updates\n        self._opacity_frame_counter = 0\n\n        gui_reset_up = self.server.gui.add_button(\n            \"Reset up direction\",\n            hint=\"Set the camera control 'up' direction to the current camera's 'up'.\",\n        )\n\n        @gui_reset_up.on_click\n        def _(event: viser.GuiEvent) -> None:\n            client = event.client\n            assert client is not None\n            client.camera.up_direction = tf.SO3(client.camera.wxyz) @ np.array(\n                [0.0, -1.0, 0.0]\n            )\n\n        button3 = self.server.gui.add_button(\"4D (Only Show Current Frame)\")\n        button4 = self.server.gui.add_button(\"3D (Show All Frames)\")\n        button5 = self.server.gui.add_button(\"Hybrid (Current SMPL + All Points)\")\n        self.is_render = False\n        self.fourd = False\n        self.hybrid_mode = False\n\n        @button3.on_click\n        def _(event: viser.GuiEvent) -> None:\n            self.fourd = True\n            self.hybrid_mode = False\n\n        @button4.on_click\n        def _(event: viser.GuiEvent) -> None:\n            self.fourd = False\n            self.hybrid_mode = False\n\n        @button5.on_click\n        def _(event: viser.GuiEvent) -> None:\n            self.fourd = False\n            self.hybrid_mode = True\n\n        self.gui_show_background = self.server.add_gui_checkbox(\n            \"Show Background\", True)\n\n        @self.gui_show_background.on_update\n        def _(_) -> None:\n            self.show_background = self.gui_show_background.value\n            self._update_point_clouds()\n\n        self.gui_show_foreground = self.server.add_gui_checkbox(\n            \"Show Foreground\", False)\n\n        @self.gui_show_foreground.on_update\n        def _(_) -> None:\n            self.show_foreground = self.gui_show_foreground.value\n            self._update_point_clouds()\n\n        self.gui_show_smpl = self.server.add_gui_checkbox(\"Show SMPL\", True)\n\n        @self.gui_show_smpl.on_update\n        def _(_) -> None:\n            # Update SMPL visibility considering downsampling factor\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'frame_nodes'):\n                self._update_smpl_visibility()\n\n        self.focal_slider = self.server.add_gui_slider(\n            \"Focal Length\",\n            min=0.1,\n            max=99999,\n            step=1,\n            initial_value=533,\n        )\n\n        self.psize_slider = self.server.add_gui_slider(\n            \"Point Size\",\n            min=0.0001,\n            max=0.1,\n            step=0.0001,\n            initial_value=0.005,\n        )\n        self.camsize_slider = self.server.add_gui_slider(\n            \"Camera Size\",\n            min=0.01,\n            max=0.5,\n            step=0.01,\n            initial_value=0.1,\n        )\n\n        self.downsample_slider = self.server.add_gui_slider(\n            \"Downsample Factor\",\n            min=1,\n            max=1000,\n            step=1,\n            initial_value=downsample_factor,\n        )\n        self.vis_threshold_slider = self.server.add_gui_slider(\n            \"Visibility Threshold\",\n            min=0.1,\n            max=50.0,\n            step=0.1,\n            initial_value=self.vis_threshold,\n        )\n        \n        self.mask_morph_slider = self.server.add_gui_slider(\n            \"Mask Morphology\",\n            min=-20.0,\n            max=100.0,\n            step=0.1,\n            initial_value=self.mask_morph,\n        )\n\n        # SMPL downsampling controls\n        self.smpl_downsample_slider = self.server.add_gui_slider(\n            \"SMPL Downsample\",\n            min=1,\n            max=200,\n            step=1,\n            initial_value=smpl_downsample_factor,\n        )\n        \n        # Camera downsampling controls\n        self.camera_downsample_slider = self.server.add_gui_slider(\n            \"Camera Downsample\",\n            min=1,\n            max=200,\n            step=1,\n            initial_value=camera_downsample_factor,\n        )\n        \n        @self.camera_downsample_slider.on_update\n        def _(_) -> None:\n            # Apply camera downsampling changes immediately\n            if hasattr(self, 'cam_handles'):\n                self._update_camera_visibility()\n            if hasattr(self, 'gt_cam_handles'):\n                self._update_gt_camera_visibility()\n\n        # Mesh opacity by time controls\n        self.mesh_time_opacity_checkbox = self.server.add_gui_checkbox(\n            \"Mesh Opacity by Time\", False\n        )\n        self.min_mesh_opacity_slider = self.server.add_gui_slider(\n            \"Min Mesh Opacity\",\n            min=0.0,\n            max=1.0,\n            step=0.05,\n            initial_value=0.1,\n        )\n\n        @self.mesh_time_opacity_checkbox.on_update\n        def _(_) -> None:\n            # Apply opacity changes to existing meshes immediately\n            if hasattr(self, 'mesh_handles'):\n                # If dynamic opacity is enabled, prefer dynamic update\n                if self.dynamic_opacity_checkbox.value and hasattr(self, 'current_step_index'):\n                    self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                else:\n                    self._update_mesh_opacities()\n\n        @self.min_mesh_opacity_slider.on_update\n        def _(_) -> None:\n            # Apply opacity changes to existing meshes immediately\n            if hasattr(self, 'mesh_handles'):\n                # If dynamic opacity is enabled, prefer dynamic update so min opacity affects decay baseline\n                if self.dynamic_opacity_checkbox.value and hasattr(self, 'current_step_index'):\n                    self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                else:\n                    self._update_mesh_opacities()\n\n        # Dynamic opacity controls\n        self.dynamic_opacity_checkbox = self.server.add_gui_checkbox(\n            \"Dynamic Opacity\", False\n        )\n        self.opacity_decay_len_slider = self.server.add_gui_slider(\n            \"Decay Length\",\n            min=1,\n            max=max(10, len(self.all_steps)),\n            step=1,\n            initial_value=min(10, max(1, len(self.all_steps))),\n        )\n        \n        # Performance control for dynamic opacity\n        self.opacity_update_throttle_slider = self.server.add_gui_slider(\n            \"Opacity Update Throttle\",\n            min=0,\n            max=10,\n            step=1,\n            initial_value=0,\n            hint=\"Skip N frames between opacity updates (0=update every frame, higher=better performance)\"\n        )\n        \n        # Performance preset buttons\n        self.performance_preset_buttons = self.server.add_gui_button_group(\n            \"Performance Presets\", \n            (\"High Quality\", \"Balanced\")\n        )\n\n        @self.dynamic_opacity_checkbox.on_update\n        def _(_) -> None:\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'current_step_index'):\n                if self.dynamic_opacity_checkbox.value:\n                    # Turning on dynamic opacity: update immediately\n                    self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                else:\n                    # Turning off dynamic opacity: fall back to static-by-time if enabled,\n                    # otherwise set all to fully opaque\n                    if self.mesh_time_opacity_checkbox.value:\n                        self._update_mesh_opacities()\n                    else:\n                        for handle in self.mesh_handles:\n                            handle.opacity = 1.0\n                        for handle in self.gt_mesh_handles:\n                            handle.opacity = 1.0\n\n        @self.opacity_decay_len_slider.on_update\n        def _(_) -> None:\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'current_step_index'):\n                self._update_dynamic_opacities(self.current_step_index, force_update=True)\n                \n        @self.opacity_update_throttle_slider.on_update\n        def _(_) -> None:\n            self._opacity_update_throttle = int(self.opacity_update_throttle_slider.value)\n            self._opacity_frame_counter = 0  # Reset counter when throttle changes\n            # Force update to apply changes immediately\n            if hasattr(self, 'current_step_index') and self.dynamic_opacity_checkbox.value:\n                self._update_dynamic_opacities(self.current_step_index, force_update=True)\n        \n        @self.performance_preset_buttons.on_click\n        def _(_) -> None:\n            preset = self.performance_preset_buttons.value\n            if preset == \"High Quality\":\n                # Best visual quality, may be slower\n                self.opacity_update_throttle_slider.value = 0\n                self._opacity_update_throttle = 0\n            elif preset == \"Balanced\":\n                # Good balance of quality and performance\n                self.opacity_update_throttle_slider.value = 2\n                self._opacity_update_throttle = 2\n                \n            self._opacity_frame_counter = 0\n            # Apply preset immediately\n            if hasattr(self, 'current_step_index') and self.dynamic_opacity_checkbox.value:\n                self._update_dynamic_opacities(self.current_step_index, force_update=True)\n\n        self.show_camera_checkbox = self.server.add_gui_checkbox(\n            \"Show Camera\", \n            initial_value=self.show_camera\n        )\n\n        self.show_gt_camera_checkbox = self.server.add_gui_checkbox(\n            \"Show GT Camera\", \n            initial_value=self.show_gt_camera\n        )\n\n        self.show_gt_smpl_checkbox = self.server.add_gui_checkbox(\n            \"Show GT SMPL\", \n            initial_value=self.show_gt_smpl\n        )\n   \n        self.pc_handles = []\n        self.cam_handles = []\n        self.gt_cam_handles = []\n        self.mesh_handles = []\n        self.mesh_step_mapping = []\n        self.gt_mesh_handles = []\n        self.gt_mesh_step_mapping = []\n\n        @self.psize_slider.on_update\n        def _(_) -> None:\n            for handle in self.pc_handles:\n                handle.point_size = self.psize_slider.value\n\n        @self.camsize_slider.on_update\n        def _(_) -> None:\n            for handle in self.cam_handles:\n                handle.scale = self.camsize_slider.value\n                handle.line_thickness = 0.03 * handle.scale\n            for handle in self.gt_cam_handles:\n                handle.scale = self.camsize_slider.value\n                handle.line_thickness = 0.03 * handle.scale\n\n        @self.downsample_slider.on_update\n        def _(_) -> None:\n            # when downsampling factor changes, regenerate all point clouds\n            if hasattr(self, 'frame_nodes'):\n                self._update_point_clouds()\n\n        @self.show_camera_checkbox.on_update\n        def _(_) -> None:\n            # update internal state\n            self.show_camera = self.show_camera_checkbox.value\n            \n            if self.show_camera:\n                # if camera display is enabled, ensure all cameras are visible with downsampling\n                self._update_camera_visibility()\n                    \n                # check if any cameras are missing\n                if hasattr(self, 'frame_nodes') and len(self.cam_handles) < len(self.frame_nodes):\n                    for i in range(len(self.cam_handles), len(self.frame_nodes)):\n                        if i < len(self.all_steps):\n                            step = self.all_steps[i]\n                            self.add_camera(step)\n                    # Apply downsampling to newly added cameras\n                    self._update_camera_visibility()\n            else:\n                # if camera display is disabled, hide all cameras\n                for handle in self.cam_handles:\n                    handle.visible = False\n\n        @self.show_gt_camera_checkbox.on_update\n        def _(_) -> None:\n            # update internal state\n            self.show_gt_camera = self.show_gt_camera_checkbox.value\n            \n            if self.show_gt_camera and self.gt_cam_dict is not None:\n                # if GT camera display is enabled, ensure all GT cameras are visible with downsampling\n                self._update_gt_camera_visibility()\n                    \n                # check if any GT cameras are missing\n                if hasattr(self, 'frame_nodes') and len(self.gt_cam_handles) < len(self.frame_nodes):\n                    for i in range(len(self.gt_cam_handles), len(self.frame_nodes)):\n                        if i < len(self.all_steps):\n                            step = self.all_steps[i]\n                            self.add_gt_camera(step)\n                    # Apply downsampling to newly added GT cameras\n                    self._update_gt_camera_visibility()\n            else:\n                # if GT camera display is disabled, hide all GT cameras\n                for handle in self.gt_cam_handles:\n                    handle.visible = False\n\n        @self.show_gt_smpl_checkbox.on_update\n        def _(_) -> None:\n            # update internal state\n            self.show_gt_smpl = self.show_gt_smpl_checkbox.value\n            \n            if self.show_gt_smpl and self.gt_smpl_verts is not None:\n                # if GT SMPL display is enabled, ensure all GT meshes are visible\n                for handle in self.gt_mesh_handles:\n                    handle.visible = True\n                    \n                # check if any GT meshes are missing\n                if hasattr(self, 'frame_nodes') and len(self.gt_mesh_handles) < len(self.frame_nodes):\n                    for i in range(len(self.gt_mesh_handles), len(self.frame_nodes)):\n                        if i < len(self.all_steps):\n                            step = self.all_steps[i]\n                            self.add_gt_smpl(step)\n            else:\n                # if GT SMPL display is disabled, hide all GT meshes\n                for handle in self.gt_mesh_handles:\n                    handle.visible = False\n\n        @self.vis_threshold_slider.on_update\n        def _(_) -> None:\n            # when visibility threshold changes, update threshold and regenerate point clouds\n            self.vis_threshold = self.vis_threshold_slider.value\n            if hasattr(self, 'frame_nodes'):\n                self._update_point_clouds()\n        \n        @self.mask_morph_slider.on_update\n        def _(_) -> None:\n            # when mask morphology changes, update morphology and regenerate point clouds\n            self.mask_morph = self.mask_morph_slider.value\n            if hasattr(self, 'frame_nodes'):\n                self._update_point_clouds()\n\n        @self.smpl_downsample_slider.on_update\n        def _(_) -> None:\n            # when SMPL downsampling factor changes, update SMPL visibility\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'frame_nodes'):\n                self._update_smpl_visibility()\n\n        self.server.on_client_connect(self._connect_client)\n\n    def get_camera_state(self, client: viser.ClientHandle) -> CameraState:\n        camera = client.camera\n        c2w = np.concatenate(\n            [\n                np.concatenate(\n                    [tf.SO3(camera.wxyz).as_matrix(), camera.position[:, None]], 1\n                ),\n                [[0, 0, 0, 1]],\n            ],\n            0,\n        )\n        return CameraState(\n            fov=camera.fov,\n            aspect=camera.aspect,\n            c2w=c2w,\n        )\n\n    @staticmethod\n    def generate_pseudo_intrinsics(h, w):","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.get_camera_state","uri":"program://Human3R/function/viser_utils.get_camera_state#L765-L780","kind":"function","name":"get_camera_state","path":"viser_utils.py","language":"python","start_line":765,"end_line":780,"context_start_line":745,"context_end_line":800,"code":"            # when visibility threshold changes, update threshold and regenerate point clouds\n            self.vis_threshold = self.vis_threshold_slider.value\n            if hasattr(self, 'frame_nodes'):\n                self._update_point_clouds()\n        \n        @self.mask_morph_slider.on_update\n        def _(_) -> None:\n            # when mask morphology changes, update morphology and regenerate point clouds\n            self.mask_morph = self.mask_morph_slider.value\n            if hasattr(self, 'frame_nodes'):\n                self._update_point_clouds()\n\n        @self.smpl_downsample_slider.on_update\n        def _(_) -> None:\n            # when SMPL downsampling factor changes, update SMPL visibility\n            if hasattr(self, 'mesh_handles') and hasattr(self, 'frame_nodes'):\n                self._update_smpl_visibility()\n\n        self.server.on_client_connect(self._connect_client)\n\n    def get_camera_state(self, client: viser.ClientHandle) -> CameraState:\n        camera = client.camera\n        c2w = np.concatenate(\n            [\n                np.concatenate(\n                    [tf.SO3(camera.wxyz).as_matrix(), camera.position[:, None]], 1\n                ),\n                [[0, 0, 0, 1]],\n            ],\n            0,\n        )\n        return CameraState(\n            fov=camera.fov,\n            aspect=camera.aspect,\n            c2w=c2w,\n        )\n\n    @staticmethod\n    def generate_pseudo_intrinsics(h, w):\n        focal = (h**2 + w**2) ** 0.5\n        return np.array([[focal, 0, w // 2], [0, focal, h // 2], [0, 0, 1]]).astype(\n            np.float32\n        )\n\n    def get_ray_map(self, c2w, h, w, intrinsics=None):\n        if intrinsics is None:\n            intrinsics = self.generate_pseudo_intrinsics(h, w)\n        i, j = np.meshgrid(np.arange(w), np.arange(h), indexing=\"xy\")\n        grid = np.stack([i, j, np.ones_like(i)], axis=-1)\n        ro = c2w[:3, 3]\n        rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n        rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n        rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n        ro = np.broadcast_to(ro, (h, w, 3))\n        ray_map = np.concatenate([ro, rd], axis=-1)\n        return ray_map","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.generate_pseudo_intrinsics","uri":"program://Human3R/function/viser_utils.generate_pseudo_intrinsics#L783-L787","kind":"function","name":"generate_pseudo_intrinsics","path":"viser_utils.py","language":"python","start_line":783,"end_line":787,"context_start_line":763,"context_end_line":807,"code":"        self.server.on_client_connect(self._connect_client)\n\n    def get_camera_state(self, client: viser.ClientHandle) -> CameraState:\n        camera = client.camera\n        c2w = np.concatenate(\n            [\n                np.concatenate(\n                    [tf.SO3(camera.wxyz).as_matrix(), camera.position[:, None]], 1\n                ),\n                [[0, 0, 0, 1]],\n            ],\n            0,\n        )\n        return CameraState(\n            fov=camera.fov,\n            aspect=camera.aspect,\n            c2w=c2w,\n        )\n\n    @staticmethod\n    def generate_pseudo_intrinsics(h, w):\n        focal = (h**2 + w**2) ** 0.5\n        return np.array([[focal, 0, w // 2], [0, focal, h // 2], [0, 0, 1]]).astype(\n            np.float32\n        )\n\n    def get_ray_map(self, c2w, h, w, intrinsics=None):\n        if intrinsics is None:\n            intrinsics = self.generate_pseudo_intrinsics(h, w)\n        i, j = np.meshgrid(np.arange(w), np.arange(h), indexing=\"xy\")\n        grid = np.stack([i, j, np.ones_like(i)], axis=-1)\n        ro = c2w[:3, 3]\n        rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n        rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n        rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n        ro = np.broadcast_to(ro, (h, w, 3))\n        ray_map = np.concatenate([ro, rd], axis=-1)\n        return ray_map\n\n    def set_camera_loc(camera, pose, K):\n        \"\"\"\n        pose: 4x4 matrix\n        K: 3x3 matrix\n        \"\"\"\n        fx, fy = K[0, 0], K[1, 1]","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.get_ray_map","uri":"program://Human3R/function/viser_utils.get_ray_map#L789-L800","kind":"function","name":"get_ray_map","path":"viser_utils.py","language":"python","start_line":789,"end_line":800,"context_start_line":769,"context_end_line":820,"code":"                np.concatenate(\n                    [tf.SO3(camera.wxyz).as_matrix(), camera.position[:, None]], 1\n                ),\n                [[0, 0, 0, 1]],\n            ],\n            0,\n        )\n        return CameraState(\n            fov=camera.fov,\n            aspect=camera.aspect,\n            c2w=c2w,\n        )\n\n    @staticmethod\n    def generate_pseudo_intrinsics(h, w):\n        focal = (h**2 + w**2) ** 0.5\n        return np.array([[focal, 0, w // 2], [0, focal, h // 2], [0, 0, 1]]).astype(\n            np.float32\n        )\n\n    def get_ray_map(self, c2w, h, w, intrinsics=None):\n        if intrinsics is None:\n            intrinsics = self.generate_pseudo_intrinsics(h, w)\n        i, j = np.meshgrid(np.arange(w), np.arange(h), indexing=\"xy\")\n        grid = np.stack([i, j, np.ones_like(i)], axis=-1)\n        ro = c2w[:3, 3]\n        rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n        rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n        rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n        ro = np.broadcast_to(ro, (h, w, 3))\n        ray_map = np.concatenate([ro, rd], axis=-1)\n        return ray_map\n\n    def set_camera_loc(camera, pose, K):\n        \"\"\"\n        pose: 4x4 matrix\n        K: 3x3 matrix\n        \"\"\"\n        fx, fy = K[0, 0], K[1, 1]\n        cx, cy = K[0, 2], K[1, 2]\n        aspect = float(cx) / float(cy)\n        fov = 2 * np.arctan(2 * cx / fx)\n        wxyz_xyz = tf.SE3.from_matrix(pose).wxyz_xyz\n        wxyz = wxyz_xyz[:4]\n        xyz = wxyz_xyz[4:]\n        camera.wxyz = wxyz\n        camera.position = xyz\n        camera.fov = fov\n        camera.aspect = aspect\n\n    def _connect_client(self, client: viser.ClientHandle):\n        from src.dust3r.inference import inference_step","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.set_camera_loc","uri":"program://Human3R/function/viser_utils.set_camera_loc#L802-L817","kind":"function","name":"set_camera_loc","path":"viser_utils.py","language":"python","start_line":802,"end_line":817,"context_start_line":782,"context_end_line":837,"code":"    @staticmethod\n    def generate_pseudo_intrinsics(h, w):\n        focal = (h**2 + w**2) ** 0.5\n        return np.array([[focal, 0, w // 2], [0, focal, h // 2], [0, 0, 1]]).astype(\n            np.float32\n        )\n\n    def get_ray_map(self, c2w, h, w, intrinsics=None):\n        if intrinsics is None:\n            intrinsics = self.generate_pseudo_intrinsics(h, w)\n        i, j = np.meshgrid(np.arange(w), np.arange(h), indexing=\"xy\")\n        grid = np.stack([i, j, np.ones_like(i)], axis=-1)\n        ro = c2w[:3, 3]\n        rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n        rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n        rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n        ro = np.broadcast_to(ro, (h, w, 3))\n        ray_map = np.concatenate([ro, rd], axis=-1)\n        return ray_map\n\n    def set_camera_loc(camera, pose, K):\n        \"\"\"\n        pose: 4x4 matrix\n        K: 3x3 matrix\n        \"\"\"\n        fx, fy = K[0, 0], K[1, 1]\n        cx, cy = K[0, 2], K[1, 2]\n        aspect = float(cx) / float(cy)\n        fov = 2 * np.arctan(2 * cx / fx)\n        wxyz_xyz = tf.SE3.from_matrix(pose).wxyz_xyz\n        wxyz = wxyz_xyz[:4]\n        xyz = wxyz_xyz[4:]\n        camera.wxyz = wxyz\n        camera.position = xyz\n        camera.fov = fov\n        camera.aspect = aspect\n\n    def _connect_client(self, client: viser.ClientHandle):\n        from src.dust3r.inference import inference_step\n        from src.dust3r.utils.geometry import geotrf\n\n        wxyz_panel = client.gui.add_text(\"wxyz:\", f\"{client.camera.wxyz}\")\n        position_panel = client.gui.add_text(\"position:\", f\"{client.camera.position}\")\n        fov_panel = client.gui.add_text(\n            \"fov:\", f\"{2 * np.arctan(self.size/self.focal_slider.value) * 180 / np.pi}\"\n        )\n        aspect_panel = client.gui.add_text(\"aspect:\", \"1.0\")\n\n        @client.camera.on_update\n        def _(_: viser.CameraHandle):\n            with self.server.atomic():\n                wxyz_panel.value = f\"{client.camera.wxyz}\"\n                position_panel.value = f\"{client.camera.position}\"\n                fov_panel.value = (\n                    f\"{2 * np.arctan(self.size/self.focal_slider.value) * 180 / np.pi}\"\n                )","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._connect_client","uri":"program://Human3R/function/viser_utils._connect_client#L819-L838","kind":"function","name":"_connect_client","path":"viser_utils.py","language":"python","start_line":819,"end_line":838,"context_start_line":799,"context_end_line":858,"code":"        ray_map = np.concatenate([ro, rd], axis=-1)\n        return ray_map\n\n    def set_camera_loc(camera, pose, K):\n        \"\"\"\n        pose: 4x4 matrix\n        K: 3x3 matrix\n        \"\"\"\n        fx, fy = K[0, 0], K[1, 1]\n        cx, cy = K[0, 2], K[1, 2]\n        aspect = float(cx) / float(cy)\n        fov = 2 * np.arctan(2 * cx / fx)\n        wxyz_xyz = tf.SE3.from_matrix(pose).wxyz_xyz\n        wxyz = wxyz_xyz[:4]\n        xyz = wxyz_xyz[4:]\n        camera.wxyz = wxyz\n        camera.position = xyz\n        camera.fov = fov\n        camera.aspect = aspect\n\n    def _connect_client(self, client: viser.ClientHandle):\n        from src.dust3r.inference import inference_step\n        from src.dust3r.utils.geometry import geotrf\n\n        wxyz_panel = client.gui.add_text(\"wxyz:\", f\"{client.camera.wxyz}\")\n        position_panel = client.gui.add_text(\"position:\", f\"{client.camera.position}\")\n        fov_panel = client.gui.add_text(\n            \"fov:\", f\"{2 * np.arctan(self.size/self.focal_slider.value) * 180 / np.pi}\"\n        )\n        aspect_panel = client.gui.add_text(\"aspect:\", \"1.0\")\n\n        @client.camera.on_update\n        def _(_: viser.CameraHandle):\n            with self.server.atomic():\n                wxyz_panel.value = f\"{client.camera.wxyz}\"\n                position_panel.value = f\"{client.camera.position}\"\n                fov_panel.value = (\n                    f\"{2 * np.arctan(self.size/self.focal_slider.value) * 180 / np.pi}\"\n                )\n                aspect_panel.value = \"1.0\"\n\n\n    @staticmethod\n    def set_color_border(image, border_width=5, color=[1, 0, 0]):\n\n        image[:border_width, :, 0] = color[0]  # Red channel\n        image[:border_width, :, 1] = color[1]  # Green channel\n        image[:border_width, :, 2] = color[2]  # Blue channel\n        image[-border_width:, :, 0] = color[0]\n        image[-border_width:, :, 1] = color[1]\n        image[-border_width:, :, 2] = color[2]\n\n        image[:, :border_width, 0] = color[0]\n        image[:, :border_width, 1] = color[1]\n        image[:, :border_width, 2] = color[2]\n        image[:, -border_width:, 0] = color[0]\n        image[:, -border_width:, 1] = color[1]\n        image[:, -border_width:, 2] = color[2]\n\n        return image","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.set_color_border","uri":"program://Human3R/function/viser_utils.set_color_border#L842-L858","kind":"function","name":"set_color_border","path":"viser_utils.py","language":"python","start_line":842,"end_line":858,"context_start_line":822,"context_end_line":878,"code":"\n        wxyz_panel = client.gui.add_text(\"wxyz:\", f\"{client.camera.wxyz}\")\n        position_panel = client.gui.add_text(\"position:\", f\"{client.camera.position}\")\n        fov_panel = client.gui.add_text(\n            \"fov:\", f\"{2 * np.arctan(self.size/self.focal_slider.value) * 180 / np.pi}\"\n        )\n        aspect_panel = client.gui.add_text(\"aspect:\", \"1.0\")\n\n        @client.camera.on_update\n        def _(_: viser.CameraHandle):\n            with self.server.atomic():\n                wxyz_panel.value = f\"{client.camera.wxyz}\"\n                position_panel.value = f\"{client.camera.position}\"\n                fov_panel.value = (\n                    f\"{2 * np.arctan(self.size/self.focal_slider.value) * 180 / np.pi}\"\n                )\n                aspect_panel.value = \"1.0\"\n\n\n    @staticmethod\n    def set_color_border(image, border_width=5, color=[1, 0, 0]):\n\n        image[:border_width, :, 0] = color[0]  # Red channel\n        image[:border_width, :, 1] = color[1]  # Green channel\n        image[:border_width, :, 2] = color[2]  # Blue channel\n        image[-border_width:, :, 0] = color[0]\n        image[-border_width:, :, 1] = color[1]\n        image[-border_width:, :, 2] = color[2]\n\n        image[:, :border_width, 0] = color[0]\n        image[:, :border_width, 1] = color[1]\n        image[:, :border_width, 2] = color[2]\n        image[:, -border_width:, 0] = color[0]\n        image[:, -border_width:, 1] = color[1]\n        image[:, -border_width:, 2] = color[2]\n\n        return image\n\n    def read_data(\n            self, \n            pc_list, \n            color_list, \n            conf_list,\n            msk_list,\n            all_smpl_verts, \n            smpl_faces, \n            smpl_id,\n            edge_color_list=None,\n            gt_smpl_verts=None):\n        pcs = {}\n        step_list = []\n        for i, pc in enumerate(pc_list):\n            step = i\n            pcs.update(\n                {\n                    step: {\n                        \"pc\": pc,","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.read_data","uri":"program://Human3R/function/viser_utils.read_data#L860-L901","kind":"function","name":"read_data","path":"viser_utils.py","language":"python","start_line":860,"end_line":901,"context_start_line":840,"context_end_line":921,"code":"\n    @staticmethod\n    def set_color_border(image, border_width=5, color=[1, 0, 0]):\n\n        image[:border_width, :, 0] = color[0]  # Red channel\n        image[:border_width, :, 1] = color[1]  # Green channel\n        image[:border_width, :, 2] = color[2]  # Blue channel\n        image[-border_width:, :, 0] = color[0]\n        image[-border_width:, :, 1] = color[1]\n        image[-border_width:, :, 2] = color[2]\n\n        image[:, :border_width, 0] = color[0]\n        image[:, :border_width, 1] = color[1]\n        image[:, :border_width, 2] = color[2]\n        image[:, -border_width:, 0] = color[0]\n        image[:, -border_width:, 1] = color[1]\n        image[:, -border_width:, 2] = color[2]\n\n        return image\n\n    def read_data(\n            self, \n            pc_list, \n            color_list, \n            conf_list,\n            msk_list,\n            all_smpl_verts, \n            smpl_faces, \n            smpl_id,\n            edge_color_list=None,\n            gt_smpl_verts=None):\n        pcs = {}\n        step_list = []\n        for i, pc in enumerate(pc_list):\n            step = i\n            pcs.update(\n                {\n                    step: {\n                        \"pc\": pc,\n                        \"color\": color_list[i],\n                        \"conf\": conf_list[i],\n                        \"msk\": msk_list[i],\n                        \"verts\": all_smpl_verts[i],\n                        \"faces\": smpl_faces,\n                        \"smpl_id\": smpl_id[i],\n                        \"edge_color\": (\n                            None if edge_color_list[i] is None else edge_color_list[i]\n                        ),\n                        \"gt_verts\": (\n                            None if gt_smpl_verts is None else gt_smpl_verts[i]\n                        ),\n                    }\n                }\n            )\n            step_list.append(step)\n        normalized_indices = (\n            np.array(list(range(len(pc_list))))\n            / np.array(list(range(len(pc_list)))).max()\n        )\n        cmap = cm.viridis\n        self.camera_colors = cmap(normalized_indices)\n        return pcs, step_list\n\n    def parse_pc_data(\n        self,\n        pc,\n        color,\n        conf=None,\n        msk=None,\n        edge_color=[0.251, 0.702, 0.902],\n        set_border_color=False,\n        downsample_factor=1,\n    ):\n\n        pred_pts = pc.reshape(-1, 3)  # [N, 3]\n\n        if set_border_color and edge_color is not None:\n            color = self.set_color_border(color[0], color=edge_color)\n        if np.isnan(color).any():\n            color = np.zeros((pred_pts.shape[0], 3))\n            color[:, 2] = 1\n        else:","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.parse_pc_data","uri":"program://Human3R/function/viser_utils.parse_pc_data#L903-L965","kind":"function","name":"parse_pc_data","path":"viser_utils.py","language":"python","start_line":903,"end_line":965,"context_start_line":883,"context_end_line":985,"code":"                        \"faces\": smpl_faces,\n                        \"smpl_id\": smpl_id[i],\n                        \"edge_color\": (\n                            None if edge_color_list[i] is None else edge_color_list[i]\n                        ),\n                        \"gt_verts\": (\n                            None if gt_smpl_verts is None else gt_smpl_verts[i]\n                        ),\n                    }\n                }\n            )\n            step_list.append(step)\n        normalized_indices = (\n            np.array(list(range(len(pc_list))))\n            / np.array(list(range(len(pc_list)))).max()\n        )\n        cmap = cm.viridis\n        self.camera_colors = cmap(normalized_indices)\n        return pcs, step_list\n\n    def parse_pc_data(\n        self,\n        pc,\n        color,\n        conf=None,\n        msk=None,\n        edge_color=[0.251, 0.702, 0.902],\n        set_border_color=False,\n        downsample_factor=1,\n    ):\n\n        pred_pts = pc.reshape(-1, 3)  # [N, 3]\n\n        if set_border_color and edge_color is not None:\n            color = self.set_color_border(color[0], color=edge_color)\n        if np.isnan(color).any():\n            color = np.zeros((pred_pts.shape[0], 3))\n            color[:, 2] = 1\n        else:\n            color = color.reshape(-1, 3)\n\n        if msk is not None:\n            msk_2d = msk[0].copy()\n            fg_mask = msk_2d > self.msk_threshold\n            \n            if abs(self.mask_morph) > 0:\n                if self.mask_morph > 0:\n                    fg_mask = binary_dilation(fg_mask, disk(abs(self.mask_morph)))\n                else:\n                    fg_mask = binary_erosion(fg_mask, disk(abs(self.mask_morph)))\n            \n            fg_mask = fg_mask.reshape(-1)\n            \n            if self.show_foreground and self.show_background:\n                display_mask = np.ones_like(fg_mask, dtype=bool)\n            elif self.show_foreground and not self.show_background:\n                display_mask = fg_mask\n            elif not self.show_foreground and self.show_background:\n                display_mask = ~fg_mask\n            else:\n                return np.array([]).reshape(0, 3), np.array([]).reshape(0, 3)\n            \n            if conf is not None:\n                conf = conf[0].reshape(-1)\n                final_mask = display_mask & (conf > self.vis_threshold)\n            else:\n                final_mask = display_mask\n                \n            pred_pts = pred_pts[final_mask]\n            color = color[final_mask]\n            \n        elif conf is not None:\n            conf = conf[0].reshape(-1)\n            pred_pts = pred_pts[conf > self.vis_threshold]\n            color = color[conf > self.vis_threshold]\n\n        # apply downsampling\n        if downsample_factor > 1 and len(pred_pts) > 0:\n            indices = np.arange(0, len(pred_pts), downsample_factor)\n            pred_pts = pred_pts[indices]\n            color = color[indices]\n\n        return pred_pts, color\n\n    def _update_point_clouds(self):\n        # prevent duplicate update\n        if self._updating_point_clouds:\n            return\n        \n        self._updating_point_clouds = True\n        try:\n            # safely clear existing point cloud display\n            for handle in self.pc_handles:\n                try:\n                    handle.remove()\n                except (KeyError, AttributeError):\n                    # ignore already deleted or non-existent handles\n                    pass\n            self.pc_handles.clear()\n            \n            for i, step in enumerate(self.all_steps):\n                if hasattr(self, 'frame_nodes') and i < len(self.frame_nodes):\n                    self._add_pc_for_step(step)","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._update_point_clouds","uri":"program://Human3R/function/viser_utils._update_point_clouds#L967-L987","kind":"function","name":"_update_point_clouds","path":"viser_utils.py","language":"python","start_line":967,"end_line":987,"context_start_line":947,"context_end_line":1007,"code":"                final_mask = display_mask & (conf > self.vis_threshold)\n            else:\n                final_mask = display_mask\n                \n            pred_pts = pred_pts[final_mask]\n            color = color[final_mask]\n            \n        elif conf is not None:\n            conf = conf[0].reshape(-1)\n            pred_pts = pred_pts[conf > self.vis_threshold]\n            color = color[conf > self.vis_threshold]\n\n        # apply downsampling\n        if downsample_factor > 1 and len(pred_pts) > 0:\n            indices = np.arange(0, len(pred_pts), downsample_factor)\n            pred_pts = pred_pts[indices]\n            color = color[indices]\n\n        return pred_pts, color\n\n    def _update_point_clouds(self):\n        # prevent duplicate update\n        if self._updating_point_clouds:\n            return\n        \n        self._updating_point_clouds = True\n        try:\n            # safely clear existing point cloud display\n            for handle in self.pc_handles:\n                try:\n                    handle.remove()\n                except (KeyError, AttributeError):\n                    # ignore already deleted or non-existent handles\n                    pass\n            self.pc_handles.clear()\n            \n            for i, step in enumerate(self.all_steps):\n                if hasattr(self, 'frame_nodes') and i < len(self.frame_nodes):\n                    self._add_pc_for_step(step)\n        finally:\n            self._updating_point_clouds = False\n\n    def _add_pc_for_step(self, step):\n        pc = self.pcs[step][\"pc\"]\n        color = self.pcs[step][\"color\"]\n        conf = self.pcs[step][\"conf\"]\n        msk = self.pcs[step][\"msk\"]\n        edge_color = self.pcs[step].get(\"edge_color\", None)\n\n        pred_pts, color = self.parse_pc_data(\n            pc, color, conf, msk, edge_color, set_border_color=True,\n            downsample_factor=self.downsample_slider.value\n        )\n\n        self.pc_handles.append(\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts\",\n                points=pred_pts,\n                colors=color,\n                point_size=0.005,\n            )","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._add_pc_for_step","uri":"program://Human3R/function/viser_utils._add_pc_for_step#L989-L1008","kind":"function","name":"_add_pc_for_step","path":"viser_utils.py","language":"python","start_line":989,"end_line":1008,"context_start_line":969,"context_end_line":1028,"code":"        if self._updating_point_clouds:\n            return\n        \n        self._updating_point_clouds = True\n        try:\n            # safely clear existing point cloud display\n            for handle in self.pc_handles:\n                try:\n                    handle.remove()\n                except (KeyError, AttributeError):\n                    # ignore already deleted or non-existent handles\n                    pass\n            self.pc_handles.clear()\n            \n            for i, step in enumerate(self.all_steps):\n                if hasattr(self, 'frame_nodes') and i < len(self.frame_nodes):\n                    self._add_pc_for_step(step)\n        finally:\n            self._updating_point_clouds = False\n\n    def _add_pc_for_step(self, step):\n        pc = self.pcs[step][\"pc\"]\n        color = self.pcs[step][\"color\"]\n        conf = self.pcs[step][\"conf\"]\n        msk = self.pcs[step][\"msk\"]\n        edge_color = self.pcs[step].get(\"edge_color\", None)\n\n        pred_pts, color = self.parse_pc_data(\n            pc, color, conf, msk, edge_color, set_border_color=True,\n            downsample_factor=self.downsample_slider.value\n        )\n\n        self.pc_handles.append(\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts\",\n                points=pred_pts,\n                colors=color,\n                point_size=0.005,\n            )\n        )\n\n    def _compute_opacity_for_index(self, step_index):\n        \"\"\"Compute opacity so that earlier frames are more transparent, later frames more opaque.\n        - step_index = 0   => opacity close to min_opacity (most transparent)\n        - step_index = N-1 => opacity close to 1.0 (most opaque)\n        \"\"\"\n        if not self.mesh_time_opacity_checkbox.value or self.num_frames <= 1:\n            return None\n        ratio = step_index / (self.num_frames - 1)\n        min_opacity = float(self.min_mesh_opacity_slider.value)\n        return min_opacity + ratio * (1.0 - min_opacity)\n\n    def _update_mesh_opacities(self):\n        # If dynamic opacity is enabled, avoid overriding dynamic values\n        if self.dynamic_opacity_checkbox.value:\n            return\n        # Update predicted meshes\n        for mesh_idx, handle in enumerate(self.mesh_handles):\n            if mesh_idx < len(self.mesh_step_mapping):\n                step = self.mesh_step_mapping[mesh_idx]","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._compute_opacity_for_index","uri":"program://Human3R/function/viser_utils._compute_opacity_for_index#L1010-L1019","kind":"function","name":"_compute_opacity_for_index","path":"viser_utils.py","language":"python","start_line":1010,"end_line":1019,"context_start_line":990,"context_end_line":1039,"code":"        pc = self.pcs[step][\"pc\"]\n        color = self.pcs[step][\"color\"]\n        conf = self.pcs[step][\"conf\"]\n        msk = self.pcs[step][\"msk\"]\n        edge_color = self.pcs[step].get(\"edge_color\", None)\n\n        pred_pts, color = self.parse_pc_data(\n            pc, color, conf, msk, edge_color, set_border_color=True,\n            downsample_factor=self.downsample_slider.value\n        )\n\n        self.pc_handles.append(\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts\",\n                points=pred_pts,\n                colors=color,\n                point_size=0.005,\n            )\n        )\n\n    def _compute_opacity_for_index(self, step_index):\n        \"\"\"Compute opacity so that earlier frames are more transparent, later frames more opaque.\n        - step_index = 0   => opacity close to min_opacity (most transparent)\n        - step_index = N-1 => opacity close to 1.0 (most opaque)\n        \"\"\"\n        if not self.mesh_time_opacity_checkbox.value or self.num_frames <= 1:\n            return None\n        ratio = step_index / (self.num_frames - 1)\n        min_opacity = float(self.min_mesh_opacity_slider.value)\n        return min_opacity + ratio * (1.0 - min_opacity)\n\n    def _update_mesh_opacities(self):\n        # If dynamic opacity is enabled, avoid overriding dynamic values\n        if self.dynamic_opacity_checkbox.value:\n            return\n        # Update predicted meshes\n        for mesh_idx, handle in enumerate(self.mesh_handles):\n            if mesh_idx < len(self.mesh_step_mapping):\n                step = self.mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                handle.opacity = self._compute_opacity_for_index(step_index)\n        # Update GT meshes\n        for mesh_idx, handle in enumerate(self.gt_mesh_handles):\n            if mesh_idx < len(self.gt_mesh_step_mapping):\n                step = self.gt_mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                handle.opacity = self._compute_opacity_for_index(step_index)\n\n    def _compute_dynamic_opacity(self, age: int):\n        \"\"\"Compute dynamic opacity by age (0 for current frame).","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._update_mesh_opacities","uri":"program://Human3R/function/viser_utils._update_mesh_opacities#L1021-L1036","kind":"function","name":"_update_mesh_opacities","path":"viser_utils.py","language":"python","start_line":1021,"end_line":1036,"context_start_line":1001,"context_end_line":1056,"code":"        self.pc_handles.append(\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts\",\n                points=pred_pts,\n                colors=color,\n                point_size=0.005,\n            )\n        )\n\n    def _compute_opacity_for_index(self, step_index):\n        \"\"\"Compute opacity so that earlier frames are more transparent, later frames more opaque.\n        - step_index = 0   => opacity close to min_opacity (most transparent)\n        - step_index = N-1 => opacity close to 1.0 (most opaque)\n        \"\"\"\n        if not self.mesh_time_opacity_checkbox.value or self.num_frames <= 1:\n            return None\n        ratio = step_index / (self.num_frames - 1)\n        min_opacity = float(self.min_mesh_opacity_slider.value)\n        return min_opacity + ratio * (1.0 - min_opacity)\n\n    def _update_mesh_opacities(self):\n        # If dynamic opacity is enabled, avoid overriding dynamic values\n        if self.dynamic_opacity_checkbox.value:\n            return\n        # Update predicted meshes\n        for mesh_idx, handle in enumerate(self.mesh_handles):\n            if mesh_idx < len(self.mesh_step_mapping):\n                step = self.mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                handle.opacity = self._compute_opacity_for_index(step_index)\n        # Update GT meshes\n        for mesh_idx, handle in enumerate(self.gt_mesh_handles):\n            if mesh_idx < len(self.gt_mesh_step_mapping):\n                step = self.gt_mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                handle.opacity = self._compute_opacity_for_index(step_index)\n\n    def _compute_dynamic_opacity(self, age: int):\n        \"\"\"Compute dynamic opacity by age (0 for current frame).\n        Linear decay to min opacity over decay_len frames.\n        \"\"\"\n        min_opacity = float(self.min_mesh_opacity_slider.value)\n        decay_len = max(1, int(self.opacity_decay_len_slider.value))\n        if age <= 0:\n            return 1.0\n        ratio = min(1.0, age / decay_len)\n        return 1.0 - ratio * (1.0 - min_opacity)\n\n    def _update_dynamic_opacities(self, current_step_index: int, force_update: bool = False):\n        if not self.dynamic_opacity_checkbox.value:\n            return\n            \n        # Performance optimization: skip updates based on throttle setting\n        if not force_update and hasattr(self, '_opacity_update_throttle'):\n            if self._last_opacity_update_step == current_step_index:\n                return  # Already updated for this step","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._compute_dynamic_opacity","uri":"program://Human3R/function/viser_utils._compute_dynamic_opacity#L1038-L1047","kind":"function","name":"_compute_dynamic_opacity","path":"viser_utils.py","language":"python","start_line":1038,"end_line":1047,"context_start_line":1018,"context_end_line":1067,"code":"        min_opacity = float(self.min_mesh_opacity_slider.value)\n        return min_opacity + ratio * (1.0 - min_opacity)\n\n    def _update_mesh_opacities(self):\n        # If dynamic opacity is enabled, avoid overriding dynamic values\n        if self.dynamic_opacity_checkbox.value:\n            return\n        # Update predicted meshes\n        for mesh_idx, handle in enumerate(self.mesh_handles):\n            if mesh_idx < len(self.mesh_step_mapping):\n                step = self.mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                handle.opacity = self._compute_opacity_for_index(step_index)\n        # Update GT meshes\n        for mesh_idx, handle in enumerate(self.gt_mesh_handles):\n            if mesh_idx < len(self.gt_mesh_step_mapping):\n                step = self.gt_mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                handle.opacity = self._compute_opacity_for_index(step_index)\n\n    def _compute_dynamic_opacity(self, age: int):\n        \"\"\"Compute dynamic opacity by age (0 for current frame).\n        Linear decay to min opacity over decay_len frames.\n        \"\"\"\n        min_opacity = float(self.min_mesh_opacity_slider.value)\n        decay_len = max(1, int(self.opacity_decay_len_slider.value))\n        if age <= 0:\n            return 1.0\n        ratio = min(1.0, age / decay_len)\n        return 1.0 - ratio * (1.0 - min_opacity)\n\n    def _update_dynamic_opacities(self, current_step_index: int, force_update: bool = False):\n        if not self.dynamic_opacity_checkbox.value:\n            return\n            \n        # Performance optimization: skip updates based on throttle setting\n        if not force_update and hasattr(self, '_opacity_update_throttle'):\n            if self._last_opacity_update_step == current_step_index:\n                return  # Already updated for this step\n                \n            if self._opacity_update_throttle > 0:\n                if self._opacity_frame_counter < self._opacity_update_throttle:\n                    self._opacity_frame_counter += 1\n                    return\n                else:\n                    self._opacity_frame_counter = 0  # Reset counter\n        \n        self._last_opacity_update_step = current_step_index\n        \n        # predicted meshes","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._update_dynamic_opacities","uri":"program://Human3R/function/viser_utils._update_dynamic_opacities#L1049-L1087","kind":"function","name":"_update_dynamic_opacities","path":"viser_utils.py","language":"python","start_line":1049,"end_line":1087,"context_start_line":1029,"context_end_line":1107,"code":"                step_index = self.step_to_index.get(step, 0)\n                handle.opacity = self._compute_opacity_for_index(step_index)\n        # Update GT meshes\n        for mesh_idx, handle in enumerate(self.gt_mesh_handles):\n            if mesh_idx < len(self.gt_mesh_step_mapping):\n                step = self.gt_mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                handle.opacity = self._compute_opacity_for_index(step_index)\n\n    def _compute_dynamic_opacity(self, age: int):\n        \"\"\"Compute dynamic opacity by age (0 for current frame).\n        Linear decay to min opacity over decay_len frames.\n        \"\"\"\n        min_opacity = float(self.min_mesh_opacity_slider.value)\n        decay_len = max(1, int(self.opacity_decay_len_slider.value))\n        if age <= 0:\n            return 1.0\n        ratio = min(1.0, age / decay_len)\n        return 1.0 - ratio * (1.0 - min_opacity)\n\n    def _update_dynamic_opacities(self, current_step_index: int, force_update: bool = False):\n        if not self.dynamic_opacity_checkbox.value:\n            return\n            \n        # Performance optimization: skip updates based on throttle setting\n        if not force_update and hasattr(self, '_opacity_update_throttle'):\n            if self._last_opacity_update_step == current_step_index:\n                return  # Already updated for this step\n                \n            if self._opacity_update_throttle > 0:\n                if self._opacity_frame_counter < self._opacity_update_throttle:\n                    self._opacity_frame_counter += 1\n                    return\n                else:\n                    self._opacity_frame_counter = 0  # Reset counter\n        \n        self._last_opacity_update_step = current_step_index\n        \n        # predicted meshes\n        for mesh_idx, handle in enumerate(self.mesh_handles):\n            if mesh_idx < len(self.mesh_step_mapping):\n                step = self.mesh_step_mapping[mesh_idx]\n                idx = self.step_to_index.get(step, 0)\n                age = current_step_index - idx\n                if age < 0:\n                    # future frames (if visible) keep opaque\n                    handle.opacity = 1.0\n                else:\n                    handle.opacity = self._compute_dynamic_opacity(age)\n        # GT meshes\n        for mesh_idx, handle in enumerate(self.gt_mesh_handles):\n            if mesh_idx < len(self.gt_mesh_step_mapping):\n                step = self.gt_mesh_step_mapping[mesh_idx]\n                idx = self.step_to_index.get(step, 0)\n                age = current_step_index - idx\n                if age < 0:\n                    handle.opacity = 1.0\n                else:\n                    handle.opacity = self._compute_dynamic_opacity(age)\n\n    def _update_smpl_visibility(self):\n        \"\"\"Update SMPL mesh visibility based on downsampling factor\"\"\"\n        downsample_factor = int(self.smpl_downsample_slider.value)\n        \n        # Update predicted meshes\n        for mesh_idx, handle in enumerate(self.mesh_handles):\n            if mesh_idx < len(self.mesh_step_mapping):\n                step = self.mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                # Only show meshes at intervals defined by downsample_factor\n                should_show = (step_index % downsample_factor == 0)\n                # Only hide if SMPL is enabled and this mesh should be hidden\n                if self.gui_show_smpl.value:\n                    handle.visible = should_show\n                else:\n                    handle.visible = False\n        \n        # Update GT meshes\n        for mesh_idx, handle in enumerate(self.gt_mesh_handles):","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._update_smpl_visibility","uri":"program://Human3R/function/viser_utils._update_smpl_visibility#L1089-L1117","kind":"function","name":"_update_smpl_visibility","path":"viser_utils.py","language":"python","start_line":1089,"end_line":1117,"context_start_line":1069,"context_end_line":1137,"code":"            if mesh_idx < len(self.mesh_step_mapping):\n                step = self.mesh_step_mapping[mesh_idx]\n                idx = self.step_to_index.get(step, 0)\n                age = current_step_index - idx\n                if age < 0:\n                    # future frames (if visible) keep opaque\n                    handle.opacity = 1.0\n                else:\n                    handle.opacity = self._compute_dynamic_opacity(age)\n        # GT meshes\n        for mesh_idx, handle in enumerate(self.gt_mesh_handles):\n            if mesh_idx < len(self.gt_mesh_step_mapping):\n                step = self.gt_mesh_step_mapping[mesh_idx]\n                idx = self.step_to_index.get(step, 0)\n                age = current_step_index - idx\n                if age < 0:\n                    handle.opacity = 1.0\n                else:\n                    handle.opacity = self._compute_dynamic_opacity(age)\n\n    def _update_smpl_visibility(self):\n        \"\"\"Update SMPL mesh visibility based on downsampling factor\"\"\"\n        downsample_factor = int(self.smpl_downsample_slider.value)\n        \n        # Update predicted meshes\n        for mesh_idx, handle in enumerate(self.mesh_handles):\n            if mesh_idx < len(self.mesh_step_mapping):\n                step = self.mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                # Only show meshes at intervals defined by downsample_factor\n                should_show = (step_index % downsample_factor == 0)\n                # Only hide if SMPL is enabled and this mesh should be hidden\n                if self.gui_show_smpl.value:\n                    handle.visible = should_show\n                else:\n                    handle.visible = False\n        \n        # Update GT meshes\n        for mesh_idx, handle in enumerate(self.gt_mesh_handles):\n            if mesh_idx < len(self.gt_mesh_step_mapping):\n                step = self.gt_mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                # Only show meshes at intervals defined by downsample_factor\n                should_show = (step_index % downsample_factor == 0)\n                # Only hide if GT SMPL is enabled and this mesh should be hidden\n                if self.show_gt_smpl:\n                    handle.visible = should_show\n                else:\n                    handle.visible = False\n\n    def add_pc(self, step):\n        pc = self.pcs[step][\"pc\"]\n        color = self.pcs[step][\"color\"]\n        conf = self.pcs[step][\"conf\"]\n        msk = self.pcs[step][\"msk\"]\n        verts = self.pcs[step][\"verts\"]\n        faces = self.pcs[step][\"faces\"]\n        smpl_id = self.pcs[step][\"smpl_id\"]\n        edge_color = self.pcs[step].get(\"edge_color\", None)\n\n        pred_pts, color = self.parse_pc_data(\n            pc, color, conf, msk, edge_color, set_border_color=True,\n            downsample_factor=self.downsample_slider.value\n        )\n\n        self.vis_pts_list.append(pred_pts)\n        self.pc_handles.append(\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts\",","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.add_pc","uri":"program://Human3R/function/viser_utils.add_pc#L1119-L1161","kind":"function","name":"add_pc","path":"viser_utils.py","language":"python","start_line":1119,"end_line":1161,"context_start_line":1099,"context_end_line":1181,"code":"                should_show = (step_index % downsample_factor == 0)\n                # Only hide if SMPL is enabled and this mesh should be hidden\n                if self.gui_show_smpl.value:\n                    handle.visible = should_show\n                else:\n                    handle.visible = False\n        \n        # Update GT meshes\n        for mesh_idx, handle in enumerate(self.gt_mesh_handles):\n            if mesh_idx < len(self.gt_mesh_step_mapping):\n                step = self.gt_mesh_step_mapping[mesh_idx]\n                step_index = self.step_to_index.get(step, 0)\n                # Only show meshes at intervals defined by downsample_factor\n                should_show = (step_index % downsample_factor == 0)\n                # Only hide if GT SMPL is enabled and this mesh should be hidden\n                if self.show_gt_smpl:\n                    handle.visible = should_show\n                else:\n                    handle.visible = False\n\n    def add_pc(self, step):\n        pc = self.pcs[step][\"pc\"]\n        color = self.pcs[step][\"color\"]\n        conf = self.pcs[step][\"conf\"]\n        msk = self.pcs[step][\"msk\"]\n        verts = self.pcs[step][\"verts\"]\n        faces = self.pcs[step][\"faces\"]\n        smpl_id = self.pcs[step][\"smpl_id\"]\n        edge_color = self.pcs[step].get(\"edge_color\", None)\n\n        pred_pts, color = self.parse_pc_data(\n            pc, color, conf, msk, edge_color, set_border_color=True,\n            downsample_factor=self.downsample_slider.value\n        )\n\n        self.vis_pts_list.append(pred_pts)\n        self.pc_handles.append(\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts\",\n                points=pred_pts,\n                colors=color,\n                point_size=0.005,\n            )\n        )\n        if len(verts) > 0:\n            for tid, vert in enumerate(verts):\n                step_idx = self.step_to_index.get(step, 0)\n                mesh_opacity = self._compute_opacity_for_index(step_idx)\n                mesh_handle = self.server.scene.add_mesh_simple(\n                    name=f\"/frames/{step}/human_{tid}\",\n                    vertices=vert,\n                    faces=faces,\n                    flat_shading=False,\n                    wireframe=False,\n                    opacity=mesh_opacity,\n                    color=get_color(smpl_id[tid]),\n                )\n                self.mesh_handles.append(mesh_handle)\n                self.mesh_step_mapping.append(step)\n        \n        # Add GT SMPL mesh if enabled\n        if self.show_gt_smpl:\n            self.add_gt_smpl(step)\n\n    def add_camera(self, step):\n        cam = self.cam_dict\n        focal = cam[\"focal\"][step]\n        pp = cam[\"pp\"][step]\n        R = cam[\"R\"][step]\n        t = cam[\"t\"][step]\n\n        q = tf.SO3.from_matrix(R).wxyz\n        fov = 2 * np.arctan(pp[0] / focal)\n        aspect = pp[0] / pp[1]\n        self.traj_list.append((q, t))\n        self.cam_handles.append(\n            self.server.add_camera_frustum(\n                name=f\"/frames/{step}/camera\",\n                fov=fov,\n                aspect=aspect,\n                wxyz=q,\n                position=t,\n                scale=0.1,","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.add_camera","uri":"program://Human3R/function/viser_utils.add_camera#L1163-L1184","kind":"function","name":"add_camera","path":"viser_utils.py","language":"python","start_line":1163,"end_line":1184,"context_start_line":1143,"context_end_line":1204,"code":"        if len(verts) > 0:\n            for tid, vert in enumerate(verts):\n                step_idx = self.step_to_index.get(step, 0)\n                mesh_opacity = self._compute_opacity_for_index(step_idx)\n                mesh_handle = self.server.scene.add_mesh_simple(\n                    name=f\"/frames/{step}/human_{tid}\",\n                    vertices=vert,\n                    faces=faces,\n                    flat_shading=False,\n                    wireframe=False,\n                    opacity=mesh_opacity,\n                    color=get_color(smpl_id[tid]),\n                )\n                self.mesh_handles.append(mesh_handle)\n                self.mesh_step_mapping.append(step)\n        \n        # Add GT SMPL mesh if enabled\n        if self.show_gt_smpl:\n            self.add_gt_smpl(step)\n\n    def add_camera(self, step):\n        cam = self.cam_dict\n        focal = cam[\"focal\"][step]\n        pp = cam[\"pp\"][step]\n        R = cam[\"R\"][step]\n        t = cam[\"t\"][step]\n\n        q = tf.SO3.from_matrix(R).wxyz\n        fov = 2 * np.arctan(pp[0] / focal)\n        aspect = pp[0] / pp[1]\n        self.traj_list.append((q, t))\n        self.cam_handles.append(\n            self.server.add_camera_frustum(\n                name=f\"/frames/{step}/camera\",\n                fov=fov,\n                aspect=aspect,\n                wxyz=q,\n                position=t,\n                scale=0.1,\n                color=tuple(self.camera_colors[step][:3]),\n            )\n        )\n\n    def add_gt_camera(self, step):\n        if self.gt_cam_dict is None:\n            return\n            \n        cam = self.gt_cam_dict\n        focal = cam[\"focal\"][step]\n        pp = cam[\"pp\"][step]\n        R = cam[\"R\"][step]\n        t = cam[\"t\"][step]\n\n        q = tf.SO3.from_matrix(R).wxyz\n        fov = 2 * np.arctan(pp[0] / focal)\n        aspect = pp[0] / pp[1]\n        self.gt_cam_handles.append(\n            self.server.add_camera_frustum(\n                name=f\"/frames/{step}/gt_camera\",\n                fov=fov,\n                aspect=aspect,\n                wxyz=q,","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.add_gt_camera","uri":"program://Human3R/function/viser_utils.add_gt_camera#L1186-L1209","kind":"function","name":"add_gt_camera","path":"viser_utils.py","language":"python","start_line":1186,"end_line":1209,"context_start_line":1166,"context_end_line":1229,"code":"        pp = cam[\"pp\"][step]\n        R = cam[\"R\"][step]\n        t = cam[\"t\"][step]\n\n        q = tf.SO3.from_matrix(R).wxyz\n        fov = 2 * np.arctan(pp[0] / focal)\n        aspect = pp[0] / pp[1]\n        self.traj_list.append((q, t))\n        self.cam_handles.append(\n            self.server.add_camera_frustum(\n                name=f\"/frames/{step}/camera\",\n                fov=fov,\n                aspect=aspect,\n                wxyz=q,\n                position=t,\n                scale=0.1,\n                color=tuple(self.camera_colors[step][:3]),\n            )\n        )\n\n    def add_gt_camera(self, step):\n        if self.gt_cam_dict is None:\n            return\n            \n        cam = self.gt_cam_dict\n        focal = cam[\"focal\"][step]\n        pp = cam[\"pp\"][step]\n        R = cam[\"R\"][step]\n        t = cam[\"t\"][step]\n\n        q = tf.SO3.from_matrix(R).wxyz\n        fov = 2 * np.arctan(pp[0] / focal)\n        aspect = pp[0] / pp[1]\n        self.gt_cam_handles.append(\n            self.server.add_camera_frustum(\n                name=f\"/frames/{step}/gt_camera\",\n                fov=fov,\n                aspect=aspect,\n                wxyz=q,\n                position=t,\n                scale=0.1,\n                color=(166, 166, 166),\n            )\n        )\n    \n    def _update_camera_visibility(self):\n        \"\"\"Update camera visibility based on downsample factor\"\"\"\n        if not hasattr(self, 'cam_handles') or not self.show_camera:\n            return\n            \n        camera_downsample_factor = int(self.camera_downsample_slider.value)\n        \n        for i, handle in enumerate(self.cam_handles):\n            if i < len(self.all_steps):\n                step_index = self.step_to_index.get(self.all_steps[i], i)\n                should_show = (step_index % camera_downsample_factor == 0)\n                handle.visible = should_show\n    \n    def _update_gt_camera_visibility(self):\n        \"\"\"Update GT camera visibility based on downsample factor\"\"\"\n        if not hasattr(self, 'gt_cam_handles') or not self.show_gt_camera:\n            return\n            \n        camera_downsample_factor = int(self.camera_downsample_slider.value)","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._update_camera_visibility","uri":"program://Human3R/function/viser_utils._update_camera_visibility#L1211-L1222","kind":"function","name":"_update_camera_visibility","path":"viser_utils.py","language":"python","start_line":1211,"end_line":1222,"context_start_line":1191,"context_end_line":1242,"code":"        focal = cam[\"focal\"][step]\n        pp = cam[\"pp\"][step]\n        R = cam[\"R\"][step]\n        t = cam[\"t\"][step]\n\n        q = tf.SO3.from_matrix(R).wxyz\n        fov = 2 * np.arctan(pp[0] / focal)\n        aspect = pp[0] / pp[1]\n        self.gt_cam_handles.append(\n            self.server.add_camera_frustum(\n                name=f\"/frames/{step}/gt_camera\",\n                fov=fov,\n                aspect=aspect,\n                wxyz=q,\n                position=t,\n                scale=0.1,\n                color=(166, 166, 166),\n            )\n        )\n    \n    def _update_camera_visibility(self):\n        \"\"\"Update camera visibility based on downsample factor\"\"\"\n        if not hasattr(self, 'cam_handles') or not self.show_camera:\n            return\n            \n        camera_downsample_factor = int(self.camera_downsample_slider.value)\n        \n        for i, handle in enumerate(self.cam_handles):\n            if i < len(self.all_steps):\n                step_index = self.step_to_index.get(self.all_steps[i], i)\n                should_show = (step_index % camera_downsample_factor == 0)\n                handle.visible = should_show\n    \n    def _update_gt_camera_visibility(self):\n        \"\"\"Update GT camera visibility based on downsample factor\"\"\"\n        if not hasattr(self, 'gt_cam_handles') or not self.show_gt_camera:\n            return\n            \n        camera_downsample_factor = int(self.camera_downsample_slider.value)\n        \n        for i, handle in enumerate(self.gt_cam_handles):\n            if i < len(self.all_steps):\n                step_index = self.step_to_index.get(self.all_steps[i], i)\n                should_show = (step_index % camera_downsample_factor == 0)\n                handle.visible = should_show\n\n    def add_gt_smpl(self, step):\n        if self.gt_smpl_verts is None:\n            return\n            \n        gt_verts = self.pcs[step][\"gt_verts\"]\n        faces = self.pcs[step][\"faces\"]","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._update_gt_camera_visibility","uri":"program://Human3R/function/viser_utils._update_gt_camera_visibility#L1224-L1235","kind":"function","name":"_update_gt_camera_visibility","path":"viser_utils.py","language":"python","start_line":1224,"end_line":1235,"context_start_line":1204,"context_end_line":1255,"code":"                wxyz=q,\n                position=t,\n                scale=0.1,\n                color=(166, 166, 166),\n            )\n        )\n    \n    def _update_camera_visibility(self):\n        \"\"\"Update camera visibility based on downsample factor\"\"\"\n        if not hasattr(self, 'cam_handles') or not self.show_camera:\n            return\n            \n        camera_downsample_factor = int(self.camera_downsample_slider.value)\n        \n        for i, handle in enumerate(self.cam_handles):\n            if i < len(self.all_steps):\n                step_index = self.step_to_index.get(self.all_steps[i], i)\n                should_show = (step_index % camera_downsample_factor == 0)\n                handle.visible = should_show\n    \n    def _update_gt_camera_visibility(self):\n        \"\"\"Update GT camera visibility based on downsample factor\"\"\"\n        if not hasattr(self, 'gt_cam_handles') or not self.show_gt_camera:\n            return\n            \n        camera_downsample_factor = int(self.camera_downsample_slider.value)\n        \n        for i, handle in enumerate(self.gt_cam_handles):\n            if i < len(self.all_steps):\n                step_index = self.step_to_index.get(self.all_steps[i], i)\n                should_show = (step_index % camera_downsample_factor == 0)\n                handle.visible = should_show\n\n    def add_gt_smpl(self, step):\n        if self.gt_smpl_verts is None:\n            return\n            \n        gt_verts = self.pcs[step][\"gt_verts\"]\n        faces = self.pcs[step][\"faces\"]\n        smpl_id = 51\n        \n        if gt_verts is not None and len(gt_verts) > 0:\n            for tid, vert in enumerate(gt_verts):\n                step_idx = self.step_to_index.get(step, 0)\n                mesh_opacity = self._compute_opacity_for_index(step_idx)\n                mesh_handle = self.server.scene.add_mesh_simple(\n                    name=f\"/frames/{step}/gt_human_{tid}\",\n                    vertices=vert,\n                    faces=faces,\n                    flat_shading=False,\n                    wireframe=False,\n                    opacity=mesh_opacity,","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.add_gt_smpl","uri":"program://Human3R/function/viser_utils.add_gt_smpl#L1237-L1259","kind":"function","name":"add_gt_smpl","path":"viser_utils.py","language":"python","start_line":1237,"end_line":1259,"context_start_line":1217,"context_end_line":1279,"code":"        \n        for i, handle in enumerate(self.cam_handles):\n            if i < len(self.all_steps):\n                step_index = self.step_to_index.get(self.all_steps[i], i)\n                should_show = (step_index % camera_downsample_factor == 0)\n                handle.visible = should_show\n    \n    def _update_gt_camera_visibility(self):\n        \"\"\"Update GT camera visibility based on downsample factor\"\"\"\n        if not hasattr(self, 'gt_cam_handles') or not self.show_gt_camera:\n            return\n            \n        camera_downsample_factor = int(self.camera_downsample_slider.value)\n        \n        for i, handle in enumerate(self.gt_cam_handles):\n            if i < len(self.all_steps):\n                step_index = self.step_to_index.get(self.all_steps[i], i)\n                should_show = (step_index % camera_downsample_factor == 0)\n                handle.visible = should_show\n\n    def add_gt_smpl(self, step):\n        if self.gt_smpl_verts is None:\n            return\n            \n        gt_verts = self.pcs[step][\"gt_verts\"]\n        faces = self.pcs[step][\"faces\"]\n        smpl_id = 51\n        \n        if gt_verts is not None and len(gt_verts) > 0:\n            for tid, vert in enumerate(gt_verts):\n                step_idx = self.step_to_index.get(step, 0)\n                mesh_opacity = self._compute_opacity_for_index(step_idx)\n                mesh_handle = self.server.scene.add_mesh_simple(\n                    name=f\"/frames/{step}/gt_human_{tid}\",\n                    vertices=vert,\n                    faces=faces,\n                    flat_shading=False,\n                    wireframe=False,\n                    opacity=mesh_opacity,\n                    color=(100, 100, 100),\n                )\n                self.gt_mesh_handles.append(mesh_handle)\n                self.gt_mesh_step_mapping.append(step)\n\n    def animate(self):\n        with self.server.add_gui_folder(\"Playback\"):\n            gui_timestep = self.server.add_gui_slider(\n                \"Train Step\",\n                min=0,\n                max=self.num_frames - 1,\n                step=1,\n                initial_value=0,\n                disabled=False,\n            )\n            gui_next_frame = self.server.add_gui_button(\"Next Step\", disabled=False)\n            gui_prev_frame = self.server.add_gui_button(\"Prev Step\", disabled=False)\n            gui_playing = self.server.add_gui_checkbox(\"Playing\", False)\n            gui_framerate = self.server.add_gui_slider(\n                \"FPS\", min=1, max=60, step=0.1, initial_value=1\n            )\n            gui_framerate_options = self.server.add_gui_button_group(\n                \"FPS options\", (\"10\", \"20\", \"30\", \"60\")\n            )","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.animate","uri":"program://Human3R/function/viser_utils.animate#L1261-L1387","kind":"function","name":"animate","path":"viser_utils.py","language":"python","start_line":1261,"end_line":1387,"context_start_line":1241,"context_end_line":1392,"code":"        gt_verts = self.pcs[step][\"gt_verts\"]\n        faces = self.pcs[step][\"faces\"]\n        smpl_id = 51\n        \n        if gt_verts is not None and len(gt_verts) > 0:\n            for tid, vert in enumerate(gt_verts):\n                step_idx = self.step_to_index.get(step, 0)\n                mesh_opacity = self._compute_opacity_for_index(step_idx)\n                mesh_handle = self.server.scene.add_mesh_simple(\n                    name=f\"/frames/{step}/gt_human_{tid}\",\n                    vertices=vert,\n                    faces=faces,\n                    flat_shading=False,\n                    wireframe=False,\n                    opacity=mesh_opacity,\n                    color=(100, 100, 100),\n                )\n                self.gt_mesh_handles.append(mesh_handle)\n                self.gt_mesh_step_mapping.append(step)\n\n    def animate(self):\n        with self.server.add_gui_folder(\"Playback\"):\n            gui_timestep = self.server.add_gui_slider(\n                \"Train Step\",\n                min=0,\n                max=self.num_frames - 1,\n                step=1,\n                initial_value=0,\n                disabled=False,\n            )\n            gui_next_frame = self.server.add_gui_button(\"Next Step\", disabled=False)\n            gui_prev_frame = self.server.add_gui_button(\"Prev Step\", disabled=False)\n            gui_playing = self.server.add_gui_checkbox(\"Playing\", False)\n            gui_framerate = self.server.add_gui_slider(\n                \"FPS\", min=1, max=60, step=0.1, initial_value=1\n            )\n            gui_framerate_options = self.server.add_gui_button_group(\n                \"FPS options\", (\"10\", \"20\", \"30\", \"60\")\n            )\n            \n        @gui_next_frame.on_click\n        def _(_) -> None:\n            gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n\n        @gui_prev_frame.on_click\n        def _(_) -> None:\n            gui_timestep.value = (gui_timestep.value - 1) % self.num_frames\n\n        @gui_playing.on_update\n        def _(_) -> None:\n            gui_timestep.disabled = gui_playing.value\n            gui_next_frame.disabled = gui_playing.value\n            gui_prev_frame.disabled = gui_playing.value\n\n        @gui_framerate_options.on_click\n        def _(_) -> None:\n            gui_framerate.value = int(gui_framerate_options.value)\n\n        prev_timestep = gui_timestep.value\n        self.current_step_index = prev_timestep\n\n        @gui_timestep.on_update\n        def _(_) -> None:\n            nonlocal prev_timestep\n            current_timestep = gui_timestep.value\n            with self.server.atomic():\n                self.frame_nodes[current_timestep].visible = True\n                self.frame_nodes[prev_timestep].visible = False\n            prev_timestep = current_timestep\n            self.current_step_index = current_timestep\n            # dynamic opacity update on step change\n            if self.dynamic_opacity_checkbox.value:\n                self._update_dynamic_opacities(current_timestep)\n            self.server.flush()  # Optional!\n\n        self.server.add_frame(\n            \"/frames\",\n            show_axes=False,\n        )\n        self.frame_nodes = []\n        for i in range(self.num_frames):\n            step = self.all_steps[i]\n            self.frame_nodes.append(\n                self.server.add_frame(\n                    f\"/frames/{step}\",\n                    show_axes=False,\n                )\n            )\n            self.add_pc(step)\n            if self.show_camera:\n                self.add_camera(step)\n            if self.show_gt_camera:\n                self.add_gt_camera(step)\n\n        prev_timestep = gui_timestep.value\n        while True:\n            if self.on_replay:\n                pass\n            else:\n                if gui_playing.value:\n                    gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n\n            current_step = gui_timestep.value\n            self.current_step_index = current_step\n            \n            for i, frame_node in enumerate(self.frame_nodes):\n                if self.hybrid_mode:\n                    frame_node.visible = i <= current_step\n                else:\n                    frame_node.visible = i <= current_step if not self.fourd else i == current_step\n            \n            # When playing, continuously update dynamic opacities (with throttling for performance)\n            if self.dynamic_opacity_checkbox.value:\n                self._update_dynamic_opacities(current_step)\n\n            show_mesh = self.gui_show_smpl.value\n            show_gt_mesh = self.show_gt_smpl\n            downsample_factor = int(self.smpl_downsample_slider.value)\n            \n            if self.hybrid_mode:\n                for mesh_idx, mesh_handle in enumerate(self.mesh_handles):\n                    if mesh_idx < len(self.mesh_step_mapping):\n                        mesh_step = self.mesh_step_mapping[mesh_idx]\n                        step_index = self.step_to_index.get(mesh_step, 0)\n                        should_show_by_downsample = (step_index % downsample_factor == 0)\n                        mesh_handle.visible = (mesh_step == current_step) and show_mesh and should_show_by_downsample\n                for mesh_idx, gt_mesh_handle in enumerate(self.gt_mesh_handles):\n                    if mesh_idx < len(self.gt_mesh_step_mapping):\n                        gt_mesh_step = self.gt_mesh_step_mapping[mesh_idx]\n                        step_index = self.step_to_index.get(gt_mesh_step, 0)\n                        should_show_by_downsample = (step_index % downsample_factor == 0)\n                        gt_mesh_handle.visible = (gt_mesh_step == current_step) and show_gt_mesh and should_show_by_downsample\n            else:\n                for mesh_idx, mesh_handle in enumerate(self.mesh_handles):\n                    if mesh_idx < len(self.mesh_step_mapping):\n                        step = self.mesh_step_mapping[mesh_idx]\n                        step_index = self.step_to_index.get(step, 0)\n                        should_show_by_downsample = (step_index % downsample_factor == 0)\n                        mesh_handle.visible = show_mesh and should_show_by_downsample\n                for mesh_idx, gt_mesh_handle in enumerate(self.gt_mesh_handles):\n                    if mesh_idx < len(self.gt_mesh_step_mapping):\n                        step = self.gt_mesh_step_mapping[mesh_idx]\n                        step_index = self.step_to_index.get(step, 0)\n                        should_show_by_downsample = (step_index % downsample_factor == 0)\n                        gt_mesh_handle.visible = show_gt_mesh and should_show_by_downsample\n\n            time.sleep(1.0 / gui_framerate.value)\n\n    def run(self):\n        self.animate()\n        while True:\n            time.sleep(10.0)","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils.run","uri":"program://Human3R/function/viser_utils.run#L1389-L1392","kind":"function","name":"run","path":"viser_utils.py","language":"python","start_line":1389,"end_line":1392,"context_start_line":1369,"context_end_line":1392,"code":"                        gt_mesh_step = self.gt_mesh_step_mapping[mesh_idx]\n                        step_index = self.step_to_index.get(gt_mesh_step, 0)\n                        should_show_by_downsample = (step_index % downsample_factor == 0)\n                        gt_mesh_handle.visible = (gt_mesh_step == current_step) and show_gt_mesh and should_show_by_downsample\n            else:\n                for mesh_idx, mesh_handle in enumerate(self.mesh_handles):\n                    if mesh_idx < len(self.mesh_step_mapping):\n                        step = self.mesh_step_mapping[mesh_idx]\n                        step_index = self.step_to_index.get(step, 0)\n                        should_show_by_downsample = (step_index % downsample_factor == 0)\n                        mesh_handle.visible = show_mesh and should_show_by_downsample\n                for mesh_idx, gt_mesh_handle in enumerate(self.gt_mesh_handles):\n                    if mesh_idx < len(self.gt_mesh_step_mapping):\n                        step = self.gt_mesh_step_mapping[mesh_idx]\n                        step_index = self.step_to_index.get(step, 0)\n                        should_show_by_downsample = (step_index % downsample_factor == 0)\n                        gt_mesh_handle.visible = show_gt_mesh and should_show_by_downsample\n\n            time.sleep(1.0 / gui_framerate.value)\n\n    def run(self):\n        self.animate()\n        while True:\n            time.sleep(10.0)","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:viser_utils._","uri":"program://Human3R/function/viser_utils._#L1303-L1314","kind":"function","name":"_","path":"viser_utils.py","language":"python","start_line":1303,"end_line":1314,"context_start_line":1283,"context_end_line":1334,"code":"            gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n\n        @gui_prev_frame.on_click\n        def _(_) -> None:\n            gui_timestep.value = (gui_timestep.value - 1) % self.num_frames\n\n        @gui_playing.on_update\n        def _(_) -> None:\n            gui_timestep.disabled = gui_playing.value\n            gui_next_frame.disabled = gui_playing.value\n            gui_prev_frame.disabled = gui_playing.value\n\n        @gui_framerate_options.on_click\n        def _(_) -> None:\n            gui_framerate.value = int(gui_framerate_options.value)\n\n        prev_timestep = gui_timestep.value\n        self.current_step_index = prev_timestep\n\n        @gui_timestep.on_update\n        def _(_) -> None:\n            nonlocal prev_timestep\n            current_timestep = gui_timestep.value\n            with self.server.atomic():\n                self.frame_nodes[current_timestep].visible = True\n                self.frame_nodes[prev_timestep].visible = False\n            prev_timestep = current_timestep\n            self.current_step_index = current_timestep\n            # dynamic opacity update on step change\n            if self.dynamic_opacity_checkbox.value:\n                self._update_dynamic_opacities(current_timestep)\n            self.server.flush()  # Optional!\n\n        self.server.add_frame(\n            \"/frames\",\n            show_axes=False,\n        )\n        self.frame_nodes = []\n        for i in range(self.num_frames):\n            step = self.all_steps[i]\n            self.frame_nodes.append(\n                self.server.add_frame(\n                    f\"/frames/{step}\",\n                    show_axes=False,\n                )\n            )\n            self.add_pc(step)\n            if self.show_camera:\n                self.add_camera(step)\n            if self.show_gt_camera:\n                self.add_gt_camera(step)\n","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.dataset.prepare_3dpw","uri":"program://Human3R/module/eval.dataset.prepare_3dpw#L1-L189","kind":"module","name":"eval.dataset.prepare_3dpw","path":"eval/dataset/prepare_3dpw.py","language":"python","start_line":1,"end_line":189,"context_start_line":1,"context_end_line":189,"code":"# Modified from Multi-HMR\n# Preprocess 3DPW dataset for human evaluation, and saves the annotations\n# files (i.e., 3dpw_test.pkl) in [ANNOT_DIR].\n# Usage: python -m eval.dataset.prepare_3dpw \"create_annots()\"\n\nimport os\nos.environ[\"PYOPENGL_PLATFORM\"] = \"egl\"\nos.environ['EGL_DEVICE_ID'] = '0'\n\nimport pickle\nimport torch\nimport smplx\nfrom tqdm import tqdm\nimport sys\nimport numpy as np\nfrom PIL import Image, ImageFile\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nsys.path.append(os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(__file__))), 'src'))\nfrom dust3r.smpl_model import SMPLX_DIR\n\nImageFile.LOAD_TRUNCATED_IMAGES = True # to avoid \"OSError: image file is truncated\"\nfrom torch.utils.data import Dataset\nimport roma\n\nANNOT_DIR = \"eval/global_human/annots/\"\nTHREEDPW_DIR = \"/path/to/3DPW\"\n\nclass THREEDPW(Dataset):\n    def __init__(self,\n                 split='test',\n                 training=False,\n                 img_size=512,\n                 root_dir=THREEDPW_DIR,\n                 force_build_dataset=0,\n                 subsample=1,\n                 *args, **kwargs\n                 ):\n        super().__init__()\n        \n        self.name = '3dpw'\n        self.annotations_dir = ANNOT_DIR\n        self.training = training\n        self.img_size = img_size\n        self.subsample = subsample\n\n        assert split in ['test']\n        \n        self.root_dir = root_dir\n        self.split = split\n        self.image_dir = os.path.join(self.root_dir, 'imageFiles')\n        self.annot_file = os.path.join(self.annotations_dir, f\"{self.name}_{split}.pkl\")\n        self.force_build_dataset = force_build_dataset\n\n        self.annots = None\n        if self.force_build_dataset or not os.path.isfile(self.annot_file):\n            self.annots = self.build_dataset()\n        if self.annots is None:\n            with open(self.annot_file, 'rb') as f:\n                self.annots = pickle.load(f)\n\n        self.imagenames = list(self.annots.keys())\n        self.imagenames.sort()\n\n        if self.subsample > 1:\n            self.imagenames = [self.imagenames[k] for k in np.arange(0,len(self.imagenames),self.subsample).tolist()]\n\n    @torch.no_grad()\n    def build_dataset(self):\n        print(f\"Bulding annotation file for: {self.name} - {self.split}\")\n\n        imagename2annot = {}\n\n        smpl_layer_male = smplx.create(SMPLX_DIR, 'smpl', gender='male')\n        smpl_layer_female = smplx.create(SMPLX_DIR, 'smpl', gender='female')\n\n        # Filenames\n        fns = os.listdir(os.path.join(self.root_dir, 'sequenceFiles', self.split))\n        fns.sort()\n\n        error = 0\n        max_human = 0\n        # Loop across sequence\n        for fn in tqdm(fns):\n            # Metadata\n            with open(os.path.join(self.root_dir, 'sequenceFiles', self.split, fn), 'rb') as f:\n                metadata = pickle.load(f, encoding='latin1')\n\n            # Camera intrinsics\n            K = metadata['cam_intrinsics']\n            focal = np.asarray([K[0,0],K[1,1]])\n            princpt = np.asarray([K[0,-1],K[1,-1]])\n\n            # Loop across time\n            seq_len = len(metadata['poses'][0])\n            n_person = len(metadata['genders'])\n            for k in range(seq_len):\n                # Image\n                seq_name = fn.replace('.pkl', '')\n                img_path = os.path.join(seq_name, f\"image_{k:05d}.jpg\")\n\n                # Resolution\n                width, height = Image.open(os.path.join(self.image_dir, img_path)).size\n\n                # Camera extrinsics\n                T = metadata['cam_poses'][k]\n                R, t = T[:3,:3], T[:3,-1]\n                \n                # Loop across person\n                persons = []\n                for i in range(n_person):\n                    # gt\n                    valid = metadata['campose_valid'][i][k]\n                    if valid == 0:\n                        continue\n                    poses = metadata['poses'][i][k].reshape(24,3)\n                    trans = metadata['trans'][i][k]\n                    shape = metadata['betas'][i][:10]\n                    gender = metadata['genders'][i]\n                    poses2d = metadata['poses2d'][i].transpose(0, 2, 1)[k] # [18,3] - openpose\n                    idx_valid2d = np.where(poses2d[:,-1] > 0.5)[0]\n                    poses2d = poses2d[idx_valid2d,:2]\n                    gender_ = 'male' if gender == 'm' else 'female'\n\n                    # apply camera extrinsic (rotation)\n                    body_pose = poses[1:]\n                    root_pose = poses[0]\n                    root_pose = roma.rotvec_to_rotmat(torch.Tensor(root_pose)).numpy()\n                    root_pose = R @ root_pose\n                    root_pose = roma.rotmat_to_rotvec(torch.Tensor(root_pose)).numpy()\n\n                    # get mesh w/0 transl\n                    smpl_layer_ = smpl_layer_male if gender_ == 'male' else smpl_layer_female\n                    out = smpl_layer_(global_orient=torch.from_numpy(root_pose).reshape(1,-1).float(),\n                                      body_pose=torch.from_numpy(body_pose).reshape(1,-1).float(),\n                                      betas=torch.from_numpy(shape).reshape(1,-1).float()\n                                    )\n                    # apply trans\n                    v3d, j3d = out.vertices.numpy().reshape(-1,3), out.joints.numpy().reshape(-1,3)\n                    mesh_cam, joint_cam = v3d + trans.reshape(1,3), j3d + trans.reshape(1,3)\n\n                    # apply camera exrinsic (translation) - it will compenstate rotation (translation from origin to root joint was not canceled)\n                    root_cam = joint_cam[0,None,:]\n                    mesh_cam = mesh_cam - root_cam + np.dot(R, root_cam.transpose(1,0)).transpose(1,0) + t # camera-centered coordinate system\n\n                    # find real transl in camera coordinate system\n                    trans = (mesh_cam - v3d)[0]\n\n                    # Append\n                    person = {\n                        # SMPL pseudo-GT\n                        'smpl_root_pose': root_pose.reshape(1,3).astype(np.float32), # axis-angle\n                        'smpl_body_pose': body_pose.reshape(23,3).astype(np.float32), # axis-angle\n                        'smpl_shape': shape.reshape(10).astype(np.float32),\n                        'smpl_transl': trans.reshape(3).astype(np.float32),\n                        'smpl_gender': gender_,\n                    }\n                    persons.append(person)\n\n                if len(persons) > max_human:\n                    max_human = len(persons)     \n\n                # Append\n                if len(persons) > 0:\n                    imagename2annot[img_path] = {\n                            # Camera\n                            'focal': focal.astype(np.float32).reshape(2),\n                            'princpt': princpt.astype(np.float32).reshape(2),\n                            'size': np.asarray([width, height]).astype(np.int32).reshape(2),\n                            # Humans\n                            'humans': persons,\n                            'cam_poses': T.astype(np.float32),\n                        }\n                    \n        print(f\"max_human: {max_human}\")\n    \n        # Saving\n        os.makedirs(os.path.dirname(self.annot_file), exist_ok=True)\n        print(f\"Saving into {self.annot_file}\")\n        with open(self.annot_file, 'wb') as f:\n            pickle.dump(imagename2annot, f, protocol=pickle.HIGHEST_PROTOCOL)\n\n        return imagename2annot\n\ndef create_annots():\n    dataset = THREEDPW(split='test', force_build_dataset=1)\n\nif __name__ == \"__main__\":\n    exec(sys.argv[1])","source_hash":"eefff66a8480dc590d2e891a31922dff5fed33d1a2119d898a226143535d8fef","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.dataset.prepare_3dpw.THREEDPW","uri":"program://Human3R/class/eval.dataset.prepare_3dpw.THREEDPW#L29-L183","kind":"class","name":"THREEDPW","path":"eval/dataset/prepare_3dpw.py","language":"python","start_line":29,"end_line":183,"context_start_line":9,"context_end_line":189,"code":"\nimport pickle\nimport torch\nimport smplx\nfrom tqdm import tqdm\nimport sys\nimport numpy as np\nfrom PIL import Image, ImageFile\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nsys.path.append(os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(__file__))), 'src'))\nfrom dust3r.smpl_model import SMPLX_DIR\n\nImageFile.LOAD_TRUNCATED_IMAGES = True # to avoid \"OSError: image file is truncated\"\nfrom torch.utils.data import Dataset\nimport roma\n\nANNOT_DIR = \"eval/global_human/annots/\"\nTHREEDPW_DIR = \"/path/to/3DPW\"\n\nclass THREEDPW(Dataset):\n    def __init__(self,\n                 split='test',\n                 training=False,\n                 img_size=512,\n                 root_dir=THREEDPW_DIR,\n                 force_build_dataset=0,\n                 subsample=1,\n                 *args, **kwargs\n                 ):\n        super().__init__()\n        \n        self.name = '3dpw'\n        self.annotations_dir = ANNOT_DIR\n        self.training = training\n        self.img_size = img_size\n        self.subsample = subsample\n\n        assert split in ['test']\n        \n        self.root_dir = root_dir\n        self.split = split\n        self.image_dir = os.path.join(self.root_dir, 'imageFiles')\n        self.annot_file = os.path.join(self.annotations_dir, f\"{self.name}_{split}.pkl\")\n        self.force_build_dataset = force_build_dataset\n\n        self.annots = None\n        if self.force_build_dataset or not os.path.isfile(self.annot_file):\n            self.annots = self.build_dataset()\n        if self.annots is None:\n            with open(self.annot_file, 'rb') as f:\n                self.annots = pickle.load(f)\n\n        self.imagenames = list(self.annots.keys())\n        self.imagenames.sort()\n\n        if self.subsample > 1:\n            self.imagenames = [self.imagenames[k] for k in np.arange(0,len(self.imagenames),self.subsample).tolist()]\n\n    @torch.no_grad()\n    def build_dataset(self):\n        print(f\"Bulding annotation file for: {self.name} - {self.split}\")\n\n        imagename2annot = {}\n\n        smpl_layer_male = smplx.create(SMPLX_DIR, 'smpl', gender='male')\n        smpl_layer_female = smplx.create(SMPLX_DIR, 'smpl', gender='female')\n\n        # Filenames\n        fns = os.listdir(os.path.join(self.root_dir, 'sequenceFiles', self.split))\n        fns.sort()\n\n        error = 0\n        max_human = 0\n        # Loop across sequence\n        for fn in tqdm(fns):\n            # Metadata\n            with open(os.path.join(self.root_dir, 'sequenceFiles', self.split, fn), 'rb') as f:\n                metadata = pickle.load(f, encoding='latin1')\n\n            # Camera intrinsics\n            K = metadata['cam_intrinsics']\n            focal = np.asarray([K[0,0],K[1,1]])\n            princpt = np.asarray([K[0,-1],K[1,-1]])\n\n            # Loop across time\n            seq_len = len(metadata['poses'][0])\n            n_person = len(metadata['genders'])\n            for k in range(seq_len):\n                # Image\n                seq_name = fn.replace('.pkl', '')\n                img_path = os.path.join(seq_name, f\"image_{k:05d}.jpg\")\n\n                # Resolution\n                width, height = Image.open(os.path.join(self.image_dir, img_path)).size\n\n                # Camera extrinsics\n                T = metadata['cam_poses'][k]\n                R, t = T[:3,:3], T[:3,-1]\n                \n                # Loop across person\n                persons = []\n                for i in range(n_person):\n                    # gt\n                    valid = metadata['campose_valid'][i][k]\n                    if valid == 0:\n                        continue\n                    poses = metadata['poses'][i][k].reshape(24,3)\n                    trans = metadata['trans'][i][k]\n                    shape = metadata['betas'][i][:10]\n                    gender = metadata['genders'][i]\n                    poses2d = metadata['poses2d'][i].transpose(0, 2, 1)[k] # [18,3] - openpose\n                    idx_valid2d = np.where(poses2d[:,-1] > 0.5)[0]\n                    poses2d = poses2d[idx_valid2d,:2]\n                    gender_ = 'male' if gender == 'm' else 'female'\n\n                    # apply camera extrinsic (rotation)\n                    body_pose = poses[1:]\n                    root_pose = poses[0]\n                    root_pose = roma.rotvec_to_rotmat(torch.Tensor(root_pose)).numpy()\n                    root_pose = R @ root_pose\n                    root_pose = roma.rotmat_to_rotvec(torch.Tensor(root_pose)).numpy()\n\n                    # get mesh w/0 transl\n                    smpl_layer_ = smpl_layer_male if gender_ == 'male' else smpl_layer_female\n                    out = smpl_layer_(global_orient=torch.from_numpy(root_pose).reshape(1,-1).float(),\n                                      body_pose=torch.from_numpy(body_pose).reshape(1,-1).float(),\n                                      betas=torch.from_numpy(shape).reshape(1,-1).float()\n                                    )\n                    # apply trans\n                    v3d, j3d = out.vertices.numpy().reshape(-1,3), out.joints.numpy().reshape(-1,3)\n                    mesh_cam, joint_cam = v3d + trans.reshape(1,3), j3d + trans.reshape(1,3)\n\n                    # apply camera exrinsic (translation) - it will compenstate rotation (translation from origin to root joint was not canceled)\n                    root_cam = joint_cam[0,None,:]\n                    mesh_cam = mesh_cam - root_cam + np.dot(R, root_cam.transpose(1,0)).transpose(1,0) + t # camera-centered coordinate system\n\n                    # find real transl in camera coordinate system\n                    trans = (mesh_cam - v3d)[0]\n\n                    # Append\n                    person = {\n                        # SMPL pseudo-GT\n                        'smpl_root_pose': root_pose.reshape(1,3).astype(np.float32), # axis-angle\n                        'smpl_body_pose': body_pose.reshape(23,3).astype(np.float32), # axis-angle\n                        'smpl_shape': shape.reshape(10).astype(np.float32),\n                        'smpl_transl': trans.reshape(3).astype(np.float32),\n                        'smpl_gender': gender_,\n                    }\n                    persons.append(person)\n\n                if len(persons) > max_human:\n                    max_human = len(persons)     \n\n                # Append\n                if len(persons) > 0:\n                    imagename2annot[img_path] = {\n                            # Camera\n                            'focal': focal.astype(np.float32).reshape(2),\n                            'princpt': princpt.astype(np.float32).reshape(2),\n                            'size': np.asarray([width, height]).astype(np.int32).reshape(2),\n                            # Humans\n                            'humans': persons,\n                            'cam_poses': T.astype(np.float32),\n                        }\n                    \n        print(f\"max_human: {max_human}\")\n    \n        # Saving\n        os.makedirs(os.path.dirname(self.annot_file), exist_ok=True)\n        print(f\"Saving into {self.annot_file}\")\n        with open(self.annot_file, 'wb') as f:\n            pickle.dump(imagename2annot, f, protocol=pickle.HIGHEST_PROTOCOL)\n\n        return imagename2annot\n\ndef create_annots():\n    dataset = THREEDPW(split='test', force_build_dataset=1)\n\nif __name__ == \"__main__\":\n    exec(sys.argv[1])","source_hash":"eefff66a8480dc590d2e891a31922dff5fed33d1a2119d898a226143535d8fef","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.dataset.prepare_3dpw.create_annots","uri":"program://Human3R/function/eval.dataset.prepare_3dpw.create_annots#L185-L186","kind":"function","name":"create_annots","path":"eval/dataset/prepare_3dpw.py","language":"python","start_line":185,"end_line":186,"context_start_line":165,"context_end_line":189,"code":"                    imagename2annot[img_path] = {\n                            # Camera\n                            'focal': focal.astype(np.float32).reshape(2),\n                            'princpt': princpt.astype(np.float32).reshape(2),\n                            'size': np.asarray([width, height]).astype(np.int32).reshape(2),\n                            # Humans\n                            'humans': persons,\n                            'cam_poses': T.astype(np.float32),\n                        }\n                    \n        print(f\"max_human: {max_human}\")\n    \n        # Saving\n        os.makedirs(os.path.dirname(self.annot_file), exist_ok=True)\n        print(f\"Saving into {self.annot_file}\")\n        with open(self.annot_file, 'wb') as f:\n            pickle.dump(imagename2annot, f, protocol=pickle.HIGHEST_PROTOCOL)\n\n        return imagename2annot\n\ndef create_annots():\n    dataset = THREEDPW(split='test', force_build_dataset=1)\n\nif __name__ == \"__main__\":\n    exec(sys.argv[1])","source_hash":"eefff66a8480dc590d2e891a31922dff5fed33d1a2119d898a226143535d8fef","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.dataset.prepare_3dpw.__init__","uri":"program://Human3R/function/eval.dataset.prepare_3dpw.__init__#L30-L66","kind":"function","name":"__init__","path":"eval/dataset/prepare_3dpw.py","language":"python","start_line":30,"end_line":66,"context_start_line":10,"context_end_line":86,"code":"import pickle\nimport torch\nimport smplx\nfrom tqdm import tqdm\nimport sys\nimport numpy as np\nfrom PIL import Image, ImageFile\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nsys.path.append(os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(__file__))), 'src'))\nfrom dust3r.smpl_model import SMPLX_DIR\n\nImageFile.LOAD_TRUNCATED_IMAGES = True # to avoid \"OSError: image file is truncated\"\nfrom torch.utils.data import Dataset\nimport roma\n\nANNOT_DIR = \"eval/global_human/annots/\"\nTHREEDPW_DIR = \"/path/to/3DPW\"\n\nclass THREEDPW(Dataset):\n    def __init__(self,\n                 split='test',\n                 training=False,\n                 img_size=512,\n                 root_dir=THREEDPW_DIR,\n                 force_build_dataset=0,\n                 subsample=1,\n                 *args, **kwargs\n                 ):\n        super().__init__()\n        \n        self.name = '3dpw'\n        self.annotations_dir = ANNOT_DIR\n        self.training = training\n        self.img_size = img_size\n        self.subsample = subsample\n\n        assert split in ['test']\n        \n        self.root_dir = root_dir\n        self.split = split\n        self.image_dir = os.path.join(self.root_dir, 'imageFiles')\n        self.annot_file = os.path.join(self.annotations_dir, f\"{self.name}_{split}.pkl\")\n        self.force_build_dataset = force_build_dataset\n\n        self.annots = None\n        if self.force_build_dataset or not os.path.isfile(self.annot_file):\n            self.annots = self.build_dataset()\n        if self.annots is None:\n            with open(self.annot_file, 'rb') as f:\n                self.annots = pickle.load(f)\n\n        self.imagenames = list(self.annots.keys())\n        self.imagenames.sort()\n\n        if self.subsample > 1:\n            self.imagenames = [self.imagenames[k] for k in np.arange(0,len(self.imagenames),self.subsample).tolist()]\n\n    @torch.no_grad()\n    def build_dataset(self):\n        print(f\"Bulding annotation file for: {self.name} - {self.split}\")\n\n        imagename2annot = {}\n\n        smpl_layer_male = smplx.create(SMPLX_DIR, 'smpl', gender='male')\n        smpl_layer_female = smplx.create(SMPLX_DIR, 'smpl', gender='female')\n\n        # Filenames\n        fns = os.listdir(os.path.join(self.root_dir, 'sequenceFiles', self.split))\n        fns.sort()\n\n        error = 0\n        max_human = 0\n        # Loop across sequence\n        for fn in tqdm(fns):\n            # Metadata\n            with open(os.path.join(self.root_dir, 'sequenceFiles', self.split, fn), 'rb') as f:","source_hash":"eefff66a8480dc590d2e891a31922dff5fed33d1a2119d898a226143535d8fef","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.dataset.prepare_3dpw.build_dataset","uri":"program://Human3R/function/eval.dataset.prepare_3dpw.build_dataset#L69-L183","kind":"function","name":"build_dataset","path":"eval/dataset/prepare_3dpw.py","language":"python","start_line":69,"end_line":183,"context_start_line":49,"context_end_line":189,"code":"        self.root_dir = root_dir\n        self.split = split\n        self.image_dir = os.path.join(self.root_dir, 'imageFiles')\n        self.annot_file = os.path.join(self.annotations_dir, f\"{self.name}_{split}.pkl\")\n        self.force_build_dataset = force_build_dataset\n\n        self.annots = None\n        if self.force_build_dataset or not os.path.isfile(self.annot_file):\n            self.annots = self.build_dataset()\n        if self.annots is None:\n            with open(self.annot_file, 'rb') as f:\n                self.annots = pickle.load(f)\n\n        self.imagenames = list(self.annots.keys())\n        self.imagenames.sort()\n\n        if self.subsample > 1:\n            self.imagenames = [self.imagenames[k] for k in np.arange(0,len(self.imagenames),self.subsample).tolist()]\n\n    @torch.no_grad()\n    def build_dataset(self):\n        print(f\"Bulding annotation file for: {self.name} - {self.split}\")\n\n        imagename2annot = {}\n\n        smpl_layer_male = smplx.create(SMPLX_DIR, 'smpl', gender='male')\n        smpl_layer_female = smplx.create(SMPLX_DIR, 'smpl', gender='female')\n\n        # Filenames\n        fns = os.listdir(os.path.join(self.root_dir, 'sequenceFiles', self.split))\n        fns.sort()\n\n        error = 0\n        max_human = 0\n        # Loop across sequence\n        for fn in tqdm(fns):\n            # Metadata\n            with open(os.path.join(self.root_dir, 'sequenceFiles', self.split, fn), 'rb') as f:\n                metadata = pickle.load(f, encoding='latin1')\n\n            # Camera intrinsics\n            K = metadata['cam_intrinsics']\n            focal = np.asarray([K[0,0],K[1,1]])\n            princpt = np.asarray([K[0,-1],K[1,-1]])\n\n            # Loop across time\n            seq_len = len(metadata['poses'][0])\n            n_person = len(metadata['genders'])\n            for k in range(seq_len):\n                # Image\n                seq_name = fn.replace('.pkl', '')\n                img_path = os.path.join(seq_name, f\"image_{k:05d}.jpg\")\n\n                # Resolution\n                width, height = Image.open(os.path.join(self.image_dir, img_path)).size\n\n                # Camera extrinsics\n                T = metadata['cam_poses'][k]\n                R, t = T[:3,:3], T[:3,-1]\n                \n                # Loop across person\n                persons = []\n                for i in range(n_person):\n                    # gt\n                    valid = metadata['campose_valid'][i][k]\n                    if valid == 0:\n                        continue\n                    poses = metadata['poses'][i][k].reshape(24,3)\n                    trans = metadata['trans'][i][k]\n                    shape = metadata['betas'][i][:10]\n                    gender = metadata['genders'][i]\n                    poses2d = metadata['poses2d'][i].transpose(0, 2, 1)[k] # [18,3] - openpose\n                    idx_valid2d = np.where(poses2d[:,-1] > 0.5)[0]\n                    poses2d = poses2d[idx_valid2d,:2]\n                    gender_ = 'male' if gender == 'm' else 'female'\n\n                    # apply camera extrinsic (rotation)\n                    body_pose = poses[1:]\n                    root_pose = poses[0]\n                    root_pose = roma.rotvec_to_rotmat(torch.Tensor(root_pose)).numpy()\n                    root_pose = R @ root_pose\n                    root_pose = roma.rotmat_to_rotvec(torch.Tensor(root_pose)).numpy()\n\n                    # get mesh w/0 transl\n                    smpl_layer_ = smpl_layer_male if gender_ == 'male' else smpl_layer_female\n                    out = smpl_layer_(global_orient=torch.from_numpy(root_pose).reshape(1,-1).float(),\n                                      body_pose=torch.from_numpy(body_pose).reshape(1,-1).float(),\n                                      betas=torch.from_numpy(shape).reshape(1,-1).float()\n                                    )\n                    # apply trans\n                    v3d, j3d = out.vertices.numpy().reshape(-1,3), out.joints.numpy().reshape(-1,3)\n                    mesh_cam, joint_cam = v3d + trans.reshape(1,3), j3d + trans.reshape(1,3)\n\n                    # apply camera exrinsic (translation) - it will compenstate rotation (translation from origin to root joint was not canceled)\n                    root_cam = joint_cam[0,None,:]\n                    mesh_cam = mesh_cam - root_cam + np.dot(R, root_cam.transpose(1,0)).transpose(1,0) + t # camera-centered coordinate system\n\n                    # find real transl in camera coordinate system\n                    trans = (mesh_cam - v3d)[0]\n\n                    # Append\n                    person = {\n                        # SMPL pseudo-GT\n                        'smpl_root_pose': root_pose.reshape(1,3).astype(np.float32), # axis-angle\n                        'smpl_body_pose': body_pose.reshape(23,3).astype(np.float32), # axis-angle\n                        'smpl_shape': shape.reshape(10).astype(np.float32),\n                        'smpl_transl': trans.reshape(3).astype(np.float32),\n                        'smpl_gender': gender_,\n                    }\n                    persons.append(person)\n\n                if len(persons) > max_human:\n                    max_human = len(persons)     \n\n                # Append\n                if len(persons) > 0:\n                    imagename2annot[img_path] = {\n                            # Camera\n                            'focal': focal.astype(np.float32).reshape(2),\n                            'princpt': princpt.astype(np.float32).reshape(2),\n                            'size': np.asarray([width, height]).astype(np.int32).reshape(2),\n                            # Humans\n                            'humans': persons,\n                            'cam_poses': T.astype(np.float32),\n                        }\n                    \n        print(f\"max_human: {max_human}\")\n    \n        # Saving\n        os.makedirs(os.path.dirname(self.annot_file), exist_ok=True)\n        print(f\"Saving into {self.annot_file}\")\n        with open(self.annot_file, 'wb') as f:\n            pickle.dump(imagename2annot, f, protocol=pickle.HIGHEST_PROTOCOL)\n\n        return imagename2annot\n\ndef create_annots():\n    dataset = THREEDPW(split='test', force_build_dataset=1)\n\nif __name__ == \"__main__\":\n    exec(sys.argv[1])","source_hash":"eefff66a8480dc590d2e891a31922dff5fed33d1a2119d898a226143535d8fef","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.metadata","uri":"program://Human3R/module/eval.relpose.metadata#L1-L252","kind":"module","name":"eval.relpose.metadata","path":"eval/relpose/metadata.py","language":"python","start_line":1,"end_line":252,"context_start_line":1,"context_end_line":252,"code":"import os\nimport glob\nfrom tqdm import tqdm\n\n# Define the merged dataset metadata dictionary\ndataset_metadata = {\n    \"davis\": {\n        \"img_path\": \"data/davis/DAVIS/JPEGImages/480p\",\n        \"mask_path\": \"data/davis/DAVIS/masked_images/480p\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: None,\n        \"traj_format\": None,\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: os.path.join(mask_path, seq),\n        \"skip_condition\": None,\n        \"process_func\": None,  # Not used in mono depth estimation\n    },\n    \"kitti\": {\n        \"img_path\": \"data/kitti/depth_selection/val_selection_cropped/image_gathered\",  # Default path\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: None,\n        \"traj_format\": None,\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": lambda args, img_path: process_kitti(args, img_path),\n    },\n    \"bonn\": {\n        \"img_path\": \"/path/to/rgbd_bonn_dataset\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(\n            img_path, f\"rgbd_bonn_{seq}\", \"rgb_110\"\n        ),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, f\"rgbd_bonn_{seq}\", \"groundtruth_110.txt\"\n        ),\n        \"traj_format\": \"tum\",\n        \"seq_list\": [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"],\n        \"full_seq\": False,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": lambda args, img_path: process_bonn(args, img_path),\n    },\n    \"nyu\": {\n        \"img_path\": \"data/nyu-v2/val/nyu_images\",\n        \"mask_path\": None,\n        \"process_func\": lambda args, img_path: process_nyu(args, img_path),\n    },\n    \"scannet\": {\n        \"img_path\": \"data/scannetv2\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq, \"color_90\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, seq, \"pose_90.txt\"\n        ),\n        \"traj_format\": \"replica\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,  # lambda save_dir, seq: os.path.exists(os.path.join(save_dir, seq)),\n        \"process_func\": lambda args, img_path: process_scannet(args, img_path),\n    },\n    \"scannet-257\": {\n        \"img_path\": \"data/scannetv2_3_257\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq, \"color_90\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, seq, \"pose_90.txt\"\n        ),\n        \"traj_format\": \"replica\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,  # lambda save_dir, seq: os.path.exists(os.path.join(save_dir, seq)),\n        \"process_func\": lambda args, img_path: process_scannet(args, img_path),\n    },\n    \"scannet-129\": {\n        \"img_path\": \"data/scannetv2_3_129\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq, \"color_90\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, seq, \"pose_90.txt\"\n        ),\n        \"traj_format\": \"replica\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,  # lambda save_dir, seq: os.path.exists(os.path.join(save_dir, seq)),\n        \"process_func\": lambda args, img_path: process_scannet(args, img_path),\n    },\n    \"scannet-65\": {\n        \"img_path\": \"data/scannetv2_3_65\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq, \"color_90\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, seq, \"pose_90.txt\"\n        ),\n        \"traj_format\": \"replica\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,  # lambda save_dir, seq: os.path.exists(os.path.join(save_dir, seq)),\n        \"process_func\": lambda args, img_path: process_scannet(args, img_path),\n    },\n    \"scannet-33\": {\n        \"img_path\": \"data/scannetv2_3_33\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq, \"color_90\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, seq, \"pose_90.txt\"\n        ),\n        \"traj_format\": \"replica\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,  # lambda save_dir, seq: os.path.exists(os.path.join(save_dir, seq)),\n        \"process_func\": lambda args, img_path: process_scannet(args, img_path),\n    },\n    \"tum\": {\n        \"img_path\": \"/path/to/tum\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq, \"rgb_90\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, seq, \"groundtruth_90.txt\"\n        ),\n        \"traj_format\": \"tum\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": None,\n    },\n    \"sintel\": {\n        \"img_path\": \"/path/to/sintel/training/final\",\n        \"anno_path\": \"/path/to/sintel/training/camdata_left\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(anno_path, seq),\n        \"traj_format\": None,\n        \"seq_list\": [\n            \"alley_2\",\n            \"ambush_4\",\n            \"ambush_5\",\n            \"ambush_6\",\n            \"cave_2\",\n            \"cave_4\",\n            \"market_2\",\n            \"market_5\",\n            \"market_6\",\n            \"shaman_3\",\n            \"sleeping_1\",\n            \"sleeping_2\",\n            \"temple_2\",\n            \"temple_3\",\n        ],\n        \"full_seq\": False,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": lambda args, img_path: process_sintel(args, img_path),\n    },\n}\n\ntum_numbers = [50, 100, 150, 200, 300, 400, 500, 600, 700, 800, 900, 1000]\ntum_configs = {\n    f\"tum_{num}\": {\n        \"img_path\": \"/path/to/long_tum\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq, num=num: os.path.join(img_path, seq, f\"rgb_{num}\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq, num=num: os.path.join(\n            img_path, seq, f\"groundtruth_{num}.txt\"\n        ),\n        \"traj_format\": \"tum\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": None,\n        }\n    for num in tum_numbers\n}\ndataset_metadata.update(tum_configs)\n\n# Define processing functions for each dataset\ndef process_kitti(args, img_path):\n    for dir in tqdm(sorted(glob.glob(f\"{img_path}/*\"))):\n        filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(dir)}\"\n        yield filelist, save_dir\n\n\ndef process_bonn(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/rgb/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = [\n            \"alley_2\",\n            \"ambush_4\",\n            \"ambush_5\",\n            \"ambush_6\",\n            \"cave_2\",\n            \"cave_4\",\n            \"market_2\",\n            \"market_5\",\n            \"market_6\",\n            \"shaman_3\",\n            \"sleeping_1\",\n            \"sleeping_2\",\n            \"temple_2\",\n            \"temple_3\",\n        ]\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/{seq}/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir","source_hash":"fed1eca7775b0155dd84b26e08fca02e97e4ac8bdd84a396f5767f62c64cd478","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.metadata.process_kitti","uri":"program://Human3R/function/eval.relpose.metadata.process_kitti#L187-L191","kind":"function","name":"process_kitti","path":"eval/relpose/metadata.py","language":"python","start_line":187,"end_line":191,"context_start_line":167,"context_end_line":211,"code":"tum_configs = {\n    f\"tum_{num}\": {\n        \"img_path\": \"/path/to/long_tum\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq, num=num: os.path.join(img_path, seq, f\"rgb_{num}\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq, num=num: os.path.join(\n            img_path, seq, f\"groundtruth_{num}.txt\"\n        ),\n        \"traj_format\": \"tum\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": None,\n        }\n    for num in tum_numbers\n}\ndataset_metadata.update(tum_configs)\n\n# Define processing functions for each dataset\ndef process_kitti(args, img_path):\n    for dir in tqdm(sorted(glob.glob(f\"{img_path}/*\"))):\n        filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(dir)}\"\n        yield filelist, save_dir\n\n\ndef process_bonn(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/rgb/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n","source_hash":"fed1eca7775b0155dd84b26e08fca02e97e4ac8bdd84a396f5767f62c64cd478","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.metadata.process_bonn","uri":"program://Human3R/function/eval.relpose.metadata.process_bonn#L194-L209","kind":"function","name":"process_bonn","path":"eval/relpose/metadata.py","language":"python","start_line":194,"end_line":209,"context_start_line":174,"context_end_line":229,"code":"        ),\n        \"traj_format\": \"tum\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": None,\n        }\n    for num in tum_numbers\n}\ndataset_metadata.update(tum_configs)\n\n# Define processing functions for each dataset\ndef process_kitti(args, img_path):\n    for dir in tqdm(sorted(glob.glob(f\"{img_path}/*\"))):\n        filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(dir)}\"\n        yield filelist, save_dir\n\n\ndef process_bonn(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/rgb/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))","source_hash":"fed1eca7775b0155dd84b26e08fca02e97e4ac8bdd84a396f5767f62c64cd478","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.metadata.process_nyu","uri":"program://Human3R/function/eval.relpose.metadata.process_nyu#L212-L215","kind":"function","name":"process_nyu","path":"eval/relpose/metadata.py","language":"python","start_line":212,"end_line":215,"context_start_line":192,"context_end_line":235,"code":"\n\ndef process_bonn(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/rgb/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = [\n            \"alley_2\",\n            \"ambush_4\",","source_hash":"fed1eca7775b0155dd84b26e08fca02e97e4ac8bdd84a396f5767f62c64cd478","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.metadata.process_scannet","uri":"program://Human3R/function/eval.relpose.metadata.process_scannet#L218-L223","kind":"function","name":"process_scannet","path":"eval/relpose/metadata.py","language":"python","start_line":218,"end_line":223,"context_start_line":198,"context_end_line":243,"code":"            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = [\n            \"alley_2\",\n            \"ambush_4\",\n            \"ambush_5\",\n            \"ambush_6\",\n            \"cave_2\",\n            \"cave_4\",\n            \"market_2\",\n            \"market_5\",\n            \"market_6\",\n            \"shaman_3\",","source_hash":"fed1eca7775b0155dd84b26e08fca02e97e4ac8bdd84a396f5767f62c64cd478","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.metadata.process_sintel","uri":"program://Human3R/function/eval.relpose.metadata.process_sintel#L226-L252","kind":"function","name":"process_sintel","path":"eval/relpose/metadata.py","language":"python","start_line":226,"end_line":252,"context_start_line":206,"context_end_line":252,"code":"        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = [\n            \"alley_2\",\n            \"ambush_4\",\n            \"ambush_5\",\n            \"ambush_6\",\n            \"cave_2\",\n            \"cave_4\",\n            \"market_2\",\n            \"market_5\",\n            \"market_6\",\n            \"shaman_3\",\n            \"sleeping_1\",\n            \"sleeping_2\",\n            \"temple_2\",\n            \"temple_3\",\n        ]\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/{seq}/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir","source_hash":"fed1eca7775b0155dd84b26e08fca02e97e4ac8bdd84a396f5767f62c64cd478","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils","uri":"program://Human3R/module/eval.relpose.utils#L1-L311","kind":"module","name":"eval.relpose.utils","path":"eval/relpose/utils.py","language":"python","start_line":1,"end_line":311,"context_start_line":1,"context_end_line":311,"code":"from copy import deepcopy\nimport cv2\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport roma\nfrom copy import deepcopy\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom eval.relpose.evo_utils import *\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\n\n# from checkpoints.dust3r.viz import colorize_np, colorize\n\n\ndef todevice(batch, device, callback=None, non_blocking=False):\n    \"\"\"Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy).\n\n    batch: list, tuple, dict of tensors or other things\n    device: pytorch device or 'numpy'\n    callback: function that would be called on every sub-elements.\n    \"\"\"\n    if callback:\n        batch = callback(batch)\n\n    if isinstance(batch, dict):\n        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef c2w_to_tumpose(c2w):\n    \"\"\"\n    Convert a camera-to-world matrix to a tuple of translation and rotation\n\n    input: c2w: 4x4 matrix\n    output: tuple of translation and rotation (x y z qw qx qy qz)\n    \"\"\"\n    # convert input to numpy\n    c2w = to_numpy(c2w)\n    xyz = c2w[:3, -1]\n    rot = Rotation.from_matrix(c2w[:3, :3])\n    qx, qy, qz, qw = rot.as_quat()\n    tum_pose = np.concatenate([xyz, [qw, qx, qy, qz]])\n    return tum_pose\n\n\ndef get_tum_poses(poses):\n    \"\"\"\n    poses: list of 4x4 arrays\n    \"\"\"\n    tt = np.arange(len(poses)).astype(float)\n    tum_poses = [c2w_to_tumpose(p) for p in poses]\n    tum_poses = np.stack(tum_poses, 0)\n    return [tum_poses, tt]\n\n\ndef save_tum_poses(poses, path):\n    traj = get_tum_poses(poses)\n    save_trajectory_tum_format(traj, path)\n    return traj[0]  # return the poses\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)\n    min_depth = depth_maps.min()  # float(torch.quantile(out, 0.01))\n    max_depth = depth_maps.max()  # float(torch.quantile(out, 0.99))\n    colored_depth = colorize(\n        depth_maps,\n        cmap_name=\"Spectral_r\",\n        range=(min_depth, max_depth),\n        append_cbar=True,\n    )\n    images = []\n\n    if conf_self is not None:\n        conf_selfs = torch.concat(conf_self, 0)\n        min_conf = torch.log(conf_selfs.min())  # float(torch.quantile(out, 0.01))\n        max_conf = torch.log(conf_selfs.max())  # float(torch.quantile(out, 0.99))\n        colored_conf = colorize(\n            torch.log(conf_selfs),\n            cmap_name=\"jet\",\n            range=(min_conf, max_conf),\n            append_cbar=True,\n        )\n\n    for i, depth_map in enumerate(colored_depth):\n        # Apply color map to depth map\n        img_path = f\"{path}/frame_{(i):04d}.png\"\n        if conf_self is None:\n            to_save = (depth_map * 255).detach().cpu().numpy().astype(np.uint8)\n        else:\n            to_save = torch.cat([depth_map, colored_conf[i]], dim=1)\n            to_save = (to_save * 255).detach().cpu().numpy().astype(np.uint8)\n        iio.imwrite(img_path, to_save)\n        images.append(Image.open(img_path))\n        np.save(f\"{path}/frame_{(i):04d}.npy\", depth_maps[i].detach().cpu().numpy())\n\n    images[0].save(\n        f\"{path}/_depth_maps.gif\",\n        save_all=True,\n        append_images=images[1:],\n        duration=100,\n        loop=0,\n    )\n\n    return depth_maps\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    # Do some plotting.\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n\n    tick_cnt = 6\n    tick_loc = np.linspace(vmin, vmax, tick_cnt)\n    cb1 = mpl.colorbar.ColorbarBase(\n        ax, cmap=cmap, norm=norm, ticks=tick_loc, orientation=\"vertical\"\n    )\n\n    tick_label = [str(np.round(x, cbar_precision)) for x in tick_loc]\n    if cbar_precision == 0:\n        tick_label = [x[:-2] for x in tick_label]\n\n    cb1.set_ticklabels(tick_label)\n\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    # fig.tight_layout()\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]\n    \"\"\"\n    if range is not None:\n        vmin, vmax = range\n    elif mask is not None:\n        # vmin, vmax = np.percentile(x[mask], (2, 100))\n        vmin = np.min(x[mask][np.nonzero(x[mask])])\n        vmax = np.max(x[mask])\n        # vmin = vmin - np.abs(vmin) * 0.01\n        x[np.logical_not(mask)] = vmin\n        # print(vmin, vmax)\n    else:\n        vmin, vmax = np.percentile(x, (1, 100))\n        vmax += 1e-6\n\n    x = np.clip(x, vmin, vmax)\n    x = (x - vmin) / (vmax - vmin)\n    # x = np.clip(x, 0., 1.)\n\n    cmap = cm.get_cmap(cmap_name)\n    x_new = cmap(x)[:, :, :3]\n\n    if mask is not None:\n        mask = np.float32(mask[:, :, np.newaxis])\n        x_new = x_new * mask + np.ones_like(x_new) * (1.0 - mask)\n\n    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\n# tensor\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.todevice","uri":"program://Human3R/function/eval.relpose.utils.todevice#L23-L48","kind":"function","name":"todevice","path":"eval/relpose/utils.py","language":"python","start_line":23,"end_line":48,"context_start_line":3,"context_end_line":68,"code":"\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport roma\nfrom copy import deepcopy\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom eval.relpose.evo_utils import *\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\n\n# from checkpoints.dust3r.viz import colorize_np, colorize\n\n\ndef todevice(batch, device, callback=None, non_blocking=False):\n    \"\"\"Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy).\n\n    batch: list, tuple, dict of tensors or other things\n    device: pytorch device or 'numpy'\n    callback: function that would be called on every sub-elements.\n    \"\"\"\n    if callback:\n        batch = callback(batch)\n\n    if isinstance(batch, dict):\n        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef c2w_to_tumpose(c2w):\n    \"\"\"\n    Convert a camera-to-world matrix to a tuple of translation and rotation\n\n    input: c2w: 4x4 matrix\n    output: tuple of translation and rotation (x y z qw qx qy qz)\n    \"\"\"\n    # convert input to numpy\n    c2w = to_numpy(c2w)\n    xyz = c2w[:3, -1]\n    rot = Rotation.from_matrix(c2w[:3, :3])","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.to_numpy","uri":"program://Human3R/function/eval.relpose.utils.to_numpy#L54-L55","kind":"function","name":"to_numpy","path":"eval/relpose/utils.py","language":"python","start_line":54,"end_line":55,"context_start_line":34,"context_end_line":75,"code":"        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef c2w_to_tumpose(c2w):\n    \"\"\"\n    Convert a camera-to-world matrix to a tuple of translation and rotation\n\n    input: c2w: 4x4 matrix\n    output: tuple of translation and rotation (x y z qw qx qy qz)\n    \"\"\"\n    # convert input to numpy\n    c2w = to_numpy(c2w)\n    xyz = c2w[:3, -1]\n    rot = Rotation.from_matrix(c2w[:3, :3])\n    qx, qy, qz, qw = rot.as_quat()\n    tum_pose = np.concatenate([xyz, [qw, qx, qy, qz]])\n    return tum_pose\n\n\ndef get_tum_poses(poses):\n    \"\"\"","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.c2w_to_tumpose","uri":"program://Human3R/function/eval.relpose.utils.c2w_to_tumpose#L58-L71","kind":"function","name":"c2w_to_tumpose","path":"eval/relpose/utils.py","language":"python","start_line":58,"end_line":71,"context_start_line":38,"context_end_line":91,"code":"\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef c2w_to_tumpose(c2w):\n    \"\"\"\n    Convert a camera-to-world matrix to a tuple of translation and rotation\n\n    input: c2w: 4x4 matrix\n    output: tuple of translation and rotation (x y z qw qx qy qz)\n    \"\"\"\n    # convert input to numpy\n    c2w = to_numpy(c2w)\n    xyz = c2w[:3, -1]\n    rot = Rotation.from_matrix(c2w[:3, :3])\n    qx, qy, qz, qw = rot.as_quat()\n    tum_pose = np.concatenate([xyz, [qw, qx, qy, qz]])\n    return tum_pose\n\n\ndef get_tum_poses(poses):\n    \"\"\"\n    poses: list of 4x4 arrays\n    \"\"\"\n    tt = np.arange(len(poses)).astype(float)\n    tum_poses = [c2w_to_tumpose(p) for p in poses]\n    tum_poses = np.stack(tum_poses, 0)\n    return [tum_poses, tt]\n\n\ndef save_tum_poses(poses, path):\n    traj = get_tum_poses(poses)\n    save_trajectory_tum_format(traj, path)\n    return traj[0]  # return the poses\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.get_tum_poses","uri":"program://Human3R/function/eval.relpose.utils.get_tum_poses#L74-L81","kind":"function","name":"get_tum_poses","path":"eval/relpose/utils.py","language":"python","start_line":74,"end_line":81,"context_start_line":54,"context_end_line":101,"code":"def to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef c2w_to_tumpose(c2w):\n    \"\"\"\n    Convert a camera-to-world matrix to a tuple of translation and rotation\n\n    input: c2w: 4x4 matrix\n    output: tuple of translation and rotation (x y z qw qx qy qz)\n    \"\"\"\n    # convert input to numpy\n    c2w = to_numpy(c2w)\n    xyz = c2w[:3, -1]\n    rot = Rotation.from_matrix(c2w[:3, :3])\n    qx, qy, qz, qw = rot.as_quat()\n    tum_pose = np.concatenate([xyz, [qw, qx, qy, qz]])\n    return tum_pose\n\n\ndef get_tum_poses(poses):\n    \"\"\"\n    poses: list of 4x4 arrays\n    \"\"\"\n    tt = np.arange(len(poses)).astype(float)\n    tum_poses = [c2w_to_tumpose(p) for p in poses]\n    tum_poses = np.stack(tum_poses, 0)\n    return [tum_poses, tt]\n\n\ndef save_tum_poses(poses, path):\n    traj = get_tum_poses(poses)\n    save_trajectory_tum_format(traj, path)\n    return traj[0]  # return the poses\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.save_tum_poses","uri":"program://Human3R/function/eval.relpose.utils.save_tum_poses#L84-L87","kind":"function","name":"save_tum_poses","path":"eval/relpose/utils.py","language":"python","start_line":84,"end_line":87,"context_start_line":64,"context_end_line":107,"code":"    \"\"\"\n    # convert input to numpy\n    c2w = to_numpy(c2w)\n    xyz = c2w[:3, -1]\n    rot = Rotation.from_matrix(c2w[:3, :3])\n    qx, qy, qz, qw = rot.as_quat()\n    tum_pose = np.concatenate([xyz, [qw, qx, qy, qz]])\n    return tum_pose\n\n\ndef get_tum_poses(poses):\n    \"\"\"\n    poses: list of 4x4 arrays\n    \"\"\"\n    tt = np.arange(len(poses)).astype(float)\n    tum_poses = [c2w_to_tumpose(p) for p in poses]\n    tum_poses = np.stack(tum_poses, 0)\n    return [tum_poses, tt]\n\n\ndef save_tum_poses(poses, path):\n    traj = get_tum_poses(poses)\n    save_trajectory_tum_format(traj, path)\n    return traj[0]  # return the poses\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.save_focals","uri":"program://Human3R/function/eval.relpose.utils.save_focals#L90-L94","kind":"function","name":"save_focals","path":"eval/relpose/utils.py","language":"python","start_line":90,"end_line":94,"context_start_line":70,"context_end_line":114,"code":"    tum_pose = np.concatenate([xyz, [qw, qx, qy, qz]])\n    return tum_pose\n\n\ndef get_tum_poses(poses):\n    \"\"\"\n    poses: list of 4x4 arrays\n    \"\"\"\n    tt = np.arange(len(poses)).astype(float)\n    tum_poses = [c2w_to_tumpose(p) for p in poses]\n    tum_poses = np.stack(tum_poses, 0)\n    return [tum_poses, tt]\n\n\ndef save_tum_poses(poses, path):\n    traj = get_tum_poses(poses)\n    save_trajectory_tum_format(traj, path)\n    return traj[0]  # return the poses\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.save_intrinsics","uri":"program://Human3R/function/eval.relpose.utils.save_intrinsics#L97-L104","kind":"function","name":"save_intrinsics","path":"eval/relpose/utils.py","language":"python","start_line":97,"end_line":104,"context_start_line":77,"context_end_line":124,"code":"    \"\"\"\n    tt = np.arange(len(poses)).astype(float)\n    tum_poses = [c2w_to_tumpose(p) for p in poses]\n    tum_poses = np.stack(tum_poses, 0)\n    return [tum_poses, tt]\n\n\ndef save_tum_poses(poses, path):\n    traj = get_tum_poses(poses)\n    save_trajectory_tum_format(traj, path)\n    return traj[0]  # return the poses\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.save_conf_maps","uri":"program://Human3R/function/eval.relpose.utils.save_conf_maps#L107-L110","kind":"function","name":"save_conf_maps","path":"eval/relpose/utils.py","language":"python","start_line":107,"end_line":110,"context_start_line":87,"context_end_line":130,"code":"    return traj[0]  # return the poses\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)\n    min_depth = depth_maps.min()  # float(torch.quantile(out, 0.01))\n    max_depth = depth_maps.max()  # float(torch.quantile(out, 0.99))\n    colored_depth = colorize(\n        depth_maps,\n        cmap_name=\"Spectral_r\",\n        range=(min_depth, max_depth),","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.save_rgb_imgs","uri":"program://Human3R/function/eval.relpose.utils.save_rgb_imgs#L113-L120","kind":"function","name":"save_rgb_imgs","path":"eval/relpose/utils.py","language":"python","start_line":113,"end_line":120,"context_start_line":93,"context_end_line":140,"code":"    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)\n    min_depth = depth_maps.min()  # float(torch.quantile(out, 0.01))\n    max_depth = depth_maps.max()  # float(torch.quantile(out, 0.99))\n    colored_depth = colorize(\n        depth_maps,\n        cmap_name=\"Spectral_r\",\n        range=(min_depth, max_depth),\n        append_cbar=True,\n    )\n    images = []\n\n    if conf_self is not None:\n        conf_selfs = torch.concat(conf_self, 0)\n        min_conf = torch.log(conf_selfs.min())  # float(torch.quantile(out, 0.01))\n        max_conf = torch.log(conf_selfs.max())  # float(torch.quantile(out, 0.99))\n        colored_conf = colorize(\n            torch.log(conf_selfs),","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.save_depth_maps","uri":"program://Human3R/function/eval.relpose.utils.save_depth_maps#L123-L166","kind":"function","name":"save_depth_maps","path":"eval/relpose/utils.py","language":"python","start_line":123,"end_line":166,"context_start_line":103,"context_end_line":186,"code":"    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)\n    min_depth = depth_maps.min()  # float(torch.quantile(out, 0.01))\n    max_depth = depth_maps.max()  # float(torch.quantile(out, 0.99))\n    colored_depth = colorize(\n        depth_maps,\n        cmap_name=\"Spectral_r\",\n        range=(min_depth, max_depth),\n        append_cbar=True,\n    )\n    images = []\n\n    if conf_self is not None:\n        conf_selfs = torch.concat(conf_self, 0)\n        min_conf = torch.log(conf_selfs.min())  # float(torch.quantile(out, 0.01))\n        max_conf = torch.log(conf_selfs.max())  # float(torch.quantile(out, 0.99))\n        colored_conf = colorize(\n            torch.log(conf_selfs),\n            cmap_name=\"jet\",\n            range=(min_conf, max_conf),\n            append_cbar=True,\n        )\n\n    for i, depth_map in enumerate(colored_depth):\n        # Apply color map to depth map\n        img_path = f\"{path}/frame_{(i):04d}.png\"\n        if conf_self is None:\n            to_save = (depth_map * 255).detach().cpu().numpy().astype(np.uint8)\n        else:\n            to_save = torch.cat([depth_map, colored_conf[i]], dim=1)\n            to_save = (to_save * 255).detach().cpu().numpy().astype(np.uint8)\n        iio.imwrite(img_path, to_save)\n        images.append(Image.open(img_path))\n        np.save(f\"{path}/frame_{(i):04d}.npy\", depth_maps[i].detach().cpu().numpy())\n\n    images[0].save(\n        f\"{path}/_depth_maps.gif\",\n        save_all=True,\n        append_images=images[1:],\n        duration=100,\n        loop=0,\n    )\n\n    return depth_maps\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    # Do some plotting.\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.get_vertical_colorbar","uri":"program://Human3R/function/eval.relpose.utils.get_vertical_colorbar#L169-L216","kind":"function","name":"get_vertical_colorbar","path":"eval/relpose/utils.py","language":"python","start_line":169,"end_line":216,"context_start_line":149,"context_end_line":236,"code":"        if conf_self is None:\n            to_save = (depth_map * 255).detach().cpu().numpy().astype(np.uint8)\n        else:\n            to_save = torch.cat([depth_map, colored_conf[i]], dim=1)\n            to_save = (to_save * 255).detach().cpu().numpy().astype(np.uint8)\n        iio.imwrite(img_path, to_save)\n        images.append(Image.open(img_path))\n        np.save(f\"{path}/frame_{(i):04d}.npy\", depth_maps[i].detach().cpu().numpy())\n\n    images[0].save(\n        f\"{path}/_depth_maps.gif\",\n        save_all=True,\n        append_images=images[1:],\n        duration=100,\n        loop=0,\n    )\n\n    return depth_maps\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    # Do some plotting.\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n\n    tick_cnt = 6\n    tick_loc = np.linspace(vmin, vmax, tick_cnt)\n    cb1 = mpl.colorbar.ColorbarBase(\n        ax, cmap=cmap, norm=norm, ticks=tick_loc, orientation=\"vertical\"\n    )\n\n    tick_label = [str(np.round(x, cbar_precision)) for x in tick_loc]\n    if cbar_precision == 0:\n        tick_label = [x[:-2] for x in tick_label]\n\n    cb1.set_ticklabels(tick_label)\n\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    # fig.tight_layout()\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.colorize_np","uri":"program://Human3R/function/eval.relpose.utils.colorize_np#L219-L279","kind":"function","name":"colorize_np","path":"eval/relpose/utils.py","language":"python","start_line":219,"end_line":279,"context_start_line":199,"context_end_line":299,"code":"\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    # fig.tight_layout()\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]\n    \"\"\"\n    if range is not None:\n        vmin, vmax = range\n    elif mask is not None:\n        # vmin, vmax = np.percentile(x[mask], (2, 100))\n        vmin = np.min(x[mask][np.nonzero(x[mask])])\n        vmax = np.max(x[mask])\n        # vmin = vmin - np.abs(vmin) * 0.01\n        x[np.logical_not(mask)] = vmin\n        # print(vmin, vmax)\n    else:\n        vmin, vmax = np.percentile(x, (1, 100))\n        vmax += 1e-6\n\n    x = np.clip(x, vmin, vmax)\n    x = (x - vmin) / (vmax - vmin)\n    # x = np.clip(x, 0., 1.)\n\n    cmap = cm.get_cmap(cmap_name)\n    x_new = cmap(x)[:, :, :3]\n\n    if mask is not None:\n        mask = np.float32(mask[:, :, np.newaxis])\n        x_new = x_new * mask + np.ones_like(x_new) * (1.0 - mask)\n\n    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\n# tensor\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.utils.colorize","uri":"program://Human3R/function/eval.relpose.utils.colorize#L283-L311","kind":"function","name":"colorize","path":"eval/relpose/utils.py","language":"python","start_line":283,"end_line":311,"context_start_line":263,"context_end_line":311,"code":"        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\n# tensor\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.launch","uri":"program://Human3R/module/eval.relpose.launch#L1-L492","kind":"module","name":"eval.relpose.launch","path":"eval/relpose/launch.py","language":"python","start_line":1,"end_line":492,"context_start_line":1,"context_end_line":492,"code":"import os\nimport sys\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nimport math\nimport cv2\nimport numpy as np\nimport torch\nimport argparse\n\nfrom copy import deepcopy\nfrom eval.relpose.metadata import dataset_metadata\nfrom eval.relpose.utils import *\n\nfrom accelerate import PartialState\nfrom add_ckpt_path import add_path_to_dust3r\n\nfrom tqdm import tqdm\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--weights\",\n        type=str,\n        help=\"path to the model weights\",\n        default=\"\",\n    )\n\n    parser.add_argument(\"--device\", type=str, default=\"cuda\", help=\"pytorch device\")\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"\",\n        help=\"value for outdir\",\n    )\n    parser.add_argument(\n        \"--no_crop\", type=bool, default=True, help=\"whether to crop input data\"\n    )\n\n    parser.add_argument(\n        \"--eval_dataset\",\n        type=str,\n        default=\"sintel\",\n        choices=list(dataset_metadata.keys()),\n    )\n    parser.add_argument(\"--size\", type=int, default=\"224\")\n\n    parser.add_argument(\n        \"--pose_eval_stride\", default=1, type=int, help=\"stride for pose evaluation\"\n    )\n    parser.add_argument(\"--shuffle\", action=\"store_true\", default=False)\n    parser.add_argument(\n        \"--full_seq\",\n        action=\"store_true\",\n        default=False,\n        help=\"use full sequence for pose evaluation\",\n    )\n    parser.add_argument(\n        \"--seq_list\",\n        nargs=\"+\",\n        default=None,\n        help=\"list of sequences for pose evaluation\",\n    )\n\n    parser.add_argument(\"--revisit\", type=int, default=1)\n    parser.add_argument(\"--freeze_state\", action=\"store_true\", default=False)\n    parser.add_argument(\"--solve_pose\", action=\"store_true\", default=False)\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    return parser\n\n\ndef eval_pose_estimation(args, model, save_dir=None):\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mask_path = metadata[\"mask_path\"]\n\n    ate_mean, rpe_trans_mean, rpe_rot_mean = eval_pose_estimation_dist(\n        args, model, save_dir=save_dir, img_path=img_path, mask_path=mask_path\n    )\n    return ate_mean, rpe_trans_mean, rpe_rot_mean\n\n\ndef eval_pose_estimation_dist(args, model, img_path, save_dir=None, mask_path=None):\n    from dust3r.inference import inference, inference_recurrent_lighter\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    anno_path = metadata.get(\"anno_path\", None)\n\n    seq_list = args.seq_list\n    if seq_list is None:\n        if metadata.get(\"full_seq\", False):\n            args.full_seq = True\n        else:\n            seq_list = metadata.get(\"seq_list\", [])\n        if args.full_seq:\n            seq_list = os.listdir(img_path)\n            seq_list = [\n                seq for seq in seq_list if os.path.isdir(os.path.join(img_path, seq))\n            ]\n        seq_list = sorted(seq_list)\n\n    if save_dir is None:\n        save_dir = args.output_dir\n\n    distributed_state = PartialState()\n    model.to(distributed_state.device)\n    device = distributed_state.device\n    img_res = getattr(model, 'mhmr_img_res', None)\n\n    with distributed_state.split_between_processes(seq_list) as seqs:\n        ate_list = []\n        rpe_trans_list = []\n        rpe_rot_list = []\n        load_img_size = args.size\n        error_log_path = f\"{save_dir}/_error_log_{distributed_state.process_index}.txt\"  # Unique log file per process\n        bug = False\n        for seq in tqdm(seqs):\n            try:\n                dir_path = metadata[\"dir_path_func\"](img_path, seq)\n\n                # Handle skip_condition\n                skip_condition = metadata.get(\"skip_condition\", None)\n                if skip_condition is not None and skip_condition(save_dir, seq):\n                    continue\n\n                mask_path_seq_func = metadata.get(\n                    \"mask_path_seq_func\", lambda mask_path, seq: None\n                )\n                mask_path_seq = mask_path_seq_func(mask_path, seq)\n\n                filelist = [\n                    os.path.join(dir_path, name) for name in os.listdir(dir_path)\n                ]\n                filelist.sort()\n                filelist = filelist[:: args.pose_eval_stride]\n\n                views = prepare_input(\n                    filelist,\n                    [True for _ in filelist],\n                    size=load_img_size,\n                    crop=not args.no_crop,\n                    revisit=args.revisit,\n                    update=not args.freeze_state,\n                    img_res=img_res,\n                    reset_interval=args.reset_interval,\n                )\n                # outputs, _ = inference(views, model, device)\n                outputs, _ = inference_recurrent_lighter(\n                    views, model, device, verbose=False, use_ttt3r=args.use_ttt3r)\n                \n                (\n                    colors,\n                    pts3ds_self,\n                    pts3ds_other,\n                    conf_self,\n                    conf_other,\n                    cam_dict,\n                    pr_poses,\n                ) = prepare_output(\n                    outputs, revisit=args.revisit, solve_pose=args.solve_pose\n                )\n\n                pred_traj = get_tum_poses(pr_poses)\n                os.makedirs(f\"{save_dir}/{seq}\", exist_ok=True)\n                save_tum_poses(pr_poses, f\"{save_dir}/{seq}/pred_traj.txt\")\n                save_focals(cam_dict, f\"{save_dir}/{seq}/pred_focal.txt\")\n                save_intrinsics(cam_dict, f\"{save_dir}/{seq}/pred_intrinsics.txt\")\n                # save_depth_maps(pts3ds_self,f'{save_dir}/{seq}', conf_self=conf_self)\n                # save_conf_maps(conf_self,f'{save_dir}/{seq}')\n                # save_rgb_imgs(colors,f'{save_dir}/{seq}')\n\n                gt_traj_file = metadata[\"gt_traj_func\"](img_path, anno_path, seq)\n                traj_format = metadata.get(\"traj_format\", None)\n\n                if args.eval_dataset == \"sintel\":\n                    gt_traj = load_traj(\n                        gt_traj_file=gt_traj_file, stride=args.pose_eval_stride\n                    )\n                elif traj_format is not None:\n                    gt_traj = load_traj(\n                        gt_traj_file=gt_traj_file,\n                        traj_format=traj_format,\n                        stride=args.pose_eval_stride,\n                    )\n                else:\n                    gt_traj = None\n\n                if gt_traj is not None:\n                    ate, rpe_trans, rpe_rot = eval_metrics(\n                        pred_traj,\n                        gt_traj,\n                        seq=seq,\n                        filename=f\"{save_dir}/{seq}_eval_metric.txt\",\n                    )\n                    plot_trajectory(\n                        pred_traj, gt_traj, title=seq, filename=f\"{save_dir}/{seq}.png\"\n                    )\n                else:\n                    ate, rpe_trans, rpe_rot = 0, 0, 0\n                    bug = True\n\n                ate_list.append(ate)\n                rpe_trans_list.append(rpe_trans)\n                rpe_rot_list.append(rpe_rot)\n\n                # Write to error log after each sequence\n                with open(error_log_path, \"a\") as f:\n                    f.write(\n                        f\"{args.eval_dataset}-{seq: <16} | ATE: {ate:.5f}, RPE trans: {rpe_trans:.5f}, RPE rot: {rpe_rot:.5f}\\n\"\n                    )\n                    f.write(f\"{ate:.5f}\\n\")\n                    f.write(f\"{rpe_trans:.5f}\\n\")\n                    f.write(f\"{rpe_rot:.5f}\\n\")\n\n            except Exception as e:\n                if \"out of memory\" in str(e):\n                    # Handle OOM\n                    torch.cuda.empty_cache()  # Clear the CUDA memory\n                    with open(error_log_path, \"a\") as f:\n                        f.write(\n                            f\"OOM error in sequence {seq}, skipping this sequence.\\n\"\n                        )\n                    print(f\"OOM error in sequence {seq}, skipping...\")\n                elif \"Degenerate covariance rank\" in str(\n                    e\n                ) or \"Eigenvalues did not converge\" in str(e):\n                    # Handle Degenerate covariance rank exception and Eigenvalues did not converge exception\n                    with open(error_log_path, \"a\") as f:\n                        f.write(f\"Exception in sequence {seq}: {str(e)}\\n\")\n                    print(f\"Traj evaluation error in sequence {seq}, skipping.\")\n                else:\n                    raise e  # Rethrow if it's not an expected exception\n\n    distributed_state.wait_for_everyone()\n\n    results = process_directory(save_dir)\n    avg_ate, avg_rpe_trans, avg_rpe_rot = calculate_averages(results)\n\n    # Write the averages to the error log (only on the main process)\n    if distributed_state.is_main_process:\n        with open(f\"{save_dir}/_error_log.txt\", \"a\") as f:\n            # Copy the error log from each process to the main error log\n            for i in range(distributed_state.num_processes):\n                if not os.path.exists(f\"{save_dir}/_error_log_{i}.txt\"):\n                    break\n                with open(f\"{save_dir}/_error_log_{i}.txt\", \"r\") as f_sub:\n                    f.write(f_sub.read())\n            f.write(\n                f\"Average ATE: {avg_ate:.5f}, Average RPE trans: {avg_rpe_trans:.5f}, Average RPE rot: {avg_rpe_rot:.5f}\\n\"\n            )\n\n    return avg_ate, avg_rpe_trans, avg_rpe_rot\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    add_path_to_dust3r(args.weights)\n    from dust3r.utils.image import load_images_for_eval as load_images\n    from src.dust3r.utils.image import pad_image\n    from dust3r.post_process import estimate_focal_knowing_depth\n    from dust3r.model import ARCroco3DStereo\n    from dust3r.utils.camera import pose_encoding_to_camera\n    from dust3r.utils.geometry import weighted_procrustes, geotrf, matrix_cumprod, get_camera_parameters\n\n    args.full_seq = False\n    args.no_crop = False\n\n    def recover_cam_params(pts3ds_self, pts3ds_other, conf_self, conf_other):\n        B, H, W, _ = pts3ds_self.shape\n        pp = (\n            torch.tensor([W // 2, H // 2], device=pts3ds_self.device)\n            .float()\n            .repeat(B, 1)\n            .reshape(B, 1, 2)\n        )\n        focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n        pts3ds_self = pts3ds_self.reshape(B, -1, 3)\n        pts3ds_other = pts3ds_other.reshape(B, -1, 3)\n        conf_self = conf_self.reshape(B, -1)\n        conf_other = conf_other.reshape(B, -1)\n        # weighted procrustes\n        c2w = weighted_procrustes(\n            pts3ds_self,\n            pts3ds_other,\n            torch.log(conf_self) * torch.log(conf_other),\n            use_weights=True,\n            return_T=True,\n        )\n        return c2w, focal, pp.reshape(B, 2)\n\n    def prepare_input(\n        img_paths,\n        img_mask,\n        size,\n        raymaps=None,\n        raymap_mask=None,\n        revisit=1,\n        update=True,\n        crop=True,\n        img_res=None, \n        reset_interval=100,\n    ):\n        images = load_images(img_paths, size=size, crop=crop)\n        if img_res is not None:\n            K_mhmr = get_camera_parameters(img_res, device=\"cpu\") # if use pseudo K\n\n        views = []\n        if raymaps is None and raymap_mask is None:\n            # Only images are provided.\n            num_views = len(images)\n            for i in range(num_views):\n                view = {\n                    \"img\": images[i][\"img\"],\n                    \"ray_map\": torch.full(\n                        (\n                            images[i][\"img\"].shape[0],\n                            6,\n                            images[i][\"img\"].shape[-2],\n                            images[i][\"img\"].shape[-1],\n                        ),\n                        torch.nan,\n                    ),\n                    \"true_shape\": torch.from_numpy(images[i][\"true_shape\"]),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        np.eye(4).astype(np.float32)\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(True).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(False).unsqueeze(0),\n                    \"update\": torch.tensor(True).unsqueeze(0),\n                    # \"reset\": torch.tensor(False).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n                if img_res is not None:\n                    view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                    view[\"K_mhmr\"] = K_mhmr\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n        else:\n            # Combine images and raymaps.\n            num_views = len(images) + len(raymaps)\n            assert len(img_mask) == len(raymap_mask) == num_views\n            assert sum(img_mask) == len(images) and sum(raymap_mask) == len(raymaps)\n\n            j = 0\n            k = 0\n            for i in range(num_views):\n                view = {\n                    \"img\": (\n                        images[j][\"img\"]\n                        if img_mask[i]\n                        else torch.full_like(images[0][\"img\"], torch.nan)\n                    ),\n                    \"ray_map\": (\n                        raymaps[k]\n                        if raymap_mask[i]\n                        else torch.full_like(raymaps[0], torch.nan)\n                    ),\n                    \"true_shape\": (\n                        torch.from_numpy(images[j][\"true_shape\"])\n                        if img_mask[i]\n                        else torch.from_numpy(np.int32([raymaps[k].shape[1:-1][::-1]]))\n                    ),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        np.eye(4).astype(np.float32)\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(raymap_mask[i]).unsqueeze(0),\n                    \"update\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    # \"reset\": torch.tensor(False).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n                if img_res is not None:\n                    view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                    view[\"K_mhmr\"] = K_mhmr\n                if img_mask[i]:\n                    j += 1\n                if raymap_mask[i]:\n                    k += 1\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n            assert j == len(images) and k == len(raymaps)\n\n        if revisit > 1:\n            # repeat input for 'revisit' times\n            new_views = []\n            for r in range(revisit):\n                for i in range(len(views)):\n                    new_view = deepcopy(views[i])\n                    new_view[\"idx\"] = r * len(views) + i\n                    new_view[\"instance\"] = str(r * len(views) + i)\n                    if r > 0:\n                        if not update:\n                            new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                    new_views.append(new_view)\n            return new_views\n        return views\n\n    def prepare_output(outputs, revisit=1, solve_pose=False):\n        valid_length = len(outputs[\"pred\"]) // revisit\n        outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n        outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n        # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n        reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n        shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n        outputs[\"pred\"] = [\n            pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n        reset_mask = reset_mask[~shifted_reset_mask]\n\n        if solve_pose:\n            pts3ds_self = [\n                output[\"pts3d_in_self_view\"].cpu() for output in outputs[\"pred\"]\n            ]\n            pts3ds_other = [\n                output[\"pts3d_in_other_view\"].cpu() for output in outputs[\"pred\"]\n            ]\n            conf_self = [output[\"conf_self\"].cpu() for output in outputs[\"pred\"]]\n            conf_other = [output[\"conf\"].cpu() for output in outputs[\"pred\"]]\n            pr_poses, focal, pp = recover_cam_params(\n                torch.cat(pts3ds_self, 0),\n                torch.cat(pts3ds_other, 0),\n                torch.cat(conf_self, 0),\n                torch.cat(conf_other, 0),\n            )\n            pts3ds_self = torch.cat(pts3ds_self, 0)\n        else:\n\n            pts3ds_self = [\n                output[\"pts3d_in_self_view\"].cpu() for output in outputs[\"pred\"]\n            ]\n            pts3ds_other = [\n                output[\"pts3d_in_other_view\"].cpu() for output in outputs[\"pred\"]\n            ]\n            conf_self = [output[\"conf_self\"].cpu() for output in outputs[\"pred\"]]\n            conf_other = [output[\"conf\"].cpu() for output in outputs[\"pred\"]]\n            pts3ds_self = torch.cat(pts3ds_self, 0)\n            pr_poses = [\n                pose_encoding_to_camera(pred[\"camera_pose\"].clone()).cpu()\n                for pred in outputs[\"pred\"]\n            ]\n            pr_poses = torch.cat(pr_poses, 0)\n\n            B, H, W, _ = pts3ds_self.shape\n            pp = (\n                torch.tensor([W // 2, H // 2], device=pts3ds_self.device)\n                .float()\n                .repeat(B, 1)\n                .reshape(B, 2)\n            )\n            focal = estimate_focal_knowing_depth(\n                pts3ds_self, pp, focal_mode=\"weiszfeld\"\n            )\n\n        if reset_mask.any():\n            identity = torch.eye(4, device=pr_poses.device)\n            reset_poses = torch.where(reset_mask.unsqueeze(-1).unsqueeze(-1), pr_poses, identity)\n            cumulative_bases = matrix_cumprod(reset_poses)\n            shifted_bases = torch.cat([identity.unsqueeze(0), cumulative_bases[:-1]], dim=0)\n            pr_poses = torch.einsum('bij,bjk->bik', shifted_bases, pr_poses)\n\n        colors = [0.5 * (output[\"rgb\"][0] + 1.0) for output in outputs[\"pred\"]]\n        cam_dict = {\n            \"focal\": focal.cpu().numpy(),\n            \"pp\": pp.cpu().numpy(),\n        }\n        return (\n            colors,\n            pts3ds_self,\n            pts3ds_other,\n            conf_self,\n            conf_other,\n            cam_dict,\n            pr_poses,\n        )\n\n    model = ARCroco3DStereo.from_pretrained(args.weights)\n    eval_pose_estimation(args, model, save_dir=args.output_dir)","source_hash":"9427c203a9f9a49ee297b62876dd88027e5fc38e89eafb54f2042f1aee1bb8b7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.launch.get_args_parser","uri":"program://Human3R/function/eval.relpose.launch.get_args_parser#L21-L72","kind":"function","name":"get_args_parser","path":"eval/relpose/launch.py","language":"python","start_line":21,"end_line":72,"context_start_line":1,"context_end_line":92,"code":"import os\nimport sys\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nimport math\nimport cv2\nimport numpy as np\nimport torch\nimport argparse\n\nfrom copy import deepcopy\nfrom eval.relpose.metadata import dataset_metadata\nfrom eval.relpose.utils import *\n\nfrom accelerate import PartialState\nfrom add_ckpt_path import add_path_to_dust3r\n\nfrom tqdm import tqdm\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--weights\",\n        type=str,\n        help=\"path to the model weights\",\n        default=\"\",\n    )\n\n    parser.add_argument(\"--device\", type=str, default=\"cuda\", help=\"pytorch device\")\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"\",\n        help=\"value for outdir\",\n    )\n    parser.add_argument(\n        \"--no_crop\", type=bool, default=True, help=\"whether to crop input data\"\n    )\n\n    parser.add_argument(\n        \"--eval_dataset\",\n        type=str,\n        default=\"sintel\",\n        choices=list(dataset_metadata.keys()),\n    )\n    parser.add_argument(\"--size\", type=int, default=\"224\")\n\n    parser.add_argument(\n        \"--pose_eval_stride\", default=1, type=int, help=\"stride for pose evaluation\"\n    )\n    parser.add_argument(\"--shuffle\", action=\"store_true\", default=False)\n    parser.add_argument(\n        \"--full_seq\",\n        action=\"store_true\",\n        default=False,\n        help=\"use full sequence for pose evaluation\",\n    )\n    parser.add_argument(\n        \"--seq_list\",\n        nargs=\"+\",\n        default=None,\n        help=\"list of sequences for pose evaluation\",\n    )\n\n    parser.add_argument(\"--revisit\", type=int, default=1)\n    parser.add_argument(\"--freeze_state\", action=\"store_true\", default=False)\n    parser.add_argument(\"--solve_pose\", action=\"store_true\", default=False)\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    return parser\n\n\ndef eval_pose_estimation(args, model, save_dir=None):\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mask_path = metadata[\"mask_path\"]\n\n    ate_mean, rpe_trans_mean, rpe_rot_mean = eval_pose_estimation_dist(\n        args, model, save_dir=save_dir, img_path=img_path, mask_path=mask_path\n    )\n    return ate_mean, rpe_trans_mean, rpe_rot_mean\n\n\ndef eval_pose_estimation_dist(args, model, img_path, save_dir=None, mask_path=None):\n    from dust3r.inference import inference, inference_recurrent_lighter\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    anno_path = metadata.get(\"anno_path\", None)\n\n    seq_list = args.seq_list","source_hash":"9427c203a9f9a49ee297b62876dd88027e5fc38e89eafb54f2042f1aee1bb8b7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.launch.eval_pose_estimation","uri":"program://Human3R/function/eval.relpose.launch.eval_pose_estimation#L75-L83","kind":"function","name":"eval_pose_estimation","path":"eval/relpose/launch.py","language":"python","start_line":75,"end_line":83,"context_start_line":55,"context_end_line":103,"code":"        \"--full_seq\",\n        action=\"store_true\",\n        default=False,\n        help=\"use full sequence for pose evaluation\",\n    )\n    parser.add_argument(\n        \"--seq_list\",\n        nargs=\"+\",\n        default=None,\n        help=\"list of sequences for pose evaluation\",\n    )\n\n    parser.add_argument(\"--revisit\", type=int, default=1)\n    parser.add_argument(\"--freeze_state\", action=\"store_true\", default=False)\n    parser.add_argument(\"--solve_pose\", action=\"store_true\", default=False)\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    return parser\n\n\ndef eval_pose_estimation(args, model, save_dir=None):\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mask_path = metadata[\"mask_path\"]\n\n    ate_mean, rpe_trans_mean, rpe_rot_mean = eval_pose_estimation_dist(\n        args, model, save_dir=save_dir, img_path=img_path, mask_path=mask_path\n    )\n    return ate_mean, rpe_trans_mean, rpe_rot_mean\n\n\ndef eval_pose_estimation_dist(args, model, img_path, save_dir=None, mask_path=None):\n    from dust3r.inference import inference, inference_recurrent_lighter\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    anno_path = metadata.get(\"anno_path\", None)\n\n    seq_list = args.seq_list\n    if seq_list is None:\n        if metadata.get(\"full_seq\", False):\n            args.full_seq = True\n        else:\n            seq_list = metadata.get(\"seq_list\", [])\n        if args.full_seq:\n            seq_list = os.listdir(img_path)\n            seq_list = [\n                seq for seq in seq_list if os.path.isdir(os.path.join(img_path, seq))\n            ]\n        seq_list = sorted(seq_list)","source_hash":"9427c203a9f9a49ee297b62876dd88027e5fc38e89eafb54f2042f1aee1bb8b7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.launch.eval_pose_estimation_dist","uri":"program://Human3R/function/eval.relpose.launch.eval_pose_estimation_dist#L86-L255","kind":"function","name":"eval_pose_estimation_dist","path":"eval/relpose/launch.py","language":"python","start_line":86,"end_line":255,"context_start_line":66,"context_end_line":275,"code":"\n    parser.add_argument(\"--revisit\", type=int, default=1)\n    parser.add_argument(\"--freeze_state\", action=\"store_true\", default=False)\n    parser.add_argument(\"--solve_pose\", action=\"store_true\", default=False)\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    return parser\n\n\ndef eval_pose_estimation(args, model, save_dir=None):\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mask_path = metadata[\"mask_path\"]\n\n    ate_mean, rpe_trans_mean, rpe_rot_mean = eval_pose_estimation_dist(\n        args, model, save_dir=save_dir, img_path=img_path, mask_path=mask_path\n    )\n    return ate_mean, rpe_trans_mean, rpe_rot_mean\n\n\ndef eval_pose_estimation_dist(args, model, img_path, save_dir=None, mask_path=None):\n    from dust3r.inference import inference, inference_recurrent_lighter\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    anno_path = metadata.get(\"anno_path\", None)\n\n    seq_list = args.seq_list\n    if seq_list is None:\n        if metadata.get(\"full_seq\", False):\n            args.full_seq = True\n        else:\n            seq_list = metadata.get(\"seq_list\", [])\n        if args.full_seq:\n            seq_list = os.listdir(img_path)\n            seq_list = [\n                seq for seq in seq_list if os.path.isdir(os.path.join(img_path, seq))\n            ]\n        seq_list = sorted(seq_list)\n\n    if save_dir is None:\n        save_dir = args.output_dir\n\n    distributed_state = PartialState()\n    model.to(distributed_state.device)\n    device = distributed_state.device\n    img_res = getattr(model, 'mhmr_img_res', None)\n\n    with distributed_state.split_between_processes(seq_list) as seqs:\n        ate_list = []\n        rpe_trans_list = []\n        rpe_rot_list = []\n        load_img_size = args.size\n        error_log_path = f\"{save_dir}/_error_log_{distributed_state.process_index}.txt\"  # Unique log file per process\n        bug = False\n        for seq in tqdm(seqs):\n            try:\n                dir_path = metadata[\"dir_path_func\"](img_path, seq)\n\n                # Handle skip_condition\n                skip_condition = metadata.get(\"skip_condition\", None)\n                if skip_condition is not None and skip_condition(save_dir, seq):\n                    continue\n\n                mask_path_seq_func = metadata.get(\n                    \"mask_path_seq_func\", lambda mask_path, seq: None\n                )\n                mask_path_seq = mask_path_seq_func(mask_path, seq)\n\n                filelist = [\n                    os.path.join(dir_path, name) for name in os.listdir(dir_path)\n                ]\n                filelist.sort()\n                filelist = filelist[:: args.pose_eval_stride]\n\n                views = prepare_input(\n                    filelist,\n                    [True for _ in filelist],\n                    size=load_img_size,\n                    crop=not args.no_crop,\n                    revisit=args.revisit,\n                    update=not args.freeze_state,\n                    img_res=img_res,\n                    reset_interval=args.reset_interval,\n                )\n                # outputs, _ = inference(views, model, device)\n                outputs, _ = inference_recurrent_lighter(\n                    views, model, device, verbose=False, use_ttt3r=args.use_ttt3r)\n                \n                (\n                    colors,\n                    pts3ds_self,\n                    pts3ds_other,\n                    conf_self,\n                    conf_other,\n                    cam_dict,\n                    pr_poses,\n                ) = prepare_output(\n                    outputs, revisit=args.revisit, solve_pose=args.solve_pose\n                )\n\n                pred_traj = get_tum_poses(pr_poses)\n                os.makedirs(f\"{save_dir}/{seq}\", exist_ok=True)\n                save_tum_poses(pr_poses, f\"{save_dir}/{seq}/pred_traj.txt\")\n                save_focals(cam_dict, f\"{save_dir}/{seq}/pred_focal.txt\")\n                save_intrinsics(cam_dict, f\"{save_dir}/{seq}/pred_intrinsics.txt\")\n                # save_depth_maps(pts3ds_self,f'{save_dir}/{seq}', conf_self=conf_self)\n                # save_conf_maps(conf_self,f'{save_dir}/{seq}')\n                # save_rgb_imgs(colors,f'{save_dir}/{seq}')\n\n                gt_traj_file = metadata[\"gt_traj_func\"](img_path, anno_path, seq)\n                traj_format = metadata.get(\"traj_format\", None)\n\n                if args.eval_dataset == \"sintel\":\n                    gt_traj = load_traj(\n                        gt_traj_file=gt_traj_file, stride=args.pose_eval_stride\n                    )\n                elif traj_format is not None:\n                    gt_traj = load_traj(\n                        gt_traj_file=gt_traj_file,\n                        traj_format=traj_format,\n                        stride=args.pose_eval_stride,\n                    )\n                else:\n                    gt_traj = None\n\n                if gt_traj is not None:\n                    ate, rpe_trans, rpe_rot = eval_metrics(\n                        pred_traj,\n                        gt_traj,\n                        seq=seq,\n                        filename=f\"{save_dir}/{seq}_eval_metric.txt\",\n                    )\n                    plot_trajectory(\n                        pred_traj, gt_traj, title=seq, filename=f\"{save_dir}/{seq}.png\"\n                    )\n                else:\n                    ate, rpe_trans, rpe_rot = 0, 0, 0\n                    bug = True\n\n                ate_list.append(ate)\n                rpe_trans_list.append(rpe_trans)\n                rpe_rot_list.append(rpe_rot)\n\n                # Write to error log after each sequence\n                with open(error_log_path, \"a\") as f:\n                    f.write(\n                        f\"{args.eval_dataset}-{seq: <16} | ATE: {ate:.5f}, RPE trans: {rpe_trans:.5f}, RPE rot: {rpe_rot:.5f}\\n\"\n                    )\n                    f.write(f\"{ate:.5f}\\n\")\n                    f.write(f\"{rpe_trans:.5f}\\n\")\n                    f.write(f\"{rpe_rot:.5f}\\n\")\n\n            except Exception as e:\n                if \"out of memory\" in str(e):\n                    # Handle OOM\n                    torch.cuda.empty_cache()  # Clear the CUDA memory\n                    with open(error_log_path, \"a\") as f:\n                        f.write(\n                            f\"OOM error in sequence {seq}, skipping this sequence.\\n\"\n                        )\n                    print(f\"OOM error in sequence {seq}, skipping...\")\n                elif \"Degenerate covariance rank\" in str(\n                    e\n                ) or \"Eigenvalues did not converge\" in str(e):\n                    # Handle Degenerate covariance rank exception and Eigenvalues did not converge exception\n                    with open(error_log_path, \"a\") as f:\n                        f.write(f\"Exception in sequence {seq}: {str(e)}\\n\")\n                    print(f\"Traj evaluation error in sequence {seq}, skipping.\")\n                else:\n                    raise e  # Rethrow if it's not an expected exception\n\n    distributed_state.wait_for_everyone()\n\n    results = process_directory(save_dir)\n    avg_ate, avg_rpe_trans, avg_rpe_rot = calculate_averages(results)\n\n    # Write the averages to the error log (only on the main process)\n    if distributed_state.is_main_process:\n        with open(f\"{save_dir}/_error_log.txt\", \"a\") as f:\n            # Copy the error log from each process to the main error log\n            for i in range(distributed_state.num_processes):\n                if not os.path.exists(f\"{save_dir}/_error_log_{i}.txt\"):\n                    break\n                with open(f\"{save_dir}/_error_log_{i}.txt\", \"r\") as f_sub:\n                    f.write(f_sub.read())\n            f.write(\n                f\"Average ATE: {avg_ate:.5f}, Average RPE trans: {avg_rpe_trans:.5f}, Average RPE rot: {avg_rpe_rot:.5f}\\n\"\n            )\n\n    return avg_ate, avg_rpe_trans, avg_rpe_rot\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    add_path_to_dust3r(args.weights)\n    from dust3r.utils.image import load_images_for_eval as load_images\n    from src.dust3r.utils.image import pad_image\n    from dust3r.post_process import estimate_focal_knowing_depth\n    from dust3r.model import ARCroco3DStereo\n    from dust3r.utils.camera import pose_encoding_to_camera\n    from dust3r.utils.geometry import weighted_procrustes, geotrf, matrix_cumprod, get_camera_parameters\n\n    args.full_seq = False\n    args.no_crop = False\n\n    def recover_cam_params(pts3ds_self, pts3ds_other, conf_self, conf_other):\n        B, H, W, _ = pts3ds_self.shape\n        pp = (\n            torch.tensor([W // 2, H // 2], device=pts3ds_self.device)","source_hash":"9427c203a9f9a49ee297b62876dd88027e5fc38e89eafb54f2042f1aee1bb8b7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.launch.recover_cam_params","uri":"program://Human3R/function/eval.relpose.launch.recover_cam_params#L272-L294","kind":"function","name":"recover_cam_params","path":"eval/relpose/launch.py","language":"python","start_line":272,"end_line":294,"context_start_line":252,"context_end_line":314,"code":"                f\"Average ATE: {avg_ate:.5f}, Average RPE trans: {avg_rpe_trans:.5f}, Average RPE rot: {avg_rpe_rot:.5f}\\n\"\n            )\n\n    return avg_ate, avg_rpe_trans, avg_rpe_rot\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    add_path_to_dust3r(args.weights)\n    from dust3r.utils.image import load_images_for_eval as load_images\n    from src.dust3r.utils.image import pad_image\n    from dust3r.post_process import estimate_focal_knowing_depth\n    from dust3r.model import ARCroco3DStereo\n    from dust3r.utils.camera import pose_encoding_to_camera\n    from dust3r.utils.geometry import weighted_procrustes, geotrf, matrix_cumprod, get_camera_parameters\n\n    args.full_seq = False\n    args.no_crop = False\n\n    def recover_cam_params(pts3ds_self, pts3ds_other, conf_self, conf_other):\n        B, H, W, _ = pts3ds_self.shape\n        pp = (\n            torch.tensor([W // 2, H // 2], device=pts3ds_self.device)\n            .float()\n            .repeat(B, 1)\n            .reshape(B, 1, 2)\n        )\n        focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n        pts3ds_self = pts3ds_self.reshape(B, -1, 3)\n        pts3ds_other = pts3ds_other.reshape(B, -1, 3)\n        conf_self = conf_self.reshape(B, -1)\n        conf_other = conf_other.reshape(B, -1)\n        # weighted procrustes\n        c2w = weighted_procrustes(\n            pts3ds_self,\n            pts3ds_other,\n            torch.log(conf_self) * torch.log(conf_other),\n            use_weights=True,\n            return_T=True,\n        )\n        return c2w, focal, pp.reshape(B, 2)\n\n    def prepare_input(\n        img_paths,\n        img_mask,\n        size,\n        raymaps=None,\n        raymap_mask=None,\n        revisit=1,\n        update=True,\n        crop=True,\n        img_res=None, \n        reset_interval=100,\n    ):\n        images = load_images(img_paths, size=size, crop=crop)\n        if img_res is not None:\n            K_mhmr = get_camera_parameters(img_res, device=\"cpu\") # if use pseudo K\n\n        views = []\n        if raymaps is None and raymap_mask is None:\n            # Only images are provided.","source_hash":"9427c203a9f9a49ee297b62876dd88027e5fc38e89eafb54f2042f1aee1bb8b7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.launch.prepare_input","uri":"program://Human3R/function/eval.relpose.launch.prepare_input#L296-L411","kind":"function","name":"prepare_input","path":"eval/relpose/launch.py","language":"python","start_line":296,"end_line":411,"context_start_line":276,"context_end_line":431,"code":"            .float()\n            .repeat(B, 1)\n            .reshape(B, 1, 2)\n        )\n        focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n        pts3ds_self = pts3ds_self.reshape(B, -1, 3)\n        pts3ds_other = pts3ds_other.reshape(B, -1, 3)\n        conf_self = conf_self.reshape(B, -1)\n        conf_other = conf_other.reshape(B, -1)\n        # weighted procrustes\n        c2w = weighted_procrustes(\n            pts3ds_self,\n            pts3ds_other,\n            torch.log(conf_self) * torch.log(conf_other),\n            use_weights=True,\n            return_T=True,\n        )\n        return c2w, focal, pp.reshape(B, 2)\n\n    def prepare_input(\n        img_paths,\n        img_mask,\n        size,\n        raymaps=None,\n        raymap_mask=None,\n        revisit=1,\n        update=True,\n        crop=True,\n        img_res=None, \n        reset_interval=100,\n    ):\n        images = load_images(img_paths, size=size, crop=crop)\n        if img_res is not None:\n            K_mhmr = get_camera_parameters(img_res, device=\"cpu\") # if use pseudo K\n\n        views = []\n        if raymaps is None and raymap_mask is None:\n            # Only images are provided.\n            num_views = len(images)\n            for i in range(num_views):\n                view = {\n                    \"img\": images[i][\"img\"],\n                    \"ray_map\": torch.full(\n                        (\n                            images[i][\"img\"].shape[0],\n                            6,\n                            images[i][\"img\"].shape[-2],\n                            images[i][\"img\"].shape[-1],\n                        ),\n                        torch.nan,\n                    ),\n                    \"true_shape\": torch.from_numpy(images[i][\"true_shape\"]),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        np.eye(4).astype(np.float32)\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(True).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(False).unsqueeze(0),\n                    \"update\": torch.tensor(True).unsqueeze(0),\n                    # \"reset\": torch.tensor(False).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n                if img_res is not None:\n                    view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                    view[\"K_mhmr\"] = K_mhmr\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n        else:\n            # Combine images and raymaps.\n            num_views = len(images) + len(raymaps)\n            assert len(img_mask) == len(raymap_mask) == num_views\n            assert sum(img_mask) == len(images) and sum(raymap_mask) == len(raymaps)\n\n            j = 0\n            k = 0\n            for i in range(num_views):\n                view = {\n                    \"img\": (\n                        images[j][\"img\"]\n                        if img_mask[i]\n                        else torch.full_like(images[0][\"img\"], torch.nan)\n                    ),\n                    \"ray_map\": (\n                        raymaps[k]\n                        if raymap_mask[i]\n                        else torch.full_like(raymaps[0], torch.nan)\n                    ),\n                    \"true_shape\": (\n                        torch.from_numpy(images[j][\"true_shape\"])\n                        if img_mask[i]\n                        else torch.from_numpy(np.int32([raymaps[k].shape[1:-1][::-1]]))\n                    ),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        np.eye(4).astype(np.float32)\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(raymap_mask[i]).unsqueeze(0),\n                    \"update\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    # \"reset\": torch.tensor(False).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n                if img_res is not None:\n                    view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                    view[\"K_mhmr\"] = K_mhmr\n                if img_mask[i]:\n                    j += 1\n                if raymap_mask[i]:\n                    k += 1\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n            assert j == len(images) and k == len(raymaps)\n\n        if revisit > 1:\n            # repeat input for 'revisit' times\n            new_views = []\n            for r in range(revisit):\n                for i in range(len(views)):\n                    new_view = deepcopy(views[i])\n                    new_view[\"idx\"] = r * len(views) + i\n                    new_view[\"instance\"] = str(r * len(views) + i)\n                    if r > 0:\n                        if not update:\n                            new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                    new_views.append(new_view)\n            return new_views\n        return views\n\n    def prepare_output(outputs, revisit=1, solve_pose=False):\n        valid_length = len(outputs[\"pred\"]) // revisit\n        outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n        outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n        # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n        reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n        shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n        outputs[\"pred\"] = [\n            pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n        reset_mask = reset_mask[~shifted_reset_mask]\n\n        if solve_pose:\n            pts3ds_self = [\n                output[\"pts3d_in_self_view\"].cpu() for output in outputs[\"pred\"]\n            ]\n            pts3ds_other = [\n                output[\"pts3d_in_other_view\"].cpu() for output in outputs[\"pred\"]\n            ]","source_hash":"9427c203a9f9a49ee297b62876dd88027e5fc38e89eafb54f2042f1aee1bb8b7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.launch.prepare_output","uri":"program://Human3R/function/eval.relpose.launch.prepare_output#L413-L489","kind":"function","name":"prepare_output","path":"eval/relpose/launch.py","language":"python","start_line":413,"end_line":489,"context_start_line":393,"context_end_line":492,"code":"                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n            assert j == len(images) and k == len(raymaps)\n\n        if revisit > 1:\n            # repeat input for 'revisit' times\n            new_views = []\n            for r in range(revisit):\n                for i in range(len(views)):\n                    new_view = deepcopy(views[i])\n                    new_view[\"idx\"] = r * len(views) + i\n                    new_view[\"instance\"] = str(r * len(views) + i)\n                    if r > 0:\n                        if not update:\n                            new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                    new_views.append(new_view)\n            return new_views\n        return views\n\n    def prepare_output(outputs, revisit=1, solve_pose=False):\n        valid_length = len(outputs[\"pred\"]) // revisit\n        outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n        outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n        # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n        reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n        shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n        outputs[\"pred\"] = [\n            pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n        reset_mask = reset_mask[~shifted_reset_mask]\n\n        if solve_pose:\n            pts3ds_self = [\n                output[\"pts3d_in_self_view\"].cpu() for output in outputs[\"pred\"]\n            ]\n            pts3ds_other = [\n                output[\"pts3d_in_other_view\"].cpu() for output in outputs[\"pred\"]\n            ]\n            conf_self = [output[\"conf_self\"].cpu() for output in outputs[\"pred\"]]\n            conf_other = [output[\"conf\"].cpu() for output in outputs[\"pred\"]]\n            pr_poses, focal, pp = recover_cam_params(\n                torch.cat(pts3ds_self, 0),\n                torch.cat(pts3ds_other, 0),\n                torch.cat(conf_self, 0),\n                torch.cat(conf_other, 0),\n            )\n            pts3ds_self = torch.cat(pts3ds_self, 0)\n        else:\n\n            pts3ds_self = [\n                output[\"pts3d_in_self_view\"].cpu() for output in outputs[\"pred\"]\n            ]\n            pts3ds_other = [\n                output[\"pts3d_in_other_view\"].cpu() for output in outputs[\"pred\"]\n            ]\n            conf_self = [output[\"conf_self\"].cpu() for output in outputs[\"pred\"]]\n            conf_other = [output[\"conf\"].cpu() for output in outputs[\"pred\"]]\n            pts3ds_self = torch.cat(pts3ds_self, 0)\n            pr_poses = [\n                pose_encoding_to_camera(pred[\"camera_pose\"].clone()).cpu()\n                for pred in outputs[\"pred\"]\n            ]\n            pr_poses = torch.cat(pr_poses, 0)\n\n            B, H, W, _ = pts3ds_self.shape\n            pp = (\n                torch.tensor([W // 2, H // 2], device=pts3ds_self.device)\n                .float()\n                .repeat(B, 1)\n                .reshape(B, 2)\n            )\n            focal = estimate_focal_knowing_depth(\n                pts3ds_self, pp, focal_mode=\"weiszfeld\"\n            )\n\n        if reset_mask.any():\n            identity = torch.eye(4, device=pr_poses.device)\n            reset_poses = torch.where(reset_mask.unsqueeze(-1).unsqueeze(-1), pr_poses, identity)\n            cumulative_bases = matrix_cumprod(reset_poses)\n            shifted_bases = torch.cat([identity.unsqueeze(0), cumulative_bases[:-1]], dim=0)\n            pr_poses = torch.einsum('bij,bjk->bik', shifted_bases, pr_poses)\n\n        colors = [0.5 * (output[\"rgb\"][0] + 1.0) for output in outputs[\"pred\"]]\n        cam_dict = {\n            \"focal\": focal.cpu().numpy(),\n            \"pp\": pp.cpu().numpy(),\n        }\n        return (\n            colors,\n            pts3ds_self,\n            pts3ds_other,\n            conf_self,\n            conf_other,\n            cam_dict,\n            pr_poses,\n        )\n\n    model = ARCroco3DStereo.from_pretrained(args.weights)\n    eval_pose_estimation(args, model, save_dir=args.output_dir)","source_hash":"9427c203a9f9a49ee297b62876dd88027e5fc38e89eafb54f2042f1aee1bb8b7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils","uri":"program://Human3R/module/eval.relpose.evo_utils#L1-L430","kind":"module","name":"eval.relpose.evo_utils","path":"eval/relpose/evo_utils.py","language":"python","start_line":1,"end_line":430,"context_start_line":1,"context_end_line":430,"code":"\nimport os\nimport re\nfrom copy import deepcopy\nfrom pathlib import Path\n\nimport evo.main_ape as main_ape\nimport evo.main_rpe as main_rpe\n\nimport matplotlib.pyplot as plt\nimport numpy as np\nfrom evo.core import sync\nfrom evo.core.metrics import PoseRelation, Unit\nfrom evo.core.trajectory import PosePath3D, PoseTrajectory3D\nfrom evo.tools import file_interface, plot\nfrom scipy.spatial.transform import Rotation\nfrom evo.core import metrics\n\n\ndef sintel_cam_read(filename):\n    \"\"\"Read camera data, return (M,N) tuple.\n\n    M is the intrinsic matrix, N is the extrinsic matrix, so that\n\n    x = M*N*X,\n    with x being a point in homogeneous image pixel coordinates, X being a\n    point in homogeneous world coordinates.\n    \"\"\"\n    TAG_FLOAT = 202021.25\n\n    f = open(filename, \"rb\")\n    check = np.fromfile(f, dtype=np.float32, count=1)[0]\n    assert (\n        check == TAG_FLOAT\n    ), \" cam_read:: Wrong tag in flow file (should be: {0}, is: {1}). Big-endian machine? \".format(\n        TAG_FLOAT, check\n    )\n    M = np.fromfile(f, dtype=\"float64\", count=9).reshape((3, 3))\n    N = np.fromfile(f, dtype=\"float64\", count=12).reshape((3, 4))\n    return M, N\n\n\ndef load_replica_traj(gt_file):\n    traj_w_c = np.loadtxt(gt_file)\n    assert traj_w_c.shape[1] == 12 or traj_w_c.shape[1] == 16\n    poses = [\n        np.array(\n            [\n                [r[0], r[1], r[2], r[3]],\n                [r[4], r[5], r[6], r[7]],\n                [r[8], r[9], r[10], r[11]],\n                [0, 0, 0, 1],\n            ]\n        )\n        for r in traj_w_c\n    ]\n\n    pose_path = PosePath3D(poses_se3=poses)\n    timestamps_mat = np.arange(traj_w_c.shape[0]).astype(float)\n\n    traj = PoseTrajectory3D(poses_se3=pose_path.poses_se3, timestamps=timestamps_mat)\n    xyz = traj.positions_xyz\n    # shift -1 column -> w in back column\n    # quat = np.roll(traj.orientations_quat_wxyz, -1, axis=1)\n    # uncomment this line if the quaternion is in scalar-first format\n    quat = traj.orientations_quat_wxyz\n\n    traj_tum = np.column_stack((xyz, quat))\n    return (traj_tum, timestamps_mat)\n\n\ndef load_sintel_traj(gt_file):\n    # Refer to ParticleSfM\n    gt_pose_lists = sorted(os.listdir(gt_file))\n    gt_pose_lists = [\n        os.path.join(gt_file, x) for x in gt_pose_lists if x.endswith(\".cam\")\n    ]\n    tstamps = [float(x.split(\"/\")[-1][:-4].split(\"_\")[-1]) for x in gt_pose_lists]\n    gt_poses = [\n        sintel_cam_read(f)[1] for f in gt_pose_lists\n    ]  # [1] means get the extrinsic\n    xyzs, wxyzs = [], []\n    tum_gt_poses = []\n    for gt_pose in gt_poses:\n        gt_pose = np.concatenate([gt_pose, np.array([[0, 0, 0, 1]])], 0)\n        gt_pose_inv = np.linalg.inv(gt_pose)  # world2cam -> cam2world\n        xyz = gt_pose_inv[:3, -1]\n        xyzs.append(xyz)\n        R = Rotation.from_matrix(gt_pose_inv[:3, :3])\n        xyzw = R.as_quat()  # scalar-last for scipy\n        wxyz = np.array([xyzw[-1], xyzw[0], xyzw[1], xyzw[2]])\n        wxyzs.append(wxyz)\n        tum_gt_pose = np.concatenate([xyz, wxyz], 0)  # TODO: check if this is correct\n        tum_gt_poses.append(tum_gt_pose)\n\n    tum_gt_poses = np.stack(tum_gt_poses, 0)\n    tum_gt_poses[:, :3] = tum_gt_poses[:, :3] - np.mean(\n        tum_gt_poses[:, :3], 0, keepdims=True\n    )\n    tt = np.expand_dims(np.stack(tstamps, 0), -1)\n    return tum_gt_poses, tt\n\n\ndef load_traj(gt_traj_file, traj_format=\"sintel\", skip=0, stride=1, num_frames=None):\n    \"\"\"Read trajectory format. Return in TUM-RGBD format.\n    Returns:\n        traj_tum (N, 7): camera to world poses in (x,y,z,qx,qy,qz,qw)\n        timestamps_mat (N, 1): timestamps\n    \"\"\"\n    if traj_format == \"replica\":\n        traj_tum, timestamps_mat = load_replica_traj(gt_traj_file)\n    elif traj_format == \"sintel\":\n        traj_tum, timestamps_mat = load_sintel_traj(gt_traj_file)\n    elif traj_format in [\"tum\", \"tartanair\"]:\n        traj = file_interface.read_tum_trajectory_file(gt_traj_file)\n        xyz = traj.positions_xyz\n        quat = traj.orientations_quat_wxyz\n        timestamps_mat = traj.timestamps\n        traj_tum = np.column_stack((xyz, quat))\n    else:\n        raise NotImplementedError\n\n    traj_tum = traj_tum[skip::stride]\n    timestamps_mat = timestamps_mat[skip::stride]\n    if num_frames is not None:\n        traj_tum = traj_tum[:num_frames]\n        timestamps_mat = timestamps_mat[:num_frames]\n    return traj_tum, timestamps_mat\n\n\ndef update_timestamps(gt_file, traj_format, skip=0, stride=1):\n    \"\"\"Update timestamps given a\"\"\"\n    if traj_format == \"tum\":\n        traj_t_map_file = gt_file.replace(\"groundtruth.txt\", \"rgb.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n    elif traj_format == \"tartanair\":\n        traj_t_map_file = gt_file.replace(\"gt_pose.txt\", \"times.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n\n\ndef load_timestamps(time_file, traj_format=\"replica\"):\n    if traj_format in [\"tum\", \"tartanair\"]:\n        with open(time_file, \"r+\") as f:\n            lines = f.readlines()\n        timestamps_mat = [\n            float(x.split(\" \")[0]) for x in lines if not x.startswith(\"#\")\n        ]\n        return timestamps_mat\n\n\ndef make_traj(args) -> PoseTrajectory3D:\n    if isinstance(args, tuple) or isinstance(args, list):\n        traj, tstamps = args\n        return PoseTrajectory3D(\n            positions_xyz=traj[:, :3],\n            orientations_quat_wxyz=traj[:, 3:],\n            timestamps=tstamps,\n        )\n    assert isinstance(args, PoseTrajectory3D), type(args)\n    return deepcopy(args)\n\n\ndef eval_metrics(pred_traj, gt_traj=None, seq=\"\", filename=\"\", sample_stride=1):\n\n    if sample_stride > 1:\n        pred_traj[0] = pred_traj[0][::sample_stride]\n        pred_traj[1] = pred_traj[1][::sample_stride]\n        if gt_traj is not None:\n            updated_gt_traj = []\n            updated_gt_traj.append(gt_traj[0][::sample_stride])\n            updated_gt_traj.append(gt_traj[1][::sample_stride])\n            gt_traj = updated_gt_traj\n\n    pred_traj = make_traj(pred_traj)\n\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps\n        else:\n            print(pred_traj.timestamps.shape[0], gt_traj.timestamps.shape[0])\n\n        gt_traj, pred_traj = sync.associate_trajectories(gt_traj, pred_traj)\n\n    # ATE\n    traj_ref = gt_traj\n    traj_est = pred_traj\n\n    ate_result = main_ape.ape(\n        traj_ref,\n        traj_est,\n        est_name=\"traj\",\n        pose_relation=PoseRelation.translation_part,\n        align=True,\n        correct_scale=True,\n    )\n\n    ate = ate_result.stats[\"rmse\"]\n    # print(ate_result.np_arrays['error_array'])\n    # exit()\n\n    # RPE rotation and translation\n    delta_list = [1]\n    rpe_rots, rpe_transs = [], []\n    for delta in delta_list:\n        rpe_rots_result = main_rpe.rpe(\n            traj_ref,\n            traj_est,\n            est_name=\"traj\",\n            pose_relation=PoseRelation.rotation_angle_deg,\n            align=True,\n            correct_scale=True,\n            delta=delta,\n            delta_unit=Unit.frames,\n            rel_delta_tol=0.01,\n            all_pairs=True,\n        )\n\n        rot = rpe_rots_result.stats[\"rmse\"]\n        rpe_rots.append(rot)\n\n    for delta in delta_list:\n        rpe_transs_result = main_rpe.rpe(\n            traj_ref,\n            traj_est,\n            est_name=\"traj\",\n            pose_relation=PoseRelation.translation_part,\n            align=True,\n            correct_scale=True,\n            delta=delta,\n            delta_unit=Unit.frames,\n            rel_delta_tol=0.01,\n            all_pairs=True,\n        )\n\n        trans = rpe_transs_result.stats[\"rmse\"]\n        rpe_transs.append(trans)\n\n    rpe_trans, rpe_rot = np.mean(rpe_transs), np.mean(rpe_rots)\n    with open(filename, \"w+\") as f:\n        f.write(f\"Seq: {seq} \\n\\n\")\n        f.write(f\"{ate_result}\")\n        f.write(f\"{rpe_rots_result}\")\n        f.write(f\"{rpe_transs_result}\")\n\n    print(f\"Save results to {filename}\")\n    return ate, rpe_trans, rpe_rot\n\n\ndef eval_metrics_first_pose_align_last_pose(\n    pred_traj, gt_traj=None, seq=\"\", filename=\"\", figpath=\"\", sample_stride=1\n):\n    if sample_stride > 1:\n        pred_traj[0] = pred_traj[0][::sample_stride]\n        pred_traj[1] = pred_traj[1][::sample_stride]\n        if gt_traj is not None:\n            gt_traj = [gt_traj[0][::sample_stride], gt_traj[1][::sample_stride]]\n    pred_traj = make_traj(pred_traj)\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps\n        else:\n            print(\n                \"Different number of poses:\",\n                pred_traj.timestamps.shape[0],\n                gt_traj.timestamps.shape[0],\n            )\n\n        gt_traj, pred_traj = sync.associate_trajectories(gt_traj, pred_traj)\n\n    if gt_traj is not None and pred_traj is not None:\n        if len(gt_traj.poses_se3) > 0 and len(pred_traj.poses_se3) > 0:\n            first_gt_pose = gt_traj.poses_se3[0]\n            first_pred_pose = pred_traj.poses_se3[0]\n            # T = (first_gt_pose) * inv(first_pred_pose)\n            T = first_gt_pose @ np.linalg.inv(first_pred_pose)\n\n            # Apply T to every predicted pose\n            aligned_pred_poses = []\n            for pose in pred_traj.poses_se3:\n                aligned_pred_poses.append(T @ pose)\n            aligned_pred_traj = PoseTrajectory3D(\n                poses_se3=aligned_pred_poses,\n                timestamps=np.array(pred_traj.timestamps),\n                # optionally copy other fields if your make_traj object has them\n            )\n            pred_traj = aligned_pred_traj  # .poses_se3 = aligned_pred_poses\n        plot_trajectory(\n            pred_traj,\n            gt_traj,\n            title=seq,\n            filename=figpath,\n            align=False,\n            correct_scale=False,\n        )\n\n    if gt_traj is not None and len(gt_traj.poses_se3) > 0:\n        gt_traj = PoseTrajectory3D(\n            poses_se3=[gt_traj.poses_se3[-1]], timestamps=[gt_traj.timestamps[-1]]\n        )\n    if pred_traj is not None and len(pred_traj.poses_se3) > 0:\n        pred_traj = PoseTrajectory3D(\n            poses_se3=[pred_traj.poses_se3[-1]], timestamps=[pred_traj.timestamps[-1]]\n        )\n\n    ate_result = main_ape.ape(\n        gt_traj,\n        pred_traj,\n        est_name=\"traj\",\n        pose_relation=PoseRelation.translation_part,\n        align=False,  # <-- important\n        correct_scale=False,  # <-- important\n    )\n    ate = ate_result.stats[\"rmse\"]\n    with open(filename, \"w+\") as f:\n        f.write(f\"Seq: {seq}\\n\\n\")\n        f.write(f\"{ate_result}\")\n\n    print(f\"Save results to {filename}\")\n\n    return ate\n\n\ndef best_plotmode(traj):\n    _, i1, i2 = np.argsort(np.var(traj.positions_xyz, axis=0))\n    plot_axes = \"xyz\"[i2] + \"xyz\"[i1]\n    return getattr(plot.PlotMode, plot_axes)\n\n\ndef plot_trajectory(\n    pred_traj, gt_traj=None, title=\"\", filename=\"\", align=True, correct_scale=True\n):\n    pred_traj = make_traj(pred_traj)\n\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps\n        else:\n            print(\"WARNING\", pred_traj.timestamps.shape[0], gt_traj.timestamps.shape[0])\n\n        gt_traj, pred_traj = sync.associate_trajectories(gt_traj, pred_traj)\n\n        if align:\n            pred_traj.align(gt_traj, correct_scale=correct_scale)\n\n    plot_collection = plot.PlotCollection(\"PlotCol\")\n    fig = plt.figure(figsize=(8, 8))\n    plot_mode = best_plotmode(gt_traj if (gt_traj is not None) else pred_traj)\n    ax = plot.prepare_axis(fig, plot_mode)\n    ax.set_title(title)\n    if gt_traj is not None:\n        plot.traj(ax, plot_mode, gt_traj, \"--\", \"gray\", \"Ground Truth\")\n    plot.traj(ax, plot_mode, pred_traj, \"-\", \"blue\", \"Predicted\")\n    plot_collection.add_figure(\"traj_error\", fig)\n    plot_collection.export(filename, confirm_overwrite=False)\n    plt.close(fig=fig)\n    print(f\"Saved trajectory to {filename.replace('.png','')}_traj_error.png\")\n\n\ndef save_trajectory_tum_format(traj, filename):\n    traj = make_traj(traj)\n    tostr = lambda a: \" \".join(map(str, a))\n    with Path(filename).open(\"w\") as f:\n        for i in range(traj.num_poses):\n            f.write(\n                f\"{traj.timestamps[i]} {tostr(traj.positions_xyz[i])} {tostr(traj.orientations_quat_wxyz[i][[0,1,2,3]])}\\n\"\n            )\n    print(f\"Saved trajectory to {filename}\")\n\n\ndef extract_metrics(file_path):\n    with open(file_path, \"r\") as file:\n        content = file.read()\n\n    # Extract metrics using regex\n    ate_match = re.search(\n        r\"APE w.r.t. translation part \\(m\\).*?rmse\\s+([0-9.]+)\", content, re.DOTALL\n    )\n    rpe_trans_match = re.search(\n        r\"RPE w.r.t. translation part \\(m\\).*?rmse\\s+([0-9.]+)\", content, re.DOTALL\n    )\n    rpe_rot_match = re.search(\n        r\"RPE w.r.t. rotation angle in degrees \\(deg\\).*?rmse\\s+([0-9.]+)\",\n        content,\n        re.DOTALL,\n    )\n\n    ate = float(ate_match.group(1)) if ate_match else 0.0\n    rpe_trans = float(rpe_trans_match.group(1)) if rpe_trans_match else 0.0\n    rpe_rot = float(rpe_rot_match.group(1)) if rpe_rot_match else 0.0\n\n    return ate, rpe_trans, rpe_rot\n\n\ndef process_directory(directory):\n    results = []\n    for root, _, files in os.walk(directory):\n        if files is not None:\n            files = sorted(files)\n        for file in files:\n            if file.endswith(\"_metric.txt\"):\n                file_path = os.path.join(root, file)\n                seq_name = file.replace(\"_eval_metric.txt\", \"\")\n                ate, rpe_trans, rpe_rot = extract_metrics(file_path)\n                results.append((seq_name, ate, rpe_trans, rpe_rot))\n\n    return results\n\n\ndef calculate_averages(results):\n    total_ate = sum(r[1] for r in results)\n    total_rpe_trans = sum(r[2] for r in results)\n    total_rpe_rot = sum(r[3] for r in results)\n    count = len(results)\n\n    if count == 0:\n        return 0.0, 0.0, 0.0\n\n    avg_ate = total_ate / count\n    avg_rpe_trans = total_rpe_trans / count\n    avg_rpe_rot = total_rpe_rot / count\n\n    return avg_ate, avg_rpe_trans, avg_rpe_rot\n","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.sintel_cam_read","uri":"program://Human3R/function/eval.relpose.evo_utils.sintel_cam_read#L20-L40","kind":"function","name":"sintel_cam_read","path":"eval/relpose/evo_utils.py","language":"python","start_line":20,"end_line":40,"context_start_line":1,"context_end_line":60,"code":"\nimport os\nimport re\nfrom copy import deepcopy\nfrom pathlib import Path\n\nimport evo.main_ape as main_ape\nimport evo.main_rpe as main_rpe\n\nimport matplotlib.pyplot as plt\nimport numpy as np\nfrom evo.core import sync\nfrom evo.core.metrics import PoseRelation, Unit\nfrom evo.core.trajectory import PosePath3D, PoseTrajectory3D\nfrom evo.tools import file_interface, plot\nfrom scipy.spatial.transform import Rotation\nfrom evo.core import metrics\n\n\ndef sintel_cam_read(filename):\n    \"\"\"Read camera data, return (M,N) tuple.\n\n    M is the intrinsic matrix, N is the extrinsic matrix, so that\n\n    x = M*N*X,\n    with x being a point in homogeneous image pixel coordinates, X being a\n    point in homogeneous world coordinates.\n    \"\"\"\n    TAG_FLOAT = 202021.25\n\n    f = open(filename, \"rb\")\n    check = np.fromfile(f, dtype=np.float32, count=1)[0]\n    assert (\n        check == TAG_FLOAT\n    ), \" cam_read:: Wrong tag in flow file (should be: {0}, is: {1}). Big-endian machine? \".format(\n        TAG_FLOAT, check\n    )\n    M = np.fromfile(f, dtype=\"float64\", count=9).reshape((3, 3))\n    N = np.fromfile(f, dtype=\"float64\", count=12).reshape((3, 4))\n    return M, N\n\n\ndef load_replica_traj(gt_file):\n    traj_w_c = np.loadtxt(gt_file)\n    assert traj_w_c.shape[1] == 12 or traj_w_c.shape[1] == 16\n    poses = [\n        np.array(\n            [\n                [r[0], r[1], r[2], r[3]],\n                [r[4], r[5], r[6], r[7]],\n                [r[8], r[9], r[10], r[11]],\n                [0, 0, 0, 1],\n            ]\n        )\n        for r in traj_w_c\n    ]\n\n    pose_path = PosePath3D(poses_se3=poses)\n    timestamps_mat = np.arange(traj_w_c.shape[0]).astype(float)\n","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.load_replica_traj","uri":"program://Human3R/function/eval.relpose.evo_utils.load_replica_traj#L43-L69","kind":"function","name":"load_replica_traj","path":"eval/relpose/evo_utils.py","language":"python","start_line":43,"end_line":69,"context_start_line":23,"context_end_line":89,"code":"    M is the intrinsic matrix, N is the extrinsic matrix, so that\n\n    x = M*N*X,\n    with x being a point in homogeneous image pixel coordinates, X being a\n    point in homogeneous world coordinates.\n    \"\"\"\n    TAG_FLOAT = 202021.25\n\n    f = open(filename, \"rb\")\n    check = np.fromfile(f, dtype=np.float32, count=1)[0]\n    assert (\n        check == TAG_FLOAT\n    ), \" cam_read:: Wrong tag in flow file (should be: {0}, is: {1}). Big-endian machine? \".format(\n        TAG_FLOAT, check\n    )\n    M = np.fromfile(f, dtype=\"float64\", count=9).reshape((3, 3))\n    N = np.fromfile(f, dtype=\"float64\", count=12).reshape((3, 4))\n    return M, N\n\n\ndef load_replica_traj(gt_file):\n    traj_w_c = np.loadtxt(gt_file)\n    assert traj_w_c.shape[1] == 12 or traj_w_c.shape[1] == 16\n    poses = [\n        np.array(\n            [\n                [r[0], r[1], r[2], r[3]],\n                [r[4], r[5], r[6], r[7]],\n                [r[8], r[9], r[10], r[11]],\n                [0, 0, 0, 1],\n            ]\n        )\n        for r in traj_w_c\n    ]\n\n    pose_path = PosePath3D(poses_se3=poses)\n    timestamps_mat = np.arange(traj_w_c.shape[0]).astype(float)\n\n    traj = PoseTrajectory3D(poses_se3=pose_path.poses_se3, timestamps=timestamps_mat)\n    xyz = traj.positions_xyz\n    # shift -1 column -> w in back column\n    # quat = np.roll(traj.orientations_quat_wxyz, -1, axis=1)\n    # uncomment this line if the quaternion is in scalar-first format\n    quat = traj.orientations_quat_wxyz\n\n    traj_tum = np.column_stack((xyz, quat))\n    return (traj_tum, timestamps_mat)\n\n\ndef load_sintel_traj(gt_file):\n    # Refer to ParticleSfM\n    gt_pose_lists = sorted(os.listdir(gt_file))\n    gt_pose_lists = [\n        os.path.join(gt_file, x) for x in gt_pose_lists if x.endswith(\".cam\")\n    ]\n    tstamps = [float(x.split(\"/\")[-1][:-4].split(\"_\")[-1]) for x in gt_pose_lists]\n    gt_poses = [\n        sintel_cam_read(f)[1] for f in gt_pose_lists\n    ]  # [1] means get the extrinsic\n    xyzs, wxyzs = [], []\n    tum_gt_poses = []\n    for gt_pose in gt_poses:\n        gt_pose = np.concatenate([gt_pose, np.array([[0, 0, 0, 1]])], 0)\n        gt_pose_inv = np.linalg.inv(gt_pose)  # world2cam -> cam2world\n        xyz = gt_pose_inv[:3, -1]\n        xyzs.append(xyz)\n        R = Rotation.from_matrix(gt_pose_inv[:3, :3])","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.load_sintel_traj","uri":"program://Human3R/function/eval.relpose.evo_utils.load_sintel_traj#L72-L101","kind":"function","name":"load_sintel_traj","path":"eval/relpose/evo_utils.py","language":"python","start_line":72,"end_line":101,"context_start_line":52,"context_end_line":121,"code":"                [0, 0, 0, 1],\n            ]\n        )\n        for r in traj_w_c\n    ]\n\n    pose_path = PosePath3D(poses_se3=poses)\n    timestamps_mat = np.arange(traj_w_c.shape[0]).astype(float)\n\n    traj = PoseTrajectory3D(poses_se3=pose_path.poses_se3, timestamps=timestamps_mat)\n    xyz = traj.positions_xyz\n    # shift -1 column -> w in back column\n    # quat = np.roll(traj.orientations_quat_wxyz, -1, axis=1)\n    # uncomment this line if the quaternion is in scalar-first format\n    quat = traj.orientations_quat_wxyz\n\n    traj_tum = np.column_stack((xyz, quat))\n    return (traj_tum, timestamps_mat)\n\n\ndef load_sintel_traj(gt_file):\n    # Refer to ParticleSfM\n    gt_pose_lists = sorted(os.listdir(gt_file))\n    gt_pose_lists = [\n        os.path.join(gt_file, x) for x in gt_pose_lists if x.endswith(\".cam\")\n    ]\n    tstamps = [float(x.split(\"/\")[-1][:-4].split(\"_\")[-1]) for x in gt_pose_lists]\n    gt_poses = [\n        sintel_cam_read(f)[1] for f in gt_pose_lists\n    ]  # [1] means get the extrinsic\n    xyzs, wxyzs = [], []\n    tum_gt_poses = []\n    for gt_pose in gt_poses:\n        gt_pose = np.concatenate([gt_pose, np.array([[0, 0, 0, 1]])], 0)\n        gt_pose_inv = np.linalg.inv(gt_pose)  # world2cam -> cam2world\n        xyz = gt_pose_inv[:3, -1]\n        xyzs.append(xyz)\n        R = Rotation.from_matrix(gt_pose_inv[:3, :3])\n        xyzw = R.as_quat()  # scalar-last for scipy\n        wxyz = np.array([xyzw[-1], xyzw[0], xyzw[1], xyzw[2]])\n        wxyzs.append(wxyz)\n        tum_gt_pose = np.concatenate([xyz, wxyz], 0)  # TODO: check if this is correct\n        tum_gt_poses.append(tum_gt_pose)\n\n    tum_gt_poses = np.stack(tum_gt_poses, 0)\n    tum_gt_poses[:, :3] = tum_gt_poses[:, :3] - np.mean(\n        tum_gt_poses[:, :3], 0, keepdims=True\n    )\n    tt = np.expand_dims(np.stack(tstamps, 0), -1)\n    return tum_gt_poses, tt\n\n\ndef load_traj(gt_traj_file, traj_format=\"sintel\", skip=0, stride=1, num_frames=None):\n    \"\"\"Read trajectory format. Return in TUM-RGBD format.\n    Returns:\n        traj_tum (N, 7): camera to world poses in (x,y,z,qx,qy,qz,qw)\n        timestamps_mat (N, 1): timestamps\n    \"\"\"\n    if traj_format == \"replica\":\n        traj_tum, timestamps_mat = load_replica_traj(gt_traj_file)\n    elif traj_format == \"sintel\":\n        traj_tum, timestamps_mat = load_sintel_traj(gt_traj_file)\n    elif traj_format in [\"tum\", \"tartanair\"]:\n        traj = file_interface.read_tum_trajectory_file(gt_traj_file)\n        xyz = traj.positions_xyz\n        quat = traj.orientations_quat_wxyz\n        timestamps_mat = traj.timestamps\n        traj_tum = np.column_stack((xyz, quat))\n    else:\n        raise NotImplementedError","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.load_traj","uri":"program://Human3R/function/eval.relpose.evo_utils.load_traj#L104-L128","kind":"function","name":"load_traj","path":"eval/relpose/evo_utils.py","language":"python","start_line":104,"end_line":128,"context_start_line":84,"context_end_line":148,"code":"    for gt_pose in gt_poses:\n        gt_pose = np.concatenate([gt_pose, np.array([[0, 0, 0, 1]])], 0)\n        gt_pose_inv = np.linalg.inv(gt_pose)  # world2cam -> cam2world\n        xyz = gt_pose_inv[:3, -1]\n        xyzs.append(xyz)\n        R = Rotation.from_matrix(gt_pose_inv[:3, :3])\n        xyzw = R.as_quat()  # scalar-last for scipy\n        wxyz = np.array([xyzw[-1], xyzw[0], xyzw[1], xyzw[2]])\n        wxyzs.append(wxyz)\n        tum_gt_pose = np.concatenate([xyz, wxyz], 0)  # TODO: check if this is correct\n        tum_gt_poses.append(tum_gt_pose)\n\n    tum_gt_poses = np.stack(tum_gt_poses, 0)\n    tum_gt_poses[:, :3] = tum_gt_poses[:, :3] - np.mean(\n        tum_gt_poses[:, :3], 0, keepdims=True\n    )\n    tt = np.expand_dims(np.stack(tstamps, 0), -1)\n    return tum_gt_poses, tt\n\n\ndef load_traj(gt_traj_file, traj_format=\"sintel\", skip=0, stride=1, num_frames=None):\n    \"\"\"Read trajectory format. Return in TUM-RGBD format.\n    Returns:\n        traj_tum (N, 7): camera to world poses in (x,y,z,qx,qy,qz,qw)\n        timestamps_mat (N, 1): timestamps\n    \"\"\"\n    if traj_format == \"replica\":\n        traj_tum, timestamps_mat = load_replica_traj(gt_traj_file)\n    elif traj_format == \"sintel\":\n        traj_tum, timestamps_mat = load_sintel_traj(gt_traj_file)\n    elif traj_format in [\"tum\", \"tartanair\"]:\n        traj = file_interface.read_tum_trajectory_file(gt_traj_file)\n        xyz = traj.positions_xyz\n        quat = traj.orientations_quat_wxyz\n        timestamps_mat = traj.timestamps\n        traj_tum = np.column_stack((xyz, quat))\n    else:\n        raise NotImplementedError\n\n    traj_tum = traj_tum[skip::stride]\n    timestamps_mat = timestamps_mat[skip::stride]\n    if num_frames is not None:\n        traj_tum = traj_tum[:num_frames]\n        timestamps_mat = timestamps_mat[:num_frames]\n    return traj_tum, timestamps_mat\n\n\ndef update_timestamps(gt_file, traj_format, skip=0, stride=1):\n    \"\"\"Update timestamps given a\"\"\"\n    if traj_format == \"tum\":\n        traj_t_map_file = gt_file.replace(\"groundtruth.txt\", \"rgb.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n    elif traj_format == \"tartanair\":\n        traj_t_map_file = gt_file.replace(\"gt_pose.txt\", \"times.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n\n\ndef load_timestamps(time_file, traj_format=\"replica\"):\n    if traj_format in [\"tum\", \"tartanair\"]:\n        with open(time_file, \"r+\") as f:\n            lines = f.readlines()\n        timestamps_mat = [\n            float(x.split(\" \")[0]) for x in lines if not x.startswith(\"#\")","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.update_timestamps","uri":"program://Human3R/function/eval.relpose.evo_utils.update_timestamps#L131-L140","kind":"function","name":"update_timestamps","path":"eval/relpose/evo_utils.py","language":"python","start_line":131,"end_line":140,"context_start_line":111,"context_end_line":160,"code":"        traj_tum, timestamps_mat = load_replica_traj(gt_traj_file)\n    elif traj_format == \"sintel\":\n        traj_tum, timestamps_mat = load_sintel_traj(gt_traj_file)\n    elif traj_format in [\"tum\", \"tartanair\"]:\n        traj = file_interface.read_tum_trajectory_file(gt_traj_file)\n        xyz = traj.positions_xyz\n        quat = traj.orientations_quat_wxyz\n        timestamps_mat = traj.timestamps\n        traj_tum = np.column_stack((xyz, quat))\n    else:\n        raise NotImplementedError\n\n    traj_tum = traj_tum[skip::stride]\n    timestamps_mat = timestamps_mat[skip::stride]\n    if num_frames is not None:\n        traj_tum = traj_tum[:num_frames]\n        timestamps_mat = timestamps_mat[:num_frames]\n    return traj_tum, timestamps_mat\n\n\ndef update_timestamps(gt_file, traj_format, skip=0, stride=1):\n    \"\"\"Update timestamps given a\"\"\"\n    if traj_format == \"tum\":\n        traj_t_map_file = gt_file.replace(\"groundtruth.txt\", \"rgb.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n    elif traj_format == \"tartanair\":\n        traj_t_map_file = gt_file.replace(\"gt_pose.txt\", \"times.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n\n\ndef load_timestamps(time_file, traj_format=\"replica\"):\n    if traj_format in [\"tum\", \"tartanair\"]:\n        with open(time_file, \"r+\") as f:\n            lines = f.readlines()\n        timestamps_mat = [\n            float(x.split(\" \")[0]) for x in lines if not x.startswith(\"#\")\n        ]\n        return timestamps_mat\n\n\ndef make_traj(args) -> PoseTrajectory3D:\n    if isinstance(args, tuple) or isinstance(args, list):\n        traj, tstamps = args\n        return PoseTrajectory3D(\n            positions_xyz=traj[:, :3],\n            orientations_quat_wxyz=traj[:, 3:],\n            timestamps=tstamps,\n        )","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.load_timestamps","uri":"program://Human3R/function/eval.relpose.evo_utils.load_timestamps#L143-L150","kind":"function","name":"load_timestamps","path":"eval/relpose/evo_utils.py","language":"python","start_line":143,"end_line":150,"context_start_line":123,"context_end_line":170,"code":"    traj_tum = traj_tum[skip::stride]\n    timestamps_mat = timestamps_mat[skip::stride]\n    if num_frames is not None:\n        traj_tum = traj_tum[:num_frames]\n        timestamps_mat = timestamps_mat[:num_frames]\n    return traj_tum, timestamps_mat\n\n\ndef update_timestamps(gt_file, traj_format, skip=0, stride=1):\n    \"\"\"Update timestamps given a\"\"\"\n    if traj_format == \"tum\":\n        traj_t_map_file = gt_file.replace(\"groundtruth.txt\", \"rgb.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n    elif traj_format == \"tartanair\":\n        traj_t_map_file = gt_file.replace(\"gt_pose.txt\", \"times.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n\n\ndef load_timestamps(time_file, traj_format=\"replica\"):\n    if traj_format in [\"tum\", \"tartanair\"]:\n        with open(time_file, \"r+\") as f:\n            lines = f.readlines()\n        timestamps_mat = [\n            float(x.split(\" \")[0]) for x in lines if not x.startswith(\"#\")\n        ]\n        return timestamps_mat\n\n\ndef make_traj(args) -> PoseTrajectory3D:\n    if isinstance(args, tuple) or isinstance(args, list):\n        traj, tstamps = args\n        return PoseTrajectory3D(\n            positions_xyz=traj[:, :3],\n            orientations_quat_wxyz=traj[:, 3:],\n            timestamps=tstamps,\n        )\n    assert isinstance(args, PoseTrajectory3D), type(args)\n    return deepcopy(args)\n\n\ndef eval_metrics(pred_traj, gt_traj=None, seq=\"\", filename=\"\", sample_stride=1):\n\n    if sample_stride > 1:\n        pred_traj[0] = pred_traj[0][::sample_stride]\n        pred_traj[1] = pred_traj[1][::sample_stride]\n        if gt_traj is not None:","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.make_traj","uri":"program://Human3R/function/eval.relpose.evo_utils.make_traj#L153-L162","kind":"function","name":"make_traj","path":"eval/relpose/evo_utils.py","language":"python","start_line":153,"end_line":162,"context_start_line":133,"context_end_line":182,"code":"    if traj_format == \"tum\":\n        traj_t_map_file = gt_file.replace(\"groundtruth.txt\", \"rgb.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n    elif traj_format == \"tartanair\":\n        traj_t_map_file = gt_file.replace(\"gt_pose.txt\", \"times.txt\")\n        timestamps = load_timestamps(traj_t_map_file, traj_format)\n        return timestamps[skip::stride]\n\n\ndef load_timestamps(time_file, traj_format=\"replica\"):\n    if traj_format in [\"tum\", \"tartanair\"]:\n        with open(time_file, \"r+\") as f:\n            lines = f.readlines()\n        timestamps_mat = [\n            float(x.split(\" \")[0]) for x in lines if not x.startswith(\"#\")\n        ]\n        return timestamps_mat\n\n\ndef make_traj(args) -> PoseTrajectory3D:\n    if isinstance(args, tuple) or isinstance(args, list):\n        traj, tstamps = args\n        return PoseTrajectory3D(\n            positions_xyz=traj[:, :3],\n            orientations_quat_wxyz=traj[:, 3:],\n            timestamps=tstamps,\n        )\n    assert isinstance(args, PoseTrajectory3D), type(args)\n    return deepcopy(args)\n\n\ndef eval_metrics(pred_traj, gt_traj=None, seq=\"\", filename=\"\", sample_stride=1):\n\n    if sample_stride > 1:\n        pred_traj[0] = pred_traj[0][::sample_stride]\n        pred_traj[1] = pred_traj[1][::sample_stride]\n        if gt_traj is not None:\n            updated_gt_traj = []\n            updated_gt_traj.append(gt_traj[0][::sample_stride])\n            updated_gt_traj.append(gt_traj[1][::sample_stride])\n            gt_traj = updated_gt_traj\n\n    pred_traj = make_traj(pred_traj)\n\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.eval_metrics","uri":"program://Human3R/function/eval.relpose.evo_utils.eval_metrics#L165-L250","kind":"function","name":"eval_metrics","path":"eval/relpose/evo_utils.py","language":"python","start_line":165,"end_line":250,"context_start_line":145,"context_end_line":270,"code":"        with open(time_file, \"r+\") as f:\n            lines = f.readlines()\n        timestamps_mat = [\n            float(x.split(\" \")[0]) for x in lines if not x.startswith(\"#\")\n        ]\n        return timestamps_mat\n\n\ndef make_traj(args) -> PoseTrajectory3D:\n    if isinstance(args, tuple) or isinstance(args, list):\n        traj, tstamps = args\n        return PoseTrajectory3D(\n            positions_xyz=traj[:, :3],\n            orientations_quat_wxyz=traj[:, 3:],\n            timestamps=tstamps,\n        )\n    assert isinstance(args, PoseTrajectory3D), type(args)\n    return deepcopy(args)\n\n\ndef eval_metrics(pred_traj, gt_traj=None, seq=\"\", filename=\"\", sample_stride=1):\n\n    if sample_stride > 1:\n        pred_traj[0] = pred_traj[0][::sample_stride]\n        pred_traj[1] = pred_traj[1][::sample_stride]\n        if gt_traj is not None:\n            updated_gt_traj = []\n            updated_gt_traj.append(gt_traj[0][::sample_stride])\n            updated_gt_traj.append(gt_traj[1][::sample_stride])\n            gt_traj = updated_gt_traj\n\n    pred_traj = make_traj(pred_traj)\n\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps\n        else:\n            print(pred_traj.timestamps.shape[0], gt_traj.timestamps.shape[0])\n\n        gt_traj, pred_traj = sync.associate_trajectories(gt_traj, pred_traj)\n\n    # ATE\n    traj_ref = gt_traj\n    traj_est = pred_traj\n\n    ate_result = main_ape.ape(\n        traj_ref,\n        traj_est,\n        est_name=\"traj\",\n        pose_relation=PoseRelation.translation_part,\n        align=True,\n        correct_scale=True,\n    )\n\n    ate = ate_result.stats[\"rmse\"]\n    # print(ate_result.np_arrays['error_array'])\n    # exit()\n\n    # RPE rotation and translation\n    delta_list = [1]\n    rpe_rots, rpe_transs = [], []\n    for delta in delta_list:\n        rpe_rots_result = main_rpe.rpe(\n            traj_ref,\n            traj_est,\n            est_name=\"traj\",\n            pose_relation=PoseRelation.rotation_angle_deg,\n            align=True,\n            correct_scale=True,\n            delta=delta,\n            delta_unit=Unit.frames,\n            rel_delta_tol=0.01,\n            all_pairs=True,\n        )\n\n        rot = rpe_rots_result.stats[\"rmse\"]\n        rpe_rots.append(rot)\n\n    for delta in delta_list:\n        rpe_transs_result = main_rpe.rpe(\n            traj_ref,\n            traj_est,\n            est_name=\"traj\",\n            pose_relation=PoseRelation.translation_part,\n            align=True,\n            correct_scale=True,\n            delta=delta,\n            delta_unit=Unit.frames,\n            rel_delta_tol=0.01,\n            all_pairs=True,\n        )\n\n        trans = rpe_transs_result.stats[\"rmse\"]\n        rpe_transs.append(trans)\n\n    rpe_trans, rpe_rot = np.mean(rpe_transs), np.mean(rpe_rots)\n    with open(filename, \"w+\") as f:\n        f.write(f\"Seq: {seq} \\n\\n\")\n        f.write(f\"{ate_result}\")\n        f.write(f\"{rpe_rots_result}\")\n        f.write(f\"{rpe_transs_result}\")\n\n    print(f\"Save results to {filename}\")\n    return ate, rpe_trans, rpe_rot\n\n\ndef eval_metrics_first_pose_align_last_pose(\n    pred_traj, gt_traj=None, seq=\"\", filename=\"\", figpath=\"\", sample_stride=1\n):\n    if sample_stride > 1:\n        pred_traj[0] = pred_traj[0][::sample_stride]\n        pred_traj[1] = pred_traj[1][::sample_stride]\n        if gt_traj is not None:\n            gt_traj = [gt_traj[0][::sample_stride], gt_traj[1][::sample_stride]]\n    pred_traj = make_traj(pred_traj)\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps\n        else:\n            print(\n                \"Different number of poses:\",\n                pred_traj.timestamps.shape[0],","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.eval_metrics_first_pose_align_last_pose","uri":"program://Human3R/function/eval.relpose.evo_utils.eval_metrics_first_pose_align_last_pose#L253-L326","kind":"function","name":"eval_metrics_first_pose_align_last_pose","path":"eval/relpose/evo_utils.py","language":"python","start_line":253,"end_line":326,"context_start_line":233,"context_end_line":346,"code":"            delta=delta,\n            delta_unit=Unit.frames,\n            rel_delta_tol=0.01,\n            all_pairs=True,\n        )\n\n        trans = rpe_transs_result.stats[\"rmse\"]\n        rpe_transs.append(trans)\n\n    rpe_trans, rpe_rot = np.mean(rpe_transs), np.mean(rpe_rots)\n    with open(filename, \"w+\") as f:\n        f.write(f\"Seq: {seq} \\n\\n\")\n        f.write(f\"{ate_result}\")\n        f.write(f\"{rpe_rots_result}\")\n        f.write(f\"{rpe_transs_result}\")\n\n    print(f\"Save results to {filename}\")\n    return ate, rpe_trans, rpe_rot\n\n\ndef eval_metrics_first_pose_align_last_pose(\n    pred_traj, gt_traj=None, seq=\"\", filename=\"\", figpath=\"\", sample_stride=1\n):\n    if sample_stride > 1:\n        pred_traj[0] = pred_traj[0][::sample_stride]\n        pred_traj[1] = pred_traj[1][::sample_stride]\n        if gt_traj is not None:\n            gt_traj = [gt_traj[0][::sample_stride], gt_traj[1][::sample_stride]]\n    pred_traj = make_traj(pred_traj)\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps\n        else:\n            print(\n                \"Different number of poses:\",\n                pred_traj.timestamps.shape[0],\n                gt_traj.timestamps.shape[0],\n            )\n\n        gt_traj, pred_traj = sync.associate_trajectories(gt_traj, pred_traj)\n\n    if gt_traj is not None and pred_traj is not None:\n        if len(gt_traj.poses_se3) > 0 and len(pred_traj.poses_se3) > 0:\n            first_gt_pose = gt_traj.poses_se3[0]\n            first_pred_pose = pred_traj.poses_se3[0]\n            # T = (first_gt_pose) * inv(first_pred_pose)\n            T = first_gt_pose @ np.linalg.inv(first_pred_pose)\n\n            # Apply T to every predicted pose\n            aligned_pred_poses = []\n            for pose in pred_traj.poses_se3:\n                aligned_pred_poses.append(T @ pose)\n            aligned_pred_traj = PoseTrajectory3D(\n                poses_se3=aligned_pred_poses,\n                timestamps=np.array(pred_traj.timestamps),\n                # optionally copy other fields if your make_traj object has them\n            )\n            pred_traj = aligned_pred_traj  # .poses_se3 = aligned_pred_poses\n        plot_trajectory(\n            pred_traj,\n            gt_traj,\n            title=seq,\n            filename=figpath,\n            align=False,\n            correct_scale=False,\n        )\n\n    if gt_traj is not None and len(gt_traj.poses_se3) > 0:\n        gt_traj = PoseTrajectory3D(\n            poses_se3=[gt_traj.poses_se3[-1]], timestamps=[gt_traj.timestamps[-1]]\n        )\n    if pred_traj is not None and len(pred_traj.poses_se3) > 0:\n        pred_traj = PoseTrajectory3D(\n            poses_se3=[pred_traj.poses_se3[-1]], timestamps=[pred_traj.timestamps[-1]]\n        )\n\n    ate_result = main_ape.ape(\n        gt_traj,\n        pred_traj,\n        est_name=\"traj\",\n        pose_relation=PoseRelation.translation_part,\n        align=False,  # <-- important\n        correct_scale=False,  # <-- important\n    )\n    ate = ate_result.stats[\"rmse\"]\n    with open(filename, \"w+\") as f:\n        f.write(f\"Seq: {seq}\\n\\n\")\n        f.write(f\"{ate_result}\")\n\n    print(f\"Save results to {filename}\")\n\n    return ate\n\n\ndef best_plotmode(traj):\n    _, i1, i2 = np.argsort(np.var(traj.positions_xyz, axis=0))\n    plot_axes = \"xyz\"[i2] + \"xyz\"[i1]\n    return getattr(plot.PlotMode, plot_axes)\n\n\ndef plot_trajectory(\n    pred_traj, gt_traj=None, title=\"\", filename=\"\", align=True, correct_scale=True\n):\n    pred_traj = make_traj(pred_traj)\n\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps\n        else:\n            print(\"WARNING\", pred_traj.timestamps.shape[0], gt_traj.timestamps.shape[0])\n","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.best_plotmode","uri":"program://Human3R/function/eval.relpose.evo_utils.best_plotmode#L329-L332","kind":"function","name":"best_plotmode","path":"eval/relpose/evo_utils.py","language":"python","start_line":329,"end_line":332,"context_start_line":309,"context_end_line":352,"code":"        )\n\n    ate_result = main_ape.ape(\n        gt_traj,\n        pred_traj,\n        est_name=\"traj\",\n        pose_relation=PoseRelation.translation_part,\n        align=False,  # <-- important\n        correct_scale=False,  # <-- important\n    )\n    ate = ate_result.stats[\"rmse\"]\n    with open(filename, \"w+\") as f:\n        f.write(f\"Seq: {seq}\\n\\n\")\n        f.write(f\"{ate_result}\")\n\n    print(f\"Save results to {filename}\")\n\n    return ate\n\n\ndef best_plotmode(traj):\n    _, i1, i2 = np.argsort(np.var(traj.positions_xyz, axis=0))\n    plot_axes = \"xyz\"[i2] + \"xyz\"[i1]\n    return getattr(plot.PlotMode, plot_axes)\n\n\ndef plot_trajectory(\n    pred_traj, gt_traj=None, title=\"\", filename=\"\", align=True, correct_scale=True\n):\n    pred_traj = make_traj(pred_traj)\n\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps\n        else:\n            print(\"WARNING\", pred_traj.timestamps.shape[0], gt_traj.timestamps.shape[0])\n\n        gt_traj, pred_traj = sync.associate_trajectories(gt_traj, pred_traj)\n\n        if align:\n            pred_traj.align(gt_traj, correct_scale=correct_scale)\n\n    plot_collection = plot.PlotCollection(\"PlotCol\")","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.plot_trajectory","uri":"program://Human3R/function/eval.relpose.evo_utils.plot_trajectory#L335-L363","kind":"function","name":"plot_trajectory","path":"eval/relpose/evo_utils.py","language":"python","start_line":335,"end_line":363,"context_start_line":315,"context_end_line":383,"code":"        pose_relation=PoseRelation.translation_part,\n        align=False,  # <-- important\n        correct_scale=False,  # <-- important\n    )\n    ate = ate_result.stats[\"rmse\"]\n    with open(filename, \"w+\") as f:\n        f.write(f\"Seq: {seq}\\n\\n\")\n        f.write(f\"{ate_result}\")\n\n    print(f\"Save results to {filename}\")\n\n    return ate\n\n\ndef best_plotmode(traj):\n    _, i1, i2 = np.argsort(np.var(traj.positions_xyz, axis=0))\n    plot_axes = \"xyz\"[i2] + \"xyz\"[i1]\n    return getattr(plot.PlotMode, plot_axes)\n\n\ndef plot_trajectory(\n    pred_traj, gt_traj=None, title=\"\", filename=\"\", align=True, correct_scale=True\n):\n    pred_traj = make_traj(pred_traj)\n\n    if gt_traj is not None:\n        gt_traj = make_traj(gt_traj)\n        if pred_traj.timestamps.shape[0] == gt_traj.timestamps.shape[0]:\n            pred_traj.timestamps = gt_traj.timestamps\n        else:\n            print(\"WARNING\", pred_traj.timestamps.shape[0], gt_traj.timestamps.shape[0])\n\n        gt_traj, pred_traj = sync.associate_trajectories(gt_traj, pred_traj)\n\n        if align:\n            pred_traj.align(gt_traj, correct_scale=correct_scale)\n\n    plot_collection = plot.PlotCollection(\"PlotCol\")\n    fig = plt.figure(figsize=(8, 8))\n    plot_mode = best_plotmode(gt_traj if (gt_traj is not None) else pred_traj)\n    ax = plot.prepare_axis(fig, plot_mode)\n    ax.set_title(title)\n    if gt_traj is not None:\n        plot.traj(ax, plot_mode, gt_traj, \"--\", \"gray\", \"Ground Truth\")\n    plot.traj(ax, plot_mode, pred_traj, \"-\", \"blue\", \"Predicted\")\n    plot_collection.add_figure(\"traj_error\", fig)\n    plot_collection.export(filename, confirm_overwrite=False)\n    plt.close(fig=fig)\n    print(f\"Saved trajectory to {filename.replace('.png','')}_traj_error.png\")\n\n\ndef save_trajectory_tum_format(traj, filename):\n    traj = make_traj(traj)\n    tostr = lambda a: \" \".join(map(str, a))\n    with Path(filename).open(\"w\") as f:\n        for i in range(traj.num_poses):\n            f.write(\n                f\"{traj.timestamps[i]} {tostr(traj.positions_xyz[i])} {tostr(traj.orientations_quat_wxyz[i][[0,1,2,3]])}\\n\"\n            )\n    print(f\"Saved trajectory to {filename}\")\n\n\ndef extract_metrics(file_path):\n    with open(file_path, \"r\") as file:\n        content = file.read()\n\n    # Extract metrics using regex\n    ate_match = re.search(\n        r\"APE w.r.t. translation part \\(m\\).*?rmse\\s+([0-9.]+)\", content, re.DOTALL","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.save_trajectory_tum_format","uri":"program://Human3R/function/eval.relpose.evo_utils.save_trajectory_tum_format#L366-L374","kind":"function","name":"save_trajectory_tum_format","path":"eval/relpose/evo_utils.py","language":"python","start_line":366,"end_line":374,"context_start_line":346,"context_end_line":394,"code":"\n        gt_traj, pred_traj = sync.associate_trajectories(gt_traj, pred_traj)\n\n        if align:\n            pred_traj.align(gt_traj, correct_scale=correct_scale)\n\n    plot_collection = plot.PlotCollection(\"PlotCol\")\n    fig = plt.figure(figsize=(8, 8))\n    plot_mode = best_plotmode(gt_traj if (gt_traj is not None) else pred_traj)\n    ax = plot.prepare_axis(fig, plot_mode)\n    ax.set_title(title)\n    if gt_traj is not None:\n        plot.traj(ax, plot_mode, gt_traj, \"--\", \"gray\", \"Ground Truth\")\n    plot.traj(ax, plot_mode, pred_traj, \"-\", \"blue\", \"Predicted\")\n    plot_collection.add_figure(\"traj_error\", fig)\n    plot_collection.export(filename, confirm_overwrite=False)\n    plt.close(fig=fig)\n    print(f\"Saved trajectory to {filename.replace('.png','')}_traj_error.png\")\n\n\ndef save_trajectory_tum_format(traj, filename):\n    traj = make_traj(traj)\n    tostr = lambda a: \" \".join(map(str, a))\n    with Path(filename).open(\"w\") as f:\n        for i in range(traj.num_poses):\n            f.write(\n                f\"{traj.timestamps[i]} {tostr(traj.positions_xyz[i])} {tostr(traj.orientations_quat_wxyz[i][[0,1,2,3]])}\\n\"\n            )\n    print(f\"Saved trajectory to {filename}\")\n\n\ndef extract_metrics(file_path):\n    with open(file_path, \"r\") as file:\n        content = file.read()\n\n    # Extract metrics using regex\n    ate_match = re.search(\n        r\"APE w.r.t. translation part \\(m\\).*?rmse\\s+([0-9.]+)\", content, re.DOTALL\n    )\n    rpe_trans_match = re.search(\n        r\"RPE w.r.t. translation part \\(m\\).*?rmse\\s+([0-9.]+)\", content, re.DOTALL\n    )\n    rpe_rot_match = re.search(\n        r\"RPE w.r.t. rotation angle in degrees \\(deg\\).*?rmse\\s+([0-9.]+)\",\n        content,\n        re.DOTALL,\n    )\n\n    ate = float(ate_match.group(1)) if ate_match else 0.0","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.extract_metrics","uri":"program://Human3R/function/eval.relpose.evo_utils.extract_metrics#L377-L398","kind":"function","name":"extract_metrics","path":"eval/relpose/evo_utils.py","language":"python","start_line":377,"end_line":398,"context_start_line":357,"context_end_line":418,"code":"    if gt_traj is not None:\n        plot.traj(ax, plot_mode, gt_traj, \"--\", \"gray\", \"Ground Truth\")\n    plot.traj(ax, plot_mode, pred_traj, \"-\", \"blue\", \"Predicted\")\n    plot_collection.add_figure(\"traj_error\", fig)\n    plot_collection.export(filename, confirm_overwrite=False)\n    plt.close(fig=fig)\n    print(f\"Saved trajectory to {filename.replace('.png','')}_traj_error.png\")\n\n\ndef save_trajectory_tum_format(traj, filename):\n    traj = make_traj(traj)\n    tostr = lambda a: \" \".join(map(str, a))\n    with Path(filename).open(\"w\") as f:\n        for i in range(traj.num_poses):\n            f.write(\n                f\"{traj.timestamps[i]} {tostr(traj.positions_xyz[i])} {tostr(traj.orientations_quat_wxyz[i][[0,1,2,3]])}\\n\"\n            )\n    print(f\"Saved trajectory to {filename}\")\n\n\ndef extract_metrics(file_path):\n    with open(file_path, \"r\") as file:\n        content = file.read()\n\n    # Extract metrics using regex\n    ate_match = re.search(\n        r\"APE w.r.t. translation part \\(m\\).*?rmse\\s+([0-9.]+)\", content, re.DOTALL\n    )\n    rpe_trans_match = re.search(\n        r\"RPE w.r.t. translation part \\(m\\).*?rmse\\s+([0-9.]+)\", content, re.DOTALL\n    )\n    rpe_rot_match = re.search(\n        r\"RPE w.r.t. rotation angle in degrees \\(deg\\).*?rmse\\s+([0-9.]+)\",\n        content,\n        re.DOTALL,\n    )\n\n    ate = float(ate_match.group(1)) if ate_match else 0.0\n    rpe_trans = float(rpe_trans_match.group(1)) if rpe_trans_match else 0.0\n    rpe_rot = float(rpe_rot_match.group(1)) if rpe_rot_match else 0.0\n\n    return ate, rpe_trans, rpe_rot\n\n\ndef process_directory(directory):\n    results = []\n    for root, _, files in os.walk(directory):\n        if files is not None:\n            files = sorted(files)\n        for file in files:\n            if file.endswith(\"_metric.txt\"):\n                file_path = os.path.join(root, file)\n                seq_name = file.replace(\"_eval_metric.txt\", \"\")\n                ate, rpe_trans, rpe_rot = extract_metrics(file_path)\n                results.append((seq_name, ate, rpe_trans, rpe_rot))\n\n    return results\n\n\ndef calculate_averages(results):\n    total_ate = sum(r[1] for r in results)\n    total_rpe_trans = sum(r[2] for r in results)","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.process_directory","uri":"program://Human3R/function/eval.relpose.evo_utils.process_directory#L401-L413","kind":"function","name":"process_directory","path":"eval/relpose/evo_utils.py","language":"python","start_line":401,"end_line":413,"context_start_line":381,"context_end_line":430,"code":"    # Extract metrics using regex\n    ate_match = re.search(\n        r\"APE w.r.t. translation part \\(m\\).*?rmse\\s+([0-9.]+)\", content, re.DOTALL\n    )\n    rpe_trans_match = re.search(\n        r\"RPE w.r.t. translation part \\(m\\).*?rmse\\s+([0-9.]+)\", content, re.DOTALL\n    )\n    rpe_rot_match = re.search(\n        r\"RPE w.r.t. rotation angle in degrees \\(deg\\).*?rmse\\s+([0-9.]+)\",\n        content,\n        re.DOTALL,\n    )\n\n    ate = float(ate_match.group(1)) if ate_match else 0.0\n    rpe_trans = float(rpe_trans_match.group(1)) if rpe_trans_match else 0.0\n    rpe_rot = float(rpe_rot_match.group(1)) if rpe_rot_match else 0.0\n\n    return ate, rpe_trans, rpe_rot\n\n\ndef process_directory(directory):\n    results = []\n    for root, _, files in os.walk(directory):\n        if files is not None:\n            files = sorted(files)\n        for file in files:\n            if file.endswith(\"_metric.txt\"):\n                file_path = os.path.join(root, file)\n                seq_name = file.replace(\"_eval_metric.txt\", \"\")\n                ate, rpe_trans, rpe_rot = extract_metrics(file_path)\n                results.append((seq_name, ate, rpe_trans, rpe_rot))\n\n    return results\n\n\ndef calculate_averages(results):\n    total_ate = sum(r[1] for r in results)\n    total_rpe_trans = sum(r[2] for r in results)\n    total_rpe_rot = sum(r[3] for r in results)\n    count = len(results)\n\n    if count == 0:\n        return 0.0, 0.0, 0.0\n\n    avg_ate = total_ate / count\n    avg_rpe_trans = total_rpe_trans / count\n    avg_rpe_rot = total_rpe_rot / count\n\n    return avg_ate, avg_rpe_trans, avg_rpe_rot\n","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.relpose.evo_utils.calculate_averages","uri":"program://Human3R/function/eval.relpose.evo_utils.calculate_averages#L416-L429","kind":"function","name":"calculate_averages","path":"eval/relpose/evo_utils.py","language":"python","start_line":416,"end_line":429,"context_start_line":396,"context_end_line":430,"code":"    rpe_rot = float(rpe_rot_match.group(1)) if rpe_rot_match else 0.0\n\n    return ate, rpe_trans, rpe_rot\n\n\ndef process_directory(directory):\n    results = []\n    for root, _, files in os.walk(directory):\n        if files is not None:\n            files = sorted(files)\n        for file in files:\n            if file.endswith(\"_metric.txt\"):\n                file_path = os.path.join(root, file)\n                seq_name = file.replace(\"_eval_metric.txt\", \"\")\n                ate, rpe_trans, rpe_rot = extract_metrics(file_path)\n                results.append((seq_name, ate, rpe_trans, rpe_rot))\n\n    return results\n\n\ndef calculate_averages(results):\n    total_ate = sum(r[1] for r in results)\n    total_rpe_trans = sum(r[2] for r in results)\n    total_rpe_rot = sum(r[3] for r in results)\n    count = len(results)\n\n    if count == 0:\n        return 0.0, 0.0, 0.0\n\n    avg_ate = total_ate / count\n    avg_rpe_trans = total_rpe_trans / count\n    avg_rpe_rot = total_rpe_rot / count\n\n    return avg_ate, avg_rpe_trans, avg_rpe_rot\n","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.metadata","uri":"program://Human3R/module/eval.video_depth.metadata#L1-L199","kind":"module","name":"eval.video_depth.metadata","path":"eval/video_depth/metadata.py","language":"python","start_line":1,"end_line":199,"context_start_line":1,"context_end_line":199,"code":"import os\nimport glob\nfrom tqdm import tqdm\n\n# Define the merged dataset metadata dictionary\ndataset_metadata = {\n    \"davis\": {\n        \"img_path\": \"data/davis/DAVIS/JPEGImages/480p\",\n        \"mask_path\": \"data/davis/DAVIS/masked_images/480p\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: None,\n        \"traj_format\": None,\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: os.path.join(mask_path, seq),\n        \"skip_condition\": None,\n        \"process_func\": None,  # Not used in mono depth estimation\n    },\n    \"kitti\": {\n        \"img_path\": \"data/kitti/depth_selection/val_selection_cropped/image_gathered\",  # Default path\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: None,\n        \"traj_format\": None,\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": lambda args, img_path: process_kitti(args, img_path),\n    },\n    \"bonn\": {\n        \"img_path\": \"/path/to/rgbd_bonn_dataset\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(\n            img_path, f\"rgbd_bonn_{seq}\", \"rgb_110\"\n        ),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, f\"rgbd_bonn_{seq}\", \"groundtruth_110.txt\"\n        ),\n        \"traj_format\": \"tum\",\n        \"seq_list\": [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"],\n        \"full_seq\": False,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": lambda args, img_path: process_bonn(args, img_path),\n    },\n    \"nyu\": {\n        \"img_path\": \"data/nyu-v2/val/nyu_images\",\n        \"mask_path\": None,\n        \"process_func\": lambda args, img_path: process_nyu(args, img_path),\n    },\n    \"scannet\": {\n        \"img_path\": \"data/scannetv2\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq, \"color_90\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, seq, \"pose_90.txt\"\n        ),\n        \"traj_format\": \"replica\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,  # lambda save_dir, seq: os.path.exists(os.path.join(save_dir, seq)),\n        \"process_func\": lambda args, img_path: process_scannet(args, img_path),\n    },\n    \"tum\": {\n        \"img_path\": \"data/tum\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq, \"rgb_90\"),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(\n            img_path, seq, \"groundtruth_90.txt\"\n        ),\n        \"traj_format\": \"tum\",\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": None,\n    },\n    \"sintel\": {\n        \"img_path\": \"/path/to/sintel/training/final\",\n        \"anno_path\": \"/path/to/sintel/training/camdata_left\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: os.path.join(anno_path, seq),\n        \"traj_format\": None,\n        \"seq_list\": [\n            \"alley_2\",\n            \"ambush_4\",\n            \"ambush_5\",\n            \"ambush_6\",\n            \"cave_2\",\n            \"cave_4\",\n            \"market_2\",\n            \"market_5\",\n            \"market_6\",\n            \"shaman_3\",\n            \"sleeping_1\",\n            \"sleeping_2\",\n            \"temple_2\",\n            \"temple_3\",\n        ],\n        \"full_seq\": False,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": lambda args, img_path: process_sintel(args, img_path),\n    },\n}\n\nbonn_numbers = [50, 100, 150, 200, 250, 300, 350, 400, 450, 500]\nbonn_configs = {\n    f\"bonn_{num}\": {\n        \"img_path\": \"/path/to/long_bonn\",\n        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq, num=num: os.path.join(\n            img_path, f\"rgbd_bonn_{seq}\", f\"rgb_{num}\"\n        ),\n        \"gt_traj_func\": lambda img_path, anno_path, seq, num=num: os.path.join(\n            img_path, f\"rgbd_bonn_{seq}\", f\"groundtruth_{num}.txt\"\n        ),\n        \"traj_format\": \"tum\",\n        \"seq_list\": [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"],\n        \"full_seq\": False,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": lambda args, img_path: process_bonn(args, img_path),\n    }\n    for num in bonn_numbers\n}\n# then update dataset_metadata\ndataset_metadata.update(bonn_configs)\n\n# Define processing functions for each dataset\ndef process_kitti(args, img_path):\n    for dir in tqdm(sorted(glob.glob(f\"{img_path}/*\"))):\n        filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(dir)}\"\n        yield filelist, save_dir\n\n\ndef process_bonn(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/rgb/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = [\n            \"alley_2\",\n            \"ambush_4\",\n            \"ambush_5\",\n            \"ambush_6\",\n            \"cave_2\",\n            \"cave_4\",\n            \"market_2\",\n            \"market_5\",\n            \"market_6\",\n            \"shaman_3\",\n            \"sleeping_1\",\n            \"sleeping_2\",\n            \"temple_2\",\n            \"temple_3\",\n        ]\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/{seq}/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir","source_hash":"ffd74aa6378d590ead9dc876de366133264ba355cc9575aa84610d3c63c51c0c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.metadata.process_kitti","uri":"program://Human3R/function/eval.video_depth.metadata.process_kitti#L134-L138","kind":"function","name":"process_kitti","path":"eval/video_depth/metadata.py","language":"python","start_line":134,"end_line":138,"context_start_line":114,"context_end_line":158,"code":"        \"mask_path\": None,\n        \"dir_path_func\": lambda img_path, seq, num=num: os.path.join(\n            img_path, f\"rgbd_bonn_{seq}\", f\"rgb_{num}\"\n        ),\n        \"gt_traj_func\": lambda img_path, anno_path, seq, num=num: os.path.join(\n            img_path, f\"rgbd_bonn_{seq}\", f\"groundtruth_{num}.txt\"\n        ),\n        \"traj_format\": \"tum\",\n        \"seq_list\": [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"],\n        \"full_seq\": False,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": lambda args, img_path: process_bonn(args, img_path),\n    }\n    for num in bonn_numbers\n}\n# then update dataset_metadata\ndataset_metadata.update(bonn_configs)\n\n# Define processing functions for each dataset\ndef process_kitti(args, img_path):\n    for dir in tqdm(sorted(glob.glob(f\"{img_path}/*\"))):\n        filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(dir)}\"\n        yield filelist, save_dir\n\n\ndef process_bonn(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/rgb/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n","source_hash":"ffd74aa6378d590ead9dc876de366133264ba355cc9575aa84610d3c63c51c0c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.metadata.process_bonn","uri":"program://Human3R/function/eval.video_depth.metadata.process_bonn#L141-L156","kind":"function","name":"process_bonn","path":"eval/video_depth/metadata.py","language":"python","start_line":141,"end_line":156,"context_start_line":121,"context_end_line":176,"code":"        \"traj_format\": \"tum\",\n        \"seq_list\": [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"],\n        \"full_seq\": False,\n        \"mask_path_seq_func\": lambda mask_path, seq: None,\n        \"skip_condition\": None,\n        \"process_func\": lambda args, img_path: process_bonn(args, img_path),\n    }\n    for num in bonn_numbers\n}\n# then update dataset_metadata\ndataset_metadata.update(bonn_configs)\n\n# Define processing functions for each dataset\ndef process_kitti(args, img_path):\n    for dir in tqdm(sorted(glob.glob(f\"{img_path}/*\"))):\n        filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(dir)}\"\n        yield filelist, save_dir\n\n\ndef process_bonn(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/rgb/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))","source_hash":"ffd74aa6378d590ead9dc876de366133264ba355cc9575aa84610d3c63c51c0c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.metadata.process_nyu","uri":"program://Human3R/function/eval.video_depth.metadata.process_nyu#L159-L162","kind":"function","name":"process_nyu","path":"eval/video_depth/metadata.py","language":"python","start_line":159,"end_line":162,"context_start_line":139,"context_end_line":182,"code":"\n\ndef process_bonn(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/rgb/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = [\n            \"alley_2\",\n            \"ambush_4\",","source_hash":"ffd74aa6378d590ead9dc876de366133264ba355cc9575aa84610d3c63c51c0c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.metadata.process_scannet","uri":"program://Human3R/function/eval.video_depth.metadata.process_scannet#L165-L170","kind":"function","name":"process_scannet","path":"eval/video_depth/metadata.py","language":"python","start_line":165,"end_line":170,"context_start_line":145,"context_end_line":190,"code":"            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = (\n            [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n            if args.seq_list is None\n            else args.seq_list\n        )\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = [\n            \"alley_2\",\n            \"ambush_4\",\n            \"ambush_5\",\n            \"ambush_6\",\n            \"cave_2\",\n            \"cave_4\",\n            \"market_2\",\n            \"market_5\",\n            \"market_6\",\n            \"shaman_3\",","source_hash":"ffd74aa6378d590ead9dc876de366133264ba355cc9575aa84610d3c63c51c0c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.metadata.process_sintel","uri":"program://Human3R/function/eval.video_depth.metadata.process_sintel#L173-L199","kind":"function","name":"process_sintel","path":"eval/video_depth/metadata.py","language":"python","start_line":173,"end_line":199,"context_start_line":153,"context_end_line":199,"code":"        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/rgbd_bonn_{seq}/rgb_110/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir\n\n\ndef process_nyu(args, img_path):\n    filelist = sorted(glob.glob(f\"{img_path}/*.png\"))\n    save_dir = f\"{args.output_dir}\"\n    yield filelist, save_dir\n\n\ndef process_scannet(args, img_path):\n    seq_list = sorted(glob.glob(f\"{img_path}/*\"))\n    for seq in tqdm(seq_list):\n        filelist = sorted(glob.glob(f\"{seq}/color_90/*.jpg\"))\n        save_dir = f\"{args.output_dir}/{os.path.basename(seq)}\"\n        yield filelist, save_dir\n\n\ndef process_sintel(args, img_path):\n    if args.full_seq:\n        for dir in tqdm(sorted(glob.glob(f\"{img_path}/*/\"))):\n            filelist = sorted(glob.glob(f\"{dir}/*.png\"))\n            save_dir = f\"{args.output_dir}/{os.path.basename(os.path.dirname(dir))}\"\n            yield filelist, save_dir\n    else:\n        seq_list = [\n            \"alley_2\",\n            \"ambush_4\",\n            \"ambush_5\",\n            \"ambush_6\",\n            \"cave_2\",\n            \"cave_4\",\n            \"market_2\",\n            \"market_5\",\n            \"market_6\",\n            \"shaman_3\",\n            \"sleeping_1\",\n            \"sleeping_2\",\n            \"temple_2\",\n            \"temple_3\",\n        ]\n        for seq in tqdm(seq_list):\n            filelist = sorted(glob.glob(f\"{img_path}/{seq}/*.png\"))\n            save_dir = f\"{args.output_dir}/{seq}\"\n            yield filelist, save_dir","source_hash":"ffd74aa6378d590ead9dc876de366133264ba355cc9575aa84610d3c63c51c0c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.utils","uri":"program://Human3R/module/eval.video_depth.utils#L1-L236","kind":"module","name":"eval.video_depth.utils","path":"eval/video_depth/utils.py","language":"python","start_line":1,"end_line":236,"context_start_line":1,"context_end_line":236,"code":"from copy import deepcopy\nimport cv2\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport roma\nfrom copy import deepcopy\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)\n    min_depth = depth_maps.min()  # float(torch.quantile(out, 0.01))\n    max_depth = depth_maps.max()  # float(torch.quantile(out, 0.99))\n    colored_depth = colorize(\n        depth_maps,\n        cmap_name=\"Spectral_r\",\n        range=(min_depth, max_depth),\n        append_cbar=True,\n    )\n    images = []\n\n    if conf_self is not None:\n        conf_selfs = torch.concat(conf_self, 0)\n        min_conf = torch.log(conf_selfs.min())  # float(torch.quantile(out, 0.01))\n        max_conf = torch.log(conf_selfs.max())  # float(torch.quantile(out, 0.99))\n        colored_conf = colorize(\n            torch.log(conf_selfs),\n            cmap_name=\"jet\",\n            range=(min_conf, max_conf),\n            append_cbar=True,\n        )\n\n    for i, depth_map in enumerate(colored_depth):\n        # Apply color map to depth map\n        img_path = f\"{path}/frame_{(i):04d}.png\"\n        if conf_self is None:\n            to_save = (depth_map * 255).detach().cpu().numpy().astype(np.uint8)\n        else:\n            to_save = torch.cat([depth_map, colored_conf[i]], dim=1)\n            to_save = (to_save * 255).detach().cpu().numpy().astype(np.uint8)\n        iio.imwrite(img_path, to_save)\n        images.append(Image.open(img_path))\n        np.save(f\"{path}/frame_{(i):04d}.npy\", depth_maps[i].detach().cpu().numpy())\n\n    # comment this as it may fail sometimes\n    # images[0].save(f'{path}/_depth_maps.gif', save_all=True, append_images=images[1:], duration=100, loop=0)\n\n    return depth_maps\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    # Do some plotting.\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n\n    tick_cnt = 6\n    tick_loc = np.linspace(vmin, vmax, tick_cnt)\n    cb1 = mpl.colorbar.ColorbarBase(\n        ax, cmap=cmap, norm=norm, ticks=tick_loc, orientation=\"vertical\"\n    )\n\n    tick_label = [str(np.round(x, cbar_precision)) for x in tick_loc]\n    if cbar_precision == 0:\n        tick_label = [x[:-2] for x in tick_label]\n\n    cb1.set_ticklabels(tick_label)\n\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    # fig.tight_layout()\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]\n    \"\"\"\n    if range is not None:\n        vmin, vmax = range\n    elif mask is not None:\n        # vmin, vmax = np.percentile(x[mask], (2, 100))\n        vmin = np.min(x[mask][np.nonzero(x[mask])])\n        vmax = np.max(x[mask])\n        # vmin = vmin - np.abs(vmin) * 0.01\n        x[np.logical_not(mask)] = vmin\n        # print(vmin, vmax)\n    else:\n        vmin, vmax = np.percentile(x, (1, 100))\n        vmax += 1e-6\n\n    x = np.clip(x, vmin, vmax)\n    x = (x - vmin) / (vmax - vmin)\n    # x = np.clip(x, 0., 1.)\n\n    cmap = cm.get_cmap(cmap_name)\n    x_new = cmap(x)[:, :, :3]\n\n    if mask is not None:\n        mask = np.float32(mask[:, :, np.newaxis])\n        x_new = x_new * mask + np.ones_like(x_new) * (1.0 - mask)\n\n    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\n# tensor\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.utils.save_focals","uri":"program://Human3R/function/eval.video_depth.utils.save_focals#L20-L24","kind":"function","name":"save_focals","path":"eval/video_depth/utils.py","language":"python","start_line":20,"end_line":24,"context_start_line":1,"context_end_line":44,"code":"from copy import deepcopy\nimport cv2\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport roma\nfrom copy import deepcopy\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.utils.save_intrinsics","uri":"program://Human3R/function/eval.video_depth.utils.save_intrinsics#L27-L34","kind":"function","name":"save_intrinsics","path":"eval/video_depth/utils.py","language":"python","start_line":27,"end_line":34,"context_start_line":7,"context_end_line":54,"code":"import roma\nfrom copy import deepcopy\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.utils.save_conf_maps","uri":"program://Human3R/function/eval.video_depth.utils.save_conf_maps#L37-L40","kind":"function","name":"save_conf_maps","path":"eval/video_depth/utils.py","language":"python","start_line":37,"end_line":40,"context_start_line":17,"context_end_line":60,"code":"from matplotlib.figure import Figure\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt\n    focals = cam_dict[\"focal\"]\n    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)\n    min_depth = depth_maps.min()  # float(torch.quantile(out, 0.01))\n    max_depth = depth_maps.max()  # float(torch.quantile(out, 0.99))\n    colored_depth = colorize(\n        depth_maps,\n        cmap_name=\"Spectral_r\",\n        range=(min_depth, max_depth),","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.utils.save_rgb_imgs","uri":"program://Human3R/function/eval.video_depth.utils.save_rgb_imgs#L43-L50","kind":"function","name":"save_rgb_imgs","path":"eval/video_depth/utils.py","language":"python","start_line":43,"end_line":50,"context_start_line":23,"context_end_line":70,"code":"    np.savetxt(path, focals, fmt=\"%.6f\")\n    return focals\n\n\ndef save_intrinsics(cam_dict, path):\n    K_raw = np.eye(3)[None].repeat(len(cam_dict[\"focal\"]), axis=0)\n    K_raw[:, 0, 0] = cam_dict[\"focal\"]\n    K_raw[:, 1, 1] = cam_dict[\"focal\"]\n    K_raw[:, :2, 2] = cam_dict[\"pp\"]\n    K = K_raw.reshape(-1, 9)\n    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)\n    min_depth = depth_maps.min()  # float(torch.quantile(out, 0.01))\n    max_depth = depth_maps.max()  # float(torch.quantile(out, 0.99))\n    colored_depth = colorize(\n        depth_maps,\n        cmap_name=\"Spectral_r\",\n        range=(min_depth, max_depth),\n        append_cbar=True,\n    )\n    images = []\n\n    if conf_self is not None:\n        conf_selfs = torch.concat(conf_self, 0)\n        min_conf = torch.log(conf_selfs.min())  # float(torch.quantile(out, 0.01))\n        max_conf = torch.log(conf_selfs.max())  # float(torch.quantile(out, 0.99))\n        colored_conf = colorize(\n            torch.log(conf_selfs),","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.utils.save_depth_maps","uri":"program://Human3R/function/eval.video_depth.utils.save_depth_maps#L53-L91","kind":"function","name":"save_depth_maps","path":"eval/video_depth/utils.py","language":"python","start_line":53,"end_line":91,"context_start_line":33,"context_end_line":111,"code":"    np.savetxt(path, K, fmt=\"%.6f\")\n    return K_raw\n\n\ndef save_conf_maps(conf, path):\n    for i, c in enumerate(conf):\n        np.save(f\"{path}/conf_{i}.npy\", c.detach().cpu().numpy())\n    return conf\n\n\ndef save_rgb_imgs(colors, path):\n    imgs = colors\n    for i, img in enumerate(imgs):\n        # convert from rgb to bgr\n        iio.imwrite(\n            f\"{path}/frame_{i:04d}.jpg\", (img.cpu().numpy() * 255).astype(np.uint8)\n        )\n    return imgs\n\n\ndef save_depth_maps(pts3ds_self, path, conf_self=None):\n    depth_maps = torch.stack([pts3d_self[..., -1] for pts3d_self in pts3ds_self], 0)\n    min_depth = depth_maps.min()  # float(torch.quantile(out, 0.01))\n    max_depth = depth_maps.max()  # float(torch.quantile(out, 0.99))\n    colored_depth = colorize(\n        depth_maps,\n        cmap_name=\"Spectral_r\",\n        range=(min_depth, max_depth),\n        append_cbar=True,\n    )\n    images = []\n\n    if conf_self is not None:\n        conf_selfs = torch.concat(conf_self, 0)\n        min_conf = torch.log(conf_selfs.min())  # float(torch.quantile(out, 0.01))\n        max_conf = torch.log(conf_selfs.max())  # float(torch.quantile(out, 0.99))\n        colored_conf = colorize(\n            torch.log(conf_selfs),\n            cmap_name=\"jet\",\n            range=(min_conf, max_conf),\n            append_cbar=True,\n        )\n\n    for i, depth_map in enumerate(colored_depth):\n        # Apply color map to depth map\n        img_path = f\"{path}/frame_{(i):04d}.png\"\n        if conf_self is None:\n            to_save = (depth_map * 255).detach().cpu().numpy().astype(np.uint8)\n        else:\n            to_save = torch.cat([depth_map, colored_conf[i]], dim=1)\n            to_save = (to_save * 255).detach().cpu().numpy().astype(np.uint8)\n        iio.imwrite(img_path, to_save)\n        images.append(Image.open(img_path))\n        np.save(f\"{path}/frame_{(i):04d}.npy\", depth_maps[i].detach().cpu().numpy())\n\n    # comment this as it may fail sometimes\n    # images[0].save(f'{path}/_depth_maps.gif', save_all=True, append_images=images[1:], duration=100, loop=0)\n\n    return depth_maps\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    # Do some plotting.\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.utils.get_vertical_colorbar","uri":"program://Human3R/function/eval.video_depth.utils.get_vertical_colorbar#L94-L141","kind":"function","name":"get_vertical_colorbar","path":"eval/video_depth/utils.py","language":"python","start_line":94,"end_line":141,"context_start_line":74,"context_end_line":161,"code":"        )\n\n    for i, depth_map in enumerate(colored_depth):\n        # Apply color map to depth map\n        img_path = f\"{path}/frame_{(i):04d}.png\"\n        if conf_self is None:\n            to_save = (depth_map * 255).detach().cpu().numpy().astype(np.uint8)\n        else:\n            to_save = torch.cat([depth_map, colored_conf[i]], dim=1)\n            to_save = (to_save * 255).detach().cpu().numpy().astype(np.uint8)\n        iio.imwrite(img_path, to_save)\n        images.append(Image.open(img_path))\n        np.save(f\"{path}/frame_{(i):04d}.npy\", depth_maps[i].detach().cpu().numpy())\n\n    # comment this as it may fail sometimes\n    # images[0].save(f'{path}/_depth_maps.gif', save_all=True, append_images=images[1:], duration=100, loop=0)\n\n    return depth_maps\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    # Do some plotting.\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n\n    tick_cnt = 6\n    tick_loc = np.linspace(vmin, vmax, tick_cnt)\n    cb1 = mpl.colorbar.ColorbarBase(\n        ax, cmap=cmap, norm=norm, ticks=tick_loc, orientation=\"vertical\"\n    )\n\n    tick_label = [str(np.round(x, cbar_precision)) for x in tick_loc]\n    if cbar_precision == 0:\n        tick_label = [x[:-2] for x in tick_label]\n\n    cb1.set_ticklabels(tick_label)\n\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    # fig.tight_layout()\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.utils.colorize_np","uri":"program://Human3R/function/eval.video_depth.utils.colorize_np#L144-L204","kind":"function","name":"colorize_np","path":"eval/video_depth/utils.py","language":"python","start_line":144,"end_line":204,"context_start_line":124,"context_end_line":224,"code":"\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    # fig.tight_layout()\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]\n    \"\"\"\n    if range is not None:\n        vmin, vmax = range\n    elif mask is not None:\n        # vmin, vmax = np.percentile(x[mask], (2, 100))\n        vmin = np.min(x[mask][np.nonzero(x[mask])])\n        vmax = np.max(x[mask])\n        # vmin = vmin - np.abs(vmin) * 0.01\n        x[np.logical_not(mask)] = vmin\n        # print(vmin, vmax)\n    else:\n        vmin, vmax = np.percentile(x, (1, 100))\n        vmax += 1e-6\n\n    x = np.clip(x, vmin, vmax)\n    x = (x - vmin) / (vmax - vmin)\n    # x = np.clip(x, 0., 1.)\n\n    cmap = cm.get_cmap(cmap_name)\n    x_new = cmap(x)[:, :, :3]\n\n    if mask is not None:\n        mask = np.float32(mask[:, :, np.newaxis])\n        x_new = x_new * mask + np.ones_like(x_new) * (1.0 - mask)\n\n    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\n# tensor\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.utils.colorize","uri":"program://Human3R/function/eval.video_depth.utils.colorize#L208-L236","kind":"function","name":"colorize","path":"eval/video_depth/utils.py","language":"python","start_line":208,"end_line":236,"context_start_line":188,"context_end_line":236,"code":"        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\n# tensor\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.launch","uri":"program://Human3R/module/eval.video_depth.launch#L1-L375","kind":"module","name":"eval.video_depth.launch","path":"eval/video_depth/launch.py","language":"python","start_line":1,"end_line":375,"context_start_line":1,"context_end_line":375,"code":"import os\nimport sys\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nimport math\nimport cv2\nimport numpy as np\nimport torch\nimport argparse\n\nfrom copy import deepcopy\nfrom eval.video_depth.metadata import dataset_metadata\nfrom eval.video_depth.utils import save_depth_maps\nfrom accelerate import PartialState\nfrom add_ckpt_path import add_path_to_dust3r\nimport time\nfrom tqdm import tqdm\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--weights\",\n        type=str,\n        help=\"path to the model weights\",\n        default=\"\",\n    )\n\n    parser.add_argument(\"--device\", type=str, default=\"cuda\", help=\"pytorch device\")\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"\",\n        help=\"value for outdir\",\n    )\n    parser.add_argument(\n        \"--no_crop\", type=bool, default=True, help=\"whether to crop input data\"\n    )\n\n    parser.add_argument(\n        \"--eval_dataset\",\n        type=str,\n        default=\"sintel\",\n        choices=list(dataset_metadata.keys()),\n    )\n    parser.add_argument(\"--size\", type=int, default=\"224\")\n\n    parser.add_argument(\n        \"--pose_eval_stride\", default=1, type=int, help=\"stride for pose evaluation\"\n    )\n    parser.add_argument(\n        \"--full_seq\",\n        action=\"store_true\",\n        default=False,\n        help=\"use full sequence for pose evaluation\",\n    )\n    parser.add_argument(\n        \"--seq_list\",\n        nargs=\"+\",\n        default=None,\n        help=\"list of sequences for pose evaluation\",\n    )\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    return parser\n\n\ndef eval_pose_estimation(args, model, save_dir=None):\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mask_path = metadata[\"mask_path\"]\n\n    ate_mean, rpe_trans_mean, rpe_rot_mean = eval_pose_estimation_dist(\n        args, model, save_dir=save_dir, img_path=img_path, mask_path=mask_path\n    )\n    return ate_mean, rpe_trans_mean, rpe_rot_mean\n\n\ndef eval_pose_estimation_dist(args, model, img_path, save_dir=None, mask_path=None):\n    from dust3r.inference import inference, inference_recurrent_lighter\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    anno_path = metadata.get(\"anno_path\", None)\n\n    seq_list = args.seq_list\n    if seq_list is None:\n        if metadata.get(\"full_seq\", False):\n            args.full_seq = True\n        else:\n            seq_list = metadata.get(\"seq_list\", [])\n        if args.full_seq:\n            seq_list = os.listdir(img_path)\n            seq_list = [\n                seq for seq in seq_list if os.path.isdir(os.path.join(img_path, seq))\n            ]\n        seq_list = sorted(seq_list)\n\n    if save_dir is None:\n        save_dir = args.output_dir\n\n    distributed_state = PartialState()\n    model.to(distributed_state.device)\n    device = distributed_state.device\n    img_res = getattr(model, 'mhmr_img_res', None)\n\n    with distributed_state.split_between_processes(seq_list) as seqs:\n        ate_list = []\n        rpe_trans_list = []\n        rpe_rot_list = []\n        load_img_size = args.size\n        assert load_img_size == 512\n        error_log_path = f\"{save_dir}/_error_log_{distributed_state.process_index}.txt\"  # Unique log file per process\n        bug = False\n        for seq in tqdm(seqs):\n            try:\n                dir_path = metadata[\"dir_path_func\"](img_path, seq)\n\n                # Handle skip_condition\n                skip_condition = metadata.get(\"skip_condition\", None)\n                if skip_condition is not None and skip_condition(save_dir, seq):\n                    continue\n\n                mask_path_seq_func = metadata.get(\n                    \"mask_path_seq_func\", lambda mask_path, seq: None\n                )\n                mask_path_seq = mask_path_seq_func(mask_path, seq)\n\n                filelist = [\n                    os.path.join(dir_path, name) for name in os.listdir(dir_path)\n                ]\n                filelist.sort()\n                filelist = filelist[:: args.pose_eval_stride]\n\n                views = prepare_input(\n                    filelist,\n                    [True for _ in filelist],\n                    size=load_img_size,\n                    crop=not args.no_crop,\n                    img_res=img_res,\n                    reset_interval=args.reset_interval,\n                )\n                start = time.time()\n                # outputs, _ = inference(views, model, device)\n                outputs, _ = inference_recurrent_lighter(\n                    views, model, device, verbose=False, use_ttt3r=args.use_ttt3r)\n                end = time.time()\n                fps = len(filelist) / (end - start)\n\n                (\n                    colors,\n                    pts3ds_self,\n                    pts3ds_other,\n                    conf_self,\n                    conf_other,\n                    cam_dict,\n                    pr_poses,\n                ) = prepare_output(outputs)\n\n                os.makedirs(f\"{save_dir}/{seq}\", exist_ok=True)\n                save_depth_maps(pts3ds_self, f\"{save_dir}/{seq}\", conf_self=conf_self)\n\n            except Exception as e:\n                if \"out of memory\" in str(e):\n                    # Handle OOM\n                    torch.cuda.empty_cache()  # Clear the CUDA memory\n                    with open(error_log_path, \"a\") as f:\n                        f.write(\n                            f\"OOM error in sequence {seq}, skipping this sequence.\\n\"\n                        )\n                    print(f\"OOM error in sequence {seq}, skipping...\")\n                elif \"Degenerate covariance rank\" in str(\n                    e\n                ) or \"Eigenvalues did not converge\" in str(e):\n                    # Handle Degenerate covariance rank exception and Eigenvalues did not converge exception\n                    with open(error_log_path, \"a\") as f:\n                        f.write(f\"Exception in sequence {seq}: {str(e)}\\n\")\n                    print(f\"Traj evaluation error in sequence {seq}, skipping.\")\n                else:\n                    raise e  # Rethrow if it's not an expected exception\n    return None, None, None\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    add_path_to_dust3r(args.weights)\n    from dust3r.utils.image import load_images_for_eval as load_images\n    from src.dust3r.utils.image import pad_image\n    from dust3r.post_process import estimate_focal_knowing_depth\n    from dust3r.model import ARCroco3DStereo\n    from dust3r.utils.camera import pose_encoding_to_camera\n    from dust3r.utils.geometry import matrix_cumprod, get_camera_parameters\n\n    if args.eval_dataset == \"sintel\":\n        args.full_seq = True\n    else:\n        args.full_seq = False\n    args.no_crop = True\n\n    def prepare_input(\n        img_paths,\n        img_mask,\n        size,\n        raymaps=None,\n        raymap_mask=None,\n        revisit=1,\n        update=True,\n        crop=True,\n        img_res=None, \n        reset_interval=100,\n    ):\n        images = load_images(img_paths, size=size, crop=crop)\n        if img_res is not None:\n            K_mhmr = get_camera_parameters(img_res, device=\"cpu\") # if use pseudo K\n\n        views = []\n        if raymaps is None and raymap_mask is None:\n            num_views = len(images)\n            # Only images are provided.\n            for i in range(num_views):\n                view = {\n                    \"img\": images[i][\"img\"],\n                    \"ray_map\": torch.full(\n                        (\n                            images[i][\"img\"].shape[0],\n                            6,\n                            images[i][\"img\"].shape[-2],\n                            images[i][\"img\"].shape[-1],\n                        ),\n                        torch.nan,\n                    ),\n                    \"true_shape\": torch.from_numpy(images[i][\"true_shape\"]),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        np.eye(4).astype(np.float32)\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(True).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(False).unsqueeze(0),\n                    \"update\": torch.tensor(True).unsqueeze(0),\n                    # \"reset\": torch.tensor(False).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n                if img_res is not None:\n                    view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                    view[\"K_mhmr\"] = K_mhmr\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n        else:\n            # Combine images and raymaps.\n            num_views = len(images) + len(raymaps)\n            assert len(img_mask) == len(raymap_mask) == num_views\n            assert sum(img_mask) == len(images) and sum(raymap_mask) == len(raymaps)\n\n            j = 0\n            k = 0\n            for i in range(num_views):\n                view = {\n                    \"img\": (\n                        images[j][\"img\"]\n                        if img_mask[i]\n                        else torch.full_like(images[0][\"img\"], torch.nan)\n                    ),\n                    \"ray_map\": (\n                        raymaps[k]\n                        if raymap_mask[i]\n                        else torch.full_like(raymaps[0], torch.nan)\n                    ),\n                    \"true_shape\": (\n                        torch.from_numpy(images[j][\"true_shape\"])\n                        if img_mask[i]\n                        else torch.from_numpy(np.int32([raymaps[k].shape[1:-1][::-1]]))\n                    ),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        np.eye(4).astype(np.float32)\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(raymap_mask[i]).unsqueeze(0),\n                    \"update\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    # \"reset\": torch.tensor(False).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n                if img_res is not None:\n                    view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                    view[\"K_mhmr\"] = K_mhmr\n                if img_mask[i]:\n                    j += 1\n                if raymap_mask[i]:\n                    k += 1\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n            assert j == len(images) and k == len(raymaps)\n\n        if revisit > 1:\n            # repeat input for 'revisit' times\n            new_views = []\n            for r in range(revisit):\n                for i in range(len(views)):\n                    new_view = deepcopy(views[i])\n                    new_view[\"idx\"] = r * len(views) + i\n                    new_view[\"instance\"] = str(r * len(views) + i)\n                    if r > 0:\n                        if not update:\n                            new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                    new_views.append(new_view)\n            return new_views\n        return views\n\n    def prepare_output(outputs, revisit=1):\n        valid_length = len(outputs[\"pred\"]) // revisit\n        outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n        outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n        # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n        reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n        shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n        outputs[\"pred\"] = [\n            pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n        reset_mask = reset_mask[~shifted_reset_mask]\n\n        pts3ds_self = [output[\"pts3d_in_self_view\"].cpu() for output in outputs[\"pred\"]]\n        pts3ds_other = [\n            output[\"pts3d_in_other_view\"].cpu() for output in outputs[\"pred\"]\n        ]\n        conf_self = [output[\"conf_self\"].cpu() for output in outputs[\"pred\"]]\n        conf_other = [output[\"conf\"].cpu() for output in outputs[\"pred\"]]\n        pts3ds_self = torch.cat(pts3ds_self, 0)\n        pr_poses = [\n            pose_encoding_to_camera(pred[\"camera_pose\"].clone()).cpu()\n            for pred in outputs[\"pred\"]\n        ]\n        pr_poses = torch.cat(pr_poses, 0)\n\n        B, H, W, _ = pts3ds_self.shape\n        pp = (\n            torch.tensor([W // 2, H // 2], device=pts3ds_self.device)\n            .float()\n            .repeat(B, 1)\n            .reshape(B, 2)\n        )\n        focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n        if reset_mask.any():\n            identity = torch.eye(4, device=pr_poses.device)\n            reset_poses = torch.where(reset_mask.unsqueeze(-1).unsqueeze(-1), pr_poses, identity)\n            cumulative_bases = matrix_cumprod(reset_poses)\n            shifted_bases = torch.cat([identity.unsqueeze(0), cumulative_bases[:-1]], dim=0)\n            pr_poses = torch.einsum('bij,bjk->bik', shifted_bases, pr_poses)\n\n        colors = [0.5 * (output[\"rgb\"][0] + 1.0) for output in outputs[\"pred\"]]\n        cam_dict = {\n            \"focal\": focal.cpu().numpy(),\n            \"pp\": pp.cpu().numpy(),\n        }\n        return (\n            colors,\n            pts3ds_self,\n            pts3ds_other,\n            conf_self,\n            conf_other,\n            cam_dict,\n            pr_poses,\n        )\n\n    model = ARCroco3DStereo.from_pretrained(args.weights)\n    eval_pose_estimation(args, model, save_dir=args.output_dir)","source_hash":"3f359473cd914d2450b9679c39667f98d970a03cdaf1d817c168b6c112cbfc1f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.launch.get_args_parser","uri":"program://Human3R/function/eval.video_depth.launch.get_args_parser#L20-L66","kind":"function","name":"get_args_parser","path":"eval/video_depth/launch.py","language":"python","start_line":20,"end_line":66,"context_start_line":1,"context_end_line":86,"code":"import os\nimport sys\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nimport math\nimport cv2\nimport numpy as np\nimport torch\nimport argparse\n\nfrom copy import deepcopy\nfrom eval.video_depth.metadata import dataset_metadata\nfrom eval.video_depth.utils import save_depth_maps\nfrom accelerate import PartialState\nfrom add_ckpt_path import add_path_to_dust3r\nimport time\nfrom tqdm import tqdm\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--weights\",\n        type=str,\n        help=\"path to the model weights\",\n        default=\"\",\n    )\n\n    parser.add_argument(\"--device\", type=str, default=\"cuda\", help=\"pytorch device\")\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"\",\n        help=\"value for outdir\",\n    )\n    parser.add_argument(\n        \"--no_crop\", type=bool, default=True, help=\"whether to crop input data\"\n    )\n\n    parser.add_argument(\n        \"--eval_dataset\",\n        type=str,\n        default=\"sintel\",\n        choices=list(dataset_metadata.keys()),\n    )\n    parser.add_argument(\"--size\", type=int, default=\"224\")\n\n    parser.add_argument(\n        \"--pose_eval_stride\", default=1, type=int, help=\"stride for pose evaluation\"\n    )\n    parser.add_argument(\n        \"--full_seq\",\n        action=\"store_true\",\n        default=False,\n        help=\"use full sequence for pose evaluation\",\n    )\n    parser.add_argument(\n        \"--seq_list\",\n        nargs=\"+\",\n        default=None,\n        help=\"list of sequences for pose evaluation\",\n    )\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    return parser\n\n\ndef eval_pose_estimation(args, model, save_dir=None):\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mask_path = metadata[\"mask_path\"]\n\n    ate_mean, rpe_trans_mean, rpe_rot_mean = eval_pose_estimation_dist(\n        args, model, save_dir=save_dir, img_path=img_path, mask_path=mask_path\n    )\n    return ate_mean, rpe_trans_mean, rpe_rot_mean\n\n\ndef eval_pose_estimation_dist(args, model, img_path, save_dir=None, mask_path=None):\n    from dust3r.inference import inference, inference_recurrent_lighter\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    anno_path = metadata.get(\"anno_path\", None)\n\n    seq_list = args.seq_list","source_hash":"3f359473cd914d2450b9679c39667f98d970a03cdaf1d817c168b6c112cbfc1f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.launch.eval_pose_estimation","uri":"program://Human3R/function/eval.video_depth.launch.eval_pose_estimation#L69-L77","kind":"function","name":"eval_pose_estimation","path":"eval/video_depth/launch.py","language":"python","start_line":69,"end_line":77,"context_start_line":49,"context_end_line":97,"code":"    parser.add_argument(\n        \"--pose_eval_stride\", default=1, type=int, help=\"stride for pose evaluation\"\n    )\n    parser.add_argument(\n        \"--full_seq\",\n        action=\"store_true\",\n        default=False,\n        help=\"use full sequence for pose evaluation\",\n    )\n    parser.add_argument(\n        \"--seq_list\",\n        nargs=\"+\",\n        default=None,\n        help=\"list of sequences for pose evaluation\",\n    )\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    return parser\n\n\ndef eval_pose_estimation(args, model, save_dir=None):\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mask_path = metadata[\"mask_path\"]\n\n    ate_mean, rpe_trans_mean, rpe_rot_mean = eval_pose_estimation_dist(\n        args, model, save_dir=save_dir, img_path=img_path, mask_path=mask_path\n    )\n    return ate_mean, rpe_trans_mean, rpe_rot_mean\n\n\ndef eval_pose_estimation_dist(args, model, img_path, save_dir=None, mask_path=None):\n    from dust3r.inference import inference, inference_recurrent_lighter\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    anno_path = metadata.get(\"anno_path\", None)\n\n    seq_list = args.seq_list\n    if seq_list is None:\n        if metadata.get(\"full_seq\", False):\n            args.full_seq = True\n        else:\n            seq_list = metadata.get(\"seq_list\", [])\n        if args.full_seq:\n            seq_list = os.listdir(img_path)\n            seq_list = [\n                seq for seq in seq_list if os.path.isdir(os.path.join(img_path, seq))\n            ]\n        seq_list = sorted(seq_list)","source_hash":"3f359473cd914d2450b9679c39667f98d970a03cdaf1d817c168b6c112cbfc1f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.launch.eval_pose_estimation_dist","uri":"program://Human3R/function/eval.video_depth.launch.eval_pose_estimation_dist#L80-L181","kind":"function","name":"eval_pose_estimation_dist","path":"eval/video_depth/launch.py","language":"python","start_line":80,"end_line":181,"context_start_line":60,"context_end_line":201,"code":"        nargs=\"+\",\n        default=None,\n        help=\"list of sequences for pose evaluation\",\n    )\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    return parser\n\n\ndef eval_pose_estimation(args, model, save_dir=None):\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mask_path = metadata[\"mask_path\"]\n\n    ate_mean, rpe_trans_mean, rpe_rot_mean = eval_pose_estimation_dist(\n        args, model, save_dir=save_dir, img_path=img_path, mask_path=mask_path\n    )\n    return ate_mean, rpe_trans_mean, rpe_rot_mean\n\n\ndef eval_pose_estimation_dist(args, model, img_path, save_dir=None, mask_path=None):\n    from dust3r.inference import inference, inference_recurrent_lighter\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    anno_path = metadata.get(\"anno_path\", None)\n\n    seq_list = args.seq_list\n    if seq_list is None:\n        if metadata.get(\"full_seq\", False):\n            args.full_seq = True\n        else:\n            seq_list = metadata.get(\"seq_list\", [])\n        if args.full_seq:\n            seq_list = os.listdir(img_path)\n            seq_list = [\n                seq for seq in seq_list if os.path.isdir(os.path.join(img_path, seq))\n            ]\n        seq_list = sorted(seq_list)\n\n    if save_dir is None:\n        save_dir = args.output_dir\n\n    distributed_state = PartialState()\n    model.to(distributed_state.device)\n    device = distributed_state.device\n    img_res = getattr(model, 'mhmr_img_res', None)\n\n    with distributed_state.split_between_processes(seq_list) as seqs:\n        ate_list = []\n        rpe_trans_list = []\n        rpe_rot_list = []\n        load_img_size = args.size\n        assert load_img_size == 512\n        error_log_path = f\"{save_dir}/_error_log_{distributed_state.process_index}.txt\"  # Unique log file per process\n        bug = False\n        for seq in tqdm(seqs):\n            try:\n                dir_path = metadata[\"dir_path_func\"](img_path, seq)\n\n                # Handle skip_condition\n                skip_condition = metadata.get(\"skip_condition\", None)\n                if skip_condition is not None and skip_condition(save_dir, seq):\n                    continue\n\n                mask_path_seq_func = metadata.get(\n                    \"mask_path_seq_func\", lambda mask_path, seq: None\n                )\n                mask_path_seq = mask_path_seq_func(mask_path, seq)\n\n                filelist = [\n                    os.path.join(dir_path, name) for name in os.listdir(dir_path)\n                ]\n                filelist.sort()\n                filelist = filelist[:: args.pose_eval_stride]\n\n                views = prepare_input(\n                    filelist,\n                    [True for _ in filelist],\n                    size=load_img_size,\n                    crop=not args.no_crop,\n                    img_res=img_res,\n                    reset_interval=args.reset_interval,\n                )\n                start = time.time()\n                # outputs, _ = inference(views, model, device)\n                outputs, _ = inference_recurrent_lighter(\n                    views, model, device, verbose=False, use_ttt3r=args.use_ttt3r)\n                end = time.time()\n                fps = len(filelist) / (end - start)\n\n                (\n                    colors,\n                    pts3ds_self,\n                    pts3ds_other,\n                    conf_self,\n                    conf_other,\n                    cam_dict,\n                    pr_poses,\n                ) = prepare_output(outputs)\n\n                os.makedirs(f\"{save_dir}/{seq}\", exist_ok=True)\n                save_depth_maps(pts3ds_self, f\"{save_dir}/{seq}\", conf_self=conf_self)\n\n            except Exception as e:\n                if \"out of memory\" in str(e):\n                    # Handle OOM\n                    torch.cuda.empty_cache()  # Clear the CUDA memory\n                    with open(error_log_path, \"a\") as f:\n                        f.write(\n                            f\"OOM error in sequence {seq}, skipping this sequence.\\n\"\n                        )\n                    print(f\"OOM error in sequence {seq}, skipping...\")\n                elif \"Degenerate covariance rank\" in str(\n                    e\n                ) or \"Eigenvalues did not converge\" in str(e):\n                    # Handle Degenerate covariance rank exception and Eigenvalues did not converge exception\n                    with open(error_log_path, \"a\") as f:\n                        f.write(f\"Exception in sequence {seq}: {str(e)}\\n\")\n                    print(f\"Traj evaluation error in sequence {seq}, skipping.\")\n                else:\n                    raise e  # Rethrow if it's not an expected exception\n    return None, None, None\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    add_path_to_dust3r(args.weights)\n    from dust3r.utils.image import load_images_for_eval as load_images\n    from src.dust3r.utils.image import pad_image\n    from dust3r.post_process import estimate_focal_knowing_depth\n    from dust3r.model import ARCroco3DStereo\n    from dust3r.utils.camera import pose_encoding_to_camera\n    from dust3r.utils.geometry import matrix_cumprod, get_camera_parameters\n\n    if args.eval_dataset == \"sintel\":\n        args.full_seq = True\n    else:\n        args.full_seq = False\n    args.no_crop = True\n\n    def prepare_input(","source_hash":"3f359473cd914d2450b9679c39667f98d970a03cdaf1d817c168b6c112cbfc1f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.launch.prepare_input","uri":"program://Human3R/function/eval.video_depth.launch.prepare_input#L201-L316","kind":"function","name":"prepare_input","path":"eval/video_depth/launch.py","language":"python","start_line":201,"end_line":316,"context_start_line":181,"context_end_line":336,"code":"    return None, None, None\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    add_path_to_dust3r(args.weights)\n    from dust3r.utils.image import load_images_for_eval as load_images\n    from src.dust3r.utils.image import pad_image\n    from dust3r.post_process import estimate_focal_knowing_depth\n    from dust3r.model import ARCroco3DStereo\n    from dust3r.utils.camera import pose_encoding_to_camera\n    from dust3r.utils.geometry import matrix_cumprod, get_camera_parameters\n\n    if args.eval_dataset == \"sintel\":\n        args.full_seq = True\n    else:\n        args.full_seq = False\n    args.no_crop = True\n\n    def prepare_input(\n        img_paths,\n        img_mask,\n        size,\n        raymaps=None,\n        raymap_mask=None,\n        revisit=1,\n        update=True,\n        crop=True,\n        img_res=None, \n        reset_interval=100,\n    ):\n        images = load_images(img_paths, size=size, crop=crop)\n        if img_res is not None:\n            K_mhmr = get_camera_parameters(img_res, device=\"cpu\") # if use pseudo K\n\n        views = []\n        if raymaps is None and raymap_mask is None:\n            num_views = len(images)\n            # Only images are provided.\n            for i in range(num_views):\n                view = {\n                    \"img\": images[i][\"img\"],\n                    \"ray_map\": torch.full(\n                        (\n                            images[i][\"img\"].shape[0],\n                            6,\n                            images[i][\"img\"].shape[-2],\n                            images[i][\"img\"].shape[-1],\n                        ),\n                        torch.nan,\n                    ),\n                    \"true_shape\": torch.from_numpy(images[i][\"true_shape\"]),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        np.eye(4).astype(np.float32)\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(True).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(False).unsqueeze(0),\n                    \"update\": torch.tensor(True).unsqueeze(0),\n                    # \"reset\": torch.tensor(False).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n                if img_res is not None:\n                    view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                    view[\"K_mhmr\"] = K_mhmr\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n        else:\n            # Combine images and raymaps.\n            num_views = len(images) + len(raymaps)\n            assert len(img_mask) == len(raymap_mask) == num_views\n            assert sum(img_mask) == len(images) and sum(raymap_mask) == len(raymaps)\n\n            j = 0\n            k = 0\n            for i in range(num_views):\n                view = {\n                    \"img\": (\n                        images[j][\"img\"]\n                        if img_mask[i]\n                        else torch.full_like(images[0][\"img\"], torch.nan)\n                    ),\n                    \"ray_map\": (\n                        raymaps[k]\n                        if raymap_mask[i]\n                        else torch.full_like(raymaps[0], torch.nan)\n                    ),\n                    \"true_shape\": (\n                        torch.from_numpy(images[j][\"true_shape\"])\n                        if img_mask[i]\n                        else torch.from_numpy(np.int32([raymaps[k].shape[1:-1][::-1]]))\n                    ),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        np.eye(4).astype(np.float32)\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(raymap_mask[i]).unsqueeze(0),\n                    \"update\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    # \"reset\": torch.tensor(False).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n                if img_res is not None:\n                    view[\"img_mhmr\"] = pad_image(view[\"img\"], img_res)\n                    view[\"K_mhmr\"] = K_mhmr\n                if img_mask[i]:\n                    j += 1\n                if raymap_mask[i]:\n                    k += 1\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n            assert j == len(images) and k == len(raymaps)\n\n        if revisit > 1:\n            # repeat input for 'revisit' times\n            new_views = []\n            for r in range(revisit):\n                for i in range(len(views)):\n                    new_view = deepcopy(views[i])\n                    new_view[\"idx\"] = r * len(views) + i\n                    new_view[\"instance\"] = str(r * len(views) + i)\n                    if r > 0:\n                        if not update:\n                            new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                    new_views.append(new_view)\n            return new_views\n        return views\n\n    def prepare_output(outputs, revisit=1):\n        valid_length = len(outputs[\"pred\"]) // revisit\n        outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n        outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n        # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n        reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n        shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n        outputs[\"pred\"] = [\n            pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n        reset_mask = reset_mask[~shifted_reset_mask]\n\n        pts3ds_self = [output[\"pts3d_in_self_view\"].cpu() for output in outputs[\"pred\"]]\n        pts3ds_other = [\n            output[\"pts3d_in_other_view\"].cpu() for output in outputs[\"pred\"]\n        ]\n        conf_self = [output[\"conf_self\"].cpu() for output in outputs[\"pred\"]]\n        conf_other = [output[\"conf\"].cpu() for output in outputs[\"pred\"]]\n        pts3ds_self = torch.cat(pts3ds_self, 0)","source_hash":"3f359473cd914d2450b9679c39667f98d970a03cdaf1d817c168b6c112cbfc1f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.launch.prepare_output","uri":"program://Human3R/function/eval.video_depth.launch.prepare_output#L318-L372","kind":"function","name":"prepare_output","path":"eval/video_depth/launch.py","language":"python","start_line":318,"end_line":372,"context_start_line":298,"context_end_line":375,"code":"                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n            assert j == len(images) and k == len(raymaps)\n\n        if revisit > 1:\n            # repeat input for 'revisit' times\n            new_views = []\n            for r in range(revisit):\n                for i in range(len(views)):\n                    new_view = deepcopy(views[i])\n                    new_view[\"idx\"] = r * len(views) + i\n                    new_view[\"instance\"] = str(r * len(views) + i)\n                    if r > 0:\n                        if not update:\n                            new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                    new_views.append(new_view)\n            return new_views\n        return views\n\n    def prepare_output(outputs, revisit=1):\n        valid_length = len(outputs[\"pred\"]) // revisit\n        outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n        outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n        # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n        reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n        shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n        outputs[\"pred\"] = [\n            pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n        reset_mask = reset_mask[~shifted_reset_mask]\n\n        pts3ds_self = [output[\"pts3d_in_self_view\"].cpu() for output in outputs[\"pred\"]]\n        pts3ds_other = [\n            output[\"pts3d_in_other_view\"].cpu() for output in outputs[\"pred\"]\n        ]\n        conf_self = [output[\"conf_self\"].cpu() for output in outputs[\"pred\"]]\n        conf_other = [output[\"conf\"].cpu() for output in outputs[\"pred\"]]\n        pts3ds_self = torch.cat(pts3ds_self, 0)\n        pr_poses = [\n            pose_encoding_to_camera(pred[\"camera_pose\"].clone()).cpu()\n            for pred in outputs[\"pred\"]\n        ]\n        pr_poses = torch.cat(pr_poses, 0)\n\n        B, H, W, _ = pts3ds_self.shape\n        pp = (\n            torch.tensor([W // 2, H // 2], device=pts3ds_self.device)\n            .float()\n            .repeat(B, 1)\n            .reshape(B, 2)\n        )\n        focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n        if reset_mask.any():\n            identity = torch.eye(4, device=pr_poses.device)\n            reset_poses = torch.where(reset_mask.unsqueeze(-1).unsqueeze(-1), pr_poses, identity)\n            cumulative_bases = matrix_cumprod(reset_poses)\n            shifted_bases = torch.cat([identity.unsqueeze(0), cumulative_bases[:-1]], dim=0)\n            pr_poses = torch.einsum('bij,bjk->bik', shifted_bases, pr_poses)\n\n        colors = [0.5 * (output[\"rgb\"][0] + 1.0) for output in outputs[\"pred\"]]\n        cam_dict = {\n            \"focal\": focal.cpu().numpy(),\n            \"pp\": pp.cpu().numpy(),\n        }\n        return (\n            colors,\n            pts3ds_self,\n            pts3ds_other,\n            conf_self,\n            conf_other,\n            cam_dict,\n            pr_poses,\n        )\n\n    model = ARCroco3DStereo.from_pretrained(args.weights)\n    eval_pose_estimation(args, model, save_dir=args.output_dir)","source_hash":"3f359473cd914d2450b9679c39667f98d970a03cdaf1d817c168b6c112cbfc1f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.tools","uri":"program://Human3R/module/eval.video_depth.tools#L1-L399","kind":"module","name":"eval.video_depth.tools","path":"eval/video_depth/tools.py","language":"python","start_line":1,"end_line":399,"context_start_line":1,"context_end_line":399,"code":"import torch\nimport numpy as np\nimport cv2\nimport glob\nimport argparse\nfrom pathlib import Path\nfrom tqdm import tqdm\nfrom copy import deepcopy\nfrom scipy.optimize import minimize\nimport os\nfrom collections import defaultdict\n\n\ndef group_by_directory(pathes, idx=-1):\n    \"\"\"\n    Groups the file paths based on the second-to-last directory in their paths.\n\n    Parameters:\n    - pathes (list): List of file paths.\n\n    Returns:\n    - dict: A dictionary where keys are the second-to-last directory names and values are lists of file paths.\n    \"\"\"\n    grouped_pathes = defaultdict(list)\n\n    for path in pathes:\n        # Extract the second-to-last directory\n        dir_name = os.path.dirname(path).split(\"/\")[idx]\n        grouped_pathes[dir_name].append(path)\n\n    return grouped_pathes\n\n\ndef depth2disparity(depth, return_mask=False):\n    if isinstance(depth, torch.Tensor):\n        disparity = torch.zeros_like(depth)\n    elif isinstance(depth, np.ndarray):\n        disparity = np.zeros_like(depth)\n    non_negtive_mask = depth > 0\n    disparity[non_negtive_mask] = 1.0 / depth[non_negtive_mask]\n    if return_mask:\n        return disparity, non_negtive_mask\n    else:\n        return disparity\n\n\ndef absolute_error_loss(params, predicted_depth, ground_truth_depth):\n    s, t = params\n\n    predicted_aligned = s * predicted_depth + t\n\n    abs_error = np.abs(predicted_aligned - ground_truth_depth)\n    return np.sum(abs_error)\n\n\ndef absolute_value_scaling(predicted_depth, ground_truth_depth, s=1, t=0):\n    predicted_depth_np = predicted_depth.cpu().numpy().reshape(-1)\n    ground_truth_depth_np = ground_truth_depth.cpu().numpy().reshape(-1)\n\n    initial_params = [s, t]  # s = 1, t = 0\n\n    result = minimize(\n        absolute_error_loss,\n        initial_params,\n        args=(predicted_depth_np, ground_truth_depth_np),\n    )\n\n    s, t = result.x\n    return s, t\n\n\ndef absolute_value_scaling2(\n    predicted_depth,\n    ground_truth_depth,\n    s_init=1.0,\n    t_init=0.0,\n    lr=1e-4,\n    max_iters=1000,\n    tol=1e-6,\n):\n    # Initialize s and t as torch tensors with requires_grad=True\n    s = torch.tensor(\n        [s_init],\n        requires_grad=True,\n        device=predicted_depth.device,\n        dtype=predicted_depth.dtype,\n    )\n    t = torch.tensor(\n        [t_init],\n        requires_grad=True,\n        device=predicted_depth.device,\n        dtype=predicted_depth.dtype,\n    )\n\n    optimizer = torch.optim.Adam([s, t], lr=lr)\n\n    prev_loss = None\n\n    for i in range(max_iters):\n        optimizer.zero_grad()\n\n        # Compute predicted aligned depth\n        predicted_aligned = s * predicted_depth + t\n\n        # Compute absolute error\n        abs_error = torch.abs(predicted_aligned - ground_truth_depth)\n\n        # Compute loss\n        loss = torch.sum(abs_error)\n\n        # Backpropagate\n        loss.backward()\n\n        # Update parameters\n        optimizer.step()\n\n        # Check convergence\n        if prev_loss is not None and torch.abs(prev_loss - loss) < tol:\n            break\n\n        prev_loss = loss.item()\n\n    return s.detach().item(), t.detach().item()\n\n\ndef depth_evaluation(\n    predicted_depth_original,\n    ground_truth_depth_original,\n    max_depth=80,\n    custom_mask=None,\n    post_clip_min=None,\n    post_clip_max=None,\n    pre_clip_min=None,\n    pre_clip_max=None,\n    align_with_lstsq=False,\n    align_with_lad=False,\n    align_with_lad2=False,\n    metric_scale=False,\n    lr=1e-4,\n    max_iters=1000,\n    use_gpu=False,\n    align_with_scale=False,\n    disp_input=False,\n):\n    \"\"\"\n    Evaluate the depth map using various metrics and return a depth error parity map, with an option for least squares alignment.\n\n    Args:\n        predicted_depth (numpy.ndarray or torch.Tensor): The predicted depth map.\n        ground_truth_depth (numpy.ndarray or torch.Tensor): The ground truth depth map.\n        max_depth (float): The maximum depth value to consider. Default is 80 meters.\n        align_with_lstsq (bool): If True, perform least squares alignment of the predicted depth with ground truth.\n\n    Returns:\n        dict: A dictionary containing the evaluation metrics.\n        torch.Tensor: The depth error parity map.\n    \"\"\"\n    if isinstance(predicted_depth_original, np.ndarray):\n        predicted_depth_original = torch.from_numpy(predicted_depth_original)\n    if isinstance(ground_truth_depth_original, np.ndarray):\n        ground_truth_depth_original = torch.from_numpy(ground_truth_depth_original)\n    if custom_mask is not None and isinstance(custom_mask, np.ndarray):\n        custom_mask = torch.from_numpy(custom_mask)\n\n    # if the dimension is 3, flatten to 2d along the batch dimension\n    if predicted_depth_original.dim() == 3:\n        _, h, w = predicted_depth_original.shape\n        predicted_depth_original = predicted_depth_original.view(-1, w)\n        ground_truth_depth_original = ground_truth_depth_original.view(-1, w)\n        if custom_mask is not None:\n            custom_mask = custom_mask.view(-1, w)\n\n    # put to device\n    if use_gpu:\n        predicted_depth_original = predicted_depth_original.cuda()\n        ground_truth_depth_original = ground_truth_depth_original.cuda()\n\n    # Filter out depths greater than max_depth\n    if max_depth is not None:\n        mask = (ground_truth_depth_original > 0) & (\n            ground_truth_depth_original < max_depth\n        )\n    else:\n        mask = ground_truth_depth_original > 0\n    predicted_depth = predicted_depth_original[mask]\n    ground_truth_depth = ground_truth_depth_original[mask]\n\n    # Clip the depth values\n    if pre_clip_min is not None:\n        predicted_depth = torch.clamp(predicted_depth, min=pre_clip_min)\n    if pre_clip_max is not None:\n        predicted_depth = torch.clamp(predicted_depth, max=pre_clip_max)\n\n    if disp_input:  # align the pred to gt in the disparity space\n        real_gt = ground_truth_depth.clone()\n        ground_truth_depth = 1 / (ground_truth_depth + 1e-8)\n\n    # various alignment methods\n    if metric_scale:\n        predicted_depth = predicted_depth\n    elif align_with_lstsq:\n        # Convert to numpy for lstsq\n        predicted_depth_np = predicted_depth.cpu().numpy().reshape(-1, 1)\n        ground_truth_depth_np = ground_truth_depth.cpu().numpy().reshape(-1, 1)\n\n        # Add a column of ones for the shift term\n        A = np.hstack([predicted_depth_np, np.ones_like(predicted_depth_np)])\n\n        # Solve for scale (s) and shift (t) using least squares\n        result = np.linalg.lstsq(A, ground_truth_depth_np, rcond=None)\n        s, t = result[0][0], result[0][1]\n\n        # convert to torch tensor\n        s = torch.tensor(s, device=predicted_depth_original.device)\n        t = torch.tensor(t, device=predicted_depth_original.device)\n\n        # Apply scale and shift\n        predicted_depth = s * predicted_depth + t\n    elif align_with_lad:\n        s, t = absolute_value_scaling(\n            predicted_depth,\n            ground_truth_depth,\n            s=torch.median(ground_truth_depth) / torch.median(predicted_depth),\n        )\n        predicted_depth = s * predicted_depth + t\n    elif align_with_lad2:\n        s_init = (\n            torch.median(ground_truth_depth) / torch.median(predicted_depth)\n        ).item()\n        s, t = absolute_value_scaling2(\n            predicted_depth,\n            ground_truth_depth,\n            s_init=s_init,\n            lr=lr,\n            max_iters=max_iters,\n        )\n        predicted_depth = s * predicted_depth + t\n    elif align_with_scale:\n        # Compute initial scale factor 's' using the closed-form solution (L2 norm)\n        dot_pred_gt = torch.nanmean(ground_truth_depth)\n        dot_pred_pred = torch.nanmean(predicted_depth)\n        s = dot_pred_gt / dot_pred_pred\n\n        # Iterative reweighted least squares using the Weiszfeld method\n        for _ in range(10):\n            # Compute residuals between scaled predictions and ground truth\n            residuals = s * predicted_depth - ground_truth_depth\n            abs_residuals = (\n                residuals.abs() + 1e-8\n            )  # Add small constant to avoid division by zero\n\n            # Compute weights inversely proportional to the residuals\n            weights = 1.0 / abs_residuals\n\n            # Update 's' using weighted sums\n            weighted_dot_pred_gt = torch.sum(\n                weights * predicted_depth * ground_truth_depth\n            )\n            weighted_dot_pred_pred = torch.sum(weights * predicted_depth**2)\n            s = weighted_dot_pred_gt / weighted_dot_pred_pred\n\n        # Optionally clip 's' to prevent extreme scaling\n        s = s.clamp(min=1e-3)\n\n        # Detach 's' if you want to stop gradients from flowing through it\n        s = s.detach()\n\n        # Apply the scale factor to the predicted depth\n        predicted_depth = s * predicted_depth\n\n    else:\n        # Align the predicted depth with the ground truth using median scaling\n        scale_factor = torch.median(ground_truth_depth) / torch.median(predicted_depth)\n        predicted_depth *= scale_factor\n\n    if disp_input:\n        # convert back to depth\n        ground_truth_depth = real_gt\n        predicted_depth = depth2disparity(predicted_depth)\n\n    # Clip the predicted depth values\n    if post_clip_min is not None:\n        predicted_depth = torch.clamp(predicted_depth, min=post_clip_min)\n    if post_clip_max is not None:\n        predicted_depth = torch.clamp(predicted_depth, max=post_clip_max)\n\n    if custom_mask is not None:\n        assert custom_mask.shape == ground_truth_depth_original.shape\n        mask_within_mask = custom_mask.cpu()[mask]\n        predicted_depth = predicted_depth[mask_within_mask]\n        ground_truth_depth = ground_truth_depth[mask_within_mask]\n\n    # Calculate the metrics\n    abs_rel = torch.mean(\n        torch.abs(predicted_depth - ground_truth_depth) / ground_truth_depth\n    ).item()\n    sq_rel = torch.mean(\n        ((predicted_depth - ground_truth_depth) ** 2) / ground_truth_depth\n    ).item()\n\n    # Correct RMSE calculation\n    rmse = torch.sqrt(torch.mean((predicted_depth - ground_truth_depth) ** 2)).item()\n\n    # Clip the depth values to avoid log(0)\n    predicted_depth = torch.clamp(predicted_depth, min=1e-5)\n    log_rmse = torch.sqrt(\n        torch.mean((torch.log(predicted_depth) - torch.log(ground_truth_depth)) ** 2)\n    ).item()\n\n    # Calculate the accuracy thresholds\n    max_ratio = torch.maximum(\n        predicted_depth / ground_truth_depth, ground_truth_depth / predicted_depth\n    )\n    threshold_0 = torch.mean((max_ratio < 1.0).float()).item()\n    threshold_1 = torch.mean((max_ratio < 1.25).float()).item()\n    threshold_2 = torch.mean((max_ratio < 1.25**2).float()).item()\n    threshold_3 = torch.mean((max_ratio < 1.25**3).float()).item()\n\n    # Compute the depth error parity map\n    if metric_scale:\n        predicted_depth_original = predicted_depth_original\n        if disp_input:\n            predicted_depth_original = depth2disparity(predicted_depth_original)\n        depth_error_parity_map = (\n            torch.abs(predicted_depth_original - ground_truth_depth_original)\n            / ground_truth_depth_original\n        )\n    elif align_with_lstsq or align_with_lad or align_with_lad2:\n        predicted_depth_original = predicted_depth_original * s + t\n        if disp_input:\n            predicted_depth_original = depth2disparity(predicted_depth_original)\n        depth_error_parity_map = (\n            torch.abs(predicted_depth_original - ground_truth_depth_original)\n            / ground_truth_depth_original\n        )\n    elif align_with_scale:\n        predicted_depth_original = predicted_depth_original * s\n        if disp_input:\n            predicted_depth_original = depth2disparity(predicted_depth_original)\n        depth_error_parity_map = (\n            torch.abs(predicted_depth_original - ground_truth_depth_original)\n            / ground_truth_depth_original\n        )\n    else:\n        predicted_depth_original = predicted_depth_original * scale_factor\n        if disp_input:\n            predicted_depth_original = depth2disparity(predicted_depth_original)\n        depth_error_parity_map = (\n            torch.abs(predicted_depth_original - ground_truth_depth_original)\n            / ground_truth_depth_original\n        )\n\n    # Reshape the depth_error_parity_map back to the original image size\n    depth_error_parity_map_full = torch.zeros_like(ground_truth_depth_original)\n    depth_error_parity_map_full = torch.where(\n        mask, depth_error_parity_map, depth_error_parity_map_full\n    )\n\n    predict_depth_map_full = predicted_depth_original\n    gt_depth_map_full = torch.zeros_like(ground_truth_depth_original)\n    gt_depth_map_full = torch.where(\n        mask, ground_truth_depth_original, gt_depth_map_full\n    )\n\n    num_valid_pixels = (\n        torch.sum(mask).item()\n        if custom_mask is None\n        else torch.sum(mask_within_mask).item()\n    )\n    if num_valid_pixels == 0:\n        (\n            abs_rel,\n            sq_rel,\n            rmse,\n            log_rmse,\n            threshold_0,\n            threshold_1,\n            threshold_2,\n            threshold_3,\n        ) = (0, 0, 0, 0, 0, 0, 0, 0)\n\n    results = {\n        \"Abs Rel\": abs_rel,\n        \"Sq Rel\": sq_rel,\n        \"RMSE\": rmse,\n        \"Log RMSE\": log_rmse,\n        \"δ < 1.\": threshold_0,\n        \"δ < 1.25\": threshold_1,\n        \"δ < 1.25^2\": threshold_2,\n        \"δ < 1.25^3\": threshold_3,\n        \"valid_pixels\": num_valid_pixels,\n    }\n\n    return (\n        results,\n        depth_error_parity_map_full,\n        predict_depth_map_full,\n        gt_depth_map_full,\n    )","source_hash":"079171f863616133446644309f22940572375c7a1ce61131084e01dbd38b0863","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.tools.group_by_directory","uri":"program://Human3R/function/eval.video_depth.tools.group_by_directory#L14-L31","kind":"function","name":"group_by_directory","path":"eval/video_depth/tools.py","language":"python","start_line":14,"end_line":31,"context_start_line":1,"context_end_line":51,"code":"import torch\nimport numpy as np\nimport cv2\nimport glob\nimport argparse\nfrom pathlib import Path\nfrom tqdm import tqdm\nfrom copy import deepcopy\nfrom scipy.optimize import minimize\nimport os\nfrom collections import defaultdict\n\n\ndef group_by_directory(pathes, idx=-1):\n    \"\"\"\n    Groups the file paths based on the second-to-last directory in their paths.\n\n    Parameters:\n    - pathes (list): List of file paths.\n\n    Returns:\n    - dict: A dictionary where keys are the second-to-last directory names and values are lists of file paths.\n    \"\"\"\n    grouped_pathes = defaultdict(list)\n\n    for path in pathes:\n        # Extract the second-to-last directory\n        dir_name = os.path.dirname(path).split(\"/\")[idx]\n        grouped_pathes[dir_name].append(path)\n\n    return grouped_pathes\n\n\ndef depth2disparity(depth, return_mask=False):\n    if isinstance(depth, torch.Tensor):\n        disparity = torch.zeros_like(depth)\n    elif isinstance(depth, np.ndarray):\n        disparity = np.zeros_like(depth)\n    non_negtive_mask = depth > 0\n    disparity[non_negtive_mask] = 1.0 / depth[non_negtive_mask]\n    if return_mask:\n        return disparity, non_negtive_mask\n    else:\n        return disparity\n\n\ndef absolute_error_loss(params, predicted_depth, ground_truth_depth):\n    s, t = params\n\n    predicted_aligned = s * predicted_depth + t\n","source_hash":"079171f863616133446644309f22940572375c7a1ce61131084e01dbd38b0863","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.tools.depth2disparity","uri":"program://Human3R/function/eval.video_depth.tools.depth2disparity#L34-L44","kind":"function","name":"depth2disparity","path":"eval/video_depth/tools.py","language":"python","start_line":34,"end_line":44,"context_start_line":14,"context_end_line":64,"code":"def group_by_directory(pathes, idx=-1):\n    \"\"\"\n    Groups the file paths based on the second-to-last directory in their paths.\n\n    Parameters:\n    - pathes (list): List of file paths.\n\n    Returns:\n    - dict: A dictionary where keys are the second-to-last directory names and values are lists of file paths.\n    \"\"\"\n    grouped_pathes = defaultdict(list)\n\n    for path in pathes:\n        # Extract the second-to-last directory\n        dir_name = os.path.dirname(path).split(\"/\")[idx]\n        grouped_pathes[dir_name].append(path)\n\n    return grouped_pathes\n\n\ndef depth2disparity(depth, return_mask=False):\n    if isinstance(depth, torch.Tensor):\n        disparity = torch.zeros_like(depth)\n    elif isinstance(depth, np.ndarray):\n        disparity = np.zeros_like(depth)\n    non_negtive_mask = depth > 0\n    disparity[non_negtive_mask] = 1.0 / depth[non_negtive_mask]\n    if return_mask:\n        return disparity, non_negtive_mask\n    else:\n        return disparity\n\n\ndef absolute_error_loss(params, predicted_depth, ground_truth_depth):\n    s, t = params\n\n    predicted_aligned = s * predicted_depth + t\n\n    abs_error = np.abs(predicted_aligned - ground_truth_depth)\n    return np.sum(abs_error)\n\n\ndef absolute_value_scaling(predicted_depth, ground_truth_depth, s=1, t=0):\n    predicted_depth_np = predicted_depth.cpu().numpy().reshape(-1)\n    ground_truth_depth_np = ground_truth_depth.cpu().numpy().reshape(-1)\n\n    initial_params = [s, t]  # s = 1, t = 0\n\n    result = minimize(\n        absolute_error_loss,\n        initial_params,","source_hash":"079171f863616133446644309f22940572375c7a1ce61131084e01dbd38b0863","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.tools.absolute_error_loss","uri":"program://Human3R/function/eval.video_depth.tools.absolute_error_loss#L47-L53","kind":"function","name":"absolute_error_loss","path":"eval/video_depth/tools.py","language":"python","start_line":47,"end_line":53,"context_start_line":27,"context_end_line":73,"code":"        # Extract the second-to-last directory\n        dir_name = os.path.dirname(path).split(\"/\")[idx]\n        grouped_pathes[dir_name].append(path)\n\n    return grouped_pathes\n\n\ndef depth2disparity(depth, return_mask=False):\n    if isinstance(depth, torch.Tensor):\n        disparity = torch.zeros_like(depth)\n    elif isinstance(depth, np.ndarray):\n        disparity = np.zeros_like(depth)\n    non_negtive_mask = depth > 0\n    disparity[non_negtive_mask] = 1.0 / depth[non_negtive_mask]\n    if return_mask:\n        return disparity, non_negtive_mask\n    else:\n        return disparity\n\n\ndef absolute_error_loss(params, predicted_depth, ground_truth_depth):\n    s, t = params\n\n    predicted_aligned = s * predicted_depth + t\n\n    abs_error = np.abs(predicted_aligned - ground_truth_depth)\n    return np.sum(abs_error)\n\n\ndef absolute_value_scaling(predicted_depth, ground_truth_depth, s=1, t=0):\n    predicted_depth_np = predicted_depth.cpu().numpy().reshape(-1)\n    ground_truth_depth_np = ground_truth_depth.cpu().numpy().reshape(-1)\n\n    initial_params = [s, t]  # s = 1, t = 0\n\n    result = minimize(\n        absolute_error_loss,\n        initial_params,\n        args=(predicted_depth_np, ground_truth_depth_np),\n    )\n\n    s, t = result.x\n    return s, t\n\n\ndef absolute_value_scaling2(\n    predicted_depth,","source_hash":"079171f863616133446644309f22940572375c7a1ce61131084e01dbd38b0863","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.tools.absolute_value_scaling","uri":"program://Human3R/function/eval.video_depth.tools.absolute_value_scaling#L56-L69","kind":"function","name":"absolute_value_scaling","path":"eval/video_depth/tools.py","language":"python","start_line":56,"end_line":69,"context_start_line":36,"context_end_line":89,"code":"        disparity = torch.zeros_like(depth)\n    elif isinstance(depth, np.ndarray):\n        disparity = np.zeros_like(depth)\n    non_negtive_mask = depth > 0\n    disparity[non_negtive_mask] = 1.0 / depth[non_negtive_mask]\n    if return_mask:\n        return disparity, non_negtive_mask\n    else:\n        return disparity\n\n\ndef absolute_error_loss(params, predicted_depth, ground_truth_depth):\n    s, t = params\n\n    predicted_aligned = s * predicted_depth + t\n\n    abs_error = np.abs(predicted_aligned - ground_truth_depth)\n    return np.sum(abs_error)\n\n\ndef absolute_value_scaling(predicted_depth, ground_truth_depth, s=1, t=0):\n    predicted_depth_np = predicted_depth.cpu().numpy().reshape(-1)\n    ground_truth_depth_np = ground_truth_depth.cpu().numpy().reshape(-1)\n\n    initial_params = [s, t]  # s = 1, t = 0\n\n    result = minimize(\n        absolute_error_loss,\n        initial_params,\n        args=(predicted_depth_np, ground_truth_depth_np),\n    )\n\n    s, t = result.x\n    return s, t\n\n\ndef absolute_value_scaling2(\n    predicted_depth,\n    ground_truth_depth,\n    s_init=1.0,\n    t_init=0.0,\n    lr=1e-4,\n    max_iters=1000,\n    tol=1e-6,\n):\n    # Initialize s and t as torch tensors with requires_grad=True\n    s = torch.tensor(\n        [s_init],\n        requires_grad=True,\n        device=predicted_depth.device,\n        dtype=predicted_depth.dtype,\n    )\n    t = torch.tensor(\n        [t_init],","source_hash":"079171f863616133446644309f22940572375c7a1ce61131084e01dbd38b0863","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.tools.absolute_value_scaling2","uri":"program://Human3R/function/eval.video_depth.tools.absolute_value_scaling2#L72-L123","kind":"function","name":"absolute_value_scaling2","path":"eval/video_depth/tools.py","language":"python","start_line":72,"end_line":123,"context_start_line":52,"context_end_line":143,"code":"    abs_error = np.abs(predicted_aligned - ground_truth_depth)\n    return np.sum(abs_error)\n\n\ndef absolute_value_scaling(predicted_depth, ground_truth_depth, s=1, t=0):\n    predicted_depth_np = predicted_depth.cpu().numpy().reshape(-1)\n    ground_truth_depth_np = ground_truth_depth.cpu().numpy().reshape(-1)\n\n    initial_params = [s, t]  # s = 1, t = 0\n\n    result = minimize(\n        absolute_error_loss,\n        initial_params,\n        args=(predicted_depth_np, ground_truth_depth_np),\n    )\n\n    s, t = result.x\n    return s, t\n\n\ndef absolute_value_scaling2(\n    predicted_depth,\n    ground_truth_depth,\n    s_init=1.0,\n    t_init=0.0,\n    lr=1e-4,\n    max_iters=1000,\n    tol=1e-6,\n):\n    # Initialize s and t as torch tensors with requires_grad=True\n    s = torch.tensor(\n        [s_init],\n        requires_grad=True,\n        device=predicted_depth.device,\n        dtype=predicted_depth.dtype,\n    )\n    t = torch.tensor(\n        [t_init],\n        requires_grad=True,\n        device=predicted_depth.device,\n        dtype=predicted_depth.dtype,\n    )\n\n    optimizer = torch.optim.Adam([s, t], lr=lr)\n\n    prev_loss = None\n\n    for i in range(max_iters):\n        optimizer.zero_grad()\n\n        # Compute predicted aligned depth\n        predicted_aligned = s * predicted_depth + t\n\n        # Compute absolute error\n        abs_error = torch.abs(predicted_aligned - ground_truth_depth)\n\n        # Compute loss\n        loss = torch.sum(abs_error)\n\n        # Backpropagate\n        loss.backward()\n\n        # Update parameters\n        optimizer.step()\n\n        # Check convergence\n        if prev_loss is not None and torch.abs(prev_loss - loss) < tol:\n            break\n\n        prev_loss = loss.item()\n\n    return s.detach().item(), t.detach().item()\n\n\ndef depth_evaluation(\n    predicted_depth_original,\n    ground_truth_depth_original,\n    max_depth=80,\n    custom_mask=None,\n    post_clip_min=None,\n    post_clip_max=None,\n    pre_clip_min=None,\n    pre_clip_max=None,\n    align_with_lstsq=False,\n    align_with_lad=False,\n    align_with_lad2=False,\n    metric_scale=False,\n    lr=1e-4,\n    max_iters=1000,\n    use_gpu=False,\n    align_with_scale=False,\n    disp_input=False,","source_hash":"079171f863616133446644309f22940572375c7a1ce61131084e01dbd38b0863","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.tools.depth_evaluation","uri":"program://Human3R/function/eval.video_depth.tools.depth_evaluation#L126-L399","kind":"function","name":"depth_evaluation","path":"eval/video_depth/tools.py","language":"python","start_line":126,"end_line":399,"context_start_line":106,"context_end_line":399,"code":"        abs_error = torch.abs(predicted_aligned - ground_truth_depth)\n\n        # Compute loss\n        loss = torch.sum(abs_error)\n\n        # Backpropagate\n        loss.backward()\n\n        # Update parameters\n        optimizer.step()\n\n        # Check convergence\n        if prev_loss is not None and torch.abs(prev_loss - loss) < tol:\n            break\n\n        prev_loss = loss.item()\n\n    return s.detach().item(), t.detach().item()\n\n\ndef depth_evaluation(\n    predicted_depth_original,\n    ground_truth_depth_original,\n    max_depth=80,\n    custom_mask=None,\n    post_clip_min=None,\n    post_clip_max=None,\n    pre_clip_min=None,\n    pre_clip_max=None,\n    align_with_lstsq=False,\n    align_with_lad=False,\n    align_with_lad2=False,\n    metric_scale=False,\n    lr=1e-4,\n    max_iters=1000,\n    use_gpu=False,\n    align_with_scale=False,\n    disp_input=False,\n):\n    \"\"\"\n    Evaluate the depth map using various metrics and return a depth error parity map, with an option for least squares alignment.\n\n    Args:\n        predicted_depth (numpy.ndarray or torch.Tensor): The predicted depth map.\n        ground_truth_depth (numpy.ndarray or torch.Tensor): The ground truth depth map.\n        max_depth (float): The maximum depth value to consider. Default is 80 meters.\n        align_with_lstsq (bool): If True, perform least squares alignment of the predicted depth with ground truth.\n\n    Returns:\n        dict: A dictionary containing the evaluation metrics.\n        torch.Tensor: The depth error parity map.\n    \"\"\"\n    if isinstance(predicted_depth_original, np.ndarray):\n        predicted_depth_original = torch.from_numpy(predicted_depth_original)\n    if isinstance(ground_truth_depth_original, np.ndarray):\n        ground_truth_depth_original = torch.from_numpy(ground_truth_depth_original)\n    if custom_mask is not None and isinstance(custom_mask, np.ndarray):\n        custom_mask = torch.from_numpy(custom_mask)\n\n    # if the dimension is 3, flatten to 2d along the batch dimension\n    if predicted_depth_original.dim() == 3:\n        _, h, w = predicted_depth_original.shape\n        predicted_depth_original = predicted_depth_original.view(-1, w)\n        ground_truth_depth_original = ground_truth_depth_original.view(-1, w)\n        if custom_mask is not None:\n            custom_mask = custom_mask.view(-1, w)\n\n    # put to device\n    if use_gpu:\n        predicted_depth_original = predicted_depth_original.cuda()\n        ground_truth_depth_original = ground_truth_depth_original.cuda()\n\n    # Filter out depths greater than max_depth\n    if max_depth is not None:\n        mask = (ground_truth_depth_original > 0) & (\n            ground_truth_depth_original < max_depth\n        )\n    else:\n        mask = ground_truth_depth_original > 0\n    predicted_depth = predicted_depth_original[mask]\n    ground_truth_depth = ground_truth_depth_original[mask]\n\n    # Clip the depth values\n    if pre_clip_min is not None:\n        predicted_depth = torch.clamp(predicted_depth, min=pre_clip_min)\n    if pre_clip_max is not None:\n        predicted_depth = torch.clamp(predicted_depth, max=pre_clip_max)\n\n    if disp_input:  # align the pred to gt in the disparity space\n        real_gt = ground_truth_depth.clone()\n        ground_truth_depth = 1 / (ground_truth_depth + 1e-8)\n\n    # various alignment methods\n    if metric_scale:\n        predicted_depth = predicted_depth\n    elif align_with_lstsq:\n        # Convert to numpy for lstsq\n        predicted_depth_np = predicted_depth.cpu().numpy().reshape(-1, 1)\n        ground_truth_depth_np = ground_truth_depth.cpu().numpy().reshape(-1, 1)\n\n        # Add a column of ones for the shift term\n        A = np.hstack([predicted_depth_np, np.ones_like(predicted_depth_np)])\n\n        # Solve for scale (s) and shift (t) using least squares\n        result = np.linalg.lstsq(A, ground_truth_depth_np, rcond=None)\n        s, t = result[0][0], result[0][1]\n\n        # convert to torch tensor\n        s = torch.tensor(s, device=predicted_depth_original.device)\n        t = torch.tensor(t, device=predicted_depth_original.device)\n\n        # Apply scale and shift\n        predicted_depth = s * predicted_depth + t\n    elif align_with_lad:\n        s, t = absolute_value_scaling(\n            predicted_depth,\n            ground_truth_depth,\n            s=torch.median(ground_truth_depth) / torch.median(predicted_depth),\n        )\n        predicted_depth = s * predicted_depth + t\n    elif align_with_lad2:\n        s_init = (\n            torch.median(ground_truth_depth) / torch.median(predicted_depth)\n        ).item()\n        s, t = absolute_value_scaling2(\n            predicted_depth,\n            ground_truth_depth,\n            s_init=s_init,\n            lr=lr,\n            max_iters=max_iters,\n        )\n        predicted_depth = s * predicted_depth + t\n    elif align_with_scale:\n        # Compute initial scale factor 's' using the closed-form solution (L2 norm)\n        dot_pred_gt = torch.nanmean(ground_truth_depth)\n        dot_pred_pred = torch.nanmean(predicted_depth)\n        s = dot_pred_gt / dot_pred_pred\n\n        # Iterative reweighted least squares using the Weiszfeld method\n        for _ in range(10):\n            # Compute residuals between scaled predictions and ground truth\n            residuals = s * predicted_depth - ground_truth_depth\n            abs_residuals = (\n                residuals.abs() + 1e-8\n            )  # Add small constant to avoid division by zero\n\n            # Compute weights inversely proportional to the residuals\n            weights = 1.0 / abs_residuals\n\n            # Update 's' using weighted sums\n            weighted_dot_pred_gt = torch.sum(\n                weights * predicted_depth * ground_truth_depth\n            )\n            weighted_dot_pred_pred = torch.sum(weights * predicted_depth**2)\n            s = weighted_dot_pred_gt / weighted_dot_pred_pred\n\n        # Optionally clip 's' to prevent extreme scaling\n        s = s.clamp(min=1e-3)\n\n        # Detach 's' if you want to stop gradients from flowing through it\n        s = s.detach()\n\n        # Apply the scale factor to the predicted depth\n        predicted_depth = s * predicted_depth\n\n    else:\n        # Align the predicted depth with the ground truth using median scaling\n        scale_factor = torch.median(ground_truth_depth) / torch.median(predicted_depth)\n        predicted_depth *= scale_factor\n\n    if disp_input:\n        # convert back to depth\n        ground_truth_depth = real_gt\n        predicted_depth = depth2disparity(predicted_depth)\n\n    # Clip the predicted depth values\n    if post_clip_min is not None:\n        predicted_depth = torch.clamp(predicted_depth, min=post_clip_min)\n    if post_clip_max is not None:\n        predicted_depth = torch.clamp(predicted_depth, max=post_clip_max)\n\n    if custom_mask is not None:\n        assert custom_mask.shape == ground_truth_depth_original.shape\n        mask_within_mask = custom_mask.cpu()[mask]\n        predicted_depth = predicted_depth[mask_within_mask]\n        ground_truth_depth = ground_truth_depth[mask_within_mask]\n\n    # Calculate the metrics\n    abs_rel = torch.mean(\n        torch.abs(predicted_depth - ground_truth_depth) / ground_truth_depth\n    ).item()\n    sq_rel = torch.mean(\n        ((predicted_depth - ground_truth_depth) ** 2) / ground_truth_depth\n    ).item()\n\n    # Correct RMSE calculation\n    rmse = torch.sqrt(torch.mean((predicted_depth - ground_truth_depth) ** 2)).item()\n\n    # Clip the depth values to avoid log(0)\n    predicted_depth = torch.clamp(predicted_depth, min=1e-5)\n    log_rmse = torch.sqrt(\n        torch.mean((torch.log(predicted_depth) - torch.log(ground_truth_depth)) ** 2)\n    ).item()\n\n    # Calculate the accuracy thresholds\n    max_ratio = torch.maximum(\n        predicted_depth / ground_truth_depth, ground_truth_depth / predicted_depth\n    )\n    threshold_0 = torch.mean((max_ratio < 1.0).float()).item()\n    threshold_1 = torch.mean((max_ratio < 1.25).float()).item()\n    threshold_2 = torch.mean((max_ratio < 1.25**2).float()).item()\n    threshold_3 = torch.mean((max_ratio < 1.25**3).float()).item()\n\n    # Compute the depth error parity map\n    if metric_scale:\n        predicted_depth_original = predicted_depth_original\n        if disp_input:\n            predicted_depth_original = depth2disparity(predicted_depth_original)\n        depth_error_parity_map = (\n            torch.abs(predicted_depth_original - ground_truth_depth_original)\n            / ground_truth_depth_original\n        )\n    elif align_with_lstsq or align_with_lad or align_with_lad2:\n        predicted_depth_original = predicted_depth_original * s + t\n        if disp_input:\n            predicted_depth_original = depth2disparity(predicted_depth_original)\n        depth_error_parity_map = (\n            torch.abs(predicted_depth_original - ground_truth_depth_original)\n            / ground_truth_depth_original\n        )\n    elif align_with_scale:\n        predicted_depth_original = predicted_depth_original * s\n        if disp_input:\n            predicted_depth_original = depth2disparity(predicted_depth_original)\n        depth_error_parity_map = (\n            torch.abs(predicted_depth_original - ground_truth_depth_original)\n            / ground_truth_depth_original\n        )\n    else:\n        predicted_depth_original = predicted_depth_original * scale_factor\n        if disp_input:\n            predicted_depth_original = depth2disparity(predicted_depth_original)\n        depth_error_parity_map = (\n            torch.abs(predicted_depth_original - ground_truth_depth_original)\n            / ground_truth_depth_original\n        )\n\n    # Reshape the depth_error_parity_map back to the original image size\n    depth_error_parity_map_full = torch.zeros_like(ground_truth_depth_original)\n    depth_error_parity_map_full = torch.where(\n        mask, depth_error_parity_map, depth_error_parity_map_full\n    )\n\n    predict_depth_map_full = predicted_depth_original\n    gt_depth_map_full = torch.zeros_like(ground_truth_depth_original)\n    gt_depth_map_full = torch.where(\n        mask, ground_truth_depth_original, gt_depth_map_full\n    )\n\n    num_valid_pixels = (\n        torch.sum(mask).item()\n        if custom_mask is None\n        else torch.sum(mask_within_mask).item()\n    )\n    if num_valid_pixels == 0:\n        (\n            abs_rel,\n            sq_rel,\n            rmse,\n            log_rmse,\n            threshold_0,\n            threshold_1,\n            threshold_2,\n            threshold_3,\n        ) = (0, 0, 0, 0, 0, 0, 0, 0)\n\n    results = {\n        \"Abs Rel\": abs_rel,\n        \"Sq Rel\": sq_rel,\n        \"RMSE\": rmse,\n        \"Log RMSE\": log_rmse,\n        \"δ < 1.\": threshold_0,\n        \"δ < 1.25\": threshold_1,\n        \"δ < 1.25^2\": threshold_2,\n        \"δ < 1.25^3\": threshold_3,\n        \"valid_pixels\": num_valid_pixels,\n    }\n\n    return (\n        results,\n        depth_error_parity_map_full,\n        predict_depth_map_full,\n        gt_depth_map_full,\n    )","source_hash":"079171f863616133446644309f22940572375c7a1ce61131084e01dbd38b0863","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.eval_depth","uri":"program://Human3R/module/eval.video_depth.eval_depth#L1-L390","kind":"module","name":"eval.video_depth.eval_depth","path":"eval/video_depth/eval_depth.py","language":"python","start_line":1,"end_line":390,"context_start_line":1,"context_end_line":390,"code":"import os\nimport sys\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nfrom eval.video_depth.tools import depth_evaluation, group_by_directory\nimport numpy as np\nimport cv2\nfrom tqdm import tqdm\nimport glob\nfrom PIL import Image\nimport argparse\nimport json\nfrom eval.video_depth.metadata import dataset_metadata\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"\",\n        help=\"value for outdir\",\n    )\n    parser.add_argument(\n        \"--eval_dataset\", type=str, default=\"nyu\", choices=list(dataset_metadata.keys())\n    )\n    parser.add_argument(\n        \"--align\",\n        type=str,\n        default=\"scale&shift\",\n        choices=[\"scale&shift\", \"scale\", \"metric\"],\n    )\n    return parser\n\n\ndef main(args):\n    if args.eval_dataset == \"sintel\":\n        TAG_FLOAT = 202021.25\n\n        def depth_read(filename):\n            \"\"\"Read depth data from file, return as numpy array.\"\"\"\n            f = open(filename, \"rb\")\n            check = np.fromfile(f, dtype=np.float32, count=1)[0]\n            assert (\n                check == TAG_FLOAT\n            ), \" depth_read:: Wrong tag in flow file (should be: {0}, is: {1}). Big-endian machine? \".format(\n                TAG_FLOAT, check\n            )\n            width = np.fromfile(f, dtype=np.int32, count=1)[0]\n            height = np.fromfile(f, dtype=np.int32, count=1)[0]\n            size = width * height\n            assert (\n                width > 0 and height > 0 and size > 1 and size < 100000000\n            ), \" depth_read:: Wrong input size (width = {0}, height = {1}).\".format(\n                width, height\n            )\n            depth = np.fromfile(f, dtype=np.float32, count=-1).reshape((height, width))\n            return depth\n\n        pred_pathes = glob.glob(\n            f\"{args.output_dir}/*/frame_*.npy\"\n        )  # TODO: update the path to your prediction\n        pred_pathes = sorted(pred_pathes)\n\n        if len(pred_pathes) > 643:\n            full = True\n        else:\n            full = False\n\n        if full:\n            depth_pathes = glob.glob(f\"/path/to/sintel/training/depth/*/*.dpt\")\n            depth_pathes = sorted(depth_pathes)\n        else:\n            seq_list = [\n                \"alley_2\",\n                \"ambush_4\",\n                \"ambush_5\",\n                \"ambush_6\",\n                \"cave_2\",\n                \"cave_4\",\n                \"market_2\",\n                \"market_5\",\n                \"market_6\",\n                \"shaman_3\",\n                \"sleeping_1\",\n                \"sleeping_2\",\n                \"temple_2\",\n                \"temple_3\",\n            ]\n            depth_pathes_folder = [\n                f\"/path/to/sintel/training/depth/{seq}\" for seq in seq_list\n            ]\n            depth_pathes = []\n            for depth_pathes_folder_i in depth_pathes_folder:\n                depth_pathes += glob.glob(depth_pathes_folder_i + \"/*.dpt\")\n            depth_pathes = sorted(depth_pathes)\n\n        def get_video_results():\n            grouped_pred_depth = group_by_directory(pred_pathes)\n\n            grouped_gt_depth = group_by_directory(depth_pathes)\n            gathered_depth_metrics = []\n\n            for key in tqdm(grouped_pred_depth.keys()):\n                pd_pathes = grouped_pred_depth[key]\n                gt_pathes = grouped_gt_depth[key.replace(\"_pred_depth\", \"\")]\n\n                gt_depth = np.stack(\n                    [depth_read(gt_path) for gt_path in gt_pathes], axis=0\n                )\n                pr_depth = np.stack(\n                    [\n                        cv2.resize(\n                            np.load(pd_path),\n                            (gt_depth.shape[2], gt_depth.shape[1]),\n                            interpolation=cv2.INTER_CUBIC,\n                        )\n                        for pd_path in pd_pathes\n                    ],\n                    axis=0,\n                )\n                # for depth eval, set align_with_lad2=False to use median alignment; set align_with_lad2=True to use scale&shift alignment\n                if args.align == \"scale&shift\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            align_with_lad2=True,\n                            use_gpu=True,\n                            post_clip_max=70,\n                        )\n                    )\n                elif args.align == \"scale\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            align_with_scale=True,\n                            use_gpu=True,\n                            post_clip_max=70,\n                        )\n                    )\n                elif args.align == \"metric\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            metric_scale=True,\n                            use_gpu=True,\n                            post_clip_max=70,\n                        )\n                    )\n                gathered_depth_metrics.append(depth_results)\n\n            depth_log_path = f\"{args.output_dir}/result_{args.align}.json\"\n            average_metrics = {\n                key: np.average(\n                    [metrics[key] for metrics in gathered_depth_metrics],\n                    weights=[\n                        metrics[\"valid_pixels\"] for metrics in gathered_depth_metrics\n                    ],\n                )\n                for key in gathered_depth_metrics[0].keys()\n                if key != \"valid_pixels\"\n            }\n            print(\"Average depth evaluation metrics:\", average_metrics)\n            with open(depth_log_path, \"w\") as f:\n                f.write(json.dumps(average_metrics))\n\n        get_video_results()\n    elif args.eval_dataset.startswith(\"bonn\"):\n\n        def depth_read(filename):\n            # loads depth map D from png file\n            # and returns it as a numpy array\n            depth_png = np.asarray(Image.open(filename))\n            # make sure we have a proper 16bit depth map here.. not 8bit!\n            assert np.max(depth_png) > 255\n            depth = depth_png.astype(np.float64) / 5000.0\n            depth[depth_png == 0] = -1.0\n            return depth\n\n        seq_list = [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n\n        if \"_\" in args.eval_dataset:\n            bonn_number = args.eval_dataset.split(\"_\")[-1]\n        else:\n            bonn_number = \"110\"  # default value\n\n        img_pathes_folder = [\n            f\"/path/to/long_bonn/rgbd_bonn_{seq}/rgb_{bonn_number}/*.png\"\n            for seq in seq_list\n        ]\n        img_pathes = []\n        for img_pathes_folder_i in img_pathes_folder:\n            img_pathes += glob.glob(img_pathes_folder_i)\n        img_pathes = sorted(img_pathes)\n        depth_pathes_folder = [\n            f\"/path/to/long_bonn/rgbd_bonn_{seq}/depth_{bonn_number}/*.png\"\n            for seq in seq_list\n        ]\n        depth_pathes = []\n        for depth_pathes_folder_i in depth_pathes_folder:\n            depth_pathes += glob.glob(depth_pathes_folder_i)\n        depth_pathes = sorted(depth_pathes)\n        pred_pathes = glob.glob(\n            f\"{args.output_dir}/*/frame*.npy\"\n        )  # TODO: update the path to your prediction\n        pred_pathes = sorted(pred_pathes)\n\n        def get_video_results():\n            grouped_pred_depth = group_by_directory(pred_pathes)\n            grouped_gt_depth = group_by_directory(depth_pathes, idx=-2)\n            gathered_depth_metrics = []\n            for key in tqdm(grouped_gt_depth.keys()):\n                pd_pathes = grouped_pred_depth[key[10:]]\n                gt_pathes = grouped_gt_depth[key]\n                gt_depth = np.stack(\n                    [depth_read(gt_path) for gt_path in gt_pathes], axis=0\n                )\n                pr_depth = np.stack(\n                    [\n                        cv2.resize(\n                            np.load(pd_path),\n                            (gt_depth.shape[2], gt_depth.shape[1]),\n                            interpolation=cv2.INTER_CUBIC,\n                        )\n                        for pd_path in pd_pathes\n                    ],\n                    axis=0,\n                )\n                # for depth eval, set align_with_lad2=False to use median alignment; set align_with_lad2=True to use scale&shift alignment\n                if args.align == \"scale&shift\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            align_with_lad2=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"scale\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            align_with_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"metric\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            metric_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                gathered_depth_metrics.append(depth_results)\n\n                # seq_len = gt_depth.shape[0]\n                # error_map = error_map.reshape(seq_len, -1, error_map.shape[-1]).cpu()\n                # error_map_colored = colorize(error_map, range=(error_map.min(), error_map.max()), append_cbar=True)\n                # ImageSequenceClip([x for x in (error_map_colored.numpy()*255).astype(np.uint8)], fps=10).write_videofile(f'{args.output_dir}/errormap_{key}_{args.align}.mp4', fps=10)\n\n            depth_log_path = f\"{args.output_dir}/result_{args.align}.json\"\n            average_metrics = {\n                key: np.average(\n                    [metrics[key] for metrics in gathered_depth_metrics],\n                    weights=[\n                        metrics[\"valid_pixels\"] for metrics in gathered_depth_metrics\n                    ],\n                )\n                for key in gathered_depth_metrics[0].keys()\n                if key != \"valid_pixels\"\n            }\n            print(\"Average depth evaluation metrics:\", average_metrics)\n            with open(depth_log_path, \"w\") as f:\n                f.write(json.dumps(average_metrics))\n\n        get_video_results()\n    elif args.eval_dataset == \"kitti\":\n\n        def depth_read(filename):\n            # loads depth map D from png file\n            # and returns it as a numpy array,\n            # for details see readme.txt\n            img_pil = Image.open(filename)\n            depth_png = np.array(img_pil, dtype=int)\n            # make sure we have a proper 16bit depth map here.. not 8bit!\n            assert np.max(depth_png) > 255\n\n            depth = depth_png.astype(float) / 256.0\n            depth[depth_png == 0] = -1.0\n            return depth\n\n        depth_pathes = glob.glob(\n            \"data/kitti/depth_selection/val_selection_cropped/groundtruth_depth_gathered/*/*.png\"\n        )\n        depth_pathes = sorted(depth_pathes)\n        pred_pathes = glob.glob(\n            f\"{args.output_dir}/*/frame_*.npy\"\n        )  # TODO: update the path to your prediction\n        pred_pathes = sorted(pred_pathes)\n\n        def get_video_results():\n            grouped_pred_depth = group_by_directory(pred_pathes)\n            grouped_gt_depth = group_by_directory(depth_pathes)\n            gathered_depth_metrics = []\n            for key in tqdm(grouped_pred_depth.keys()):\n                pd_pathes = grouped_pred_depth[key]\n                gt_pathes = grouped_gt_depth[key]\n                gt_depth = np.stack(\n                    [depth_read(gt_path) for gt_path in gt_pathes], axis=0\n                )\n                pr_depth = np.stack(\n                    [\n                        cv2.resize(\n                            np.load(pd_path),\n                            (gt_depth.shape[2], gt_depth.shape[1]),\n                            interpolation=cv2.INTER_CUBIC,\n                        )\n                        for pd_path in pd_pathes\n                    ],\n                    axis=0,\n                )\n\n                # for depth eval, set align_with_lad2=False to use median alignment; set align_with_lad2=True to use scale&shift alignment\n                if args.align == \"scale&shift\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=None,\n                            align_with_lad2=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"scale\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=None,\n                            align_with_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"metric\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=None,\n                            metric_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                gathered_depth_metrics.append(depth_results)\n\n            depth_log_path = f\"{args.output_dir}/result_{args.align}.json\"\n            average_metrics = {\n                key: np.average(\n                    [metrics[key] for metrics in gathered_depth_metrics],\n                    weights=[\n                        metrics[\"valid_pixels\"] for metrics in gathered_depth_metrics\n                    ],\n                )\n                for key in gathered_depth_metrics[0].keys()\n                if key != \"valid_pixels\"\n            }\n            print(\"Average depth evaluation metrics:\", average_metrics)\n            with open(depth_log_path, \"w\") as f:\n                f.write(json.dumps(average_metrics))\n\n        get_video_results()\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    main(args)","source_hash":"924683ea2cb1871eefa5d1266bb3aed3d887aaabf173d23b4a20b537325d0b34","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.eval_depth.get_args_parser","uri":"program://Human3R/function/eval.video_depth.eval_depth.get_args_parser#L16-L34","kind":"function","name":"get_args_parser","path":"eval/video_depth/eval_depth.py","language":"python","start_line":16,"end_line":34,"context_start_line":1,"context_end_line":54,"code":"import os\nimport sys\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nfrom eval.video_depth.tools import depth_evaluation, group_by_directory\nimport numpy as np\nimport cv2\nfrom tqdm import tqdm\nimport glob\nfrom PIL import Image\nimport argparse\nimport json\nfrom eval.video_depth.metadata import dataset_metadata\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"\",\n        help=\"value for outdir\",\n    )\n    parser.add_argument(\n        \"--eval_dataset\", type=str, default=\"nyu\", choices=list(dataset_metadata.keys())\n    )\n    parser.add_argument(\n        \"--align\",\n        type=str,\n        default=\"scale&shift\",\n        choices=[\"scale&shift\", \"scale\", \"metric\"],\n    )\n    return parser\n\n\ndef main(args):\n    if args.eval_dataset == \"sintel\":\n        TAG_FLOAT = 202021.25\n\n        def depth_read(filename):\n            \"\"\"Read depth data from file, return as numpy array.\"\"\"\n            f = open(filename, \"rb\")\n            check = np.fromfile(f, dtype=np.float32, count=1)[0]\n            assert (\n                check == TAG_FLOAT\n            ), \" depth_read:: Wrong tag in flow file (should be: {0}, is: {1}). Big-endian machine? \".format(\n                TAG_FLOAT, check\n            )\n            width = np.fromfile(f, dtype=np.int32, count=1)[0]\n            height = np.fromfile(f, dtype=np.int32, count=1)[0]\n            size = width * height\n            assert (\n                width > 0 and height > 0 and size > 1 and size < 100000000","source_hash":"924683ea2cb1871eefa5d1266bb3aed3d887aaabf173d23b4a20b537325d0b34","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.eval_depth.main","uri":"program://Human3R/function/eval.video_depth.eval_depth.main#L37-L384","kind":"function","name":"main","path":"eval/video_depth/eval_depth.py","language":"python","start_line":37,"end_line":384,"context_start_line":17,"context_end_line":390,"code":"    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"\",\n        help=\"value for outdir\",\n    )\n    parser.add_argument(\n        \"--eval_dataset\", type=str, default=\"nyu\", choices=list(dataset_metadata.keys())\n    )\n    parser.add_argument(\n        \"--align\",\n        type=str,\n        default=\"scale&shift\",\n        choices=[\"scale&shift\", \"scale\", \"metric\"],\n    )\n    return parser\n\n\ndef main(args):\n    if args.eval_dataset == \"sintel\":\n        TAG_FLOAT = 202021.25\n\n        def depth_read(filename):\n            \"\"\"Read depth data from file, return as numpy array.\"\"\"\n            f = open(filename, \"rb\")\n            check = np.fromfile(f, dtype=np.float32, count=1)[0]\n            assert (\n                check == TAG_FLOAT\n            ), \" depth_read:: Wrong tag in flow file (should be: {0}, is: {1}). Big-endian machine? \".format(\n                TAG_FLOAT, check\n            )\n            width = np.fromfile(f, dtype=np.int32, count=1)[0]\n            height = np.fromfile(f, dtype=np.int32, count=1)[0]\n            size = width * height\n            assert (\n                width > 0 and height > 0 and size > 1 and size < 100000000\n            ), \" depth_read:: Wrong input size (width = {0}, height = {1}).\".format(\n                width, height\n            )\n            depth = np.fromfile(f, dtype=np.float32, count=-1).reshape((height, width))\n            return depth\n\n        pred_pathes = glob.glob(\n            f\"{args.output_dir}/*/frame_*.npy\"\n        )  # TODO: update the path to your prediction\n        pred_pathes = sorted(pred_pathes)\n\n        if len(pred_pathes) > 643:\n            full = True\n        else:\n            full = False\n\n        if full:\n            depth_pathes = glob.glob(f\"/path/to/sintel/training/depth/*/*.dpt\")\n            depth_pathes = sorted(depth_pathes)\n        else:\n            seq_list = [\n                \"alley_2\",\n                \"ambush_4\",\n                \"ambush_5\",\n                \"ambush_6\",\n                \"cave_2\",\n                \"cave_4\",\n                \"market_2\",\n                \"market_5\",\n                \"market_6\",\n                \"shaman_3\",\n                \"sleeping_1\",\n                \"sleeping_2\",\n                \"temple_2\",\n                \"temple_3\",\n            ]\n            depth_pathes_folder = [\n                f\"/path/to/sintel/training/depth/{seq}\" for seq in seq_list\n            ]\n            depth_pathes = []\n            for depth_pathes_folder_i in depth_pathes_folder:\n                depth_pathes += glob.glob(depth_pathes_folder_i + \"/*.dpt\")\n            depth_pathes = sorted(depth_pathes)\n\n        def get_video_results():\n            grouped_pred_depth = group_by_directory(pred_pathes)\n\n            grouped_gt_depth = group_by_directory(depth_pathes)\n            gathered_depth_metrics = []\n\n            for key in tqdm(grouped_pred_depth.keys()):\n                pd_pathes = grouped_pred_depth[key]\n                gt_pathes = grouped_gt_depth[key.replace(\"_pred_depth\", \"\")]\n\n                gt_depth = np.stack(\n                    [depth_read(gt_path) for gt_path in gt_pathes], axis=0\n                )\n                pr_depth = np.stack(\n                    [\n                        cv2.resize(\n                            np.load(pd_path),\n                            (gt_depth.shape[2], gt_depth.shape[1]),\n                            interpolation=cv2.INTER_CUBIC,\n                        )\n                        for pd_path in pd_pathes\n                    ],\n                    axis=0,\n                )\n                # for depth eval, set align_with_lad2=False to use median alignment; set align_with_lad2=True to use scale&shift alignment\n                if args.align == \"scale&shift\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            align_with_lad2=True,\n                            use_gpu=True,\n                            post_clip_max=70,\n                        )\n                    )\n                elif args.align == \"scale\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            align_with_scale=True,\n                            use_gpu=True,\n                            post_clip_max=70,\n                        )\n                    )\n                elif args.align == \"metric\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            metric_scale=True,\n                            use_gpu=True,\n                            post_clip_max=70,\n                        )\n                    )\n                gathered_depth_metrics.append(depth_results)\n\n            depth_log_path = f\"{args.output_dir}/result_{args.align}.json\"\n            average_metrics = {\n                key: np.average(\n                    [metrics[key] for metrics in gathered_depth_metrics],\n                    weights=[\n                        metrics[\"valid_pixels\"] for metrics in gathered_depth_metrics\n                    ],\n                )\n                for key in gathered_depth_metrics[0].keys()\n                if key != \"valid_pixels\"\n            }\n            print(\"Average depth evaluation metrics:\", average_metrics)\n            with open(depth_log_path, \"w\") as f:\n                f.write(json.dumps(average_metrics))\n\n        get_video_results()\n    elif args.eval_dataset.startswith(\"bonn\"):\n\n        def depth_read(filename):\n            # loads depth map D from png file\n            # and returns it as a numpy array\n            depth_png = np.asarray(Image.open(filename))\n            # make sure we have a proper 16bit depth map here.. not 8bit!\n            assert np.max(depth_png) > 255\n            depth = depth_png.astype(np.float64) / 5000.0\n            depth[depth_png == 0] = -1.0\n            return depth\n\n        seq_list = [\"balloon2\", \"crowd2\", \"crowd3\", \"person_tracking2\", \"synchronous\"]\n\n        if \"_\" in args.eval_dataset:\n            bonn_number = args.eval_dataset.split(\"_\")[-1]\n        else:\n            bonn_number = \"110\"  # default value\n\n        img_pathes_folder = [\n            f\"/path/to/long_bonn/rgbd_bonn_{seq}/rgb_{bonn_number}/*.png\"\n            for seq in seq_list\n        ]\n        img_pathes = []\n        for img_pathes_folder_i in img_pathes_folder:\n            img_pathes += glob.glob(img_pathes_folder_i)\n        img_pathes = sorted(img_pathes)\n        depth_pathes_folder = [\n            f\"/path/to/long_bonn/rgbd_bonn_{seq}/depth_{bonn_number}/*.png\"\n            for seq in seq_list\n        ]\n        depth_pathes = []\n        for depth_pathes_folder_i in depth_pathes_folder:\n            depth_pathes += glob.glob(depth_pathes_folder_i)\n        depth_pathes = sorted(depth_pathes)\n        pred_pathes = glob.glob(\n            f\"{args.output_dir}/*/frame*.npy\"\n        )  # TODO: update the path to your prediction\n        pred_pathes = sorted(pred_pathes)\n\n        def get_video_results():\n            grouped_pred_depth = group_by_directory(pred_pathes)\n            grouped_gt_depth = group_by_directory(depth_pathes, idx=-2)\n            gathered_depth_metrics = []\n            for key in tqdm(grouped_gt_depth.keys()):\n                pd_pathes = grouped_pred_depth[key[10:]]\n                gt_pathes = grouped_gt_depth[key]\n                gt_depth = np.stack(\n                    [depth_read(gt_path) for gt_path in gt_pathes], axis=0\n                )\n                pr_depth = np.stack(\n                    [\n                        cv2.resize(\n                            np.load(pd_path),\n                            (gt_depth.shape[2], gt_depth.shape[1]),\n                            interpolation=cv2.INTER_CUBIC,\n                        )\n                        for pd_path in pd_pathes\n                    ],\n                    axis=0,\n                )\n                # for depth eval, set align_with_lad2=False to use median alignment; set align_with_lad2=True to use scale&shift alignment\n                if args.align == \"scale&shift\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            align_with_lad2=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"scale\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            align_with_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"metric\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=70,\n                            metric_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                gathered_depth_metrics.append(depth_results)\n\n                # seq_len = gt_depth.shape[0]\n                # error_map = error_map.reshape(seq_len, -1, error_map.shape[-1]).cpu()\n                # error_map_colored = colorize(error_map, range=(error_map.min(), error_map.max()), append_cbar=True)\n                # ImageSequenceClip([x for x in (error_map_colored.numpy()*255).astype(np.uint8)], fps=10).write_videofile(f'{args.output_dir}/errormap_{key}_{args.align}.mp4', fps=10)\n\n            depth_log_path = f\"{args.output_dir}/result_{args.align}.json\"\n            average_metrics = {\n                key: np.average(\n                    [metrics[key] for metrics in gathered_depth_metrics],\n                    weights=[\n                        metrics[\"valid_pixels\"] for metrics in gathered_depth_metrics\n                    ],\n                )\n                for key in gathered_depth_metrics[0].keys()\n                if key != \"valid_pixels\"\n            }\n            print(\"Average depth evaluation metrics:\", average_metrics)\n            with open(depth_log_path, \"w\") as f:\n                f.write(json.dumps(average_metrics))\n\n        get_video_results()\n    elif args.eval_dataset == \"kitti\":\n\n        def depth_read(filename):\n            # loads depth map D from png file\n            # and returns it as a numpy array,\n            # for details see readme.txt\n            img_pil = Image.open(filename)\n            depth_png = np.array(img_pil, dtype=int)\n            # make sure we have a proper 16bit depth map here.. not 8bit!\n            assert np.max(depth_png) > 255\n\n            depth = depth_png.astype(float) / 256.0\n            depth[depth_png == 0] = -1.0\n            return depth\n\n        depth_pathes = glob.glob(\n            \"data/kitti/depth_selection/val_selection_cropped/groundtruth_depth_gathered/*/*.png\"\n        )\n        depth_pathes = sorted(depth_pathes)\n        pred_pathes = glob.glob(\n            f\"{args.output_dir}/*/frame_*.npy\"\n        )  # TODO: update the path to your prediction\n        pred_pathes = sorted(pred_pathes)\n\n        def get_video_results():\n            grouped_pred_depth = group_by_directory(pred_pathes)\n            grouped_gt_depth = group_by_directory(depth_pathes)\n            gathered_depth_metrics = []\n            for key in tqdm(grouped_pred_depth.keys()):\n                pd_pathes = grouped_pred_depth[key]\n                gt_pathes = grouped_gt_depth[key]\n                gt_depth = np.stack(\n                    [depth_read(gt_path) for gt_path in gt_pathes], axis=0\n                )\n                pr_depth = np.stack(\n                    [\n                        cv2.resize(\n                            np.load(pd_path),\n                            (gt_depth.shape[2], gt_depth.shape[1]),\n                            interpolation=cv2.INTER_CUBIC,\n                        )\n                        for pd_path in pd_pathes\n                    ],\n                    axis=0,\n                )\n\n                # for depth eval, set align_with_lad2=False to use median alignment; set align_with_lad2=True to use scale&shift alignment\n                if args.align == \"scale&shift\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=None,\n                            align_with_lad2=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"scale\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=None,\n                            align_with_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"metric\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=None,\n                            metric_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                gathered_depth_metrics.append(depth_results)\n\n            depth_log_path = f\"{args.output_dir}/result_{args.align}.json\"\n            average_metrics = {\n                key: np.average(\n                    [metrics[key] for metrics in gathered_depth_metrics],\n                    weights=[\n                        metrics[\"valid_pixels\"] for metrics in gathered_depth_metrics\n                    ],\n                )\n                for key in gathered_depth_metrics[0].keys()\n                if key != \"valid_pixels\"\n            }\n            print(\"Average depth evaluation metrics:\", average_metrics)\n            with open(depth_log_path, \"w\") as f:\n                f.write(json.dumps(average_metrics))\n\n        get_video_results()\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    main(args)","source_hash":"924683ea2cb1871eefa5d1266bb3aed3d887aaabf173d23b4a20b537325d0b34","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.eval_depth.depth_read","uri":"program://Human3R/function/eval.video_depth.eval_depth.depth_read#L292-L303","kind":"function","name":"depth_read","path":"eval/video_depth/eval_depth.py","language":"python","start_line":292,"end_line":303,"context_start_line":272,"context_end_line":323,"code":"                # ImageSequenceClip([x for x in (error_map_colored.numpy()*255).astype(np.uint8)], fps=10).write_videofile(f'{args.output_dir}/errormap_{key}_{args.align}.mp4', fps=10)\n\n            depth_log_path = f\"{args.output_dir}/result_{args.align}.json\"\n            average_metrics = {\n                key: np.average(\n                    [metrics[key] for metrics in gathered_depth_metrics],\n                    weights=[\n                        metrics[\"valid_pixels\"] for metrics in gathered_depth_metrics\n                    ],\n                )\n                for key in gathered_depth_metrics[0].keys()\n                if key != \"valid_pixels\"\n            }\n            print(\"Average depth evaluation metrics:\", average_metrics)\n            with open(depth_log_path, \"w\") as f:\n                f.write(json.dumps(average_metrics))\n\n        get_video_results()\n    elif args.eval_dataset == \"kitti\":\n\n        def depth_read(filename):\n            # loads depth map D from png file\n            # and returns it as a numpy array,\n            # for details see readme.txt\n            img_pil = Image.open(filename)\n            depth_png = np.array(img_pil, dtype=int)\n            # make sure we have a proper 16bit depth map here.. not 8bit!\n            assert np.max(depth_png) > 255\n\n            depth = depth_png.astype(float) / 256.0\n            depth[depth_png == 0] = -1.0\n            return depth\n\n        depth_pathes = glob.glob(\n            \"data/kitti/depth_selection/val_selection_cropped/groundtruth_depth_gathered/*/*.png\"\n        )\n        depth_pathes = sorted(depth_pathes)\n        pred_pathes = glob.glob(\n            f\"{args.output_dir}/*/frame_*.npy\"\n        )  # TODO: update the path to your prediction\n        pred_pathes = sorted(pred_pathes)\n\n        def get_video_results():\n            grouped_pred_depth = group_by_directory(pred_pathes)\n            grouped_gt_depth = group_by_directory(depth_pathes)\n            gathered_depth_metrics = []\n            for key in tqdm(grouped_pred_depth.keys()):\n                pd_pathes = grouped_pred_depth[key]\n                gt_pathes = grouped_gt_depth[key]\n                gt_depth = np.stack(\n                    [depth_read(gt_path) for gt_path in gt_pathes], axis=0\n                )","source_hash":"924683ea2cb1871eefa5d1266bb3aed3d887aaabf173d23b4a20b537325d0b34","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.video_depth.eval_depth.get_video_results","uri":"program://Human3R/function/eval.video_depth.eval_depth.get_video_results#L314-L382","kind":"function","name":"get_video_results","path":"eval/video_depth/eval_depth.py","language":"python","start_line":314,"end_line":382,"context_start_line":294,"context_end_line":390,"code":"            # and returns it as a numpy array,\n            # for details see readme.txt\n            img_pil = Image.open(filename)\n            depth_png = np.array(img_pil, dtype=int)\n            # make sure we have a proper 16bit depth map here.. not 8bit!\n            assert np.max(depth_png) > 255\n\n            depth = depth_png.astype(float) / 256.0\n            depth[depth_png == 0] = -1.0\n            return depth\n\n        depth_pathes = glob.glob(\n            \"data/kitti/depth_selection/val_selection_cropped/groundtruth_depth_gathered/*/*.png\"\n        )\n        depth_pathes = sorted(depth_pathes)\n        pred_pathes = glob.glob(\n            f\"{args.output_dir}/*/frame_*.npy\"\n        )  # TODO: update the path to your prediction\n        pred_pathes = sorted(pred_pathes)\n\n        def get_video_results():\n            grouped_pred_depth = group_by_directory(pred_pathes)\n            grouped_gt_depth = group_by_directory(depth_pathes)\n            gathered_depth_metrics = []\n            for key in tqdm(grouped_pred_depth.keys()):\n                pd_pathes = grouped_pred_depth[key]\n                gt_pathes = grouped_gt_depth[key]\n                gt_depth = np.stack(\n                    [depth_read(gt_path) for gt_path in gt_pathes], axis=0\n                )\n                pr_depth = np.stack(\n                    [\n                        cv2.resize(\n                            np.load(pd_path),\n                            (gt_depth.shape[2], gt_depth.shape[1]),\n                            interpolation=cv2.INTER_CUBIC,\n                        )\n                        for pd_path in pd_pathes\n                    ],\n                    axis=0,\n                )\n\n                # for depth eval, set align_with_lad2=False to use median alignment; set align_with_lad2=True to use scale&shift alignment\n                if args.align == \"scale&shift\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=None,\n                            align_with_lad2=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"scale\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=None,\n                            align_with_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                elif args.align == \"metric\":\n                    depth_results, error_map, depth_predict, depth_gt = (\n                        depth_evaluation(\n                            pr_depth,\n                            gt_depth,\n                            max_depth=None,\n                            metric_scale=True,\n                            use_gpu=True,\n                        )\n                    )\n                gathered_depth_metrics.append(depth_results)\n\n            depth_log_path = f\"{args.output_dir}/result_{args.align}.json\"\n            average_metrics = {\n                key: np.average(\n                    [metrics[key] for metrics in gathered_depth_metrics],\n                    weights=[\n                        metrics[\"valid_pixels\"] for metrics in gathered_depth_metrics\n                    ],\n                )\n                for key in gathered_depth_metrics[0].keys()\n                if key != \"valid_pixels\"\n            }\n            print(\"Average depth evaluation metrics:\", average_metrics)\n            with open(depth_log_path, \"w\") as f:\n                f.write(json.dumps(average_metrics))\n\n        get_video_results()\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    main(args)","source_hash":"924683ea2cb1871eefa5d1266bb3aed3d887aaabf173d23b4a20b537325d0b34","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata","uri":"program://Human3R/module/eval.global_human.metadata#L1-L366","kind":"module","name":"eval.global_human.metadata","path":"eval/global_human/metadata.py","language":"python","start_line":1,"end_line":366,"context_start_line":1,"context_end_line":366,"code":"import os\nimport torch\nimport numpy as np\nimport pickle\nfrom eval.global_human.data_utils import *\n\n# Define the merged dataset metadata dictionary\n\ndef create_emdb(split):\n    return {\n        \"img_path\": \"/path/to/EMDB\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/{seq}/images\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [],\n        \"split\": split,\n        \"subsample\": 1,\n        \"get_view_func\": lambda inputs: load_view_emdb(*inputs),\n        \"get_seq_func\": lambda img_path, split, annots: get_seq_emdb(img_path, split, annots),\n        \"get_annot_func\": lambda img_path, split: get_annot_emdb(img_path, split),\n        \"is_global\": lambda split: {1: False, 2: True}[split],\n    }\n\ndataset_metadata = {\n    \"bedlam\": {\n        \"img_path\": \"/path/to/processed_bedlam\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/Test/{seq}/rgb\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [f.replace(\"/rgb/\", \"/mask/\") for f in filelist],\n        \"split\": \"Test\",\n        \"subsample\": 25, # 25 is used in Multi-HMR's code\n        \"get_view_func\": lambda inputs: load_view_bedlam(*inputs[:2]),\n        \"get_seq_func\": None,\n        \"is_global\": lambda split: False,\n    },\n    \"3dpw\": {\n        \"img_path\": \"/path/to/3DPW\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/imageFiles/{seq}\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [],\n        \"split\": \"test\",\n        \"subsample\": 1,\n        \"get_view_func\": lambda inputs: load_view_3dpw(*inputs[:3]),\n        \"get_seq_func\": lambda img_path, split, annots: get_seq_3dpw(img_path, split, annots),\n        \"get_annot_func\": lambda img_path, split: get_annot(img_path, split, \"3dpw\"),\n        \"is_global\": lambda split: False,\n    },\n    \"emdb1\": create_emdb(1),\n    \"emdb2\": create_emdb(2),\n    \"rich\": {\n        \"img_path\": \"/path/to/RICH\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/{seq}\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [],\n        \"split\": \"test\",\n        \"subsample\": 1,\n        \"get_view_func\": lambda inputs: load_view_rich(*inputs),\n        \"get_seq_func\": lambda img_path, split, annots: get_seq_rich(img_path, split, annots),\n        \"get_annot_func\": lambda img_path, split: get_annot_rich(),\n        \"is_global\": lambda split: True,\n    }\n}\n\n\ndef get_annot(img_path, split, dataset):\n    annot_file = os.path.join(f\"eval/global_human/annots/{dataset}_{split}.pkl\")\n    with open(annot_file, 'rb') as f:\n        annots = pickle.load(f)\n    return annots\n\ndef get_annot_emdb(img_path, split):\n    annots = {}\n    for pkl_name in EMDB_LIST[split]:\n        data = load_pkl(os.path.join(img_path, pkl_name))\n        annots[data[\"name\"]] = data\n    return annots\n\ndef get_annot_rich():\n    \"\"\"\n    For annotations of RICH dataset, \n    please download from https://github.com/zju3dv/GVHMR/blob/main/docs/INSTALL.md#inputs--outputs\n    \"\"\"\n    annot_path = os.path.join(os.path.dirname(__file__), \"annots/RICH\")\n    annots = torch.load(os.path.join(annot_path, \"hmr4d_support/rich_test_labels.pt\"))\n    cam_params = torch.load(os.path.join(annot_path, \"resource/cam2params.pt\"))\n\n    for vid in list(annots.keys()):\n        _, sname, cname = vid.split(\"/\")\n        scene = sname.split(\"_\")[0]\n        cid = int(cname.split(\"_\")[1])\n        cam_key = f\"{scene}_{cid}\"\n        annots[vid][\"T_w2c\"], annots[vid][\"K\"] = cam_params[cam_key]\n    \n    return annots\n\ndef get_seq_emdb(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot_emdb(img_path, split)\n\n    seq_list = list(annots.keys())\n    seq_to_images = {}\n    for seq in seq_list:\n        mask = annots[seq][\"mask\"]\n        seq_images_dir = os.path.join(img_path, seq, \"images\")\n        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_rich(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot_rich()\n\n    seq_list = list(annots.keys())\n\n    seq_to_images = {}\n    for seq in seq_list:\n        mask = annots[seq][\"frame_id\"]\n        seq_images_dir = os.path.join(img_path, seq)\n        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_3dpw(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot(img_path, split, \"3dpw\")\n\n    imagenames = sorted(annots.keys())\n\n    seq_to_images = {}\n    for imgname in imagenames:\n        seq, image = imgname.split('/')\n        seq_to_images.setdefault(seq, []).append(image)\n\n    return list(seq_to_images.keys()), seq_to_images\n\ndef load_view_bedlam(img_paths, images):\n    max_humans = 10\n    smpl_key2shape = {\n        'smplx_root_pose': (1, 3), \n        'smplx_body_pose': (21, 3), \n        'smplx_jaw_pose': (1, 3), \n        'smplx_leye_pose': (1, 3), \n        'smplx_reye_pose': (1, 3), \n        'smplx_left_hand_pose': (15, 3), \n        'smplx_right_hand_pose': (15, 3), \n        'smplx_shape': (11,), \n        'smplx_transl': (3,), \n        'smplx_gender_id': (),\n        }\n    for img_path, image in zip(img_paths, images):\n        # load camera\n        cam_path = img_path.replace(\"/rgb/\", \"/cam/\").replace(\".png\", \".npz\")\n        cam = np.load(cam_path)\n        camera_pose = cam[\"pose\"]\n        intrinsics = cam[\"intrinsics\"]\n\n        # calculate scale factor\n        fy_scale, fx_scale = (image['true_shape'] / image['ori_shape'])[0]\n\n        # update intrinsics\n        intrinsics[0, 0] *= fx_scale  # fx\n        intrinsics[1, 1] *= fy_scale  # fy\n        intrinsics[0, 2] *= fx_scale  # cx\n        intrinsics[1, 2] *= fy_scale  # cy\n\n        # load smpl\n        annot_file = img_path.replace(\"/rgb/\", \"/smpl/\").replace(\".png\", \".pkl\")\n        annots = []\n        smpl_mask = np.zeros(max_humans, dtype=np.bool_)\n\n        if os.path.isfile(annot_file):\n            with open(annot_file, 'rb') as f:\n                annots = pickle.load(f)\n            humans = [hum for hum in annots]\n            # humans = [hum for hum in annots if hum['smplx_transl'][-1] > 0.01] # the person should be in front of the camera\n            if len(humans) > 0:\n                smpl_mask[:len(humans)] = 1.\n                l_dist = [hum['smplx_transl'][-1] for hum in humans]\n                indexed_lst = list(enumerate(l_dist))\n                sorted_indexed = sorted(indexed_lst, key=lambda x: x[1], reverse=False)\n                sorted_indices = [index for index, _ in sorted_indexed]\n                annots = [humans[h_idx] for h_idx in sorted_indices]\n\n                # Update smplx_gender - 0=neutral - 1=male - 2=female - kids?\n                for hum in annots:\n                    hum['smplx_gender_id'] = np.asarray({'neutral': 0}[hum['smplx_gender']])\n\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict[k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            if len(humans) > 0:\n                for h in range(len(humans)):\n                    smpl_dict[k][h] = annots[h][k].astype(np.float32)\n\n        image.update({\n            'camera_pose': camera_pose.astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            **smpl_dict,\n        })\n\n    return images\n\ndef load_view_3dpw(img_paths, images, annots):\n    max_humans = 2\n    smpl_key2shape = {\n        'smpl_root_pose': (1, 3), \n        'smpl_body_pose': (23, 3), \n        'smpl_shape': (10,), \n        'smpl_transl': (3,), \n        'smpl_gender_id': (),\n        }\n\n    for img_path, image in zip(img_paths, images):\n        img_key = '/'.join(img_path.split('/')[-2:])\n\n        # load camera\n        T_w2c = annots[img_key][\"cam_poses\"]\n        camera_pose = np.linalg.inv(T_w2c) # T_c2w\n        focal = annots[img_key][\"focal\"]\n        princpt = annots[img_key][\"princpt\"]\n        intrinsics = np.array([\n            [focal[0], 0, princpt[0]],\n            [0, focal[1], princpt[1]],\n            [0, 0, 1]\n        ], dtype=np.float32)\n\n        # calculate scale factor\n        fy_scale, fx_scale = (image['true_shape'] / image['ori_shape'])[0]\n\n        # update intrinsics\n        intrinsics[0, 0] *= fx_scale  # fx\n        intrinsics[1, 1] *= fy_scale  # fy\n        intrinsics[0, 2] *= fx_scale  # cx\n        intrinsics[1, 2] *= fy_scale  # cy\n\n        # load smpl\n        humans = annots[img_key][\"humans\"]\n        smpl_mask = np.zeros(max_humans, dtype=np.bool_)\n\n        if len(humans) > 0:\n            smpl_mask[:len(humans)] = 1.\n            l_dist = [hum['smpl_transl'][-1] for hum in humans]\n            indexed_lst = list(enumerate(l_dist))\n            sorted_indexed = sorted(indexed_lst, key=lambda x: x[1], reverse=False)\n            sorted_indices = [index for index, _ in sorted_indexed]\n            annots_humans = [humans[h_idx] for h_idx in sorted_indices]\n\n            # Update smplx_gender - 0=neutral - 1=male - 2=female - kids?\n            for hum in annots_humans:\n                hum['smpl_gender_id'] = np.asarray({'male': 1, 'female': 2}[hum['smpl_gender']])\n\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict[k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            if len(humans) > 0:\n                for h in range(len(humans)):\n                    smpl_dict[k][h] = annots_humans[h][k].astype(np.float32)\n\n        image.update({\n            'camera_pose': camera_pose.astype(np.float32),\n            'T_w2c': T_w2c.astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            **smpl_dict,\n        })\n\n    return images\n\ndef load_view_emdb(img_paths, images, annots, indices):\n    seq = '/'.join(img_paths[0].split('/')[-4:-2])\n    gender = annots[seq]['gender']\n    masks = annots[seq]['mask']\n\n    # load camera\n    T_w2c = annots[seq]['T_w2c'][masks] # (n_frame,4,4)\n    camera_poses = np.linalg.inv(T_w2c) # T_c2w: (n_frame,4,4)\n    intrinsics = annots[seq]['K_fullimg'].copy() # (3,3)\n\n    # update intrinsics\n    fy_scale, fx_scale = (\n        images[0]['true_shape'] / images[0]['ori_shape'])[0]\n    intrinsics[0, 0] *= fx_scale  # fx\n    intrinsics[1, 1] *= fy_scale  # fy\n    intrinsics[0, 2] *= fx_scale  # cx\n    intrinsics[1, 2] *= fy_scale  # cy\n\n    # load smpl\n    smpl_params = annots[seq]['smpl_params']    # world space\n\n    max_humans = 1\n    smpl_key2shape = {\n        'smpl_root_pose_w': (1, 3), \n        'smpl_body_pose': (23, 3), \n        'smpl_shape': (10,), \n        'smpl_transl_w': (3,), \n        }\n\n    for img_id, image in zip(indices, images):\n        smpl_mask = np.ones(max_humans, dtype=np.bool_)\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict[k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            smpl_dict[k][0] = smpl_params[k][masks][img_id].reshape(*shape).astype(np.float32)\n\n        image.update({\n            'T_w2c': T_w2c[img_id].astype(np.float32),\n            'camera_pose': camera_poses[img_id].astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            'smpl_gender_id': np.asarray({'male': 1, 'female': 2}[gender]),\n            **smpl_dict,\n        })\n\n    return images\n\ndef load_view_rich(img_paths, images, annots, indices):\n    seq = '/'.join(img_paths[0].split('/')[-4:-1])\n    gender = annots[seq]['gender']\n\n    # load camera\n    T_w2c = annots[seq]['T_w2c'].numpy() # (4,4)\n    camera_poses = np.linalg.inv(T_w2c) # T_c2w: (4,4)\n    intrinsics = annots[seq]['K'].numpy().copy() # (3,3)\n\n    # update intrinsics\n    fy_scale, fx_scale = (\n        images[0]['true_shape'] / images[0]['ori_shape'])[0]\n    intrinsics[0, 0] *= fx_scale  # fx\n    intrinsics[1, 1] *= fy_scale  # fy\n    intrinsics[0, 2] *= fx_scale  # cx\n    intrinsics[1, 2] *= fy_scale  # cy\n\n    # load smpl\n    smpl_params = annots[seq]['gt_smplx_params']    # world space\n\n    max_humans = 1\n    smpl_key2shape = {\n        'global_orient': (1, 3), \n        'body_pose': (21, 3), \n        'betas': (10,), \n        'transl': (3,), \n        }\n\n    for img_id, image in zip(indices, images):\n        smpl_mask = np.ones(max_humans, dtype=np.bool_)\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict['smplx_'+k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            smpl_dict['smplx_'+k][0] = smpl_params[k][img_id].reshape(*shape).numpy().astype(np.float32)\n\n        image.update({\n            'T_w2c': T_w2c.astype(np.float32),\n            'camera_pose': camera_poses.astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            'smplx_gender_id': np.asarray({'male': 1, 'female': 2}[gender]),\n            **smpl_dict,\n        })\n\n    return images","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.create_emdb","uri":"program://Human3R/function/eval.global_human.metadata.create_emdb#L9-L22","kind":"function","name":"create_emdb","path":"eval/global_human/metadata.py","language":"python","start_line":9,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os\nimport torch\nimport numpy as np\nimport pickle\nfrom eval.global_human.data_utils import *\n\n# Define the merged dataset metadata dictionary\n\ndef create_emdb(split):\n    return {\n        \"img_path\": \"/path/to/EMDB\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/{seq}/images\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [],\n        \"split\": split,\n        \"subsample\": 1,\n        \"get_view_func\": lambda inputs: load_view_emdb(*inputs),\n        \"get_seq_func\": lambda img_path, split, annots: get_seq_emdb(img_path, split, annots),\n        \"get_annot_func\": lambda img_path, split: get_annot_emdb(img_path, split),\n        \"is_global\": lambda split: {1: False, 2: True}[split],\n    }\n\ndataset_metadata = {\n    \"bedlam\": {\n        \"img_path\": \"/path/to/processed_bedlam\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/Test/{seq}/rgb\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [f.replace(\"/rgb/\", \"/mask/\") for f in filelist],\n        \"split\": \"Test\",\n        \"subsample\": 25, # 25 is used in Multi-HMR's code\n        \"get_view_func\": lambda inputs: load_view_bedlam(*inputs[:2]),\n        \"get_seq_func\": None,\n        \"is_global\": lambda split: False,\n    },\n    \"3dpw\": {\n        \"img_path\": \"/path/to/3DPW\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/imageFiles/{seq}\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [],","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.get_annot","uri":"program://Human3R/function/eval.global_human.metadata.get_annot#L68-L72","kind":"function","name":"get_annot","path":"eval/global_human/metadata.py","language":"python","start_line":68,"end_line":72,"context_start_line":48,"context_end_line":92,"code":"        \"is_global\": lambda split: False,\n    },\n    \"emdb1\": create_emdb(1),\n    \"emdb2\": create_emdb(2),\n    \"rich\": {\n        \"img_path\": \"/path/to/RICH\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/{seq}\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [],\n        \"split\": \"test\",\n        \"subsample\": 1,\n        \"get_view_func\": lambda inputs: load_view_rich(*inputs),\n        \"get_seq_func\": lambda img_path, split, annots: get_seq_rich(img_path, split, annots),\n        \"get_annot_func\": lambda img_path, split: get_annot_rich(),\n        \"is_global\": lambda split: True,\n    }\n}\n\n\ndef get_annot(img_path, split, dataset):\n    annot_file = os.path.join(f\"eval/global_human/annots/{dataset}_{split}.pkl\")\n    with open(annot_file, 'rb') as f:\n        annots = pickle.load(f)\n    return annots\n\ndef get_annot_emdb(img_path, split):\n    annots = {}\n    for pkl_name in EMDB_LIST[split]:\n        data = load_pkl(os.path.join(img_path, pkl_name))\n        annots[data[\"name\"]] = data\n    return annots\n\ndef get_annot_rich():\n    \"\"\"\n    For annotations of RICH dataset, \n    please download from https://github.com/zju3dv/GVHMR/blob/main/docs/INSTALL.md#inputs--outputs\n    \"\"\"\n    annot_path = os.path.join(os.path.dirname(__file__), \"annots/RICH\")\n    annots = torch.load(os.path.join(annot_path, \"hmr4d_support/rich_test_labels.pt\"))\n    cam_params = torch.load(os.path.join(annot_path, \"resource/cam2params.pt\"))\n\n    for vid in list(annots.keys()):\n        _, sname, cname = vid.split(\"/\")\n        scene = sname.split(\"_\")[0]","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.get_annot_emdb","uri":"program://Human3R/function/eval.global_human.metadata.get_annot_emdb#L74-L79","kind":"function","name":"get_annot_emdb","path":"eval/global_human/metadata.py","language":"python","start_line":74,"end_line":79,"context_start_line":54,"context_end_line":99,"code":"        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/{seq}\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [],\n        \"split\": \"test\",\n        \"subsample\": 1,\n        \"get_view_func\": lambda inputs: load_view_rich(*inputs),\n        \"get_seq_func\": lambda img_path, split, annots: get_seq_rich(img_path, split, annots),\n        \"get_annot_func\": lambda img_path, split: get_annot_rich(),\n        \"is_global\": lambda split: True,\n    }\n}\n\n\ndef get_annot(img_path, split, dataset):\n    annot_file = os.path.join(f\"eval/global_human/annots/{dataset}_{split}.pkl\")\n    with open(annot_file, 'rb') as f:\n        annots = pickle.load(f)\n    return annots\n\ndef get_annot_emdb(img_path, split):\n    annots = {}\n    for pkl_name in EMDB_LIST[split]:\n        data = load_pkl(os.path.join(img_path, pkl_name))\n        annots[data[\"name\"]] = data\n    return annots\n\ndef get_annot_rich():\n    \"\"\"\n    For annotations of RICH dataset, \n    please download from https://github.com/zju3dv/GVHMR/blob/main/docs/INSTALL.md#inputs--outputs\n    \"\"\"\n    annot_path = os.path.join(os.path.dirname(__file__), \"annots/RICH\")\n    annots = torch.load(os.path.join(annot_path, \"hmr4d_support/rich_test_labels.pt\"))\n    cam_params = torch.load(os.path.join(annot_path, \"resource/cam2params.pt\"))\n\n    for vid in list(annots.keys()):\n        _, sname, cname = vid.split(\"/\")\n        scene = sname.split(\"_\")[0]\n        cid = int(cname.split(\"_\")[1])\n        cam_key = f\"{scene}_{cid}\"\n        annots[vid][\"T_w2c\"], annots[vid][\"K\"] = cam_params[cam_key]\n    \n    return annots\n\ndef get_seq_emdb(img_path, split, annots=None):","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.get_annot_rich","uri":"program://Human3R/function/eval.global_human.metadata.get_annot_rich#L81-L97","kind":"function","name":"get_annot_rich","path":"eval/global_human/metadata.py","language":"python","start_line":81,"end_line":97,"context_start_line":61,"context_end_line":117,"code":"        \"get_seq_func\": lambda img_path, split, annots: get_seq_rich(img_path, split, annots),\n        \"get_annot_func\": lambda img_path, split: get_annot_rich(),\n        \"is_global\": lambda split: True,\n    }\n}\n\n\ndef get_annot(img_path, split, dataset):\n    annot_file = os.path.join(f\"eval/global_human/annots/{dataset}_{split}.pkl\")\n    with open(annot_file, 'rb') as f:\n        annots = pickle.load(f)\n    return annots\n\ndef get_annot_emdb(img_path, split):\n    annots = {}\n    for pkl_name in EMDB_LIST[split]:\n        data = load_pkl(os.path.join(img_path, pkl_name))\n        annots[data[\"name\"]] = data\n    return annots\n\ndef get_annot_rich():\n    \"\"\"\n    For annotations of RICH dataset, \n    please download from https://github.com/zju3dv/GVHMR/blob/main/docs/INSTALL.md#inputs--outputs\n    \"\"\"\n    annot_path = os.path.join(os.path.dirname(__file__), \"annots/RICH\")\n    annots = torch.load(os.path.join(annot_path, \"hmr4d_support/rich_test_labels.pt\"))\n    cam_params = torch.load(os.path.join(annot_path, \"resource/cam2params.pt\"))\n\n    for vid in list(annots.keys()):\n        _, sname, cname = vid.split(\"/\")\n        scene = sname.split(\"_\")[0]\n        cid = int(cname.split(\"_\")[1])\n        cam_key = f\"{scene}_{cid}\"\n        annots[vid][\"T_w2c\"], annots[vid][\"K\"] = cam_params[cam_key]\n    \n    return annots\n\ndef get_seq_emdb(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot_emdb(img_path, split)\n\n    seq_list = list(annots.keys())\n    seq_to_images = {}\n    for seq in seq_list:\n        mask = annots[seq][\"mask\"]\n        seq_images_dir = os.path.join(img_path, seq, \"images\")\n        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_rich(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot_rich()\n\n    seq_list = list(annots.keys())","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.get_seq_emdb","uri":"program://Human3R/function/eval.global_human.metadata.get_seq_emdb#L99-L111","kind":"function","name":"get_seq_emdb","path":"eval/global_human/metadata.py","language":"python","start_line":99,"end_line":111,"context_start_line":79,"context_end_line":131,"code":"    return annots\n\ndef get_annot_rich():\n    \"\"\"\n    For annotations of RICH dataset, \n    please download from https://github.com/zju3dv/GVHMR/blob/main/docs/INSTALL.md#inputs--outputs\n    \"\"\"\n    annot_path = os.path.join(os.path.dirname(__file__), \"annots/RICH\")\n    annots = torch.load(os.path.join(annot_path, \"hmr4d_support/rich_test_labels.pt\"))\n    cam_params = torch.load(os.path.join(annot_path, \"resource/cam2params.pt\"))\n\n    for vid in list(annots.keys()):\n        _, sname, cname = vid.split(\"/\")\n        scene = sname.split(\"_\")[0]\n        cid = int(cname.split(\"_\")[1])\n        cam_key = f\"{scene}_{cid}\"\n        annots[vid][\"T_w2c\"], annots[vid][\"K\"] = cam_params[cam_key]\n    \n    return annots\n\ndef get_seq_emdb(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot_emdb(img_path, split)\n\n    seq_list = list(annots.keys())\n    seq_to_images = {}\n    for seq in seq_list:\n        mask = annots[seq][\"mask\"]\n        seq_images_dir = os.path.join(img_path, seq, \"images\")\n        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_rich(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot_rich()\n\n    seq_list = list(annots.keys())\n\n    seq_to_images = {}\n    for seq in seq_list:\n        mask = annots[seq][\"frame_id\"]\n        seq_images_dir = os.path.join(img_path, seq)\n        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_3dpw(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot(img_path, split, \"3dpw\")\n","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.get_seq_rich","uri":"program://Human3R/function/eval.global_human.metadata.get_seq_rich#L113-L126","kind":"function","name":"get_seq_rich","path":"eval/global_human/metadata.py","language":"python","start_line":113,"end_line":126,"context_start_line":93,"context_end_line":146,"code":"        cid = int(cname.split(\"_\")[1])\n        cam_key = f\"{scene}_{cid}\"\n        annots[vid][\"T_w2c\"], annots[vid][\"K\"] = cam_params[cam_key]\n    \n    return annots\n\ndef get_seq_emdb(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot_emdb(img_path, split)\n\n    seq_list = list(annots.keys())\n    seq_to_images = {}\n    for seq in seq_list:\n        mask = annots[seq][\"mask\"]\n        seq_images_dir = os.path.join(img_path, seq, \"images\")\n        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_rich(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot_rich()\n\n    seq_list = list(annots.keys())\n\n    seq_to_images = {}\n    for seq in seq_list:\n        mask = annots[seq][\"frame_id\"]\n        seq_images_dir = os.path.join(img_path, seq)\n        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_3dpw(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot(img_path, split, \"3dpw\")\n\n    imagenames = sorted(annots.keys())\n\n    seq_to_images = {}\n    for imgname in imagenames:\n        seq, image = imgname.split('/')\n        seq_to_images.setdefault(seq, []).append(image)\n\n    return list(seq_to_images.keys()), seq_to_images\n\ndef load_view_bedlam(img_paths, images):\n    max_humans = 10\n    smpl_key2shape = {\n        'smplx_root_pose': (1, 3), \n        'smplx_body_pose': (21, 3), \n        'smplx_jaw_pose': (1, 3), ","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.get_seq_3dpw","uri":"program://Human3R/function/eval.global_human.metadata.get_seq_3dpw#L128-L139","kind":"function","name":"get_seq_3dpw","path":"eval/global_human/metadata.py","language":"python","start_line":128,"end_line":139,"context_start_line":108,"context_end_line":159,"code":"        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_rich(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot_rich()\n\n    seq_list = list(annots.keys())\n\n    seq_to_images = {}\n    for seq in seq_list:\n        mask = annots[seq][\"frame_id\"]\n        seq_images_dir = os.path.join(img_path, seq)\n        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_3dpw(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot(img_path, split, \"3dpw\")\n\n    imagenames = sorted(annots.keys())\n\n    seq_to_images = {}\n    for imgname in imagenames:\n        seq, image = imgname.split('/')\n        seq_to_images.setdefault(seq, []).append(image)\n\n    return list(seq_to_images.keys()), seq_to_images\n\ndef load_view_bedlam(img_paths, images):\n    max_humans = 10\n    smpl_key2shape = {\n        'smplx_root_pose': (1, 3), \n        'smplx_body_pose': (21, 3), \n        'smplx_jaw_pose': (1, 3), \n        'smplx_leye_pose': (1, 3), \n        'smplx_reye_pose': (1, 3), \n        'smplx_left_hand_pose': (15, 3), \n        'smplx_right_hand_pose': (15, 3), \n        'smplx_shape': (11,), \n        'smplx_transl': (3,), \n        'smplx_gender_id': (),\n        }\n    for img_path, image in zip(img_paths, images):\n        # load camera\n        cam_path = img_path.replace(\"/rgb/\", \"/cam/\").replace(\".png\", \".npz\")\n        cam = np.load(cam_path)\n        camera_pose = cam[\"pose\"]","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.load_view_bedlam","uri":"program://Human3R/function/eval.global_human.metadata.load_view_bedlam#L141-L207","kind":"function","name":"load_view_bedlam","path":"eval/global_human/metadata.py","language":"python","start_line":141,"end_line":207,"context_start_line":121,"context_end_line":227,"code":"        mask = annots[seq][\"frame_id\"]\n        seq_images_dir = os.path.join(img_path, seq)\n        all_images = np.array(sorted(os.listdir(seq_images_dir)))\n        seq_to_images[seq] = all_images[mask].tolist()\n\n    return seq_list, seq_to_images\n\ndef get_seq_3dpw(img_path, split, annots=None):\n    if annots is None:\n        annots = get_annot(img_path, split, \"3dpw\")\n\n    imagenames = sorted(annots.keys())\n\n    seq_to_images = {}\n    for imgname in imagenames:\n        seq, image = imgname.split('/')\n        seq_to_images.setdefault(seq, []).append(image)\n\n    return list(seq_to_images.keys()), seq_to_images\n\ndef load_view_bedlam(img_paths, images):\n    max_humans = 10\n    smpl_key2shape = {\n        'smplx_root_pose': (1, 3), \n        'smplx_body_pose': (21, 3), \n        'smplx_jaw_pose': (1, 3), \n        'smplx_leye_pose': (1, 3), \n        'smplx_reye_pose': (1, 3), \n        'smplx_left_hand_pose': (15, 3), \n        'smplx_right_hand_pose': (15, 3), \n        'smplx_shape': (11,), \n        'smplx_transl': (3,), \n        'smplx_gender_id': (),\n        }\n    for img_path, image in zip(img_paths, images):\n        # load camera\n        cam_path = img_path.replace(\"/rgb/\", \"/cam/\").replace(\".png\", \".npz\")\n        cam = np.load(cam_path)\n        camera_pose = cam[\"pose\"]\n        intrinsics = cam[\"intrinsics\"]\n\n        # calculate scale factor\n        fy_scale, fx_scale = (image['true_shape'] / image['ori_shape'])[0]\n\n        # update intrinsics\n        intrinsics[0, 0] *= fx_scale  # fx\n        intrinsics[1, 1] *= fy_scale  # fy\n        intrinsics[0, 2] *= fx_scale  # cx\n        intrinsics[1, 2] *= fy_scale  # cy\n\n        # load smpl\n        annot_file = img_path.replace(\"/rgb/\", \"/smpl/\").replace(\".png\", \".pkl\")\n        annots = []\n        smpl_mask = np.zeros(max_humans, dtype=np.bool_)\n\n        if os.path.isfile(annot_file):\n            with open(annot_file, 'rb') as f:\n                annots = pickle.load(f)\n            humans = [hum for hum in annots]\n            # humans = [hum for hum in annots if hum['smplx_transl'][-1] > 0.01] # the person should be in front of the camera\n            if len(humans) > 0:\n                smpl_mask[:len(humans)] = 1.\n                l_dist = [hum['smplx_transl'][-1] for hum in humans]\n                indexed_lst = list(enumerate(l_dist))\n                sorted_indexed = sorted(indexed_lst, key=lambda x: x[1], reverse=False)\n                sorted_indices = [index for index, _ in sorted_indexed]\n                annots = [humans[h_idx] for h_idx in sorted_indices]\n\n                # Update smplx_gender - 0=neutral - 1=male - 2=female - kids?\n                for hum in annots:\n                    hum['smplx_gender_id'] = np.asarray({'neutral': 0}[hum['smplx_gender']])\n\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict[k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            if len(humans) > 0:\n                for h in range(len(humans)):\n                    smpl_dict[k][h] = annots[h][k].astype(np.float32)\n\n        image.update({\n            'camera_pose': camera_pose.astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            **smpl_dict,\n        })\n\n    return images\n\ndef load_view_3dpw(img_paths, images, annots):\n    max_humans = 2\n    smpl_key2shape = {\n        'smpl_root_pose': (1, 3), \n        'smpl_body_pose': (23, 3), \n        'smpl_shape': (10,), \n        'smpl_transl': (3,), \n        'smpl_gender_id': (),\n        }\n\n    for img_path, image in zip(img_paths, images):\n        img_key = '/'.join(img_path.split('/')[-2:])\n\n        # load camera\n        T_w2c = annots[img_key][\"cam_poses\"]\n        camera_pose = np.linalg.inv(T_w2c) # T_c2w\n        focal = annots[img_key][\"focal\"]\n        princpt = annots[img_key][\"princpt\"]\n        intrinsics = np.array([","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.load_view_3dpw","uri":"program://Human3R/function/eval.global_human.metadata.load_view_3dpw#L209-L273","kind":"function","name":"load_view_3dpw","path":"eval/global_human/metadata.py","language":"python","start_line":209,"end_line":273,"context_start_line":189,"context_end_line":293,"code":"                # Update smplx_gender - 0=neutral - 1=male - 2=female - kids?\n                for hum in annots:\n                    hum['smplx_gender_id'] = np.asarray({'neutral': 0}[hum['smplx_gender']])\n\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict[k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            if len(humans) > 0:\n                for h in range(len(humans)):\n                    smpl_dict[k][h] = annots[h][k].astype(np.float32)\n\n        image.update({\n            'camera_pose': camera_pose.astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            **smpl_dict,\n        })\n\n    return images\n\ndef load_view_3dpw(img_paths, images, annots):\n    max_humans = 2\n    smpl_key2shape = {\n        'smpl_root_pose': (1, 3), \n        'smpl_body_pose': (23, 3), \n        'smpl_shape': (10,), \n        'smpl_transl': (3,), \n        'smpl_gender_id': (),\n        }\n\n    for img_path, image in zip(img_paths, images):\n        img_key = '/'.join(img_path.split('/')[-2:])\n\n        # load camera\n        T_w2c = annots[img_key][\"cam_poses\"]\n        camera_pose = np.linalg.inv(T_w2c) # T_c2w\n        focal = annots[img_key][\"focal\"]\n        princpt = annots[img_key][\"princpt\"]\n        intrinsics = np.array([\n            [focal[0], 0, princpt[0]],\n            [0, focal[1], princpt[1]],\n            [0, 0, 1]\n        ], dtype=np.float32)\n\n        # calculate scale factor\n        fy_scale, fx_scale = (image['true_shape'] / image['ori_shape'])[0]\n\n        # update intrinsics\n        intrinsics[0, 0] *= fx_scale  # fx\n        intrinsics[1, 1] *= fy_scale  # fy\n        intrinsics[0, 2] *= fx_scale  # cx\n        intrinsics[1, 2] *= fy_scale  # cy\n\n        # load smpl\n        humans = annots[img_key][\"humans\"]\n        smpl_mask = np.zeros(max_humans, dtype=np.bool_)\n\n        if len(humans) > 0:\n            smpl_mask[:len(humans)] = 1.\n            l_dist = [hum['smpl_transl'][-1] for hum in humans]\n            indexed_lst = list(enumerate(l_dist))\n            sorted_indexed = sorted(indexed_lst, key=lambda x: x[1], reverse=False)\n            sorted_indices = [index for index, _ in sorted_indexed]\n            annots_humans = [humans[h_idx] for h_idx in sorted_indices]\n\n            # Update smplx_gender - 0=neutral - 1=male - 2=female - kids?\n            for hum in annots_humans:\n                hum['smpl_gender_id'] = np.asarray({'male': 1, 'female': 2}[hum['smpl_gender']])\n\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict[k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            if len(humans) > 0:\n                for h in range(len(humans)):\n                    smpl_dict[k][h] = annots_humans[h][k].astype(np.float32)\n\n        image.update({\n            'camera_pose': camera_pose.astype(np.float32),\n            'T_w2c': T_w2c.astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            **smpl_dict,\n        })\n\n    return images\n\ndef load_view_emdb(img_paths, images, annots, indices):\n    seq = '/'.join(img_paths[0].split('/')[-4:-2])\n    gender = annots[seq]['gender']\n    masks = annots[seq]['mask']\n\n    # load camera\n    T_w2c = annots[seq]['T_w2c'][masks] # (n_frame,4,4)\n    camera_poses = np.linalg.inv(T_w2c) # T_c2w: (n_frame,4,4)\n    intrinsics = annots[seq]['K_fullimg'].copy() # (3,3)\n\n    # update intrinsics\n    fy_scale, fx_scale = (\n        images[0]['true_shape'] / images[0]['ori_shape'])[0]\n    intrinsics[0, 0] *= fx_scale  # fx\n    intrinsics[1, 1] *= fy_scale  # fy\n    intrinsics[0, 2] *= fx_scale  # cx\n    intrinsics[1, 2] *= fy_scale  # cy\n\n    # load smpl","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.load_view_emdb","uri":"program://Human3R/function/eval.global_human.metadata.load_view_emdb#L275-L320","kind":"function","name":"load_view_emdb","path":"eval/global_human/metadata.py","language":"python","start_line":275,"end_line":320,"context_start_line":255,"context_end_line":340,"code":"            for hum in annots_humans:\n                hum['smpl_gender_id'] = np.asarray({'male': 1, 'female': 2}[hum['smpl_gender']])\n\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict[k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            if len(humans) > 0:\n                for h in range(len(humans)):\n                    smpl_dict[k][h] = annots_humans[h][k].astype(np.float32)\n\n        image.update({\n            'camera_pose': camera_pose.astype(np.float32),\n            'T_w2c': T_w2c.astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            **smpl_dict,\n        })\n\n    return images\n\ndef load_view_emdb(img_paths, images, annots, indices):\n    seq = '/'.join(img_paths[0].split('/')[-4:-2])\n    gender = annots[seq]['gender']\n    masks = annots[seq]['mask']\n\n    # load camera\n    T_w2c = annots[seq]['T_w2c'][masks] # (n_frame,4,4)\n    camera_poses = np.linalg.inv(T_w2c) # T_c2w: (n_frame,4,4)\n    intrinsics = annots[seq]['K_fullimg'].copy() # (3,3)\n\n    # update intrinsics\n    fy_scale, fx_scale = (\n        images[0]['true_shape'] / images[0]['ori_shape'])[0]\n    intrinsics[0, 0] *= fx_scale  # fx\n    intrinsics[1, 1] *= fy_scale  # fy\n    intrinsics[0, 2] *= fx_scale  # cx\n    intrinsics[1, 2] *= fy_scale  # cy\n\n    # load smpl\n    smpl_params = annots[seq]['smpl_params']    # world space\n\n    max_humans = 1\n    smpl_key2shape = {\n        'smpl_root_pose_w': (1, 3), \n        'smpl_body_pose': (23, 3), \n        'smpl_shape': (10,), \n        'smpl_transl_w': (3,), \n        }\n\n    for img_id, image in zip(indices, images):\n        smpl_mask = np.ones(max_humans, dtype=np.bool_)\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict[k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            smpl_dict[k][0] = smpl_params[k][masks][img_id].reshape(*shape).astype(np.float32)\n\n        image.update({\n            'T_w2c': T_w2c[img_id].astype(np.float32),\n            'camera_pose': camera_poses[img_id].astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            'smpl_gender_id': np.asarray({'male': 1, 'female': 2}[gender]),\n            **smpl_dict,\n        })\n\n    return images\n\ndef load_view_rich(img_paths, images, annots, indices):\n    seq = '/'.join(img_paths[0].split('/')[-4:-1])\n    gender = annots[seq]['gender']\n\n    # load camera\n    T_w2c = annots[seq]['T_w2c'].numpy() # (4,4)\n    camera_poses = np.linalg.inv(T_w2c) # T_c2w: (4,4)\n    intrinsics = annots[seq]['K'].numpy().copy() # (3,3)\n\n    # update intrinsics\n    fy_scale, fx_scale = (\n        images[0]['true_shape'] / images[0]['ori_shape'])[0]\n    intrinsics[0, 0] *= fx_scale  # fx\n    intrinsics[1, 1] *= fy_scale  # fy\n    intrinsics[0, 2] *= fx_scale  # cx\n    intrinsics[1, 2] *= fy_scale  # cy\n\n    # load smpl\n    smpl_params = annots[seq]['gt_smplx_params']    # world space","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.metadata.load_view_rich","uri":"program://Human3R/function/eval.global_human.metadata.load_view_rich#L322-L366","kind":"function","name":"load_view_rich","path":"eval/global_human/metadata.py","language":"python","start_line":322,"end_line":366,"context_start_line":302,"context_end_line":366,"code":"        }\n\n    for img_id, image in zip(indices, images):\n        smpl_mask = np.ones(max_humans, dtype=np.bool_)\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict[k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            smpl_dict[k][0] = smpl_params[k][masks][img_id].reshape(*shape).astype(np.float32)\n\n        image.update({\n            'T_w2c': T_w2c[img_id].astype(np.float32),\n            'camera_pose': camera_poses[img_id].astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            'smpl_gender_id': np.asarray({'male': 1, 'female': 2}[gender]),\n            **smpl_dict,\n        })\n\n    return images\n\ndef load_view_rich(img_paths, images, annots, indices):\n    seq = '/'.join(img_paths[0].split('/')[-4:-1])\n    gender = annots[seq]['gender']\n\n    # load camera\n    T_w2c = annots[seq]['T_w2c'].numpy() # (4,4)\n    camera_poses = np.linalg.inv(T_w2c) # T_c2w: (4,4)\n    intrinsics = annots[seq]['K'].numpy().copy() # (3,3)\n\n    # update intrinsics\n    fy_scale, fx_scale = (\n        images[0]['true_shape'] / images[0]['ori_shape'])[0]\n    intrinsics[0, 0] *= fx_scale  # fx\n    intrinsics[1, 1] *= fy_scale  # fy\n    intrinsics[0, 2] *= fx_scale  # cx\n    intrinsics[1, 2] *= fy_scale  # cy\n\n    # load smpl\n    smpl_params = annots[seq]['gt_smplx_params']    # world space\n\n    max_humans = 1\n    smpl_key2shape = {\n        'global_orient': (1, 3), \n        'body_pose': (21, 3), \n        'betas': (10,), \n        'transl': (3,), \n        }\n\n    for img_id, image in zip(indices, images):\n        smpl_mask = np.ones(max_humans, dtype=np.bool_)\n        smpl_dict = {}\n        for k, shape in smpl_key2shape.items():\n            smpl_dict['smplx_'+k] = np.zeros((max_humans, *shape), dtype=np.float32)\n            smpl_dict['smplx_'+k][0] = smpl_params[k][img_id].reshape(*shape).numpy().astype(np.float32)\n\n        image.update({\n            'T_w2c': T_w2c.astype(np.float32),\n            'camera_pose': camera_poses.astype(np.float32),\n            'intrinsics': intrinsics.astype(np.float32),\n            'smpl_mask': smpl_mask,\n            'smplx_gender_id': np.asarray({'male': 1, 'female': 2}[gender]),\n            **smpl_dict,\n        })\n\n    return images","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.data_utils","uri":"program://Human3R/module/eval.global_human.data_utils#L1-L90","kind":"module","name":"eval.global_human.data_utils","path":"eval/global_human/data_utils.py","language":"python","start_line":1,"end_line":90,"context_start_line":1,"context_end_line":90,"code":"# Modified from GVHMR [https://github.com/zju3dv/GVHMR].\n# Load EMDB data\n\nimport pickle\nimport torch\n\nEMDB1_LIST = [\n    'P8/69_outdoor_cartwheel/P8_69_outdoor_cartwheel_data.pkl', # 656\n    'P5/42_indoor_dancing/P5_42_indoor_dancing_data.pkl', # 1291\n    'P6/51_outdoor_dancing/P6_51_outdoor_dancing_data.pkl',  # 1427\n    'P2/23_outdoor_hug_tree/P2_23_outdoor_hug_tree_data.pkl',  # 1086\n    'P6/49_outdoor_big_stairs_down/P6_49_outdoor_big_stairs_down_data.pkl',  # DUPLICATE 1559\n\n    'P7/59_outdoor_rom/P7_59_outdoor_rom_data.pkl', # 1839\n    'P3/31_outdoor_workout/P3_31_outdoor_workout_data.pkl',  # 1216\n    'P3/33_outdoor_soccer_warmup_b/P3_33_outdoor_soccer_warmup_b_data.pkl',  # 1433\n    'P7/57_outdoor_rock_chair/P7_57_outdoor_rock_chair_data.pkl', # DUPLICATE 1558\n\n    'P3/32_outdoor_soccer_warmup_a/P3_32_outdoor_soccer_warmup_a_data.pkl',  # 1084\n    'P8/64_outdoor_skateboard/P8_64_outdoor_skateboard_data.pkl',   # DUPLICATE 1704\n    'P7/60_outdoor_workout/P7_60_outdoor_workout_data.pkl',  # 1693\n    'P6/50_outdoor_workout/P6_50_outdoor_workout_data.pkl', # 1532\n\n    'P8/68_outdoor_handstand/P8_68_outdoor_handstand_data.pkl',  # 1606\n    'P9/76_outdoor_sitting/P9_76_outdoor_sitting_data.pkl',  # 1768\n    'P1/14_outdoor_climb/P1_14_outdoor_climb_data.pkl',  # 1284\n    'P5/44_indoor_rom/P5_44_indoor_rom_data.pkl', # 1381\n]\nEMDB1_NAMES = [\"_\".join(p.split(\"/\")[:2]) for p in EMDB1_LIST]\n\nEMDB2_LIST = [\n    'P2/19_indoor_walk_off_mvs/P2_19_indoor_walk_off_mvs_data.pkl',  # 1299\n    'P3/29_outdoor_stairs_up/P3_29_outdoor_stairs_up_data.pkl', # 1205\n    'P4/35_indoor_walk/P4_35_indoor_walk_data.pkl',  # 1226\n    'P7/55_outdoor_walk/P7_55_outdoor_walk_data.pkl',  # 2179\n    'P9/80_outdoor_walk_big_circle/P9_80_outdoor_walk_big_circle_data.pkl',  # 2240\n    'P9/77_outdoor_stairs_up/P9_77_outdoor_stairs_up_data.pkl',  # DUPLICATE 728\n    'P9/79_outdoor_walk_rectangle/P9_79_outdoor_walk_rectangle_data.pkl',  # 1917\n\n    'P7/57_outdoor_rock_chair/P7_57_outdoor_rock_chair_data.pkl',  # DUPLICATE 1558\n    'P2/24_outdoor_long_walk/P2_24_outdoor_long_walk_data.pkl',  # 3280\n    'P3/30_outdoor_stairs_down/P3_30_outdoor_stairs_down_data.pkl',  #1137\n    'P4/36_outdoor_long_walk/P4_36_outdoor_long_walk_data.pkl',  # 2160\n    'P6/49_outdoor_big_stairs_down/P6_49_outdoor_big_stairs_down_data.pkl',   # DUPLICATE 1559\n    'P9/78_outdoor_stairs_up_down/P9_78_outdoor_stairs_up_down_data.pkl',  # 1083\n\n    'P7/56_outdoor_stairs_up_down/P7_56_outdoor_stairs_up_down_data.pkl', # 1120\n    'P2/20_outdoor_walk/P2_20_outdoor_walk_data.pkl',  # 2713\n    'P3/27_indoor_walk_off_mvs/P3_27_indoor_walk_off_mvs_data.pkl',  # 1448\n    'P4/37_outdoor_run_circle/P4_37_outdoor_run_circle_data.pkl', # 881\n    'P5/40_indoor_walk_big_circle/P5_40_indoor_walk_big_circle_data.pkl', # 2661\n    'P6/48_outdoor_walk_downhill/P6_48_outdoor_walk_downhill_data.pkl',  # 1959\n\n    'P0/09_outdoor_walk/P0_09_outdoor_walk_data.pkl',  # 2009\n    'P3/28_outdoor_walk_lunges/P3_28_outdoor_walk_lunges_data.pkl',  # 1836\n    'P7/58_outdoor_parcours/P7_58_outdoor_parcours_data.pkl',  # 1332\n    'P7/61_outdoor_sit_lie_walk/P7_61_outdoor_sit_lie_walk_data.pkl', # 1914\n    'P8/64_outdoor_skateboard/P8_64_outdoor_skateboard_data.pkl',   # DUPLICATE 1704\n    'P8/65_outdoor_walk_straight/P8_65_outdoor_walk_straight_data.pkl',  # 1981\n]\n\nEMDB2_NAMES = [\"_\".join(p.split(\"/\")[:2]) for p in EMDB2_LIST]\nEMDB_NAMES = {1: EMDB1_NAMES, 2: EMDB2_NAMES}\nEMDB_LIST = {1: EMDB1_LIST, 2: EMDB2_LIST}\n\n\ndef load_pkl(fp):\n    annot = pickle.load(open(fp, \"rb\"))\n    # ['gender', 'name', 'emdb1', 'emdb2', 'n_frames', 'good_frames_mask', 'camera', 'smpl', 'kp2d', 'bboxes', 'subfolder']\n    data = {}\n\n    F = annot[\"n_frames\"]\n    data[\"smpl_params\"] = {\n        \"smpl_body_pose\": annot[\"smpl\"][\"poses_body\"],  # (F, 69)\n        \"smpl_shape\": annot[\"smpl\"][\"betas\"][None].repeat(F, axis=0),  # (F, 10)\n        \"smpl_root_pose_w\": annot[\"smpl\"][\"poses_root\"],  # (F, 3)\n        \"smpl_transl_w\": annot[\"smpl\"][\"trans\"],  # (F, 3)\n    }\n\n    data[\"name\"] = annot[\"name\"].replace('_', '/', 1)\n    data[\"gender\"] = annot[\"gender\"]\n    data[\"mask\"] = annot[\"good_frames_mask\"]   # (L,)\n    data[\"K_fullimg\"] = annot[\"camera\"][\"intrinsics\"]  # (3, 3)\n    data[\"T_w2c\"] = annot[\"camera\"][\"extrinsics\"]  # (L, 4, 4)\n\n    return data\n\ndef to_tensor(ndarray):\n    tensor = torch.from_numpy(ndarray).float()\n    return tensor","source_hash":"2a695c1ada5890d3f4d5a97b6195f7c76acea26e923ca9a1fceeb91cc7860c14","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.data_utils.load_pkl","uri":"program://Human3R/function/eval.global_human.data_utils.load_pkl#L67-L86","kind":"function","name":"load_pkl","path":"eval/global_human/data_utils.py","language":"python","start_line":67,"end_line":86,"context_start_line":47,"context_end_line":90,"code":"    'P7/56_outdoor_stairs_up_down/P7_56_outdoor_stairs_up_down_data.pkl', # 1120\n    'P2/20_outdoor_walk/P2_20_outdoor_walk_data.pkl',  # 2713\n    'P3/27_indoor_walk_off_mvs/P3_27_indoor_walk_off_mvs_data.pkl',  # 1448\n    'P4/37_outdoor_run_circle/P4_37_outdoor_run_circle_data.pkl', # 881\n    'P5/40_indoor_walk_big_circle/P5_40_indoor_walk_big_circle_data.pkl', # 2661\n    'P6/48_outdoor_walk_downhill/P6_48_outdoor_walk_downhill_data.pkl',  # 1959\n\n    'P0/09_outdoor_walk/P0_09_outdoor_walk_data.pkl',  # 2009\n    'P3/28_outdoor_walk_lunges/P3_28_outdoor_walk_lunges_data.pkl',  # 1836\n    'P7/58_outdoor_parcours/P7_58_outdoor_parcours_data.pkl',  # 1332\n    'P7/61_outdoor_sit_lie_walk/P7_61_outdoor_sit_lie_walk_data.pkl', # 1914\n    'P8/64_outdoor_skateboard/P8_64_outdoor_skateboard_data.pkl',   # DUPLICATE 1704\n    'P8/65_outdoor_walk_straight/P8_65_outdoor_walk_straight_data.pkl',  # 1981\n]\n\nEMDB2_NAMES = [\"_\".join(p.split(\"/\")[:2]) for p in EMDB2_LIST]\nEMDB_NAMES = {1: EMDB1_NAMES, 2: EMDB2_NAMES}\nEMDB_LIST = {1: EMDB1_LIST, 2: EMDB2_LIST}\n\n\ndef load_pkl(fp):\n    annot = pickle.load(open(fp, \"rb\"))\n    # ['gender', 'name', 'emdb1', 'emdb2', 'n_frames', 'good_frames_mask', 'camera', 'smpl', 'kp2d', 'bboxes', 'subfolder']\n    data = {}\n\n    F = annot[\"n_frames\"]\n    data[\"smpl_params\"] = {\n        \"smpl_body_pose\": annot[\"smpl\"][\"poses_body\"],  # (F, 69)\n        \"smpl_shape\": annot[\"smpl\"][\"betas\"][None].repeat(F, axis=0),  # (F, 10)\n        \"smpl_root_pose_w\": annot[\"smpl\"][\"poses_root\"],  # (F, 3)\n        \"smpl_transl_w\": annot[\"smpl\"][\"trans\"],  # (F, 3)\n    }\n\n    data[\"name\"] = annot[\"name\"].replace('_', '/', 1)\n    data[\"gender\"] = annot[\"gender\"]\n    data[\"mask\"] = annot[\"good_frames_mask\"]   # (L,)\n    data[\"K_fullimg\"] = annot[\"camera\"][\"intrinsics\"]  # (3, 3)\n    data[\"T_w2c\"] = annot[\"camera\"][\"extrinsics\"]  # (L, 4, 4)\n\n    return data\n\ndef to_tensor(ndarray):\n    tensor = torch.from_numpy(ndarray).float()\n    return tensor","source_hash":"2a695c1ada5890d3f4d5a97b6195f7c76acea26e923ca9a1fceeb91cc7860c14","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.data_utils.to_tensor","uri":"program://Human3R/function/eval.global_human.data_utils.to_tensor#L88-L90","kind":"function","name":"to_tensor","path":"eval/global_human/data_utils.py","language":"python","start_line":88,"end_line":90,"context_start_line":68,"context_end_line":90,"code":"    annot = pickle.load(open(fp, \"rb\"))\n    # ['gender', 'name', 'emdb1', 'emdb2', 'n_frames', 'good_frames_mask', 'camera', 'smpl', 'kp2d', 'bboxes', 'subfolder']\n    data = {}\n\n    F = annot[\"n_frames\"]\n    data[\"smpl_params\"] = {\n        \"smpl_body_pose\": annot[\"smpl\"][\"poses_body\"],  # (F, 69)\n        \"smpl_shape\": annot[\"smpl\"][\"betas\"][None].repeat(F, axis=0),  # (F, 10)\n        \"smpl_root_pose_w\": annot[\"smpl\"][\"poses_root\"],  # (F, 3)\n        \"smpl_transl_w\": annot[\"smpl\"][\"trans\"],  # (F, 3)\n    }\n\n    data[\"name\"] = annot[\"name\"].replace('_', '/', 1)\n    data[\"gender\"] = annot[\"gender\"]\n    data[\"mask\"] = annot[\"good_frames_mask\"]   # (L,)\n    data[\"K_fullimg\"] = annot[\"camera\"][\"intrinsics\"]  # (3, 3)\n    data[\"T_w2c\"] = annot[\"camera\"][\"extrinsics\"]  # (L, 4, 4)\n\n    return data\n\ndef to_tensor(ndarray):\n    tensor = torch.from_numpy(ndarray).float()\n    return tensor","source_hash":"2a695c1ada5890d3f4d5a97b6195f7c76acea26e923ca9a1fceeb91cc7860c14","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils","uri":"program://Human3R/module/eval.global_human.utils#L1-L754","kind":"module","name":"eval.global_human.utils","path":"eval/global_human/utils.py","language":"python","start_line":1,"end_line":754,"context_start_line":1,"context_end_line":754,"code":"from copy import deepcopy\nimport cv2\n\nimport os\nimport re\nimport numpy as np\nimport torch\nimport roma\nimport math\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom eval.relpose.evo_utils import *\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\nfrom itertools import product\n\ndef todevice(batch, device, callback=None, non_blocking=False):\n    \"\"\"Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy).\n\n    batch: list, tuple, dict of tensors or other things\n    device: pytorch device or 'numpy'\n    callback: function that would be called on every sub-elements.\n    \"\"\"\n    if callback:\n        batch = callback(batch)\n\n    if isinstance(batch, dict):\n        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef compute_prf1(count, miss, fp):\n    \"\"\"\n    Code modified from https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/evaluation/RH_evaluation/evaluation.py#L90\n    \"\"\"\n    if count == 0:\n        return 0, 0, 0\n    all_tp = count - miss\n    all_fp = fp\n    all_fn = miss\n    if all_tp == 0:\n        return 0., 0., 0.\n    all_f1_score = round(all_tp / (all_tp + 0.5 * (all_fp + all_fn)), 2)\n    all_recall = round(all_tp / (all_tp + all_fn), 2)\n    all_precision = round(all_tp / (all_tp + all_fp), 2)\n    return 100. * all_precision, 100.* all_recall, 100. * all_f1_score\n\n\ndef match_2d_greedy(\n        pred_kps,\n        gtkp,\n        valid_mask,\n        imgPath=None,\n        baseline=None,\n        iou_thresh=0.05,\n        valid=None,\n        ind=-1):\n    '''\n    Code modified from: https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/trace2/evaluation/eval_3DPW.py#L232\n    matches groundtruth keypoints to the detection by considering all possible matchings.\n    :return: best possible matching, a list of tuples, where each tuple corresponds to one match of pred_person.to gt_person.\n            the order within one tuple is as follows (idx_pred_kps, idx_gt_kps)\n    '''\n    predList = np.arange(len(pred_kps))\n    gtList = np.arange(len(gtkp))\n    # get all pairs of elements in pred_kps, gtkp\n    # all combinations of 2 elements from l1 and l2\n    combs = list(product(predList, gtList))\n\n    errors_per_pair = {}\n    errors_per_pair_list = []\n    for comb in combs:\n        vmask = valid_mask[comb[1]]\n        assert vmask.sum()>0, print('no valid points')\n        errors_per_pair[str(comb)] = np.linalg.norm(pred_kps[comb[0]][vmask, :2] - gtkp[comb[1]][vmask, :2], 2)\n        errors_per_pair_list.append(errors_per_pair[str(comb)])\n\n    gtAssigned = np.zeros((len(gtkp),), dtype=bool)\n    opAssigned = np.zeros((len(pred_kps),), dtype=bool)\n    errors_per_pair_list = np.array(errors_per_pair_list)\n\n    bestMatch = []\n    excludedGtBecauseInvalid = []\n    falsePositiveCounter = 0\n    while np.sum(gtAssigned) < len(gtAssigned) and np.sum(\n            opAssigned) + falsePositiveCounter < len(pred_kps):\n        found = False\n        falsePositive = False\n        while not(found):\n            if sum(np.inf == errors_per_pair_list) == len(\n                    errors_per_pair_list):\n                print('something went wrong here')\n\n            minIdx = np.argmin(errors_per_pair_list)\n            minComb = combs[minIdx]\n            # compute IOU\n            iou = get_bbx_overlap(\n                pred_kps[minComb[0]], gtkp[minComb[1]]) #, imgPath, baseline)\n            # if neither prediction nor ground truth has been matched before and iou\n            # is larger than threshold\n            if not(opAssigned[minComb[0]]) and not(\n                    gtAssigned[minComb[1]]) and iou >= iou_thresh:\n                #print(imgPath + ': found matching')\n                found = True\n                errors_per_pair_list[minIdx] = np.inf\n            else:\n                errors_per_pair_list[minIdx] = np.inf\n                # if errors_per_pair_list[minIdx] >\n                # matching_threshold*headBboxs[combs[minIdx][1]]:\n                if iou < iou_thresh:\n                    #print(\n                    #   imgPath + ': false positive detected using threshold')\n                    found = True\n                    falsePositive = True\n                    falsePositiveCounter += 1\n\n        # if ground truth of combination is valid keep the match, else exclude\n        # gt from matching\n        if not(valid is None):\n            if valid[minComb[1]]:\n                if not falsePositive:\n                    bestMatch.append(minComb)\n                    opAssigned[minComb[0]] = True\n                    gtAssigned[minComb[1]] = True\n            else:\n                gtAssigned[minComb[1]] = True\n                excludedGtBecauseInvalid.append(minComb[1])\n\n        elif not falsePositive:\n            # same as above but without checking for valid\n            bestMatch.append(minComb)\n            opAssigned[minComb[0]] = True\n            gtAssigned[minComb[1]] = True\n\n    bestMatch = np.array(bestMatch)\n    # add false positives and false negatives to the matching\n    # find which elements have been successfully assigned\n    opAssigned = []\n    gtAssigned = []\n    for pair in bestMatch:\n        opAssigned.append(pair[0])\n        gtAssigned.append(pair[1])\n    opAssigned.sort()\n    gtAssigned.sort()\n\n    falsePositives = []\n    misses = []\n\n    # handle false positives\n    opIds = np.arange(len(pred_kps))\n    # returns values of oIds that are not in opAssigned\n    notAssignedIds = np.setdiff1d(opIds, opAssigned)\n    for notAssignedId in notAssignedIds:\n        falsePositives.append(notAssignedId)\n    gtIds = np.arange(len(gtList))\n    # returns values of gtIds that are not in gtAssigned\n    notAssignedIdsGt = np.setdiff1d(gtIds, gtAssigned)\n\n    # handle false negatives/misses\n    for notAssignedIdGt in notAssignedIdsGt:\n        if not(valid is None):  # if using the new matching\n            if valid[notAssignedIdGt]:\n                #print(imgPath + ': miss')\n                misses.append(notAssignedIdGt)\n            else:\n                excludedGtBecauseInvalid.append(notAssignedIdGt)\n        else:\n            #print(imgPath + ': miss')\n            misses.append(notAssignedIdGt)\n\n    return bestMatch, falsePositives, misses  # tuples are (idx_pred_kps, idx_gt_kps)\n\ndef get_bbx_overlap(p1, p2):\n    \"\"\"\n    Code modifed from https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/trace2/evaluation/eval_3DPW.py#L185\n    \"\"\"\n    min_p1 = np.min(p1, axis=0)\n    min_p2 = np.min(p2, axis=0)\n    max_p1 = np.max(p1, axis=0)\n    max_p2 = np.max(p2, axis=0)\n\n    bb1 = {}\n    bb2 = {}\n\n    bb1['x1'] = min_p1[0]\n    bb1['x2'] = max_p1[0]\n    bb1['y1'] = min_p1[1]\n    bb1['y2'] = max_p1[1]\n    bb2['x1'] = min_p2[0]\n    bb2['x2'] = max_p2[0]\n    bb2['y1'] = min_p2[1]\n    bb2['y2'] = max_p2[1]\n\n    assert bb1['x1'] < bb1['x2']\n    assert bb1['y1'] < bb1['y2']\n    assert bb2['x1'] < bb2['x2']\n    assert bb2['y1'] < bb2['y2']\n    # determine the coordinates of the intersection rectangle\n    x_left = max(bb1['x1'], bb2['x1'])\n    y_top = max(bb1['y1'], bb2['y1'])\n    x_right = min(bb1['x2'], bb2['x2'])\n    y_bottom = min(bb1['y2'], bb2['y2'])\n\n    # The intersection of two axis-aligned bounding boxes is always an\n    # axis-aligned bounding box\n    intersection_area = max(0, x_right - x_left + 1) * \\\n        max(0, y_bottom - y_top + 1)\n\n    # compute the area of both AABBs\n    bb1_area = (bb1['x2'] - bb1['x1'] + 1) * (bb1['y2'] - bb1['y1'] + 1)\n    bb2_area = (bb2['x2'] - bb2['x1'] + 1) * (bb2['y2'] - bb2['y1'] + 1)\n\n    # compute the intersection over union by taking the intersection\n    # area and dividing it by the sum of prediction + ground-truth\n    # areas - the interesection area\n    iou = intersection_area / float(bb1_area + bb2_area - intersection_area)\n\n    return iou\n\n    \ndef avg_per_human(lst, default=0.0):\n    values = [x for arr in lst for x in np.array(arr).flatten()]\n    return np.mean(values) if values else default\n\ndef extract_metrics(file_path):\n    with open(file_path, \"r\") as file:\n        content = file.read()\n\n    matches = re.findall(\n        r'n_human: (\\d+).*?PVE: ([\\d.]+), PA-PVE: ([\\d.]+), Metric-PVE: ([\\d.]+), MPJPE: ([\\d.]+), PA-MPJPE: ([\\d.]+), Metric-MPJPE: ([\\d.]+), RootError: ([\\d.]+), W-MPJPE: ([\\d.]+), WA-MPJPE: ([\\d.]+), RTE: ([\\d.]+), Scaled-RTE: ([\\d.]+), Jitter: ([\\d.]+), Foot-Sliding: ([\\d.]+), Precision: ([\\d.]+), Recall: ([\\d.]+), F1-Score: ([\\d.]+)',\n        content\n    )\n\n    if not matches:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    total_humans = 0\n    weighted_sums = [0.0] * 16\n    \n    for match in matches:\n        n_human = int(match[0])\n        metrics = [float(x) for x in match[1:]]\n        \n        total_humans += n_human\n        for i, metric in enumerate(metrics):\n            weighted_sums[i] += metric * n_human\n    \n    if total_humans == 0:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    averages = [ws / total_humans for ws in weighted_sums]\n    \n    return total_humans, *averages\n\n\ndef process_directory(directory):\n    results = []\n    for root, _, files in os.walk(directory):\n        for file in sorted(files):\n            if file.endswith(\".txt\"):\n                file_path = os.path.join(root, file)\n                result = extract_metrics(file_path)\n                if result[0] > 0:\n                    results.append(result)\n    return results\n\n\ndef calculate_averages(results):\n    if not results:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    total_humans = sum(r[0] for r in results)\n    if total_humans == 0:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    weighted_sums = [sum(r[i] * r[0] for r in results) for i in range(1, 17)]\n    averages = [ws / total_humans for ws in weighted_sums]\n    \n    return total_humans, *averages\n\ndef visualize(\n    save_dir, img_path, view, gt_v3d_c, pred_v3d_c, \n    K_to_proj, gt_K, bestMatch, smpl_face\n):\n    from dust3r.utils import render_meshes, denormalize_rgb\n    from viser_utils import get_color\n    from PIL import Image\n    \n    os.makedirs(save_dir, exist_ok=True)\n    n_humans_i = pred_v3d_c.shape[0]\n\n    # image\n    img_array = denormalize_rgb(view.cpu().numpy())\n\n    focal = K_to_proj[[0,1],[0,1]].cpu().numpy()\n    princpt = K_to_proj[[0,1],[-1,-1]].cpu().numpy()\n    gt_focal = gt_K[[0,1],[0,1]].cpu().numpy()\n    gt_princpt = gt_K[[0,1],[-1,-1]].cpu().numpy()\n\n    gt_color_indices = list(range(gt_v3d_c.shape[0]))\n    pred_color_indices = [-1] * n_humans_i\n    for (pid, gid) in bestMatch:\n        pred_color_indices[pid] = gt_color_indices[gid]\n\n    next_color_idx = max(gt_color_indices) + 1 if gt_color_indices else 0\n    for i in range(len(pred_color_indices)):\n        if pred_color_indices[i] == -1:\n            pred_color_indices[i] = next_color_idx\n            next_color_idx += 1\n\n    # gt\n    gt_verts, gt_faces = [], []\n    for j in range(gt_v3d_c.shape[0]):\n        gt_verts.append(gt_v3d_c[j].cpu().numpy().reshape(-1,3))\n        gt_faces.append(smpl_face)\n\n    gt_colors = [get_color(gt_color_indices[j])/255 for j in range(len(gt_verts))]\n    gt_rend_array = render_meshes(img_array.copy(), \n                                    gt_verts, \n                                    gt_faces,\n                                    {'focal': gt_focal, 'princpt': gt_princpt},\n                                    color=gt_colors)\n    \n    # pred\n    pred_verts, pred_faces = [], []\n    for j in range(n_humans_i):\n        pred_verts.append(pred_v3d_c[j].cpu().numpy().reshape(-1,3))\n        pred_faces.append(smpl_face)\n    \n    pred_colors = [get_color(pred_color_indices[j])/255 for j in range(len(pred_verts))]\n    pred_rend_array = render_meshes(img_array.copy(), \n                                    pred_verts, \n                                    pred_faces,\n                                    {'focal': focal, 'princpt': princpt},\n                                    color=pred_colors)\n\n    img = np.concatenate([img_array, pred_rend_array, gt_rend_array], 1)\n    Image.fromarray(img).save(\n        os.path.join(\n            f\"{save_dir}/{os.path.splitext(os.path.basename(img_path))[0]}.jpg\"\n            ))\n\ndef write_log(log_path, dataset, seq, counter, metrics):\n    with open(log_path, \"a\") as f:\n        f.write(\n            f\"{dataset}-{seq: <16} |  \"\n            f\"n_human: {counter['n_human']:06d} | \"\n            f\"PVE: {avg_per_human(metrics['ca_pve']):.1f}, \"\n            f\"PA-PVE: {avg_per_human(metrics['pa_pve']):.1f}, \"\n            f\"Metric-PVE: {avg_per_human(metrics['me_pve']):.1f}, \"\n            f\"MPJPE: {avg_per_human(metrics['ca_mpjpe']):.1f}, \"\n            f\"PA-MPJPE: {avg_per_human(metrics['pa_mpjpe']):.1f}, \"\n            f\"Metric-MPJPE: {avg_per_human(metrics['me_mpjpe']):.1f}, \"\n            f\"RootError: {avg_per_human(metrics['rt_error']):.1f}, \"\n            f\"W-MPJPE: {avg_per_human(metrics['wa2_mpjpe']):.1f}, \"\n            f\"WA-MPJPE: {avg_per_human(metrics['waa_mpjpe']):.1f}, \"\n            f\"RTE: {avg_per_human(metrics['rte']):.1f}, \"\n            f\"Scaled-RTE: {avg_per_human(metrics['rte_scaled']):.1f}, \"\n            f\"Jitter: {avg_per_human(metrics['jitter']):.1f}, \"\n            f\"Foot-Sliding: {avg_per_human(metrics['fs']):.1f}, \"\n            f\"Precision: {metrics['precision']:.1f}, \"\n            f\"Recall: {metrics['recall']:.1f}, \"\n            f\"F1-Score: {metrics['f1_score']:.1f}\\n\"\n        )\n\ndef get_summary_log(summary):\n    \"\"\"Generate summary log for evaluation results\"\"\"\n    return (\n        f\"EVALUATION SUMMARY\\n\"\n        f\"{'='*7}EVALUATION SUMMARY{'='*7}\\n\"\n        f\"Total Humans: {summary[0]}\\n\"\n        f\"\\n\"\n        f\"Camera Coordinate Metrics (mm):\\n\"\n        f\"  PVE:         {summary[1]:6.1f}\\n\"\n        f\"  PA-PVE:      {summary[2]:6.1f}\\n\"\n        f\"  Metric-PVE:  {summary[3]:6.1f}\\n\"\n        f\"\\n\"\n        f\"  MPJPE:       {summary[4]:6.1f}\\n\"\n        f\"  PA-MPJPE:    {summary[5]:6.1f}\\n\"\n        f\"  Metric-MPJPE: {summary[6]:6.1f}\\n\"\n        f\"  Root-Error:  {summary[7]:6.1f}\\n\"\n        f\"\\n\"\n        f\"Global Coordinate Metrics (cm):\\n\"\n        f\"  W-MPJPE:     {summary[8]:6.1f}\\n\"\n        f\"  WA-MPJPE:    {summary[9]:6.1f}\\n\"\n        f\"  RTE:         {summary[10]:6.1f}\\n\"\n        f\"  Scaled-RTE:  {summary[11]:6.1f}\\n\"\n        f\"  Jitter:      {summary[12]:6.1f}\\n\"\n        f\"  Foot-Sliding: {summary[13]:6.1f}\\n\"\n        f\"\\n\"\n        f\"Detection Metrics (%):\\n\"\n        f\"  Precision:   {summary[14]:6.1f}\\n\"\n        f\"  Recall:      {summary[15]:6.1f}\\n\"\n        f\"  F1-Score:    {summary[16]:6.1f}\\n\"\n        f\"{'='*32}\\n\"\n    )\n    \n\n# Evaluation metrics\n# Code modifed from https://github.com/zju3dv/GVHMR/blob/088caff492aa38c2d82cea363b78a3c65a83118f/hmr4d/utils/eval/eval_utils.py\n\ndef compute_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).mean(dim=-1).numpy()\n\ndef compute_perjoint_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).numpy()\n\ndef batch_compute_similarity_transform_torch(S1, S2):\n    '''\n    Computes a similarity transform (sR, t) that takes\n    a set of 3D points S1 (3 x N) closest to a set of 3D points S2,\n    where R is an 3x3 rotation matrix, t 3x1 translation, s scale.\n    i.e. solves the orthogonal Procrutes problem.\n    '''\n    S1 = S1.permute(0,2,1)\n    S2 = S2.permute(0,2,1)\n    transposed = True\n\n    # 1. Remove mean.\n    mu1 = S1.mean(axis=-1, keepdims=True)\n    mu2 = S2.mean(axis=-1, keepdims=True)\n\n    X1 = S1 - mu1\n    X2 = S2 - mu2\n\n    # 2. Compute variance of X1 used for scale.\n    var1 = torch.sum(X1**2, dim=1).sum(dim=1)\n\n    # 3. The outer product of X1 and X2.\n    K = X1.bmm(X2.permute(0,2,1))\n\n    # 4. Solution that Maximizes trace(R'K) is R=U*V', where U, V are\n    # singular vectors of K.\n    U, s, V = torch.svd(K)\n\n    # Construct Z that fixes the orientation of R to get det(R)=1.\n    Z = torch.eye(U.shape[1], device=S1.device).unsqueeze(0)\n    Z = Z.repeat(U.shape[0],1,1)\n    Z[:,-1, -1] *= torch.sign(torch.det(U.bmm(V.permute(0,2,1))))\n\n    # Construct R.\n    R = V.bmm(Z.bmm(U.permute(0,2,1)))\n\n    # 5. Recover scale.\n    scale = torch.cat([torch.trace(x).unsqueeze(0) for x in R.bmm(K)]) / var1\n\n    # 6. Recover translation.\n    t = mu2 - (scale.unsqueeze(-1).unsqueeze(-1) * (R.bmm(mu1)))\n\n    # 7. Error:\n    S1_hat = scale.unsqueeze(-1).unsqueeze(-1) * R.bmm(S1) + t\n\n    if transposed:\n        S1_hat = S1_hat.permute(0,2,1)\n\n    return S1_hat\n\ndef batch_align_by_pelvis(data_list, pelvis_idxs):\n    \"\"\"\n    Assumes data is given as [pred_j3d, target_j3d, pred_verts, target_verts].\n    Each data is in shape of (batch, num_points, 3)\n    Pelvis is notated as one / two joints indices.\n    Align all data to the corresponding pelvis location.\n    \"\"\"\n\n    pred_j3d, target_j3d, pred_verts, target_verts = data_list\n    \n    pred_pelvis = pred_j3d[:, pelvis_idxs].mean(dim=1, keepdims=True).clone()\n    target_pelvis = target_j3d[:, pelvis_idxs].mean(dim=1, keepdims=True).clone()\n    \n    # Align to the pelvis\n    pred_j3d = pred_j3d - pred_pelvis\n    target_j3d = target_j3d - target_pelvis\n    pred_verts = pred_verts - pred_pelvis\n    target_verts = target_verts - target_pelvis\n    \n    return (pred_j3d, target_j3d, pred_verts, target_verts, pred_pelvis, target_pelvis)\n\ndef align_pcl(Y, X, weight=None, fixed_scale=False):\n    \"\"\"\n    align similarity transform to align X with Y using umeyama method\n    X' = s * R * X + t is aligned with Y\n    :param Y (*, N, 3) first trajectory\n    :param X (*, N, 3) second trajectory\n    :param weight (*, N, 1) optional weight of valid correspondences\n    :returns s (*, 1), R (*, 3, 3), t (*, 3)\n    \"\"\"\n    *dims, N, _ = Y.shape\n    N = torch.ones(*dims, 1, 1) * N\n\n    if weight is not None:\n        Y = Y * weight\n        X = X * weight\n        N = weight.sum(dim=-2, keepdim=True)  # (*, 1, 1)\n\n    # subtract mean\n    my = Y.sum(dim=-2) / N[..., 0]  # (*, 3)\n    mx = X.sum(dim=-2) / N[..., 0]\n    y0 = Y - my[..., None, :]  # (*, N, 3)\n    x0 = X - mx[..., None, :]\n\n    if weight is not None:\n        y0 = y0 * weight\n        x0 = x0 * weight\n\n    # correlation\n    C = torch.matmul(y0.transpose(-1, -2), x0) / N  # (*, 3, 3)\n    U, D, Vh = torch.linalg.svd(C)  # (*, 3, 3), (*, 3), (*, 3, 3)\n\n    S = torch.eye(3).reshape(*(1,) * (len(dims)), 3, 3).repeat(*dims, 1, 1)\n    neg = torch.det(U) * torch.det(Vh.transpose(-1, -2)) < 0\n    S[neg, 2, 2] = -1\n\n    R = torch.matmul(U, torch.matmul(S, Vh))  # (*, 3, 3)\n\n    D = torch.diag_embed(D)  # (*, 3, 3)\n    if fixed_scale:\n        s = torch.ones(*dims, 1, device=Y.device, dtype=torch.float32)\n    else:\n        var = torch.sum(torch.square(x0), dim=(-1, -2), keepdim=True) / N  # (*, 1, 1)\n        s = torch.diagonal(torch.matmul(D, S), dim1=-2, dim2=-1).sum(dim=-1, keepdim=True) / var[..., 0]  # (*, 1)\n\n    t = my - s * torch.matmul(R, mx[..., None])[..., 0]  # (*, 3)\n\n    return s, R, t\n\ndef global_align_joints(gt_joints, pred_joints):\n    \"\"\"\n    :param gt_joints (T, J, 3)\n    :param pred_joints (T, J, 3)\n    \"\"\"\n    s_glob, R_glob, t_glob = align_pcl(gt_joints.reshape(-1, 3), pred_joints.reshape(-1, 3))\n    pred_glob = s_glob * torch.einsum(\"ij,tnj->tn\n# ... truncated ...","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":true}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.todevice","uri":"program://Human3R/function/eval.global_human.utils.todevice#L22-L47","kind":"function","name":"todevice","path":"eval/global_human/utils.py","language":"python","start_line":22,"end_line":47,"context_start_line":2,"context_end_line":67,"code":"import cv2\n\nimport os\nimport re\nimport numpy as np\nimport torch\nimport roma\nimport math\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom eval.relpose.evo_utils import *\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\nfrom itertools import product\n\ndef todevice(batch, device, callback=None, non_blocking=False):\n    \"\"\"Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy).\n\n    batch: list, tuple, dict of tensors or other things\n    device: pytorch device or 'numpy'\n    callback: function that would be called on every sub-elements.\n    \"\"\"\n    if callback:\n        batch = callback(batch)\n\n    if isinstance(batch, dict):\n        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef compute_prf1(count, miss, fp):\n    \"\"\"\n    Code modified from https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/evaluation/RH_evaluation/evaluation.py#L90\n    \"\"\"\n    if count == 0:\n        return 0, 0, 0\n    all_tp = count - miss\n    all_fp = fp\n    all_fn = miss\n    if all_tp == 0:\n        return 0., 0., 0.","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.to_numpy","uri":"program://Human3R/function/eval.global_human.utils.to_numpy#L53-L54","kind":"function","name":"to_numpy","path":"eval/global_human/utils.py","language":"python","start_line":53,"end_line":54,"context_start_line":33,"context_end_line":74,"code":"        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef compute_prf1(count, miss, fp):\n    \"\"\"\n    Code modified from https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/evaluation/RH_evaluation/evaluation.py#L90\n    \"\"\"\n    if count == 0:\n        return 0, 0, 0\n    all_tp = count - miss\n    all_fp = fp\n    all_fn = miss\n    if all_tp == 0:\n        return 0., 0., 0.\n    all_f1_score = round(all_tp / (all_tp + 0.5 * (all_fp + all_fn)), 2)\n    all_recall = round(all_tp / (all_tp + all_fn), 2)\n    all_precision = round(all_tp / (all_tp + all_fp), 2)\n    return 100. * all_precision, 100.* all_recall, 100. * all_f1_score\n\n\ndef match_2d_greedy(","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.compute_prf1","uri":"program://Human3R/function/eval.global_human.utils.compute_prf1#L57-L71","kind":"function","name":"compute_prf1","path":"eval/global_human/utils.py","language":"python","start_line":57,"end_line":71,"context_start_line":37,"context_end_line":91,"code":"\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef compute_prf1(count, miss, fp):\n    \"\"\"\n    Code modified from https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/evaluation/RH_evaluation/evaluation.py#L90\n    \"\"\"\n    if count == 0:\n        return 0, 0, 0\n    all_tp = count - miss\n    all_fp = fp\n    all_fn = miss\n    if all_tp == 0:\n        return 0., 0., 0.\n    all_f1_score = round(all_tp / (all_tp + 0.5 * (all_fp + all_fn)), 2)\n    all_recall = round(all_tp / (all_tp + all_fn), 2)\n    all_precision = round(all_tp / (all_tp + all_fp), 2)\n    return 100. * all_precision, 100.* all_recall, 100. * all_f1_score\n\n\ndef match_2d_greedy(\n        pred_kps,\n        gtkp,\n        valid_mask,\n        imgPath=None,\n        baseline=None,\n        iou_thresh=0.05,\n        valid=None,\n        ind=-1):\n    '''\n    Code modified from: https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/trace2/evaluation/eval_3DPW.py#L232\n    matches groundtruth keypoints to the detection by considering all possible matchings.\n    :return: best possible matching, a list of tuples, where each tuple corresponds to one match of pred_person.to gt_person.\n            the order within one tuple is as follows (idx_pred_kps, idx_gt_kps)\n    '''\n    predList = np.arange(len(pred_kps))\n    gtList = np.arange(len(gtkp))\n    # get all pairs of elements in pred_kps, gtkp","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.match_2d_greedy","uri":"program://Human3R/function/eval.global_human.utils.match_2d_greedy#L74-L196","kind":"function","name":"match_2d_greedy","path":"eval/global_human/utils.py","language":"python","start_line":74,"end_line":196,"context_start_line":54,"context_end_line":216,"code":"    return todevice(x, \"numpy\")\n\n\ndef compute_prf1(count, miss, fp):\n    \"\"\"\n    Code modified from https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/evaluation/RH_evaluation/evaluation.py#L90\n    \"\"\"\n    if count == 0:\n        return 0, 0, 0\n    all_tp = count - miss\n    all_fp = fp\n    all_fn = miss\n    if all_tp == 0:\n        return 0., 0., 0.\n    all_f1_score = round(all_tp / (all_tp + 0.5 * (all_fp + all_fn)), 2)\n    all_recall = round(all_tp / (all_tp + all_fn), 2)\n    all_precision = round(all_tp / (all_tp + all_fp), 2)\n    return 100. * all_precision, 100.* all_recall, 100. * all_f1_score\n\n\ndef match_2d_greedy(\n        pred_kps,\n        gtkp,\n        valid_mask,\n        imgPath=None,\n        baseline=None,\n        iou_thresh=0.05,\n        valid=None,\n        ind=-1):\n    '''\n    Code modified from: https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/trace2/evaluation/eval_3DPW.py#L232\n    matches groundtruth keypoints to the detection by considering all possible matchings.\n    :return: best possible matching, a list of tuples, where each tuple corresponds to one match of pred_person.to gt_person.\n            the order within one tuple is as follows (idx_pred_kps, idx_gt_kps)\n    '''\n    predList = np.arange(len(pred_kps))\n    gtList = np.arange(len(gtkp))\n    # get all pairs of elements in pred_kps, gtkp\n    # all combinations of 2 elements from l1 and l2\n    combs = list(product(predList, gtList))\n\n    errors_per_pair = {}\n    errors_per_pair_list = []\n    for comb in combs:\n        vmask = valid_mask[comb[1]]\n        assert vmask.sum()>0, print('no valid points')\n        errors_per_pair[str(comb)] = np.linalg.norm(pred_kps[comb[0]][vmask, :2] - gtkp[comb[1]][vmask, :2], 2)\n        errors_per_pair_list.append(errors_per_pair[str(comb)])\n\n    gtAssigned = np.zeros((len(gtkp),), dtype=bool)\n    opAssigned = np.zeros((len(pred_kps),), dtype=bool)\n    errors_per_pair_list = np.array(errors_per_pair_list)\n\n    bestMatch = []\n    excludedGtBecauseInvalid = []\n    falsePositiveCounter = 0\n    while np.sum(gtAssigned) < len(gtAssigned) and np.sum(\n            opAssigned) + falsePositiveCounter < len(pred_kps):\n        found = False\n        falsePositive = False\n        while not(found):\n            if sum(np.inf == errors_per_pair_list) == len(\n                    errors_per_pair_list):\n                print('something went wrong here')\n\n            minIdx = np.argmin(errors_per_pair_list)\n            minComb = combs[minIdx]\n            # compute IOU\n            iou = get_bbx_overlap(\n                pred_kps[minComb[0]], gtkp[minComb[1]]) #, imgPath, baseline)\n            # if neither prediction nor ground truth has been matched before and iou\n            # is larger than threshold\n            if not(opAssigned[minComb[0]]) and not(\n                    gtAssigned[minComb[1]]) and iou >= iou_thresh:\n                #print(imgPath + ': found matching')\n                found = True\n                errors_per_pair_list[minIdx] = np.inf\n            else:\n                errors_per_pair_list[minIdx] = np.inf\n                # if errors_per_pair_list[minIdx] >\n                # matching_threshold*headBboxs[combs[minIdx][1]]:\n                if iou < iou_thresh:\n                    #print(\n                    #   imgPath + ': false positive detected using threshold')\n                    found = True\n                    falsePositive = True\n                    falsePositiveCounter += 1\n\n        # if ground truth of combination is valid keep the match, else exclude\n        # gt from matching\n        if not(valid is None):\n            if valid[minComb[1]]:\n                if not falsePositive:\n                    bestMatch.append(minComb)\n                    opAssigned[minComb[0]] = True\n                    gtAssigned[minComb[1]] = True\n            else:\n                gtAssigned[minComb[1]] = True\n                excludedGtBecauseInvalid.append(minComb[1])\n\n        elif not falsePositive:\n            # same as above but without checking for valid\n            bestMatch.append(minComb)\n            opAssigned[minComb[0]] = True\n            gtAssigned[minComb[1]] = True\n\n    bestMatch = np.array(bestMatch)\n    # add false positives and false negatives to the matching\n    # find which elements have been successfully assigned\n    opAssigned = []\n    gtAssigned = []\n    for pair in bestMatch:\n        opAssigned.append(pair[0])\n        gtAssigned.append(pair[1])\n    opAssigned.sort()\n    gtAssigned.sort()\n\n    falsePositives = []\n    misses = []\n\n    # handle false positives\n    opIds = np.arange(len(pred_kps))\n    # returns values of oIds that are not in opAssigned\n    notAssignedIds = np.setdiff1d(opIds, opAssigned)\n    for notAssignedId in notAssignedIds:\n        falsePositives.append(notAssignedId)\n    gtIds = np.arange(len(gtList))\n    # returns values of gtIds that are not in gtAssigned\n    notAssignedIdsGt = np.setdiff1d(gtIds, gtAssigned)\n\n    # handle false negatives/misses\n    for notAssignedIdGt in notAssignedIdsGt:\n        if not(valid is None):  # if using the new matching\n            if valid[notAssignedIdGt]:\n                #print(imgPath + ': miss')\n                misses.append(notAssignedIdGt)\n            else:\n                excludedGtBecauseInvalid.append(notAssignedIdGt)\n        else:\n            #print(imgPath + ': miss')\n            misses.append(notAssignedIdGt)\n\n    return bestMatch, falsePositives, misses  # tuples are (idx_pred_kps, idx_gt_kps)\n\ndef get_bbx_overlap(p1, p2):\n    \"\"\"\n    Code modifed from https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/trace2/evaluation/eval_3DPW.py#L185\n    \"\"\"\n    min_p1 = np.min(p1, axis=0)\n    min_p2 = np.min(p2, axis=0)\n    max_p1 = np.max(p1, axis=0)\n    max_p2 = np.max(p2, axis=0)\n\n    bb1 = {}\n    bb2 = {}\n\n    bb1['x1'] = min_p1[0]\n    bb1['x2'] = max_p1[0]\n    bb1['y1'] = min_p1[1]\n    bb1['y2'] = max_p1[1]\n    bb2['x1'] = min_p2[0]\n    bb2['x2'] = max_p2[0]\n    bb2['y1'] = min_p2[1]","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.get_bbx_overlap","uri":"program://Human3R/function/eval.global_human.utils.get_bbx_overlap#L198-L243","kind":"function","name":"get_bbx_overlap","path":"eval/global_human/utils.py","language":"python","start_line":198,"end_line":243,"context_start_line":178,"context_end_line":263,"code":"    for notAssignedId in notAssignedIds:\n        falsePositives.append(notAssignedId)\n    gtIds = np.arange(len(gtList))\n    # returns values of gtIds that are not in gtAssigned\n    notAssignedIdsGt = np.setdiff1d(gtIds, gtAssigned)\n\n    # handle false negatives/misses\n    for notAssignedIdGt in notAssignedIdsGt:\n        if not(valid is None):  # if using the new matching\n            if valid[notAssignedIdGt]:\n                #print(imgPath + ': miss')\n                misses.append(notAssignedIdGt)\n            else:\n                excludedGtBecauseInvalid.append(notAssignedIdGt)\n        else:\n            #print(imgPath + ': miss')\n            misses.append(notAssignedIdGt)\n\n    return bestMatch, falsePositives, misses  # tuples are (idx_pred_kps, idx_gt_kps)\n\ndef get_bbx_overlap(p1, p2):\n    \"\"\"\n    Code modifed from https://github.com/Arthur151/ROMP/blob/4eebd3647f57d291d26423e51f0d514ff7197cb3/simple_romp/trace2/evaluation/eval_3DPW.py#L185\n    \"\"\"\n    min_p1 = np.min(p1, axis=0)\n    min_p2 = np.min(p2, axis=0)\n    max_p1 = np.max(p1, axis=0)\n    max_p2 = np.max(p2, axis=0)\n\n    bb1 = {}\n    bb2 = {}\n\n    bb1['x1'] = min_p1[0]\n    bb1['x2'] = max_p1[0]\n    bb1['y1'] = min_p1[1]\n    bb1['y2'] = max_p1[1]\n    bb2['x1'] = min_p2[0]\n    bb2['x2'] = max_p2[0]\n    bb2['y1'] = min_p2[1]\n    bb2['y2'] = max_p2[1]\n\n    assert bb1['x1'] < bb1['x2']\n    assert bb1['y1'] < bb1['y2']\n    assert bb2['x1'] < bb2['x2']\n    assert bb2['y1'] < bb2['y2']\n    # determine the coordinates of the intersection rectangle\n    x_left = max(bb1['x1'], bb2['x1'])\n    y_top = max(bb1['y1'], bb2['y1'])\n    x_right = min(bb1['x2'], bb2['x2'])\n    y_bottom = min(bb1['y2'], bb2['y2'])\n\n    # The intersection of two axis-aligned bounding boxes is always an\n    # axis-aligned bounding box\n    intersection_area = max(0, x_right - x_left + 1) * \\\n        max(0, y_bottom - y_top + 1)\n\n    # compute the area of both AABBs\n    bb1_area = (bb1['x2'] - bb1['x1'] + 1) * (bb1['y2'] - bb1['y1'] + 1)\n    bb2_area = (bb2['x2'] - bb2['x1'] + 1) * (bb2['y2'] - bb2['y1'] + 1)\n\n    # compute the intersection over union by taking the intersection\n    # area and dividing it by the sum of prediction + ground-truth\n    # areas - the interesection area\n    iou = intersection_area / float(bb1_area + bb2_area - intersection_area)\n\n    return iou\n\n    \ndef avg_per_human(lst, default=0.0):\n    values = [x for arr in lst for x in np.array(arr).flatten()]\n    return np.mean(values) if values else default\n\ndef extract_metrics(file_path):\n    with open(file_path, \"r\") as file:\n        content = file.read()\n\n    matches = re.findall(\n        r'n_human: (\\d+).*?PVE: ([\\d.]+), PA-PVE: ([\\d.]+), Metric-PVE: ([\\d.]+), MPJPE: ([\\d.]+), PA-MPJPE: ([\\d.]+), Metric-MPJPE: ([\\d.]+), RootError: ([\\d.]+), W-MPJPE: ([\\d.]+), WA-MPJPE: ([\\d.]+), RTE: ([\\d.]+), Scaled-RTE: ([\\d.]+), Jitter: ([\\d.]+), Foot-Sliding: ([\\d.]+), Precision: ([\\d.]+), Recall: ([\\d.]+), F1-Score: ([\\d.]+)',\n        content\n    )\n\n    if not matches:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    total_humans = 0\n    weighted_sums = [0.0] * 16","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.avg_per_human","uri":"program://Human3R/function/eval.global_human.utils.avg_per_human#L246-L248","kind":"function","name":"avg_per_human","path":"eval/global_human/utils.py","language":"python","start_line":246,"end_line":248,"context_start_line":226,"context_end_line":268,"code":"    x_right = min(bb1['x2'], bb2['x2'])\n    y_bottom = min(bb1['y2'], bb2['y2'])\n\n    # The intersection of two axis-aligned bounding boxes is always an\n    # axis-aligned bounding box\n    intersection_area = max(0, x_right - x_left + 1) * \\\n        max(0, y_bottom - y_top + 1)\n\n    # compute the area of both AABBs\n    bb1_area = (bb1['x2'] - bb1['x1'] + 1) * (bb1['y2'] - bb1['y1'] + 1)\n    bb2_area = (bb2['x2'] - bb2['x1'] + 1) * (bb2['y2'] - bb2['y1'] + 1)\n\n    # compute the intersection over union by taking the intersection\n    # area and dividing it by the sum of prediction + ground-truth\n    # areas - the interesection area\n    iou = intersection_area / float(bb1_area + bb2_area - intersection_area)\n\n    return iou\n\n    \ndef avg_per_human(lst, default=0.0):\n    values = [x for arr in lst for x in np.array(arr).flatten()]\n    return np.mean(values) if values else default\n\ndef extract_metrics(file_path):\n    with open(file_path, \"r\") as file:\n        content = file.read()\n\n    matches = re.findall(\n        r'n_human: (\\d+).*?PVE: ([\\d.]+), PA-PVE: ([\\d.]+), Metric-PVE: ([\\d.]+), MPJPE: ([\\d.]+), PA-MPJPE: ([\\d.]+), Metric-MPJPE: ([\\d.]+), RootError: ([\\d.]+), W-MPJPE: ([\\d.]+), WA-MPJPE: ([\\d.]+), RTE: ([\\d.]+), Scaled-RTE: ([\\d.]+), Jitter: ([\\d.]+), Foot-Sliding: ([\\d.]+), Precision: ([\\d.]+), Recall: ([\\d.]+), F1-Score: ([\\d.]+)',\n        content\n    )\n\n    if not matches:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    total_humans = 0\n    weighted_sums = [0.0] * 16\n    \n    for match in matches:\n        n_human = int(match[0])\n        metrics = [float(x) for x in match[1:]]\n        ","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.extract_metrics","uri":"program://Human3R/function/eval.global_human.utils.extract_metrics#L250-L278","kind":"function","name":"extract_metrics","path":"eval/global_human/utils.py","language":"python","start_line":250,"end_line":278,"context_start_line":230,"context_end_line":298,"code":"    # axis-aligned bounding box\n    intersection_area = max(0, x_right - x_left + 1) * \\\n        max(0, y_bottom - y_top + 1)\n\n    # compute the area of both AABBs\n    bb1_area = (bb1['x2'] - bb1['x1'] + 1) * (bb1['y2'] - bb1['y1'] + 1)\n    bb2_area = (bb2['x2'] - bb2['x1'] + 1) * (bb2['y2'] - bb2['y1'] + 1)\n\n    # compute the intersection over union by taking the intersection\n    # area and dividing it by the sum of prediction + ground-truth\n    # areas - the interesection area\n    iou = intersection_area / float(bb1_area + bb2_area - intersection_area)\n\n    return iou\n\n    \ndef avg_per_human(lst, default=0.0):\n    values = [x for arr in lst for x in np.array(arr).flatten()]\n    return np.mean(values) if values else default\n\ndef extract_metrics(file_path):\n    with open(file_path, \"r\") as file:\n        content = file.read()\n\n    matches = re.findall(\n        r'n_human: (\\d+).*?PVE: ([\\d.]+), PA-PVE: ([\\d.]+), Metric-PVE: ([\\d.]+), MPJPE: ([\\d.]+), PA-MPJPE: ([\\d.]+), Metric-MPJPE: ([\\d.]+), RootError: ([\\d.]+), W-MPJPE: ([\\d.]+), WA-MPJPE: ([\\d.]+), RTE: ([\\d.]+), Scaled-RTE: ([\\d.]+), Jitter: ([\\d.]+), Foot-Sliding: ([\\d.]+), Precision: ([\\d.]+), Recall: ([\\d.]+), F1-Score: ([\\d.]+)',\n        content\n    )\n\n    if not matches:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    total_humans = 0\n    weighted_sums = [0.0] * 16\n    \n    for match in matches:\n        n_human = int(match[0])\n        metrics = [float(x) for x in match[1:]]\n        \n        total_humans += n_human\n        for i, metric in enumerate(metrics):\n            weighted_sums[i] += metric * n_human\n    \n    if total_humans == 0:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    averages = [ws / total_humans for ws in weighted_sums]\n    \n    return total_humans, *averages\n\n\ndef process_directory(directory):\n    results = []\n    for root, _, files in os.walk(directory):\n        for file in sorted(files):\n            if file.endswith(\".txt\"):\n                file_path = os.path.join(root, file)\n                result = extract_metrics(file_path)\n                if result[0] > 0:\n                    results.append(result)\n    return results\n\n\ndef calculate_averages(results):\n    if not results:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    total_humans = sum(r[0] for r in results)\n    if total_humans == 0:","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.process_directory","uri":"program://Human3R/function/eval.global_human.utils.process_directory#L281-L290","kind":"function","name":"process_directory","path":"eval/global_human/utils.py","language":"python","start_line":281,"end_line":290,"context_start_line":261,"context_end_line":310,"code":"    \n    total_humans = 0\n    weighted_sums = [0.0] * 16\n    \n    for match in matches:\n        n_human = int(match[0])\n        metrics = [float(x) for x in match[1:]]\n        \n        total_humans += n_human\n        for i, metric in enumerate(metrics):\n            weighted_sums[i] += metric * n_human\n    \n    if total_humans == 0:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    averages = [ws / total_humans for ws in weighted_sums]\n    \n    return total_humans, *averages\n\n\ndef process_directory(directory):\n    results = []\n    for root, _, files in os.walk(directory):\n        for file in sorted(files):\n            if file.endswith(\".txt\"):\n                file_path = os.path.join(root, file)\n                result = extract_metrics(file_path)\n                if result[0] > 0:\n                    results.append(result)\n    return results\n\n\ndef calculate_averages(results):\n    if not results:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    total_humans = sum(r[0] for r in results)\n    if total_humans == 0:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    weighted_sums = [sum(r[i] * r[0] for r in results) for i in range(1, 17)]\n    averages = [ws / total_humans for ws in weighted_sums]\n    \n    return total_humans, *averages\n\ndef visualize(\n    save_dir, img_path, view, gt_v3d_c, pred_v3d_c, \n    K_to_proj, gt_K, bestMatch, smpl_face\n):\n    from dust3r.utils import render_meshes, denormalize_rgb","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.calculate_averages","uri":"program://Human3R/function/eval.global_human.utils.calculate_averages#L293-L304","kind":"function","name":"calculate_averages","path":"eval/global_human/utils.py","language":"python","start_line":293,"end_line":304,"context_start_line":273,"context_end_line":324,"code":"    if total_humans == 0:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    averages = [ws / total_humans for ws in weighted_sums]\n    \n    return total_humans, *averages\n\n\ndef process_directory(directory):\n    results = []\n    for root, _, files in os.walk(directory):\n        for file in sorted(files):\n            if file.endswith(\".txt\"):\n                file_path = os.path.join(root, file)\n                result = extract_metrics(file_path)\n                if result[0] > 0:\n                    results.append(result)\n    return results\n\n\ndef calculate_averages(results):\n    if not results:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    total_humans = sum(r[0] for r in results)\n    if total_humans == 0:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    weighted_sums = [sum(r[i] * r[0] for r in results) for i in range(1, 17)]\n    averages = [ws / total_humans for ws in weighted_sums]\n    \n    return total_humans, *averages\n\ndef visualize(\n    save_dir, img_path, view, gt_v3d_c, pred_v3d_c, \n    K_to_proj, gt_K, bestMatch, smpl_face\n):\n    from dust3r.utils import render_meshes, denormalize_rgb\n    from viser_utils import get_color\n    from PIL import Image\n    \n    os.makedirs(save_dir, exist_ok=True)\n    n_humans_i = pred_v3d_c.shape[0]\n\n    # image\n    img_array = denormalize_rgb(view.cpu().numpy())\n\n    focal = K_to_proj[[0,1],[0,1]].cpu().numpy()\n    princpt = K_to_proj[[0,1],[-1,-1]].cpu().numpy()\n    gt_focal = gt_K[[0,1],[0,1]].cpu().numpy()\n    gt_princpt = gt_K[[0,1],[-1,-1]].cpu().numpy()\n","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.visualize","uri":"program://Human3R/function/eval.global_human.utils.visualize#L306-L366","kind":"function","name":"visualize","path":"eval/global_human/utils.py","language":"python","start_line":306,"end_line":366,"context_start_line":286,"context_end_line":386,"code":"                file_path = os.path.join(root, file)\n                result = extract_metrics(file_path)\n                if result[0] > 0:\n                    results.append(result)\n    return results\n\n\ndef calculate_averages(results):\n    if not results:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    total_humans = sum(r[0] for r in results)\n    if total_humans == 0:\n        return 0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0\n    \n    weighted_sums = [sum(r[i] * r[0] for r in results) for i in range(1, 17)]\n    averages = [ws / total_humans for ws in weighted_sums]\n    \n    return total_humans, *averages\n\ndef visualize(\n    save_dir, img_path, view, gt_v3d_c, pred_v3d_c, \n    K_to_proj, gt_K, bestMatch, smpl_face\n):\n    from dust3r.utils import render_meshes, denormalize_rgb\n    from viser_utils import get_color\n    from PIL import Image\n    \n    os.makedirs(save_dir, exist_ok=True)\n    n_humans_i = pred_v3d_c.shape[0]\n\n    # image\n    img_array = denormalize_rgb(view.cpu().numpy())\n\n    focal = K_to_proj[[0,1],[0,1]].cpu().numpy()\n    princpt = K_to_proj[[0,1],[-1,-1]].cpu().numpy()\n    gt_focal = gt_K[[0,1],[0,1]].cpu().numpy()\n    gt_princpt = gt_K[[0,1],[-1,-1]].cpu().numpy()\n\n    gt_color_indices = list(range(gt_v3d_c.shape[0]))\n    pred_color_indices = [-1] * n_humans_i\n    for (pid, gid) in bestMatch:\n        pred_color_indices[pid] = gt_color_indices[gid]\n\n    next_color_idx = max(gt_color_indices) + 1 if gt_color_indices else 0\n    for i in range(len(pred_color_indices)):\n        if pred_color_indices[i] == -1:\n            pred_color_indices[i] = next_color_idx\n            next_color_idx += 1\n\n    # gt\n    gt_verts, gt_faces = [], []\n    for j in range(gt_v3d_c.shape[0]):\n        gt_verts.append(gt_v3d_c[j].cpu().numpy().reshape(-1,3))\n        gt_faces.append(smpl_face)\n\n    gt_colors = [get_color(gt_color_indices[j])/255 for j in range(len(gt_verts))]\n    gt_rend_array = render_meshes(img_array.copy(), \n                                    gt_verts, \n                                    gt_faces,\n                                    {'focal': gt_focal, 'princpt': gt_princpt},\n                                    color=gt_colors)\n    \n    # pred\n    pred_verts, pred_faces = [], []\n    for j in range(n_humans_i):\n        pred_verts.append(pred_v3d_c[j].cpu().numpy().reshape(-1,3))\n        pred_faces.append(smpl_face)\n    \n    pred_colors = [get_color(pred_color_indices[j])/255 for j in range(len(pred_verts))]\n    pred_rend_array = render_meshes(img_array.copy(), \n                                    pred_verts, \n                                    pred_faces,\n                                    {'focal': focal, 'princpt': princpt},\n                                    color=pred_colors)\n\n    img = np.concatenate([img_array, pred_rend_array, gt_rend_array], 1)\n    Image.fromarray(img).save(\n        os.path.join(\n            f\"{save_dir}/{os.path.splitext(os.path.basename(img_path))[0]}.jpg\"\n            ))\n\ndef write_log(log_path, dataset, seq, counter, metrics):\n    with open(log_path, \"a\") as f:\n        f.write(\n            f\"{dataset}-{seq: <16} |  \"\n            f\"n_human: {counter['n_human']:06d} | \"\n            f\"PVE: {avg_per_human(metrics['ca_pve']):.1f}, \"\n            f\"PA-PVE: {avg_per_human(metrics['pa_pve']):.1f}, \"\n            f\"Metric-PVE: {avg_per_human(metrics['me_pve']):.1f}, \"\n            f\"MPJPE: {avg_per_human(metrics['ca_mpjpe']):.1f}, \"\n            f\"PA-MPJPE: {avg_per_human(metrics['pa_mpjpe']):.1f}, \"\n            f\"Metric-MPJPE: {avg_per_human(metrics['me_mpjpe']):.1f}, \"\n            f\"RootError: {avg_per_human(metrics['rt_error']):.1f}, \"\n            f\"W-MPJPE: {avg_per_human(metrics['wa2_mpjpe']):.1f}, \"\n            f\"WA-MPJPE: {avg_per_human(metrics['waa_mpjpe']):.1f}, \"\n            f\"RTE: {avg_per_human(metrics['rte']):.1f}, \"\n            f\"Scaled-RTE: {avg_per_human(metrics['rte_scaled']):.1f}, \"\n            f\"Jitter: {avg_per_human(metrics['jitter']):.1f}, \"\n            f\"Foot-Sliding: {avg_per_human(metrics['fs']):.1f}, \"\n            f\"Precision: {metrics['precision']:.1f}, \"","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.write_log","uri":"program://Human3R/function/eval.global_human.utils.write_log#L368-L389","kind":"function","name":"write_log","path":"eval/global_human/utils.py","language":"python","start_line":368,"end_line":389,"context_start_line":348,"context_end_line":409,"code":"    \n    # pred\n    pred_verts, pred_faces = [], []\n    for j in range(n_humans_i):\n        pred_verts.append(pred_v3d_c[j].cpu().numpy().reshape(-1,3))\n        pred_faces.append(smpl_face)\n    \n    pred_colors = [get_color(pred_color_indices[j])/255 for j in range(len(pred_verts))]\n    pred_rend_array = render_meshes(img_array.copy(), \n                                    pred_verts, \n                                    pred_faces,\n                                    {'focal': focal, 'princpt': princpt},\n                                    color=pred_colors)\n\n    img = np.concatenate([img_array, pred_rend_array, gt_rend_array], 1)\n    Image.fromarray(img).save(\n        os.path.join(\n            f\"{save_dir}/{os.path.splitext(os.path.basename(img_path))[0]}.jpg\"\n            ))\n\ndef write_log(log_path, dataset, seq, counter, metrics):\n    with open(log_path, \"a\") as f:\n        f.write(\n            f\"{dataset}-{seq: <16} |  \"\n            f\"n_human: {counter['n_human']:06d} | \"\n            f\"PVE: {avg_per_human(metrics['ca_pve']):.1f}, \"\n            f\"PA-PVE: {avg_per_human(metrics['pa_pve']):.1f}, \"\n            f\"Metric-PVE: {avg_per_human(metrics['me_pve']):.1f}, \"\n            f\"MPJPE: {avg_per_human(metrics['ca_mpjpe']):.1f}, \"\n            f\"PA-MPJPE: {avg_per_human(metrics['pa_mpjpe']):.1f}, \"\n            f\"Metric-MPJPE: {avg_per_human(metrics['me_mpjpe']):.1f}, \"\n            f\"RootError: {avg_per_human(metrics['rt_error']):.1f}, \"\n            f\"W-MPJPE: {avg_per_human(metrics['wa2_mpjpe']):.1f}, \"\n            f\"WA-MPJPE: {avg_per_human(metrics['waa_mpjpe']):.1f}, \"\n            f\"RTE: {avg_per_human(metrics['rte']):.1f}, \"\n            f\"Scaled-RTE: {avg_per_human(metrics['rte_scaled']):.1f}, \"\n            f\"Jitter: {avg_per_human(metrics['jitter']):.1f}, \"\n            f\"Foot-Sliding: {avg_per_human(metrics['fs']):.1f}, \"\n            f\"Precision: {metrics['precision']:.1f}, \"\n            f\"Recall: {metrics['recall']:.1f}, \"\n            f\"F1-Score: {metrics['f1_score']:.1f}\\n\"\n        )\n\ndef get_summary_log(summary):\n    \"\"\"Generate summary log for evaluation results\"\"\"\n    return (\n        f\"EVALUATION SUMMARY\\n\"\n        f\"{'='*7}EVALUATION SUMMARY{'='*7}\\n\"\n        f\"Total Humans: {summary[0]}\\n\"\n        f\"\\n\"\n        f\"Camera Coordinate Metrics (mm):\\n\"\n        f\"  PVE:         {summary[1]:6.1f}\\n\"\n        f\"  PA-PVE:      {summary[2]:6.1f}\\n\"\n        f\"  Metric-PVE:  {summary[3]:6.1f}\\n\"\n        f\"\\n\"\n        f\"  MPJPE:       {summary[4]:6.1f}\\n\"\n        f\"  PA-MPJPE:    {summary[5]:6.1f}\\n\"\n        f\"  Metric-MPJPE: {summary[6]:6.1f}\\n\"\n        f\"  Root-Error:  {summary[7]:6.1f}\\n\"\n        f\"\\n\"\n        f\"Global Coordinate Metrics (cm):\\n\"\n        f\"  W-MPJPE:     {summary[8]:6.1f}\\n\"","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.get_summary_log","uri":"program://Human3R/function/eval.global_human.utils.get_summary_log#L391-L421","kind":"function","name":"get_summary_log","path":"eval/global_human/utils.py","language":"python","start_line":391,"end_line":421,"context_start_line":371,"context_end_line":441,"code":"            f\"{dataset}-{seq: <16} |  \"\n            f\"n_human: {counter['n_human']:06d} | \"\n            f\"PVE: {avg_per_human(metrics['ca_pve']):.1f}, \"\n            f\"PA-PVE: {avg_per_human(metrics['pa_pve']):.1f}, \"\n            f\"Metric-PVE: {avg_per_human(metrics['me_pve']):.1f}, \"\n            f\"MPJPE: {avg_per_human(metrics['ca_mpjpe']):.1f}, \"\n            f\"PA-MPJPE: {avg_per_human(metrics['pa_mpjpe']):.1f}, \"\n            f\"Metric-MPJPE: {avg_per_human(metrics['me_mpjpe']):.1f}, \"\n            f\"RootError: {avg_per_human(metrics['rt_error']):.1f}, \"\n            f\"W-MPJPE: {avg_per_human(metrics['wa2_mpjpe']):.1f}, \"\n            f\"WA-MPJPE: {avg_per_human(metrics['waa_mpjpe']):.1f}, \"\n            f\"RTE: {avg_per_human(metrics['rte']):.1f}, \"\n            f\"Scaled-RTE: {avg_per_human(metrics['rte_scaled']):.1f}, \"\n            f\"Jitter: {avg_per_human(metrics['jitter']):.1f}, \"\n            f\"Foot-Sliding: {avg_per_human(metrics['fs']):.1f}, \"\n            f\"Precision: {metrics['precision']:.1f}, \"\n            f\"Recall: {metrics['recall']:.1f}, \"\n            f\"F1-Score: {metrics['f1_score']:.1f}\\n\"\n        )\n\ndef get_summary_log(summary):\n    \"\"\"Generate summary log for evaluation results\"\"\"\n    return (\n        f\"EVALUATION SUMMARY\\n\"\n        f\"{'='*7}EVALUATION SUMMARY{'='*7}\\n\"\n        f\"Total Humans: {summary[0]}\\n\"\n        f\"\\n\"\n        f\"Camera Coordinate Metrics (mm):\\n\"\n        f\"  PVE:         {summary[1]:6.1f}\\n\"\n        f\"  PA-PVE:      {summary[2]:6.1f}\\n\"\n        f\"  Metric-PVE:  {summary[3]:6.1f}\\n\"\n        f\"\\n\"\n        f\"  MPJPE:       {summary[4]:6.1f}\\n\"\n        f\"  PA-MPJPE:    {summary[5]:6.1f}\\n\"\n        f\"  Metric-MPJPE: {summary[6]:6.1f}\\n\"\n        f\"  Root-Error:  {summary[7]:6.1f}\\n\"\n        f\"\\n\"\n        f\"Global Coordinate Metrics (cm):\\n\"\n        f\"  W-MPJPE:     {summary[8]:6.1f}\\n\"\n        f\"  WA-MPJPE:    {summary[9]:6.1f}\\n\"\n        f\"  RTE:         {summary[10]:6.1f}\\n\"\n        f\"  Scaled-RTE:  {summary[11]:6.1f}\\n\"\n        f\"  Jitter:      {summary[12]:6.1f}\\n\"\n        f\"  Foot-Sliding: {summary[13]:6.1f}\\n\"\n        f\"\\n\"\n        f\"Detection Metrics (%):\\n\"\n        f\"  Precision:   {summary[14]:6.1f}\\n\"\n        f\"  Recall:      {summary[15]:6.1f}\\n\"\n        f\"  F1-Score:    {summary[16]:6.1f}\\n\"\n        f\"{'='*32}\\n\"\n    )\n    \n\n# Evaluation metrics\n# Code modifed from https://github.com/zju3dv/GVHMR/blob/088caff492aa38c2d82cea363b78a3c65a83118f/hmr4d/utils/eval/eval_utils.py\n\ndef compute_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).mean(dim=-1).numpy()\n\ndef compute_perjoint_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).numpy()\n\ndef batch_compute_similarity_transform_torch(S1, S2):\n    '''\n    Computes a similarity transform (sR, t) that takes\n    a set of 3D points S1 (3 x N) closest to a set of 3D points S2,\n    where R is an 3x3 rotation matrix, t 3x1 translation, s scale.\n    i.e. solves the orthogonal Procrutes problem.\n    '''\n    S1 = S1.permute(0,2,1)\n    S2 = S2.permute(0,2,1)","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.compute_jpe","uri":"program://Human3R/function/eval.global_human.utils.compute_jpe#L427-L428","kind":"function","name":"compute_jpe","path":"eval/global_human/utils.py","language":"python","start_line":427,"end_line":428,"context_start_line":407,"context_end_line":448,"code":"        f\"\\n\"\n        f\"Global Coordinate Metrics (cm):\\n\"\n        f\"  W-MPJPE:     {summary[8]:6.1f}\\n\"\n        f\"  WA-MPJPE:    {summary[9]:6.1f}\\n\"\n        f\"  RTE:         {summary[10]:6.1f}\\n\"\n        f\"  Scaled-RTE:  {summary[11]:6.1f}\\n\"\n        f\"  Jitter:      {summary[12]:6.1f}\\n\"\n        f\"  Foot-Sliding: {summary[13]:6.1f}\\n\"\n        f\"\\n\"\n        f\"Detection Metrics (%):\\n\"\n        f\"  Precision:   {summary[14]:6.1f}\\n\"\n        f\"  Recall:      {summary[15]:6.1f}\\n\"\n        f\"  F1-Score:    {summary[16]:6.1f}\\n\"\n        f\"{'='*32}\\n\"\n    )\n    \n\n# Evaluation metrics\n# Code modifed from https://github.com/zju3dv/GVHMR/blob/088caff492aa38c2d82cea363b78a3c65a83118f/hmr4d/utils/eval/eval_utils.py\n\ndef compute_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).mean(dim=-1).numpy()\n\ndef compute_perjoint_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).numpy()\n\ndef batch_compute_similarity_transform_torch(S1, S2):\n    '''\n    Computes a similarity transform (sR, t) that takes\n    a set of 3D points S1 (3 x N) closest to a set of 3D points S2,\n    where R is an 3x3 rotation matrix, t 3x1 translation, s scale.\n    i.e. solves the orthogonal Procrutes problem.\n    '''\n    S1 = S1.permute(0,2,1)\n    S2 = S2.permute(0,2,1)\n    transposed = True\n\n    # 1. Remove mean.\n    mu1 = S1.mean(axis=-1, keepdims=True)\n    mu2 = S2.mean(axis=-1, keepdims=True)\n\n    X1 = S1 - mu1","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.compute_perjoint_jpe","uri":"program://Human3R/function/eval.global_human.utils.compute_perjoint_jpe#L430-L431","kind":"function","name":"compute_perjoint_jpe","path":"eval/global_human/utils.py","language":"python","start_line":430,"end_line":431,"context_start_line":410,"context_end_line":451,"code":"        f\"  WA-MPJPE:    {summary[9]:6.1f}\\n\"\n        f\"  RTE:         {summary[10]:6.1f}\\n\"\n        f\"  Scaled-RTE:  {summary[11]:6.1f}\\n\"\n        f\"  Jitter:      {summary[12]:6.1f}\\n\"\n        f\"  Foot-Sliding: {summary[13]:6.1f}\\n\"\n        f\"\\n\"\n        f\"Detection Metrics (%):\\n\"\n        f\"  Precision:   {summary[14]:6.1f}\\n\"\n        f\"  Recall:      {summary[15]:6.1f}\\n\"\n        f\"  F1-Score:    {summary[16]:6.1f}\\n\"\n        f\"{'='*32}\\n\"\n    )\n    \n\n# Evaluation metrics\n# Code modifed from https://github.com/zju3dv/GVHMR/blob/088caff492aa38c2d82cea363b78a3c65a83118f/hmr4d/utils/eval/eval_utils.py\n\ndef compute_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).mean(dim=-1).numpy()\n\ndef compute_perjoint_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).numpy()\n\ndef batch_compute_similarity_transform_torch(S1, S2):\n    '''\n    Computes a similarity transform (sR, t) that takes\n    a set of 3D points S1 (3 x N) closest to a set of 3D points S2,\n    where R is an 3x3 rotation matrix, t 3x1 translation, s scale.\n    i.e. solves the orthogonal Procrutes problem.\n    '''\n    S1 = S1.permute(0,2,1)\n    S2 = S2.permute(0,2,1)\n    transposed = True\n\n    # 1. Remove mean.\n    mu1 = S1.mean(axis=-1, keepdims=True)\n    mu2 = S2.mean(axis=-1, keepdims=True)\n\n    X1 = S1 - mu1\n    X2 = S2 - mu2\n\n    # 2. Compute variance of X1 used for scale.","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.batch_compute_similarity_transform_torch","uri":"program://Human3R/function/eval.global_human.utils.batch_compute_similarity_transform_torch#L433-L481","kind":"function","name":"batch_compute_similarity_transform_torch","path":"eval/global_human/utils.py","language":"python","start_line":433,"end_line":481,"context_start_line":413,"context_end_line":501,"code":"        f\"  Jitter:      {summary[12]:6.1f}\\n\"\n        f\"  Foot-Sliding: {summary[13]:6.1f}\\n\"\n        f\"\\n\"\n        f\"Detection Metrics (%):\\n\"\n        f\"  Precision:   {summary[14]:6.1f}\\n\"\n        f\"  Recall:      {summary[15]:6.1f}\\n\"\n        f\"  F1-Score:    {summary[16]:6.1f}\\n\"\n        f\"{'='*32}\\n\"\n    )\n    \n\n# Evaluation metrics\n# Code modifed from https://github.com/zju3dv/GVHMR/blob/088caff492aa38c2d82cea363b78a3c65a83118f/hmr4d/utils/eval/eval_utils.py\n\ndef compute_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).mean(dim=-1).numpy()\n\ndef compute_perjoint_jpe(S1, S2):\n    return torch.sqrt(((S1 - S2) ** 2).sum(dim=-1)).numpy()\n\ndef batch_compute_similarity_transform_torch(S1, S2):\n    '''\n    Computes a similarity transform (sR, t) that takes\n    a set of 3D points S1 (3 x N) closest to a set of 3D points S2,\n    where R is an 3x3 rotation matrix, t 3x1 translation, s scale.\n    i.e. solves the orthogonal Procrutes problem.\n    '''\n    S1 = S1.permute(0,2,1)\n    S2 = S2.permute(0,2,1)\n    transposed = True\n\n    # 1. Remove mean.\n    mu1 = S1.mean(axis=-1, keepdims=True)\n    mu2 = S2.mean(axis=-1, keepdims=True)\n\n    X1 = S1 - mu1\n    X2 = S2 - mu2\n\n    # 2. Compute variance of X1 used for scale.\n    var1 = torch.sum(X1**2, dim=1).sum(dim=1)\n\n    # 3. The outer product of X1 and X2.\n    K = X1.bmm(X2.permute(0,2,1))\n\n    # 4. Solution that Maximizes trace(R'K) is R=U*V', where U, V are\n    # singular vectors of K.\n    U, s, V = torch.svd(K)\n\n    # Construct Z that fixes the orientation of R to get det(R)=1.\n    Z = torch.eye(U.shape[1], device=S1.device).unsqueeze(0)\n    Z = Z.repeat(U.shape[0],1,1)\n    Z[:,-1, -1] *= torch.sign(torch.det(U.bmm(V.permute(0,2,1))))\n\n    # Construct R.\n    R = V.bmm(Z.bmm(U.permute(0,2,1)))\n\n    # 5. Recover scale.\n    scale = torch.cat([torch.trace(x).unsqueeze(0) for x in R.bmm(K)]) / var1\n\n    # 6. Recover translation.\n    t = mu2 - (scale.unsqueeze(-1).unsqueeze(-1) * (R.bmm(mu1)))\n\n    # 7. Error:\n    S1_hat = scale.unsqueeze(-1).unsqueeze(-1) * R.bmm(S1) + t\n\n    if transposed:\n        S1_hat = S1_hat.permute(0,2,1)\n\n    return S1_hat\n\ndef batch_align_by_pelvis(data_list, pelvis_idxs):\n    \"\"\"\n    Assumes data is given as [pred_j3d, target_j3d, pred_verts, target_verts].\n    Each data is in shape of (batch, num_points, 3)\n    Pelvis is notated as one / two joints indices.\n    Align all data to the corresponding pelvis location.\n    \"\"\"\n\n    pred_j3d, target_j3d, pred_verts, target_verts = data_list\n    \n    pred_pelvis = pred_j3d[:, pelvis_idxs].mean(dim=1, keepdims=True).clone()\n    target_pelvis = target_j3d[:, pelvis_idxs].mean(dim=1, keepdims=True).clone()\n    \n    # Align to the pelvis\n    pred_j3d = pred_j3d - pred_pelvis\n    target_j3d = target_j3d - target_pelvis\n    pred_verts = pred_verts - pred_pelvis\n    target_verts = target_verts - target_pelvis\n    ","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.batch_align_by_pelvis","uri":"program://Human3R/function/eval.global_human.utils.batch_align_by_pelvis#L483-L502","kind":"function","name":"batch_align_by_pelvis","path":"eval/global_human/utils.py","language":"python","start_line":483,"end_line":502,"context_start_line":463,"context_end_line":522,"code":"    Z = Z.repeat(U.shape[0],1,1)\n    Z[:,-1, -1] *= torch.sign(torch.det(U.bmm(V.permute(0,2,1))))\n\n    # Construct R.\n    R = V.bmm(Z.bmm(U.permute(0,2,1)))\n\n    # 5. Recover scale.\n    scale = torch.cat([torch.trace(x).unsqueeze(0) for x in R.bmm(K)]) / var1\n\n    # 6. Recover translation.\n    t = mu2 - (scale.unsqueeze(-1).unsqueeze(-1) * (R.bmm(mu1)))\n\n    # 7. Error:\n    S1_hat = scale.unsqueeze(-1).unsqueeze(-1) * R.bmm(S1) + t\n\n    if transposed:\n        S1_hat = S1_hat.permute(0,2,1)\n\n    return S1_hat\n\ndef batch_align_by_pelvis(data_list, pelvis_idxs):\n    \"\"\"\n    Assumes data is given as [pred_j3d, target_j3d, pred_verts, target_verts].\n    Each data is in shape of (batch, num_points, 3)\n    Pelvis is notated as one / two joints indices.\n    Align all data to the corresponding pelvis location.\n    \"\"\"\n\n    pred_j3d, target_j3d, pred_verts, target_verts = data_list\n    \n    pred_pelvis = pred_j3d[:, pelvis_idxs].mean(dim=1, keepdims=True).clone()\n    target_pelvis = target_j3d[:, pelvis_idxs].mean(dim=1, keepdims=True).clone()\n    \n    # Align to the pelvis\n    pred_j3d = pred_j3d - pred_pelvis\n    target_j3d = target_j3d - target_pelvis\n    pred_verts = pred_verts - pred_pelvis\n    target_verts = target_verts - target_pelvis\n    \n    return (pred_j3d, target_j3d, pred_verts, target_verts, pred_pelvis, target_pelvis)\n\ndef align_pcl(Y, X, weight=None, fixed_scale=False):\n    \"\"\"\n    align similarity transform to align X with Y using umeyama method\n    X' = s * R * X + t is aligned with Y\n    :param Y (*, N, 3) first trajectory\n    :param X (*, N, 3) second trajectory\n    :param weight (*, N, 1) optional weight of valid correspondences\n    :returns s (*, 1), R (*, 3, 3), t (*, 3)\n    \"\"\"\n    *dims, N, _ = Y.shape\n    N = torch.ones(*dims, 1, 1) * N\n\n    if weight is not None:\n        Y = Y * weight\n        X = X * weight\n        N = weight.sum(dim=-2, keepdim=True)  # (*, 1, 1)\n\n    # subtract mean\n    my = Y.sum(dim=-2) / N[..., 0]  # (*, 3)","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.align_pcl","uri":"program://Human3R/function/eval.global_human.utils.align_pcl#L504-L550","kind":"function","name":"align_pcl","path":"eval/global_human/utils.py","language":"python","start_line":504,"end_line":550,"context_start_line":484,"context_end_line":570,"code":"    \"\"\"\n    Assumes data is given as [pred_j3d, target_j3d, pred_verts, target_verts].\n    Each data is in shape of (batch, num_points, 3)\n    Pelvis is notated as one / two joints indices.\n    Align all data to the corresponding pelvis location.\n    \"\"\"\n\n    pred_j3d, target_j3d, pred_verts, target_verts = data_list\n    \n    pred_pelvis = pred_j3d[:, pelvis_idxs].mean(dim=1, keepdims=True).clone()\n    target_pelvis = target_j3d[:, pelvis_idxs].mean(dim=1, keepdims=True).clone()\n    \n    # Align to the pelvis\n    pred_j3d = pred_j3d - pred_pelvis\n    target_j3d = target_j3d - target_pelvis\n    pred_verts = pred_verts - pred_pelvis\n    target_verts = target_verts - target_pelvis\n    \n    return (pred_j3d, target_j3d, pred_verts, target_verts, pred_pelvis, target_pelvis)\n\ndef align_pcl(Y, X, weight=None, fixed_scale=False):\n    \"\"\"\n    align similarity transform to align X with Y using umeyama method\n    X' = s * R * X + t is aligned with Y\n    :param Y (*, N, 3) first trajectory\n    :param X (*, N, 3) second trajectory\n    :param weight (*, N, 1) optional weight of valid correspondences\n    :returns s (*, 1), R (*, 3, 3), t (*, 3)\n    \"\"\"\n    *dims, N, _ = Y.shape\n    N = torch.ones(*dims, 1, 1) * N\n\n    if weight is not None:\n        Y = Y * weight\n        X = X * weight\n        N = weight.sum(dim=-2, keepdim=True)  # (*, 1, 1)\n\n    # subtract mean\n    my = Y.sum(dim=-2) / N[..., 0]  # (*, 3)\n    mx = X.sum(dim=-2) / N[..., 0]\n    y0 = Y - my[..., None, :]  # (*, N, 3)\n    x0 = X - mx[..., None, :]\n\n    if weight is not None:\n        y0 = y0 * weight\n        x0 = x0 * weight\n\n    # correlation\n    C = torch.matmul(y0.transpose(-1, -2), x0) / N  # (*, 3, 3)\n    U, D, Vh = torch.linalg.svd(C)  # (*, 3, 3), (*, 3), (*, 3, 3)\n\n    S = torch.eye(3).reshape(*(1,) * (len(dims)), 3, 3).repeat(*dims, 1, 1)\n    neg = torch.det(U) * torch.det(Vh.transpose(-1, -2)) < 0\n    S[neg, 2, 2] = -1\n\n    R = torch.matmul(U, torch.matmul(S, Vh))  # (*, 3, 3)\n\n    D = torch.diag_embed(D)  # (*, 3, 3)\n    if fixed_scale:\n        s = torch.ones(*dims, 1, device=Y.device, dtype=torch.float32)\n    else:\n        var = torch.sum(torch.square(x0), dim=(-1, -2), keepdim=True) / N  # (*, 1, 1)\n        s = torch.diagonal(torch.matmul(D, S), dim1=-2, dim2=-1).sum(dim=-1, keepdim=True) / var[..., 0]  # (*, 1)\n\n    t = my - s * torch.matmul(R, mx[..., None])[..., 0]  # (*, 3)\n\n    return s, R, t\n\ndef global_align_joints(gt_joints, pred_joints):\n    \"\"\"\n    :param gt_joints (T, J, 3)\n    :param pred_joints (T, J, 3)\n    \"\"\"\n    s_glob, R_glob, t_glob = align_pcl(gt_joints.reshape(-1, 3), pred_joints.reshape(-1, 3))\n    pred_glob = s_glob * torch.einsum(\"ij,tnj->tni\", R_glob, pred_joints) + t_glob[None, None]\n    return pred_glob\n\ndef first_align_joints(gt_joints, pred_joints):\n    \"\"\"\n    align the first two frames\n    :param gt_joints (T, J, 3)\n    :param pred_joints (T, J, 3)\n    \"\"\"\n    # (1, 1), (1, 3, 3), (1, 3)\n    s_first, R_first, t_first = align_pcl(gt_joints[:2].reshape(1, -1, 3), pred_joints[:2].reshape(1, -1, 3))\n    pred_first = s_first * torch.einsum(\"tij,tnj->tni\", R_first, pred_joints) + t_first[:, None]\n    return pred_first","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.global_align_joints","uri":"program://Human3R/function/eval.global_human.utils.global_align_joints#L552-L559","kind":"function","name":"global_align_joints","path":"eval/global_human/utils.py","language":"python","start_line":552,"end_line":559,"context_start_line":532,"context_end_line":579,"code":"    C = torch.matmul(y0.transpose(-1, -2), x0) / N  # (*, 3, 3)\n    U, D, Vh = torch.linalg.svd(C)  # (*, 3, 3), (*, 3), (*, 3, 3)\n\n    S = torch.eye(3).reshape(*(1,) * (len(dims)), 3, 3).repeat(*dims, 1, 1)\n    neg = torch.det(U) * torch.det(Vh.transpose(-1, -2)) < 0\n    S[neg, 2, 2] = -1\n\n    R = torch.matmul(U, torch.matmul(S, Vh))  # (*, 3, 3)\n\n    D = torch.diag_embed(D)  # (*, 3, 3)\n    if fixed_scale:\n        s = torch.ones(*dims, 1, device=Y.device, dtype=torch.float32)\n    else:\n        var = torch.sum(torch.square(x0), dim=(-1, -2), keepdim=True) / N  # (*, 1, 1)\n        s = torch.diagonal(torch.matmul(D, S), dim1=-2, dim2=-1).sum(dim=-1, keepdim=True) / var[..., 0]  # (*, 1)\n\n    t = my - s * torch.matmul(R, mx[..., None])[..., 0]  # (*, 3)\n\n    return s, R, t\n\ndef global_align_joints(gt_joints, pred_joints):\n    \"\"\"\n    :param gt_joints (T, J, 3)\n    :param pred_joints (T, J, 3)\n    \"\"\"\n    s_glob, R_glob, t_glob = align_pcl(gt_joints.reshape(-1, 3), pred_joints.reshape(-1, 3))\n    pred_glob = s_glob * torch.einsum(\"ij,tnj->tni\", R_glob, pred_joints) + t_glob[None, None]\n    return pred_glob\n\ndef first_align_joints(gt_joints, pred_joints):\n    \"\"\"\n    align the first two frames\n    :param gt_joints (T, J, 3)\n    :param pred_joints (T, J, 3)\n    \"\"\"\n    # (1, 1), (1, 3, 3), (1, 3)\n    s_first, R_first, t_first = align_pcl(gt_joints[:2].reshape(1, -1, 3), pred_joints[:2].reshape(1, -1, 3))\n    pred_first = s_first * torch.einsum(\"tij,tnj->tni\", R_first, pred_joints) + t_first[:, None]\n    return pred_first\n\ndef compute_rte(target_trans, pred_trans, fixed_scale=True):\n    # Compute the global alignment\n    scale, rot, trans = align_pcl(target_trans[None, :], pred_trans[None, :], fixed_scale=fixed_scale)\n    pred_trans_hat = (scale * torch.einsum(\"tij,tnj->tni\", rot, pred_trans[None, :]) + trans[None, :])[0]\n\n    # Compute the entire displacement of ground truth trajectory\n    disps, disp = [], 0\n    for p1, p2 in zip(target_trans, target_trans[1:]):","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.first_align_joints","uri":"program://Human3R/function/eval.global_human.utils.first_align_joints#L561-L570","kind":"function","name":"first_align_joints","path":"eval/global_human/utils.py","language":"python","start_line":561,"end_line":570,"context_start_line":541,"context_end_line":590,"code":"    D = torch.diag_embed(D)  # (*, 3, 3)\n    if fixed_scale:\n        s = torch.ones(*dims, 1, device=Y.device, dtype=torch.float32)\n    else:\n        var = torch.sum(torch.square(x0), dim=(-1, -2), keepdim=True) / N  # (*, 1, 1)\n        s = torch.diagonal(torch.matmul(D, S), dim1=-2, dim2=-1).sum(dim=-1, keepdim=True) / var[..., 0]  # (*, 1)\n\n    t = my - s * torch.matmul(R, mx[..., None])[..., 0]  # (*, 3)\n\n    return s, R, t\n\ndef global_align_joints(gt_joints, pred_joints):\n    \"\"\"\n    :param gt_joints (T, J, 3)\n    :param pred_joints (T, J, 3)\n    \"\"\"\n    s_glob, R_glob, t_glob = align_pcl(gt_joints.reshape(-1, 3), pred_joints.reshape(-1, 3))\n    pred_glob = s_glob * torch.einsum(\"ij,tnj->tni\", R_glob, pred_joints) + t_glob[None, None]\n    return pred_glob\n\ndef first_align_joints(gt_joints, pred_joints):\n    \"\"\"\n    align the first two frames\n    :param gt_joints (T, J, 3)\n    :param pred_joints (T, J, 3)\n    \"\"\"\n    # (1, 1), (1, 3, 3), (1, 3)\n    s_first, R_first, t_first = align_pcl(gt_joints[:2].reshape(1, -1, 3), pred_joints[:2].reshape(1, -1, 3))\n    pred_first = s_first * torch.einsum(\"tij,tnj->tni\", R_first, pred_joints) + t_first[:, None]\n    return pred_first\n\ndef compute_rte(target_trans, pred_trans, fixed_scale=True):\n    # Compute the global alignment\n    scale, rot, trans = align_pcl(target_trans[None, :], pred_trans[None, :], fixed_scale=fixed_scale)\n    pred_trans_hat = (scale * torch.einsum(\"tij,tnj->tni\", rot, pred_trans[None, :]) + trans[None, :])[0]\n\n    # Compute the entire displacement of ground truth trajectory\n    disps, disp = [], 0\n    for p1, p2 in zip(target_trans, target_trans[1:]):\n        delta = (p2 - p1).norm(2, dim=-1)\n        disp += delta\n        disps.append(disp)\n\n    # Compute absolute root-translation-error (RTE)\n    rte = torch.norm(target_trans - pred_trans_hat, 2, dim=-1)\n\n    # Normalize it to the displacement\n    return (rte / disp).numpy()\n\n","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.compute_rte","uri":"program://Human3R/function/eval.global_human.utils.compute_rte#L572-L588","kind":"function","name":"compute_rte","path":"eval/global_human/utils.py","language":"python","start_line":572,"end_line":588,"context_start_line":552,"context_end_line":608,"code":"def global_align_joints(gt_joints, pred_joints):\n    \"\"\"\n    :param gt_joints (T, J, 3)\n    :param pred_joints (T, J, 3)\n    \"\"\"\n    s_glob, R_glob, t_glob = align_pcl(gt_joints.reshape(-1, 3), pred_joints.reshape(-1, 3))\n    pred_glob = s_glob * torch.einsum(\"ij,tnj->tni\", R_glob, pred_joints) + t_glob[None, None]\n    return pred_glob\n\ndef first_align_joints(gt_joints, pred_joints):\n    \"\"\"\n    align the first two frames\n    :param gt_joints (T, J, 3)\n    :param pred_joints (T, J, 3)\n    \"\"\"\n    # (1, 1), (1, 3, 3), (1, 3)\n    s_first, R_first, t_first = align_pcl(gt_joints[:2].reshape(1, -1, 3), pred_joints[:2].reshape(1, -1, 3))\n    pred_first = s_first * torch.einsum(\"tij,tnj->tni\", R_first, pred_joints) + t_first[:, None]\n    return pred_first\n\ndef compute_rte(target_trans, pred_trans, fixed_scale=True):\n    # Compute the global alignment\n    scale, rot, trans = align_pcl(target_trans[None, :], pred_trans[None, :], fixed_scale=fixed_scale)\n    pred_trans_hat = (scale * torch.einsum(\"tij,tnj->tni\", rot, pred_trans[None, :]) + trans[None, :])[0]\n\n    # Compute the entire displacement of ground truth trajectory\n    disps, disp = [], 0\n    for p1, p2 in zip(target_trans, target_trans[1:]):\n        delta = (p2 - p1).norm(2, dim=-1)\n        disp += delta\n        disps.append(disp)\n\n    # Compute absolute root-translation-error (RTE)\n    rte = torch.norm(target_trans - pred_trans_hat, 2, dim=-1)\n\n    # Normalize it to the displacement\n    return (rte / disp).numpy()\n\n\ndef compute_jitter(joints, fps=30):\n    \"\"\"compute jitter of the motion\n    Args:\n        joints (N, J, 3).\n        fps (float).\n    Returns:\n        jitter (N-3).\n    \"\"\"\n    pred_jitter = torch.norm(\n        (joints[3:] - 3 * joints[2:-1] + 3 * joints[1:-2] - joints[:-3]) * (fps**3),\n        dim=2,\n    ).mean(dim=-1)\n\n    return pred_jitter.cpu().numpy() / 10.0\n\n\ndef compute_foot_sliding(target_verts, pred_verts, thr=1e-2):\n    \"\"\"compute foot sliding error","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.compute_jitter","uri":"program://Human3R/function/eval.global_human.utils.compute_jitter#L591-L604","kind":"function","name":"compute_jitter","path":"eval/global_human/utils.py","language":"python","start_line":591,"end_line":604,"context_start_line":571,"context_end_line":624,"code":"\ndef compute_rte(target_trans, pred_trans, fixed_scale=True):\n    # Compute the global alignment\n    scale, rot, trans = align_pcl(target_trans[None, :], pred_trans[None, :], fixed_scale=fixed_scale)\n    pred_trans_hat = (scale * torch.einsum(\"tij,tnj->tni\", rot, pred_trans[None, :]) + trans[None, :])[0]\n\n    # Compute the entire displacement of ground truth trajectory\n    disps, disp = [], 0\n    for p1, p2 in zip(target_trans, target_trans[1:]):\n        delta = (p2 - p1).norm(2, dim=-1)\n        disp += delta\n        disps.append(disp)\n\n    # Compute absolute root-translation-error (RTE)\n    rte = torch.norm(target_trans - pred_trans_hat, 2, dim=-1)\n\n    # Normalize it to the displacement\n    return (rte / disp).numpy()\n\n\ndef compute_jitter(joints, fps=30):\n    \"\"\"compute jitter of the motion\n    Args:\n        joints (N, J, 3).\n        fps (float).\n    Returns:\n        jitter (N-3).\n    \"\"\"\n    pred_jitter = torch.norm(\n        (joints[3:] - 3 * joints[2:-1] + 3 * joints[1:-2] - joints[:-3]) * (fps**3),\n        dim=2,\n    ).mean(dim=-1)\n\n    return pred_jitter.cpu().numpy() / 10.0\n\n\ndef compute_foot_sliding(target_verts, pred_verts, thr=1e-2):\n    \"\"\"compute foot sliding error\n    The foot ground contact label is computed by the threshold of 1 cm/frame\n    Args:\n        target_verts (N, 6890, 3).\n        pred_verts (N, 6890, 3).\n    Returns:\n        error (N frames in contact).\n    \"\"\"\n    assert target_verts.shape == pred_verts.shape\n    assert target_verts.shape[-2] == 6890\n\n    # Foot vertices idxs\n    foot_idxs = [3216, 3387, 6617, 6787]\n\n    # Compute contact label\n    foot_loc = target_verts[:, foot_idxs]\n    foot_disp = (foot_loc[1:] - foot_loc[:-1]).norm(2, dim=-1)","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.compute_foot_sliding","uri":"program://Human3R/function/eval.global_human.utils.compute_foot_sliding#L607-L632","kind":"function","name":"compute_foot_sliding","path":"eval/global_human/utils.py","language":"python","start_line":607,"end_line":632,"context_start_line":587,"context_end_line":652,"code":"    # Normalize it to the displacement\n    return (rte / disp).numpy()\n\n\ndef compute_jitter(joints, fps=30):\n    \"\"\"compute jitter of the motion\n    Args:\n        joints (N, J, 3).\n        fps (float).\n    Returns:\n        jitter (N-3).\n    \"\"\"\n    pred_jitter = torch.norm(\n        (joints[3:] - 3 * joints[2:-1] + 3 * joints[1:-2] - joints[:-3]) * (fps**3),\n        dim=2,\n    ).mean(dim=-1)\n\n    return pred_jitter.cpu().numpy() / 10.0\n\n\ndef compute_foot_sliding(target_verts, pred_verts, thr=1e-2):\n    \"\"\"compute foot sliding error\n    The foot ground contact label is computed by the threshold of 1 cm/frame\n    Args:\n        target_verts (N, 6890, 3).\n        pred_verts (N, 6890, 3).\n    Returns:\n        error (N frames in contact).\n    \"\"\"\n    assert target_verts.shape == pred_verts.shape\n    assert target_verts.shape[-2] == 6890\n\n    # Foot vertices idxs\n    foot_idxs = [3216, 3387, 6617, 6787]\n\n    # Compute contact label\n    foot_loc = target_verts[:, foot_idxs]\n    foot_disp = (foot_loc[1:] - foot_loc[:-1]).norm(2, dim=-1)\n    contact = foot_disp[:] < thr\n\n    pred_feet_loc = pred_verts[:, foot_idxs]\n    pred_disp = (pred_feet_loc[1:] - pred_feet_loc[:-1]).norm(2, dim=-1)\n\n    error = pred_disp[contact]\n\n    return error.cpu().numpy()\n\ndef eval_camcoord(batch, pelvis_idxs=[1, 2], fps=30):\n    \"\"\"\n    Args:\n        batch (dict): {\n            \"pred_j3d\": (..., J, 3) tensor\n            \"target_j3d\":\n            \"pred_v3d\":\n            \"target_v3d\":\n        }\n    Returns:\n        cam_coord_metrics (dict): {\n            \"pa_mpjpe\": (..., ) numpy array\n            \"mpjpe\":\n            \"pve\":\n            \"accel\":\n        }\n    \"\"\"\n    # All data is in camera coordinates\n    pred_j3d = batch[\"pred_j3d\"]  # (..., J, 3)","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.eval_camcoord","uri":"program://Human3R/function/eval.global_human.utils.eval_camcoord#L634-L691","kind":"function","name":"eval_camcoord","path":"eval/global_human/utils.py","language":"python","start_line":634,"end_line":691,"context_start_line":614,"context_end_line":711,"code":"        error (N frames in contact).\n    \"\"\"\n    assert target_verts.shape == pred_verts.shape\n    assert target_verts.shape[-2] == 6890\n\n    # Foot vertices idxs\n    foot_idxs = [3216, 3387, 6617, 6787]\n\n    # Compute contact label\n    foot_loc = target_verts[:, foot_idxs]\n    foot_disp = (foot_loc[1:] - foot_loc[:-1]).norm(2, dim=-1)\n    contact = foot_disp[:] < thr\n\n    pred_feet_loc = pred_verts[:, foot_idxs]\n    pred_disp = (pred_feet_loc[1:] - pred_feet_loc[:-1]).norm(2, dim=-1)\n\n    error = pred_disp[contact]\n\n    return error.cpu().numpy()\n\ndef eval_camcoord(batch, pelvis_idxs=[1, 2], fps=30):\n    \"\"\"\n    Args:\n        batch (dict): {\n            \"pred_j3d\": (..., J, 3) tensor\n            \"target_j3d\":\n            \"pred_v3d\":\n            \"target_v3d\":\n        }\n    Returns:\n        cam_coord_metrics (dict): {\n            \"pa_mpjpe\": (..., ) numpy array\n            \"mpjpe\":\n            \"pve\":\n            \"accel\":\n        }\n    \"\"\"\n    # All data is in camera coordinates\n    pred_j3d = batch[\"pred_j3d\"]  # (..., J, 3)\n    target_j3d = batch[\"target_j3d\"]\n    pred_verts = batch[\"pred_v3d\"]\n    target_verts = batch[\"target_v3d\"]\n\n    # Center Align by pelvis\n    (   ca_pred_j3d, ca_target_j3d, ca_pred_verts, ca_target_verts, pred_pelvis, target_pelvis\n     ) = batch_align_by_pelvis(\n        [pred_j3d, target_j3d, pred_verts, target_verts], pelvis_idxs\n    )\n\n    pa_pred_j3d = batch_compute_similarity_transform_torch(ca_pred_j3d, ca_target_j3d)\n    pa_pred_verts = batch_compute_similarity_transform_torch(ca_pred_verts, ca_target_verts)\n\n    # Metrics\n    m2mm = 1000\n\n    # metric scale\n    rt_error = compute_jpe(pred_pelvis, target_pelvis) * m2mm\n    me_mpjpe = compute_jpe(pred_j3d, target_j3d) * m2mm\n    me_pve = compute_jpe(pred_verts, target_verts) * m2mm\n\n    # center aligned\n    ca_mpjpe = compute_jpe(ca_pred_j3d, ca_target_j3d) * m2mm\n    ca_pve = compute_jpe(ca_pred_verts, ca_target_verts) * m2mm\n   \n    # procrustes aligned\n    pa_mpjpe = compute_jpe(pa_pred_j3d, ca_target_j3d) * m2mm\n    pa_pve = compute_jpe(pa_pred_verts, ca_target_verts) * m2mm\n    \n    camcoord_metrics = {\n        \"me_mpjpe\": me_mpjpe,\n        \"ca_mpjpe\": ca_mpjpe,\n        \"pa_mpjpe\": pa_mpjpe,\n        \"me_pve\": me_pve,\n        \"ca_pve\": ca_pve,\n        \"pa_pve\": pa_pve,\n        \"rt_error\": rt_error,\n    }\n    return camcoord_metrics\n\ndef eval_global(batch, subsample=1):\n    \"\"\"Follow WHAM, the input has skipped invalid frames\n    Args:\n        batch (dict): {\n            \"pred_j3d\": (F, J, 3) tensor\n            \"target_j3d\":\n            \"pred_v3d\":\n            \"target_v3d\":\n        }\n    Returns:\n        global_metrics (dict): {\n            \"wa2_mpjpe\": (F, ) numpy array\n            \"waa_mpjpe\":\n            \"rte\":\n            \"jitter\":\n            \"fs\":\n        }\n    \"\"\"\n    # All data is in global coordinates","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.utils.eval_global","uri":"program://Human3R/function/eval.global_human.utils.eval_global#L693-L754","kind":"function","name":"eval_global","path":"eval/global_human/utils.py","language":"python","start_line":693,"end_line":754,"context_start_line":673,"context_end_line":754,"code":"\n    # center aligned\n    ca_mpjpe = compute_jpe(ca_pred_j3d, ca_target_j3d) * m2mm\n    ca_pve = compute_jpe(ca_pred_verts, ca_target_verts) * m2mm\n   \n    # procrustes aligned\n    pa_mpjpe = compute_jpe(pa_pred_j3d, ca_target_j3d) * m2mm\n    pa_pve = compute_jpe(pa_pred_verts, ca_target_verts) * m2mm\n    \n    camcoord_metrics = {\n        \"me_mpjpe\": me_mpjpe,\n        \"ca_mpjpe\": ca_mpjpe,\n        \"pa_mpjpe\": pa_mpjpe,\n        \"me_pve\": me_pve,\n        \"ca_pve\": ca_pve,\n        \"pa_pve\": pa_pve,\n        \"rt_error\": rt_error,\n    }\n    return camcoord_metrics\n\ndef eval_global(batch, subsample=1):\n    \"\"\"Follow WHAM, the input has skipped invalid frames\n    Args:\n        batch (dict): {\n            \"pred_j3d\": (F, J, 3) tensor\n            \"target_j3d\":\n            \"pred_v3d\":\n            \"target_v3d\":\n        }\n    Returns:\n        global_metrics (dict): {\n            \"wa2_mpjpe\": (F, ) numpy array\n            \"waa_mpjpe\":\n            \"rte\":\n            \"jitter\":\n            \"fs\":\n        }\n    \"\"\"\n    # All data is in global coordinates\n    pred_j3d_glob = batch[\"pred_j3d\"]  # (..., J, 3)\n    target_j3d_glob = batch[\"target_j3d\"]\n    pred_verts_glob = batch[\"pred_v3d\"]\n    target_verts_glob = batch[\"target_v3d\"]\n\n    seq_length = pred_j3d_glob.shape[0]\n\n    # Use chunk to compare\n    chunk_length = int(100 / subsample)\n    wa2_mpjpe, waa_mpjpe = [], []\n    for start in range(0, seq_length, chunk_length):\n        end = min(seq_length, start + chunk_length)\n\n        target_j3d = target_j3d_glob[start:end].clone().cpu()\n        pred_j3d = pred_j3d_glob[start:end].clone().cpu()\n\n        w_j3d = first_align_joints(target_j3d, pred_j3d)\n        wa_j3d = global_align_joints(target_j3d, pred_j3d)\n\n        wa2_mpjpe.append(compute_jpe(target_j3d, w_j3d))\n        waa_mpjpe.append(compute_jpe(target_j3d, wa_j3d))\n\n    # Metrics\n    m2mm = 1000\n    wa2_mpjpe = np.concatenate(wa2_mpjpe) * m2mm\n    waa_mpjpe = np.concatenate(waa_mpjpe) * m2mm\n\n    # Additional Metrics\n    rte = compute_rte(target_j3d_glob[:, 0].cpu(), pred_j3d_glob[:, 0].cpu()) * 1e2\n    rte_scaled = compute_rte(\n        target_j3d_glob[:, 0].cpu(), pred_j3d_glob[:, 0].cpu(), fixed_scale=False) * 1e2\n    jitter = compute_jitter(pred_j3d_glob, fps=30)\n    foot_sliding = compute_foot_sliding(target_verts_glob, pred_verts_glob) * m2mm\n\n    global_metrics = {\n        \"wa2_mpjpe\": wa2_mpjpe,\n        \"waa_mpjpe\": waa_mpjpe,\n        \"rte\": rte,\n        \"rte_scaled\": rte_scaled,\n        \"jitter\": jitter,\n        \"fs\": foot_sliding,\n    }\n    return global_metrics","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch","uri":"program://Human3R/module/eval.global_human.launch#L1-L765","kind":"module","name":"eval.global_human.launch","path":"eval/global_human/launch.py","language":"python","start_line":1,"end_line":765,"context_start_line":1,"context_end_line":765,"code":"import os\nimport sys\nimport gc\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nimport numpy as np\nimport torch\nimport argparse\n\nfrom copy import deepcopy\nfrom eval.global_human.metadata import dataset_metadata\nfrom eval.global_human.utils import *\n\nfrom accelerate import PartialState\nfrom add_ckpt_path import add_path_to_dust3r\n\nfrom collections import defaultdict, Counter\nfrom tqdm import tqdm\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--weights\",\n        type=str,\n        help=\"path to the model weights\",\n        default=\"\",\n    )\n    parser.add_argument(\"--device\", type=str, default=\"cuda\", help=\"pytorch device\")\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"\",\n        help=\"value for outdir\",\n    )\n    parser.add_argument(\n        \"--no_crop\", type=bool, default=True, help=\"whether to crop input data\"\n    )\n    parser.add_argument(\n        \"--eval_dataset\",\n        type=str,\n        default=\"bedlam\",\n        choices=list(dataset_metadata.keys()),\n    )\n\n    parser.add_argument(\"--crop_res\", type=int, nargs=2, metavar=(\"W\", \"H\"), default=None)\n    parser.add_argument(\"--size\", type=int, default=\"224\")\n    parser.add_argument(\"--shuffle\", action=\"store_true\", default=False)\n    parser.add_argument(\"--revisit\", type=int, default=1)\n    parser.add_argument(\"--freeze_state\", action=\"store_true\", default=False)\n    parser.add_argument(\"--solve_pose\", action=\"store_true\", default=False)\n    parser.add_argument(\"--vis\", action=\"store_true\", default=False)\n    parser.add_argument(\"--save\", action=\"store_true\", default=False)\n    parser.add_argument(\"--is_naive\", action=\"store_true\", default=False)\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    parser.add_argument(\"--use_fake_K\", action=\"store_true\", default=False)\n    return parser\n\ndef get_seq_list(metadata, img_path):\n    get_seq_func = metadata.get(\"get_seq_func\", None)\n    split = metadata.get(\"split\", \"\")\n    \n    if get_seq_func:\n        annots = metadata[\"get_annot_func\"](img_path, split)\n        seq_list, seq_to_images = get_seq_func(img_path, split, annots)\n        return seq_list, seq_to_images, annots\n    \n    if metadata.get(\"full_seq\", False):\n        seq_dir = f\"{img_path}/{split}\"\n        seq_list = [seq for seq in os.listdir(seq_dir) \n                   if os.path.isdir(os.path.join(seq_dir, seq))]\n        return sorted(seq_list), None, None\n    else:\n        return  sorted(metadata.get(\"seq_list\", [])), None, None\n\ndef get_file_list(metadata, img_path, seq, seq_to_images=None):\n    dir_path = metadata[\"dir_path_func\"](img_path, seq)\n    subsample = metadata.get(\"subsample\", 1)\n    max_frames = metadata.get(\"max_frames\", None)\n\n    if seq_to_images is not None:\n        filelist = [os.path.join(dir_path, name) for name in seq_to_images[seq]]\n    else:\n        filelist = [os.path.join(dir_path, name) for name in os.listdir(dir_path)]\n\n    filelist.sort()\n    \n    if max_frames is not None:\n        filelist = filelist[:max_frames]\n\n    sampled_indices = list(range(0, len(filelist), subsample))\n    filelist = filelist[::subsample]\n    return filelist, sampled_indices\n\ndef run_inference(\n        device, args, metadata, filelist, sampled_indices, annots, model, smpl_model):\n    from dust3r.inference import inference_recurrent_lighter\n    \n    get_view_func=metadata.get(\"get_view_func\", None)\n    mask_path_func = metadata.get(\"mask_path_func\", None)    \n    mask_path_list = mask_path_func(filelist) if mask_path_func is not None else []\n\n    views = prepare_input(\n        filelist,\n        [True for _ in filelist],\n        msk_paths=mask_path_list,\n        size=args.size,\n        crop=not args.no_crop,\n        revisit=args.revisit,\n        update=not args.freeze_state,\n        load_func=get_view_func,\n        annots=annots,\n        sampled_indices=sampled_indices,\n        reset_interval=args.reset_interval,\n        crop_res=args.crop_res\n    )\n    with torch.no_grad():\n        smpl_model.update_smpl_gt_eval(views, args.eval_dataset)     \n    gt = prepare_gt(views)\n\n    keep_keys = set(\n        [\"img\", \"img_mask\", \"true_shape\", \"img_mhmr\", \"reset\", \"update\", \"K_mhmr\"])\n    for i, view in enumerate(views):\n        views[i] = {key: view[key] for key in keep_keys if key in view}\n\n    outputs, _ = inference_recurrent_lighter(\n        views, model, device, verbose=False, is_naive=args.is_naive, use_ttt3r=args.use_ttt3r)\n\n    pred = prepare_output(\n        outputs, revisit=args.revisit, solve_pose=args.solve_pose, is_save=args.save)\n\n    del outputs, views\n    gc.collect()\n    torch.cuda.empty_cache()\n\n    return gt, pred\n\ndef get_pred_smpl(pred, gt, f_id, is_naive, smpl_layer, mhmr_img_res, K_to_proj):\n    n_humans_i = pred['shape'][f_id].shape[0]\n    expand = lambda x: x.expand(n_humans_i, -1, -1)\n\n    with torch.no_grad():\n        if is_naive:\n            dist = pred['transl'][f_id][:, 0].unsqueeze(-1)\n            dist = to_euclidean_dist(\n                mhmr_img_res, dist, expand(gt['K_mhmr'][f_id]))\n            smpl_out = smpl_layer(\n                pred['rotvec'][f_id], \n                pred['shape'][f_id], \n                None, \n                pred['loc'][f_id], \n                dist, \n                K=expand(gt['K_mhmr'][f_id]), \n                expression=pred['expression'][f_id],\n                K_to_proj=expand(gt['K'][f_id]),\n                )\n            pred['transl'][f_id] = smpl_out['smpl_transl']\n        else:\n            smpl_out = smpl_layer(\n                pred['rotvec'][f_id], \n                pred['shape'][f_id], \n                pred['transl'][f_id], \n                None, None, \n                K=expand(K_to_proj[f_id]), \n                expression=pred['expression'][f_id])\n    \n    return smpl_out['smpl_v3d']\n\ndef match_2d(pr_j2d, gt_j2d):\n    # match pred to gt - based on 2d bbox\n    gt_j2d = gt_j2d.numpy()\n    bestMatch, falsePositives, misses = match_2d_greedy(\n        pr_j2d.numpy()[:,:gt_j2d.shape[1]], \n        gt_j2d, \n        np.ones_like(gt_j2d[...,0]).astype(np.bool_))\n\n    update = {\n        'count': len(gt_j2d),\n        'miss': len(misses), \n        'fp': len(falsePositives)\n    }\n\n    return bestMatch, update\n\n\ndef eval_smpl_error(args, model, smpl_model, smpl_layer, save_dir=None):\n    from dust3r.utils.geometry import perspective_projection, geotrf\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mhmr_img_res = getattr(model, \"mhmr_img_res\", None)\n    subsample = metadata.get(\"subsample\", 1)\n    is_global = metadata[\"is_global\"](metadata.get(\"split\", \"\"))\n    pelvis_idx = smpl_model.pelvis_idx\n\n    seq_list, seq_to_images, annots = get_seq_list(metadata, img_path)\n\n    if save_dir is None:\n        save_dir = args.output_dir\n\n    distributed_state = PartialState()\n    model.to(distributed_state.device)\n    device = distributed_state.device\n\n    with distributed_state.split_between_processes(seq_list) as seqs:\n        if len(seq_list) < distributed_state.num_processes:\n            if distributed_state.process_index >= len(seq_list):\n                seqs = []\n        error_log_path = f\"{save_dir}/_error_log_{distributed_state.process_index}.txt\"\n\n        for seq_idx, seq in enumerate(tqdm(seqs)):\n            try:\n                print(f\"Evaluating sequence: {seq}\")\n                filelist, sampled_indices = get_file_list(metadata, img_path, seq, seq_to_images)\n                gt, pred = run_inference(\n                    device, args, metadata, filelist, sampled_indices, annots, model, smpl_model)\n\n                K_to_proj = gt['K'] if args.is_naive else pred['K'] # CHOOSE ONE in: pred['K'] or gt['K']\n                T_c2w = pred['T_c2w'] # CHOOSE ONE in: pred['T_c2w'] or gt['T_c2w']\n\n                global_batch = []\n                metrics = defaultdict(list)\n                counter = Counter()\n                for f_id in range(len(filelist)):\n                    n_humans_i = pred['shape'][f_id].shape[0]\n                    if n_humans_i > 0:\n                        pred_v3d_c = get_pred_smpl(\n                            pred, gt, f_id, args.is_naive, smpl_layer, mhmr_img_res, K_to_proj)\n\n                        pred_v3d_c = smpl_model.smplx2smpl @ pred_v3d_c\n                        pred_j3d_c = smpl_model.j_regressor @ pred_v3d_c\n                        pr_j2d = perspective_projection(\n                            pred_j3d_c, K_to_proj[f_id].expand(n_humans_i, -1 , -1))\n                    else:\n                        pred_v3d_c = torch.empty(0, 6890, 3, dtype=torch.float32)\n                        pr_j2d = torch.empty(0, 24, 2, dtype=torch.float32)\n                    \n                    bestMatch, update = match_2d(pr_j2d, gt['j2d'][f_id])\n                    counter.update(update)\n\n                    # 3d metrics\n                    if len(bestMatch) > 0:\n                        counter.update({'n_human': len(bestMatch)})\n                        pid, gid = bestMatch[:, 0], bestMatch[:, 1]\n\n                        # camera coordinate metrics\n                        camcoord_batch = {\n                            \"pred_j3d\": pred_j3d_c[pid],\n                            \"target_j3d\": gt['j3d_c'][f_id][gid],\n                            \"pred_v3d\": pred_v3d_c[pid],\n                            \"target_v3d\": gt['v3d_c'][f_id][gid],\n                        }\n                        camcoord_metrics = eval_camcoord(camcoord_batch, pelvis_idx)\n\n                        for k, v in camcoord_metrics.items():\n                            metrics[k].append(v)\n\n                        if is_global:\n                            expand = lambda x: x.expand(len(bestMatch), -1, -1)\n                            global_batch.append({\n                                \"pred_j3d\": geotrf(expand(T_c2w[f_id]), pred_j3d_c[pid]),\n                                \"target_j3d\": gt['j3d_w'][f_id][gid],\n                                \"pred_v3d\": geotrf(expand(T_c2w[f_id]), pred_v3d_c[pid]),\n                                \"target_v3d\": gt['v3d_w'][f_id][gid],\n                            })\n\n                    if args.save:\n                        color = 0.5 * (gt['img'][f_id].permute(1, 2, 0) + 1.0)\n                        out_dir = f\"{save_dir}/{seq}/{f_id:06d}\"\n                        for k in [\"pts3d\", \"conf\", \"color\", \"camera\", \"smpl\", \"mask\"]:\n                            os.makedirs(os.path.join(out_dir, k), exist_ok=True)\n                        np.save(os.path.join(out_dir, \"pts3d\", f\"{f_id:06d}.npy\"), pred['pts3d_self'][f_id])\n                        np.save(os.path.join(out_dir, \"conf\", f\"pred_{f_id:06d}.npy\"), pred['conf_self'][f_id])\n                        np.save(os.path.join(out_dir, \"color\", f\"{f_id:06d}.npy\"), color)\n                        np.save(os.path.join(out_dir, \"mask\", f\"pred_{f_id:06d}.npy\"), pred['msk'][f_id])\n                        np.savez(os.path.join(out_dir, \"camera\", f\"pred_{f_id:06d}.npz\"), \n                                pose=pred['T_c2w'][f_id], K=pred['K'][f_id])\n                        np.savez(os.path.join(out_dir, \"camera\", f\"gt_{f_id:06d}.npz\"), \n                                pose=gt['T_c2w'][f_id], K=gt['K'][f_id])\n                        if len(bestMatch) > 0:\n                            np.save(os.path.join(out_dir, \"smpl\", f\"pred_{f_id:06d}.npy\"), pred_v3d_c[pid])\n                            np.save(os.path.join(out_dir, \"smpl\", f\"gt_{f_id:06d}.npy\"), gt['v3d_c'][f_id][gid])\n                        else:\n                            np.save(os.path.join(out_dir, \"smpl\", f\"pred_{f_id:06d}.npy\"), pred_v3d_c)\n                            np.save(os.path.join(out_dir, \"smpl\", f\"gt_{f_id:06d}.npy\"), gt['v3d_c'][f_id])\n\n                    if args.vis:\n                        visualize(\n                            save_dir=f\"{save_dir}/{seq}\",\n                            img_path=filelist[f_id],\n                            view=gt['img'][f_id],\n                            gt_v3d_c=gt['v3d_c'][f_id],\n                            pred_v3d_c=pred_v3d_c,\n                            K_to_proj=K_to_proj[f_id],\n                            gt_K=gt['K'][f_id],\n                            bestMatch=bestMatch,\n                            smpl_face=smpl_model.smpl_faces['smpl'],\n                            )\n\n                metrics['precision'], metrics['recall'], metrics['f1_score']= compute_prf1(\n                    counter['count'], counter['miss'], counter['fp'])\n                 \n                # global coordinate metrics\n                if is_global:\n                    global_batch = {k: torch.cat([b[k] for b in global_batch]) for k in global_batch[0]}\n                    global_metrics = eval_global(global_batch, subsample)\n                    for k, v in global_metrics.items():\n                        metrics[k].append(v)\n                        \n                torch.cuda.empty_cache()\n\n                # Write to error log after each sequence\n                os.makedirs(save_dir, exist_ok=True)\n                write_log(error_log_path, args.eval_dataset, seq, counter, metrics)\n\n                del gt, pred, filelist, sampled_indices\n                del global_batch, metrics, counter\n                if 'global_metrics' in locals():\n                    del global_metrics\n                gc.collect()\n\n\n            except Exception as e:\n                print(f\"Exception in sequence {seq}: {str(e)}\")\n                if \"out of memory\" in str(e):\n                    torch.cuda.empty_cache()\n                    with open(error_log_path, \"a\") as f:\n                        f.write(\n                            f\"OOM error in sequence {seq}, skipping this sequence.\\n\"\n                        )\n                    print(f\"OOM error in sequence {seq}, skipping...\")\n                elif \"Degenerate covariance rank\" in str(\n                    e\n                ) or \"Eigenvalues did not converge\" in str(e):\n                    with open(error_log_path, \"a\") as f:\n                        f.write(f\"Exception in sequence {seq}: {str(e)}\\n\")\n                    print(f\"Traj evaluation error in sequence {seq}, skipping.\")\n                else:\n                    raise e\n\n    distributed_state.wait_for_everyone()\n    torch.cuda.empty_cache()\n\n    results = process_directory(save_dir)\n    summary = calculate_averages(results)\n\n    if distributed_state.is_main_process:\n        with open(f\"{save_dir}/_error_log.txt\", \"a\") as f:\n            for i in range(distributed_state.num_processes):\n                if not os.path.exists(f\"{save_dir}/_error_log_{i}.txt\"):\n                    break\n                with open(f\"{save_dir}/_error_log_{i}.txt\", \"r\") as f_sub:\n                    f.write(f_sub.read())\n\n            log = get_summary_log(summary)\n            f.write(log) \n    \n        print(log.strip())\n        \n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    add_path_to_dust3r(args.weights)\n    from dust3r.utils.image import load_images_for_eval as load_images\n    from dust3r.utils.image import load_masks_for_eval as load_masks\n    from dust3r.post_process import estimate_focal_knowing_depth\n    from dust3r.model import ARCroco3DStereo\n    from dust3r.utils.camera import pose_encoding_to_camera\n    from dust3r.utils.geometry import weighted_procrustes, to_euclidean_dist, matrix_cumprod\n    from dust3r.smpl_model import SMPLModel\n    from dust3r.utils import SMPL_Layer\n    from dust3r.utils.image import unpad_image\n\n    args.no_crop = False\n\n    def recover_cam_params(pts3ds_self, pts3ds_other, conf_self, conf_other):\n        B, H, W, _ = pts3ds_self.shape\n        pp = (\n            torch.tensor([W // 2, H // 2], device=pts3ds_self.device)\n            .float()\n            .repeat(B, 1)\n            .reshape(B, 1, 2)\n        )\n        focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n        pts3ds_self = pts3ds_self.reshape(B, -1, 3)\n        pts3ds_other = pts3ds_other.reshape(B, -1, 3)\n        conf_self = conf_self.reshape(B, -1)\n        conf_other = conf_other.reshape(B, -1)\n        # weighted procrustes\n        c2w = weighted_procrustes(\n            pts3ds_self,\n            pts3ds_other,\n            torch.log(conf_self) * torch.log(conf_other),\n            use_weights=True,\n            return_T=True,\n        )\n        return c2w, focal, pp.reshape(B, 2)\n    \n    def _crop_resize(image, intrinsics, crop_res):\n        import dust3r.datasets.utils.cropping as cropping\n        from dust3r.utils.image import ImgNorm\n        \n        # image is a tensor in CHW with values in [-1, 1]; convert to HWC uint8 for PIL\n        img_device = image.device if isinstance(image, torch.Tensor) else None\n        had_batch_dim = False\n        if isinstance(image, torch.Tensor):\n            # accept [3,H,W] or [1,3,H,W]; squeeze batch if present\n            if image.dim() == 4:\n                assert image.shape[0] == 1, \"_crop_resize expects a single image; got a batch\"\n                image = image.squeeze(0)\n                had_batch_dim = True\n            elif image.dim() != 3:\n                raise RuntimeError(f\"Unexpected image tensor shape {tuple(image.shape)}; expected CHW or 1xCHW\")\n\n            image_np = (\n                (image.detach().cpu().permute(1, 2, 0).numpy() * 0.5 +0.5) * 255\n            ).clip(0, 255).astype(np.uint8)\n        else:\n            image_np = image\n\n        target_resolution = np.array(crop_res)\n        image_pil, _, intrinsics = cropping.rescale_image_depthmap(\n            image_np, None, intrinsics, target_resolution\n        )\n\n        # actual cropping (if necessary) with bilinear interpolation\n        intrinsics2 = cropping.camera_matrix_of_crop(\n            intrinsics, image_pil.size, crop_res, offset_factor=0.5\n        )\n        crop_bbox = cropping.bbox_from_intrinsics_in_out(\n            intrinsics, intrinsics2, crop_res\n        )\n        image_pil, _, intrinsics2 = cropping.crop_image_depthmap(\n            image_pil, None, intrinsics, crop_bbox\n        )\n\n        # convert back to normalized CHW tensor on original device\n        image_arr = np.array(image_pil)\n        if image_arr.ndim == 2:\n            image_arr = np.repeat(image_arr[..., None], 3, axis=2)\n        image_tensor = ImgNorm(image_pil)  # CHW, [-1, 1]\n        if had_batch_dim:\n            image_tensor = image_tensor.unsqueeze(0)\n        if img_device is not None:\n            image_tensor = image_tensor.to(img_device)\n\n        # return intrinsics as torch tensor with batch dim like upstream expects\n        intrinsics_tensor = torch.from_numpy(intrinsics2).unsqueeze(0)\n\n        return image_tensor, intrinsics_tensor\n\n    def prepare_input(\n        img_paths,\n        img_mask,\n        msk_paths,\n        size,\n        raymaps=None,\n        raymap_mask=None,\n        revisit=1,\n        update=True,\n        crop=True,\n        load_func=None,\n        annots=None,\n        sampled_indices=None,\n        reset_interval=100,\n        crop_res=None\n    ):\n        images = load_images(img_paths, size=size, verbose=False, crop=crop)\n        images = load_func((img_paths, images, annots, sampled_indices))\n\n        has_msk = len(msk_paths) > 0\n        if has_msk:\n            msks = load_masks(msk_paths, size=size, verbose=False, crop=crop)\n\n        views = []\n        if raymaps is None and raymap_mask is None:\n            num_views = len(images)\n\n            for i in range(num_views):\n                view = {\n                    \"img\": images[i][\"img\"],\n                    \"ray_map\": torch.full(\n                        (\n                            images[i][\"img\"].shape[0],\n                            6,\n                            images[i][\"img\"].shape[-2],\n                            images[i][\"img\"].shape[-1],\n                        ),\n                        torch.nan,\n# ... truncated ...","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":true}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.get_args_parser","uri":"program://Human3R/function/eval.global_human.launch.get_args_parser#L22-L60","kind":"function","name":"get_args_parser","path":"eval/global_human/launch.py","language":"python","start_line":22,"end_line":60,"context_start_line":2,"context_end_line":80,"code":"import sys\nimport gc\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nimport numpy as np\nimport torch\nimport argparse\n\nfrom copy import deepcopy\nfrom eval.global_human.metadata import dataset_metadata\nfrom eval.global_human.utils import *\n\nfrom accelerate import PartialState\nfrom add_ckpt_path import add_path_to_dust3r\n\nfrom collections import defaultdict, Counter\nfrom tqdm import tqdm\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--weights\",\n        type=str,\n        help=\"path to the model weights\",\n        default=\"\",\n    )\n    parser.add_argument(\"--device\", type=str, default=\"cuda\", help=\"pytorch device\")\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,\n        default=\"\",\n        help=\"value for outdir\",\n    )\n    parser.add_argument(\n        \"--no_crop\", type=bool, default=True, help=\"whether to crop input data\"\n    )\n    parser.add_argument(\n        \"--eval_dataset\",\n        type=str,\n        default=\"bedlam\",\n        choices=list(dataset_metadata.keys()),\n    )\n\n    parser.add_argument(\"--crop_res\", type=int, nargs=2, metavar=(\"W\", \"H\"), default=None)\n    parser.add_argument(\"--size\", type=int, default=\"224\")\n    parser.add_argument(\"--shuffle\", action=\"store_true\", default=False)\n    parser.add_argument(\"--revisit\", type=int, default=1)\n    parser.add_argument(\"--freeze_state\", action=\"store_true\", default=False)\n    parser.add_argument(\"--solve_pose\", action=\"store_true\", default=False)\n    parser.add_argument(\"--vis\", action=\"store_true\", default=False)\n    parser.add_argument(\"--save\", action=\"store_true\", default=False)\n    parser.add_argument(\"--is_naive\", action=\"store_true\", default=False)\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    parser.add_argument(\"--use_fake_K\", action=\"store_true\", default=False)\n    return parser\n\ndef get_seq_list(metadata, img_path):\n    get_seq_func = metadata.get(\"get_seq_func\", None)\n    split = metadata.get(\"split\", \"\")\n    \n    if get_seq_func:\n        annots = metadata[\"get_annot_func\"](img_path, split)\n        seq_list, seq_to_images = get_seq_func(img_path, split, annots)\n        return seq_list, seq_to_images, annots\n    \n    if metadata.get(\"full_seq\", False):\n        seq_dir = f\"{img_path}/{split}\"\n        seq_list = [seq for seq in os.listdir(seq_dir) \n                   if os.path.isdir(os.path.join(seq_dir, seq))]\n        return sorted(seq_list), None, None\n    else:\n        return  sorted(metadata.get(\"seq_list\", [])), None, None\n\ndef get_file_list(metadata, img_path, seq, seq_to_images=None):\n    dir_path = metadata[\"dir_path_func\"](img_path, seq)","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.get_seq_list","uri":"program://Human3R/function/eval.global_human.launch.get_seq_list#L62-L77","kind":"function","name":"get_seq_list","path":"eval/global_human/launch.py","language":"python","start_line":62,"end_line":77,"context_start_line":42,"context_end_line":97,"code":"        \"--eval_dataset\",\n        type=str,\n        default=\"bedlam\",\n        choices=list(dataset_metadata.keys()),\n    )\n\n    parser.add_argument(\"--crop_res\", type=int, nargs=2, metavar=(\"W\", \"H\"), default=None)\n    parser.add_argument(\"--size\", type=int, default=\"224\")\n    parser.add_argument(\"--shuffle\", action=\"store_true\", default=False)\n    parser.add_argument(\"--revisit\", type=int, default=1)\n    parser.add_argument(\"--freeze_state\", action=\"store_true\", default=False)\n    parser.add_argument(\"--solve_pose\", action=\"store_true\", default=False)\n    parser.add_argument(\"--vis\", action=\"store_true\", default=False)\n    parser.add_argument(\"--save\", action=\"store_true\", default=False)\n    parser.add_argument(\"--is_naive\", action=\"store_true\", default=False)\n    parser.add_argument(\"--use_ttt3r\", action=\"store_true\", default=False)\n    parser.add_argument(\"--reset_interval\", type=int, default=100000000)\n    parser.add_argument(\"--use_fake_K\", action=\"store_true\", default=False)\n    return parser\n\ndef get_seq_list(metadata, img_path):\n    get_seq_func = metadata.get(\"get_seq_func\", None)\n    split = metadata.get(\"split\", \"\")\n    \n    if get_seq_func:\n        annots = metadata[\"get_annot_func\"](img_path, split)\n        seq_list, seq_to_images = get_seq_func(img_path, split, annots)\n        return seq_list, seq_to_images, annots\n    \n    if metadata.get(\"full_seq\", False):\n        seq_dir = f\"{img_path}/{split}\"\n        seq_list = [seq for seq in os.listdir(seq_dir) \n                   if os.path.isdir(os.path.join(seq_dir, seq))]\n        return sorted(seq_list), None, None\n    else:\n        return  sorted(metadata.get(\"seq_list\", [])), None, None\n\ndef get_file_list(metadata, img_path, seq, seq_to_images=None):\n    dir_path = metadata[\"dir_path_func\"](img_path, seq)\n    subsample = metadata.get(\"subsample\", 1)\n    max_frames = metadata.get(\"max_frames\", None)\n\n    if seq_to_images is not None:\n        filelist = [os.path.join(dir_path, name) for name in seq_to_images[seq]]\n    else:\n        filelist = [os.path.join(dir_path, name) for name in os.listdir(dir_path)]\n\n    filelist.sort()\n    \n    if max_frames is not None:\n        filelist = filelist[:max_frames]\n\n    sampled_indices = list(range(0, len(filelist), subsample))\n    filelist = filelist[::subsample]\n    return filelist, sampled_indices\n","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.get_file_list","uri":"program://Human3R/function/eval.global_human.launch.get_file_list#L79-L96","kind":"function","name":"get_file_list","path":"eval/global_human/launch.py","language":"python","start_line":79,"end_line":96,"context_start_line":59,"context_end_line":116,"code":"    parser.add_argument(\"--use_fake_K\", action=\"store_true\", default=False)\n    return parser\n\ndef get_seq_list(metadata, img_path):\n    get_seq_func = metadata.get(\"get_seq_func\", None)\n    split = metadata.get(\"split\", \"\")\n    \n    if get_seq_func:\n        annots = metadata[\"get_annot_func\"](img_path, split)\n        seq_list, seq_to_images = get_seq_func(img_path, split, annots)\n        return seq_list, seq_to_images, annots\n    \n    if metadata.get(\"full_seq\", False):\n        seq_dir = f\"{img_path}/{split}\"\n        seq_list = [seq for seq in os.listdir(seq_dir) \n                   if os.path.isdir(os.path.join(seq_dir, seq))]\n        return sorted(seq_list), None, None\n    else:\n        return  sorted(metadata.get(\"seq_list\", [])), None, None\n\ndef get_file_list(metadata, img_path, seq, seq_to_images=None):\n    dir_path = metadata[\"dir_path_func\"](img_path, seq)\n    subsample = metadata.get(\"subsample\", 1)\n    max_frames = metadata.get(\"max_frames\", None)\n\n    if seq_to_images is not None:\n        filelist = [os.path.join(dir_path, name) for name in seq_to_images[seq]]\n    else:\n        filelist = [os.path.join(dir_path, name) for name in os.listdir(dir_path)]\n\n    filelist.sort()\n    \n    if max_frames is not None:\n        filelist = filelist[:max_frames]\n\n    sampled_indices = list(range(0, len(filelist), subsample))\n    filelist = filelist[::subsample]\n    return filelist, sampled_indices\n\ndef run_inference(\n        device, args, metadata, filelist, sampled_indices, annots, model, smpl_model):\n    from dust3r.inference import inference_recurrent_lighter\n    \n    get_view_func=metadata.get(\"get_view_func\", None)\n    mask_path_func = metadata.get(\"mask_path_func\", None)    \n    mask_path_list = mask_path_func(filelist) if mask_path_func is not None else []\n\n    views = prepare_input(\n        filelist,\n        [True for _ in filelist],\n        msk_paths=mask_path_list,\n        size=args.size,\n        crop=not args.no_crop,\n        revisit=args.revisit,\n        update=not args.freeze_state,\n        load_func=get_view_func,\n        annots=annots,\n        sampled_indices=sampled_indices,","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.run_inference","uri":"program://Human3R/function/eval.global_human.launch.run_inference#L98-L139","kind":"function","name":"run_inference","path":"eval/global_human/launch.py","language":"python","start_line":98,"end_line":139,"context_start_line":78,"context_end_line":159,"code":"\ndef get_file_list(metadata, img_path, seq, seq_to_images=None):\n    dir_path = metadata[\"dir_path_func\"](img_path, seq)\n    subsample = metadata.get(\"subsample\", 1)\n    max_frames = metadata.get(\"max_frames\", None)\n\n    if seq_to_images is not None:\n        filelist = [os.path.join(dir_path, name) for name in seq_to_images[seq]]\n    else:\n        filelist = [os.path.join(dir_path, name) for name in os.listdir(dir_path)]\n\n    filelist.sort()\n    \n    if max_frames is not None:\n        filelist = filelist[:max_frames]\n\n    sampled_indices = list(range(0, len(filelist), subsample))\n    filelist = filelist[::subsample]\n    return filelist, sampled_indices\n\ndef run_inference(\n        device, args, metadata, filelist, sampled_indices, annots, model, smpl_model):\n    from dust3r.inference import inference_recurrent_lighter\n    \n    get_view_func=metadata.get(\"get_view_func\", None)\n    mask_path_func = metadata.get(\"mask_path_func\", None)    \n    mask_path_list = mask_path_func(filelist) if mask_path_func is not None else []\n\n    views = prepare_input(\n        filelist,\n        [True for _ in filelist],\n        msk_paths=mask_path_list,\n        size=args.size,\n        crop=not args.no_crop,\n        revisit=args.revisit,\n        update=not args.freeze_state,\n        load_func=get_view_func,\n        annots=annots,\n        sampled_indices=sampled_indices,\n        reset_interval=args.reset_interval,\n        crop_res=args.crop_res\n    )\n    with torch.no_grad():\n        smpl_model.update_smpl_gt_eval(views, args.eval_dataset)     \n    gt = prepare_gt(views)\n\n    keep_keys = set(\n        [\"img\", \"img_mask\", \"true_shape\", \"img_mhmr\", \"reset\", \"update\", \"K_mhmr\"])\n    for i, view in enumerate(views):\n        views[i] = {key: view[key] for key in keep_keys if key in view}\n\n    outputs, _ = inference_recurrent_lighter(\n        views, model, device, verbose=False, is_naive=args.is_naive, use_ttt3r=args.use_ttt3r)\n\n    pred = prepare_output(\n        outputs, revisit=args.revisit, solve_pose=args.solve_pose, is_save=args.save)\n\n    del outputs, views\n    gc.collect()\n    torch.cuda.empty_cache()\n\n    return gt, pred\n\ndef get_pred_smpl(pred, gt, f_id, is_naive, smpl_layer, mhmr_img_res, K_to_proj):\n    n_humans_i = pred['shape'][f_id].shape[0]\n    expand = lambda x: x.expand(n_humans_i, -1, -1)\n\n    with torch.no_grad():\n        if is_naive:\n            dist = pred['transl'][f_id][:, 0].unsqueeze(-1)\n            dist = to_euclidean_dist(\n                mhmr_img_res, dist, expand(gt['K_mhmr'][f_id]))\n            smpl_out = smpl_layer(\n                pred['rotvec'][f_id], \n                pred['shape'][f_id], \n                None, \n                pred['loc'][f_id], \n                dist, \n                K=expand(gt['K_mhmr'][f_id]), \n                expression=pred['expression'][f_id],\n                K_to_proj=expand(gt['K'][f_id]),\n                )","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.get_pred_smpl","uri":"program://Human3R/function/eval.global_human.launch.get_pred_smpl#L141-L170","kind":"function","name":"get_pred_smpl","path":"eval/global_human/launch.py","language":"python","start_line":141,"end_line":170,"context_start_line":121,"context_end_line":190,"code":"        smpl_model.update_smpl_gt_eval(views, args.eval_dataset)     \n    gt = prepare_gt(views)\n\n    keep_keys = set(\n        [\"img\", \"img_mask\", \"true_shape\", \"img_mhmr\", \"reset\", \"update\", \"K_mhmr\"])\n    for i, view in enumerate(views):\n        views[i] = {key: view[key] for key in keep_keys if key in view}\n\n    outputs, _ = inference_recurrent_lighter(\n        views, model, device, verbose=False, is_naive=args.is_naive, use_ttt3r=args.use_ttt3r)\n\n    pred = prepare_output(\n        outputs, revisit=args.revisit, solve_pose=args.solve_pose, is_save=args.save)\n\n    del outputs, views\n    gc.collect()\n    torch.cuda.empty_cache()\n\n    return gt, pred\n\ndef get_pred_smpl(pred, gt, f_id, is_naive, smpl_layer, mhmr_img_res, K_to_proj):\n    n_humans_i = pred['shape'][f_id].shape[0]\n    expand = lambda x: x.expand(n_humans_i, -1, -1)\n\n    with torch.no_grad():\n        if is_naive:\n            dist = pred['transl'][f_id][:, 0].unsqueeze(-1)\n            dist = to_euclidean_dist(\n                mhmr_img_res, dist, expand(gt['K_mhmr'][f_id]))\n            smpl_out = smpl_layer(\n                pred['rotvec'][f_id], \n                pred['shape'][f_id], \n                None, \n                pred['loc'][f_id], \n                dist, \n                K=expand(gt['K_mhmr'][f_id]), \n                expression=pred['expression'][f_id],\n                K_to_proj=expand(gt['K'][f_id]),\n                )\n            pred['transl'][f_id] = smpl_out['smpl_transl']\n        else:\n            smpl_out = smpl_layer(\n                pred['rotvec'][f_id], \n                pred['shape'][f_id], \n                pred['transl'][f_id], \n                None, None, \n                K=expand(K_to_proj[f_id]), \n                expression=pred['expression'][f_id])\n    \n    return smpl_out['smpl_v3d']\n\ndef match_2d(pr_j2d, gt_j2d):\n    # match pred to gt - based on 2d bbox\n    gt_j2d = gt_j2d.numpy()\n    bestMatch, falsePositives, misses = match_2d_greedy(\n        pr_j2d.numpy()[:,:gt_j2d.shape[1]], \n        gt_j2d, \n        np.ones_like(gt_j2d[...,0]).astype(np.bool_))\n\n    update = {\n        'count': len(gt_j2d),\n        'miss': len(misses), \n        'fp': len(falsePositives)\n    }\n\n    return bestMatch, update\n\n\ndef eval_smpl_error(args, model, smpl_model, smpl_layer, save_dir=None):\n    from dust3r.utils.geometry import perspective_projection, geotrf","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.match_2d","uri":"program://Human3R/function/eval.global_human.launch.match_2d#L172-L186","kind":"function","name":"match_2d","path":"eval/global_human/launch.py","language":"python","start_line":172,"end_line":186,"context_start_line":152,"context_end_line":206,"code":"                pred['shape'][f_id], \n                None, \n                pred['loc'][f_id], \n                dist, \n                K=expand(gt['K_mhmr'][f_id]), \n                expression=pred['expression'][f_id],\n                K_to_proj=expand(gt['K'][f_id]),\n                )\n            pred['transl'][f_id] = smpl_out['smpl_transl']\n        else:\n            smpl_out = smpl_layer(\n                pred['rotvec'][f_id], \n                pred['shape'][f_id], \n                pred['transl'][f_id], \n                None, None, \n                K=expand(K_to_proj[f_id]), \n                expression=pred['expression'][f_id])\n    \n    return smpl_out['smpl_v3d']\n\ndef match_2d(pr_j2d, gt_j2d):\n    # match pred to gt - based on 2d bbox\n    gt_j2d = gt_j2d.numpy()\n    bestMatch, falsePositives, misses = match_2d_greedy(\n        pr_j2d.numpy()[:,:gt_j2d.shape[1]], \n        gt_j2d, \n        np.ones_like(gt_j2d[...,0]).astype(np.bool_))\n\n    update = {\n        'count': len(gt_j2d),\n        'miss': len(misses), \n        'fp': len(falsePositives)\n    }\n\n    return bestMatch, update\n\n\ndef eval_smpl_error(args, model, smpl_model, smpl_layer, save_dir=None):\n    from dust3r.utils.geometry import perspective_projection, geotrf\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mhmr_img_res = getattr(model, \"mhmr_img_res\", None)\n    subsample = metadata.get(\"subsample\", 1)\n    is_global = metadata[\"is_global\"](metadata.get(\"split\", \"\"))\n    pelvis_idx = smpl_model.pelvis_idx\n\n    seq_list, seq_to_images, annots = get_seq_list(metadata, img_path)\n\n    if save_dir is None:\n        save_dir = args.output_dir\n\n    distributed_state = PartialState()\n    model.to(distributed_state.device)\n    device = distributed_state.device","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.eval_smpl_error","uri":"program://Human3R/function/eval.global_human.launch.eval_smpl_error#L189-L361","kind":"function","name":"eval_smpl_error","path":"eval/global_human/launch.py","language":"python","start_line":189,"end_line":361,"context_start_line":169,"context_end_line":381,"code":"    \n    return smpl_out['smpl_v3d']\n\ndef match_2d(pr_j2d, gt_j2d):\n    # match pred to gt - based on 2d bbox\n    gt_j2d = gt_j2d.numpy()\n    bestMatch, falsePositives, misses = match_2d_greedy(\n        pr_j2d.numpy()[:,:gt_j2d.shape[1]], \n        gt_j2d, \n        np.ones_like(gt_j2d[...,0]).astype(np.bool_))\n\n    update = {\n        'count': len(gt_j2d),\n        'miss': len(misses), \n        'fp': len(falsePositives)\n    }\n\n    return bestMatch, update\n\n\ndef eval_smpl_error(args, model, smpl_model, smpl_layer, save_dir=None):\n    from dust3r.utils.geometry import perspective_projection, geotrf\n\n    metadata = dataset_metadata.get(args.eval_dataset)\n    img_path = metadata[\"img_path\"]\n    mhmr_img_res = getattr(model, \"mhmr_img_res\", None)\n    subsample = metadata.get(\"subsample\", 1)\n    is_global = metadata[\"is_global\"](metadata.get(\"split\", \"\"))\n    pelvis_idx = smpl_model.pelvis_idx\n\n    seq_list, seq_to_images, annots = get_seq_list(metadata, img_path)\n\n    if save_dir is None:\n        save_dir = args.output_dir\n\n    distributed_state = PartialState()\n    model.to(distributed_state.device)\n    device = distributed_state.device\n\n    with distributed_state.split_between_processes(seq_list) as seqs:\n        if len(seq_list) < distributed_state.num_processes:\n            if distributed_state.process_index >= len(seq_list):\n                seqs = []\n        error_log_path = f\"{save_dir}/_error_log_{distributed_state.process_index}.txt\"\n\n        for seq_idx, seq in enumerate(tqdm(seqs)):\n            try:\n                print(f\"Evaluating sequence: {seq}\")\n                filelist, sampled_indices = get_file_list(metadata, img_path, seq, seq_to_images)\n                gt, pred = run_inference(\n                    device, args, metadata, filelist, sampled_indices, annots, model, smpl_model)\n\n                K_to_proj = gt['K'] if args.is_naive else pred['K'] # CHOOSE ONE in: pred['K'] or gt['K']\n                T_c2w = pred['T_c2w'] # CHOOSE ONE in: pred['T_c2w'] or gt['T_c2w']\n\n                global_batch = []\n                metrics = defaultdict(list)\n                counter = Counter()\n                for f_id in range(len(filelist)):\n                    n_humans_i = pred['shape'][f_id].shape[0]\n                    if n_humans_i > 0:\n                        pred_v3d_c = get_pred_smpl(\n                            pred, gt, f_id, args.is_naive, smpl_layer, mhmr_img_res, K_to_proj)\n\n                        pred_v3d_c = smpl_model.smplx2smpl @ pred_v3d_c\n                        pred_j3d_c = smpl_model.j_regressor @ pred_v3d_c\n                        pr_j2d = perspective_projection(\n                            pred_j3d_c, K_to_proj[f_id].expand(n_humans_i, -1 , -1))\n                    else:\n                        pred_v3d_c = torch.empty(0, 6890, 3, dtype=torch.float32)\n                        pr_j2d = torch.empty(0, 24, 2, dtype=torch.float32)\n                    \n                    bestMatch, update = match_2d(pr_j2d, gt['j2d'][f_id])\n                    counter.update(update)\n\n                    # 3d metrics\n                    if len(bestMatch) > 0:\n                        counter.update({'n_human': len(bestMatch)})\n                        pid, gid = bestMatch[:, 0], bestMatch[:, 1]\n\n                        # camera coordinate metrics\n                        camcoord_batch = {\n                            \"pred_j3d\": pred_j3d_c[pid],\n                            \"target_j3d\": gt['j3d_c'][f_id][gid],\n                            \"pred_v3d\": pred_v3d_c[pid],\n                            \"target_v3d\": gt['v3d_c'][f_id][gid],\n                        }\n                        camcoord_metrics = eval_camcoord(camcoord_batch, pelvis_idx)\n\n                        for k, v in camcoord_metrics.items():\n                            metrics[k].append(v)\n\n                        if is_global:\n                            expand = lambda x: x.expand(len(bestMatch), -1, -1)\n                            global_batch.append({\n                                \"pred_j3d\": geotrf(expand(T_c2w[f_id]), pred_j3d_c[pid]),\n                                \"target_j3d\": gt['j3d_w'][f_id][gid],\n                                \"pred_v3d\": geotrf(expand(T_c2w[f_id]), pred_v3d_c[pid]),\n                                \"target_v3d\": gt['v3d_w'][f_id][gid],\n                            })\n\n                    if args.save:\n                        color = 0.5 * (gt['img'][f_id].permute(1, 2, 0) + 1.0)\n                        out_dir = f\"{save_dir}/{seq}/{f_id:06d}\"\n                        for k in [\"pts3d\", \"conf\", \"color\", \"camera\", \"smpl\", \"mask\"]:\n                            os.makedirs(os.path.join(out_dir, k), exist_ok=True)\n                        np.save(os.path.join(out_dir, \"pts3d\", f\"{f_id:06d}.npy\"), pred['pts3d_self'][f_id])\n                        np.save(os.path.join(out_dir, \"conf\", f\"pred_{f_id:06d}.npy\"), pred['conf_self'][f_id])\n                        np.save(os.path.join(out_dir, \"color\", f\"{f_id:06d}.npy\"), color)\n                        np.save(os.path.join(out_dir, \"mask\", f\"pred_{f_id:06d}.npy\"), pred['msk'][f_id])\n                        np.savez(os.path.join(out_dir, \"camera\", f\"pred_{f_id:06d}.npz\"), \n                                pose=pred['T_c2w'][f_id], K=pred['K'][f_id])\n                        np.savez(os.path.join(out_dir, \"camera\", f\"gt_{f_id:06d}.npz\"), \n                                pose=gt['T_c2w'][f_id], K=gt['K'][f_id])\n                        if len(bestMatch) > 0:\n                            np.save(os.path.join(out_dir, \"smpl\", f\"pred_{f_id:06d}.npy\"), pred_v3d_c[pid])\n                            np.save(os.path.join(out_dir, \"smpl\", f\"gt_{f_id:06d}.npy\"), gt['v3d_c'][f_id][gid])\n                        else:\n                            np.save(os.path.join(out_dir, \"smpl\", f\"pred_{f_id:06d}.npy\"), pred_v3d_c)\n                            np.save(os.path.join(out_dir, \"smpl\", f\"gt_{f_id:06d}.npy\"), gt['v3d_c'][f_id])\n\n                    if args.vis:\n                        visualize(\n                            save_dir=f\"{save_dir}/{seq}\",\n                            img_path=filelist[f_id],\n                            view=gt['img'][f_id],\n                            gt_v3d_c=gt['v3d_c'][f_id],\n                            pred_v3d_c=pred_v3d_c,\n                            K_to_proj=K_to_proj[f_id],\n                            gt_K=gt['K'][f_id],\n                            bestMatch=bestMatch,\n                            smpl_face=smpl_model.smpl_faces['smpl'],\n                            )\n\n                metrics['precision'], metrics['recall'], metrics['f1_score']= compute_prf1(\n                    counter['count'], counter['miss'], counter['fp'])\n                 \n                # global coordinate metrics\n                if is_global:\n                    global_batch = {k: torch.cat([b[k] for b in global_batch]) for k in global_batch[0]}\n                    global_metrics = eval_global(global_batch, subsample)\n                    for k, v in global_metrics.items():\n                        metrics[k].append(v)\n                        \n                torch.cuda.empty_cache()\n\n                # Write to error log after each sequence\n                os.makedirs(save_dir, exist_ok=True)\n                write_log(error_log_path, args.eval_dataset, seq, counter, metrics)\n\n                del gt, pred, filelist, sampled_indices\n                del global_batch, metrics, counter\n                if 'global_metrics' in locals():\n                    del global_metrics\n                gc.collect()\n\n\n            except Exception as e:\n                print(f\"Exception in sequence {seq}: {str(e)}\")\n                if \"out of memory\" in str(e):\n                    torch.cuda.empty_cache()\n                    with open(error_log_path, \"a\") as f:\n                        f.write(\n                            f\"OOM error in sequence {seq}, skipping this sequence.\\n\"\n                        )\n                    print(f\"OOM error in sequence {seq}, skipping...\")\n                elif \"Degenerate covariance rank\" in str(\n                    e\n                ) or \"Eigenvalues did not converge\" in str(e):\n                    with open(error_log_path, \"a\") as f:\n                        f.write(f\"Exception in sequence {seq}: {str(e)}\\n\")\n                    print(f\"Traj evaluation error in sequence {seq}, skipping.\")\n                else:\n                    raise e\n\n    distributed_state.wait_for_everyone()\n    torch.cuda.empty_cache()\n\n    results = process_directory(save_dir)\n    summary = calculate_averages(results)\n\n    if distributed_state.is_main_process:\n        with open(f\"{save_dir}/_error_log.txt\", \"a\") as f:\n            for i in range(distributed_state.num_processes):\n                if not os.path.exists(f\"{save_dir}/_error_log_{i}.txt\"):\n                    break\n                with open(f\"{save_dir}/_error_log_{i}.txt\", \"r\") as f_sub:\n                    f.write(f_sub.read())\n\n            log = get_summary_log(summary)\n            f.write(log) \n    \n        print(log.strip())\n        \n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    add_path_to_dust3r(args.weights)\n    from dust3r.utils.image import load_images_for_eval as load_images\n    from dust3r.utils.image import load_masks_for_eval as load_masks\n    from dust3r.post_process import estimate_focal_knowing_depth\n    from dust3r.model import ARCroco3DStereo\n    from dust3r.utils.camera import pose_encoding_to_camera\n    from dust3r.utils.geometry import weighted_procrustes, to_euclidean_dist, matrix_cumprod\n    from dust3r.smpl_model import SMPLModel\n    from dust3r.utils import SMPL_Layer\n    from dust3r.utils.image import unpad_image\n\n    args.no_crop = False\n\n    def recover_cam_params(pts3ds_self, pts3ds_other, conf_self, conf_other):","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.recover_cam_params","uri":"program://Human3R/function/eval.global_human.launch.recover_cam_params#L381-L403","kind":"function","name":"recover_cam_params","path":"eval/global_human/launch.py","language":"python","start_line":381,"end_line":403,"context_start_line":361,"context_end_line":423,"code":"        print(log.strip())\n        \n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    add_path_to_dust3r(args.weights)\n    from dust3r.utils.image import load_images_for_eval as load_images\n    from dust3r.utils.image import load_masks_for_eval as load_masks\n    from dust3r.post_process import estimate_focal_knowing_depth\n    from dust3r.model import ARCroco3DStereo\n    from dust3r.utils.camera import pose_encoding_to_camera\n    from dust3r.utils.geometry import weighted_procrustes, to_euclidean_dist, matrix_cumprod\n    from dust3r.smpl_model import SMPLModel\n    from dust3r.utils import SMPL_Layer\n    from dust3r.utils.image import unpad_image\n\n    args.no_crop = False\n\n    def recover_cam_params(pts3ds_self, pts3ds_other, conf_self, conf_other):\n        B, H, W, _ = pts3ds_self.shape\n        pp = (\n            torch.tensor([W // 2, H // 2], device=pts3ds_self.device)\n            .float()\n            .repeat(B, 1)\n            .reshape(B, 1, 2)\n        )\n        focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n        pts3ds_self = pts3ds_self.reshape(B, -1, 3)\n        pts3ds_other = pts3ds_other.reshape(B, -1, 3)\n        conf_self = conf_self.reshape(B, -1)\n        conf_other = conf_other.reshape(B, -1)\n        # weighted procrustes\n        c2w = weighted_procrustes(\n            pts3ds_self,\n            pts3ds_other,\n            torch.log(conf_self) * torch.log(conf_other),\n            use_weights=True,\n            return_T=True,\n        )\n        return c2w, focal, pp.reshape(B, 2)\n    \n    def _crop_resize(image, intrinsics, crop_res):\n        import dust3r.datasets.utils.cropping as cropping\n        from dust3r.utils.image import ImgNorm\n        \n        # image is a tensor in CHW with values in [-1, 1]; convert to HWC uint8 for PIL\n        img_device = image.device if isinstance(image, torch.Tensor) else None\n        had_batch_dim = False\n        if isinstance(image, torch.Tensor):\n            # accept [3,H,W] or [1,3,H,W]; squeeze batch if present\n            if image.dim() == 4:\n                assert image.shape[0] == 1, \"_crop_resize expects a single image; got a batch\"\n                image = image.squeeze(0)\n                had_batch_dim = True\n            elif image.dim() != 3:\n                raise RuntimeError(f\"Unexpected image tensor shape {tuple(image.shape)}; expected CHW or 1xCHW\")\n\n            image_np = (\n                (image.detach().cpu().permute(1, 2, 0).numpy() * 0.5 +0.5) * 255\n            ).clip(0, 255).astype(np.uint8)","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch._crop_resize","uri":"program://Human3R/function/eval.global_human.launch._crop_resize#L405-L456","kind":"function","name":"_crop_resize","path":"eval/global_human/launch.py","language":"python","start_line":405,"end_line":456,"context_start_line":385,"context_end_line":476,"code":"            .float()\n            .repeat(B, 1)\n            .reshape(B, 1, 2)\n        )\n        focal = estimate_focal_knowing_depth(pts3ds_self, pp, focal_mode=\"weiszfeld\")\n\n        pts3ds_self = pts3ds_self.reshape(B, -1, 3)\n        pts3ds_other = pts3ds_other.reshape(B, -1, 3)\n        conf_self = conf_self.reshape(B, -1)\n        conf_other = conf_other.reshape(B, -1)\n        # weighted procrustes\n        c2w = weighted_procrustes(\n            pts3ds_self,\n            pts3ds_other,\n            torch.log(conf_self) * torch.log(conf_other),\n            use_weights=True,\n            return_T=True,\n        )\n        return c2w, focal, pp.reshape(B, 2)\n    \n    def _crop_resize(image, intrinsics, crop_res):\n        import dust3r.datasets.utils.cropping as cropping\n        from dust3r.utils.image import ImgNorm\n        \n        # image is a tensor in CHW with values in [-1, 1]; convert to HWC uint8 for PIL\n        img_device = image.device if isinstance(image, torch.Tensor) else None\n        had_batch_dim = False\n        if isinstance(image, torch.Tensor):\n            # accept [3,H,W] or [1,3,H,W]; squeeze batch if present\n            if image.dim() == 4:\n                assert image.shape[0] == 1, \"_crop_resize expects a single image; got a batch\"\n                image = image.squeeze(0)\n                had_batch_dim = True\n            elif image.dim() != 3:\n                raise RuntimeError(f\"Unexpected image tensor shape {tuple(image.shape)}; expected CHW or 1xCHW\")\n\n            image_np = (\n                (image.detach().cpu().permute(1, 2, 0).numpy() * 0.5 +0.5) * 255\n            ).clip(0, 255).astype(np.uint8)\n        else:\n            image_np = image\n\n        target_resolution = np.array(crop_res)\n        image_pil, _, intrinsics = cropping.rescale_image_depthmap(\n            image_np, None, intrinsics, target_resolution\n        )\n\n        # actual cropping (if necessary) with bilinear interpolation\n        intrinsics2 = cropping.camera_matrix_of_crop(\n            intrinsics, image_pil.size, crop_res, offset_factor=0.5\n        )\n        crop_bbox = cropping.bbox_from_intrinsics_in_out(\n            intrinsics, intrinsics2, crop_res\n        )\n        image_pil, _, intrinsics2 = cropping.crop_image_depthmap(\n            image_pil, None, intrinsics, crop_bbox\n        )\n\n        # convert back to normalized CHW tensor on original device\n        image_arr = np.array(image_pil)\n        if image_arr.ndim == 2:\n            image_arr = np.repeat(image_arr[..., None], 3, axis=2)\n        image_tensor = ImgNorm(image_pil)  # CHW, [-1, 1]\n        if had_batch_dim:\n            image_tensor = image_tensor.unsqueeze(0)\n        if img_device is not None:\n            image_tensor = image_tensor.to(img_device)\n\n        # return intrinsics as torch tensor with batch dim like upstream expects\n        intrinsics_tensor = torch.from_numpy(intrinsics2).unsqueeze(0)\n\n        return image_tensor, intrinsics_tensor\n\n    def prepare_input(\n        img_paths,\n        img_mask,\n        msk_paths,\n        size,\n        raymaps=None,\n        raymap_mask=None,\n        revisit=1,\n        update=True,\n        crop=True,\n        load_func=None,\n        annots=None,\n        sampled_indices=None,\n        reset_interval=100,\n        crop_res=None\n    ):\n        images = load_images(img_paths, size=size, verbose=False, crop=crop)\n        images = load_func((img_paths, images, annots, sampled_indices))\n","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.prepare_input","uri":"program://Human3R/function/eval.global_human.launch.prepare_input#L458-L610","kind":"function","name":"prepare_input","path":"eval/global_human/launch.py","language":"python","start_line":458,"end_line":610,"context_start_line":438,"context_end_line":630,"code":"        )\n        image_pil, _, intrinsics2 = cropping.crop_image_depthmap(\n            image_pil, None, intrinsics, crop_bbox\n        )\n\n        # convert back to normalized CHW tensor on original device\n        image_arr = np.array(image_pil)\n        if image_arr.ndim == 2:\n            image_arr = np.repeat(image_arr[..., None], 3, axis=2)\n        image_tensor = ImgNorm(image_pil)  # CHW, [-1, 1]\n        if had_batch_dim:\n            image_tensor = image_tensor.unsqueeze(0)\n        if img_device is not None:\n            image_tensor = image_tensor.to(img_device)\n\n        # return intrinsics as torch tensor with batch dim like upstream expects\n        intrinsics_tensor = torch.from_numpy(intrinsics2).unsqueeze(0)\n\n        return image_tensor, intrinsics_tensor\n\n    def prepare_input(\n        img_paths,\n        img_mask,\n        msk_paths,\n        size,\n        raymaps=None,\n        raymap_mask=None,\n        revisit=1,\n        update=True,\n        crop=True,\n        load_func=None,\n        annots=None,\n        sampled_indices=None,\n        reset_interval=100,\n        crop_res=None\n    ):\n        images = load_images(img_paths, size=size, verbose=False, crop=crop)\n        images = load_func((img_paths, images, annots, sampled_indices))\n\n        has_msk = len(msk_paths) > 0\n        if has_msk:\n            msks = load_masks(msk_paths, size=size, verbose=False, crop=crop)\n\n        views = []\n        if raymaps is None and raymap_mask is None:\n            num_views = len(images)\n\n            for i in range(num_views):\n                view = {\n                    \"img\": images[i][\"img\"],\n                    \"ray_map\": torch.full(\n                        (\n                            images[i][\"img\"].shape[0],\n                            6,\n                            images[i][\"img\"].shape[-2],\n                            images[i][\"img\"].shape[-1],\n                        ),\n                        torch.nan,\n                    ),\n                    \"true_shape\": torch.from_numpy(images[i][\"true_shape\"]),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        images[i][\"camera_pose\"]\n                    ).unsqueeze(0),\n                    \"camera_intrinsics\": torch.from_numpy(\n                        images[i][\"intrinsics\"]\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(True).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(False).unsqueeze(0),\n                    \"update\": torch.tensor(True).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n                \n                if crop_res is not None:\n                    view[\"img\"], view[\"camera_intrinsics\"] = _crop_resize(\n                        images[i][\"img\"], images[i][\"intrinsics\"], crop_res)\n                    # update true_shape to reflect cropped image size (H, W)\n                    view[\"true_shape\"] = torch.tensor([view[\"img\"].shape[-2:]], dtype=torch.int32)\n\n                if has_msk:\n                    view[\"msk\"] = msks[i]\n                for key in images[i].keys():\n                    if key.startswith((\"smpl\", \"T_w2c\")):\n                        view[key] = torch.tensor(images[i][key]).unsqueeze(0)\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n\n        else:\n\n            num_views = len(images) + len(raymaps)\n            assert len(img_mask) == len(raymap_mask) == num_views\n            assert sum(img_mask) == len(images) and sum(raymap_mask) == len(raymaps)\n\n            j = 0\n            k = 0\n            for i in range(num_views):\n                view = {\n                    \"img\": (\n                        images[j][\"img\"]\n                        if img_mask[i]\n                        else torch.full_like(images[0][\"img\"], torch.nan)\n                    ),\n                    \"ray_map\": (\n                        raymaps[k]\n                        if raymap_mask[i]\n                        else torch.full_like(raymaps[0], torch.nan)\n                    ),\n                    \"true_shape\": (\n                        torch.from_numpy(images[j][\"true_shape\"])\n                        if img_mask[i]\n                        else torch.from_numpy(np.int32([raymaps[k].shape[1:-1][::-1]]))\n                    ),\n                    \"idx\": i,\n                    \"instance\": str(i),\n                    \"camera_pose\": torch.from_numpy(\n                        images[i][\"camera_pose\"]\n                    ).unsqueeze(0),\n                    \"camera_intrinsics\": torch.from_numpy(\n                        images[i][\"intrinsics\"]\n                    ).unsqueeze(0),\n                    \"img_mask\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    \"ray_mask\": torch.tensor(raymap_mask[i]).unsqueeze(0),\n                    \"update\": torch.tensor(img_mask[i]).unsqueeze(0),\n                    \"reset\": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),\n                }\n\n                if crop_res is not None:\n                    view[\"img\"], view[\"camera_intrinsics\"] = _crop_resize(\n                        images[i][\"img\"], images[i][\"intrinsics\"], crop_res)\n                    # update true_shape to reflect cropped image size (H, W)\n                    view[\"true_shape\"] = torch.tensor([view[\"img\"].shape[-2:]], dtype=torch.int32)\n\n                if has_msk:\n                    view[\"msk\"] = (\n                        msks[j]\n                        if img_mask[i]\n                        else torch.full_like(images[0][\"img\"], torch.nan)\n                    )\n                for key in images[i].keys():\n                    if key.startswith((\"smpl\", \"T_w2c\")):\n                        view[key] = torch.tensor(images[i][key]).unsqueeze(0)\n                \n                if img_mask[i]:\n                    j += 1\n                if raymap_mask[i]:\n                    k += 1\n                views.append(view)\n                if (i+1) % reset_interval == 0:\n                    overlap_view = deepcopy(view)\n                    overlap_view[\"reset\"] = torch.tensor(False).unsqueeze(0)\n                    views.append(overlap_view)\n            assert j == len(images) and k == len(raymaps)\n\n        if revisit > 1:\n            # repeat input for 'revisit' times\n            new_views = []\n            for r in range(revisit):\n                for i in range(len(views)):\n                    new_view = deepcopy(views[i])\n                    new_view[\"idx\"] = r * len(views) + i\n                    new_view[\"instance\"] = str(r * len(views) + i)\n                    if r > 0:\n                        if not update:\n                            new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                    new_views.append(new_view)\n            return new_views\n        \n        del images\n        return views\n\n    def prepare_gt(gts, revisit=1):\n        target_out = defaultdict(list)\n\n        valid_length = len(gts) // revisit\n        gts = gts[-valid_length:]\n\n        # delet overlaps: reset_mask=True\n        gts = [gt for gt in gts if not gt[\"reset\"]]\n\n        intrinsics = [gt[\"camera_intrinsics\"] for gt in gts]\n        K_mhmr = [gt[\"K_mhmr\"] for gt in gts]\n        camera_pose = [gt[\"camera_pose\"] for gt in gts]\n        imgs = [gt[\"img\"] for gt in gts]\n        target_out['K'] = torch.cat(intrinsics, 0)\n        target_out['K_mhmr'] = torch.cat(K_mhmr, 0)\n        target_out['T_c2w'] = torch.cat(camera_pose, 0)\n        target_out['img'] = torch.cat(imgs, 0)\n        if 'T_w2c' in gts[0]:\n            T_w2c_list = [gt[\"T_w2c\"] for gt in gts]","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.prepare_gt","uri":"program://Human3R/function/eval.global_human.launch.prepare_gt#L612-L643","kind":"function","name":"prepare_gt","path":"eval/global_human/launch.py","language":"python","start_line":612,"end_line":643,"context_start_line":592,"context_end_line":663,"code":"                    views.append(overlap_view)\n            assert j == len(images) and k == len(raymaps)\n\n        if revisit > 1:\n            # repeat input for 'revisit' times\n            new_views = []\n            for r in range(revisit):\n                for i in range(len(views)):\n                    new_view = deepcopy(views[i])\n                    new_view[\"idx\"] = r * len(views) + i\n                    new_view[\"instance\"] = str(r * len(views) + i)\n                    if r > 0:\n                        if not update:\n                            new_view[\"update\"] = torch.tensor(False).unsqueeze(0)\n                    new_views.append(new_view)\n            return new_views\n        \n        del images\n        return views\n\n    def prepare_gt(gts, revisit=1):\n        target_out = defaultdict(list)\n\n        valid_length = len(gts) // revisit\n        gts = gts[-valid_length:]\n\n        # delet overlaps: reset_mask=True\n        gts = [gt for gt in gts if not gt[\"reset\"]]\n\n        intrinsics = [gt[\"camera_intrinsics\"] for gt in gts]\n        K_mhmr = [gt[\"K_mhmr\"] for gt in gts]\n        camera_pose = [gt[\"camera_pose\"] for gt in gts]\n        imgs = [gt[\"img\"] for gt in gts]\n        target_out['K'] = torch.cat(intrinsics, 0)\n        target_out['K_mhmr'] = torch.cat(K_mhmr, 0)\n        target_out['T_c2w'] = torch.cat(camera_pose, 0)\n        target_out['img'] = torch.cat(imgs, 0)\n        if 'T_w2c' in gts[0]:\n            T_w2c_list = [gt[\"T_w2c\"] for gt in gts]\n            target_out['T_w2c'] = torch.cat(T_w2c_list, 0)\n\n        smpl_mask_list = [gt[\"smpl_mask\"] for gt in gts]\n        for gt, smpl_mask in zip(gts, smpl_mask_list):\n            target_out['v3d_c'].append(gt[\"smpl_v3d_c\"][smpl_mask])\n            target_out['j3d_c'].append(gt[\"smpl_j3d_c\"][smpl_mask])\n            target_out['v3d_w'].append(gt[\"smpl_v3d_w\"][smpl_mask])\n            target_out['j3d_w'].append(gt[\"smpl_j3d_w\"][smpl_mask])\n            target_out['v2d'].append(gt[\"smpl_v2d\"][smpl_mask])\n            target_out['j2d'].append(gt[\"smpl_j2d\"][smpl_mask])\n        \n        del gts\n        return target_out\n    \n    def prepare_output(outputs, revisit=1, solve_pose=False, is_save=False):\n        pred_out = defaultdict(list)\n\n        valid_length = len(outputs[\"pred\"]) // revisit\n        outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n        outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n        # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n        reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n        shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n        outputs[\"pred\"] = [\n            pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n        reset_mask = reset_mask[~shifted_reset_mask]\n\n        if solve_pose:\n            pts3ds_self_to_save = [\n                output[\"pts3d_in_self_view\"] for output in outputs[\"pred\"]\n            ]\n            pts3ds_other = [","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:eval.global_human.launch.prepare_output","uri":"program://Human3R/function/eval.global_human.launch.prepare_output#L645-L741","kind":"function","name":"prepare_output","path":"eval/global_human/launch.py","language":"python","start_line":645,"end_line":741,"context_start_line":625,"context_end_line":761,"code":"        target_out['K'] = torch.cat(intrinsics, 0)\n        target_out['K_mhmr'] = torch.cat(K_mhmr, 0)\n        target_out['T_c2w'] = torch.cat(camera_pose, 0)\n        target_out['img'] = torch.cat(imgs, 0)\n        if 'T_w2c' in gts[0]:\n            T_w2c_list = [gt[\"T_w2c\"] for gt in gts]\n            target_out['T_w2c'] = torch.cat(T_w2c_list, 0)\n\n        smpl_mask_list = [gt[\"smpl_mask\"] for gt in gts]\n        for gt, smpl_mask in zip(gts, smpl_mask_list):\n            target_out['v3d_c'].append(gt[\"smpl_v3d_c\"][smpl_mask])\n            target_out['j3d_c'].append(gt[\"smpl_j3d_c\"][smpl_mask])\n            target_out['v3d_w'].append(gt[\"smpl_v3d_w\"][smpl_mask])\n            target_out['j3d_w'].append(gt[\"smpl_j3d_w\"][smpl_mask])\n            target_out['v2d'].append(gt[\"smpl_v2d\"][smpl_mask])\n            target_out['j2d'].append(gt[\"smpl_j2d\"][smpl_mask])\n        \n        del gts\n        return target_out\n    \n    def prepare_output(outputs, revisit=1, solve_pose=False, is_save=False):\n        pred_out = defaultdict(list)\n\n        valid_length = len(outputs[\"pred\"]) // revisit\n        outputs[\"pred\"] = outputs[\"pred\"][-valid_length:]\n        outputs[\"views\"] = outputs[\"views\"][-valid_length:]\n\n        # delet overlaps: reset_mask=True outputs[\"pred\"] and outputs[\"views\"]\n        reset_mask = torch.cat([view[\"reset\"] for view in outputs[\"views\"]], 0)\n        shifted_reset_mask = torch.cat([torch.tensor(False).unsqueeze(0), reset_mask[:-1]], dim=0)\n        outputs[\"pred\"] = [\n            pred for pred, mask in zip(outputs[\"pred\"], shifted_reset_mask) if not mask]\n        reset_mask = reset_mask[~shifted_reset_mask]\n\n        if solve_pose:\n            pts3ds_self_to_save = [\n                output[\"pts3d_in_self_view\"] for output in outputs[\"pred\"]\n            ]\n            pts3ds_other = [\n                output[\"pts3d_in_other_view\"] for output in outputs[\"pred\"]\n            ]\n            conf_self = [output[\"conf_self\"] for output in outputs[\"pred\"]]\n            conf_other = [output[\"conf\"] for output in outputs[\"pred\"]]\n            pr_poses, focal, pp = recover_cam_params(\n                torch.cat(pts3ds_self_to_save, 0),\n                torch.cat(pts3ds_other, 0),\n                torch.cat(conf_self, 0),\n                torch.cat(conf_other, 0),\n            )\n            pts3ds_self = torch.cat(pts3ds_self_to_save, 0)\n        else:\n            pts3ds_self_to_save = [\n                output[\"pts3d_in_self_view\"] for output in outputs[\"pred\"]\n            ]\n            pts3ds_other = [\n                output[\"pts3d_in_other_view\"] for output in outputs[\"pred\"]\n            ]\n            conf_self = [output[\"conf_self\"] for output in outputs[\"pred\"]]\n            conf_other = [output[\"conf\"] for output in outputs[\"pred\"]]\n            pts3ds_self = torch.cat(pts3ds_self_to_save, 0)\n            pr_poses = [\n                pose_encoding_to_camera(pred[\"camera_pose\"].clone())\n                for pred in outputs[\"pred\"]\n            ]\n            pr_poses = torch.cat(pr_poses, 0)\n\n            B, H, W, _ = pts3ds_self.shape\n            pp = (\n                torch.tensor([W // 2, H // 2], device=pts3ds_self.device)\n                .float()\n                .repeat(B, 1)\n                .reshape(B, 2)\n            )\n            focal = estimate_focal_knowing_depth(\n                pts3ds_self, pp, focal_mode=\"weiszfeld\"\n            )\n\n        if is_save:\n            has_mask = \"msk\" in outputs[\"pred\"][0]\n            if has_mask:\n                msks = [output[\"msk\"][...,0] for output in outputs[\"pred\"]]\n                msks = [unpad_image(m, [H, W]) for m in msks]\n            else:\n                msks = [torch.zeros(1, H, W) for _ in range(B)]\n\n            pred_out[\"pts3d_self\"] = pts3ds_self_to_save\n            pred_out[\"conf_self\"] = conf_self\n            pred_out[\"msk\"] = msks\n\n        if reset_mask.any():\n            identity = torch.eye(4, device=pr_poses.device)\n            reset_poses = torch.where(reset_mask.unsqueeze(-1).unsqueeze(-1), pr_poses, identity)\n            cumulative_bases = matrix_cumprod(reset_poses)\n            shifted_bases = torch.cat([identity.unsqueeze(0), cumulative_bases[:-1]], dim=0)\n            pr_poses = torch.einsum('bij,bjk->bik', shifted_bases, pr_poses)\n        pred_out['T_c2w'] = pr_poses\n\n        intrinsics = torch.eye(3, device=pp.device).unsqueeze(0).repeat(B, 1, 1)\n        intrinsics[:, 0, 0] = focal  # fx\n        intrinsics[:, 1, 1] = focal  # fy\n        intrinsics[:, [0, 1], 2] = pp\n        pred_out['K'] = intrinsics\n\n        # get SMPL parameters from outputs\n        pred_out['shape'] = [output.get(\n            \"smpl_shape\", torch.empty(1,0,10))[0] for output in outputs[\"pred\"]]\n        pred_out['rotvec'] = [roma.rotmat_to_rotvec(output.get(\n                \"smpl_rotmat\", torch.empty(1,0,53,3,3))[0]) for output in outputs[\"pred\"]]\n        pred_out['transl'] = [output.get(\n            \"smpl_transl\", torch.empty(1,0,3))[0] for output in outputs[\"pred\"]]\n        pred_out['expression'] = [output.get(\n            \"smpl_expression\", [None])[0] for output in outputs[\"pred\"]]\n        pred_out['loc'] = [output.get(\n            \"smpl_loc\", torch.empty(1,0,2))[0] for output in outputs[\"pred\"]]\n        \n        del outputs\n        return pred_out\n\n\n    model = ARCroco3DStereo.from_pretrained(args.weights)\n    # SMPL model for gt\n    smpl_model = SMPLModel(\n        \"cpu\", \n        model_args={\n            'patch_size': model.croco_args['patch_size'], \n            'mhmr_img_res': model.mhmr_img_res, \n            'bb_patch_size': model.bb_patch_size\n        },\n        eval_args={\n            'dataset': args.eval_dataset,\n            'use_fake_K': args.use_fake_K\n        }\n    )\n    # SMPL layer for pred\n    smpl_layer = SMPL_Layer(type='smplx',\n                            gender='neutral',\n                            num_betas=10,","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam","uri":"program://Human3R/module/datasets_preprocess.preprocess_bedlam#L1-L530","kind":"module","name":"datasets_preprocess.preprocess_bedlam","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":1,"end_line":530,"context_start_line":1,"context_end_line":530,"code":"#!/usr/bin/env python3\n\"\"\"\nModified from CUT3R [https://github.com/CUT3R/CUT3R].\n\nProcess Bedlam scenes by computing camera intrinsics and extrinsics\nfrom extracted data. The script reads per-scene CSV and image/depth files,\ncomputes the necessary camera parameters, and saves the resulting camera\nfiles (as .npz files) in an output directory.\nNote: CUT3R filtered out HDRI scenes and closeup scenes.\nWe also filter out the sequences without SMPLX annotations following Multi-HMR.\n\nUsage:\n    python preprocess_bedlam.py --root /path/to/extracted_data \\\n                             --outdir /path/to/processed_bedlam \\\n                             --annot_dir /path/to/bedlam/processed_labels\n\"\"\"\n\nimport os\nimport sys\nimport cv2\nimport numpy as np\nimport pandas as pd\nfrom glob import glob\nimport shutil\nimport OpenEXR  # Ensure OpenEXR is installed\nfrom concurrent.futures import ProcessPoolExecutor, as_completed\nfrom tqdm import tqdm\nimport argparse\nimport pickle\n\n\n# Enable OpenEXR support in OpenCV.\nos.environ[\"OPENCV_IO_ENABLE_OPENEXR\"] = \"1\"\n\n# Global constants\nIMG_FORMAT = \".png\"\nrotate_flag = False\nSENSOR_W = 36\nSENSOR_H = 20.25\nIMG_W = 1280\nIMG_H = 720\n\ntest_list = [\n    \"20221018_1_250_batch01hand_zoom_suburb_b\",\n    \"20221018_3_250_batch01hand_orbit_archVizUI3_time15\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium\",\n    # \"20221018_3-8_250_batch01hand\", # filtered out by CUT3R, but Multi-HMR involves it\n]\n\n# -----------------------------------------------------------------------------\n# Helper functions for camera parameter conversion\n# -----------------------------------------------------------------------------\n\n\ndef focalLength_mm2px(focalLength, dslr_sens, focalPoint):\n    focal_pixel = (focalLength / dslr_sens) * focalPoint * 2\n    return focal_pixel\n\n\ndef get_cam_int(fl, sens_w, sens_h, cx, cy):\n    flx = focalLength_mm2px(fl, sens_w, cx)\n    fly = focalLength_mm2px(fl, sens_h, cy)\n    cam_mat = np.array([[flx, 0, cx], [0, fly, cy], [0, 0, 1]])\n    return cam_mat\n\n\ndef unreal2cv2(points):\n    # Permute coordinates: x --> y, y --> z, z --> x\n    points = np.roll(points, 2, axis=1)\n    # Invert the y-axis\n    points = points * np.array([1.0, -1.0, 1.0])\n    return points\n\n\ndef get_cam_trans(body_trans, cam_trans):\n    cam_trans = np.array(cam_trans) / 100\n    cam_trans = unreal2cv2(np.reshape(cam_trans, (1, 3)))\n    body_trans = np.array(body_trans) / 100\n    body_trans = unreal2cv2(np.reshape(body_trans, (1, 3)))\n    trans = body_trans - cam_trans\n    return trans\n\n\ndef get_cam_rotmat(pitch, yaw, roll):\n    rotmat_yaw, _ = cv2.Rodrigues(np.array([[0, (yaw / 180) * np.pi, 0]], dtype=float))\n    rotmat_pitch, _ = cv2.Rodrigues(np.array([pitch / 180 * np.pi, 0, 0]).reshape(3, 1))\n    rotmat_roll, _ = cv2.Rodrigues(np.array([0, 0, roll / 180 * np.pi]).reshape(3, 1))\n    final_rotmat = rotmat_roll @ (rotmat_pitch @ rotmat_yaw)\n    return final_rotmat\n\n\ndef get_global_orient(cam_pitch, cam_yaw, cam_roll):\n    pitch_rotmat, _ = cv2.Rodrigues(\n        np.array([cam_pitch / 180 * np.pi, 0, 0]).reshape(3, 1)\n    )\n    roll_rotmat, _ = cv2.Rodrigues(\n        np.array([0, 0, cam_roll / 180 * np.pi]).reshape(3, 1)\n    )\n    final_rotmat = roll_rotmat @ pitch_rotmat\n    return final_rotmat\n\n\ndef convert_translation_to_opencv(x, y, z):\n    t_cv = np.array([y, -z, x])\n    return t_cv\n\n\ndef rotation_matrix_unreal(yaw, pitch, roll):\n    yaw_rad = np.deg2rad(yaw)\n    pitch_rad = np.deg2rad(pitch)\n    roll_rad = np.deg2rad(roll)\n    # Yaw (left-handed)\n    R_yaw = np.array(\n        [\n            [np.cos(-yaw_rad), -np.sin(-yaw_rad), 0],\n            [np.sin(-yaw_rad), np.cos(-yaw_rad), 0],\n            [0, 0, 1],\n        ]\n    )\n    # Pitch (right-handed)\n    R_pitch = np.array(\n        [\n            [np.cos(pitch_rad), 0, np.sin(pitch_rad)],\n            [0, 1, 0],\n            [-np.sin(pitch_rad), 0, np.cos(pitch_rad)],\n        ]\n    )\n    # Roll (right-handed)\n    R_roll = np.array(\n        [\n            [1, 0, 0],\n            [0, np.cos(roll_rad), -np.sin(roll_rad)],\n            [0, np.sin(roll_rad), np.cos(roll_rad)],\n        ]\n    )\n    R_unreal = R_roll @ R_pitch @ R_yaw\n    return R_unreal\n\n\ndef convert_rotation_to_opencv(R_unreal):\n    # Transformation matrix from Unreal to OpenCV coordinate system.\n    C = np.array([[0, 1, 0], [0, 0, -1], [1, 0, 0]])\n    R_cv = C @ R_unreal @ C.T\n    return R_cv\n\n\ndef get_rot_unreal(yaw, pitch, roll):\n    yaw_rad = np.deg2rad(yaw)\n    pitch_rad = np.deg2rad(pitch)\n    roll_rad = np.deg2rad(roll)\n    R_yaw = np.array(\n        [\n            [np.cos(yaw_rad), -np.sin(yaw_rad), 0],\n            [np.sin(yaw_rad), np.cos(yaw_rad), 0],\n            [0, 0, 1],\n        ]\n    )\n    R_pitch = np.array(\n        [\n            [np.cos(pitch_rad), 0, -np.sin(pitch_rad)],\n            [0, 1, 0],\n            [np.sin(pitch_rad), 0, np.cos(pitch_rad)],\n        ]\n    )\n    R_roll = np.array(\n        [\n            [1, 0, 0],\n            [0, np.cos(roll_rad), np.sin(roll_rad)],\n            [0, -np.sin(roll_rad), np.cos(roll_rad)],\n        ]\n    )\n    R_unreal = R_yaw @ R_pitch @ R_roll\n    return R_unreal\n\n\ndef get_extrinsics_unreal(R_unreal, t_unreal):\n    cam_trans = np.array(t_unreal)\n    ext = np.eye(4)\n    ext[:3, :3] = R_unreal\n    ext[:3, 3] = cam_trans.reshape(1, 3)\n    return ext\n\n\ndef get_extrinsics_opencv(yaw, pitch, roll, x, y, z):\n    R_unreal = get_rot_unreal(yaw, pitch, roll)\n    t_unreal = np.array([x / 100.0, y / 100.0, z / 100.0])\n    T_u2wu = get_extrinsics_unreal(R_unreal, t_unreal)\n    T_opencv2unreal = np.array(\n        [[0, 0, -1, 0], [1, 0, 0, 0], [0, -1, 0, 0], [0, 0, 0, 1]], dtype=np.float32\n    )\n    T_wu2ou = np.array(\n        [[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]], dtype=np.float32\n    )\n    return np.linalg.inv(T_opencv2unreal @ T_u2wu @ T_wu2ou)\n\n\n# -----------------------------------------------------------------------------\n# Get camera parameters from the extracted images and CSV data.\n# -----------------------------------------------------------------------------\n\n\ndef get_params(\n    image_folder,\n    fl,\n    trans_body,\n    cam_x,\n    cam_y,\n    cam_z,\n    fps,\n    cam_pitch_,\n    cam_roll_,\n    cam_yaw_,\n):\n    all_images = sorted(glob(os.path.join(image_folder, \"*\" + IMG_FORMAT)))\n    imgnames, cam_ext, cam_int = [], [], []\n\n    for img_ind, image_path in enumerate(all_images):\n        # Process every 5th frame.\n        if img_ind % 5 != 0:\n            continue\n        cam_ind = img_ind\n\n        cam_pitch_ind = cam_pitch_[cam_ind]\n        cam_yaw_ind = cam_yaw_[cam_ind]\n        cam_roll_ind = cam_roll_[cam_ind]\n\n        CAM_INT = get_cam_int(fl[cam_ind], SENSOR_W, SENSOR_H, IMG_W / 2.0, IMG_H / 2.0)\n\n        rot_unreal = rotation_matrix_unreal(cam_yaw_ind, cam_pitch_ind, cam_roll_ind)\n        rot_cv = convert_rotation_to_opencv(rot_unreal)\n        trans_cv = convert_translation_to_opencv(\n            cam_x[cam_ind] / 100.0, cam_y[cam_ind] / 100.0, cam_z[cam_ind] / 100.0\n        )\n        cam_ext_ = np.eye(4)\n        cam_ext_[:3, :3] = rot_cv\n        # The camera pose is computed as the inverse of the transformed translation.\n        cam_ext_[:3, 3] = -rot_cv @ trans_cv\n\n        imgnames.append(\n            os.path.join(image_path.split(\"/\")[-2], image_path.split(\"/\")[-1])\n        )\n        cam_ext.append(cam_ext_)  # camera_pose: c2w\n        cam_int.append(CAM_INT)\n    return imgnames, cam_ext, cam_int\n\n\n# -----------------------------------------------------------------------------\n# Processing per sequence.\n# -----------------------------------------------------------------------------\n\n\ndef process_seq(args):\n    \"\"\"\n    Process a single sequence task. For each image, load the corresponding\n    depth and image files, and save the computed camera intrinsics and the inverse\n    of the extrinsic matrix (i.e. the camera pose in world coordinates) as an NPZ file.\n    \"\"\"\n    (\n        scene,\n        seq_name,\n        outdir,\n        image_folder_base,\n        depth_folder_base,\n        mask_folder_base,\n        imgnames,\n        cam_ext,\n        cam_int,\n        humans,\n        imgname_array,\n    ) = args\n\n    split = 'Test' if scene in test_list else 'Training'\n    \n    out_rgb_dir = os.path.join(outdir, split, '_'.join([scene, seq_name]), 'rgb')\n    out_depth_dir = os.path.join(outdir, split, '_'.join([scene, seq_name]), 'depth')\n    out_cam_dir = os.path.join(outdir, split, \"_\".join([scene, seq_name]), \"cam\")\n    out_smpl_dir = os.path.join(outdir, split, \"_\".join([scene, seq_name]), \"smpl\")\n    out_mask_dir = os.path.join(outdir, split, '_'.join([scene, seq_name]), 'mask')\n    os.makedirs(out_rgb_dir, exist_ok=True)\n    os.makedirs(out_depth_dir, exist_ok=True)\n    os.makedirs(out_cam_dir, exist_ok=True)\n    os.makedirs(out_smpl_dir, exist_ok=True)\n    os.makedirs(out_mask_dir, exist_ok=True)\n\n    assert (\n        len(imgnames) == len(cam_ext) == len(cam_int)\n    ), f\"Inconsistent lengths for {scene}_{seq_name}\"\n    \n    invalid_images = []\n    for imgname, ext, intr, human in zip(imgnames, cam_ext, cam_int, humans):\n        if imgname not in imgname_array:\n            invalid_images.append(f\"Invalid image: {scene}/{imgname}\")\n            continue\n        \n        depthname = imgname.replace(\".png\", \"_depth.exr\")\n        maskname = imgname.replace(\".png\", \"_env.png\")\n        imgpath = os.path.join(image_folder_base, imgname)\n        depthpath = os.path.join(depth_folder_base, depthname)\n        maskpath = os.path.join(mask_folder_base, maskname)\n        depth= OpenEXR.File(depthpath).parts[0].channels['Depth'].pixels\n        depth = depth.astype(np.float32)/100.0\n\n        mask = cv2.imread(maskpath, cv2.IMREAD_GRAYSCALE)\n        if mask is None:\n            # if mask file not exist, create a zero mask\n            mask = np.zeros((IMG_H, IMG_W), dtype=np.uint8)\n        else:\n            # invert the mask\n            mask = 255 - mask      \n        \n        outimg_path = os.path.join(out_rgb_dir, os.path.basename(imgpath))\n        outmask_path = os.path.join(out_mask_dir, os.path.basename(imgpath))\n        outdepth_path = os.path.join(out_depth_dir, os.path.basename(imgpath).replace('.png','.npy'))\n        outcam_path = os.path.join(\n            out_cam_dir, os.path.basename(imgpath).replace(\".png\", \".npz\")\n        )\n\n        out_smpl_path = os.path.join(out_smpl_dir, os.path.basename(imgpath).replace(\".png\", \".pkl\"))\n\n        shutil.copy(imgpath, outimg_path)\n        cv2.imwrite(outmask_path, mask)\n        np.save(outdepth_path, depth)\n        np.savez(outcam_path, intrinsics=intr, pose=np.linalg.inv(ext))  # pose: w2c\n        with open(out_smpl_path, 'wb') as f:\n            pickle.dump(human, f, protocol=pickle.HIGHEST_PROTOCOL)\n    return invalid_images\n\n\n# -----------------------------------------------------------------------------\n# Main entry point.\n# -----------------------------------------------------------------------------\n\n\ndef main():\n    parser = argparse.ArgumentParser(\n        description=\"Process Bedlam scenes: compute camera intrinsics and extrinsics, \"\n        \"and save processed camera files.\"\n    )\n    parser.add_argument(\n        \"--root\",\n        type=str,\n        required=True,\n        help=\"Root directory of the extracted data (scenes).\",\n    )\n    parser.add_argument(\n        \"--outdir\", type=str, required=True, help=\"Output directory for processed data.\"\n    )\n    parser.add_argument(\n        \"--annot_dir\", type=str, required=True, help=\"Annotation directory.\"\n    )\n    parser.add_argument(\n        \"--num_workers\",\n        type=int,\n        default=None,\n        help=\"Number of worker processes (default: os.cpu_count()//2).\",\n    )\n    args = parser.parse_args()\n\n    root = args.root\n    outdir = args.outdir\n    annot_dir = args.annot_dir\n    num_workers = (\n        args.num_workers if args.num_workers is not None else (os.cpu_count() or 4) // 2\n    )\n\n    invalid_list = []\n\n    # Get scene directories from the root folder.\n    scenes = sorted(\n        [d for d in os.listdir(root) if os.path.isdir(os.path.join(root, d))]\n    )\n    # Exclude HDRI scenes.\n    hdri_scenes = [\n        \"20221010_3_1000_batch01hand\",\n        \"20221017_3_1000_batch01hand\",\n        \"20221018_3-8_250_batch01hand\", # filtered out by CUT3R, but Multi-HMR involves it\n        \"20221019_3_250_highbmihand\",\n    ]\n    scenes = np.setdiff1d(scenes, hdri_scenes)\n\n    max_human = 0\n\n    tasks = []\n    for scene in tqdm(scenes, desc=\"Collecting tasks\"):\n        # Skip closeup scenes.\n        if \"closeup\" in scene:\n            continue\n        base_folder = os.path.join(root, scene)\n        image_folder_base = os.path.join(root, scene, \"png\")\n        depth_folder_base = os.path.join(root, scene, \"depth\")\n        mask_folder_base = os.path.join(root, scene, \"masks\")\n        csv_path = os.path.join(base_folder, \"be_seq.csv\")\n        annot_path = os.path.join(annot_dir, scene + \"_6fps.npz\")\n        if not os.path.exists(annot_path):\n            annot_path = os.path.join(annot_dir, scene + \"_30fps.npz\")\n        if not os.path.exists(csv_path):\n            continue\n        csv_data = pd.read_csv(csv_path)\n        csv_data = csv_data.to_dict(\"list\")\n        cam_csv_base = os.path.join(base_folder, \"ground_truth\", \"camera\")\n        annot_x = np.load(annot_path)\n\n        # Retrieving SMPL parameters once\n        pose_cam_array = annot_x['pose_cam']\n        H_array = annot_x['cam_ext']\n        shape_array = annot_x['shape']\n        imgname_array = annot_x['imgname']\n        trans_cam_array = annot_x['trans_cam']\n        \n        seq_count = 0\n        max_seq_per_scene = 1500000000000\n\n        # Look for a row in the CSV with a \"sequence_name\" comment.\n        for idx, comment in enumerate(csv_data.get(\"Comment\", [])):\n            if \"sequence_name\" in comment:\n                if seq_count >= max_seq_per_scene:\n                    break\n                    \n                seq_name = comment.split(\";\")[0].split(\"=\")[-1]\n                if not np.any(np.char.startswith(imgname_array, seq_name + '/')):\n                    invalid_list.append(f\"Invalid sequence: {scene}/{seq_name}\")\n                    continue\n                  \n                cam_csv_path = os.path.join(cam_csv_base, seq_name + \"_camera.csv\")\n                if not os.path.exists(cam_csv_path):\n                    continue\n                cam_csv_data = pd.read_csv(cam_csv_path)\n                cam_csv_data = cam_csv_data.to_dict(\"list\")\n                cam_x = cam_csv_data[\"x\"]\n                cam_y = cam_csv_data[\"y\"]\n                cam_z = cam_csv_data[\"z\"]\n                cam_yaw_ = cam_csv_data[\"yaw\"]\n                cam_pitch_ = cam_csv_data[\"pitch\"]\n                cam_roll_ = cam_csv_data[\"roll\"]\n                fl = cam_csv_data[\"focal_length\"]\n                image_folder = os.path.join(image_folder_base, seq_name)\n                trans_body = None  # Not used here.\n                imgnames, cam_ext, cam_int = get_params(\n                    image_folder,\n                    fl,\n                    trans_body,\n                    cam_x,\n                    cam_y,\n                    cam_z,\n                    6,\n                    cam_pitch_=cam_pitch_,\n                    cam_roll_=cam_roll_,\n                    cam_yaw_=cam_yaw_,\n                )\n                humans = [] # humans for each sequence\n                for imgname in imgnames:\n                    idxs = np.where(imgname == imgname_array)[0]\n\n                    if len(idxs) > max_human:\n                        max_human = len(idxs)\n\n                    persons_per_img = [] # persons for each image\n                    for i in idxs:\n                        sys.stdout.flush()\n\n                        # SMPLX params\n                        pose = pose_cam_array[i]\n                        root_pose = pose[:3]\n                        body_pose=pose[3:66]\n                        jaw_pose=pose[66:69]\n                        leye_pose=pose[69:72]\n                        reye_pose=pose[72:75]\n                        left_hand_pose=pose[75:120]\n                        right_hand_pose=pose[120:165]\n                        betas=shape_array[i]\n                        transl = trans_cam_array[i] + H_array[i][:, 3][:3]\n\n                        person = {\n                            # SMPL GT in camera coordinates system\n                            'smplx_root_pose': root_pose.reshape(1,3), # axis-angle\n                            'smplx_body_pose': body_pose.reshape(21,3), # axis-angle\n                            'smplx_jaw_pose': jaw_pose.reshape(1,3), # axis-angle\n                            'smplx_leye_pose': leye_pose.reshape(1,3), # axis-angle\n                            'smplx_reye_pose': reye_pose.reshape(1,3), # axis-angle\n                            'smplx_left_hand_pose': left_hand_pose.reshape(15,3), # axis-angle\n                            'smplx_right_hand_pose': right_hand_pose.reshape(15,3), # axis-angle\n                            'smplx_shape': betas.reshape(11),\n                            'smplx_gender': 'neutral',\n                            'smplx_transl': transl.reshape(3),\n                        }\n                        persons_per_img.append(person)\n                    humans.append(persons_per_img)\n\n                tasks.append(\n                    (\n                        scene,\n                        seq_name,\n                        outdir,\n                        image_folder_base,\n                        depth_folder_base,\n                        mask_folder_base,\n                        imgnames,\n                        cam_ext,\n                        cam_int,\n                        humans,\n                        imgname_array,\n                    )\n                )\n\n                seq_count += 1\n\n    print(f\"max_human: {max_human}\")\n    \n    # Process each task in parallel.\n    with ProcessPoolExecutor(max_workers=num_workers) as executor:\n        futures = {executor.submit(process_seq, task): task for task in tasks}\n        for future in tqdm(\n            as_completed(futures), total=len(futures), desc=\"Processing sequences\"\n        ):\n            # error = future.result()\n            # if error:\n            #     print(error)\n            invalid_images = future.result()\n            if invalid_images:\n                invalid_list.extend(invalid_images)\n\n    # print invalid items\n    if invalid_list:\n        print(\"\\nInvalid sequences and images:\")\n        for item in invalid_list:\n            print(item)\n\n\nif __name__ == \"__main__\":\n    main()","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.focalLength_mm2px","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.focalLength_mm2px#L55-L57","kind":"function","name":"focalLength_mm2px","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":55,"end_line":57,"context_start_line":35,"context_end_line":77,"code":"# Global constants\nIMG_FORMAT = \".png\"\nrotate_flag = False\nSENSOR_W = 36\nSENSOR_H = 20.25\nIMG_W = 1280\nIMG_H = 720\n\ntest_list = [\n    \"20221018_1_250_batch01hand_zoom_suburb_b\",\n    \"20221018_3_250_batch01hand_orbit_archVizUI3_time15\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium\",\n    # \"20221018_3-8_250_batch01hand\", # filtered out by CUT3R, but Multi-HMR involves it\n]\n\n# -----------------------------------------------------------------------------\n# Helper functions for camera parameter conversion\n# -----------------------------------------------------------------------------\n\n\ndef focalLength_mm2px(focalLength, dslr_sens, focalPoint):\n    focal_pixel = (focalLength / dslr_sens) * focalPoint * 2\n    return focal_pixel\n\n\ndef get_cam_int(fl, sens_w, sens_h, cx, cy):\n    flx = focalLength_mm2px(fl, sens_w, cx)\n    fly = focalLength_mm2px(fl, sens_h, cy)\n    cam_mat = np.array([[flx, 0, cx], [0, fly, cy], [0, 0, 1]])\n    return cam_mat\n\n\ndef unreal2cv2(points):\n    # Permute coordinates: x --> y, y --> z, z --> x\n    points = np.roll(points, 2, axis=1)\n    # Invert the y-axis\n    points = points * np.array([1.0, -1.0, 1.0])\n    return points\n\n\ndef get_cam_trans(body_trans, cam_trans):\n    cam_trans = np.array(cam_trans) / 100\n    cam_trans = unreal2cv2(np.reshape(cam_trans, (1, 3)))","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.get_cam_int","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.get_cam_int#L60-L64","kind":"function","name":"get_cam_int","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":60,"end_line":64,"context_start_line":40,"context_end_line":84,"code":"IMG_W = 1280\nIMG_H = 720\n\ntest_list = [\n    \"20221018_1_250_batch01hand_zoom_suburb_b\",\n    \"20221018_3_250_batch01hand_orbit_archVizUI3_time15\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium\",\n    # \"20221018_3-8_250_batch01hand\", # filtered out by CUT3R, but Multi-HMR involves it\n]\n\n# -----------------------------------------------------------------------------\n# Helper functions for camera parameter conversion\n# -----------------------------------------------------------------------------\n\n\ndef focalLength_mm2px(focalLength, dslr_sens, focalPoint):\n    focal_pixel = (focalLength / dslr_sens) * focalPoint * 2\n    return focal_pixel\n\n\ndef get_cam_int(fl, sens_w, sens_h, cx, cy):\n    flx = focalLength_mm2px(fl, sens_w, cx)\n    fly = focalLength_mm2px(fl, sens_h, cy)\n    cam_mat = np.array([[flx, 0, cx], [0, fly, cy], [0, 0, 1]])\n    return cam_mat\n\n\ndef unreal2cv2(points):\n    # Permute coordinates: x --> y, y --> z, z --> x\n    points = np.roll(points, 2, axis=1)\n    # Invert the y-axis\n    points = points * np.array([1.0, -1.0, 1.0])\n    return points\n\n\ndef get_cam_trans(body_trans, cam_trans):\n    cam_trans = np.array(cam_trans) / 100\n    cam_trans = unreal2cv2(np.reshape(cam_trans, (1, 3)))\n    body_trans = np.array(body_trans) / 100\n    body_trans = unreal2cv2(np.reshape(body_trans, (1, 3)))\n    trans = body_trans - cam_trans\n    return trans\n\n\ndef get_cam_rotmat(pitch, yaw, roll):","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.unreal2cv2","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.unreal2cv2#L67-L72","kind":"function","name":"unreal2cv2","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":67,"end_line":72,"context_start_line":47,"context_end_line":92,"code":"    # \"20221018_3-8_250_batch01hand\", # filtered out by CUT3R, but Multi-HMR involves it\n]\n\n# -----------------------------------------------------------------------------\n# Helper functions for camera parameter conversion\n# -----------------------------------------------------------------------------\n\n\ndef focalLength_mm2px(focalLength, dslr_sens, focalPoint):\n    focal_pixel = (focalLength / dslr_sens) * focalPoint * 2\n    return focal_pixel\n\n\ndef get_cam_int(fl, sens_w, sens_h, cx, cy):\n    flx = focalLength_mm2px(fl, sens_w, cx)\n    fly = focalLength_mm2px(fl, sens_h, cy)\n    cam_mat = np.array([[flx, 0, cx], [0, fly, cy], [0, 0, 1]])\n    return cam_mat\n\n\ndef unreal2cv2(points):\n    # Permute coordinates: x --> y, y --> z, z --> x\n    points = np.roll(points, 2, axis=1)\n    # Invert the y-axis\n    points = points * np.array([1.0, -1.0, 1.0])\n    return points\n\n\ndef get_cam_trans(body_trans, cam_trans):\n    cam_trans = np.array(cam_trans) / 100\n    cam_trans = unreal2cv2(np.reshape(cam_trans, (1, 3)))\n    body_trans = np.array(body_trans) / 100\n    body_trans = unreal2cv2(np.reshape(body_trans, (1, 3)))\n    trans = body_trans - cam_trans\n    return trans\n\n\ndef get_cam_rotmat(pitch, yaw, roll):\n    rotmat_yaw, _ = cv2.Rodrigues(np.array([[0, (yaw / 180) * np.pi, 0]], dtype=float))\n    rotmat_pitch, _ = cv2.Rodrigues(np.array([pitch / 180 * np.pi, 0, 0]).reshape(3, 1))\n    rotmat_roll, _ = cv2.Rodrigues(np.array([0, 0, roll / 180 * np.pi]).reshape(3, 1))\n    final_rotmat = rotmat_roll @ (rotmat_pitch @ rotmat_yaw)\n    return final_rotmat\n\n\ndef get_global_orient(cam_pitch, cam_yaw, cam_roll):","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.get_cam_trans","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.get_cam_trans#L75-L81","kind":"function","name":"get_cam_trans","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":75,"end_line":81,"context_start_line":55,"context_end_line":101,"code":"def focalLength_mm2px(focalLength, dslr_sens, focalPoint):\n    focal_pixel = (focalLength / dslr_sens) * focalPoint * 2\n    return focal_pixel\n\n\ndef get_cam_int(fl, sens_w, sens_h, cx, cy):\n    flx = focalLength_mm2px(fl, sens_w, cx)\n    fly = focalLength_mm2px(fl, sens_h, cy)\n    cam_mat = np.array([[flx, 0, cx], [0, fly, cy], [0, 0, 1]])\n    return cam_mat\n\n\ndef unreal2cv2(points):\n    # Permute coordinates: x --> y, y --> z, z --> x\n    points = np.roll(points, 2, axis=1)\n    # Invert the y-axis\n    points = points * np.array([1.0, -1.0, 1.0])\n    return points\n\n\ndef get_cam_trans(body_trans, cam_trans):\n    cam_trans = np.array(cam_trans) / 100\n    cam_trans = unreal2cv2(np.reshape(cam_trans, (1, 3)))\n    body_trans = np.array(body_trans) / 100\n    body_trans = unreal2cv2(np.reshape(body_trans, (1, 3)))\n    trans = body_trans - cam_trans\n    return trans\n\n\ndef get_cam_rotmat(pitch, yaw, roll):\n    rotmat_yaw, _ = cv2.Rodrigues(np.array([[0, (yaw / 180) * np.pi, 0]], dtype=float))\n    rotmat_pitch, _ = cv2.Rodrigues(np.array([pitch / 180 * np.pi, 0, 0]).reshape(3, 1))\n    rotmat_roll, _ = cv2.Rodrigues(np.array([0, 0, roll / 180 * np.pi]).reshape(3, 1))\n    final_rotmat = rotmat_roll @ (rotmat_pitch @ rotmat_yaw)\n    return final_rotmat\n\n\ndef get_global_orient(cam_pitch, cam_yaw, cam_roll):\n    pitch_rotmat, _ = cv2.Rodrigues(\n        np.array([cam_pitch / 180 * np.pi, 0, 0]).reshape(3, 1)\n    )\n    roll_rotmat, _ = cv2.Rodrigues(\n        np.array([0, 0, cam_roll / 180 * np.pi]).reshape(3, 1)\n    )\n    final_rotmat = roll_rotmat @ pitch_rotmat\n    return final_rotmat\n","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.get_cam_rotmat","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.get_cam_rotmat#L84-L89","kind":"function","name":"get_cam_rotmat","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":84,"end_line":89,"context_start_line":64,"context_end_line":109,"code":"    return cam_mat\n\n\ndef unreal2cv2(points):\n    # Permute coordinates: x --> y, y --> z, z --> x\n    points = np.roll(points, 2, axis=1)\n    # Invert the y-axis\n    points = points * np.array([1.0, -1.0, 1.0])\n    return points\n\n\ndef get_cam_trans(body_trans, cam_trans):\n    cam_trans = np.array(cam_trans) / 100\n    cam_trans = unreal2cv2(np.reshape(cam_trans, (1, 3)))\n    body_trans = np.array(body_trans) / 100\n    body_trans = unreal2cv2(np.reshape(body_trans, (1, 3)))\n    trans = body_trans - cam_trans\n    return trans\n\n\ndef get_cam_rotmat(pitch, yaw, roll):\n    rotmat_yaw, _ = cv2.Rodrigues(np.array([[0, (yaw / 180) * np.pi, 0]], dtype=float))\n    rotmat_pitch, _ = cv2.Rodrigues(np.array([pitch / 180 * np.pi, 0, 0]).reshape(3, 1))\n    rotmat_roll, _ = cv2.Rodrigues(np.array([0, 0, roll / 180 * np.pi]).reshape(3, 1))\n    final_rotmat = rotmat_roll @ (rotmat_pitch @ rotmat_yaw)\n    return final_rotmat\n\n\ndef get_global_orient(cam_pitch, cam_yaw, cam_roll):\n    pitch_rotmat, _ = cv2.Rodrigues(\n        np.array([cam_pitch / 180 * np.pi, 0, 0]).reshape(3, 1)\n    )\n    roll_rotmat, _ = cv2.Rodrigues(\n        np.array([0, 0, cam_roll / 180 * np.pi]).reshape(3, 1)\n    )\n    final_rotmat = roll_rotmat @ pitch_rotmat\n    return final_rotmat\n\n\ndef convert_translation_to_opencv(x, y, z):\n    t_cv = np.array([y, -z, x])\n    return t_cv\n\n\ndef rotation_matrix_unreal(yaw, pitch, roll):\n    yaw_rad = np.deg2rad(yaw)","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.get_global_orient","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.get_global_orient#L92-L100","kind":"function","name":"get_global_orient","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":92,"end_line":100,"context_start_line":72,"context_end_line":120,"code":"    return points\n\n\ndef get_cam_trans(body_trans, cam_trans):\n    cam_trans = np.array(cam_trans) / 100\n    cam_trans = unreal2cv2(np.reshape(cam_trans, (1, 3)))\n    body_trans = np.array(body_trans) / 100\n    body_trans = unreal2cv2(np.reshape(body_trans, (1, 3)))\n    trans = body_trans - cam_trans\n    return trans\n\n\ndef get_cam_rotmat(pitch, yaw, roll):\n    rotmat_yaw, _ = cv2.Rodrigues(np.array([[0, (yaw / 180) * np.pi, 0]], dtype=float))\n    rotmat_pitch, _ = cv2.Rodrigues(np.array([pitch / 180 * np.pi, 0, 0]).reshape(3, 1))\n    rotmat_roll, _ = cv2.Rodrigues(np.array([0, 0, roll / 180 * np.pi]).reshape(3, 1))\n    final_rotmat = rotmat_roll @ (rotmat_pitch @ rotmat_yaw)\n    return final_rotmat\n\n\ndef get_global_orient(cam_pitch, cam_yaw, cam_roll):\n    pitch_rotmat, _ = cv2.Rodrigues(\n        np.array([cam_pitch / 180 * np.pi, 0, 0]).reshape(3, 1)\n    )\n    roll_rotmat, _ = cv2.Rodrigues(\n        np.array([0, 0, cam_roll / 180 * np.pi]).reshape(3, 1)\n    )\n    final_rotmat = roll_rotmat @ pitch_rotmat\n    return final_rotmat\n\n\ndef convert_translation_to_opencv(x, y, z):\n    t_cv = np.array([y, -z, x])\n    return t_cv\n\n\ndef rotation_matrix_unreal(yaw, pitch, roll):\n    yaw_rad = np.deg2rad(yaw)\n    pitch_rad = np.deg2rad(pitch)\n    roll_rad = np.deg2rad(roll)\n    # Yaw (left-handed)\n    R_yaw = np.array(\n        [\n            [np.cos(-yaw_rad), -np.sin(-yaw_rad), 0],\n            [np.sin(-yaw_rad), np.cos(-yaw_rad), 0],\n            [0, 0, 1],\n        ]\n    )\n    # Pitch (right-handed)","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.convert_translation_to_opencv","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.convert_translation_to_opencv#L103-L105","kind":"function","name":"convert_translation_to_opencv","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":103,"end_line":105,"context_start_line":83,"context_end_line":125,"code":"\ndef get_cam_rotmat(pitch, yaw, roll):\n    rotmat_yaw, _ = cv2.Rodrigues(np.array([[0, (yaw / 180) * np.pi, 0]], dtype=float))\n    rotmat_pitch, _ = cv2.Rodrigues(np.array([pitch / 180 * np.pi, 0, 0]).reshape(3, 1))\n    rotmat_roll, _ = cv2.Rodrigues(np.array([0, 0, roll / 180 * np.pi]).reshape(3, 1))\n    final_rotmat = rotmat_roll @ (rotmat_pitch @ rotmat_yaw)\n    return final_rotmat\n\n\ndef get_global_orient(cam_pitch, cam_yaw, cam_roll):\n    pitch_rotmat, _ = cv2.Rodrigues(\n        np.array([cam_pitch / 180 * np.pi, 0, 0]).reshape(3, 1)\n    )\n    roll_rotmat, _ = cv2.Rodrigues(\n        np.array([0, 0, cam_roll / 180 * np.pi]).reshape(3, 1)\n    )\n    final_rotmat = roll_rotmat @ pitch_rotmat\n    return final_rotmat\n\n\ndef convert_translation_to_opencv(x, y, z):\n    t_cv = np.array([y, -z, x])\n    return t_cv\n\n\ndef rotation_matrix_unreal(yaw, pitch, roll):\n    yaw_rad = np.deg2rad(yaw)\n    pitch_rad = np.deg2rad(pitch)\n    roll_rad = np.deg2rad(roll)\n    # Yaw (left-handed)\n    R_yaw = np.array(\n        [\n            [np.cos(-yaw_rad), -np.sin(-yaw_rad), 0],\n            [np.sin(-yaw_rad), np.cos(-yaw_rad), 0],\n            [0, 0, 1],\n        ]\n    )\n    # Pitch (right-handed)\n    R_pitch = np.array(\n        [\n            [np.cos(pitch_rad), 0, np.sin(pitch_rad)],\n            [0, 1, 0],\n            [-np.sin(pitch_rad), 0, np.cos(pitch_rad)],","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.rotation_matrix_unreal","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.rotation_matrix_unreal#L108-L137","kind":"function","name":"rotation_matrix_unreal","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":108,"end_line":137,"context_start_line":88,"context_end_line":157,"code":"    final_rotmat = rotmat_roll @ (rotmat_pitch @ rotmat_yaw)\n    return final_rotmat\n\n\ndef get_global_orient(cam_pitch, cam_yaw, cam_roll):\n    pitch_rotmat, _ = cv2.Rodrigues(\n        np.array([cam_pitch / 180 * np.pi, 0, 0]).reshape(3, 1)\n    )\n    roll_rotmat, _ = cv2.Rodrigues(\n        np.array([0, 0, cam_roll / 180 * np.pi]).reshape(3, 1)\n    )\n    final_rotmat = roll_rotmat @ pitch_rotmat\n    return final_rotmat\n\n\ndef convert_translation_to_opencv(x, y, z):\n    t_cv = np.array([y, -z, x])\n    return t_cv\n\n\ndef rotation_matrix_unreal(yaw, pitch, roll):\n    yaw_rad = np.deg2rad(yaw)\n    pitch_rad = np.deg2rad(pitch)\n    roll_rad = np.deg2rad(roll)\n    # Yaw (left-handed)\n    R_yaw = np.array(\n        [\n            [np.cos(-yaw_rad), -np.sin(-yaw_rad), 0],\n            [np.sin(-yaw_rad), np.cos(-yaw_rad), 0],\n            [0, 0, 1],\n        ]\n    )\n    # Pitch (right-handed)\n    R_pitch = np.array(\n        [\n            [np.cos(pitch_rad), 0, np.sin(pitch_rad)],\n            [0, 1, 0],\n            [-np.sin(pitch_rad), 0, np.cos(pitch_rad)],\n        ]\n    )\n    # Roll (right-handed)\n    R_roll = np.array(\n        [\n            [1, 0, 0],\n            [0, np.cos(roll_rad), -np.sin(roll_rad)],\n            [0, np.sin(roll_rad), np.cos(roll_rad)],\n        ]\n    )\n    R_unreal = R_roll @ R_pitch @ R_yaw\n    return R_unreal\n\n\ndef convert_rotation_to_opencv(R_unreal):\n    # Transformation matrix from Unreal to OpenCV coordinate system.\n    C = np.array([[0, 1, 0], [0, 0, -1], [1, 0, 0]])\n    R_cv = C @ R_unreal @ C.T\n    return R_cv\n\n\ndef get_rot_unreal(yaw, pitch, roll):\n    yaw_rad = np.deg2rad(yaw)\n    pitch_rad = np.deg2rad(pitch)\n    roll_rad = np.deg2rad(roll)\n    R_yaw = np.array(\n        [\n            [np.cos(yaw_rad), -np.sin(yaw_rad), 0],\n            [np.sin(yaw_rad), np.cos(yaw_rad), 0],\n            [0, 0, 1],\n        ]\n    )","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.convert_rotation_to_opencv","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.convert_rotation_to_opencv#L140-L144","kind":"function","name":"convert_rotation_to_opencv","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":140,"end_line":144,"context_start_line":120,"context_end_line":164,"code":"    # Pitch (right-handed)\n    R_pitch = np.array(\n        [\n            [np.cos(pitch_rad), 0, np.sin(pitch_rad)],\n            [0, 1, 0],\n            [-np.sin(pitch_rad), 0, np.cos(pitch_rad)],\n        ]\n    )\n    # Roll (right-handed)\n    R_roll = np.array(\n        [\n            [1, 0, 0],\n            [0, np.cos(roll_rad), -np.sin(roll_rad)],\n            [0, np.sin(roll_rad), np.cos(roll_rad)],\n        ]\n    )\n    R_unreal = R_roll @ R_pitch @ R_yaw\n    return R_unreal\n\n\ndef convert_rotation_to_opencv(R_unreal):\n    # Transformation matrix from Unreal to OpenCV coordinate system.\n    C = np.array([[0, 1, 0], [0, 0, -1], [1, 0, 0]])\n    R_cv = C @ R_unreal @ C.T\n    return R_cv\n\n\ndef get_rot_unreal(yaw, pitch, roll):\n    yaw_rad = np.deg2rad(yaw)\n    pitch_rad = np.deg2rad(pitch)\n    roll_rad = np.deg2rad(roll)\n    R_yaw = np.array(\n        [\n            [np.cos(yaw_rad), -np.sin(yaw_rad), 0],\n            [np.sin(yaw_rad), np.cos(yaw_rad), 0],\n            [0, 0, 1],\n        ]\n    )\n    R_pitch = np.array(\n        [\n            [np.cos(pitch_rad), 0, -np.sin(pitch_rad)],\n            [0, 1, 0],\n            [np.sin(pitch_rad), 0, np.cos(pitch_rad)],\n        ]\n    )","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.get_rot_unreal","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.get_rot_unreal#L147-L173","kind":"function","name":"get_rot_unreal","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":147,"end_line":173,"context_start_line":127,"context_end_line":193,"code":"    )\n    # Roll (right-handed)\n    R_roll = np.array(\n        [\n            [1, 0, 0],\n            [0, np.cos(roll_rad), -np.sin(roll_rad)],\n            [0, np.sin(roll_rad), np.cos(roll_rad)],\n        ]\n    )\n    R_unreal = R_roll @ R_pitch @ R_yaw\n    return R_unreal\n\n\ndef convert_rotation_to_opencv(R_unreal):\n    # Transformation matrix from Unreal to OpenCV coordinate system.\n    C = np.array([[0, 1, 0], [0, 0, -1], [1, 0, 0]])\n    R_cv = C @ R_unreal @ C.T\n    return R_cv\n\n\ndef get_rot_unreal(yaw, pitch, roll):\n    yaw_rad = np.deg2rad(yaw)\n    pitch_rad = np.deg2rad(pitch)\n    roll_rad = np.deg2rad(roll)\n    R_yaw = np.array(\n        [\n            [np.cos(yaw_rad), -np.sin(yaw_rad), 0],\n            [np.sin(yaw_rad), np.cos(yaw_rad), 0],\n            [0, 0, 1],\n        ]\n    )\n    R_pitch = np.array(\n        [\n            [np.cos(pitch_rad), 0, -np.sin(pitch_rad)],\n            [0, 1, 0],\n            [np.sin(pitch_rad), 0, np.cos(pitch_rad)],\n        ]\n    )\n    R_roll = np.array(\n        [\n            [1, 0, 0],\n            [0, np.cos(roll_rad), np.sin(roll_rad)],\n            [0, -np.sin(roll_rad), np.cos(roll_rad)],\n        ]\n    )\n    R_unreal = R_yaw @ R_pitch @ R_roll\n    return R_unreal\n\n\ndef get_extrinsics_unreal(R_unreal, t_unreal):\n    cam_trans = np.array(t_unreal)\n    ext = np.eye(4)\n    ext[:3, :3] = R_unreal\n    ext[:3, 3] = cam_trans.reshape(1, 3)\n    return ext\n\n\ndef get_extrinsics_opencv(yaw, pitch, roll, x, y, z):\n    R_unreal = get_rot_unreal(yaw, pitch, roll)\n    t_unreal = np.array([x / 100.0, y / 100.0, z / 100.0])\n    T_u2wu = get_extrinsics_unreal(R_unreal, t_unreal)\n    T_opencv2unreal = np.array(\n        [[0, 0, -1, 0], [1, 0, 0, 0], [0, -1, 0, 0], [0, 0, 0, 1]], dtype=np.float32\n    )\n    T_wu2ou = np.array(\n        [[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]], dtype=np.float32\n    )","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.get_extrinsics_unreal","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.get_extrinsics_unreal#L176-L181","kind":"function","name":"get_extrinsics_unreal","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":176,"end_line":181,"context_start_line":156,"context_end_line":201,"code":"        ]\n    )\n    R_pitch = np.array(\n        [\n            [np.cos(pitch_rad), 0, -np.sin(pitch_rad)],\n            [0, 1, 0],\n            [np.sin(pitch_rad), 0, np.cos(pitch_rad)],\n        ]\n    )\n    R_roll = np.array(\n        [\n            [1, 0, 0],\n            [0, np.cos(roll_rad), np.sin(roll_rad)],\n            [0, -np.sin(roll_rad), np.cos(roll_rad)],\n        ]\n    )\n    R_unreal = R_yaw @ R_pitch @ R_roll\n    return R_unreal\n\n\ndef get_extrinsics_unreal(R_unreal, t_unreal):\n    cam_trans = np.array(t_unreal)\n    ext = np.eye(4)\n    ext[:3, :3] = R_unreal\n    ext[:3, 3] = cam_trans.reshape(1, 3)\n    return ext\n\n\ndef get_extrinsics_opencv(yaw, pitch, roll, x, y, z):\n    R_unreal = get_rot_unreal(yaw, pitch, roll)\n    t_unreal = np.array([x / 100.0, y / 100.0, z / 100.0])\n    T_u2wu = get_extrinsics_unreal(R_unreal, t_unreal)\n    T_opencv2unreal = np.array(\n        [[0, 0, -1, 0], [1, 0, 0, 0], [0, -1, 0, 0], [0, 0, 0, 1]], dtype=np.float32\n    )\n    T_wu2ou = np.array(\n        [[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]], dtype=np.float32\n    )\n    return np.linalg.inv(T_opencv2unreal @ T_u2wu @ T_wu2ou)\n\n\n# -----------------------------------------------------------------------------\n# Get camera parameters from the extracted images and CSV data.\n# -----------------------------------------------------------------------------\n\n","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.get_extrinsics_opencv","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.get_extrinsics_opencv#L184-L194","kind":"function","name":"get_extrinsics_opencv","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":184,"end_line":194,"context_start_line":164,"context_end_line":214,"code":"    )\n    R_roll = np.array(\n        [\n            [1, 0, 0],\n            [0, np.cos(roll_rad), np.sin(roll_rad)],\n            [0, -np.sin(roll_rad), np.cos(roll_rad)],\n        ]\n    )\n    R_unreal = R_yaw @ R_pitch @ R_roll\n    return R_unreal\n\n\ndef get_extrinsics_unreal(R_unreal, t_unreal):\n    cam_trans = np.array(t_unreal)\n    ext = np.eye(4)\n    ext[:3, :3] = R_unreal\n    ext[:3, 3] = cam_trans.reshape(1, 3)\n    return ext\n\n\ndef get_extrinsics_opencv(yaw, pitch, roll, x, y, z):\n    R_unreal = get_rot_unreal(yaw, pitch, roll)\n    t_unreal = np.array([x / 100.0, y / 100.0, z / 100.0])\n    T_u2wu = get_extrinsics_unreal(R_unreal, t_unreal)\n    T_opencv2unreal = np.array(\n        [[0, 0, -1, 0], [1, 0, 0, 0], [0, -1, 0, 0], [0, 0, 0, 1]], dtype=np.float32\n    )\n    T_wu2ou = np.array(\n        [[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]], dtype=np.float32\n    )\n    return np.linalg.inv(T_opencv2unreal @ T_u2wu @ T_wu2ou)\n\n\n# -----------------------------------------------------------------------------\n# Get camera parameters from the extracted images and CSV data.\n# -----------------------------------------------------------------------------\n\n\ndef get_params(\n    image_folder,\n    fl,\n    trans_body,\n    cam_x,\n    cam_y,\n    cam_z,\n    fps,\n    cam_pitch_,\n    cam_roll_,\n    cam_yaw_,\n):\n    all_images = sorted(glob(os.path.join(image_folder, \"*\" + IMG_FORMAT)))","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.get_params","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.get_params#L202-L244","kind":"function","name":"get_params","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":202,"end_line":244,"context_start_line":182,"context_end_line":264,"code":"\n\ndef get_extrinsics_opencv(yaw, pitch, roll, x, y, z):\n    R_unreal = get_rot_unreal(yaw, pitch, roll)\n    t_unreal = np.array([x / 100.0, y / 100.0, z / 100.0])\n    T_u2wu = get_extrinsics_unreal(R_unreal, t_unreal)\n    T_opencv2unreal = np.array(\n        [[0, 0, -1, 0], [1, 0, 0, 0], [0, -1, 0, 0], [0, 0, 0, 1]], dtype=np.float32\n    )\n    T_wu2ou = np.array(\n        [[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]], dtype=np.float32\n    )\n    return np.linalg.inv(T_opencv2unreal @ T_u2wu @ T_wu2ou)\n\n\n# -----------------------------------------------------------------------------\n# Get camera parameters from the extracted images and CSV data.\n# -----------------------------------------------------------------------------\n\n\ndef get_params(\n    image_folder,\n    fl,\n    trans_body,\n    cam_x,\n    cam_y,\n    cam_z,\n    fps,\n    cam_pitch_,\n    cam_roll_,\n    cam_yaw_,\n):\n    all_images = sorted(glob(os.path.join(image_folder, \"*\" + IMG_FORMAT)))\n    imgnames, cam_ext, cam_int = [], [], []\n\n    for img_ind, image_path in enumerate(all_images):\n        # Process every 5th frame.\n        if img_ind % 5 != 0:\n            continue\n        cam_ind = img_ind\n\n        cam_pitch_ind = cam_pitch_[cam_ind]\n        cam_yaw_ind = cam_yaw_[cam_ind]\n        cam_roll_ind = cam_roll_[cam_ind]\n\n        CAM_INT = get_cam_int(fl[cam_ind], SENSOR_W, SENSOR_H, IMG_W / 2.0, IMG_H / 2.0)\n\n        rot_unreal = rotation_matrix_unreal(cam_yaw_ind, cam_pitch_ind, cam_roll_ind)\n        rot_cv = convert_rotation_to_opencv(rot_unreal)\n        trans_cv = convert_translation_to_opencv(\n            cam_x[cam_ind] / 100.0, cam_y[cam_ind] / 100.0, cam_z[cam_ind] / 100.0\n        )\n        cam_ext_ = np.eye(4)\n        cam_ext_[:3, :3] = rot_cv\n        # The camera pose is computed as the inverse of the transformed translation.\n        cam_ext_[:3, 3] = -rot_cv @ trans_cv\n\n        imgnames.append(\n            os.path.join(image_path.split(\"/\")[-2], image_path.split(\"/\")[-1])\n        )\n        cam_ext.append(cam_ext_)  # camera_pose: c2w\n        cam_int.append(CAM_INT)\n    return imgnames, cam_ext, cam_int\n\n\n# -----------------------------------------------------------------------------\n# Processing per sequence.\n# -----------------------------------------------------------------------------\n\n\ndef process_seq(args):\n    \"\"\"\n    Process a single sequence task. For each image, load the corresponding\n    depth and image files, and save the computed camera intrinsics and the inverse\n    of the extrinsic matrix (i.e. the camera pose in world coordinates) as an NPZ file.\n    \"\"\"\n    (\n        scene,\n        seq_name,\n        outdir,\n        image_folder_base,\n        depth_folder_base,\n        mask_folder_base,","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.process_seq","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.process_seq#L252-L326","kind":"function","name":"process_seq","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":252,"end_line":326,"context_start_line":232,"context_end_line":346,"code":"            cam_x[cam_ind] / 100.0, cam_y[cam_ind] / 100.0, cam_z[cam_ind] / 100.0\n        )\n        cam_ext_ = np.eye(4)\n        cam_ext_[:3, :3] = rot_cv\n        # The camera pose is computed as the inverse of the transformed translation.\n        cam_ext_[:3, 3] = -rot_cv @ trans_cv\n\n        imgnames.append(\n            os.path.join(image_path.split(\"/\")[-2], image_path.split(\"/\")[-1])\n        )\n        cam_ext.append(cam_ext_)  # camera_pose: c2w\n        cam_int.append(CAM_INT)\n    return imgnames, cam_ext, cam_int\n\n\n# -----------------------------------------------------------------------------\n# Processing per sequence.\n# -----------------------------------------------------------------------------\n\n\ndef process_seq(args):\n    \"\"\"\n    Process a single sequence task. For each image, load the corresponding\n    depth and image files, and save the computed camera intrinsics and the inverse\n    of the extrinsic matrix (i.e. the camera pose in world coordinates) as an NPZ file.\n    \"\"\"\n    (\n        scene,\n        seq_name,\n        outdir,\n        image_folder_base,\n        depth_folder_base,\n        mask_folder_base,\n        imgnames,\n        cam_ext,\n        cam_int,\n        humans,\n        imgname_array,\n    ) = args\n\n    split = 'Test' if scene in test_list else 'Training'\n    \n    out_rgb_dir = os.path.join(outdir, split, '_'.join([scene, seq_name]), 'rgb')\n    out_depth_dir = os.path.join(outdir, split, '_'.join([scene, seq_name]), 'depth')\n    out_cam_dir = os.path.join(outdir, split, \"_\".join([scene, seq_name]), \"cam\")\n    out_smpl_dir = os.path.join(outdir, split, \"_\".join([scene, seq_name]), \"smpl\")\n    out_mask_dir = os.path.join(outdir, split, '_'.join([scene, seq_name]), 'mask')\n    os.makedirs(out_rgb_dir, exist_ok=True)\n    os.makedirs(out_depth_dir, exist_ok=True)\n    os.makedirs(out_cam_dir, exist_ok=True)\n    os.makedirs(out_smpl_dir, exist_ok=True)\n    os.makedirs(out_mask_dir, exist_ok=True)\n\n    assert (\n        len(imgnames) == len(cam_ext) == len(cam_int)\n    ), f\"Inconsistent lengths for {scene}_{seq_name}\"\n    \n    invalid_images = []\n    for imgname, ext, intr, human in zip(imgnames, cam_ext, cam_int, humans):\n        if imgname not in imgname_array:\n            invalid_images.append(f\"Invalid image: {scene}/{imgname}\")\n            continue\n        \n        depthname = imgname.replace(\".png\", \"_depth.exr\")\n        maskname = imgname.replace(\".png\", \"_env.png\")\n        imgpath = os.path.join(image_folder_base, imgname)\n        depthpath = os.path.join(depth_folder_base, depthname)\n        maskpath = os.path.join(mask_folder_base, maskname)\n        depth= OpenEXR.File(depthpath).parts[0].channels['Depth'].pixels\n        depth = depth.astype(np.float32)/100.0\n\n        mask = cv2.imread(maskpath, cv2.IMREAD_GRAYSCALE)\n        if mask is None:\n            # if mask file not exist, create a zero mask\n            mask = np.zeros((IMG_H, IMG_W), dtype=np.uint8)\n        else:\n            # invert the mask\n            mask = 255 - mask      \n        \n        outimg_path = os.path.join(out_rgb_dir, os.path.basename(imgpath))\n        outmask_path = os.path.join(out_mask_dir, os.path.basename(imgpath))\n        outdepth_path = os.path.join(out_depth_dir, os.path.basename(imgpath).replace('.png','.npy'))\n        outcam_path = os.path.join(\n            out_cam_dir, os.path.basename(imgpath).replace(\".png\", \".npz\")\n        )\n\n        out_smpl_path = os.path.join(out_smpl_dir, os.path.basename(imgpath).replace(\".png\", \".pkl\"))\n\n        shutil.copy(imgpath, outimg_path)\n        cv2.imwrite(outmask_path, mask)\n        np.save(outdepth_path, depth)\n        np.savez(outcam_path, intrinsics=intr, pose=np.linalg.inv(ext))  # pose: w2c\n        with open(out_smpl_path, 'wb') as f:\n            pickle.dump(human, f, protocol=pickle.HIGHEST_PROTOCOL)\n    return invalid_images\n\n\n# -----------------------------------------------------------------------------\n# Main entry point.\n# -----------------------------------------------------------------------------\n\n\ndef main():\n    parser = argparse.ArgumentParser(\n        description=\"Process Bedlam scenes: compute camera intrinsics and extrinsics, \"\n        \"and save processed camera files.\"\n    )\n    parser.add_argument(\n        \"--root\",\n        type=str,\n        required=True,\n        help=\"Root directory of the extracted data (scenes).\",\n    )\n    parser.add_argument(\n        \"--outdir\", type=str, required=True, help=\"Output directory for processed data.\"","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:datasets_preprocess.preprocess_bedlam.main","uri":"program://Human3R/function/datasets_preprocess.preprocess_bedlam.main#L334-L526","kind":"function","name":"main","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":334,"end_line":526,"context_start_line":314,"context_end_line":530,"code":"        outcam_path = os.path.join(\n            out_cam_dir, os.path.basename(imgpath).replace(\".png\", \".npz\")\n        )\n\n        out_smpl_path = os.path.join(out_smpl_dir, os.path.basename(imgpath).replace(\".png\", \".pkl\"))\n\n        shutil.copy(imgpath, outimg_path)\n        cv2.imwrite(outmask_path, mask)\n        np.save(outdepth_path, depth)\n        np.savez(outcam_path, intrinsics=intr, pose=np.linalg.inv(ext))  # pose: w2c\n        with open(out_smpl_path, 'wb') as f:\n            pickle.dump(human, f, protocol=pickle.HIGHEST_PROTOCOL)\n    return invalid_images\n\n\n# -----------------------------------------------------------------------------\n# Main entry point.\n# -----------------------------------------------------------------------------\n\n\ndef main():\n    parser = argparse.ArgumentParser(\n        description=\"Process Bedlam scenes: compute camera intrinsics and extrinsics, \"\n        \"and save processed camera files.\"\n    )\n    parser.add_argument(\n        \"--root\",\n        type=str,\n        required=True,\n        help=\"Root directory of the extracted data (scenes).\",\n    )\n    parser.add_argument(\n        \"--outdir\", type=str, required=True, help=\"Output directory for processed data.\"\n    )\n    parser.add_argument(\n        \"--annot_dir\", type=str, required=True, help=\"Annotation directory.\"\n    )\n    parser.add_argument(\n        \"--num_workers\",\n        type=int,\n        default=None,\n        help=\"Number of worker processes (default: os.cpu_count()//2).\",\n    )\n    args = parser.parse_args()\n\n    root = args.root\n    outdir = args.outdir\n    annot_dir = args.annot_dir\n    num_workers = (\n        args.num_workers if args.num_workers is not None else (os.cpu_count() or 4) // 2\n    )\n\n    invalid_list = []\n\n    # Get scene directories from the root folder.\n    scenes = sorted(\n        [d for d in os.listdir(root) if os.path.isdir(os.path.join(root, d))]\n    )\n    # Exclude HDRI scenes.\n    hdri_scenes = [\n        \"20221010_3_1000_batch01hand\",\n        \"20221017_3_1000_batch01hand\",\n        \"20221018_3-8_250_batch01hand\", # filtered out by CUT3R, but Multi-HMR involves it\n        \"20221019_3_250_highbmihand\",\n    ]\n    scenes = np.setdiff1d(scenes, hdri_scenes)\n\n    max_human = 0\n\n    tasks = []\n    for scene in tqdm(scenes, desc=\"Collecting tasks\"):\n        # Skip closeup scenes.\n        if \"closeup\" in scene:\n            continue\n        base_folder = os.path.join(root, scene)\n        image_folder_base = os.path.join(root, scene, \"png\")\n        depth_folder_base = os.path.join(root, scene, \"depth\")\n        mask_folder_base = os.path.join(root, scene, \"masks\")\n        csv_path = os.path.join(base_folder, \"be_seq.csv\")\n        annot_path = os.path.join(annot_dir, scene + \"_6fps.npz\")\n        if not os.path.exists(annot_path):\n            annot_path = os.path.join(annot_dir, scene + \"_30fps.npz\")\n        if not os.path.exists(csv_path):\n            continue\n        csv_data = pd.read_csv(csv_path)\n        csv_data = csv_data.to_dict(\"list\")\n        cam_csv_base = os.path.join(base_folder, \"ground_truth\", \"camera\")\n        annot_x = np.load(annot_path)\n\n        # Retrieving SMPL parameters once\n        pose_cam_array = annot_x['pose_cam']\n        H_array = annot_x['cam_ext']\n        shape_array = annot_x['shape']\n        imgname_array = annot_x['imgname']\n        trans_cam_array = annot_x['trans_cam']\n        \n        seq_count = 0\n        max_seq_per_scene = 1500000000000\n\n        # Look for a row in the CSV with a \"sequence_name\" comment.\n        for idx, comment in enumerate(csv_data.get(\"Comment\", [])):\n            if \"sequence_name\" in comment:\n                if seq_count >= max_seq_per_scene:\n                    break\n                    \n                seq_name = comment.split(\";\")[0].split(\"=\")[-1]\n                if not np.any(np.char.startswith(imgname_array, seq_name + '/')):\n                    invalid_list.append(f\"Invalid sequence: {scene}/{seq_name}\")\n                    continue\n                  \n                cam_csv_path = os.path.join(cam_csv_base, seq_name + \"_camera.csv\")\n                if not os.path.exists(cam_csv_path):\n                    continue\n                cam_csv_data = pd.read_csv(cam_csv_path)\n                cam_csv_data = cam_csv_data.to_dict(\"list\")\n                cam_x = cam_csv_data[\"x\"]\n                cam_y = cam_csv_data[\"y\"]\n                cam_z = cam_csv_data[\"z\"]\n                cam_yaw_ = cam_csv_data[\"yaw\"]\n                cam_pitch_ = cam_csv_data[\"pitch\"]\n                cam_roll_ = cam_csv_data[\"roll\"]\n                fl = cam_csv_data[\"focal_length\"]\n                image_folder = os.path.join(image_folder_base, seq_name)\n                trans_body = None  # Not used here.\n                imgnames, cam_ext, cam_int = get_params(\n                    image_folder,\n                    fl,\n                    trans_body,\n                    cam_x,\n                    cam_y,\n                    cam_z,\n                    6,\n                    cam_pitch_=cam_pitch_,\n                    cam_roll_=cam_roll_,\n                    cam_yaw_=cam_yaw_,\n                )\n                humans = [] # humans for each sequence\n                for imgname in imgnames:\n                    idxs = np.where(imgname == imgname_array)[0]\n\n                    if len(idxs) > max_human:\n                        max_human = len(idxs)\n\n                    persons_per_img = [] # persons for each image\n                    for i in idxs:\n                        sys.stdout.flush()\n\n                        # SMPLX params\n                        pose = pose_cam_array[i]\n                        root_pose = pose[:3]\n                        body_pose=pose[3:66]\n                        jaw_pose=pose[66:69]\n                        leye_pose=pose[69:72]\n                        reye_pose=pose[72:75]\n                        left_hand_pose=pose[75:120]\n                        right_hand_pose=pose[120:165]\n                        betas=shape_array[i]\n                        transl = trans_cam_array[i] + H_array[i][:, 3][:3]\n\n                        person = {\n                            # SMPL GT in camera coordinates system\n                            'smplx_root_pose': root_pose.reshape(1,3), # axis-angle\n                            'smplx_body_pose': body_pose.reshape(21,3), # axis-angle\n                            'smplx_jaw_pose': jaw_pose.reshape(1,3), # axis-angle\n                            'smplx_leye_pose': leye_pose.reshape(1,3), # axis-angle\n                            'smplx_reye_pose': reye_pose.reshape(1,3), # axis-angle\n                            'smplx_left_hand_pose': left_hand_pose.reshape(15,3), # axis-angle\n                            'smplx_right_hand_pose': right_hand_pose.reshape(15,3), # axis-angle\n                            'smplx_shape': betas.reshape(11),\n                            'smplx_gender': 'neutral',\n                            'smplx_transl': transl.reshape(3),\n                        }\n                        persons_per_img.append(person)\n                    humans.append(persons_per_img)\n\n                tasks.append(\n                    (\n                        scene,\n                        seq_name,\n                        outdir,\n                        image_folder_base,\n                        depth_folder_base,\n                        mask_folder_base,\n                        imgnames,\n                        cam_ext,\n                        cam_int,\n                        humans,\n                        imgname_array,\n                    )\n                )\n\n                seq_count += 1\n\n    print(f\"max_human: {max_human}\")\n    \n    # Process each task in parallel.\n    with ProcessPoolExecutor(max_workers=num_workers) as executor:\n        futures = {executor.submit(process_seq, task): task for task in tasks}\n        for future in tqdm(\n            as_completed(futures), total=len(futures), desc=\"Processing sequences\"\n        ):\n            # error = future.result()\n            # if error:\n            #     print(error)\n            invalid_images = future.result()\n            if invalid_images:\n                invalid_list.extend(invalid_images)\n\n    # print invalid items\n    if invalid_list:\n        print(\"\\nInvalid sequences and images:\")\n        for item in invalid_list:\n            print(item)\n\n\nif __name__ == \"__main__\":\n    main()","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train","uri":"program://Human3R/module/src.train#L1-L1049","kind":"module","name":"src.train","path":"src/train.py","language":"python","start_line":1,"end_line":1049,"context_start_line":1,"context_end_line":1049,"code":"# --------------------------------------------------------\n# training code for Human3R\n# --------------------------------------------------------\n# References:\n# CUT3R: https://github.com/CUT3R/CUT3R\n# DUSt3R: https://github.com/naver/dust3r\n# --------------------------------------------------------\nimport argparse\nimport datetime\nimport json\nimport numpy as np\nimport os\nimport sys\nimport time\nimport math\nfrom collections import defaultdict\nfrom pathlib import Path\nfrom typing import Sized\n\nimport torch\nimport torch.backends.cudnn as cudnn\nimport torch.nn.functional as F\nfrom torch.utils.tensorboard import SummaryWriter\n\nfrom dust3r.utils.device import todevice\n\ntorch.backends.cuda.matmul.allow_tf32 = True  # for gpu >= Ampere and pytorch >= 1.12\n\nfrom dust3r.model import (\n    PreTrainedModel,\n    ARCroco3DStereo,\n    ARCroco3DStereoConfig,\n    inf,\n    strip_module,\n    strip_module_mhmr,\n)  # noqa: F401, needed when loading the model\nfrom dust3r.smpl_model import SMPLModel\nfrom dust3r.datasets import get_data_loader\nfrom dust3r.losses import *  # noqa: F401, needed when loading the model\nfrom dust3r.inference import loss_of_one_batch  # noqa\nfrom dust3r.viz import colorize\nfrom dust3r.utils.render import get_render_results, get_render_smpl\nimport dust3r.utils.path_to_croco  # noqa: F401\nimport croco.utils.misc as misc  # noqa\nfrom croco.utils.misc import NativeScalerWithGradNormCount as NativeScaler  # noqa\n\nimport hydra\nfrom omegaconf import OmegaConf\nimport logging\nimport pathlib\nfrom tqdm import tqdm\nimport random\nimport builtins\nimport shutil\n\nfrom accelerate import Accelerator\nfrom accelerate import DistributedDataParallelKwargs, InitProcessGroupKwargs\nfrom accelerate.logging import get_logger\nfrom datetime import timedelta\nimport torch.multiprocessing\n\ntorch.multiprocessing.set_sharing_strategy(\"file_system\")\n\nprinter = get_logger(__name__, log_level=\"DEBUG\")\n\n\ndef setup_for_distributed(accelerator: Accelerator):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (accelerator.num_processes > 8)\n        if accelerator.is_main_process or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef save_current_code(outdir):\n    now = datetime.datetime.now()  # current date and time\n    date_time = now.strftime(\"%m_%d-%H:%M:%S\")\n    src_dir = \".\"\n    dst_dir = os.path.join(outdir, \"code\", \"{}\".format(date_time))\n    shutil.copytree(\n        src_dir,\n        dst_dir,\n        ignore=shutil.ignore_patterns(\n            \".vscode*\",\n            \"assets*\",\n            \"example*\",\n            \"checkpoints*\",\n            \"OLD*\",\n            \"logs*\",\n            \"out*\",\n            \"runs*\",\n            \"*.png\",\n            \"*.mp4\",\n            \"*__pycache__*\",\n            \"*.git*\",\n            \"*.idea*\",\n            \"*.zip\",\n            \"*.jpg\",\n            \"*.pth\",\n            \"*.pt\",\n            \"*.npy\",\n            \"*.npz\",\n            \"*.pkl\",\n        ),\n        dirs_exist_ok=True,\n    )\n    return dst_dir\n\n\ndef train(args):\n\n    accelerator = Accelerator(\n        gradient_accumulation_steps=args.accum_iter,\n        mixed_precision=\"bf16\",\n        kwargs_handlers=[\n            DistributedDataParallelKwargs(find_unused_parameters=True),\n            InitProcessGroupKwargs(timeout=timedelta(seconds=6000)),\n        ],\n    )\n    device = accelerator.device\n\n    setup_for_distributed(accelerator)\n\n    printer.info(\"output_dir: \" + args.output_dir)\n    if args.output_dir:\n        Path(args.output_dir).mkdir(parents=True, exist_ok=True)\n\n    if accelerator.is_main_process:\n        dst_dir = save_current_code(outdir=args.output_dir)\n        printer.info(f\"Saving current code to {dst_dir}\")\n\n    # auto resume\n    if not args.resume:\n        last_ckpt_fname = os.path.join(args.output_dir, f\"checkpoint-last.pth\")\n        args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None\n\n    printer.info(\"job dir: {}\".format(os.path.dirname(os.path.realpath(__file__))))\n\n    # fix the seed\n    seed = args.seed + accelerator.state.process_index\n    printer.info(\n        f\"Setting seed to {seed} for process {accelerator.state.process_index}\"\n    )\n    torch.manual_seed(seed)\n    np.random.seed(seed)\n    random.seed(seed)\n    cudnn.benchmark = args.benchmark\n\n    # training dataset and loader\n    printer.info(\"Building train dataset %s\", args.train_dataset)\n    #  dataset and loader\n    data_loader_train = build_dataset(\n        args.train_dataset,\n        args.batch_size,\n        args.num_workers,\n        accelerator=accelerator,\n        test=False,\n        fixed_length=args.fixed_length\n    )\n    printer.info(\"Building test dataset %s\", args.test_dataset)\n    data_loader_test = {\n        dataset.split(\"(\")[0]: build_dataset(\n            dataset,\n            args.batch_size,\n            args.num_workers,\n            accelerator=accelerator,\n            test=True,\n            fixed_length=True\n        )\n        for dataset in args.test_dataset.split(\"+\")\n    }\n\n    # model\n    printer.info(\"Loading model: %s\", args.model)\n    model: PreTrainedModel = eval(args.model)\n    smpl_model: SMPLModel = SMPLModel(\n        device, \n        model_args={\n            'patch_size': model.croco_args['patch_size'], \n            'mhmr_img_res': model.mhmr_img_res, \n            'bb_patch_size': model.bb_patch_size\n        })\n    printer.info(f\"All model parameters: {sum(p.numel() for p in model.parameters())}\")\n    printer.info(\n        f\"Encoder parameters: {sum(p.numel() for p in model.enc_blocks.parameters())}\"\n    )\n    printer.info(\n        f\"Decoder parameters: {sum(p.numel() for p in model.dec_blocks.parameters())}\"\n    )\n\n    printer.info(f\">> Creating train criterion = {args.train_criterion}\")\n    train_criterion = eval(args.train_criterion).to(device)\n    printer.info(\n        f\">> Creating test criterion = {args.test_criterion or args.train_criterion}\"\n    )\n    test_criterion = eval(args.test_criterion or args.criterion).to(device)\n\n    model.to(device)\n\n    if args.gradient_checkpointing:\n        model.gradient_checkpointing_enable()\n    if args.long_context:\n        model.fixed_input_length = False\n\n    if args.pretrained and not args.resume:\n        printer.info(f\"Loading pretrained: {args.pretrained}\")\n        ckpt = torch.load(args.pretrained, map_location=device)\n        load_only_encoder = getattr(args, \"load_only_encoder\", False)\n        if load_only_encoder:\n            filtered_state_dict = {\n                k: v\n                for k, v in ckpt[\"model\"].items()\n                if \"enc_blocks\" in k or \"patch_embed\" in k\n            }\n            merge_state_dict = strip_module(filtered_state_dict)\n        else:\n            merge_state_dict = strip_module(ckpt[\"model\"])\n        del ckpt  # in case it occupies memory\n\n        if args.pretrained_mhmr:\n            printer.info(f\"Loading Multi-HMR pretrained: {args.pretrained_mhmr}\")\n            ckpt_mhmr = torch.load(args.pretrained_mhmr, map_location=device)\n            merge_state_dict.update(strip_module_mhmr(ckpt_mhmr[\"model_state_dict\"]))\n            del ckpt_mhmr  # in case it occupies memory\n\n        printer.info(\n            model.load_state_dict(merge_state_dict, strict=False)\n        )\n\n    # # following timm: set wd as 0 for bias and norm layers\n    param_groups = misc.get_parameter_groups(model, args.weight_decay)\n    optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))\n    # print(optimizer)\n    loss_scaler = NativeScaler(accelerator=accelerator)\n\n    accelerator.even_batches = False\n    optimizer, model, data_loader_train = accelerator.prepare(\n        optimizer, model, data_loader_train\n    )\n\n    def write_log_stats(epoch, train_stats, test_stats):\n        if accelerator.is_main_process:\n            if log_writer is not None:\n                log_writer.flush()\n\n            log_stats = dict(\n                epoch=epoch, **{f\"train_{k}\": v for k, v in train_stats.items()}\n            )\n            for test_name in data_loader_test:\n                if test_name not in test_stats:\n                    continue\n                log_stats.update(\n                    {test_name + \"_\" + k: v for k, v in test_stats[test_name].items()}\n                )\n\n            with open(\n                os.path.join(args.output_dir, \"log.txt\"), mode=\"a\", encoding=\"utf-8\"\n            ) as f:\n                f.write(json.dumps(log_stats) + \"\\n\")\n\n    def save_model(epoch, fname, best_so_far):\n        misc.save_model(\n            accelerator=accelerator,\n            args=args,\n            model_without_ddp=model,\n            optimizer=optimizer,\n            loss_scaler=loss_scaler,\n            epoch=epoch,\n            fname=fname,\n            best_so_far=best_so_far,\n        )\n\n    best_so_far = misc.load_model(\n        args=args, model_without_ddp=model, optimizer=optimizer, loss_scaler=loss_scaler\n    )\n    if best_so_far is None:\n        best_so_far = float(\"inf\")\n    log_writer = (\n        SummaryWriter(log_dir=args.output_dir) if accelerator.is_main_process else None\n    )\n\n    printer.info(f\"Start training for {args.epochs} epochs\")\n    start_time = time.time()\n    train_stats = test_stats = {}\n\n    for epoch in range(args.start_epoch, args.epochs + 1):\n\n        # Save immediately the last checkpoint\n        if epoch > args.start_epoch:\n            if (\n                args.save_freq\n                and np.allclose(epoch / args.save_freq, int(epoch / args.save_freq))\n                or epoch == args.epochs\n            ):\n                save_model(epoch - 1, \"last\", best_so_far)\n\n        # Test on multiple datasets\n        new_best = False\n        if epoch >= 0 and args.eval_freq > 0 and epoch % args.eval_freq == 0:\n            test_stats = {}\n            for test_name, testset in data_loader_test.items():\n                stats = test_one_epoch(\n                    model,\n                    test_criterion,\n                    testset,\n                    accelerator,\n                    device,\n                    epoch,\n                    log_writer=log_writer,\n                    args=args,\n                    prefix=test_name,\n                    smpl_model=smpl_model,\n                )\n                test_stats[test_name] = stats\n\n                # Save best of all\n                if stats[\"loss_med\"] < best_so_far:\n                    best_so_far = stats[\"loss_med\"]\n                    new_best = True\n        # Save more stuff\n        write_log_stats(epoch, train_stats, test_stats)\n\n        if epoch > args.start_epoch:\n            if args.keep_freq and epoch % args.keep_freq == 0:\n                save_model(epoch - 1, str(epoch), best_so_far)\n            if new_best:\n                save_model(epoch - 1, \"best\", best_so_far)\n        if epoch >= args.epochs:\n            break  # exit after writing last test to disk\n\n        # Train\n        train_stats = train_one_epoch(\n            model,\n            train_criterion,\n            data_loader_train,\n            optimizer,\n            accelerator,\n            epoch,\n            loss_scaler,\n            log_writer=log_writer,\n            args=args,\n            smpl_model=smpl_model,\n        )\n\n    total_time = time.time() - start_time\n    total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n    printer.info(\"Training time {}\".format(total_time_str))\n\n    save_final_model(accelerator, args, args.epochs, model, best_so_far=best_so_far)\n\n\ndef save_final_model(accelerator, args, epoch, model_without_ddp, best_so_far=None):\n    output_dir = Path(args.output_dir)\n    checkpoint_path = output_dir / \"checkpoint-final.pth\"\n    to_save = {\n        \"args\": args,\n        \"model\": (\n            model_without_ddp\n            if isinstance(model_without_ddp, dict)\n            else model_without_ddp.cpu().state_dict()\n        ),\n        \"epoch\": epoch,\n    }\n    if best_so_far is not None:\n        to_save[\"best_so_far\"] = best_so_far\n    printer.info(f\">> Saving model to {checkpoint_path} ...\")\n    misc.save_on_master(accelerator, to_save, checkpoint_path)\n\n\ndef build_dataset(dataset, batch_size, num_workers, accelerator, test=False, fixed_length=False):\n    split = [\"Train\", \"Test\"][test]\n    printer.info(f\"Building {split} Data loader for dataset: {dataset}\")\n    loader = get_data_loader(\n        dataset,\n        batch_size=batch_size,\n        num_workers=num_workers,\n        pin_mem=True,\n        shuffle=not (test),\n        drop_last=not (test),\n        accelerator=accelerator,\n        fixed_length=fixed_length\n    )\n    return loader\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    data_loader: Sized,\n    optimizer: torch.optim.Optimizer,\n    accelerator: Accelerator,\n    epoch: int,\n    loss_scaler,\n    args,\n    log_writer=None,\n    smpl_model: SMPLModel = None\n):\n    assert torch.backends.cuda.matmul.allow_tf32 == True\n\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)\n    accum_iter = args.accum_iter\n\n    def save_model(epoch, fname, best_so_far):\n        misc.save_model(\n            accelerator=accelerator,\n            args=args,\n            model_without_ddp=model,\n            optimizer=optimizer,\n            loss_scaler=loss_scaler,\n            epoch=epoch,\n            fname=fname,\n            best_so_far=best_so_far,\n        )\n\n    if log_writer is not None:\n        printer.info(\"log_dir: {}\".format(log_writer.log_dir))\n\n    if hasattr(data_loader, \"dataset\") and hasattr(data_loader.dataset, \"set_epoch\"):\n        data_loader.dataset.set_epoch(epoch)\n    if (\n        hasattr(data_loader, \"batch_sampler\")\n        and hasattr(data_loader.batch_sampler, \"batch_sampler\")\n        and hasattr(data_loader.batch_sampler.batch_sampler, \"set_epoch\")\n    ):\n        data_loader.batch_sampler.batch_sampler.set_epoch(epoch)\n\n    optimizer.zero_grad()\n\n    for data_iter_step, batch in enumerate(\n        metric_logger.log_every(data_loader, args.print_freq, accelerator, header)\n    ):\n        with accelerator.accumulate(model):\n            epoch_f = epoch + data_iter_step / len(data_loader)\n            step = int(epoch_f * len(data_loader))\n            # we use a per iteration (instead of per epoch) lr scheduler\n            if data_iter_step % accum_iter == 0:\n                misc.adjust_learning_rate(optimizer, epoch_f, args)\n            if not args.long_context:\n                result = loss_of_one_batch(\n                    batch,\n                    model,\n                    criterion,\n                    accelerator,\n                    symmetrize_batch=False,\n                    use_amp=bool(args.amp),\n                    smpl_model=smpl_model\n                )\n            else:\n                NotImplementedError(\"Long context is not supported\")\n            has_msk = \"msk\" in result[\"pred\"][0]\n            loss, loss_details = result[\"loss\"]  # criterion returns two values\n            loss_value = float(loss)\n\n            if not math.isfinite(loss_value):\n                print(\n                    f\"Loss is {loss_value}, stopping training, loss details: {loss_details}\"\n                )\n                sys.exit(1)\n            if not result.get(\"already_backprop\", False):\n                loss_scaler(\n                    loss,\n                    optimizer,\n                    parameters=model.parameters(),\n                    update_grad=True,\n                    clip_grad=1.0,\n                )\n                optimizer.zero_grad()\n\n            is_metric = batch[0][\"is_metric\"]\n            curr_num_view = len(batch)\n\n            del loss\n            tb_vis_img = (data_iter_step + 1) % accum_iter == 0 and (\n                (step + 1) % (args.print_img_freq)\n            ) == 0\n            if not tb_vis_img:\n                del batch\n            else:\n                torch.cuda.empty_cache()\n\n            lr = optimizer.param_groups[0][\"lr\"]\n            metric_logger.update(epoch=epoch_f)\n            metric_logger.update(lr=lr)\n            metric_logger.update(step=step)\n\n            metric_logger.update(loss=loss_value, **loss_details)\n\n            if (data_iter_step + 1) % accum_iter == 0 and (\n                (data_iter_step + 1) % (accum_iter * args.print_freq)\n            ) == 0:\n                loss_value_reduce = accelerator.gather(\n                    torch.tensor(loss_value).to(accelerator.device)\n                ).mean()  # MUST BE EXECUTED BY ALL NODES\n\n                if log_writer is None:\n                    continue\n                \"\"\" We use epoch_1000x as the x-axis in tensorboard.\n                This calibrates different curves when batch size changes.\n                \"\"\"\n                epoch_1000x = int(epoch_f * 1000)\n                log_writer.add_scalar(\"train_loss\", loss_value_reduce, step)\n                log_writer.add_scalar(\"train_lr\", lr, step)\n                log_writer.add_scalar(\"train_iter\", epoch_1000x, step)\n                for name, val in loss_details.items():\n                    if isinstance(val, torch.Tensor):\n                        if val.ndim > 0:\n                            continue\n                    if isinstance(val, dict):\n                        continue\n                    log_writer.add_scalar(\"train_\" + name, val, step)\n\n            if tb_vis_img:\n                if log_writer is None:\n                    continue\n                with torch.no_grad():\n                    depths_self, gt_depths_self = get_render_results(\n                        batch, result[\"pred\"], self_view=True\n                    )\n                    depths_cross, gt_depths_cross = get_render_results(\n                        batch, result[\"pred\"], self_view=False\n                    )\n                    gt_msks, pr_msks, gt_hms, pr_hms, gt_smpls, pr_smpls = get_render_smpl(\n                        batch, result[\"pred\"], smpl_model, loss_details, has_msk=has_msk\n                    )\n                    for k in range(len(batch)):\n                        loss_details[f\"self_pred_depth_{k+1}\"] = depths_self[k].detach().cpu()\n                        loss_details[f\"self_gt_depth_{k+1}\"] = gt_depths_self[k].detach().cpu()\n                        loss_details[f\"pred_depth_{k+1}\"] = depths_cross[k].detach().cpu()\n                        loss_details[f\"gt_depth_{k+1}\"] = gt_depths_cross[k].detach().cpu()           \n                        loss_details[f\"pred_hm_{k+1}\"] = pr_hms[k].detach().cpu()\n                        loss_details[f\"gt_hm_{k+1}\"] = gt_hms[k].detach().cpu()\n                        loss_details[f\"pred_smpl_rend_{k+1}\"] = pr_smpls[k].detach().cpu()\n                        loss_details[f\"gt_smpl_rend_{k+1}\"] = gt_smpls[k].detach().cpu()\n                        if has_msk:\n                            loss_details[f\"pred_msk_{k+1}\"] = pr_msks[k].detach().cpu()\n                            loss_details[f\"gt_msk_{k+1}\"] = gt_msks[k].detach().cpu()\n\n                imgs_stacked_dict = get_vis_imgs_new(\n                    loss_details, \n                    args.num_imgs_vis, \n                    curr_num_view, \n                    is_metric=is_metric, \n                    has_msk=has_msk)\n                for name, imgs_stacked in imgs_stacked_dict.items():\n                    log_writer.add_images(\n                        \"train\" + \"/\" + name, imgs_stacked, step, dataformats=\"HWC\"\n                    )\n                del batch\n\n        if (\n            data_iter_step % int(args.save_freq * len(data_loader)) == 0\n            and data_iter_step != 0\n            and data_iter_step != len(data_loader) - 1\n        ):\n            print(\"saving at step\", data_iter_step)\n            sa\n# ... truncated ...","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.train.setup_for_distributed","uri":"program://Human3R/function/src.train.setup_for_distributed#L67-L81","kind":"function","name":"setup_for_distributed","path":"src/train.py","language":"python","start_line":67,"end_line":81,"context_start_line":47,"context_end_line":101,"code":"import hydra\nfrom omegaconf import OmegaConf\nimport logging\nimport pathlib\nfrom tqdm import tqdm\nimport random\nimport builtins\nimport shutil\n\nfrom accelerate import Accelerator\nfrom accelerate import DistributedDataParallelKwargs, InitProcessGroupKwargs\nfrom accelerate.logging import get_logger\nfrom datetime import timedelta\nimport torch.multiprocessing\n\ntorch.multiprocessing.set_sharing_strategy(\"file_system\")\n\nprinter = get_logger(__name__, log_level=\"DEBUG\")\n\n\ndef setup_for_distributed(accelerator: Accelerator):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (accelerator.num_processes > 8)\n        if accelerator.is_main_process or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef save_current_code(outdir):\n    now = datetime.datetime.now()  # current date and time\n    date_time = now.strftime(\"%m_%d-%H:%M:%S\")\n    src_dir = \".\"\n    dst_dir = os.path.join(outdir, \"code\", \"{}\".format(date_time))\n    shutil.copytree(\n        src_dir,\n        dst_dir,\n        ignore=shutil.ignore_patterns(\n            \".vscode*\",\n            \"assets*\",\n            \"example*\",\n            \"checkpoints*\",\n            \"OLD*\",\n            \"logs*\",\n            \"out*\",\n            \"runs*\",\n            \"*.png\",","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.save_current_code","uri":"program://Human3R/function/src.train.save_current_code#L84-L116","kind":"function","name":"save_current_code","path":"src/train.py","language":"python","start_line":84,"end_line":116,"context_start_line":64,"context_end_line":136,"code":"printer = get_logger(__name__, log_level=\"DEBUG\")\n\n\ndef setup_for_distributed(accelerator: Accelerator):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (accelerator.num_processes > 8)\n        if accelerator.is_main_process or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef save_current_code(outdir):\n    now = datetime.datetime.now()  # current date and time\n    date_time = now.strftime(\"%m_%d-%H:%M:%S\")\n    src_dir = \".\"\n    dst_dir = os.path.join(outdir, \"code\", \"{}\".format(date_time))\n    shutil.copytree(\n        src_dir,\n        dst_dir,\n        ignore=shutil.ignore_patterns(\n            \".vscode*\",\n            \"assets*\",\n            \"example*\",\n            \"checkpoints*\",\n            \"OLD*\",\n            \"logs*\",\n            \"out*\",\n            \"runs*\",\n            \"*.png\",\n            \"*.mp4\",\n            \"*__pycache__*\",\n            \"*.git*\",\n            \"*.idea*\",\n            \"*.zip\",\n            \"*.jpg\",\n            \"*.pth\",\n            \"*.pt\",\n            \"*.npy\",\n            \"*.npz\",\n            \"*.pkl\",\n        ),\n        dirs_exist_ok=True,\n    )\n    return dst_dir\n\n\ndef train(args):\n\n    accelerator = Accelerator(\n        gradient_accumulation_steps=args.accum_iter,\n        mixed_precision=\"bf16\",\n        kwargs_handlers=[\n            DistributedDataParallelKwargs(find_unused_parameters=True),\n            InitProcessGroupKwargs(timeout=timedelta(seconds=6000)),\n        ],\n    )\n    device = accelerator.device\n\n    setup_for_distributed(accelerator)\n\n    printer.info(\"output_dir: \" + args.output_dir)\n    if args.output_dir:\n        Path(args.output_dir).mkdir(parents=True, exist_ok=True)\n","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.train","uri":"program://Human3R/function/src.train.train#L119-L358","kind":"function","name":"train","path":"src/train.py","language":"python","start_line":119,"end_line":358,"context_start_line":99,"context_end_line":378,"code":"            \"out*\",\n            \"runs*\",\n            \"*.png\",\n            \"*.mp4\",\n            \"*__pycache__*\",\n            \"*.git*\",\n            \"*.idea*\",\n            \"*.zip\",\n            \"*.jpg\",\n            \"*.pth\",\n            \"*.pt\",\n            \"*.npy\",\n            \"*.npz\",\n            \"*.pkl\",\n        ),\n        dirs_exist_ok=True,\n    )\n    return dst_dir\n\n\ndef train(args):\n\n    accelerator = Accelerator(\n        gradient_accumulation_steps=args.accum_iter,\n        mixed_precision=\"bf16\",\n        kwargs_handlers=[\n            DistributedDataParallelKwargs(find_unused_parameters=True),\n            InitProcessGroupKwargs(timeout=timedelta(seconds=6000)),\n        ],\n    )\n    device = accelerator.device\n\n    setup_for_distributed(accelerator)\n\n    printer.info(\"output_dir: \" + args.output_dir)\n    if args.output_dir:\n        Path(args.output_dir).mkdir(parents=True, exist_ok=True)\n\n    if accelerator.is_main_process:\n        dst_dir = save_current_code(outdir=args.output_dir)\n        printer.info(f\"Saving current code to {dst_dir}\")\n\n    # auto resume\n    if not args.resume:\n        last_ckpt_fname = os.path.join(args.output_dir, f\"checkpoint-last.pth\")\n        args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None\n\n    printer.info(\"job dir: {}\".format(os.path.dirname(os.path.realpath(__file__))))\n\n    # fix the seed\n    seed = args.seed + accelerator.state.process_index\n    printer.info(\n        f\"Setting seed to {seed} for process {accelerator.state.process_index}\"\n    )\n    torch.manual_seed(seed)\n    np.random.seed(seed)\n    random.seed(seed)\n    cudnn.benchmark = args.benchmark\n\n    # training dataset and loader\n    printer.info(\"Building train dataset %s\", args.train_dataset)\n    #  dataset and loader\n    data_loader_train = build_dataset(\n        args.train_dataset,\n        args.batch_size,\n        args.num_workers,\n        accelerator=accelerator,\n        test=False,\n        fixed_length=args.fixed_length\n    )\n    printer.info(\"Building test dataset %s\", args.test_dataset)\n    data_loader_test = {\n        dataset.split(\"(\")[0]: build_dataset(\n            dataset,\n            args.batch_size,\n            args.num_workers,\n            accelerator=accelerator,\n            test=True,\n            fixed_length=True\n        )\n        for dataset in args.test_dataset.split(\"+\")\n    }\n\n    # model\n    printer.info(\"Loading model: %s\", args.model)\n    model: PreTrainedModel = eval(args.model)\n    smpl_model: SMPLModel = SMPLModel(\n        device, \n        model_args={\n            'patch_size': model.croco_args['patch_size'], \n            'mhmr_img_res': model.mhmr_img_res, \n            'bb_patch_size': model.bb_patch_size\n        })\n    printer.info(f\"All model parameters: {sum(p.numel() for p in model.parameters())}\")\n    printer.info(\n        f\"Encoder parameters: {sum(p.numel() for p in model.enc_blocks.parameters())}\"\n    )\n    printer.info(\n        f\"Decoder parameters: {sum(p.numel() for p in model.dec_blocks.parameters())}\"\n    )\n\n    printer.info(f\">> Creating train criterion = {args.train_criterion}\")\n    train_criterion = eval(args.train_criterion).to(device)\n    printer.info(\n        f\">> Creating test criterion = {args.test_criterion or args.train_criterion}\"\n    )\n    test_criterion = eval(args.test_criterion or args.criterion).to(device)\n\n    model.to(device)\n\n    if args.gradient_checkpointing:\n        model.gradient_checkpointing_enable()\n    if args.long_context:\n        model.fixed_input_length = False\n\n    if args.pretrained and not args.resume:\n        printer.info(f\"Loading pretrained: {args.pretrained}\")\n        ckpt = torch.load(args.pretrained, map_location=device)\n        load_only_encoder = getattr(args, \"load_only_encoder\", False)\n        if load_only_encoder:\n            filtered_state_dict = {\n                k: v\n                for k, v in ckpt[\"model\"].items()\n                if \"enc_blocks\" in k or \"patch_embed\" in k\n            }\n            merge_state_dict = strip_module(filtered_state_dict)\n        else:\n            merge_state_dict = strip_module(ckpt[\"model\"])\n        del ckpt  # in case it occupies memory\n\n        if args.pretrained_mhmr:\n            printer.info(f\"Loading Multi-HMR pretrained: {args.pretrained_mhmr}\")\n            ckpt_mhmr = torch.load(args.pretrained_mhmr, map_location=device)\n            merge_state_dict.update(strip_module_mhmr(ckpt_mhmr[\"model_state_dict\"]))\n            del ckpt_mhmr  # in case it occupies memory\n\n        printer.info(\n            model.load_state_dict(merge_state_dict, strict=False)\n        )\n\n    # # following timm: set wd as 0 for bias and norm layers\n    param_groups = misc.get_parameter_groups(model, args.weight_decay)\n    optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))\n    # print(optimizer)\n    loss_scaler = NativeScaler(accelerator=accelerator)\n\n    accelerator.even_batches = False\n    optimizer, model, data_loader_train = accelerator.prepare(\n        optimizer, model, data_loader_train\n    )\n\n    def write_log_stats(epoch, train_stats, test_stats):\n        if accelerator.is_main_process:\n            if log_writer is not None:\n                log_writer.flush()\n\n            log_stats = dict(\n                epoch=epoch, **{f\"train_{k}\": v for k, v in train_stats.items()}\n            )\n            for test_name in data_loader_test:\n                if test_name not in test_stats:\n                    continue\n                log_stats.update(\n                    {test_name + \"_\" + k: v for k, v in test_stats[test_name].items()}\n                )\n\n            with open(\n                os.path.join(args.output_dir, \"log.txt\"), mode=\"a\", encoding=\"utf-8\"\n            ) as f:\n                f.write(json.dumps(log_stats) + \"\\n\")\n\n    def save_model(epoch, fname, best_so_far):\n        misc.save_model(\n            accelerator=accelerator,\n            args=args,\n            model_without_ddp=model,\n            optimizer=optimizer,\n            loss_scaler=loss_scaler,\n            epoch=epoch,\n            fname=fname,\n            best_so_far=best_so_far,\n        )\n\n    best_so_far = misc.load_model(\n        args=args, model_without_ddp=model, optimizer=optimizer, loss_scaler=loss_scaler\n    )\n    if best_so_far is None:\n        best_so_far = float(\"inf\")\n    log_writer = (\n        SummaryWriter(log_dir=args.output_dir) if accelerator.is_main_process else None\n    )\n\n    printer.info(f\"Start training for {args.epochs} epochs\")\n    start_time = time.time()\n    train_stats = test_stats = {}\n\n    for epoch in range(args.start_epoch, args.epochs + 1):\n\n        # Save immediately the last checkpoint\n        if epoch > args.start_epoch:\n            if (\n                args.save_freq\n                and np.allclose(epoch / args.save_freq, int(epoch / args.save_freq))\n                or epoch == args.epochs\n            ):\n                save_model(epoch - 1, \"last\", best_so_far)\n\n        # Test on multiple datasets\n        new_best = False\n        if epoch >= 0 and args.eval_freq > 0 and epoch % args.eval_freq == 0:\n            test_stats = {}\n            for test_name, testset in data_loader_test.items():\n                stats = test_one_epoch(\n                    model,\n                    test_criterion,\n                    testset,\n                    accelerator,\n                    device,\n                    epoch,\n                    log_writer=log_writer,\n                    args=args,\n                    prefix=test_name,\n                    smpl_model=smpl_model,\n                )\n                test_stats[test_name] = stats\n\n                # Save best of all\n                if stats[\"loss_med\"] < best_so_far:\n                    best_so_far = stats[\"loss_med\"]\n                    new_best = True\n        # Save more stuff\n        write_log_stats(epoch, train_stats, test_stats)\n\n        if epoch > args.start_epoch:\n            if args.keep_freq and epoch % args.keep_freq == 0:\n                save_model(epoch - 1, str(epoch), best_so_far)\n            if new_best:\n                save_model(epoch - 1, \"best\", best_so_far)\n        if epoch >= args.epochs:\n            break  # exit after writing last test to disk\n\n        # Train\n        train_stats = train_one_epoch(\n            model,\n            train_criterion,\n            data_loader_train,\n            optimizer,\n            accelerator,\n            epoch,\n            loss_scaler,\n            log_writer=log_writer,\n            args=args,\n            smpl_model=smpl_model,\n        )\n\n    total_time = time.time() - start_time\n    total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n    printer.info(\"Training time {}\".format(total_time_str))\n\n    save_final_model(accelerator, args, args.epochs, model, best_so_far=best_so_far)\n\n\ndef save_final_model(accelerator, args, epoch, model_without_ddp, best_so_far=None):\n    output_dir = Path(args.output_dir)\n    checkpoint_path = output_dir / \"checkpoint-final.pth\"\n    to_save = {\n        \"args\": args,\n        \"model\": (\n            model_without_ddp\n            if isinstance(model_without_ddp, dict)\n            else model_without_ddp.cpu().state_dict()\n        ),\n        \"epoch\": epoch,\n    }\n    if best_so_far is not None:\n        to_save[\"best_so_far\"] = best_so_far\n    printer.info(f\">> Saving model to {checkpoint_path} ...\")\n    misc.save_on_master(accelerator, to_save, checkpoint_path)\n\n","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.save_final_model","uri":"program://Human3R/function/src.train.save_final_model#L361-L376","kind":"function","name":"save_final_model","path":"src/train.py","language":"python","start_line":361,"end_line":376,"context_start_line":341,"context_end_line":396,"code":"        train_stats = train_one_epoch(\n            model,\n            train_criterion,\n            data_loader_train,\n            optimizer,\n            accelerator,\n            epoch,\n            loss_scaler,\n            log_writer=log_writer,\n            args=args,\n            smpl_model=smpl_model,\n        )\n\n    total_time = time.time() - start_time\n    total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n    printer.info(\"Training time {}\".format(total_time_str))\n\n    save_final_model(accelerator, args, args.epochs, model, best_so_far=best_so_far)\n\n\ndef save_final_model(accelerator, args, epoch, model_without_ddp, best_so_far=None):\n    output_dir = Path(args.output_dir)\n    checkpoint_path = output_dir / \"checkpoint-final.pth\"\n    to_save = {\n        \"args\": args,\n        \"model\": (\n            model_without_ddp\n            if isinstance(model_without_ddp, dict)\n            else model_without_ddp.cpu().state_dict()\n        ),\n        \"epoch\": epoch,\n    }\n    if best_so_far is not None:\n        to_save[\"best_so_far\"] = best_so_far\n    printer.info(f\">> Saving model to {checkpoint_path} ...\")\n    misc.save_on_master(accelerator, to_save, checkpoint_path)\n\n\ndef build_dataset(dataset, batch_size, num_workers, accelerator, test=False, fixed_length=False):\n    split = [\"Train\", \"Test\"][test]\n    printer.info(f\"Building {split} Data loader for dataset: {dataset}\")\n    loader = get_data_loader(\n        dataset,\n        batch_size=batch_size,\n        num_workers=num_workers,\n        pin_mem=True,\n        shuffle=not (test),\n        drop_last=not (test),\n        accelerator=accelerator,\n        fixed_length=fixed_length\n    )\n    return loader\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.build_dataset","uri":"program://Human3R/function/src.train.build_dataset#L379-L392","kind":"function","name":"build_dataset","path":"src/train.py","language":"python","start_line":379,"end_line":392,"context_start_line":359,"context_end_line":412,"code":"\n\ndef save_final_model(accelerator, args, epoch, model_without_ddp, best_so_far=None):\n    output_dir = Path(args.output_dir)\n    checkpoint_path = output_dir / \"checkpoint-final.pth\"\n    to_save = {\n        \"args\": args,\n        \"model\": (\n            model_without_ddp\n            if isinstance(model_without_ddp, dict)\n            else model_without_ddp.cpu().state_dict()\n        ),\n        \"epoch\": epoch,\n    }\n    if best_so_far is not None:\n        to_save[\"best_so_far\"] = best_so_far\n    printer.info(f\">> Saving model to {checkpoint_path} ...\")\n    misc.save_on_master(accelerator, to_save, checkpoint_path)\n\n\ndef build_dataset(dataset, batch_size, num_workers, accelerator, test=False, fixed_length=False):\n    split = [\"Train\", \"Test\"][test]\n    printer.info(f\"Building {split} Data loader for dataset: {dataset}\")\n    loader = get_data_loader(\n        dataset,\n        batch_size=batch_size,\n        num_workers=num_workers,\n        pin_mem=True,\n        shuffle=not (test),\n        drop_last=not (test),\n        accelerator=accelerator,\n        fixed_length=fixed_length\n    )\n    return loader\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    data_loader: Sized,\n    optimizer: torch.optim.Optimizer,\n    accelerator: Accelerator,\n    epoch: int,\n    loss_scaler,\n    args,\n    log_writer=None,\n    smpl_model: SMPLModel = None\n):\n    assert torch.backends.cuda.matmul.allow_tf32 == True\n\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.train_one_epoch","uri":"program://Human3R/function/src.train.train_one_epoch#L395-L573","kind":"function","name":"train_one_epoch","path":"src/train.py","language":"python","start_line":395,"end_line":573,"context_start_line":375,"context_end_line":593,"code":"    printer.info(f\">> Saving model to {checkpoint_path} ...\")\n    misc.save_on_master(accelerator, to_save, checkpoint_path)\n\n\ndef build_dataset(dataset, batch_size, num_workers, accelerator, test=False, fixed_length=False):\n    split = [\"Train\", \"Test\"][test]\n    printer.info(f\"Building {split} Data loader for dataset: {dataset}\")\n    loader = get_data_loader(\n        dataset,\n        batch_size=batch_size,\n        num_workers=num_workers,\n        pin_mem=True,\n        shuffle=not (test),\n        drop_last=not (test),\n        accelerator=accelerator,\n        fixed_length=fixed_length\n    )\n    return loader\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    data_loader: Sized,\n    optimizer: torch.optim.Optimizer,\n    accelerator: Accelerator,\n    epoch: int,\n    loss_scaler,\n    args,\n    log_writer=None,\n    smpl_model: SMPLModel = None\n):\n    assert torch.backends.cuda.matmul.allow_tf32 == True\n\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)\n    accum_iter = args.accum_iter\n\n    def save_model(epoch, fname, best_so_far):\n        misc.save_model(\n            accelerator=accelerator,\n            args=args,\n            model_without_ddp=model,\n            optimizer=optimizer,\n            loss_scaler=loss_scaler,\n            epoch=epoch,\n            fname=fname,\n            best_so_far=best_so_far,\n        )\n\n    if log_writer is not None:\n        printer.info(\"log_dir: {}\".format(log_writer.log_dir))\n\n    if hasattr(data_loader, \"dataset\") and hasattr(data_loader.dataset, \"set_epoch\"):\n        data_loader.dataset.set_epoch(epoch)\n    if (\n        hasattr(data_loader, \"batch_sampler\")\n        and hasattr(data_loader.batch_sampler, \"batch_sampler\")\n        and hasattr(data_loader.batch_sampler.batch_sampler, \"set_epoch\")\n    ):\n        data_loader.batch_sampler.batch_sampler.set_epoch(epoch)\n\n    optimizer.zero_grad()\n\n    for data_iter_step, batch in enumerate(\n        metric_logger.log_every(data_loader, args.print_freq, accelerator, header)\n    ):\n        with accelerator.accumulate(model):\n            epoch_f = epoch + data_iter_step / len(data_loader)\n            step = int(epoch_f * len(data_loader))\n            # we use a per iteration (instead of per epoch) lr scheduler\n            if data_iter_step % accum_iter == 0:\n                misc.adjust_learning_rate(optimizer, epoch_f, args)\n            if not args.long_context:\n                result = loss_of_one_batch(\n                    batch,\n                    model,\n                    criterion,\n                    accelerator,\n                    symmetrize_batch=False,\n                    use_amp=bool(args.amp),\n                    smpl_model=smpl_model\n                )\n            else:\n                NotImplementedError(\"Long context is not supported\")\n            has_msk = \"msk\" in result[\"pred\"][0]\n            loss, loss_details = result[\"loss\"]  # criterion returns two values\n            loss_value = float(loss)\n\n            if not math.isfinite(loss_value):\n                print(\n                    f\"Loss is {loss_value}, stopping training, loss details: {loss_details}\"\n                )\n                sys.exit(1)\n            if not result.get(\"already_backprop\", False):\n                loss_scaler(\n                    loss,\n                    optimizer,\n                    parameters=model.parameters(),\n                    update_grad=True,\n                    clip_grad=1.0,\n                )\n                optimizer.zero_grad()\n\n            is_metric = batch[0][\"is_metric\"]\n            curr_num_view = len(batch)\n\n            del loss\n            tb_vis_img = (data_iter_step + 1) % accum_iter == 0 and (\n                (step + 1) % (args.print_img_freq)\n            ) == 0\n            if not tb_vis_img:\n                del batch\n            else:\n                torch.cuda.empty_cache()\n\n            lr = optimizer.param_groups[0][\"lr\"]\n            metric_logger.update(epoch=epoch_f)\n            metric_logger.update(lr=lr)\n            metric_logger.update(step=step)\n\n            metric_logger.update(loss=loss_value, **loss_details)\n\n            if (data_iter_step + 1) % accum_iter == 0 and (\n                (data_iter_step + 1) % (accum_iter * args.print_freq)\n            ) == 0:\n                loss_value_reduce = accelerator.gather(\n                    torch.tensor(loss_value).to(accelerator.device)\n                ).mean()  # MUST BE EXECUTED BY ALL NODES\n\n                if log_writer is None:\n                    continue\n                \"\"\" We use epoch_1000x as the x-axis in tensorboard.\n                This calibrates different curves when batch size changes.\n                \"\"\"\n                epoch_1000x = int(epoch_f * 1000)\n                log_writer.add_scalar(\"train_loss\", loss_value_reduce, step)\n                log_writer.add_scalar(\"train_lr\", lr, step)\n                log_writer.add_scalar(\"train_iter\", epoch_1000x, step)\n                for name, val in loss_details.items():\n                    if isinstance(val, torch.Tensor):\n                        if val.ndim > 0:\n                            continue\n                    if isinstance(val, dict):\n                        continue\n                    log_writer.add_scalar(\"train_\" + name, val, step)\n\n            if tb_vis_img:\n                if log_writer is None:\n                    continue\n                with torch.no_grad():\n                    depths_self, gt_depths_self = get_render_results(\n                        batch, result[\"pred\"], self_view=True\n                    )\n                    depths_cross, gt_depths_cross = get_render_results(\n                        batch, result[\"pred\"], self_view=False\n                    )\n                    gt_msks, pr_msks, gt_hms, pr_hms, gt_smpls, pr_smpls = get_render_smpl(\n                        batch, result[\"pred\"], smpl_model, loss_details, has_msk=has_msk\n                    )\n                    for k in range(len(batch)):\n                        loss_details[f\"self_pred_depth_{k+1}\"] = depths_self[k].detach().cpu()\n                        loss_details[f\"self_gt_depth_{k+1}\"] = gt_depths_self[k].detach().cpu()\n                        loss_details[f\"pred_depth_{k+1}\"] = depths_cross[k].detach().cpu()\n                        loss_details[f\"gt_depth_{k+1}\"] = gt_depths_cross[k].detach().cpu()           \n                        loss_details[f\"pred_hm_{k+1}\"] = pr_hms[k].detach().cpu()\n                        loss_details[f\"gt_hm_{k+1}\"] = gt_hms[k].detach().cpu()\n                        loss_details[f\"pred_smpl_rend_{k+1}\"] = pr_smpls[k].detach().cpu()\n                        loss_details[f\"gt_smpl_rend_{k+1}\"] = gt_smpls[k].detach().cpu()\n                        if has_msk:\n                            loss_details[f\"pred_msk_{k+1}\"] = pr_msks[k].detach().cpu()\n                            loss_details[f\"gt_msk_{k+1}\"] = gt_msks[k].detach().cpu()\n\n                imgs_stacked_dict = get_vis_imgs_new(\n                    loss_details, \n                    args.num_imgs_vis, \n                    curr_num_view, \n                    is_metric=is_metric, \n                    has_msk=has_msk)\n                for name, imgs_stacked in imgs_stacked_dict.items():\n                    log_writer.add_images(\n                        \"train\" + \"/\" + name, imgs_stacked, step, dataformats=\"HWC\"\n                    )\n                del batch\n\n        if (\n            data_iter_step % int(args.save_freq * len(data_loader)) == 0\n            and data_iter_step != 0\n            and data_iter_step != len(data_loader) - 1\n        ):\n            print(\"saving at step\", data_iter_step)\n            save_model(epoch - 1, \"last\", float(\"inf\"))\n\n    # gather the stats from all processes\n    metric_logger.synchronize_between_processes(accelerator)\n    printer.info(\"Averaged stats: %s\", metric_logger)\n    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}\n\n\n@torch.no_grad()\ndef test_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    data_loader: Sized,\n    accelerator: Accelerator,\n    device: torch.device,\n    epoch: int,\n    args,\n    log_writer=None,\n    prefix=\"test\",\n    smpl_model: SMPLModel = None\n):\n\n    model.eval()\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.meters = defaultdict(lambda: misc.SmoothedValue(window_size=9**9))\n    header = \"Test Epoch: [{}]\".format(epoch)","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.test_one_epoch","uri":"program://Human3R/function/src.train.test_one_epoch#L577-L680","kind":"function","name":"test_one_epoch","path":"src/train.py","language":"python","start_line":577,"end_line":680,"context_start_line":557,"context_end_line":700,"code":"                    log_writer.add_images(\n                        \"train\" + \"/\" + name, imgs_stacked, step, dataformats=\"HWC\"\n                    )\n                del batch\n\n        if (\n            data_iter_step % int(args.save_freq * len(data_loader)) == 0\n            and data_iter_step != 0\n            and data_iter_step != len(data_loader) - 1\n        ):\n            print(\"saving at step\", data_iter_step)\n            save_model(epoch - 1, \"last\", float(\"inf\"))\n\n    # gather the stats from all processes\n    metric_logger.synchronize_between_processes(accelerator)\n    printer.info(\"Averaged stats: %s\", metric_logger)\n    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}\n\n\n@torch.no_grad()\ndef test_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    data_loader: Sized,\n    accelerator: Accelerator,\n    device: torch.device,\n    epoch: int,\n    args,\n    log_writer=None,\n    prefix=\"test\",\n    smpl_model: SMPLModel = None\n):\n\n    model.eval()\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.meters = defaultdict(lambda: misc.SmoothedValue(window_size=9**9))\n    header = \"Test Epoch: [{}]\".format(epoch)\n\n    if log_writer is not None:\n        printer.info(\"log_dir: {}\".format(log_writer.log_dir))\n\n    if hasattr(data_loader, \"dataset\") and hasattr(data_loader.dataset, \"set_epoch\"):\n        data_loader.dataset.set_epoch(0)\n    if (\n        hasattr(data_loader, \"batch_sampler\")\n        and hasattr(data_loader.batch_sampler, \"batch_sampler\")\n        and hasattr(data_loader.batch_sampler.batch_sampler, \"set_epoch\")\n    ):\n        data_loader.batch_sampler.batch_sampler.set_epoch(0)\n\n    for _, batch in enumerate(\n        metric_logger.log_every(data_loader, args.print_freq, accelerator, header)\n    ):\n        batch = todevice(batch, device)\n        result = loss_of_one_batch(\n            batch,\n            model,\n            criterion,\n            accelerator,\n            symmetrize_batch=False,\n            use_amp=bool(args.amp),\n            smpl_model=smpl_model\n        )\n\n        has_msk = \"msk\" in result[\"pred\"][0]\n        loss_value, loss_details = result[\"loss\"]  # criterion returns two values\n        metric_logger.update(loss=float(loss_value), **loss_details)\n\n    printer.info(\"Averaged stats: %s\", metric_logger)\n\n    aggs = [(\"avg\", \"global_avg\"), (\"med\", \"median\")]\n    results = {\n        f\"{k}_{tag}\": getattr(meter, attr)\n        for k, meter in metric_logger.meters.items()\n        for tag, attr in aggs\n    }\n\n    if log_writer is not None:\n        for name, val in results.items():\n            if isinstance(val, torch.Tensor):\n                if val.ndim > 0:\n                    continue\n            if isinstance(val, dict):\n                continue\n            log_writer.add_scalar(prefix + \"_\" + name, val, 1000 * epoch)\n\n        depths_self, gt_depths_self = get_render_results(\n            batch, result[\"pred\"], self_view=True\n        )\n        depths_cross, gt_depths_cross = get_render_results(\n            batch, result[\"pred\"], self_view=False\n        )\n        gt_msks, pr_msks, gt_hms, pr_hms, gt_smpls, pr_smpls = get_render_smpl(\n            batch, result[\"pred\"], smpl_model, loss_details, has_msk=has_msk\n        )\n        for k in range(len(batch)):\n            loss_details[f\"self_pred_depth_{k+1}\"] = depths_self[k].detach().cpu()\n            loss_details[f\"self_gt_depth_{k+1}\"] = gt_depths_self[k].detach().cpu()\n            loss_details[f\"pred_depth_{k+1}\"] = depths_cross[k].detach().cpu()\n            loss_details[f\"gt_depth_{k+1}\"] = gt_depths_cross[k].detach().cpu()\n            loss_details[f\"pred_hm_{k+1}\"] = pr_hms[k].detach().cpu()\n            loss_details[f\"gt_hm_{k+1}\"] = gt_hms[k].detach().cpu()\n            loss_details[f\"pred_smpl_rend_{k+1}\"] = pr_smpls[k].detach().cpu()\n            loss_details[f\"gt_smpl_rend_{k+1}\"] = gt_smpls[k].detach().cpu()\n            if has_msk:\n                loss_details[f\"pred_msk_{k+1}\"] = pr_msks[k].detach().cpu()\n                loss_details[f\"gt_msk_{k+1}\"] = gt_msks[k].detach().cpu()\n\n        imgs_stacked_dict = get_vis_imgs_new(\n            loss_details,\n            args.num_imgs_vis,\n            args.num_test_views,\n            is_metric=batch[0][\"is_metric\"],\n            has_msk=has_msk\n        )\n        for name, imgs_stacked in imgs_stacked_dict.items():\n            log_writer.add_images(\n                prefix + \"/\" + name, imgs_stacked, 1000 * epoch, dataformats=\"HWC\"\n            )\n\n    del loss_details, loss_value, batch\n    torch.cuda.empty_cache()\n\n    return results\n\n\ndef batch_append(original_list, new_list):\n    for sublist, new_item in zip(original_list, new_list):\n        sublist.append(new_item)\n    return original_list\n\n\ndef gen_mask_indicator(img_mask_list, ray_mask_list, num_views, h, w):\n    output = []\n    for img_mask, ray_mask in zip(img_mask_list, ray_mask_list):\n        out = torch.zeros((h, w * num_views, 3))\n        for i in range(num_views):\n            if img_mask[i] and not ray_mask[i]:\n                offset = 0\n            elif not img_mask[i] and ray_mask[i]:\n                offset = 1\n            else:\n                offset = 0.5\n            out[:, i * w : (i + 1) * w] += offset","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.batch_append","uri":"program://Human3R/function/src.train.batch_append#L683-L686","kind":"function","name":"batch_append","path":"src/train.py","language":"python","start_line":683,"end_line":686,"context_start_line":663,"context_end_line":706,"code":"                loss_details[f\"gt_msk_{k+1}\"] = gt_msks[k].detach().cpu()\n\n        imgs_stacked_dict = get_vis_imgs_new(\n            loss_details,\n            args.num_imgs_vis,\n            args.num_test_views,\n            is_metric=batch[0][\"is_metric\"],\n            has_msk=has_msk\n        )\n        for name, imgs_stacked in imgs_stacked_dict.items():\n            log_writer.add_images(\n                prefix + \"/\" + name, imgs_stacked, 1000 * epoch, dataformats=\"HWC\"\n            )\n\n    del loss_details, loss_value, batch\n    torch.cuda.empty_cache()\n\n    return results\n\n\ndef batch_append(original_list, new_list):\n    for sublist, new_item in zip(original_list, new_list):\n        sublist.append(new_item)\n    return original_list\n\n\ndef gen_mask_indicator(img_mask_list, ray_mask_list, num_views, h, w):\n    output = []\n    for img_mask, ray_mask in zip(img_mask_list, ray_mask_list):\n        out = torch.zeros((h, w * num_views, 3))\n        for i in range(num_views):\n            if img_mask[i] and not ray_mask[i]:\n                offset = 0\n            elif not img_mask[i] and ray_mask[i]:\n                offset = 1\n            else:\n                offset = 0.5\n            out[:, i * w : (i + 1) * w] += offset\n        output.append(out)\n    return output\n\n\ndef vis_and_cat(\n    gt_imgs,","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.gen_mask_indicator","uri":"program://Human3R/function/src.train.gen_mask_indicator#L689-L702","kind":"function","name":"gen_mask_indicator","path":"src/train.py","language":"python","start_line":689,"end_line":702,"context_start_line":669,"context_end_line":722,"code":"            is_metric=batch[0][\"is_metric\"],\n            has_msk=has_msk\n        )\n        for name, imgs_stacked in imgs_stacked_dict.items():\n            log_writer.add_images(\n                prefix + \"/\" + name, imgs_stacked, 1000 * epoch, dataformats=\"HWC\"\n            )\n\n    del loss_details, loss_value, batch\n    torch.cuda.empty_cache()\n\n    return results\n\n\ndef batch_append(original_list, new_list):\n    for sublist, new_item in zip(original_list, new_list):\n        sublist.append(new_item)\n    return original_list\n\n\ndef gen_mask_indicator(img_mask_list, ray_mask_list, num_views, h, w):\n    output = []\n    for img_mask, ray_mask in zip(img_mask_list, ray_mask_list):\n        out = torch.zeros((h, w * num_views, 3))\n        for i in range(num_views):\n            if img_mask[i] and not ray_mask[i]:\n                offset = 0\n            elif not img_mask[i] and ray_mask[i]:\n                offset = 1\n            else:\n                offset = 0.5\n            out[:, i * w : (i + 1) * w] += offset\n        output.append(out)\n    return output\n\n\ndef vis_and_cat(\n    gt_imgs,\n    pred_imgs,\n    gt_msks,\n    pred_msks,\n    gt_hms,\n    pred_hms,\n    gt_smpl_rends,\n    pred_smpl_rends,\n    cross_gt_depths,\n    cross_pred_depths,\n    self_gt_depths,\n    self_pred_depths,\n    cross_conf,\n    self_conf,\n    ray_indicator,\n    is_metric,\n    has_msk=False","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.vis_and_cat","uri":"program://Human3R/function/src.train.vis_and_cat#L705-L847","kind":"function","name":"vis_and_cat","path":"src/train.py","language":"python","start_line":705,"end_line":847,"context_start_line":685,"context_end_line":867,"code":"        sublist.append(new_item)\n    return original_list\n\n\ndef gen_mask_indicator(img_mask_list, ray_mask_list, num_views, h, w):\n    output = []\n    for img_mask, ray_mask in zip(img_mask_list, ray_mask_list):\n        out = torch.zeros((h, w * num_views, 3))\n        for i in range(num_views):\n            if img_mask[i] and not ray_mask[i]:\n                offset = 0\n            elif not img_mask[i] and ray_mask[i]:\n                offset = 1\n            else:\n                offset = 0.5\n            out[:, i * w : (i + 1) * w] += offset\n        output.append(out)\n    return output\n\n\ndef vis_and_cat(\n    gt_imgs,\n    pred_imgs,\n    gt_msks,\n    pred_msks,\n    gt_hms,\n    pred_hms,\n    gt_smpl_rends,\n    pred_smpl_rends,\n    cross_gt_depths,\n    cross_pred_depths,\n    self_gt_depths,\n    self_pred_depths,\n    cross_conf,\n    self_conf,\n    ray_indicator,\n    is_metric,\n    has_msk=False\n):\n    cross_depth_gt_min = torch.quantile(cross_gt_depths, 0.01).item()\n    cross_depth_gt_max = torch.quantile(cross_gt_depths, 0.99).item()\n    cross_depth_pred_min = torch.quantile(cross_pred_depths, 0.01).item()\n    cross_depth_pred_max = torch.quantile(cross_pred_depths, 0.99).item()\n    cross_depth_min = min(cross_depth_gt_min, cross_depth_pred_min)\n    cross_depth_max = max(cross_depth_gt_max, cross_depth_pred_max)\n\n    cross_gt_depths_vis = colorize(\n        cross_gt_depths,\n        range=(\n            (cross_depth_min, cross_depth_max)\n            if is_metric\n            else (cross_depth_gt_min, cross_depth_gt_max)\n        ),\n        append_cbar=True,\n    )\n    cross_pred_depths_vis = colorize(\n        cross_pred_depths,\n        range=(\n            (cross_depth_min, cross_depth_max)\n            if is_metric\n            else (cross_depth_pred_min, cross_depth_pred_max)\n        ),\n        append_cbar=True,\n    )\n\n    self_depth_gt_min = torch.quantile(self_gt_depths, 0.01).item()\n    self_depth_gt_max = torch.quantile(self_gt_depths, 0.99).item()\n    self_depth_pred_min = torch.quantile(self_pred_depths, 0.01).item()\n    self_depth_pred_max = torch.quantile(self_pred_depths, 0.99).item()\n    self_depth_min = min(self_depth_gt_min, self_depth_pred_min)\n    self_depth_max = max(self_depth_gt_max, self_depth_pred_max)\n\n    self_gt_depths_vis = colorize(\n        self_gt_depths,\n        range=(\n            (self_depth_min, self_depth_max)\n            if is_metric\n            else (self_depth_gt_min, self_depth_gt_max)\n        ),\n        append_cbar=True,\n    )\n    self_pred_depths_vis = colorize(\n        self_pred_depths,\n        range=(\n            (self_depth_min, self_depth_max)\n            if is_metric\n            else (self_depth_pred_min, self_depth_pred_max)\n        ),\n        append_cbar=True,\n    )\n    if len(cross_conf) > 0:\n        cross_conf_vis = colorize(cross_conf, append_cbar=True)\n    if len(self_conf) > 0:\n        self_conf_vis = colorize(self_conf, append_cbar=True)\n    gt_imgs_vis = torch.zeros_like(cross_gt_depths_vis)\n    gt_imgs_vis[: gt_imgs.shape[0], : gt_imgs.shape[1]] = gt_imgs\n    pred_imgs_vis = torch.zeros_like(cross_gt_depths_vis)\n    pred_imgs_vis[: pred_imgs.shape[0], : pred_imgs.shape[1]] = pred_imgs\n    if has_msk:\n        gt_msks_vis = torch.zeros_like(cross_gt_depths_vis)\n        gt_msks_vis[: gt_msks.shape[0], : gt_msks.shape[1]] = gt_msks\n        pred_msks_vis = torch.zeros_like(cross_gt_depths_vis)\n        pred_msks_vis[: pred_msks.shape[0], : pred_msks.shape[1]] = pred_msks\n    gt_hms_vis = torch.zeros_like(cross_gt_depths_vis)\n    gt_hms_vis[: gt_hms.shape[0], : gt_hms.shape[1]] = gt_hms\n    pred_hms_vis = torch.zeros_like(cross_gt_depths_vis)\n    pred_hms_vis[: pred_hms.shape[0], : pred_hms.shape[1]] = pred_hms\n    gt_smpl_rends_vis = torch.zeros_like(cross_gt_depths_vis)\n    gt_smpl_rends_vis[: gt_smpl_rends.shape[0], : gt_smpl_rends.shape[1]] = gt_smpl_rends\n    pred_smpl_rends_vis = torch.zeros_like(cross_gt_depths_vis)\n    pred_smpl_rends_vis[: pred_smpl_rends.shape[0], : pred_smpl_rends.shape[1]] = pred_smpl_rends\n    ray_indicator_vis = torch.cat(\n        [\n            ray_indicator,\n            torch.zeros(\n                ray_indicator.shape[0],\n                cross_pred_depths_vis.shape[1] - ray_indicator.shape[1],\n                3,\n            ),\n        ],\n        dim=1,\n    )\n    if has_msk:\n        out = torch.cat(\n            [\n                ray_indicator_vis,\n                gt_imgs_vis,\n                pred_imgs_vis,\n                gt_msks_vis,\n                pred_msks_vis,\n                gt_hms_vis,\n                pred_hms_vis,\n                gt_smpl_rends_vis,\n                pred_smpl_rends_vis,\n                self_gt_depths_vis,\n                self_pred_depths_vis,\n                self_conf_vis,\n                cross_gt_depths_vis,\n                cross_pred_depths_vis,\n                cross_conf_vis,\n            ],\n            dim=0,\n        )\n    else:\n        out = torch.cat(\n            [\n                ray_indicator_vis,\n                gt_imgs_vis,\n                pred_imgs_vis,\n                gt_hms_vis,\n                pred_hms_vis,\n                gt_smpl_rends_vis,\n                pred_smpl_rends_vis,\n                self_gt_depths_vis,\n                self_pred_depths_vis,\n                self_conf_vis,\n                cross_gt_depths_vis,\n                cross_pred_depths_vis,\n                cross_conf_vis,\n            ],\n            dim=0,\n        )\n    return out\n\n\ndef get_vis_imgs_new(loss_details, num_imgs_vis, num_views, is_metric, has_msk=False):\n    ret_dict = {}\n    gt_img_list = [[] for _ in range(num_imgs_vis)]\n    pred_img_list = [[] for _ in range(num_imgs_vis)]\n\n    cross_gt_depth_list = [[] for _ in range(num_imgs_vis)]\n    cross_pred_depth_list = [[] for _ in range(num_imgs_vis)]\n\n    self_gt_depth_list = [[] for _ in range(num_imgs_vis)]\n    self_pred_depth_list = [[] for _ in range(num_imgs_vis)]\n\n    gt_msk_list = [[] for _ in range(num_imgs_vis)]\n    pred_msk_list = [[] for _ in range(num_imgs_vis)]\n    gt_hm_list = [[] for _ in range(num_imgs_vis)]\n    pred_hm_list = [[] for _ in range(num_imgs_vis)]\n    gt_smpl_rend_list = [[] for _ in range(num_imgs_vis)]\n    pred_smpl_rend_list = [[] for _ in range(num_imgs_vis)]\n","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.get_vis_imgs_new","uri":"program://Human3R/function/src.train.get_vis_imgs_new#L850-L1033","kind":"function","name":"get_vis_imgs_new","path":"src/train.py","language":"python","start_line":850,"end_line":1033,"context_start_line":830,"context_end_line":1049,"code":"            [\n                ray_indicator_vis,\n                gt_imgs_vis,\n                pred_imgs_vis,\n                gt_hms_vis,\n                pred_hms_vis,\n                gt_smpl_rends_vis,\n                pred_smpl_rends_vis,\n                self_gt_depths_vis,\n                self_pred_depths_vis,\n                self_conf_vis,\n                cross_gt_depths_vis,\n                cross_pred_depths_vis,\n                cross_conf_vis,\n            ],\n            dim=0,\n        )\n    return out\n\n\ndef get_vis_imgs_new(loss_details, num_imgs_vis, num_views, is_metric, has_msk=False):\n    ret_dict = {}\n    gt_img_list = [[] for _ in range(num_imgs_vis)]\n    pred_img_list = [[] for _ in range(num_imgs_vis)]\n\n    cross_gt_depth_list = [[] for _ in range(num_imgs_vis)]\n    cross_pred_depth_list = [[] for _ in range(num_imgs_vis)]\n\n    self_gt_depth_list = [[] for _ in range(num_imgs_vis)]\n    self_pred_depth_list = [[] for _ in range(num_imgs_vis)]\n\n    gt_msk_list = [[] for _ in range(num_imgs_vis)]\n    pred_msk_list = [[] for _ in range(num_imgs_vis)]\n    gt_hm_list = [[] for _ in range(num_imgs_vis)]\n    pred_hm_list = [[] for _ in range(num_imgs_vis)]\n    gt_smpl_rend_list = [[] for _ in range(num_imgs_vis)]\n    pred_smpl_rend_list = [[] for _ in range(num_imgs_vis)]\n\n    cross_view_conf_list = [[] for _ in range(num_imgs_vis)]\n    self_view_conf_list = [[] for _ in range(num_imgs_vis)]\n    cross_view_conf_exits = False\n    self_view_conf_exits = False\n\n    img_mask_list = [[] for _ in range(num_imgs_vis)]\n    ray_mask_list = [[] for _ in range(num_imgs_vis)]\n\n    if num_views > 30:\n        stride = 5\n    elif num_views > 20:\n        stride = 3\n    elif num_views > 10:\n        stride = 2\n    else:\n        stride = 1\n    for i in range(0, num_views, stride):\n        gt_imgs = 0.5 * (loss_details[f\"gt_img{i+1}\"] + 1)[:num_imgs_vis].detach().cpu()\n        width = gt_imgs.shape[2]\n        pred_imgs = (\n            0.5 * (loss_details[f\"pred_rgb_{i+1}\"] + 1)[:num_imgs_vis].detach().cpu()\n        )\n        gt_img_list = batch_append(gt_img_list, gt_imgs.unbind(dim=0))\n        pred_img_list = batch_append(pred_img_list, pred_imgs.unbind(dim=0))\n\n        cross_pred_depths = (\n            loss_details[f\"pred_depth_{i+1}\"][:num_imgs_vis].detach().cpu()\n        )\n        cross_gt_depths = (\n            loss_details[f\"gt_depth_{i+1}\"]\n            .to(gt_imgs.device)[:num_imgs_vis]\n            .detach()\n            .cpu()\n        )\n        cross_pred_depth_list = batch_append(\n            cross_pred_depth_list, cross_pred_depths.unbind(dim=0)\n        )\n        cross_gt_depth_list = batch_append(\n            cross_gt_depth_list, cross_gt_depths.unbind(dim=0)\n        )\n\n        self_gt_depths = (\n            loss_details[f\"self_gt_depth_{i+1}\"][:num_imgs_vis].detach().cpu()\n        )\n        self_pred_depths = (\n            loss_details[f\"self_pred_depth_{i+1}\"][:num_imgs_vis].detach().cpu()\n        )\n        self_gt_depth_list = batch_append(\n            self_gt_depth_list, self_gt_depths.unbind(dim=0)\n        )\n        self_pred_depth_list = batch_append(\n            self_pred_depth_list, self_pred_depths.unbind(dim=0)\n        )\n\n        if has_msk:\n            gt_msks = (\n                loss_details[f\"gt_msk_{i+1}\"][:num_imgs_vis].detach().cpu()\n            )\n            pred_msks = (\n                loss_details[f\"pred_msk_{i+1}\"][:num_imgs_vis].detach().cpu()\n            )\n        gt_hms = (\n            loss_details[f\"gt_hm_{i+1}\"][:num_imgs_vis].detach().cpu()\n        )\n        pred_hms = (\n            loss_details[f\"pred_hm_{i+1}\"][:num_imgs_vis].detach().cpu()\n        )\n        gt_smpl_rends = (\n            loss_details[f\"gt_smpl_rend_{i+1}\"][:num_imgs_vis].detach().cpu()\n        )\n        pred_smpl_rends = (\n            loss_details[f\"pred_smpl_rend_{i+1}\"][:num_imgs_vis].detach().cpu()\n        )\n        if has_msk:\n            gt_msk_list = batch_append(gt_msk_list, gt_msks.unbind(dim=0))\n            pred_msk_list = batch_append(pred_msk_list, pred_msks.unbind(dim=0))\n        gt_hm_list = batch_append(gt_hm_list, gt_hms.unbind(dim=0))\n        pred_hm_list = batch_append(pred_hm_list, pred_hms.unbind(dim=0))\n        gt_smpl_rend_list = batch_append(\n            gt_smpl_rend_list, gt_smpl_rends.unbind(dim=0))\n        pred_smpl_rend_list = batch_append(\n            pred_smpl_rend_list, pred_smpl_rends.unbind(dim=0))\n\n        if f\"conf_{i+1}\" in loss_details:\n            cross_view_conf = loss_details[f\"conf_{i+1}\"][:num_imgs_vis].detach().cpu()\n            cross_view_conf_list = batch_append(\n                cross_view_conf_list, cross_view_conf.unbind(dim=0)\n            )\n            cross_view_conf_exits = True\n\n        if f\"self_conf_{i+1}\" in loss_details:\n            self_view_conf = (\n                loss_details[f\"self_conf_{i+1}\"][:num_imgs_vis].detach().cpu()\n            )\n            self_view_conf_list = batch_append(\n                self_view_conf_list, self_view_conf.unbind(dim=0)\n            )\n            self_view_conf_exits = True\n\n        img_mask_list = batch_append(\n            img_mask_list,\n            loss_details[f\"img_mask_{i+1}\"][:num_imgs_vis].detach().cpu().unbind(dim=0),\n        )\n        ray_mask_list = batch_append(\n            ray_mask_list,\n            loss_details[f\"ray_mask_{i+1}\"][:num_imgs_vis].detach().cpu().unbind(dim=0),\n        )\n\n    # each element in the list is [H, num_views * W, (3)], the size of the list is num_imgs_vis\n    gt_img_list = [torch.cat(sublist, dim=1) for sublist in gt_img_list]\n    pred_img_list = [torch.cat(sublist, dim=1) for sublist in pred_img_list]\n    cross_pred_depth_list = [\n        torch.cat(sublist, dim=1) for sublist in cross_pred_depth_list\n    ]\n    cross_gt_depth_list = [torch.cat(sublist, dim=1) for sublist in cross_gt_depth_list]\n    self_gt_depth_list = [torch.cat(sublist, dim=1) for sublist in self_gt_depth_list]\n    self_pred_depth_list = [\n        torch.cat(sublist, dim=1) for sublist in self_pred_depth_list\n    ]\n    if has_msk:\n        gt_msk_list = [torch.cat(sublist, dim=1) for sublist in gt_msk_list]\n        pred_msk_list = [torch.cat(sublist, dim=1) for sublist in pred_msk_list]\n    gt_hm_list = [torch.cat(sublist, dim=1) for sublist in gt_hm_list]\n    pred_hm_list = [torch.cat(sublist, dim=1) for sublist in pred_hm_list]\n    gt_smpl_rend_list = [torch.cat(sublist, dim=1) for sublist in gt_smpl_rend_list]\n    pred_smpl_rend_list = [torch.cat(sublist, dim=1) for sublist in pred_smpl_rend_list]\n    cross_view_conf_list = (\n        [torch.cat(sublist, dim=1) for sublist in cross_view_conf_list]\n        if cross_view_conf_exits\n        else []\n    )\n    self_view_conf_list = (\n        [torch.cat(sublist, dim=1) for sublist in self_view_conf_list]\n        if self_view_conf_exits\n        else []\n    )\n    # each elment in the list is [num_views,], the size of the list is num_imgs_vis\n    img_mask_list = [torch.stack(sublist, dim=0) for sublist in img_mask_list]\n    ray_mask_list = [torch.stack(sublist, dim=0) for sublist in ray_mask_list]\n\n    ray_indicator = gen_mask_indicator(\n        img_mask_list, ray_mask_list, len(img_mask_list[0]), 30, width\n    )\n\n    for i in range(num_imgs_vis):\n        out = vis_and_cat(\n            gt_img_list[i],\n            pred_img_list[i],\n            gt_msk_list[i],\n            pred_msk_list[i],\n            gt_hm_list[i],\n            pred_hm_list[i],\n            gt_smpl_rend_list[i],\n            pred_smpl_rend_list[i],\n            cross_gt_depth_list[i],\n            cross_pred_depth_list[i],\n            self_gt_depth_list[i],\n            self_pred_depth_list[i],\n            cross_view_conf_list[i],\n            self_view_conf_list[i],\n            ray_indicator[i],\n            is_metric[i],\n            has_msk=has_msk\n        )\n        ret_dict[f\"imgs_{i}\"] = out\n    return ret_dict\n\n\n@hydra.main(\n    version_base=None,\n    config_path=str(os.path.dirname(os.path.abspath(__file__))) + \"/../config\",\n    config_name=\"train.yaml\",\n)\ndef run(cfg: OmegaConf):\n    OmegaConf.resolve(cfg)\n    logdir = pathlib.Path(cfg.logdir)\n    logdir.mkdir(parents=True, exist_ok=True)\n    train(cfg)\n\n\nif __name__ == \"__main__\":\n    run()","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.run","uri":"program://Human3R/function/src.train.run#L1041-L1045","kind":"function","name":"run","path":"src/train.py","language":"python","start_line":1041,"end_line":1045,"context_start_line":1021,"context_end_line":1049,"code":"            pred_smpl_rend_list[i],\n            cross_gt_depth_list[i],\n            cross_pred_depth_list[i],\n            self_gt_depth_list[i],\n            self_pred_depth_list[i],\n            cross_view_conf_list[i],\n            self_view_conf_list[i],\n            ray_indicator[i],\n            is_metric[i],\n            has_msk=has_msk\n        )\n        ret_dict[f\"imgs_{i}\"] = out\n    return ret_dict\n\n\n@hydra.main(\n    version_base=None,\n    config_path=str(os.path.dirname(os.path.abspath(__file__))) + \"/../config\",\n    config_name=\"train.yaml\",\n)\ndef run(cfg: OmegaConf):\n    OmegaConf.resolve(cfg)\n    logdir = pathlib.Path(cfg.logdir)\n    logdir.mkdir(parents=True, exist_ok=True)\n    train(cfg)\n\n\nif __name__ == \"__main__\":\n    run()","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.print","uri":"program://Human3R/function/src.train.print#L73-L79","kind":"function","name":"print","path":"src/train.py","language":"python","start_line":73,"end_line":79,"context_start_line":53,"context_end_line":99,"code":"import builtins\nimport shutil\n\nfrom accelerate import Accelerator\nfrom accelerate import DistributedDataParallelKwargs, InitProcessGroupKwargs\nfrom accelerate.logging import get_logger\nfrom datetime import timedelta\nimport torch.multiprocessing\n\ntorch.multiprocessing.set_sharing_strategy(\"file_system\")\n\nprinter = get_logger(__name__, log_level=\"DEBUG\")\n\n\ndef setup_for_distributed(accelerator: Accelerator):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (accelerator.num_processes > 8)\n        if accelerator.is_main_process or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef save_current_code(outdir):\n    now = datetime.datetime.now()  # current date and time\n    date_time = now.strftime(\"%m_%d-%H:%M:%S\")\n    src_dir = \".\"\n    dst_dir = os.path.join(outdir, \"code\", \"{}\".format(date_time))\n    shutil.copytree(\n        src_dir,\n        dst_dir,\n        ignore=shutil.ignore_patterns(\n            \".vscode*\",\n            \"assets*\",\n            \"example*\",\n            \"checkpoints*\",\n            \"OLD*\",\n            \"logs*\",\n            \"out*\",","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.write_log_stats","uri":"program://Human3R/function/src.train.write_log_stats#L250-L268","kind":"function","name":"write_log_stats","path":"src/train.py","language":"python","start_line":250,"end_line":268,"context_start_line":230,"context_end_line":288,"code":"            printer.info(f\"Loading Multi-HMR pretrained: {args.pretrained_mhmr}\")\n            ckpt_mhmr = torch.load(args.pretrained_mhmr, map_location=device)\n            merge_state_dict.update(strip_module_mhmr(ckpt_mhmr[\"model_state_dict\"]))\n            del ckpt_mhmr  # in case it occupies memory\n\n        printer.info(\n            model.load_state_dict(merge_state_dict, strict=False)\n        )\n\n    # # following timm: set wd as 0 for bias and norm layers\n    param_groups = misc.get_parameter_groups(model, args.weight_decay)\n    optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))\n    # print(optimizer)\n    loss_scaler = NativeScaler(accelerator=accelerator)\n\n    accelerator.even_batches = False\n    optimizer, model, data_loader_train = accelerator.prepare(\n        optimizer, model, data_loader_train\n    )\n\n    def write_log_stats(epoch, train_stats, test_stats):\n        if accelerator.is_main_process:\n            if log_writer is not None:\n                log_writer.flush()\n\n            log_stats = dict(\n                epoch=epoch, **{f\"train_{k}\": v for k, v in train_stats.items()}\n            )\n            for test_name in data_loader_test:\n                if test_name not in test_stats:\n                    continue\n                log_stats.update(\n                    {test_name + \"_\" + k: v for k, v in test_stats[test_name].items()}\n                )\n\n            with open(\n                os.path.join(args.output_dir, \"log.txt\"), mode=\"a\", encoding=\"utf-8\"\n            ) as f:\n                f.write(json.dumps(log_stats) + \"\\n\")\n\n    def save_model(epoch, fname, best_so_far):\n        misc.save_model(\n            accelerator=accelerator,\n            args=args,\n            model_without_ddp=model,\n            optimizer=optimizer,\n            loss_scaler=loss_scaler,\n            epoch=epoch,\n            fname=fname,\n            best_so_far=best_so_far,\n        )\n\n    best_so_far = misc.load_model(\n        args=args, model_without_ddp=model, optimizer=optimizer, loss_scaler=loss_scaler\n    )\n    if best_so_far is None:\n        best_so_far = float(\"inf\")\n    log_writer = (\n        SummaryWriter(log_dir=args.output_dir) if accelerator.is_main_process else None","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.train.save_model","uri":"program://Human3R/function/src.train.save_model#L415-L425","kind":"function","name":"save_model","path":"src/train.py","language":"python","start_line":415,"end_line":425,"context_start_line":395,"context_end_line":445,"code":"def train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    data_loader: Sized,\n    optimizer: torch.optim.Optimizer,\n    accelerator: Accelerator,\n    epoch: int,\n    loss_scaler,\n    args,\n    log_writer=None,\n    smpl_model: SMPLModel = None\n):\n    assert torch.backends.cuda.matmul.allow_tf32 == True\n\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)\n    accum_iter = args.accum_iter\n\n    def save_model(epoch, fname, best_so_far):\n        misc.save_model(\n            accelerator=accelerator,\n            args=args,\n            model_without_ddp=model,\n            optimizer=optimizer,\n            loss_scaler=loss_scaler,\n            epoch=epoch,\n            fname=fname,\n            best_so_far=best_so_far,\n        )\n\n    if log_writer is not None:\n        printer.info(\"log_dir: {}\".format(log_writer.log_dir))\n\n    if hasattr(data_loader, \"dataset\") and hasattr(data_loader.dataset, \"set_epoch\"):\n        data_loader.dataset.set_epoch(epoch)\n    if (\n        hasattr(data_loader, \"batch_sampler\")\n        and hasattr(data_loader.batch_sampler, \"batch_sampler\")\n        and hasattr(data_loader.batch_sampler.batch_sampler, \"set_epoch\")\n    ):\n        data_loader.batch_sampler.batch_sampler.set_epoch(epoch)\n\n    optimizer.zero_grad()\n\n    for data_iter_step, batch in enumerate(\n        metric_logger.log_every(data_loader, args.print_freq, accelerator, header)\n    ):\n        with accelerator.accumulate(model):\n            epoch_f = epoch + data_iter_step / len(data_loader)","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.camera_embed","uri":"program://Human3R/module/src.mhmr.blocks.camera_embed#L1-L58","kind":"module","name":"src.mhmr.blocks.camera_embed","path":"src/mhmr/blocks/camera_embed.py","language":"python","start_line":1,"end_line":58,"context_start_line":1,"context_end_line":58,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\nimport numpy as np\n\nclass FourierPositionEncoding(nn.Module):\n    def __init__(self, n, num_bands, max_resolution):\n        \"\"\"\n        Module that generate Fourier encoding - no learning involved\n        \"\"\"\n        super().__init__()\n\n        self.num_bands = num_bands\n        self.max_resolution = [max_resolution] * n\n    \n    @property\n    def channels(self):\n        \"\"\"\n        Return the output dimension\n        \"\"\"        \n        num_dims = len(self.max_resolution)\n        encoding_size = self.num_bands * num_dims\n        encoding_size *= 2 # sin-cos\n        encoding_size += num_dims # concat\n\n        return encoding_size\n    \n    def forward(self, pos):\n        \"\"\"\n        Forward pass that take rays as input and generate Fourier positional encodings\n        \"\"\"\n        fourier_pos_enc = _generate_fourier_features(pos, num_bands=self.num_bands, max_resolution=self.max_resolution)\n        return fourier_pos_enc\n    \n\ndef _generate_fourier_features(pos, num_bands, max_resolution):\n    \"\"\"Generate fourier features from a given set of positions and frequencies\"\"\"\n    b, n = pos.shape[:2]\n    device = pos.device\n\n    # Linear frequency sampling\n    min_freq = 1.0\n    freq_bands = torch.stack([torch.linspace(start=min_freq, end=res / 2, steps=num_bands, device=device) for res in max_resolution], dim=0)\n\n    # Stacking\n    per_pos_features = torch.stack([pos[i, :, :][:, :, None] * freq_bands[None, :, :] for i in range(b)], 0)\n    per_pos_features = per_pos_features.reshape(b, n, -1)\n\n    # Sin-Cos\n    per_pos_features = torch.cat([torch.sin(np.pi * per_pos_features), torch.cos(np.pi * per_pos_features)], dim=-1)\n\n    # Concat with initial pos\n    per_pos_features = torch.cat([pos, per_pos_features], dim=-1)\n\n    return per_pos_features","source_hash":"756359c5c2186151b8e64c7ff20d4581419015dee1f45e8fc689f7c5ec0a7b7b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.camera_embed.FourierPositionEncoding","uri":"program://Human3R/class/src.mhmr.blocks.camera_embed.FourierPositionEncoding#L9-L36","kind":"class","name":"FourierPositionEncoding","path":"src/mhmr/blocks/camera_embed.py","language":"python","start_line":9,"end_line":36,"context_start_line":1,"context_end_line":56,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\nimport numpy as np\n\nclass FourierPositionEncoding(nn.Module):\n    def __init__(self, n, num_bands, max_resolution):\n        \"\"\"\n        Module that generate Fourier encoding - no learning involved\n        \"\"\"\n        super().__init__()\n\n        self.num_bands = num_bands\n        self.max_resolution = [max_resolution] * n\n    \n    @property\n    def channels(self):\n        \"\"\"\n        Return the output dimension\n        \"\"\"        \n        num_dims = len(self.max_resolution)\n        encoding_size = self.num_bands * num_dims\n        encoding_size *= 2 # sin-cos\n        encoding_size += num_dims # concat\n\n        return encoding_size\n    \n    def forward(self, pos):\n        \"\"\"\n        Forward pass that take rays as input and generate Fourier positional encodings\n        \"\"\"\n        fourier_pos_enc = _generate_fourier_features(pos, num_bands=self.num_bands, max_resolution=self.max_resolution)\n        return fourier_pos_enc\n    \n\ndef _generate_fourier_features(pos, num_bands, max_resolution):\n    \"\"\"Generate fourier features from a given set of positions and frequencies\"\"\"\n    b, n = pos.shape[:2]\n    device = pos.device\n\n    # Linear frequency sampling\n    min_freq = 1.0\n    freq_bands = torch.stack([torch.linspace(start=min_freq, end=res / 2, steps=num_bands, device=device) for res in max_resolution], dim=0)\n\n    # Stacking\n    per_pos_features = torch.stack([pos[i, :, :][:, :, None] * freq_bands[None, :, :] for i in range(b)], 0)\n    per_pos_features = per_pos_features.reshape(b, n, -1)\n\n    # Sin-Cos\n    per_pos_features = torch.cat([torch.sin(np.pi * per_pos_features), torch.cos(np.pi * per_pos_features)], dim=-1)\n\n    # Concat with initial pos\n    per_pos_features = torch.cat([pos, per_pos_features], dim=-1)","source_hash":"756359c5c2186151b8e64c7ff20d4581419015dee1f45e8fc689f7c5ec0a7b7b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.camera_embed._generate_fourier_features","uri":"program://Human3R/function/src.mhmr.blocks.camera_embed._generate_fourier_features#L39-L58","kind":"function","name":"_generate_fourier_features","path":"src/mhmr/blocks/camera_embed.py","language":"python","start_line":39,"end_line":58,"context_start_line":19,"context_end_line":58,"code":"    @property\n    def channels(self):\n        \"\"\"\n        Return the output dimension\n        \"\"\"        \n        num_dims = len(self.max_resolution)\n        encoding_size = self.num_bands * num_dims\n        encoding_size *= 2 # sin-cos\n        encoding_size += num_dims # concat\n\n        return encoding_size\n    \n    def forward(self, pos):\n        \"\"\"\n        Forward pass that take rays as input and generate Fourier positional encodings\n        \"\"\"\n        fourier_pos_enc = _generate_fourier_features(pos, num_bands=self.num_bands, max_resolution=self.max_resolution)\n        return fourier_pos_enc\n    \n\ndef _generate_fourier_features(pos, num_bands, max_resolution):\n    \"\"\"Generate fourier features from a given set of positions and frequencies\"\"\"\n    b, n = pos.shape[:2]\n    device = pos.device\n\n    # Linear frequency sampling\n    min_freq = 1.0\n    freq_bands = torch.stack([torch.linspace(start=min_freq, end=res / 2, steps=num_bands, device=device) for res in max_resolution], dim=0)\n\n    # Stacking\n    per_pos_features = torch.stack([pos[i, :, :][:, :, None] * freq_bands[None, :, :] for i in range(b)], 0)\n    per_pos_features = per_pos_features.reshape(b, n, -1)\n\n    # Sin-Cos\n    per_pos_features = torch.cat([torch.sin(np.pi * per_pos_features), torch.cos(np.pi * per_pos_features)], dim=-1)\n\n    # Concat with initial pos\n    per_pos_features = torch.cat([pos, per_pos_features], dim=-1)\n\n    return per_pos_features","source_hash":"756359c5c2186151b8e64c7ff20d4581419015dee1f45e8fc689f7c5ec0a7b7b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.camera_embed.__init__","uri":"program://Human3R/function/src.mhmr.blocks.camera_embed.__init__#L10-L17","kind":"function","name":"__init__","path":"src/mhmr/blocks/camera_embed.py","language":"python","start_line":10,"end_line":17,"context_start_line":1,"context_end_line":37,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\nimport numpy as np\n\nclass FourierPositionEncoding(nn.Module):\n    def __init__(self, n, num_bands, max_resolution):\n        \"\"\"\n        Module that generate Fourier encoding - no learning involved\n        \"\"\"\n        super().__init__()\n\n        self.num_bands = num_bands\n        self.max_resolution = [max_resolution] * n\n    \n    @property\n    def channels(self):\n        \"\"\"\n        Return the output dimension\n        \"\"\"        \n        num_dims = len(self.max_resolution)\n        encoding_size = self.num_bands * num_dims\n        encoding_size *= 2 # sin-cos\n        encoding_size += num_dims # concat\n\n        return encoding_size\n    \n    def forward(self, pos):\n        \"\"\"\n        Forward pass that take rays as input and generate Fourier positional encodings\n        \"\"\"\n        fourier_pos_enc = _generate_fourier_features(pos, num_bands=self.num_bands, max_resolution=self.max_resolution)\n        return fourier_pos_enc\n    ","source_hash":"756359c5c2186151b8e64c7ff20d4581419015dee1f45e8fc689f7c5ec0a7b7b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.camera_embed.channels","uri":"program://Human3R/function/src.mhmr.blocks.camera_embed.channels#L20-L29","kind":"function","name":"channels","path":"src/mhmr/blocks/camera_embed.py","language":"python","start_line":20,"end_line":29,"context_start_line":1,"context_end_line":49,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\nimport numpy as np\n\nclass FourierPositionEncoding(nn.Module):\n    def __init__(self, n, num_bands, max_resolution):\n        \"\"\"\n        Module that generate Fourier encoding - no learning involved\n        \"\"\"\n        super().__init__()\n\n        self.num_bands = num_bands\n        self.max_resolution = [max_resolution] * n\n    \n    @property\n    def channels(self):\n        \"\"\"\n        Return the output dimension\n        \"\"\"        \n        num_dims = len(self.max_resolution)\n        encoding_size = self.num_bands * num_dims\n        encoding_size *= 2 # sin-cos\n        encoding_size += num_dims # concat\n\n        return encoding_size\n    \n    def forward(self, pos):\n        \"\"\"\n        Forward pass that take rays as input and generate Fourier positional encodings\n        \"\"\"\n        fourier_pos_enc = _generate_fourier_features(pos, num_bands=self.num_bands, max_resolution=self.max_resolution)\n        return fourier_pos_enc\n    \n\ndef _generate_fourier_features(pos, num_bands, max_resolution):\n    \"\"\"Generate fourier features from a given set of positions and frequencies\"\"\"\n    b, n = pos.shape[:2]\n    device = pos.device\n\n    # Linear frequency sampling\n    min_freq = 1.0\n    freq_bands = torch.stack([torch.linspace(start=min_freq, end=res / 2, steps=num_bands, device=device) for res in max_resolution], dim=0)\n\n    # Stacking\n    per_pos_features = torch.stack([pos[i, :, :][:, :, None] * freq_bands[None, :, :] for i in range(b)], 0)","source_hash":"756359c5c2186151b8e64c7ff20d4581419015dee1f45e8fc689f7c5ec0a7b7b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.camera_embed.forward","uri":"program://Human3R/function/src.mhmr.blocks.camera_embed.forward#L31-L36","kind":"function","name":"forward","path":"src/mhmr/blocks/camera_embed.py","language":"python","start_line":31,"end_line":36,"context_start_line":11,"context_end_line":56,"code":"        \"\"\"\n        Module that generate Fourier encoding - no learning involved\n        \"\"\"\n        super().__init__()\n\n        self.num_bands = num_bands\n        self.max_resolution = [max_resolution] * n\n    \n    @property\n    def channels(self):\n        \"\"\"\n        Return the output dimension\n        \"\"\"        \n        num_dims = len(self.max_resolution)\n        encoding_size = self.num_bands * num_dims\n        encoding_size *= 2 # sin-cos\n        encoding_size += num_dims # concat\n\n        return encoding_size\n    \n    def forward(self, pos):\n        \"\"\"\n        Forward pass that take rays as input and generate Fourier positional encodings\n        \"\"\"\n        fourier_pos_enc = _generate_fourier_features(pos, num_bands=self.num_bands, max_resolution=self.max_resolution)\n        return fourier_pos_enc\n    \n\ndef _generate_fourier_features(pos, num_bands, max_resolution):\n    \"\"\"Generate fourier features from a given set of positions and frequencies\"\"\"\n    b, n = pos.shape[:2]\n    device = pos.device\n\n    # Linear frequency sampling\n    min_freq = 1.0\n    freq_bands = torch.stack([torch.linspace(start=min_freq, end=res / 2, steps=num_bands, device=device) for res in max_resolution], dim=0)\n\n    # Stacking\n    per_pos_features = torch.stack([pos[i, :, :][:, :, None] * freq_bands[None, :, :] for i in range(b)], 0)\n    per_pos_features = per_pos_features.reshape(b, n, -1)\n\n    # Sin-Cos\n    per_pos_features = torch.cat([torch.sin(np.pi * per_pos_features), torch.cos(np.pi * per_pos_features)], dim=-1)\n\n    # Concat with initial pos\n    per_pos_features = torch.cat([pos, per_pos_features], dim=-1)","source_hash":"756359c5c2186151b8e64c7ff20d4581419015dee1f45e8fc689f7c5ec0a7b7b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer","uri":"program://Human3R/module/src.mhmr.blocks.cross_attn_transformer#L1-L359","kind":"module","name":"src.mhmr.blocks.cross_attn_transformer","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":1,"end_line":359,"context_start_line":1,"context_end_line":359,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nfrom typing import Callable, Optional\nimport torch\nfrom torch import nn\nfrom inspect import isfunction\nfrom einops import rearrange\n\nclass AdaptiveLayerNorm1D(torch.nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/t_cond_mlp.py#L7\n    \"\"\"\n    def __init__(self, data_dim: int, norm_cond_dim: int):\n        super().__init__()\n        if data_dim <= 0:\n            raise ValueError(f\"data_dim must be positive, but got {data_dim}\")\n        if norm_cond_dim <= 0:\n            raise ValueError(f\"norm_cond_dim must be positive, but got {norm_cond_dim}\")\n        self.norm = torch.nn.LayerNorm(\n            data_dim\n        )  # TODO: Check if elementwise_affine=True is correct\n        self.linear = torch.nn.Linear(norm_cond_dim, 2 * data_dim)\n        torch.nn.init.zeros_(self.linear.weight)\n        torch.nn.init.zeros_(self.linear.bias)\n\n    def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor:\n        # x: (batch, ..., data_dim)\n        # t: (batch, norm_cond_dim)\n        # return: (batch, data_dim)\n        x = self.norm(x)\n        alpha, beta = self.linear(t).chunk(2, dim=-1)\n\n        # Add singleton dimensions to alpha and beta\n        if x.dim() > 2:\n            alpha = alpha.view(alpha.shape[0], *([1] * (x.dim() - 2)), alpha.shape[1])\n            beta = beta.view(beta.shape[0], *([1] * (x.dim() - 2)), beta.shape[1])\n\n        return x * (1 + alpha) + beta\n\n\ndef normalization_layer(norm: Optional[str], dim: int, norm_cond_dim: int = -1):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/t_cond_mlp.py#L48\n    \"\"\"\n    if norm == \"batch\":\n        return torch.nn.BatchNorm1d(dim)\n    elif norm == \"layer\":\n        return torch.nn.LayerNorm(dim)\n    elif norm == \"ada\":\n        assert norm_cond_dim > 0, f\"norm_cond_dim must be positive, got {norm_cond_dim}\"\n        return AdaptiveLayerNorm1D(dim, norm_cond_dim)\n    elif norm is None:\n        return torch.nn.Identity()\n    else:\n        raise ValueError(f\"Unknown norm: {norm}\")\n\n\ndef exists(val):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L17\"\n    return val is not None\n\n\ndef default(val, d):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L21\"\n    if exists(val):\n        return val\n    return d() if isfunction(d) else d\n\n\nclass PreNorm(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L27\n    \"\"\"\n    def __init__(self, dim: int, fn: Callable, norm: str = \"layer\", norm_cond_dim: int = -1):\n        super().__init__()\n        self.norm = normalization_layer(norm, dim, norm_cond_dim)\n        self.fn = fn\n\n    def forward(self, x: torch.Tensor, *args, **kwargs):\n        if isinstance(self.norm, AdaptiveLayerNorm1D):\n            return self.fn(self.norm(x, *args), **kwargs)\n        else:\n            return self.fn(self.norm(x), **kwargs)\n\n\nclass FeedForward(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L40\n    \"\"\"\n    def __init__(self, dim, hidden_dim, dropout=0.0):\n        super().__init__()\n        self.net = nn.Sequential(\n            nn.Linear(dim, hidden_dim),\n            nn.GELU(),\n            nn.Dropout(dropout),\n            nn.Linear(hidden_dim, dim),\n            nn.Dropout(dropout),\n        )\n\n    def forward(self, x):\n        return self.net(x)\n\n\nclass Attention(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L55\n    \"\"\"\n    def __init__(self, dim, heads=8, dim_head=64, dropout=0.0):\n        super().__init__()\n        inner_dim = dim_head * heads\n        project_out = not (heads == 1 and dim_head == dim)\n\n        self.heads = heads\n        self.scale = dim_head**-0.5\n\n        self.attend = nn.Softmax(dim=-1)\n        self.dropout = nn.Dropout(dropout)\n\n        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias=False)\n\n        self.to_out = (\n            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))\n            if project_out\n            else nn.Identity()\n        )\n\n    def forward(self, x, mask=None):\n\n        qkv = self.to_qkv(x).chunk(3, dim=-1)\n        # n --> the num query dimension\n\n        # TODO reshape b into b2 n and mask.\n        q, k, v = map(lambda t: rearrange(t, \"b n (h d) -> b h n d\", h=self.heads), qkv)\n\n        if mask is not None:\n            q, k, v = [x * mask[:, None, :, None] for x in [q, k, v]]\n        \n        # q, k, v: [13:51:03.400365] torch.Size([22, 1, 256])\n        #q, k ,vk after reshape: torch.Size([16, 8, 1, 32])\n        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale\n\n        if mask is not None:\n            dots = dots - (1 - mask)[:, None, None, :] * 10e10\n\n        attn = self.attend(dots)\n\n        if mask is not None: # Just for good measure; this is probably overkill\n            attn = attn * mask[:, None, None, :]\n\n        attn = self.dropout(attn)\n\n        out = torch.matmul(attn, v)\n\n        # out shape :torch.Size([16, 8, 1, 32])\n\n        out = rearrange(out, \"b h n d -> b n (h d)\")\n        return self.to_out(out)\n\n\nclass CrossAttention(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L89\"\n    def __init__(self, dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):\n        super().__init__()\n        inner_dim = dim_head * heads\n        project_out = not (heads == 1 and dim_head == dim)\n\n        self.heads = heads\n        self.scale = dim_head**-0.5\n\n        self.attend = nn.Softmax(dim=-1)\n        self.dropout = nn.Dropout(dropout)\n\n        context_dim = default(context_dim, dim)\n        self.to_kv = nn.Linear(context_dim, inner_dim * 2, bias=False)\n        self.to_q = nn.Linear(dim, inner_dim, bias=False)\n\n        self.to_out = (\n            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))\n            if project_out\n            else nn.Identity()\n        )\n\n    def forward(self, x, context=None, mask=None):\n\n        context = default(context, x)\n        k, v = self.to_kv(context).chunk(2, dim=-1)\n        q = self.to_q(x)\n        q, k, v = map(lambda t: rearrange(t, \"b n (h d) -> b h n d\", h=self.heads), [q, k, v])\n\n        if mask is not None:\n            q = q * mask[:, None, :, None]\n        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale\n        if mask is not None:\n            dots = dots - (1 - mask).float()[:, None, :, None] * 1e6\n        attn = self.attend(dots)\n        attn = self.dropout(attn)\n\n        out = torch.matmul(attn, v)\n\n        if mask is not None: # Just for good measure; this is probably overkill\n            out = out * mask[:, None, :,  None]\n        out = rearrange(out, \"b h n d -> b n (h d)\")\n        return self.to_out(out)\n\nclass TransformerCrossAttn(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L160\"\n    def __init__(\n        self,\n        dim: int,\n        depth: int,\n        heads: int,\n        dim_head: int,\n        mlp_dim: int,\n        dropout: float = 0.0,\n        norm: str = \"layer\",\n        norm_cond_dim: int = -1,\n        context_dim: Optional[int] = None,\n    ):\n        super().__init__()\n        self.layers = nn.ModuleList([])\n        for _ in range(depth):\n            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)\n            ca = CrossAttention(\n                dim, context_dim=context_dim, heads=heads, dim_head=dim_head, dropout=dropout\n            )\n            ff = FeedForward(dim, mlp_dim, dropout=dropout)\n            self.layers.append(\n                nn.ModuleList(\n                    [\n                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),\n                        PreNorm(dim, ca, norm=norm, norm_cond_dim=norm_cond_dim),\n                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),\n                    ]\n                )\n            )\n\n    def forward(self, x: torch.Tensor, *args, context=None, context_list=None, mask=None):\n\n        if context_list is None:\n            context_list = [context] * len(self.layers)\n\n        if len(context_list) != len(self.layers):\n            raise ValueError(f\"len(context_list) != len(self.layers) ({len(context_list)} != {len(self.layers)})\")\n\n        for i, (self_attn, cross_attn, ff) in enumerate(self.layers):\n            if mask is not None:\n                try:\n                    x = x * mask[:, :, None]\n                except:\n                    print(\"see \")\n                    import pdb; pdb.set_trace()\n            x = self_attn(x, mask=mask, *args) + x\n            x = cross_attn(x, mask=mask, *args, context=context_list[i]) + x\n            x = ff(x, *args) + x\n\n        if mask is not None:\n            x = x * mask[:, :, None]\n\n        return x\n\nclass DropTokenDropout(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L204\"\n    def __init__(self, p: float = 0.1):\n        super().__init__()\n        if p < 0 or p > 1:\n            raise ValueError(\n                \"dropout probability has to be between 0 and 1, \" \"but got {}\".format(p)\n            )\n        self.p = p\n\n    def forward(self, x: torch.Tensor):\n        # x: (batch_size, seq_len, dim)\n        if self.training and self.p > 0:\n            zero_mask = torch.full_like(x[0, :, 0], self.p).bernoulli().bool()\n            # TODO: permutation idx for each batch using torch.argsort\n            if zero_mask.any():\n                x = x[:, ~zero_mask, :]\n        return x\n\n\nclass ZeroTokenDropout(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L223\"\n    def __init__(self, p: float = 0.1):\n        super().__init__()\n        if p < 0 or p > 1:\n            raise ValueError(\n                \"dropout probability has to be between 0 and 1, \" \"but got {}\".format(p)\n            )\n        self.p = p\n\n    def forward(self, x: torch.Tensor):\n        # x: (batch_size, seq_len, dim)\n        if self.training and self.p > 0:\n            zero_mask = torch.full_like(x[:, :, 0], self.p).bernoulli().bool()\n            # Zero-out the masked tokens\n            x[zero_mask, :] = 0\n        return x\n\n\nclass TransformerDecoder(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L301\"\n    def __init__(\n        self,\n        num_tokens: int,\n        token_dim: int,\n        dim: int,\n        depth: int,\n        heads: int,\n        mlp_dim: int,\n        dim_head: int = 64,\n        dropout: float = 0.0,\n        emb_dropout: float = 0.0,\n        emb_dropout_type: str = 'drop',\n        norm: str = \"layer\",\n        norm_cond_dim: int = -1,\n        context_dim: Optional[int] = None,\n        skip_token_embedding: bool = False,\n    ):\n        super().__init__()\n        if not skip_token_embedding:\n            self.to_token_embedding = nn.Linear(token_dim, dim)\n        else:\n            self.to_token_embedding = nn.Identity()\n            if token_dim != dim:\n                raise ValueError(\n                    f\"token_dim ({token_dim}) != dim ({dim}) when skip_token_embedding is True\"\n                )\n\n        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))\n        if emb_dropout_type == \"drop\":\n            self.dropout = DropTokenDropout(emb_dropout)\n        elif emb_dropout_type == \"zero\":\n            self.dropout = ZeroTokenDropout(emb_dropout)\n        elif emb_dropout_type == \"normal\":\n            self.dropout = nn.Dropout(emb_dropout)\n\n        self.transformer = TransformerCrossAttn(\n            dim,\n            depth,\n            heads,\n            dim_head,\n            mlp_dim,\n            dropout,\n            norm=norm,\n            norm_cond_dim=norm_cond_dim,\n            context_dim=context_dim,\n        )\n\n    def forward(self, inp: torch.Tensor, *args, context=None, context_list=None, mask=None):\n        x = self.to_token_embedding(inp)\n        b, n, _ = x.shape\n\n        x = self.dropout(x)\n        #x += self.pos_embedding[:, :n]\n        x += self.pos_embedding[:, 0][:, None, :] # For now, we don't wish to embed a position. We might in future versions though.\n        x = self.transformer(x, *args, context=context, context_list=context_list, mask=mask)\n        return x","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.AdaptiveLayerNorm1D","uri":"program://Human3R/class/src.mhmr.blocks.cross_attn_transformer.AdaptiveLayerNorm1D#L11-L40","kind":"class","name":"AdaptiveLayerNorm1D","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":11,"end_line":40,"context_start_line":1,"context_end_line":60,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nfrom typing import Callable, Optional\nimport torch\nfrom torch import nn\nfrom inspect import isfunction\nfrom einops import rearrange\n\nclass AdaptiveLayerNorm1D(torch.nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/t_cond_mlp.py#L7\n    \"\"\"\n    def __init__(self, data_dim: int, norm_cond_dim: int):\n        super().__init__()\n        if data_dim <= 0:\n            raise ValueError(f\"data_dim must be positive, but got {data_dim}\")\n        if norm_cond_dim <= 0:\n            raise ValueError(f\"norm_cond_dim must be positive, but got {norm_cond_dim}\")\n        self.norm = torch.nn.LayerNorm(\n            data_dim\n        )  # TODO: Check if elementwise_affine=True is correct\n        self.linear = torch.nn.Linear(norm_cond_dim, 2 * data_dim)\n        torch.nn.init.zeros_(self.linear.weight)\n        torch.nn.init.zeros_(self.linear.bias)\n\n    def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor:\n        # x: (batch, ..., data_dim)\n        # t: (batch, norm_cond_dim)\n        # return: (batch, data_dim)\n        x = self.norm(x)\n        alpha, beta = self.linear(t).chunk(2, dim=-1)\n\n        # Add singleton dimensions to alpha and beta\n        if x.dim() > 2:\n            alpha = alpha.view(alpha.shape[0], *([1] * (x.dim() - 2)), alpha.shape[1])\n            beta = beta.view(beta.shape[0], *([1] * (x.dim() - 2)), beta.shape[1])\n\n        return x * (1 + alpha) + beta\n\n\ndef normalization_layer(norm: Optional[str], dim: int, norm_cond_dim: int = -1):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/t_cond_mlp.py#L48\n    \"\"\"\n    if norm == \"batch\":\n        return torch.nn.BatchNorm1d(dim)\n    elif norm == \"layer\":\n        return torch.nn.LayerNorm(dim)\n    elif norm == \"ada\":\n        assert norm_cond_dim > 0, f\"norm_cond_dim must be positive, got {norm_cond_dim}\"\n        return AdaptiveLayerNorm1D(dim, norm_cond_dim)\n    elif norm is None:\n        return torch.nn.Identity()\n    else:\n        raise ValueError(f\"Unknown norm: {norm}\")\n\n\ndef exists(val):","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.normalization_layer","uri":"program://Human3R/function/src.mhmr.blocks.cross_attn_transformer.normalization_layer#L43-L57","kind":"function","name":"normalization_layer","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":43,"end_line":57,"context_start_line":23,"context_end_line":77,"code":"        )  # TODO: Check if elementwise_affine=True is correct\n        self.linear = torch.nn.Linear(norm_cond_dim, 2 * data_dim)\n        torch.nn.init.zeros_(self.linear.weight)\n        torch.nn.init.zeros_(self.linear.bias)\n\n    def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor:\n        # x: (batch, ..., data_dim)\n        # t: (batch, norm_cond_dim)\n        # return: (batch, data_dim)\n        x = self.norm(x)\n        alpha, beta = self.linear(t).chunk(2, dim=-1)\n\n        # Add singleton dimensions to alpha and beta\n        if x.dim() > 2:\n            alpha = alpha.view(alpha.shape[0], *([1] * (x.dim() - 2)), alpha.shape[1])\n            beta = beta.view(beta.shape[0], *([1] * (x.dim() - 2)), beta.shape[1])\n\n        return x * (1 + alpha) + beta\n\n\ndef normalization_layer(norm: Optional[str], dim: int, norm_cond_dim: int = -1):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/t_cond_mlp.py#L48\n    \"\"\"\n    if norm == \"batch\":\n        return torch.nn.BatchNorm1d(dim)\n    elif norm == \"layer\":\n        return torch.nn.LayerNorm(dim)\n    elif norm == \"ada\":\n        assert norm_cond_dim > 0, f\"norm_cond_dim must be positive, got {norm_cond_dim}\"\n        return AdaptiveLayerNorm1D(dim, norm_cond_dim)\n    elif norm is None:\n        return torch.nn.Identity()\n    else:\n        raise ValueError(f\"Unknown norm: {norm}\")\n\n\ndef exists(val):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L17\"\n    return val is not None\n\n\ndef default(val, d):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L21\"\n    if exists(val):\n        return val\n    return d() if isfunction(d) else d\n\n\nclass PreNorm(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L27\n    \"\"\"\n    def __init__(self, dim: int, fn: Callable, norm: str = \"layer\", norm_cond_dim: int = -1):\n        super().__init__()","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.exists","uri":"program://Human3R/function/src.mhmr.blocks.cross_attn_transformer.exists#L60-L62","kind":"function","name":"exists","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":60,"end_line":62,"context_start_line":40,"context_end_line":82,"code":"        return x * (1 + alpha) + beta\n\n\ndef normalization_layer(norm: Optional[str], dim: int, norm_cond_dim: int = -1):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/t_cond_mlp.py#L48\n    \"\"\"\n    if norm == \"batch\":\n        return torch.nn.BatchNorm1d(dim)\n    elif norm == \"layer\":\n        return torch.nn.LayerNorm(dim)\n    elif norm == \"ada\":\n        assert norm_cond_dim > 0, f\"norm_cond_dim must be positive, got {norm_cond_dim}\"\n        return AdaptiveLayerNorm1D(dim, norm_cond_dim)\n    elif norm is None:\n        return torch.nn.Identity()\n    else:\n        raise ValueError(f\"Unknown norm: {norm}\")\n\n\ndef exists(val):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L17\"\n    return val is not None\n\n\ndef default(val, d):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L21\"\n    if exists(val):\n        return val\n    return d() if isfunction(d) else d\n\n\nclass PreNorm(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L27\n    \"\"\"\n    def __init__(self, dim: int, fn: Callable, norm: str = \"layer\", norm_cond_dim: int = -1):\n        super().__init__()\n        self.norm = normalization_layer(norm, dim, norm_cond_dim)\n        self.fn = fn\n\n    def forward(self, x: torch.Tensor, *args, **kwargs):\n        if isinstance(self.norm, AdaptiveLayerNorm1D):","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.default","uri":"program://Human3R/function/src.mhmr.blocks.cross_attn_transformer.default#L65-L69","kind":"function","name":"default","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":65,"end_line":69,"context_start_line":45,"context_end_line":89,"code":"    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/t_cond_mlp.py#L48\n    \"\"\"\n    if norm == \"batch\":\n        return torch.nn.BatchNorm1d(dim)\n    elif norm == \"layer\":\n        return torch.nn.LayerNorm(dim)\n    elif norm == \"ada\":\n        assert norm_cond_dim > 0, f\"norm_cond_dim must be positive, got {norm_cond_dim}\"\n        return AdaptiveLayerNorm1D(dim, norm_cond_dim)\n    elif norm is None:\n        return torch.nn.Identity()\n    else:\n        raise ValueError(f\"Unknown norm: {norm}\")\n\n\ndef exists(val):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L17\"\n    return val is not None\n\n\ndef default(val, d):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L21\"\n    if exists(val):\n        return val\n    return d() if isfunction(d) else d\n\n\nclass PreNorm(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L27\n    \"\"\"\n    def __init__(self, dim: int, fn: Callable, norm: str = \"layer\", norm_cond_dim: int = -1):\n        super().__init__()\n        self.norm = normalization_layer(norm, dim, norm_cond_dim)\n        self.fn = fn\n\n    def forward(self, x: torch.Tensor, *args, **kwargs):\n        if isinstance(self.norm, AdaptiveLayerNorm1D):\n            return self.fn(self.norm(x, *args), **kwargs)\n        else:\n            return self.fn(self.norm(x), **kwargs)\n\n\nclass FeedForward(nn.Module):\n    \"\"\"","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.PreNorm","uri":"program://Human3R/class/src.mhmr.blocks.cross_attn_transformer.PreNorm#L72-L85","kind":"class","name":"PreNorm","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":72,"end_line":85,"context_start_line":52,"context_end_line":105,"code":"        assert norm_cond_dim > 0, f\"norm_cond_dim must be positive, got {norm_cond_dim}\"\n        return AdaptiveLayerNorm1D(dim, norm_cond_dim)\n    elif norm is None:\n        return torch.nn.Identity()\n    else:\n        raise ValueError(f\"Unknown norm: {norm}\")\n\n\ndef exists(val):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L17\"\n    return val is not None\n\n\ndef default(val, d):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L21\"\n    if exists(val):\n        return val\n    return d() if isfunction(d) else d\n\n\nclass PreNorm(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L27\n    \"\"\"\n    def __init__(self, dim: int, fn: Callable, norm: str = \"layer\", norm_cond_dim: int = -1):\n        super().__init__()\n        self.norm = normalization_layer(norm, dim, norm_cond_dim)\n        self.fn = fn\n\n    def forward(self, x: torch.Tensor, *args, **kwargs):\n        if isinstance(self.norm, AdaptiveLayerNorm1D):\n            return self.fn(self.norm(x, *args), **kwargs)\n        else:\n            return self.fn(self.norm(x), **kwargs)\n\n\nclass FeedForward(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L40\n    \"\"\"\n    def __init__(self, dim, hidden_dim, dropout=0.0):\n        super().__init__()\n        self.net = nn.Sequential(\n            nn.Linear(dim, hidden_dim),\n            nn.GELU(),\n            nn.Dropout(dropout),\n            nn.Linear(hidden_dim, dim),\n            nn.Dropout(dropout),\n        )\n\n    def forward(self, x):\n        return self.net(x)\n\n","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.FeedForward","uri":"program://Human3R/class/src.mhmr.blocks.cross_attn_transformer.FeedForward#L88-L103","kind":"class","name":"FeedForward","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":88,"end_line":103,"context_start_line":68,"context_end_line":123,"code":"        return val\n    return d() if isfunction(d) else d\n\n\nclass PreNorm(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L27\n    \"\"\"\n    def __init__(self, dim: int, fn: Callable, norm: str = \"layer\", norm_cond_dim: int = -1):\n        super().__init__()\n        self.norm = normalization_layer(norm, dim, norm_cond_dim)\n        self.fn = fn\n\n    def forward(self, x: torch.Tensor, *args, **kwargs):\n        if isinstance(self.norm, AdaptiveLayerNorm1D):\n            return self.fn(self.norm(x, *args), **kwargs)\n        else:\n            return self.fn(self.norm(x), **kwargs)\n\n\nclass FeedForward(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L40\n    \"\"\"\n    def __init__(self, dim, hidden_dim, dropout=0.0):\n        super().__init__()\n        self.net = nn.Sequential(\n            nn.Linear(dim, hidden_dim),\n            nn.GELU(),\n            nn.Dropout(dropout),\n            nn.Linear(hidden_dim, dim),\n            nn.Dropout(dropout),\n        )\n\n    def forward(self, x):\n        return self.net(x)\n\n\nclass Attention(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L55\n    \"\"\"\n    def __init__(self, dim, heads=8, dim_head=64, dropout=0.0):\n        super().__init__()\n        inner_dim = dim_head * heads\n        project_out = not (heads == 1 and dim_head == dim)\n\n        self.heads = heads\n        self.scale = dim_head**-0.5\n\n        self.attend = nn.Softmax(dim=-1)\n        self.dropout = nn.Dropout(dropout)\n\n        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias=False)\n\n        self.to_out = (","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.Attention","uri":"program://Human3R/class/src.mhmr.blocks.cross_attn_transformer.Attention#L106-L159","kind":"class","name":"Attention","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":106,"end_line":159,"context_start_line":86,"context_end_line":179,"code":"\n\nclass FeedForward(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L40\n    \"\"\"\n    def __init__(self, dim, hidden_dim, dropout=0.0):\n        super().__init__()\n        self.net = nn.Sequential(\n            nn.Linear(dim, hidden_dim),\n            nn.GELU(),\n            nn.Dropout(dropout),\n            nn.Linear(hidden_dim, dim),\n            nn.Dropout(dropout),\n        )\n\n    def forward(self, x):\n        return self.net(x)\n\n\nclass Attention(nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L55\n    \"\"\"\n    def __init__(self, dim, heads=8, dim_head=64, dropout=0.0):\n        super().__init__()\n        inner_dim = dim_head * heads\n        project_out = not (heads == 1 and dim_head == dim)\n\n        self.heads = heads\n        self.scale = dim_head**-0.5\n\n        self.attend = nn.Softmax(dim=-1)\n        self.dropout = nn.Dropout(dropout)\n\n        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias=False)\n\n        self.to_out = (\n            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))\n            if project_out\n            else nn.Identity()\n        )\n\n    def forward(self, x, mask=None):\n\n        qkv = self.to_qkv(x).chunk(3, dim=-1)\n        # n --> the num query dimension\n\n        # TODO reshape b into b2 n and mask.\n        q, k, v = map(lambda t: rearrange(t, \"b n (h d) -> b h n d\", h=self.heads), qkv)\n\n        if mask is not None:\n            q, k, v = [x * mask[:, None, :, None] for x in [q, k, v]]\n        \n        # q, k, v: [13:51:03.400365] torch.Size([22, 1, 256])\n        #q, k ,vk after reshape: torch.Size([16, 8, 1, 32])\n        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale\n\n        if mask is not None:\n            dots = dots - (1 - mask)[:, None, None, :] * 10e10\n\n        attn = self.attend(dots)\n\n        if mask is not None: # Just for good measure; this is probably overkill\n            attn = attn * mask[:, None, None, :]\n\n        attn = self.dropout(attn)\n\n        out = torch.matmul(attn, v)\n\n        # out shape :torch.Size([16, 8, 1, 32])\n\n        out = rearrange(out, \"b h n d -> b n (h d)\")\n        return self.to_out(out)\n\n\nclass CrossAttention(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L89\"\n    def __init__(self, dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):\n        super().__init__()\n        inner_dim = dim_head * heads\n        project_out = not (heads == 1 and dim_head == dim)\n\n        self.heads = heads\n        self.scale = dim_head**-0.5\n\n        self.attend = nn.Softmax(dim=-1)\n        self.dropout = nn.Dropout(dropout)\n\n        context_dim = default(context_dim, dim)\n        self.to_kv = nn.Linear(context_dim, inner_dim * 2, bias=False)\n        self.to_q = nn.Linear(dim, inner_dim, bias=False)\n\n        self.to_out = (","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.CrossAttention","uri":"program://Human3R/class/src.mhmr.blocks.cross_attn_transformer.CrossAttention#L162-L205","kind":"class","name":"CrossAttention","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":162,"end_line":205,"context_start_line":142,"context_end_line":225,"code":"        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale\n\n        if mask is not None:\n            dots = dots - (1 - mask)[:, None, None, :] * 10e10\n\n        attn = self.attend(dots)\n\n        if mask is not None: # Just for good measure; this is probably overkill\n            attn = attn * mask[:, None, None, :]\n\n        attn = self.dropout(attn)\n\n        out = torch.matmul(attn, v)\n\n        # out shape :torch.Size([16, 8, 1, 32])\n\n        out = rearrange(out, \"b h n d -> b n (h d)\")\n        return self.to_out(out)\n\n\nclass CrossAttention(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L89\"\n    def __init__(self, dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):\n        super().__init__()\n        inner_dim = dim_head * heads\n        project_out = not (heads == 1 and dim_head == dim)\n\n        self.heads = heads\n        self.scale = dim_head**-0.5\n\n        self.attend = nn.Softmax(dim=-1)\n        self.dropout = nn.Dropout(dropout)\n\n        context_dim = default(context_dim, dim)\n        self.to_kv = nn.Linear(context_dim, inner_dim * 2, bias=False)\n        self.to_q = nn.Linear(dim, inner_dim, bias=False)\n\n        self.to_out = (\n            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))\n            if project_out\n            else nn.Identity()\n        )\n\n    def forward(self, x, context=None, mask=None):\n\n        context = default(context, x)\n        k, v = self.to_kv(context).chunk(2, dim=-1)\n        q = self.to_q(x)\n        q, k, v = map(lambda t: rearrange(t, \"b n (h d) -> b h n d\", h=self.heads), [q, k, v])\n\n        if mask is not None:\n            q = q * mask[:, None, :, None]\n        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale\n        if mask is not None:\n            dots = dots - (1 - mask).float()[:, None, :, None] * 1e6\n        attn = self.attend(dots)\n        attn = self.dropout(attn)\n\n        out = torch.matmul(attn, v)\n\n        if mask is not None: # Just for good measure; this is probably overkill\n            out = out * mask[:, None, :,  None]\n        out = rearrange(out, \"b h n d -> b n (h d)\")\n        return self.to_out(out)\n\nclass TransformerCrossAttn(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L160\"\n    def __init__(\n        self,\n        dim: int,\n        depth: int,\n        heads: int,\n        dim_head: int,\n        mlp_dim: int,\n        dropout: float = 0.0,\n        norm: str = \"layer\",\n        norm_cond_dim: int = -1,\n        context_dim: Optional[int] = None,\n    ):\n        super().__init__()\n        self.layers = nn.ModuleList([])\n        for _ in range(depth):\n            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)\n            ca = CrossAttention(","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.TransformerCrossAttn","uri":"program://Human3R/class/src.mhmr.blocks.cross_attn_transformer.TransformerCrossAttn#L207-L261","kind":"class","name":"TransformerCrossAttn","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":207,"end_line":261,"context_start_line":187,"context_end_line":281,"code":"        context = default(context, x)\n        k, v = self.to_kv(context).chunk(2, dim=-1)\n        q = self.to_q(x)\n        q, k, v = map(lambda t: rearrange(t, \"b n (h d) -> b h n d\", h=self.heads), [q, k, v])\n\n        if mask is not None:\n            q = q * mask[:, None, :, None]\n        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale\n        if mask is not None:\n            dots = dots - (1 - mask).float()[:, None, :, None] * 1e6\n        attn = self.attend(dots)\n        attn = self.dropout(attn)\n\n        out = torch.matmul(attn, v)\n\n        if mask is not None: # Just for good measure; this is probably overkill\n            out = out * mask[:, None, :,  None]\n        out = rearrange(out, \"b h n d -> b n (h d)\")\n        return self.to_out(out)\n\nclass TransformerCrossAttn(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L160\"\n    def __init__(\n        self,\n        dim: int,\n        depth: int,\n        heads: int,\n        dim_head: int,\n        mlp_dim: int,\n        dropout: float = 0.0,\n        norm: str = \"layer\",\n        norm_cond_dim: int = -1,\n        context_dim: Optional[int] = None,\n    ):\n        super().__init__()\n        self.layers = nn.ModuleList([])\n        for _ in range(depth):\n            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)\n            ca = CrossAttention(\n                dim, context_dim=context_dim, heads=heads, dim_head=dim_head, dropout=dropout\n            )\n            ff = FeedForward(dim, mlp_dim, dropout=dropout)\n            self.layers.append(\n                nn.ModuleList(\n                    [\n                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),\n                        PreNorm(dim, ca, norm=norm, norm_cond_dim=norm_cond_dim),\n                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),\n                    ]\n                )\n            )\n\n    def forward(self, x: torch.Tensor, *args, context=None, context_list=None, mask=None):\n\n        if context_list is None:\n            context_list = [context] * len(self.layers)\n\n        if len(context_list) != len(self.layers):\n            raise ValueError(f\"len(context_list) != len(self.layers) ({len(context_list)} != {len(self.layers)})\")\n\n        for i, (self_attn, cross_attn, ff) in enumerate(self.layers):\n            if mask is not None:\n                try:\n                    x = x * mask[:, :, None]\n                except:\n                    print(\"see \")\n                    import pdb; pdb.set_trace()\n            x = self_attn(x, mask=mask, *args) + x\n            x = cross_attn(x, mask=mask, *args, context=context_list[i]) + x\n            x = ff(x, *args) + x\n\n        if mask is not None:\n            x = x * mask[:, :, None]\n\n        return x\n\nclass DropTokenDropout(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L204\"\n    def __init__(self, p: float = 0.1):\n        super().__init__()\n        if p < 0 or p > 1:\n            raise ValueError(\n                \"dropout probability has to be between 0 and 1, \" \"but got {}\".format(p)\n            )\n        self.p = p\n\n    def forward(self, x: torch.Tensor):\n        # x: (batch_size, seq_len, dim)\n        if self.training and self.p > 0:\n            zero_mask = torch.full_like(x[0, :, 0], self.p).bernoulli().bool()\n            # TODO: permutation idx for each batch using torch.argsort\n            if zero_mask.any():\n                x = x[:, ~zero_mask, :]\n        return x\n","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.DropTokenDropout","uri":"program://Human3R/class/src.mhmr.blocks.cross_attn_transformer.DropTokenDropout#L263-L280","kind":"class","name":"DropTokenDropout","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":263,"end_line":280,"context_start_line":243,"context_end_line":300,"code":"\n        if len(context_list) != len(self.layers):\n            raise ValueError(f\"len(context_list) != len(self.layers) ({len(context_list)} != {len(self.layers)})\")\n\n        for i, (self_attn, cross_attn, ff) in enumerate(self.layers):\n            if mask is not None:\n                try:\n                    x = x * mask[:, :, None]\n                except:\n                    print(\"see \")\n                    import pdb; pdb.set_trace()\n            x = self_attn(x, mask=mask, *args) + x\n            x = cross_attn(x, mask=mask, *args, context=context_list[i]) + x\n            x = ff(x, *args) + x\n\n        if mask is not None:\n            x = x * mask[:, :, None]\n\n        return x\n\nclass DropTokenDropout(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L204\"\n    def __init__(self, p: float = 0.1):\n        super().__init__()\n        if p < 0 or p > 1:\n            raise ValueError(\n                \"dropout probability has to be between 0 and 1, \" \"but got {}\".format(p)\n            )\n        self.p = p\n\n    def forward(self, x: torch.Tensor):\n        # x: (batch_size, seq_len, dim)\n        if self.training and self.p > 0:\n            zero_mask = torch.full_like(x[0, :, 0], self.p).bernoulli().bool()\n            # TODO: permutation idx for each batch using torch.argsort\n            if zero_mask.any():\n                x = x[:, ~zero_mask, :]\n        return x\n\n\nclass ZeroTokenDropout(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L223\"\n    def __init__(self, p: float = 0.1):\n        super().__init__()\n        if p < 0 or p > 1:\n            raise ValueError(\n                \"dropout probability has to be between 0 and 1, \" \"but got {}\".format(p)\n            )\n        self.p = p\n\n    def forward(self, x: torch.Tensor):\n        # x: (batch_size, seq_len, dim)\n        if self.training and self.p > 0:\n            zero_mask = torch.full_like(x[:, :, 0], self.p).bernoulli().bool()\n            # Zero-out the masked tokens\n            x[zero_mask, :] = 0\n        return x\n","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.ZeroTokenDropout","uri":"program://Human3R/class/src.mhmr.blocks.cross_attn_transformer.ZeroTokenDropout#L283-L299","kind":"class","name":"ZeroTokenDropout","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":283,"end_line":299,"context_start_line":263,"context_end_line":319,"code":"class DropTokenDropout(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L204\"\n    def __init__(self, p: float = 0.1):\n        super().__init__()\n        if p < 0 or p > 1:\n            raise ValueError(\n                \"dropout probability has to be between 0 and 1, \" \"but got {}\".format(p)\n            )\n        self.p = p\n\n    def forward(self, x: torch.Tensor):\n        # x: (batch_size, seq_len, dim)\n        if self.training and self.p > 0:\n            zero_mask = torch.full_like(x[0, :, 0], self.p).bernoulli().bool()\n            # TODO: permutation idx for each batch using torch.argsort\n            if zero_mask.any():\n                x = x[:, ~zero_mask, :]\n        return x\n\n\nclass ZeroTokenDropout(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L223\"\n    def __init__(self, p: float = 0.1):\n        super().__init__()\n        if p < 0 or p > 1:\n            raise ValueError(\n                \"dropout probability has to be between 0 and 1, \" \"but got {}\".format(p)\n            )\n        self.p = p\n\n    def forward(self, x: torch.Tensor):\n        # x: (batch_size, seq_len, dim)\n        if self.training and self.p > 0:\n            zero_mask = torch.full_like(x[:, :, 0], self.p).bernoulli().bool()\n            # Zero-out the masked tokens\n            x[zero_mask, :] = 0\n        return x\n\n\nclass TransformerDecoder(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L301\"\n    def __init__(\n        self,\n        num_tokens: int,\n        token_dim: int,\n        dim: int,\n        depth: int,\n        heads: int,\n        mlp_dim: int,\n        dim_head: int = 64,\n        dropout: float = 0.0,\n        emb_dropout: float = 0.0,\n        emb_dropout_type: str = 'drop',\n        norm: str = \"layer\",\n        norm_cond_dim: int = -1,\n        context_dim: Optional[int] = None,\n        skip_token_embedding: bool = False,","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.TransformerDecoder","uri":"program://Human3R/class/src.mhmr.blocks.cross_attn_transformer.TransformerDecoder#L302-L359","kind":"class","name":"TransformerDecoder","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":302,"end_line":359,"context_start_line":282,"context_end_line":359,"code":"\nclass ZeroTokenDropout(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L223\"\n    def __init__(self, p: float = 0.1):\n        super().__init__()\n        if p < 0 or p > 1:\n            raise ValueError(\n                \"dropout probability has to be between 0 and 1, \" \"but got {}\".format(p)\n            )\n        self.p = p\n\n    def forward(self, x: torch.Tensor):\n        # x: (batch_size, seq_len, dim)\n        if self.training and self.p > 0:\n            zero_mask = torch.full_like(x[:, :, 0], self.p).bernoulli().bool()\n            # Zero-out the masked tokens\n            x[zero_mask, :] = 0\n        return x\n\n\nclass TransformerDecoder(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L301\"\n    def __init__(\n        self,\n        num_tokens: int,\n        token_dim: int,\n        dim: int,\n        depth: int,\n        heads: int,\n        mlp_dim: int,\n        dim_head: int = 64,\n        dropout: float = 0.0,\n        emb_dropout: float = 0.0,\n        emb_dropout_type: str = 'drop',\n        norm: str = \"layer\",\n        norm_cond_dim: int = -1,\n        context_dim: Optional[int] = None,\n        skip_token_embedding: bool = False,\n    ):\n        super().__init__()\n        if not skip_token_embedding:\n            self.to_token_embedding = nn.Linear(token_dim, dim)\n        else:\n            self.to_token_embedding = nn.Identity()\n            if token_dim != dim:\n                raise ValueError(\n                    f\"token_dim ({token_dim}) != dim ({dim}) when skip_token_embedding is True\"\n                )\n\n        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))\n        if emb_dropout_type == \"drop\":\n            self.dropout = DropTokenDropout(emb_dropout)\n        elif emb_dropout_type == \"zero\":\n            self.dropout = ZeroTokenDropout(emb_dropout)\n        elif emb_dropout_type == \"normal\":\n            self.dropout = nn.Dropout(emb_dropout)\n\n        self.transformer = TransformerCrossAttn(\n            dim,\n            depth,\n            heads,\n            dim_head,\n            mlp_dim,\n            dropout,\n            norm=norm,\n            norm_cond_dim=norm_cond_dim,\n            context_dim=context_dim,\n        )\n\n    def forward(self, inp: torch.Tensor, *args, context=None, context_list=None, mask=None):\n        x = self.to_token_embedding(inp)\n        b, n, _ = x.shape\n\n        x = self.dropout(x)\n        #x += self.pos_embedding[:, :n]\n        x += self.pos_embedding[:, 0][:, None, :] # For now, we don't wish to embed a position. We might in future versions though.\n        x = self.transformer(x, *args, context=context, context_list=context_list, mask=mask)\n        return x","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.__init__","uri":"program://Human3R/function/src.mhmr.blocks.cross_attn_transformer.__init__#L304-L349","kind":"function","name":"__init__","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":304,"end_line":349,"context_start_line":284,"context_end_line":359,"code":"    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L223\"\n    def __init__(self, p: float = 0.1):\n        super().__init__()\n        if p < 0 or p > 1:\n            raise ValueError(\n                \"dropout probability has to be between 0 and 1, \" \"but got {}\".format(p)\n            )\n        self.p = p\n\n    def forward(self, x: torch.Tensor):\n        # x: (batch_size, seq_len, dim)\n        if self.training and self.p > 0:\n            zero_mask = torch.full_like(x[:, :, 0], self.p).bernoulli().bool()\n            # Zero-out the masked tokens\n            x[zero_mask, :] = 0\n        return x\n\n\nclass TransformerDecoder(nn.Module):\n    \"Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/pose_transformer.py#L301\"\n    def __init__(\n        self,\n        num_tokens: int,\n        token_dim: int,\n        dim: int,\n        depth: int,\n        heads: int,\n        mlp_dim: int,\n        dim_head: int = 64,\n        dropout: float = 0.0,\n        emb_dropout: float = 0.0,\n        emb_dropout_type: str = 'drop',\n        norm: str = \"layer\",\n        norm_cond_dim: int = -1,\n        context_dim: Optional[int] = None,\n        skip_token_embedding: bool = False,\n    ):\n        super().__init__()\n        if not skip_token_embedding:\n            self.to_token_embedding = nn.Linear(token_dim, dim)\n        else:\n            self.to_token_embedding = nn.Identity()\n            if token_dim != dim:\n                raise ValueError(\n                    f\"token_dim ({token_dim}) != dim ({dim}) when skip_token_embedding is True\"\n                )\n\n        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))\n        if emb_dropout_type == \"drop\":\n            self.dropout = DropTokenDropout(emb_dropout)\n        elif emb_dropout_type == \"zero\":\n            self.dropout = ZeroTokenDropout(emb_dropout)\n        elif emb_dropout_type == \"normal\":\n            self.dropout = nn.Dropout(emb_dropout)\n\n        self.transformer = TransformerCrossAttn(\n            dim,\n            depth,\n            heads,\n            dim_head,\n            mlp_dim,\n            dropout,\n            norm=norm,\n            norm_cond_dim=norm_cond_dim,\n            context_dim=context_dim,\n        )\n\n    def forward(self, inp: torch.Tensor, *args, context=None, context_list=None, mask=None):\n        x = self.to_token_embedding(inp)\n        b, n, _ = x.shape\n\n        x = self.dropout(x)\n        #x += self.pos_embedding[:, :n]\n        x += self.pos_embedding[:, 0][:, None, :] # For now, we don't wish to embed a position. We might in future versions though.\n        x = self.transformer(x, *args, context=context, context_list=context_list, mask=mask)\n        return x","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.cross_attn_transformer.forward","uri":"program://Human3R/function/src.mhmr.blocks.cross_attn_transformer.forward#L351-L359","kind":"function","name":"forward","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":351,"end_line":359,"context_start_line":331,"context_end_line":359,"code":"        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))\n        if emb_dropout_type == \"drop\":\n            self.dropout = DropTokenDropout(emb_dropout)\n        elif emb_dropout_type == \"zero\":\n            self.dropout = ZeroTokenDropout(emb_dropout)\n        elif emb_dropout_type == \"normal\":\n            self.dropout = nn.Dropout(emb_dropout)\n\n        self.transformer = TransformerCrossAttn(\n            dim,\n            depth,\n            heads,\n            dim_head,\n            mlp_dim,\n            dropout,\n            norm=norm,\n            norm_cond_dim=norm_cond_dim,\n            context_dim=context_dim,\n        )\n\n    def forward(self, inp: torch.Tensor, *args, context=None, context_list=None, mask=None):\n        x = self.to_token_embedding(inp)\n        b, n, _ = x.shape\n\n        x = self.dropout(x)\n        #x += self.pos_embedding[:, :n]\n        x += self.pos_embedding[:, 0][:, None, :] # For now, we don't wish to embed a position. We might in future versions though.\n        x = self.transformer(x, *args, context=context, context_list=context_list, mask=mask)\n        return x","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.dinov2","uri":"program://Human3R/module/src.mhmr.blocks.dinov2#L1-L27","kind":"module","name":"src.mhmr.blocks.dinov2","path":"src/mhmr/blocks/dinov2.py","language":"python","start_line":1,"end_line":27,"context_start_line":1,"context_end_line":27,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\n\nclass Dinov2Backbone(nn.Module):\n    def __init__(self, name='dinov2_vitb14', pretrained=False, *args, **kwargs):\n        super().__init__()\n        self.name = name\n        self.encoder = torch.hub.load('facebookresearch/dinov2', self.name, pretrained=pretrained)\n        self.patch_size = self.encoder.patch_size\n        self.embed_dim = self.encoder.embed_dim\n\n    def forward(self, x):\n        \"\"\"\n        Encode a RGB image using a ViT-backbone\n        Args:\n            - x: torch.Tensor of shape [bs,3,w,h]\n        Return:\n            - y: torch.Tensor of shape [bs,k,d] - image in patchified mode\n        \"\"\"\n        assert len(x.shape) == 4\n        y = self.encoder.get_intermediate_layers(x)[0] # ViT-L+896x896: [bs,4096,1024] - [bs,nb_patches,emb]\n        return y\n","source_hash":"4b239e93d13eb9ee975f749374c8b36531fb220fb767a8488946258cdd3e630b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.dinov2.Dinov2Backbone","uri":"program://Human3R/class/src.mhmr.blocks.dinov2.Dinov2Backbone#L8-L26","kind":"class","name":"Dinov2Backbone","path":"src/mhmr/blocks/dinov2.py","language":"python","start_line":8,"end_line":26,"context_start_line":1,"context_end_line":27,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\n\nclass Dinov2Backbone(nn.Module):\n    def __init__(self, name='dinov2_vitb14', pretrained=False, *args, **kwargs):\n        super().__init__()\n        self.name = name\n        self.encoder = torch.hub.load('facebookresearch/dinov2', self.name, pretrained=pretrained)\n        self.patch_size = self.encoder.patch_size\n        self.embed_dim = self.encoder.embed_dim\n\n    def forward(self, x):\n        \"\"\"\n        Encode a RGB image using a ViT-backbone\n        Args:\n            - x: torch.Tensor of shape [bs,3,w,h]\n        Return:\n            - y: torch.Tensor of shape [bs,k,d] - image in patchified mode\n        \"\"\"\n        assert len(x.shape) == 4\n        y = self.encoder.get_intermediate_layers(x)[0] # ViT-L+896x896: [bs,4096,1024] - [bs,nb_patches,emb]\n        return y\n","source_hash":"4b239e93d13eb9ee975f749374c8b36531fb220fb767a8488946258cdd3e630b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.dinov2.__init__","uri":"program://Human3R/function/src.mhmr.blocks.dinov2.__init__#L9-L14","kind":"function","name":"__init__","path":"src/mhmr/blocks/dinov2.py","language":"python","start_line":9,"end_line":14,"context_start_line":1,"context_end_line":27,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\n\nclass Dinov2Backbone(nn.Module):\n    def __init__(self, name='dinov2_vitb14', pretrained=False, *args, **kwargs):\n        super().__init__()\n        self.name = name\n        self.encoder = torch.hub.load('facebookresearch/dinov2', self.name, pretrained=pretrained)\n        self.patch_size = self.encoder.patch_size\n        self.embed_dim = self.encoder.embed_dim\n\n    def forward(self, x):\n        \"\"\"\n        Encode a RGB image using a ViT-backbone\n        Args:\n            - x: torch.Tensor of shape [bs,3,w,h]\n        Return:\n            - y: torch.Tensor of shape [bs,k,d] - image in patchified mode\n        \"\"\"\n        assert len(x.shape) == 4\n        y = self.encoder.get_intermediate_layers(x)[0] # ViT-L+896x896: [bs,4096,1024] - [bs,nb_patches,emb]\n        return y\n","source_hash":"4b239e93d13eb9ee975f749374c8b36531fb220fb767a8488946258cdd3e630b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.mhmr.blocks.dinov2.forward","uri":"program://Human3R/function/src.mhmr.blocks.dinov2.forward#L16-L26","kind":"function","name":"forward","path":"src/mhmr/blocks/dinov2.py","language":"python","start_line":16,"end_line":26,"context_start_line":1,"context_end_line":27,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\n\nclass Dinov2Backbone(nn.Module):\n    def __init__(self, name='dinov2_vitb14', pretrained=False, *args, **kwargs):\n        super().__init__()\n        self.name = name\n        self.encoder = torch.hub.load('facebookresearch/dinov2', self.name, pretrained=pretrained)\n        self.patch_size = self.encoder.patch_size\n        self.embed_dim = self.encoder.embed_dim\n\n    def forward(self, x):\n        \"\"\"\n        Encode a RGB image using a ViT-backbone\n        Args:\n            - x: torch.Tensor of shape [bs,3,w,h]\n        Return:\n            - y: torch.Tensor of shape [bs,k,d] - image in patchified mode\n        \"\"\"\n        assert len(x.shape) == 4\n        y = self.encoder.get_intermediate_layers(x)[0] # ViT-L+896x896: [bs,4096,1024] - [bs,nb_patches,emb]\n        return y\n","source_hash":"4b239e93d13eb9ee975f749374c8b36531fb220fb767a8488946258cdd3e630b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.pretrain","uri":"program://Human3R/module/src.croco.pretrain#L1-L391","kind":"module","name":"src.croco.pretrain","path":"src/croco/pretrain.py","language":"python","start_line":1,"end_line":391,"context_start_line":1,"context_end_line":391,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# Pre-training CroCo\n# --------------------------------------------------------\n# References:\n# MAE: https://github.com/facebookresearch/mae\n# DeiT: https://github.com/facebookresearch/deit\n# BEiT: https://github.com/microsoft/unilm/tree/master/beit\n# --------------------------------------------------------\nimport argparse\nimport datetime\nimport json\nimport numpy as np\nimport os\nimport sys\nimport time\nimport math\nfrom pathlib import Path\nfrom typing import Iterable\n\nimport torch\nimport torch.distributed as dist\nimport torch.backends.cudnn as cudnn\nfrom torch.utils.tensorboard import SummaryWriter\nimport torchvision.transforms as transforms\nimport torchvision.datasets as datasets\n\nimport utils.misc as misc\nfrom utils.misc import NativeScalerWithGradNormCount as NativeScaler\nfrom models.croco import CroCoNet\nfrom models.criterion import MaskedMSE\nfrom datasets.pairs_dataset import PairsDataset\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser(\"CroCo pre-training\", add_help=False)\n    # model and criterion\n    parser.add_argument(\n        \"--model\",\n        default=\"CroCoNet()\",\n        type=str,\n        help=\"string containing the model to build\",\n    )\n    parser.add_argument(\n        \"--norm_pix_loss\",\n        default=1,\n        choices=[0, 1],\n        help=\"apply per-patch mean/std normalization before applying the loss\",\n    )\n    # dataset\n    parser.add_argument(\n        \"--dataset\", default=\"habitat_release\", type=str, help=\"training set\"\n    )\n    parser.add_argument(\n        \"--transforms\", default=\"crop224+acolor\", type=str, help=\"transforms to apply\"\n    )  # in the paper, we also use some homography and rotation, but find later that they were not useful or even harmful\n    # training\n    parser.add_argument(\"--seed\", default=0, type=int, help=\"Random seed\")\n    parser.add_argument(\n        \"--batch_size\",\n        default=64,\n        type=int,\n        help=\"Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus\",\n    )\n    parser.add_argument(\n        \"--epochs\",\n        default=800,\n        type=int,\n        help=\"Maximum number of epochs for the scheduler\",\n    )\n    parser.add_argument(\n        \"--max_epoch\", default=400, type=int, help=\"Stop training at this epoch\"\n    )\n    parser.add_argument(\n        \"--accum_iter\",\n        default=1,\n        type=int,\n        help=\"Accumulate gradient iterations (for increasing the effective batch size under memory constraints)\",\n    )\n    parser.add_argument(\n        \"--weight_decay\", type=float, default=0.05, help=\"weight decay (default: 0.05)\"\n    )\n    parser.add_argument(\n        \"--lr\",\n        type=float,\n        default=None,\n        metavar=\"LR\",\n        help=\"learning rate (absolute lr)\",\n    )\n    parser.add_argument(\n        \"--blr\",\n        type=float,\n        default=1.5e-4,\n        metavar=\"LR\",\n        help=\"base learning rate: absolute_lr = base_lr * total_batch_size / 256\",\n    )\n    parser.add_argument(\n        \"--min_lr\",\n        type=float,\n        default=0.0,\n        metavar=\"LR\",\n        help=\"lower lr bound for cyclic schedulers that hit 0\",\n    )\n    parser.add_argument(\n        \"--warmup_epochs\", type=int, default=40, metavar=\"N\", help=\"epochs to warmup LR\"\n    )\n    parser.add_argument(\n        \"--amp\",\n        type=int,\n        default=1,\n        choices=[0, 1],\n        help=\"Use Automatic Mixed Precision for pretraining\",\n    )\n    # others\n    parser.add_argument(\"--num_workers\", default=8, type=int)\n    parser.add_argument(\n        \"--world_size\", default=1, type=int, help=\"number of distributed processes\"\n    )\n    parser.add_argument(\"--local_rank\", default=-1, type=int)\n    parser.add_argument(\n        \"--dist_url\", default=\"env://\", help=\"url used to set up distributed training\"\n    )\n    parser.add_argument(\n        \"--save_freq\",\n        default=1,\n        type=int,\n        help=\"frequence (number of epochs) to save checkpoint in checkpoint-last.pth\",\n    )\n    parser.add_argument(\n        \"--keep_freq\",\n        default=20,\n        type=int,\n        help=\"frequence (number of epochs) to save checkpoint in checkpoint-%d.pth\",\n    )\n    parser.add_argument(\n        \"--print_freq\",\n        default=20,\n        type=int,\n        help=\"frequence (number of iterations) to print infos while training\",\n    )\n    # paths\n    parser.add_argument(\n        \"--output_dir\",\n        default=\"./output/\",\n        type=str,\n        help=\"path where to save the output\",\n    )\n    parser.add_argument(\n        \"--data_dir\", default=\"./data/\", type=str, help=\"path where data are stored\"\n    )\n    return parser\n\n\ndef main(args):\n    misc.init_distributed_mode(args)\n    global_rank = misc.get_rank()\n    world_size = misc.get_world_size()\n\n    print(\"output_dir: \" + args.output_dir)\n    if args.output_dir:\n        Path(args.output_dir).mkdir(parents=True, exist_ok=True)\n\n    # auto resume\n    last_ckpt_fname = os.path.join(args.output_dir, f\"checkpoint-last.pth\")\n    args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None\n\n    print(\"job dir: {}\".format(os.path.dirname(os.path.realpath(__file__))))\n    print(\"{}\".format(args).replace(\", \", \",\\n\"))\n\n    device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n    device = torch.device(device)\n\n    # fix the seed\n    seed = args.seed + misc.get_rank()\n    torch.manual_seed(seed)\n    np.random.seed(seed)\n\n    cudnn.benchmark = True\n\n    ## training dataset and loader\n    print(\n        \"Building dataset for {:s} with transforms {:s}\".format(\n            args.dataset, args.transforms\n        )\n    )\n    dataset = PairsDataset(args.dataset, trfs=args.transforms, data_dir=args.data_dir)\n    if world_size > 1:\n        sampler_train = torch.utils.data.DistributedSampler(\n            dataset, num_replicas=world_size, rank=global_rank, shuffle=True\n        )\n        print(\"Sampler_train = %s\" % str(sampler_train))\n    else:\n        sampler_train = torch.utils.data.RandomSampler(dataset)\n    data_loader_train = torch.utils.data.DataLoader(\n        dataset,\n        sampler=sampler_train,\n        batch_size=args.batch_size,\n        num_workers=args.num_workers,\n        pin_memory=True,\n        drop_last=True,\n    )\n\n    ## model\n    print(\"Loading model: {:s}\".format(args.model))\n    model = eval(args.model)\n    print(\n        \"Loading criterion: MaskedMSE(norm_pix_loss={:s})\".format(\n            str(bool(args.norm_pix_loss))\n        )\n    )\n    criterion = MaskedMSE(norm_pix_loss=bool(args.norm_pix_loss))\n\n    model.to(device)\n    model_without_ddp = model\n    print(\"Model = %s\" % str(model_without_ddp))\n\n    eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size()\n    if args.lr is None:  # only base_lr is specified\n        args.lr = args.blr * eff_batch_size / 256\n    print(\"base lr: %.2e\" % (args.lr * 256 / eff_batch_size))\n    print(\"actual lr: %.2e\" % args.lr)\n    print(\"accumulate grad iterations: %d\" % args.accum_iter)\n    print(\"effective batch size: %d\" % eff_batch_size)\n\n    if args.distributed:\n        model = torch.nn.parallel.DistributedDataParallel(\n            model, device_ids=[args.gpu], find_unused_parameters=True, static_graph=True\n        )\n        model_without_ddp = model.module\n\n    param_groups = misc.get_parameter_groups(\n        model_without_ddp, args.weight_decay\n    )  # following timm: set wd as 0 for bias and norm layers\n    optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))\n    print(optimizer)\n    loss_scaler = NativeScaler()\n\n    misc.load_model(\n        args=args,\n        model_without_ddp=model_without_ddp,\n        optimizer=optimizer,\n        loss_scaler=loss_scaler,\n    )\n\n    if global_rank == 0 and args.output_dir is not None:\n        log_writer = SummaryWriter(log_dir=args.output_dir)\n    else:\n        log_writer = None\n\n    print(f\"Start training until {args.max_epoch} epochs\")\n    start_time = time.time()\n    for epoch in range(args.start_epoch, args.max_epoch):\n        if world_size > 1:\n            data_loader_train.sampler.set_epoch(epoch)\n\n        train_stats = train_one_epoch(\n            model,\n            criterion,\n            data_loader_train,\n            optimizer,\n            device,\n            epoch,\n            loss_scaler,\n            log_writer=log_writer,\n            args=args,\n        )\n\n        if args.output_dir and epoch % args.save_freq == 0:\n            misc.save_model(\n                args=args,\n                model_without_ddp=model_without_ddp,\n                optimizer=optimizer,\n                loss_scaler=loss_scaler,\n                epoch=epoch,\n                fname=\"last\",\n            )\n\n        if (\n            args.output_dir\n            and (epoch % args.keep_freq == 0 or epoch + 1 == args.max_epoch)\n            and (epoch > 0 or args.max_epoch == 1)\n        ):\n            misc.save_model(\n                args=args,\n                model_without_ddp=model_without_ddp,\n                optimizer=optimizer,\n                loss_scaler=loss_scaler,\n                epoch=epoch,\n            )\n\n        log_stats = {\n            **{f\"train_{k}\": v for k, v in train_stats.items()},\n            \"epoch\": epoch,\n        }\n\n        if args.output_dir and misc.is_main_process():\n            if log_writer is not None:\n                log_writer.flush()\n            with open(\n                os.path.join(args.output_dir, \"log.txt\"), mode=\"a\", encoding=\"utf-8\"\n            ) as f:\n                f.write(json.dumps(log_stats) + \"\\n\")\n\n    total_time = time.time() - start_time\n    total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n    print(\"Training time {}\".format(total_time_str))\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    data_loader: Iterable,\n    optimizer: torch.optim.Optimizer,\n    device: torch.device,\n    epoch: int,\n    loss_scaler,\n    log_writer=None,\n    args=None,\n):\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)\n    accum_iter = args.accum_iter\n\n    optimizer.zero_grad()\n\n    if log_writer is not None:\n        print(\"log_dir: {}\".format(log_writer.log_dir))\n\n    for data_iter_step, (image1, image2) in enumerate(\n        metric_logger.log_every(data_loader, args.print_freq, header)\n    ):\n\n        # we use a per iteration  lr scheduler\n        if data_iter_step % accum_iter == 0:\n            misc.adjust_learning_rate(\n                optimizer, data_iter_step / len(data_loader) + epoch, args\n            )\n\n        image1 = image1.to(device, non_blocking=True)\n        image2 = image2.to(device, non_blocking=True)\n        with torch.cuda.amp.autocast(enabled=bool(args.amp)):\n            out, mask, target = model(image1, image2)\n            loss = criterion(out, mask, target)\n\n        loss_value = loss.item()\n\n        if not math.isfinite(loss_value):\n            print(\"Loss is {}, stopping training\".format(loss_value))\n            sys.exit(1)\n\n        loss /= accum_iter\n        loss_scaler(\n            loss,\n            optimizer,\n            parameters=model.parameters(),\n            update_grad=(data_iter_step + 1) % accum_iter == 0,\n        )\n        if (data_iter_step + 1) % accum_iter == 0:\n            optimizer.zero_grad()\n\n        torch.cuda.synchronize()\n\n        metric_logger.update(loss=loss_value)\n\n        lr = optimizer.param_groups[0][\"lr\"]\n        metric_logger.update(lr=lr)\n\n        loss_value_reduce = misc.all_reduce_mean(loss_value)\n        if (\n            log_writer is not None\n            and ((data_iter_step + 1) % (accum_iter * args.print_freq)) == 0\n        ):\n            # x-axis is based on epoch_1000x in the tensorboard, calibrating differences curves when batch size changes\n            epoch_1000x = int((data_iter_step / len(data_loader) + epoch) * 1000)\n            log_writer.add_scalar(\"train_loss\", loss_value_reduce, epoch_1000x)\n            log_writer.add_scalar(\"lr\", lr, epoch_1000x)\n\n    # gather the stats from all processes\n    metric_logger.synchronize_between_processes()\n    print(\"Averaged stats:\", metric_logger)\n    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    main(args)","source_hash":"d9ab7f1f3c1d4175e5eb4be181de5b2941ddb31fe6a9904212936beaa0d4ca1a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.pretrain.get_args_parser","uri":"program://Human3R/function/src.croco.pretrain.get_args_parser#L37-L153","kind":"function","name":"get_args_parser","path":"src/croco/pretrain.py","language":"python","start_line":37,"end_line":153,"context_start_line":17,"context_end_line":173,"code":"import sys\nimport time\nimport math\nfrom pathlib import Path\nfrom typing import Iterable\n\nimport torch\nimport torch.distributed as dist\nimport torch.backends.cudnn as cudnn\nfrom torch.utils.tensorboard import SummaryWriter\nimport torchvision.transforms as transforms\nimport torchvision.datasets as datasets\n\nimport utils.misc as misc\nfrom utils.misc import NativeScalerWithGradNormCount as NativeScaler\nfrom models.croco import CroCoNet\nfrom models.criterion import MaskedMSE\nfrom datasets.pairs_dataset import PairsDataset\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser(\"CroCo pre-training\", add_help=False)\n    # model and criterion\n    parser.add_argument(\n        \"--model\",\n        default=\"CroCoNet()\",\n        type=str,\n        help=\"string containing the model to build\",\n    )\n    parser.add_argument(\n        \"--norm_pix_loss\",\n        default=1,\n        choices=[0, 1],\n        help=\"apply per-patch mean/std normalization before applying the loss\",\n    )\n    # dataset\n    parser.add_argument(\n        \"--dataset\", default=\"habitat_release\", type=str, help=\"training set\"\n    )\n    parser.add_argument(\n        \"--transforms\", default=\"crop224+acolor\", type=str, help=\"transforms to apply\"\n    )  # in the paper, we also use some homography and rotation, but find later that they were not useful or even harmful\n    # training\n    parser.add_argument(\"--seed\", default=0, type=int, help=\"Random seed\")\n    parser.add_argument(\n        \"--batch_size\",\n        default=64,\n        type=int,\n        help=\"Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus\",\n    )\n    parser.add_argument(\n        \"--epochs\",\n        default=800,\n        type=int,\n        help=\"Maximum number of epochs for the scheduler\",\n    )\n    parser.add_argument(\n        \"--max_epoch\", default=400, type=int, help=\"Stop training at this epoch\"\n    )\n    parser.add_argument(\n        \"--accum_iter\",\n        default=1,\n        type=int,\n        help=\"Accumulate gradient iterations (for increasing the effective batch size under memory constraints)\",\n    )\n    parser.add_argument(\n        \"--weight_decay\", type=float, default=0.05, help=\"weight decay (default: 0.05)\"\n    )\n    parser.add_argument(\n        \"--lr\",\n        type=float,\n        default=None,\n        metavar=\"LR\",\n        help=\"learning rate (absolute lr)\",\n    )\n    parser.add_argument(\n        \"--blr\",\n        type=float,\n        default=1.5e-4,\n        metavar=\"LR\",\n        help=\"base learning rate: absolute_lr = base_lr * total_batch_size / 256\",\n    )\n    parser.add_argument(\n        \"--min_lr\",\n        type=float,\n        default=0.0,\n        metavar=\"LR\",\n        help=\"lower lr bound for cyclic schedulers that hit 0\",\n    )\n    parser.add_argument(\n        \"--warmup_epochs\", type=int, default=40, metavar=\"N\", help=\"epochs to warmup LR\"\n    )\n    parser.add_argument(\n        \"--amp\",\n        type=int,\n        default=1,\n        choices=[0, 1],\n        help=\"Use Automatic Mixed Precision for pretraining\",\n    )\n    # others\n    parser.add_argument(\"--num_workers\", default=8, type=int)\n    parser.add_argument(\n        \"--world_size\", default=1, type=int, help=\"number of distributed processes\"\n    )\n    parser.add_argument(\"--local_rank\", default=-1, type=int)\n    parser.add_argument(\n        \"--dist_url\", default=\"env://\", help=\"url used to set up distributed training\"\n    )\n    parser.add_argument(\n        \"--save_freq\",\n        default=1,\n        type=int,\n        help=\"frequence (number of epochs) to save checkpoint in checkpoint-last.pth\",\n    )\n    parser.add_argument(\n        \"--keep_freq\",\n        default=20,\n        type=int,\n        help=\"frequence (number of epochs) to save checkpoint in checkpoint-%d.pth\",\n    )\n    parser.add_argument(\n        \"--print_freq\",\n        default=20,\n        type=int,\n        help=\"frequence (number of iterations) to print infos while training\",\n    )\n    # paths\n    parser.add_argument(\n        \"--output_dir\",\n        default=\"./output/\",\n        type=str,\n        help=\"path where to save the output\",\n    )\n    parser.add_argument(\n        \"--data_dir\", default=\"./data/\", type=str, help=\"path where data are stored\"\n    )\n    return parser\n\n\ndef main(args):\n    misc.init_distributed_mode(args)\n    global_rank = misc.get_rank()\n    world_size = misc.get_world_size()\n\n    print(\"output_dir: \" + args.output_dir)\n    if args.output_dir:\n        Path(args.output_dir).mkdir(parents=True, exist_ok=True)\n\n    # auto resume\n    last_ckpt_fname = os.path.join(args.output_dir, f\"checkpoint-last.pth\")\n    args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None\n\n    print(\"job dir: {}\".format(os.path.dirname(os.path.realpath(__file__))))\n    print(\"{}\".format(args).replace(\", \", \",\\n\"))\n\n    device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n    device = torch.device(device)","source_hash":"d9ab7f1f3c1d4175e5eb4be181de5b2941ddb31fe6a9904212936beaa0d4ca1a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.pretrain.main","uri":"program://Human3R/function/src.croco.pretrain.main#L156-L308","kind":"function","name":"main","path":"src/croco/pretrain.py","language":"python","start_line":156,"end_line":308,"context_start_line":136,"context_end_line":328,"code":"    )\n    parser.add_argument(\n        \"--print_freq\",\n        default=20,\n        type=int,\n        help=\"frequence (number of iterations) to print infos while training\",\n    )\n    # paths\n    parser.add_argument(\n        \"--output_dir\",\n        default=\"./output/\",\n        type=str,\n        help=\"path where to save the output\",\n    )\n    parser.add_argument(\n        \"--data_dir\", default=\"./data/\", type=str, help=\"path where data are stored\"\n    )\n    return parser\n\n\ndef main(args):\n    misc.init_distributed_mode(args)\n    global_rank = misc.get_rank()\n    world_size = misc.get_world_size()\n\n    print(\"output_dir: \" + args.output_dir)\n    if args.output_dir:\n        Path(args.output_dir).mkdir(parents=True, exist_ok=True)\n\n    # auto resume\n    last_ckpt_fname = os.path.join(args.output_dir, f\"checkpoint-last.pth\")\n    args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None\n\n    print(\"job dir: {}\".format(os.path.dirname(os.path.realpath(__file__))))\n    print(\"{}\".format(args).replace(\", \", \",\\n\"))\n\n    device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n    device = torch.device(device)\n\n    # fix the seed\n    seed = args.seed + misc.get_rank()\n    torch.manual_seed(seed)\n    np.random.seed(seed)\n\n    cudnn.benchmark = True\n\n    ## training dataset and loader\n    print(\n        \"Building dataset for {:s} with transforms {:s}\".format(\n            args.dataset, args.transforms\n        )\n    )\n    dataset = PairsDataset(args.dataset, trfs=args.transforms, data_dir=args.data_dir)\n    if world_size > 1:\n        sampler_train = torch.utils.data.DistributedSampler(\n            dataset, num_replicas=world_size, rank=global_rank, shuffle=True\n        )\n        print(\"Sampler_train = %s\" % str(sampler_train))\n    else:\n        sampler_train = torch.utils.data.RandomSampler(dataset)\n    data_loader_train = torch.utils.data.DataLoader(\n        dataset,\n        sampler=sampler_train,\n        batch_size=args.batch_size,\n        num_workers=args.num_workers,\n        pin_memory=True,\n        drop_last=True,\n    )\n\n    ## model\n    print(\"Loading model: {:s}\".format(args.model))\n    model = eval(args.model)\n    print(\n        \"Loading criterion: MaskedMSE(norm_pix_loss={:s})\".format(\n            str(bool(args.norm_pix_loss))\n        )\n    )\n    criterion = MaskedMSE(norm_pix_loss=bool(args.norm_pix_loss))\n\n    model.to(device)\n    model_without_ddp = model\n    print(\"Model = %s\" % str(model_without_ddp))\n\n    eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size()\n    if args.lr is None:  # only base_lr is specified\n        args.lr = args.blr * eff_batch_size / 256\n    print(\"base lr: %.2e\" % (args.lr * 256 / eff_batch_size))\n    print(\"actual lr: %.2e\" % args.lr)\n    print(\"accumulate grad iterations: %d\" % args.accum_iter)\n    print(\"effective batch size: %d\" % eff_batch_size)\n\n    if args.distributed:\n        model = torch.nn.parallel.DistributedDataParallel(\n            model, device_ids=[args.gpu], find_unused_parameters=True, static_graph=True\n        )\n        model_without_ddp = model.module\n\n    param_groups = misc.get_parameter_groups(\n        model_without_ddp, args.weight_decay\n    )  # following timm: set wd as 0 for bias and norm layers\n    optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))\n    print(optimizer)\n    loss_scaler = NativeScaler()\n\n    misc.load_model(\n        args=args,\n        model_without_ddp=model_without_ddp,\n        optimizer=optimizer,\n        loss_scaler=loss_scaler,\n    )\n\n    if global_rank == 0 and args.output_dir is not None:\n        log_writer = SummaryWriter(log_dir=args.output_dir)\n    else:\n        log_writer = None\n\n    print(f\"Start training until {args.max_epoch} epochs\")\n    start_time = time.time()\n    for epoch in range(args.start_epoch, args.max_epoch):\n        if world_size > 1:\n            data_loader_train.sampler.set_epoch(epoch)\n\n        train_stats = train_one_epoch(\n            model,\n            criterion,\n            data_loader_train,\n            optimizer,\n            device,\n            epoch,\n            loss_scaler,\n            log_writer=log_writer,\n            args=args,\n        )\n\n        if args.output_dir and epoch % args.save_freq == 0:\n            misc.save_model(\n                args=args,\n                model_without_ddp=model_without_ddp,\n                optimizer=optimizer,\n                loss_scaler=loss_scaler,\n                epoch=epoch,\n                fname=\"last\",\n            )\n\n        if (\n            args.output_dir\n            and (epoch % args.keep_freq == 0 or epoch + 1 == args.max_epoch)\n            and (epoch > 0 or args.max_epoch == 1)\n        ):\n            misc.save_model(\n                args=args,\n                model_without_ddp=model_without_ddp,\n                optimizer=optimizer,\n                loss_scaler=loss_scaler,\n                epoch=epoch,\n            )\n\n        log_stats = {\n            **{f\"train_{k}\": v for k, v in train_stats.items()},\n            \"epoch\": epoch,\n        }\n\n        if args.output_dir and misc.is_main_process():\n            if log_writer is not None:\n                log_writer.flush()\n            with open(\n                os.path.join(args.output_dir, \"log.txt\"), mode=\"a\", encoding=\"utf-8\"\n            ) as f:\n                f.write(json.dumps(log_stats) + \"\\n\")\n\n    total_time = time.time() - start_time\n    total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n    print(\"Training time {}\".format(total_time_str))\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    data_loader: Iterable,\n    optimizer: torch.optim.Optimizer,\n    device: torch.device,\n    epoch: int,\n    loss_scaler,\n    log_writer=None,\n    args=None,\n):\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)\n    accum_iter = args.accum_iter\n\n    optimizer.zero_grad()","source_hash":"d9ab7f1f3c1d4175e5eb4be181de5b2941ddb31fe6a9904212936beaa0d4ca1a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.pretrain.train_one_epoch","uri":"program://Human3R/function/src.croco.pretrain.train_one_epoch#L311-L385","kind":"function","name":"train_one_epoch","path":"src/croco/pretrain.py","language":"python","start_line":311,"end_line":385,"context_start_line":291,"context_end_line":391,"code":"            )\n\n        log_stats = {\n            **{f\"train_{k}\": v for k, v in train_stats.items()},\n            \"epoch\": epoch,\n        }\n\n        if args.output_dir and misc.is_main_process():\n            if log_writer is not None:\n                log_writer.flush()\n            with open(\n                os.path.join(args.output_dir, \"log.txt\"), mode=\"a\", encoding=\"utf-8\"\n            ) as f:\n                f.write(json.dumps(log_stats) + \"\\n\")\n\n    total_time = time.time() - start_time\n    total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n    print(\"Training time {}\".format(total_time_str))\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    data_loader: Iterable,\n    optimizer: torch.optim.Optimizer,\n    device: torch.device,\n    epoch: int,\n    loss_scaler,\n    log_writer=None,\n    args=None,\n):\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)\n    accum_iter = args.accum_iter\n\n    optimizer.zero_grad()\n\n    if log_writer is not None:\n        print(\"log_dir: {}\".format(log_writer.log_dir))\n\n    for data_iter_step, (image1, image2) in enumerate(\n        metric_logger.log_every(data_loader, args.print_freq, header)\n    ):\n\n        # we use a per iteration  lr scheduler\n        if data_iter_step % accum_iter == 0:\n            misc.adjust_learning_rate(\n                optimizer, data_iter_step / len(data_loader) + epoch, args\n            )\n\n        image1 = image1.to(device, non_blocking=True)\n        image2 = image2.to(device, non_blocking=True)\n        with torch.cuda.amp.autocast(enabled=bool(args.amp)):\n            out, mask, target = model(image1, image2)\n            loss = criterion(out, mask, target)\n\n        loss_value = loss.item()\n\n        if not math.isfinite(loss_value):\n            print(\"Loss is {}, stopping training\".format(loss_value))\n            sys.exit(1)\n\n        loss /= accum_iter\n        loss_scaler(\n            loss,\n            optimizer,\n            parameters=model.parameters(),\n            update_grad=(data_iter_step + 1) % accum_iter == 0,\n        )\n        if (data_iter_step + 1) % accum_iter == 0:\n            optimizer.zero_grad()\n\n        torch.cuda.synchronize()\n\n        metric_logger.update(loss=loss_value)\n\n        lr = optimizer.param_groups[0][\"lr\"]\n        metric_logger.update(lr=lr)\n\n        loss_value_reduce = misc.all_reduce_mean(loss_value)\n        if (\n            log_writer is not None\n            and ((data_iter_step + 1) % (accum_iter * args.print_freq)) == 0\n        ):\n            # x-axis is based on epoch_1000x in the tensorboard, calibrating differences curves when batch size changes\n            epoch_1000x = int((data_iter_step / len(data_loader) + epoch) * 1000)\n            log_writer.add_scalar(\"train_loss\", loss_value_reduce, epoch_1000x)\n            log_writer.add_scalar(\"lr\", lr, epoch_1000x)\n\n    # gather the stats from all processes\n    metric_logger.synchronize_between_processes()\n    print(\"Averaged stats:\", metric_logger)\n    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    main(args)","source_hash":"d9ab7f1f3c1d4175e5eb4be181de5b2941ddb31fe6a9904212936beaa0d4ca1a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks","uri":"program://Human3R/module/src.croco.models.blocks#L1-L440","kind":"module","name":"src.croco.models.blocks","path":"src/croco/models/blocks.py","language":"python","start_line":1,"end_line":440,"context_start_line":1,"context_end_line":440,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# Main encoder/decoder blocks\n# --------------------------------------------------------\n# References:\n# timm\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/helpers.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/mlp.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/patch_embed.py\n\n\nimport torch\nimport torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc\nfrom torch.nn.functional import scaled_dot_product_attention\n\n\ndef _ntuple(n):\n    def parse(x):\n        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):\n            return x\n        return tuple(repeat(x, n))\n\n    return parse\n\n\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n\n        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])\n        self.act = act_layer()\n        self.drop1 = nn.Dropout(drop_probs[0])\n        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])\n        self.drop2 = nn.Dropout(drop_probs[1])\n\n    def forward(self, x):\n        return self.drop2(self.fc2(self.drop1(self.act(self.fc1(x)))))\n\n\nclass Mlp_flex(nn.Module):\n    \"\"\"Modified MLP with flexible number of layers\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        num_layers=2,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n        hidden_dims=None,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        \n        # process bias and dropout parameters\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n        \n        # build layer list\n        self.layers = nn.ModuleList()\n        \n        # if hidden_dims is specified, use it; otherwise use hidden_features\n        if hidden_dims is not None:\n            assert len(hidden_dims) == num_layers - 1\n            layer_dims = [in_features] + hidden_dims + [out_features]\n        else:\n            # use the same hidden_features\n            layer_dims = [in_features] + [hidden_features] * (num_layers - 1) + [out_features]\n        \n        # create each layer\n        for i in range(num_layers):\n            in_dim = layer_dims[i]\n            out_dim = layer_dims[i + 1]\n            \n            # linear layer\n            layer = nn.Linear(in_dim, out_dim, bias=bias[i % len(bias)] if isinstance(bias, (list, tuple)) else bias)\n            self.layers.append(layer)\n            \n            # add activation function and dropout except the last layer\n            if i < num_layers - 1:\n                self.layers.append(act_layer())\n                dropout_prob = drop_probs[i % len(drop_probs)] if isinstance(drop_probs, (list, tuple)) else drop_probs\n                if dropout_prob > 0:\n                    self.layers.append(nn.Dropout(dropout_prob))\n\n    def forward(self, x):\n        for layer in self.layers:\n            x = layer(x)\n        return x\n\n\nclass Attention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, x, xpos):\n        B, N, C = x.shape\n\n        qkv = (\n            self.qkv(x)\n            .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n            .transpose(1, 3)\n        )\n        q, k, v = [qkv[:, :, i] for i in range(3)]\n        # q,k,v = qkv.unbind(2)  # make torchscript happy (cannot use tensor as tuple)\n\n        q_type = q.dtype\n        k_type = k.dtype\n        if self.rope is not None:\n            q = q.to(torch.float16)\n            k = k.to(torch.float16)\n            with torch.autocast(device_type=\"cuda\", enabled=False):\n                q = self.rope(q, xpos)\n                k = self.rope(k, xpos)\n            q = q.to(q_type)\n            k = k.to(k_type)\n\n        # attn = (q @ k.transpose(-2, -1)) * self.scale\n        # attn = attn.softmax(dim=-1)\n        # attn = self.attn_drop(attn)\n\n        # x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n        # x = memory_efficient_attention(query=q.permute(0, 2, 1, 3), key=k.permute(0, 2, 1, 3), value=v.permute(0, 2, 1, 3), p=self.attn_drop.p, scale=self.scale).reshape(B, N, C)\n        x = (\n            scaled_dot_product_attention(\n                query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n            )\n            .transpose(1, 2)\n            .reshape(B, N, C)\n        )\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        return x\n\n\nclass Block(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x)))\n        return x\n\n\nclass CrossAttention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n\n        self.projq = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projk = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projv = nn.Linear(dim, dim, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, query, key, value, qpos, kpos):\n        B, Nq, C = query.shape\n        Nk = key.shape[1]\n        Nv = value.shape[1]\n\n        q = (\n            self.projq(query)\n            .reshape(B, Nq, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n        k = (\n            self.projk(key)\n            .reshape(B, Nk, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n        v = (\n            self.projv(value)\n            .reshape(B, Nv, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n\n        q_type = q.dtype\n        k_type = k.dtype\n        if self.rope is not None:\n            if qpos is not None:\n                q = q.to(torch.float16)\n                with torch.autocast(device_type=\"cuda\", enabled=False):\n                    q = self.rope(q, qpos)\n                q = q.to(q_type)\n\n            if kpos is not None:\n                k = k.to(torch.float16)\n                with torch.autocast(device_type=\"cuda\", enabled=False):\n                    k = self.rope(k, kpos)\n                k = k.to(k_type)\n\n        # attn = (q @ k.transpose(-2, -1)) * self.scale\n        # attn = attn.softmax(dim=-1)\n        # attn = self.attn_drop(attn)\n\n        # x = (attn @ v).transpose(1, 2).reshape(B, Nq, C)\n\n        # x = memory_efficient_attention(query=q.permute(0, 2, 1, 3), key=k.permute(0, 2, 1, 3), value=v.permute(0, 2, 1, 3), p=self.attn_drop.p, scale=self.scale).reshape(B, Nq, C)\n        x = (\n            scaled_dot_product_attention(\n                query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n            )\n            .transpose(1, 2)\n            .reshape(B, Nq, C)\n        )\n\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        return x\n\n\nclass DecoderBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.cross_attn = CrossAttention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        self.norm3 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, xpos, ypos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, y_, xpos, ypos))\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y\n\n\n# patch embedding\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n        self.num_patches = self.grid_size[0] * self.grid_size[1]\n        self.flatten = flatten\n\n        self.proj = nn.Conv2d(\n            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n        )\n        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()\n\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data\n        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks._ntuple","uri":"program://Human3R/function/src.croco.models.blocks._ntuple#L25-L31","kind":"function","name":"_ntuple","path":"src/croco/models/blocks.py","language":"python","start_line":25,"end_line":31,"context_start_line":5,"context_end_line":51,"code":"# --------------------------------------------------------\n# Main encoder/decoder blocks\n# --------------------------------------------------------\n# References:\n# timm\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/helpers.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/mlp.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/patch_embed.py\n\n\nimport torch\nimport torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc\nfrom torch.nn.functional import scaled_dot_product_attention\n\n\ndef _ntuple(n):\n    def parse(x):\n        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):\n            return x\n        return tuple(repeat(x, n))\n\n    return parse\n\n\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.drop_path","uri":"program://Human3R/function/src.croco.models.blocks.drop_path#L37-L50","kind":"function","name":"drop_path","path":"src/croco/models/blocks.py","language":"python","start_line":37,"end_line":50,"context_start_line":17,"context_end_line":70,"code":"import torch\nimport torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc\nfrom torch.nn.functional import scaled_dot_product_attention\n\n\ndef _ntuple(n):\n    def parse(x):\n        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):\n            return x\n        return tuple(repeat(x, n))\n\n    return parse\n\n\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.DropPath","uri":"program://Human3R/class/src.croco.models.blocks.DropPath#L53-L65","kind":"class","name":"DropPath","path":"src/croco/models/blocks.py","language":"python","start_line":53,"end_line":65,"context_start_line":33,"context_end_line":85,"code":"\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.Mlp","uri":"program://Human3R/class/src.croco.models.blocks.Mlp#L68-L93","kind":"class","name":"Mlp","path":"src/croco/models/blocks.py","language":"python","start_line":68,"end_line":93,"context_start_line":48,"context_end_line":113,"code":"    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n\n        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])\n        self.act = act_layer()\n        self.drop1 = nn.Dropout(drop_probs[0])\n        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])\n        self.drop2 = nn.Dropout(drop_probs[1])\n\n    def forward(self, x):\n        return self.drop2(self.fc2(self.drop1(self.act(self.fc1(x)))))\n\n\nclass Mlp_flex(nn.Module):\n    \"\"\"Modified MLP with flexible number of layers\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        num_layers=2,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n        hidden_dims=None,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        ","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.Mlp_flex","uri":"program://Human3R/class/src.croco.models.blocks.Mlp_flex#L96-L148","kind":"class","name":"Mlp_flex","path":"src/croco/models/blocks.py","language":"python","start_line":96,"end_line":148,"context_start_line":76,"context_end_line":168,"code":"        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n\n        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])\n        self.act = act_layer()\n        self.drop1 = nn.Dropout(drop_probs[0])\n        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])\n        self.drop2 = nn.Dropout(drop_probs[1])\n\n    def forward(self, x):\n        return self.drop2(self.fc2(self.drop1(self.act(self.fc1(x)))))\n\n\nclass Mlp_flex(nn.Module):\n    \"\"\"Modified MLP with flexible number of layers\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        num_layers=2,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n        hidden_dims=None,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        \n        # process bias and dropout parameters\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n        \n        # build layer list\n        self.layers = nn.ModuleList()\n        \n        # if hidden_dims is specified, use it; otherwise use hidden_features\n        if hidden_dims is not None:\n            assert len(hidden_dims) == num_layers - 1\n            layer_dims = [in_features] + hidden_dims + [out_features]\n        else:\n            # use the same hidden_features\n            layer_dims = [in_features] + [hidden_features] * (num_layers - 1) + [out_features]\n        \n        # create each layer\n        for i in range(num_layers):\n            in_dim = layer_dims[i]\n            out_dim = layer_dims[i + 1]\n            \n            # linear layer\n            layer = nn.Linear(in_dim, out_dim, bias=bias[i % len(bias)] if isinstance(bias, (list, tuple)) else bias)\n            self.layers.append(layer)\n            \n            # add activation function and dropout except the last layer\n            if i < num_layers - 1:\n                self.layers.append(act_layer())\n                dropout_prob = drop_probs[i % len(drop_probs)] if isinstance(drop_probs, (list, tuple)) else drop_probs\n                if dropout_prob > 0:\n                    self.layers.append(nn.Dropout(dropout_prob))\n\n    def forward(self, x):\n        for layer in self.layers:\n            x = layer(x)\n        return x\n\n\nclass Attention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, x, xpos):\n        B, N, C = x.shape\n","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.Attention","uri":"program://Human3R/class/src.croco.models.blocks.Attention#L151-L203","kind":"class","name":"Attention","path":"src/croco/models/blocks.py","language":"python","start_line":151,"end_line":203,"context_start_line":131,"context_end_line":223,"code":"            in_dim = layer_dims[i]\n            out_dim = layer_dims[i + 1]\n            \n            # linear layer\n            layer = nn.Linear(in_dim, out_dim, bias=bias[i % len(bias)] if isinstance(bias, (list, tuple)) else bias)\n            self.layers.append(layer)\n            \n            # add activation function and dropout except the last layer\n            if i < num_layers - 1:\n                self.layers.append(act_layer())\n                dropout_prob = drop_probs[i % len(drop_probs)] if isinstance(drop_probs, (list, tuple)) else drop_probs\n                if dropout_prob > 0:\n                    self.layers.append(nn.Dropout(dropout_prob))\n\n    def forward(self, x):\n        for layer in self.layers:\n            x = layer(x)\n        return x\n\n\nclass Attention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, x, xpos):\n        B, N, C = x.shape\n\n        qkv = (\n            self.qkv(x)\n            .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n            .transpose(1, 3)\n        )\n        q, k, v = [qkv[:, :, i] for i in range(3)]\n        # q,k,v = qkv.unbind(2)  # make torchscript happy (cannot use tensor as tuple)\n\n        q_type = q.dtype\n        k_type = k.dtype\n        if self.rope is not None:\n            q = q.to(torch.float16)\n            k = k.to(torch.float16)\n            with torch.autocast(device_type=\"cuda\", enabled=False):\n                q = self.rope(q, xpos)\n                k = self.rope(k, xpos)\n            q = q.to(q_type)\n            k = k.to(k_type)\n\n        # attn = (q @ k.transpose(-2, -1)) * self.scale\n        # attn = attn.softmax(dim=-1)\n        # attn = self.attn_drop(attn)\n\n        # x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n        # x = memory_efficient_attention(query=q.permute(0, 2, 1, 3), key=k.permute(0, 2, 1, 3), value=v.permute(0, 2, 1, 3), p=self.attn_drop.p, scale=self.scale).reshape(B, N, C)\n        x = (\n            scaled_dot_product_attention(\n                query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n            )\n            .transpose(1, 2)\n            .reshape(B, N, C)\n        )\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        return x\n\n\nclass Block(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.Block","uri":"program://Human3R/class/src.croco.models.blocks.Block#L206-L245","kind":"class","name":"Block","path":"src/croco/models/blocks.py","language":"python","start_line":206,"end_line":245,"context_start_line":186,"context_end_line":265,"code":"            k = k.to(k_type)\n\n        # attn = (q @ k.transpose(-2, -1)) * self.scale\n        # attn = attn.softmax(dim=-1)\n        # attn = self.attn_drop(attn)\n\n        # x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n        # x = memory_efficient_attention(query=q.permute(0, 2, 1, 3), key=k.permute(0, 2, 1, 3), value=v.permute(0, 2, 1, 3), p=self.attn_drop.p, scale=self.scale).reshape(B, N, C)\n        x = (\n            scaled_dot_product_attention(\n                query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n            )\n            .transpose(1, 2)\n            .reshape(B, N, C)\n        )\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        return x\n\n\nclass Block(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x)))\n        return x\n\n\nclass CrossAttention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n\n        self.projq = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projk = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projv = nn.Linear(dim, dim, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n\n        self.rope = rope.float() if rope is not None else None","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.CrossAttention","uri":"program://Human3R/class/src.croco.models.blocks.CrossAttention#L248-L320","kind":"class","name":"CrossAttention","path":"src/croco/models/blocks.py","language":"python","start_line":248,"end_line":320,"context_start_line":228,"context_end_line":340,"code":"            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x)))\n        return x\n\n\nclass CrossAttention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n\n        self.projq = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projk = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projv = nn.Linear(dim, dim, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, query, key, value, qpos, kpos):\n        B, Nq, C = query.shape\n        Nk = key.shape[1]\n        Nv = value.shape[1]\n\n        q = (\n            self.projq(query)\n            .reshape(B, Nq, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n        k = (\n            self.projk(key)\n            .reshape(B, Nk, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n        v = (\n            self.projv(value)\n            .reshape(B, Nv, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n\n        q_type = q.dtype\n        k_type = k.dtype\n        if self.rope is not None:\n            if qpos is not None:\n                q = q.to(torch.float16)\n                with torch.autocast(device_type=\"cuda\", enabled=False):\n                    q = self.rope(q, qpos)\n                q = q.to(q_type)\n\n            if kpos is not None:\n                k = k.to(torch.float16)\n                with torch.autocast(device_type=\"cuda\", enabled=False):\n                    k = self.rope(k, kpos)\n                k = k.to(k_type)\n\n        # attn = (q @ k.transpose(-2, -1)) * self.scale\n        # attn = attn.softmax(dim=-1)\n        # attn = self.attn_drop(attn)\n\n        # x = (attn @ v).transpose(1, 2).reshape(B, Nq, C)\n\n        # x = memory_efficient_attention(query=q.permute(0, 2, 1, 3), key=k.permute(0, 2, 1, 3), value=v.permute(0, 2, 1, 3), p=self.attn_drop.p, scale=self.scale).reshape(B, Nq, C)\n        x = (\n            scaled_dot_product_attention(\n                query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n            )\n            .transpose(1, 2)\n            .reshape(B, Nq, C)\n        )\n\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        return x\n\n\nclass DecoderBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.DecoderBlock","uri":"program://Human3R/class/src.croco.models.blocks.DecoderBlock#L323-L374","kind":"class","name":"DecoderBlock","path":"src/croco/models/blocks.py","language":"python","start_line":323,"end_line":374,"context_start_line":303,"context_end_line":394,"code":"        # attn = (q @ k.transpose(-2, -1)) * self.scale\n        # attn = attn.softmax(dim=-1)\n        # attn = self.attn_drop(attn)\n\n        # x = (attn @ v).transpose(1, 2).reshape(B, Nq, C)\n\n        # x = memory_efficient_attention(query=q.permute(0, 2, 1, 3), key=k.permute(0, 2, 1, 3), value=v.permute(0, 2, 1, 3), p=self.attn_drop.p, scale=self.scale).reshape(B, Nq, C)\n        x = (\n            scaled_dot_product_attention(\n                query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n            )\n            .transpose(1, 2)\n            .reshape(B, Nq, C)\n        )\n\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        return x\n\n\nclass DecoderBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.cross_attn = CrossAttention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        self.norm3 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, xpos, ypos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, y_, xpos, ypos))\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y\n\n\n# patch embedding\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.PositionGetter","uri":"program://Human3R/class/src.croco.models.blocks.PositionGetter#L378-L390","kind":"class","name":"PositionGetter","path":"src/croco/models/blocks.py","language":"python","start_line":378,"end_line":390,"context_start_line":358,"context_end_line":410,"code":"        self.norm2 = norm_layer(dim)\n        self.norm3 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, xpos, ypos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, y_, xpos, ypos))\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y\n\n\n# patch embedding\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.PatchEmbed","uri":"program://Human3R/class/src.croco.models.blocks.PatchEmbed#L393-L440","kind":"class","name":"PatchEmbed","path":"src/croco/models/blocks.py","language":"python","start_line":393,"end_line":440,"context_start_line":373,"context_end_line":440,"code":"        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y\n\n\n# patch embedding\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n        self.num_patches = self.grid_size[0] * self.grid_size[1]\n        self.flatten = flatten\n\n        self.proj = nn.Conv2d(\n            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n        )\n        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()\n\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data\n        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.parse","uri":"program://Human3R/function/src.croco.models.blocks.parse#L26-L29","kind":"function","name":"parse","path":"src/croco/models/blocks.py","language":"python","start_line":26,"end_line":29,"context_start_line":6,"context_end_line":49,"code":"# Main encoder/decoder blocks\n# --------------------------------------------------------\n# References:\n# timm\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/helpers.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/mlp.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/patch_embed.py\n\n\nimport torch\nimport torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc\nfrom torch.nn.functional import scaled_dot_product_attention\n\n\ndef _ntuple(n):\n    def parse(x):\n        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):\n            return x\n        return tuple(repeat(x, n))\n\n    return parse\n\n\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.__init__","uri":"program://Human3R/function/src.croco.models.blocks.__init__#L396-L419","kind":"function","name":"__init__","path":"src/croco/models/blocks.py","language":"python","start_line":396,"end_line":419,"context_start_line":376,"context_end_line":439,"code":"\n# patch embedding\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n        self.num_patches = self.grid_size[0] * self.grid_size[1]\n        self.flatten = flatten\n\n        self.proj = nn.Conv2d(\n            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n        )\n        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()\n\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.forward","uri":"program://Human3R/function/src.croco.models.blocks.forward#L421-L436","kind":"function","name":"forward","path":"src/croco/models/blocks.py","language":"python","start_line":421,"end_line":436,"context_start_line":401,"context_end_line":440,"code":"        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n        self.num_patches = self.grid_size[0] * self.grid_size[1]\n        self.flatten = flatten\n\n        self.proj = nn.Conv2d(\n            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n        )\n        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()\n\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data\n        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.extra_repr","uri":"program://Human3R/function/src.croco.models.blocks.extra_repr#L64-L65","kind":"function","name":"extra_repr","path":"src/croco/models/blocks.py","language":"python","start_line":64,"end_line":65,"context_start_line":44,"context_end_line":85,"code":"    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks.__call__","uri":"program://Human3R/function/src.croco.models.blocks.__call__#L384-L390","kind":"function","name":"__call__","path":"src/croco/models/blocks.py","language":"python","start_line":384,"end_line":390,"context_start_line":364,"context_end_line":410,"code":"            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, xpos, ypos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, y_, xpos, ypos))\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y\n\n\n# patch embedding\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.blocks._init_weights","uri":"program://Human3R/function/src.croco.models.blocks._init_weights#L438-L440","kind":"function","name":"_init_weights","path":"src/croco/models/blocks.py","language":"python","start_line":438,"end_line":440,"context_start_line":418,"context_end_line":440,"code":"\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data\n        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.criterion","uri":"program://Human3R/module/src.croco.models.criterion#L1-L38","kind":"module","name":"src.croco.models.criterion","path":"src/croco/models/criterion.py","language":"python","start_line":1,"end_line":38,"context_start_line":1,"context_end_line":38,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# Criterion to train CroCo\n# --------------------------------------------------------\n# References:\n# MAE: https://github.com/facebookresearch/mae\n# --------------------------------------------------------\n\nimport torch\n\n\nclass MaskedMSE(torch.nn.Module):\n\n    def __init__(self, norm_pix_loss=False, masked=True):\n        \"\"\"\n        norm_pix_loss: normalize each patch by their pixel mean and variance\n        masked: compute loss over the masked patches only\n        \"\"\"\n        super().__init__()\n        self.norm_pix_loss = norm_pix_loss\n        self.masked = masked\n\n    def forward(self, pred, mask, target):\n\n        if self.norm_pix_loss:\n            mean = target.mean(dim=-1, keepdim=True)\n            var = target.var(dim=-1, keepdim=True)\n            target = (target - mean) / (var + 1.0e-6) ** 0.5\n\n        loss = (pred - target) ** 2\n        loss = loss.mean(dim=-1)  # [N, L], mean loss per patch\n        if self.masked:\n            loss = (loss * mask).sum() / mask.sum()  # mean loss on masked patches\n        else:\n            loss = loss.mean()  # mean loss\n        return loss","source_hash":"19d837dec5326843d0e4c03ffae6cab0f9274f0c3354894e7fcd9762c97ef47c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.criterion.MaskedMSE","uri":"program://Human3R/class/src.croco.models.criterion.MaskedMSE#L14-L38","kind":"class","name":"MaskedMSE","path":"src/croco/models/criterion.py","language":"python","start_line":14,"end_line":38,"context_start_line":1,"context_end_line":38,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# Criterion to train CroCo\n# --------------------------------------------------------\n# References:\n# MAE: https://github.com/facebookresearch/mae\n# --------------------------------------------------------\n\nimport torch\n\n\nclass MaskedMSE(torch.nn.Module):\n\n    def __init__(self, norm_pix_loss=False, masked=True):\n        \"\"\"\n        norm_pix_loss: normalize each patch by their pixel mean and variance\n        masked: compute loss over the masked patches only\n        \"\"\"\n        super().__init__()\n        self.norm_pix_loss = norm_pix_loss\n        self.masked = masked\n\n    def forward(self, pred, mask, target):\n\n        if self.norm_pix_loss:\n            mean = target.mean(dim=-1, keepdim=True)\n            var = target.var(dim=-1, keepdim=True)\n            target = (target - mean) / (var + 1.0e-6) ** 0.5\n\n        loss = (pred - target) ** 2\n        loss = loss.mean(dim=-1)  # [N, L], mean loss per patch\n        if self.masked:\n            loss = (loss * mask).sum() / mask.sum()  # mean loss on masked patches\n        else:\n            loss = loss.mean()  # mean loss\n        return loss","source_hash":"19d837dec5326843d0e4c03ffae6cab0f9274f0c3354894e7fcd9762c97ef47c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.criterion.__init__","uri":"program://Human3R/function/src.croco.models.criterion.__init__#L16-L23","kind":"function","name":"__init__","path":"src/croco/models/criterion.py","language":"python","start_line":16,"end_line":23,"context_start_line":1,"context_end_line":38,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# Criterion to train CroCo\n# --------------------------------------------------------\n# References:\n# MAE: https://github.com/facebookresearch/mae\n# --------------------------------------------------------\n\nimport torch\n\n\nclass MaskedMSE(torch.nn.Module):\n\n    def __init__(self, norm_pix_loss=False, masked=True):\n        \"\"\"\n        norm_pix_loss: normalize each patch by their pixel mean and variance\n        masked: compute loss over the masked patches only\n        \"\"\"\n        super().__init__()\n        self.norm_pix_loss = norm_pix_loss\n        self.masked = masked\n\n    def forward(self, pred, mask, target):\n\n        if self.norm_pix_loss:\n            mean = target.mean(dim=-1, keepdim=True)\n            var = target.var(dim=-1, keepdim=True)\n            target = (target - mean) / (var + 1.0e-6) ** 0.5\n\n        loss = (pred - target) ** 2\n        loss = loss.mean(dim=-1)  # [N, L], mean loss per patch\n        if self.masked:\n            loss = (loss * mask).sum() / mask.sum()  # mean loss on masked patches\n        else:\n            loss = loss.mean()  # mean loss\n        return loss","source_hash":"19d837dec5326843d0e4c03ffae6cab0f9274f0c3354894e7fcd9762c97ef47c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.criterion.forward","uri":"program://Human3R/function/src.croco.models.criterion.forward#L25-L38","kind":"function","name":"forward","path":"src/croco/models/criterion.py","language":"python","start_line":25,"end_line":38,"context_start_line":5,"context_end_line":38,"code":"# Criterion to train CroCo\n# --------------------------------------------------------\n# References:\n# MAE: https://github.com/facebookresearch/mae\n# --------------------------------------------------------\n\nimport torch\n\n\nclass MaskedMSE(torch.nn.Module):\n\n    def __init__(self, norm_pix_loss=False, masked=True):\n        \"\"\"\n        norm_pix_loss: normalize each patch by their pixel mean and variance\n        masked: compute loss over the masked patches only\n        \"\"\"\n        super().__init__()\n        self.norm_pix_loss = norm_pix_loss\n        self.masked = masked\n\n    def forward(self, pred, mask, target):\n\n        if self.norm_pix_loss:\n            mean = target.mean(dim=-1, keepdim=True)\n            var = target.var(dim=-1, keepdim=True)\n            target = (target - mean) / (var + 1.0e-6) ** 0.5\n\n        loss = (pred - target) ** 2\n        loss = loss.mean(dim=-1)  # [N, L], mean loss per patch\n        if self.masked:\n            loss = (loss * mask).sum() / mask.sum()  # mean loss on masked patches\n        else:\n            loss = loss.mean()  # mean loss\n        return loss","source_hash":"19d837dec5326843d0e4c03ffae6cab0f9274f0c3354894e7fcd9762c97ef47c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream","uri":"program://Human3R/module/src.croco.models.croco_downstream#L1-L141","kind":"module","name":"src.croco.models.croco_downstream","path":"src/croco/models/croco_downstream.py","language":"python","start_line":1,"end_line":141,"context_start_line":1,"context_end_line":141,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# CroCo model for downstream tasks\n# --------------------------------------------------------\n\nimport torch\n\nfrom .croco import CroCoNet\n\n\ndef croco_args_from_ckpt(ckpt):\n    if \"croco_kwargs\" in ckpt:  # CroCo v2 released models\n        return ckpt[\"croco_kwargs\"]\n    elif \"args\" in ckpt and hasattr(\n        ckpt[\"args\"], \"model\"\n    ):  # pretrained using the official code release\n        s = ckpt[\n            \"args\"\n        ].model  # eg \"CroCoNet(enc_embed_dim=1024, enc_num_heads=16, enc_depth=24)\"\n        assert s.startswith(\"CroCoNet(\")\n        return eval(\n            \"dict\" + s[len(\"CroCoNet\") :]\n        )  # transform it into the string of a dictionary and evaluate it\n    else:  # CroCo v1 released models\n        return dict()\n\n\nclass CroCoDownstreamMonocularEncoder(CroCoNet):\n\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for monocular downstream task, only using the encoder.\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        NOTE: It works by *calling super().__init__() but with redefined setters\n\n        \"\"\"\n        super(CroCoDownstreamMonocularEncoder, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_decoder(self, *args, **kwargs):\n        \"\"\"No decoder\"\"\"\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No 'prediction head' for downstream tasks.\"\"\"\n        return\n\n    def forward(self, img):\n        \"\"\"\n        img if of size batch_size x 3 x h x w\n        \"\"\"\n        B, C, H, W = img.size()\n        img_info = {\"height\": H, \"width\": W}\n        need_all_layers = (\n            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, _, _ = self._encode_image(\n            img, do_mask=False, return_all_blocks=need_all_layers\n        )\n        return self.head(out, img_info)\n\n\nclass CroCoDownstreamBinocular(CroCoNet):\n\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for binocular downstream task\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        \"\"\"\n        super(CroCoDownstreamBinocular, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head for downstream tasks, define your own head\"\"\"\n        return\n\n    def encode_image_pairs(self, img1, img2, return_all_blocks=False):\n        \"\"\"run encoder for a pair of images\n        it is actually ~5% faster to concatenate the images along the batch dimension\n         than to encode them separately\n        \"\"\"\n        ## the two commented lines below is the naive version with separate encoding\n        # out, pos, _ = self._encode_image(img1, do_mask=False, return_all_blocks=return_all_blocks)\n        # out2, pos2, _ = self._encode_image(img2, do_mask=False, return_all_blocks=False)\n        ## and now the faster version\n        out, pos, _ = self._encode_image(\n            torch.cat((img1, img2), dim=0),\n            do_mask=False,\n            return_all_blocks=return_all_blocks,\n        )\n        if return_all_blocks:\n            out, out2 = list(map(list, zip(*[o.chunk(2, dim=0) for o in out])))\n            out2 = out2[-1]\n        else:\n            out, out2 = out.chunk(2, dim=0)\n        pos, pos2 = pos.chunk(2, dim=0)\n        return out, out2, pos, pos2\n\n    def forward(self, img1, img2):\n        B, C, H, W = img1.size()\n        img_info = {\"height\": H, \"width\": W}\n        return_all_blocks = (\n            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, out2, pos, pos2 = self.encode_image_pairs(\n            img1, img2, return_all_blocks=return_all_blocks\n        )\n        if return_all_blocks:\n            decout = self._decoder(\n                out[-1], pos, None, out2, pos2, return_all_blocks=return_all_blocks\n            )\n            decout = out + decout\n        else:\n            decout = self._decoder(\n                out, pos, None, out2, pos2, return_all_blocks=return_all_blocks\n            )\n        return self.head(decout, img_info)","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream.croco_args_from_ckpt","uri":"program://Human3R/function/src.croco.models.croco_downstream.croco_args_from_ckpt#L13-L27","kind":"function","name":"croco_args_from_ckpt","path":"src/croco/models/croco_downstream.py","language":"python","start_line":13,"end_line":27,"context_start_line":1,"context_end_line":47,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# CroCo model for downstream tasks\n# --------------------------------------------------------\n\nimport torch\n\nfrom .croco import CroCoNet\n\n\ndef croco_args_from_ckpt(ckpt):\n    if \"croco_kwargs\" in ckpt:  # CroCo v2 released models\n        return ckpt[\"croco_kwargs\"]\n    elif \"args\" in ckpt and hasattr(\n        ckpt[\"args\"], \"model\"\n    ):  # pretrained using the official code release\n        s = ckpt[\n            \"args\"\n        ].model  # eg \"CroCoNet(enc_embed_dim=1024, enc_num_heads=16, enc_depth=24)\"\n        assert s.startswith(\"CroCoNet(\")\n        return eval(\n            \"dict\" + s[len(\"CroCoNet\") :]\n        )  # transform it into the string of a dictionary and evaluate it\n    else:  # CroCo v1 released models\n        return dict()\n\n\nclass CroCoDownstreamMonocularEncoder(CroCoNet):\n\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for monocular downstream task, only using the encoder.\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        NOTE: It works by *calling super().__init__() but with redefined setters\n\n        \"\"\"\n        super(CroCoDownstreamMonocularEncoder, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream.CroCoDownstreamMonocularEncoder","uri":"program://Human3R/class/src.croco.models.croco_downstream.CroCoDownstreamMonocularEncoder#L30-L73","kind":"class","name":"CroCoDownstreamMonocularEncoder","path":"src/croco/models/croco_downstream.py","language":"python","start_line":30,"end_line":73,"context_start_line":10,"context_end_line":93,"code":"from .croco import CroCoNet\n\n\ndef croco_args_from_ckpt(ckpt):\n    if \"croco_kwargs\" in ckpt:  # CroCo v2 released models\n        return ckpt[\"croco_kwargs\"]\n    elif \"args\" in ckpt and hasattr(\n        ckpt[\"args\"], \"model\"\n    ):  # pretrained using the official code release\n        s = ckpt[\n            \"args\"\n        ].model  # eg \"CroCoNet(enc_embed_dim=1024, enc_num_heads=16, enc_depth=24)\"\n        assert s.startswith(\"CroCoNet(\")\n        return eval(\n            \"dict\" + s[len(\"CroCoNet\") :]\n        )  # transform it into the string of a dictionary and evaluate it\n    else:  # CroCo v1 released models\n        return dict()\n\n\nclass CroCoDownstreamMonocularEncoder(CroCoNet):\n\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for monocular downstream task, only using the encoder.\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        NOTE: It works by *calling super().__init__() but with redefined setters\n\n        \"\"\"\n        super(CroCoDownstreamMonocularEncoder, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_decoder(self, *args, **kwargs):\n        \"\"\"No decoder\"\"\"\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No 'prediction head' for downstream tasks.\"\"\"\n        return\n\n    def forward(self, img):\n        \"\"\"\n        img if of size batch_size x 3 x h x w\n        \"\"\"\n        B, C, H, W = img.size()\n        img_info = {\"height\": H, \"width\": W}\n        need_all_layers = (\n            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, _, _ = self._encode_image(\n            img, do_mask=False, return_all_blocks=need_all_layers\n        )\n        return self.head(out, img_info)\n\n\nclass CroCoDownstreamBinocular(CroCoNet):\n\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for binocular downstream task\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        \"\"\"\n        super(CroCoDownstreamBinocular, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream.CroCoDownstreamBinocular","uri":"program://Human3R/class/src.croco.models.croco_downstream.CroCoDownstreamBinocular#L76-L141","kind":"class","name":"CroCoDownstreamBinocular","path":"src/croco/models/croco_downstream.py","language":"python","start_line":76,"end_line":141,"context_start_line":56,"context_end_line":141,"code":"\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No 'prediction head' for downstream tasks.\"\"\"\n        return\n\n    def forward(self, img):\n        \"\"\"\n        img if of size batch_size x 3 x h x w\n        \"\"\"\n        B, C, H, W = img.size()\n        img_info = {\"height\": H, \"width\": W}\n        need_all_layers = (\n            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, _, _ = self._encode_image(\n            img, do_mask=False, return_all_blocks=need_all_layers\n        )\n        return self.head(out, img_info)\n\n\nclass CroCoDownstreamBinocular(CroCoNet):\n\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for binocular downstream task\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        \"\"\"\n        super(CroCoDownstreamBinocular, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head for downstream tasks, define your own head\"\"\"\n        return\n\n    def encode_image_pairs(self, img1, img2, return_all_blocks=False):\n        \"\"\"run encoder for a pair of images\n        it is actually ~5% faster to concatenate the images along the batch dimension\n         than to encode them separately\n        \"\"\"\n        ## the two commented lines below is the naive version with separate encoding\n        # out, pos, _ = self._encode_image(img1, do_mask=False, return_all_blocks=return_all_blocks)\n        # out2, pos2, _ = self._encode_image(img2, do_mask=False, return_all_blocks=False)\n        ## and now the faster version\n        out, pos, _ = self._encode_image(\n            torch.cat((img1, img2), dim=0),\n            do_mask=False,\n            return_all_blocks=return_all_blocks,\n        )\n        if return_all_blocks:\n            out, out2 = list(map(list, zip(*[o.chunk(2, dim=0) for o in out])))\n            out2 = out2[-1]\n        else:\n            out, out2 = out.chunk(2, dim=0)\n        pos, pos2 = pos.chunk(2, dim=0)\n        return out, out2, pos, pos2\n\n    def forward(self, img1, img2):\n        B, C, H, W = img1.size()\n        img_info = {\"height\": H, \"width\": W}\n        return_all_blocks = (\n            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, out2, pos, pos2 = self.encode_image_pairs(\n            img1, img2, return_all_blocks=return_all_blocks\n        )\n        if return_all_blocks:\n            decout = self._decoder(\n                out[-1], pos, None, out2, pos2, return_all_blocks=return_all_blocks\n            )\n            decout = out + decout\n        else:\n            decout = self._decoder(\n                out, pos, None, out2, pos2, return_all_blocks=return_all_blocks\n            )\n        return self.head(decout, img_info)","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream.__init__","uri":"program://Human3R/function/src.croco.models.croco_downstream.__init__#L78-L86","kind":"function","name":"__init__","path":"src/croco/models/croco_downstream.py","language":"python","start_line":78,"end_line":86,"context_start_line":58,"context_end_line":106,"code":"        \"\"\"No 'prediction head' for downstream tasks.\"\"\"\n        return\n\n    def forward(self, img):\n        \"\"\"\n        img if of size batch_size x 3 x h x w\n        \"\"\"\n        B, C, H, W = img.size()\n        img_info = {\"height\": H, \"width\": W}\n        need_all_layers = (\n            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, _, _ = self._encode_image(\n            img, do_mask=False, return_all_blocks=need_all_layers\n        )\n        return self.head(out, img_info)\n\n\nclass CroCoDownstreamBinocular(CroCoNet):\n\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for binocular downstream task\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        \"\"\"\n        super(CroCoDownstreamBinocular, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head for downstream tasks, define your own head\"\"\"\n        return\n\n    def encode_image_pairs(self, img1, img2, return_all_blocks=False):\n        \"\"\"run encoder for a pair of images\n        it is actually ~5% faster to concatenate the images along the batch dimension\n         than to encode them separately\n        \"\"\"\n        ## the two commented lines below is the naive version with separate encoding","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream._set_mask_generator","uri":"program://Human3R/function/src.croco.models.croco_downstream._set_mask_generator#L88-L90","kind":"function","name":"_set_mask_generator","path":"src/croco/models/croco_downstream.py","language":"python","start_line":88,"end_line":90,"context_start_line":68,"context_end_line":110,"code":"            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, _, _ = self._encode_image(\n            img, do_mask=False, return_all_blocks=need_all_layers\n        )\n        return self.head(out, img_info)\n\n\nclass CroCoDownstreamBinocular(CroCoNet):\n\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for binocular downstream task\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        \"\"\"\n        super(CroCoDownstreamBinocular, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head for downstream tasks, define your own head\"\"\"\n        return\n\n    def encode_image_pairs(self, img1, img2, return_all_blocks=False):\n        \"\"\"run encoder for a pair of images\n        it is actually ~5% faster to concatenate the images along the batch dimension\n         than to encode them separately\n        \"\"\"\n        ## the two commented lines below is the naive version with separate encoding\n        # out, pos, _ = self._encode_image(img1, do_mask=False, return_all_blocks=return_all_blocks)\n        # out2, pos2, _ = self._encode_image(img2, do_mask=False, return_all_blocks=False)\n        ## and now the faster version\n        out, pos, _ = self._encode_image(","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream._set_mask_token","uri":"program://Human3R/function/src.croco.models.croco_downstream._set_mask_token#L92-L95","kind":"function","name":"_set_mask_token","path":"src/croco/models/croco_downstream.py","language":"python","start_line":92,"end_line":95,"context_start_line":72,"context_end_line":115,"code":"        )\n        return self.head(out, img_info)\n\n\nclass CroCoDownstreamBinocular(CroCoNet):\n\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for binocular downstream task\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        \"\"\"\n        super(CroCoDownstreamBinocular, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head for downstream tasks, define your own head\"\"\"\n        return\n\n    def encode_image_pairs(self, img1, img2, return_all_blocks=False):\n        \"\"\"run encoder for a pair of images\n        it is actually ~5% faster to concatenate the images along the batch dimension\n         than to encode them separately\n        \"\"\"\n        ## the two commented lines below is the naive version with separate encoding\n        # out, pos, _ = self._encode_image(img1, do_mask=False, return_all_blocks=return_all_blocks)\n        # out2, pos2, _ = self._encode_image(img2, do_mask=False, return_all_blocks=False)\n        ## and now the faster version\n        out, pos, _ = self._encode_image(\n            torch.cat((img1, img2), dim=0),\n            do_mask=False,\n            return_all_blocks=return_all_blocks,\n        )\n        if return_all_blocks:","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream._set_decoder","uri":"program://Human3R/function/src.croco.models.croco_downstream._set_decoder#L53-L55","kind":"function","name":"_set_decoder","path":"src/croco/models/croco_downstream.py","language":"python","start_line":53,"end_line":55,"context_start_line":33,"context_end_line":75,"code":"        \"\"\"Build network for monocular downstream task, only using the encoder.\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        NOTE: It works by *calling super().__init__() but with redefined setters\n\n        \"\"\"\n        super(CroCoDownstreamMonocularEncoder, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_decoder(self, *args, **kwargs):\n        \"\"\"No decoder\"\"\"\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No 'prediction head' for downstream tasks.\"\"\"\n        return\n\n    def forward(self, img):\n        \"\"\"\n        img if of size batch_size x 3 x h x w\n        \"\"\"\n        B, C, H, W = img.size()\n        img_info = {\"height\": H, \"width\": W}\n        need_all_layers = (\n            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, _, _ = self._encode_image(\n            img, do_mask=False, return_all_blocks=need_all_layers\n        )\n        return self.head(out, img_info)\n\n","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream._set_prediction_head","uri":"program://Human3R/function/src.croco.models.croco_downstream._set_prediction_head#L97-L99","kind":"function","name":"_set_prediction_head","path":"src/croco/models/croco_downstream.py","language":"python","start_line":97,"end_line":99,"context_start_line":77,"context_end_line":119,"code":"\n    def __init__(self, head, **kwargs):\n        \"\"\"Build network for binocular downstream task\n        It takes an extra argument head, that is called with the features\n          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        \"\"\"\n        super(CroCoDownstreamBinocular, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head for downstream tasks, define your own head\"\"\"\n        return\n\n    def encode_image_pairs(self, img1, img2, return_all_blocks=False):\n        \"\"\"run encoder for a pair of images\n        it is actually ~5% faster to concatenate the images along the batch dimension\n         than to encode them separately\n        \"\"\"\n        ## the two commented lines below is the naive version with separate encoding\n        # out, pos, _ = self._encode_image(img1, do_mask=False, return_all_blocks=return_all_blocks)\n        # out2, pos2, _ = self._encode_image(img2, do_mask=False, return_all_blocks=False)\n        ## and now the faster version\n        out, pos, _ = self._encode_image(\n            torch.cat((img1, img2), dim=0),\n            do_mask=False,\n            return_all_blocks=return_all_blocks,\n        )\n        if return_all_blocks:\n            out, out2 = list(map(list, zip(*[o.chunk(2, dim=0) for o in out])))\n            out2 = out2[-1]\n        else:\n            out, out2 = out.chunk(2, dim=0)","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream.forward","uri":"program://Human3R/function/src.croco.models.croco_downstream.forward#L123-L141","kind":"function","name":"forward","path":"src/croco/models/croco_downstream.py","language":"python","start_line":123,"end_line":141,"context_start_line":103,"context_end_line":141,"code":"        it is actually ~5% faster to concatenate the images along the batch dimension\n         than to encode them separately\n        \"\"\"\n        ## the two commented lines below is the naive version with separate encoding\n        # out, pos, _ = self._encode_image(img1, do_mask=False, return_all_blocks=return_all_blocks)\n        # out2, pos2, _ = self._encode_image(img2, do_mask=False, return_all_blocks=False)\n        ## and now the faster version\n        out, pos, _ = self._encode_image(\n            torch.cat((img1, img2), dim=0),\n            do_mask=False,\n            return_all_blocks=return_all_blocks,\n        )\n        if return_all_blocks:\n            out, out2 = list(map(list, zip(*[o.chunk(2, dim=0) for o in out])))\n            out2 = out2[-1]\n        else:\n            out, out2 = out.chunk(2, dim=0)\n        pos, pos2 = pos.chunk(2, dim=0)\n        return out, out2, pos, pos2\n\n    def forward(self, img1, img2):\n        B, C, H, W = img1.size()\n        img_info = {\"height\": H, \"width\": W}\n        return_all_blocks = (\n            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, out2, pos, pos2 = self.encode_image_pairs(\n            img1, img2, return_all_blocks=return_all_blocks\n        )\n        if return_all_blocks:\n            decout = self._decoder(\n                out[-1], pos, None, out2, pos2, return_all_blocks=return_all_blocks\n            )\n            decout = out + decout\n        else:\n            decout = self._decoder(\n                out, pos, None, out2, pos2, return_all_blocks=return_all_blocks\n            )\n        return self.head(decout, img_info)","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco_downstream.encode_image_pairs","uri":"program://Human3R/function/src.croco.models.croco_downstream.encode_image_pairs#L101-L121","kind":"function","name":"encode_image_pairs","path":"src/croco/models/croco_downstream.py","language":"python","start_line":101,"end_line":121,"context_start_line":81,"context_end_line":141,"code":"          and a dictionary img_info containing 'width' and 'height' keys\n        The head is setup with the croconet arguments in this init function\n        \"\"\"\n        super(CroCoDownstreamBinocular, self).__init__(**kwargs)\n        head.setup(self)\n        self.head = head\n\n    def _set_mask_generator(self, *args, **kwargs):\n        \"\"\"No mask generator\"\"\"\n        return\n\n    def _set_mask_token(self, *args, **kwargs):\n        \"\"\"No mask token\"\"\"\n        self.mask_token = None\n        return\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head for downstream tasks, define your own head\"\"\"\n        return\n\n    def encode_image_pairs(self, img1, img2, return_all_blocks=False):\n        \"\"\"run encoder for a pair of images\n        it is actually ~5% faster to concatenate the images along the batch dimension\n         than to encode them separately\n        \"\"\"\n        ## the two commented lines below is the naive version with separate encoding\n        # out, pos, _ = self._encode_image(img1, do_mask=False, return_all_blocks=return_all_blocks)\n        # out2, pos2, _ = self._encode_image(img2, do_mask=False, return_all_blocks=False)\n        ## and now the faster version\n        out, pos, _ = self._encode_image(\n            torch.cat((img1, img2), dim=0),\n            do_mask=False,\n            return_all_blocks=return_all_blocks,\n        )\n        if return_all_blocks:\n            out, out2 = list(map(list, zip(*[o.chunk(2, dim=0) for o in out])))\n            out2 = out2[-1]\n        else:\n            out, out2 = out.chunk(2, dim=0)\n        pos, pos2 = pos.chunk(2, dim=0)\n        return out, out2, pos, pos2\n\n    def forward(self, img1, img2):\n        B, C, H, W = img1.size()\n        img_info = {\"height\": H, \"width\": W}\n        return_all_blocks = (\n            hasattr(self.head, \"return_all_blocks\") and self.head.return_all_blocks\n        )\n        out, out2, pos, pos2 = self.encode_image_pairs(\n            img1, img2, return_all_blocks=return_all_blocks\n        )\n        if return_all_blocks:\n            decout = self._decoder(\n                out[-1], pos, None, out2, pos2, return_all_blocks=return_all_blocks\n            )\n            decout = out + decout\n        else:\n            decout = self._decoder(\n                out, pos, None, out2, pos2, return_all_blocks=return_all_blocks\n            )\n        return self.head(decout, img_info)","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block","uri":"program://Human3R/module/src.croco.models.dpt_block#L1-L513","kind":"module","name":"src.croco.models.dpt_block","path":"src/croco/models/dpt_block.py","language":"python","start_line":1,"end_line":513,"context_start_line":1,"context_end_line":513,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# DPT head for ViTs\n# --------------------------------------------------------\n# References:\n# https://github.com/isl-org/DPT\n# https://github.com/EPFL-VILAB/MultiMAE/blob/main/multimae/output_adapters.py\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom einops import rearrange, repeat\nfrom typing import Union, Tuple, Iterable, List, Optional, Dict\n\n\ndef pair(t):\n    return t if isinstance(t, tuple) else (t, t)\n\n\ndef make_scratch(in_shape, out_shape, groups=1, expand=False):\n    scratch = nn.Module()\n\n    out_shape1 = out_shape\n    out_shape2 = out_shape\n    out_shape3 = out_shape\n    out_shape4 = out_shape\n    if expand == True:\n        out_shape1 = out_shape\n        out_shape2 = out_shape * 2\n        out_shape3 = out_shape * 4\n        out_shape4 = out_shape * 8\n\n    scratch.layer1_rn = nn.Conv2d(\n        in_shape[0],\n        out_shape1,\n        kernel_size=3,\n        stride=1,\n        padding=1,\n        bias=False,\n        groups=groups,\n    )\n    scratch.layer2_rn = nn.Conv2d(\n        in_shape[1],\n        out_shape2,\n        kernel_size=3,\n        stride=1,\n        padding=1,\n        bias=False,\n        groups=groups,\n    )\n    scratch.layer3_rn = nn.Conv2d(\n        in_shape[2],\n        out_shape3,\n        kernel_size=3,\n        stride=1,\n        padding=1,\n        bias=False,\n        groups=groups,\n    )\n    scratch.layer4_rn = nn.Conv2d(\n        in_shape[3],\n        out_shape4,\n        kernel_size=3,\n        stride=1,\n        padding=1,\n        bias=False,\n        groups=groups,\n    )\n\n    scratch.layer_rn = nn.ModuleList(\n        [\n            scratch.layer1_rn,\n            scratch.layer2_rn,\n            scratch.layer3_rn,\n            scratch.layer4_rn,\n        ]\n    )\n\n    return scratch\n\n\nclass ResidualConvUnit_custom(nn.Module):\n    \"\"\"Residual convolution module.\"\"\"\n\n    def __init__(self, features, activation, bn):\n        \"\"\"Init.\n        Args:\n            features (int): number of features\n        \"\"\"\n        super().__init__()\n\n        self.bn = bn\n\n        self.groups = 1\n\n        self.conv1 = nn.Conv2d(\n            features,\n            features,\n            kernel_size=3,\n            stride=1,\n            padding=1,\n            bias=not self.bn,\n            groups=self.groups,\n        )\n\n        self.conv2 = nn.Conv2d(\n            features,\n            features,\n            kernel_size=3,\n            stride=1,\n            padding=1,\n            bias=not self.bn,\n            groups=self.groups,\n        )\n\n        if self.bn == True:\n            self.bn1 = nn.BatchNorm2d(features)\n            self.bn2 = nn.BatchNorm2d(features)\n\n        self.activation = activation\n\n        self.skip_add = nn.quantized.FloatFunctional()\n\n    def forward(self, x):\n        \"\"\"Forward pass.\n        Args:\n            x (tensor): input\n        Returns:\n            tensor: output\n        \"\"\"\n\n        out = self.activation(x)\n        out = self.conv1(out)\n        if self.bn == True:\n            out = self.bn1(out)\n\n        out = self.activation(out)\n        out = self.conv2(out)\n        if self.bn == True:\n            out = self.bn2(out)\n\n        if self.groups > 1:\n            out = self.conv_merge(out)\n\n        return self.skip_add.add(out, x)\n\n\nclass FeatureFusionBlock_custom(nn.Module):\n    \"\"\"Feature fusion block.\"\"\"\n\n    def __init__(\n        self,\n        features,\n        activation,\n        deconv=False,\n        bn=False,\n        expand=False,\n        align_corners=True,\n        width_ratio=1,\n    ):\n        \"\"\"Init.\n        Args:\n            features (int): number of features\n        \"\"\"\n        super(FeatureFusionBlock_custom, self).__init__()\n        self.width_ratio = width_ratio\n\n        self.deconv = deconv\n        self.align_corners = align_corners\n\n        self.groups = 1\n\n        self.expand = expand\n        out_features = features\n        if self.expand == True:\n            out_features = features // 2\n\n        self.out_conv = nn.Conv2d(\n            features,\n            out_features,\n            kernel_size=1,\n            stride=1,\n            padding=0,\n            bias=True,\n            groups=1,\n        )\n\n        self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn)\n        self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn)\n\n        self.skip_add = nn.quantized.FloatFunctional()\n\n    def forward(self, *xs):\n        \"\"\"Forward pass.\n        Returns:\n            tensor: output\n        \"\"\"\n        output = xs[0]\n\n        if len(xs) == 2:\n            res = self.resConfUnit1(xs[1])\n            if self.width_ratio != 1:\n                res = F.interpolate(\n                    res, size=(output.shape[2], output.shape[3]), mode=\"bilinear\"\n                )\n\n            output = self.skip_add.add(output, res)\n            # output += res\n\n        output = self.resConfUnit2(output)\n\n        if self.width_ratio != 1:\n            # and output.shape[3] < self.width_ratio * output.shape[2]\n            # size=(image.shape[])\n            if (output.shape[3] / output.shape[2]) < (2 / 3) * self.width_ratio:\n                shape = 3 * output.shape[3]\n            else:\n                shape = int(self.width_ratio * 2 * output.shape[2])\n            output = F.interpolate(\n                output, size=(2 * output.shape[2], shape), mode=\"bilinear\"\n            )\n        else:\n            output = nn.functional.interpolate(\n                output,\n                scale_factor=2,\n                mode=\"bilinear\",\n                align_corners=self.align_corners,\n            )\n        output = self.out_conv(output)\n        return output\n\n\ndef make_fusion_block(features, use_bn, width_ratio=1):\n    return FeatureFusionBlock_custom(\n        features,\n        nn.ReLU(False),\n        deconv=False,\n        bn=use_bn,\n        expand=False,\n        align_corners=True,\n        width_ratio=width_ratio,\n    )\n\n\nclass Interpolate(nn.Module):\n    \"\"\"Interpolation module.\"\"\"\n\n    def __init__(self, scale_factor, mode, align_corners=False):\n        \"\"\"Init.\n        Args:\n            scale_factor (float): scaling\n            mode (str): interpolation mode\n        \"\"\"\n        super(Interpolate, self).__init__()\n\n        self.interp = nn.functional.interpolate\n        self.scale_factor = scale_factor\n        self.mode = mode\n        self.align_corners = align_corners\n\n    def forward(self, x):\n        \"\"\"Forward pass.\n        Args:\n            x (tensor): input\n        Returns:\n            tensor: interpolated data\n        \"\"\"\n\n        x = self.interp(\n            x,\n            scale_factor=self.scale_factor,\n            mode=self.mode,\n            align_corners=self.align_corners,\n        )\n\n        return x\n\n\nclass DPTOutputAdapter(nn.Module):\n    \"\"\"DPT output adapter.\n\n    :param num_cahnnels: Number of output channels\n    :param stride_level: tride level compared to the full-sized image.\n        E.g. 4 for 1/4th the size of the image.\n    :param patch_size_full: Int or tuple of the patch size over the full image size.\n        Patch size for smaller inputs will be computed accordingly.\n    :param hooks: Index of intermediate layers\n    :param layer_dims: Dimension of intermediate layers\n    :param feature_dim: Feature dimension\n    :param last_dim: out_channels/in_channels for the last two Conv2d when head_type == regression\n    :param use_bn: If set to True, activates batch norm\n    :param dim_tokens_enc:  Dimension of tokens coming from encoder\n    \"\"\"\n\n    def __init__(\n        self,\n        num_channels: int = 1,\n        stride_level: int = 1,\n        patch_size: Union[int, Tuple[int, int]] = 16,\n        main_tasks: Iterable[str] = (\"rgb\",),\n        hooks: List[int] = [2, 5, 8, 11],\n        layer_dims: List[int] = [96, 192, 384, 768],\n        feature_dim: int = 256,\n        last_dim: int = 32,\n        use_bn: bool = False,\n        dim_tokens_enc: Optional[int] = None,\n        head_type: str = \"regression\",\n        output_width_ratio=1,\n        **kwargs\n    ):\n        super().__init__()\n        self.num_channels = num_channels\n        self.stride_level = stride_level\n        self.patch_size = pair(patch_size)\n        self.main_tasks = main_tasks\n        self.hooks = hooks\n        self.layer_dims = layer_dims\n        self.feature_dim = feature_dim\n        self.dim_tokens_enc = (\n            dim_tokens_enc * len(self.main_tasks)\n            if dim_tokens_enc is not None\n            else None\n        )\n        self.head_type = head_type\n\n        # Actual patch height and width, taking into account stride of input\n        self.P_H = max(1, self.patch_size[0] // stride_level)\n        self.P_W = max(1, self.patch_size[1] // stride_level)\n\n        self.scratch = make_scratch(layer_dims, feature_dim, groups=1, expand=False)\n\n        self.scratch.refinenet1 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n        self.scratch.refinenet2 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n        self.scratch.refinenet3 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n        self.scratch.refinenet4 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n\n        if self.head_type == \"regression\":\n            # The \"DPTDepthModel\" head\n            self.head = nn.Sequential(\n                nn.Conv2d(\n                    feature_dim, feature_dim // 2, kernel_size=3, stride=1, padding=1\n                ),\n                Interpolate(scale_factor=2, mode=\"bilinear\", align_corners=True),\n                nn.Conv2d(\n                    feature_dim // 2, last_dim, kernel_size=3, stride=1, padding=1\n                ),\n                nn.ReLU(True),\n                nn.Conv2d(\n                    last_dim, self.num_channels, kernel_size=1, stride=1, padding=0\n                ),\n            )\n        elif self.head_type == \"semseg\":\n            # The \"DPTSegmentationModel\" head\n            self.head = nn.Sequential(\n                nn.Conv2d(\n                    feature_dim, feature_dim, kernel_size=3, padding=1, bias=False\n                ),\n                nn.BatchNorm2d(feature_dim) if use_bn else nn.Identity(),\n                nn.ReLU(True),\n                nn.Dropout(0.1, False),\n                nn.Conv2d(feature_dim, self.num_channels, kernel_size=1),\n                Interpolate(scale_factor=2, mode=\"bilinear\", align_corners=True),\n            )\n        else:\n            raise ValueError('DPT head_type must be \"regression\" or \"semseg\".')\n\n        if self.dim_tokens_enc is not None:\n            self.init(dim_tokens_enc=dim_tokens_enc)\n\n    def init(self, dim_tokens_enc=768):\n        \"\"\"\n        Initialize parts of decoder that are dependent on dimension of encoder tokens.\n        Should be called when setting up MultiMAE.\n\n        :param dim_tokens_enc: Dimension of tokens coming from encoder\n        \"\"\"\n        # print(dim_tokens_enc)\n\n        # Set up activation postprocessing layers\n        if isinstance(dim_tokens_enc, int):\n            dim_tokens_enc = 4 * [dim_tokens_enc]\n\n        self.dim_tokens_enc = [dt * len(self.main_tasks) for dt in dim_tokens_enc]\n\n        self.act_1_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[0],\n                out_channels=self.layer_dims[0],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            ),\n            nn.ConvTranspose2d(\n                in_channels=self.layer_dims[0],\n                out_channels=self.layer_dims[0],\n                kernel_size=4,\n                stride=4,\n                padding=0,\n                bias=True,\n                dilation=1,\n                groups=1,\n            ),\n        )\n\n        self.act_2_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[1],\n                out_channels=self.layer_dims[1],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            ),\n            nn.ConvTranspose2d(\n                in_channels=self.layer_dims[1],\n                out_channels=self.layer_dims[1],\n                kernel_size=2,\n                stride=2,\n                padding=0,\n                bias=True,\n                dilation=1,\n                groups=1,\n            ),\n        )\n\n        self.act_3_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[2],\n                out_channels=self.layer_dims[2],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            )\n        )\n\n        self.act_4_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[3],\n                out_channels=self.layer_dims[3],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            ),\n            nn.Conv2d(\n                in_channels=self.layer_dims[3],\n                out_channels=self.layer_dims[3],\n                kernel_size=3,\n                stride=2,\n                padding=1,\n            ),\n        )\n\n        self.act_postprocess = nn.ModuleList(\n            [\n                self.act_1_postprocess,\n                self.act_2_postprocess,\n                self.act_3_postprocess,\n                self.act_4_postprocess,\n            ]\n        )\n\n    def adapt_tokens(self, encoder_tokens):\n        # Adapt tokens\n        x = []\n        x.append(encoder_tokens[:, :])\n        x = torch.cat(x, dim=-1)\n        return x\n\n    def forward(self, encoder_tokens: List[torch.Tensor], image_size):\n        # input_info: Dict):\n        assert (\n            self.dim_tokens_enc is not None\n        ), \"Need to call init(dim_tokens_enc) function first\"\n        H, W = image_size\n\n        # Number of patches in height and width\n        N_H = H // (self.stride_level * self.P_H)\n        N_W = W // (self.stride_level * self.P_W)\n\n        # Hook decoder onto 4 layers from specified ViT layers\n        layers = [encoder_tokens[hook] for hook in self.hooks]\n\n        # Extract only task-relevant tokens and ignore global tokens.\n        layers = [self.adapt_tokens(l) for l in layers]\n\n        # Reshape tokens to spatial representation\n        layers = [\n            rearrange(l, \"b (nh nw) c -> b c nh nw\", nh=N_H, nw=N_W) for l in layers\n        ]\n\n        layers = [self.act_postprocess[idx](l) for idx, l in enumerate(layers)]\n        # Project layers to chosen feature dim\n        layers = [self.scratch.layer_rn[idx](l) for idx, l in enumerate(layers)]\n\n        # Fuse layers using refinement stages\n        path_4 = self.scratch.refinenet4(layers[3])\n        path_3 = self.scratch.refinenet3(path_4, layers[2])\n        path_2 = self.scratch.refinenet2(path_3, layers[1])\n        path_1 = self.scratch.refinenet1(path_2, layers[0])\n\n        # Output head\n        out = self.head(path_1)\n\n        return out","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.pair","uri":"program://Human3R/function/src.croco.models.dpt_block.pair#L18-L19","kind":"function","name":"pair","path":"src/croco/models/dpt_block.py","language":"python","start_line":18,"end_line":19,"context_start_line":1,"context_end_line":39,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# DPT head for ViTs\n# --------------------------------------------------------\n# References:\n# https://github.com/isl-org/DPT\n# https://github.com/EPFL-VILAB/MultiMAE/blob/main/multimae/output_adapters.py\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom einops import rearrange, repeat\nfrom typing import Union, Tuple, Iterable, List, Optional, Dict\n\n\ndef pair(t):\n    return t if isinstance(t, tuple) else (t, t)\n\n\ndef make_scratch(in_shape, out_shape, groups=1, expand=False):\n    scratch = nn.Module()\n\n    out_shape1 = out_shape\n    out_shape2 = out_shape\n    out_shape3 = out_shape\n    out_shape4 = out_shape\n    if expand == True:\n        out_shape1 = out_shape\n        out_shape2 = out_shape * 2\n        out_shape3 = out_shape * 4\n        out_shape4 = out_shape * 8\n\n    scratch.layer1_rn = nn.Conv2d(\n        in_shape[0],\n        out_shape1,\n        kernel_size=3,\n        stride=1,","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.make_scratch","uri":"program://Human3R/function/src.croco.models.dpt_block.make_scratch#L22-L81","kind":"function","name":"make_scratch","path":"src/croco/models/dpt_block.py","language":"python","start_line":22,"end_line":81,"context_start_line":2,"context_end_line":101,"code":"# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# DPT head for ViTs\n# --------------------------------------------------------\n# References:\n# https://github.com/isl-org/DPT\n# https://github.com/EPFL-VILAB/MultiMAE/blob/main/multimae/output_adapters.py\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom einops import rearrange, repeat\nfrom typing import Union, Tuple, Iterable, List, Optional, Dict\n\n\ndef pair(t):\n    return t if isinstance(t, tuple) else (t, t)\n\n\ndef make_scratch(in_shape, out_shape, groups=1, expand=False):\n    scratch = nn.Module()\n\n    out_shape1 = out_shape\n    out_shape2 = out_shape\n    out_shape3 = out_shape\n    out_shape4 = out_shape\n    if expand == True:\n        out_shape1 = out_shape\n        out_shape2 = out_shape * 2\n        out_shape3 = out_shape * 4\n        out_shape4 = out_shape * 8\n\n    scratch.layer1_rn = nn.Conv2d(\n        in_shape[0],\n        out_shape1,\n        kernel_size=3,\n        stride=1,\n        padding=1,\n        bias=False,\n        groups=groups,\n    )\n    scratch.layer2_rn = nn.Conv2d(\n        in_shape[1],\n        out_shape2,\n        kernel_size=3,\n        stride=1,\n        padding=1,\n        bias=False,\n        groups=groups,\n    )\n    scratch.layer3_rn = nn.Conv2d(\n        in_shape[2],\n        out_shape3,\n        kernel_size=3,\n        stride=1,\n        padding=1,\n        bias=False,\n        groups=groups,\n    )\n    scratch.layer4_rn = nn.Conv2d(\n        in_shape[3],\n        out_shape4,\n        kernel_size=3,\n        stride=1,\n        padding=1,\n        bias=False,\n        groups=groups,\n    )\n\n    scratch.layer_rn = nn.ModuleList(\n        [\n            scratch.layer1_rn,\n            scratch.layer2_rn,\n            scratch.layer3_rn,\n            scratch.layer4_rn,\n        ]\n    )\n\n    return scratch\n\n\nclass ResidualConvUnit_custom(nn.Module):\n    \"\"\"Residual convolution module.\"\"\"\n\n    def __init__(self, features, activation, bn):\n        \"\"\"Init.\n        Args:\n            features (int): number of features\n        \"\"\"\n        super().__init__()\n\n        self.bn = bn\n\n        self.groups = 1\n\n        self.conv1 = nn.Conv2d(\n            features,\n            features,\n            kernel_size=3,","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.ResidualConvUnit_custom","uri":"program://Human3R/class/src.croco.models.dpt_block.ResidualConvUnit_custom#L84-L147","kind":"class","name":"ResidualConvUnit_custom","path":"src/croco/models/dpt_block.py","language":"python","start_line":84,"end_line":147,"context_start_line":64,"context_end_line":167,"code":"        out_shape4,\n        kernel_size=3,\n        stride=1,\n        padding=1,\n        bias=False,\n        groups=groups,\n    )\n\n    scratch.layer_rn = nn.ModuleList(\n        [\n            scratch.layer1_rn,\n            scratch.layer2_rn,\n            scratch.layer3_rn,\n            scratch.layer4_rn,\n        ]\n    )\n\n    return scratch\n\n\nclass ResidualConvUnit_custom(nn.Module):\n    \"\"\"Residual convolution module.\"\"\"\n\n    def __init__(self, features, activation, bn):\n        \"\"\"Init.\n        Args:\n            features (int): number of features\n        \"\"\"\n        super().__init__()\n\n        self.bn = bn\n\n        self.groups = 1\n\n        self.conv1 = nn.Conv2d(\n            features,\n            features,\n            kernel_size=3,\n            stride=1,\n            padding=1,\n            bias=not self.bn,\n            groups=self.groups,\n        )\n\n        self.conv2 = nn.Conv2d(\n            features,\n            features,\n            kernel_size=3,\n            stride=1,\n            padding=1,\n            bias=not self.bn,\n            groups=self.groups,\n        )\n\n        if self.bn == True:\n            self.bn1 = nn.BatchNorm2d(features)\n            self.bn2 = nn.BatchNorm2d(features)\n\n        self.activation = activation\n\n        self.skip_add = nn.quantized.FloatFunctional()\n\n    def forward(self, x):\n        \"\"\"Forward pass.\n        Args:\n            x (tensor): input\n        Returns:\n            tensor: output\n        \"\"\"\n\n        out = self.activation(x)\n        out = self.conv1(out)\n        if self.bn == True:\n            out = self.bn1(out)\n\n        out = self.activation(out)\n        out = self.conv2(out)\n        if self.bn == True:\n            out = self.bn2(out)\n\n        if self.groups > 1:\n            out = self.conv_merge(out)\n\n        return self.skip_add.add(out, x)\n\n\nclass FeatureFusionBlock_custom(nn.Module):\n    \"\"\"Feature fusion block.\"\"\"\n\n    def __init__(\n        self,\n        features,\n        activation,\n        deconv=False,\n        bn=False,\n        expand=False,\n        align_corners=True,\n        width_ratio=1,\n    ):\n        \"\"\"Init.\n        Args:\n            features (int): number of features\n        \"\"\"\n        super(FeatureFusionBlock_custom, self).__init__()","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.FeatureFusionBlock_custom","uri":"program://Human3R/class/src.croco.models.dpt_block.FeatureFusionBlock_custom#L150-L232","kind":"class","name":"FeatureFusionBlock_custom","path":"src/croco/models/dpt_block.py","language":"python","start_line":150,"end_line":232,"context_start_line":130,"context_end_line":252,"code":"        Returns:\n            tensor: output\n        \"\"\"\n\n        out = self.activation(x)\n        out = self.conv1(out)\n        if self.bn == True:\n            out = self.bn1(out)\n\n        out = self.activation(out)\n        out = self.conv2(out)\n        if self.bn == True:\n            out = self.bn2(out)\n\n        if self.groups > 1:\n            out = self.conv_merge(out)\n\n        return self.skip_add.add(out, x)\n\n\nclass FeatureFusionBlock_custom(nn.Module):\n    \"\"\"Feature fusion block.\"\"\"\n\n    def __init__(\n        self,\n        features,\n        activation,\n        deconv=False,\n        bn=False,\n        expand=False,\n        align_corners=True,\n        width_ratio=1,\n    ):\n        \"\"\"Init.\n        Args:\n            features (int): number of features\n        \"\"\"\n        super(FeatureFusionBlock_custom, self).__init__()\n        self.width_ratio = width_ratio\n\n        self.deconv = deconv\n        self.align_corners = align_corners\n\n        self.groups = 1\n\n        self.expand = expand\n        out_features = features\n        if self.expand == True:\n            out_features = features // 2\n\n        self.out_conv = nn.Conv2d(\n            features,\n            out_features,\n            kernel_size=1,\n            stride=1,\n            padding=0,\n            bias=True,\n            groups=1,\n        )\n\n        self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn)\n        self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn)\n\n        self.skip_add = nn.quantized.FloatFunctional()\n\n    def forward(self, *xs):\n        \"\"\"Forward pass.\n        Returns:\n            tensor: output\n        \"\"\"\n        output = xs[0]\n\n        if len(xs) == 2:\n            res = self.resConfUnit1(xs[1])\n            if self.width_ratio != 1:\n                res = F.interpolate(\n                    res, size=(output.shape[2], output.shape[3]), mode=\"bilinear\"\n                )\n\n            output = self.skip_add.add(output, res)\n            # output += res\n\n        output = self.resConfUnit2(output)\n\n        if self.width_ratio != 1:\n            # and output.shape[3] < self.width_ratio * output.shape[2]\n            # size=(image.shape[])\n            if (output.shape[3] / output.shape[2]) < (2 / 3) * self.width_ratio:\n                shape = 3 * output.shape[3]\n            else:\n                shape = int(self.width_ratio * 2 * output.shape[2])\n            output = F.interpolate(\n                output, size=(2 * output.shape[2], shape), mode=\"bilinear\"\n            )\n        else:\n            output = nn.functional.interpolate(\n                output,\n                scale_factor=2,\n                mode=\"bilinear\",\n                align_corners=self.align_corners,\n            )\n        output = self.out_conv(output)\n        return output\n\n\ndef make_fusion_block(features, use_bn, width_ratio=1):\n    return FeatureFusionBlock_custom(\n        features,\n        nn.ReLU(False),\n        deconv=False,\n        bn=use_bn,\n        expand=False,\n        align_corners=True,\n        width_ratio=width_ratio,\n    )\n\n\nclass Interpolate(nn.Module):\n    \"\"\"Interpolation module.\"\"\"\n\n    def __init__(self, scale_factor, mode, align_corners=False):\n        \"\"\"Init.\n        Args:","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.make_fusion_block","uri":"program://Human3R/function/src.croco.models.dpt_block.make_fusion_block#L235-L244","kind":"function","name":"make_fusion_block","path":"src/croco/models/dpt_block.py","language":"python","start_line":235,"end_line":244,"context_start_line":215,"context_end_line":264,"code":"            # and output.shape[3] < self.width_ratio * output.shape[2]\n            # size=(image.shape[])\n            if (output.shape[3] / output.shape[2]) < (2 / 3) * self.width_ratio:\n                shape = 3 * output.shape[3]\n            else:\n                shape = int(self.width_ratio * 2 * output.shape[2])\n            output = F.interpolate(\n                output, size=(2 * output.shape[2], shape), mode=\"bilinear\"\n            )\n        else:\n            output = nn.functional.interpolate(\n                output,\n                scale_factor=2,\n                mode=\"bilinear\",\n                align_corners=self.align_corners,\n            )\n        output = self.out_conv(output)\n        return output\n\n\ndef make_fusion_block(features, use_bn, width_ratio=1):\n    return FeatureFusionBlock_custom(\n        features,\n        nn.ReLU(False),\n        deconv=False,\n        bn=use_bn,\n        expand=False,\n        align_corners=True,\n        width_ratio=width_ratio,\n    )\n\n\nclass Interpolate(nn.Module):\n    \"\"\"Interpolation module.\"\"\"\n\n    def __init__(self, scale_factor, mode, align_corners=False):\n        \"\"\"Init.\n        Args:\n            scale_factor (float): scaling\n            mode (str): interpolation mode\n        \"\"\"\n        super(Interpolate, self).__init__()\n\n        self.interp = nn.functional.interpolate\n        self.scale_factor = scale_factor\n        self.mode = mode\n        self.align_corners = align_corners\n\n    def forward(self, x):\n        \"\"\"Forward pass.","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.Interpolate","uri":"program://Human3R/class/src.croco.models.dpt_block.Interpolate#L247-L278","kind":"class","name":"Interpolate","path":"src/croco/models/dpt_block.py","language":"python","start_line":247,"end_line":278,"context_start_line":227,"context_end_line":298,"code":"                scale_factor=2,\n                mode=\"bilinear\",\n                align_corners=self.align_corners,\n            )\n        output = self.out_conv(output)\n        return output\n\n\ndef make_fusion_block(features, use_bn, width_ratio=1):\n    return FeatureFusionBlock_custom(\n        features,\n        nn.ReLU(False),\n        deconv=False,\n        bn=use_bn,\n        expand=False,\n        align_corners=True,\n        width_ratio=width_ratio,\n    )\n\n\nclass Interpolate(nn.Module):\n    \"\"\"Interpolation module.\"\"\"\n\n    def __init__(self, scale_factor, mode, align_corners=False):\n        \"\"\"Init.\n        Args:\n            scale_factor (float): scaling\n            mode (str): interpolation mode\n        \"\"\"\n        super(Interpolate, self).__init__()\n\n        self.interp = nn.functional.interpolate\n        self.scale_factor = scale_factor\n        self.mode = mode\n        self.align_corners = align_corners\n\n    def forward(self, x):\n        \"\"\"Forward pass.\n        Args:\n            x (tensor): input\n        Returns:\n            tensor: interpolated data\n        \"\"\"\n\n        x = self.interp(\n            x,\n            scale_factor=self.scale_factor,\n            mode=self.mode,\n            align_corners=self.align_corners,\n        )\n\n        return x\n\n\nclass DPTOutputAdapter(nn.Module):\n    \"\"\"DPT output adapter.\n\n    :param num_cahnnels: Number of output channels\n    :param stride_level: tride level compared to the full-sized image.\n        E.g. 4 for 1/4th the size of the image.\n    :param patch_size_full: Int or tuple of the patch size over the full image size.\n        Patch size for smaller inputs will be computed accordingly.\n    :param hooks: Index of intermediate layers\n    :param layer_dims: Dimension of intermediate layers\n    :param feature_dim: Feature dimension\n    :param last_dim: out_channels/in_channels for the last two Conv2d when head_type == regression\n    :param use_bn: If set to True, activates batch norm\n    :param dim_tokens_enc:  Dimension of tokens coming from encoder\n    \"\"\"\n\n    def __init__(\n        self,","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.DPTOutputAdapter","uri":"program://Human3R/class/src.croco.models.dpt_block.DPTOutputAdapter#L281-L513","kind":"class","name":"DPTOutputAdapter","path":"src/croco/models/dpt_block.py","language":"python","start_line":281,"end_line":513,"context_start_line":261,"context_end_line":513,"code":"        self.align_corners = align_corners\n\n    def forward(self, x):\n        \"\"\"Forward pass.\n        Args:\n            x (tensor): input\n        Returns:\n            tensor: interpolated data\n        \"\"\"\n\n        x = self.interp(\n            x,\n            scale_factor=self.scale_factor,\n            mode=self.mode,\n            align_corners=self.align_corners,\n        )\n\n        return x\n\n\nclass DPTOutputAdapter(nn.Module):\n    \"\"\"DPT output adapter.\n\n    :param num_cahnnels: Number of output channels\n    :param stride_level: tride level compared to the full-sized image.\n        E.g. 4 for 1/4th the size of the image.\n    :param patch_size_full: Int or tuple of the patch size over the full image size.\n        Patch size for smaller inputs will be computed accordingly.\n    :param hooks: Index of intermediate layers\n    :param layer_dims: Dimension of intermediate layers\n    :param feature_dim: Feature dimension\n    :param last_dim: out_channels/in_channels for the last two Conv2d when head_type == regression\n    :param use_bn: If set to True, activates batch norm\n    :param dim_tokens_enc:  Dimension of tokens coming from encoder\n    \"\"\"\n\n    def __init__(\n        self,\n        num_channels: int = 1,\n        stride_level: int = 1,\n        patch_size: Union[int, Tuple[int, int]] = 16,\n        main_tasks: Iterable[str] = (\"rgb\",),\n        hooks: List[int] = [2, 5, 8, 11],\n        layer_dims: List[int] = [96, 192, 384, 768],\n        feature_dim: int = 256,\n        last_dim: int = 32,\n        use_bn: bool = False,\n        dim_tokens_enc: Optional[int] = None,\n        head_type: str = \"regression\",\n        output_width_ratio=1,\n        **kwargs\n    ):\n        super().__init__()\n        self.num_channels = num_channels\n        self.stride_level = stride_level\n        self.patch_size = pair(patch_size)\n        self.main_tasks = main_tasks\n        self.hooks = hooks\n        self.layer_dims = layer_dims\n        self.feature_dim = feature_dim\n        self.dim_tokens_enc = (\n            dim_tokens_enc * len(self.main_tasks)\n            if dim_tokens_enc is not None\n            else None\n        )\n        self.head_type = head_type\n\n        # Actual patch height and width, taking into account stride of input\n        self.P_H = max(1, self.patch_size[0] // stride_level)\n        self.P_W = max(1, self.patch_size[1] // stride_level)\n\n        self.scratch = make_scratch(layer_dims, feature_dim, groups=1, expand=False)\n\n        self.scratch.refinenet1 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n        self.scratch.refinenet2 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n        self.scratch.refinenet3 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n        self.scratch.refinenet4 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n\n        if self.head_type == \"regression\":\n            # The \"DPTDepthModel\" head\n            self.head = nn.Sequential(\n                nn.Conv2d(\n                    feature_dim, feature_dim // 2, kernel_size=3, stride=1, padding=1\n                ),\n                Interpolate(scale_factor=2, mode=\"bilinear\", align_corners=True),\n                nn.Conv2d(\n                    feature_dim // 2, last_dim, kernel_size=3, stride=1, padding=1\n                ),\n                nn.ReLU(True),\n                nn.Conv2d(\n                    last_dim, self.num_channels, kernel_size=1, stride=1, padding=0\n                ),\n            )\n        elif self.head_type == \"semseg\":\n            # The \"DPTSegmentationModel\" head\n            self.head = nn.Sequential(\n                nn.Conv2d(\n                    feature_dim, feature_dim, kernel_size=3, padding=1, bias=False\n                ),\n                nn.BatchNorm2d(feature_dim) if use_bn else nn.Identity(),\n                nn.ReLU(True),\n                nn.Dropout(0.1, False),\n                nn.Conv2d(feature_dim, self.num_channels, kernel_size=1),\n                Interpolate(scale_factor=2, mode=\"bilinear\", align_corners=True),\n            )\n        else:\n            raise ValueError('DPT head_type must be \"regression\" or \"semseg\".')\n\n        if self.dim_tokens_enc is not None:\n            self.init(dim_tokens_enc=dim_tokens_enc)\n\n    def init(self, dim_tokens_enc=768):\n        \"\"\"\n        Initialize parts of decoder that are dependent on dimension of encoder tokens.\n        Should be called when setting up MultiMAE.\n\n        :param dim_tokens_enc: Dimension of tokens coming from encoder\n        \"\"\"\n        # print(dim_tokens_enc)\n\n        # Set up activation postprocessing layers\n        if isinstance(dim_tokens_enc, int):\n            dim_tokens_enc = 4 * [dim_tokens_enc]\n\n        self.dim_tokens_enc = [dt * len(self.main_tasks) for dt in dim_tokens_enc]\n\n        self.act_1_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[0],\n                out_channels=self.layer_dims[0],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            ),\n            nn.ConvTranspose2d(\n                in_channels=self.layer_dims[0],\n                out_channels=self.layer_dims[0],\n                kernel_size=4,\n                stride=4,\n                padding=0,\n                bias=True,\n                dilation=1,\n                groups=1,\n            ),\n        )\n\n        self.act_2_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[1],\n                out_channels=self.layer_dims[1],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            ),\n            nn.ConvTranspose2d(\n                in_channels=self.layer_dims[1],\n                out_channels=self.layer_dims[1],\n                kernel_size=2,\n                stride=2,\n                padding=0,\n                bias=True,\n                dilation=1,\n                groups=1,\n            ),\n        )\n\n        self.act_3_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[2],\n                out_channels=self.layer_dims[2],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            )\n        )\n\n        self.act_4_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[3],\n                out_channels=self.layer_dims[3],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            ),\n            nn.Conv2d(\n                in_channels=self.layer_dims[3],\n                out_channels=self.layer_dims[3],\n                kernel_size=3,\n                stride=2,\n                padding=1,\n            ),\n        )\n\n        self.act_postprocess = nn.ModuleList(\n            [\n                self.act_1_postprocess,\n                self.act_2_postprocess,\n                self.act_3_postprocess,\n                self.act_4_postprocess,\n            ]\n        )\n\n    def adapt_tokens(self, encoder_tokens):\n        # Adapt tokens\n        x = []\n        x.append(encoder_tokens[:, :])\n        x = torch.cat(x, dim=-1)\n        return x\n\n    def forward(self, encoder_tokens: List[torch.Tensor], image_size):\n        # input_info: Dict):\n        assert (\n            self.dim_tokens_enc is not None\n        ), \"Need to call init(dim_tokens_enc) function first\"\n        H, W = image_size\n\n        # Number of patches in height and width\n        N_H = H // (self.stride_level * self.P_H)\n        N_W = W // (self.stride_level * self.P_W)\n\n        # Hook decoder onto 4 layers from specified ViT layers\n        layers = [encoder_tokens[hook] for hook in self.hooks]\n\n        # Extract only task-relevant tokens and ignore global tokens.\n        layers = [self.adapt_tokens(l) for l in layers]\n\n        # Reshape tokens to spatial representation\n        layers = [\n            rearrange(l, \"b (nh nw) c -> b c nh nw\", nh=N_H, nw=N_W) for l in layers\n        ]\n\n        layers = [self.act_postprocess[idx](l) for idx, l in enumerate(layers)]\n        # Project layers to chosen feature dim\n        layers = [self.scratch.layer_rn[idx](l) for idx, l in enumerate(layers)]\n\n        # Fuse layers using refinement stages\n        path_4 = self.scratch.refinenet4(layers[3])\n        path_3 = self.scratch.refinenet3(path_4, layers[2])\n        path_2 = self.scratch.refinenet2(path_3, layers[1])\n        path_1 = self.scratch.refinenet1(path_2, layers[0])\n\n        # Output head\n        out = self.head(path_1)\n\n        return out","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.__init__","uri":"program://Human3R/function/src.croco.models.dpt_block.__init__#L297-L378","kind":"function","name":"__init__","path":"src/croco/models/dpt_block.py","language":"python","start_line":297,"end_line":378,"context_start_line":277,"context_end_line":398,"code":"\n        return x\n\n\nclass DPTOutputAdapter(nn.Module):\n    \"\"\"DPT output adapter.\n\n    :param num_cahnnels: Number of output channels\n    :param stride_level: tride level compared to the full-sized image.\n        E.g. 4 for 1/4th the size of the image.\n    :param patch_size_full: Int or tuple of the patch size over the full image size.\n        Patch size for smaller inputs will be computed accordingly.\n    :param hooks: Index of intermediate layers\n    :param layer_dims: Dimension of intermediate layers\n    :param feature_dim: Feature dimension\n    :param last_dim: out_channels/in_channels for the last two Conv2d when head_type == regression\n    :param use_bn: If set to True, activates batch norm\n    :param dim_tokens_enc:  Dimension of tokens coming from encoder\n    \"\"\"\n\n    def __init__(\n        self,\n        num_channels: int = 1,\n        stride_level: int = 1,\n        patch_size: Union[int, Tuple[int, int]] = 16,\n        main_tasks: Iterable[str] = (\"rgb\",),\n        hooks: List[int] = [2, 5, 8, 11],\n        layer_dims: List[int] = [96, 192, 384, 768],\n        feature_dim: int = 256,\n        last_dim: int = 32,\n        use_bn: bool = False,\n        dim_tokens_enc: Optional[int] = None,\n        head_type: str = \"regression\",\n        output_width_ratio=1,\n        **kwargs\n    ):\n        super().__init__()\n        self.num_channels = num_channels\n        self.stride_level = stride_level\n        self.patch_size = pair(patch_size)\n        self.main_tasks = main_tasks\n        self.hooks = hooks\n        self.layer_dims = layer_dims\n        self.feature_dim = feature_dim\n        self.dim_tokens_enc = (\n            dim_tokens_enc * len(self.main_tasks)\n            if dim_tokens_enc is not None\n            else None\n        )\n        self.head_type = head_type\n\n        # Actual patch height and width, taking into account stride of input\n        self.P_H = max(1, self.patch_size[0] // stride_level)\n        self.P_W = max(1, self.patch_size[1] // stride_level)\n\n        self.scratch = make_scratch(layer_dims, feature_dim, groups=1, expand=False)\n\n        self.scratch.refinenet1 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n        self.scratch.refinenet2 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n        self.scratch.refinenet3 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n        self.scratch.refinenet4 = make_fusion_block(\n            feature_dim, use_bn, output_width_ratio\n        )\n\n        if self.head_type == \"regression\":\n            # The \"DPTDepthModel\" head\n            self.head = nn.Sequential(\n                nn.Conv2d(\n                    feature_dim, feature_dim // 2, kernel_size=3, stride=1, padding=1\n                ),\n                Interpolate(scale_factor=2, mode=\"bilinear\", align_corners=True),\n                nn.Conv2d(\n                    feature_dim // 2, last_dim, kernel_size=3, stride=1, padding=1\n                ),\n                nn.ReLU(True),\n                nn.Conv2d(\n                    last_dim, self.num_channels, kernel_size=1, stride=1, padding=0\n                ),\n            )\n        elif self.head_type == \"semseg\":\n            # The \"DPTSegmentationModel\" head\n            self.head = nn.Sequential(\n                nn.Conv2d(\n                    feature_dim, feature_dim, kernel_size=3, padding=1, bias=False\n                ),\n                nn.BatchNorm2d(feature_dim) if use_bn else nn.Identity(),\n                nn.ReLU(True),\n                nn.Dropout(0.1, False),\n                nn.Conv2d(feature_dim, self.num_channels, kernel_size=1),\n                Interpolate(scale_factor=2, mode=\"bilinear\", align_corners=True),\n            )\n        else:\n            raise ValueError('DPT head_type must be \"regression\" or \"semseg\".')\n\n        if self.dim_tokens_enc is not None:\n            self.init(dim_tokens_enc=dim_tokens_enc)\n\n    def init(self, dim_tokens_enc=768):\n        \"\"\"\n        Initialize parts of decoder that are dependent on dimension of encoder tokens.\n        Should be called when setting up MultiMAE.\n\n        :param dim_tokens_enc: Dimension of tokens coming from encoder\n        \"\"\"\n        # print(dim_tokens_enc)\n\n        # Set up activation postprocessing layers\n        if isinstance(dim_tokens_enc, int):\n            dim_tokens_enc = 4 * [dim_tokens_enc]\n\n        self.dim_tokens_enc = [dt * len(self.main_tasks) for dt in dim_tokens_enc]\n\n        self.act_1_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[0],\n                out_channels=self.layer_dims[0],","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.forward","uri":"program://Human3R/function/src.croco.models.dpt_block.forward#L478-L513","kind":"function","name":"forward","path":"src/croco/models/dpt_block.py","language":"python","start_line":478,"end_line":513,"context_start_line":458,"context_end_line":513,"code":"                padding=1,\n            ),\n        )\n\n        self.act_postprocess = nn.ModuleList(\n            [\n                self.act_1_postprocess,\n                self.act_2_postprocess,\n                self.act_3_postprocess,\n                self.act_4_postprocess,\n            ]\n        )\n\n    def adapt_tokens(self, encoder_tokens):\n        # Adapt tokens\n        x = []\n        x.append(encoder_tokens[:, :])\n        x = torch.cat(x, dim=-1)\n        return x\n\n    def forward(self, encoder_tokens: List[torch.Tensor], image_size):\n        # input_info: Dict):\n        assert (\n            self.dim_tokens_enc is not None\n        ), \"Need to call init(dim_tokens_enc) function first\"\n        H, W = image_size\n\n        # Number of patches in height and width\n        N_H = H // (self.stride_level * self.P_H)\n        N_W = W // (self.stride_level * self.P_W)\n\n        # Hook decoder onto 4 layers from specified ViT layers\n        layers = [encoder_tokens[hook] for hook in self.hooks]\n\n        # Extract only task-relevant tokens and ignore global tokens.\n        layers = [self.adapt_tokens(l) for l in layers]\n\n        # Reshape tokens to spatial representation\n        layers = [\n            rearrange(l, \"b (nh nw) c -> b c nh nw\", nh=N_H, nw=N_W) for l in layers\n        ]\n\n        layers = [self.act_postprocess[idx](l) for idx, l in enumerate(layers)]\n        # Project layers to chosen feature dim\n        layers = [self.scratch.layer_rn[idx](l) for idx, l in enumerate(layers)]\n\n        # Fuse layers using refinement stages\n        path_4 = self.scratch.refinenet4(layers[3])\n        path_3 = self.scratch.refinenet3(path_4, layers[2])\n        path_2 = self.scratch.refinenet2(path_3, layers[1])\n        path_1 = self.scratch.refinenet1(path_2, layers[0])\n\n        # Output head\n        out = self.head(path_1)\n\n        return out","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.init","uri":"program://Human3R/function/src.croco.models.dpt_block.init#L380-L469","kind":"function","name":"init","path":"src/croco/models/dpt_block.py","language":"python","start_line":380,"end_line":469,"context_start_line":360,"context_end_line":489,"code":"                ),\n            )\n        elif self.head_type == \"semseg\":\n            # The \"DPTSegmentationModel\" head\n            self.head = nn.Sequential(\n                nn.Conv2d(\n                    feature_dim, feature_dim, kernel_size=3, padding=1, bias=False\n                ),\n                nn.BatchNorm2d(feature_dim) if use_bn else nn.Identity(),\n                nn.ReLU(True),\n                nn.Dropout(0.1, False),\n                nn.Conv2d(feature_dim, self.num_channels, kernel_size=1),\n                Interpolate(scale_factor=2, mode=\"bilinear\", align_corners=True),\n            )\n        else:\n            raise ValueError('DPT head_type must be \"regression\" or \"semseg\".')\n\n        if self.dim_tokens_enc is not None:\n            self.init(dim_tokens_enc=dim_tokens_enc)\n\n    def init(self, dim_tokens_enc=768):\n        \"\"\"\n        Initialize parts of decoder that are dependent on dimension of encoder tokens.\n        Should be called when setting up MultiMAE.\n\n        :param dim_tokens_enc: Dimension of tokens coming from encoder\n        \"\"\"\n        # print(dim_tokens_enc)\n\n        # Set up activation postprocessing layers\n        if isinstance(dim_tokens_enc, int):\n            dim_tokens_enc = 4 * [dim_tokens_enc]\n\n        self.dim_tokens_enc = [dt * len(self.main_tasks) for dt in dim_tokens_enc]\n\n        self.act_1_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[0],\n                out_channels=self.layer_dims[0],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            ),\n            nn.ConvTranspose2d(\n                in_channels=self.layer_dims[0],\n                out_channels=self.layer_dims[0],\n                kernel_size=4,\n                stride=4,\n                padding=0,\n                bias=True,\n                dilation=1,\n                groups=1,\n            ),\n        )\n\n        self.act_2_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[1],\n                out_channels=self.layer_dims[1],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            ),\n            nn.ConvTranspose2d(\n                in_channels=self.layer_dims[1],\n                out_channels=self.layer_dims[1],\n                kernel_size=2,\n                stride=2,\n                padding=0,\n                bias=True,\n                dilation=1,\n                groups=1,\n            ),\n        )\n\n        self.act_3_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[2],\n                out_channels=self.layer_dims[2],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            )\n        )\n\n        self.act_4_postprocess = nn.Sequential(\n            nn.Conv2d(\n                in_channels=self.dim_tokens_enc[3],\n                out_channels=self.layer_dims[3],\n                kernel_size=1,\n                stride=1,\n                padding=0,\n            ),\n            nn.Conv2d(\n                in_channels=self.layer_dims[3],\n                out_channels=self.layer_dims[3],\n                kernel_size=3,\n                stride=2,\n                padding=1,\n            ),\n        )\n\n        self.act_postprocess = nn.ModuleList(\n            [\n                self.act_1_postprocess,\n                self.act_2_postprocess,\n                self.act_3_postprocess,\n                self.act_4_postprocess,\n            ]\n        )\n\n    def adapt_tokens(self, encoder_tokens):\n        # Adapt tokens\n        x = []\n        x.append(encoder_tokens[:, :])\n        x = torch.cat(x, dim=-1)\n        return x\n\n    def forward(self, encoder_tokens: List[torch.Tensor], image_size):\n        # input_info: Dict):\n        assert (\n            self.dim_tokens_enc is not None\n        ), \"Need to call init(dim_tokens_enc) function first\"\n        H, W = image_size\n\n        # Number of patches in height and width\n        N_H = H // (self.stride_level * self.P_H)\n        N_W = W // (self.stride_level * self.P_W)\n\n        # Hook decoder onto 4 layers from specified ViT layers","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.dpt_block.adapt_tokens","uri":"program://Human3R/function/src.croco.models.dpt_block.adapt_tokens#L471-L476","kind":"function","name":"adapt_tokens","path":"src/croco/models/dpt_block.py","language":"python","start_line":471,"end_line":476,"context_start_line":451,"context_end_line":496,"code":"                padding=0,\n            ),\n            nn.Conv2d(\n                in_channels=self.layer_dims[3],\n                out_channels=self.layer_dims[3],\n                kernel_size=3,\n                stride=2,\n                padding=1,\n            ),\n        )\n\n        self.act_postprocess = nn.ModuleList(\n            [\n                self.act_1_postprocess,\n                self.act_2_postprocess,\n                self.act_3_postprocess,\n                self.act_4_postprocess,\n            ]\n        )\n\n    def adapt_tokens(self, encoder_tokens):\n        # Adapt tokens\n        x = []\n        x.append(encoder_tokens[:, :])\n        x = torch.cat(x, dim=-1)\n        return x\n\n    def forward(self, encoder_tokens: List[torch.Tensor], image_size):\n        # input_info: Dict):\n        assert (\n            self.dim_tokens_enc is not None\n        ), \"Need to call init(dim_tokens_enc) function first\"\n        H, W = image_size\n\n        # Number of patches in height and width\n        N_H = H // (self.stride_level * self.P_H)\n        N_W = W // (self.stride_level * self.P_W)\n\n        # Hook decoder onto 4 layers from specified ViT layers\n        layers = [encoder_tokens[hook] for hook in self.hooks]\n\n        # Extract only task-relevant tokens and ignore global tokens.\n        layers = [self.adapt_tokens(l) for l in layers]\n\n        # Reshape tokens to spatial representation\n        layers = [","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed","uri":"program://Human3R/module/src.croco.models.pos_embed#L1-L179","kind":"module","name":"src.croco.models.pos_embed","path":"src/croco/models/pos_embed.py","language":"python","start_line":1,"end_line":179,"context_start_line":1,"context_end_line":179,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# Position embedding utils\n# --------------------------------------------------------\n\n\nimport numpy as np\n\nimport torch\n\n\n# --------------------------------------------------------\n# 2D sine-cosine position embedding\n# References:\n# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py\n# Transformer: https://github.com/tensorflow/models/blob/master/official/nlp/transformer/model_utils.py\n# MoCo v3: https://github.com/facebookresearch/moco-v3\n# --------------------------------------------------------\ndef get_2d_sincos_pos_embed(embed_dim, grid_size, n_cls_token=0):\n    \"\"\"\n    grid_size: int of the grid height and width\n    return:\n    pos_embed: [grid_size*grid_size, embed_dim] or [n_cls_token+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)\n    \"\"\"\n    grid_h = np.arange(grid_size, dtype=np.float32)\n    grid_w = np.arange(grid_size, dtype=np.float32)\n    grid = np.meshgrid(grid_w, grid_h)  # here w goes first\n    grid = np.stack(grid, axis=0)\n\n    grid = grid.reshape([2, 1, grid_size, grid_size])\n    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)\n    if n_cls_token > 0:\n        pos_embed = np.concatenate(\n            [np.zeros([n_cls_token, embed_dim]), pos_embed], axis=0\n        )\n    return pos_embed\n\n\ndef get_2d_sincos_pos_embed_from_grid(embed_dim, grid):\n    assert embed_dim % 2 == 0\n\n    # use half of dimensions to encode grid_h\n    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)\n    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)\n\n    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)\n    return emb\n\n\ndef get_1d_sincos_pos_embed_from_grid(embed_dim, pos):\n    \"\"\"\n    embed_dim: output dimension for each position\n    pos: a list of positions to be encoded: size (M,)\n    out: (M, D)\n    \"\"\"\n    assert embed_dim % 2 == 0\n    omega = np.arange(embed_dim // 2, dtype=float)\n    omega /= embed_dim / 2.0\n    omega = 1.0 / 10000**omega  # (D/2,)\n\n    pos = pos.reshape(-1)  # (M,)\n    out = np.einsum(\"m,d->md\", pos, omega)  # (M, D/2), outer product\n\n    emb_sin = np.sin(out)  # (M, D/2)\n    emb_cos = np.cos(out)  # (M, D/2)\n\n    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)\n    return emb\n\n\n# --------------------------------------------------------\n# Interpolate position embeddings for high-resolution\n# References:\n# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py\n# DeiT: https://github.com/facebookresearch/deit\n# --------------------------------------------------------\ndef interpolate_pos_embed(model, checkpoint_model):\n    if \"pos_embed\" in checkpoint_model:\n        pos_embed_checkpoint = checkpoint_model[\"pos_embed\"]\n        embedding_size = pos_embed_checkpoint.shape[-1]\n        num_patches = model.patch_embed.num_patches\n        num_extra_tokens = model.pos_embed.shape[-2] - num_patches\n        # height (== width) for the checkpoint position embedding\n        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n        # height (== width) for the new position embedding\n        new_size = int(num_patches**0.5)\n        # class_token and dist_token are kept unchanged\n        if orig_size != new_size:\n            print(\n                \"Position interpolate from %dx%d to %dx%d\"\n                % (orig_size, orig_size, new_size, new_size)\n            )\n            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n            # only the position tokens are interpolated\n            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n            pos_tokens = pos_tokens.reshape(\n                -1, orig_size, orig_size, embedding_size\n            ).permute(0, 3, 1, 2)\n            pos_tokens = torch.nn.functional.interpolate(\n                pos_tokens,\n                size=(new_size, new_size),\n                mode=\"bicubic\",\n                align_corners=False,\n            )\n            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n            checkpoint_model[\"pos_embed\"] = new_pos_embed\n\n\n# ----------------------------------------------------------\n# RoPE2D: RoPE implementation in 2D\n# ----------------------------------------------------------\n\ntry:\n    from models.curope import cuRoPE2D\n\n    RoPE2D = cuRoPE2D\nexcept ImportError:\n    print(\n        \"Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead\"\n    )\n\n    class RoPE2D(torch.nn.Module):\n\n        def __init__(self, freq=100.0, F0=1.0):\n            super().__init__()\n            self.base = freq\n            self.F0 = F0\n            self.cache = {}\n\n        def get_cos_sin(self, D, seq_len, device, dtype):\n            if (D, seq_len, device, dtype) not in self.cache:\n                inv_freq = 1.0 / (\n                    self.base ** (torch.arange(0, D, 2).float().to(device) / D)\n                )\n                t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)\n                freqs = torch.einsum(\"i,j->ij\", t, inv_freq).to(dtype)\n                freqs = torch.cat((freqs, freqs), dim=-1)\n                cos = freqs.cos()  # (Seq, Dim)\n                sin = freqs.sin()\n                self.cache[D, seq_len, device, dtype] = (cos, sin)\n            return self.cache[D, seq_len, device, dtype]\n\n        @staticmethod\n        def rotate_half(x):\n            x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]\n            return torch.cat((-x2, x1), dim=-1)\n\n        def apply_rope1d(self, tokens, pos1d, cos, sin):\n            assert pos1d.ndim == 2\n            cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]\n            sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]\n            return (tokens * cos) + (self.rotate_half(tokens) * sin)\n\n        def forward(self, tokens, positions):\n            \"\"\"\n            input:\n                * tokens: batch_size x nheads x ntokens x dim\n                * positions: batch_size x ntokens x 2 (y and x position of each token)\n            output:\n                * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)\n            \"\"\"\n            assert (\n                tokens.size(3) % 2 == 0\n            ), \"number of dimensions should be a multiple of two\"\n            D = tokens.size(3) // 2\n            assert positions.ndim == 3 and positions.shape[-1] == 2  # Batch, Seq, 2\n            cos, sin = self.get_cos_sin(\n                D, int(positions.max()) + 1, tokens.device, tokens.dtype\n            )\n            # split features into two along the feature dimension, and apply rope1d on each half\n            y, x = tokens.chunk(2, dim=-1)\n            y = self.apply_rope1d(y, positions[:, :, 0], cos, sin)\n            x = self.apply_rope1d(x, positions[:, :, 1], cos, sin)\n            tokens = torch.cat((y, x), dim=-1)\n            return tokens","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.get_2d_sincos_pos_embed","uri":"program://Human3R/function/src.croco.models.pos_embed.get_2d_sincos_pos_embed#L22-L39","kind":"function","name":"get_2d_sincos_pos_embed","path":"src/croco/models/pos_embed.py","language":"python","start_line":22,"end_line":39,"context_start_line":2,"context_end_line":59,"code":"# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# Position embedding utils\n# --------------------------------------------------------\n\n\nimport numpy as np\n\nimport torch\n\n\n# --------------------------------------------------------\n# 2D sine-cosine position embedding\n# References:\n# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py\n# Transformer: https://github.com/tensorflow/models/blob/master/official/nlp/transformer/model_utils.py\n# MoCo v3: https://github.com/facebookresearch/moco-v3\n# --------------------------------------------------------\ndef get_2d_sincos_pos_embed(embed_dim, grid_size, n_cls_token=0):\n    \"\"\"\n    grid_size: int of the grid height and width\n    return:\n    pos_embed: [grid_size*grid_size, embed_dim] or [n_cls_token+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)\n    \"\"\"\n    grid_h = np.arange(grid_size, dtype=np.float32)\n    grid_w = np.arange(grid_size, dtype=np.float32)\n    grid = np.meshgrid(grid_w, grid_h)  # here w goes first\n    grid = np.stack(grid, axis=0)\n\n    grid = grid.reshape([2, 1, grid_size, grid_size])\n    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)\n    if n_cls_token > 0:\n        pos_embed = np.concatenate(\n            [np.zeros([n_cls_token, embed_dim]), pos_embed], axis=0\n        )\n    return pos_embed\n\n\ndef get_2d_sincos_pos_embed_from_grid(embed_dim, grid):\n    assert embed_dim % 2 == 0\n\n    # use half of dimensions to encode grid_h\n    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)\n    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)\n\n    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)\n    return emb\n\n\ndef get_1d_sincos_pos_embed_from_grid(embed_dim, pos):\n    \"\"\"\n    embed_dim: output dimension for each position\n    pos: a list of positions to be encoded: size (M,)\n    out: (M, D)\n    \"\"\"\n    assert embed_dim % 2 == 0","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.get_2d_sincos_pos_embed_from_grid","uri":"program://Human3R/function/src.croco.models.pos_embed.get_2d_sincos_pos_embed_from_grid#L42-L50","kind":"function","name":"get_2d_sincos_pos_embed_from_grid","path":"src/croco/models/pos_embed.py","language":"python","start_line":42,"end_line":50,"context_start_line":22,"context_end_line":70,"code":"def get_2d_sincos_pos_embed(embed_dim, grid_size, n_cls_token=0):\n    \"\"\"\n    grid_size: int of the grid height and width\n    return:\n    pos_embed: [grid_size*grid_size, embed_dim] or [n_cls_token+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)\n    \"\"\"\n    grid_h = np.arange(grid_size, dtype=np.float32)\n    grid_w = np.arange(grid_size, dtype=np.float32)\n    grid = np.meshgrid(grid_w, grid_h)  # here w goes first\n    grid = np.stack(grid, axis=0)\n\n    grid = grid.reshape([2, 1, grid_size, grid_size])\n    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)\n    if n_cls_token > 0:\n        pos_embed = np.concatenate(\n            [np.zeros([n_cls_token, embed_dim]), pos_embed], axis=0\n        )\n    return pos_embed\n\n\ndef get_2d_sincos_pos_embed_from_grid(embed_dim, grid):\n    assert embed_dim % 2 == 0\n\n    # use half of dimensions to encode grid_h\n    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)\n    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)\n\n    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)\n    return emb\n\n\ndef get_1d_sincos_pos_embed_from_grid(embed_dim, pos):\n    \"\"\"\n    embed_dim: output dimension for each position\n    pos: a list of positions to be encoded: size (M,)\n    out: (M, D)\n    \"\"\"\n    assert embed_dim % 2 == 0\n    omega = np.arange(embed_dim // 2, dtype=float)\n    omega /= embed_dim / 2.0\n    omega = 1.0 / 10000**omega  # (D/2,)\n\n    pos = pos.reshape(-1)  # (M,)\n    out = np.einsum(\"m,d->md\", pos, omega)  # (M, D/2), outer product\n\n    emb_sin = np.sin(out)  # (M, D/2)\n    emb_cos = np.cos(out)  # (M, D/2)\n\n    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.get_1d_sincos_pos_embed_from_grid","uri":"program://Human3R/function/src.croco.models.pos_embed.get_1d_sincos_pos_embed_from_grid#L53-L71","kind":"function","name":"get_1d_sincos_pos_embed_from_grid","path":"src/croco/models/pos_embed.py","language":"python","start_line":53,"end_line":71,"context_start_line":33,"context_end_line":91,"code":"    grid = grid.reshape([2, 1, grid_size, grid_size])\n    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)\n    if n_cls_token > 0:\n        pos_embed = np.concatenate(\n            [np.zeros([n_cls_token, embed_dim]), pos_embed], axis=0\n        )\n    return pos_embed\n\n\ndef get_2d_sincos_pos_embed_from_grid(embed_dim, grid):\n    assert embed_dim % 2 == 0\n\n    # use half of dimensions to encode grid_h\n    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)\n    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)\n\n    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)\n    return emb\n\n\ndef get_1d_sincos_pos_embed_from_grid(embed_dim, pos):\n    \"\"\"\n    embed_dim: output dimension for each position\n    pos: a list of positions to be encoded: size (M,)\n    out: (M, D)\n    \"\"\"\n    assert embed_dim % 2 == 0\n    omega = np.arange(embed_dim // 2, dtype=float)\n    omega /= embed_dim / 2.0\n    omega = 1.0 / 10000**omega  # (D/2,)\n\n    pos = pos.reshape(-1)  # (M,)\n    out = np.einsum(\"m,d->md\", pos, omega)  # (M, D/2), outer product\n\n    emb_sin = np.sin(out)  # (M, D/2)\n    emb_cos = np.cos(out)  # (M, D/2)\n\n    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)\n    return emb\n\n\n# --------------------------------------------------------\n# Interpolate position embeddings for high-resolution\n# References:\n# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py\n# DeiT: https://github.com/facebookresearch/deit\n# --------------------------------------------------------\ndef interpolate_pos_embed(model, checkpoint_model):\n    if \"pos_embed\" in checkpoint_model:\n        pos_embed_checkpoint = checkpoint_model[\"pos_embed\"]\n        embedding_size = pos_embed_checkpoint.shape[-1]\n        num_patches = model.patch_embed.num_patches\n        num_extra_tokens = model.pos_embed.shape[-2] - num_patches\n        # height (== width) for the checkpoint position embedding\n        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n        # height (== width) for the new position embedding\n        new_size = int(num_patches**0.5)\n        # class_token and dist_token are kept unchanged\n        if orig_size != new_size:","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.interpolate_pos_embed","uri":"program://Human3R/function/src.croco.models.pos_embed.interpolate_pos_embed#L80-L110","kind":"function","name":"interpolate_pos_embed","path":"src/croco/models/pos_embed.py","language":"python","start_line":80,"end_line":110,"context_start_line":60,"context_end_line":130,"code":"    omega = np.arange(embed_dim // 2, dtype=float)\n    omega /= embed_dim / 2.0\n    omega = 1.0 / 10000**omega  # (D/2,)\n\n    pos = pos.reshape(-1)  # (M,)\n    out = np.einsum(\"m,d->md\", pos, omega)  # (M, D/2), outer product\n\n    emb_sin = np.sin(out)  # (M, D/2)\n    emb_cos = np.cos(out)  # (M, D/2)\n\n    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)\n    return emb\n\n\n# --------------------------------------------------------\n# Interpolate position embeddings for high-resolution\n# References:\n# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py\n# DeiT: https://github.com/facebookresearch/deit\n# --------------------------------------------------------\ndef interpolate_pos_embed(model, checkpoint_model):\n    if \"pos_embed\" in checkpoint_model:\n        pos_embed_checkpoint = checkpoint_model[\"pos_embed\"]\n        embedding_size = pos_embed_checkpoint.shape[-1]\n        num_patches = model.patch_embed.num_patches\n        num_extra_tokens = model.pos_embed.shape[-2] - num_patches\n        # height (== width) for the checkpoint position embedding\n        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n        # height (== width) for the new position embedding\n        new_size = int(num_patches**0.5)\n        # class_token and dist_token are kept unchanged\n        if orig_size != new_size:\n            print(\n                \"Position interpolate from %dx%d to %dx%d\"\n                % (orig_size, orig_size, new_size, new_size)\n            )\n            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n            # only the position tokens are interpolated\n            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n            pos_tokens = pos_tokens.reshape(\n                -1, orig_size, orig_size, embedding_size\n            ).permute(0, 3, 1, 2)\n            pos_tokens = torch.nn.functional.interpolate(\n                pos_tokens,\n                size=(new_size, new_size),\n                mode=\"bicubic\",\n                align_corners=False,\n            )\n            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n            checkpoint_model[\"pos_embed\"] = new_pos_embed\n\n\n# ----------------------------------------------------------\n# RoPE2D: RoPE implementation in 2D\n# ----------------------------------------------------------\n\ntry:\n    from models.curope import cuRoPE2D\n\n    RoPE2D = cuRoPE2D\nexcept ImportError:\n    print(\n        \"Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead\"\n    )\n\n    class RoPE2D(torch.nn.Module):\n\n        def __init__(self, freq=100.0, F0=1.0):\n            super().__init__()\n            self.base = freq","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.RoPE2D","uri":"program://Human3R/class/src.croco.models.pos_embed.RoPE2D#L126-L179","kind":"class","name":"RoPE2D","path":"src/croco/models/pos_embed.py","language":"python","start_line":126,"end_line":179,"context_start_line":106,"context_end_line":179,"code":"                align_corners=False,\n            )\n            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n            checkpoint_model[\"pos_embed\"] = new_pos_embed\n\n\n# ----------------------------------------------------------\n# RoPE2D: RoPE implementation in 2D\n# ----------------------------------------------------------\n\ntry:\n    from models.curope import cuRoPE2D\n\n    RoPE2D = cuRoPE2D\nexcept ImportError:\n    print(\n        \"Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead\"\n    )\n\n    class RoPE2D(torch.nn.Module):\n\n        def __init__(self, freq=100.0, F0=1.0):\n            super().__init__()\n            self.base = freq\n            self.F0 = F0\n            self.cache = {}\n\n        def get_cos_sin(self, D, seq_len, device, dtype):\n            if (D, seq_len, device, dtype) not in self.cache:\n                inv_freq = 1.0 / (\n                    self.base ** (torch.arange(0, D, 2).float().to(device) / D)\n                )\n                t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)\n                freqs = torch.einsum(\"i,j->ij\", t, inv_freq).to(dtype)\n                freqs = torch.cat((freqs, freqs), dim=-1)\n                cos = freqs.cos()  # (Seq, Dim)\n                sin = freqs.sin()\n                self.cache[D, seq_len, device, dtype] = (cos, sin)\n            return self.cache[D, seq_len, device, dtype]\n\n        @staticmethod\n        def rotate_half(x):\n            x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]\n            return torch.cat((-x2, x1), dim=-1)\n\n        def apply_rope1d(self, tokens, pos1d, cos, sin):\n            assert pos1d.ndim == 2\n            cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]\n            sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]\n            return (tokens * cos) + (self.rotate_half(tokens) * sin)\n\n        def forward(self, tokens, positions):\n            \"\"\"\n            input:\n                * tokens: batch_size x nheads x ntokens x dim\n                * positions: batch_size x ntokens x 2 (y and x position of each token)\n            output:\n                * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)\n            \"\"\"\n            assert (\n                tokens.size(3) % 2 == 0\n            ), \"number of dimensions should be a multiple of two\"\n            D = tokens.size(3) // 2\n            assert positions.ndim == 3 and positions.shape[-1] == 2  # Batch, Seq, 2\n            cos, sin = self.get_cos_sin(\n                D, int(positions.max()) + 1, tokens.device, tokens.dtype\n            )\n            # split features into two along the feature dimension, and apply rope1d on each half\n            y, x = tokens.chunk(2, dim=-1)\n            y = self.apply_rope1d(y, positions[:, :, 0], cos, sin)\n            x = self.apply_rope1d(x, positions[:, :, 1], cos, sin)\n            tokens = torch.cat((y, x), dim=-1)\n            return tokens","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.__init__","uri":"program://Human3R/function/src.croco.models.pos_embed.__init__#L128-L132","kind":"function","name":"__init__","path":"src/croco/models/pos_embed.py","language":"python","start_line":128,"end_line":132,"context_start_line":108,"context_end_line":152,"code":"            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n            checkpoint_model[\"pos_embed\"] = new_pos_embed\n\n\n# ----------------------------------------------------------\n# RoPE2D: RoPE implementation in 2D\n# ----------------------------------------------------------\n\ntry:\n    from models.curope import cuRoPE2D\n\n    RoPE2D = cuRoPE2D\nexcept ImportError:\n    print(\n        \"Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead\"\n    )\n\n    class RoPE2D(torch.nn.Module):\n\n        def __init__(self, freq=100.0, F0=1.0):\n            super().__init__()\n            self.base = freq\n            self.F0 = F0\n            self.cache = {}\n\n        def get_cos_sin(self, D, seq_len, device, dtype):\n            if (D, seq_len, device, dtype) not in self.cache:\n                inv_freq = 1.0 / (\n                    self.base ** (torch.arange(0, D, 2).float().to(device) / D)\n                )\n                t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)\n                freqs = torch.einsum(\"i,j->ij\", t, inv_freq).to(dtype)\n                freqs = torch.cat((freqs, freqs), dim=-1)\n                cos = freqs.cos()  # (Seq, Dim)\n                sin = freqs.sin()\n                self.cache[D, seq_len, device, dtype] = (cos, sin)\n            return self.cache[D, seq_len, device, dtype]\n\n        @staticmethod\n        def rotate_half(x):\n            x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]\n            return torch.cat((-x2, x1), dim=-1)\n\n        def apply_rope1d(self, tokens, pos1d, cos, sin):","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.get_cos_sin","uri":"program://Human3R/function/src.croco.models.pos_embed.get_cos_sin#L134-L145","kind":"function","name":"get_cos_sin","path":"src/croco/models/pos_embed.py","language":"python","start_line":134,"end_line":145,"context_start_line":114,"context_end_line":165,"code":"# RoPE2D: RoPE implementation in 2D\n# ----------------------------------------------------------\n\ntry:\n    from models.curope import cuRoPE2D\n\n    RoPE2D = cuRoPE2D\nexcept ImportError:\n    print(\n        \"Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead\"\n    )\n\n    class RoPE2D(torch.nn.Module):\n\n        def __init__(self, freq=100.0, F0=1.0):\n            super().__init__()\n            self.base = freq\n            self.F0 = F0\n            self.cache = {}\n\n        def get_cos_sin(self, D, seq_len, device, dtype):\n            if (D, seq_len, device, dtype) not in self.cache:\n                inv_freq = 1.0 / (\n                    self.base ** (torch.arange(0, D, 2).float().to(device) / D)\n                )\n                t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)\n                freqs = torch.einsum(\"i,j->ij\", t, inv_freq).to(dtype)\n                freqs = torch.cat((freqs, freqs), dim=-1)\n                cos = freqs.cos()  # (Seq, Dim)\n                sin = freqs.sin()\n                self.cache[D, seq_len, device, dtype] = (cos, sin)\n            return self.cache[D, seq_len, device, dtype]\n\n        @staticmethod\n        def rotate_half(x):\n            x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]\n            return torch.cat((-x2, x1), dim=-1)\n\n        def apply_rope1d(self, tokens, pos1d, cos, sin):\n            assert pos1d.ndim == 2\n            cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]\n            sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]\n            return (tokens * cos) + (self.rotate_half(tokens) * sin)\n\n        def forward(self, tokens, positions):\n            \"\"\"\n            input:\n                * tokens: batch_size x nheads x ntokens x dim\n                * positions: batch_size x ntokens x 2 (y and x position of each token)\n            output:\n                * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)\n            \"\"\"","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.rotate_half","uri":"program://Human3R/function/src.croco.models.pos_embed.rotate_half#L148-L150","kind":"function","name":"rotate_half","path":"src/croco/models/pos_embed.py","language":"python","start_line":148,"end_line":150,"context_start_line":128,"context_end_line":170,"code":"        def __init__(self, freq=100.0, F0=1.0):\n            super().__init__()\n            self.base = freq\n            self.F0 = F0\n            self.cache = {}\n\n        def get_cos_sin(self, D, seq_len, device, dtype):\n            if (D, seq_len, device, dtype) not in self.cache:\n                inv_freq = 1.0 / (\n                    self.base ** (torch.arange(0, D, 2).float().to(device) / D)\n                )\n                t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)\n                freqs = torch.einsum(\"i,j->ij\", t, inv_freq).to(dtype)\n                freqs = torch.cat((freqs, freqs), dim=-1)\n                cos = freqs.cos()  # (Seq, Dim)\n                sin = freqs.sin()\n                self.cache[D, seq_len, device, dtype] = (cos, sin)\n            return self.cache[D, seq_len, device, dtype]\n\n        @staticmethod\n        def rotate_half(x):\n            x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]\n            return torch.cat((-x2, x1), dim=-1)\n\n        def apply_rope1d(self, tokens, pos1d, cos, sin):\n            assert pos1d.ndim == 2\n            cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]\n            sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]\n            return (tokens * cos) + (self.rotate_half(tokens) * sin)\n\n        def forward(self, tokens, positions):\n            \"\"\"\n            input:\n                * tokens: batch_size x nheads x ntokens x dim\n                * positions: batch_size x ntokens x 2 (y and x position of each token)\n            output:\n                * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)\n            \"\"\"\n            assert (\n                tokens.size(3) % 2 == 0\n            ), \"number of dimensions should be a multiple of two\"\n            D = tokens.size(3) // 2\n            assert positions.ndim == 3 and positions.shape[-1] == 2  # Batch, Seq, 2","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.apply_rope1d","uri":"program://Human3R/function/src.croco.models.pos_embed.apply_rope1d#L152-L156","kind":"function","name":"apply_rope1d","path":"src/croco/models/pos_embed.py","language":"python","start_line":152,"end_line":156,"context_start_line":132,"context_end_line":176,"code":"            self.cache = {}\n\n        def get_cos_sin(self, D, seq_len, device, dtype):\n            if (D, seq_len, device, dtype) not in self.cache:\n                inv_freq = 1.0 / (\n                    self.base ** (torch.arange(0, D, 2).float().to(device) / D)\n                )\n                t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)\n                freqs = torch.einsum(\"i,j->ij\", t, inv_freq).to(dtype)\n                freqs = torch.cat((freqs, freqs), dim=-1)\n                cos = freqs.cos()  # (Seq, Dim)\n                sin = freqs.sin()\n                self.cache[D, seq_len, device, dtype] = (cos, sin)\n            return self.cache[D, seq_len, device, dtype]\n\n        @staticmethod\n        def rotate_half(x):\n            x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]\n            return torch.cat((-x2, x1), dim=-1)\n\n        def apply_rope1d(self, tokens, pos1d, cos, sin):\n            assert pos1d.ndim == 2\n            cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]\n            sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]\n            return (tokens * cos) + (self.rotate_half(tokens) * sin)\n\n        def forward(self, tokens, positions):\n            \"\"\"\n            input:\n                * tokens: batch_size x nheads x ntokens x dim\n                * positions: batch_size x ntokens x 2 (y and x position of each token)\n            output:\n                * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)\n            \"\"\"\n            assert (\n                tokens.size(3) % 2 == 0\n            ), \"number of dimensions should be a multiple of two\"\n            D = tokens.size(3) // 2\n            assert positions.ndim == 3 and positions.shape[-1] == 2  # Batch, Seq, 2\n            cos, sin = self.get_cos_sin(\n                D, int(positions.max()) + 1, tokens.device, tokens.dtype\n            )\n            # split features into two along the feature dimension, and apply rope1d on each half\n            y, x = tokens.chunk(2, dim=-1)\n            y = self.apply_rope1d(y, positions[:, :, 0], cos, sin)","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.pos_embed.forward","uri":"program://Human3R/function/src.croco.models.pos_embed.forward#L158-L179","kind":"function","name":"forward","path":"src/croco/models/pos_embed.py","language":"python","start_line":158,"end_line":179,"context_start_line":138,"context_end_line":179,"code":"                )\n                t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)\n                freqs = torch.einsum(\"i,j->ij\", t, inv_freq).to(dtype)\n                freqs = torch.cat((freqs, freqs), dim=-1)\n                cos = freqs.cos()  # (Seq, Dim)\n                sin = freqs.sin()\n                self.cache[D, seq_len, device, dtype] = (cos, sin)\n            return self.cache[D, seq_len, device, dtype]\n\n        @staticmethod\n        def rotate_half(x):\n            x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]\n            return torch.cat((-x2, x1), dim=-1)\n\n        def apply_rope1d(self, tokens, pos1d, cos, sin):\n            assert pos1d.ndim == 2\n            cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]\n            sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]\n            return (tokens * cos) + (self.rotate_half(tokens) * sin)\n\n        def forward(self, tokens, positions):\n            \"\"\"\n            input:\n                * tokens: batch_size x nheads x ntokens x dim\n                * positions: batch_size x ntokens x 2 (y and x position of each token)\n            output:\n                * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)\n            \"\"\"\n            assert (\n                tokens.size(3) % 2 == 0\n            ), \"number of dimensions should be a multiple of two\"\n            D = tokens.size(3) // 2\n            assert positions.ndim == 3 and positions.shape[-1] == 2  # Batch, Seq, 2\n            cos, sin = self.get_cos_sin(\n                D, int(positions.max()) + 1, tokens.device, tokens.dtype\n            )\n            # split features into two along the feature dimension, and apply rope1d on each half\n            y, x = tokens.chunk(2, dim=-1)\n            y = self.apply_rope1d(y, positions[:, :, 0], cos, sin)\n            x = self.apply_rope1d(x, positions[:, :, 1], cos, sin)\n            tokens = torch.cat((y, x), dim=-1)\n            return tokens","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco","uri":"program://Human3R/module/src.croco.models.croco#L1-L330","kind":"module","name":"src.croco.models.croco","path":"src/croco/models/croco.py","language":"python","start_line":1,"end_line":330,"context_start_line":1,"context_end_line":330,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# CroCo model during pretraining\n# --------------------------------------------------------\n\n\nimport torch\nimport torch.nn as nn\n\ntorch.backends.cuda.matmul.allow_tf32 = True  # for gpu >= Ampere and pytorch >= 1.12\nfrom functools import partial\n\nfrom models.blocks import Block, DecoderBlock, PatchEmbed\nfrom models.pos_embed import get_2d_sincos_pos_embed, RoPE2D\nfrom models.masking import RandomMask\n\nfrom transformers import PretrainedConfig\nfrom transformers import PreTrainedModel\n\n\nclass CrocoConfig(PretrainedConfig):\n    model_type = \"croco\"\n\n    def __init__(\n        self,\n        img_size=224,  # input image size\n        patch_size=16,  # patch_size\n        mask_ratio=0.9,  # ratios of masked tokens\n        enc_embed_dim=768,  # encoder feature dimension\n        enc_depth=12,  # encoder depth\n        enc_num_heads=12,  # encoder number of heads in the transformer block\n        dec_embed_dim=512,  # decoder feature dimension\n        dec_depth=8,  # decoder depth\n        dec_num_heads=16,  # decoder number of heads in the transformer block\n        mlp_ratio=4,\n        norm_layer=partial(nn.LayerNorm, eps=1e-6),\n        norm_im2_in_dec=True,  # whether to apply normalization of the 'memory' = (second image) in the decoder\n        pos_embed=\"cosine\",  # positional embedding (either cosine or RoPE100)\n    ):\n        super().__init__()\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.mask_ratio = mask_ratio\n        self.enc_embed_dim = enc_embed_dim\n        self.enc_depth = enc_depth\n        self.enc_num_heads = enc_num_heads\n        self.dec_embed_dim = dec_embed_dim\n        self.dec_depth = dec_depth\n        self.dec_num_heads = dec_num_heads\n        self.mlp_ratio = mlp_ratio\n        self.norm_layer = norm_layer\n        self.norm_im2_in_dec = norm_im2_in_dec\n        self.pos_embed = pos_embed\n\n\nclass CroCoNet(PreTrainedModel):\n\n    config_class = CrocoConfig\n    base_model_prefix = \"croco\"\n\n    def __init__(self, config: CrocoConfig):\n\n        super().__init__(config)\n\n        # patch embeddings  (with initialization done as in MAE)\n        self._set_patch_embed(config.img_size, config.patch_size, config.enc_embed_dim)\n\n        # mask generations\n        self._set_mask_generator(self.patch_embed.num_patches, config.mask_ratio)\n\n        self.pos_embed = config.pos_embed\n        if config.pos_embed == \"cosine\":\n            # positional embedding of the encoder\n            enc_pos_embed = get_2d_sincos_pos_embed(\n                config.enc_embed_dim,\n                int(self.patch_embed.num_patches**0.5),\n                n_cls_token=0,\n            )\n            self.register_buffer(\n                \"enc_pos_embed\", torch.from_numpy(enc_pos_embed).float()\n            )\n            # positional embedding of the decoder\n            dec_pos_embed = get_2d_sincos_pos_embed(\n                config.dec_embed_dim,\n                int(self.patch_embed.num_patches**0.5),\n                n_cls_token=0,\n            )\n            self.register_buffer(\n                \"dec_pos_embed\", torch.from_numpy(dec_pos_embed).float()\n            )\n            # pos embedding in each block\n            self.rope = None  # nothing for cosine\n        elif config.pos_embed.startswith(\"RoPE\"):  # eg RoPE100\n            self.enc_pos_embed = None  # nothing to add in the encoder with RoPE\n            self.dec_pos_embed = None  # nothing to add in the decoder with RoPE\n            if RoPE2D is None:\n                raise ImportError(\n                    \"Cannot find cuRoPE2D, please install it following the README instructions\"\n                )\n            freq = float(config.pos_embed[len(\"RoPE\") :])\n            self.rope = RoPE2D(freq=freq)\n        else:\n            raise NotImplementedError(\"Unknown pos_embed \" + config.pos_embed)\n\n        # transformer for the encoder\n        self.enc_depth = config.enc_depth\n        self.enc_embed_dim = config.enc_embed_dim\n        self.enc_blocks = nn.ModuleList(\n            [\n                Block(\n                    config.enc_embed_dim,\n                    config.enc_num_heads,\n                    config.mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=config.norm_layer,\n                    rope=self.rope,\n                )\n                for i in range(config.enc_depth)\n            ]\n        )\n        self.enc_norm = config.norm_layer(config.enc_embed_dim)\n\n        # masked tokens\n        # self._set_mask_token(config.dec_embed_dim)\n        self.mask_token = None\n\n        # decoder\n        self._set_decoder(\n            config.enc_embed_dim,\n            config.dec_embed_dim,\n            config.dec_num_heads,\n            config.dec_depth,\n            config.mlp_ratio,\n            config.norm_layer,\n            config.norm_im2_in_dec,\n        )\n\n        # prediction head\n        self._set_prediction_head(config.dec_embed_dim, config.patch_size)\n\n        # initializer weights\n        self.initialize_weights()\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = PatchEmbed(img_size, patch_size, 3, enc_embed_dim)\n\n    def _set_mask_generator(self, num_patches, mask_ratio):\n        self.mask_generator = RandomMask(num_patches, mask_ratio)\n\n    def _set_mask_token(self, dec_embed_dim):\n        self.mask_token = nn.Parameter(torch.zeros(1, 1, dec_embed_dim))\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        # transfer from encoder to decoder\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        # transformer for the decoder\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        # final norm layer\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_prediction_head(self, dec_embed_dim, patch_size):\n        self.prediction_head = nn.Linear(dec_embed_dim, patch_size**2 * 3, bias=True)\n\n    def initialize_weights(self):\n        # patch embed\n        self.patch_embed._init_weights()\n        # mask tokens\n        if self.mask_token is not None:\n            torch.nn.init.normal_(self.mask_token, std=0.02)\n        # linears and layer norms\n        self.apply(self._init_weights)\n\n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            # we use xavier_uniform following official JAX ViT:\n            torch.nn.init.xavier_uniform_(m.weight)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n\n    def _encode_image(self, image, do_mask=False, return_all_blocks=False):\n        \"\"\"\n        image has B x 3 x img_size x img_size\n        do_mask: whether to perform masking or not\n        return_all_blocks: if True, return the features at the end of every block\n                           instead of just the features from the last block (eg for some prediction heads)\n        \"\"\"\n        # embed the image into patches  (x has size B x Npatches x C)\n        # and get position if each return patch (pos has size B x Npatches x 2)\n        x, pos = self.patch_embed(image)\n        # add positional embedding without cls token\n        if self.enc_pos_embed is not None:\n            x = x + self.enc_pos_embed[None, ...]\n        # apply masking\n        B, N, C = x.size()\n        if do_mask:\n            masks = self.mask_generator(x)\n            x = x[~masks].view(B, -1, C)\n            posvis = pos[~masks].view(B, -1, 2)\n        else:\n            B, N, C = x.size()\n            masks = torch.zeros((B, N), dtype=bool)\n            posvis = pos\n        # now apply the transformer encoder and normalization\n        if return_all_blocks:\n            out = []\n            for blk in self.enc_blocks:\n                x = blk(x, posvis)\n                out.append(x)\n            out[-1] = self.enc_norm(out[-1])\n            return out, pos, masks\n        else:\n            for blk in self.enc_blocks:\n                x = blk(x, posvis)\n            x = self.enc_norm(x)\n            return x, pos, masks\n\n    def _decoder(self, feat1, pos1, masks1, feat2, pos2, return_all_blocks=False):\n        \"\"\"\n        return_all_blocks: if True, return the features at the end of every block\n                           instead of just the features from the last block (eg for some prediction heads)\n\n        masks1 can be None => assume image1 fully visible\n        \"\"\"\n        # encoder to decoder layer\n        visf1 = self.decoder_embed(feat1)\n        f2 = self.decoder_embed(feat2)\n        # append masked tokens to the sequence\n        B, Nenc, C = visf1.size()\n        if masks1 is None:  # downstreams\n            f1_ = visf1\n        else:  # pretraining\n            Ntotal = masks1.size(1)\n            f1_ = self.mask_token.repeat(B, Ntotal, 1).to(dtype=visf1.dtype)\n            f1_[~masks1] = visf1.view(B * Nenc, C)\n        # add positional embedding\n        if self.dec_pos_embed is not None:\n            f1_ = f1_ + self.dec_pos_embed\n            f2 = f2 + self.dec_pos_embed\n        # apply Transformer blocks\n        out = f1_\n        out2 = f2\n        if return_all_blocks:\n            _out, out = out, []\n            for blk in self.dec_blocks:\n                _out, out2 = blk(_out, out2, pos1, pos2)\n                out.append(_out)\n            out[-1] = self.dec_norm(out[-1])\n        else:\n            for blk in self.dec_blocks:\n                out, out2 = blk(out, out2, pos1, pos2)\n            out = self.dec_norm(out)\n        return out\n\n    def patchify(self, imgs):\n        \"\"\"\n        imgs: (B, 3, H, W)\n        x: (B, L, patch_size**2 *3)\n        \"\"\"\n        p = self.patch_embed.patch_size[0]\n        assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0\n\n        h = w = imgs.shape[2] // p\n        x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))\n        x = torch.einsum(\"nchpwq->nhwpqc\", x)\n        x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))\n\n        return x\n\n    def unpatchify(self, x, channels=3):\n        \"\"\"\n        x: (N, L, patch_size**2 *channels)\n        imgs: (N, 3, H, W)\n        \"\"\"\n        patch_size = self.patch_embed.patch_size[0]\n        h = w = int(x.shape[1] ** 0.5)\n        assert h * w == x.shape[1]\n        x = x.reshape(shape=(x.shape[0], h, w, patch_size, patch_size, channels))\n        x = torch.einsum(\"nhwpqc->nchpwq\", x)\n        imgs = x.reshape(shape=(x.shape[0], channels, h * patch_size, h * patch_size))\n        return imgs\n\n    # def forward(self, img1, img2):\n    # \"\"\"\n    # img1: tensor of size B x 3 x img_size x img_size\n    # img2: tensor of size B x 3 x img_size x img_size\n\n    # out will be    B x N x (3*patch_size*patch_size)\n    # masks are also returned as B x N just in case\n    # \"\"\"\n    # # encoder of the masked first image\n    # feat1, pos1, mask1 = self._encode_image(img1, do_mask=True)\n    # # encoder of the second image\n    # feat2, pos2, _ = self._encode_image(img2, do_mask=False)\n    # # decoder\n    # decfeat = self._decoder(feat1, pos1, mask1, feat2, pos2)\n    # # prediction head\n    # out = self.prediction_head(decfeat)\n    # # get target\n    # target = self.patchify(img1)\n    # return out, mask1, target","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco.CrocoConfig","uri":"program://Human3R/class/src.croco.models.croco.CrocoConfig#L24-L56","kind":"class","name":"CrocoConfig","path":"src/croco/models/croco.py","language":"python","start_line":24,"end_line":56,"context_start_line":4,"context_end_line":76,"code":"\n# --------------------------------------------------------\n# CroCo model during pretraining\n# --------------------------------------------------------\n\n\nimport torch\nimport torch.nn as nn\n\ntorch.backends.cuda.matmul.allow_tf32 = True  # for gpu >= Ampere and pytorch >= 1.12\nfrom functools import partial\n\nfrom models.blocks import Block, DecoderBlock, PatchEmbed\nfrom models.pos_embed import get_2d_sincos_pos_embed, RoPE2D\nfrom models.masking import RandomMask\n\nfrom transformers import PretrainedConfig\nfrom transformers import PreTrainedModel\n\n\nclass CrocoConfig(PretrainedConfig):\n    model_type = \"croco\"\n\n    def __init__(\n        self,\n        img_size=224,  # input image size\n        patch_size=16,  # patch_size\n        mask_ratio=0.9,  # ratios of masked tokens\n        enc_embed_dim=768,  # encoder feature dimension\n        enc_depth=12,  # encoder depth\n        enc_num_heads=12,  # encoder number of heads in the transformer block\n        dec_embed_dim=512,  # decoder feature dimension\n        dec_depth=8,  # decoder depth\n        dec_num_heads=16,  # decoder number of heads in the transformer block\n        mlp_ratio=4,\n        norm_layer=partial(nn.LayerNorm, eps=1e-6),\n        norm_im2_in_dec=True,  # whether to apply normalization of the 'memory' = (second image) in the decoder\n        pos_embed=\"cosine\",  # positional embedding (either cosine or RoPE100)\n    ):\n        super().__init__()\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.mask_ratio = mask_ratio\n        self.enc_embed_dim = enc_embed_dim\n        self.enc_depth = enc_depth\n        self.enc_num_heads = enc_num_heads\n        self.dec_embed_dim = dec_embed_dim\n        self.dec_depth = dec_depth\n        self.dec_num_heads = dec_num_heads\n        self.mlp_ratio = mlp_ratio\n        self.norm_layer = norm_layer\n        self.norm_im2_in_dec = norm_im2_in_dec\n        self.pos_embed = pos_embed\n\n\nclass CroCoNet(PreTrainedModel):\n\n    config_class = CrocoConfig\n    base_model_prefix = \"croco\"\n\n    def __init__(self, config: CrocoConfig):\n\n        super().__init__(config)\n\n        # patch embeddings  (with initialization done as in MAE)\n        self._set_patch_embed(config.img_size, config.patch_size, config.enc_embed_dim)\n\n        # mask generations\n        self._set_mask_generator(self.patch_embed.num_patches, config.mask_ratio)\n\n        self.pos_embed = config.pos_embed\n        if config.pos_embed == \"cosine\":\n            # positional embedding of the encoder","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco.CroCoNet","uri":"program://Human3R/class/src.croco.models.croco.CroCoNet#L59-L310","kind":"class","name":"CroCoNet","path":"src/croco/models/croco.py","language":"python","start_line":59,"end_line":310,"context_start_line":39,"context_end_line":330,"code":"        norm_layer=partial(nn.LayerNorm, eps=1e-6),\n        norm_im2_in_dec=True,  # whether to apply normalization of the 'memory' = (second image) in the decoder\n        pos_embed=\"cosine\",  # positional embedding (either cosine or RoPE100)\n    ):\n        super().__init__()\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.mask_ratio = mask_ratio\n        self.enc_embed_dim = enc_embed_dim\n        self.enc_depth = enc_depth\n        self.enc_num_heads = enc_num_heads\n        self.dec_embed_dim = dec_embed_dim\n        self.dec_depth = dec_depth\n        self.dec_num_heads = dec_num_heads\n        self.mlp_ratio = mlp_ratio\n        self.norm_layer = norm_layer\n        self.norm_im2_in_dec = norm_im2_in_dec\n        self.pos_embed = pos_embed\n\n\nclass CroCoNet(PreTrainedModel):\n\n    config_class = CrocoConfig\n    base_model_prefix = \"croco\"\n\n    def __init__(self, config: CrocoConfig):\n\n        super().__init__(config)\n\n        # patch embeddings  (with initialization done as in MAE)\n        self._set_patch_embed(config.img_size, config.patch_size, config.enc_embed_dim)\n\n        # mask generations\n        self._set_mask_generator(self.patch_embed.num_patches, config.mask_ratio)\n\n        self.pos_embed = config.pos_embed\n        if config.pos_embed == \"cosine\":\n            # positional embedding of the encoder\n            enc_pos_embed = get_2d_sincos_pos_embed(\n                config.enc_embed_dim,\n                int(self.patch_embed.num_patches**0.5),\n                n_cls_token=0,\n            )\n            self.register_buffer(\n                \"enc_pos_embed\", torch.from_numpy(enc_pos_embed).float()\n            )\n            # positional embedding of the decoder\n            dec_pos_embed = get_2d_sincos_pos_embed(\n                config.dec_embed_dim,\n                int(self.patch_embed.num_patches**0.5),\n                n_cls_token=0,\n            )\n            self.register_buffer(\n                \"dec_pos_embed\", torch.from_numpy(dec_pos_embed).float()\n            )\n            # pos embedding in each block\n            self.rope = None  # nothing for cosine\n        elif config.pos_embed.startswith(\"RoPE\"):  # eg RoPE100\n            self.enc_pos_embed = None  # nothing to add in the encoder with RoPE\n            self.dec_pos_embed = None  # nothing to add in the decoder with RoPE\n            if RoPE2D is None:\n                raise ImportError(\n                    \"Cannot find cuRoPE2D, please install it following the README instructions\"\n                )\n            freq = float(config.pos_embed[len(\"RoPE\") :])\n            self.rope = RoPE2D(freq=freq)\n        else:\n            raise NotImplementedError(\"Unknown pos_embed \" + config.pos_embed)\n\n        # transformer for the encoder\n        self.enc_depth = config.enc_depth\n        self.enc_embed_dim = config.enc_embed_dim\n        self.enc_blocks = nn.ModuleList(\n            [\n                Block(\n                    config.enc_embed_dim,\n                    config.enc_num_heads,\n                    config.mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=config.norm_layer,\n                    rope=self.rope,\n                )\n                for i in range(config.enc_depth)\n            ]\n        )\n        self.enc_norm = config.norm_layer(config.enc_embed_dim)\n\n        # masked tokens\n        # self._set_mask_token(config.dec_embed_dim)\n        self.mask_token = None\n\n        # decoder\n        self._set_decoder(\n            config.enc_embed_dim,\n            config.dec_embed_dim,\n            config.dec_num_heads,\n            config.dec_depth,\n            config.mlp_ratio,\n            config.norm_layer,\n            config.norm_im2_in_dec,\n        )\n\n        # prediction head\n        self._set_prediction_head(config.dec_embed_dim, config.patch_size)\n\n        # initializer weights\n        self.initialize_weights()\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = PatchEmbed(img_size, patch_size, 3, enc_embed_dim)\n\n    def _set_mask_generator(self, num_patches, mask_ratio):\n        self.mask_generator = RandomMask(num_patches, mask_ratio)\n\n    def _set_mask_token(self, dec_embed_dim):\n        self.mask_token = nn.Parameter(torch.zeros(1, 1, dec_embed_dim))\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        # transfer from encoder to decoder\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        # transformer for the decoder\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        # final norm layer\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_prediction_head(self, dec_embed_dim, patch_size):\n        self.prediction_head = nn.Linear(dec_embed_dim, patch_size**2 * 3, bias=True)\n\n    def initialize_weights(self):\n        # patch embed\n        self.patch_embed._init_weights()\n        # mask tokens\n        if self.mask_token is not None:\n            torch.nn.init.normal_(self.mask_token, std=0.02)\n        # linears and layer norms\n        self.apply(self._init_weights)\n\n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            # we use xavier_uniform following official JAX ViT:\n            torch.nn.init.xavier_uniform_(m.weight)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n\n    def _encode_image(self, image, do_mask=False, return_all_blocks=False):\n        \"\"\"\n        image has B x 3 x img_size x img_size\n        do_mask: whether to perform masking or not\n        return_all_blocks: if True, return the features at the end of every block\n                           instead of just the features from the last block (eg for some prediction heads)\n        \"\"\"\n        # embed the image into patches  (x has size B x Npatches x C)\n        # and get position if each return patch (pos has size B x Npatches x 2)\n        x, pos = self.patch_embed(image)\n        # add positional embedding without cls token\n        if self.enc_pos_embed is not None:\n            x = x + self.enc_pos_embed[None, ...]\n        # apply masking\n        B, N, C = x.size()\n        if do_mask:\n            masks = self.mask_generator(x)\n            x = x[~masks].view(B, -1, C)\n            posvis = pos[~masks].view(B, -1, 2)\n        else:\n            B, N, C = x.size()\n            masks = torch.zeros((B, N), dtype=bool)\n            posvis = pos\n        # now apply the transformer encoder and normalization\n        if return_all_blocks:\n            out = []\n            for blk in self.enc_blocks:\n                x = blk(x, posvis)\n                out.append(x)\n            out[-1] = self.enc_norm(out[-1])\n            return out, pos, masks\n        else:\n            for blk in self.enc_blocks:\n                x = blk(x, posvis)\n            x = self.enc_norm(x)\n            return x, pos, masks\n\n    def _decoder(self, feat1, pos1, masks1, feat2, pos2, return_all_blocks=False):\n        \"\"\"\n        return_all_blocks: if True, return the features at the end of every block\n                           instead of just the features from the last block (eg for some prediction heads)\n\n        masks1 can be None => assume image1 fully visible\n        \"\"\"\n        # encoder to decoder layer\n        visf1 = self.decoder_embed(feat1)\n        f2 = self.decoder_embed(feat2)\n        # append masked tokens to the sequence\n        B, Nenc, C = visf1.size()\n        if masks1 is None:  # downstreams\n            f1_ = visf1\n        else:  # pretraining\n            Ntotal = masks1.size(1)\n            f1_ = self.mask_token.repeat(B, Ntotal, 1).to(dtype=visf1.dtype)\n            f1_[~masks1] = visf1.view(B * Nenc, C)\n        # add positional embedding\n        if self.dec_pos_embed is not None:\n            f1_ = f1_ + self.dec_pos_embed\n            f2 = f2 + self.dec_pos_embed\n        # apply Transformer blocks\n        out = f1_\n        out2 = f2\n        if return_all_blocks:\n            _out, out = out, []\n            for blk in self.dec_blocks:\n                _out, out2 = blk(_out, out2, pos1, pos2)\n                out.append(_out)\n            out[-1] = self.dec_norm(out[-1])\n        else:\n            for blk in self.dec_blocks:\n                out, out2 = blk(out, out2, pos1, pos2)\n            out = self.dec_norm(out)\n        return out\n\n    def patchify(self, imgs):\n        \"\"\"\n        imgs: (B, 3, H, W)\n        x: (B, L, patch_size**2 *3)\n        \"\"\"\n        p = self.patch_embed.patch_size[0]\n        assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0\n\n        h = w = imgs.shape[2] // p\n        x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))\n        x = torch.einsum(\"nchpwq->nhwpqc\", x)\n        x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))\n\n        return x\n\n    def unpatchify(self, x, channels=3):\n        \"\"\"\n        x: (N, L, patch_size**2 *channels)\n        imgs: (N, 3, H, W)\n        \"\"\"\n        patch_size = self.patch_embed.patch_size[0]\n        h = w = int(x.shape[1] ** 0.5)\n        assert h * w == x.shape[1]\n        x = x.reshape(shape=(x.shape[0], h, w, patch_size, patch_size, channels))\n        x = torch.einsum(\"nhwpqc->nchpwq\", x)\n        imgs = x.reshape(shape=(x.shape[0], channels, h * patch_size, h * patch_size))\n        return imgs\n\n    # def forward(self, img1, img2):\n    # \"\"\"\n    # img1: tensor of size B x 3 x img_size x img_size\n    # img2: tensor of size B x 3 x img_size x img_size\n\n    # out will be    B x N x (3*patch_size*patch_size)\n    # masks are also returned as B x N just in case\n    # \"\"\"\n    # # encoder of the masked first image\n    # feat1, pos1, mask1 = self._encode_image(img1, do_mask=True)\n    # # encoder of the second image\n    # feat2, pos2, _ = self._encode_image(img2, do_mask=False)\n    # # decoder\n    # decfeat = self._decoder(feat1, pos1, mask1, feat2, pos2)\n    # # prediction head\n    # out = self.prediction_head(decfeat)\n    # # get target\n    # target = self.patchify(img1)\n    # return out, mask1, target","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco.__init__","uri":"program://Human3R/function/src.croco.models.croco.__init__#L64-L145","kind":"function","name":"__init__","path":"src/croco/models/croco.py","language":"python","start_line":64,"end_line":145,"context_start_line":44,"context_end_line":165,"code":"        self.img_size = img_size\n        self.patch_size = patch_size\n        self.mask_ratio = mask_ratio\n        self.enc_embed_dim = enc_embed_dim\n        self.enc_depth = enc_depth\n        self.enc_num_heads = enc_num_heads\n        self.dec_embed_dim = dec_embed_dim\n        self.dec_depth = dec_depth\n        self.dec_num_heads = dec_num_heads\n        self.mlp_ratio = mlp_ratio\n        self.norm_layer = norm_layer\n        self.norm_im2_in_dec = norm_im2_in_dec\n        self.pos_embed = pos_embed\n\n\nclass CroCoNet(PreTrainedModel):\n\n    config_class = CrocoConfig\n    base_model_prefix = \"croco\"\n\n    def __init__(self, config: CrocoConfig):\n\n        super().__init__(config)\n\n        # patch embeddings  (with initialization done as in MAE)\n        self._set_patch_embed(config.img_size, config.patch_size, config.enc_embed_dim)\n\n        # mask generations\n        self._set_mask_generator(self.patch_embed.num_patches, config.mask_ratio)\n\n        self.pos_embed = config.pos_embed\n        if config.pos_embed == \"cosine\":\n            # positional embedding of the encoder\n            enc_pos_embed = get_2d_sincos_pos_embed(\n                config.enc_embed_dim,\n                int(self.patch_embed.num_patches**0.5),\n                n_cls_token=0,\n            )\n            self.register_buffer(\n                \"enc_pos_embed\", torch.from_numpy(enc_pos_embed).float()\n            )\n            # positional embedding of the decoder\n            dec_pos_embed = get_2d_sincos_pos_embed(\n                config.dec_embed_dim,\n                int(self.patch_embed.num_patches**0.5),\n                n_cls_token=0,\n            )\n            self.register_buffer(\n                \"dec_pos_embed\", torch.from_numpy(dec_pos_embed).float()\n            )\n            # pos embedding in each block\n            self.rope = None  # nothing for cosine\n        elif config.pos_embed.startswith(\"RoPE\"):  # eg RoPE100\n            self.enc_pos_embed = None  # nothing to add in the encoder with RoPE\n            self.dec_pos_embed = None  # nothing to add in the decoder with RoPE\n            if RoPE2D is None:\n                raise ImportError(\n                    \"Cannot find cuRoPE2D, please install it following the README instructions\"\n                )\n            freq = float(config.pos_embed[len(\"RoPE\") :])\n            self.rope = RoPE2D(freq=freq)\n        else:\n            raise NotImplementedError(\"Unknown pos_embed \" + config.pos_embed)\n\n        # transformer for the encoder\n        self.enc_depth = config.enc_depth\n        self.enc_embed_dim = config.enc_embed_dim\n        self.enc_blocks = nn.ModuleList(\n            [\n                Block(\n                    config.enc_embed_dim,\n                    config.enc_num_heads,\n                    config.mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=config.norm_layer,\n                    rope=self.rope,\n                )\n                for i in range(config.enc_depth)\n            ]\n        )\n        self.enc_norm = config.norm_layer(config.enc_embed_dim)\n\n        # masked tokens\n        # self._set_mask_token(config.dec_embed_dim)\n        self.mask_token = None\n\n        # decoder\n        self._set_decoder(\n            config.enc_embed_dim,\n            config.dec_embed_dim,\n            config.dec_num_heads,\n            config.dec_depth,\n            config.mlp_ratio,\n            config.norm_layer,\n            config.norm_im2_in_dec,\n        )\n\n        # prediction head\n        self._set_prediction_head(config.dec_embed_dim, config.patch_size)\n\n        # initializer weights\n        self.initialize_weights()\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = PatchEmbed(img_size, patch_size, 3, enc_embed_dim)\n\n    def _set_mask_generator(self, num_patches, mask_ratio):\n        self.mask_generator = RandomMask(num_patches, mask_ratio)\n\n    def _set_mask_token(self, dec_embed_dim):\n        self.mask_token = nn.Parameter(torch.zeros(1, 1, dec_embed_dim))\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco._set_patch_embed","uri":"program://Human3R/function/src.croco.models.croco._set_patch_embed#L147-L148","kind":"function","name":"_set_patch_embed","path":"src/croco/models/croco.py","language":"python","start_line":147,"end_line":148,"context_start_line":127,"context_end_line":168,"code":"        # self._set_mask_token(config.dec_embed_dim)\n        self.mask_token = None\n\n        # decoder\n        self._set_decoder(\n            config.enc_embed_dim,\n            config.dec_embed_dim,\n            config.dec_num_heads,\n            config.dec_depth,\n            config.mlp_ratio,\n            config.norm_layer,\n            config.norm_im2_in_dec,\n        )\n\n        # prediction head\n        self._set_prediction_head(config.dec_embed_dim, config.patch_size)\n\n        # initializer weights\n        self.initialize_weights()\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = PatchEmbed(img_size, patch_size, 3, enc_embed_dim)\n\n    def _set_mask_generator(self, num_patches, mask_ratio):\n        self.mask_generator = RandomMask(num_patches, mask_ratio)\n\n    def _set_mask_token(self, dec_embed_dim):\n        self.mask_token = nn.Parameter(torch.zeros(1, 1, dec_embed_dim))\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        # transfer from encoder to decoder","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco._set_mask_generator","uri":"program://Human3R/function/src.croco.models.croco._set_mask_generator#L150-L151","kind":"function","name":"_set_mask_generator","path":"src/croco/models/croco.py","language":"python","start_line":150,"end_line":151,"context_start_line":130,"context_end_line":171,"code":"        # decoder\n        self._set_decoder(\n            config.enc_embed_dim,\n            config.dec_embed_dim,\n            config.dec_num_heads,\n            config.dec_depth,\n            config.mlp_ratio,\n            config.norm_layer,\n            config.norm_im2_in_dec,\n        )\n\n        # prediction head\n        self._set_prediction_head(config.dec_embed_dim, config.patch_size)\n\n        # initializer weights\n        self.initialize_weights()\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = PatchEmbed(img_size, patch_size, 3, enc_embed_dim)\n\n    def _set_mask_generator(self, num_patches, mask_ratio):\n        self.mask_generator = RandomMask(num_patches, mask_ratio)\n\n    def _set_mask_token(self, dec_embed_dim):\n        self.mask_token = nn.Parameter(torch.zeros(1, 1, dec_embed_dim))\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        # transfer from encoder to decoder\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        # transformer for the decoder\n        self.dec_blocks = nn.ModuleList(","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco._set_mask_token","uri":"program://Human3R/function/src.croco.models.croco._set_mask_token#L153-L154","kind":"function","name":"_set_mask_token","path":"src/croco/models/croco.py","language":"python","start_line":153,"end_line":154,"context_start_line":133,"context_end_line":174,"code":"            config.dec_embed_dim,\n            config.dec_num_heads,\n            config.dec_depth,\n            config.mlp_ratio,\n            config.norm_layer,\n            config.norm_im2_in_dec,\n        )\n\n        # prediction head\n        self._set_prediction_head(config.dec_embed_dim, config.patch_size)\n\n        # initializer weights\n        self.initialize_weights()\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = PatchEmbed(img_size, patch_size, 3, enc_embed_dim)\n\n    def _set_mask_generator(self, num_patches, mask_ratio):\n        self.mask_generator = RandomMask(num_patches, mask_ratio)\n\n    def _set_mask_token(self, dec_embed_dim):\n        self.mask_token = nn.Parameter(torch.zeros(1, 1, dec_embed_dim))\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        # transfer from encoder to decoder\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        # transformer for the decoder\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco._set_decoder","uri":"program://Human3R/function/src.croco.models.croco._set_decoder#L156-L186","kind":"function","name":"_set_decoder","path":"src/croco/models/croco.py","language":"python","start_line":156,"end_line":186,"context_start_line":136,"context_end_line":206,"code":"            config.mlp_ratio,\n            config.norm_layer,\n            config.norm_im2_in_dec,\n        )\n\n        # prediction head\n        self._set_prediction_head(config.dec_embed_dim, config.patch_size)\n\n        # initializer weights\n        self.initialize_weights()\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = PatchEmbed(img_size, patch_size, 3, enc_embed_dim)\n\n    def _set_mask_generator(self, num_patches, mask_ratio):\n        self.mask_generator = RandomMask(num_patches, mask_ratio)\n\n    def _set_mask_token(self, dec_embed_dim):\n        self.mask_token = nn.Parameter(torch.zeros(1, 1, dec_embed_dim))\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        # transfer from encoder to decoder\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        # transformer for the decoder\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        # final norm layer\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_prediction_head(self, dec_embed_dim, patch_size):\n        self.prediction_head = nn.Linear(dec_embed_dim, patch_size**2 * 3, bias=True)\n\n    def initialize_weights(self):\n        # patch embed\n        self.patch_embed._init_weights()\n        # mask tokens\n        if self.mask_token is not None:\n            torch.nn.init.normal_(self.mask_token, std=0.02)\n        # linears and layer norms\n        self.apply(self._init_weights)\n\n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            # we use xavier_uniform following official JAX ViT:\n            torch.nn.init.xavier_uniform_(m.weight)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco._set_prediction_head","uri":"program://Human3R/function/src.croco.models.croco._set_prediction_head#L188-L189","kind":"function","name":"_set_prediction_head","path":"src/croco/models/croco.py","language":"python","start_line":188,"end_line":189,"context_start_line":168,"context_end_line":209,"code":"        # transfer from encoder to decoder\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        # transformer for the decoder\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        # final norm layer\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_prediction_head(self, dec_embed_dim, patch_size):\n        self.prediction_head = nn.Linear(dec_embed_dim, patch_size**2 * 3, bias=True)\n\n    def initialize_weights(self):\n        # patch embed\n        self.patch_embed._init_weights()\n        # mask tokens\n        if self.mask_token is not None:\n            torch.nn.init.normal_(self.mask_token, std=0.02)\n        # linears and layer norms\n        self.apply(self._init_weights)\n\n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            # we use xavier_uniform following official JAX ViT:\n            torch.nn.init.xavier_uniform_(m.weight)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco.initialize_weights","uri":"program://Human3R/function/src.croco.models.croco.initialize_weights#L191-L198","kind":"function","name":"initialize_weights","path":"src/croco/models/croco.py","language":"python","start_line":191,"end_line":198,"context_start_line":171,"context_end_line":218,"code":"        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        # final norm layer\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_prediction_head(self, dec_embed_dim, patch_size):\n        self.prediction_head = nn.Linear(dec_embed_dim, patch_size**2 * 3, bias=True)\n\n    def initialize_weights(self):\n        # patch embed\n        self.patch_embed._init_weights()\n        # mask tokens\n        if self.mask_token is not None:\n            torch.nn.init.normal_(self.mask_token, std=0.02)\n        # linears and layer norms\n        self.apply(self._init_weights)\n\n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            # we use xavier_uniform following official JAX ViT:\n            torch.nn.init.xavier_uniform_(m.weight)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n\n    def _encode_image(self, image, do_mask=False, return_all_blocks=False):\n        \"\"\"\n        image has B x 3 x img_size x img_size\n        do_mask: whether to perform masking or not\n        return_all_blocks: if True, return the features at the end of every block\n                           instead of just the features from the last block (eg for some prediction heads)\n        \"\"\"\n        # embed the image into patches  (x has size B x Npatches x C)\n        # and get position if each return patch (pos has size B x Npatches x 2)","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco._init_weights","uri":"program://Human3R/function/src.croco.models.croco._init_weights#L200-L208","kind":"function","name":"_init_weights","path":"src/croco/models/croco.py","language":"python","start_line":200,"end_line":208,"context_start_line":180,"context_end_line":228,"code":"                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        # final norm layer\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_prediction_head(self, dec_embed_dim, patch_size):\n        self.prediction_head = nn.Linear(dec_embed_dim, patch_size**2 * 3, bias=True)\n\n    def initialize_weights(self):\n        # patch embed\n        self.patch_embed._init_weights()\n        # mask tokens\n        if self.mask_token is not None:\n            torch.nn.init.normal_(self.mask_token, std=0.02)\n        # linears and layer norms\n        self.apply(self._init_weights)\n\n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            # we use xavier_uniform following official JAX ViT:\n            torch.nn.init.xavier_uniform_(m.weight)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n\n    def _encode_image(self, image, do_mask=False, return_all_blocks=False):\n        \"\"\"\n        image has B x 3 x img_size x img_size\n        do_mask: whether to perform masking or not\n        return_all_blocks: if True, return the features at the end of every block\n                           instead of just the features from the last block (eg for some prediction heads)\n        \"\"\"\n        # embed the image into patches  (x has size B x Npatches x C)\n        # and get position if each return patch (pos has size B x Npatches x 2)\n        x, pos = self.patch_embed(image)\n        # add positional embedding without cls token\n        if self.enc_pos_embed is not None:\n            x = x + self.enc_pos_embed[None, ...]\n        # apply masking\n        B, N, C = x.size()\n        if do_mask:\n            masks = self.mask_generator(x)\n            x = x[~masks].view(B, -1, C)\n            posvis = pos[~masks].view(B, -1, 2)","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco._encode_image","uri":"program://Human3R/function/src.croco.models.croco._encode_image#L210-L245","kind":"function","name":"_encode_image","path":"src/croco/models/croco.py","language":"python","start_line":210,"end_line":245,"context_start_line":190,"context_end_line":265,"code":"\n    def initialize_weights(self):\n        # patch embed\n        self.patch_embed._init_weights()\n        # mask tokens\n        if self.mask_token is not None:\n            torch.nn.init.normal_(self.mask_token, std=0.02)\n        # linears and layer norms\n        self.apply(self._init_weights)\n\n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            # we use xavier_uniform following official JAX ViT:\n            torch.nn.init.xavier_uniform_(m.weight)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n\n    def _encode_image(self, image, do_mask=False, return_all_blocks=False):\n        \"\"\"\n        image has B x 3 x img_size x img_size\n        do_mask: whether to perform masking or not\n        return_all_blocks: if True, return the features at the end of every block\n                           instead of just the features from the last block (eg for some prediction heads)\n        \"\"\"\n        # embed the image into patches  (x has size B x Npatches x C)\n        # and get position if each return patch (pos has size B x Npatches x 2)\n        x, pos = self.patch_embed(image)\n        # add positional embedding without cls token\n        if self.enc_pos_embed is not None:\n            x = x + self.enc_pos_embed[None, ...]\n        # apply masking\n        B, N, C = x.size()\n        if do_mask:\n            masks = self.mask_generator(x)\n            x = x[~masks].view(B, -1, C)\n            posvis = pos[~masks].view(B, -1, 2)\n        else:\n            B, N, C = x.size()\n            masks = torch.zeros((B, N), dtype=bool)\n            posvis = pos\n        # now apply the transformer encoder and normalization\n        if return_all_blocks:\n            out = []\n            for blk in self.enc_blocks:\n                x = blk(x, posvis)\n                out.append(x)\n            out[-1] = self.enc_norm(out[-1])\n            return out, pos, masks\n        else:\n            for blk in self.enc_blocks:\n                x = blk(x, posvis)\n            x = self.enc_norm(x)\n            return x, pos, masks\n\n    def _decoder(self, feat1, pos1, masks1, feat2, pos2, return_all_blocks=False):\n        \"\"\"\n        return_all_blocks: if True, return the features at the end of every block\n                           instead of just the features from the last block (eg for some prediction heads)\n\n        masks1 can be None => assume image1 fully visible\n        \"\"\"\n        # encoder to decoder layer\n        visf1 = self.decoder_embed(feat1)\n        f2 = self.decoder_embed(feat2)\n        # append masked tokens to the sequence\n        B, Nenc, C = visf1.size()\n        if masks1 is None:  # downstreams\n            f1_ = visf1\n        else:  # pretraining\n            Ntotal = masks1.size(1)\n            f1_ = self.mask_token.repeat(B, Ntotal, 1).to(dtype=visf1.dtype)\n            f1_[~masks1] = visf1.view(B * Nenc, C)\n        # add positional embedding","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco._decoder","uri":"program://Human3R/function/src.croco.models.croco._decoder#L247-L282","kind":"function","name":"_decoder","path":"src/croco/models/croco.py","language":"python","start_line":247,"end_line":282,"context_start_line":227,"context_end_line":302,"code":"            x = x[~masks].view(B, -1, C)\n            posvis = pos[~masks].view(B, -1, 2)\n        else:\n            B, N, C = x.size()\n            masks = torch.zeros((B, N), dtype=bool)\n            posvis = pos\n        # now apply the transformer encoder and normalization\n        if return_all_blocks:\n            out = []\n            for blk in self.enc_blocks:\n                x = blk(x, posvis)\n                out.append(x)\n            out[-1] = self.enc_norm(out[-1])\n            return out, pos, masks\n        else:\n            for blk in self.enc_blocks:\n                x = blk(x, posvis)\n            x = self.enc_norm(x)\n            return x, pos, masks\n\n    def _decoder(self, feat1, pos1, masks1, feat2, pos2, return_all_blocks=False):\n        \"\"\"\n        return_all_blocks: if True, return the features at the end of every block\n                           instead of just the features from the last block (eg for some prediction heads)\n\n        masks1 can be None => assume image1 fully visible\n        \"\"\"\n        # encoder to decoder layer\n        visf1 = self.decoder_embed(feat1)\n        f2 = self.decoder_embed(feat2)\n        # append masked tokens to the sequence\n        B, Nenc, C = visf1.size()\n        if masks1 is None:  # downstreams\n            f1_ = visf1\n        else:  # pretraining\n            Ntotal = masks1.size(1)\n            f1_ = self.mask_token.repeat(B, Ntotal, 1).to(dtype=visf1.dtype)\n            f1_[~masks1] = visf1.view(B * Nenc, C)\n        # add positional embedding\n        if self.dec_pos_embed is not None:\n            f1_ = f1_ + self.dec_pos_embed\n            f2 = f2 + self.dec_pos_embed\n        # apply Transformer blocks\n        out = f1_\n        out2 = f2\n        if return_all_blocks:\n            _out, out = out, []\n            for blk in self.dec_blocks:\n                _out, out2 = blk(_out, out2, pos1, pos2)\n                out.append(_out)\n            out[-1] = self.dec_norm(out[-1])\n        else:\n            for blk in self.dec_blocks:\n                out, out2 = blk(out, out2, pos1, pos2)\n            out = self.dec_norm(out)\n        return out\n\n    def patchify(self, imgs):\n        \"\"\"\n        imgs: (B, 3, H, W)\n        x: (B, L, patch_size**2 *3)\n        \"\"\"\n        p = self.patch_embed.patch_size[0]\n        assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0\n\n        h = w = imgs.shape[2] // p\n        x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))\n        x = torch.einsum(\"nchpwq->nhwpqc\", x)\n        x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))\n\n        return x\n\n    def unpatchify(self, x, channels=3):\n        \"\"\"\n        x: (N, L, patch_size**2 *channels)\n        imgs: (N, 3, H, W)","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco.patchify","uri":"program://Human3R/function/src.croco.models.croco.patchify#L284-L297","kind":"function","name":"patchify","path":"src/croco/models/croco.py","language":"python","start_line":284,"end_line":297,"context_start_line":264,"context_end_line":317,"code":"            f1_[~masks1] = visf1.view(B * Nenc, C)\n        # add positional embedding\n        if self.dec_pos_embed is not None:\n            f1_ = f1_ + self.dec_pos_embed\n            f2 = f2 + self.dec_pos_embed\n        # apply Transformer blocks\n        out = f1_\n        out2 = f2\n        if return_all_blocks:\n            _out, out = out, []\n            for blk in self.dec_blocks:\n                _out, out2 = blk(_out, out2, pos1, pos2)\n                out.append(_out)\n            out[-1] = self.dec_norm(out[-1])\n        else:\n            for blk in self.dec_blocks:\n                out, out2 = blk(out, out2, pos1, pos2)\n            out = self.dec_norm(out)\n        return out\n\n    def patchify(self, imgs):\n        \"\"\"\n        imgs: (B, 3, H, W)\n        x: (B, L, patch_size**2 *3)\n        \"\"\"\n        p = self.patch_embed.patch_size[0]\n        assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0\n\n        h = w = imgs.shape[2] // p\n        x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))\n        x = torch.einsum(\"nchpwq->nhwpqc\", x)\n        x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))\n\n        return x\n\n    def unpatchify(self, x, channels=3):\n        \"\"\"\n        x: (N, L, patch_size**2 *channels)\n        imgs: (N, 3, H, W)\n        \"\"\"\n        patch_size = self.patch_embed.patch_size[0]\n        h = w = int(x.shape[1] ** 0.5)\n        assert h * w == x.shape[1]\n        x = x.reshape(shape=(x.shape[0], h, w, patch_size, patch_size, channels))\n        x = torch.einsum(\"nhwpqc->nchpwq\", x)\n        imgs = x.reshape(shape=(x.shape[0], channels, h * patch_size, h * patch_size))\n        return imgs\n\n    # def forward(self, img1, img2):\n    # \"\"\"\n    # img1: tensor of size B x 3 x img_size x img_size\n    # img2: tensor of size B x 3 x img_size x img_size\n\n    # out will be    B x N x (3*patch_size*patch_size)","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.croco.unpatchify","uri":"program://Human3R/function/src.croco.models.croco.unpatchify#L299-L310","kind":"function","name":"unpatchify","path":"src/croco/models/croco.py","language":"python","start_line":299,"end_line":310,"context_start_line":279,"context_end_line":330,"code":"            for blk in self.dec_blocks:\n                out, out2 = blk(out, out2, pos1, pos2)\n            out = self.dec_norm(out)\n        return out\n\n    def patchify(self, imgs):\n        \"\"\"\n        imgs: (B, 3, H, W)\n        x: (B, L, patch_size**2 *3)\n        \"\"\"\n        p = self.patch_embed.patch_size[0]\n        assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0\n\n        h = w = imgs.shape[2] // p\n        x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))\n        x = torch.einsum(\"nchpwq->nhwpqc\", x)\n        x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))\n\n        return x\n\n    def unpatchify(self, x, channels=3):\n        \"\"\"\n        x: (N, L, patch_size**2 *channels)\n        imgs: (N, 3, H, W)\n        \"\"\"\n        patch_size = self.patch_embed.patch_size[0]\n        h = w = int(x.shape[1] ** 0.5)\n        assert h * w == x.shape[1]\n        x = x.reshape(shape=(x.shape[0], h, w, patch_size, patch_size, channels))\n        x = torch.einsum(\"nhwpqc->nchpwq\", x)\n        imgs = x.reshape(shape=(x.shape[0], channels, h * patch_size, h * patch_size))\n        return imgs\n\n    # def forward(self, img1, img2):\n    # \"\"\"\n    # img1: tensor of size B x 3 x img_size x img_size\n    # img2: tensor of size B x 3 x img_size x img_size\n\n    # out will be    B x N x (3*patch_size*patch_size)\n    # masks are also returned as B x N just in case\n    # \"\"\"\n    # # encoder of the masked first image\n    # feat1, pos1, mask1 = self._encode_image(img1, do_mask=True)\n    # # encoder of the second image\n    # feat2, pos2, _ = self._encode_image(img2, do_mask=False)\n    # # decoder\n    # decfeat = self._decoder(feat1, pos1, mask1, feat2, pos2)\n    # # prediction head\n    # out = self.prediction_head(decfeat)\n    # # get target\n    # target = self.patchify(img1)\n    # return out, mask1, target","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.head_downstream","uri":"program://Human3R/module/src.croco.models.head_downstream#L1-L83","kind":"module","name":"src.croco.models.head_downstream","path":"src/croco/models/head_downstream.py","language":"python","start_line":1,"end_line":83,"context_start_line":1,"context_end_line":83,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Heads for downstream tasks\n# --------------------------------------------------------\n\n\"\"\"\nA head is a module where the __init__ defines only the head hyperparameters.\nA method setup(croconet) takes a CroCoNet and set all layers according to the head and croconet attributes.\nThe forward takes the features as well as a dictionary img_info containing the keys 'width' and 'height'\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom .dpt_block import DPTOutputAdapter\n\n\nclass PixelwiseTaskWithDPT(nn.Module):\n    \"\"\"DPT module for CroCo.\n    by default, hooks_idx will be equal to:\n    * for encoder-only: 4 equally spread layers\n    * for encoder+decoder: last encoder + 3 equally spread layers of the decoder\n    \"\"\"\n\n    def __init__(\n        self,\n        *,\n        hooks_idx=None,\n        layer_dims=[96, 192, 384, 768],\n        output_width_ratio=1,\n        num_channels=1,\n        postprocess=None,\n        **kwargs,\n    ):\n        super(PixelwiseTaskWithDPT, self).__init__()\n        self.return_all_blocks = True  # backbone needs to return all layers\n        self.postprocess = postprocess\n        self.output_width_ratio = output_width_ratio\n        self.num_channels = num_channels\n        self.hooks_idx = hooks_idx\n        self.layer_dims = layer_dims\n\n    def setup(self, croconet):\n        dpt_args = {\n            \"output_width_ratio\": self.output_width_ratio,\n            \"num_channels\": self.num_channels,\n        }\n        if self.hooks_idx is None:\n            if hasattr(croconet, \"dec_blocks\"):  # encoder + decoder\n                step = {8: 3, 12: 4, 24: 8}[croconet.dec_depth]\n                hooks_idx = [\n                    croconet.dec_depth + croconet.enc_depth - 1 - i * step\n                    for i in range(3, -1, -1)\n                ]\n            else:  # encoder only\n                step = croconet.enc_depth // 4\n                hooks_idx = [\n                    croconet.enc_depth - 1 - i * step for i in range(3, -1, -1)\n                ]\n            self.hooks_idx = hooks_idx\n            print(\n                f\"  PixelwiseTaskWithDPT: automatically setting hook_idxs={self.hooks_idx}\"\n            )\n        dpt_args[\"hooks\"] = self.hooks_idx\n        dpt_args[\"layer_dims\"] = self.layer_dims\n        self.dpt = DPTOutputAdapter(**dpt_args)\n        dim_tokens = [\n            (\n                croconet.enc_embed_dim\n                if hook < croconet.enc_depth\n                else croconet.dec_embed_dim\n            )\n            for hook in self.hooks_idx\n        ]\n        dpt_init_args = {\"dim_tokens_enc\": dim_tokens}\n        self.dpt.init(**dpt_init_args)\n\n    def forward(self, x, img_info):\n        out = self.dpt(x, image_size=(img_info[\"height\"], img_info[\"width\"]))\n        if self.postprocess:\n            out = self.postprocess(out)\n        return out","source_hash":"b02fdbfeb55dd3ffa53c2442a566df2f33f43bda7ffb84237665ef78f615dad8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.head_downstream.PixelwiseTaskWithDPT","uri":"program://Human3R/class/src.croco.models.head_downstream.PixelwiseTaskWithDPT#L19-L83","kind":"class","name":"PixelwiseTaskWithDPT","path":"src/croco/models/head_downstream.py","language":"python","start_line":19,"end_line":83,"context_start_line":1,"context_end_line":83,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Heads for downstream tasks\n# --------------------------------------------------------\n\n\"\"\"\nA head is a module where the __init__ defines only the head hyperparameters.\nA method setup(croconet) takes a CroCoNet and set all layers according to the head and croconet attributes.\nThe forward takes the features as well as a dictionary img_info containing the keys 'width' and 'height'\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom .dpt_block import DPTOutputAdapter\n\n\nclass PixelwiseTaskWithDPT(nn.Module):\n    \"\"\"DPT module for CroCo.\n    by default, hooks_idx will be equal to:\n    * for encoder-only: 4 equally spread layers\n    * for encoder+decoder: last encoder + 3 equally spread layers of the decoder\n    \"\"\"\n\n    def __init__(\n        self,\n        *,\n        hooks_idx=None,\n        layer_dims=[96, 192, 384, 768],\n        output_width_ratio=1,\n        num_channels=1,\n        postprocess=None,\n        **kwargs,\n    ):\n        super(PixelwiseTaskWithDPT, self).__init__()\n        self.return_all_blocks = True  # backbone needs to return all layers\n        self.postprocess = postprocess\n        self.output_width_ratio = output_width_ratio\n        self.num_channels = num_channels\n        self.hooks_idx = hooks_idx\n        self.layer_dims = layer_dims\n\n    def setup(self, croconet):\n        dpt_args = {\n            \"output_width_ratio\": self.output_width_ratio,\n            \"num_channels\": self.num_channels,\n        }\n        if self.hooks_idx is None:\n            if hasattr(croconet, \"dec_blocks\"):  # encoder + decoder\n                step = {8: 3, 12: 4, 24: 8}[croconet.dec_depth]\n                hooks_idx = [\n                    croconet.dec_depth + croconet.enc_depth - 1 - i * step\n                    for i in range(3, -1, -1)\n                ]\n            else:  # encoder only\n                step = croconet.enc_depth // 4\n                hooks_idx = [\n                    croconet.enc_depth - 1 - i * step for i in range(3, -1, -1)\n                ]\n            self.hooks_idx = hooks_idx\n            print(\n                f\"  PixelwiseTaskWithDPT: automatically setting hook_idxs={self.hooks_idx}\"\n            )\n        dpt_args[\"hooks\"] = self.hooks_idx\n        dpt_args[\"layer_dims\"] = self.layer_dims\n        self.dpt = DPTOutputAdapter(**dpt_args)\n        dim_tokens = [\n            (\n                croconet.enc_embed_dim\n                if hook < croconet.enc_depth\n                else croconet.dec_embed_dim\n            )\n            for hook in self.hooks_idx\n        ]\n        dpt_init_args = {\"dim_tokens_enc\": dim_tokens}\n        self.dpt.init(**dpt_init_args)\n\n    def forward(self, x, img_info):\n        out = self.dpt(x, image_size=(img_info[\"height\"], img_info[\"width\"]))\n        if self.postprocess:\n            out = self.postprocess(out)\n        return out","source_hash":"b02fdbfeb55dd3ffa53c2442a566df2f33f43bda7ffb84237665ef78f615dad8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.head_downstream.__init__","uri":"program://Human3R/function/src.croco.models.head_downstream.__init__#L26-L42","kind":"function","name":"__init__","path":"src/croco/models/head_downstream.py","language":"python","start_line":26,"end_line":42,"context_start_line":6,"context_end_line":62,"code":"# --------------------------------------------------------\n\n\"\"\"\nA head is a module where the __init__ defines only the head hyperparameters.\nA method setup(croconet) takes a CroCoNet and set all layers according to the head and croconet attributes.\nThe forward takes the features as well as a dictionary img_info containing the keys 'width' and 'height'\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom .dpt_block import DPTOutputAdapter\n\n\nclass PixelwiseTaskWithDPT(nn.Module):\n    \"\"\"DPT module for CroCo.\n    by default, hooks_idx will be equal to:\n    * for encoder-only: 4 equally spread layers\n    * for encoder+decoder: last encoder + 3 equally spread layers of the decoder\n    \"\"\"\n\n    def __init__(\n        self,\n        *,\n        hooks_idx=None,\n        layer_dims=[96, 192, 384, 768],\n        output_width_ratio=1,\n        num_channels=1,\n        postprocess=None,\n        **kwargs,\n    ):\n        super(PixelwiseTaskWithDPT, self).__init__()\n        self.return_all_blocks = True  # backbone needs to return all layers\n        self.postprocess = postprocess\n        self.output_width_ratio = output_width_ratio\n        self.num_channels = num_channels\n        self.hooks_idx = hooks_idx\n        self.layer_dims = layer_dims\n\n    def setup(self, croconet):\n        dpt_args = {\n            \"output_width_ratio\": self.output_width_ratio,\n            \"num_channels\": self.num_channels,\n        }\n        if self.hooks_idx is None:\n            if hasattr(croconet, \"dec_blocks\"):  # encoder + decoder\n                step = {8: 3, 12: 4, 24: 8}[croconet.dec_depth]\n                hooks_idx = [\n                    croconet.dec_depth + croconet.enc_depth - 1 - i * step\n                    for i in range(3, -1, -1)\n                ]\n            else:  # encoder only\n                step = croconet.enc_depth // 4\n                hooks_idx = [\n                    croconet.enc_depth - 1 - i * step for i in range(3, -1, -1)\n                ]\n            self.hooks_idx = hooks_idx\n            print(","source_hash":"b02fdbfeb55dd3ffa53c2442a566df2f33f43bda7ffb84237665ef78f615dad8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.head_downstream.setup","uri":"program://Human3R/function/src.croco.models.head_downstream.setup#L44-L77","kind":"function","name":"setup","path":"src/croco/models/head_downstream.py","language":"python","start_line":44,"end_line":77,"context_start_line":24,"context_end_line":83,"code":"    \"\"\"\n\n    def __init__(\n        self,\n        *,\n        hooks_idx=None,\n        layer_dims=[96, 192, 384, 768],\n        output_width_ratio=1,\n        num_channels=1,\n        postprocess=None,\n        **kwargs,\n    ):\n        super(PixelwiseTaskWithDPT, self).__init__()\n        self.return_all_blocks = True  # backbone needs to return all layers\n        self.postprocess = postprocess\n        self.output_width_ratio = output_width_ratio\n        self.num_channels = num_channels\n        self.hooks_idx = hooks_idx\n        self.layer_dims = layer_dims\n\n    def setup(self, croconet):\n        dpt_args = {\n            \"output_width_ratio\": self.output_width_ratio,\n            \"num_channels\": self.num_channels,\n        }\n        if self.hooks_idx is None:\n            if hasattr(croconet, \"dec_blocks\"):  # encoder + decoder\n                step = {8: 3, 12: 4, 24: 8}[croconet.dec_depth]\n                hooks_idx = [\n                    croconet.dec_depth + croconet.enc_depth - 1 - i * step\n                    for i in range(3, -1, -1)\n                ]\n            else:  # encoder only\n                step = croconet.enc_depth // 4\n                hooks_idx = [\n                    croconet.enc_depth - 1 - i * step for i in range(3, -1, -1)\n                ]\n            self.hooks_idx = hooks_idx\n            print(\n                f\"  PixelwiseTaskWithDPT: automatically setting hook_idxs={self.hooks_idx}\"\n            )\n        dpt_args[\"hooks\"] = self.hooks_idx\n        dpt_args[\"layer_dims\"] = self.layer_dims\n        self.dpt = DPTOutputAdapter(**dpt_args)\n        dim_tokens = [\n            (\n                croconet.enc_embed_dim\n                if hook < croconet.enc_depth\n                else croconet.dec_embed_dim\n            )\n            for hook in self.hooks_idx\n        ]\n        dpt_init_args = {\"dim_tokens_enc\": dim_tokens}\n        self.dpt.init(**dpt_init_args)\n\n    def forward(self, x, img_info):\n        out = self.dpt(x, image_size=(img_info[\"height\"], img_info[\"width\"]))\n        if self.postprocess:\n            out = self.postprocess(out)\n        return out","source_hash":"b02fdbfeb55dd3ffa53c2442a566df2f33f43bda7ffb84237665ef78f615dad8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.head_downstream.forward","uri":"program://Human3R/function/src.croco.models.head_downstream.forward#L79-L83","kind":"function","name":"forward","path":"src/croco/models/head_downstream.py","language":"python","start_line":79,"end_line":83,"context_start_line":59,"context_end_line":83,"code":"                    croconet.enc_depth - 1 - i * step for i in range(3, -1, -1)\n                ]\n            self.hooks_idx = hooks_idx\n            print(\n                f\"  PixelwiseTaskWithDPT: automatically setting hook_idxs={self.hooks_idx}\"\n            )\n        dpt_args[\"hooks\"] = self.hooks_idx\n        dpt_args[\"layer_dims\"] = self.layer_dims\n        self.dpt = DPTOutputAdapter(**dpt_args)\n        dim_tokens = [\n            (\n                croconet.enc_embed_dim\n                if hook < croconet.enc_depth\n                else croconet.dec_embed_dim\n            )\n            for hook in self.hooks_idx\n        ]\n        dpt_init_args = {\"dim_tokens_enc\": dim_tokens}\n        self.dpt.init(**dpt_init_args)\n\n    def forward(self, x, img_info):\n        out = self.dpt(x, image_size=(img_info[\"height\"], img_info[\"width\"]))\n        if self.postprocess:\n            out = self.postprocess(out)\n        return out","source_hash":"b02fdbfeb55dd3ffa53c2442a566df2f33f43bda7ffb84237665ef78f615dad8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.masking","uri":"program://Human3R/module/src.croco.models.masking#L1-L26","kind":"module","name":"src.croco.models.masking","path":"src/croco/models/masking.py","language":"python","start_line":1,"end_line":26,"context_start_line":1,"context_end_line":26,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# Masking utils\n# --------------------------------------------------------\n\nimport torch\nimport torch.nn as nn\n\n\nclass RandomMask(nn.Module):\n    \"\"\"\n    random masking\n    \"\"\"\n\n    def __init__(self, num_patches, mask_ratio):\n        super().__init__()\n        self.num_patches = num_patches\n        self.num_mask = int(mask_ratio * self.num_patches)\n\n    def __call__(self, x):\n        noise = torch.rand(x.size(0), self.num_patches, device=x.device)\n        argsort = torch.argsort(noise, dim=1)\n        return argsort < self.num_mask","source_hash":"541cf1fac15d432d40fd89502e820103309d59aa1bb8e54d1088aea1fe72a460","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.masking.RandomMask","uri":"program://Human3R/class/src.croco.models.masking.RandomMask#L13-L26","kind":"class","name":"RandomMask","path":"src/croco/models/masking.py","language":"python","start_line":13,"end_line":26,"context_start_line":1,"context_end_line":26,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# Masking utils\n# --------------------------------------------------------\n\nimport torch\nimport torch.nn as nn\n\n\nclass RandomMask(nn.Module):\n    \"\"\"\n    random masking\n    \"\"\"\n\n    def __init__(self, num_patches, mask_ratio):\n        super().__init__()\n        self.num_patches = num_patches\n        self.num_mask = int(mask_ratio * self.num_patches)\n\n    def __call__(self, x):\n        noise = torch.rand(x.size(0), self.num_patches, device=x.device)\n        argsort = torch.argsort(noise, dim=1)\n        return argsort < self.num_mask","source_hash":"541cf1fac15d432d40fd89502e820103309d59aa1bb8e54d1088aea1fe72a460","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.masking.__init__","uri":"program://Human3R/function/src.croco.models.masking.__init__#L18-L21","kind":"function","name":"__init__","path":"src/croco/models/masking.py","language":"python","start_line":18,"end_line":21,"context_start_line":1,"context_end_line":26,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# Masking utils\n# --------------------------------------------------------\n\nimport torch\nimport torch.nn as nn\n\n\nclass RandomMask(nn.Module):\n    \"\"\"\n    random masking\n    \"\"\"\n\n    def __init__(self, num_patches, mask_ratio):\n        super().__init__()\n        self.num_patches = num_patches\n        self.num_mask = int(mask_ratio * self.num_patches)\n\n    def __call__(self, x):\n        noise = torch.rand(x.size(0), self.num_patches, device=x.device)\n        argsort = torch.argsort(noise, dim=1)\n        return argsort < self.num_mask","source_hash":"541cf1fac15d432d40fd89502e820103309d59aa1bb8e54d1088aea1fe72a460","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.masking.__call__","uri":"program://Human3R/function/src.croco.models.masking.__call__#L23-L26","kind":"function","name":"__call__","path":"src/croco/models/masking.py","language":"python","start_line":23,"end_line":26,"context_start_line":3,"context_end_line":26,"code":"\n\n# --------------------------------------------------------\n# Masking utils\n# --------------------------------------------------------\n\nimport torch\nimport torch.nn as nn\n\n\nclass RandomMask(nn.Module):\n    \"\"\"\n    random masking\n    \"\"\"\n\n    def __init__(self, num_patches, mask_ratio):\n        super().__init__()\n        self.num_patches = num_patches\n        self.num_mask = int(mask_ratio * self.num_patches)\n\n    def __call__(self, x):\n        noise = torch.rand(x.size(0), self.num_patches, device=x.device)\n        argsort = torch.argsort(noise, dim=1)\n        return argsort < self.num_mask","source_hash":"541cf1fac15d432d40fd89502e820103309d59aa1bb8e54d1088aea1fe72a460","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.curope.setup","uri":"program://Human3R/module/src.croco.models.curope.setup#L1-L34","kind":"module","name":"src.croco.models.curope.setup","path":"src/croco/models/curope/setup.py","language":"python","start_line":1,"end_line":34,"context_start_line":1,"context_end_line":34,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nfrom setuptools import setup\nfrom torch import cuda\nfrom torch.utils.cpp_extension import BuildExtension, CUDAExtension\n\n# compile for all possible CUDA architectures\nall_cuda_archs = cuda.get_gencode_flags().replace(\"compute=\", \"arch=\").split()\n# alternatively, you can list cuda archs that you want, eg:\n# all_cuda_archs = [\n# '-gencode', 'arch=compute_70,code=sm_70',\n# '-gencode', 'arch=compute_75,code=sm_75',\n# '-gencode', 'arch=compute_80,code=sm_80',\n# '-gencode', 'arch=compute_86,code=sm_86'\n# ]\n\nsetup(\n    name=\"curope\",\n    ext_modules=[\n        CUDAExtension(\n            name=\"curope\",\n            sources=[\n                \"curope.cpp\",\n                \"kernels.cu\",\n            ],\n            extra_compile_args=dict(\n                nvcc=[\"-O3\", \"--ptxas-options=-v\", \"--use_fast_math\"] + all_cuda_archs,\n                cxx=[\"-O3\"],\n            ),\n        )\n    ],\n    cmdclass={\"build_ext\": BuildExtension},\n)","source_hash":"7b85998c132be2ece2eef307d088355e73c39b326c874518862b34f1fd209057","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.curope.curope2d","uri":"program://Human3R/module/src.croco.models.curope.curope2d#L1-L40","kind":"module","name":"src.croco.models.curope.curope2d","path":"src/croco/models/curope/curope2d.py","language":"python","start_line":1,"end_line":40,"context_start_line":1,"context_end_line":40,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport torch\n\ntry:\n    import curope as _kernels  # run `python setup.py install`\nexcept ModuleNotFoundError:\n    from . import curope as _kernels  # run `python setup.py build_ext --inplace`\n\n\nclass cuRoPE2D_func(torch.autograd.Function):\n\n    @staticmethod\n    def forward(ctx, tokens, positions, base, F0=1):\n        ctx.save_for_backward(positions)\n        ctx.saved_base = base\n        ctx.saved_F0 = F0\n        # tokens = tokens.clone() # uncomment this if inplace doesn't work\n        _kernels.rope_2d(tokens, positions, base, F0)\n        ctx.mark_dirty(tokens)\n        return tokens\n\n    @staticmethod\n    def backward(ctx, grad_res):\n        positions, base, F0 = ctx.saved_tensors[0], ctx.saved_base, ctx.saved_F0\n        _kernels.rope_2d(grad_res, positions, base, -F0)\n        ctx.mark_dirty(grad_res)\n        return grad_res, None, None, None\n\n\nclass cuRoPE2D(torch.nn.Module):\n    def __init__(self, freq=100.0, F0=1.0):\n        super().__init__()\n        self.base = freq\n        self.F0 = F0\n\n    def forward(self, tokens, positions):\n        cuRoPE2D_func.apply(tokens.transpose(1, 2), positions, self.base, self.F0)\n        return tokens","source_hash":"dc49348fc210445a9b12252e1faf2ec678346e3ae3ef9735867eb39556739377","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.curope.curope2d.cuRoPE2D_func","uri":"program://Human3R/class/src.croco.models.curope.curope2d.cuRoPE2D_func#L12-L29","kind":"class","name":"cuRoPE2D_func","path":"src/croco/models/curope/curope2d.py","language":"python","start_line":12,"end_line":29,"context_start_line":1,"context_end_line":40,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport torch\n\ntry:\n    import curope as _kernels  # run `python setup.py install`\nexcept ModuleNotFoundError:\n    from . import curope as _kernels  # run `python setup.py build_ext --inplace`\n\n\nclass cuRoPE2D_func(torch.autograd.Function):\n\n    @staticmethod\n    def forward(ctx, tokens, positions, base, F0=1):\n        ctx.save_for_backward(positions)\n        ctx.saved_base = base\n        ctx.saved_F0 = F0\n        # tokens = tokens.clone() # uncomment this if inplace doesn't work\n        _kernels.rope_2d(tokens, positions, base, F0)\n        ctx.mark_dirty(tokens)\n        return tokens\n\n    @staticmethod\n    def backward(ctx, grad_res):\n        positions, base, F0 = ctx.saved_tensors[0], ctx.saved_base, ctx.saved_F0\n        _kernels.rope_2d(grad_res, positions, base, -F0)\n        ctx.mark_dirty(grad_res)\n        return grad_res, None, None, None\n\n\nclass cuRoPE2D(torch.nn.Module):\n    def __init__(self, freq=100.0, F0=1.0):\n        super().__init__()\n        self.base = freq\n        self.F0 = F0\n\n    def forward(self, tokens, positions):\n        cuRoPE2D_func.apply(tokens.transpose(1, 2), positions, self.base, self.F0)\n        return tokens","source_hash":"dc49348fc210445a9b12252e1faf2ec678346e3ae3ef9735867eb39556739377","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.curope.curope2d.cuRoPE2D","uri":"program://Human3R/class/src.croco.models.curope.curope2d.cuRoPE2D#L32-L40","kind":"class","name":"cuRoPE2D","path":"src/croco/models/curope/curope2d.py","language":"python","start_line":32,"end_line":40,"context_start_line":12,"context_end_line":40,"code":"class cuRoPE2D_func(torch.autograd.Function):\n\n    @staticmethod\n    def forward(ctx, tokens, positions, base, F0=1):\n        ctx.save_for_backward(positions)\n        ctx.saved_base = base\n        ctx.saved_F0 = F0\n        # tokens = tokens.clone() # uncomment this if inplace doesn't work\n        _kernels.rope_2d(tokens, positions, base, F0)\n        ctx.mark_dirty(tokens)\n        return tokens\n\n    @staticmethod\n    def backward(ctx, grad_res):\n        positions, base, F0 = ctx.saved_tensors[0], ctx.saved_base, ctx.saved_F0\n        _kernels.rope_2d(grad_res, positions, base, -F0)\n        ctx.mark_dirty(grad_res)\n        return grad_res, None, None, None\n\n\nclass cuRoPE2D(torch.nn.Module):\n    def __init__(self, freq=100.0, F0=1.0):\n        super().__init__()\n        self.base = freq\n        self.F0 = F0\n\n    def forward(self, tokens, positions):\n        cuRoPE2D_func.apply(tokens.transpose(1, 2), positions, self.base, self.F0)\n        return tokens","source_hash":"dc49348fc210445a9b12252e1faf2ec678346e3ae3ef9735867eb39556739377","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.curope.curope2d.forward","uri":"program://Human3R/function/src.croco.models.curope.curope2d.forward#L38-L40","kind":"function","name":"forward","path":"src/croco/models/curope/curope2d.py","language":"python","start_line":38,"end_line":40,"context_start_line":18,"context_end_line":40,"code":"        ctx.saved_F0 = F0\n        # tokens = tokens.clone() # uncomment this if inplace doesn't work\n        _kernels.rope_2d(tokens, positions, base, F0)\n        ctx.mark_dirty(tokens)\n        return tokens\n\n    @staticmethod\n    def backward(ctx, grad_res):\n        positions, base, F0 = ctx.saved_tensors[0], ctx.saved_base, ctx.saved_F0\n        _kernels.rope_2d(grad_res, positions, base, -F0)\n        ctx.mark_dirty(grad_res)\n        return grad_res, None, None, None\n\n\nclass cuRoPE2D(torch.nn.Module):\n    def __init__(self, freq=100.0, F0=1.0):\n        super().__init__()\n        self.base = freq\n        self.F0 = F0\n\n    def forward(self, tokens, positions):\n        cuRoPE2D_func.apply(tokens.transpose(1, 2), positions, self.base, self.F0)\n        return tokens","source_hash":"dc49348fc210445a9b12252e1faf2ec678346e3ae3ef9735867eb39556739377","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.curope.curope2d.backward","uri":"program://Human3R/function/src.croco.models.curope.curope2d.backward#L25-L29","kind":"function","name":"backward","path":"src/croco/models/curope/curope2d.py","language":"python","start_line":25,"end_line":29,"context_start_line":5,"context_end_line":40,"code":"\ntry:\n    import curope as _kernels  # run `python setup.py install`\nexcept ModuleNotFoundError:\n    from . import curope as _kernels  # run `python setup.py build_ext --inplace`\n\n\nclass cuRoPE2D_func(torch.autograd.Function):\n\n    @staticmethod\n    def forward(ctx, tokens, positions, base, F0=1):\n        ctx.save_for_backward(positions)\n        ctx.saved_base = base\n        ctx.saved_F0 = F0\n        # tokens = tokens.clone() # uncomment this if inplace doesn't work\n        _kernels.rope_2d(tokens, positions, base, F0)\n        ctx.mark_dirty(tokens)\n        return tokens\n\n    @staticmethod\n    def backward(ctx, grad_res):\n        positions, base, F0 = ctx.saved_tensors[0], ctx.saved_base, ctx.saved_F0\n        _kernels.rope_2d(grad_res, positions, base, -F0)\n        ctx.mark_dirty(grad_res)\n        return grad_res, None, None, None\n\n\nclass cuRoPE2D(torch.nn.Module):\n    def __init__(self, freq=100.0, F0=1.0):\n        super().__init__()\n        self.base = freq\n        self.F0 = F0\n\n    def forward(self, tokens, positions):\n        cuRoPE2D_func.apply(tokens.transpose(1, 2), positions, self.base, self.F0)\n        return tokens","source_hash":"dc49348fc210445a9b12252e1faf2ec678346e3ae3ef9735867eb39556739377","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.models.curope.curope2d.__init__","uri":"program://Human3R/function/src.croco.models.curope.curope2d.__init__#L33-L36","kind":"function","name":"__init__","path":"src/croco/models/curope/curope2d.py","language":"python","start_line":33,"end_line":36,"context_start_line":13,"context_end_line":40,"code":"\n    @staticmethod\n    def forward(ctx, tokens, positions, base, F0=1):\n        ctx.save_for_backward(positions)\n        ctx.saved_base = base\n        ctx.saved_F0 = F0\n        # tokens = tokens.clone() # uncomment this if inplace doesn't work\n        _kernels.rope_2d(tokens, positions, base, F0)\n        ctx.mark_dirty(tokens)\n        return tokens\n\n    @staticmethod\n    def backward(ctx, grad_res):\n        positions, base, F0 = ctx.saved_tensors[0], ctx.saved_base, ctx.saved_F0\n        _kernels.rope_2d(grad_res, positions, base, -F0)\n        ctx.mark_dirty(grad_res)\n        return grad_res, None, None, None\n\n\nclass cuRoPE2D(torch.nn.Module):\n    def __init__(self, freq=100.0, F0=1.0):\n        super().__init__()\n        self.base = freq\n        self.F0 = F0\n\n    def forward(self, tokens, positions):\n        cuRoPE2D_func.apply(tokens.transpose(1, 2), positions, self.base, self.F0)\n        return tokens","source_hash":"dc49348fc210445a9b12252e1faf2ec678346e3ae3ef9735867eb39556739377","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc","uri":"program://Human3R/module/src.croco.utils.misc#L1-L603","kind":"module","name":"src.croco.utils.misc","path":"src/croco/utils/misc.py","language":"python","start_line":1,"end_line":603,"context_start_line":1,"context_end_line":603,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# utilitary functions for CroCo\n# --------------------------------------------------------\n# References:\n# MAE: https://github.com/facebookresearch/mae\n# DeiT: https://github.com/facebookresearch/deit\n# BEiT: https://github.com/microsoft/unilm/tree/master/beit\n# --------------------------------------------------------\n\nimport builtins\nimport datetime\nimport os\nimport time\nimport math\nimport json\nfrom collections import defaultdict, deque\nfrom pathlib import Path\nimport numpy as np\n\nimport torch\nimport torch.distributed as dist\nfrom torch import inf\nfrom accelerate import Accelerator\nfrom accelerate.logging import get_logger\n\nprinter = get_logger(__name__, log_level=\"DEBUG\")\n\n\nclass SmoothedValue(object):\n    \"\"\"Track a series of values and provide access to smoothed values.\"\"\"\n\n    def __init__(self, window_size=20, fmt=None):\n        if fmt is None:\n            fmt = \"{median:.4f} ({global_avg:.4f})\"\n        self.deque = deque(maxlen=window_size)\n        self.total = 0.0\n        self.count = 0\n        self.fmt = fmt\n\n    def update(self, value, n=1):\n        self.deque.append(value)\n        self.count += n\n        self.total += value * n\n\n    def synchronize_between_processes(self, accelerator: Accelerator):\n        \"\"\"Synchronize the count and total across all processes.\"\"\"\n        if accelerator.num_processes == 1:\n            return\n        t = torch.tensor(\n            [self.count, self.total], dtype=torch.float64, device=accelerator.device\n        )\n        accelerator.wait_for_everyone()\n        accelerator.reduce(t, reduction=\"sum\")\n        t = t.tolist()\n        self.count = int(t[0])\n        self.total = t[1]\n\n    @property\n    def median(self):\n        return torch.tensor(list(self.deque)).median().item()\n\n    @property\n    def avg(self):\n        return torch.tensor(list(self.deque), dtype=torch.float32).mean().item()\n\n    @property\n    def global_avg(self):\n        return self.total / self.count\n\n    @property\n    def max(self):\n        return max(self.deque)\n\n    @property\n    def value(self):\n        return self.deque[-1]\n\n    def __str__(self):\n        return self.fmt.format(\n            median=self.median,\n            avg=self.avg,\n            global_avg=self.global_avg,\n            max=self.max,\n            value=self.value,\n        )\n\n\nclass MetricLogger(object):\n    def __init__(self, delimiter=\"\\t\"):\n        self.meters = defaultdict(SmoothedValue)\n        self.delimiter = delimiter\n\n    def update(self, **kwargs):\n        for k, v in kwargs.items():\n            if v is None:\n                continue\n            if isinstance(v, torch.Tensor):\n                if v.ndim > 0:\n                    continue\n                v = v.item()\n            if isinstance(v, list):\n                continue\n            assert isinstance(v, (float, int))\n            self.meters[k].update(v)\n\n    def __getattr__(self, attr):\n        if attr in self.meters:\n            return self.meters[attr]\n        if attr in self.__dict__:\n            return self.__dict__[attr]\n        raise AttributeError(\n            \"'{}' object has no attribute '{}'\".format(type(self).__name__, attr)\n        )\n\n    def __str__(self):\n        loss_str = []\n        for name, meter in self.meters.items():\n            loss_str.append(\"{}: {}\".format(name, str(meter)))\n        return self.delimiter.join(loss_str)\n\n    def synchronize_between_processes(self, accelerator):\n        for meter in self.meters.values():\n            meter.synchronize_between_processes(accelerator)\n\n    def add_meter(self, name, meter):\n        self.meters[name] = meter\n\n    def log_every(\n        self, iterable, print_freq, accelerator: Accelerator, header=None, max_iter=None\n    ):\n        i = 0\n        if not header:\n            header = \"\"\n        start_time = time.time()\n        end = time.time()\n        iter_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        data_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        len_iterable = min(len(iterable), max_iter) if max_iter else len(iterable)\n        space_fmt = \":\" + str(len(str(len_iterable))) + \"d\"\n        log_msg = [\n            header,\n            \"[{0\" + space_fmt + \"}/{1}]\",\n            \"eta: {eta}\",\n            \"{meters}\",\n            \"time: {time}\",\n            \"data: {data}\",\n        ]\n        if torch.cuda.is_available():\n            log_msg.append(\"max mem: {memory:.0f}\")\n        log_msg = self.delimiter.join(log_msg)\n        MB = 1024.0 * 1024.0\n        for it, obj in enumerate(iterable):\n            data_time.update(time.time() - end)\n            yield obj\n            iter_time.update(time.time() - end)\n            if i % print_freq == 0 or i == len_iterable - 1:\n                eta_seconds = iter_time.global_avg * (len_iterable - i)\n                eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))\n                if torch.cuda.is_available():\n                    if accelerator.is_main_process:\n                        printer.info(\n                            log_msg.format(\n                                i,\n                                len_iterable,\n                                eta=eta_string,\n                                meters=str(self),\n                                time=str(iter_time),\n                                data=str(data_time),\n                                memory=torch.cuda.max_memory_allocated() / MB,\n                            )\n                        )\n                else:\n                    if accelerator.is_main_process:\n                        printer.info(\n                            log_msg.format(\n                                i,\n                                len_iterable,\n                                eta=eta_string,\n                                meters=str(self),\n                                time=str(iter_time),\n                                data=str(data_time),\n                            )\n                        )\n            i += 1\n            end = time.time()\n            if max_iter and it >= max_iter:\n                break\n        total_time = time.time() - start_time\n        total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n        if accelerator.is_main_process:\n            printer.info(\n                \"{} Total time: {} ({:.4f} s / it)\".format(\n                    header, total_time_str, total_time / len_iterable\n                )\n            )\n\n\ndef setup_for_distributed(is_master):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (get_world_size() > 8)\n        if is_master or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef is_dist_avail_and_initialized():\n    if not dist.is_available():\n        return False\n    if not dist.is_initialized():\n        return False\n    return True\n\n\ndef get_world_size():\n    if not is_dist_avail_and_initialized():\n        return 1\n    return dist.get_world_size()\n\n\ndef get_rank():\n    if not is_dist_avail_and_initialized():\n        return 0\n    return dist.get_rank()\n\n\ndef is_main_process(accelerator: Accelerator):\n    return accelerator.is_main_process\n\n\ndef save_on_master(accelerator: Accelerator, *args, **kwargs):\n    if is_main_process(accelerator):\n        torch.save(*args, **kwargs)\n\n\ndef init_distributed_mode(args):\n    nodist = args.nodist if hasattr(args, \"nodist\") else False\n    if \"RANK\" in os.environ and \"WORLD_SIZE\" in os.environ and not nodist:\n        args.rank = int(os.environ[\"RANK\"])\n        args.world_size = int(os.environ[\"WORLD_SIZE\"])\n        args.gpu = int(os.environ[\"LOCAL_RANK\"])\n    else:\n        print(\"Not using distributed mode\")\n        setup_for_distributed(is_master=True)  # hack\n        args.distributed = False\n        return\n\n    args.distributed = True\n\n    torch.cuda.set_device(args.gpu)\n    args.dist_backend = \"nccl\"\n    print(\n        \"| distributed init (rank {}): {}, gpu {}\".format(\n            args.rank, args.dist_url, args.gpu\n        ),\n        flush=True,\n    )\n    torch.distributed.init_process_group(\n        backend=args.dist_backend,\n        init_method=args.dist_url,\n        world_size=args.world_size,\n        rank=args.rank,\n    )\n    torch.distributed.barrier()\n    setup_for_distributed(args.rank == 0)\n\n\nclass NativeScalerWithGradNormCount:\n    state_dict_key = \"amp_scaler\"\n\n    def __init__(self, enabled=True, accelerator: Accelerator = None):\n        self.accelerator = accelerator\n\n    def __call__(\n        self,\n        loss,\n        optimizer,\n        clip_grad=None,\n        parameters=None,\n        create_graph=False,\n        update_grad=True,\n    ):\n        self.accelerator.backward(\n            loss, create_graph=create_graph\n        )  # .backward(create_graph=create_graph)\n        if update_grad:\n            if clip_grad is not None:\n                assert parameters is not None\n                # self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n                norm = self.accelerator.clip_grad_norm_(parameters, clip_grad)\n            else:\n                if self.accelerator.scaler is not None:\n                    self.accelerator.unscale_gradients()\n                norm = get_grad_norm_(parameters)\n            optimizer.step()\n        else:\n            norm = None\n        return norm\n\n    def state_dict(self):\n        if self.accelerator.scaler is not None:\n            return self.accelerator.scaler.state_dict()\n        else:\n            return {}\n\n    def load_state_dict(self, state_dict):\n        if self.accelerator.scaler is not None:\n            self.accelerator.scaler.load_state_dict(state_dict)\n\n\n# class NativeScalerWithGradNormCount:\n#     state_dict_key = \"amp_scaler\"\n\n#     def __init__(self, enabled=True, accelerator:Accelerator=None):\n#         self._scaler = torch.cuda.amp.GradScaler(enabled=enabled)\n#         self.accelerator = accelerator\n\n#     def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True):\n#         # self.accelerator.backward(loss, create_graph=create_graph) #.backward(create_graph=create_graph)\n#         self._scaler.scale(loss).backward(create_graph=create_graph)\n#         if update_grad:\n#             if clip_grad is not None:\n#                 assert parameters is not None\n#                 # #self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n#                 # norm = self.accelerator.clip_grad_norm_(parameters, clip_grad)\n#                 self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n#                 norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad)\n#             else:\n#                 # if self.accelerator.scaler is not None:\n#                 #     self.accelerator.unscale_gradients()\n#                 # norm = get_grad_norm_(parameters)\n#                 self._scaler.unscale_(optimizer)\n#                 norm = get_grad_norm_(parameters)\n#             # optimizer.step()\n#             self._scaler.step(optimizer)\n#             self._scaler.update()\n#         else:\n#             norm = None\n#         return norm\n\n#     # def state_dict(self):\n#     #     if self.accelerator.scaler is not None:\n#     #         return self.accelerator.scaler.state_dict()\n#     #     else:\n#     #         return {}\n\n#     # def load_state_dict(self, state_dict):\n#     #     if self.accelerator.scaler is not None:\n#     #         self.accelerator.scaler.load_state_dict(state_dict)\n\n#     def state_dict(self):\n#         return self._scaler.state_dict()\n\n#     def load_state_dict(self, state_dict):\n#         self._scaler.load_state_dict(state_dict)\n\n\ndef get_grad_norm_(parameters, norm_type: float = 2.0) -> torch.Tensor:\n    if isinstance(parameters, torch.Tensor):\n        parameters = [parameters]\n    parameters = [p for p in parameters if p.grad is not None]\n    norm_type = float(norm_type)\n    if len(parameters) == 0:\n        return torch.tensor(0.0)\n    device = parameters[0].grad.device\n    if norm_type == inf:\n        total_norm = max(p.grad.detach().abs().max().to(device) for p in parameters)\n    else:\n        total_norm = torch.norm(\n            torch.stack(\n                [torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters]\n            ),\n            norm_type,\n        )\n    return total_norm\n\n\ndef save_model(\n    accelerator,\n    args,\n    epoch,\n    model_without_ddp,\n    optimizer,\n    loss_scaler,\n    fname=None,\n    best_so_far=None,\n):\n    if accelerator.is_main_process:\n        output_dir = Path(args.output_dir)\n        if fname is None:\n            fname = str(epoch)\n        checkpoint_path = output_dir / (\"checkpoint-%s.pth\" % fname)\n        to_save = {\n            \"model\": model_without_ddp.state_dict(),\n            \"optimizer\": optimizer.state_dict(),\n            \"scaler\": loss_scaler.state_dict(),\n            \"args\": args,\n            \"epoch\": epoch,\n        }\n        if best_so_far is not None:\n            to_save[\"best_so_far\"] = best_so_far\n        print(f\">> Saving model to {checkpoint_path} ...\")\n        save_on_master(accelerator, to_save, checkpoint_path)\n\n\ndef load_model(args, model_without_ddp, optimizer, loss_scaler):\n    args.start_epoch = 0\n    best_so_far = None\n    if args.resume is not None:\n        if args.resume.startswith(\"https\"):\n            checkpoint = torch.hub.load_state_dict_from_url(\n                args.resume, map_location=\"cpu\", check_hash=True\n            )\n        else:\n            checkpoint = torch.load(args.resume, map_location=\"cpu\")\n        printer.info(\"Resume checkpoint %s\" % args.resume)\n        model_without_ddp.load_state_dict(checkpoint[\"model\"], strict=False)\n        args.start_epoch = checkpoint[\"epoch\"] + 1\n        optimizer.load_state_dict(checkpoint[\"optimizer\"])\n        if \"scaler\" in checkpoint:\n            loss_scaler.load_state_dict(checkpoint[\"scaler\"])\n        if \"best_so_far\" in checkpoint:\n            best_so_far = checkpoint[\"best_so_far\"]\n            printer.info(\" & best_so_far={:g}\".format(best_so_far))\n        else:\n            printer.info(\"\")\n        printer.info(\"With optim & sched! start_epoch={:d}\".format(args.start_epoch))\n    return best_so_far\n\n\ndef all_reduce_mean(x, accelerator):\n    \"\"\"Use accelerator to all-reduce and compute mean.\"\"\"\n    if accelerator.state.num_processes > 1:\n        x_reduce = torch.tensor(x).cuda()\n        accelerator.reduce(x_reduce, reduce_op=\"SUM\")\n        x_reduce /= accelerator.state.num_processes\n        return x_reduce.item()\n    else:\n        return x\n\n\ndef _replace(text, src, tgt, rm=\"\"):\n    \"\"\"Advanced string replacement.\n    Given a text:\n    - replace all elements in src by the corresponding element in tgt\n    - remove all elements in rm\n    \"\"\"\n    if len(tgt) == 1:\n        tgt = tgt * len(src)\n    assert len(src) == len(tgt), f\"'{src}' and '{tgt}' should have the same len\"\n    for s, t in zip(src, tgt):\n        text = text.replace(s, t)\n    for c in rm:\n        text = text.replace(c, \"\")\n    return text\n\n\ndef filename(obj):\n    \"\"\"transform a python obj or cmd into a proper filename.\n    - \\1 gets replaced by slash '/'\n    - \\2 gets replaced by comma ','\n    \"\"\"\n    if not isinstance(obj, str):\n        obj = repr(obj)\n    obj = str(obj).replace(\"()\", \"\")\n    obj = _replace(obj, \"_,(*/\\1\\2\", \"-__x%/,\", rm=\" )'\\\"\")\n    assert all(len(s) < 256 for s in obj.split(os.sep)), (\n        \"filename too long (>256 characters):\\n\" + obj\n    )\n    return obj\n\n\ndef _get_num_layer_for_vit(var_name, enc_depth, dec_depth):\n    if var_name in (\"cls_token\", \"mask_token\", \"pos_embed\", \"global_tokens\"):\n        return 0\n    elif var_name.startswith(\"patch_embed\"):\n        return 0\n    elif var_name.startswith(\"enc_blocks\"):\n        layer_id = int(var_name.split(\".\")[1])\n        return layer_id + 1\n    elif var_name.startswith(\"decoder_embed\") or var_name.startswith(\n        \"enc_norm\"\n    ):  # part of the last black\n        return enc_depth\n    elif var_name.startswith(\"dec_blocks\"):\n        layer_id = int(var_name.split(\".\")[1])\n        return enc_depth + layer_id + 1\n    elif var_name.startswith(\"dec_norm\"):  # part of the last block\n        return enc_depth + dec_depth\n    elif any(var_name.startswith(k) for k in [\"head\", \"prediction_head\"]):\n        return enc_depth + dec_depth + 1\n    else:\n        raise NotImplementedError(var_name)\n\n\ndef get_parameter_groups(\n    model, weight_decay, layer_decay=1.0, skip_list=(), no_lr_scale_list=[]\n):\n    parameter_group_names = {}\n    parameter_group_vars = {}\n    enc_depth, dec_depth = None, None\n    # prepare layer decay values\n    assert layer_decay == 1.0 or 0.0 < layer_decay < 1.0\n    if layer_decay < 1.0:\n        enc_depth = model.enc_depth\n        dec_depth = model.dec_depth if hasattr(model, \"dec_blocks\") else 0\n        num_layers = enc_depth + dec_depth\n        layer_decay_values = list(\n            layer_decay ** (num_layers + 1 - i) for i in range(num_layers + 2)\n        )\n\n    for name, param in model.named_parameters():\n        if not param.requires_grad:\n            continue  # frozen weights\n\n        if getattr(param, '_is_frozen', False):\n            continue  # frozen weights\n\n        # Assign weight decay values\n        if len(param.shape) == 1 or name.endswith(\".bias\") or name in skip_list:\n            if \"enc_blocks\" in name:\n                group_name = \"no_decay_enc_blocks\"\n            else:\n                group_name = \"no_decay\"\n            this_weight_decay = 0.0\n        else:\n            if \"enc_blocks\" in name:\n                group_name = \"decay_enc_blocks\"\n            else:\n                group_name = \"decay\"\n            this_weight_decay = weight_decay\n\n        # Assign layer ID for LR scaling\n        if layer_decay < 1.0:\n            skip_scale = False\n            layer_id = _get_num_layer_for_vit(name, enc_depth, dec_depth)\n            group_name = \"layer_%d_%s\" % (layer_id, group_name)\n            if name in no_lr_scale_list:\n                skip_scale = True\n                group_name = f\"{group_name}_no_lr_scale\"\n        else:\n            layer_id = 0\n            skip_scale = True\n\n        if group_name not in parameter_group_names:\n            if not skip_scale:\n                scale = layer_decay_values[layer_id]\n            else:\n                scale = 1.0\n\n            if \"enc_blocks\" in group_name:\n                scale *= 1.0\n            parameter_group_names[group_name] = {\n                \"weight_decay\": this_weight_decay,\n                \"params\": [],\n                \"lr_scale\": scale,\n            }\n            parameter_group_vars[group_name] = {\n                \"weight_decay\": this_weight_decay,\n                \"params\": [],\n                \"lr_scale\": scale,\n            }\n\n        parameter_group_vars[group_name][\"params\"].append(param)\n        parameter_group_names[group_name][\"params\"].append(name)\n    printer.info(\"Param groups = %s\" % json.dumps(parameter_group_names, indent=2))\n    return list(parameter_group_vars.values())\n\n\ndef adjust_learning_rate(optimizer, epoch, args):\n    \"\"\"Decay the learning rate with half-cycle cosine after warmup\"\"\"\n\n    if epoch < args.warmup_epochs:\n        lr = args.lr * epoch / args.warmup_epochs\n    else:\n        # lr = args.lr\n        lr = args.min_lr + (args.lr - args.min_lr) * 0.5 * (\n            1.0\n            + math.cos(\n                math.pi\n                * (epoch - args.warmup_epochs)\n                / (args.epochs - args.warmup_epochs)\n            )\n        )\n\n    for param_group in optimizer.param_groups:\n        if \"lr_scale\" in\n# ... truncated ...","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.SmoothedValue","uri":"program://Human3R/class/src.croco.utils.misc.SmoothedValue#L32-L88","kind":"class","name":"SmoothedValue","path":"src/croco/utils/misc.py","language":"python","start_line":32,"end_line":88,"context_start_line":12,"context_end_line":108,"code":"\nimport builtins\nimport datetime\nimport os\nimport time\nimport math\nimport json\nfrom collections import defaultdict, deque\nfrom pathlib import Path\nimport numpy as np\n\nimport torch\nimport torch.distributed as dist\nfrom torch import inf\nfrom accelerate import Accelerator\nfrom accelerate.logging import get_logger\n\nprinter = get_logger(__name__, log_level=\"DEBUG\")\n\n\nclass SmoothedValue(object):\n    \"\"\"Track a series of values and provide access to smoothed values.\"\"\"\n\n    def __init__(self, window_size=20, fmt=None):\n        if fmt is None:\n            fmt = \"{median:.4f} ({global_avg:.4f})\"\n        self.deque = deque(maxlen=window_size)\n        self.total = 0.0\n        self.count = 0\n        self.fmt = fmt\n\n    def update(self, value, n=1):\n        self.deque.append(value)\n        self.count += n\n        self.total += value * n\n\n    def synchronize_between_processes(self, accelerator: Accelerator):\n        \"\"\"Synchronize the count and total across all processes.\"\"\"\n        if accelerator.num_processes == 1:\n            return\n        t = torch.tensor(\n            [self.count, self.total], dtype=torch.float64, device=accelerator.device\n        )\n        accelerator.wait_for_everyone()\n        accelerator.reduce(t, reduction=\"sum\")\n        t = t.tolist()\n        self.count = int(t[0])\n        self.total = t[1]\n\n    @property\n    def median(self):\n        return torch.tensor(list(self.deque)).median().item()\n\n    @property\n    def avg(self):\n        return torch.tensor(list(self.deque), dtype=torch.float32).mean().item()\n\n    @property\n    def global_avg(self):\n        return self.total / self.count\n\n    @property\n    def max(self):\n        return max(self.deque)\n\n    @property\n    def value(self):\n        return self.deque[-1]\n\n    def __str__(self):\n        return self.fmt.format(\n            median=self.median,\n            avg=self.avg,\n            global_avg=self.global_avg,\n            max=self.max,\n            value=self.value,\n        )\n\n\nclass MetricLogger(object):\n    def __init__(self, delimiter=\"\\t\"):\n        self.meters = defaultdict(SmoothedValue)\n        self.delimiter = delimiter\n\n    def update(self, **kwargs):\n        for k, v in kwargs.items():\n            if v is None:\n                continue\n            if isinstance(v, torch.Tensor):\n                if v.ndim > 0:\n                    continue\n                v = v.item()\n            if isinstance(v, list):\n                continue\n            assert isinstance(v, (float, int))\n            self.meters[k].update(v)\n","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.MetricLogger","uri":"program://Human3R/class/src.croco.utils.misc.MetricLogger#L91-L198","kind":"class","name":"MetricLogger","path":"src/croco/utils/misc.py","language":"python","start_line":91,"end_line":198,"context_start_line":71,"context_end_line":218,"code":"        return self.total / self.count\n\n    @property\n    def max(self):\n        return max(self.deque)\n\n    @property\n    def value(self):\n        return self.deque[-1]\n\n    def __str__(self):\n        return self.fmt.format(\n            median=self.median,\n            avg=self.avg,\n            global_avg=self.global_avg,\n            max=self.max,\n            value=self.value,\n        )\n\n\nclass MetricLogger(object):\n    def __init__(self, delimiter=\"\\t\"):\n        self.meters = defaultdict(SmoothedValue)\n        self.delimiter = delimiter\n\n    def update(self, **kwargs):\n        for k, v in kwargs.items():\n            if v is None:\n                continue\n            if isinstance(v, torch.Tensor):\n                if v.ndim > 0:\n                    continue\n                v = v.item()\n            if isinstance(v, list):\n                continue\n            assert isinstance(v, (float, int))\n            self.meters[k].update(v)\n\n    def __getattr__(self, attr):\n        if attr in self.meters:\n            return self.meters[attr]\n        if attr in self.__dict__:\n            return self.__dict__[attr]\n        raise AttributeError(\n            \"'{}' object has no attribute '{}'\".format(type(self).__name__, attr)\n        )\n\n    def __str__(self):\n        loss_str = []\n        for name, meter in self.meters.items():\n            loss_str.append(\"{}: {}\".format(name, str(meter)))\n        return self.delimiter.join(loss_str)\n\n    def synchronize_between_processes(self, accelerator):\n        for meter in self.meters.values():\n            meter.synchronize_between_processes(accelerator)\n\n    def add_meter(self, name, meter):\n        self.meters[name] = meter\n\n    def log_every(\n        self, iterable, print_freq, accelerator: Accelerator, header=None, max_iter=None\n    ):\n        i = 0\n        if not header:\n            header = \"\"\n        start_time = time.time()\n        end = time.time()\n        iter_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        data_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        len_iterable = min(len(iterable), max_iter) if max_iter else len(iterable)\n        space_fmt = \":\" + str(len(str(len_iterable))) + \"d\"\n        log_msg = [\n            header,\n            \"[{0\" + space_fmt + \"}/{1}]\",\n            \"eta: {eta}\",\n            \"{meters}\",\n            \"time: {time}\",\n            \"data: {data}\",\n        ]\n        if torch.cuda.is_available():\n            log_msg.append(\"max mem: {memory:.0f}\")\n        log_msg = self.delimiter.join(log_msg)\n        MB = 1024.0 * 1024.0\n        for it, obj in enumerate(iterable):\n            data_time.update(time.time() - end)\n            yield obj\n            iter_time.update(time.time() - end)\n            if i % print_freq == 0 or i == len_iterable - 1:\n                eta_seconds = iter_time.global_avg * (len_iterable - i)\n                eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))\n                if torch.cuda.is_available():\n                    if accelerator.is_main_process:\n                        printer.info(\n                            log_msg.format(\n                                i,\n                                len_iterable,\n                                eta=eta_string,\n                                meters=str(self),\n                                time=str(iter_time),\n                                data=str(data_time),\n                                memory=torch.cuda.max_memory_allocated() / MB,\n                            )\n                        )\n                else:\n                    if accelerator.is_main_process:\n                        printer.info(\n                            log_msg.format(\n                                i,\n                                len_iterable,\n                                eta=eta_string,\n                                meters=str(self),\n                                time=str(iter_time),\n                                data=str(data_time),\n                            )\n                        )\n            i += 1\n            end = time.time()\n            if max_iter and it >= max_iter:\n                break\n        total_time = time.time() - start_time\n        total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n        if accelerator.is_main_process:\n            printer.info(\n                \"{} Total time: {} ({:.4f} s / it)\".format(\n                    header, total_time_str, total_time / len_iterable\n                )\n            )\n\n\ndef setup_for_distributed(is_master):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (get_world_size() > 8)\n        if is_master or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef is_dist_avail_and_initialized():","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.setup_for_distributed","uri":"program://Human3R/function/src.croco.utils.misc.setup_for_distributed#L201-L215","kind":"function","name":"setup_for_distributed","path":"src/croco/utils/misc.py","language":"python","start_line":201,"end_line":215,"context_start_line":181,"context_end_line":235,"code":"                                eta=eta_string,\n                                meters=str(self),\n                                time=str(iter_time),\n                                data=str(data_time),\n                            )\n                        )\n            i += 1\n            end = time.time()\n            if max_iter and it >= max_iter:\n                break\n        total_time = time.time() - start_time\n        total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n        if accelerator.is_main_process:\n            printer.info(\n                \"{} Total time: {} ({:.4f} s / it)\".format(\n                    header, total_time_str, total_time / len_iterable\n                )\n            )\n\n\ndef setup_for_distributed(is_master):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (get_world_size() > 8)\n        if is_master or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef is_dist_avail_and_initialized():\n    if not dist.is_available():\n        return False\n    if not dist.is_initialized():\n        return False\n    return True\n\n\ndef get_world_size():\n    if not is_dist_avail_and_initialized():\n        return 1\n    return dist.get_world_size()\n\n\ndef get_rank():\n    if not is_dist_avail_and_initialized():\n        return 0\n    return dist.get_rank()","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.is_dist_avail_and_initialized","uri":"program://Human3R/function/src.croco.utils.misc.is_dist_avail_and_initialized#L218-L223","kind":"function","name":"is_dist_avail_and_initialized","path":"src/croco/utils/misc.py","language":"python","start_line":218,"end_line":223,"context_start_line":198,"context_end_line":243,"code":"            )\n\n\ndef setup_for_distributed(is_master):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (get_world_size() > 8)\n        if is_master or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef is_dist_avail_and_initialized():\n    if not dist.is_available():\n        return False\n    if not dist.is_initialized():\n        return False\n    return True\n\n\ndef get_world_size():\n    if not is_dist_avail_and_initialized():\n        return 1\n    return dist.get_world_size()\n\n\ndef get_rank():\n    if not is_dist_avail_and_initialized():\n        return 0\n    return dist.get_rank()\n\n\ndef is_main_process(accelerator: Accelerator):\n    return accelerator.is_main_process\n\n\ndef save_on_master(accelerator: Accelerator, *args, **kwargs):\n    if is_main_process(accelerator):","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.get_world_size","uri":"program://Human3R/function/src.croco.utils.misc.get_world_size#L226-L229","kind":"function","name":"get_world_size","path":"src/croco/utils/misc.py","language":"python","start_line":226,"end_line":229,"context_start_line":206,"context_end_line":249,"code":"\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (get_world_size() > 8)\n        if is_master or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef is_dist_avail_and_initialized():\n    if not dist.is_available():\n        return False\n    if not dist.is_initialized():\n        return False\n    return True\n\n\ndef get_world_size():\n    if not is_dist_avail_and_initialized():\n        return 1\n    return dist.get_world_size()\n\n\ndef get_rank():\n    if not is_dist_avail_and_initialized():\n        return 0\n    return dist.get_rank()\n\n\ndef is_main_process(accelerator: Accelerator):\n    return accelerator.is_main_process\n\n\ndef save_on_master(accelerator: Accelerator, *args, **kwargs):\n    if is_main_process(accelerator):\n        torch.save(*args, **kwargs)\n\n\ndef init_distributed_mode(args):\n    nodist = args.nodist if hasattr(args, \"nodist\") else False\n    if \"RANK\" in os.environ and \"WORLD_SIZE\" in os.environ and not nodist:","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.get_rank","uri":"program://Human3R/function/src.croco.utils.misc.get_rank#L232-L235","kind":"function","name":"get_rank","path":"src/croco/utils/misc.py","language":"python","start_line":232,"end_line":235,"context_start_line":212,"context_end_line":255,"code":"            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef is_dist_avail_and_initialized():\n    if not dist.is_available():\n        return False\n    if not dist.is_initialized():\n        return False\n    return True\n\n\ndef get_world_size():\n    if not is_dist_avail_and_initialized():\n        return 1\n    return dist.get_world_size()\n\n\ndef get_rank():\n    if not is_dist_avail_and_initialized():\n        return 0\n    return dist.get_rank()\n\n\ndef is_main_process(accelerator: Accelerator):\n    return accelerator.is_main_process\n\n\ndef save_on_master(accelerator: Accelerator, *args, **kwargs):\n    if is_main_process(accelerator):\n        torch.save(*args, **kwargs)\n\n\ndef init_distributed_mode(args):\n    nodist = args.nodist if hasattr(args, \"nodist\") else False\n    if \"RANK\" in os.environ and \"WORLD_SIZE\" in os.environ and not nodist:\n        args.rank = int(os.environ[\"RANK\"])\n        args.world_size = int(os.environ[\"WORLD_SIZE\"])\n        args.gpu = int(os.environ[\"LOCAL_RANK\"])\n    else:\n        print(\"Not using distributed mode\")\n        setup_for_distributed(is_master=True)  # hack","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.is_main_process","uri":"program://Human3R/function/src.croco.utils.misc.is_main_process#L238-L239","kind":"function","name":"is_main_process","path":"src/croco/utils/misc.py","language":"python","start_line":238,"end_line":239,"context_start_line":218,"context_end_line":259,"code":"def is_dist_avail_and_initialized():\n    if not dist.is_available():\n        return False\n    if not dist.is_initialized():\n        return False\n    return True\n\n\ndef get_world_size():\n    if not is_dist_avail_and_initialized():\n        return 1\n    return dist.get_world_size()\n\n\ndef get_rank():\n    if not is_dist_avail_and_initialized():\n        return 0\n    return dist.get_rank()\n\n\ndef is_main_process(accelerator: Accelerator):\n    return accelerator.is_main_process\n\n\ndef save_on_master(accelerator: Accelerator, *args, **kwargs):\n    if is_main_process(accelerator):\n        torch.save(*args, **kwargs)\n\n\ndef init_distributed_mode(args):\n    nodist = args.nodist if hasattr(args, \"nodist\") else False\n    if \"RANK\" in os.environ and \"WORLD_SIZE\" in os.environ and not nodist:\n        args.rank = int(os.environ[\"RANK\"])\n        args.world_size = int(os.environ[\"WORLD_SIZE\"])\n        args.gpu = int(os.environ[\"LOCAL_RANK\"])\n    else:\n        print(\"Not using distributed mode\")\n        setup_for_distributed(is_master=True)  # hack\n        args.distributed = False\n        return\n\n    args.distributed = True","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.save_on_master","uri":"program://Human3R/function/src.croco.utils.misc.save_on_master#L242-L244","kind":"function","name":"save_on_master","path":"src/croco/utils/misc.py","language":"python","start_line":242,"end_line":244,"context_start_line":222,"context_end_line":264,"code":"        return False\n    return True\n\n\ndef get_world_size():\n    if not is_dist_avail_and_initialized():\n        return 1\n    return dist.get_world_size()\n\n\ndef get_rank():\n    if not is_dist_avail_and_initialized():\n        return 0\n    return dist.get_rank()\n\n\ndef is_main_process(accelerator: Accelerator):\n    return accelerator.is_main_process\n\n\ndef save_on_master(accelerator: Accelerator, *args, **kwargs):\n    if is_main_process(accelerator):\n        torch.save(*args, **kwargs)\n\n\ndef init_distributed_mode(args):\n    nodist = args.nodist if hasattr(args, \"nodist\") else False\n    if \"RANK\" in os.environ and \"WORLD_SIZE\" in os.environ and not nodist:\n        args.rank = int(os.environ[\"RANK\"])\n        args.world_size = int(os.environ[\"WORLD_SIZE\"])\n        args.gpu = int(os.environ[\"LOCAL_RANK\"])\n    else:\n        print(\"Not using distributed mode\")\n        setup_for_distributed(is_master=True)  # hack\n        args.distributed = False\n        return\n\n    args.distributed = True\n\n    torch.cuda.set_device(args.gpu)\n    args.dist_backend = \"nccl\"\n    print(\n        \"| distributed init (rank {}): {}, gpu {}\".format(","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.init_distributed_mode","uri":"program://Human3R/function/src.croco.utils.misc.init_distributed_mode#L247-L276","kind":"function","name":"init_distributed_mode","path":"src/croco/utils/misc.py","language":"python","start_line":247,"end_line":276,"context_start_line":227,"context_end_line":296,"code":"    if not is_dist_avail_and_initialized():\n        return 1\n    return dist.get_world_size()\n\n\ndef get_rank():\n    if not is_dist_avail_and_initialized():\n        return 0\n    return dist.get_rank()\n\n\ndef is_main_process(accelerator: Accelerator):\n    return accelerator.is_main_process\n\n\ndef save_on_master(accelerator: Accelerator, *args, **kwargs):\n    if is_main_process(accelerator):\n        torch.save(*args, **kwargs)\n\n\ndef init_distributed_mode(args):\n    nodist = args.nodist if hasattr(args, \"nodist\") else False\n    if \"RANK\" in os.environ and \"WORLD_SIZE\" in os.environ and not nodist:\n        args.rank = int(os.environ[\"RANK\"])\n        args.world_size = int(os.environ[\"WORLD_SIZE\"])\n        args.gpu = int(os.environ[\"LOCAL_RANK\"])\n    else:\n        print(\"Not using distributed mode\")\n        setup_for_distributed(is_master=True)  # hack\n        args.distributed = False\n        return\n\n    args.distributed = True\n\n    torch.cuda.set_device(args.gpu)\n    args.dist_backend = \"nccl\"\n    print(\n        \"| distributed init (rank {}): {}, gpu {}\".format(\n            args.rank, args.dist_url, args.gpu\n        ),\n        flush=True,\n    )\n    torch.distributed.init_process_group(\n        backend=args.dist_backend,\n        init_method=args.dist_url,\n        world_size=args.world_size,\n        rank=args.rank,\n    )\n    torch.distributed.barrier()\n    setup_for_distributed(args.rank == 0)\n\n\nclass NativeScalerWithGradNormCount:\n    state_dict_key = \"amp_scaler\"\n\n    def __init__(self, enabled=True, accelerator: Accelerator = None):\n        self.accelerator = accelerator\n\n    def __call__(\n        self,\n        loss,\n        optimizer,\n        clip_grad=None,\n        parameters=None,\n        create_graph=False,\n        update_grad=True,\n    ):\n        self.accelerator.backward(\n            loss, create_graph=create_graph\n        )  # .backward(create_graph=create_graph)","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.NativeScalerWithGradNormCount","uri":"program://Human3R/class/src.croco.utils.misc.NativeScalerWithGradNormCount#L279-L319","kind":"class","name":"NativeScalerWithGradNormCount","path":"src/croco/utils/misc.py","language":"python","start_line":279,"end_line":319,"context_start_line":259,"context_end_line":339,"code":"    args.distributed = True\n\n    torch.cuda.set_device(args.gpu)\n    args.dist_backend = \"nccl\"\n    print(\n        \"| distributed init (rank {}): {}, gpu {}\".format(\n            args.rank, args.dist_url, args.gpu\n        ),\n        flush=True,\n    )\n    torch.distributed.init_process_group(\n        backend=args.dist_backend,\n        init_method=args.dist_url,\n        world_size=args.world_size,\n        rank=args.rank,\n    )\n    torch.distributed.barrier()\n    setup_for_distributed(args.rank == 0)\n\n\nclass NativeScalerWithGradNormCount:\n    state_dict_key = \"amp_scaler\"\n\n    def __init__(self, enabled=True, accelerator: Accelerator = None):\n        self.accelerator = accelerator\n\n    def __call__(\n        self,\n        loss,\n        optimizer,\n        clip_grad=None,\n        parameters=None,\n        create_graph=False,\n        update_grad=True,\n    ):\n        self.accelerator.backward(\n            loss, create_graph=create_graph\n        )  # .backward(create_graph=create_graph)\n        if update_grad:\n            if clip_grad is not None:\n                assert parameters is not None\n                # self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n                norm = self.accelerator.clip_grad_norm_(parameters, clip_grad)\n            else:\n                if self.accelerator.scaler is not None:\n                    self.accelerator.unscale_gradients()\n                norm = get_grad_norm_(parameters)\n            optimizer.step()\n        else:\n            norm = None\n        return norm\n\n    def state_dict(self):\n        if self.accelerator.scaler is not None:\n            return self.accelerator.scaler.state_dict()\n        else:\n            return {}\n\n    def load_state_dict(self, state_dict):\n        if self.accelerator.scaler is not None:\n            self.accelerator.scaler.load_state_dict(state_dict)\n\n\n# class NativeScalerWithGradNormCount:\n#     state_dict_key = \"amp_scaler\"\n\n#     def __init__(self, enabled=True, accelerator:Accelerator=None):\n#         self._scaler = torch.cuda.amp.GradScaler(enabled=enabled)\n#         self.accelerator = accelerator\n\n#     def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True):\n#         # self.accelerator.backward(loss, create_graph=create_graph) #.backward(create_graph=create_graph)\n#         self._scaler.scale(loss).backward(create_graph=create_graph)\n#         if update_grad:\n#             if clip_grad is not None:\n#                 assert parameters is not None\n#                 # #self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n#                 # norm = self.accelerator.clip_grad_norm_(parameters, clip_grad)\n#                 self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n#                 norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad)\n#             else:","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.get_grad_norm_","uri":"program://Human3R/function/src.croco.utils.misc.get_grad_norm_#L369-L386","kind":"function","name":"get_grad_norm_","path":"src/croco/utils/misc.py","language":"python","start_line":369,"end_line":386,"context_start_line":349,"context_end_line":406,"code":"#             norm = None\n#         return norm\n\n#     # def state_dict(self):\n#     #     if self.accelerator.scaler is not None:\n#     #         return self.accelerator.scaler.state_dict()\n#     #     else:\n#     #         return {}\n\n#     # def load_state_dict(self, state_dict):\n#     #     if self.accelerator.scaler is not None:\n#     #         self.accelerator.scaler.load_state_dict(state_dict)\n\n#     def state_dict(self):\n#         return self._scaler.state_dict()\n\n#     def load_state_dict(self, state_dict):\n#         self._scaler.load_state_dict(state_dict)\n\n\ndef get_grad_norm_(parameters, norm_type: float = 2.0) -> torch.Tensor:\n    if isinstance(parameters, torch.Tensor):\n        parameters = [parameters]\n    parameters = [p for p in parameters if p.grad is not None]\n    norm_type = float(norm_type)\n    if len(parameters) == 0:\n        return torch.tensor(0.0)\n    device = parameters[0].grad.device\n    if norm_type == inf:\n        total_norm = max(p.grad.detach().abs().max().to(device) for p in parameters)\n    else:\n        total_norm = torch.norm(\n            torch.stack(\n                [torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters]\n            ),\n            norm_type,\n        )\n    return total_norm\n\n\ndef save_model(\n    accelerator,\n    args,\n    epoch,\n    model_without_ddp,\n    optimizer,\n    loss_scaler,\n    fname=None,\n    best_so_far=None,\n):\n    if accelerator.is_main_process:\n        output_dir = Path(args.output_dir)\n        if fname is None:\n            fname = str(epoch)\n        checkpoint_path = output_dir / (\"checkpoint-%s.pth\" % fname)\n        to_save = {\n            \"model\": model_without_ddp.state_dict(),\n            \"optimizer\": optimizer.state_dict(),","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.save_model","uri":"program://Human3R/function/src.croco.utils.misc.save_model#L389-L414","kind":"function","name":"save_model","path":"src/croco/utils/misc.py","language":"python","start_line":389,"end_line":414,"context_start_line":369,"context_end_line":434,"code":"def get_grad_norm_(parameters, norm_type: float = 2.0) -> torch.Tensor:\n    if isinstance(parameters, torch.Tensor):\n        parameters = [parameters]\n    parameters = [p for p in parameters if p.grad is not None]\n    norm_type = float(norm_type)\n    if len(parameters) == 0:\n        return torch.tensor(0.0)\n    device = parameters[0].grad.device\n    if norm_type == inf:\n        total_norm = max(p.grad.detach().abs().max().to(device) for p in parameters)\n    else:\n        total_norm = torch.norm(\n            torch.stack(\n                [torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters]\n            ),\n            norm_type,\n        )\n    return total_norm\n\n\ndef save_model(\n    accelerator,\n    args,\n    epoch,\n    model_without_ddp,\n    optimizer,\n    loss_scaler,\n    fname=None,\n    best_so_far=None,\n):\n    if accelerator.is_main_process:\n        output_dir = Path(args.output_dir)\n        if fname is None:\n            fname = str(epoch)\n        checkpoint_path = output_dir / (\"checkpoint-%s.pth\" % fname)\n        to_save = {\n            \"model\": model_without_ddp.state_dict(),\n            \"optimizer\": optimizer.state_dict(),\n            \"scaler\": loss_scaler.state_dict(),\n            \"args\": args,\n            \"epoch\": epoch,\n        }\n        if best_so_far is not None:\n            to_save[\"best_so_far\"] = best_so_far\n        print(f\">> Saving model to {checkpoint_path} ...\")\n        save_on_master(accelerator, to_save, checkpoint_path)\n\n\ndef load_model(args, model_without_ddp, optimizer, loss_scaler):\n    args.start_epoch = 0\n    best_so_far = None\n    if args.resume is not None:\n        if args.resume.startswith(\"https\"):\n            checkpoint = torch.hub.load_state_dict_from_url(\n                args.resume, map_location=\"cpu\", check_hash=True\n            )\n        else:\n            checkpoint = torch.load(args.resume, map_location=\"cpu\")\n        printer.info(\"Resume checkpoint %s\" % args.resume)\n        model_without_ddp.load_state_dict(checkpoint[\"model\"], strict=False)\n        args.start_epoch = checkpoint[\"epoch\"] + 1\n        optimizer.load_state_dict(checkpoint[\"optimizer\"])\n        if \"scaler\" in checkpoint:\n            loss_scaler.load_state_dict(checkpoint[\"scaler\"])\n        if \"best_so_far\" in checkpoint:\n            best_so_far = checkpoint[\"best_so_far\"]","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.load_model","uri":"program://Human3R/function/src.croco.utils.misc.load_model#L417-L439","kind":"function","name":"load_model","path":"src/croco/utils/misc.py","language":"python","start_line":417,"end_line":439,"context_start_line":397,"context_end_line":459,"code":"    best_so_far=None,\n):\n    if accelerator.is_main_process:\n        output_dir = Path(args.output_dir)\n        if fname is None:\n            fname = str(epoch)\n        checkpoint_path = output_dir / (\"checkpoint-%s.pth\" % fname)\n        to_save = {\n            \"model\": model_without_ddp.state_dict(),\n            \"optimizer\": optimizer.state_dict(),\n            \"scaler\": loss_scaler.state_dict(),\n            \"args\": args,\n            \"epoch\": epoch,\n        }\n        if best_so_far is not None:\n            to_save[\"best_so_far\"] = best_so_far\n        print(f\">> Saving model to {checkpoint_path} ...\")\n        save_on_master(accelerator, to_save, checkpoint_path)\n\n\ndef load_model(args, model_without_ddp, optimizer, loss_scaler):\n    args.start_epoch = 0\n    best_so_far = None\n    if args.resume is not None:\n        if args.resume.startswith(\"https\"):\n            checkpoint = torch.hub.load_state_dict_from_url(\n                args.resume, map_location=\"cpu\", check_hash=True\n            )\n        else:\n            checkpoint = torch.load(args.resume, map_location=\"cpu\")\n        printer.info(\"Resume checkpoint %s\" % args.resume)\n        model_without_ddp.load_state_dict(checkpoint[\"model\"], strict=False)\n        args.start_epoch = checkpoint[\"epoch\"] + 1\n        optimizer.load_state_dict(checkpoint[\"optimizer\"])\n        if \"scaler\" in checkpoint:\n            loss_scaler.load_state_dict(checkpoint[\"scaler\"])\n        if \"best_so_far\" in checkpoint:\n            best_so_far = checkpoint[\"best_so_far\"]\n            printer.info(\" & best_so_far={:g}\".format(best_so_far))\n        else:\n            printer.info(\"\")\n        printer.info(\"With optim & sched! start_epoch={:d}\".format(args.start_epoch))\n    return best_so_far\n\n\ndef all_reduce_mean(x, accelerator):\n    \"\"\"Use accelerator to all-reduce and compute mean.\"\"\"\n    if accelerator.state.num_processes > 1:\n        x_reduce = torch.tensor(x).cuda()\n        accelerator.reduce(x_reduce, reduce_op=\"SUM\")\n        x_reduce /= accelerator.state.num_processes\n        return x_reduce.item()\n    else:\n        return x\n\n\ndef _replace(text, src, tgt, rm=\"\"):\n    \"\"\"Advanced string replacement.\n    Given a text:\n    - replace all elements in src by the corresponding element in tgt\n    - remove all elements in rm\n    \"\"\"\n    if len(tgt) == 1:","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.all_reduce_mean","uri":"program://Human3R/function/src.croco.utils.misc.all_reduce_mean#L442-L450","kind":"function","name":"all_reduce_mean","path":"src/croco/utils/misc.py","language":"python","start_line":442,"end_line":450,"context_start_line":422,"context_end_line":470,"code":"            checkpoint = torch.hub.load_state_dict_from_url(\n                args.resume, map_location=\"cpu\", check_hash=True\n            )\n        else:\n            checkpoint = torch.load(args.resume, map_location=\"cpu\")\n        printer.info(\"Resume checkpoint %s\" % args.resume)\n        model_without_ddp.load_state_dict(checkpoint[\"model\"], strict=False)\n        args.start_epoch = checkpoint[\"epoch\"] + 1\n        optimizer.load_state_dict(checkpoint[\"optimizer\"])\n        if \"scaler\" in checkpoint:\n            loss_scaler.load_state_dict(checkpoint[\"scaler\"])\n        if \"best_so_far\" in checkpoint:\n            best_so_far = checkpoint[\"best_so_far\"]\n            printer.info(\" & best_so_far={:g}\".format(best_so_far))\n        else:\n            printer.info(\"\")\n        printer.info(\"With optim & sched! start_epoch={:d}\".format(args.start_epoch))\n    return best_so_far\n\n\ndef all_reduce_mean(x, accelerator):\n    \"\"\"Use accelerator to all-reduce and compute mean.\"\"\"\n    if accelerator.state.num_processes > 1:\n        x_reduce = torch.tensor(x).cuda()\n        accelerator.reduce(x_reduce, reduce_op=\"SUM\")\n        x_reduce /= accelerator.state.num_processes\n        return x_reduce.item()\n    else:\n        return x\n\n\ndef _replace(text, src, tgt, rm=\"\"):\n    \"\"\"Advanced string replacement.\n    Given a text:\n    - replace all elements in src by the corresponding element in tgt\n    - remove all elements in rm\n    \"\"\"\n    if len(tgt) == 1:\n        tgt = tgt * len(src)\n    assert len(src) == len(tgt), f\"'{src}' and '{tgt}' should have the same len\"\n    for s, t in zip(src, tgt):\n        text = text.replace(s, t)\n    for c in rm:\n        text = text.replace(c, \"\")\n    return text\n\n\ndef filename(obj):\n    \"\"\"transform a python obj or cmd into a proper filename.","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc._replace","uri":"program://Human3R/function/src.croco.utils.misc._replace#L453-L466","kind":"function","name":"_replace","path":"src/croco/utils/misc.py","language":"python","start_line":453,"end_line":466,"context_start_line":433,"context_end_line":486,"code":"        if \"best_so_far\" in checkpoint:\n            best_so_far = checkpoint[\"best_so_far\"]\n            printer.info(\" & best_so_far={:g}\".format(best_so_far))\n        else:\n            printer.info(\"\")\n        printer.info(\"With optim & sched! start_epoch={:d}\".format(args.start_epoch))\n    return best_so_far\n\n\ndef all_reduce_mean(x, accelerator):\n    \"\"\"Use accelerator to all-reduce and compute mean.\"\"\"\n    if accelerator.state.num_processes > 1:\n        x_reduce = torch.tensor(x).cuda()\n        accelerator.reduce(x_reduce, reduce_op=\"SUM\")\n        x_reduce /= accelerator.state.num_processes\n        return x_reduce.item()\n    else:\n        return x\n\n\ndef _replace(text, src, tgt, rm=\"\"):\n    \"\"\"Advanced string replacement.\n    Given a text:\n    - replace all elements in src by the corresponding element in tgt\n    - remove all elements in rm\n    \"\"\"\n    if len(tgt) == 1:\n        tgt = tgt * len(src)\n    assert len(src) == len(tgt), f\"'{src}' and '{tgt}' should have the same len\"\n    for s, t in zip(src, tgt):\n        text = text.replace(s, t)\n    for c in rm:\n        text = text.replace(c, \"\")\n    return text\n\n\ndef filename(obj):\n    \"\"\"transform a python obj or cmd into a proper filename.\n    - \\1 gets replaced by slash '/'\n    - \\2 gets replaced by comma ','\n    \"\"\"\n    if not isinstance(obj, str):\n        obj = repr(obj)\n    obj = str(obj).replace(\"()\", \"\")\n    obj = _replace(obj, \"_,(*/\\1\\2\", \"-__x%/,\", rm=\" )'\\\"\")\n    assert all(len(s) < 256 for s in obj.split(os.sep)), (\n        \"filename too long (>256 characters):\\n\" + obj\n    )\n    return obj\n\n\ndef _get_num_layer_for_vit(var_name, enc_depth, dec_depth):\n    if var_name in (\"cls_token\", \"mask_token\", \"pos_embed\", \"global_tokens\"):\n        return 0","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.filename","uri":"program://Human3R/function/src.croco.utils.misc.filename#L469-L481","kind":"function","name":"filename","path":"src/croco/utils/misc.py","language":"python","start_line":469,"end_line":481,"context_start_line":449,"context_end_line":501,"code":"    else:\n        return x\n\n\ndef _replace(text, src, tgt, rm=\"\"):\n    \"\"\"Advanced string replacement.\n    Given a text:\n    - replace all elements in src by the corresponding element in tgt\n    - remove all elements in rm\n    \"\"\"\n    if len(tgt) == 1:\n        tgt = tgt * len(src)\n    assert len(src) == len(tgt), f\"'{src}' and '{tgt}' should have the same len\"\n    for s, t in zip(src, tgt):\n        text = text.replace(s, t)\n    for c in rm:\n        text = text.replace(c, \"\")\n    return text\n\n\ndef filename(obj):\n    \"\"\"transform a python obj or cmd into a proper filename.\n    - \\1 gets replaced by slash '/'\n    - \\2 gets replaced by comma ','\n    \"\"\"\n    if not isinstance(obj, str):\n        obj = repr(obj)\n    obj = str(obj).replace(\"()\", \"\")\n    obj = _replace(obj, \"_,(*/\\1\\2\", \"-__x%/,\", rm=\" )'\\\"\")\n    assert all(len(s) < 256 for s in obj.split(os.sep)), (\n        \"filename too long (>256 characters):\\n\" + obj\n    )\n    return obj\n\n\ndef _get_num_layer_for_vit(var_name, enc_depth, dec_depth):\n    if var_name in (\"cls_token\", \"mask_token\", \"pos_embed\", \"global_tokens\"):\n        return 0\n    elif var_name.startswith(\"patch_embed\"):\n        return 0\n    elif var_name.startswith(\"enc_blocks\"):\n        layer_id = int(var_name.split(\".\")[1])\n        return layer_id + 1\n    elif var_name.startswith(\"decoder_embed\") or var_name.startswith(\n        \"enc_norm\"\n    ):  # part of the last black\n        return enc_depth\n    elif var_name.startswith(\"dec_blocks\"):\n        layer_id = int(var_name.split(\".\")[1])\n        return enc_depth + layer_id + 1\n    elif var_name.startswith(\"dec_norm\"):  # part of the last block\n        return enc_depth + dec_depth\n    elif any(var_name.startswith(k) for k in [\"head\", \"prediction_head\"]):","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc._get_num_layer_for_vit","uri":"program://Human3R/function/src.croco.utils.misc._get_num_layer_for_vit#L484-L504","kind":"function","name":"_get_num_layer_for_vit","path":"src/croco/utils/misc.py","language":"python","start_line":484,"end_line":504,"context_start_line":464,"context_end_line":524,"code":"    for c in rm:\n        text = text.replace(c, \"\")\n    return text\n\n\ndef filename(obj):\n    \"\"\"transform a python obj or cmd into a proper filename.\n    - \\1 gets replaced by slash '/'\n    - \\2 gets replaced by comma ','\n    \"\"\"\n    if not isinstance(obj, str):\n        obj = repr(obj)\n    obj = str(obj).replace(\"()\", \"\")\n    obj = _replace(obj, \"_,(*/\\1\\2\", \"-__x%/,\", rm=\" )'\\\"\")\n    assert all(len(s) < 256 for s in obj.split(os.sep)), (\n        \"filename too long (>256 characters):\\n\" + obj\n    )\n    return obj\n\n\ndef _get_num_layer_for_vit(var_name, enc_depth, dec_depth):\n    if var_name in (\"cls_token\", \"mask_token\", \"pos_embed\", \"global_tokens\"):\n        return 0\n    elif var_name.startswith(\"patch_embed\"):\n        return 0\n    elif var_name.startswith(\"enc_blocks\"):\n        layer_id = int(var_name.split(\".\")[1])\n        return layer_id + 1\n    elif var_name.startswith(\"decoder_embed\") or var_name.startswith(\n        \"enc_norm\"\n    ):  # part of the last black\n        return enc_depth\n    elif var_name.startswith(\"dec_blocks\"):\n        layer_id = int(var_name.split(\".\")[1])\n        return enc_depth + layer_id + 1\n    elif var_name.startswith(\"dec_norm\"):  # part of the last block\n        return enc_depth + dec_depth\n    elif any(var_name.startswith(k) for k in [\"head\", \"prediction_head\"]):\n        return enc_depth + dec_depth + 1\n    else:\n        raise NotImplementedError(var_name)\n\n\ndef get_parameter_groups(\n    model, weight_decay, layer_decay=1.0, skip_list=(), no_lr_scale_list=[]\n):\n    parameter_group_names = {}\n    parameter_group_vars = {}\n    enc_depth, dec_depth = None, None\n    # prepare layer decay values\n    assert layer_decay == 1.0 or 0.0 < layer_decay < 1.0\n    if layer_decay < 1.0:\n        enc_depth = model.enc_depth\n        dec_depth = model.dec_depth if hasattr(model, \"dec_blocks\") else 0\n        num_layers = enc_depth + dec_depth\n        layer_decay_values = list(\n            layer_decay ** (num_layers + 1 - i) for i in range(num_layers + 2)\n        )\n\n    for name, param in model.named_parameters():\n        if not param.requires_grad:","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.get_parameter_groups","uri":"program://Human3R/function/src.croco.utils.misc.get_parameter_groups#L507-L578","kind":"function","name":"get_parameter_groups","path":"src/croco/utils/misc.py","language":"python","start_line":507,"end_line":578,"context_start_line":487,"context_end_line":598,"code":"    elif var_name.startswith(\"patch_embed\"):\n        return 0\n    elif var_name.startswith(\"enc_blocks\"):\n        layer_id = int(var_name.split(\".\")[1])\n        return layer_id + 1\n    elif var_name.startswith(\"decoder_embed\") or var_name.startswith(\n        \"enc_norm\"\n    ):  # part of the last black\n        return enc_depth\n    elif var_name.startswith(\"dec_blocks\"):\n        layer_id = int(var_name.split(\".\")[1])\n        return enc_depth + layer_id + 1\n    elif var_name.startswith(\"dec_norm\"):  # part of the last block\n        return enc_depth + dec_depth\n    elif any(var_name.startswith(k) for k in [\"head\", \"prediction_head\"]):\n        return enc_depth + dec_depth + 1\n    else:\n        raise NotImplementedError(var_name)\n\n\ndef get_parameter_groups(\n    model, weight_decay, layer_decay=1.0, skip_list=(), no_lr_scale_list=[]\n):\n    parameter_group_names = {}\n    parameter_group_vars = {}\n    enc_depth, dec_depth = None, None\n    # prepare layer decay values\n    assert layer_decay == 1.0 or 0.0 < layer_decay < 1.0\n    if layer_decay < 1.0:\n        enc_depth = model.enc_depth\n        dec_depth = model.dec_depth if hasattr(model, \"dec_blocks\") else 0\n        num_layers = enc_depth + dec_depth\n        layer_decay_values = list(\n            layer_decay ** (num_layers + 1 - i) for i in range(num_layers + 2)\n        )\n\n    for name, param in model.named_parameters():\n        if not param.requires_grad:\n            continue  # frozen weights\n\n        if getattr(param, '_is_frozen', False):\n            continue  # frozen weights\n\n        # Assign weight decay values\n        if len(param.shape) == 1 or name.endswith(\".bias\") or name in skip_list:\n            if \"enc_blocks\" in name:\n                group_name = \"no_decay_enc_blocks\"\n            else:\n                group_name = \"no_decay\"\n            this_weight_decay = 0.0\n        else:\n            if \"enc_blocks\" in name:\n                group_name = \"decay_enc_blocks\"\n            else:\n                group_name = \"decay\"\n            this_weight_decay = weight_decay\n\n        # Assign layer ID for LR scaling\n        if layer_decay < 1.0:\n            skip_scale = False\n            layer_id = _get_num_layer_for_vit(name, enc_depth, dec_depth)\n            group_name = \"layer_%d_%s\" % (layer_id, group_name)\n            if name in no_lr_scale_list:\n                skip_scale = True\n                group_name = f\"{group_name}_no_lr_scale\"\n        else:\n            layer_id = 0\n            skip_scale = True\n\n        if group_name not in parameter_group_names:\n            if not skip_scale:\n                scale = layer_decay_values[layer_id]\n            else:\n                scale = 1.0\n\n            if \"enc_blocks\" in group_name:\n                scale *= 1.0\n            parameter_group_names[group_name] = {\n                \"weight_decay\": this_weight_decay,\n                \"params\": [],\n                \"lr_scale\": scale,\n            }\n            parameter_group_vars[group_name] = {\n                \"weight_decay\": this_weight_decay,\n                \"params\": [],\n                \"lr_scale\": scale,\n            }\n\n        parameter_group_vars[group_name][\"params\"].append(param)\n        parameter_group_names[group_name][\"params\"].append(name)\n    printer.info(\"Param groups = %s\" % json.dumps(parameter_group_names, indent=2))\n    return list(parameter_group_vars.values())\n\n\ndef adjust_learning_rate(optimizer, epoch, args):\n    \"\"\"Decay the learning rate with half-cycle cosine after warmup\"\"\"\n\n    if epoch < args.warmup_epochs:\n        lr = args.lr * epoch / args.warmup_epochs\n    else:\n        # lr = args.lr\n        lr = args.min_lr + (args.lr - args.min_lr) * 0.5 * (\n            1.0\n            + math.cos(\n                math.pi\n                * (epoch - args.warmup_epochs)\n                / (args.epochs - args.warmup_epochs)\n            )\n        )\n\n    for param_group in optimizer.param_groups:\n        if \"lr_scale\" in param_group:","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.adjust_learning_rate","uri":"program://Human3R/function/src.croco.utils.misc.adjust_learning_rate#L581-L603","kind":"function","name":"adjust_learning_rate","path":"src/croco/utils/misc.py","language":"python","start_line":581,"end_line":603,"context_start_line":561,"context_end_line":603,"code":"\n            if \"enc_blocks\" in group_name:\n                scale *= 1.0\n            parameter_group_names[group_name] = {\n                \"weight_decay\": this_weight_decay,\n                \"params\": [],\n                \"lr_scale\": scale,\n            }\n            parameter_group_vars[group_name] = {\n                \"weight_decay\": this_weight_decay,\n                \"params\": [],\n                \"lr_scale\": scale,\n            }\n\n        parameter_group_vars[group_name][\"params\"].append(param)\n        parameter_group_names[group_name][\"params\"].append(name)\n    printer.info(\"Param groups = %s\" % json.dumps(parameter_group_names, indent=2))\n    return list(parameter_group_vars.values())\n\n\ndef adjust_learning_rate(optimizer, epoch, args):\n    \"\"\"Decay the learning rate with half-cycle cosine after warmup\"\"\"\n\n    if epoch < args.warmup_epochs:\n        lr = args.lr * epoch / args.warmup_epochs\n    else:\n        # lr = args.lr\n        lr = args.min_lr + (args.lr - args.min_lr) * 0.5 * (\n            1.0\n            + math.cos(\n                math.pi\n                * (epoch - args.warmup_epochs)\n                / (args.epochs - args.warmup_epochs)\n            )\n        )\n\n    for param_group in optimizer.param_groups:\n        if \"lr_scale\" in param_group:\n            param_group[\"lr\"] = lr * param_group[\"lr_scale\"]\n        else:\n            param_group[\"lr\"] = lr\n\n    return lr","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.__init__","uri":"program://Human3R/function/src.croco.utils.misc.__init__#L282-L283","kind":"function","name":"__init__","path":"src/croco/utils/misc.py","language":"python","start_line":282,"end_line":283,"context_start_line":262,"context_end_line":303,"code":"    args.dist_backend = \"nccl\"\n    print(\n        \"| distributed init (rank {}): {}, gpu {}\".format(\n            args.rank, args.dist_url, args.gpu\n        ),\n        flush=True,\n    )\n    torch.distributed.init_process_group(\n        backend=args.dist_backend,\n        init_method=args.dist_url,\n        world_size=args.world_size,\n        rank=args.rank,\n    )\n    torch.distributed.barrier()\n    setup_for_distributed(args.rank == 0)\n\n\nclass NativeScalerWithGradNormCount:\n    state_dict_key = \"amp_scaler\"\n\n    def __init__(self, enabled=True, accelerator: Accelerator = None):\n        self.accelerator = accelerator\n\n    def __call__(\n        self,\n        loss,\n        optimizer,\n        clip_grad=None,\n        parameters=None,\n        create_graph=False,\n        update_grad=True,\n    ):\n        self.accelerator.backward(\n            loss, create_graph=create_graph\n        )  # .backward(create_graph=create_graph)\n        if update_grad:\n            if clip_grad is not None:\n                assert parameters is not None\n                # self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n                norm = self.accelerator.clip_grad_norm_(parameters, clip_grad)\n            else:\n                if self.accelerator.scaler is not None:","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.update","uri":"program://Human3R/function/src.croco.utils.misc.update#L96-L107","kind":"function","name":"update","path":"src/croco/utils/misc.py","language":"python","start_line":96,"end_line":107,"context_start_line":76,"context_end_line":127,"code":"\n    @property\n    def value(self):\n        return self.deque[-1]\n\n    def __str__(self):\n        return self.fmt.format(\n            median=self.median,\n            avg=self.avg,\n            global_avg=self.global_avg,\n            max=self.max,\n            value=self.value,\n        )\n\n\nclass MetricLogger(object):\n    def __init__(self, delimiter=\"\\t\"):\n        self.meters = defaultdict(SmoothedValue)\n        self.delimiter = delimiter\n\n    def update(self, **kwargs):\n        for k, v in kwargs.items():\n            if v is None:\n                continue\n            if isinstance(v, torch.Tensor):\n                if v.ndim > 0:\n                    continue\n                v = v.item()\n            if isinstance(v, list):\n                continue\n            assert isinstance(v, (float, int))\n            self.meters[k].update(v)\n\n    def __getattr__(self, attr):\n        if attr in self.meters:\n            return self.meters[attr]\n        if attr in self.__dict__:\n            return self.__dict__[attr]\n        raise AttributeError(\n            \"'{}' object has no attribute '{}'\".format(type(self).__name__, attr)\n        )\n\n    def __str__(self):\n        loss_str = []\n        for name, meter in self.meters.items():\n            loss_str.append(\"{}: {}\".format(name, str(meter)))\n        return self.delimiter.join(loss_str)\n\n    def synchronize_between_processes(self, accelerator):\n        for meter in self.meters.values():\n            meter.synchronize_between_processes(accelerator)\n","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.synchronize_between_processes","uri":"program://Human3R/function/src.croco.utils.misc.synchronize_between_processes#L124-L126","kind":"function","name":"synchronize_between_processes","path":"src/croco/utils/misc.py","language":"python","start_line":124,"end_line":126,"context_start_line":104,"context_end_line":146,"code":"            if isinstance(v, list):\n                continue\n            assert isinstance(v, (float, int))\n            self.meters[k].update(v)\n\n    def __getattr__(self, attr):\n        if attr in self.meters:\n            return self.meters[attr]\n        if attr in self.__dict__:\n            return self.__dict__[attr]\n        raise AttributeError(\n            \"'{}' object has no attribute '{}'\".format(type(self).__name__, attr)\n        )\n\n    def __str__(self):\n        loss_str = []\n        for name, meter in self.meters.items():\n            loss_str.append(\"{}: {}\".format(name, str(meter)))\n        return self.delimiter.join(loss_str)\n\n    def synchronize_between_processes(self, accelerator):\n        for meter in self.meters.values():\n            meter.synchronize_between_processes(accelerator)\n\n    def add_meter(self, name, meter):\n        self.meters[name] = meter\n\n    def log_every(\n        self, iterable, print_freq, accelerator: Accelerator, header=None, max_iter=None\n    ):\n        i = 0\n        if not header:\n            header = \"\"\n        start_time = time.time()\n        end = time.time()\n        iter_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        data_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        len_iterable = min(len(iterable), max_iter) if max_iter else len(iterable)\n        space_fmt = \":\" + str(len(str(len_iterable))) + \"d\"\n        log_msg = [\n            header,\n            \"[{0\" + space_fmt + \"}/{1}]\",\n            \"eta: {eta}\",","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.median","uri":"program://Human3R/function/src.croco.utils.misc.median#L62-L63","kind":"function","name":"median","path":"src/croco/utils/misc.py","language":"python","start_line":62,"end_line":63,"context_start_line":42,"context_end_line":83,"code":"\n    def update(self, value, n=1):\n        self.deque.append(value)\n        self.count += n\n        self.total += value * n\n\n    def synchronize_between_processes(self, accelerator: Accelerator):\n        \"\"\"Synchronize the count and total across all processes.\"\"\"\n        if accelerator.num_processes == 1:\n            return\n        t = torch.tensor(\n            [self.count, self.total], dtype=torch.float64, device=accelerator.device\n        )\n        accelerator.wait_for_everyone()\n        accelerator.reduce(t, reduction=\"sum\")\n        t = t.tolist()\n        self.count = int(t[0])\n        self.total = t[1]\n\n    @property\n    def median(self):\n        return torch.tensor(list(self.deque)).median().item()\n\n    @property\n    def avg(self):\n        return torch.tensor(list(self.deque), dtype=torch.float32).mean().item()\n\n    @property\n    def global_avg(self):\n        return self.total / self.count\n\n    @property\n    def max(self):\n        return max(self.deque)\n\n    @property\n    def value(self):\n        return self.deque[-1]\n\n    def __str__(self):\n        return self.fmt.format(\n            median=self.median,","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.avg","uri":"program://Human3R/function/src.croco.utils.misc.avg#L66-L67","kind":"function","name":"avg","path":"src/croco/utils/misc.py","language":"python","start_line":66,"end_line":67,"context_start_line":46,"context_end_line":87,"code":"        self.total += value * n\n\n    def synchronize_between_processes(self, accelerator: Accelerator):\n        \"\"\"Synchronize the count and total across all processes.\"\"\"\n        if accelerator.num_processes == 1:\n            return\n        t = torch.tensor(\n            [self.count, self.total], dtype=torch.float64, device=accelerator.device\n        )\n        accelerator.wait_for_everyone()\n        accelerator.reduce(t, reduction=\"sum\")\n        t = t.tolist()\n        self.count = int(t[0])\n        self.total = t[1]\n\n    @property\n    def median(self):\n        return torch.tensor(list(self.deque)).median().item()\n\n    @property\n    def avg(self):\n        return torch.tensor(list(self.deque), dtype=torch.float32).mean().item()\n\n    @property\n    def global_avg(self):\n        return self.total / self.count\n\n    @property\n    def max(self):\n        return max(self.deque)\n\n    @property\n    def value(self):\n        return self.deque[-1]\n\n    def __str__(self):\n        return self.fmt.format(\n            median=self.median,\n            avg=self.avg,\n            global_avg=self.global_avg,\n            max=self.max,\n            value=self.value,","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.global_avg","uri":"program://Human3R/function/src.croco.utils.misc.global_avg#L70-L71","kind":"function","name":"global_avg","path":"src/croco/utils/misc.py","language":"python","start_line":70,"end_line":71,"context_start_line":50,"context_end_line":91,"code":"        if accelerator.num_processes == 1:\n            return\n        t = torch.tensor(\n            [self.count, self.total], dtype=torch.float64, device=accelerator.device\n        )\n        accelerator.wait_for_everyone()\n        accelerator.reduce(t, reduction=\"sum\")\n        t = t.tolist()\n        self.count = int(t[0])\n        self.total = t[1]\n\n    @property\n    def median(self):\n        return torch.tensor(list(self.deque)).median().item()\n\n    @property\n    def avg(self):\n        return torch.tensor(list(self.deque), dtype=torch.float32).mean().item()\n\n    @property\n    def global_avg(self):\n        return self.total / self.count\n\n    @property\n    def max(self):\n        return max(self.deque)\n\n    @property\n    def value(self):\n        return self.deque[-1]\n\n    def __str__(self):\n        return self.fmt.format(\n            median=self.median,\n            avg=self.avg,\n            global_avg=self.global_avg,\n            max=self.max,\n            value=self.value,\n        )\n\n\nclass MetricLogger(object):","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.max","uri":"program://Human3R/function/src.croco.utils.misc.max#L74-L75","kind":"function","name":"max","path":"src/croco/utils/misc.py","language":"python","start_line":74,"end_line":75,"context_start_line":54,"context_end_line":95,"code":"        )\n        accelerator.wait_for_everyone()\n        accelerator.reduce(t, reduction=\"sum\")\n        t = t.tolist()\n        self.count = int(t[0])\n        self.total = t[1]\n\n    @property\n    def median(self):\n        return torch.tensor(list(self.deque)).median().item()\n\n    @property\n    def avg(self):\n        return torch.tensor(list(self.deque), dtype=torch.float32).mean().item()\n\n    @property\n    def global_avg(self):\n        return self.total / self.count\n\n    @property\n    def max(self):\n        return max(self.deque)\n\n    @property\n    def value(self):\n        return self.deque[-1]\n\n    def __str__(self):\n        return self.fmt.format(\n            median=self.median,\n            avg=self.avg,\n            global_avg=self.global_avg,\n            max=self.max,\n            value=self.value,\n        )\n\n\nclass MetricLogger(object):\n    def __init__(self, delimiter=\"\\t\"):\n        self.meters = defaultdict(SmoothedValue)\n        self.delimiter = delimiter\n","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.value","uri":"program://Human3R/function/src.croco.utils.misc.value#L78-L79","kind":"function","name":"value","path":"src/croco/utils/misc.py","language":"python","start_line":78,"end_line":79,"context_start_line":58,"context_end_line":99,"code":"        self.count = int(t[0])\n        self.total = t[1]\n\n    @property\n    def median(self):\n        return torch.tensor(list(self.deque)).median().item()\n\n    @property\n    def avg(self):\n        return torch.tensor(list(self.deque), dtype=torch.float32).mean().item()\n\n    @property\n    def global_avg(self):\n        return self.total / self.count\n\n    @property\n    def max(self):\n        return max(self.deque)\n\n    @property\n    def value(self):\n        return self.deque[-1]\n\n    def __str__(self):\n        return self.fmt.format(\n            median=self.median,\n            avg=self.avg,\n            global_avg=self.global_avg,\n            max=self.max,\n            value=self.value,\n        )\n\n\nclass MetricLogger(object):\n    def __init__(self, delimiter=\"\\t\"):\n        self.meters = defaultdict(SmoothedValue)\n        self.delimiter = delimiter\n\n    def update(self, **kwargs):\n        for k, v in kwargs.items():\n            if v is None:\n                continue","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.__str__","uri":"program://Human3R/function/src.croco.utils.misc.__str__#L118-L122","kind":"function","name":"__str__","path":"src/croco/utils/misc.py","language":"python","start_line":118,"end_line":122,"context_start_line":98,"context_end_line":142,"code":"            if v is None:\n                continue\n            if isinstance(v, torch.Tensor):\n                if v.ndim > 0:\n                    continue\n                v = v.item()\n            if isinstance(v, list):\n                continue\n            assert isinstance(v, (float, int))\n            self.meters[k].update(v)\n\n    def __getattr__(self, attr):\n        if attr in self.meters:\n            return self.meters[attr]\n        if attr in self.__dict__:\n            return self.__dict__[attr]\n        raise AttributeError(\n            \"'{}' object has no attribute '{}'\".format(type(self).__name__, attr)\n        )\n\n    def __str__(self):\n        loss_str = []\n        for name, meter in self.meters.items():\n            loss_str.append(\"{}: {}\".format(name, str(meter)))\n        return self.delimiter.join(loss_str)\n\n    def synchronize_between_processes(self, accelerator):\n        for meter in self.meters.values():\n            meter.synchronize_between_processes(accelerator)\n\n    def add_meter(self, name, meter):\n        self.meters[name] = meter\n\n    def log_every(\n        self, iterable, print_freq, accelerator: Accelerator, header=None, max_iter=None\n    ):\n        i = 0\n        if not header:\n            header = \"\"\n        start_time = time.time()\n        end = time.time()\n        iter_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        data_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        len_iterable = min(len(iterable), max_iter) if max_iter else len(iterable)\n        space_fmt = \":\" + str(len(str(len_iterable))) + \"d\"","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.__getattr__","uri":"program://Human3R/function/src.croco.utils.misc.__getattr__#L109-L116","kind":"function","name":"__getattr__","path":"src/croco/utils/misc.py","language":"python","start_line":109,"end_line":116,"context_start_line":89,"context_end_line":136,"code":"\n\nclass MetricLogger(object):\n    def __init__(self, delimiter=\"\\t\"):\n        self.meters = defaultdict(SmoothedValue)\n        self.delimiter = delimiter\n\n    def update(self, **kwargs):\n        for k, v in kwargs.items():\n            if v is None:\n                continue\n            if isinstance(v, torch.Tensor):\n                if v.ndim > 0:\n                    continue\n                v = v.item()\n            if isinstance(v, list):\n                continue\n            assert isinstance(v, (float, int))\n            self.meters[k].update(v)\n\n    def __getattr__(self, attr):\n        if attr in self.meters:\n            return self.meters[attr]\n        if attr in self.__dict__:\n            return self.__dict__[attr]\n        raise AttributeError(\n            \"'{}' object has no attribute '{}'\".format(type(self).__name__, attr)\n        )\n\n    def __str__(self):\n        loss_str = []\n        for name, meter in self.meters.items():\n            loss_str.append(\"{}: {}\".format(name, str(meter)))\n        return self.delimiter.join(loss_str)\n\n    def synchronize_between_processes(self, accelerator):\n        for meter in self.meters.values():\n            meter.synchronize_between_processes(accelerator)\n\n    def add_meter(self, name, meter):\n        self.meters[name] = meter\n\n    def log_every(\n        self, iterable, print_freq, accelerator: Accelerator, header=None, max_iter=None\n    ):\n        i = 0\n        if not header:\n            header = \"\"","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.add_meter","uri":"program://Human3R/function/src.croco.utils.misc.add_meter#L128-L129","kind":"function","name":"add_meter","path":"src/croco/utils/misc.py","language":"python","start_line":128,"end_line":129,"context_start_line":108,"context_end_line":149,"code":"\n    def __getattr__(self, attr):\n        if attr in self.meters:\n            return self.meters[attr]\n        if attr in self.__dict__:\n            return self.__dict__[attr]\n        raise AttributeError(\n            \"'{}' object has no attribute '{}'\".format(type(self).__name__, attr)\n        )\n\n    def __str__(self):\n        loss_str = []\n        for name, meter in self.meters.items():\n            loss_str.append(\"{}: {}\".format(name, str(meter)))\n        return self.delimiter.join(loss_str)\n\n    def synchronize_between_processes(self, accelerator):\n        for meter in self.meters.values():\n            meter.synchronize_between_processes(accelerator)\n\n    def add_meter(self, name, meter):\n        self.meters[name] = meter\n\n    def log_every(\n        self, iterable, print_freq, accelerator: Accelerator, header=None, max_iter=None\n    ):\n        i = 0\n        if not header:\n            header = \"\"\n        start_time = time.time()\n        end = time.time()\n        iter_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        data_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        len_iterable = min(len(iterable), max_iter) if max_iter else len(iterable)\n        space_fmt = \":\" + str(len(str(len_iterable))) + \"d\"\n        log_msg = [\n            header,\n            \"[{0\" + space_fmt + \"}/{1}]\",\n            \"eta: {eta}\",\n            \"{meters}\",\n            \"time: {time}\",\n            \"data: {data}\",","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.log_every","uri":"program://Human3R/function/src.croco.utils.misc.log_every#L131-L198","kind":"function","name":"log_every","path":"src/croco/utils/misc.py","language":"python","start_line":131,"end_line":198,"context_start_line":111,"context_end_line":218,"code":"            return self.meters[attr]\n        if attr in self.__dict__:\n            return self.__dict__[attr]\n        raise AttributeError(\n            \"'{}' object has no attribute '{}'\".format(type(self).__name__, attr)\n        )\n\n    def __str__(self):\n        loss_str = []\n        for name, meter in self.meters.items():\n            loss_str.append(\"{}: {}\".format(name, str(meter)))\n        return self.delimiter.join(loss_str)\n\n    def synchronize_between_processes(self, accelerator):\n        for meter in self.meters.values():\n            meter.synchronize_between_processes(accelerator)\n\n    def add_meter(self, name, meter):\n        self.meters[name] = meter\n\n    def log_every(\n        self, iterable, print_freq, accelerator: Accelerator, header=None, max_iter=None\n    ):\n        i = 0\n        if not header:\n            header = \"\"\n        start_time = time.time()\n        end = time.time()\n        iter_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        data_time = SmoothedValue(fmt=\"{avg:.4f}\")\n        len_iterable = min(len(iterable), max_iter) if max_iter else len(iterable)\n        space_fmt = \":\" + str(len(str(len_iterable))) + \"d\"\n        log_msg = [\n            header,\n            \"[{0\" + space_fmt + \"}/{1}]\",\n            \"eta: {eta}\",\n            \"{meters}\",\n            \"time: {time}\",\n            \"data: {data}\",\n        ]\n        if torch.cuda.is_available():\n            log_msg.append(\"max mem: {memory:.0f}\")\n        log_msg = self.delimiter.join(log_msg)\n        MB = 1024.0 * 1024.0\n        for it, obj in enumerate(iterable):\n            data_time.update(time.time() - end)\n            yield obj\n            iter_time.update(time.time() - end)\n            if i % print_freq == 0 or i == len_iterable - 1:\n                eta_seconds = iter_time.global_avg * (len_iterable - i)\n                eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))\n                if torch.cuda.is_available():\n                    if accelerator.is_main_process:\n                        printer.info(\n                            log_msg.format(\n                                i,\n                                len_iterable,\n                                eta=eta_string,\n                                meters=str(self),\n                                time=str(iter_time),\n                                data=str(data_time),\n                                memory=torch.cuda.max_memory_allocated() / MB,\n                            )\n                        )\n                else:\n                    if accelerator.is_main_process:\n                        printer.info(\n                            log_msg.format(\n                                i,\n                                len_iterable,\n                                eta=eta_string,\n                                meters=str(self),\n                                time=str(iter_time),\n                                data=str(data_time),\n                            )\n                        )\n            i += 1\n            end = time.time()\n            if max_iter and it >= max_iter:\n                break\n        total_time = time.time() - start_time\n        total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n        if accelerator.is_main_process:\n            printer.info(\n                \"{} Total time: {} ({:.4f} s / it)\".format(\n                    header, total_time_str, total_time / len_iterable\n                )\n            )\n\n\ndef setup_for_distributed(is_master):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (get_world_size() > 8)\n        if is_master or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef is_dist_avail_and_initialized():","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.print","uri":"program://Human3R/function/src.croco.utils.misc.print#L207-L213","kind":"function","name":"print","path":"src/croco/utils/misc.py","language":"python","start_line":207,"end_line":213,"context_start_line":187,"context_end_line":233,"code":"            i += 1\n            end = time.time()\n            if max_iter and it >= max_iter:\n                break\n        total_time = time.time() - start_time\n        total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n        if accelerator.is_main_process:\n            printer.info(\n                \"{} Total time: {} ({:.4f} s / it)\".format(\n                    header, total_time_str, total_time / len_iterable\n                )\n            )\n\n\ndef setup_for_distributed(is_master):\n    \"\"\"\n    This function disables printing when not in master process\n    \"\"\"\n    builtin_print = builtins.print\n\n    def print(*args, **kwargs):\n        force = kwargs.pop(\"force\", False)\n        force = force or (get_world_size() > 8)\n        if is_master or force:\n            now = datetime.datetime.now().time()\n            builtin_print(\"[{}] \".format(now), end=\"\")  # print with time stamp\n            builtin_print(*args, **kwargs)\n\n    builtins.print = print\n\n\ndef is_dist_avail_and_initialized():\n    if not dist.is_available():\n        return False\n    if not dist.is_initialized():\n        return False\n    return True\n\n\ndef get_world_size():\n    if not is_dist_avail_and_initialized():\n        return 1\n    return dist.get_world_size()\n\n\ndef get_rank():\n    if not is_dist_avail_and_initialized():","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.__call__","uri":"program://Human3R/function/src.croco.utils.misc.__call__#L285-L309","kind":"function","name":"__call__","path":"src/croco/utils/misc.py","language":"python","start_line":285,"end_line":309,"context_start_line":265,"context_end_line":329,"code":"            args.rank, args.dist_url, args.gpu\n        ),\n        flush=True,\n    )\n    torch.distributed.init_process_group(\n        backend=args.dist_backend,\n        init_method=args.dist_url,\n        world_size=args.world_size,\n        rank=args.rank,\n    )\n    torch.distributed.barrier()\n    setup_for_distributed(args.rank == 0)\n\n\nclass NativeScalerWithGradNormCount:\n    state_dict_key = \"amp_scaler\"\n\n    def __init__(self, enabled=True, accelerator: Accelerator = None):\n        self.accelerator = accelerator\n\n    def __call__(\n        self,\n        loss,\n        optimizer,\n        clip_grad=None,\n        parameters=None,\n        create_graph=False,\n        update_grad=True,\n    ):\n        self.accelerator.backward(\n            loss, create_graph=create_graph\n        )  # .backward(create_graph=create_graph)\n        if update_grad:\n            if clip_grad is not None:\n                assert parameters is not None\n                # self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n                norm = self.accelerator.clip_grad_norm_(parameters, clip_grad)\n            else:\n                if self.accelerator.scaler is not None:\n                    self.accelerator.unscale_gradients()\n                norm = get_grad_norm_(parameters)\n            optimizer.step()\n        else:\n            norm = None\n        return norm\n\n    def state_dict(self):\n        if self.accelerator.scaler is not None:\n            return self.accelerator.scaler.state_dict()\n        else:\n            return {}\n\n    def load_state_dict(self, state_dict):\n        if self.accelerator.scaler is not None:\n            self.accelerator.scaler.load_state_dict(state_dict)\n\n\n# class NativeScalerWithGradNormCount:\n#     state_dict_key = \"amp_scaler\"\n\n#     def __init__(self, enabled=True, accelerator:Accelerator=None):\n#         self._scaler = torch.cuda.amp.GradScaler(enabled=enabled)\n#         self.accelerator = accelerator\n\n#     def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True):","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.state_dict","uri":"program://Human3R/function/src.croco.utils.misc.state_dict#L311-L315","kind":"function","name":"state_dict","path":"src/croco/utils/misc.py","language":"python","start_line":311,"end_line":315,"context_start_line":291,"context_end_line":335,"code":"        create_graph=False,\n        update_grad=True,\n    ):\n        self.accelerator.backward(\n            loss, create_graph=create_graph\n        )  # .backward(create_graph=create_graph)\n        if update_grad:\n            if clip_grad is not None:\n                assert parameters is not None\n                # self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n                norm = self.accelerator.clip_grad_norm_(parameters, clip_grad)\n            else:\n                if self.accelerator.scaler is not None:\n                    self.accelerator.unscale_gradients()\n                norm = get_grad_norm_(parameters)\n            optimizer.step()\n        else:\n            norm = None\n        return norm\n\n    def state_dict(self):\n        if self.accelerator.scaler is not None:\n            return self.accelerator.scaler.state_dict()\n        else:\n            return {}\n\n    def load_state_dict(self, state_dict):\n        if self.accelerator.scaler is not None:\n            self.accelerator.scaler.load_state_dict(state_dict)\n\n\n# class NativeScalerWithGradNormCount:\n#     state_dict_key = \"amp_scaler\"\n\n#     def __init__(self, enabled=True, accelerator:Accelerator=None):\n#         self._scaler = torch.cuda.amp.GradScaler(enabled=enabled)\n#         self.accelerator = accelerator\n\n#     def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True):\n#         # self.accelerator.backward(loss, create_graph=create_graph) #.backward(create_graph=create_graph)\n#         self._scaler.scale(loss).backward(create_graph=create_graph)\n#         if update_grad:\n#             if clip_grad is not None:\n#                 assert parameters is not None\n#                 # #self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.utils.misc.load_state_dict","uri":"program://Human3R/function/src.croco.utils.misc.load_state_dict#L317-L319","kind":"function","name":"load_state_dict","path":"src/croco/utils/misc.py","language":"python","start_line":317,"end_line":319,"context_start_line":297,"context_end_line":339,"code":"        if update_grad:\n            if clip_grad is not None:\n                assert parameters is not None\n                # self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n                norm = self.accelerator.clip_grad_norm_(parameters, clip_grad)\n            else:\n                if self.accelerator.scaler is not None:\n                    self.accelerator.unscale_gradients()\n                norm = get_grad_norm_(parameters)\n            optimizer.step()\n        else:\n            norm = None\n        return norm\n\n    def state_dict(self):\n        if self.accelerator.scaler is not None:\n            return self.accelerator.scaler.state_dict()\n        else:\n            return {}\n\n    def load_state_dict(self, state_dict):\n        if self.accelerator.scaler is not None:\n            self.accelerator.scaler.load_state_dict(state_dict)\n\n\n# class NativeScalerWithGradNormCount:\n#     state_dict_key = \"amp_scaler\"\n\n#     def __init__(self, enabled=True, accelerator:Accelerator=None):\n#         self._scaler = torch.cuda.amp.GradScaler(enabled=enabled)\n#         self.accelerator = accelerator\n\n#     def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True):\n#         # self.accelerator.backward(loss, create_graph=create_graph) #.backward(create_graph=create_graph)\n#         self._scaler.scale(loss).backward(create_graph=create_graph)\n#         if update_grad:\n#             if clip_grad is not None:\n#                 assert parameters is not None\n#                 # #self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n#                 # norm = self.accelerator.clip_grad_norm_(parameters, clip_grad)\n#                 self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place\n#                 norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad)\n#             else:","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion","uri":"program://Human3R/module/src.croco.stereoflow.criterion#L1-L351","kind":"module","name":"src.croco.stereoflow.criterion","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":1,"end_line":351,"context_start_line":1,"context_end_line":351,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Losses, metrics per batch, metrics per dataset\n# --------------------------------------------------------\n\nimport torch\nfrom torch import nn\nimport torch.nn.functional as F\n\n\ndef _get_gtnorm(gt):\n    if gt.size(1) == 1:  # stereo\n        return gt\n    # flow\n    return torch.sqrt(torch.sum(gt**2, dim=1, keepdims=True))  # Bx1xHxW\n\n\n############ losses without confidence\n\n\nclass L1Loss(nn.Module):\n\n    def __init__(self, max_gtnorm=None):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = False\n\n    def _error(self, gt, predictions):\n        return torch.abs(gt - predictions)\n\n    def forward(self, predictions, gt, inspect=False):\n        mask = torch.isfinite(gt)\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt).expand(-1, gt.size(1), -1, -1) < self.max_gtnorm\n        if inspect:\n            return self._error(gt, predictions)\n        return self._error(gt[mask], predictions[mask]).mean()\n\n\n############## losses with confience\n## there are several parametrizations\n\n\nclass LaplacianLoss(nn.Module):  # used for CroCo-Stereo on ETH3D, d'=exp(d)\n\n    def __init__(self, max_gtnorm=None):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / torch.exp(conf[mask])\n            + conf[mask]\n        ).mean()  # + torch.log(2) => which is a constant\n\n\nclass LaplacianLossBounded(\n    nn.Module\n):  # used for CroCo-Flow ; in the equation of the paper, we have a=1/b\n    def __init__(self, max_gtnorm=10000.0, a=0.25, b=4.0):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n        self.a, self.b = a, b\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        conf = (self.b - self.a) * torch.sigmoid(conf) + self.a\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / conf[mask]\n            + torch.log(conf)[mask]\n        ).mean()  # + torch.log(2) => which is a constant\n\n\nclass LaplacianLossBounded2(\n    nn.Module\n):  # used for CroCo-Stereo (except for ETH3D) ; in the equation of the paper, we have a=b\n    def __init__(self, max_gtnorm=None, a=3.0, b=3.0):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n        self.a, self.b = a, b\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        conf = 2 * self.a * (torch.sigmoid(conf / self.b) - 0.5)\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / torch.exp(conf[mask])\n            + conf[mask]\n        ).mean()  # + torch.log(2) => which is a constant\n\n\n############## metrics per batch\n\n\nclass StereoMetrics(nn.Module):\n\n    def __init__(self, do_quantile=False):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 2, 3]\n        self.do_quantile = do_quantile\n\n    def forward(self, predictions, gt):\n        B = predictions.size(0)\n        metrics = {}\n        gtcopy = gt.clone()\n        mask = torch.isfinite(gtcopy)\n        gtcopy[~mask] = (\n            999999.0  # we make a copy and put a non-infinite value, such that it does not become nan once multiplied by the mask value 0\n        )\n        Npx = mask.view(B, -1).sum(dim=1)\n        L1error = (torch.abs(gtcopy - predictions) * mask).view(B, -1)\n        L2error = (torch.square(gtcopy - predictions) * mask).view(B, -1)\n        # avgerr\n        metrics[\"avgerr\"] = torch.mean(L1error.sum(dim=1) / Npx)\n        # rmse\n        metrics[\"rmse\"] = torch.sqrt(L2error.sum(dim=1) / Npx).mean(dim=0)\n        # err > t for t in [0.5,1,2,3]\n        for ths in self.bad_ths:\n            metrics[\"bad@{:.1f}\".format(ths)] = (\n                ((L1error > ths) * mask.view(B, -1)).sum(dim=1) / Npx\n            ).mean(dim=0) * 100\n        return metrics\n\n\nclass FlowMetrics(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [1, 3, 5]\n\n    def forward(self, predictions, gt):\n        B = predictions.size(0)\n        metrics = {}\n        mask = torch.isfinite(gt[:, 0, :, :])  # both x and y would be infinite\n        Npx = mask.view(B, -1).sum(dim=1)\n        gtcopy = (\n            gt.clone()\n        )  # to compute L1/L2 error, we need to have non-infinite value, the error computed at this locations will be ignored\n        gtcopy[:, 0, :, :][~mask] = 999999.0\n        gtcopy[:, 1, :, :][~mask] = 999999.0\n        L1error = (torch.abs(gtcopy - predictions).sum(dim=1) * mask).view(B, -1)\n        L2error = (\n            torch.sqrt(torch.sum(torch.square(gtcopy - predictions), dim=1)) * mask\n        ).view(B, -1)\n        metrics[\"L1err\"] = torch.mean(L1error.sum(dim=1) / Npx)\n        metrics[\"EPE\"] = torch.mean(L2error.sum(dim=1) / Npx)\n        for ths in self.bad_ths:\n            metrics[\"bad@{:.1f}\".format(ths)] = (\n                ((L2error > ths) * mask.view(B, -1)).sum(dim=1) / Npx\n            ).mean(dim=0) * 100\n        return metrics\n\n\n############## metrics per dataset\n## we update the average and maintain the number of pixels while adding data batch per batch\n## at the beggining, call reset()\n## after each batch, call add_batch(...)\n## at the end: call get_results()\n\n\nclass StereoDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 2, 3]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self._metrics = None\n\n    def add_batch(self, predictions, gt):\n        assert predictions.size(1) == 1, predictions.size()\n        assert gt.size(1) == 1, gt.size()\n        if (\n            gt.size(2) == predictions.size(2) * 2\n            and gt.size(3) == predictions.size(3) * 2\n        ):  # special case for Spring ...\n            L1err = torch.minimum(\n                torch.minimum(\n                    torch.minimum(\n                        torch.sum(torch.abs(gt[:, :, 0::2, 0::2] - predictions), dim=1),\n                        torch.sum(torch.abs(gt[:, :, 1::2, 0::2] - predictions), dim=1),\n                    ),\n                    torch.sum(torch.abs(gt[:, :, 0::2, 1::2] - predictions), dim=1),\n                ),\n                torch.sum(torch.abs(gt[:, :, 1::2, 1::2] - predictions), dim=1),\n            )\n            valid = torch.isfinite(L1err)\n        else:\n            valid = torch.isfinite(gt[:, 0, :, :])  # both x and y would be infinite\n            L1err = torch.sum(torch.abs(gt - predictions), dim=1)\n        N = valid.sum()\n        Nnew = self.agg_N + N\n        self.agg_L1err = (\n            float(self.agg_N) / Nnew * self.agg_L1err\n            + L1err[valid].mean().cpu() * float(N) / Nnew\n        )\n        self.agg_N = Nnew\n        for i, th in enumerate(self.bad_ths):\n            self.agg_Nbad[i] += (L1err[valid] > th).sum().cpu()\n\n    def _compute_metrics(self):\n        if self._metrics is not None:\n            return\n        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics\n\n\nclass FlowDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 3, 5]\n        self.speed_ths = [(0, 10), (10, 40), (40, torch.inf)]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_L2err = torch.tensor(0.0)  # L2 (=EPE) error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self.agg_EPEspeed = [\n            torch.tensor(0.0) for _ in self.speed_ths\n        ]  # EPE per speed bin so far\n        self.agg_Nspeed = [0 for _ in self.speed_ths]  # N pixels per speed bin so far\n        self._metrics = None\n        self.pairname_results = {}\n\n    def add_batch(self, predictions, gt):\n        assert predictions.size(1) == 2, predictions.size()\n        assert gt.size(1) == 2, gt.size()\n        if (\n            gt.size(2) == predictions.size(2) * 2\n            and gt.size(3) == predictions.size(3) * 2\n        ):  # special case for Spring ...\n            L1err = torch.minimum(\n                torch.minimum(\n                    torch.minimum(\n                        torch.sum(torch.abs(gt[:, :, 0::2, 0::2] - predictions), dim=1),\n                        torch.sum(torch.abs(gt[:, :, 1::2, 0::2] - predictions), dim=1),\n                    ),\n                    torch.sum(torch.abs(gt[:, :, 0::2, 1::2] - predictions), dim=1),\n                ),\n                torch.sum(torch.abs(gt[:, :, 1::2, 1::2] - predictions), dim=1),\n            )\n            L2err = torch.minimum(\n                torch.minimum(\n                    torch.minimum(\n                        torch.sqrt(\n                            torch.sum(\n                                torch.square(gt[:, :, 0::2, 0::2] - predictions), dim=1\n                            )\n                        ),\n                        torch.sqrt(\n                            torch.sum(\n                                torch.square(gt[:, :, 1::2, 0::2] - predictions), dim=1\n                            )\n                        ),\n                    ),\n                    torch.sqrt(\n                        torch.sum(\n                            torch.square(gt[:, :, 0::2, 1::2] - predictions), dim=1\n                        )\n                    ),\n                ),\n                torch.sqrt(\n                    torch.sum(torch.square(gt[:, :, 1::2, 1::2] - predictions), dim=1)\n                ),\n            )\n            valid = torch.isfinite(L1err)\n            gtspeed = (\n                torch.sqrt(torch.sum(torch.square(gt[:, :, 0::2, 0::2]), dim=1))\n                + torch.sqrt(torch.sum(torch.square(gt[:, :, 0::2, 1::2]), dim=1))\n                + torch.sqrt(torch.sum(torch.square(gt[:, :, 1::2, 0::2]), dim=1))\n                + torch.sqrt(torch.sum(torch.square(gt[:, :, 1::2, 1::2]), dim=1))\n            ) / 4.0  # let's just average them\n        else:\n            valid = torch.isfinite(gt[:, 0, :, :])  # both x and y would be infinite\n            L1err = torch.sum(torch.abs(gt - predictions), dim=1)\n            L2err = torch.sqrt(torch.sum(torch.square(gt - predictions), dim=1))\n            gtspeed = torch.sqrt(torch.sum(torch.square(gt), dim=1))\n        N = valid.sum()\n        Nnew = self.agg_N + N\n        self.agg_L1err = (\n            float(self.agg_N) / Nnew * self.agg_L1err\n            + L1err[valid].mean().cpu() * float(N) / Nnew\n        )\n        self.agg_L2err = (\n            float(self.agg_N) / Nnew * self.agg_L2err\n            + L2err[valid].mean().cpu() * float(N) / Nnew\n        )\n        self.agg_N = Nnew\n        for i, th in enumerate(self.bad_ths):\n            self.agg_Nbad[i] += (L2err[valid] > th).sum().cpu()\n        for i, (th1, th2) in enumerate(self.speed_ths):\n            vv = (gtspeed[valid] >= th1) * (gtspeed[valid] < th2)\n            iNspeed = vv.sum()\n            if iNspeed == 0:\n                continue\n            iNnew = self.agg_Nspeed[i] + iNspeed\n            self.agg_EPEspeed[i] = (\n                float(self.agg_Nspeed[i]) / iNnew * self.agg_EPEspeed[i]\n                + float(iNspeed) / iNnew * L2err[valid][vv].mean().cpu()\n            )\n            self.agg_Nspeed[i] = iNnew\n\n    def _compute_metrics(self):\n        if self._metrics is not None:\n            return\n        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        out[\"EPE\"] = self.agg_L2err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        for i, (th1, th2) in enumerate(self.speed_ths):\n            out[\"s{:d}{:s}\".format(th1, \"-\" + str(th2) if th2 < torch.inf else \"+\")] = (\n                self.agg_EPEspeed[i].item()\n            )\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion._get_gtnorm","uri":"program://Human3R/function/src.croco.stereoflow.criterion._get_gtnorm#L13-L17","kind":"function","name":"_get_gtnorm","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":13,"end_line":17,"context_start_line":1,"context_end_line":37,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Losses, metrics per batch, metrics per dataset\n# --------------------------------------------------------\n\nimport torch\nfrom torch import nn\nimport torch.nn.functional as F\n\n\ndef _get_gtnorm(gt):\n    if gt.size(1) == 1:  # stereo\n        return gt\n    # flow\n    return torch.sqrt(torch.sum(gt**2, dim=1, keepdims=True))  # Bx1xHxW\n\n\n############ losses without confidence\n\n\nclass L1Loss(nn.Module):\n\n    def __init__(self, max_gtnorm=None):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = False\n\n    def _error(self, gt, predictions):\n        return torch.abs(gt - predictions)\n\n    def forward(self, predictions, gt, inspect=False):\n        mask = torch.isfinite(gt)\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt).expand(-1, gt.size(1), -1, -1) < self.max_gtnorm\n        if inspect:","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.L1Loss","uri":"program://Human3R/class/src.croco.stereoflow.criterion.L1Loss#L23-L39","kind":"class","name":"L1Loss","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":23,"end_line":39,"context_start_line":3,"context_end_line":59,"code":"\n# --------------------------------------------------------\n# Losses, metrics per batch, metrics per dataset\n# --------------------------------------------------------\n\nimport torch\nfrom torch import nn\nimport torch.nn.functional as F\n\n\ndef _get_gtnorm(gt):\n    if gt.size(1) == 1:  # stereo\n        return gt\n    # flow\n    return torch.sqrt(torch.sum(gt**2, dim=1, keepdims=True))  # Bx1xHxW\n\n\n############ losses without confidence\n\n\nclass L1Loss(nn.Module):\n\n    def __init__(self, max_gtnorm=None):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = False\n\n    def _error(self, gt, predictions):\n        return torch.abs(gt - predictions)\n\n    def forward(self, predictions, gt, inspect=False):\n        mask = torch.isfinite(gt)\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt).expand(-1, gt.size(1), -1, -1) < self.max_gtnorm\n        if inspect:\n            return self._error(gt, predictions)\n        return self._error(gt[mask], predictions[mask]).mean()\n\n\n############## losses with confience\n## there are several parametrizations\n\n\nclass LaplacianLoss(nn.Module):  # used for CroCo-Stereo on ETH3D, d'=exp(d)\n\n    def __init__(self, max_gtnorm=None):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        return (","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.LaplacianLoss","uri":"program://Human3R/class/src.croco.stereoflow.criterion.LaplacianLoss#L46-L62","kind":"class","name":"LaplacianLoss","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":46,"end_line":62,"context_start_line":26,"context_end_line":82,"code":"        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = False\n\n    def _error(self, gt, predictions):\n        return torch.abs(gt - predictions)\n\n    def forward(self, predictions, gt, inspect=False):\n        mask = torch.isfinite(gt)\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt).expand(-1, gt.size(1), -1, -1) < self.max_gtnorm\n        if inspect:\n            return self._error(gt, predictions)\n        return self._error(gt[mask], predictions[mask]).mean()\n\n\n############## losses with confience\n## there are several parametrizations\n\n\nclass LaplacianLoss(nn.Module):  # used for CroCo-Stereo on ETH3D, d'=exp(d)\n\n    def __init__(self, max_gtnorm=None):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / torch.exp(conf[mask])\n            + conf[mask]\n        ).mean()  # + torch.log(2) => which is a constant\n\n\nclass LaplacianLossBounded(\n    nn.Module\n):  # used for CroCo-Flow ; in the equation of the paper, we have a=1/b\n    def __init__(self, max_gtnorm=10000.0, a=0.25, b=4.0):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n        self.a, self.b = a, b\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        conf = (self.b - self.a) * torch.sigmoid(conf) + self.a\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / conf[mask]","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.LaplacianLossBounded","uri":"program://Human3R/class/src.croco.stereoflow.criterion.LaplacianLossBounded#L65-L84","kind":"class","name":"LaplacianLossBounded","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":65,"end_line":84,"context_start_line":45,"context_end_line":104,"code":"\nclass LaplacianLoss(nn.Module):  # used for CroCo-Stereo on ETH3D, d'=exp(d)\n\n    def __init__(self, max_gtnorm=None):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / torch.exp(conf[mask])\n            + conf[mask]\n        ).mean()  # + torch.log(2) => which is a constant\n\n\nclass LaplacianLossBounded(\n    nn.Module\n):  # used for CroCo-Flow ; in the equation of the paper, we have a=1/b\n    def __init__(self, max_gtnorm=10000.0, a=0.25, b=4.0):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n        self.a, self.b = a, b\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        conf = (self.b - self.a) * torch.sigmoid(conf) + self.a\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / conf[mask]\n            + torch.log(conf)[mask]\n        ).mean()  # + torch.log(2) => which is a constant\n\n\nclass LaplacianLossBounded2(\n    nn.Module\n):  # used for CroCo-Stereo (except for ETH3D) ; in the equation of the paper, we have a=b\n    def __init__(self, max_gtnorm=None, a=3.0, b=3.0):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n        self.a, self.b = a, b\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        conf = 2 * self.a * (torch.sigmoid(conf / self.b) - 0.5)\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / torch.exp(conf[mask])","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.LaplacianLossBounded2","uri":"program://Human3R/class/src.croco.stereoflow.criterion.LaplacianLossBounded2#L87-L106","kind":"class","name":"LaplacianLossBounded2","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":87,"end_line":106,"context_start_line":67,"context_end_line":126,"code":"):  # used for CroCo-Flow ; in the equation of the paper, we have a=1/b\n    def __init__(self, max_gtnorm=10000.0, a=0.25, b=4.0):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n        self.a, self.b = a, b\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        conf = (self.b - self.a) * torch.sigmoid(conf) + self.a\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / conf[mask]\n            + torch.log(conf)[mask]\n        ).mean()  # + torch.log(2) => which is a constant\n\n\nclass LaplacianLossBounded2(\n    nn.Module\n):  # used for CroCo-Stereo (except for ETH3D) ; in the equation of the paper, we have a=b\n    def __init__(self, max_gtnorm=None, a=3.0, b=3.0):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n        self.a, self.b = a, b\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        conf = 2 * self.a * (torch.sigmoid(conf / self.b) - 0.5)\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / torch.exp(conf[mask])\n            + conf[mask]\n        ).mean()  # + torch.log(2) => which is a constant\n\n\n############## metrics per batch\n\n\nclass StereoMetrics(nn.Module):\n\n    def __init__(self, do_quantile=False):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 2, 3]\n        self.do_quantile = do_quantile\n\n    def forward(self, predictions, gt):\n        B = predictions.size(0)\n        metrics = {}\n        gtcopy = gt.clone()\n        mask = torch.isfinite(gtcopy)\n        gtcopy[~mask] = (\n            999999.0  # we make a copy and put a non-infinite value, such that it does not become nan once multiplied by the mask value 0\n        )","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.StereoMetrics","uri":"program://Human3R/class/src.croco.stereoflow.criterion.StereoMetrics#L112-L139","kind":"class","name":"StereoMetrics","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":112,"end_line":139,"context_start_line":92,"context_end_line":159,"code":"        self.max_gtnorm = max_gtnorm\n        self.with_conf = True\n        self.a, self.b = a, b\n\n    def forward(self, predictions, gt, conf):\n        mask = torch.isfinite(gt)\n        mask = mask[:, 0, :, :]\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm\n        conf = conf.squeeze(1)\n        conf = 2 * self.a * (torch.sigmoid(conf / self.b) - 0.5)\n        return (\n            torch.abs(gt - predictions).sum(dim=1)[mask] / torch.exp(conf[mask])\n            + conf[mask]\n        ).mean()  # + torch.log(2) => which is a constant\n\n\n############## metrics per batch\n\n\nclass StereoMetrics(nn.Module):\n\n    def __init__(self, do_quantile=False):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 2, 3]\n        self.do_quantile = do_quantile\n\n    def forward(self, predictions, gt):\n        B = predictions.size(0)\n        metrics = {}\n        gtcopy = gt.clone()\n        mask = torch.isfinite(gtcopy)\n        gtcopy[~mask] = (\n            999999.0  # we make a copy and put a non-infinite value, such that it does not become nan once multiplied by the mask value 0\n        )\n        Npx = mask.view(B, -1).sum(dim=1)\n        L1error = (torch.abs(gtcopy - predictions) * mask).view(B, -1)\n        L2error = (torch.square(gtcopy - predictions) * mask).view(B, -1)\n        # avgerr\n        metrics[\"avgerr\"] = torch.mean(L1error.sum(dim=1) / Npx)\n        # rmse\n        metrics[\"rmse\"] = torch.sqrt(L2error.sum(dim=1) / Npx).mean(dim=0)\n        # err > t for t in [0.5,1,2,3]\n        for ths in self.bad_ths:\n            metrics[\"bad@{:.1f}\".format(ths)] = (\n                ((L1error > ths) * mask.view(B, -1)).sum(dim=1) / Npx\n            ).mean(dim=0) * 100\n        return metrics\n\n\nclass FlowMetrics(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [1, 3, 5]\n\n    def forward(self, predictions, gt):\n        B = predictions.size(0)\n        metrics = {}\n        mask = torch.isfinite(gt[:, 0, :, :])  # both x and y would be infinite\n        Npx = mask.view(B, -1).sum(dim=1)\n        gtcopy = (\n            gt.clone()\n        )  # to compute L1/L2 error, we need to have non-infinite value, the error computed at this locations will be ignored\n        gtcopy[:, 0, :, :][~mask] = 999999.0\n        gtcopy[:, 1, :, :][~mask] = 999999.0\n        L1error = (torch.abs(gtcopy - predictions).sum(dim=1) * mask).view(B, -1)\n        L2error = (\n            torch.sqrt(torch.sum(torch.square(gtcopy - predictions), dim=1)) * mask","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.FlowMetrics","uri":"program://Human3R/class/src.croco.stereoflow.criterion.FlowMetrics#L142-L167","kind":"class","name":"FlowMetrics","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":142,"end_line":167,"context_start_line":122,"context_end_line":187,"code":"        gtcopy = gt.clone()\n        mask = torch.isfinite(gtcopy)\n        gtcopy[~mask] = (\n            999999.0  # we make a copy and put a non-infinite value, such that it does not become nan once multiplied by the mask value 0\n        )\n        Npx = mask.view(B, -1).sum(dim=1)\n        L1error = (torch.abs(gtcopy - predictions) * mask).view(B, -1)\n        L2error = (torch.square(gtcopy - predictions) * mask).view(B, -1)\n        # avgerr\n        metrics[\"avgerr\"] = torch.mean(L1error.sum(dim=1) / Npx)\n        # rmse\n        metrics[\"rmse\"] = torch.sqrt(L2error.sum(dim=1) / Npx).mean(dim=0)\n        # err > t for t in [0.5,1,2,3]\n        for ths in self.bad_ths:\n            metrics[\"bad@{:.1f}\".format(ths)] = (\n                ((L1error > ths) * mask.view(B, -1)).sum(dim=1) / Npx\n            ).mean(dim=0) * 100\n        return metrics\n\n\nclass FlowMetrics(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [1, 3, 5]\n\n    def forward(self, predictions, gt):\n        B = predictions.size(0)\n        metrics = {}\n        mask = torch.isfinite(gt[:, 0, :, :])  # both x and y would be infinite\n        Npx = mask.view(B, -1).sum(dim=1)\n        gtcopy = (\n            gt.clone()\n        )  # to compute L1/L2 error, we need to have non-infinite value, the error computed at this locations will be ignored\n        gtcopy[:, 0, :, :][~mask] = 999999.0\n        gtcopy[:, 1, :, :][~mask] = 999999.0\n        L1error = (torch.abs(gtcopy - predictions).sum(dim=1) * mask).view(B, -1)\n        L2error = (\n            torch.sqrt(torch.sum(torch.square(gtcopy - predictions), dim=1)) * mask\n        ).view(B, -1)\n        metrics[\"L1err\"] = torch.mean(L1error.sum(dim=1) / Npx)\n        metrics[\"EPE\"] = torch.mean(L2error.sum(dim=1) / Npx)\n        for ths in self.bad_ths:\n            metrics[\"bad@{:.1f}\".format(ths)] = (\n                ((L2error > ths) * mask.view(B, -1)).sum(dim=1) / Npx\n            ).mean(dim=0) * 100\n        return metrics\n\n\n############## metrics per dataset\n## we update the average and maintain the number of pixels while adding data batch per batch\n## at the beggining, call reset()\n## after each batch, call add_batch(...)\n## at the end: call get_results()\n\n\nclass StereoDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 2, 3]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self._metrics = None","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.StereoDatasetMetrics","uri":"program://Human3R/class/src.croco.stereoflow.criterion.StereoDatasetMetrics#L177-L233","kind":"class","name":"StereoDatasetMetrics","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":177,"end_line":233,"context_start_line":157,"context_end_line":253,"code":"        L1error = (torch.abs(gtcopy - predictions).sum(dim=1) * mask).view(B, -1)\n        L2error = (\n            torch.sqrt(torch.sum(torch.square(gtcopy - predictions), dim=1)) * mask\n        ).view(B, -1)\n        metrics[\"L1err\"] = torch.mean(L1error.sum(dim=1) / Npx)\n        metrics[\"EPE\"] = torch.mean(L2error.sum(dim=1) / Npx)\n        for ths in self.bad_ths:\n            metrics[\"bad@{:.1f}\".format(ths)] = (\n                ((L2error > ths) * mask.view(B, -1)).sum(dim=1) / Npx\n            ).mean(dim=0) * 100\n        return metrics\n\n\n############## metrics per dataset\n## we update the average and maintain the number of pixels while adding data batch per batch\n## at the beggining, call reset()\n## after each batch, call add_batch(...)\n## at the end: call get_results()\n\n\nclass StereoDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 2, 3]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self._metrics = None\n\n    def add_batch(self, predictions, gt):\n        assert predictions.size(1) == 1, predictions.size()\n        assert gt.size(1) == 1, gt.size()\n        if (\n            gt.size(2) == predictions.size(2) * 2\n            and gt.size(3) == predictions.size(3) * 2\n        ):  # special case for Spring ...\n            L1err = torch.minimum(\n                torch.minimum(\n                    torch.minimum(\n                        torch.sum(torch.abs(gt[:, :, 0::2, 0::2] - predictions), dim=1),\n                        torch.sum(torch.abs(gt[:, :, 1::2, 0::2] - predictions), dim=1),\n                    ),\n                    torch.sum(torch.abs(gt[:, :, 0::2, 1::2] - predictions), dim=1),\n                ),\n                torch.sum(torch.abs(gt[:, :, 1::2, 1::2] - predictions), dim=1),\n            )\n            valid = torch.isfinite(L1err)\n        else:\n            valid = torch.isfinite(gt[:, 0, :, :])  # both x and y would be infinite\n            L1err = torch.sum(torch.abs(gt - predictions), dim=1)\n        N = valid.sum()\n        Nnew = self.agg_N + N\n        self.agg_L1err = (\n            float(self.agg_N) / Nnew * self.agg_L1err\n            + L1err[valid].mean().cpu() * float(N) / Nnew\n        )\n        self.agg_N = Nnew\n        for i, th in enumerate(self.bad_ths):\n            self.agg_Nbad[i] += (L1err[valid] > th).sum().cpu()\n\n    def _compute_metrics(self):\n        if self._metrics is not None:\n            return\n        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics\n\n\nclass FlowDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 3, 5]\n        self.speed_ths = [(0, 10), (10, 40), (40, torch.inf)]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_L2err = torch.tensor(0.0)  # L2 (=EPE) error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self.agg_EPEspeed = [\n            torch.tensor(0.0) for _ in self.speed_ths\n        ]  # EPE per speed bin so far\n        self.agg_Nspeed = [0 for _ in self.speed_ths]  # N pixels per speed bin so far\n        self._metrics = None\n        self.pairname_results = {}","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.FlowDatasetMetrics","uri":"program://Human3R/class/src.croco.stereoflow.criterion.FlowDatasetMetrics#L236-L351","kind":"class","name":"FlowDatasetMetrics","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":236,"end_line":351,"context_start_line":216,"context_end_line":351,"code":"        self.agg_N = Nnew\n        for i, th in enumerate(self.bad_ths):\n            self.agg_Nbad[i] += (L1err[valid] > th).sum().cpu()\n\n    def _compute_metrics(self):\n        if self._metrics is not None:\n            return\n        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics\n\n\nclass FlowDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 3, 5]\n        self.speed_ths = [(0, 10), (10, 40), (40, torch.inf)]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_L2err = torch.tensor(0.0)  # L2 (=EPE) error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self.agg_EPEspeed = [\n            torch.tensor(0.0) for _ in self.speed_ths\n        ]  # EPE per speed bin so far\n        self.agg_Nspeed = [0 for _ in self.speed_ths]  # N pixels per speed bin so far\n        self._metrics = None\n        self.pairname_results = {}\n\n    def add_batch(self, predictions, gt):\n        assert predictions.size(1) == 2, predictions.size()\n        assert gt.size(1) == 2, gt.size()\n        if (\n            gt.size(2) == predictions.size(2) * 2\n            and gt.size(3) == predictions.size(3) * 2\n        ):  # special case for Spring ...\n            L1err = torch.minimum(\n                torch.minimum(\n                    torch.minimum(\n                        torch.sum(torch.abs(gt[:, :, 0::2, 0::2] - predictions), dim=1),\n                        torch.sum(torch.abs(gt[:, :, 1::2, 0::2] - predictions), dim=1),\n                    ),\n                    torch.sum(torch.abs(gt[:, :, 0::2, 1::2] - predictions), dim=1),\n                ),\n                torch.sum(torch.abs(gt[:, :, 1::2, 1::2] - predictions), dim=1),\n            )\n            L2err = torch.minimum(\n                torch.minimum(\n                    torch.minimum(\n                        torch.sqrt(\n                            torch.sum(\n                                torch.square(gt[:, :, 0::2, 0::2] - predictions), dim=1\n                            )\n                        ),\n                        torch.sqrt(\n                            torch.sum(\n                                torch.square(gt[:, :, 1::2, 0::2] - predictions), dim=1\n                            )\n                        ),\n                    ),\n                    torch.sqrt(\n                        torch.sum(\n                            torch.square(gt[:, :, 0::2, 1::2] - predictions), dim=1\n                        )\n                    ),\n                ),\n                torch.sqrt(\n                    torch.sum(torch.square(gt[:, :, 1::2, 1::2] - predictions), dim=1)\n                ),\n            )\n            valid = torch.isfinite(L1err)\n            gtspeed = (\n                torch.sqrt(torch.sum(torch.square(gt[:, :, 0::2, 0::2]), dim=1))\n                + torch.sqrt(torch.sum(torch.square(gt[:, :, 0::2, 1::2]), dim=1))\n                + torch.sqrt(torch.sum(torch.square(gt[:, :, 1::2, 0::2]), dim=1))\n                + torch.sqrt(torch.sum(torch.square(gt[:, :, 1::2, 1::2]), dim=1))\n            ) / 4.0  # let's just average them\n        else:\n            valid = torch.isfinite(gt[:, 0, :, :])  # both x and y would be infinite\n            L1err = torch.sum(torch.abs(gt - predictions), dim=1)\n            L2err = torch.sqrt(torch.sum(torch.square(gt - predictions), dim=1))\n            gtspeed = torch.sqrt(torch.sum(torch.square(gt), dim=1))\n        N = valid.sum()\n        Nnew = self.agg_N + N\n        self.agg_L1err = (\n            float(self.agg_N) / Nnew * self.agg_L1err\n            + L1err[valid].mean().cpu() * float(N) / Nnew\n        )\n        self.agg_L2err = (\n            float(self.agg_N) / Nnew * self.agg_L2err\n            + L2err[valid].mean().cpu() * float(N) / Nnew\n        )\n        self.agg_N = Nnew\n        for i, th in enumerate(self.bad_ths):\n            self.agg_Nbad[i] += (L2err[valid] > th).sum().cpu()\n        for i, (th1, th2) in enumerate(self.speed_ths):\n            vv = (gtspeed[valid] >= th1) * (gtspeed[valid] < th2)\n            iNspeed = vv.sum()\n            if iNspeed == 0:\n                continue\n            iNnew = self.agg_Nspeed[i] + iNspeed\n            self.agg_EPEspeed[i] = (\n                float(self.agg_Nspeed[i]) / iNnew * self.agg_EPEspeed[i]\n                + float(iNspeed) / iNnew * L2err[valid][vv].mean().cpu()\n            )\n            self.agg_Nspeed[i] = iNnew\n\n    def _compute_metrics(self):\n        if self._metrics is not None:\n            return\n        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        out[\"EPE\"] = self.agg_L2err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        for i, (th1, th2) in enumerate(self.speed_ths):\n            out[\"s{:d}{:s}\".format(th1, \"-\" + str(th2) if th2 < torch.inf else \"+\")] = (\n                self.agg_EPEspeed[i].item()\n            )\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.__init__","uri":"program://Human3R/function/src.croco.stereoflow.criterion.__init__#L238-L241","kind":"function","name":"__init__","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":238,"end_line":241,"context_start_line":218,"context_end_line":261,"code":"            self.agg_Nbad[i] += (L1err[valid] > th).sum().cpu()\n\n    def _compute_metrics(self):\n        if self._metrics is not None:\n            return\n        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics\n\n\nclass FlowDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 3, 5]\n        self.speed_ths = [(0, 10), (10, 40), (40, torch.inf)]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_L2err = torch.tensor(0.0)  # L2 (=EPE) error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self.agg_EPEspeed = [\n            torch.tensor(0.0) for _ in self.speed_ths\n        ]  # EPE per speed bin so far\n        self.agg_Nspeed = [0 for _ in self.speed_ths]  # N pixels per speed bin so far\n        self._metrics = None\n        self.pairname_results = {}\n\n    def add_batch(self, predictions, gt):\n        assert predictions.size(1) == 2, predictions.size()\n        assert gt.size(1) == 2, gt.size()\n        if (\n            gt.size(2) == predictions.size(2) * 2\n            and gt.size(3) == predictions.size(3) * 2\n        ):  # special case for Spring ...","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion._error","uri":"program://Human3R/function/src.croco.stereoflow.criterion._error#L30-L31","kind":"function","name":"_error","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":30,"end_line":31,"context_start_line":10,"context_end_line":51,"code":"import torch.nn.functional as F\n\n\ndef _get_gtnorm(gt):\n    if gt.size(1) == 1:  # stereo\n        return gt\n    # flow\n    return torch.sqrt(torch.sum(gt**2, dim=1, keepdims=True))  # Bx1xHxW\n\n\n############ losses without confidence\n\n\nclass L1Loss(nn.Module):\n\n    def __init__(self, max_gtnorm=None):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = False\n\n    def _error(self, gt, predictions):\n        return torch.abs(gt - predictions)\n\n    def forward(self, predictions, gt, inspect=False):\n        mask = torch.isfinite(gt)\n        if self.max_gtnorm is not None:\n            mask *= _get_gtnorm(gt).expand(-1, gt.size(1), -1, -1) < self.max_gtnorm\n        if inspect:\n            return self._error(gt, predictions)\n        return self._error(gt[mask], predictions[mask]).mean()\n\n\n############## losses with confience\n## there are several parametrizations\n\n\nclass LaplacianLoss(nn.Module):  # used for CroCo-Stereo on ETH3D, d'=exp(d)\n\n    def __init__(self, max_gtnorm=None):\n        super().__init__()\n        self.max_gtnorm = max_gtnorm\n        self.with_conf = True","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.forward","uri":"program://Human3R/function/src.croco.stereoflow.criterion.forward#L147-L167","kind":"function","name":"forward","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":147,"end_line":167,"context_start_line":127,"context_end_line":187,"code":"        Npx = mask.view(B, -1).sum(dim=1)\n        L1error = (torch.abs(gtcopy - predictions) * mask).view(B, -1)\n        L2error = (torch.square(gtcopy - predictions) * mask).view(B, -1)\n        # avgerr\n        metrics[\"avgerr\"] = torch.mean(L1error.sum(dim=1) / Npx)\n        # rmse\n        metrics[\"rmse\"] = torch.sqrt(L2error.sum(dim=1) / Npx).mean(dim=0)\n        # err > t for t in [0.5,1,2,3]\n        for ths in self.bad_ths:\n            metrics[\"bad@{:.1f}\".format(ths)] = (\n                ((L1error > ths) * mask.view(B, -1)).sum(dim=1) / Npx\n            ).mean(dim=0) * 100\n        return metrics\n\n\nclass FlowMetrics(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [1, 3, 5]\n\n    def forward(self, predictions, gt):\n        B = predictions.size(0)\n        metrics = {}\n        mask = torch.isfinite(gt[:, 0, :, :])  # both x and y would be infinite\n        Npx = mask.view(B, -1).sum(dim=1)\n        gtcopy = (\n            gt.clone()\n        )  # to compute L1/L2 error, we need to have non-infinite value, the error computed at this locations will be ignored\n        gtcopy[:, 0, :, :][~mask] = 999999.0\n        gtcopy[:, 1, :, :][~mask] = 999999.0\n        L1error = (torch.abs(gtcopy - predictions).sum(dim=1) * mask).view(B, -1)\n        L2error = (\n            torch.sqrt(torch.sum(torch.square(gtcopy - predictions), dim=1)) * mask\n        ).view(B, -1)\n        metrics[\"L1err\"] = torch.mean(L1error.sum(dim=1) / Npx)\n        metrics[\"EPE\"] = torch.mean(L2error.sum(dim=1) / Npx)\n        for ths in self.bad_ths:\n            metrics[\"bad@{:.1f}\".format(ths)] = (\n                ((L2error > ths) * mask.view(B, -1)).sum(dim=1) / Npx\n            ).mean(dim=0) * 100\n        return metrics\n\n\n############## metrics per dataset\n## we update the average and maintain the number of pixels while adding data batch per batch\n## at the beggining, call reset()\n## after each batch, call add_batch(...)\n## at the end: call get_results()\n\n\nclass StereoDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 2, 3]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self._metrics = None","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.reset","uri":"program://Human3R/function/src.croco.stereoflow.criterion.reset#L243-L253","kind":"function","name":"reset","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":243,"end_line":253,"context_start_line":223,"context_end_line":273,"code":"        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics\n\n\nclass FlowDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 3, 5]\n        self.speed_ths = [(0, 10), (10, 40), (40, torch.inf)]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_L2err = torch.tensor(0.0)  # L2 (=EPE) error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self.agg_EPEspeed = [\n            torch.tensor(0.0) for _ in self.speed_ths\n        ]  # EPE per speed bin so far\n        self.agg_Nspeed = [0 for _ in self.speed_ths]  # N pixels per speed bin so far\n        self._metrics = None\n        self.pairname_results = {}\n\n    def add_batch(self, predictions, gt):\n        assert predictions.size(1) == 2, predictions.size()\n        assert gt.size(1) == 2, gt.size()\n        if (\n            gt.size(2) == predictions.size(2) * 2\n            and gt.size(3) == predictions.size(3) * 2\n        ):  # special case for Spring ...\n            L1err = torch.minimum(\n                torch.minimum(\n                    torch.minimum(\n                        torch.sum(torch.abs(gt[:, :, 0::2, 0::2] - predictions), dim=1),\n                        torch.sum(torch.abs(gt[:, :, 1::2, 0::2] - predictions), dim=1),\n                    ),\n                    torch.sum(torch.abs(gt[:, :, 0::2, 1::2] - predictions), dim=1),\n                ),\n                torch.sum(torch.abs(gt[:, :, 1::2, 1::2] - predictions), dim=1),\n            )\n            L2err = torch.minimum(\n                torch.minimum(","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.add_batch","uri":"program://Human3R/function/src.croco.stereoflow.criterion.add_batch#L255-L331","kind":"function","name":"add_batch","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":255,"end_line":331,"context_start_line":235,"context_end_line":351,"code":"\nclass FlowDatasetMetrics(nn.Module):\n\n    def __init__(self):\n        super().__init__()\n        self.bad_ths = [0.5, 1, 3, 5]\n        self.speed_ths = [(0, 10), (10, 40), (40, torch.inf)]\n\n    def reset(self):\n        self.agg_N = 0  # number of pixels so far\n        self.agg_L1err = torch.tensor(0.0)  # L1 error so far\n        self.agg_L2err = torch.tensor(0.0)  # L2 (=EPE) error so far\n        self.agg_Nbad = [0 for _ in self.bad_ths]  # counter of bad pixels\n        self.agg_EPEspeed = [\n            torch.tensor(0.0) for _ in self.speed_ths\n        ]  # EPE per speed bin so far\n        self.agg_Nspeed = [0 for _ in self.speed_ths]  # N pixels per speed bin so far\n        self._metrics = None\n        self.pairname_results = {}\n\n    def add_batch(self, predictions, gt):\n        assert predictions.size(1) == 2, predictions.size()\n        assert gt.size(1) == 2, gt.size()\n        if (\n            gt.size(2) == predictions.size(2) * 2\n            and gt.size(3) == predictions.size(3) * 2\n        ):  # special case for Spring ...\n            L1err = torch.minimum(\n                torch.minimum(\n                    torch.minimum(\n                        torch.sum(torch.abs(gt[:, :, 0::2, 0::2] - predictions), dim=1),\n                        torch.sum(torch.abs(gt[:, :, 1::2, 0::2] - predictions), dim=1),\n                    ),\n                    torch.sum(torch.abs(gt[:, :, 0::2, 1::2] - predictions), dim=1),\n                ),\n                torch.sum(torch.abs(gt[:, :, 1::2, 1::2] - predictions), dim=1),\n            )\n            L2err = torch.minimum(\n                torch.minimum(\n                    torch.minimum(\n                        torch.sqrt(\n                            torch.sum(\n                                torch.square(gt[:, :, 0::2, 0::2] - predictions), dim=1\n                            )\n                        ),\n                        torch.sqrt(\n                            torch.sum(\n                                torch.square(gt[:, :, 1::2, 0::2] - predictions), dim=1\n                            )\n                        ),\n                    ),\n                    torch.sqrt(\n                        torch.sum(\n                            torch.square(gt[:, :, 0::2, 1::2] - predictions), dim=1\n                        )\n                    ),\n                ),\n                torch.sqrt(\n                    torch.sum(torch.square(gt[:, :, 1::2, 1::2] - predictions), dim=1)\n                ),\n            )\n            valid = torch.isfinite(L1err)\n            gtspeed = (\n                torch.sqrt(torch.sum(torch.square(gt[:, :, 0::2, 0::2]), dim=1))\n                + torch.sqrt(torch.sum(torch.square(gt[:, :, 0::2, 1::2]), dim=1))\n                + torch.sqrt(torch.sum(torch.square(gt[:, :, 1::2, 0::2]), dim=1))\n                + torch.sqrt(torch.sum(torch.square(gt[:, :, 1::2, 1::2]), dim=1))\n            ) / 4.0  # let's just average them\n        else:\n            valid = torch.isfinite(gt[:, 0, :, :])  # both x and y would be infinite\n            L1err = torch.sum(torch.abs(gt - predictions), dim=1)\n            L2err = torch.sqrt(torch.sum(torch.square(gt - predictions), dim=1))\n            gtspeed = torch.sqrt(torch.sum(torch.square(gt), dim=1))\n        N = valid.sum()\n        Nnew = self.agg_N + N\n        self.agg_L1err = (\n            float(self.agg_N) / Nnew * self.agg_L1err\n            + L1err[valid].mean().cpu() * float(N) / Nnew\n        )\n        self.agg_L2err = (\n            float(self.agg_N) / Nnew * self.agg_L2err\n            + L2err[valid].mean().cpu() * float(N) / Nnew\n        )\n        self.agg_N = Nnew\n        for i, th in enumerate(self.bad_ths):\n            self.agg_Nbad[i] += (L2err[valid] > th).sum().cpu()\n        for i, (th1, th2) in enumerate(self.speed_ths):\n            vv = (gtspeed[valid] >= th1) * (gtspeed[valid] < th2)\n            iNspeed = vv.sum()\n            if iNspeed == 0:\n                continue\n            iNnew = self.agg_Nspeed[i] + iNspeed\n            self.agg_EPEspeed[i] = (\n                float(self.agg_Nspeed[i]) / iNnew * self.agg_EPEspeed[i]\n                + float(iNspeed) / iNnew * L2err[valid][vv].mean().cpu()\n            )\n            self.agg_Nspeed[i] = iNnew\n\n    def _compute_metrics(self):\n        if self._metrics is not None:\n            return\n        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        out[\"EPE\"] = self.agg_L2err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        for i, (th1, th2) in enumerate(self.speed_ths):\n            out[\"s{:d}{:s}\".format(th1, \"-\" + str(th2) if th2 < torch.inf else \"+\")] = (\n                self.agg_EPEspeed[i].item()\n            )\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion._compute_metrics","uri":"program://Human3R/function/src.croco.stereoflow.criterion._compute_metrics#L333-L347","kind":"function","name":"_compute_metrics","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":333,"end_line":347,"context_start_line":313,"context_end_line":351,"code":"        )\n        self.agg_L2err = (\n            float(self.agg_N) / Nnew * self.agg_L2err\n            + L2err[valid].mean().cpu() * float(N) / Nnew\n        )\n        self.agg_N = Nnew\n        for i, th in enumerate(self.bad_ths):\n            self.agg_Nbad[i] += (L2err[valid] > th).sum().cpu()\n        for i, (th1, th2) in enumerate(self.speed_ths):\n            vv = (gtspeed[valid] >= th1) * (gtspeed[valid] < th2)\n            iNspeed = vv.sum()\n            if iNspeed == 0:\n                continue\n            iNnew = self.agg_Nspeed[i] + iNspeed\n            self.agg_EPEspeed[i] = (\n                float(self.agg_Nspeed[i]) / iNnew * self.agg_EPEspeed[i]\n                + float(iNspeed) / iNnew * L2err[valid][vv].mean().cpu()\n            )\n            self.agg_Nspeed[i] = iNnew\n\n    def _compute_metrics(self):\n        if self._metrics is not None:\n            return\n        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        out[\"EPE\"] = self.agg_L2err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        for i, (th1, th2) in enumerate(self.speed_ths):\n            out[\"s{:d}{:s}\".format(th1, \"-\" + str(th2) if th2 < torch.inf else \"+\")] = (\n                self.agg_EPEspeed[i].item()\n            )\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.criterion.get_results","uri":"program://Human3R/function/src.croco.stereoflow.criterion.get_results#L349-L351","kind":"function","name":"get_results","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":349,"end_line":351,"context_start_line":329,"context_end_line":351,"code":"                + float(iNspeed) / iNnew * L2err[valid][vv].mean().cpu()\n            )\n            self.agg_Nspeed[i] = iNnew\n\n    def _compute_metrics(self):\n        if self._metrics is not None:\n            return\n        out = {}\n        out[\"L1err\"] = self.agg_L1err.item()\n        out[\"EPE\"] = self.agg_L2err.item()\n        for i, th in enumerate(self.bad_ths):\n            out[\"bad@{:.1f}\".format(th)] = (\n                float(self.agg_Nbad[i]) / self.agg_N\n            ).item() * 100.0\n        for i, (th1, th2) in enumerate(self.speed_ths):\n            out[\"s{:d}{:s}\".format(th1, \"-\" + str(th2) if th2 < torch.inf else \"+\")] = (\n                self.agg_EPEspeed[i].item()\n            )\n        self._metrics = out\n\n    def get_results(self):\n        self._compute_metrics()  # to avoid recompute them multiple times\n        return self._metrics","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor","uri":"program://Human3R/module/src.croco.stereoflow.augmentor#L1-L396","kind":"module","name":"src.croco.stereoflow.augmentor","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":1,"end_line":396,"context_start_line":1,"context_end_line":396,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Data augmentation for training stereo and flow\n# --------------------------------------------------------\n\n# References\n# https://github.com/autonomousvision/unimatch/blob/master/dataloader/stereo/transforms.py\n# https://github.com/autonomousvision/unimatch/blob/master/dataloader/flow/transforms.py\n\n\nimport numpy as np\nimport random\nfrom PIL import Image\n\nimport cv2\n\ncv2.setNumThreads(0)\ncv2.ocl.setUseOpenCL(False)\n\nimport torch\nfrom torchvision.transforms import ColorJitter\nimport torchvision.transforms.functional as FF\n\n\nclass StereoAugmentor(object):\n\n    def __init__(\n        self,\n        crop_size,\n        scale_prob=0.5,\n        scale_xonly=True,\n        lhth=800.0,\n        lminscale=0.0,\n        lmaxscale=1.0,\n        hminscale=-0.2,\n        hmaxscale=0.4,\n        scale_interp_nearest=True,\n        rightjitterprob=0.5,\n        v_flip_prob=0.5,\n        color_aug_asym=True,\n        color_choice_prob=0.5,\n    ):\n        self.crop_size = crop_size\n        self.scale_prob = scale_prob\n        self.scale_xonly = scale_xonly\n        self.lhth = lhth\n        self.lminscale = lminscale\n        self.lmaxscale = lmaxscale\n        self.hminscale = hminscale\n        self.hmaxscale = hmaxscale\n        self.scale_interp_nearest = scale_interp_nearest\n        self.rightjitterprob = rightjitterprob\n        self.v_flip_prob = v_flip_prob\n        self.color_aug_asym = color_aug_asym\n        self.color_choice_prob = color_choice_prob\n\n    def _random_scale(self, img1, img2, disp):\n        ch, cw = self.crop_size\n        h, w = img1.shape[:2]\n        if self.scale_prob > 0.0 and np.random.rand() < self.scale_prob:\n            min_scale, max_scale = (\n                (self.lminscale, self.lmaxscale)\n                if min(h, w) < self.lhth\n                else (self.hminscale, self.hmaxscale)\n            )\n            scale_x = 2.0 ** np.random.uniform(min_scale, max_scale)\n            scale_x = np.clip(scale_x, (cw + 8) / float(w), None)\n            scale_y = 1.0\n            if not self.scale_xonly:\n                scale_y = scale_x\n                scale_y = np.clip(scale_y, (ch + 8) / float(h), None)\n            img1 = cv2.resize(\n                img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            img2 = cv2.resize(\n                img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            disp = (\n                cv2.resize(\n                    disp,\n                    None,\n                    fx=scale_x,\n                    fy=scale_y,\n                    interpolation=(\n                        cv2.INTER_LINEAR\n                        if not self.scale_interp_nearest\n                        else cv2.INTER_NEAREST\n                    ),\n                )\n                * scale_x\n            )\n        else:  # check if we need to resize to be able to crop\n            h, w = img1.shape[:2]\n            clip_scale = (cw + 8) / float(w)\n            if clip_scale > 1.0:\n                scale_x = clip_scale\n                scale_y = scale_x if not self.scale_xonly else 1.0\n                img1 = cv2.resize(\n                    img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n                )\n                img2 = cv2.resize(\n                    img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n                )\n                disp = (\n                    cv2.resize(\n                        disp,\n                        None,\n                        fx=scale_x,\n                        fy=scale_y,\n                        interpolation=(\n                            cv2.INTER_LINEAR\n                            if not self.scale_interp_nearest\n                            else cv2.INTER_NEAREST\n                        ),\n                    )\n                    * scale_x\n                )\n        return img1, img2, disp\n\n    def _random_crop(self, img1, img2, disp):\n        h, w = img1.shape[:2]\n        ch, cw = self.crop_size\n        assert ch <= h and cw <= w, (img1.shape, h, w, ch, cw)\n        offset_x = np.random.randint(w - cw + 1)\n        offset_y = np.random.randint(h - ch + 1)\n        img1 = img1[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        img2 = img2[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        disp = disp[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        return img1, img2, disp\n\n    def _random_vflip(self, img1, img2, disp):\n        # vertical flip\n        if self.v_flip_prob > 0 and np.random.rand() < self.v_flip_prob:\n            img1 = np.copy(np.flipud(img1))\n            img2 = np.copy(np.flipud(img2))\n            disp = np.copy(np.flipud(disp))\n        return img1, img2, disp\n\n    def _random_rotate_shift_right(self, img2):\n        if self.rightjitterprob > 0.0 and np.random.rand() < self.rightjitterprob:\n            angle, pixel = 0.1, 2\n            px = np.random.uniform(-pixel, pixel)\n            ag = np.random.uniform(-angle, angle)\n            image_center = (\n                np.random.uniform(0, img2.shape[0]),\n                np.random.uniform(0, img2.shape[1]),\n            )\n            rot_mat = cv2.getRotationMatrix2D(image_center, ag, 1.0)\n            img2 = cv2.warpAffine(\n                img2, rot_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n            trans_mat = np.float32([[1, 0, 0], [0, 1, px]])\n            img2 = cv2.warpAffine(\n                img2, trans_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n        return img2\n\n    def _random_color_contrast(self, img1, img2):\n        if np.random.random() < 0.5:\n            contrast_factor = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_contrast(img1, contrast_factor)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                contrast_factor = np.random.uniform(0.8, 1.2)\n            img2 = FF.adjust_contrast(img2, contrast_factor)\n        return img1, img2\n\n    def _random_color_gamma(self, img1, img2):\n        if np.random.random() < 0.5:\n            gamma = np.random.uniform(0.7, 1.5)\n            img1 = FF.adjust_gamma(img1, gamma)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                gamma = np.random.uniform(0.7, 1.5)\n            img2 = FF.adjust_gamma(img2, gamma)\n        return img1, img2\n\n    def _random_color_brightness(self, img1, img2):\n        if np.random.random() < 0.5:\n            brightness = np.random.uniform(0.5, 2.0)\n            img1 = FF.adjust_brightness(img1, brightness)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                brightness = np.random.uniform(0.5, 2.0)\n            img2 = FF.adjust_brightness(img2, brightness)\n        return img1, img2\n\n    def _random_color_hue(self, img1, img2):\n        if np.random.random() < 0.5:\n            hue = np.random.uniform(-0.1, 0.1)\n            img1 = FF.adjust_hue(img1, hue)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                hue = np.random.uniform(-0.1, 0.1)\n            img2 = FF.adjust_hue(img2, hue)\n        return img1, img2\n\n    def _random_color_saturation(self, img1, img2):\n        if np.random.random() < 0.5:\n            saturation = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_saturation(img1, saturation)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                saturation = np.random.uniform(-0.8, 1.2)\n            img2 = FF.adjust_saturation(img2, saturation)\n        return img1, img2\n\n    def _random_color(self, img1, img2):\n        trfs = [\n            self._random_color_contrast,\n            self._random_color_gamma,\n            self._random_color_brightness,\n            self._random_color_hue,\n            self._random_color_saturation,\n        ]\n        img1 = Image.fromarray(img1.astype(\"uint8\"))\n        img2 = Image.fromarray(img2.astype(\"uint8\"))\n        if np.random.random() < self.color_choice_prob:\n            # A single transform\n            t = random.choice(trfs)\n            img1, img2 = t(img1, img2)\n        else:\n            # Combination of trfs\n            # Random order\n            random.shuffle(trfs)\n            for t in trfs:\n                img1, img2 = t(img1, img2)\n        img1 = np.array(img1).astype(np.float32)\n        img2 = np.array(img2).astype(np.float32)\n        return img1, img2\n\n    def __call__(self, img1, img2, disp, dataset_name):\n        img1, img2, disp = self._random_scale(img1, img2, disp)\n        img1, img2, disp = self._random_crop(img1, img2, disp)\n        img1, img2, disp = self._random_vflip(img1, img2, disp)\n        img2 = self._random_rotate_shift_right(img2)\n        img1, img2 = self._random_color(img1, img2)\n        return img1, img2, disp\n\n\nclass FlowAugmentor:\n\n    def __init__(\n        self,\n        crop_size,\n        min_scale=-0.2,\n        max_scale=0.5,\n        spatial_aug_prob=0.8,\n        stretch_prob=0.8,\n        max_stretch=0.2,\n        h_flip_prob=0.5,\n        v_flip_prob=0.1,\n        asymmetric_color_aug_prob=0.2,\n    ):\n\n        # spatial augmentation params\n        self.crop_size = crop_size\n        self.min_scale = min_scale\n        self.max_scale = max_scale\n        self.spatial_aug_prob = spatial_aug_prob\n        self.stretch_prob = stretch_prob\n        self.max_stretch = max_stretch\n\n        # flip augmentation params\n        self.h_flip_prob = h_flip_prob\n        self.v_flip_prob = v_flip_prob\n\n        # photometric augmentation params\n        self.photo_aug = ColorJitter(\n            brightness=0.4, contrast=0.4, saturation=0.4, hue=0.5 / 3.14\n        )\n\n        self.asymmetric_color_aug_prob = asymmetric_color_aug_prob\n\n    def color_transform(self, img1, img2):\n        \"\"\"Photometric augmentation\"\"\"\n\n        # asymmetric\n        if np.random.rand() < self.asymmetric_color_aug_prob:\n            img1 = np.array(self.photo_aug(Image.fromarray(img1)), dtype=np.uint8)\n            img2 = np.array(self.photo_aug(Image.fromarray(img2)), dtype=np.uint8)\n\n        # symmetric\n        else:\n            image_stack = np.concatenate([img1, img2], axis=0)\n            image_stack = np.array(\n                self.photo_aug(Image.fromarray(image_stack)), dtype=np.uint8\n            )\n            img1, img2 = np.split(image_stack, 2, axis=0)\n\n        return img1, img2\n\n    def _resize_flow(self, flow, scale_x, scale_y, factor=1.0):\n        if np.all(np.isfinite(flow)):\n            flow = cv2.resize(\n                flow,\n                None,\n                fx=scale_x / factor,\n                fy=scale_y / factor,\n                interpolation=cv2.INTER_LINEAR,\n            )\n            flow = flow * [scale_x, scale_y]\n        else:  # sparse version\n            fx, fy = scale_x, scale_y\n            ht, wd = flow.shape[:2]\n            coords = np.meshgrid(np.arange(wd), np.arange(ht))\n            coords = np.stack(coords, axis=-1)\n\n            coords = coords.reshape(-1, 2).astype(np.float32)\n            flow = flow.reshape(-1, 2).astype(np.float32)\n            valid = np.isfinite(flow[:, 0])\n\n            coords0 = coords[valid]\n            flow0 = flow[valid]\n\n            ht1 = int(round(ht * fy / factor))\n            wd1 = int(round(wd * fx / factor))\n\n            rescale = np.expand_dims(np.array([fx, fy]), axis=0)\n            coords1 = coords0 * rescale / factor\n            flow1 = flow0 * rescale\n\n            xx = np.round(coords1[:, 0]).astype(np.int32)\n            yy = np.round(coords1[:, 1]).astype(np.int32)\n\n            v = (xx > 0) & (xx < wd1) & (yy > 0) & (yy < ht1)\n            xx = xx[v]\n            yy = yy[v]\n            flow1 = flow1[v]\n\n            flow = np.inf * np.ones(\n                [ht1, wd1, 2], dtype=np.float32\n            )  # invalid value every where, before we fill it with the correct ones\n            flow[yy, xx] = flow1\n        return flow\n\n    def spatial_transform(self, img1, img2, flow, dname):\n\n        if np.random.rand() < self.spatial_aug_prob:\n            # randomly sample scale\n            ht, wd = img1.shape[:2]\n            clip_min_scale = np.maximum(\n                (self.crop_size[0] + 8) / float(ht), (self.crop_size[1] + 8) / float(wd)\n            )\n            min_scale, max_scale = self.min_scale, self.max_scale\n            scale = 2 ** np.random.uniform(self.min_scale, self.max_scale)\n            scale_x = scale\n            scale_y = scale\n            if np.random.rand() < self.stretch_prob:\n                scale_x *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)\n                scale_y *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)\n            scale_x = np.clip(scale_x, clip_min_scale, None)\n            scale_y = np.clip(scale_y, clip_min_scale, None)\n            # rescale the images\n            img1 = cv2.resize(\n                img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            img2 = cv2.resize(\n                img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            flow = self._resize_flow(\n                flow, scale_x, scale_y, factor=2.0 if dname == \"Spring\" else 1.0\n            )\n        elif dname == \"Spring\":\n            flow = self._resize_flow(flow, 1.0, 1.0, factor=2.0)\n\n        if self.h_flip_prob > 0.0 and np.random.rand() < self.h_flip_prob:  # h-flip\n            img1 = img1[:, ::-1]\n            img2 = img2[:, ::-1]\n            flow = flow[:, ::-1] * [-1.0, 1.0]\n\n        if self.v_flip_prob > 0.0 and np.random.rand() < self.v_flip_prob:  # v-flip\n            img1 = img1[::-1, :]\n            img2 = img2[::-1, :]\n            flow = flow[::-1, :] * [1.0, -1.0]\n\n        # In case no cropping\n        if img1.shape[0] - self.crop_size[0] > 0:\n            y0 = np.random.randint(0, img1.shape[0] - self.crop_size[0])\n        else:\n            y0 = 0\n        if img1.shape[1] - self.crop_size[1] > 0:\n            x0 = np.random.randint(0, img1.shape[1] - self.crop_size[1])\n        else:\n            x0 = 0\n\n        img1 = img1[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n        img2 = img2[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n        flow = flow[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n\n        return img1, img2, flow\n\n    def __call__(self, img1, img2, flow, dname):\n        img1, img2, flow = self.spatial_transform(img1, img2, flow, dname)\n        img1, img2 = self.color_transform(img1, img2)\n        img1 = np.ascontiguousarray(img1)\n        img2 = np.ascontiguousarray(img2)\n        flow = np.ascontiguousarray(flow)\n        return img1, img2, flow","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor.StereoAugmentor","uri":"program://Human3R/class/src.croco.stereoflow.augmentor.StereoAugmentor#L27-L235","kind":"class","name":"StereoAugmentor","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":27,"end_line":235,"context_start_line":7,"context_end_line":255,"code":"\n# References\n# https://github.com/autonomousvision/unimatch/blob/master/dataloader/stereo/transforms.py\n# https://github.com/autonomousvision/unimatch/blob/master/dataloader/flow/transforms.py\n\n\nimport numpy as np\nimport random\nfrom PIL import Image\n\nimport cv2\n\ncv2.setNumThreads(0)\ncv2.ocl.setUseOpenCL(False)\n\nimport torch\nfrom torchvision.transforms import ColorJitter\nimport torchvision.transforms.functional as FF\n\n\nclass StereoAugmentor(object):\n\n    def __init__(\n        self,\n        crop_size,\n        scale_prob=0.5,\n        scale_xonly=True,\n        lhth=800.0,\n        lminscale=0.0,\n        lmaxscale=1.0,\n        hminscale=-0.2,\n        hmaxscale=0.4,\n        scale_interp_nearest=True,\n        rightjitterprob=0.5,\n        v_flip_prob=0.5,\n        color_aug_asym=True,\n        color_choice_prob=0.5,\n    ):\n        self.crop_size = crop_size\n        self.scale_prob = scale_prob\n        self.scale_xonly = scale_xonly\n        self.lhth = lhth\n        self.lminscale = lminscale\n        self.lmaxscale = lmaxscale\n        self.hminscale = hminscale\n        self.hmaxscale = hmaxscale\n        self.scale_interp_nearest = scale_interp_nearest\n        self.rightjitterprob = rightjitterprob\n        self.v_flip_prob = v_flip_prob\n        self.color_aug_asym = color_aug_asym\n        self.color_choice_prob = color_choice_prob\n\n    def _random_scale(self, img1, img2, disp):\n        ch, cw = self.crop_size\n        h, w = img1.shape[:2]\n        if self.scale_prob > 0.0 and np.random.rand() < self.scale_prob:\n            min_scale, max_scale = (\n                (self.lminscale, self.lmaxscale)\n                if min(h, w) < self.lhth\n                else (self.hminscale, self.hmaxscale)\n            )\n            scale_x = 2.0 ** np.random.uniform(min_scale, max_scale)\n            scale_x = np.clip(scale_x, (cw + 8) / float(w), None)\n            scale_y = 1.0\n            if not self.scale_xonly:\n                scale_y = scale_x\n                scale_y = np.clip(scale_y, (ch + 8) / float(h), None)\n            img1 = cv2.resize(\n                img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            img2 = cv2.resize(\n                img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            disp = (\n                cv2.resize(\n                    disp,\n                    None,\n                    fx=scale_x,\n                    fy=scale_y,\n                    interpolation=(\n                        cv2.INTER_LINEAR\n                        if not self.scale_interp_nearest\n                        else cv2.INTER_NEAREST\n                    ),\n                )\n                * scale_x\n            )\n        else:  # check if we need to resize to be able to crop\n            h, w = img1.shape[:2]\n            clip_scale = (cw + 8) / float(w)\n            if clip_scale > 1.0:\n                scale_x = clip_scale\n                scale_y = scale_x if not self.scale_xonly else 1.0\n                img1 = cv2.resize(\n                    img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n                )\n                img2 = cv2.resize(\n                    img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n                )\n                disp = (\n                    cv2.resize(\n                        disp,\n                        None,\n                        fx=scale_x,\n                        fy=scale_y,\n                        interpolation=(\n                            cv2.INTER_LINEAR\n                            if not self.scale_interp_nearest\n                            else cv2.INTER_NEAREST\n                        ),\n                    )\n                    * scale_x\n                )\n        return img1, img2, disp\n\n    def _random_crop(self, img1, img2, disp):\n        h, w = img1.shape[:2]\n        ch, cw = self.crop_size\n        assert ch <= h and cw <= w, (img1.shape, h, w, ch, cw)\n        offset_x = np.random.randint(w - cw + 1)\n        offset_y = np.random.randint(h - ch + 1)\n        img1 = img1[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        img2 = img2[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        disp = disp[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        return img1, img2, disp\n\n    def _random_vflip(self, img1, img2, disp):\n        # vertical flip\n        if self.v_flip_prob > 0 and np.random.rand() < self.v_flip_prob:\n            img1 = np.copy(np.flipud(img1))\n            img2 = np.copy(np.flipud(img2))\n            disp = np.copy(np.flipud(disp))\n        return img1, img2, disp\n\n    def _random_rotate_shift_right(self, img2):\n        if self.rightjitterprob > 0.0 and np.random.rand() < self.rightjitterprob:\n            angle, pixel = 0.1, 2\n            px = np.random.uniform(-pixel, pixel)\n            ag = np.random.uniform(-angle, angle)\n            image_center = (\n                np.random.uniform(0, img2.shape[0]),\n                np.random.uniform(0, img2.shape[1]),\n            )\n            rot_mat = cv2.getRotationMatrix2D(image_center, ag, 1.0)\n            img2 = cv2.warpAffine(\n                img2, rot_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n            trans_mat = np.float32([[1, 0, 0], [0, 1, px]])\n            img2 = cv2.warpAffine(\n                img2, trans_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n        return img2\n\n    def _random_color_contrast(self, img1, img2):\n        if np.random.random() < 0.5:\n            contrast_factor = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_contrast(img1, contrast_factor)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                contrast_factor = np.random.uniform(0.8, 1.2)\n            img2 = FF.adjust_contrast(img2, contrast_factor)\n        return img1, img2\n\n    def _random_color_gamma(self, img1, img2):\n        if np.random.random() < 0.5:\n            gamma = np.random.uniform(0.7, 1.5)\n            img1 = FF.adjust_gamma(img1, gamma)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                gamma = np.random.uniform(0.7, 1.5)\n            img2 = FF.adjust_gamma(img2, gamma)\n        return img1, img2\n\n    def _random_color_brightness(self, img1, img2):\n        if np.random.random() < 0.5:\n            brightness = np.random.uniform(0.5, 2.0)\n            img1 = FF.adjust_brightness(img1, brightness)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                brightness = np.random.uniform(0.5, 2.0)\n            img2 = FF.adjust_brightness(img2, brightness)\n        return img1, img2\n\n    def _random_color_hue(self, img1, img2):\n        if np.random.random() < 0.5:\n            hue = np.random.uniform(-0.1, 0.1)\n            img1 = FF.adjust_hue(img1, hue)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                hue = np.random.uniform(-0.1, 0.1)\n            img2 = FF.adjust_hue(img2, hue)\n        return img1, img2\n\n    def _random_color_saturation(self, img1, img2):\n        if np.random.random() < 0.5:\n            saturation = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_saturation(img1, saturation)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                saturation = np.random.uniform(-0.8, 1.2)\n            img2 = FF.adjust_saturation(img2, saturation)\n        return img1, img2\n\n    def _random_color(self, img1, img2):\n        trfs = [\n            self._random_color_contrast,\n            self._random_color_gamma,\n            self._random_color_brightness,\n            self._random_color_hue,\n            self._random_color_saturation,\n        ]\n        img1 = Image.fromarray(img1.astype(\"uint8\"))\n        img2 = Image.fromarray(img2.astype(\"uint8\"))\n        if np.random.random() < self.color_choice_prob:\n            # A single transform\n            t = random.choice(trfs)\n            img1, img2 = t(img1, img2)\n        else:\n            # Combination of trfs\n            # Random order\n            random.shuffle(trfs)\n            for t in trfs:\n                img1, img2 = t(img1, img2)\n        img1 = np.array(img1).astype(np.float32)\n        img2 = np.array(img2).astype(np.float32)\n        return img1, img2\n\n    def __call__(self, img1, img2, disp, dataset_name):\n        img1, img2, disp = self._random_scale(img1, img2, disp)\n        img1, img2, disp = self._random_crop(img1, img2, disp)\n        img1, img2, disp = self._random_vflip(img1, img2, disp)\n        img2 = self._random_rotate_shift_right(img2)\n        img1, img2 = self._random_color(img1, img2)\n        return img1, img2, disp\n\n\nclass FlowAugmentor:\n\n    def __init__(\n        self,\n        crop_size,\n        min_scale=-0.2,\n        max_scale=0.5,\n        spatial_aug_prob=0.8,\n        stretch_prob=0.8,\n        max_stretch=0.2,\n        h_flip_prob=0.5,\n        v_flip_prob=0.1,\n        asymmetric_color_aug_prob=0.2,\n    ):\n\n        # spatial augmentation params\n        self.crop_size = crop_size\n        self.min_scale = min_scale","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor.FlowAugmentor","uri":"program://Human3R/class/src.croco.stereoflow.augmentor.FlowAugmentor#L238-L396","kind":"class","name":"FlowAugmentor","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":238,"end_line":396,"context_start_line":218,"context_end_line":396,"code":"            img1, img2 = t(img1, img2)\n        else:\n            # Combination of trfs\n            # Random order\n            random.shuffle(trfs)\n            for t in trfs:\n                img1, img2 = t(img1, img2)\n        img1 = np.array(img1).astype(np.float32)\n        img2 = np.array(img2).astype(np.float32)\n        return img1, img2\n\n    def __call__(self, img1, img2, disp, dataset_name):\n        img1, img2, disp = self._random_scale(img1, img2, disp)\n        img1, img2, disp = self._random_crop(img1, img2, disp)\n        img1, img2, disp = self._random_vflip(img1, img2, disp)\n        img2 = self._random_rotate_shift_right(img2)\n        img1, img2 = self._random_color(img1, img2)\n        return img1, img2, disp\n\n\nclass FlowAugmentor:\n\n    def __init__(\n        self,\n        crop_size,\n        min_scale=-0.2,\n        max_scale=0.5,\n        spatial_aug_prob=0.8,\n        stretch_prob=0.8,\n        max_stretch=0.2,\n        h_flip_prob=0.5,\n        v_flip_prob=0.1,\n        asymmetric_color_aug_prob=0.2,\n    ):\n\n        # spatial augmentation params\n        self.crop_size = crop_size\n        self.min_scale = min_scale\n        self.max_scale = max_scale\n        self.spatial_aug_prob = spatial_aug_prob\n        self.stretch_prob = stretch_prob\n        self.max_stretch = max_stretch\n\n        # flip augmentation params\n        self.h_flip_prob = h_flip_prob\n        self.v_flip_prob = v_flip_prob\n\n        # photometric augmentation params\n        self.photo_aug = ColorJitter(\n            brightness=0.4, contrast=0.4, saturation=0.4, hue=0.5 / 3.14\n        )\n\n        self.asymmetric_color_aug_prob = asymmetric_color_aug_prob\n\n    def color_transform(self, img1, img2):\n        \"\"\"Photometric augmentation\"\"\"\n\n        # asymmetric\n        if np.random.rand() < self.asymmetric_color_aug_prob:\n            img1 = np.array(self.photo_aug(Image.fromarray(img1)), dtype=np.uint8)\n            img2 = np.array(self.photo_aug(Image.fromarray(img2)), dtype=np.uint8)\n\n        # symmetric\n        else:\n            image_stack = np.concatenate([img1, img2], axis=0)\n            image_stack = np.array(\n                self.photo_aug(Image.fromarray(image_stack)), dtype=np.uint8\n            )\n            img1, img2 = np.split(image_stack, 2, axis=0)\n\n        return img1, img2\n\n    def _resize_flow(self, flow, scale_x, scale_y, factor=1.0):\n        if np.all(np.isfinite(flow)):\n            flow = cv2.resize(\n                flow,\n                None,\n                fx=scale_x / factor,\n                fy=scale_y / factor,\n                interpolation=cv2.INTER_LINEAR,\n            )\n            flow = flow * [scale_x, scale_y]\n        else:  # sparse version\n            fx, fy = scale_x, scale_y\n            ht, wd = flow.shape[:2]\n            coords = np.meshgrid(np.arange(wd), np.arange(ht))\n            coords = np.stack(coords, axis=-1)\n\n            coords = coords.reshape(-1, 2).astype(np.float32)\n            flow = flow.reshape(-1, 2).astype(np.float32)\n            valid = np.isfinite(flow[:, 0])\n\n            coords0 = coords[valid]\n            flow0 = flow[valid]\n\n            ht1 = int(round(ht * fy / factor))\n            wd1 = int(round(wd * fx / factor))\n\n            rescale = np.expand_dims(np.array([fx, fy]), axis=0)\n            coords1 = coords0 * rescale / factor\n            flow1 = flow0 * rescale\n\n            xx = np.round(coords1[:, 0]).astype(np.int32)\n            yy = np.round(coords1[:, 1]).astype(np.int32)\n\n            v = (xx > 0) & (xx < wd1) & (yy > 0) & (yy < ht1)\n            xx = xx[v]\n            yy = yy[v]\n            flow1 = flow1[v]\n\n            flow = np.inf * np.ones(\n                [ht1, wd1, 2], dtype=np.float32\n            )  # invalid value every where, before we fill it with the correct ones\n            flow[yy, xx] = flow1\n        return flow\n\n    def spatial_transform(self, img1, img2, flow, dname):\n\n        if np.random.rand() < self.spatial_aug_prob:\n            # randomly sample scale\n            ht, wd = img1.shape[:2]\n            clip_min_scale = np.maximum(\n                (self.crop_size[0] + 8) / float(ht), (self.crop_size[1] + 8) / float(wd)\n            )\n            min_scale, max_scale = self.min_scale, self.max_scale\n            scale = 2 ** np.random.uniform(self.min_scale, self.max_scale)\n            scale_x = scale\n            scale_y = scale\n            if np.random.rand() < self.stretch_prob:\n                scale_x *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)\n                scale_y *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)\n            scale_x = np.clip(scale_x, clip_min_scale, None)\n            scale_y = np.clip(scale_y, clip_min_scale, None)\n            # rescale the images\n            img1 = cv2.resize(\n                img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            img2 = cv2.resize(\n                img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            flow = self._resize_flow(\n                flow, scale_x, scale_y, factor=2.0 if dname == \"Spring\" else 1.0\n            )\n        elif dname == \"Spring\":\n            flow = self._resize_flow(flow, 1.0, 1.0, factor=2.0)\n\n        if self.h_flip_prob > 0.0 and np.random.rand() < self.h_flip_prob:  # h-flip\n            img1 = img1[:, ::-1]\n            img2 = img2[:, ::-1]\n            flow = flow[:, ::-1] * [-1.0, 1.0]\n\n        if self.v_flip_prob > 0.0 and np.random.rand() < self.v_flip_prob:  # v-flip\n            img1 = img1[::-1, :]\n            img2 = img2[::-1, :]\n            flow = flow[::-1, :] * [1.0, -1.0]\n\n        # In case no cropping\n        if img1.shape[0] - self.crop_size[0] > 0:\n            y0 = np.random.randint(0, img1.shape[0] - self.crop_size[0])\n        else:\n            y0 = 0\n        if img1.shape[1] - self.crop_size[1] > 0:\n            x0 = np.random.randint(0, img1.shape[1] - self.crop_size[1])\n        else:\n            x0 = 0\n\n        img1 = img1[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n        img2 = img2[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n        flow = flow[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n\n        return img1, img2, flow\n\n    def __call__(self, img1, img2, flow, dname):\n        img1, img2, flow = self.spatial_transform(img1, img2, flow, dname)\n        img1, img2 = self.color_transform(img1, img2)\n        img1 = np.ascontiguousarray(img1)\n        img2 = np.ascontiguousarray(img2)\n        flow = np.ascontiguousarray(flow)\n        return img1, img2, flow","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor.__init__","uri":"program://Human3R/function/src.croco.stereoflow.augmentor.__init__#L240-L270","kind":"function","name":"__init__","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":240,"end_line":270,"context_start_line":220,"context_end_line":290,"code":"            # Combination of trfs\n            # Random order\n            random.shuffle(trfs)\n            for t in trfs:\n                img1, img2 = t(img1, img2)\n        img1 = np.array(img1).astype(np.float32)\n        img2 = np.array(img2).astype(np.float32)\n        return img1, img2\n\n    def __call__(self, img1, img2, disp, dataset_name):\n        img1, img2, disp = self._random_scale(img1, img2, disp)\n        img1, img2, disp = self._random_crop(img1, img2, disp)\n        img1, img2, disp = self._random_vflip(img1, img2, disp)\n        img2 = self._random_rotate_shift_right(img2)\n        img1, img2 = self._random_color(img1, img2)\n        return img1, img2, disp\n\n\nclass FlowAugmentor:\n\n    def __init__(\n        self,\n        crop_size,\n        min_scale=-0.2,\n        max_scale=0.5,\n        spatial_aug_prob=0.8,\n        stretch_prob=0.8,\n        max_stretch=0.2,\n        h_flip_prob=0.5,\n        v_flip_prob=0.1,\n        asymmetric_color_aug_prob=0.2,\n    ):\n\n        # spatial augmentation params\n        self.crop_size = crop_size\n        self.min_scale = min_scale\n        self.max_scale = max_scale\n        self.spatial_aug_prob = spatial_aug_prob\n        self.stretch_prob = stretch_prob\n        self.max_stretch = max_stretch\n\n        # flip augmentation params\n        self.h_flip_prob = h_flip_prob\n        self.v_flip_prob = v_flip_prob\n\n        # photometric augmentation params\n        self.photo_aug = ColorJitter(\n            brightness=0.4, contrast=0.4, saturation=0.4, hue=0.5 / 3.14\n        )\n\n        self.asymmetric_color_aug_prob = asymmetric_color_aug_prob\n\n    def color_transform(self, img1, img2):\n        \"\"\"Photometric augmentation\"\"\"\n\n        # asymmetric\n        if np.random.rand() < self.asymmetric_color_aug_prob:\n            img1 = np.array(self.photo_aug(Image.fromarray(img1)), dtype=np.uint8)\n            img2 = np.array(self.photo_aug(Image.fromarray(img2)), dtype=np.uint8)\n\n        # symmetric\n        else:\n            image_stack = np.concatenate([img1, img2], axis=0)\n            image_stack = np.array(\n                self.photo_aug(Image.fromarray(image_stack)), dtype=np.uint8\n            )\n            img1, img2 = np.split(image_stack, 2, axis=0)\n\n        return img1, img2\n\n    def _resize_flow(self, flow, scale_x, scale_y, factor=1.0):","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_scale","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_scale#L59-L120","kind":"function","name":"_random_scale","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":59,"end_line":120,"context_start_line":39,"context_end_line":140,"code":"        scale_interp_nearest=True,\n        rightjitterprob=0.5,\n        v_flip_prob=0.5,\n        color_aug_asym=True,\n        color_choice_prob=0.5,\n    ):\n        self.crop_size = crop_size\n        self.scale_prob = scale_prob\n        self.scale_xonly = scale_xonly\n        self.lhth = lhth\n        self.lminscale = lminscale\n        self.lmaxscale = lmaxscale\n        self.hminscale = hminscale\n        self.hmaxscale = hmaxscale\n        self.scale_interp_nearest = scale_interp_nearest\n        self.rightjitterprob = rightjitterprob\n        self.v_flip_prob = v_flip_prob\n        self.color_aug_asym = color_aug_asym\n        self.color_choice_prob = color_choice_prob\n\n    def _random_scale(self, img1, img2, disp):\n        ch, cw = self.crop_size\n        h, w = img1.shape[:2]\n        if self.scale_prob > 0.0 and np.random.rand() < self.scale_prob:\n            min_scale, max_scale = (\n                (self.lminscale, self.lmaxscale)\n                if min(h, w) < self.lhth\n                else (self.hminscale, self.hmaxscale)\n            )\n            scale_x = 2.0 ** np.random.uniform(min_scale, max_scale)\n            scale_x = np.clip(scale_x, (cw + 8) / float(w), None)\n            scale_y = 1.0\n            if not self.scale_xonly:\n                scale_y = scale_x\n                scale_y = np.clip(scale_y, (ch + 8) / float(h), None)\n            img1 = cv2.resize(\n                img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            img2 = cv2.resize(\n                img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            disp = (\n                cv2.resize(\n                    disp,\n                    None,\n                    fx=scale_x,\n                    fy=scale_y,\n                    interpolation=(\n                        cv2.INTER_LINEAR\n                        if not self.scale_interp_nearest\n                        else cv2.INTER_NEAREST\n                    ),\n                )\n                * scale_x\n            )\n        else:  # check if we need to resize to be able to crop\n            h, w = img1.shape[:2]\n            clip_scale = (cw + 8) / float(w)\n            if clip_scale > 1.0:\n                scale_x = clip_scale\n                scale_y = scale_x if not self.scale_xonly else 1.0\n                img1 = cv2.resize(\n                    img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n                )\n                img2 = cv2.resize(\n                    img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n                )\n                disp = (\n                    cv2.resize(\n                        disp,\n                        None,\n                        fx=scale_x,\n                        fy=scale_y,\n                        interpolation=(\n                            cv2.INTER_LINEAR\n                            if not self.scale_interp_nearest\n                            else cv2.INTER_NEAREST\n                        ),\n                    )\n                    * scale_x\n                )\n        return img1, img2, disp\n\n    def _random_crop(self, img1, img2, disp):\n        h, w = img1.shape[:2]\n        ch, cw = self.crop_size\n        assert ch <= h and cw <= w, (img1.shape, h, w, ch, cw)\n        offset_x = np.random.randint(w - cw + 1)\n        offset_y = np.random.randint(h - ch + 1)\n        img1 = img1[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        img2 = img2[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        disp = disp[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        return img1, img2, disp\n\n    def _random_vflip(self, img1, img2, disp):\n        # vertical flip\n        if self.v_flip_prob > 0 and np.random.rand() < self.v_flip_prob:\n            img1 = np.copy(np.flipud(img1))\n            img2 = np.copy(np.flipud(img2))\n            disp = np.copy(np.flipud(disp))\n        return img1, img2, disp\n","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_crop","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_crop#L122-L131","kind":"function","name":"_random_crop","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":122,"end_line":131,"context_start_line":102,"context_end_line":151,"code":"                )\n                img2 = cv2.resize(\n                    img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n                )\n                disp = (\n                    cv2.resize(\n                        disp,\n                        None,\n                        fx=scale_x,\n                        fy=scale_y,\n                        interpolation=(\n                            cv2.INTER_LINEAR\n                            if not self.scale_interp_nearest\n                            else cv2.INTER_NEAREST\n                        ),\n                    )\n                    * scale_x\n                )\n        return img1, img2, disp\n\n    def _random_crop(self, img1, img2, disp):\n        h, w = img1.shape[:2]\n        ch, cw = self.crop_size\n        assert ch <= h and cw <= w, (img1.shape, h, w, ch, cw)\n        offset_x = np.random.randint(w - cw + 1)\n        offset_y = np.random.randint(h - ch + 1)\n        img1 = img1[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        img2 = img2[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        disp = disp[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        return img1, img2, disp\n\n    def _random_vflip(self, img1, img2, disp):\n        # vertical flip\n        if self.v_flip_prob > 0 and np.random.rand() < self.v_flip_prob:\n            img1 = np.copy(np.flipud(img1))\n            img2 = np.copy(np.flipud(img2))\n            disp = np.copy(np.flipud(disp))\n        return img1, img2, disp\n\n    def _random_rotate_shift_right(self, img2):\n        if self.rightjitterprob > 0.0 and np.random.rand() < self.rightjitterprob:\n            angle, pixel = 0.1, 2\n            px = np.random.uniform(-pixel, pixel)\n            ag = np.random.uniform(-angle, angle)\n            image_center = (\n                np.random.uniform(0, img2.shape[0]),\n                np.random.uniform(0, img2.shape[1]),\n            )\n            rot_mat = cv2.getRotationMatrix2D(image_center, ag, 1.0)\n            img2 = cv2.warpAffine(","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_vflip","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_vflip#L133-L139","kind":"function","name":"_random_vflip","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":133,"end_line":139,"context_start_line":113,"context_end_line":159,"code":"                            cv2.INTER_LINEAR\n                            if not self.scale_interp_nearest\n                            else cv2.INTER_NEAREST\n                        ),\n                    )\n                    * scale_x\n                )\n        return img1, img2, disp\n\n    def _random_crop(self, img1, img2, disp):\n        h, w = img1.shape[:2]\n        ch, cw = self.crop_size\n        assert ch <= h and cw <= w, (img1.shape, h, w, ch, cw)\n        offset_x = np.random.randint(w - cw + 1)\n        offset_y = np.random.randint(h - ch + 1)\n        img1 = img1[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        img2 = img2[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        disp = disp[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        return img1, img2, disp\n\n    def _random_vflip(self, img1, img2, disp):\n        # vertical flip\n        if self.v_flip_prob > 0 and np.random.rand() < self.v_flip_prob:\n            img1 = np.copy(np.flipud(img1))\n            img2 = np.copy(np.flipud(img2))\n            disp = np.copy(np.flipud(disp))\n        return img1, img2, disp\n\n    def _random_rotate_shift_right(self, img2):\n        if self.rightjitterprob > 0.0 and np.random.rand() < self.rightjitterprob:\n            angle, pixel = 0.1, 2\n            px = np.random.uniform(-pixel, pixel)\n            ag = np.random.uniform(-angle, angle)\n            image_center = (\n                np.random.uniform(0, img2.shape[0]),\n                np.random.uniform(0, img2.shape[1]),\n            )\n            rot_mat = cv2.getRotationMatrix2D(image_center, ag, 1.0)\n            img2 = cv2.warpAffine(\n                img2, rot_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n            trans_mat = np.float32([[1, 0, 0], [0, 1, px]])\n            img2 = cv2.warpAffine(\n                img2, trans_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n        return img2\n","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_rotate_shift_right","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_rotate_shift_right#L141-L158","kind":"function","name":"_random_rotate_shift_right","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":141,"end_line":158,"context_start_line":121,"context_end_line":178,"code":"\n    def _random_crop(self, img1, img2, disp):\n        h, w = img1.shape[:2]\n        ch, cw = self.crop_size\n        assert ch <= h and cw <= w, (img1.shape, h, w, ch, cw)\n        offset_x = np.random.randint(w - cw + 1)\n        offset_y = np.random.randint(h - ch + 1)\n        img1 = img1[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        img2 = img2[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        disp = disp[offset_y : offset_y + ch, offset_x : offset_x + cw]\n        return img1, img2, disp\n\n    def _random_vflip(self, img1, img2, disp):\n        # vertical flip\n        if self.v_flip_prob > 0 and np.random.rand() < self.v_flip_prob:\n            img1 = np.copy(np.flipud(img1))\n            img2 = np.copy(np.flipud(img2))\n            disp = np.copy(np.flipud(disp))\n        return img1, img2, disp\n\n    def _random_rotate_shift_right(self, img2):\n        if self.rightjitterprob > 0.0 and np.random.rand() < self.rightjitterprob:\n            angle, pixel = 0.1, 2\n            px = np.random.uniform(-pixel, pixel)\n            ag = np.random.uniform(-angle, angle)\n            image_center = (\n                np.random.uniform(0, img2.shape[0]),\n                np.random.uniform(0, img2.shape[1]),\n            )\n            rot_mat = cv2.getRotationMatrix2D(image_center, ag, 1.0)\n            img2 = cv2.warpAffine(\n                img2, rot_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n            trans_mat = np.float32([[1, 0, 0], [0, 1, px]])\n            img2 = cv2.warpAffine(\n                img2, trans_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n        return img2\n\n    def _random_color_contrast(self, img1, img2):\n        if np.random.random() < 0.5:\n            contrast_factor = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_contrast(img1, contrast_factor)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                contrast_factor = np.random.uniform(0.8, 1.2)\n            img2 = FF.adjust_contrast(img2, contrast_factor)\n        return img1, img2\n\n    def _random_color_gamma(self, img1, img2):\n        if np.random.random() < 0.5:\n            gamma = np.random.uniform(0.7, 1.5)\n            img1 = FF.adjust_gamma(img1, gamma)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                gamma = np.random.uniform(0.7, 1.5)\n            img2 = FF.adjust_gamma(img2, gamma)\n        return img1, img2\n\n    def _random_color_brightness(self, img1, img2):","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_color_contrast","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_color_contrast#L160-L167","kind":"function","name":"_random_color_contrast","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":160,"end_line":167,"context_start_line":140,"context_end_line":187,"code":"\n    def _random_rotate_shift_right(self, img2):\n        if self.rightjitterprob > 0.0 and np.random.rand() < self.rightjitterprob:\n            angle, pixel = 0.1, 2\n            px = np.random.uniform(-pixel, pixel)\n            ag = np.random.uniform(-angle, angle)\n            image_center = (\n                np.random.uniform(0, img2.shape[0]),\n                np.random.uniform(0, img2.shape[1]),\n            )\n            rot_mat = cv2.getRotationMatrix2D(image_center, ag, 1.0)\n            img2 = cv2.warpAffine(\n                img2, rot_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n            trans_mat = np.float32([[1, 0, 0], [0, 1, px]])\n            img2 = cv2.warpAffine(\n                img2, trans_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n        return img2\n\n    def _random_color_contrast(self, img1, img2):\n        if np.random.random() < 0.5:\n            contrast_factor = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_contrast(img1, contrast_factor)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                contrast_factor = np.random.uniform(0.8, 1.2)\n            img2 = FF.adjust_contrast(img2, contrast_factor)\n        return img1, img2\n\n    def _random_color_gamma(self, img1, img2):\n        if np.random.random() < 0.5:\n            gamma = np.random.uniform(0.7, 1.5)\n            img1 = FF.adjust_gamma(img1, gamma)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                gamma = np.random.uniform(0.7, 1.5)\n            img2 = FF.adjust_gamma(img2, gamma)\n        return img1, img2\n\n    def _random_color_brightness(self, img1, img2):\n        if np.random.random() < 0.5:\n            brightness = np.random.uniform(0.5, 2.0)\n            img1 = FF.adjust_brightness(img1, brightness)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                brightness = np.random.uniform(0.5, 2.0)\n            img2 = FF.adjust_brightness(img2, brightness)\n        return img1, img2\n\n    def _random_color_hue(self, img1, img2):","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_color_gamma","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_color_gamma#L169-L176","kind":"function","name":"_random_color_gamma","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":169,"end_line":176,"context_start_line":149,"context_end_line":196,"code":"            )\n            rot_mat = cv2.getRotationMatrix2D(image_center, ag, 1.0)\n            img2 = cv2.warpAffine(\n                img2, rot_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n            trans_mat = np.float32([[1, 0, 0], [0, 1, px]])\n            img2 = cv2.warpAffine(\n                img2, trans_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR\n            )\n        return img2\n\n    def _random_color_contrast(self, img1, img2):\n        if np.random.random() < 0.5:\n            contrast_factor = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_contrast(img1, contrast_factor)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                contrast_factor = np.random.uniform(0.8, 1.2)\n            img2 = FF.adjust_contrast(img2, contrast_factor)\n        return img1, img2\n\n    def _random_color_gamma(self, img1, img2):\n        if np.random.random() < 0.5:\n            gamma = np.random.uniform(0.7, 1.5)\n            img1 = FF.adjust_gamma(img1, gamma)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                gamma = np.random.uniform(0.7, 1.5)\n            img2 = FF.adjust_gamma(img2, gamma)\n        return img1, img2\n\n    def _random_color_brightness(self, img1, img2):\n        if np.random.random() < 0.5:\n            brightness = np.random.uniform(0.5, 2.0)\n            img1 = FF.adjust_brightness(img1, brightness)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                brightness = np.random.uniform(0.5, 2.0)\n            img2 = FF.adjust_brightness(img2, brightness)\n        return img1, img2\n\n    def _random_color_hue(self, img1, img2):\n        if np.random.random() < 0.5:\n            hue = np.random.uniform(-0.1, 0.1)\n            img1 = FF.adjust_hue(img1, hue)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                hue = np.random.uniform(-0.1, 0.1)\n            img2 = FF.adjust_hue(img2, hue)\n        return img1, img2\n\n    def _random_color_saturation(self, img1, img2):","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_color_brightness","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_color_brightness#L178-L185","kind":"function","name":"_random_color_brightness","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":178,"end_line":185,"context_start_line":158,"context_end_line":205,"code":"        return img2\n\n    def _random_color_contrast(self, img1, img2):\n        if np.random.random() < 0.5:\n            contrast_factor = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_contrast(img1, contrast_factor)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                contrast_factor = np.random.uniform(0.8, 1.2)\n            img2 = FF.adjust_contrast(img2, contrast_factor)\n        return img1, img2\n\n    def _random_color_gamma(self, img1, img2):\n        if np.random.random() < 0.5:\n            gamma = np.random.uniform(0.7, 1.5)\n            img1 = FF.adjust_gamma(img1, gamma)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                gamma = np.random.uniform(0.7, 1.5)\n            img2 = FF.adjust_gamma(img2, gamma)\n        return img1, img2\n\n    def _random_color_brightness(self, img1, img2):\n        if np.random.random() < 0.5:\n            brightness = np.random.uniform(0.5, 2.0)\n            img1 = FF.adjust_brightness(img1, brightness)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                brightness = np.random.uniform(0.5, 2.0)\n            img2 = FF.adjust_brightness(img2, brightness)\n        return img1, img2\n\n    def _random_color_hue(self, img1, img2):\n        if np.random.random() < 0.5:\n            hue = np.random.uniform(-0.1, 0.1)\n            img1 = FF.adjust_hue(img1, hue)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                hue = np.random.uniform(-0.1, 0.1)\n            img2 = FF.adjust_hue(img2, hue)\n        return img1, img2\n\n    def _random_color_saturation(self, img1, img2):\n        if np.random.random() < 0.5:\n            saturation = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_saturation(img1, saturation)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                saturation = np.random.uniform(-0.8, 1.2)\n            img2 = FF.adjust_saturation(img2, saturation)\n        return img1, img2\n\n    def _random_color(self, img1, img2):","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_color_hue","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_color_hue#L187-L194","kind":"function","name":"_random_color_hue","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":187,"end_line":194,"context_start_line":167,"context_end_line":214,"code":"        return img1, img2\n\n    def _random_color_gamma(self, img1, img2):\n        if np.random.random() < 0.5:\n            gamma = np.random.uniform(0.7, 1.5)\n            img1 = FF.adjust_gamma(img1, gamma)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                gamma = np.random.uniform(0.7, 1.5)\n            img2 = FF.adjust_gamma(img2, gamma)\n        return img1, img2\n\n    def _random_color_brightness(self, img1, img2):\n        if np.random.random() < 0.5:\n            brightness = np.random.uniform(0.5, 2.0)\n            img1 = FF.adjust_brightness(img1, brightness)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                brightness = np.random.uniform(0.5, 2.0)\n            img2 = FF.adjust_brightness(img2, brightness)\n        return img1, img2\n\n    def _random_color_hue(self, img1, img2):\n        if np.random.random() < 0.5:\n            hue = np.random.uniform(-0.1, 0.1)\n            img1 = FF.adjust_hue(img1, hue)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                hue = np.random.uniform(-0.1, 0.1)\n            img2 = FF.adjust_hue(img2, hue)\n        return img1, img2\n\n    def _random_color_saturation(self, img1, img2):\n        if np.random.random() < 0.5:\n            saturation = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_saturation(img1, saturation)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                saturation = np.random.uniform(-0.8, 1.2)\n            img2 = FF.adjust_saturation(img2, saturation)\n        return img1, img2\n\n    def _random_color(self, img1, img2):\n        trfs = [\n            self._random_color_contrast,\n            self._random_color_gamma,\n            self._random_color_brightness,\n            self._random_color_hue,\n            self._random_color_saturation,\n        ]\n        img1 = Image.fromarray(img1.astype(\"uint8\"))\n        img2 = Image.fromarray(img2.astype(\"uint8\"))","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_color_saturation","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_color_saturation#L196-L203","kind":"function","name":"_random_color_saturation","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":196,"end_line":203,"context_start_line":176,"context_end_line":223,"code":"        return img1, img2\n\n    def _random_color_brightness(self, img1, img2):\n        if np.random.random() < 0.5:\n            brightness = np.random.uniform(0.5, 2.0)\n            img1 = FF.adjust_brightness(img1, brightness)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                brightness = np.random.uniform(0.5, 2.0)\n            img2 = FF.adjust_brightness(img2, brightness)\n        return img1, img2\n\n    def _random_color_hue(self, img1, img2):\n        if np.random.random() < 0.5:\n            hue = np.random.uniform(-0.1, 0.1)\n            img1 = FF.adjust_hue(img1, hue)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                hue = np.random.uniform(-0.1, 0.1)\n            img2 = FF.adjust_hue(img2, hue)\n        return img1, img2\n\n    def _random_color_saturation(self, img1, img2):\n        if np.random.random() < 0.5:\n            saturation = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_saturation(img1, saturation)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                saturation = np.random.uniform(-0.8, 1.2)\n            img2 = FF.adjust_saturation(img2, saturation)\n        return img1, img2\n\n    def _random_color(self, img1, img2):\n        trfs = [\n            self._random_color_contrast,\n            self._random_color_gamma,\n            self._random_color_brightness,\n            self._random_color_hue,\n            self._random_color_saturation,\n        ]\n        img1 = Image.fromarray(img1.astype(\"uint8\"))\n        img2 = Image.fromarray(img2.astype(\"uint8\"))\n        if np.random.random() < self.color_choice_prob:\n            # A single transform\n            t = random.choice(trfs)\n            img1, img2 = t(img1, img2)\n        else:\n            # Combination of trfs\n            # Random order\n            random.shuffle(trfs)\n            for t in trfs:","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._random_color","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._random_color#L205-L227","kind":"function","name":"_random_color","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":205,"end_line":227,"context_start_line":185,"context_end_line":247,"code":"        return img1, img2\n\n    def _random_color_hue(self, img1, img2):\n        if np.random.random() < 0.5:\n            hue = np.random.uniform(-0.1, 0.1)\n            img1 = FF.adjust_hue(img1, hue)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                hue = np.random.uniform(-0.1, 0.1)\n            img2 = FF.adjust_hue(img2, hue)\n        return img1, img2\n\n    def _random_color_saturation(self, img1, img2):\n        if np.random.random() < 0.5:\n            saturation = np.random.uniform(0.8, 1.2)\n            img1 = FF.adjust_saturation(img1, saturation)\n            if self.color_aug_asym and np.random.random() < 0.5:\n                saturation = np.random.uniform(-0.8, 1.2)\n            img2 = FF.adjust_saturation(img2, saturation)\n        return img1, img2\n\n    def _random_color(self, img1, img2):\n        trfs = [\n            self._random_color_contrast,\n            self._random_color_gamma,\n            self._random_color_brightness,\n            self._random_color_hue,\n            self._random_color_saturation,\n        ]\n        img1 = Image.fromarray(img1.astype(\"uint8\"))\n        img2 = Image.fromarray(img2.astype(\"uint8\"))\n        if np.random.random() < self.color_choice_prob:\n            # A single transform\n            t = random.choice(trfs)\n            img1, img2 = t(img1, img2)\n        else:\n            # Combination of trfs\n            # Random order\n            random.shuffle(trfs)\n            for t in trfs:\n                img1, img2 = t(img1, img2)\n        img1 = np.array(img1).astype(np.float32)\n        img2 = np.array(img2).astype(np.float32)\n        return img1, img2\n\n    def __call__(self, img1, img2, disp, dataset_name):\n        img1, img2, disp = self._random_scale(img1, img2, disp)\n        img1, img2, disp = self._random_crop(img1, img2, disp)\n        img1, img2, disp = self._random_vflip(img1, img2, disp)\n        img2 = self._random_rotate_shift_right(img2)\n        img1, img2 = self._random_color(img1, img2)\n        return img1, img2, disp\n\n\nclass FlowAugmentor:\n\n    def __init__(\n        self,\n        crop_size,\n        min_scale=-0.2,\n        max_scale=0.5,\n        spatial_aug_prob=0.8,\n        stretch_prob=0.8,\n        max_stretch=0.2,","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor.__call__","uri":"program://Human3R/function/src.croco.stereoflow.augmentor.__call__#L390-L396","kind":"function","name":"__call__","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":390,"end_line":396,"context_start_line":370,"context_end_line":396,"code":"            img1 = img1[::-1, :]\n            img2 = img2[::-1, :]\n            flow = flow[::-1, :] * [1.0, -1.0]\n\n        # In case no cropping\n        if img1.shape[0] - self.crop_size[0] > 0:\n            y0 = np.random.randint(0, img1.shape[0] - self.crop_size[0])\n        else:\n            y0 = 0\n        if img1.shape[1] - self.crop_size[1] > 0:\n            x0 = np.random.randint(0, img1.shape[1] - self.crop_size[1])\n        else:\n            x0 = 0\n\n        img1 = img1[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n        img2 = img2[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n        flow = flow[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n\n        return img1, img2, flow\n\n    def __call__(self, img1, img2, flow, dname):\n        img1, img2, flow = self.spatial_transform(img1, img2, flow, dname)\n        img1, img2 = self.color_transform(img1, img2)\n        img1 = np.ascontiguousarray(img1)\n        img2 = np.ascontiguousarray(img2)\n        flow = np.ascontiguousarray(flow)\n        return img1, img2, flow","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor.color_transform","uri":"program://Human3R/function/src.croco.stereoflow.augmentor.color_transform#L272-L288","kind":"function","name":"color_transform","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":272,"end_line":288,"context_start_line":252,"context_end_line":308,"code":"\n        # spatial augmentation params\n        self.crop_size = crop_size\n        self.min_scale = min_scale\n        self.max_scale = max_scale\n        self.spatial_aug_prob = spatial_aug_prob\n        self.stretch_prob = stretch_prob\n        self.max_stretch = max_stretch\n\n        # flip augmentation params\n        self.h_flip_prob = h_flip_prob\n        self.v_flip_prob = v_flip_prob\n\n        # photometric augmentation params\n        self.photo_aug = ColorJitter(\n            brightness=0.4, contrast=0.4, saturation=0.4, hue=0.5 / 3.14\n        )\n\n        self.asymmetric_color_aug_prob = asymmetric_color_aug_prob\n\n    def color_transform(self, img1, img2):\n        \"\"\"Photometric augmentation\"\"\"\n\n        # asymmetric\n        if np.random.rand() < self.asymmetric_color_aug_prob:\n            img1 = np.array(self.photo_aug(Image.fromarray(img1)), dtype=np.uint8)\n            img2 = np.array(self.photo_aug(Image.fromarray(img2)), dtype=np.uint8)\n\n        # symmetric\n        else:\n            image_stack = np.concatenate([img1, img2], axis=0)\n            image_stack = np.array(\n                self.photo_aug(Image.fromarray(image_stack)), dtype=np.uint8\n            )\n            img1, img2 = np.split(image_stack, 2, axis=0)\n\n        return img1, img2\n\n    def _resize_flow(self, flow, scale_x, scale_y, factor=1.0):\n        if np.all(np.isfinite(flow)):\n            flow = cv2.resize(\n                flow,\n                None,\n                fx=scale_x / factor,\n                fy=scale_y / factor,\n                interpolation=cv2.INTER_LINEAR,\n            )\n            flow = flow * [scale_x, scale_y]\n        else:  # sparse version\n            fx, fy = scale_x, scale_y\n            ht, wd = flow.shape[:2]\n            coords = np.meshgrid(np.arange(wd), np.arange(ht))\n            coords = np.stack(coords, axis=-1)\n\n            coords = coords.reshape(-1, 2).astype(np.float32)\n            flow = flow.reshape(-1, 2).astype(np.float32)\n            valid = np.isfinite(flow[:, 0])","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor._resize_flow","uri":"program://Human3R/function/src.croco.stereoflow.augmentor._resize_flow#L290-L332","kind":"function","name":"_resize_flow","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":290,"end_line":332,"context_start_line":270,"context_end_line":352,"code":"        self.asymmetric_color_aug_prob = asymmetric_color_aug_prob\n\n    def color_transform(self, img1, img2):\n        \"\"\"Photometric augmentation\"\"\"\n\n        # asymmetric\n        if np.random.rand() < self.asymmetric_color_aug_prob:\n            img1 = np.array(self.photo_aug(Image.fromarray(img1)), dtype=np.uint8)\n            img2 = np.array(self.photo_aug(Image.fromarray(img2)), dtype=np.uint8)\n\n        # symmetric\n        else:\n            image_stack = np.concatenate([img1, img2], axis=0)\n            image_stack = np.array(\n                self.photo_aug(Image.fromarray(image_stack)), dtype=np.uint8\n            )\n            img1, img2 = np.split(image_stack, 2, axis=0)\n\n        return img1, img2\n\n    def _resize_flow(self, flow, scale_x, scale_y, factor=1.0):\n        if np.all(np.isfinite(flow)):\n            flow = cv2.resize(\n                flow,\n                None,\n                fx=scale_x / factor,\n                fy=scale_y / factor,\n                interpolation=cv2.INTER_LINEAR,\n            )\n            flow = flow * [scale_x, scale_y]\n        else:  # sparse version\n            fx, fy = scale_x, scale_y\n            ht, wd = flow.shape[:2]\n            coords = np.meshgrid(np.arange(wd), np.arange(ht))\n            coords = np.stack(coords, axis=-1)\n\n            coords = coords.reshape(-1, 2).astype(np.float32)\n            flow = flow.reshape(-1, 2).astype(np.float32)\n            valid = np.isfinite(flow[:, 0])\n\n            coords0 = coords[valid]\n            flow0 = flow[valid]\n\n            ht1 = int(round(ht * fy / factor))\n            wd1 = int(round(wd * fx / factor))\n\n            rescale = np.expand_dims(np.array([fx, fy]), axis=0)\n            coords1 = coords0 * rescale / factor\n            flow1 = flow0 * rescale\n\n            xx = np.round(coords1[:, 0]).astype(np.int32)\n            yy = np.round(coords1[:, 1]).astype(np.int32)\n\n            v = (xx > 0) & (xx < wd1) & (yy > 0) & (yy < ht1)\n            xx = xx[v]\n            yy = yy[v]\n            flow1 = flow1[v]\n\n            flow = np.inf * np.ones(\n                [ht1, wd1, 2], dtype=np.float32\n            )  # invalid value every where, before we fill it with the correct ones\n            flow[yy, xx] = flow1\n        return flow\n\n    def spatial_transform(self, img1, img2, flow, dname):\n\n        if np.random.rand() < self.spatial_aug_prob:\n            # randomly sample scale\n            ht, wd = img1.shape[:2]\n            clip_min_scale = np.maximum(\n                (self.crop_size[0] + 8) / float(ht), (self.crop_size[1] + 8) / float(wd)\n            )\n            min_scale, max_scale = self.min_scale, self.max_scale\n            scale = 2 ** np.random.uniform(self.min_scale, self.max_scale)\n            scale_x = scale\n            scale_y = scale\n            if np.random.rand() < self.stretch_prob:\n                scale_x *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)\n                scale_y *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)\n            scale_x = np.clip(scale_x, clip_min_scale, None)\n            scale_y = np.clip(scale_y, clip_min_scale, None)\n            # rescale the images\n            img1 = cv2.resize(","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.augmentor.spatial_transform","uri":"program://Human3R/function/src.croco.stereoflow.augmentor.spatial_transform#L334-L388","kind":"function","name":"spatial_transform","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":334,"end_line":388,"context_start_line":314,"context_end_line":396,"code":"            wd1 = int(round(wd * fx / factor))\n\n            rescale = np.expand_dims(np.array([fx, fy]), axis=0)\n            coords1 = coords0 * rescale / factor\n            flow1 = flow0 * rescale\n\n            xx = np.round(coords1[:, 0]).astype(np.int32)\n            yy = np.round(coords1[:, 1]).astype(np.int32)\n\n            v = (xx > 0) & (xx < wd1) & (yy > 0) & (yy < ht1)\n            xx = xx[v]\n            yy = yy[v]\n            flow1 = flow1[v]\n\n            flow = np.inf * np.ones(\n                [ht1, wd1, 2], dtype=np.float32\n            )  # invalid value every where, before we fill it with the correct ones\n            flow[yy, xx] = flow1\n        return flow\n\n    def spatial_transform(self, img1, img2, flow, dname):\n\n        if np.random.rand() < self.spatial_aug_prob:\n            # randomly sample scale\n            ht, wd = img1.shape[:2]\n            clip_min_scale = np.maximum(\n                (self.crop_size[0] + 8) / float(ht), (self.crop_size[1] + 8) / float(wd)\n            )\n            min_scale, max_scale = self.min_scale, self.max_scale\n            scale = 2 ** np.random.uniform(self.min_scale, self.max_scale)\n            scale_x = scale\n            scale_y = scale\n            if np.random.rand() < self.stretch_prob:\n                scale_x *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)\n                scale_y *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)\n            scale_x = np.clip(scale_x, clip_min_scale, None)\n            scale_y = np.clip(scale_y, clip_min_scale, None)\n            # rescale the images\n            img1 = cv2.resize(\n                img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            img2 = cv2.resize(\n                img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR\n            )\n            flow = self._resize_flow(\n                flow, scale_x, scale_y, factor=2.0 if dname == \"Spring\" else 1.0\n            )\n        elif dname == \"Spring\":\n            flow = self._resize_flow(flow, 1.0, 1.0, factor=2.0)\n\n        if self.h_flip_prob > 0.0 and np.random.rand() < self.h_flip_prob:  # h-flip\n            img1 = img1[:, ::-1]\n            img2 = img2[:, ::-1]\n            flow = flow[:, ::-1] * [-1.0, 1.0]\n\n        if self.v_flip_prob > 0.0 and np.random.rand() < self.v_flip_prob:  # v-flip\n            img1 = img1[::-1, :]\n            img2 = img2[::-1, :]\n            flow = flow[::-1, :] * [1.0, -1.0]\n\n        # In case no cropping\n        if img1.shape[0] - self.crop_size[0] > 0:\n            y0 = np.random.randint(0, img1.shape[0] - self.crop_size[0])\n        else:\n            y0 = 0\n        if img1.shape[1] - self.crop_size[1] > 0:\n            x0 = np.random.randint(0, img1.shape[1] - self.crop_size[1])\n        else:\n            x0 = 0\n\n        img1 = img1[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n        img2 = img2[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n        flow = flow[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]\n\n        return img1, img2, flow\n\n    def __call__(self, img1, img2, flow, dname):\n        img1, img2, flow = self.spatial_transform(img1, img2, flow, dname)\n        img1, img2 = self.color_transform(img1, img2)\n        img1 = np.ascontiguousarray(img1)\n        img2 = np.ascontiguousarray(img2)\n        flow = np.ascontiguousarray(flow)\n        return img1, img2, flow","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.train","uri":"program://Human3R/module/src.croco.stereoflow.train#L1-L455","kind":"module","name":"src.croco.stereoflow.train","path":"src/croco/stereoflow/train.py","language":"python","start_line":1,"end_line":455,"context_start_line":1,"context_end_line":455,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Main training function\n# --------------------------------------------------------\n\nimport argparse\nimport datetime\nimport json\nimport numpy as np\nimport os\nimport sys\nimport time\n\nimport torch\nimport torch.distributed as dist\nimport torch.backends.cudnn as cudnn\nfrom torch.utils.tensorboard import SummaryWriter\nimport torchvision.transforms as transforms\nimport torchvision.datasets as datasets\nfrom torch.utils.data import DataLoader\n\nimport utils\nimport utils.misc as misc\nfrom utils.misc import NativeScalerWithGradNormCount as NativeScaler\nfrom models.croco_downstream import CroCoDownstreamBinocular, croco_args_from_ckpt\nfrom models.pos_embed import interpolate_pos_embed\nfrom models.head_downstream import PixelwiseTaskWithDPT\n\nfrom stereoflow.datasets_stereo import (\n    get_train_dataset_stereo,\n    get_test_datasets_stereo,\n)\nfrom stereoflow.datasets_flow import get_train_dataset_flow, get_test_datasets_flow\nfrom stereoflow.engine import train_one_epoch, validate_one_epoch\nfrom stereoflow.criterion import *\n\n\ndef get_args_parser():\n    # prepare subparsers\n    parser = argparse.ArgumentParser(\n        \"Finetuning CroCo models on stereo or flow\", add_help=False\n    )\n    subparsers = parser.add_subparsers(\n        title=\"Task (stereo or flow)\", dest=\"task\", required=True\n    )\n    parser_stereo = subparsers.add_parser(\"stereo\", help=\"Training stereo model\")\n    parser_flow = subparsers.add_parser(\"flow\", help=\"Training flow model\")\n\n    def add_arg(\n        name_or_flags, default=None, default_stereo=None, default_flow=None, **kwargs\n    ):\n        if default is not None:\n            assert (\n                default_stereo is None and default_flow is None\n            ), \"setting default makes default_stereo and default_flow disabled\"\n        parser_stereo.add_argument(\n            name_or_flags,\n            default=default if default is not None else default_stereo,\n            **kwargs,\n        )\n        parser_flow.add_argument(\n            name_or_flags,\n            default=default if default is not None else default_flow,\n            **kwargs,\n        )\n\n    # output dir\n    add_arg(\n        \"--output_dir\",\n        required=True,\n        type=str,\n        help=\"path where to save, if empty, automatically created\",\n    )\n    # model\n    add_arg(\n        \"--crop\",\n        type=int,\n        nargs=\"+\",\n        default_stereo=[352, 704],\n        default_flow=[320, 384],\n        help=\"size of the random image crops used during training.\",\n    )\n    add_arg(\n        \"--pretrained\",\n        required=True,\n        type=str,\n        help=\"Load pretrained model (required as croco arguments come from there)\",\n    )\n    # criterion\n    add_arg(\n        \"--criterion\",\n        default_stereo=\"LaplacianLossBounded2()\",\n        default_flow=\"LaplacianLossBounded()\",\n        type=str,\n        help=\"string to evaluate to get criterion\",\n    )\n    add_arg(\"--bestmetric\", default_stereo=\"avgerr\", default_flow=\"EPE\", type=str)\n    # dataset\n    add_arg(\"--dataset\", type=str, required=True, help=\"training set\")\n    # training\n    add_arg(\"--seed\", default=0, type=int, help=\"seed\")\n    add_arg(\n        \"--batch_size\",\n        default_stereo=6,\n        default_flow=8,\n        type=int,\n        help=\"Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus\",\n    )\n    add_arg(\"--epochs\", default=32, type=int, help=\"number of training epochs\")\n    add_arg(\n        \"--img_per_epoch\",\n        type=int,\n        default=None,\n        help=\"Fix the number of images seen in an epoch (None means use all training pairs)\",\n    )\n    add_arg(\n        \"--accum_iter\",\n        default=1,\n        type=int,\n        help=\"Accumulate gradient iterations (for increasing the effective batch size under memory constraints)\",\n    )\n    add_arg(\n        \"--weight_decay\", type=float, default=0.05, help=\"weight decay (default: 0.05)\"\n    )\n    add_arg(\n        \"--lr\",\n        type=float,\n        default_stereo=3e-5,\n        default_flow=2e-5,\n        metavar=\"LR\",\n        help=\"learning rate (absolute lr)\",\n    )\n    add_arg(\n        \"--min_lr\",\n        type=float,\n        default=0.0,\n        metavar=\"LR\",\n        help=\"lower lr bound for cyclic schedulers that hit 0\",\n    )\n    add_arg(\n        \"--warmup_epochs\", type=int, default=1, metavar=\"N\", help=\"epochs to warmup LR\"\n    )\n    add_arg(\n        \"--optimizer\",\n        default=\"AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))\",\n        type=str,\n        help=\"Optimizer from torch.optim [ default: AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95)) ]\",\n    )\n    add_arg(\n        \"--amp\",\n        default=0,\n        type=int,\n        choices=[0, 1],\n        help=\"enable automatic mixed precision training\",\n    )\n    # validation\n    add_arg(\n        \"--val_dataset\",\n        type=str,\n        default=\"\",\n        help=\"Validation sets, multiple separated by + (empty string means that no validation is performed)\",\n    )\n    add_arg(\n        \"--tile_conf_mode\",\n        type=str,\n        default_stereo=\"conf_expsigmoid_15_3\",\n        default_flow=\"conf_expsigmoid_10_5\",\n        help=\"Weights for tile aggregation\",\n    )\n    add_arg(\n        \"--val_overlap\", default=0.7, type=float, help=\"Overlap value for the tiling\"\n    )\n    # others\n    add_arg(\"--num_workers\", default=8, type=int)\n    add_arg(\"--eval_every\", type=int, default=1, help=\"Val loss evaluation frequency\")\n    add_arg(\"--save_every\", type=int, default=1, help=\"Save checkpoint frequency\")\n    add_arg(\n        \"--start_from\",\n        type=str,\n        default=None,\n        help=\"Start training using weights from an other model (eg for finetuning)\",\n    )\n    add_arg(\n        \"--tboard_log_step\",\n        type=int,\n        default=100,\n        help=\"Log to tboard every so many steps\",\n    )\n    add_arg(\n        \"--dist_url\", default=\"env://\", help=\"url used to set up distributed training\"\n    )\n\n    return parser\n\n\ndef main(args):\n    misc.init_distributed_mode(args)\n    global_rank = misc.get_rank()\n    num_tasks = misc.get_world_size()\n\n    assert os.path.isfile(args.pretrained)\n    print(\"output_dir: \" + args.output_dir)\n    os.makedirs(args.output_dir, exist_ok=True)\n\n    # fix the seed for reproducibility\n    seed = args.seed + misc.get_rank()\n    torch.manual_seed(seed)\n    np.random.seed(seed)\n    cudnn.benchmark = True\n\n    # Metrics / criterion\n    device = torch.device(\"cuda\") if torch.cuda.is_available() else torch.device(\"cpu\")\n    metrics = (StereoMetrics if args.task == \"stereo\" else FlowMetrics)().to(device)\n    criterion = eval(args.criterion).to(device)\n    print(\"Criterion: \", args.criterion)\n\n    # Prepare model\n    assert os.path.isfile(args.pretrained)\n    ckpt = torch.load(args.pretrained, \"cpu\")\n    croco_args = croco_args_from_ckpt(ckpt)\n    croco_args[\"img_size\"] = (args.crop[0], args.crop[1])\n    print(\"Croco args: \" + str(croco_args))\n    args.croco_args = croco_args  # saved for test time\n    # prepare head\n    num_channels = {\"stereo\": 1, \"flow\": 2}[args.task]\n    if criterion.with_conf:\n        num_channels += 1\n    print(f\"Building head PixelwiseTaskWithDPT() with {num_channels} channel(s)\")\n    head = PixelwiseTaskWithDPT()\n    head.num_channels = num_channels\n    # build model and load pretrained weights\n    model = CroCoDownstreamBinocular(head, **croco_args)\n    interpolate_pos_embed(model, ckpt[\"model\"])\n    msg = model.load_state_dict(ckpt[\"model\"], strict=False)\n    print(msg)\n\n    total_params = sum(p.numel() for p in model.parameters())\n    total_params_trainable = sum(\n        p.numel() for p in model.parameters() if p.requires_grad\n    )\n    print(f\"Total params: {total_params}\")\n    print(f\"Total params trainable: {total_params_trainable}\")\n    model_without_ddp = model.to(device)\n\n    eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size()\n    print(\"lr: %.2e\" % args.lr)\n    print(\"accumulate grad iterations: %d\" % args.accum_iter)\n    print(\"effective batch size: %d\" % eff_batch_size)\n\n    if args.distributed:\n        model = torch.nn.parallel.DistributedDataParallel(\n            model, device_ids=[args.gpu], static_graph=True\n        )\n        model_without_ddp = model.module\n\n    # following timm: set wd as 0 for bias and norm layers\n    param_groups = misc.get_parameter_groups(model_without_ddp, args.weight_decay)\n    optimizer = eval(f\"torch.optim.{args.optimizer}\")\n    print(optimizer)\n    loss_scaler = NativeScaler()\n\n    # automatic restart\n    last_ckpt_fname = os.path.join(args.output_dir, f\"checkpoint-last.pth\")\n    args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None\n\n    if not args.resume and args.start_from:\n        print(f\"Starting from an other model's weights: {args.start_from}\")\n        best_so_far = None\n        args.start_epoch = 0\n        ckpt = torch.load(args.start_from, \"cpu\")\n        msg = model_without_ddp.load_state_dict(ckpt[\"model\"], strict=False)\n        print(msg)\n    else:\n        best_so_far = misc.load_model(\n            args=args,\n            model_without_ddp=model_without_ddp,\n            optimizer=optimizer,\n            loss_scaler=loss_scaler,\n        )\n\n    if best_so_far is None:\n        best_so_far = np.inf\n\n    # tensorboard\n    log_writer = None\n    if global_rank == 0 and args.output_dir is not None:\n        log_writer = SummaryWriter(\n            log_dir=args.output_dir, purge_step=args.start_epoch * 1000\n        )\n\n    #  dataset and loader\n    print(\"Building Train Data loader for dataset: \", args.dataset)\n    train_dataset = (\n        get_train_dataset_stereo if args.task == \"stereo\" else get_train_dataset_flow\n    )(args.dataset, crop_size=args.crop)\n\n    def _print_repr_dataset(d):\n        if isinstance(d, torch.utils.data.dataset.ConcatDataset):\n            for dd in d.datasets:\n                _print_repr_dataset(dd)\n        else:\n            print(repr(d))\n\n    _print_repr_dataset(train_dataset)\n    print(\"  total length:\", len(train_dataset))\n    if args.distributed:\n        sampler_train = torch.utils.data.DistributedSampler(\n            train_dataset, num_replicas=num_tasks, rank=global_rank, shuffle=True\n        )\n    else:\n        sampler_train = torch.utils.data.RandomSampler(train_dataset)\n    data_loader_train = torch.utils.data.DataLoader(\n        train_dataset,\n        sampler=sampler_train,\n        batch_size=args.batch_size,\n        num_workers=args.num_workers,\n        pin_memory=True,\n        drop_last=True,\n    )\n    if args.val_dataset == \"\":\n        data_loaders_val = None\n    else:\n        print(\"Building Val Data loader for datasets: \", args.val_dataset)\n        val_datasets = (\n            get_test_datasets_stereo\n            if args.task == \"stereo\"\n            else get_test_datasets_flow\n        )(args.val_dataset)\n        for val_dataset in val_datasets:\n            print(repr(val_dataset))\n        data_loaders_val = [\n            DataLoader(\n                val_dataset,\n                batch_size=1,\n                shuffle=False,\n                num_workers=args.num_workers,\n                pin_memory=True,\n                drop_last=False,\n            )\n            for val_dataset in val_datasets\n        ]\n        bestmetric = (\n            \"AVG_\"\n            if len(data_loaders_val) > 1\n            else str(data_loaders_val[0].dataset) + \"_\"\n        ) + args.bestmetric\n\n    print(f\"Start training for {args.epochs} epochs\")\n    start_time = time.time()\n    # Training Loop\n    for epoch in range(args.start_epoch, args.epochs):\n\n        if args.distributed:\n            data_loader_train.sampler.set_epoch(epoch)\n\n        # Train\n        epoch_start = time.time()\n        train_stats = train_one_epoch(\n            model,\n            criterion,\n            metrics,\n            data_loader_train,\n            optimizer,\n            device,\n            epoch,\n            loss_scaler,\n            log_writer=log_writer,\n            args=args,\n        )\n        epoch_time = time.time() - epoch_start\n\n        if args.distributed:\n            dist.barrier()\n\n        # Validation (current naive implementation runs the validation on every gpu ... not smart ...)\n        if (\n            data_loaders_val is not None\n            and args.eval_every > 0\n            and (epoch + 1) % args.eval_every == 0\n        ):\n            val_epoch_start = time.time()\n            val_stats = validate_one_epoch(\n                model,\n                criterion,\n                metrics,\n                data_loaders_val,\n                device,\n                epoch,\n                log_writer=log_writer,\n                args=args,\n            )\n            val_epoch_time = time.time() - val_epoch_start\n\n            val_best = val_stats[bestmetric]\n\n            # Save best of all\n            if val_best <= best_so_far:\n                best_so_far = val_best\n                misc.save_model(\n                    args=args,\n                    model_without_ddp=model_without_ddp,\n                    optimizer=optimizer,\n                    loss_scaler=loss_scaler,\n                    epoch=epoch,\n                    best_so_far=best_so_far,\n                    fname=\"best\",\n                )\n\n            log_stats = {\n                **{f\"train_{k}\": v for k, v in train_stats.items()},\n                \"epoch\": epoch,\n                **{f\"val_{k}\": v for k, v in val_stats.items()},\n            }\n        else:\n            log_stats = {\n                **{f\"train_{k}\": v for k, v in train_stats.items()},\n                \"epoch\": epoch,\n            }\n\n        if args.distributed:\n            dist.barrier()\n\n        # Save stuff\n        if args.output_dir and (\n            (epoch + 1) % args.save_every == 0 or epoch + 1 == args.epochs\n        ):\n            misc.save_model(\n                args=args,\n                model_without_ddp=model_without_ddp,\n                optimizer=optimizer,\n                loss_scaler=loss_scaler,\n                epoch=epoch,\n                best_so_far=best_so_far,\n                fname=\"last\",\n            )\n\n        if args.output_dir:\n            if log_writer is not None:\n                log_writer.flush()\n            with open(\n                os.path.join(args.output_dir, \"log.txt\"), mode=\"a\", encoding=\"utf-8\"\n            ) as f:\n                f.write(json.dumps(log_stats) + \"\\n\")\n\n    total_time = time.time() - start_time\n    total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n    print(\"Training time {}\".format(total_time_str))\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    main(args)","source_hash":"504e0fbf6bf43cfb177a5b689add022db755bc954edd509fef0acb4b2b9693ad","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.train.get_args_parser","uri":"program://Human3R/function/src.croco.stereoflow.train.get_args_parser#L40-L195","kind":"function","name":"get_args_parser","path":"src/croco/stereoflow/train.py","language":"python","start_line":40,"end_line":195,"context_start_line":20,"context_end_line":215,"code":"import torchvision.transforms as transforms\nimport torchvision.datasets as datasets\nfrom torch.utils.data import DataLoader\n\nimport utils\nimport utils.misc as misc\nfrom utils.misc import NativeScalerWithGradNormCount as NativeScaler\nfrom models.croco_downstream import CroCoDownstreamBinocular, croco_args_from_ckpt\nfrom models.pos_embed import interpolate_pos_embed\nfrom models.head_downstream import PixelwiseTaskWithDPT\n\nfrom stereoflow.datasets_stereo import (\n    get_train_dataset_stereo,\n    get_test_datasets_stereo,\n)\nfrom stereoflow.datasets_flow import get_train_dataset_flow, get_test_datasets_flow\nfrom stereoflow.engine import train_one_epoch, validate_one_epoch\nfrom stereoflow.criterion import *\n\n\ndef get_args_parser():\n    # prepare subparsers\n    parser = argparse.ArgumentParser(\n        \"Finetuning CroCo models on stereo or flow\", add_help=False\n    )\n    subparsers = parser.add_subparsers(\n        title=\"Task (stereo or flow)\", dest=\"task\", required=True\n    )\n    parser_stereo = subparsers.add_parser(\"stereo\", help=\"Training stereo model\")\n    parser_flow = subparsers.add_parser(\"flow\", help=\"Training flow model\")\n\n    def add_arg(\n        name_or_flags, default=None, default_stereo=None, default_flow=None, **kwargs\n    ):\n        if default is not None:\n            assert (\n                default_stereo is None and default_flow is None\n            ), \"setting default makes default_stereo and default_flow disabled\"\n        parser_stereo.add_argument(\n            name_or_flags,\n            default=default if default is not None else default_stereo,\n            **kwargs,\n        )\n        parser_flow.add_argument(\n            name_or_flags,\n            default=default if default is not None else default_flow,\n            **kwargs,\n        )\n\n    # output dir\n    add_arg(\n        \"--output_dir\",\n        required=True,\n        type=str,\n        help=\"path where to save, if empty, automatically created\",\n    )\n    # model\n    add_arg(\n        \"--crop\",\n        type=int,\n        nargs=\"+\",\n        default_stereo=[352, 704],\n        default_flow=[320, 384],\n        help=\"size of the random image crops used during training.\",\n    )\n    add_arg(\n        \"--pretrained\",\n        required=True,\n        type=str,\n        help=\"Load pretrained model (required as croco arguments come from there)\",\n    )\n    # criterion\n    add_arg(\n        \"--criterion\",\n        default_stereo=\"LaplacianLossBounded2()\",\n        default_flow=\"LaplacianLossBounded()\",\n        type=str,\n        help=\"string to evaluate to get criterion\",\n    )\n    add_arg(\"--bestmetric\", default_stereo=\"avgerr\", default_flow=\"EPE\", type=str)\n    # dataset\n    add_arg(\"--dataset\", type=str, required=True, help=\"training set\")\n    # training\n    add_arg(\"--seed\", default=0, type=int, help=\"seed\")\n    add_arg(\n        \"--batch_size\",\n        default_stereo=6,\n        default_flow=8,\n        type=int,\n        help=\"Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus\",\n    )\n    add_arg(\"--epochs\", default=32, type=int, help=\"number of training epochs\")\n    add_arg(\n        \"--img_per_epoch\",\n        type=int,\n        default=None,\n        help=\"Fix the number of images seen in an epoch (None means use all training pairs)\",\n    )\n    add_arg(\n        \"--accum_iter\",\n        default=1,\n        type=int,\n        help=\"Accumulate gradient iterations (for increasing the effective batch size under memory constraints)\",\n    )\n    add_arg(\n        \"--weight_decay\", type=float, default=0.05, help=\"weight decay (default: 0.05)\"\n    )\n    add_arg(\n        \"--lr\",\n        type=float,\n        default_stereo=3e-5,\n        default_flow=2e-5,\n        metavar=\"LR\",\n        help=\"learning rate (absolute lr)\",\n    )\n    add_arg(\n        \"--min_lr\",\n        type=float,\n        default=0.0,\n        metavar=\"LR\",\n        help=\"lower lr bound for cyclic schedulers that hit 0\",\n    )\n    add_arg(\n        \"--warmup_epochs\", type=int, default=1, metavar=\"N\", help=\"epochs to warmup LR\"\n    )\n    add_arg(\n        \"--optimizer\",\n        default=\"AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))\",\n        type=str,\n        help=\"Optimizer from torch.optim [ default: AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95)) ]\",\n    )\n    add_arg(\n        \"--amp\",\n        default=0,\n        type=int,\n        choices=[0, 1],\n        help=\"enable automatic mixed precision training\",\n    )\n    # validation\n    add_arg(\n        \"--val_dataset\",\n        type=str,\n        default=\"\",\n        help=\"Validation sets, multiple separated by + (empty string means that no validation is performed)\",\n    )\n    add_arg(\n        \"--tile_conf_mode\",\n        type=str,\n        default_stereo=\"conf_expsigmoid_15_3\",\n        default_flow=\"conf_expsigmoid_10_5\",\n        help=\"Weights for tile aggregation\",\n    )\n    add_arg(\n        \"--val_overlap\", default=0.7, type=float, help=\"Overlap value for the tiling\"\n    )\n    # others\n    add_arg(\"--num_workers\", default=8, type=int)\n    add_arg(\"--eval_every\", type=int, default=1, help=\"Val loss evaluation frequency\")\n    add_arg(\"--save_every\", type=int, default=1, help=\"Save checkpoint frequency\")\n    add_arg(\n        \"--start_from\",\n        type=str,\n        default=None,\n        help=\"Start training using weights from an other model (eg for finetuning)\",\n    )\n    add_arg(\n        \"--tboard_log_step\",\n        type=int,\n        default=100,\n        help=\"Log to tboard every so many steps\",\n    )\n    add_arg(\n        \"--dist_url\", default=\"env://\", help=\"url used to set up distributed training\"\n    )\n\n    return parser\n\n\ndef main(args):\n    misc.init_distributed_mode(args)\n    global_rank = misc.get_rank()\n    num_tasks = misc.get_world_size()\n\n    assert os.path.isfile(args.pretrained)\n    print(\"output_dir: \" + args.output_dir)\n    os.makedirs(args.output_dir, exist_ok=True)\n\n    # fix the seed for reproducibility\n    seed = args.seed + misc.get_rank()\n    torch.manual_seed(seed)\n    np.random.seed(seed)\n    cudnn.benchmark = True\n\n    # Metrics / criterion\n    device = torch.device(\"cuda\") if torch.cuda.is_available() else torch.device(\"cpu\")\n    metrics = (StereoMetrics if args.task == \"stereo\" else FlowMetrics)().to(device)","source_hash":"504e0fbf6bf43cfb177a5b689add022db755bc954edd509fef0acb4b2b9693ad","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.train.main","uri":"program://Human3R/function/src.croco.stereoflow.train.main#L198-L449","kind":"function","name":"main","path":"src/croco/stereoflow/train.py","language":"python","start_line":198,"end_line":449,"context_start_line":178,"context_end_line":455,"code":"    add_arg(\"--save_every\", type=int, default=1, help=\"Save checkpoint frequency\")\n    add_arg(\n        \"--start_from\",\n        type=str,\n        default=None,\n        help=\"Start training using weights from an other model (eg for finetuning)\",\n    )\n    add_arg(\n        \"--tboard_log_step\",\n        type=int,\n        default=100,\n        help=\"Log to tboard every so many steps\",\n    )\n    add_arg(\n        \"--dist_url\", default=\"env://\", help=\"url used to set up distributed training\"\n    )\n\n    return parser\n\n\ndef main(args):\n    misc.init_distributed_mode(args)\n    global_rank = misc.get_rank()\n    num_tasks = misc.get_world_size()\n\n    assert os.path.isfile(args.pretrained)\n    print(\"output_dir: \" + args.output_dir)\n    os.makedirs(args.output_dir, exist_ok=True)\n\n    # fix the seed for reproducibility\n    seed = args.seed + misc.get_rank()\n    torch.manual_seed(seed)\n    np.random.seed(seed)\n    cudnn.benchmark = True\n\n    # Metrics / criterion\n    device = torch.device(\"cuda\") if torch.cuda.is_available() else torch.device(\"cpu\")\n    metrics = (StereoMetrics if args.task == \"stereo\" else FlowMetrics)().to(device)\n    criterion = eval(args.criterion).to(device)\n    print(\"Criterion: \", args.criterion)\n\n    # Prepare model\n    assert os.path.isfile(args.pretrained)\n    ckpt = torch.load(args.pretrained, \"cpu\")\n    croco_args = croco_args_from_ckpt(ckpt)\n    croco_args[\"img_size\"] = (args.crop[0], args.crop[1])\n    print(\"Croco args: \" + str(croco_args))\n    args.croco_args = croco_args  # saved for test time\n    # prepare head\n    num_channels = {\"stereo\": 1, \"flow\": 2}[args.task]\n    if criterion.with_conf:\n        num_channels += 1\n    print(f\"Building head PixelwiseTaskWithDPT() with {num_channels} channel(s)\")\n    head = PixelwiseTaskWithDPT()\n    head.num_channels = num_channels\n    # build model and load pretrained weights\n    model = CroCoDownstreamBinocular(head, **croco_args)\n    interpolate_pos_embed(model, ckpt[\"model\"])\n    msg = model.load_state_dict(ckpt[\"model\"], strict=False)\n    print(msg)\n\n    total_params = sum(p.numel() for p in model.parameters())\n    total_params_trainable = sum(\n        p.numel() for p in model.parameters() if p.requires_grad\n    )\n    print(f\"Total params: {total_params}\")\n    print(f\"Total params trainable: {total_params_trainable}\")\n    model_without_ddp = model.to(device)\n\n    eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size()\n    print(\"lr: %.2e\" % args.lr)\n    print(\"accumulate grad iterations: %d\" % args.accum_iter)\n    print(\"effective batch size: %d\" % eff_batch_size)\n\n    if args.distributed:\n        model = torch.nn.parallel.DistributedDataParallel(\n            model, device_ids=[args.gpu], static_graph=True\n        )\n        model_without_ddp = model.module\n\n    # following timm: set wd as 0 for bias and norm layers\n    param_groups = misc.get_parameter_groups(model_without_ddp, args.weight_decay)\n    optimizer = eval(f\"torch.optim.{args.optimizer}\")\n    print(optimizer)\n    loss_scaler = NativeScaler()\n\n    # automatic restart\n    last_ckpt_fname = os.path.join(args.output_dir, f\"checkpoint-last.pth\")\n    args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None\n\n    if not args.resume and args.start_from:\n        print(f\"Starting from an other model's weights: {args.start_from}\")\n        best_so_far = None\n        args.start_epoch = 0\n        ckpt = torch.load(args.start_from, \"cpu\")\n        msg = model_without_ddp.load_state_dict(ckpt[\"model\"], strict=False)\n        print(msg)\n    else:\n        best_so_far = misc.load_model(\n            args=args,\n            model_without_ddp=model_without_ddp,\n            optimizer=optimizer,\n            loss_scaler=loss_scaler,\n        )\n\n    if best_so_far is None:\n        best_so_far = np.inf\n\n    # tensorboard\n    log_writer = None\n    if global_rank == 0 and args.output_dir is not None:\n        log_writer = SummaryWriter(\n            log_dir=args.output_dir, purge_step=args.start_epoch * 1000\n        )\n\n    #  dataset and loader\n    print(\"Building Train Data loader for dataset: \", args.dataset)\n    train_dataset = (\n        get_train_dataset_stereo if args.task == \"stereo\" else get_train_dataset_flow\n    )(args.dataset, crop_size=args.crop)\n\n    def _print_repr_dataset(d):\n        if isinstance(d, torch.utils.data.dataset.ConcatDataset):\n            for dd in d.datasets:\n                _print_repr_dataset(dd)\n        else:\n            print(repr(d))\n\n    _print_repr_dataset(train_dataset)\n    print(\"  total length:\", len(train_dataset))\n    if args.distributed:\n        sampler_train = torch.utils.data.DistributedSampler(\n            train_dataset, num_replicas=num_tasks, rank=global_rank, shuffle=True\n        )\n    else:\n        sampler_train = torch.utils.data.RandomSampler(train_dataset)\n    data_loader_train = torch.utils.data.DataLoader(\n        train_dataset,\n        sampler=sampler_train,\n        batch_size=args.batch_size,\n        num_workers=args.num_workers,\n        pin_memory=True,\n        drop_last=True,\n    )\n    if args.val_dataset == \"\":\n        data_loaders_val = None\n    else:\n        print(\"Building Val Data loader for datasets: \", args.val_dataset)\n        val_datasets = (\n            get_test_datasets_stereo\n            if args.task == \"stereo\"\n            else get_test_datasets_flow\n        )(args.val_dataset)\n        for val_dataset in val_datasets:\n            print(repr(val_dataset))\n        data_loaders_val = [\n            DataLoader(\n                val_dataset,\n                batch_size=1,\n                shuffle=False,\n                num_workers=args.num_workers,\n                pin_memory=True,\n                drop_last=False,\n            )\n            for val_dataset in val_datasets\n        ]\n        bestmetric = (\n            \"AVG_\"\n            if len(data_loaders_val) > 1\n            else str(data_loaders_val[0].dataset) + \"_\"\n        ) + args.bestmetric\n\n    print(f\"Start training for {args.epochs} epochs\")\n    start_time = time.time()\n    # Training Loop\n    for epoch in range(args.start_epoch, args.epochs):\n\n        if args.distributed:\n            data_loader_train.sampler.set_epoch(epoch)\n\n        # Train\n        epoch_start = time.time()\n        train_stats = train_one_epoch(\n            model,\n            criterion,\n            metrics,\n            data_loader_train,\n            optimizer,\n            device,\n            epoch,\n            loss_scaler,\n            log_writer=log_writer,\n            args=args,\n        )\n        epoch_time = time.time() - epoch_start\n\n        if args.distributed:\n            dist.barrier()\n\n        # Validation (current naive implementation runs the validation on every gpu ... not smart ...)\n        if (\n            data_loaders_val is not None\n            and args.eval_every > 0\n            and (epoch + 1) % args.eval_every == 0\n        ):\n            val_epoch_start = time.time()\n            val_stats = validate_one_epoch(\n                model,\n                criterion,\n                metrics,\n                data_loaders_val,\n                device,\n                epoch,\n                log_writer=log_writer,\n                args=args,\n            )\n            val_epoch_time = time.time() - val_epoch_start\n\n            val_best = val_stats[bestmetric]\n\n            # Save best of all\n            if val_best <= best_so_far:\n                best_so_far = val_best\n                misc.save_model(\n                    args=args,\n                    model_without_ddp=model_without_ddp,\n                    optimizer=optimizer,\n                    loss_scaler=loss_scaler,\n                    epoch=epoch,\n                    best_so_far=best_so_far,\n                    fname=\"best\",\n                )\n\n            log_stats = {\n                **{f\"train_{k}\": v for k, v in train_stats.items()},\n                \"epoch\": epoch,\n                **{f\"val_{k}\": v for k, v in val_stats.items()},\n            }\n        else:\n            log_stats = {\n                **{f\"train_{k}\": v for k, v in train_stats.items()},\n                \"epoch\": epoch,\n            }\n\n        if args.distributed:\n            dist.barrier()\n\n        # Save stuff\n        if args.output_dir and (\n            (epoch + 1) % args.save_every == 0 or epoch + 1 == args.epochs\n        ):\n            misc.save_model(\n                args=args,\n                model_without_ddp=model_without_ddp,\n                optimizer=optimizer,\n                loss_scaler=loss_scaler,\n                epoch=epoch,\n                best_so_far=best_so_far,\n                fname=\"last\",\n            )\n\n        if args.output_dir:\n            if log_writer is not None:\n                log_writer.flush()\n            with open(\n                os.path.join(args.output_dir, \"log.txt\"), mode=\"a\", encoding=\"utf-8\"\n            ) as f:\n                f.write(json.dumps(log_stats) + \"\\n\")\n\n    total_time = time.time() - start_time\n    total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n    print(\"Training time {}\".format(total_time_str))\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    main(args)","source_hash":"504e0fbf6bf43cfb177a5b689add022db755bc954edd509fef0acb4b2b9693ad","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.train.add_arg","uri":"program://Human3R/function/src.croco.stereoflow.train.add_arg#L51-L67","kind":"function","name":"add_arg","path":"src/croco/stereoflow/train.py","language":"python","start_line":51,"end_line":67,"context_start_line":31,"context_end_line":87,"code":"from stereoflow.datasets_stereo import (\n    get_train_dataset_stereo,\n    get_test_datasets_stereo,\n)\nfrom stereoflow.datasets_flow import get_train_dataset_flow, get_test_datasets_flow\nfrom stereoflow.engine import train_one_epoch, validate_one_epoch\nfrom stereoflow.criterion import *\n\n\ndef get_args_parser():\n    # prepare subparsers\n    parser = argparse.ArgumentParser(\n        \"Finetuning CroCo models on stereo or flow\", add_help=False\n    )\n    subparsers = parser.add_subparsers(\n        title=\"Task (stereo or flow)\", dest=\"task\", required=True\n    )\n    parser_stereo = subparsers.add_parser(\"stereo\", help=\"Training stereo model\")\n    parser_flow = subparsers.add_parser(\"flow\", help=\"Training flow model\")\n\n    def add_arg(\n        name_or_flags, default=None, default_stereo=None, default_flow=None, **kwargs\n    ):\n        if default is not None:\n            assert (\n                default_stereo is None and default_flow is None\n            ), \"setting default makes default_stereo and default_flow disabled\"\n        parser_stereo.add_argument(\n            name_or_flags,\n            default=default if default is not None else default_stereo,\n            **kwargs,\n        )\n        parser_flow.add_argument(\n            name_or_flags,\n            default=default if default is not None else default_flow,\n            **kwargs,\n        )\n\n    # output dir\n    add_arg(\n        \"--output_dir\",\n        required=True,\n        type=str,\n        help=\"path where to save, if empty, automatically created\",\n    )\n    # model\n    add_arg(\n        \"--crop\",\n        type=int,\n        nargs=\"+\",\n        default_stereo=[352, 704],\n        default_flow=[320, 384],\n        help=\"size of the random image crops used during training.\",\n    )\n    add_arg(\n        \"--pretrained\",\n        required=True,","source_hash":"504e0fbf6bf43cfb177a5b689add022db755bc954edd509fef0acb4b2b9693ad","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.train._print_repr_dataset","uri":"program://Human3R/function/src.croco.stereoflow.train._print_repr_dataset#L299-L304","kind":"function","name":"_print_repr_dataset","path":"src/croco/stereoflow/train.py","language":"python","start_line":299,"end_line":304,"context_start_line":279,"context_end_line":324,"code":"            optimizer=optimizer,\n            loss_scaler=loss_scaler,\n        )\n\n    if best_so_far is None:\n        best_so_far = np.inf\n\n    # tensorboard\n    log_writer = None\n    if global_rank == 0 and args.output_dir is not None:\n        log_writer = SummaryWriter(\n            log_dir=args.output_dir, purge_step=args.start_epoch * 1000\n        )\n\n    #  dataset and loader\n    print(\"Building Train Data loader for dataset: \", args.dataset)\n    train_dataset = (\n        get_train_dataset_stereo if args.task == \"stereo\" else get_train_dataset_flow\n    )(args.dataset, crop_size=args.crop)\n\n    def _print_repr_dataset(d):\n        if isinstance(d, torch.utils.data.dataset.ConcatDataset):\n            for dd in d.datasets:\n                _print_repr_dataset(dd)\n        else:\n            print(repr(d))\n\n    _print_repr_dataset(train_dataset)\n    print(\"  total length:\", len(train_dataset))\n    if args.distributed:\n        sampler_train = torch.utils.data.DistributedSampler(\n            train_dataset, num_replicas=num_tasks, rank=global_rank, shuffle=True\n        )\n    else:\n        sampler_train = torch.utils.data.RandomSampler(train_dataset)\n    data_loader_train = torch.utils.data.DataLoader(\n        train_dataset,\n        sampler=sampler_train,\n        batch_size=args.batch_size,\n        num_workers=args.num_workers,\n        pin_memory=True,\n        drop_last=True,\n    )\n    if args.val_dataset == \"\":\n        data_loaders_val = None\n    else:","source_hash":"504e0fbf6bf43cfb177a5b689add022db755bc954edd509fef0acb4b2b9693ad","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo","uri":"program://Human3R/module/src.croco.stereoflow.datasets_stereo#L1-L991","kind":"module","name":"src.croco.stereoflow.datasets_stereo","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":1,"end_line":991,"context_start_line":1,"context_end_line":991,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Dataset structure for stereo\n# --------------------------------------------------------\n\nimport sys, os\nimport os.path as osp\nimport pickle\nimport numpy as np\nfrom PIL import Image\nimport json\nimport h5py\nfrom glob import glob\nimport cv2\n\nimport torch\nfrom torch.utils import data\n\nfrom .augmentor import StereoAugmentor\n\n\ndataset_to_root = {\n    \"CREStereo\": \"./data/stereoflow//crenet_stereo_trainset/stereo_trainset/crestereo/\",\n    \"SceneFlow\": \"./data/stereoflow//SceneFlow/\",\n    \"ETH3DLowRes\": \"./data/stereoflow/eth3d_lowres/\",\n    \"Booster\": \"./data/stereoflow/booster_gt/\",\n    \"Middlebury2021\": \"./data/stereoflow/middlebury/2021/data/\",\n    \"Middlebury2014\": \"./data/stereoflow/middlebury/2014/\",\n    \"Middlebury2006\": \"./data/stereoflow/middlebury/2006/\",\n    \"Middlebury2005\": \"./data/stereoflow/middlebury/2005/train/\",\n    \"MiddleburyEval3\": \"./data/stereoflow/middlebury/MiddEval3/\",\n    \"Spring\": \"./data/stereoflow/spring/\",\n    \"Kitti15\": \"./data/stereoflow/kitti-stereo-2015/\",\n    \"Kitti12\": \"./data/stereoflow/kitti-stereo-2012/\",\n}\ncache_dir = \"./data/stereoflow/datasets_stereo_cache/\"\n\n\nin1k_mean = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1)\nin1k_std = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1)\n\n\ndef img_to_tensor(img):\n    img = torch.from_numpy(img).permute(2, 0, 1).float() / 255.0\n    img = (img - in1k_mean) / in1k_std\n    return img\n\n\ndef disp_to_tensor(disp):\n    return torch.from_numpy(disp)[None, :, :]\n\n\nclass StereoDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = StereoAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(self.pairnames)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        Limgname = self.pairname_to_Limgname(pairname)\n        Rimgname = self.pairname_to_Rimgname(pairname)\n        Ldispname = (\n            self.pairname_to_Ldispname(pairname)\n            if self.pairname_to_Ldispname is not None\n            else None\n        )\n\n        # load images and disparities\n        Limg = _read_img(Limgname)\n        Rimg = _read_img(Rimgname)\n        disp = self.load_disparity(Ldispname) if Ldispname is not None else None\n\n        # sanity check\n        if disp is not None:\n            assert np.all(disp > 0) or self.name == \"Spring\", (\n                self.name,\n                pairname,\n                Ldispname,\n            )\n\n        # apply augmentations\n        if self.augmentor is not None:\n            Limg, Rimg, disp = self.augmentor(Limg, Rimg, disp, self.name)\n\n        if self.totensor:\n            Limg = img_to_tensor(Limg)\n            Rimg = img_to_tensor(Rimg)\n            if disp is None:\n                disp = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            else:\n                disp = disp_to_tensor(disp)\n\n        return Limg, Rimg, disp, str(pairname)\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass CREStereoDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"CREStereo\"\n        self._set_root()\n        assert self.split in [\"train\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_left.jpg\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname + \"_right.jpg\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname + \"_left.disp.png\"\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_disparity = _read_crestereo_disp\n\n    def _build_cache(self):\n        allpairs = [\n            s + \"/\" + f[: -len(\"_left.jpg\")]\n            for s in sorted(os.listdir(self.root))\n            for f in sorted(os.listdir(self.root + \"/\" + s))\n            if f.endswith(\"_left.jpg\")\n        ]\n        assert len(allpairs) == 200000, \"incorrect parsing of pairs in CreStereo\"\n        tosave = {\"train\": allpairs}\n        return tosave\n\n\nclass SceneFlowDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"SceneFlow\"\n        self._set_root()\n        assert self.split in [\n            \"train_finalpass\",\n            \"train_cleanpass\",\n            \"train_allpass\",\n            \"test_finalpass\",\n            \"test_cleanpass\",\n            \"test_allpass\",\n            \"test1of100_cleanpass\",\n            \"test1of100_finalpass\",\n        ]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname\n        ).replace(\"/left/\", \"/right/\")\n        self.pairname_to_Ldispname = (\n            lambda pairname: osp.join(self.root, pairname)\n            .replace(\"/frames_finalpass/\", \"/disparity/\")\n            .replace(\"/frames_cleanpass/\", \"/disparity/\")[:-4]\n            + \".pfm\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_sceneflow_disp\n\n    def _build_cache(self):\n        trainpairs = []\n        # driving\n        pairs = sorted(glob(self.root + \"Driving/frames_finalpass/*/*/*/left/*.png\"))\n        pairs = list(map(lambda x: x[len(self.root) :], pairs))\n        assert len(pairs) == 4400, \"incorrect parsing of pairs in SceneFlow\"\n        trainpairs += pairs\n        # monkaa\n        pairs = sorted(glob(self.root + \"Monkaa/frames_finalpass/*/left/*.png\"))\n        pairs = list(map(lambda x: x[len(self.root) :], pairs))\n        assert len(pairs) == 8664, \"incorrect parsing of pairs in SceneFlow\"\n        trainpairs += pairs\n        # flyingthings\n        pairs = sorted(\n            glob(self.root + \"FlyingThings/frames_finalpass/TRAIN/*/*/left/*.png\")\n        )\n        pairs = list(map(lambda x: x[len(self.root) :], pairs))\n        assert len(pairs) == 22390, \"incorrect parsing of pairs in SceneFlow\"\n        trainpairs += pairs\n        assert len(trainpairs) == 35454, \"incorrect parsing of pairs in SceneFlow\"\n        testpairs = sorted(\n            glob(self.root + \"FlyingThings/frames_finalpass/TEST/*/*/left/*.png\")\n        )\n        testpairs = list(map(lambda x: x[len(self.root) :], testpairs))\n        assert len(testpairs) == 4370, \"incorrect parsing of pairs in SceneFlow\"\n        test1of100pairs = testpairs[::100]\n        assert len(test1of100pairs) == 44, \"incorrect parsing of pairs in SceneFlow\"\n        # all\n        tosave = {\n            \"train_finalpass\": trainpairs,\n            \"train_cleanpass\": list(\n                map(\n                    lambda x: x.replace(\"frames_finalpass\", \"frames_cleanpass\"),\n                    trainpairs,\n                )\n            ),\n            \"test_finalpass\": testpairs,\n            \"test_cleanpass\": list(\n                map(\n                    lambda x: x.replace(\"frames_finalpass\", \"frames_cleanpass\"),\n                    testpairs,\n                )\n            ),\n            \"test1of100_finalpass\": test1of100pairs,\n            \"test1of100_cleanpass\": list(\n                map(\n                    lambda x: x.replace(\"frames_finalpass\", \"frames_cleanpass\"),\n                    test1of100pairs,\n                )\n            ),\n        }\n        tosave[\"train_allpass\"] = tosave[\"train_finalpass\"] + tosave[\"train_cleanpass\"]\n        tosave[\"test_allpass\"] = tosave[\"test_finalpass\"] + tosave[\"test_cleanpass\"]\n        return tosave\n\n\nclass Md21Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2021\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname.replace(\"/im0\", \"/im1\")\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname.split(\"/\")[0], \"disp0.pfm\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_middlebury_disp\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        trainpairs = []\n        for s in seqs:\n            # trainpairs += [s+'/im0.png'] # we should remove it, it is included as such in other lightings\n            trainpairs += [\n                s + \"/ambient/\" + b + \"/\" + a\n                for b in sorted(os.listdir(osp.join(self.root, s, \"ambient\")))\n                for a in sorted(os.listdir(osp.join(self.root, s, \"ambient\", b)))\n                if a.startswith(\"im0\")\n            ]\n        assert len(trainpairs) == 355\n        subtrainpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in seqs[:-2])\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in seqs[-2:])\n        ]\n        assert (\n            len(subtrainpairs) == 335 and len(subvalpairs) == 20\n        ), \"incorrect parsing of pairs in Middlebury 2021\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass Md14Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2014\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"im0.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"disp0.pfm\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_middlebury_disp\n        self.has_constant_resolution = False\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        trainpairs = []\n        for s in seqs:\n            trainpairs += [s + \"/im1.png\", s + \"/im1E.png\", s + \"/im1L.png\"]\n        assert len(trainpairs) == 138\n        valseqs = [\"Umbrella-imperfect\", \"Vintage-perfect\"]\n        assert all(s in seqs for s in valseqs)\n        subtrainpairs = [\n            p for p in trainpairs if not any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        assert (\n            len(subtrainpairs) == 132 and len(subvalpairs) == 6\n        ), \"incorrect parsing of pairs in Middlebury 2014\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass Md06Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2006\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"view5.png\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname.split(\"/\")[0], \"disp1.png\"\n        )\n        self.load_disparity = _read_middlebury20052006_disp\n        self.has_constant_resolution = False\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        trainpairs = []\n        for s in seqs:\n            for i in [\"Illum1\", \"Illum2\", \"Illum3\"]:\n                for e in [\"Exp0\", \"Exp1\", \"Exp2\"]:\n                    trainpairs.append(osp.join(s, i, e, \"view1.png\"))\n        assert len(trainpairs) == 189\n        valseqs = [\"Rocks1\", \"Wood2\"]\n        assert all(s in seqs for s in valseqs)\n        subtrainpairs = [\n            p for p in trainpairs if not any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        assert (\n            len(subtrainpairs) == 171 and len(subvalpairs) == 18\n        ), \"incorrect parsing of pairs in Middlebury 2006\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass Md05Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2005\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"view5.png\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname.split(\"/\")[0], \"disp1.png\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_middlebury20052006_disp\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        trainpairs = []\n        for s in seqs:\n            for i in [\"Illum1\", \"Illum2\", \"Illum3\"]:\n                for e in [\"Exp0\", \"Exp1\", \"Exp2\"]:\n                    trainpairs.append(osp.join(s, i, e, \"view1.png\"))\n        assert len(trainpairs) == 54, \"incorrect parsing of pairs in Middlebury 2005\"\n        valseqs = [\"Reindeer\"]\n        assert all(s in seqs for s in valseqs)\n        subtrainpairs = [\n            p for p in trainpairs if not any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        assert (\n            len(subtrainpairs) == 45 and len(subvalpairs) == 9\n        ), \"incorrect parsing of pairs in Middlebury 2005\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass MdEval3Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"MiddleburyEval3\"\n        self._set_root()\n        assert self.split in [\n            s + \"_\" + r\n            for s in [\"train\", \"subtrain\", \"subval\", \"test\", \"all\"]\n            for r in [\"full\", \"half\", \"quarter\"]\n        ]\n        if self.split.endswith(\"_full\"):\n            self.root = self.root.replace(\"/MiddEval3\", \"/MiddEval3_F\")\n        elif self.split.endswith(\"_half\"):\n            self.root = self.root.replace(\"/MiddEval3\", \"/MiddEval3_H\")\n        else:\n            assert self.split.endswith(\"_quarter\")\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname, \"im0.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname, \"im1.png\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: (\n            None\n            if pairname.startswith(\"test\")\n            else osp.join(self.root, pairname, \"disp0GT.pfm\")\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_disparity = _read_middlebury_disp\n        # for submission only\n        self.submission_methodname = \"CroCo-Stereo\"\n        self.submission_sresolution = (\n            \"F\"\n            if self.split.endswith(\"_full\")\n            else (\"H\" if self.split.endswith(\"_half\") else \"Q\")\n        )\n\n    def _build_cache(self):\n        trainpairs = [\"train/\" + s for s in sorted(os.listdir(self.root + \"train/\"))]\n        testpairs = [\"test/\" + s for s in sorted(os.listdir(self.root + \"test/\"))]\n        subvalpairs = trainpairs[-1:]\n        subtrainpairs = trainpairs[:-1]\n        allpairs = trainpairs + testpairs\n        assert (\n            len(trainpairs) == 15\n            and len(testpairs) == 15\n            and len(subvalpairs) == 1\n            and len(subtrainpairs) == 14\n            and len(allpairs) == 30\n        ), \"incorrect parsing of pairs in Middlebury Eval v3\"\n        tosave = {}\n        for r in [\"full\", \"half\", \"quarter\"]:\n            tosave.update(\n                **{\n                    \"train_\" + r: trainpairs,\n                    \"subtrain_\" + r: subtrainpairs,\n                    \"subval_\" + r: subvalpairs,\n                    \"test_\" + r: testpairs,\n                    \"all_\" + r: allpairs,\n                }\n            )\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(\n            outdir,\n            pairname.split(\"/\")[0].replace(\"train\", \"training\")\n            + self.submission_sresolution,\n            pairname.split(\"/\")[1],\n            \"disp0\" + self.submission_methodname + \".pfm\",\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writePFM(outfile, prediction)\n        timefile = os.path.join(\n            os.path.dirname(outfile), \"time\" + self.submission_methodname + \".txt\"\n        )\n        with open(timefile, \"w\") as fid:\n            fid.write(str(time))\n\n    def finalize_submission(self, outdir):\n        cmd = f'cd {outdir}/; zip -r \"{self.submission_methodname}.zip\" .'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/{self.submission_methodname}.zip\")\n\n\nclass ETH3DLowResDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"ETH3DLowRes\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\", \"subtrain\", \"subval\", \"all\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname, \"im0.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname, \"im1.png\"\n        )\n        self.pairname_to_Ldispname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: (\n                None\n                if pairname.startswith(\"test/\")\n                else osp.join(\n                    self.root, pairname.replace(\"train/\", \"train_gt/\"), \"disp0GT.pfm\"\n                )\n# ... truncated ...","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.img_to_tensor","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.img_to_tensor#L45-L48","kind":"function","name":"img_to_tensor","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":45,"end_line":48,"context_start_line":25,"context_end_line":68,"code":"    \"CREStereo\": \"./data/stereoflow//crenet_stereo_trainset/stereo_trainset/crestereo/\",\n    \"SceneFlow\": \"./data/stereoflow//SceneFlow/\",\n    \"ETH3DLowRes\": \"./data/stereoflow/eth3d_lowres/\",\n    \"Booster\": \"./data/stereoflow/booster_gt/\",\n    \"Middlebury2021\": \"./data/stereoflow/middlebury/2021/data/\",\n    \"Middlebury2014\": \"./data/stereoflow/middlebury/2014/\",\n    \"Middlebury2006\": \"./data/stereoflow/middlebury/2006/\",\n    \"Middlebury2005\": \"./data/stereoflow/middlebury/2005/train/\",\n    \"MiddleburyEval3\": \"./data/stereoflow/middlebury/MiddEval3/\",\n    \"Spring\": \"./data/stereoflow/spring/\",\n    \"Kitti15\": \"./data/stereoflow/kitti-stereo-2015/\",\n    \"Kitti12\": \"./data/stereoflow/kitti-stereo-2012/\",\n}\ncache_dir = \"./data/stereoflow/datasets_stereo_cache/\"\n\n\nin1k_mean = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1)\nin1k_std = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1)\n\n\ndef img_to_tensor(img):\n    img = torch.from_numpy(img).permute(2, 0, 1).float() / 255.0\n    img = (img - in1k_mean) / in1k_std\n    return img\n\n\ndef disp_to_tensor(disp):\n    return torch.from_numpy(disp)[None, :, :]\n\n\nclass StereoDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = StereoAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.disp_to_tensor","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.disp_to_tensor#L51-L52","kind":"function","name":"disp_to_tensor","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":51,"end_line":52,"context_start_line":31,"context_end_line":72,"code":"    \"Middlebury2006\": \"./data/stereoflow/middlebury/2006/\",\n    \"Middlebury2005\": \"./data/stereoflow/middlebury/2005/train/\",\n    \"MiddleburyEval3\": \"./data/stereoflow/middlebury/MiddEval3/\",\n    \"Spring\": \"./data/stereoflow/spring/\",\n    \"Kitti15\": \"./data/stereoflow/kitti-stereo-2015/\",\n    \"Kitti12\": \"./data/stereoflow/kitti-stereo-2012/\",\n}\ncache_dir = \"./data/stereoflow/datasets_stereo_cache/\"\n\n\nin1k_mean = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1)\nin1k_std = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1)\n\n\ndef img_to_tensor(img):\n    img = torch.from_numpy(img).permute(2, 0, 1).float() / 255.0\n    img = (img - in1k_mean) / in1k_std\n    return img\n\n\ndef disp_to_tensor(disp):\n    return torch.from_numpy(disp)[None, :, :]\n\n\nclass StereoDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = StereoAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.StereoDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.StereoDataset#L55-L154","kind":"class","name":"StereoDataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":55,"end_line":154,"context_start_line":35,"context_end_line":174,"code":"    \"Kitti15\": \"./data/stereoflow/kitti-stereo-2015/\",\n    \"Kitti12\": \"./data/stereoflow/kitti-stereo-2012/\",\n}\ncache_dir = \"./data/stereoflow/datasets_stereo_cache/\"\n\n\nin1k_mean = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1)\nin1k_std = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1)\n\n\ndef img_to_tensor(img):\n    img = torch.from_numpy(img).permute(2, 0, 1).float() / 255.0\n    img = (img - in1k_mean) / in1k_std\n    return img\n\n\ndef disp_to_tensor(disp):\n    return torch.from_numpy(disp)[None, :, :]\n\n\nclass StereoDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = StereoAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(self.pairnames)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        Limgname = self.pairname_to_Limgname(pairname)\n        Rimgname = self.pairname_to_Rimgname(pairname)\n        Ldispname = (\n            self.pairname_to_Ldispname(pairname)\n            if self.pairname_to_Ldispname is not None\n            else None\n        )\n\n        # load images and disparities\n        Limg = _read_img(Limgname)\n        Rimg = _read_img(Rimgname)\n        disp = self.load_disparity(Ldispname) if Ldispname is not None else None\n\n        # sanity check\n        if disp is not None:\n            assert np.all(disp > 0) or self.name == \"Spring\", (\n                self.name,\n                pairname,\n                Ldispname,\n            )\n\n        # apply augmentations\n        if self.augmentor is not None:\n            Limg, Rimg, disp = self.augmentor(Limg, Rimg, disp, self.name)\n\n        if self.totensor:\n            Limg = img_to_tensor(Limg)\n            Rimg = img_to_tensor(Rimg)\n            if disp is None:\n                disp = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            else:\n                disp = disp_to_tensor(disp)\n\n        return Limg, Rimg, disp, str(pairname)\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass CREStereoDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"CREStereo\"\n        self._set_root()\n        assert self.split in [\"train\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_left.jpg\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname + \"_right.jpg\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname + \"_left.disp.png\"\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_disparity = _read_crestereo_disp\n","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.CREStereoDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.CREStereoDataset#L157-L184","kind":"class","name":"CREStereoDataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":157,"end_line":184,"context_start_line":137,"context_end_line":204,"code":"\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass CREStereoDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"CREStereo\"\n        self._set_root()\n        assert self.split in [\"train\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_left.jpg\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname + \"_right.jpg\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname + \"_left.disp.png\"\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_disparity = _read_crestereo_disp\n\n    def _build_cache(self):\n        allpairs = [\n            s + \"/\" + f[: -len(\"_left.jpg\")]\n            for s in sorted(os.listdir(self.root))\n            for f in sorted(os.listdir(self.root + \"/\" + s))\n            if f.endswith(\"_left.jpg\")\n        ]\n        assert len(allpairs) == 200000, \"incorrect parsing of pairs in CreStereo\"\n        tosave = {\"train\": allpairs}\n        return tosave\n\n\nclass SceneFlowDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"SceneFlow\"\n        self._set_root()\n        assert self.split in [\n            \"train_finalpass\",\n            \"train_cleanpass\",\n            \"train_allpass\",\n            \"test_finalpass\",\n            \"test_cleanpass\",\n            \"test_allpass\",\n            \"test1of100_cleanpass\",\n            \"test1of100_finalpass\",\n        ]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.SceneFlowDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.SceneFlowDataset#L187-L268","kind":"class","name":"SceneFlowDataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":187,"end_line":268,"context_start_line":167,"context_end_line":288,"code":"            self.root, pairname + \"_right.jpg\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname + \"_left.disp.png\"\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_disparity = _read_crestereo_disp\n\n    def _build_cache(self):\n        allpairs = [\n            s + \"/\" + f[: -len(\"_left.jpg\")]\n            for s in sorted(os.listdir(self.root))\n            for f in sorted(os.listdir(self.root + \"/\" + s))\n            if f.endswith(\"_left.jpg\")\n        ]\n        assert len(allpairs) == 200000, \"incorrect parsing of pairs in CreStereo\"\n        tosave = {\"train\": allpairs}\n        return tosave\n\n\nclass SceneFlowDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"SceneFlow\"\n        self._set_root()\n        assert self.split in [\n            \"train_finalpass\",\n            \"train_cleanpass\",\n            \"train_allpass\",\n            \"test_finalpass\",\n            \"test_cleanpass\",\n            \"test_allpass\",\n            \"test1of100_cleanpass\",\n            \"test1of100_finalpass\",\n        ]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname\n        ).replace(\"/left/\", \"/right/\")\n        self.pairname_to_Ldispname = (\n            lambda pairname: osp.join(self.root, pairname)\n            .replace(\"/frames_finalpass/\", \"/disparity/\")\n            .replace(\"/frames_cleanpass/\", \"/disparity/\")[:-4]\n            + \".pfm\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_sceneflow_disp\n\n    def _build_cache(self):\n        trainpairs = []\n        # driving\n        pairs = sorted(glob(self.root + \"Driving/frames_finalpass/*/*/*/left/*.png\"))\n        pairs = list(map(lambda x: x[len(self.root) :], pairs))\n        assert len(pairs) == 4400, \"incorrect parsing of pairs in SceneFlow\"\n        trainpairs += pairs\n        # monkaa\n        pairs = sorted(glob(self.root + \"Monkaa/frames_finalpass/*/left/*.png\"))\n        pairs = list(map(lambda x: x[len(self.root) :], pairs))\n        assert len(pairs) == 8664, \"incorrect parsing of pairs in SceneFlow\"\n        trainpairs += pairs\n        # flyingthings\n        pairs = sorted(\n            glob(self.root + \"FlyingThings/frames_finalpass/TRAIN/*/*/left/*.png\")\n        )\n        pairs = list(map(lambda x: x[len(self.root) :], pairs))\n        assert len(pairs) == 22390, \"incorrect parsing of pairs in SceneFlow\"\n        trainpairs += pairs\n        assert len(trainpairs) == 35454, \"incorrect parsing of pairs in SceneFlow\"\n        testpairs = sorted(\n            glob(self.root + \"FlyingThings/frames_finalpass/TEST/*/*/left/*.png\")\n        )\n        testpairs = list(map(lambda x: x[len(self.root) :], testpairs))\n        assert len(testpairs) == 4370, \"incorrect parsing of pairs in SceneFlow\"\n        test1of100pairs = testpairs[::100]\n        assert len(test1of100pairs) == 44, \"incorrect parsing of pairs in SceneFlow\"\n        # all\n        tosave = {\n            \"train_finalpass\": trainpairs,\n            \"train_cleanpass\": list(\n                map(\n                    lambda x: x.replace(\"frames_finalpass\", \"frames_cleanpass\"),\n                    trainpairs,\n                )\n            ),\n            \"test_finalpass\": testpairs,\n            \"test_cleanpass\": list(\n                map(\n                    lambda x: x.replace(\"frames_finalpass\", \"frames_cleanpass\"),\n                    testpairs,\n                )\n            ),\n            \"test1of100_finalpass\": test1of100pairs,\n            \"test1of100_cleanpass\": list(\n                map(\n                    lambda x: x.replace(\"frames_finalpass\", \"frames_cleanpass\"),\n                    test1of100pairs,\n                )\n            ),\n        }\n        tosave[\"train_allpass\"] = tosave[\"train_finalpass\"] + tosave[\"train_cleanpass\"]\n        tosave[\"test_allpass\"] = tosave[\"test_finalpass\"] + tosave[\"test_cleanpass\"]\n        return tosave\n\n\nclass Md21Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2021\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname.replace(\"/im0\", \"/im1\")\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname.split(\"/\")[0], \"disp0.pfm\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_middlebury_disp\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.Md21Dataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.Md21Dataset#L271-L309","kind":"class","name":"Md21Dataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":271,"end_line":309,"context_start_line":251,"context_end_line":329,"code":"            \"test_finalpass\": testpairs,\n            \"test_cleanpass\": list(\n                map(\n                    lambda x: x.replace(\"frames_finalpass\", \"frames_cleanpass\"),\n                    testpairs,\n                )\n            ),\n            \"test1of100_finalpass\": test1of100pairs,\n            \"test1of100_cleanpass\": list(\n                map(\n                    lambda x: x.replace(\"frames_finalpass\", \"frames_cleanpass\"),\n                    test1of100pairs,\n                )\n            ),\n        }\n        tosave[\"train_allpass\"] = tosave[\"train_finalpass\"] + tosave[\"train_cleanpass\"]\n        tosave[\"test_allpass\"] = tosave[\"test_finalpass\"] + tosave[\"test_cleanpass\"]\n        return tosave\n\n\nclass Md21Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2021\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname.replace(\"/im0\", \"/im1\")\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname.split(\"/\")[0], \"disp0.pfm\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_middlebury_disp\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        trainpairs = []\n        for s in seqs:\n            # trainpairs += [s+'/im0.png'] # we should remove it, it is included as such in other lightings\n            trainpairs += [\n                s + \"/ambient/\" + b + \"/\" + a\n                for b in sorted(os.listdir(osp.join(self.root, s, \"ambient\")))\n                for a in sorted(os.listdir(osp.join(self.root, s, \"ambient\", b)))\n                if a.startswith(\"im0\")\n            ]\n        assert len(trainpairs) == 355\n        subtrainpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in seqs[:-2])\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in seqs[-2:])\n        ]\n        assert (\n            len(subtrainpairs) == 335 and len(subvalpairs) == 20\n        ), \"incorrect parsing of pairs in Middlebury 2021\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass Md14Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2014\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"im0.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"disp0.pfm\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_middlebury_disp\n        self.has_constant_resolution = False\n\n    def _build_cache(self):","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.Md14Dataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.Md14Dataset#L312-L347","kind":"class","name":"Md14Dataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":312,"end_line":347,"context_start_line":292,"context_end_line":367,"code":"            trainpairs += [\n                s + \"/ambient/\" + b + \"/\" + a\n                for b in sorted(os.listdir(osp.join(self.root, s, \"ambient\")))\n                for a in sorted(os.listdir(osp.join(self.root, s, \"ambient\", b)))\n                if a.startswith(\"im0\")\n            ]\n        assert len(trainpairs) == 355\n        subtrainpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in seqs[:-2])\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in seqs[-2:])\n        ]\n        assert (\n            len(subtrainpairs) == 335 and len(subvalpairs) == 20\n        ), \"incorrect parsing of pairs in Middlebury 2021\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass Md14Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2014\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"im0.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"disp0.pfm\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_middlebury_disp\n        self.has_constant_resolution = False\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        trainpairs = []\n        for s in seqs:\n            trainpairs += [s + \"/im1.png\", s + \"/im1E.png\", s + \"/im1L.png\"]\n        assert len(trainpairs) == 138\n        valseqs = [\"Umbrella-imperfect\", \"Vintage-perfect\"]\n        assert all(s in seqs for s in valseqs)\n        subtrainpairs = [\n            p for p in trainpairs if not any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        assert (\n            len(subtrainpairs) == 132 and len(subvalpairs) == 6\n        ), \"incorrect parsing of pairs in Middlebury 2014\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass Md06Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2006\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"view5.png\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname.split(\"/\")[0], \"disp1.png\"\n        )\n        self.load_disparity = _read_middlebury20052006_disp\n        self.has_constant_resolution = False\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.Md06Dataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.Md06Dataset#L350-L386","kind":"class","name":"Md06Dataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":350,"end_line":386,"context_start_line":330,"context_end_line":406,"code":"        seqs = sorted(os.listdir(self.root))\n        trainpairs = []\n        for s in seqs:\n            trainpairs += [s + \"/im1.png\", s + \"/im1E.png\", s + \"/im1L.png\"]\n        assert len(trainpairs) == 138\n        valseqs = [\"Umbrella-imperfect\", \"Vintage-perfect\"]\n        assert all(s in seqs for s in valseqs)\n        subtrainpairs = [\n            p for p in trainpairs if not any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        assert (\n            len(subtrainpairs) == 132 and len(subvalpairs) == 6\n        ), \"incorrect parsing of pairs in Middlebury 2014\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass Md06Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2006\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"view5.png\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname.split(\"/\")[0], \"disp1.png\"\n        )\n        self.load_disparity = _read_middlebury20052006_disp\n        self.has_constant_resolution = False\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        trainpairs = []\n        for s in seqs:\n            for i in [\"Illum1\", \"Illum2\", \"Illum3\"]:\n                for e in [\"Exp0\", \"Exp1\", \"Exp2\"]:\n                    trainpairs.append(osp.join(s, i, e, \"view1.png\"))\n        assert len(trainpairs) == 189\n        valseqs = [\"Rocks1\", \"Wood2\"]\n        assert all(s in seqs for s in valseqs)\n        subtrainpairs = [\n            p for p in trainpairs if not any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        assert (\n            len(subtrainpairs) == 171 and len(subvalpairs) == 18\n        ), \"incorrect parsing of pairs in Middlebury 2006\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass Md05Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2005\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"view5.png\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname.split(\"/\")[0], \"disp1.png\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_middlebury20052006_disp\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.Md05Dataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.Md05Dataset#L389-L425","kind":"class","name":"Md05Dataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":389,"end_line":425,"context_start_line":369,"context_end_line":445,"code":"        for s in seqs:\n            for i in [\"Illum1\", \"Illum2\", \"Illum3\"]:\n                for e in [\"Exp0\", \"Exp1\", \"Exp2\"]:\n                    trainpairs.append(osp.join(s, i, e, \"view1.png\"))\n        assert len(trainpairs) == 189\n        valseqs = [\"Rocks1\", \"Wood2\"]\n        assert all(s in seqs for s in valseqs)\n        subtrainpairs = [\n            p for p in trainpairs if not any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        assert (\n            len(subtrainpairs) == 171 and len(subvalpairs) == 18\n        ), \"incorrect parsing of pairs in Middlebury 2006\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass Md05Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Middlebury2005\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"view5.png\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname.split(\"/\")[0], \"disp1.png\"\n        )\n        self.pairname_to_str = lambda pairname: pairname[:-4]\n        self.load_disparity = _read_middlebury20052006_disp\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        trainpairs = []\n        for s in seqs:\n            for i in [\"Illum1\", \"Illum2\", \"Illum3\"]:\n                for e in [\"Exp0\", \"Exp1\", \"Exp2\"]:\n                    trainpairs.append(osp.join(s, i, e, \"view1.png\"))\n        assert len(trainpairs) == 54, \"incorrect parsing of pairs in Middlebury 2005\"\n        valseqs = [\"Reindeer\"]\n        assert all(s in seqs for s in valseqs)\n        subtrainpairs = [\n            p for p in trainpairs if not any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        assert (\n            len(subtrainpairs) == 45 and len(subvalpairs) == 9\n        ), \"incorrect parsing of pairs in Middlebury 2005\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass MdEval3Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"MiddleburyEval3\"\n        self._set_root()\n        assert self.split in [\n            s + \"_\" + r\n            for s in [\"train\", \"subtrain\", \"subval\", \"test\", \"all\"]\n            for r in [\"full\", \"half\", \"quarter\"]\n        ]\n        if self.split.endswith(\"_full\"):\n            self.root = self.root.replace(\"/MiddEval3\", \"/MiddEval3_F\")\n        elif self.split.endswith(\"_half\"):\n            self.root = self.root.replace(\"/MiddEval3\", \"/MiddEval3_H\")\n        else:\n            assert self.split.endswith(\"_quarter\")\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname, \"im0.png\"","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.MdEval3Dataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.MdEval3Dataset#L428-L513","kind":"class","name":"MdEval3Dataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":428,"end_line":513,"context_start_line":408,"context_end_line":533,"code":"        for s in seqs:\n            for i in [\"Illum1\", \"Illum2\", \"Illum3\"]:\n                for e in [\"Exp0\", \"Exp1\", \"Exp2\"]:\n                    trainpairs.append(osp.join(s, i, e, \"view1.png\"))\n        assert len(trainpairs) == 54, \"incorrect parsing of pairs in Middlebury 2005\"\n        valseqs = [\"Reindeer\"]\n        assert all(s in seqs for s in valseqs)\n        subtrainpairs = [\n            p for p in trainpairs if not any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        subvalpairs = [\n            p for p in trainpairs if any(p.startswith(s + \"/\") for s in valseqs)\n        ]\n        assert (\n            len(subtrainpairs) == 45 and len(subvalpairs) == 9\n        ), \"incorrect parsing of pairs in Middlebury 2005\"\n        tosave = {\"train\": trainpairs, \"subtrain\": subtrainpairs, \"subval\": subvalpairs}\n        return tosave\n\n\nclass MdEval3Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"MiddleburyEval3\"\n        self._set_root()\n        assert self.split in [\n            s + \"_\" + r\n            for s in [\"train\", \"subtrain\", \"subval\", \"test\", \"all\"]\n            for r in [\"full\", \"half\", \"quarter\"]\n        ]\n        if self.split.endswith(\"_full\"):\n            self.root = self.root.replace(\"/MiddEval3\", \"/MiddEval3_F\")\n        elif self.split.endswith(\"_half\"):\n            self.root = self.root.replace(\"/MiddEval3\", \"/MiddEval3_H\")\n        else:\n            assert self.split.endswith(\"_quarter\")\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname, \"im0.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname, \"im1.png\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: (\n            None\n            if pairname.startswith(\"test\")\n            else osp.join(self.root, pairname, \"disp0GT.pfm\")\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_disparity = _read_middlebury_disp\n        # for submission only\n        self.submission_methodname = \"CroCo-Stereo\"\n        self.submission_sresolution = (\n            \"F\"\n            if self.split.endswith(\"_full\")\n            else (\"H\" if self.split.endswith(\"_half\") else \"Q\")\n        )\n\n    def _build_cache(self):\n        trainpairs = [\"train/\" + s for s in sorted(os.listdir(self.root + \"train/\"))]\n        testpairs = [\"test/\" + s for s in sorted(os.listdir(self.root + \"test/\"))]\n        subvalpairs = trainpairs[-1:]\n        subtrainpairs = trainpairs[:-1]\n        allpairs = trainpairs + testpairs\n        assert (\n            len(trainpairs) == 15\n            and len(testpairs) == 15\n            and len(subvalpairs) == 1\n            and len(subtrainpairs) == 14\n            and len(allpairs) == 30\n        ), \"incorrect parsing of pairs in Middlebury Eval v3\"\n        tosave = {}\n        for r in [\"full\", \"half\", \"quarter\"]:\n            tosave.update(\n                **{\n                    \"train_\" + r: trainpairs,\n                    \"subtrain_\" + r: subtrainpairs,\n                    \"subval_\" + r: subvalpairs,\n                    \"test_\" + r: testpairs,\n                    \"all_\" + r: allpairs,\n                }\n            )\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(\n            outdir,\n            pairname.split(\"/\")[0].replace(\"train\", \"training\")\n            + self.submission_sresolution,\n            pairname.split(\"/\")[1],\n            \"disp0\" + self.submission_methodname + \".pfm\",\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writePFM(outfile, prediction)\n        timefile = os.path.join(\n            os.path.dirname(outfile), \"time\" + self.submission_methodname + \".txt\"\n        )\n        with open(timefile, \"w\") as fid:\n            fid.write(str(time))\n\n    def finalize_submission(self, outdir):\n        cmd = f'cd {outdir}/; zip -r \"{self.submission_methodname}.zip\" .'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/{self.submission_methodname}.zip\")\n\n\nclass ETH3DLowResDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"ETH3DLowRes\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\", \"subtrain\", \"subval\", \"all\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname, \"im0.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname, \"im1.png\"\n        )\n        self.pairname_to_Ldispname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: (\n                None\n                if pairname.startswith(\"test/\")","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.ETH3DLowResDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.ETH3DLowResDataset#L516-L584","kind":"class","name":"ETH3DLowResDataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":516,"end_line":584,"context_start_line":496,"context_end_line":604,"code":"            pairname.split(\"/\")[0].replace(\"train\", \"training\")\n            + self.submission_sresolution,\n            pairname.split(\"/\")[1],\n            \"disp0\" + self.submission_methodname + \".pfm\",\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writePFM(outfile, prediction)\n        timefile = os.path.join(\n            os.path.dirname(outfile), \"time\" + self.submission_methodname + \".txt\"\n        )\n        with open(timefile, \"w\") as fid:\n            fid.write(str(time))\n\n    def finalize_submission(self, outdir):\n        cmd = f'cd {outdir}/; zip -r \"{self.submission_methodname}.zip\" .'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/{self.submission_methodname}.zip\")\n\n\nclass ETH3DLowResDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"ETH3DLowRes\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\", \"subtrain\", \"subval\", \"all\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname, \"im0.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname, \"im1.png\"\n        )\n        self.pairname_to_Ldispname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: (\n                None\n                if pairname.startswith(\"test/\")\n                else osp.join(\n                    self.root, pairname.replace(\"train/\", \"train_gt/\"), \"disp0GT.pfm\"\n                )\n            )\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_disparity = _read_eth3d_disp\n        self.has_constant_resolution = False\n\n    def _build_cache(self):\n        trainpairs = [\"train/\" + s for s in sorted(os.listdir(self.root + \"train/\"))]\n        testpairs = [\"test/\" + s for s in sorted(os.listdir(self.root + \"test/\"))]\n        assert (\n            len(trainpairs) == 27 and len(testpairs) == 20\n        ), \"incorrect parsing of pairs in ETH3D Low Res\"\n        subvalpairs = [\n            \"train/delivery_area_3s\",\n            \"train/electro_3l\",\n            \"train/playground_3l\",\n        ]\n        assert all(p in trainpairs for p in subvalpairs)\n        subtrainpairs = [p for p in trainpairs if not p in subvalpairs]\n        assert (\n            len(subvalpairs) == 3 and len(subtrainpairs) == 24\n        ), \"incorrect parsing of pairs in ETH3D Low Res\"\n        tosave = {\n            \"train\": trainpairs,\n            \"test\": testpairs,\n            \"subtrain\": subtrainpairs,\n            \"subval\": subvalpairs,\n            \"all\": trainpairs + testpairs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(\n            outdir, \"low_res_two_view\", pairname.split(\"/\")[1] + \".pfm\"\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writePFM(outfile, prediction)\n        timefile = outfile[:-4] + \".txt\"\n        with open(timefile, \"w\") as fid:\n            fid.write(\"runtime \" + str(time))\n\n    def finalize_submission(self, outdir):\n        cmd = f'cd {outdir}/; zip -r \"eth3d_low_res_two_view_results.zip\" low_res_two_view'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/eth3d_low_res_two_view_results.zip\")\n\n\nclass BoosterDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Booster\"\n        self._set_root()\n        assert self.split in [\n            \"train_balanced\",\n            \"test_balanced\",\n            \"subtrain_balanced\",\n            \"subval_balanced\",\n        ]  # we use only the balanced version\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname\n        ).replace(\"/camera_00/\", \"/camera_02/\")\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"../disp_00.npy\"\n        )  # same images with different colors, same gt per sequence","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.BoosterDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.BoosterDataset#L587-L636","kind":"class","name":"BoosterDataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":587,"end_line":636,"context_start_line":567,"context_end_line":656,"code":"\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(\n            outdir, \"low_res_two_view\", pairname.split(\"/\")[1] + \".pfm\"\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writePFM(outfile, prediction)\n        timefile = outfile[:-4] + \".txt\"\n        with open(timefile, \"w\") as fid:\n            fid.write(\"runtime \" + str(time))\n\n    def finalize_submission(self, outdir):\n        cmd = f'cd {outdir}/; zip -r \"eth3d_low_res_two_view_results.zip\" low_res_two_view'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/eth3d_low_res_two_view_results.zip\")\n\n\nclass BoosterDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Booster\"\n        self._set_root()\n        assert self.split in [\n            \"train_balanced\",\n            \"test_balanced\",\n            \"subtrain_balanced\",\n            \"subval_balanced\",\n        ]  # we use only the balanced version\n        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname\n        ).replace(\"/camera_00/\", \"/camera_02/\")\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, osp.dirname(pairname), \"../disp_00.npy\"\n        )  # same images with different colors, same gt per sequence\n        self.pairname_to_str = lambda pairname: pairname[:-4].replace(\n            \"/camera_00/\", \"/\"\n        )\n        self.load_disparity = _read_booster_disp\n\n    def _build_cache(self):\n        trainseqs = sorted(os.listdir(self.root + \"train/balanced\"))\n        trainpairs = [\n            \"train/balanced/\" + s + \"/camera_00/\" + imname\n            for s in trainseqs\n            for imname in sorted(\n                os.listdir(self.root + \"train/balanced/\" + s + \"/camera_00/\")\n            )\n        ]\n        testpairs = [\n            \"test/balanced/\" + s + \"/camera_00/\" + imname\n            for s in sorted(os.listdir(self.root + \"test/balanced\"))\n            for imname in sorted(\n                os.listdir(self.root + \"test/balanced/\" + s + \"/camera_00/\")\n            )\n        ]\n        assert len(trainpairs) == 228 and len(testpairs) == 191\n        subtrainpairs = [p for p in trainpairs if any(s in p for s in trainseqs[:-2])]\n        subvalpairs = [p for p in trainpairs if any(s in p for s in trainseqs[-2:])]\n        # warning: if we do validation split, we should split scenes!!!\n        tosave = {\n            \"train_balanced\": trainpairs,\n            \"test_balanced\": testpairs,\n            \"subtrain_balanced\": subtrainpairs,\n            \"subval_balanced\": subvalpairs,\n        }\n        return tosave\n\n\nclass SpringDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Spring\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \".png\"\n        )\n        self.pairname_to_Rimgname = (\n            lambda pairname: osp.join(self.root, pairname + \".png\")\n            .replace(\"frame_right\", \"<frame_right>\")\n            .replace(\"frame_left\", \"frame_right\")\n            .replace(\"<frame_right>\", \"frame_left\")\n        )\n        self.pairname_to_Ldispname = lambda pairname: (\n            None\n            if pairname.startswith(\"test\")","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.SpringDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.SpringDataset#L639-L716","kind":"class","name":"SpringDataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":639,"end_line":716,"context_start_line":619,"context_end_line":736,"code":"        testpairs = [\n            \"test/balanced/\" + s + \"/camera_00/\" + imname\n            for s in sorted(os.listdir(self.root + \"test/balanced\"))\n            for imname in sorted(\n                os.listdir(self.root + \"test/balanced/\" + s + \"/camera_00/\")\n            )\n        ]\n        assert len(trainpairs) == 228 and len(testpairs) == 191\n        subtrainpairs = [p for p in trainpairs if any(s in p for s in trainseqs[:-2])]\n        subvalpairs = [p for p in trainpairs if any(s in p for s in trainseqs[-2:])]\n        # warning: if we do validation split, we should split scenes!!!\n        tosave = {\n            \"train_balanced\": trainpairs,\n            \"test_balanced\": testpairs,\n            \"subtrain_balanced\": subtrainpairs,\n            \"subval_balanced\": subvalpairs,\n        }\n        return tosave\n\n\nclass SpringDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Spring\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\", \"subtrain\", \"subval\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \".png\"\n        )\n        self.pairname_to_Rimgname = (\n            lambda pairname: osp.join(self.root, pairname + \".png\")\n            .replace(\"frame_right\", \"<frame_right>\")\n            .replace(\"frame_left\", \"frame_right\")\n            .replace(\"<frame_right>\", \"frame_left\")\n        )\n        self.pairname_to_Ldispname = lambda pairname: (\n            None\n            if pairname.startswith(\"test\")\n            else osp.join(self.root, pairname + \".dsp5\")\n            .replace(\"frame_left\", \"disp1_left\")\n            .replace(\"frame_right\", \"disp1_right\")\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_disparity = _read_hdf5_disp\n\n    def _build_cache(self):\n        trainseqs = sorted(os.listdir(osp.join(self.root, \"train\")))\n        trainpairs = [\n            osp.join(\"train\", s, \"frame_left\", f[:-4])\n            for s in trainseqs\n            for f in sorted(os.listdir(osp.join(self.root, \"train\", s, \"frame_left\")))\n        ]\n        testseqs = sorted(os.listdir(osp.join(self.root, \"test\")))\n        testpairs = [\n            osp.join(\"test\", s, \"frame_left\", f[:-4])\n            for s in testseqs\n            for f in sorted(os.listdir(osp.join(self.root, \"test\", s, \"frame_left\")))\n        ]\n        testpairs += [p.replace(\"frame_left\", \"frame_right\") for p in testpairs]\n        \"\"\"maxnorm = {'0001': 32.88, '0002': 228.5, '0004': 298.2, '0005': 142.5, '0006': 113.6, '0007': 27.3, '0008': 554.5, '0009': 155.6, '0010': 126.1, '0011': 87.6, '0012': 303.2, '0013': 24.14, '0014': 82.56, '0015': 98.44, '0016': 156.9, '0017': 28.17, '0018': 21.03, '0020': 178.0, '0021': 58.06, '0022': 354.2, '0023': 8.79, '0024': 97.06, '0025': 55.16, '0026': 91.9, '0027': 156.6, '0030': 200.4, '0032': 58.66, '0033': 373.5, '0036': 149.4, '0037': 5.625, '0038': 37.0, '0039': 12.2, '0041': 453.5, '0043': 457.0, '0044': 379.5, '0045': 161.8, '0047': 105.44} # => let'use 0041\"\"\"\n        subtrainpairs = [p for p in trainpairs if p.split(\"/\")[1] != \"0041\"]\n        subvalpairs = [p for p in trainpairs if p.split(\"/\")[1] == \"0041\"]\n        assert (\n            len(trainpairs) == 5000\n            and len(testpairs) == 2000\n            and len(subtrainpairs) == 4904\n            and len(subvalpairs) == 96\n        ), \"incorrect parsing of pairs in Spring\"\n        tosave = {\n            \"train\": trainpairs,\n            \"test\": testpairs,\n            \"subtrain\": subtrainpairs,\n            \"subval\": subvalpairs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = (\n            os.path.join(outdir, pairname + \".dsp5\")\n            .replace(\"frame_left\", \"disp1_left\")\n            .replace(\"frame_right\", \"disp1_right\")\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeDsp5File(prediction, outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        exe = \"{self.root}/disp1_subsampling\"\n        if os.path.isfile(exe):\n            cmd = f'cd \"{outdir}/test\"; {exe} .'\n            print(cmd)\n            os.system(cmd)\n        else:\n            print(\"Could not find disp1_subsampling executable for submission.\")\n            print(\"Please download it and run:\")\n            print(f'cd \"{outdir}/test\"; <disp1_subsampling_exe> .')\n\n\nclass Kitti12Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti12\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname.replace(\"/colored_0/\", \"/colored_1/\") + \"_10.png\"\n        )\n        self.pairname_to_Ldispname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/colored_0/\", \"/disp_occ/\") + \"_10.png\"\n            )","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.Kitti12Dataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.Kitti12Dataset#L719-L763","kind":"class","name":"Kitti12Dataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":719,"end_line":763,"context_start_line":699,"context_end_line":783,"code":"            os.path.join(outdir, pairname + \".dsp5\")\n            .replace(\"frame_left\", \"disp1_left\")\n            .replace(\"frame_right\", \"disp1_right\")\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeDsp5File(prediction, outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        exe = \"{self.root}/disp1_subsampling\"\n        if os.path.isfile(exe):\n            cmd = f'cd \"{outdir}/test\"; {exe} .'\n            print(cmd)\n            os.system(cmd)\n        else:\n            print(\"Could not find disp1_subsampling executable for submission.\")\n            print(\"Please download it and run:\")\n            print(f'cd \"{outdir}/test\"; <disp1_subsampling_exe> .')\n\n\nclass Kitti12Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti12\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname.replace(\"/colored_0/\", \"/colored_1/\") + \"_10.png\"\n        )\n        self.pairname_to_Ldispname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/colored_0/\", \"/disp_occ/\") + \"_10.png\"\n            )\n        )\n        self.pairname_to_str = lambda pairname: pairname.replace(\"/colored_0/\", \"/\")\n        self.load_disparity = _read_kitti_disp\n\n    def _build_cache(self):\n        trainseqs = [\"training/colored_0/%06d\" % (i) for i in range(194)]\n        testseqs = [\"testing/colored_0/%06d\" % (i) for i in range(195)]\n        assert (\n            len(trainseqs) == 194 and len(testseqs) == 195\n        ), \"incorrect parsing of pairs in Kitti12\"\n        tosave = {\"train\": trainseqs, \"test\": testseqs}\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(outdir, pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        img = (prediction * 256).astype(\"uint16\")\n        Image.fromarray(img).save(outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti12_results.zip\" .'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti12_results.zip\")\n\n\nclass Kitti15Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti15\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\", \"test\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname.replace(\"/image_2/\", \"/image_3/\") + \"_10.png\"\n        )\n        self.pairname_to_Ldispname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/image_2/\", \"/disp_occ_0/\") + \"_10.png\"\n            )","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.Kitti15Dataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_stereo.Kitti15Dataset#L766-L820","kind":"class","name":"Kitti15Dataset","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":766,"end_line":820,"context_start_line":746,"context_end_line":840,"code":"        ), \"incorrect parsing of pairs in Kitti12\"\n        tosave = {\"train\": trainseqs, \"test\": testseqs}\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(outdir, pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        img = (prediction * 256).astype(\"uint16\")\n        Image.fromarray(img).save(outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti12_results.zip\" .'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti12_results.zip\")\n\n\nclass Kitti15Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti15\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\", \"test\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname.replace(\"/image_2/\", \"/image_3/\") + \"_10.png\"\n        )\n        self.pairname_to_Ldispname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/image_2/\", \"/disp_occ_0/\") + \"_10.png\"\n            )\n        )\n        self.pairname_to_str = lambda pairname: pairname.replace(\"/image_2/\", \"/\")\n        self.load_disparity = _read_kitti_disp\n\n    def _build_cache(self):\n        trainseqs = [\"training/image_2/%06d\" % (i) for i in range(200)]\n        subtrainseqs = trainseqs[:-5]\n        subvalseqs = trainseqs[-5:]\n        testseqs = [\"testing/image_2/%06d\" % (i) for i in range(200)]\n        assert (\n            len(trainseqs) == 200\n            and len(subtrainseqs) == 195\n            and len(subvalseqs) == 5\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(outdir, \"disp_0\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        img = (prediction * 256).astype(\"uint16\")\n        Image.fromarray(img).save(outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_results.zip\" disp_0'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_results.zip\")\n\n\n### auxiliary functions\n\n\ndef _read_img(filename):\n    # convert to RGB for scene flow finalpass data\n    img = np.asarray(Image.open(filename).convert(\"RGB\"))\n    return img\n\n\ndef _read_booster_disp(filename):\n    disp = np.load(filename)\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_png_disp(filename, coef=1.0):\n    disp = np.asarray(Image.open(filename))\n    disp = disp.astype(np.float32) / coef","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_img","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_img#L826-L829","kind":"function","name":"_read_img","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":826,"end_line":829,"context_start_line":806,"context_end_line":849,"code":"\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(outdir, \"disp_0\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        img = (prediction * 256).astype(\"uint16\")\n        Image.fromarray(img).save(outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_results.zip\" disp_0'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_results.zip\")\n\n\n### auxiliary functions\n\n\ndef _read_img(filename):\n    # convert to RGB for scene flow finalpass data\n    img = np.asarray(Image.open(filename).convert(\"RGB\"))\n    return img\n\n\ndef _read_booster_disp(filename):\n    disp = np.load(filename)\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_png_disp(filename, coef=1.0):\n    disp = np.asarray(Image.open(filename))\n    disp = disp.astype(np.float32) / coef\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_pfm_disp(filename):\n    disp = np.ascontiguousarray(_read_pfm(filename)[0])\n    disp[disp <= 0] = (\n        np.inf\n    )  # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_booster_disp","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_booster_disp#L832-L835","kind":"function","name":"_read_booster_disp","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":832,"end_line":835,"context_start_line":812,"context_end_line":855,"code":"        img = (prediction * 256).astype(\"uint16\")\n        Image.fromarray(img).save(outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_results.zip\" disp_0'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_results.zip\")\n\n\n### auxiliary functions\n\n\ndef _read_img(filename):\n    # convert to RGB for scene flow finalpass data\n    img = np.asarray(Image.open(filename).convert(\"RGB\"))\n    return img\n\n\ndef _read_booster_disp(filename):\n    disp = np.load(filename)\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_png_disp(filename, coef=1.0):\n    disp = np.asarray(Image.open(filename))\n    disp = disp.astype(np.float32) / coef\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_pfm_disp(filename):\n    disp = np.ascontiguousarray(_read_pfm(filename)[0])\n    disp[disp <= 0] = (\n        np.inf\n    )  # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm\n    return disp\n\n\ndef _read_npy_disp(filename):\n    return np.load(filename)\n","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_png_disp","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_png_disp#L838-L842","kind":"function","name":"_read_png_disp","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":838,"end_line":842,"context_start_line":818,"context_end_line":862,"code":"        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_results.zip\")\n\n\n### auxiliary functions\n\n\ndef _read_img(filename):\n    # convert to RGB for scene flow finalpass data\n    img = np.asarray(Image.open(filename).convert(\"RGB\"))\n    return img\n\n\ndef _read_booster_disp(filename):\n    disp = np.load(filename)\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_png_disp(filename, coef=1.0):\n    disp = np.asarray(Image.open(filename))\n    disp = disp.astype(np.float32) / coef\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_pfm_disp(filename):\n    disp = np.ascontiguousarray(_read_pfm(filename)[0])\n    disp[disp <= 0] = (\n        np.inf\n    )  # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm\n    return disp\n\n\ndef _read_npy_disp(filename):\n    return np.load(filename)\n\n\ndef _read_crestereo_disp(filename):\n    return _read_png_disp(filename, coef=32.0)\n\n\ndef _read_middlebury20052006_disp(filename):\n    return _read_png_disp(filename, coef=1.0)","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_pfm_disp","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_pfm_disp#L845-L850","kind":"function","name":"_read_pfm_disp","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":845,"end_line":850,"context_start_line":825,"context_end_line":870,"code":"\ndef _read_img(filename):\n    # convert to RGB for scene flow finalpass data\n    img = np.asarray(Image.open(filename).convert(\"RGB\"))\n    return img\n\n\ndef _read_booster_disp(filename):\n    disp = np.load(filename)\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_png_disp(filename, coef=1.0):\n    disp = np.asarray(Image.open(filename))\n    disp = disp.astype(np.float32) / coef\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_pfm_disp(filename):\n    disp = np.ascontiguousarray(_read_pfm(filename)[0])\n    disp[disp <= 0] = (\n        np.inf\n    )  # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm\n    return disp\n\n\ndef _read_npy_disp(filename):\n    return np.load(filename)\n\n\ndef _read_crestereo_disp(filename):\n    return _read_png_disp(filename, coef=32.0)\n\n\ndef _read_middlebury20052006_disp(filename):\n    return _read_png_disp(filename, coef=1.0)\n\n\ndef _read_kitti_disp(filename):\n    return _read_png_disp(filename, coef=256.0)\n\n\n_read_sceneflow_disp = _read_pfm_disp\n_read_eth3d_disp = _read_pfm_disp","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_npy_disp","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_npy_disp#L853-L854","kind":"function","name":"_read_npy_disp","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":853,"end_line":854,"context_start_line":833,"context_end_line":874,"code":"    disp = np.load(filename)\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_png_disp(filename, coef=1.0):\n    disp = np.asarray(Image.open(filename))\n    disp = disp.astype(np.float32) / coef\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_pfm_disp(filename):\n    disp = np.ascontiguousarray(_read_pfm(filename)[0])\n    disp[disp <= 0] = (\n        np.inf\n    )  # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm\n    return disp\n\n\ndef _read_npy_disp(filename):\n    return np.load(filename)\n\n\ndef _read_crestereo_disp(filename):\n    return _read_png_disp(filename, coef=32.0)\n\n\ndef _read_middlebury20052006_disp(filename):\n    return _read_png_disp(filename, coef=1.0)\n\n\ndef _read_kitti_disp(filename):\n    return _read_png_disp(filename, coef=256.0)\n\n\n_read_sceneflow_disp = _read_pfm_disp\n_read_eth3d_disp = _read_pfm_disp\n_read_middlebury_disp = _read_pfm_disp\n_read_carla_disp = _read_pfm_disp\n_read_tartanair_disp = _read_npy_disp\n","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_crestereo_disp","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_crestereo_disp#L857-L858","kind":"function","name":"_read_crestereo_disp","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":857,"end_line":858,"context_start_line":837,"context_end_line":878,"code":"\ndef _read_png_disp(filename, coef=1.0):\n    disp = np.asarray(Image.open(filename))\n    disp = disp.astype(np.float32) / coef\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_pfm_disp(filename):\n    disp = np.ascontiguousarray(_read_pfm(filename)[0])\n    disp[disp <= 0] = (\n        np.inf\n    )  # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm\n    return disp\n\n\ndef _read_npy_disp(filename):\n    return np.load(filename)\n\n\ndef _read_crestereo_disp(filename):\n    return _read_png_disp(filename, coef=32.0)\n\n\ndef _read_middlebury20052006_disp(filename):\n    return _read_png_disp(filename, coef=1.0)\n\n\ndef _read_kitti_disp(filename):\n    return _read_png_disp(filename, coef=256.0)\n\n\n_read_sceneflow_disp = _read_pfm_disp\n_read_eth3d_disp = _read_pfm_disp\n_read_middlebury_disp = _read_pfm_disp\n_read_carla_disp = _read_pfm_disp\n_read_tartanair_disp = _read_npy_disp\n\n\ndef _read_hdf5_disp(filename):\n    disp = np.asarray(h5py.File(filename)[\"disparity\"])\n    disp[np.isnan(disp)] = np.inf  # make invalid values as +inf","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_middlebury20052006_disp","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_middlebury20052006_disp#L861-L862","kind":"function","name":"_read_middlebury20052006_disp","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":861,"end_line":862,"context_start_line":841,"context_end_line":882,"code":"    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_pfm_disp(filename):\n    disp = np.ascontiguousarray(_read_pfm(filename)[0])\n    disp[disp <= 0] = (\n        np.inf\n    )  # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm\n    return disp\n\n\ndef _read_npy_disp(filename):\n    return np.load(filename)\n\n\ndef _read_crestereo_disp(filename):\n    return _read_png_disp(filename, coef=32.0)\n\n\ndef _read_middlebury20052006_disp(filename):\n    return _read_png_disp(filename, coef=1.0)\n\n\ndef _read_kitti_disp(filename):\n    return _read_png_disp(filename, coef=256.0)\n\n\n_read_sceneflow_disp = _read_pfm_disp\n_read_eth3d_disp = _read_pfm_disp\n_read_middlebury_disp = _read_pfm_disp\n_read_carla_disp = _read_pfm_disp\n_read_tartanair_disp = _read_npy_disp\n\n\ndef _read_hdf5_disp(filename):\n    disp = np.asarray(h5py.File(filename)[\"disparity\"])\n    disp[np.isnan(disp)] = np.inf  # make invalid values as +inf\n    # disp[disp==0.0] = np.inf # make invalid values as +inf\n    return disp.astype(np.float32)\n\n","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_kitti_disp","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_kitti_disp#L865-L866","kind":"function","name":"_read_kitti_disp","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":865,"end_line":866,"context_start_line":845,"context_end_line":886,"code":"def _read_pfm_disp(filename):\n    disp = np.ascontiguousarray(_read_pfm(filename)[0])\n    disp[disp <= 0] = (\n        np.inf\n    )  # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm\n    return disp\n\n\ndef _read_npy_disp(filename):\n    return np.load(filename)\n\n\ndef _read_crestereo_disp(filename):\n    return _read_png_disp(filename, coef=32.0)\n\n\ndef _read_middlebury20052006_disp(filename):\n    return _read_png_disp(filename, coef=1.0)\n\n\ndef _read_kitti_disp(filename):\n    return _read_png_disp(filename, coef=256.0)\n\n\n_read_sceneflow_disp = _read_pfm_disp\n_read_eth3d_disp = _read_pfm_disp\n_read_middlebury_disp = _read_pfm_disp\n_read_carla_disp = _read_pfm_disp\n_read_tartanair_disp = _read_npy_disp\n\n\ndef _read_hdf5_disp(filename):\n    disp = np.asarray(h5py.File(filename)[\"disparity\"])\n    disp[np.isnan(disp)] = np.inf  # make invalid values as +inf\n    # disp[disp==0.0] = np.inf # make invalid values as +inf\n    return disp.astype(np.float32)\n\n\nimport re\n\n\ndef _read_pfm(file):","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_hdf5_disp","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_hdf5_disp#L876-L880","kind":"function","name":"_read_hdf5_disp","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":876,"end_line":880,"context_start_line":856,"context_end_line":900,"code":"\ndef _read_crestereo_disp(filename):\n    return _read_png_disp(filename, coef=32.0)\n\n\ndef _read_middlebury20052006_disp(filename):\n    return _read_png_disp(filename, coef=1.0)\n\n\ndef _read_kitti_disp(filename):\n    return _read_png_disp(filename, coef=256.0)\n\n\n_read_sceneflow_disp = _read_pfm_disp\n_read_eth3d_disp = _read_pfm_disp\n_read_middlebury_disp = _read_pfm_disp\n_read_carla_disp = _read_pfm_disp\n_read_tartanair_disp = _read_npy_disp\n\n\ndef _read_hdf5_disp(filename):\n    disp = np.asarray(h5py.File(filename)[\"disparity\"])\n    disp[np.isnan(disp)] = np.inf  # make invalid values as +inf\n    # disp[disp==0.0] = np.inf # make invalid values as +inf\n    return disp.astype(np.float32)\n\n\nimport re\n\n\ndef _read_pfm(file):\n    file = open(file, \"rb\")\n\n    color = None\n    width = None\n    height = None\n    scale = None\n    endian = None\n\n    header = file.readline().rstrip()\n    if header.decode(\"ascii\") == \"PF\":\n        color = True\n    elif header.decode(\"ascii\") == \"Pf\":\n        color = False\n    else:","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._read_pfm","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._read_pfm#L886-L921","kind":"function","name":"_read_pfm","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":886,"end_line":921,"context_start_line":866,"context_end_line":941,"code":"    return _read_png_disp(filename, coef=256.0)\n\n\n_read_sceneflow_disp = _read_pfm_disp\n_read_eth3d_disp = _read_pfm_disp\n_read_middlebury_disp = _read_pfm_disp\n_read_carla_disp = _read_pfm_disp\n_read_tartanair_disp = _read_npy_disp\n\n\ndef _read_hdf5_disp(filename):\n    disp = np.asarray(h5py.File(filename)[\"disparity\"])\n    disp[np.isnan(disp)] = np.inf  # make invalid values as +inf\n    # disp[disp==0.0] = np.inf # make invalid values as +inf\n    return disp.astype(np.float32)\n\n\nimport re\n\n\ndef _read_pfm(file):\n    file = open(file, \"rb\")\n\n    color = None\n    width = None\n    height = None\n    scale = None\n    endian = None\n\n    header = file.readline().rstrip()\n    if header.decode(\"ascii\") == \"PF\":\n        color = True\n    elif header.decode(\"ascii\") == \"Pf\":\n        color = False\n    else:\n        raise Exception(\"Not a PFM file.\")\n\n    dim_match = re.match(r\"^(\\d+)\\s(\\d+)\\s$\", file.readline().decode(\"ascii\"))\n    if dim_match:\n        width, height = list(map(int, dim_match.groups()))\n    else:\n        raise Exception(\"Malformed PFM header.\")\n\n    scale = float(file.readline().decode(\"ascii\").rstrip())\n    if scale < 0:  # little-endian\n        endian = \"<\"\n        scale = -scale\n    else:\n        endian = \">\"  # big-endian\n\n    data = np.fromfile(file, endian + \"f\")\n    shape = (height, width, 3) if color else (height, width)\n\n    data = np.reshape(data, shape)\n    data = np.flipud(data)\n    return data, scale\n\n\ndef writePFM(file, image, scale=1):\n    file = open(file, \"wb\")\n\n    color = None\n\n    if image.dtype.name != \"float32\":\n        raise Exception(\"Image dtype must be float32.\")\n\n    image = np.flipud(image)\n\n    if len(image.shape) == 3 and image.shape[2] == 3:  # color image\n        color = True\n    elif (\n        len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1\n    ):  # greyscale\n        color = False\n    else:\n        raise Exception(\"Image must have H x W x 3, H x W x 1 or H x W dimensions.\")","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.writePFM","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.writePFM#L924-L953","kind":"function","name":"writePFM","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":924,"end_line":953,"context_start_line":904,"context_end_line":973,"code":"    if dim_match:\n        width, height = list(map(int, dim_match.groups()))\n    else:\n        raise Exception(\"Malformed PFM header.\")\n\n    scale = float(file.readline().decode(\"ascii\").rstrip())\n    if scale < 0:  # little-endian\n        endian = \"<\"\n        scale = -scale\n    else:\n        endian = \">\"  # big-endian\n\n    data = np.fromfile(file, endian + \"f\")\n    shape = (height, width, 3) if color else (height, width)\n\n    data = np.reshape(data, shape)\n    data = np.flipud(data)\n    return data, scale\n\n\ndef writePFM(file, image, scale=1):\n    file = open(file, \"wb\")\n\n    color = None\n\n    if image.dtype.name != \"float32\":\n        raise Exception(\"Image dtype must be float32.\")\n\n    image = np.flipud(image)\n\n    if len(image.shape) == 3 and image.shape[2] == 3:  # color image\n        color = True\n    elif (\n        len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1\n    ):  # greyscale\n        color = False\n    else:\n        raise Exception(\"Image must have H x W x 3, H x W x 1 or H x W dimensions.\")\n\n    file.write(\"PF\\n\" if color else \"Pf\\n\".encode())\n    file.write(\"%d %d\\n\".encode() % (image.shape[1], image.shape[0]))\n\n    endian = image.dtype.byteorder\n\n    if endian == \"<\" or endian == \"=\" and sys.byteorder == \"little\":\n        scale = -scale\n\n    file.write(\"%f\\n\".encode() % scale)\n\n    image.tofile(file)\n\n\ndef writeDsp5File(disp, filename):\n    with h5py.File(filename, \"w\") as f:\n        f.create_dataset(\"disparity\", data=disp, compression=\"gzip\", compression_opts=5)\n\n\n# disp visualization\n\n\ndef vis_disparity(disp, m=None, M=None):\n    if m is None:\n        m = disp.min()\n    if M is None:\n        M = disp.max()\n    disp_vis = (disp - m) / (M - m) * 255.0\n    disp_vis = disp_vis.astype(\"uint8\")\n    disp_vis = cv2.applyColorMap(disp_vis, cv2.COLORMAP_INFERNO)\n    return disp_vis\n","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.writeDsp5File","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.writeDsp5File#L956-L958","kind":"function","name":"writeDsp5File","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":956,"end_line":958,"context_start_line":936,"context_end_line":978,"code":"    elif (\n        len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1\n    ):  # greyscale\n        color = False\n    else:\n        raise Exception(\"Image must have H x W x 3, H x W x 1 or H x W dimensions.\")\n\n    file.write(\"PF\\n\" if color else \"Pf\\n\".encode())\n    file.write(\"%d %d\\n\".encode() % (image.shape[1], image.shape[0]))\n\n    endian = image.dtype.byteorder\n\n    if endian == \"<\" or endian == \"=\" and sys.byteorder == \"little\":\n        scale = -scale\n\n    file.write(\"%f\\n\".encode() % scale)\n\n    image.tofile(file)\n\n\ndef writeDsp5File(disp, filename):\n    with h5py.File(filename, \"w\") as f:\n        f.create_dataset(\"disparity\", data=disp, compression=\"gzip\", compression_opts=5)\n\n\n# disp visualization\n\n\ndef vis_disparity(disp, m=None, M=None):\n    if m is None:\n        m = disp.min()\n    if M is None:\n        M = disp.max()\n    disp_vis = (disp - m) / (M - m) * 255.0\n    disp_vis = disp_vis.astype(\"uint8\")\n    disp_vis = cv2.applyColorMap(disp_vis, cv2.COLORMAP_INFERNO)\n    return disp_vis\n\n\n# dataset getter\n\n\ndef get_train_dataset_stereo(dataset_str, augmentor=True, crop_size=None):","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.vis_disparity","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.vis_disparity#L964-L972","kind":"function","name":"vis_disparity","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":964,"end_line":972,"context_start_line":944,"context_end_line":991,"code":"    file.write(\"%d %d\\n\".encode() % (image.shape[1], image.shape[0]))\n\n    endian = image.dtype.byteorder\n\n    if endian == \"<\" or endian == \"=\" and sys.byteorder == \"little\":\n        scale = -scale\n\n    file.write(\"%f\\n\".encode() % scale)\n\n    image.tofile(file)\n\n\ndef writeDsp5File(disp, filename):\n    with h5py.File(filename, \"w\") as f:\n        f.create_dataset(\"disparity\", data=disp, compression=\"gzip\", compression_opts=5)\n\n\n# disp visualization\n\n\ndef vis_disparity(disp, m=None, M=None):\n    if m is None:\n        m = disp.min()\n    if M is None:\n        M = disp.max()\n    disp_vis = (disp - m) / (M - m) * 255.0\n    disp_vis = disp_vis.astype(\"uint8\")\n    disp_vis = cv2.applyColorMap(disp_vis, cv2.COLORMAP_INFERNO)\n    return disp_vis\n\n\n# dataset getter\n\n\ndef get_train_dataset_stereo(dataset_str, augmentor=True, crop_size=None):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    if augmentor:\n        dataset_str = dataset_str.replace(\")\", \", augmentor=True)\")\n    if crop_size is not None:\n        dataset_str = dataset_str.replace(\n            \")\", \", crop_size={:s})\".format(str(crop_size))\n        )\n    return eval(dataset_str)\n\n\ndef get_test_datasets_stereo(dataset_str):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    return [eval(s) for s in dataset_str.split(\"+\")]","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.get_train_dataset_stereo","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.get_train_dataset_stereo#L978-L986","kind":"function","name":"get_train_dataset_stereo","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":978,"end_line":986,"context_start_line":958,"context_end_line":991,"code":"        f.create_dataset(\"disparity\", data=disp, compression=\"gzip\", compression_opts=5)\n\n\n# disp visualization\n\n\ndef vis_disparity(disp, m=None, M=None):\n    if m is None:\n        m = disp.min()\n    if M is None:\n        M = disp.max()\n    disp_vis = (disp - m) / (M - m) * 255.0\n    disp_vis = disp_vis.astype(\"uint8\")\n    disp_vis = cv2.applyColorMap(disp_vis, cv2.COLORMAP_INFERNO)\n    return disp_vis\n\n\n# dataset getter\n\n\ndef get_train_dataset_stereo(dataset_str, augmentor=True, crop_size=None):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    if augmentor:\n        dataset_str = dataset_str.replace(\")\", \", augmentor=True)\")\n    if crop_size is not None:\n        dataset_str = dataset_str.replace(\n            \")\", \", crop_size={:s})\".format(str(crop_size))\n        )\n    return eval(dataset_str)\n\n\ndef get_test_datasets_stereo(dataset_str):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    return [eval(s) for s in dataset_str.split(\"+\")]","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.get_test_datasets_stereo","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.get_test_datasets_stereo#L989-L991","kind":"function","name":"get_test_datasets_stereo","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":989,"end_line":991,"context_start_line":969,"context_end_line":991,"code":"    disp_vis = (disp - m) / (M - m) * 255.0\n    disp_vis = disp_vis.astype(\"uint8\")\n    disp_vis = cv2.applyColorMap(disp_vis, cv2.COLORMAP_INFERNO)\n    return disp_vis\n\n\n# dataset getter\n\n\ndef get_train_dataset_stereo(dataset_str, augmentor=True, crop_size=None):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    if augmentor:\n        dataset_str = dataset_str.replace(\")\", \", augmentor=True)\")\n    if crop_size is not None:\n        dataset_str = dataset_str.replace(\n            \")\", \", crop_size={:s})\".format(str(crop_size))\n        )\n    return eval(dataset_str)\n\n\ndef get_test_datasets_stereo(dataset_str):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    return [eval(s) for s in dataset_str.split(\"+\")]","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.__init__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.__init__#L57-L70","kind":"function","name":"__init__","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":57,"end_line":70,"context_start_line":37,"context_end_line":90,"code":"}\ncache_dir = \"./data/stereoflow/datasets_stereo_cache/\"\n\n\nin1k_mean = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1)\nin1k_std = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1)\n\n\ndef img_to_tensor(img):\n    img = torch.from_numpy(img).permute(2, 0, 1).float() / 255.0\n    img = (img - in1k_mean) / in1k_std\n    return img\n\n\ndef disp_to_tensor(disp):\n    return torch.from_numpy(disp)[None, :, :]\n\n\nclass StereoDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = StereoAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(self.pairnames)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        Limgname = self.pairname_to_Limgname(pairname)\n        Rimgname = self.pairname_to_Rimgname(pairname)\n        Ldispname = (\n            self.pairname_to_Ldispname(pairname)\n            if self.pairname_to_Ldispname is not None\n            else None","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.prepare_data","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.prepare_data#L72-L76","kind":"function","name":"prepare_data","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":72,"end_line":76,"context_start_line":52,"context_end_line":96,"code":"    return torch.from_numpy(disp)[None, :, :]\n\n\nclass StereoDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = StereoAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(self.pairnames)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        Limgname = self.pairname_to_Limgname(pairname)\n        Rimgname = self.pairname_to_Rimgname(pairname)\n        Ldispname = (\n            self.pairname_to_Ldispname(pairname)\n            if self.pairname_to_Ldispname is not None\n            else None\n        )\n\n        # load images and disparities\n        Limg = _read_img(Limgname)\n        Rimg = _read_img(Rimgname)\n        disp = self.load_disparity(Ldispname) if Ldispname is not None else None","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.__len__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.__len__#L78-L79","kind":"function","name":"__len__","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":78,"end_line":79,"context_start_line":58,"context_end_line":99,"code":"        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = StereoAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(self.pairnames)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        Limgname = self.pairname_to_Limgname(pairname)\n        Rimgname = self.pairname_to_Rimgname(pairname)\n        Ldispname = (\n            self.pairname_to_Ldispname(pairname)\n            if self.pairname_to_Ldispname is not None\n            else None\n        )\n\n        # load images and disparities\n        Limg = _read_img(Limgname)\n        Rimg = _read_img(Rimgname)\n        disp = self.load_disparity(Ldispname) if Ldispname is not None else None\n\n        # sanity check\n        if disp is not None:","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.__getitem__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.__getitem__#L81-L120","kind":"function","name":"__getitem__","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":81,"end_line":120,"context_start_line":61,"context_end_line":140,"code":"        if crop_size:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = StereoAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(self.pairnames)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        Limgname = self.pairname_to_Limgname(pairname)\n        Rimgname = self.pairname_to_Rimgname(pairname)\n        Ldispname = (\n            self.pairname_to_Ldispname(pairname)\n            if self.pairname_to_Ldispname is not None\n            else None\n        )\n\n        # load images and disparities\n        Limg = _read_img(Limgname)\n        Rimg = _read_img(Rimgname)\n        disp = self.load_disparity(Ldispname) if Ldispname is not None else None\n\n        # sanity check\n        if disp is not None:\n            assert np.all(disp > 0) or self.name == \"Spring\", (\n                self.name,\n                pairname,\n                Ldispname,\n            )\n\n        # apply augmentations\n        if self.augmentor is not None:\n            Limg, Rimg, disp = self.augmentor(Limg, Rimg, disp, self.name)\n\n        if self.totensor:\n            Limg = img_to_tensor(Limg)\n            Rimg = img_to_tensor(Rimg)\n            if disp is None:\n                disp = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            else:\n                disp = disp_to_tensor(disp)\n\n        return Limg, Rimg, disp, str(pairname)\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.__rmul__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.__rmul__#L122-L125","kind":"function","name":"__rmul__","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":122,"end_line":125,"context_start_line":102,"context_end_line":145,"code":"                pairname,\n                Ldispname,\n            )\n\n        # apply augmentations\n        if self.augmentor is not None:\n            Limg, Rimg, disp = self.augmentor(Limg, Rimg, disp, self.name)\n\n        if self.totensor:\n            Limg = img_to_tensor(Limg)\n            Rimg = img_to_tensor(Rimg)\n            if disp is None:\n                disp = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            else:\n                disp = disp_to_tensor(disp)\n\n        return Limg, Rimg, disp, str(pairname)\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.__str__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.__str__#L127-L128","kind":"function","name":"__str__","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":127,"end_line":128,"context_start_line":107,"context_end_line":148,"code":"        if self.augmentor is not None:\n            Limg, Rimg, disp = self.augmentor(Limg, Rimg, disp, self.name)\n\n        if self.totensor:\n            Limg = img_to_tensor(Limg)\n            Rimg = img_to_tensor(Rimg)\n            if disp is None:\n                disp = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            else:\n                disp = disp_to_tensor(disp)\n\n        return Limg, Rimg, disp, str(pairname)\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.__repr__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.__repr__#L130-L136","kind":"function","name":"__repr__","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":130,"end_line":136,"context_start_line":110,"context_end_line":156,"code":"        if self.totensor:\n            Limg = img_to_tensor(Limg)\n            Rimg = img_to_tensor(Rimg)\n            if disp is None:\n                disp = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            else:\n                disp = disp_to_tensor(disp)\n\n        return Limg, Rimg, disp, str(pairname)\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._set_root","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._set_root#L138-L142","kind":"function","name":"_set_root","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":138,"end_line":142,"context_start_line":118,"context_end_line":162,"code":"                disp = disp_to_tensor(disp)\n\n        return Limg, Rimg, disp, str(pairname)\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass CREStereoDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"CREStereo\"\n        self._set_root()\n        assert self.split in [\"train\"]","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._load_or_build_cache","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._load_or_build_cache#L144-L154","kind":"function","name":"_load_or_build_cache","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":144,"end_line":154,"context_start_line":124,"context_end_line":174,"code":"        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass CREStereoDataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"CREStereo\"\n        self._set_root()\n        assert self.split in [\"train\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_left.jpg\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname + \"_right.jpg\"\n        )\n        self.pairname_to_Ldispname = lambda pairname: osp.join(\n            self.root, pairname + \"_left.disp.png\"\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_disparity = _read_crestereo_disp\n","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._prepare_data","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._prepare_data#L768-L786","kind":"function","name":"_prepare_data","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":768,"end_line":786,"context_start_line":748,"context_end_line":806,"code":"        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(outdir, pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        img = (prediction * 256).astype(\"uint16\")\n        Image.fromarray(img).save(outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti12_results.zip\" .'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti12_results.zip\")\n\n\nclass Kitti15Dataset(StereoDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti15\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\", \"test\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname.replace(\"/image_2/\", \"/image_3/\") + \"_10.png\"\n        )\n        self.pairname_to_Ldispname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/image_2/\", \"/disp_occ_0/\") + \"_10.png\"\n            )\n        )\n        self.pairname_to_str = lambda pairname: pairname.replace(\"/image_2/\", \"/\")\n        self.load_disparity = _read_kitti_disp\n\n    def _build_cache(self):\n        trainseqs = [\"training/image_2/%06d\" % (i) for i in range(200)]\n        subtrainseqs = trainseqs[:-5]\n        subvalseqs = trainseqs[-5:]\n        testseqs = [\"testing/image_2/%06d\" % (i) for i in range(200)]\n        assert (\n            len(trainseqs) == 200\n            and len(subtrainseqs) == 195\n            and len(subvalseqs) == 5\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo._build_cache","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo._build_cache#L788-L805","kind":"function","name":"_build_cache","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":788,"end_line":805,"context_start_line":768,"context_end_line":825,"code":"    def _prepare_data(self):\n        self.name = \"Kitti15\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\", \"test\"]\n        self.pairname_to_Limgname = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_Rimgname = lambda pairname: osp.join(\n            self.root, pairname.replace(\"/image_2/\", \"/image_3/\") + \"_10.png\"\n        )\n        self.pairname_to_Ldispname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/image_2/\", \"/disp_occ_0/\") + \"_10.png\"\n            )\n        )\n        self.pairname_to_str = lambda pairname: pairname.replace(\"/image_2/\", \"/\")\n        self.load_disparity = _read_kitti_disp\n\n    def _build_cache(self):\n        trainseqs = [\"training/image_2/%06d\" % (i) for i in range(200)]\n        subtrainseqs = trainseqs[:-5]\n        subvalseqs = trainseqs[-5:]\n        testseqs = [\"testing/image_2/%06d\" % (i) for i in range(200)]\n        assert (\n            len(trainseqs) == 200\n            and len(subtrainseqs) == 195\n            and len(subvalseqs) == 5\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(outdir, \"disp_0\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        img = (prediction * 256).astype(\"uint16\")\n        Image.fromarray(img).save(outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_results.zip\" disp_0'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_results.zip\")\n\n\n### auxiliary functions\n\n","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.submission_save_pairname","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.submission_save_pairname#L807-L813","kind":"function","name":"submission_save_pairname","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":807,"end_line":813,"context_start_line":787,"context_end_line":833,"code":"\n    def _build_cache(self):\n        trainseqs = [\"training/image_2/%06d\" % (i) for i in range(200)]\n        subtrainseqs = trainseqs[:-5]\n        subvalseqs = trainseqs[-5:]\n        testseqs = [\"testing/image_2/%06d\" % (i) for i in range(200)]\n        assert (\n            len(trainseqs) == 200\n            and len(subtrainseqs) == 195\n            and len(subvalseqs) == 5\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(outdir, \"disp_0\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        img = (prediction * 256).astype(\"uint16\")\n        Image.fromarray(img).save(outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_results.zip\" disp_0'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_results.zip\")\n\n\n### auxiliary functions\n\n\ndef _read_img(filename):\n    # convert to RGB for scene flow finalpass data\n    img = np.asarray(Image.open(filename).convert(\"RGB\"))\n    return img\n\n\ndef _read_booster_disp(filename):\n    disp = np.load(filename)","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_stereo.finalize_submission","uri":"program://Human3R/function/src.croco.stereoflow.datasets_stereo.finalize_submission#L815-L820","kind":"function","name":"finalize_submission","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":815,"end_line":820,"context_start_line":795,"context_end_line":840,"code":"            and len(subtrainseqs) == 195\n            and len(subvalseqs) == 5\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 2\n        assert prediction.dtype == np.float32\n        outfile = os.path.join(outdir, \"disp_0\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        img = (prediction * 256).astype(\"uint16\")\n        Image.fromarray(img).save(outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_results.zip\" disp_0'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_results.zip\")\n\n\n### auxiliary functions\n\n\ndef _read_img(filename):\n    # convert to RGB for scene flow finalpass data\n    img = np.asarray(Image.open(filename).convert(\"RGB\"))\n    return img\n\n\ndef _read_booster_disp(filename):\n    disp = np.load(filename)\n    disp[disp == 0.0] = np.inf\n    return disp\n\n\ndef _read_png_disp(filename, coef=1.0):\n    disp = np.asarray(Image.open(filename))\n    disp = disp.astype(np.float32) / coef","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.test","uri":"program://Human3R/module/src.croco.stereoflow.test#L1-L303","kind":"module","name":"src.croco.stereoflow.test","path":"src/croco/stereoflow/test.py","language":"python","start_line":1,"end_line":303,"context_start_line":1,"context_end_line":303,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Main test function\n# --------------------------------------------------------\n\nimport os\nimport argparse\nimport pickle\nfrom PIL import Image\nimport numpy as np\nfrom tqdm import tqdm\n\nimport torch\nfrom torch.utils.data import DataLoader\n\nimport utils.misc as misc\nfrom models.croco_downstream import CroCoDownstreamBinocular\nfrom models.head_downstream import PixelwiseTaskWithDPT\n\nfrom stereoflow.criterion import *\nfrom stereoflow.datasets_stereo import get_test_datasets_stereo\nfrom stereoflow.datasets_flow import get_test_datasets_flow\nfrom stereoflow.engine import tiled_pred\n\nfrom stereoflow.datasets_stereo import vis_disparity\nfrom stereoflow.datasets_flow import flowToColor\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser(\"Test CroCo models on stereo/flow\", add_help=False)\n    # important argument\n    parser.add_argument(\n        \"--model\", required=True, type=str, help=\"Path to the model to evaluate\"\n    )\n    parser.add_argument(\n        \"--dataset\",\n        required=True,\n        type=str,\n        help=\"test dataset (there can be multiple dataset separated by a +)\",\n    )\n    # tiling\n    parser.add_argument(\n        \"--tile_conf_mode\",\n        type=str,\n        default=\"\",\n        help=\"Weights for the tiling aggregation based on confidence (empty means use the formula from the loaded checkpoint\",\n    )\n    parser.add_argument(\n        \"--tile_overlap\", type=float, default=0.7, help=\"overlap between tiles\"\n    )\n    # save (it will automatically go to <model_path>_<dataset_str>/<tile_str>_<save>)\n    parser.add_argument(\n        \"--save\",\n        type=str,\n        nargs=\"+\",\n        default=[],\n        help=\"what to save: \\\n                              metrics (pickle file), \\\n                              pred (raw prediction save as torch tensor), \\\n                              visu (visualization in png of each prediction), \\\n                              err10 (visualization in png of the error clamp at 10 for each prediction), \\\n                              submission (submission file)\",\n    )\n    # other (no impact)\n    parser.add_argument(\"--num_workers\", default=4, type=int)\n    return parser\n\n\ndef _load_model_and_criterion(model_path, do_load_metrics, device):\n    print(\"loading model from\", model_path)\n    assert os.path.isfile(model_path)\n    ckpt = torch.load(model_path, \"cpu\")\n\n    ckpt_args = ckpt[\"args\"]\n    task = ckpt_args.task\n    tile_conf_mode = ckpt_args.tile_conf_mode\n    num_channels = {\"stereo\": 1, \"flow\": 2}[task]\n    with_conf = eval(ckpt_args.criterion).with_conf\n    if with_conf:\n        num_channels += 1\n    print(\"head: PixelwiseTaskWithDPT()\")\n    head = PixelwiseTaskWithDPT()\n    head.num_channels = num_channels\n    print(\"croco_args:\", ckpt_args.croco_args)\n    model = CroCoDownstreamBinocular(head, **ckpt_args.croco_args)\n    msg = model.load_state_dict(ckpt[\"model\"], strict=True)\n    model.eval()\n    model = model.to(device)\n\n    if do_load_metrics:\n        if task == \"stereo\":\n            metrics = StereoDatasetMetrics().to(device)\n        else:\n            metrics = FlowDatasetMetrics().to(device)\n    else:\n        metrics = None\n\n    return model, metrics, ckpt_args.crop, with_conf, task, tile_conf_mode\n\n\ndef _save_batch(\n    pred, gt, pairnames, dataset, task, save, outdir, time, submission_dir=None\n):\n\n    for i in range(len(pairnames)):\n\n        pairname = (\n            eval(pairnames[i]) if pairnames[i].startswith(\"(\") else pairnames[i]\n        )  # unbatch pairname\n        fname = os.path.join(outdir, dataset.pairname_to_str(pairname))\n        os.makedirs(os.path.dirname(fname), exist_ok=True)\n\n        predi = pred[i, ...]\n        if gt is not None:\n            gti = gt[i, ...]\n\n        if \"pred\" in save:\n            torch.save(predi.squeeze(0).cpu(), fname + \"_pred.pth\")\n\n        if \"visu\" in save:\n            if task == \"stereo\":\n                disparity = predi.permute((1, 2, 0)).squeeze(2).cpu().numpy()\n                m, M = None\n                if gt is not None:\n                    mask = torch.isfinite(gti)\n                    m = gt[mask].min()\n                    M = gt[mask].max()\n                img_disparity = vis_disparity(disparity, m=m, M=M)\n                Image.fromarray(img_disparity).save(fname + \"_pred.png\")\n            else:\n                # normalize flowToColor according to the maxnorm of gt (or prediction if not available)\n                flowNorm = (\n                    torch.sqrt(\n                        torch.sum((gti if gt is not None else predi) ** 2, dim=0)\n                    )\n                    .max()\n                    .item()\n                )\n                imgflow = flowToColor(\n                    predi.permute((1, 2, 0)).cpu().numpy(), maxflow=flowNorm\n                )\n                Image.fromarray(imgflow).save(fname + \"_pred.png\")\n\n        if \"err10\" in save:\n            assert gt is not None\n            L2err = torch.sqrt(torch.sum((gti - predi) ** 2, dim=0))\n            valid = torch.isfinite(gti[0, :, :])\n            L2err[~valid] = 0.0\n            L2err = torch.clamp(L2err, max=10.0)\n            red = (L2err * 255.0 / 10.0).to(dtype=torch.uint8)[:, :, None]\n            zer = torch.zeros_like(red)\n            imgerr = torch.cat((red, zer, zer), dim=2).cpu().numpy()\n            Image.fromarray(imgerr).save(fname + \"_err10.png\")\n\n        if \"submission\" in save:\n            assert submission_dir is not None\n            predi_np = (\n                predi.permute(1, 2, 0).squeeze(2).cpu().numpy()\n            )  # transform into HxWx2 for flow or HxW for stereo\n            dataset.submission_save_pairname(pairname, predi_np, submission_dir, time)\n\n\ndef main(args):\n\n    # load the pretrained model and metrics\n    device = (\n        torch.device(\"cuda:0\") if torch.cuda.is_available() else torch.device(\"cpu\")\n    )\n    model, metrics, cropsize, with_conf, task, tile_conf_mode = (\n        _load_model_and_criterion(args.model, \"metrics\" in args.save, device)\n    )\n    if args.tile_conf_mode == \"\":\n        args.tile_conf_mode = tile_conf_mode\n\n    # load the datasets\n    datasets = (\n        get_test_datasets_stereo if task == \"stereo\" else get_test_datasets_flow\n    )(args.dataset)\n    dataloaders = [\n        DataLoader(\n            dataset,\n            batch_size=1,\n            shuffle=False,\n            num_workers=args.num_workers,\n            pin_memory=True,\n            drop_last=False,\n        )\n        for dataset in datasets\n    ]\n\n    # run\n    for i, dataloader in enumerate(dataloaders):\n        dataset = datasets[i]\n        dstr = args.dataset.split(\"+\")[i]\n\n        outdir = args.model + \"_\" + misc.filename(dstr)\n        if \"metrics\" in args.save and len(args.save) == 1:\n            fname = os.path.join(\n                outdir, f\"conf_{args.tile_conf_mode}_overlap_{args.tile_overlap}.pkl\"\n            )\n            if os.path.isfile(fname) and len(args.save) == 1:\n                print(\"  metrics already compute in \" + fname)\n                with open(fname, \"rb\") as fid:\n                    results = pickle.load(fid)\n                for k, v in results.items():\n                    print(\"{:s}: {:.3f}\".format(k, v))\n                continue\n\n        if \"submission\" in args.save:\n            dirname = (\n                f\"submission_conf_{args.tile_conf_mode}_overlap_{args.tile_overlap}\"\n            )\n            submission_dir = os.path.join(outdir, dirname)\n        else:\n            submission_dir = None\n\n        print(\"\")\n        print(\"saving {:s} in {:s}\".format(\"+\".join(args.save), outdir))\n        print(repr(dataset))\n\n        if metrics is not None:\n            metrics.reset()\n\n        for data_iter_step, (image1, image2, gt, pairnames) in enumerate(\n            tqdm(dataloader)\n        ):\n\n            do_flip = (\n                task == \"stereo\"\n                and dstr.startswith(\"Spring\")\n                and any(\"right\" in p for p in pairnames)\n            )  # we flip the images and will flip the prediction after as we assume img1 is on the left\n\n            image1 = image1.to(device, non_blocking=True)\n            image2 = image2.to(device, non_blocking=True)\n            gt = (\n                gt.to(device, non_blocking=True) if gt.numel() > 0 else None\n            )  # special case for test time\n            if do_flip:\n                assert all(\"right\" in p for p in pairnames)\n                image1 = image1.flip(\n                    dims=[3]\n                )  # this is already the right frame, let's flip it\n                image2 = image2.flip(dims=[3])\n                gt = gt  # that is ok\n\n            with torch.inference_mode():\n                pred, _, _, time = tiled_pred(\n                    model,\n                    None,\n                    image1,\n                    image2,\n                    None if dataset.name == \"Spring\" else gt,\n                    conf_mode=args.tile_conf_mode,\n                    overlap=args.tile_overlap,\n                    crop=cropsize,\n                    with_conf=with_conf,\n                    return_time=True,\n                )\n\n                if do_flip:\n                    pred = pred.flip(dims=[3])\n\n                if metrics is not None:\n                    metrics.add_batch(pred, gt)\n\n                if any(k in args.save for k in [\"pred\", \"visu\", \"err10\", \"submission\"]):\n                    _save_batch(\n                        pred,\n                        gt,\n                        pairnames,\n                        dataset,\n                        task,\n                        args.save,\n                        outdir,\n                        time,\n                        submission_dir=submission_dir,\n                    )\n\n        # print\n        if metrics is not None:\n            results = metrics.get_results()\n            for k, v in results.items():\n                print(\"{:s}: {:.3f}\".format(k, v))\n\n        # save if needed\n        if \"metrics\" in args.save:\n            os.makedirs(os.path.dirname(fname), exist_ok=True)\n            with open(fname, \"wb\") as fid:\n                pickle.dump(results, fid)\n            print(\"metrics saved in\", fname)\n\n        # finalize submission if needed\n        if \"submission\" in args.save:\n            dataset.finalize_submission(submission_dir)\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    main(args)","source_hash":"9b2037c221595588a9c1684ecf9bedc5786069a7378f136984998d469c1ea252","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.test.get_args_parser","uri":"program://Human3R/function/src.croco.stereoflow.test.get_args_parser#L31-L68","kind":"function","name":"get_args_parser","path":"src/croco/stereoflow/test.py","language":"python","start_line":31,"end_line":68,"context_start_line":11,"context_end_line":88,"code":"from PIL import Image\nimport numpy as np\nfrom tqdm import tqdm\n\nimport torch\nfrom torch.utils.data import DataLoader\n\nimport utils.misc as misc\nfrom models.croco_downstream import CroCoDownstreamBinocular\nfrom models.head_downstream import PixelwiseTaskWithDPT\n\nfrom stereoflow.criterion import *\nfrom stereoflow.datasets_stereo import get_test_datasets_stereo\nfrom stereoflow.datasets_flow import get_test_datasets_flow\nfrom stereoflow.engine import tiled_pred\n\nfrom stereoflow.datasets_stereo import vis_disparity\nfrom stereoflow.datasets_flow import flowToColor\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser(\"Test CroCo models on stereo/flow\", add_help=False)\n    # important argument\n    parser.add_argument(\n        \"--model\", required=True, type=str, help=\"Path to the model to evaluate\"\n    )\n    parser.add_argument(\n        \"--dataset\",\n        required=True,\n        type=str,\n        help=\"test dataset (there can be multiple dataset separated by a +)\",\n    )\n    # tiling\n    parser.add_argument(\n        \"--tile_conf_mode\",\n        type=str,\n        default=\"\",\n        help=\"Weights for the tiling aggregation based on confidence (empty means use the formula from the loaded checkpoint\",\n    )\n    parser.add_argument(\n        \"--tile_overlap\", type=float, default=0.7, help=\"overlap between tiles\"\n    )\n    # save (it will automatically go to <model_path>_<dataset_str>/<tile_str>_<save>)\n    parser.add_argument(\n        \"--save\",\n        type=str,\n        nargs=\"+\",\n        default=[],\n        help=\"what to save: \\\n                              metrics (pickle file), \\\n                              pred (raw prediction save as torch tensor), \\\n                              visu (visualization in png of each prediction), \\\n                              err10 (visualization in png of the error clamp at 10 for each prediction), \\\n                              submission (submission file)\",\n    )\n    # other (no impact)\n    parser.add_argument(\"--num_workers\", default=4, type=int)\n    return parser\n\n\ndef _load_model_and_criterion(model_path, do_load_metrics, device):\n    print(\"loading model from\", model_path)\n    assert os.path.isfile(model_path)\n    ckpt = torch.load(model_path, \"cpu\")\n\n    ckpt_args = ckpt[\"args\"]\n    task = ckpt_args.task\n    tile_conf_mode = ckpt_args.tile_conf_mode\n    num_channels = {\"stereo\": 1, \"flow\": 2}[task]\n    with_conf = eval(ckpt_args.criterion).with_conf\n    if with_conf:\n        num_channels += 1\n    print(\"head: PixelwiseTaskWithDPT()\")\n    head = PixelwiseTaskWithDPT()\n    head.num_channels = num_channels\n    print(\"croco_args:\", ckpt_args.croco_args)\n    model = CroCoDownstreamBinocular(head, **ckpt_args.croco_args)\n    msg = model.load_state_dict(ckpt[\"model\"], strict=True)","source_hash":"9b2037c221595588a9c1684ecf9bedc5786069a7378f136984998d469c1ea252","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.test._load_model_and_criterion","uri":"program://Human3R/function/src.croco.stereoflow.test._load_model_and_criterion#L71-L100","kind":"function","name":"_load_model_and_criterion","path":"src/croco/stereoflow/test.py","language":"python","start_line":71,"end_line":100,"context_start_line":51,"context_end_line":120,"code":"        \"--tile_overlap\", type=float, default=0.7, help=\"overlap between tiles\"\n    )\n    # save (it will automatically go to <model_path>_<dataset_str>/<tile_str>_<save>)\n    parser.add_argument(\n        \"--save\",\n        type=str,\n        nargs=\"+\",\n        default=[],\n        help=\"what to save: \\\n                              metrics (pickle file), \\\n                              pred (raw prediction save as torch tensor), \\\n                              visu (visualization in png of each prediction), \\\n                              err10 (visualization in png of the error clamp at 10 for each prediction), \\\n                              submission (submission file)\",\n    )\n    # other (no impact)\n    parser.add_argument(\"--num_workers\", default=4, type=int)\n    return parser\n\n\ndef _load_model_and_criterion(model_path, do_load_metrics, device):\n    print(\"loading model from\", model_path)\n    assert os.path.isfile(model_path)\n    ckpt = torch.load(model_path, \"cpu\")\n\n    ckpt_args = ckpt[\"args\"]\n    task = ckpt_args.task\n    tile_conf_mode = ckpt_args.tile_conf_mode\n    num_channels = {\"stereo\": 1, \"flow\": 2}[task]\n    with_conf = eval(ckpt_args.criterion).with_conf\n    if with_conf:\n        num_channels += 1\n    print(\"head: PixelwiseTaskWithDPT()\")\n    head = PixelwiseTaskWithDPT()\n    head.num_channels = num_channels\n    print(\"croco_args:\", ckpt_args.croco_args)\n    model = CroCoDownstreamBinocular(head, **ckpt_args.croco_args)\n    msg = model.load_state_dict(ckpt[\"model\"], strict=True)\n    model.eval()\n    model = model.to(device)\n\n    if do_load_metrics:\n        if task == \"stereo\":\n            metrics = StereoDatasetMetrics().to(device)\n        else:\n            metrics = FlowDatasetMetrics().to(device)\n    else:\n        metrics = None\n\n    return model, metrics, ckpt_args.crop, with_conf, task, tile_conf_mode\n\n\ndef _save_batch(\n    pred, gt, pairnames, dataset, task, save, outdir, time, submission_dir=None\n):\n\n    for i in range(len(pairnames)):\n\n        pairname = (\n            eval(pairnames[i]) if pairnames[i].startswith(\"(\") else pairnames[i]\n        )  # unbatch pairname\n        fname = os.path.join(outdir, dataset.pairname_to_str(pairname))\n        os.makedirs(os.path.dirname(fname), exist_ok=True)\n\n        predi = pred[i, ...]\n        if gt is not None:\n            gti = gt[i, ...]\n\n        if \"pred\" in save:\n            torch.save(predi.squeeze(0).cpu(), fname + \"_pred.pth\")","source_hash":"9b2037c221595588a9c1684ecf9bedc5786069a7378f136984998d469c1ea252","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.test._save_batch","uri":"program://Human3R/function/src.croco.stereoflow.test._save_batch#L103-L162","kind":"function","name":"_save_batch","path":"src/croco/stereoflow/test.py","language":"python","start_line":103,"end_line":162,"context_start_line":83,"context_end_line":182,"code":"    print(\"head: PixelwiseTaskWithDPT()\")\n    head = PixelwiseTaskWithDPT()\n    head.num_channels = num_channels\n    print(\"croco_args:\", ckpt_args.croco_args)\n    model = CroCoDownstreamBinocular(head, **ckpt_args.croco_args)\n    msg = model.load_state_dict(ckpt[\"model\"], strict=True)\n    model.eval()\n    model = model.to(device)\n\n    if do_load_metrics:\n        if task == \"stereo\":\n            metrics = StereoDatasetMetrics().to(device)\n        else:\n            metrics = FlowDatasetMetrics().to(device)\n    else:\n        metrics = None\n\n    return model, metrics, ckpt_args.crop, with_conf, task, tile_conf_mode\n\n\ndef _save_batch(\n    pred, gt, pairnames, dataset, task, save, outdir, time, submission_dir=None\n):\n\n    for i in range(len(pairnames)):\n\n        pairname = (\n            eval(pairnames[i]) if pairnames[i].startswith(\"(\") else pairnames[i]\n        )  # unbatch pairname\n        fname = os.path.join(outdir, dataset.pairname_to_str(pairname))\n        os.makedirs(os.path.dirname(fname), exist_ok=True)\n\n        predi = pred[i, ...]\n        if gt is not None:\n            gti = gt[i, ...]\n\n        if \"pred\" in save:\n            torch.save(predi.squeeze(0).cpu(), fname + \"_pred.pth\")\n\n        if \"visu\" in save:\n            if task == \"stereo\":\n                disparity = predi.permute((1, 2, 0)).squeeze(2).cpu().numpy()\n                m, M = None\n                if gt is not None:\n                    mask = torch.isfinite(gti)\n                    m = gt[mask].min()\n                    M = gt[mask].max()\n                img_disparity = vis_disparity(disparity, m=m, M=M)\n                Image.fromarray(img_disparity).save(fname + \"_pred.png\")\n            else:\n                # normalize flowToColor according to the maxnorm of gt (or prediction if not available)\n                flowNorm = (\n                    torch.sqrt(\n                        torch.sum((gti if gt is not None else predi) ** 2, dim=0)\n                    )\n                    .max()\n                    .item()\n                )\n                imgflow = flowToColor(\n                    predi.permute((1, 2, 0)).cpu().numpy(), maxflow=flowNorm\n                )\n                Image.fromarray(imgflow).save(fname + \"_pred.png\")\n\n        if \"err10\" in save:\n            assert gt is not None\n            L2err = torch.sqrt(torch.sum((gti - predi) ** 2, dim=0))\n            valid = torch.isfinite(gti[0, :, :])\n            L2err[~valid] = 0.0\n            L2err = torch.clamp(L2err, max=10.0)\n            red = (L2err * 255.0 / 10.0).to(dtype=torch.uint8)[:, :, None]\n            zer = torch.zeros_like(red)\n            imgerr = torch.cat((red, zer, zer), dim=2).cpu().numpy()\n            Image.fromarray(imgerr).save(fname + \"_err10.png\")\n\n        if \"submission\" in save:\n            assert submission_dir is not None\n            predi_np = (\n                predi.permute(1, 2, 0).squeeze(2).cpu().numpy()\n            )  # transform into HxWx2 for flow or HxW for stereo\n            dataset.submission_save_pairname(pairname, predi_np, submission_dir, time)\n\n\ndef main(args):\n\n    # load the pretrained model and metrics\n    device = (\n        torch.device(\"cuda:0\") if torch.cuda.is_available() else torch.device(\"cpu\")\n    )\n    model, metrics, cropsize, with_conf, task, tile_conf_mode = (\n        _load_model_and_criterion(args.model, \"metrics\" in args.save, device)\n    )\n    if args.tile_conf_mode == \"\":\n        args.tile_conf_mode = tile_conf_mode\n\n    # load the datasets\n    datasets = (\n        get_test_datasets_stereo if task == \"stereo\" else get_test_datasets_flow\n    )(args.dataset)\n    dataloaders = [\n        DataLoader(","source_hash":"9b2037c221595588a9c1684ecf9bedc5786069a7378f136984998d469c1ea252","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.test.main","uri":"program://Human3R/function/src.croco.stereoflow.test.main#L165-L297","kind":"function","name":"main","path":"src/croco/stereoflow/test.py","language":"python","start_line":165,"end_line":297,"context_start_line":145,"context_end_line":303,"code":"\n        if \"err10\" in save:\n            assert gt is not None\n            L2err = torch.sqrt(torch.sum((gti - predi) ** 2, dim=0))\n            valid = torch.isfinite(gti[0, :, :])\n            L2err[~valid] = 0.0\n            L2err = torch.clamp(L2err, max=10.0)\n            red = (L2err * 255.0 / 10.0).to(dtype=torch.uint8)[:, :, None]\n            zer = torch.zeros_like(red)\n            imgerr = torch.cat((red, zer, zer), dim=2).cpu().numpy()\n            Image.fromarray(imgerr).save(fname + \"_err10.png\")\n\n        if \"submission\" in save:\n            assert submission_dir is not None\n            predi_np = (\n                predi.permute(1, 2, 0).squeeze(2).cpu().numpy()\n            )  # transform into HxWx2 for flow or HxW for stereo\n            dataset.submission_save_pairname(pairname, predi_np, submission_dir, time)\n\n\ndef main(args):\n\n    # load the pretrained model and metrics\n    device = (\n        torch.device(\"cuda:0\") if torch.cuda.is_available() else torch.device(\"cpu\")\n    )\n    model, metrics, cropsize, with_conf, task, tile_conf_mode = (\n        _load_model_and_criterion(args.model, \"metrics\" in args.save, device)\n    )\n    if args.tile_conf_mode == \"\":\n        args.tile_conf_mode = tile_conf_mode\n\n    # load the datasets\n    datasets = (\n        get_test_datasets_stereo if task == \"stereo\" else get_test_datasets_flow\n    )(args.dataset)\n    dataloaders = [\n        DataLoader(\n            dataset,\n            batch_size=1,\n            shuffle=False,\n            num_workers=args.num_workers,\n            pin_memory=True,\n            drop_last=False,\n        )\n        for dataset in datasets\n    ]\n\n    # run\n    for i, dataloader in enumerate(dataloaders):\n        dataset = datasets[i]\n        dstr = args.dataset.split(\"+\")[i]\n\n        outdir = args.model + \"_\" + misc.filename(dstr)\n        if \"metrics\" in args.save and len(args.save) == 1:\n            fname = os.path.join(\n                outdir, f\"conf_{args.tile_conf_mode}_overlap_{args.tile_overlap}.pkl\"\n            )\n            if os.path.isfile(fname) and len(args.save) == 1:\n                print(\"  metrics already compute in \" + fname)\n                with open(fname, \"rb\") as fid:\n                    results = pickle.load(fid)\n                for k, v in results.items():\n                    print(\"{:s}: {:.3f}\".format(k, v))\n                continue\n\n        if \"submission\" in args.save:\n            dirname = (\n                f\"submission_conf_{args.tile_conf_mode}_overlap_{args.tile_overlap}\"\n            )\n            submission_dir = os.path.join(outdir, dirname)\n        else:\n            submission_dir = None\n\n        print(\"\")\n        print(\"saving {:s} in {:s}\".format(\"+\".join(args.save), outdir))\n        print(repr(dataset))\n\n        if metrics is not None:\n            metrics.reset()\n\n        for data_iter_step, (image1, image2, gt, pairnames) in enumerate(\n            tqdm(dataloader)\n        ):\n\n            do_flip = (\n                task == \"stereo\"\n                and dstr.startswith(\"Spring\")\n                and any(\"right\" in p for p in pairnames)\n            )  # we flip the images and will flip the prediction after as we assume img1 is on the left\n\n            image1 = image1.to(device, non_blocking=True)\n            image2 = image2.to(device, non_blocking=True)\n            gt = (\n                gt.to(device, non_blocking=True) if gt.numel() > 0 else None\n            )  # special case for test time\n            if do_flip:\n                assert all(\"right\" in p for p in pairnames)\n                image1 = image1.flip(\n                    dims=[3]\n                )  # this is already the right frame, let's flip it\n                image2 = image2.flip(dims=[3])\n                gt = gt  # that is ok\n\n            with torch.inference_mode():\n                pred, _, _, time = tiled_pred(\n                    model,\n                    None,\n                    image1,\n                    image2,\n                    None if dataset.name == \"Spring\" else gt,\n                    conf_mode=args.tile_conf_mode,\n                    overlap=args.tile_overlap,\n                    crop=cropsize,\n                    with_conf=with_conf,\n                    return_time=True,\n                )\n\n                if do_flip:\n                    pred = pred.flip(dims=[3])\n\n                if metrics is not None:\n                    metrics.add_batch(pred, gt)\n\n                if any(k in args.save for k in [\"pred\", \"visu\", \"err10\", \"submission\"]):\n                    _save_batch(\n                        pred,\n                        gt,\n                        pairnames,\n                        dataset,\n                        task,\n                        args.save,\n                        outdir,\n                        time,\n                        submission_dir=submission_dir,\n                    )\n\n        # print\n        if metrics is not None:\n            results = metrics.get_results()\n            for k, v in results.items():\n                print(\"{:s}: {:.3f}\".format(k, v))\n\n        # save if needed\n        if \"metrics\" in args.save:\n            os.makedirs(os.path.dirname(fname), exist_ok=True)\n            with open(fname, \"wb\") as fid:\n                pickle.dump(results, fid)\n            print(\"metrics saved in\", fname)\n\n        # finalize submission if needed\n        if \"submission\" in args.save:\n            dataset.finalize_submission(submission_dir)\n\n\nif __name__ == \"__main__\":\n    args = get_args_parser()\n    args = args.parse_args()\n    main(args)","source_hash":"9b2037c221595588a9c1684ecf9bedc5786069a7378f136984998d469c1ea252","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine","uri":"program://Human3R/module/src.croco.stereoflow.engine#L1-L367","kind":"module","name":"src.croco.stereoflow.engine","path":"src/croco/stereoflow/engine.py","language":"python","start_line":1,"end_line":367,"context_start_line":1,"context_end_line":367,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Main function for training one epoch or testing\n# --------------------------------------------------------\n\nimport math\nimport sys\nfrom typing import Iterable\nimport numpy as np\nimport torch\nimport torchvision\n\nfrom utils import misc as misc\n\n\ndef split_prediction_conf(predictions, with_conf=False):\n    if not with_conf:\n        return predictions, None\n    conf = predictions[:, -1:, :, :]\n    predictions = predictions[:, :-1, :, :]\n    return predictions, conf\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    metrics: torch.nn.Module,\n    data_loader: Iterable,\n    optimizer: torch.optim.Optimizer,\n    device: torch.device,\n    epoch: int,\n    loss_scaler,\n    log_writer=None,\n    print_freq=20,\n    args=None,\n):\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)\n\n    accum_iter = args.accum_iter\n\n    optimizer.zero_grad()\n\n    details = {}\n\n    if log_writer is not None:\n        print(\"log_dir: {}\".format(log_writer.log_dir))\n\n    if args.img_per_epoch:\n        iter_per_epoch = args.img_per_epoch // args.batch_size + int(\n            args.img_per_epoch % args.batch_size > 0\n        )\n        assert (\n            len(data_loader) >= iter_per_epoch\n        ), \"Dataset is too small for so many iterations\"\n        len_data_loader = iter_per_epoch\n    else:\n        len_data_loader, iter_per_epoch = len(data_loader), None\n\n    for data_iter_step, (image1, image2, gt, pairname) in enumerate(\n        metric_logger.log_every(\n            data_loader, print_freq, header, max_iter=iter_per_epoch\n        )\n    ):\n\n        image1 = image1.to(device, non_blocking=True)\n        image2 = image2.to(device, non_blocking=True)\n        gt = gt.to(device, non_blocking=True)\n\n        # we use a per iteration (instead of per epoch) lr scheduler\n        if data_iter_step % accum_iter == 0:\n            misc.adjust_learning_rate(\n                optimizer, data_iter_step / len_data_loader + epoch, args\n            )\n\n        with torch.cuda.amp.autocast(enabled=bool(args.amp)):\n            prediction = model(image1, image2)\n            prediction, conf = split_prediction_conf(prediction, criterion.with_conf)\n            batch_metrics = metrics(prediction.detach(), gt)\n            loss = (\n                criterion(prediction, gt)\n                if conf is None\n                else criterion(prediction, gt, conf)\n            )\n\n        loss_value = loss.item()\n        if not math.isfinite(loss_value):\n            print(\"Loss is {}, stopping training\".format(loss_value))\n            sys.exit(1)\n\n        loss /= accum_iter\n        loss_scaler(\n            loss,\n            optimizer,\n            parameters=model.parameters(),\n            update_grad=(data_iter_step + 1) % accum_iter == 0,\n        )\n        if (data_iter_step + 1) % accum_iter == 0:\n            optimizer.zero_grad()\n\n        torch.cuda.synchronize()\n\n        metric_logger.update(loss=loss_value)\n        for k, v in batch_metrics.items():\n            metric_logger.update(**{k: v.item()})\n        lr = optimizer.param_groups[0][\"lr\"]\n        metric_logger.update(lr=lr)\n\n        # if args.dsitributed: loss_value_reduce = misc.all_reduce_mean(loss_value)\n        time_to_log = (data_iter_step + 1) % (\n            args.tboard_log_step * accum_iter\n        ) == 0 or data_iter_step == len_data_loader - 1\n        loss_value_reduce = misc.all_reduce_mean(loss_value)\n        if log_writer is not None and time_to_log:\n            epoch_1000x = int((data_iter_step / len_data_loader + epoch) * 1000)\n            # We use epoch_1000x as the x-axis in tensorboard. This calibrates different curves when batch size changes.\n            log_writer.add_scalar(\"train/loss\", loss_value_reduce, epoch_1000x)\n            log_writer.add_scalar(\"lr\", lr, epoch_1000x)\n            for k, v in batch_metrics.items():\n                log_writer.add_scalar(\"train/\" + k, v.item(), epoch_1000x)\n\n    # gather the stats from all processes\n    # if args.distributed: metric_logger.synchronize_between_processes()\n    print(\"Averaged stats:\", metric_logger)\n    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}\n\n\n@torch.no_grad()\ndef validate_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    metrics: torch.nn.Module,\n    data_loaders: list[Iterable],\n    device: torch.device,\n    epoch: int,\n    log_writer=None,\n    args=None,\n):\n\n    model.eval()\n    metric_loggers = []\n    header = \"Epoch: [{}]\".format(epoch)\n    print_freq = 20\n\n    conf_mode = args.tile_conf_mode\n    crop = args.crop\n\n    if log_writer is not None:\n        print(\"log_dir: {}\".format(log_writer.log_dir))\n\n    results = {}\n    dnames = []\n    image1, image2, gt, prediction = None, None, None, None\n    for didx, data_loader in enumerate(data_loaders):\n        dname = str(data_loader.dataset)\n        dnames.append(dname)\n        metric_loggers.append(misc.MetricLogger(delimiter=\"  \"))\n        for data_iter_step, (image1, image2, gt, pairname) in enumerate(\n            metric_loggers[didx].log_every(data_loader, print_freq, header)\n        ):\n            image1 = image1.to(device, non_blocking=True)\n            image2 = image2.to(device, non_blocking=True)\n            gt = gt.to(device, non_blocking=True)\n            if dname.startswith(\"Spring\"):\n                assert (\n                    gt.size(2) == image1.size(2) * 2\n                    and gt.size(3) == image1.size(3) * 2\n                )\n                gt = (\n                    gt[:, :, 0::2, 0::2]\n                    + gt[:, :, 0::2, 1::2]\n                    + gt[:, :, 1::2, 0::2]\n                    + gt[:, :, 1::2, 1::2]\n                ) / 4.0  # we approximate the gt based on the 2x upsampled ones\n\n            with torch.inference_mode():\n                prediction, tiled_loss, c = tiled_pred(\n                    model,\n                    criterion,\n                    image1,\n                    image2,\n                    gt,\n                    conf_mode=conf_mode,\n                    overlap=args.val_overlap,\n                    crop=crop,\n                    with_conf=criterion.with_conf,\n                )\n                batch_metrics = metrics(prediction.detach(), gt)\n                loss = (\n                    criterion(prediction.detach(), gt)\n                    if not criterion.with_conf\n                    else criterion(prediction.detach(), gt, c)\n                )\n                loss_value = loss.item()\n                metric_loggers[didx].update(loss_tiled=tiled_loss.item())\n                metric_loggers[didx].update(**{f\"loss\": loss_value})\n                for k, v in batch_metrics.items():\n                    metric_loggers[didx].update(**{dname + \"_\" + k: v.item()})\n\n    results = {\n        k: meter.global_avg for ml in metric_loggers for k, meter in ml.meters.items()\n    }\n    if len(dnames) > 1:\n        for k in batch_metrics.keys():\n            results[\"AVG_\" + k] = sum(\n                results[dname + \"_\" + k] for dname in dnames\n            ) / len(dnames)\n\n    if log_writer is not None:\n        epoch_1000x = int((1 + epoch) * 1000)\n        for k, v in results.items():\n            log_writer.add_scalar(\"val/\" + k, v, epoch_1000x)\n\n    print(\"Averaged stats:\", results)\n    return results\n\n\nimport torch.nn.functional as F\n\n\ndef _resize_img(img, new_size):\n    return F.interpolate(img, size=new_size, mode=\"bicubic\", align_corners=False)\n\n\ndef _resize_stereo_or_flow(data, new_size):\n    assert data.ndim == 4\n    assert data.size(1) in [1, 2]\n    scale_x = new_size[1] / float(data.size(3))\n    out = F.interpolate(data, size=new_size, mode=\"bicubic\", align_corners=False)\n    out[:, 0, :, :] *= scale_x\n    if out.size(1) == 2:\n        scale_y = new_size[0] / float(data.size(2))\n        out[:, 1, :, :] *= scale_y\n        print(scale_x, new_size, data.shape)\n    return out\n\n\n@torch.no_grad()\ndef tiled_pred(\n    model,\n    criterion,\n    img1,\n    img2,\n    gt,\n    overlap=0.5,\n    bad_crop_thr=0.05,\n    downscale=False,\n    crop=512,\n    ret=\"loss\",\n    conf_mode=\"conf_expsigmoid_10_5\",\n    with_conf=False,\n    return_time=False,\n):\n\n    # for each image, we are going to run inference on many overlapping patches\n    # then, all predictions will be weighted-averaged\n    if gt is not None:\n        B, C, H, W = gt.shape\n    else:\n        B, _, H, W = img1.shape\n        C = model.head.num_channels - int(with_conf)\n    win_height, win_width = crop[0], crop[1]\n\n    # upscale to be larger than the crop\n    do_change_scale = H < win_height or W < win_width\n    if do_change_scale:\n        upscale_factor = max(win_width / W, win_height / W)\n        original_size = (H, W)\n        new_size = (round(H * upscale_factor), round(W * upscale_factor))\n        img1 = _resize_img(img1, new_size)\n        img2 = _resize_img(img2, new_size)\n        # resize gt just for the computation of tiled losses\n        if gt is not None:\n            gt = _resize_stereo_or_flow(gt, new_size)\n        H, W = img1.shape[2:4]\n\n    if conf_mode.startswith(\"conf_expsigmoid_\"):  # conf_expsigmoid_30_10\n        beta, betasigmoid = map(float, conf_mode[len(\"conf_expsigmoid_\") :].split(\"_\"))\n    elif conf_mode.startswith(\"conf_expbeta\"):  # conf_expbeta3\n        beta = float(conf_mode[len(\"conf_expbeta\") :])\n    else:\n        raise NotImplementedError(f\"conf_mode {conf_mode} is not implemented\")\n\n    def crop_generator():\n        for sy in _overlapping(H, win_height, overlap):\n            for sx in _overlapping(W, win_width, overlap):\n                yield sy, sx, sy, sx, True\n\n    # keep track of weighted sum of prediction*weights and weights\n    accu_pred = img1.new_zeros(\n        (B, C, H, W)\n    )  # accumulate the weighted sum of predictions\n    accu_conf = img1.new_zeros((B, H, W)) + 1e-16  # accumulate the weights\n    accu_c = img1.new_zeros(\n        (B, H, W)\n    )  # accumulate the weighted sum of confidences ; not so useful except for computing some losses\n\n    tiled_losses = []\n\n    if return_time:\n        start = torch.cuda.Event(enable_timing=True)\n        end = torch.cuda.Event(enable_timing=True)\n        start.record()\n\n    for sy1, sx1, sy2, sx2, aligned in crop_generator():\n        # compute optical flow there\n        pred = model(_crop(img1, sy1, sx1), _crop(img2, sy2, sx2))\n        pred, predconf = split_prediction_conf(pred, with_conf=with_conf)\n\n        if gt is not None:\n            gtcrop = _crop(gt, sy1, sx1)\n        if criterion is not None and gt is not None:\n            tiled_losses.append(\n                criterion(pred, gtcrop).item()\n                if predconf is None\n                else criterion(pred, gtcrop, predconf).item()\n            )\n\n        if conf_mode.startswith(\"conf_expsigmoid_\"):\n            conf = torch.exp(\n                -beta * 2 * (torch.sigmoid(predconf / betasigmoid) - 0.5)\n            ).view(B, win_height, win_width)\n        elif conf_mode.startswith(\"conf_expbeta\"):\n            conf = torch.exp(-beta * predconf).view(B, win_height, win_width)\n        else:\n            raise NotImplementedError\n\n        accu_pred[..., sy1, sx1] += pred * conf[:, None, :, :]\n        accu_conf[..., sy1, sx1] += conf\n        accu_c[..., sy1, sx1] += predconf.view(B, win_height, win_width) * conf\n\n    pred = accu_pred / accu_conf[:, None, :, :]\n    c = accu_c / accu_conf\n    assert not torch.any(torch.isnan(pred))\n\n    if return_time:\n        end.record()\n        torch.cuda.synchronize()\n        time = start.elapsed_time(end) / 1000.0  # this was in milliseconds\n\n    if do_change_scale:\n        pred = _resize_stereo_or_flow(pred, original_size)\n\n    if return_time:\n        return pred, torch.mean(torch.tensor(tiled_losses)), c, time\n    return pred, torch.mean(torch.tensor(tiled_losses)), c\n\n\ndef _overlapping(total, window, overlap=0.5):\n    assert total >= window and 0 <= overlap < 1, (total, window, overlap)\n    num_windows = 1 + int(np.ceil((total - window) / ((1 - overlap) * window)))\n    offsets = np.linspace(0, total - window, num_windows).round().astype(int)\n    yield from (slice(x, x + window) for x in offsets)\n\n\ndef _crop(img, sy, sx):\n    B, THREE, H, W = img.shape\n    if 0 <= sy.start and sy.stop <= H and 0 <= sx.start and sx.stop <= W:\n        return img[:, :, sy, sx]\n    l, r = max(0, -sx.start), max(0, sx.stop - W)\n    t, b = max(0, -sy.start), max(0, sy.stop - H)\n    img = torch.nn.functional.pad(img, (l, r, t, b), mode=\"constant\")\n    return img[:, :, slice(sy.start + t, sy.stop + t), slice(sx.start + l, sx.stop + l)]","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine.split_prediction_conf","uri":"program://Human3R/function/src.croco.stereoflow.engine.split_prediction_conf#L18-L23","kind":"function","name":"split_prediction_conf","path":"src/croco/stereoflow/engine.py","language":"python","start_line":18,"end_line":23,"context_start_line":1,"context_end_line":43,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Main function for training one epoch or testing\n# --------------------------------------------------------\n\nimport math\nimport sys\nfrom typing import Iterable\nimport numpy as np\nimport torch\nimport torchvision\n\nfrom utils import misc as misc\n\n\ndef split_prediction_conf(predictions, with_conf=False):\n    if not with_conf:\n        return predictions, None\n    conf = predictions[:, -1:, :, :]\n    predictions = predictions[:, :-1, :, :]\n    return predictions, conf\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    metrics: torch.nn.Module,\n    data_loader: Iterable,\n    optimizer: torch.optim.Optimizer,\n    device: torch.device,\n    epoch: int,\n    loss_scaler,\n    log_writer=None,\n    print_freq=20,\n    args=None,\n):\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)\n","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine.train_one_epoch","uri":"program://Human3R/function/src.croco.stereoflow.engine.train_one_epoch#L26-L129","kind":"function","name":"train_one_epoch","path":"src/croco/stereoflow/engine.py","language":"python","start_line":26,"end_line":129,"context_start_line":6,"context_end_line":149,"code":"# --------------------------------------------------------\n\nimport math\nimport sys\nfrom typing import Iterable\nimport numpy as np\nimport torch\nimport torchvision\n\nfrom utils import misc as misc\n\n\ndef split_prediction_conf(predictions, with_conf=False):\n    if not with_conf:\n        return predictions, None\n    conf = predictions[:, -1:, :, :]\n    predictions = predictions[:, :-1, :, :]\n    return predictions, conf\n\n\ndef train_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    metrics: torch.nn.Module,\n    data_loader: Iterable,\n    optimizer: torch.optim.Optimizer,\n    device: torch.device,\n    epoch: int,\n    loss_scaler,\n    log_writer=None,\n    print_freq=20,\n    args=None,\n):\n    model.train(True)\n    metric_logger = misc.MetricLogger(delimiter=\"  \")\n    metric_logger.add_meter(\"lr\", misc.SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n    header = \"Epoch: [{}]\".format(epoch)\n\n    accum_iter = args.accum_iter\n\n    optimizer.zero_grad()\n\n    details = {}\n\n    if log_writer is not None:\n        print(\"log_dir: {}\".format(log_writer.log_dir))\n\n    if args.img_per_epoch:\n        iter_per_epoch = args.img_per_epoch // args.batch_size + int(\n            args.img_per_epoch % args.batch_size > 0\n        )\n        assert (\n            len(data_loader) >= iter_per_epoch\n        ), \"Dataset is too small for so many iterations\"\n        len_data_loader = iter_per_epoch\n    else:\n        len_data_loader, iter_per_epoch = len(data_loader), None\n\n    for data_iter_step, (image1, image2, gt, pairname) in enumerate(\n        metric_logger.log_every(\n            data_loader, print_freq, header, max_iter=iter_per_epoch\n        )\n    ):\n\n        image1 = image1.to(device, non_blocking=True)\n        image2 = image2.to(device, non_blocking=True)\n        gt = gt.to(device, non_blocking=True)\n\n        # we use a per iteration (instead of per epoch) lr scheduler\n        if data_iter_step % accum_iter == 0:\n            misc.adjust_learning_rate(\n                optimizer, data_iter_step / len_data_loader + epoch, args\n            )\n\n        with torch.cuda.amp.autocast(enabled=bool(args.amp)):\n            prediction = model(image1, image2)\n            prediction, conf = split_prediction_conf(prediction, criterion.with_conf)\n            batch_metrics = metrics(prediction.detach(), gt)\n            loss = (\n                criterion(prediction, gt)\n                if conf is None\n                else criterion(prediction, gt, conf)\n            )\n\n        loss_value = loss.item()\n        if not math.isfinite(loss_value):\n            print(\"Loss is {}, stopping training\".format(loss_value))\n            sys.exit(1)\n\n        loss /= accum_iter\n        loss_scaler(\n            loss,\n            optimizer,\n            parameters=model.parameters(),\n            update_grad=(data_iter_step + 1) % accum_iter == 0,\n        )\n        if (data_iter_step + 1) % accum_iter == 0:\n            optimizer.zero_grad()\n\n        torch.cuda.synchronize()\n\n        metric_logger.update(loss=loss_value)\n        for k, v in batch_metrics.items():\n            metric_logger.update(**{k: v.item()})\n        lr = optimizer.param_groups[0][\"lr\"]\n        metric_logger.update(lr=lr)\n\n        # if args.dsitributed: loss_value_reduce = misc.all_reduce_mean(loss_value)\n        time_to_log = (data_iter_step + 1) % (\n            args.tboard_log_step * accum_iter\n        ) == 0 or data_iter_step == len_data_loader - 1\n        loss_value_reduce = misc.all_reduce_mean(loss_value)\n        if log_writer is not None and time_to_log:\n            epoch_1000x = int((data_iter_step / len_data_loader + epoch) * 1000)\n            # We use epoch_1000x as the x-axis in tensorboard. This calibrates different curves when batch size changes.\n            log_writer.add_scalar(\"train/loss\", loss_value_reduce, epoch_1000x)\n            log_writer.add_scalar(\"lr\", lr, epoch_1000x)\n            for k, v in batch_metrics.items():\n                log_writer.add_scalar(\"train/\" + k, v.item(), epoch_1000x)\n\n    # gather the stats from all processes\n    # if args.distributed: metric_logger.synchronize_between_processes()\n    print(\"Averaged stats:\", metric_logger)\n    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}\n\n\n@torch.no_grad()\ndef validate_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    metrics: torch.nn.Module,\n    data_loaders: list[Iterable],\n    device: torch.device,\n    epoch: int,\n    log_writer=None,\n    args=None,\n):\n\n    model.eval()\n    metric_loggers = []\n    header = \"Epoch: [{}]\".format(epoch)\n    print_freq = 20\n\n    conf_mode = args.tile_conf_mode","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine.validate_one_epoch","uri":"program://Human3R/function/src.croco.stereoflow.engine.validate_one_epoch#L133-L219","kind":"function","name":"validate_one_epoch","path":"src/croco/stereoflow/engine.py","language":"python","start_line":133,"end_line":219,"context_start_line":113,"context_end_line":239,"code":"        # if args.dsitributed: loss_value_reduce = misc.all_reduce_mean(loss_value)\n        time_to_log = (data_iter_step + 1) % (\n            args.tboard_log_step * accum_iter\n        ) == 0 or data_iter_step == len_data_loader - 1\n        loss_value_reduce = misc.all_reduce_mean(loss_value)\n        if log_writer is not None and time_to_log:\n            epoch_1000x = int((data_iter_step / len_data_loader + epoch) * 1000)\n            # We use epoch_1000x as the x-axis in tensorboard. This calibrates different curves when batch size changes.\n            log_writer.add_scalar(\"train/loss\", loss_value_reduce, epoch_1000x)\n            log_writer.add_scalar(\"lr\", lr, epoch_1000x)\n            for k, v in batch_metrics.items():\n                log_writer.add_scalar(\"train/\" + k, v.item(), epoch_1000x)\n\n    # gather the stats from all processes\n    # if args.distributed: metric_logger.synchronize_between_processes()\n    print(\"Averaged stats:\", metric_logger)\n    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}\n\n\n@torch.no_grad()\ndef validate_one_epoch(\n    model: torch.nn.Module,\n    criterion: torch.nn.Module,\n    metrics: torch.nn.Module,\n    data_loaders: list[Iterable],\n    device: torch.device,\n    epoch: int,\n    log_writer=None,\n    args=None,\n):\n\n    model.eval()\n    metric_loggers = []\n    header = \"Epoch: [{}]\".format(epoch)\n    print_freq = 20\n\n    conf_mode = args.tile_conf_mode\n    crop = args.crop\n\n    if log_writer is not None:\n        print(\"log_dir: {}\".format(log_writer.log_dir))\n\n    results = {}\n    dnames = []\n    image1, image2, gt, prediction = None, None, None, None\n    for didx, data_loader in enumerate(data_loaders):\n        dname = str(data_loader.dataset)\n        dnames.append(dname)\n        metric_loggers.append(misc.MetricLogger(delimiter=\"  \"))\n        for data_iter_step, (image1, image2, gt, pairname) in enumerate(\n            metric_loggers[didx].log_every(data_loader, print_freq, header)\n        ):\n            image1 = image1.to(device, non_blocking=True)\n            image2 = image2.to(device, non_blocking=True)\n            gt = gt.to(device, non_blocking=True)\n            if dname.startswith(\"Spring\"):\n                assert (\n                    gt.size(2) == image1.size(2) * 2\n                    and gt.size(3) == image1.size(3) * 2\n                )\n                gt = (\n                    gt[:, :, 0::2, 0::2]\n                    + gt[:, :, 0::2, 1::2]\n                    + gt[:, :, 1::2, 0::2]\n                    + gt[:, :, 1::2, 1::2]\n                ) / 4.0  # we approximate the gt based on the 2x upsampled ones\n\n            with torch.inference_mode():\n                prediction, tiled_loss, c = tiled_pred(\n                    model,\n                    criterion,\n                    image1,\n                    image2,\n                    gt,\n                    conf_mode=conf_mode,\n                    overlap=args.val_overlap,\n                    crop=crop,\n                    with_conf=criterion.with_conf,\n                )\n                batch_metrics = metrics(prediction.detach(), gt)\n                loss = (\n                    criterion(prediction.detach(), gt)\n                    if not criterion.with_conf\n                    else criterion(prediction.detach(), gt, c)\n                )\n                loss_value = loss.item()\n                metric_loggers[didx].update(loss_tiled=tiled_loss.item())\n                metric_loggers[didx].update(**{f\"loss\": loss_value})\n                for k, v in batch_metrics.items():\n                    metric_loggers[didx].update(**{dname + \"_\" + k: v.item()})\n\n    results = {\n        k: meter.global_avg for ml in metric_loggers for k, meter in ml.meters.items()\n    }\n    if len(dnames) > 1:\n        for k in batch_metrics.keys():\n            results[\"AVG_\" + k] = sum(\n                results[dname + \"_\" + k] for dname in dnames\n            ) / len(dnames)\n\n    if log_writer is not None:\n        epoch_1000x = int((1 + epoch) * 1000)\n        for k, v in results.items():\n            log_writer.add_scalar(\"val/\" + k, v, epoch_1000x)\n\n    print(\"Averaged stats:\", results)\n    return results\n\n\nimport torch.nn.functional as F\n\n\ndef _resize_img(img, new_size):\n    return F.interpolate(img, size=new_size, mode=\"bicubic\", align_corners=False)\n\n\ndef _resize_stereo_or_flow(data, new_size):\n    assert data.ndim == 4\n    assert data.size(1) in [1, 2]\n    scale_x = new_size[1] / float(data.size(3))\n    out = F.interpolate(data, size=new_size, mode=\"bicubic\", align_corners=False)\n    out[:, 0, :, :] *= scale_x\n    if out.size(1) == 2:\n        scale_y = new_size[0] / float(data.size(2))\n        out[:, 1, :, :] *= scale_y\n        print(scale_x, new_size, data.shape)\n    return out","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine._resize_img","uri":"program://Human3R/function/src.croco.stereoflow.engine._resize_img#L225-L226","kind":"function","name":"_resize_img","path":"src/croco/stereoflow/engine.py","language":"python","start_line":225,"end_line":226,"context_start_line":205,"context_end_line":246,"code":"        k: meter.global_avg for ml in metric_loggers for k, meter in ml.meters.items()\n    }\n    if len(dnames) > 1:\n        for k in batch_metrics.keys():\n            results[\"AVG_\" + k] = sum(\n                results[dname + \"_\" + k] for dname in dnames\n            ) / len(dnames)\n\n    if log_writer is not None:\n        epoch_1000x = int((1 + epoch) * 1000)\n        for k, v in results.items():\n            log_writer.add_scalar(\"val/\" + k, v, epoch_1000x)\n\n    print(\"Averaged stats:\", results)\n    return results\n\n\nimport torch.nn.functional as F\n\n\ndef _resize_img(img, new_size):\n    return F.interpolate(img, size=new_size, mode=\"bicubic\", align_corners=False)\n\n\ndef _resize_stereo_or_flow(data, new_size):\n    assert data.ndim == 4\n    assert data.size(1) in [1, 2]\n    scale_x = new_size[1] / float(data.size(3))\n    out = F.interpolate(data, size=new_size, mode=\"bicubic\", align_corners=False)\n    out[:, 0, :, :] *= scale_x\n    if out.size(1) == 2:\n        scale_y = new_size[0] / float(data.size(2))\n        out[:, 1, :, :] *= scale_y\n        print(scale_x, new_size, data.shape)\n    return out\n\n\n@torch.no_grad()\ndef tiled_pred(\n    model,\n    criterion,\n    img1,","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine._resize_stereo_or_flow","uri":"program://Human3R/function/src.croco.stereoflow.engine._resize_stereo_or_flow#L229-L239","kind":"function","name":"_resize_stereo_or_flow","path":"src/croco/stereoflow/engine.py","language":"python","start_line":229,"end_line":239,"context_start_line":209,"context_end_line":259,"code":"            results[\"AVG_\" + k] = sum(\n                results[dname + \"_\" + k] for dname in dnames\n            ) / len(dnames)\n\n    if log_writer is not None:\n        epoch_1000x = int((1 + epoch) * 1000)\n        for k, v in results.items():\n            log_writer.add_scalar(\"val/\" + k, v, epoch_1000x)\n\n    print(\"Averaged stats:\", results)\n    return results\n\n\nimport torch.nn.functional as F\n\n\ndef _resize_img(img, new_size):\n    return F.interpolate(img, size=new_size, mode=\"bicubic\", align_corners=False)\n\n\ndef _resize_stereo_or_flow(data, new_size):\n    assert data.ndim == 4\n    assert data.size(1) in [1, 2]\n    scale_x = new_size[1] / float(data.size(3))\n    out = F.interpolate(data, size=new_size, mode=\"bicubic\", align_corners=False)\n    out[:, 0, :, :] *= scale_x\n    if out.size(1) == 2:\n        scale_y = new_size[0] / float(data.size(2))\n        out[:, 1, :, :] *= scale_y\n        print(scale_x, new_size, data.shape)\n    return out\n\n\n@torch.no_grad()\ndef tiled_pred(\n    model,\n    criterion,\n    img1,\n    img2,\n    gt,\n    overlap=0.5,\n    bad_crop_thr=0.05,\n    downscale=False,\n    crop=512,\n    ret=\"loss\",\n    conf_mode=\"conf_expsigmoid_10_5\",\n    with_conf=False,\n    return_time=False,\n):\n\n    # for each image, we are going to run inference on many overlapping patches","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine.tiled_pred","uri":"program://Human3R/function/src.croco.stereoflow.engine.tiled_pred#L243-L350","kind":"function","name":"tiled_pred","path":"src/croco/stereoflow/engine.py","language":"python","start_line":243,"end_line":350,"context_start_line":223,"context_end_line":367,"code":"\n\ndef _resize_img(img, new_size):\n    return F.interpolate(img, size=new_size, mode=\"bicubic\", align_corners=False)\n\n\ndef _resize_stereo_or_flow(data, new_size):\n    assert data.ndim == 4\n    assert data.size(1) in [1, 2]\n    scale_x = new_size[1] / float(data.size(3))\n    out = F.interpolate(data, size=new_size, mode=\"bicubic\", align_corners=False)\n    out[:, 0, :, :] *= scale_x\n    if out.size(1) == 2:\n        scale_y = new_size[0] / float(data.size(2))\n        out[:, 1, :, :] *= scale_y\n        print(scale_x, new_size, data.shape)\n    return out\n\n\n@torch.no_grad()\ndef tiled_pred(\n    model,\n    criterion,\n    img1,\n    img2,\n    gt,\n    overlap=0.5,\n    bad_crop_thr=0.05,\n    downscale=False,\n    crop=512,\n    ret=\"loss\",\n    conf_mode=\"conf_expsigmoid_10_5\",\n    with_conf=False,\n    return_time=False,\n):\n\n    # for each image, we are going to run inference on many overlapping patches\n    # then, all predictions will be weighted-averaged\n    if gt is not None:\n        B, C, H, W = gt.shape\n    else:\n        B, _, H, W = img1.shape\n        C = model.head.num_channels - int(with_conf)\n    win_height, win_width = crop[0], crop[1]\n\n    # upscale to be larger than the crop\n    do_change_scale = H < win_height or W < win_width\n    if do_change_scale:\n        upscale_factor = max(win_width / W, win_height / W)\n        original_size = (H, W)\n        new_size = (round(H * upscale_factor), round(W * upscale_factor))\n        img1 = _resize_img(img1, new_size)\n        img2 = _resize_img(img2, new_size)\n        # resize gt just for the computation of tiled losses\n        if gt is not None:\n            gt = _resize_stereo_or_flow(gt, new_size)\n        H, W = img1.shape[2:4]\n\n    if conf_mode.startswith(\"conf_expsigmoid_\"):  # conf_expsigmoid_30_10\n        beta, betasigmoid = map(float, conf_mode[len(\"conf_expsigmoid_\") :].split(\"_\"))\n    elif conf_mode.startswith(\"conf_expbeta\"):  # conf_expbeta3\n        beta = float(conf_mode[len(\"conf_expbeta\") :])\n    else:\n        raise NotImplementedError(f\"conf_mode {conf_mode} is not implemented\")\n\n    def crop_generator():\n        for sy in _overlapping(H, win_height, overlap):\n            for sx in _overlapping(W, win_width, overlap):\n                yield sy, sx, sy, sx, True\n\n    # keep track of weighted sum of prediction*weights and weights\n    accu_pred = img1.new_zeros(\n        (B, C, H, W)\n    )  # accumulate the weighted sum of predictions\n    accu_conf = img1.new_zeros((B, H, W)) + 1e-16  # accumulate the weights\n    accu_c = img1.new_zeros(\n        (B, H, W)\n    )  # accumulate the weighted sum of confidences ; not so useful except for computing some losses\n\n    tiled_losses = []\n\n    if return_time:\n        start = torch.cuda.Event(enable_timing=True)\n        end = torch.cuda.Event(enable_timing=True)\n        start.record()\n\n    for sy1, sx1, sy2, sx2, aligned in crop_generator():\n        # compute optical flow there\n        pred = model(_crop(img1, sy1, sx1), _crop(img2, sy2, sx2))\n        pred, predconf = split_prediction_conf(pred, with_conf=with_conf)\n\n        if gt is not None:\n            gtcrop = _crop(gt, sy1, sx1)\n        if criterion is not None and gt is not None:\n            tiled_losses.append(\n                criterion(pred, gtcrop).item()\n                if predconf is None\n                else criterion(pred, gtcrop, predconf).item()\n            )\n\n        if conf_mode.startswith(\"conf_expsigmoid_\"):\n            conf = torch.exp(\n                -beta * 2 * (torch.sigmoid(predconf / betasigmoid) - 0.5)\n            ).view(B, win_height, win_width)\n        elif conf_mode.startswith(\"conf_expbeta\"):\n            conf = torch.exp(-beta * predconf).view(B, win_height, win_width)\n        else:\n            raise NotImplementedError\n\n        accu_pred[..., sy1, sx1] += pred * conf[:, None, :, :]\n        accu_conf[..., sy1, sx1] += conf\n        accu_c[..., sy1, sx1] += predconf.view(B, win_height, win_width) * conf\n\n    pred = accu_pred / accu_conf[:, None, :, :]\n    c = accu_c / accu_conf\n    assert not torch.any(torch.isnan(pred))\n\n    if return_time:\n        end.record()\n        torch.cuda.synchronize()\n        time = start.elapsed_time(end) / 1000.0  # this was in milliseconds\n\n    if do_change_scale:\n        pred = _resize_stereo_or_flow(pred, original_size)\n\n    if return_time:\n        return pred, torch.mean(torch.tensor(tiled_losses)), c, time\n    return pred, torch.mean(torch.tensor(tiled_losses)), c\n\n\ndef _overlapping(total, window, overlap=0.5):\n    assert total >= window and 0 <= overlap < 1, (total, window, overlap)\n    num_windows = 1 + int(np.ceil((total - window) / ((1 - overlap) * window)))\n    offsets = np.linspace(0, total - window, num_windows).round().astype(int)\n    yield from (slice(x, x + window) for x in offsets)\n\n\ndef _crop(img, sy, sx):\n    B, THREE, H, W = img.shape\n    if 0 <= sy.start and sy.stop <= H and 0 <= sx.start and sx.stop <= W:\n        return img[:, :, sy, sx]\n    l, r = max(0, -sx.start), max(0, sx.stop - W)\n    t, b = max(0, -sy.start), max(0, sy.stop - H)\n    img = torch.nn.functional.pad(img, (l, r, t, b), mode=\"constant\")\n    return img[:, :, slice(sy.start + t, sy.stop + t), slice(sx.start + l, sx.stop + l)]","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine._overlapping","uri":"program://Human3R/function/src.croco.stereoflow.engine._overlapping#L353-L357","kind":"function","name":"_overlapping","path":"src/croco/stereoflow/engine.py","language":"python","start_line":353,"end_line":357,"context_start_line":333,"context_end_line":367,"code":"        accu_conf[..., sy1, sx1] += conf\n        accu_c[..., sy1, sx1] += predconf.view(B, win_height, win_width) * conf\n\n    pred = accu_pred / accu_conf[:, None, :, :]\n    c = accu_c / accu_conf\n    assert not torch.any(torch.isnan(pred))\n\n    if return_time:\n        end.record()\n        torch.cuda.synchronize()\n        time = start.elapsed_time(end) / 1000.0  # this was in milliseconds\n\n    if do_change_scale:\n        pred = _resize_stereo_or_flow(pred, original_size)\n\n    if return_time:\n        return pred, torch.mean(torch.tensor(tiled_losses)), c, time\n    return pred, torch.mean(torch.tensor(tiled_losses)), c\n\n\ndef _overlapping(total, window, overlap=0.5):\n    assert total >= window and 0 <= overlap < 1, (total, window, overlap)\n    num_windows = 1 + int(np.ceil((total - window) / ((1 - overlap) * window)))\n    offsets = np.linspace(0, total - window, num_windows).round().astype(int)\n    yield from (slice(x, x + window) for x in offsets)\n\n\ndef _crop(img, sy, sx):\n    B, THREE, H, W = img.shape\n    if 0 <= sy.start and sy.stop <= H and 0 <= sx.start and sx.stop <= W:\n        return img[:, :, sy, sx]\n    l, r = max(0, -sx.start), max(0, sx.stop - W)\n    t, b = max(0, -sy.start), max(0, sy.stop - H)\n    img = torch.nn.functional.pad(img, (l, r, t, b), mode=\"constant\")\n    return img[:, :, slice(sy.start + t, sy.stop + t), slice(sx.start + l, sx.stop + l)]","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine._crop","uri":"program://Human3R/function/src.croco.stereoflow.engine._crop#L360-L367","kind":"function","name":"_crop","path":"src/croco/stereoflow/engine.py","language":"python","start_line":360,"end_line":367,"context_start_line":340,"context_end_line":367,"code":"    if return_time:\n        end.record()\n        torch.cuda.synchronize()\n        time = start.elapsed_time(end) / 1000.0  # this was in milliseconds\n\n    if do_change_scale:\n        pred = _resize_stereo_or_flow(pred, original_size)\n\n    if return_time:\n        return pred, torch.mean(torch.tensor(tiled_losses)), c, time\n    return pred, torch.mean(torch.tensor(tiled_losses)), c\n\n\ndef _overlapping(total, window, overlap=0.5):\n    assert total >= window and 0 <= overlap < 1, (total, window, overlap)\n    num_windows = 1 + int(np.ceil((total - window) / ((1 - overlap) * window)))\n    offsets = np.linspace(0, total - window, num_windows).round().astype(int)\n    yield from (slice(x, x + window) for x in offsets)\n\n\ndef _crop(img, sy, sx):\n    B, THREE, H, W = img.shape\n    if 0 <= sy.start and sy.stop <= H and 0 <= sx.start and sx.stop <= W:\n        return img[:, :, sy, sx]\n    l, r = max(0, -sx.start), max(0, sx.stop - W)\n    t, b = max(0, -sy.start), max(0, sy.stop - H)\n    img = torch.nn.functional.pad(img, (l, r, t, b), mode=\"constant\")\n    return img[:, :, slice(sy.start + t, sy.stop + t), slice(sx.start + l, sx.stop + l)]","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.engine.crop_generator","uri":"program://Human3R/function/src.croco.stereoflow.engine.crop_generator#L288-L291","kind":"function","name":"crop_generator","path":"src/croco/stereoflow/engine.py","language":"python","start_line":288,"end_line":291,"context_start_line":268,"context_end_line":311,"code":"    # upscale to be larger than the crop\n    do_change_scale = H < win_height or W < win_width\n    if do_change_scale:\n        upscale_factor = max(win_width / W, win_height / W)\n        original_size = (H, W)\n        new_size = (round(H * upscale_factor), round(W * upscale_factor))\n        img1 = _resize_img(img1, new_size)\n        img2 = _resize_img(img2, new_size)\n        # resize gt just for the computation of tiled losses\n        if gt is not None:\n            gt = _resize_stereo_or_flow(gt, new_size)\n        H, W = img1.shape[2:4]\n\n    if conf_mode.startswith(\"conf_expsigmoid_\"):  # conf_expsigmoid_30_10\n        beta, betasigmoid = map(float, conf_mode[len(\"conf_expsigmoid_\") :].split(\"_\"))\n    elif conf_mode.startswith(\"conf_expbeta\"):  # conf_expbeta3\n        beta = float(conf_mode[len(\"conf_expbeta\") :])\n    else:\n        raise NotImplementedError(f\"conf_mode {conf_mode} is not implemented\")\n\n    def crop_generator():\n        for sy in _overlapping(H, win_height, overlap):\n            for sx in _overlapping(W, win_width, overlap):\n                yield sy, sx, sy, sx, True\n\n    # keep track of weighted sum of prediction*weights and weights\n    accu_pred = img1.new_zeros(\n        (B, C, H, W)\n    )  # accumulate the weighted sum of predictions\n    accu_conf = img1.new_zeros((B, H, W)) + 1e-16  # accumulate the weights\n    accu_c = img1.new_zeros(\n        (B, H, W)\n    )  # accumulate the weighted sum of confidences ; not so useful except for computing some losses\n\n    tiled_losses = []\n\n    if return_time:\n        start = torch.cuda.Event(enable_timing=True)\n        end = torch.cuda.Event(enable_timing=True)\n        start.record()\n\n    for sy1, sx1, sy2, sx2, aligned in crop_generator():\n        # compute optical flow there\n        pred = model(_crop(img1, sy1, sx1), _crop(img2, sy2, sx2))","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow","uri":"program://Human3R/module/src.croco.stereoflow.datasets_flow#L1-L936","kind":"module","name":"src.croco.stereoflow.datasets_flow","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":1,"end_line":936,"context_start_line":1,"context_end_line":936,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Dataset structure for flow\n# --------------------------------------------------------\n\nimport os\nimport os.path as osp\nimport pickle\nimport numpy as np\nimport struct\nfrom PIL import Image\nimport json\nimport h5py\nimport torch\nfrom torch.utils import data\n\nfrom .augmentor import FlowAugmentor\nfrom .datasets_stereo import _read_img, img_to_tensor, dataset_to_root, _read_pfm\nfrom copy import deepcopy\n\ndataset_to_root = deepcopy(dataset_to_root)\n\ndataset_to_root.update(\n    **{\n        \"TartanAir\": \"./data/stereoflow/TartanAir\",\n        \"FlyingChairs\": \"./data/stereoflow/FlyingChairs/\",\n        \"FlyingThings\": osp.join(dataset_to_root[\"SceneFlow\"], \"FlyingThings\") + \"/\",\n        \"MPISintel\": \"./data/stereoflow//MPI-Sintel/\" + \"/\",\n    }\n)\ncache_dir = \"./data/stereoflow/datasets_flow_cache/\"\n\n\ndef flow_to_tensor(disp):\n    return torch.from_numpy(disp).float().permute(2, 0, 1)\n\n\nclass FlowDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size is not None:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = FlowAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(\n            self.pairnames\n        )  # each pairname is typically of the form (str, int1, int2)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        img1name = self.pairname_to_img1name(pairname)\n        img2name = self.pairname_to_img2name(pairname)\n        flowname = (\n            self.pairname_to_flowname(pairname)\n            if self.pairname_to_flowname is not None\n            else None\n        )\n\n        # load images and disparities\n        img1 = _read_img(img1name)\n        img2 = _read_img(img2name)\n        flow = self.load_flow(flowname) if flowname is not None else None\n\n        # apply augmentations\n        if self.augmentor is not None:\n            img1, img2, flow = self.augmentor(img1, img2, flow, self.name)\n\n        if self.totensor:\n            img1 = img_to_tensor(img1)\n            img2 = img_to_tensor(img2)\n            if flow is not None:\n                flow = flow_to_tensor(flow)\n            else:\n                flow = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            pairname = str(\n                pairname\n            )  # transform potential tuple to str to be able to batch it\n\n        return img1, img2, flow, pairname\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass TartanAirDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"TartanAir\"\n        self._set_root()\n        assert self.split in [\"train\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname[0], \"image_left/{:06d}_left.png\".format(pairname[1])\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname[0], \"image_left/{:06d}_left.png\".format(pairname[2])\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            \"flow/{:06d}_{:06d}_flow.npy\".format(pairname[1], pairname[2]),\n        )\n        self.pairname_to_str = lambda pairname: os.path.join(\n            pairname[0][pairname[0].find(\"/\") + 1 :],\n            \"{:06d}_{:06d}\".format(pairname[1], pairname[2]),\n        )\n        self.load_flow = _read_numpy_flow\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        pairs = [\n            (osp.join(s, s, difficulty, Pxxx), int(a[:6]), int(a[:6]) + 1)\n            for s in seqs\n            for difficulty in [\"Easy\", \"Hard\"]\n            for Pxxx in sorted(os.listdir(osp.join(self.root, s, s, difficulty)))\n            for a in sorted(\n                os.listdir(osp.join(self.root, s, s, difficulty, Pxxx, \"image_left/\"))\n            )[:-1]\n        ]\n        assert len(pairs) == 306268, \"incorrect parsing of pairs in TartanAir\"\n        tosave = {\"train\": pairs}\n        return tosave\n\n\nclass FlyingChairsDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"FlyingChairs\"\n        self._set_root()\n        assert self.split in [\"train\", \"val\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_img1.ppm\"\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_img2.ppm\"\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_flow.flo\"\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_flow = _read_flo_file\n\n    def _build_cache(self):\n        split_file = osp.join(self.root, \"chairs_split.txt\")\n        split_list = np.loadtxt(split_file, dtype=np.int32)\n        trainpairs = [\"{:05d}\".format(i) for i in np.where(split_list == 1)[0] + 1]\n        valpairs = [\"{:05d}\".format(i) for i in np.where(split_list == 2)[0] + 1]\n        assert (\n            len(trainpairs) == 22232 and len(valpairs) == 640\n        ), \"incorrect parsing of pairs in MPI-Sintel\"\n        tosave = {\"train\": trainpairs, \"val\": valpairs}\n        return tosave\n\n\nclass FlyingThingsDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"FlyingThings\"\n        self._set_root()\n        assert self.split in [\n            f\"{set_}_{pass_}pass{camstr}\"\n            for set_ in [\"train\", \"test\", \"test1024\"]\n            for camstr in [\"\", \"_rightcam\"]\n            for pass_ in [\"clean\", \"final\", \"all\"]\n        ]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root,\n            f\"frames_{pairname[3]}pass\",\n            pairname[0].replace(\"into_future\", \"\").replace(\"into_past\", \"\"),\n            \"{:04d}.png\".format(pairname[1]),\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root,\n            f\"frames_{pairname[3]}pass\",\n            pairname[0].replace(\"into_future\", \"\").replace(\"into_past\", \"\"),\n            \"{:04d}.png\".format(pairname[2]),\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root,\n            \"optical_flow\",\n            pairname[0],\n            \"OpticalFlowInto{f:s}_{i:04d}_{c:s}.pfm\".format(\n                f=\"Future\" if \"future\" in pairname[0] else \"Past\",\n                i=pairname[1],\n                c=\"L\" if \"left\" in pairname[0] else \"R\",\n            ),\n        )\n        self.pairname_to_str = lambda pairname: os.path.join(\n            pairname[3] + \"pass\",\n            pairname[0],\n            \"Into{f:s}_{i:04d}_{c:s}\".format(\n                f=\"Future\" if \"future\" in pairname[0] else \"Past\",\n                i=pairname[1],\n                c=\"L\" if \"left\" in pairname[0] else \"R\",\n            ),\n        )\n        self.load_flow = _read_pfm_flow\n\n    def _build_cache(self):\n        tosave = {}\n        # train and test splits for the different passes\n        for set_ in [\"train\", \"test\"]:\n            sroot = osp.join(self.root, \"optical_flow\", set_.upper())\n            fname_to_i = lambda f: int(\n                f[len(\"OpticalFlowIntoFuture_\") : -len(\"_L.pfm\")]\n            )\n            pp = [\n                (osp.join(set_.upper(), d, s, \"into_future/left\"), fname_to_i(fname))\n                for d in sorted(os.listdir(sroot))\n                for s in sorted(os.listdir(osp.join(sroot, d)))\n                for fname in sorted(\n                    os.listdir(osp.join(sroot, d, s, \"into_future/left\"))\n                )[:-1]\n            ]\n            pairs = [(a, i, i + 1) for a, i in pp]\n            pairs += [(a.replace(\"into_future\", \"into_past\"), i + 1, i) for a, i in pp]\n            assert (\n                len(pairs) == {\"train\": 40302, \"test\": 7866}[set_]\n            ), \"incorrect parsing of pairs Flying Things\"\n            for cam in [\"left\", \"right\"]:\n                camstr = \"\" if cam == \"left\" else f\"_{cam}cam\"\n                for pass_ in [\"final\", \"clean\"]:\n                    tosave[f\"{set_}_{pass_}pass{camstr}\"] = [\n                        (a.replace(\"left\", cam), i, j, pass_) for a, i, j in pairs\n                    ]\n                tosave[f\"{set_}_allpass{camstr}\"] = (\n                    tosave[f\"{set_}_cleanpass{camstr}\"]\n                    + tosave[f\"{set_}_finalpass{camstr}\"]\n                )\n        # test1024: this is the same split as unimatch 'validation' split\n        # see https://github.com/autonomousvision/unimatch/blob/master/dataloader/flow/datasets.py#L229\n        test1024_nsamples = 1024\n        alltest_nsamples = len(tosave[\"test_cleanpass\"])  # 7866\n        stride = alltest_nsamples // test1024_nsamples\n        remove = alltest_nsamples % test1024_nsamples\n        for cam in [\"left\", \"right\"]:\n            camstr = \"\" if cam == \"left\" else f\"_{cam}cam\"\n            for pass_ in [\"final\", \"clean\"]:\n                tosave[f\"test1024_{pass_}pass{camstr}\"] = sorted(\n                    tosave[f\"test_{pass_}pass{camstr}\"]\n                )[:-remove][\n                    ::stride\n                ]  # warning, it was not sorted before\n            assert (\n                len(tosave[\"test1024_cleanpass\"]) == 1024\n            ), \"incorrect parsing of pairs in Flying Things\"\n            tosave[f\"test1024_allpass{camstr}\"] = (\n                tosave[f\"test1024_cleanpass{camstr}\"]\n                + tosave[f\"test1024_finalpass{camstr}\"]\n            )\n        return tosave\n\n\nclass MPISintelDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"MPISintel\"\n        self._set_root()\n        assert self.split in [\n            s + \"_\" + p\n            for s in [\"train\", \"test\", \"subval\", \"subtrain\"]\n            for p in [\"cleanpass\", \"finalpass\", \"allpass\"]\n        ]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname[0], \"frame_{:04d}.png\".format(pairname[1])\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname[0], \"frame_{:04d}.png\".format(pairname[1] + 1)\n        )\n        self.pairname_to_flowname = lambda pairname: (\n            None\n            if pairname[0].startswith(\"test/\")\n            else osp.join(\n                self.root,\n                pairname[0].replace(\"/clean/\", \"/flow/\").replace(\"/final/\", \"/flow/\"),\n                \"frame_{:04d}.flo\".format(pairname[1]),\n            )\n        )\n        self.pairname_to_str = lambda pairname: osp.join(\n            pairname[0], \"frame_{:04d}\".format(pairname[1])\n        )\n        self.load_flow = _read_flo_file\n\n    def _build_cache(self):\n        trainseqs = sorted(os.listdir(self.root + \"training/clean\"))\n        trainpairs = [\n            (osp.join(\"training/clean\", s), i)\n            for s in trainseqs\n            for i in range(1, len(os.listdir(self.root + \"training/clean/\" + s)))\n        ]\n        subvalseqs = [\"temple_2\", \"temple_3\"]\n        subtrainseqs = [s for s in trainseqs if s not in subvalseqs]\n        subvalpairs = [(p, i) for p, i in trainpairs if any(s in p for s in subvalseqs)]\n        subtrainpairs = [\n            (p, i) for p, i in trainpairs if any(s in p for s in subtrainseqs)\n        ]\n        testseqs = sorted(os.listdir(self.root + \"test/clean\"))\n        testpairs = [\n            (osp.join(\"test/clean\", s), i)\n            for s in testseqs\n            for i in range(1, len(os.listdir(self.root + \"test/clean/\" + s)))\n        ]\n        assert (\n            len(trainpairs) == 1041\n            and len(testpairs) == 552\n            and len(subvalpairs) == 98\n            and len(subtrainpairs) == 943\n        ), \"incorrect parsing of pairs in MPI-Sintel\"\n        tosave = {}\n        tosave[\"train_cleanpass\"] = trainpairs\n        tosave[\"test_cleanpass\"] = testpairs\n        tosave[\"subval_cleanpass\"] = subvalpairs\n        tosave[\"subtrain_cleanpass\"] = subtrainpairs\n        for t in [\"train\", \"test\", \"subval\", \"subtrain\"]:\n            tosave[t + \"_finalpass\"] = [\n                (p.replace(\"/clean/\", \"/final/\"), i)\n                for p, i in tosave[t + \"_cleanpass\"]\n            ]\n            tosave[t + \"_allpass\"] = tosave[t + \"_cleanpass\"] + tosave[t + \"_finalpass\"]\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, _time):\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(\n            outdir, \"submission\", self.pairname_to_str(pairname) + \".flo\"\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowFile(prediction, outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test_allpass\"\n        bundle_exe = \"/nfs/data/ffs-3d/datasets/StereoFlow/MPI-Sintel/bundler/linux-x64/bundler\"  # eg <bundle_exe> <path_to_results_for_clean> <path_to_results_for_final> <output/bundled.lzma>\n        if os.path.isfile(bundle_exe):\n            cmd = f'{bundle_exe} \"{outdir}/submission/test/clean/\" \"{outdir}/submission/test/final\" \"{outdir}/submission/bundled.lzma\"'\n            print(cmd)\n            os.system(cmd)\n            print(f'Done. Submission file at: \"{outdir}/submission/bundled.lzma\"')\n        else:\n            print(\"Could not find bundler executable for submission.\")\n            print(\"Please download it and run:\")\n            print(\n                f'<bundle_exe> \"{outdir}/submission/test/clean/\" \"{outdir}/submission/test/final\" \"{outdir}/submission/bundled.lzma\"'\n            )\n\n\nclass SpringDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"Spring\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\", \"subtrain\", \"subval\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            pairname[1],\n            \"frame_\" + pairname[3],\n            \"frame_{:s}_{:04d}.png\".format(pairname[3], pairname[4]),\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            pairname[1],\n            \"frame_\" + pairname[3],\n            \"frame_{:s}_{:04d}.png\".format(\n                pairname[3], pairname[4] + (1 if pairname[2] == \"FW\" else -1)\n            ),\n        )\n        self.pairname_to_flowname = lambda pairname: (\n            None\n            if pairname[0] == \"test\"\n            else osp.join(\n                self.root,\n                pairname[0],\n                pairname[1],\n                f\"flow_{pairname[2]}_{pairname[3]}\",\n                f\"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}.flo5\",\n            )\n        )\n        self.pairname_to_str = lambda pairname: osp.join(\n            pairname[0],\n            pairname[1],\n            f\"flow_{pairname[2]}_{pairname[3]}\",\n            f\"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}\",\n        )\n        self.load_flow = _read_hdf5_flow\n\n    def _build_cache(self):\n        # train\n        trainseqs = sorted(os.listdir(osp.join(self.root, \"train\")))\n        trainpairs = []\n        for leftright in [\"left\", \"right\"]:\n            for fwbw in [\"FW\", \"BW\"]:\n                trainpairs += [\n                    (\n                        \"train\",\n                        s,\n                        fwbw,\n                        leftright,\n                        int(f[len(f\"flow_{fwbw}_{leftright}_\") : -len(\".flo5\")]),\n                    )\n                    for s in trainseqs\n                    for f in sorted(\n                        os.listdir(\n                            osp.join(self.root, \"train\", s, f\"flow_{fwbw}_{leftright}\")\n                        )\n                    )\n                ]\n        # test\n        testseqs = sorted(os.listdir(osp.join(self.root, \"test\")))\n        testpairs = []\n        for leftright in [\"left\", \"right\"]:\n            testpairs += [\n                (\n                    \"test\",\n                    s,\n                    \"FW\",\n                    leftright,\n                    int(f[len(f\"frame_{leftright}_\") : -len(\".png\")]),\n                )\n                for s in testseqs\n                for f in sorted(\n                    os.listdir(osp.join(self.root, \"test\", s, f\"frame_{leftright}\"))\n                )[:-1]\n            ]\n            testpairs += [\n                (\n                    \"test\",\n                    s,\n                    \"BW\",\n                    leftright,\n                    int(f[len(f\"frame_{leftright}_\") : -len(\".png\")]) + 1,\n                )\n                for s in testseqs\n                for f in sorted(\n                    os.listdir(osp.join(self.root, \"test\", s, f\"frame_{leftright}\"))\n                )[:-1]\n            ]\n        # subtrain / subval\n        subtrainpairs = [p for p in trainpairs if p[1] != \"0041\"]\n        subvalpairs = [p for p in trainpairs if p[1] == \"0041\"]\n        assert (\n            len(trainpairs) == 19852\n            and len(testpairs) == 3960\n            and len(subtrainpairs) == 19472\n            and len(subvalpairs) == 380\n        ), \"incorrect parsing of pairs in Spring\"\n        tosave = {\n            \"train\": trainpairs,\n            \"test\": testpairs,\n            \"subtrain\": subtrainpairs,\n            \"subval\": subvalpairs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        assert prediction.dtype == np.float32\n        outfile = osp.join(\n            outdir,\n            pairname[0],\n            pairname[1],\n            f\"flow_{pairname[2]}_{pairname[3]}\",\n            f\"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}.flo5\",\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlo5File(prediction, outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        exe = \"{self.root}/flow_subsampling\"\n        if os.path.isfile(exe):\n            cmd = f'cd \"{outdir}/test\"; {exe} .'\n            print(cmd)\n            os.system(cmd)\n            pr\n# ... truncated ...","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.flow_to_tensor","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.flow_to_tensor#L36-L37","kind":"function","name":"flow_to_tensor","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":36,"end_line":37,"context_start_line":16,"context_end_line":57,"code":"import torch\nfrom torch.utils import data\n\nfrom .augmentor import FlowAugmentor\nfrom .datasets_stereo import _read_img, img_to_tensor, dataset_to_root, _read_pfm\nfrom copy import deepcopy\n\ndataset_to_root = deepcopy(dataset_to_root)\n\ndataset_to_root.update(\n    **{\n        \"TartanAir\": \"./data/stereoflow/TartanAir\",\n        \"FlyingChairs\": \"./data/stereoflow/FlyingChairs/\",\n        \"FlyingThings\": osp.join(dataset_to_root[\"SceneFlow\"], \"FlyingThings\") + \"/\",\n        \"MPISintel\": \"./data/stereoflow//MPI-Sintel/\" + \"/\",\n    }\n)\ncache_dir = \"./data/stereoflow/datasets_flow_cache/\"\n\n\ndef flow_to_tensor(disp):\n    return torch.from_numpy(disp).float().permute(2, 0, 1)\n\n\nclass FlowDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size is not None:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = FlowAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.FlowDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_flow.FlowDataset#L40-L136","kind":"class","name":"FlowDataset","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":40,"end_line":136,"context_start_line":20,"context_end_line":156,"code":"from .datasets_stereo import _read_img, img_to_tensor, dataset_to_root, _read_pfm\nfrom copy import deepcopy\n\ndataset_to_root = deepcopy(dataset_to_root)\n\ndataset_to_root.update(\n    **{\n        \"TartanAir\": \"./data/stereoflow/TartanAir\",\n        \"FlyingChairs\": \"./data/stereoflow/FlyingChairs/\",\n        \"FlyingThings\": osp.join(dataset_to_root[\"SceneFlow\"], \"FlyingThings\") + \"/\",\n        \"MPISintel\": \"./data/stereoflow//MPI-Sintel/\" + \"/\",\n    }\n)\ncache_dir = \"./data/stereoflow/datasets_flow_cache/\"\n\n\ndef flow_to_tensor(disp):\n    return torch.from_numpy(disp).float().permute(2, 0, 1)\n\n\nclass FlowDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size is not None:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = FlowAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(\n            self.pairnames\n        )  # each pairname is typically of the form (str, int1, int2)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        img1name = self.pairname_to_img1name(pairname)\n        img2name = self.pairname_to_img2name(pairname)\n        flowname = (\n            self.pairname_to_flowname(pairname)\n            if self.pairname_to_flowname is not None\n            else None\n        )\n\n        # load images and disparities\n        img1 = _read_img(img1name)\n        img2 = _read_img(img2name)\n        flow = self.load_flow(flowname) if flowname is not None else None\n\n        # apply augmentations\n        if self.augmentor is not None:\n            img1, img2, flow = self.augmentor(img1, img2, flow, self.name)\n\n        if self.totensor:\n            img1 = img_to_tensor(img1)\n            img2 = img_to_tensor(img2)\n            if flow is not None:\n                flow = flow_to_tensor(flow)\n            else:\n                flow = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            pairname = str(\n                pairname\n            )  # transform potential tuple to str to be able to batch it\n\n        return img1, img2, flow, pairname\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass TartanAirDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"TartanAir\"\n        self._set_root()\n        assert self.split in [\"train\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname[0], \"image_left/{:06d}_left.png\".format(pairname[1])\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname[0], \"image_left/{:06d}_left.png\".format(pairname[2])\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            \"flow/{:06d}_{:06d}_flow.npy\".format(pairname[1], pairname[2]),\n        )\n        self.pairname_to_str = lambda pairname: os.path.join(","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.TartanAirDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_flow.TartanAirDataset#L139-L175","kind":"class","name":"TartanAirDataset","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":139,"end_line":175,"context_start_line":119,"context_end_line":195,"code":"\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass TartanAirDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"TartanAir\"\n        self._set_root()\n        assert self.split in [\"train\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname[0], \"image_left/{:06d}_left.png\".format(pairname[1])\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname[0], \"image_left/{:06d}_left.png\".format(pairname[2])\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            \"flow/{:06d}_{:06d}_flow.npy\".format(pairname[1], pairname[2]),\n        )\n        self.pairname_to_str = lambda pairname: os.path.join(\n            pairname[0][pairname[0].find(\"/\") + 1 :],\n            \"{:06d}_{:06d}\".format(pairname[1], pairname[2]),\n        )\n        self.load_flow = _read_numpy_flow\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        pairs = [\n            (osp.join(s, s, difficulty, Pxxx), int(a[:6]), int(a[:6]) + 1)\n            for s in seqs\n            for difficulty in [\"Easy\", \"Hard\"]\n            for Pxxx in sorted(os.listdir(osp.join(self.root, s, s, difficulty)))\n            for a in sorted(\n                os.listdir(osp.join(self.root, s, s, difficulty, Pxxx, \"image_left/\"))\n            )[:-1]\n        ]\n        assert len(pairs) == 306268, \"incorrect parsing of pairs in TartanAir\"\n        tosave = {\"train\": pairs}\n        return tosave\n\n\nclass FlyingChairsDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"FlyingChairs\"\n        self._set_root()\n        assert self.split in [\"train\", \"val\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_img1.ppm\"\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_img2.ppm\"\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_flow.flo\"\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_flow = _read_flo_file\n","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.FlyingChairsDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_flow.FlyingChairsDataset#L178-L205","kind":"class","name":"FlyingChairsDataset","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":178,"end_line":205,"context_start_line":158,"context_end_line":225,"code":"            \"{:06d}_{:06d}\".format(pairname[1], pairname[2]),\n        )\n        self.load_flow = _read_numpy_flow\n\n    def _build_cache(self):\n        seqs = sorted(os.listdir(self.root))\n        pairs = [\n            (osp.join(s, s, difficulty, Pxxx), int(a[:6]), int(a[:6]) + 1)\n            for s in seqs\n            for difficulty in [\"Easy\", \"Hard\"]\n            for Pxxx in sorted(os.listdir(osp.join(self.root, s, s, difficulty)))\n            for a in sorted(\n                os.listdir(osp.join(self.root, s, s, difficulty, Pxxx, \"image_left/\"))\n            )[:-1]\n        ]\n        assert len(pairs) == 306268, \"incorrect parsing of pairs in TartanAir\"\n        tosave = {\"train\": pairs}\n        return tosave\n\n\nclass FlyingChairsDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"FlyingChairs\"\n        self._set_root()\n        assert self.split in [\"train\", \"val\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_img1.ppm\"\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_img2.ppm\"\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_flow.flo\"\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_flow = _read_flo_file\n\n    def _build_cache(self):\n        split_file = osp.join(self.root, \"chairs_split.txt\")\n        split_list = np.loadtxt(split_file, dtype=np.int32)\n        trainpairs = [\"{:05d}\".format(i) for i in np.where(split_list == 1)[0] + 1]\n        valpairs = [\"{:05d}\".format(i) for i in np.where(split_list == 2)[0] + 1]\n        assert (\n            len(trainpairs) == 22232 and len(valpairs) == 640\n        ), \"incorrect parsing of pairs in MPI-Sintel\"\n        tosave = {\"train\": trainpairs, \"val\": valpairs}\n        return tosave\n\n\nclass FlyingThingsDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"FlyingThings\"\n        self._set_root()\n        assert self.split in [\n            f\"{set_}_{pass_}pass{camstr}\"\n            for set_ in [\"train\", \"test\", \"test1024\"]\n            for camstr in [\"\", \"_rightcam\"]\n            for pass_ in [\"clean\", \"final\", \"all\"]\n        ]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root,\n            f\"frames_{pairname[3]}pass\",\n            pairname[0].replace(\"into_future\", \"\").replace(\"into_past\", \"\"),\n            \"{:04d}.png\".format(pairname[1]),\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.FlyingThingsDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_flow.FlyingThingsDataset#L208-L304","kind":"class","name":"FlyingThingsDataset","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":208,"end_line":304,"context_start_line":188,"context_end_line":324,"code":"            self.root, \"data\", pairname + \"_img2.ppm\"\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root, \"data\", pairname + \"_flow.flo\"\n        )\n        self.pairname_to_str = lambda pairname: pairname\n        self.load_flow = _read_flo_file\n\n    def _build_cache(self):\n        split_file = osp.join(self.root, \"chairs_split.txt\")\n        split_list = np.loadtxt(split_file, dtype=np.int32)\n        trainpairs = [\"{:05d}\".format(i) for i in np.where(split_list == 1)[0] + 1]\n        valpairs = [\"{:05d}\".format(i) for i in np.where(split_list == 2)[0] + 1]\n        assert (\n            len(trainpairs) == 22232 and len(valpairs) == 640\n        ), \"incorrect parsing of pairs in MPI-Sintel\"\n        tosave = {\"train\": trainpairs, \"val\": valpairs}\n        return tosave\n\n\nclass FlyingThingsDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"FlyingThings\"\n        self._set_root()\n        assert self.split in [\n            f\"{set_}_{pass_}pass{camstr}\"\n            for set_ in [\"train\", \"test\", \"test1024\"]\n            for camstr in [\"\", \"_rightcam\"]\n            for pass_ in [\"clean\", \"final\", \"all\"]\n        ]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root,\n            f\"frames_{pairname[3]}pass\",\n            pairname[0].replace(\"into_future\", \"\").replace(\"into_past\", \"\"),\n            \"{:04d}.png\".format(pairname[1]),\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root,\n            f\"frames_{pairname[3]}pass\",\n            pairname[0].replace(\"into_future\", \"\").replace(\"into_past\", \"\"),\n            \"{:04d}.png\".format(pairname[2]),\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root,\n            \"optical_flow\",\n            pairname[0],\n            \"OpticalFlowInto{f:s}_{i:04d}_{c:s}.pfm\".format(\n                f=\"Future\" if \"future\" in pairname[0] else \"Past\",\n                i=pairname[1],\n                c=\"L\" if \"left\" in pairname[0] else \"R\",\n            ),\n        )\n        self.pairname_to_str = lambda pairname: os.path.join(\n            pairname[3] + \"pass\",\n            pairname[0],\n            \"Into{f:s}_{i:04d}_{c:s}\".format(\n                f=\"Future\" if \"future\" in pairname[0] else \"Past\",\n                i=pairname[1],\n                c=\"L\" if \"left\" in pairname[0] else \"R\",\n            ),\n        )\n        self.load_flow = _read_pfm_flow\n\n    def _build_cache(self):\n        tosave = {}\n        # train and test splits for the different passes\n        for set_ in [\"train\", \"test\"]:\n            sroot = osp.join(self.root, \"optical_flow\", set_.upper())\n            fname_to_i = lambda f: int(\n                f[len(\"OpticalFlowIntoFuture_\") : -len(\"_L.pfm\")]\n            )\n            pp = [\n                (osp.join(set_.upper(), d, s, \"into_future/left\"), fname_to_i(fname))\n                for d in sorted(os.listdir(sroot))\n                for s in sorted(os.listdir(osp.join(sroot, d)))\n                for fname in sorted(\n                    os.listdir(osp.join(sroot, d, s, \"into_future/left\"))\n                )[:-1]\n            ]\n            pairs = [(a, i, i + 1) for a, i in pp]\n            pairs += [(a.replace(\"into_future\", \"into_past\"), i + 1, i) for a, i in pp]\n            assert (\n                len(pairs) == {\"train\": 40302, \"test\": 7866}[set_]\n            ), \"incorrect parsing of pairs Flying Things\"\n            for cam in [\"left\", \"right\"]:\n                camstr = \"\" if cam == \"left\" else f\"_{cam}cam\"\n                for pass_ in [\"final\", \"clean\"]:\n                    tosave[f\"{set_}_{pass_}pass{camstr}\"] = [\n                        (a.replace(\"left\", cam), i, j, pass_) for a, i, j in pairs\n                    ]\n                tosave[f\"{set_}_allpass{camstr}\"] = (\n                    tosave[f\"{set_}_cleanpass{camstr}\"]\n                    + tosave[f\"{set_}_finalpass{camstr}\"]\n                )\n        # test1024: this is the same split as unimatch 'validation' split\n        # see https://github.com/autonomousvision/unimatch/blob/master/dataloader/flow/datasets.py#L229\n        test1024_nsamples = 1024\n        alltest_nsamples = len(tosave[\"test_cleanpass\"])  # 7866\n        stride = alltest_nsamples // test1024_nsamples\n        remove = alltest_nsamples % test1024_nsamples\n        for cam in [\"left\", \"right\"]:\n            camstr = \"\" if cam == \"left\" else f\"_{cam}cam\"\n            for pass_ in [\"final\", \"clean\"]:\n                tosave[f\"test1024_{pass_}pass{camstr}\"] = sorted(\n                    tosave[f\"test_{pass_}pass{camstr}\"]\n                )[:-remove][\n                    ::stride\n                ]  # warning, it was not sorted before\n            assert (\n                len(tosave[\"test1024_cleanpass\"]) == 1024\n            ), \"incorrect parsing of pairs in Flying Things\"\n            tosave[f\"test1024_allpass{camstr}\"] = (\n                tosave[f\"test1024_cleanpass{camstr}\"]\n                + tosave[f\"test1024_finalpass{camstr}\"]\n            )\n        return tosave\n\n\nclass MPISintelDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"MPISintel\"\n        self._set_root()\n        assert self.split in [\n            s + \"_\" + p\n            for s in [\"train\", \"test\", \"subval\", \"subtrain\"]\n            for p in [\"cleanpass\", \"finalpass\", \"allpass\"]\n        ]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname[0], \"frame_{:04d}.png\".format(pairname[1])\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname[0], \"frame_{:04d}.png\".format(pairname[1] + 1)\n        )\n        self.pairname_to_flowname = lambda pairname: (\n            None","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.MPISintelDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_flow.MPISintelDataset#L307-L396","kind":"class","name":"MPISintelDataset","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":307,"end_line":396,"context_start_line":287,"context_end_line":416,"code":"        stride = alltest_nsamples // test1024_nsamples\n        remove = alltest_nsamples % test1024_nsamples\n        for cam in [\"left\", \"right\"]:\n            camstr = \"\" if cam == \"left\" else f\"_{cam}cam\"\n            for pass_ in [\"final\", \"clean\"]:\n                tosave[f\"test1024_{pass_}pass{camstr}\"] = sorted(\n                    tosave[f\"test_{pass_}pass{camstr}\"]\n                )[:-remove][\n                    ::stride\n                ]  # warning, it was not sorted before\n            assert (\n                len(tosave[\"test1024_cleanpass\"]) == 1024\n            ), \"incorrect parsing of pairs in Flying Things\"\n            tosave[f\"test1024_allpass{camstr}\"] = (\n                tosave[f\"test1024_cleanpass{camstr}\"]\n                + tosave[f\"test1024_finalpass{camstr}\"]\n            )\n        return tosave\n\n\nclass MPISintelDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"MPISintel\"\n        self._set_root()\n        assert self.split in [\n            s + \"_\" + p\n            for s in [\"train\", \"test\", \"subval\", \"subtrain\"]\n            for p in [\"cleanpass\", \"finalpass\", \"allpass\"]\n        ]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname[0], \"frame_{:04d}.png\".format(pairname[1])\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname[0], \"frame_{:04d}.png\".format(pairname[1] + 1)\n        )\n        self.pairname_to_flowname = lambda pairname: (\n            None\n            if pairname[0].startswith(\"test/\")\n            else osp.join(\n                self.root,\n                pairname[0].replace(\"/clean/\", \"/flow/\").replace(\"/final/\", \"/flow/\"),\n                \"frame_{:04d}.flo\".format(pairname[1]),\n            )\n        )\n        self.pairname_to_str = lambda pairname: osp.join(\n            pairname[0], \"frame_{:04d}\".format(pairname[1])\n        )\n        self.load_flow = _read_flo_file\n\n    def _build_cache(self):\n        trainseqs = sorted(os.listdir(self.root + \"training/clean\"))\n        trainpairs = [\n            (osp.join(\"training/clean\", s), i)\n            for s in trainseqs\n            for i in range(1, len(os.listdir(self.root + \"training/clean/\" + s)))\n        ]\n        subvalseqs = [\"temple_2\", \"temple_3\"]\n        subtrainseqs = [s for s in trainseqs if s not in subvalseqs]\n        subvalpairs = [(p, i) for p, i in trainpairs if any(s in p for s in subvalseqs)]\n        subtrainpairs = [\n            (p, i) for p, i in trainpairs if any(s in p for s in subtrainseqs)\n        ]\n        testseqs = sorted(os.listdir(self.root + \"test/clean\"))\n        testpairs = [\n            (osp.join(\"test/clean\", s), i)\n            for s in testseqs\n            for i in range(1, len(os.listdir(self.root + \"test/clean/\" + s)))\n        ]\n        assert (\n            len(trainpairs) == 1041\n            and len(testpairs) == 552\n            and len(subvalpairs) == 98\n            and len(subtrainpairs) == 943\n        ), \"incorrect parsing of pairs in MPI-Sintel\"\n        tosave = {}\n        tosave[\"train_cleanpass\"] = trainpairs\n        tosave[\"test_cleanpass\"] = testpairs\n        tosave[\"subval_cleanpass\"] = subvalpairs\n        tosave[\"subtrain_cleanpass\"] = subtrainpairs\n        for t in [\"train\", \"test\", \"subval\", \"subtrain\"]:\n            tosave[t + \"_finalpass\"] = [\n                (p.replace(\"/clean/\", \"/final/\"), i)\n                for p, i in tosave[t + \"_cleanpass\"]\n            ]\n            tosave[t + \"_allpass\"] = tosave[t + \"_cleanpass\"] + tosave[t + \"_finalpass\"]\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, _time):\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(\n            outdir, \"submission\", self.pairname_to_str(pairname) + \".flo\"\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowFile(prediction, outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test_allpass\"\n        bundle_exe = \"/nfs/data/ffs-3d/datasets/StereoFlow/MPI-Sintel/bundler/linux-x64/bundler\"  # eg <bundle_exe> <path_to_results_for_clean> <path_to_results_for_final> <output/bundled.lzma>\n        if os.path.isfile(bundle_exe):\n            cmd = f'{bundle_exe} \"{outdir}/submission/test/clean/\" \"{outdir}/submission/test/final\" \"{outdir}/submission/bundled.lzma\"'\n            print(cmd)\n            os.system(cmd)\n            print(f'Done. Submission file at: \"{outdir}/submission/bundled.lzma\"')\n        else:\n            print(\"Could not find bundler executable for submission.\")\n            print(\"Please download it and run:\")\n            print(\n                f'<bundle_exe> \"{outdir}/submission/test/clean/\" \"{outdir}/submission/test/final\" \"{outdir}/submission/bundled.lzma\"'\n            )\n\n\nclass SpringDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"Spring\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\", \"subtrain\", \"subval\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            pairname[1],\n            \"frame_\" + pairname[3],\n            \"frame_{:s}_{:04d}.png\".format(pairname[3], pairname[4]),\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            pairname[1],\n            \"frame_\" + pairname[3],","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.SpringDataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_flow.SpringDataset#L399-L533","kind":"class","name":"SpringDataset","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":399,"end_line":533,"context_start_line":379,"context_end_line":553,"code":"        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowFile(prediction, outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test_allpass\"\n        bundle_exe = \"/nfs/data/ffs-3d/datasets/StereoFlow/MPI-Sintel/bundler/linux-x64/bundler\"  # eg <bundle_exe> <path_to_results_for_clean> <path_to_results_for_final> <output/bundled.lzma>\n        if os.path.isfile(bundle_exe):\n            cmd = f'{bundle_exe} \"{outdir}/submission/test/clean/\" \"{outdir}/submission/test/final\" \"{outdir}/submission/bundled.lzma\"'\n            print(cmd)\n            os.system(cmd)\n            print(f'Done. Submission file at: \"{outdir}/submission/bundled.lzma\"')\n        else:\n            print(\"Could not find bundler executable for submission.\")\n            print(\"Please download it and run:\")\n            print(\n                f'<bundle_exe> \"{outdir}/submission/test/clean/\" \"{outdir}/submission/test/final\" \"{outdir}/submission/bundled.lzma\"'\n            )\n\n\nclass SpringDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"Spring\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\", \"subtrain\", \"subval\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            pairname[1],\n            \"frame_\" + pairname[3],\n            \"frame_{:s}_{:04d}.png\".format(pairname[3], pairname[4]),\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            pairname[1],\n            \"frame_\" + pairname[3],\n            \"frame_{:s}_{:04d}.png\".format(\n                pairname[3], pairname[4] + (1 if pairname[2] == \"FW\" else -1)\n            ),\n        )\n        self.pairname_to_flowname = lambda pairname: (\n            None\n            if pairname[0] == \"test\"\n            else osp.join(\n                self.root,\n                pairname[0],\n                pairname[1],\n                f\"flow_{pairname[2]}_{pairname[3]}\",\n                f\"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}.flo5\",\n            )\n        )\n        self.pairname_to_str = lambda pairname: osp.join(\n            pairname[0],\n            pairname[1],\n            f\"flow_{pairname[2]}_{pairname[3]}\",\n            f\"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}\",\n        )\n        self.load_flow = _read_hdf5_flow\n\n    def _build_cache(self):\n        # train\n        trainseqs = sorted(os.listdir(osp.join(self.root, \"train\")))\n        trainpairs = []\n        for leftright in [\"left\", \"right\"]:\n            for fwbw in [\"FW\", \"BW\"]:\n                trainpairs += [\n                    (\n                        \"train\",\n                        s,\n                        fwbw,\n                        leftright,\n                        int(f[len(f\"flow_{fwbw}_{leftright}_\") : -len(\".flo5\")]),\n                    )\n                    for s in trainseqs\n                    for f in sorted(\n                        os.listdir(\n                            osp.join(self.root, \"train\", s, f\"flow_{fwbw}_{leftright}\")\n                        )\n                    )\n                ]\n        # test\n        testseqs = sorted(os.listdir(osp.join(self.root, \"test\")))\n        testpairs = []\n        for leftright in [\"left\", \"right\"]:\n            testpairs += [\n                (\n                    \"test\",\n                    s,\n                    \"FW\",\n                    leftright,\n                    int(f[len(f\"frame_{leftright}_\") : -len(\".png\")]),\n                )\n                for s in testseqs\n                for f in sorted(\n                    os.listdir(osp.join(self.root, \"test\", s, f\"frame_{leftright}\"))\n                )[:-1]\n            ]\n            testpairs += [\n                (\n                    \"test\",\n                    s,\n                    \"BW\",\n                    leftright,\n                    int(f[len(f\"frame_{leftright}_\") : -len(\".png\")]) + 1,\n                )\n                for s in testseqs\n                for f in sorted(\n                    os.listdir(osp.join(self.root, \"test\", s, f\"frame_{leftright}\"))\n                )[:-1]\n            ]\n        # subtrain / subval\n        subtrainpairs = [p for p in trainpairs if p[1] != \"0041\"]\n        subvalpairs = [p for p in trainpairs if p[1] == \"0041\"]\n        assert (\n            len(trainpairs) == 19852\n            and len(testpairs) == 3960\n            and len(subtrainpairs) == 19472\n            and len(subvalpairs) == 380\n        ), \"incorrect parsing of pairs in Spring\"\n        tosave = {\n            \"train\": trainpairs,\n            \"test\": testpairs,\n            \"subtrain\": subtrainpairs,\n            \"subval\": subvalpairs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        assert prediction.dtype == np.float32\n        outfile = osp.join(\n            outdir,\n            pairname[0],\n            pairname[1],\n            f\"flow_{pairname[2]}_{pairname[3]}\",\n            f\"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}.flo5\",\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlo5File(prediction, outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        exe = \"{self.root}/flow_subsampling\"\n        if os.path.isfile(exe):\n            cmd = f'cd \"{outdir}/test\"; {exe} .'\n            print(cmd)\n            os.system(cmd)\n            print(f\"Done. Submission file at {outdir}/test/flow_submission.hdf5\")\n        else:\n            print(\"Could not find flow_subsampling executable for submission.\")\n            print(\"Please download it and run:\")\n            print(f'cd \"{outdir}/test\"; <flow_subsampling_exe> .')\n\n\nclass Kitti12Dataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti12\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname + \"_11.png\"\n        )\n        self.pairname_to_flowname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/colored_0/\", \"/flow_occ/\") + \"_10.png\"\n            )","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.Kitti12Dataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_flow.Kitti12Dataset#L536-L579","kind":"class","name":"Kitti12Dataset","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":536,"end_line":579,"context_start_line":516,"context_end_line":599,"code":"            f\"flow_{pairname[2]}_{pairname[3]}\",\n            f\"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}.flo5\",\n        )\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlo5File(prediction, outfile)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        exe = \"{self.root}/flow_subsampling\"\n        if os.path.isfile(exe):\n            cmd = f'cd \"{outdir}/test\"; {exe} .'\n            print(cmd)\n            os.system(cmd)\n            print(f\"Done. Submission file at {outdir}/test/flow_submission.hdf5\")\n        else:\n            print(\"Could not find flow_subsampling executable for submission.\")\n            print(\"Please download it and run:\")\n            print(f'cd \"{outdir}/test\"; <flow_subsampling_exe> .')\n\n\nclass Kitti12Dataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti12\"\n        self._set_root()\n        assert self.split in [\"train\", \"test\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname + \"_11.png\"\n        )\n        self.pairname_to_flowname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/colored_0/\", \"/flow_occ/\") + \"_10.png\"\n            )\n        )\n        self.pairname_to_str = lambda pairname: pairname.replace(\"/colored_0/\", \"/\")\n        self.load_flow = _read_kitti_flow\n\n    def _build_cache(self):\n        trainseqs = [\"training/colored_0/%06d\" % (i) for i in range(194)]\n        testseqs = [\"testing/colored_0/%06d\" % (i) for i in range(195)]\n        assert (\n            len(trainseqs) == 194 and len(testseqs) == 195\n        ), \"incorrect parsing of pairs in Kitti12\"\n        tosave = {\"train\": trainseqs, \"test\": testseqs}\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(outdir, pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowKitti(outfile, prediction)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti12_flow_results.zip\" .'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti12_flow_results.zip\")\n\n\nclass Kitti15Dataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti15\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\", \"test\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname + \"_11.png\"\n        )\n        self.pairname_to_flowname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/image_2/\", \"/flow_occ/\") + \"_10.png\"\n            )","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.Kitti15Dataset","uri":"program://Human3R/class/src.croco.stereoflow.datasets_flow.Kitti15Dataset#L582-L635","kind":"class","name":"Kitti15Dataset","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":582,"end_line":635,"context_start_line":562,"context_end_line":655,"code":"            len(trainseqs) == 194 and len(testseqs) == 195\n        ), \"incorrect parsing of pairs in Kitti12\"\n        tosave = {\"train\": trainseqs, \"test\": testseqs}\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(outdir, pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowKitti(outfile, prediction)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti12_flow_results.zip\" .'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti12_flow_results.zip\")\n\n\nclass Kitti15Dataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti15\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\", \"test\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname + \"_11.png\"\n        )\n        self.pairname_to_flowname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/image_2/\", \"/flow_occ/\") + \"_10.png\"\n            )\n        )\n        self.pairname_to_str = lambda pairname: pairname.replace(\"/image_2/\", \"/\")\n        self.load_flow = _read_kitti_flow\n\n    def _build_cache(self):\n        trainseqs = [\"training/image_2/%06d\" % (i) for i in range(200)]\n        subtrainseqs = trainseqs[:-10]\n        subvalseqs = trainseqs[-10:]\n        testseqs = [\"testing/image_2/%06d\" % (i) for i in range(200)]\n        assert (\n            len(trainseqs) == 200\n            and len(subtrainseqs) == 190\n            and len(subvalseqs) == 10\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(outdir, \"flow\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowKitti(outfile, prediction)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_flow_results.zip\" flow'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_flow_results.zip\")\n\n\nimport cv2\n\n\ndef _read_numpy_flow(filename):\n    return np.load(filename)\n\n\ndef _read_pfm_flow(filename):\n    f, _ = _read_pfm(filename)\n    assert np.all(f[:, :, 2] == 0.0)\n    return np.ascontiguousarray(f[:, :, :2])\n\n\nTAG_FLOAT = 202021.25  # tag to check the sanity of the file\nTAG_STRING = \"PIEH\"  # string containing the tag\nMIN_WIDTH = 1\nMAX_WIDTH = 99999\nMIN_HEIGHT = 1","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow._read_numpy_flow","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow._read_numpy_flow#L641-L642","kind":"function","name":"_read_numpy_flow","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":641,"end_line":642,"context_start_line":621,"context_end_line":662,"code":"        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(outdir, \"flow\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowKitti(outfile, prediction)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_flow_results.zip\" flow'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_flow_results.zip\")\n\n\nimport cv2\n\n\ndef _read_numpy_flow(filename):\n    return np.load(filename)\n\n\ndef _read_pfm_flow(filename):\n    f, _ = _read_pfm(filename)\n    assert np.all(f[:, :, 2] == 0.0)\n    return np.ascontiguousarray(f[:, :, :2])\n\n\nTAG_FLOAT = 202021.25  # tag to check the sanity of the file\nTAG_STRING = \"PIEH\"  # string containing the tag\nMIN_WIDTH = 1\nMAX_WIDTH = 99999\nMIN_HEIGHT = 1\nMAX_HEIGHT = 99999\n\n\ndef readFlowFile(filename):\n    \"\"\"\n    readFlowFile(<FILENAME>) reads a flow file <FILENAME> into a 2-band np.array.\n    if <FILENAME> does not exist, an IOError is raised.","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow._read_pfm_flow","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow._read_pfm_flow#L645-L648","kind":"function","name":"_read_pfm_flow","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":645,"end_line":648,"context_start_line":625,"context_end_line":668,"code":"        assert prediction.shape[2] == 2\n        outfile = os.path.join(outdir, \"flow\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowKitti(outfile, prediction)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_flow_results.zip\" flow'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_flow_results.zip\")\n\n\nimport cv2\n\n\ndef _read_numpy_flow(filename):\n    return np.load(filename)\n\n\ndef _read_pfm_flow(filename):\n    f, _ = _read_pfm(filename)\n    assert np.all(f[:, :, 2] == 0.0)\n    return np.ascontiguousarray(f[:, :, :2])\n\n\nTAG_FLOAT = 202021.25  # tag to check the sanity of the file\nTAG_STRING = \"PIEH\"  # string containing the tag\nMIN_WIDTH = 1\nMAX_WIDTH = 99999\nMIN_HEIGHT = 1\nMAX_HEIGHT = 99999\n\n\ndef readFlowFile(filename):\n    \"\"\"\n    readFlowFile(<FILENAME>) reads a flow file <FILENAME> into a 2-band np.array.\n    if <FILENAME> does not exist, an IOError is raised.\n    if <FILENAME> does not finish by '.flo' or the tag, the width, the height or the file's size is illegal, an Expcetion is raised.\n    ---- PARAMETERS ----\n        filename: string containg the name of the file to read a flow\n    ---- OUTPUTS ----\n        a np.array of dimension (height x width x 2) containing the flow of type 'float32'\n    \"\"\"","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.readFlowFile","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.readFlowFile#L659-L700","kind":"function","name":"readFlowFile","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":659,"end_line":700,"context_start_line":639,"context_end_line":720,"code":"\n\ndef _read_numpy_flow(filename):\n    return np.load(filename)\n\n\ndef _read_pfm_flow(filename):\n    f, _ = _read_pfm(filename)\n    assert np.all(f[:, :, 2] == 0.0)\n    return np.ascontiguousarray(f[:, :, :2])\n\n\nTAG_FLOAT = 202021.25  # tag to check the sanity of the file\nTAG_STRING = \"PIEH\"  # string containing the tag\nMIN_WIDTH = 1\nMAX_WIDTH = 99999\nMIN_HEIGHT = 1\nMAX_HEIGHT = 99999\n\n\ndef readFlowFile(filename):\n    \"\"\"\n    readFlowFile(<FILENAME>) reads a flow file <FILENAME> into a 2-band np.array.\n    if <FILENAME> does not exist, an IOError is raised.\n    if <FILENAME> does not finish by '.flo' or the tag, the width, the height or the file's size is illegal, an Expcetion is raised.\n    ---- PARAMETERS ----\n        filename: string containg the name of the file to read a flow\n    ---- OUTPUTS ----\n        a np.array of dimension (height x width x 2) containing the flow of type 'float32'\n    \"\"\"\n\n    # check filename\n    if not filename.endswith(\".flo\"):\n        raise Exception(\n            \"readFlowFile({:s}): filename must finish with '.flo'\".format(filename)\n        )\n\n    # open the file and read it\n    with open(filename, \"rb\") as f:\n        # check tag\n        tag = struct.unpack(\"f\", f.read(4))[0]\n        if tag != TAG_FLOAT:\n            raise Exception(\"flow_utils.readFlowFile({:s}): wrong tag\".format(filename))\n        # read dimension\n        w, h = struct.unpack(\"ii\", f.read(8))\n        if w < MIN_WIDTH or w > MAX_WIDTH:\n            raise Exception(\n                \"flow_utils.readFlowFile({:s}: illegal width {:d}\".format(filename, w)\n            )\n        if h < MIN_HEIGHT or h > MAX_HEIGHT:\n            raise Exception(\n                \"flow_utils.readFlowFile({:s}: illegal height {:d}\".format(filename, h)\n            )\n        flow = np.fromfile(f, \"float32\")\n        if not flow.shape == (h * w * 2,):\n            raise Exception(\n                \"flow_utils.readFlowFile({:s}: illegal size of the file\".format(\n                    filename\n                )\n            )\n        flow.shape = (h, w, 2)\n        return flow\n\n\ndef writeFlowFile(flow, filename):\n    \"\"\"\n    writeFlowFile(flow,<FILENAME>) write flow to the file <FILENAME>.\n    if <FILENAME> does not exist, an IOError is raised.\n    if <FILENAME> does not finish with '.flo' or the flow has not 2 bands, an Exception is raised.\n    ---- PARAMETERS ----\n        flow: np.array of dimension (height x width x 2) containing the flow to write\n        filename: string containg the name of the file to write a flow\n    \"\"\"\n\n    # check filename\n    if not filename.endswith(\".flo\"):\n        raise Exception(\n            \"flow_utils.writeFlowFile(<flow>,{:s}): filename must finish with '.flo'\".format(\n                filename\n            )\n        )\n","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.writeFlowFile","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.writeFlowFile#L703-L736","kind":"function","name":"writeFlowFile","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":703,"end_line":736,"context_start_line":683,"context_end_line":756,"code":"        w, h = struct.unpack(\"ii\", f.read(8))\n        if w < MIN_WIDTH or w > MAX_WIDTH:\n            raise Exception(\n                \"flow_utils.readFlowFile({:s}: illegal width {:d}\".format(filename, w)\n            )\n        if h < MIN_HEIGHT or h > MAX_HEIGHT:\n            raise Exception(\n                \"flow_utils.readFlowFile({:s}: illegal height {:d}\".format(filename, h)\n            )\n        flow = np.fromfile(f, \"float32\")\n        if not flow.shape == (h * w * 2,):\n            raise Exception(\n                \"flow_utils.readFlowFile({:s}: illegal size of the file\".format(\n                    filename\n                )\n            )\n        flow.shape = (h, w, 2)\n        return flow\n\n\ndef writeFlowFile(flow, filename):\n    \"\"\"\n    writeFlowFile(flow,<FILENAME>) write flow to the file <FILENAME>.\n    if <FILENAME> does not exist, an IOError is raised.\n    if <FILENAME> does not finish with '.flo' or the flow has not 2 bands, an Exception is raised.\n    ---- PARAMETERS ----\n        flow: np.array of dimension (height x width x 2) containing the flow to write\n        filename: string containg the name of the file to write a flow\n    \"\"\"\n\n    # check filename\n    if not filename.endswith(\".flo\"):\n        raise Exception(\n            \"flow_utils.writeFlowFile(<flow>,{:s}): filename must finish with '.flo'\".format(\n                filename\n            )\n        )\n\n    if not flow.shape[2:] == (2,):\n        raise Exception(\n            \"flow_utils.writeFlowFile(<flow>,{:s}): <flow> must have 2 bands\".format(\n                filename\n            )\n        )\n\n    # open the file and write it\n    with open(filename, \"wb\") as f:\n        # write TAG\n        f.write(TAG_STRING.encode(\"utf-8\"))\n        # write dimension\n        f.write(struct.pack(\"ii\", flow.shape[1], flow.shape[0]))\n        # write the flow\n\n        flow.astype(np.float32).tofile(f)\n\n\n_read_flo_file = readFlowFile\n\n\ndef _read_kitti_flow(filename):\n    flow = cv2.imread(filename, cv2.IMREAD_ANYDEPTH | cv2.IMREAD_COLOR)\n    flow = flow[:, :, ::-1].astype(np.float32)\n    valid = flow[:, :, 2] > 0\n    flow = flow[:, :, :2]\n    flow = (flow - 2**15) / 64.0\n    flow[~valid, 0] = np.inf\n    flow[~valid, 1] = np.inf\n    return flow\n\n\n_read_hd1k_flow = _read_kitti_flow\n\n\ndef writeFlowKitti(filename, uv):","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow._read_kitti_flow","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow._read_kitti_flow#L742-L750","kind":"function","name":"_read_kitti_flow","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":742,"end_line":750,"context_start_line":722,"context_end_line":770,"code":"        raise Exception(\n            \"flow_utils.writeFlowFile(<flow>,{:s}): <flow> must have 2 bands\".format(\n                filename\n            )\n        )\n\n    # open the file and write it\n    with open(filename, \"wb\") as f:\n        # write TAG\n        f.write(TAG_STRING.encode(\"utf-8\"))\n        # write dimension\n        f.write(struct.pack(\"ii\", flow.shape[1], flow.shape[0]))\n        # write the flow\n\n        flow.astype(np.float32).tofile(f)\n\n\n_read_flo_file = readFlowFile\n\n\ndef _read_kitti_flow(filename):\n    flow = cv2.imread(filename, cv2.IMREAD_ANYDEPTH | cv2.IMREAD_COLOR)\n    flow = flow[:, :, ::-1].astype(np.float32)\n    valid = flow[:, :, 2] > 0\n    flow = flow[:, :, :2]\n    flow = (flow - 2**15) / 64.0\n    flow[~valid, 0] = np.inf\n    flow[~valid, 1] = np.inf\n    return flow\n\n\n_read_hd1k_flow = _read_kitti_flow\n\n\ndef writeFlowKitti(filename, uv):\n    uv = 64.0 * uv + 2**15\n    valid = np.ones([uv.shape[0], uv.shape[1], 1])\n    uv = np.concatenate([uv, valid], axis=-1).astype(np.uint16)\n    cv2.imwrite(filename, uv[..., ::-1])\n\n\ndef writeFlo5File(flow, filename):\n    with h5py.File(filename, \"w\") as f:\n        f.create_dataset(\"flow\", data=flow, compression=\"gzip\", compression_opts=5)\n\n\ndef _read_hdf5_flow(filename):\n    flow = np.asarray(h5py.File(filename)[\"flow\"])\n    flow[np.isnan(flow)] = np.inf  # make invalid values as +inf","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.writeFlowKitti","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.writeFlowKitti#L756-L760","kind":"function","name":"writeFlowKitti","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":756,"end_line":760,"context_start_line":736,"context_end_line":780,"code":"        flow.astype(np.float32).tofile(f)\n\n\n_read_flo_file = readFlowFile\n\n\ndef _read_kitti_flow(filename):\n    flow = cv2.imread(filename, cv2.IMREAD_ANYDEPTH | cv2.IMREAD_COLOR)\n    flow = flow[:, :, ::-1].astype(np.float32)\n    valid = flow[:, :, 2] > 0\n    flow = flow[:, :, :2]\n    flow = (flow - 2**15) / 64.0\n    flow[~valid, 0] = np.inf\n    flow[~valid, 1] = np.inf\n    return flow\n\n\n_read_hd1k_flow = _read_kitti_flow\n\n\ndef writeFlowKitti(filename, uv):\n    uv = 64.0 * uv + 2**15\n    valid = np.ones([uv.shape[0], uv.shape[1], 1])\n    uv = np.concatenate([uv, valid], axis=-1).astype(np.uint16)\n    cv2.imwrite(filename, uv[..., ::-1])\n\n\ndef writeFlo5File(flow, filename):\n    with h5py.File(filename, \"w\") as f:\n        f.create_dataset(\"flow\", data=flow, compression=\"gzip\", compression_opts=5)\n\n\ndef _read_hdf5_flow(filename):\n    flow = np.asarray(h5py.File(filename)[\"flow\"])\n    flow[np.isnan(flow)] = np.inf  # make invalid values as +inf\n    return flow.astype(np.float32)\n\n\n# flow visualization\nRY = 15\nYG = 6\nGC = 4\nCB = 11\nBM = 13\nMR = 6","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.writeFlo5File","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.writeFlo5File#L763-L765","kind":"function","name":"writeFlo5File","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":763,"end_line":765,"context_start_line":743,"context_end_line":785,"code":"    flow = cv2.imread(filename, cv2.IMREAD_ANYDEPTH | cv2.IMREAD_COLOR)\n    flow = flow[:, :, ::-1].astype(np.float32)\n    valid = flow[:, :, 2] > 0\n    flow = flow[:, :, :2]\n    flow = (flow - 2**15) / 64.0\n    flow[~valid, 0] = np.inf\n    flow[~valid, 1] = np.inf\n    return flow\n\n\n_read_hd1k_flow = _read_kitti_flow\n\n\ndef writeFlowKitti(filename, uv):\n    uv = 64.0 * uv + 2**15\n    valid = np.ones([uv.shape[0], uv.shape[1], 1])\n    uv = np.concatenate([uv, valid], axis=-1).astype(np.uint16)\n    cv2.imwrite(filename, uv[..., ::-1])\n\n\ndef writeFlo5File(flow, filename):\n    with h5py.File(filename, \"w\") as f:\n        f.create_dataset(\"flow\", data=flow, compression=\"gzip\", compression_opts=5)\n\n\ndef _read_hdf5_flow(filename):\n    flow = np.asarray(h5py.File(filename)[\"flow\"])\n    flow[np.isnan(flow)] = np.inf  # make invalid values as +inf\n    return flow.astype(np.float32)\n\n\n# flow visualization\nRY = 15\nYG = 6\nGC = 4\nCB = 11\nBM = 13\nMR = 6\nUNKNOWN_THRESH = 1e9\n\n\ndef colorTest():\n    \"\"\"","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow._read_hdf5_flow","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow._read_hdf5_flow#L768-L771","kind":"function","name":"_read_hdf5_flow","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":768,"end_line":771,"context_start_line":748,"context_end_line":791,"code":"    flow[~valid, 0] = np.inf\n    flow[~valid, 1] = np.inf\n    return flow\n\n\n_read_hd1k_flow = _read_kitti_flow\n\n\ndef writeFlowKitti(filename, uv):\n    uv = 64.0 * uv + 2**15\n    valid = np.ones([uv.shape[0], uv.shape[1], 1])\n    uv = np.concatenate([uv, valid], axis=-1).astype(np.uint16)\n    cv2.imwrite(filename, uv[..., ::-1])\n\n\ndef writeFlo5File(flow, filename):\n    with h5py.File(filename, \"w\") as f:\n        f.create_dataset(\"flow\", data=flow, compression=\"gzip\", compression_opts=5)\n\n\ndef _read_hdf5_flow(filename):\n    flow = np.asarray(h5py.File(filename)[\"flow\"])\n    flow[np.isnan(flow)] = np.inf  # make invalid values as +inf\n    return flow.astype(np.float32)\n\n\n# flow visualization\nRY = 15\nYG = 6\nGC = 4\nCB = 11\nBM = 13\nMR = 6\nUNKNOWN_THRESH = 1e9\n\n\ndef colorTest():\n    \"\"\"\n    flow_utils.colorTest(): display an example of image showing the color encoding scheme\n    \"\"\"\n    import matplotlib.pylab as plt\n\n    truerange = 1\n    h, w = 151, 151","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.colorTest","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.colorTest#L784-L805","kind":"function","name":"colorTest","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":784,"end_line":805,"context_start_line":764,"context_end_line":825,"code":"    with h5py.File(filename, \"w\") as f:\n        f.create_dataset(\"flow\", data=flow, compression=\"gzip\", compression_opts=5)\n\n\ndef _read_hdf5_flow(filename):\n    flow = np.asarray(h5py.File(filename)[\"flow\"])\n    flow[np.isnan(flow)] = np.inf  # make invalid values as +inf\n    return flow.astype(np.float32)\n\n\n# flow visualization\nRY = 15\nYG = 6\nGC = 4\nCB = 11\nBM = 13\nMR = 6\nUNKNOWN_THRESH = 1e9\n\n\ndef colorTest():\n    \"\"\"\n    flow_utils.colorTest(): display an example of image showing the color encoding scheme\n    \"\"\"\n    import matplotlib.pylab as plt\n\n    truerange = 1\n    h, w = 151, 151\n    trange = truerange * 1.04\n    s2 = round(h / 2)\n    x, y = np.meshgrid(range(w), range(h))\n    u = x * trange / s2 - trange\n    v = y * trange / s2 - trange\n    img = _computeColor(\n        np.concatenate((u[:, :, np.newaxis], v[:, :, np.newaxis]), 2)\n        / trange\n        / np.sqrt(2)\n    )\n    plt.imshow(img)\n    plt.axis(\"off\")\n    plt.axhline(round(h / 2), color=\"k\")\n    plt.axvline(round(w / 2), color=\"k\")\n\n\ndef flowToColor(flow, maxflow=None, maxmaxflow=None, saturate=False):\n    \"\"\"\n    flow_utils.flowToColor(flow): return a color code flow field, normalized based on the maximum l2-norm of the flow\n    flow_utils.flowToColor(flow,maxflow): return a color code flow field, normalized by maxflow\n    ---- PARAMETERS ----\n        flow: flow to display of shape (height x width x 2)\n        maxflow (default:None): if given, normalize the flow by its value, otherwise by the flow norm\n        maxmaxflow (default:None): if given, normalize the flow by the max of its value and the flow norm\n    ---- OUTPUT ----\n        an np.array of shape (height x width x 3) of type uint8 containing a color code of the flow\n    \"\"\"\n    h, w, n = flow.shape\n    # check size of flow\n    assert n == 2, \"flow_utils.flowToColor(flow): flow must have 2 bands\"\n    # fix unknown flow\n    unknown_idx = np.max(np.abs(flow), 2) > UNKNOWN_THRESH\n    flow[unknown_idx] = 0.0\n    # compute max flow if needed","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.flowToColor","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.flowToColor#L808-L836","kind":"function","name":"flowToColor","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":808,"end_line":836,"context_start_line":788,"context_end_line":856,"code":"    import matplotlib.pylab as plt\n\n    truerange = 1\n    h, w = 151, 151\n    trange = truerange * 1.04\n    s2 = round(h / 2)\n    x, y = np.meshgrid(range(w), range(h))\n    u = x * trange / s2 - trange\n    v = y * trange / s2 - trange\n    img = _computeColor(\n        np.concatenate((u[:, :, np.newaxis], v[:, :, np.newaxis]), 2)\n        / trange\n        / np.sqrt(2)\n    )\n    plt.imshow(img)\n    plt.axis(\"off\")\n    plt.axhline(round(h / 2), color=\"k\")\n    plt.axvline(round(w / 2), color=\"k\")\n\n\ndef flowToColor(flow, maxflow=None, maxmaxflow=None, saturate=False):\n    \"\"\"\n    flow_utils.flowToColor(flow): return a color code flow field, normalized based on the maximum l2-norm of the flow\n    flow_utils.flowToColor(flow,maxflow): return a color code flow field, normalized by maxflow\n    ---- PARAMETERS ----\n        flow: flow to display of shape (height x width x 2)\n        maxflow (default:None): if given, normalize the flow by its value, otherwise by the flow norm\n        maxmaxflow (default:None): if given, normalize the flow by the max of its value and the flow norm\n    ---- OUTPUT ----\n        an np.array of shape (height x width x 3) of type uint8 containing a color code of the flow\n    \"\"\"\n    h, w, n = flow.shape\n    # check size of flow\n    assert n == 2, \"flow_utils.flowToColor(flow): flow must have 2 bands\"\n    # fix unknown flow\n    unknown_idx = np.max(np.abs(flow), 2) > UNKNOWN_THRESH\n    flow[unknown_idx] = 0.0\n    # compute max flow if needed\n    if maxflow is None:\n        maxflow = flowMaxNorm(flow)\n    if maxmaxflow is not None:\n        maxflow = min(maxmaxflow, maxflow)\n    # normalize flow\n    eps = np.spacing(1)  # minimum positive float value to avoid division by 0\n    # compute the flow\n    img = _computeColor(flow / (maxflow + eps), saturate=saturate)\n    # put black pixels in unknown location\n    img[np.tile(unknown_idx[:, :, np.newaxis], [1, 1, 3])] = 0.0\n    return img\n\n\ndef flowMaxNorm(flow):\n    \"\"\"\n    flow_utils.flowMaxNorm(flow): return the maximum of the l2-norm of the given flow\n    ---- PARAMETERS ----\n        flow: the flow\n\n    ---- OUTPUT ----\n        a float containing the maximum of the l2-norm of the flow\n    \"\"\"\n    return np.max(np.sqrt(np.sum(np.square(flow), 2)))\n\n\ndef _computeColor(flow, saturate=True):\n    \"\"\"\n    flow_utils._computeColor(flow): compute color codes for the flow field flow\n\n    ---- PARAMETERS ----\n        flow: np.array of dimension (height x width x 2) containing the flow to display","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.flowMaxNorm","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.flowMaxNorm#L839-L848","kind":"function","name":"flowMaxNorm","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":839,"end_line":848,"context_start_line":819,"context_end_line":868,"code":"    h, w, n = flow.shape\n    # check size of flow\n    assert n == 2, \"flow_utils.flowToColor(flow): flow must have 2 bands\"\n    # fix unknown flow\n    unknown_idx = np.max(np.abs(flow), 2) > UNKNOWN_THRESH\n    flow[unknown_idx] = 0.0\n    # compute max flow if needed\n    if maxflow is None:\n        maxflow = flowMaxNorm(flow)\n    if maxmaxflow is not None:\n        maxflow = min(maxmaxflow, maxflow)\n    # normalize flow\n    eps = np.spacing(1)  # minimum positive float value to avoid division by 0\n    # compute the flow\n    img = _computeColor(flow / (maxflow + eps), saturate=saturate)\n    # put black pixels in unknown location\n    img[np.tile(unknown_idx[:, :, np.newaxis], [1, 1, 3])] = 0.0\n    return img\n\n\ndef flowMaxNorm(flow):\n    \"\"\"\n    flow_utils.flowMaxNorm(flow): return the maximum of the l2-norm of the given flow\n    ---- PARAMETERS ----\n        flow: the flow\n\n    ---- OUTPUT ----\n        a float containing the maximum of the l2-norm of the flow\n    \"\"\"\n    return np.max(np.sqrt(np.sum(np.square(flow), 2)))\n\n\ndef _computeColor(flow, saturate=True):\n    \"\"\"\n    flow_utils._computeColor(flow): compute color codes for the flow field flow\n\n    ---- PARAMETERS ----\n        flow: np.array of dimension (height x width x 2) containing the flow to display\n    ---- OUTPUTS ----\n        an np.array of dimension (height x width x 3) containing the color conversion of the flow\n    \"\"\"\n    # set nan to 0\n    nanidx = np.isnan(flow[:, :, 0])\n    flow[nanidx] = 0.0\n\n    # colorwheel\n    ncols = RY + YG + GC + CB + BM + MR\n    nchans = 3\n    colorwheel = np.zeros((ncols, nchans), \"uint8\")\n    col = 0","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow._computeColor","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow._computeColor#L851-L917","kind":"function","name":"_computeColor","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":851,"end_line":917,"context_start_line":831,"context_end_line":936,"code":"    eps = np.spacing(1)  # minimum positive float value to avoid division by 0\n    # compute the flow\n    img = _computeColor(flow / (maxflow + eps), saturate=saturate)\n    # put black pixels in unknown location\n    img[np.tile(unknown_idx[:, :, np.newaxis], [1, 1, 3])] = 0.0\n    return img\n\n\ndef flowMaxNorm(flow):\n    \"\"\"\n    flow_utils.flowMaxNorm(flow): return the maximum of the l2-norm of the given flow\n    ---- PARAMETERS ----\n        flow: the flow\n\n    ---- OUTPUT ----\n        a float containing the maximum of the l2-norm of the flow\n    \"\"\"\n    return np.max(np.sqrt(np.sum(np.square(flow), 2)))\n\n\ndef _computeColor(flow, saturate=True):\n    \"\"\"\n    flow_utils._computeColor(flow): compute color codes for the flow field flow\n\n    ---- PARAMETERS ----\n        flow: np.array of dimension (height x width x 2) containing the flow to display\n    ---- OUTPUTS ----\n        an np.array of dimension (height x width x 3) containing the color conversion of the flow\n    \"\"\"\n    # set nan to 0\n    nanidx = np.isnan(flow[:, :, 0])\n    flow[nanidx] = 0.0\n\n    # colorwheel\n    ncols = RY + YG + GC + CB + BM + MR\n    nchans = 3\n    colorwheel = np.zeros((ncols, nchans), \"uint8\")\n    col = 0\n    # RY\n    colorwheel[:RY, 0] = 255\n    colorwheel[:RY, 1] = [(255 * i) // RY for i in range(RY)]\n    col += RY\n    # YG\n    colorwheel[col : col + YG, 0] = [255 - (255 * i) // YG for i in range(YG)]\n    colorwheel[col : col + YG, 1] = 255\n    col += YG\n    # GC\n    colorwheel[col : col + GC, 1] = 255\n    colorwheel[col : col + GC, 2] = [(255 * i) // GC for i in range(GC)]\n    col += GC\n    # CB\n    colorwheel[col : col + CB, 1] = [255 - (255 * i) // CB for i in range(CB)]\n    colorwheel[col : col + CB, 2] = 255\n    col += CB\n    # BM\n    colorwheel[col : col + BM, 0] = [(255 * i) // BM for i in range(BM)]\n    colorwheel[col : col + BM, 2] = 255\n    col += BM\n    # MR\n    colorwheel[col : col + MR, 0] = 255\n    colorwheel[col : col + MR, 2] = [255 - (255 * i) // MR for i in range(MR)]\n\n    # compute utility variables\n    rad = np.sqrt(np.sum(np.square(flow), 2))  # magnitude\n    a = np.arctan2(-flow[:, :, 1], -flow[:, :, 0]) / np.pi  # angle\n    fk = (a + 1) / 2 * (ncols - 1)  # map [-1,1] to [0,ncols-1]\n    k0 = np.floor(fk).astype(\"int\")\n    k1 = k0 + 1\n    k1[k1 == ncols] = 0\n    f = fk - k0\n\n    if not saturate:\n        rad = np.minimum(rad, 1)\n\n    # compute the image\n    img = np.zeros((flow.shape[0], flow.shape[1], nchans), \"uint8\")\n    for i in range(nchans):\n        tmp = colorwheel[:, i].astype(\"float\")\n        col0 = tmp[k0] / 255\n        col1 = tmp[k1] / 255\n        col = (1 - f) * col0 + f * col1\n        idx = rad <= 1\n        col[idx] = 1 - rad[idx] * (1 - col[idx])  # increase saturation with radius\n        col[~idx] *= 0.75  # out of range\n        img[:, :, i] = (255 * col * (1 - nanidx.astype(\"float\"))).astype(\"uint8\")\n\n    return img\n\n\n# flow dataset getter\n\n\ndef get_train_dataset_flow(dataset_str, augmentor=True, crop_size=None):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    if augmentor:\n        dataset_str = dataset_str.replace(\")\", \", augmentor=True)\")\n    if crop_size is not None:\n        dataset_str = dataset_str.replace(\n            \")\", \", crop_size={:s})\".format(str(crop_size))\n        )\n    return eval(dataset_str)\n\n\ndef get_test_datasets_flow(dataset_str):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    return [eval(s) for s in dataset_str.split(\"+\")]","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.get_train_dataset_flow","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.get_train_dataset_flow#L923-L931","kind":"function","name":"get_train_dataset_flow","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":923,"end_line":931,"context_start_line":903,"context_end_line":936,"code":"        rad = np.minimum(rad, 1)\n\n    # compute the image\n    img = np.zeros((flow.shape[0], flow.shape[1], nchans), \"uint8\")\n    for i in range(nchans):\n        tmp = colorwheel[:, i].astype(\"float\")\n        col0 = tmp[k0] / 255\n        col1 = tmp[k1] / 255\n        col = (1 - f) * col0 + f * col1\n        idx = rad <= 1\n        col[idx] = 1 - rad[idx] * (1 - col[idx])  # increase saturation with radius\n        col[~idx] *= 0.75  # out of range\n        img[:, :, i] = (255 * col * (1 - nanidx.astype(\"float\"))).astype(\"uint8\")\n\n    return img\n\n\n# flow dataset getter\n\n\ndef get_train_dataset_flow(dataset_str, augmentor=True, crop_size=None):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    if augmentor:\n        dataset_str = dataset_str.replace(\")\", \", augmentor=True)\")\n    if crop_size is not None:\n        dataset_str = dataset_str.replace(\n            \")\", \", crop_size={:s})\".format(str(crop_size))\n        )\n    return eval(dataset_str)\n\n\ndef get_test_datasets_flow(dataset_str):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    return [eval(s) for s in dataset_str.split(\"+\")]","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.get_test_datasets_flow","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.get_test_datasets_flow#L934-L936","kind":"function","name":"get_test_datasets_flow","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":934,"end_line":936,"context_start_line":914,"context_end_line":936,"code":"        col[~idx] *= 0.75  # out of range\n        img[:, :, i] = (255 * col * (1 - nanidx.astype(\"float\"))).astype(\"uint8\")\n\n    return img\n\n\n# flow dataset getter\n\n\ndef get_train_dataset_flow(dataset_str, augmentor=True, crop_size=None):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    if augmentor:\n        dataset_str = dataset_str.replace(\")\", \", augmentor=True)\")\n    if crop_size is not None:\n        dataset_str = dataset_str.replace(\n            \")\", \", crop_size={:s})\".format(str(crop_size))\n        )\n    return eval(dataset_str)\n\n\ndef get_test_datasets_flow(dataset_str):\n    dataset_str = dataset_str.replace(\"(\", \"Dataset(\")\n    return [eval(s) for s in dataset_str.split(\"+\")]","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.__init__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.__init__#L42-L55","kind":"function","name":"__init__","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":42,"end_line":55,"context_start_line":22,"context_end_line":75,"code":"\ndataset_to_root = deepcopy(dataset_to_root)\n\ndataset_to_root.update(\n    **{\n        \"TartanAir\": \"./data/stereoflow/TartanAir\",\n        \"FlyingChairs\": \"./data/stereoflow/FlyingChairs/\",\n        \"FlyingThings\": osp.join(dataset_to_root[\"SceneFlow\"], \"FlyingThings\") + \"/\",\n        \"MPISintel\": \"./data/stereoflow//MPI-Sintel/\" + \"/\",\n    }\n)\ncache_dir = \"./data/stereoflow/datasets_flow_cache/\"\n\n\ndef flow_to_tensor(disp):\n    return torch.from_numpy(disp).float().permute(2, 0, 1)\n\n\nclass FlowDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size is not None:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = FlowAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(\n            self.pairnames\n        )  # each pairname is typically of the form (str, int1, int2)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        img1name = self.pairname_to_img1name(pairname)\n        img2name = self.pairname_to_img2name(pairname)\n        flowname = (\n            self.pairname_to_flowname(pairname)","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.prepare_data","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.prepare_data#L57-L61","kind":"function","name":"prepare_data","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":57,"end_line":61,"context_start_line":37,"context_end_line":81,"code":"    return torch.from_numpy(disp).float().permute(2, 0, 1)\n\n\nclass FlowDataset(data.Dataset):\n\n    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):\n        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size is not None:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = FlowAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(\n            self.pairnames\n        )  # each pairname is typically of the form (str, int1, int2)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        img1name = self.pairname_to_img1name(pairname)\n        img2name = self.pairname_to_img2name(pairname)\n        flowname = (\n            self.pairname_to_flowname(pairname)\n            if self.pairname_to_flowname is not None\n            else None\n        )\n\n        # load images and disparities\n        img1 = _read_img(img1name)","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.__len__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.__len__#L63-L66","kind":"function","name":"__len__","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":63,"end_line":66,"context_start_line":43,"context_end_line":86,"code":"        self.split = split\n        if not augmentor:\n            assert crop_size is None\n        if crop_size is not None:\n            assert augmentor\n        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = FlowAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(\n            self.pairnames\n        )  # each pairname is typically of the form (str, int1, int2)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        img1name = self.pairname_to_img1name(pairname)\n        img2name = self.pairname_to_img2name(pairname)\n        flowname = (\n            self.pairname_to_flowname(pairname)\n            if self.pairname_to_flowname is not None\n            else None\n        )\n\n        # load images and disparities\n        img1 = _read_img(img1name)\n        img2 = _read_img(img2name)\n        flow = self.load_flow(flowname) if flowname is not None else None\n\n        # apply augmentations\n        if self.augmentor is not None:","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.__getitem__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.__getitem__#L68-L102","kind":"function","name":"__getitem__","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":68,"end_line":102,"context_start_line":48,"context_end_line":122,"code":"        self.crop_size = crop_size\n        self.augmentor_str = augmentor\n        self.augmentor = FlowAugmentor(crop_size) if augmentor else None\n        self.totensor = totensor\n        self.rmul = 1  # keep track of rmul\n        self.has_constant_resolution = True  # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)\n        self._prepare_data()\n        self._load_or_build_cache()\n\n    def prepare_data(self):\n        \"\"\"\n        to be defined for each dataset\n        \"\"\"\n        raise NotImplementedError\n\n    def __len__(self):\n        return len(\n            self.pairnames\n        )  # each pairname is typically of the form (str, int1, int2)\n\n    def __getitem__(self, index):\n        pairname = self.pairnames[index]\n\n        # get filenames\n        img1name = self.pairname_to_img1name(pairname)\n        img2name = self.pairname_to_img2name(pairname)\n        flowname = (\n            self.pairname_to_flowname(pairname)\n            if self.pairname_to_flowname is not None\n            else None\n        )\n\n        # load images and disparities\n        img1 = _read_img(img1name)\n        img2 = _read_img(img2name)\n        flow = self.load_flow(flowname) if flowname is not None else None\n\n        # apply augmentations\n        if self.augmentor is not None:\n            img1, img2, flow = self.augmentor(img1, img2, flow, self.name)\n\n        if self.totensor:\n            img1 = img_to_tensor(img1)\n            img2 = img_to_tensor(img2)\n            if flow is not None:\n                flow = flow_to_tensor(flow)\n            else:\n                flow = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            pairname = str(\n                pairname\n            )  # transform potential tuple to str to be able to batch it\n\n        return img1, img2, flow, pairname\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.__rmul__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.__rmul__#L104-L107","kind":"function","name":"__rmul__","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":104,"end_line":107,"context_start_line":84,"context_end_line":127,"code":"\n        # apply augmentations\n        if self.augmentor is not None:\n            img1, img2, flow = self.augmentor(img1, img2, flow, self.name)\n\n        if self.totensor:\n            img1 = img_to_tensor(img1)\n            img2 = img_to_tensor(img2)\n            if flow is not None:\n                flow = flow_to_tensor(flow)\n            else:\n                flow = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            pairname = str(\n                pairname\n            )  # transform potential tuple to str to be able to batch it\n\n        return img1, img2, flow, pairname\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.__str__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.__str__#L109-L110","kind":"function","name":"__str__","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":109,"end_line":110,"context_start_line":89,"context_end_line":130,"code":"        if self.totensor:\n            img1 = img_to_tensor(img1)\n            img2 = img_to_tensor(img2)\n            if flow is not None:\n                flow = flow_to_tensor(flow)\n            else:\n                flow = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            pairname = str(\n                pairname\n            )  # transform potential tuple to str to be able to batch it\n\n        return img1, img2, flow, pairname\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.__repr__","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.__repr__#L112-L118","kind":"function","name":"__repr__","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":112,"end_line":118,"context_start_line":92,"context_end_line":138,"code":"            if flow is not None:\n                flow = flow_to_tensor(flow)\n            else:\n                flow = torch.tensor(\n                    []\n                )  # to allow dataloader batching with default collate_gn\n            pairname = str(\n                pairname\n            )  # transform potential tuple to str to be able to batch it\n\n        return img1, img2, flow, pairname\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow._set_root","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow._set_root#L120-L124","kind":"function","name":"_set_root","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":120,"end_line":124,"context_start_line":100,"context_end_line":144,"code":"            )  # transform potential tuple to str to be able to batch it\n\n        return img1, img2, flow, pairname\n\n    def __rmul__(self, v):\n        self.rmul *= v\n        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass TartanAirDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"TartanAir\"\n        self._set_root()\n        assert self.split in [\"train\"]","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow._load_or_build_cache","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow._load_or_build_cache#L126-L136","kind":"function","name":"_load_or_build_cache","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":126,"end_line":136,"context_start_line":106,"context_end_line":156,"code":"        self.pairnames = v * self.pairnames\n        return self\n\n    def __str__(self):\n        return f\"{self.__class__.__name__}_{self.split}\"\n\n    def __repr__(self):\n        s = f\"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})\"\n        if self.rmul == 1:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)}\"\n        else:\n            s += f\"\\n\\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})\"\n        return s\n\n    def _set_root(self):\n        self.root = dataset_to_root[self.name]\n        assert os.path.isdir(\n            self.root\n        ), f\"could not find root directory for dataset {self.name}: {self.root}\"\n\n    def _load_or_build_cache(self):\n        cache_file = osp.join(cache_dir, self.name + \".pkl\")\n        if osp.isfile(cache_file):\n            with open(cache_file, \"rb\") as fid:\n                self.pairnames = pickle.load(fid)[self.split]\n        else:\n            tosave = self._build_cache()\n            os.makedirs(cache_dir, exist_ok=True)\n            with open(cache_file, \"wb\") as fid:\n                pickle.dump(tosave, fid)\n            self.pairnames = tosave[self.split]\n\n\nclass TartanAirDataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"TartanAir\"\n        self._set_root()\n        assert self.split in [\"train\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname[0], \"image_left/{:06d}_left.png\".format(pairname[1])\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname[0], \"image_left/{:06d}_left.png\".format(pairname[2])\n        )\n        self.pairname_to_flowname = lambda pairname: osp.join(\n            self.root,\n            pairname[0],\n            \"flow/{:06d}_{:06d}_flow.npy\".format(pairname[1], pairname[2]),\n        )\n        self.pairname_to_str = lambda pairname: os.path.join(","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow._prepare_data","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow._prepare_data#L584-L602","kind":"function","name":"_prepare_data","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":584,"end_line":602,"context_start_line":564,"context_end_line":622,"code":"        tosave = {\"train\": trainseqs, \"test\": testseqs}\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(outdir, pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowKitti(outfile, prediction)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti12_flow_results.zip\" .'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti12_flow_results.zip\")\n\n\nclass Kitti15Dataset(FlowDataset):\n\n    def _prepare_data(self):\n        self.name = \"Kitti15\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\", \"test\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname + \"_11.png\"\n        )\n        self.pairname_to_flowname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/image_2/\", \"/flow_occ/\") + \"_10.png\"\n            )\n        )\n        self.pairname_to_str = lambda pairname: pairname.replace(\"/image_2/\", \"/\")\n        self.load_flow = _read_kitti_flow\n\n    def _build_cache(self):\n        trainseqs = [\"training/image_2/%06d\" % (i) for i in range(200)]\n        subtrainseqs = trainseqs[:-10]\n        subvalseqs = trainseqs[-10:]\n        testseqs = [\"testing/image_2/%06d\" % (i) for i in range(200)]\n        assert (\n            len(trainseqs) == 200\n            and len(subtrainseqs) == 190\n            and len(subvalseqs) == 10\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow._build_cache","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow._build_cache#L604-L621","kind":"function","name":"_build_cache","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":604,"end_line":621,"context_start_line":584,"context_end_line":641,"code":"    def _prepare_data(self):\n        self.name = \"Kitti15\"\n        self._set_root()\n        assert self.split in [\"train\", \"subtrain\", \"subval\", \"test\"]\n        self.pairname_to_img1name = lambda pairname: osp.join(\n            self.root, pairname + \"_10.png\"\n        )\n        self.pairname_to_img2name = lambda pairname: osp.join(\n            self.root, pairname + \"_11.png\"\n        )\n        self.pairname_to_flowname = (\n            None\n            if self.split == \"test\"\n            else lambda pairname: osp.join(\n                self.root, pairname.replace(\"/image_2/\", \"/flow_occ/\") + \"_10.png\"\n            )\n        )\n        self.pairname_to_str = lambda pairname: pairname.replace(\"/image_2/\", \"/\")\n        self.load_flow = _read_kitti_flow\n\n    def _build_cache(self):\n        trainseqs = [\"training/image_2/%06d\" % (i) for i in range(200)]\n        subtrainseqs = trainseqs[:-10]\n        subvalseqs = trainseqs[-10:]\n        testseqs = [\"testing/image_2/%06d\" % (i) for i in range(200)]\n        assert (\n            len(trainseqs) == 200\n            and len(subtrainseqs) == 190\n            and len(subvalseqs) == 10\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(outdir, \"flow\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowKitti(outfile, prediction)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_flow_results.zip\" flow'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_flow_results.zip\")\n\n\nimport cv2\n\n\ndef _read_numpy_flow(filename):","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.submission_save_pairname","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.submission_save_pairname#L623-L628","kind":"function","name":"submission_save_pairname","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":623,"end_line":628,"context_start_line":603,"context_end_line":648,"code":"\n    def _build_cache(self):\n        trainseqs = [\"training/image_2/%06d\" % (i) for i in range(200)]\n        subtrainseqs = trainseqs[:-10]\n        subvalseqs = trainseqs[-10:]\n        testseqs = [\"testing/image_2/%06d\" % (i) for i in range(200)]\n        assert (\n            len(trainseqs) == 200\n            and len(subtrainseqs) == 190\n            and len(subvalseqs) == 10\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(outdir, \"flow\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowKitti(outfile, prediction)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_flow_results.zip\" flow'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_flow_results.zip\")\n\n\nimport cv2\n\n\ndef _read_numpy_flow(filename):\n    return np.load(filename)\n\n\ndef _read_pfm_flow(filename):\n    f, _ = _read_pfm(filename)\n    assert np.all(f[:, :, 2] == 0.0)\n    return np.ascontiguousarray(f[:, :, :2])","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.stereoflow.datasets_flow.finalize_submission","uri":"program://Human3R/function/src.croco.stereoflow.datasets_flow.finalize_submission#L630-L635","kind":"function","name":"finalize_submission","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":630,"end_line":635,"context_start_line":610,"context_end_line":655,"code":"            len(trainseqs) == 200\n            and len(subtrainseqs) == 190\n            and len(subvalseqs) == 10\n            and len(testseqs) == 200\n        ), \"incorrect parsing of pairs in Kitti15\"\n        tosave = {\n            \"train\": trainseqs,\n            \"subtrain\": subtrainseqs,\n            \"subval\": subvalseqs,\n            \"test\": testseqs,\n        }\n        return tosave\n\n    def submission_save_pairname(self, pairname, prediction, outdir, time):\n        assert prediction.ndim == 3\n        assert prediction.shape[2] == 2\n        outfile = os.path.join(outdir, \"flow\", pairname.split(\"/\")[-1] + \"_10.png\")\n        os.makedirs(os.path.dirname(outfile), exist_ok=True)\n        writeFlowKitti(outfile, prediction)\n\n    def finalize_submission(self, outdir):\n        assert self.split == \"test\"\n        cmd = f'cd {outdir}/; zip -r \"kitti15_flow_results.zip\" flow'\n        print(cmd)\n        os.system(cmd)\n        print(f\"Done. Submission file at {outdir}/kitti15_flow_results.zip\")\n\n\nimport cv2\n\n\ndef _read_numpy_flow(filename):\n    return np.load(filename)\n\n\ndef _read_pfm_flow(filename):\n    f, _ = _read_pfm(filename)\n    assert np.all(f[:, :, 2] == 0.0)\n    return np.ascontiguousarray(f[:, :, :2])\n\n\nTAG_FLOAT = 202021.25  # tag to check the sanity of the file\nTAG_STRING = \"PIEH\"  # string containing the tag\nMIN_WIDTH = 1\nMAX_WIDTH = 99999\nMIN_HEIGHT = 1","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset","uri":"program://Human3R/module/src.croco.datasets.pairs_dataset#L1-L162","kind":"module","name":"src.croco.datasets.pairs_dataset","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":1,"end_line":162,"context_start_line":1,"context_end_line":162,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nfrom torch.utils.data import Dataset\nfrom PIL import Image\n\nfrom datasets.transforms import get_pair_transforms\n\n\ndef load_image(impath):\n    return Image.open(impath)\n\n\ndef load_pairs_from_cache_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:\n        lines = fid.read().strip().splitlines()\n    pairs = [\n        (os.path.join(root, l.split()[0]), os.path.join(root, l.split()[1]))\n        for l in lines\n    ]\n    return pairs\n\n\ndef load_pairs_from_list_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:\n        lines = fid.read().strip().splitlines()\n    pairs = [\n        (os.path.join(root, l + \"_1.jpg\"), os.path.join(root, l + \"_2.jpg\"))\n        for l in lines\n        if not l.startswith(\"#\")\n    ]\n    return pairs\n\n\ndef write_cache_file(fname, pairs, root=\"\"):\n    if len(root) > 0:\n        if not root.endswith(\"/\"):\n            root += \"/\"\n        assert os.path.isdir(root)\n    s = \"\"\n    for im1, im2 in pairs:\n        if len(root) > 0:\n            assert im1.startswith(root), im1\n            assert im2.startswith(root), im2\n        s += \"{:s} {:s}\\n\".format(im1[len(root) :], im2[len(root) :])\n    with open(fname, \"w\") as fid:\n        fid.write(s[:-1])\n\n\ndef parse_and_cache_all_pairs(dname, data_dir=\"./data/\"):\n    if dname == \"habitat_release\":\n        dirname = os.path.join(data_dir, \"habitat_release\")\n        assert os.path.isdir(dirname), (\n            \"cannot find folder for habitat_release pairs: \" + dirname\n        )\n        cache_file = os.path.join(dirname, \"pairs.txt\")\n        assert not os.path.isfile(cache_file), (\n            \"cache file already exists: \" + cache_file\n        )\n\n        print(\"Parsing pairs for dataset: \" + dname)\n        pairs = []\n        for root, dirs, files in os.walk(dirname):\n            if \"val\" in root:\n                continue\n            dirs.sort()\n            pairs += [\n                (\n                    os.path.join(root, f),\n                    os.path.join(root, f[: -len(\"_1.jpeg\")] + \"_2.jpeg\"),\n                )\n                for f in sorted(files)\n                if f.endswith(\"_1.jpeg\")\n            ]\n        print(\"Found {:,} pairs\".format(len(pairs)))\n        print(\"Writing cache to: \" + cache_file)\n        write_cache_file(cache_file, pairs, root=dirname)\n\n    else:\n        raise NotImplementedError(\"Unknown dataset: \" + dname)\n\n\ndef dnames_to_image_pairs(dnames, data_dir=\"./data/\"):\n    \"\"\"\n    dnames: list of datasets with image pairs, separated by +\n    \"\"\"\n    all_pairs = []\n    for dname in dnames.split(\"+\"):\n        if dname == \"habitat_release\":\n            dirname = os.path.join(data_dir, \"habitat_release\")\n            assert os.path.isdir(dirname), (\n                \"cannot find folder for habitat_release pairs: \" + dirname\n            )\n            cache_file = os.path.join(dirname, \"pairs.txt\")\n            assert os.path.isfile(cache_file), (\n                \"cannot find cache file for habitat_release pairs, please first create the cache file, see instructions. \"\n                + cache_file\n            )\n            pairs = load_pairs_from_cache_file(cache_file, root=dirname)\n        elif dname in [\"ARKitScenes\", \"MegaDepth\", \"3DStreetView\", \"IndoorVL\"]:\n            dirname = os.path.join(data_dir, dname + \"_crops\")\n            assert os.path.isdir(\n                dirname\n            ), \"cannot find folder for {:s} pairs: {:s}\".format(dname, dirname)\n            list_file = os.path.join(dirname, \"listing.txt\")\n            assert os.path.isfile(\n                list_file\n            ), \"cannot find list file for {:s} pairs, see instructions. {:s}\".format(\n                dname, list_file\n            )\n            pairs = load_pairs_from_list_file(list_file, root=dirname)\n        print(\"  {:s}: {:,} pairs\".format(dname, len(pairs)))\n        all_pairs += pairs\n    if \"+\" in dnames:\n        print(\" Total: {:,} pairs\".format(len(all_pairs)))\n    return all_pairs\n\n\nclass PairsDataset(Dataset):\n\n    def __init__(\n        self, dnames, trfs=\"\", totensor=True, normalize=True, data_dir=\"./data/\"\n    ):\n        super().__init__()\n        self.image_pairs = dnames_to_image_pairs(dnames, data_dir=data_dir)\n        self.transforms = get_pair_transforms(\n            transform_str=trfs, totensor=totensor, normalize=normalize\n        )\n\n    def __len__(self):\n        return len(self.image_pairs)\n\n    def __getitem__(self, index):\n        im1path, im2path = self.image_pairs[index]\n        im1 = load_image(im1path)\n        im2 = load_image(im2path)\n        if self.transforms is not None:\n            im1, im2 = self.transforms(im1, im2)\n        return im1, im2\n\n\nif __name__ == \"__main__\":\n    import argparse\n\n    parser = argparse.ArgumentParser(\n        prog=\"Computing and caching list of pairs for a given dataset\"\n    )\n    parser.add_argument(\n        \"--data_dir\", default=\"./data/\", type=str, help=\"path where data are stored\"\n    )\n    parser.add_argument(\n        \"--dataset\", default=\"habitat_release\", type=str, help=\"name of the dataset\"\n    )\n    args = parser.parse_args()\n    parse_and_cache_all_pairs(dname=args.dataset, data_dir=args.data_dir)","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.load_image","uri":"program://Human3R/function/src.croco.datasets.pairs_dataset.load_image#L11-L12","kind":"function","name":"load_image","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":11,"end_line":12,"context_start_line":1,"context_end_line":32,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nfrom torch.utils.data import Dataset\nfrom PIL import Image\n\nfrom datasets.transforms import get_pair_transforms\n\n\ndef load_image(impath):\n    return Image.open(impath)\n\n\ndef load_pairs_from_cache_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:\n        lines = fid.read().strip().splitlines()\n    pairs = [\n        (os.path.join(root, l.split()[0]), os.path.join(root, l.split()[1]))\n        for l in lines\n    ]\n    return pairs\n\n\ndef load_pairs_from_list_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.load_pairs_from_cache_file","uri":"program://Human3R/function/src.croco.datasets.pairs_dataset.load_pairs_from_cache_file#L15-L25","kind":"function","name":"load_pairs_from_cache_file","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":15,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nfrom torch.utils.data import Dataset\nfrom PIL import Image\n\nfrom datasets.transforms import get_pair_transforms\n\n\ndef load_image(impath):\n    return Image.open(impath)\n\n\ndef load_pairs_from_cache_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:\n        lines = fid.read().strip().splitlines()\n    pairs = [\n        (os.path.join(root, l.split()[0]), os.path.join(root, l.split()[1]))\n        for l in lines\n    ]\n    return pairs\n\n\ndef load_pairs_from_list_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:\n        lines = fid.read().strip().splitlines()\n    pairs = [\n        (os.path.join(root, l + \"_1.jpg\"), os.path.join(root, l + \"_2.jpg\"))\n        for l in lines\n        if not l.startswith(\"#\")\n    ]\n    return pairs\n\n\ndef write_cache_file(fname, pairs, root=\"\"):\n    if len(root) > 0:\n        if not root.endswith(\"/\"):\n            root += \"/\"","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.load_pairs_from_list_file","uri":"program://Human3R/function/src.croco.datasets.pairs_dataset.load_pairs_from_list_file#L28-L39","kind":"function","name":"load_pairs_from_list_file","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":28,"end_line":39,"context_start_line":8,"context_end_line":59,"code":"from datasets.transforms import get_pair_transforms\n\n\ndef load_image(impath):\n    return Image.open(impath)\n\n\ndef load_pairs_from_cache_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:\n        lines = fid.read().strip().splitlines()\n    pairs = [\n        (os.path.join(root, l.split()[0]), os.path.join(root, l.split()[1]))\n        for l in lines\n    ]\n    return pairs\n\n\ndef load_pairs_from_list_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:\n        lines = fid.read().strip().splitlines()\n    pairs = [\n        (os.path.join(root, l + \"_1.jpg\"), os.path.join(root, l + \"_2.jpg\"))\n        for l in lines\n        if not l.startswith(\"#\")\n    ]\n    return pairs\n\n\ndef write_cache_file(fname, pairs, root=\"\"):\n    if len(root) > 0:\n        if not root.endswith(\"/\"):\n            root += \"/\"\n        assert os.path.isdir(root)\n    s = \"\"\n    for im1, im2 in pairs:\n        if len(root) > 0:\n            assert im1.startswith(root), im1\n            assert im2.startswith(root), im2\n        s += \"{:s} {:s}\\n\".format(im1[len(root) :], im2[len(root) :])\n    with open(fname, \"w\") as fid:\n        fid.write(s[:-1])\n\n\ndef parse_and_cache_all_pairs(dname, data_dir=\"./data/\"):\n    if dname == \"habitat_release\":\n        dirname = os.path.join(data_dir, \"habitat_release\")","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.write_cache_file","uri":"program://Human3R/function/src.croco.datasets.pairs_dataset.write_cache_file#L42-L54","kind":"function","name":"write_cache_file","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":42,"end_line":54,"context_start_line":22,"context_end_line":74,"code":"        (os.path.join(root, l.split()[0]), os.path.join(root, l.split()[1]))\n        for l in lines\n    ]\n    return pairs\n\n\ndef load_pairs_from_list_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:\n        lines = fid.read().strip().splitlines()\n    pairs = [\n        (os.path.join(root, l + \"_1.jpg\"), os.path.join(root, l + \"_2.jpg\"))\n        for l in lines\n        if not l.startswith(\"#\")\n    ]\n    return pairs\n\n\ndef write_cache_file(fname, pairs, root=\"\"):\n    if len(root) > 0:\n        if not root.endswith(\"/\"):\n            root += \"/\"\n        assert os.path.isdir(root)\n    s = \"\"\n    for im1, im2 in pairs:\n        if len(root) > 0:\n            assert im1.startswith(root), im1\n            assert im2.startswith(root), im2\n        s += \"{:s} {:s}\\n\".format(im1[len(root) :], im2[len(root) :])\n    with open(fname, \"w\") as fid:\n        fid.write(s[:-1])\n\n\ndef parse_and_cache_all_pairs(dname, data_dir=\"./data/\"):\n    if dname == \"habitat_release\":\n        dirname = os.path.join(data_dir, \"habitat_release\")\n        assert os.path.isdir(dirname), (\n            \"cannot find folder for habitat_release pairs: \" + dirname\n        )\n        cache_file = os.path.join(dirname, \"pairs.txt\")\n        assert not os.path.isfile(cache_file), (\n            \"cache file already exists: \" + cache_file\n        )\n\n        print(\"Parsing pairs for dataset: \" + dname)\n        pairs = []\n        for root, dirs, files in os.walk(dirname):\n            if \"val\" in root:\n                continue\n            dirs.sort()\n            pairs += [","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.parse_and_cache_all_pairs","uri":"program://Human3R/function/src.croco.datasets.pairs_dataset.parse_and_cache_all_pairs#L57-L87","kind":"function","name":"parse_and_cache_all_pairs","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":57,"end_line":87,"context_start_line":37,"context_end_line":107,"code":"        if not l.startswith(\"#\")\n    ]\n    return pairs\n\n\ndef write_cache_file(fname, pairs, root=\"\"):\n    if len(root) > 0:\n        if not root.endswith(\"/\"):\n            root += \"/\"\n        assert os.path.isdir(root)\n    s = \"\"\n    for im1, im2 in pairs:\n        if len(root) > 0:\n            assert im1.startswith(root), im1\n            assert im2.startswith(root), im2\n        s += \"{:s} {:s}\\n\".format(im1[len(root) :], im2[len(root) :])\n    with open(fname, \"w\") as fid:\n        fid.write(s[:-1])\n\n\ndef parse_and_cache_all_pairs(dname, data_dir=\"./data/\"):\n    if dname == \"habitat_release\":\n        dirname = os.path.join(data_dir, \"habitat_release\")\n        assert os.path.isdir(dirname), (\n            \"cannot find folder for habitat_release pairs: \" + dirname\n        )\n        cache_file = os.path.join(dirname, \"pairs.txt\")\n        assert not os.path.isfile(cache_file), (\n            \"cache file already exists: \" + cache_file\n        )\n\n        print(\"Parsing pairs for dataset: \" + dname)\n        pairs = []\n        for root, dirs, files in os.walk(dirname):\n            if \"val\" in root:\n                continue\n            dirs.sort()\n            pairs += [\n                (\n                    os.path.join(root, f),\n                    os.path.join(root, f[: -len(\"_1.jpeg\")] + \"_2.jpeg\"),\n                )\n                for f in sorted(files)\n                if f.endswith(\"_1.jpeg\")\n            ]\n        print(\"Found {:,} pairs\".format(len(pairs)))\n        print(\"Writing cache to: \" + cache_file)\n        write_cache_file(cache_file, pairs, root=dirname)\n\n    else:\n        raise NotImplementedError(\"Unknown dataset: \" + dname)\n\n\ndef dnames_to_image_pairs(dnames, data_dir=\"./data/\"):\n    \"\"\"\n    dnames: list of datasets with image pairs, separated by +\n    \"\"\"\n    all_pairs = []\n    for dname in dnames.split(\"+\"):\n        if dname == \"habitat_release\":\n            dirname = os.path.join(data_dir, \"habitat_release\")\n            assert os.path.isdir(dirname), (\n                \"cannot find folder for habitat_release pairs: \" + dirname\n            )\n            cache_file = os.path.join(dirname, \"pairs.txt\")\n            assert os.path.isfile(cache_file), (\n                \"cannot find cache file for habitat_release pairs, please first create the cache file, see instructions. \"\n                + cache_file\n            )\n            pairs = load_pairs_from_cache_file(cache_file, root=dirname)\n        elif dname in [\"ARKitScenes\", \"MegaDepth\", \"3DStreetView\", \"IndoorVL\"]:","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.dnames_to_image_pairs","uri":"program://Human3R/function/src.croco.datasets.pairs_dataset.dnames_to_image_pairs#L90-L123","kind":"function","name":"dnames_to_image_pairs","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":90,"end_line":123,"context_start_line":70,"context_end_line":143,"code":"        for root, dirs, files in os.walk(dirname):\n            if \"val\" in root:\n                continue\n            dirs.sort()\n            pairs += [\n                (\n                    os.path.join(root, f),\n                    os.path.join(root, f[: -len(\"_1.jpeg\")] + \"_2.jpeg\"),\n                )\n                for f in sorted(files)\n                if f.endswith(\"_1.jpeg\")\n            ]\n        print(\"Found {:,} pairs\".format(len(pairs)))\n        print(\"Writing cache to: \" + cache_file)\n        write_cache_file(cache_file, pairs, root=dirname)\n\n    else:\n        raise NotImplementedError(\"Unknown dataset: \" + dname)\n\n\ndef dnames_to_image_pairs(dnames, data_dir=\"./data/\"):\n    \"\"\"\n    dnames: list of datasets with image pairs, separated by +\n    \"\"\"\n    all_pairs = []\n    for dname in dnames.split(\"+\"):\n        if dname == \"habitat_release\":\n            dirname = os.path.join(data_dir, \"habitat_release\")\n            assert os.path.isdir(dirname), (\n                \"cannot find folder for habitat_release pairs: \" + dirname\n            )\n            cache_file = os.path.join(dirname, \"pairs.txt\")\n            assert os.path.isfile(cache_file), (\n                \"cannot find cache file for habitat_release pairs, please first create the cache file, see instructions. \"\n                + cache_file\n            )\n            pairs = load_pairs_from_cache_file(cache_file, root=dirname)\n        elif dname in [\"ARKitScenes\", \"MegaDepth\", \"3DStreetView\", \"IndoorVL\"]:\n            dirname = os.path.join(data_dir, dname + \"_crops\")\n            assert os.path.isdir(\n                dirname\n            ), \"cannot find folder for {:s} pairs: {:s}\".format(dname, dirname)\n            list_file = os.path.join(dirname, \"listing.txt\")\n            assert os.path.isfile(\n                list_file\n            ), \"cannot find list file for {:s} pairs, see instructions. {:s}\".format(\n                dname, list_file\n            )\n            pairs = load_pairs_from_list_file(list_file, root=dirname)\n        print(\"  {:s}: {:,} pairs\".format(dname, len(pairs)))\n        all_pairs += pairs\n    if \"+\" in dnames:\n        print(\" Total: {:,} pairs\".format(len(all_pairs)))\n    return all_pairs\n\n\nclass PairsDataset(Dataset):\n\n    def __init__(\n        self, dnames, trfs=\"\", totensor=True, normalize=True, data_dir=\"./data/\"\n    ):\n        super().__init__()\n        self.image_pairs = dnames_to_image_pairs(dnames, data_dir=data_dir)\n        self.transforms = get_pair_transforms(\n            transform_str=trfs, totensor=totensor, normalize=normalize\n        )\n\n    def __len__(self):\n        return len(self.image_pairs)\n\n    def __getitem__(self, index):\n        im1path, im2path = self.image_pairs[index]\n        im1 = load_image(im1path)\n        im2 = load_image(im2path)","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.PairsDataset","uri":"program://Human3R/class/src.croco.datasets.pairs_dataset.PairsDataset#L126-L146","kind":"class","name":"PairsDataset","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":126,"end_line":146,"context_start_line":106,"context_end_line":162,"code":"            pairs = load_pairs_from_cache_file(cache_file, root=dirname)\n        elif dname in [\"ARKitScenes\", \"MegaDepth\", \"3DStreetView\", \"IndoorVL\"]:\n            dirname = os.path.join(data_dir, dname + \"_crops\")\n            assert os.path.isdir(\n                dirname\n            ), \"cannot find folder for {:s} pairs: {:s}\".format(dname, dirname)\n            list_file = os.path.join(dirname, \"listing.txt\")\n            assert os.path.isfile(\n                list_file\n            ), \"cannot find list file for {:s} pairs, see instructions. {:s}\".format(\n                dname, list_file\n            )\n            pairs = load_pairs_from_list_file(list_file, root=dirname)\n        print(\"  {:s}: {:,} pairs\".format(dname, len(pairs)))\n        all_pairs += pairs\n    if \"+\" in dnames:\n        print(\" Total: {:,} pairs\".format(len(all_pairs)))\n    return all_pairs\n\n\nclass PairsDataset(Dataset):\n\n    def __init__(\n        self, dnames, trfs=\"\", totensor=True, normalize=True, data_dir=\"./data/\"\n    ):\n        super().__init__()\n        self.image_pairs = dnames_to_image_pairs(dnames, data_dir=data_dir)\n        self.transforms = get_pair_transforms(\n            transform_str=trfs, totensor=totensor, normalize=normalize\n        )\n\n    def __len__(self):\n        return len(self.image_pairs)\n\n    def __getitem__(self, index):\n        im1path, im2path = self.image_pairs[index]\n        im1 = load_image(im1path)\n        im2 = load_image(im2path)\n        if self.transforms is not None:\n            im1, im2 = self.transforms(im1, im2)\n        return im1, im2\n\n\nif __name__ == \"__main__\":\n    import argparse\n\n    parser = argparse.ArgumentParser(\n        prog=\"Computing and caching list of pairs for a given dataset\"\n    )\n    parser.add_argument(\n        \"--data_dir\", default=\"./data/\", type=str, help=\"path where data are stored\"\n    )\n    parser.add_argument(\n        \"--dataset\", default=\"habitat_release\", type=str, help=\"name of the dataset\"\n    )\n    args = parser.parse_args()\n    parse_and_cache_all_pairs(dname=args.dataset, data_dir=args.data_dir)","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.__init__","uri":"program://Human3R/function/src.croco.datasets.pairs_dataset.__init__#L128-L135","kind":"function","name":"__init__","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":128,"end_line":135,"context_start_line":108,"context_end_line":155,"code":"            dirname = os.path.join(data_dir, dname + \"_crops\")\n            assert os.path.isdir(\n                dirname\n            ), \"cannot find folder for {:s} pairs: {:s}\".format(dname, dirname)\n            list_file = os.path.join(dirname, \"listing.txt\")\n            assert os.path.isfile(\n                list_file\n            ), \"cannot find list file for {:s} pairs, see instructions. {:s}\".format(\n                dname, list_file\n            )\n            pairs = load_pairs_from_list_file(list_file, root=dirname)\n        print(\"  {:s}: {:,} pairs\".format(dname, len(pairs)))\n        all_pairs += pairs\n    if \"+\" in dnames:\n        print(\" Total: {:,} pairs\".format(len(all_pairs)))\n    return all_pairs\n\n\nclass PairsDataset(Dataset):\n\n    def __init__(\n        self, dnames, trfs=\"\", totensor=True, normalize=True, data_dir=\"./data/\"\n    ):\n        super().__init__()\n        self.image_pairs = dnames_to_image_pairs(dnames, data_dir=data_dir)\n        self.transforms = get_pair_transforms(\n            transform_str=trfs, totensor=totensor, normalize=normalize\n        )\n\n    def __len__(self):\n        return len(self.image_pairs)\n\n    def __getitem__(self, index):\n        im1path, im2path = self.image_pairs[index]\n        im1 = load_image(im1path)\n        im2 = load_image(im2path)\n        if self.transforms is not None:\n            im1, im2 = self.transforms(im1, im2)\n        return im1, im2\n\n\nif __name__ == \"__main__\":\n    import argparse\n\n    parser = argparse.ArgumentParser(\n        prog=\"Computing and caching list of pairs for a given dataset\"\n    )\n    parser.add_argument(","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.__len__","uri":"program://Human3R/function/src.croco.datasets.pairs_dataset.__len__#L137-L138","kind":"function","name":"__len__","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":137,"end_line":138,"context_start_line":117,"context_end_line":158,"code":"            )\n            pairs = load_pairs_from_list_file(list_file, root=dirname)\n        print(\"  {:s}: {:,} pairs\".format(dname, len(pairs)))\n        all_pairs += pairs\n    if \"+\" in dnames:\n        print(\" Total: {:,} pairs\".format(len(all_pairs)))\n    return all_pairs\n\n\nclass PairsDataset(Dataset):\n\n    def __init__(\n        self, dnames, trfs=\"\", totensor=True, normalize=True, data_dir=\"./data/\"\n    ):\n        super().__init__()\n        self.image_pairs = dnames_to_image_pairs(dnames, data_dir=data_dir)\n        self.transforms = get_pair_transforms(\n            transform_str=trfs, totensor=totensor, normalize=normalize\n        )\n\n    def __len__(self):\n        return len(self.image_pairs)\n\n    def __getitem__(self, index):\n        im1path, im2path = self.image_pairs[index]\n        im1 = load_image(im1path)\n        im2 = load_image(im2path)\n        if self.transforms is not None:\n            im1, im2 = self.transforms(im1, im2)\n        return im1, im2\n\n\nif __name__ == \"__main__\":\n    import argparse\n\n    parser = argparse.ArgumentParser(\n        prog=\"Computing and caching list of pairs for a given dataset\"\n    )\n    parser.add_argument(\n        \"--data_dir\", default=\"./data/\", type=str, help=\"path where data are stored\"\n    )\n    parser.add_argument(","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.pairs_dataset.__getitem__","uri":"program://Human3R/function/src.croco.datasets.pairs_dataset.__getitem__#L140-L146","kind":"function","name":"__getitem__","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":140,"end_line":146,"context_start_line":120,"context_end_line":162,"code":"        all_pairs += pairs\n    if \"+\" in dnames:\n        print(\" Total: {:,} pairs\".format(len(all_pairs)))\n    return all_pairs\n\n\nclass PairsDataset(Dataset):\n\n    def __init__(\n        self, dnames, trfs=\"\", totensor=True, normalize=True, data_dir=\"./data/\"\n    ):\n        super().__init__()\n        self.image_pairs = dnames_to_image_pairs(dnames, data_dir=data_dir)\n        self.transforms = get_pair_transforms(\n            transform_str=trfs, totensor=totensor, normalize=normalize\n        )\n\n    def __len__(self):\n        return len(self.image_pairs)\n\n    def __getitem__(self, index):\n        im1path, im2path = self.image_pairs[index]\n        im1 = load_image(im1path)\n        im2 = load_image(im2path)\n        if self.transforms is not None:\n            im1, im2 = self.transforms(im1, im2)\n        return im1, im2\n\n\nif __name__ == \"__main__\":\n    import argparse\n\n    parser = argparse.ArgumentParser(\n        prog=\"Computing and caching list of pairs for a given dataset\"\n    )\n    parser.add_argument(\n        \"--data_dir\", default=\"./data/\", type=str, help=\"path where data are stored\"\n    )\n    parser.add_argument(\n        \"--dataset\", default=\"habitat_release\", type=str, help=\"name of the dataset\"\n    )\n    args = parser.parse_args()\n    parse_and_cache_all_pairs(dname=args.dataset, data_dir=args.data_dir)","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms","uri":"program://Human3R/module/src.croco.datasets.transforms#L1-L135","kind":"module","name":"src.croco.datasets.transforms","path":"src/croco/datasets/transforms.py","language":"python","start_line":1,"end_line":135,"context_start_line":1,"context_end_line":135,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport torch\nimport torchvision.transforms\nimport torchvision.transforms.functional as F\n\n# \"Pair\": apply a transform on a pair\n# \"Both\": apply the exact same transform to both images\n\n\nclass ComposePair(torchvision.transforms.Compose):\n    def __call__(self, img1, img2):\n        for t in self.transforms:\n            img1, img2 = t(img1, img2)\n        return img1, img2\n\n\nclass NormalizeBoth(torchvision.transforms.Normalize):\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ToTensorBoth(torchvision.transforms.ToTensor):\n    def __call__(self, img1, img2):\n        img1 = super().__call__(img1)\n        img2 = super().__call__(img2)\n        return img1, img2\n\n\nclass RandomCropPair(torchvision.transforms.RandomCrop):\n    # the crop will be intentionally different for the two images with this class\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ColorJitterPair(torchvision.transforms.ColorJitter):\n    # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob\n    def __init__(self, assymetric_prob, **kwargs):\n        super().__init__(**kwargs)\n        self.assymetric_prob = assymetric_prob\n\n    def jitter_one(\n        self,\n        img,\n        fn_idx,\n        brightness_factor,\n        contrast_factor,\n        saturation_factor,\n        hue_factor,\n    ):\n        for fn_id in fn_idx:\n            if fn_id == 0 and brightness_factor is not None:\n                img = F.adjust_brightness(img, brightness_factor)\n            elif fn_id == 1 and contrast_factor is not None:\n                img = F.adjust_contrast(img, contrast_factor)\n            elif fn_id == 2 and saturation_factor is not None:\n                img = F.adjust_saturation(img, saturation_factor)\n            elif fn_id == 3 and hue_factor is not None:\n                img = F.adjust_hue(img, hue_factor)\n        return img\n\n    def forward(self, img1, img2):\n\n        fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = (\n            self.get_params(self.brightness, self.contrast, self.saturation, self.hue)\n        )\n        img1 = self.jitter_one(\n            img1,\n            fn_idx,\n            brightness_factor,\n            contrast_factor,\n            saturation_factor,\n            hue_factor,\n        )\n        if torch.rand(1) < self.assymetric_prob:  # assymetric:\n            (\n                fn_idx,\n                brightness_factor,\n                contrast_factor,\n                saturation_factor,\n                hue_factor,\n            ) = self.get_params(\n                self.brightness, self.contrast, self.saturation, self.hue\n            )\n        img2 = self.jitter_one(\n            img2,\n            fn_idx,\n            brightness_factor,\n            contrast_factor,\n            saturation_factor,\n            hue_factor,\n        )\n        return img1, img2\n\n\ndef get_pair_transforms(transform_str, totensor=True, normalize=True):\n    # transform_str is eg    crop224+color\n    trfs = []\n    for s in transform_str.split(\"+\"):\n        if s.startswith(\"crop\"):\n            size = int(s[len(\"crop\") :])\n            trfs.append(RandomCropPair(size))\n        elif s == \"acolor\":\n            trfs.append(\n                ColorJitterPair(\n                    assymetric_prob=1.0,\n                    brightness=(0.6, 1.4),\n                    contrast=(0.6, 1.4),\n                    saturation=(0.6, 1.4),\n                    hue=0.0,\n                )\n            )\n        elif s == \"\":  # if transform_str was \"\"\n            pass\n        else:\n            raise NotImplementedError(\"Unknown augmentation: \" + s)\n\n    if totensor:\n        trfs.append(ToTensorBoth())\n    if normalize:\n        trfs.append(\n            NormalizeBoth(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])\n        )\n\n    if len(trfs) == 0:\n        return None\n    elif len(trfs) == 1:\n        return trfs\n    else:\n        return ComposePair(trfs)","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.ComposePair","uri":"program://Human3R/class/src.croco.datasets.transforms.ComposePair#L12-L16","kind":"class","name":"ComposePair","path":"src/croco/datasets/transforms.py","language":"python","start_line":12,"end_line":16,"context_start_line":1,"context_end_line":36,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport torch\nimport torchvision.transforms\nimport torchvision.transforms.functional as F\n\n# \"Pair\": apply a transform on a pair\n# \"Both\": apply the exact same transform to both images\n\n\nclass ComposePair(torchvision.transforms.Compose):\n    def __call__(self, img1, img2):\n        for t in self.transforms:\n            img1, img2 = t(img1, img2)\n        return img1, img2\n\n\nclass NormalizeBoth(torchvision.transforms.Normalize):\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ToTensorBoth(torchvision.transforms.ToTensor):\n    def __call__(self, img1, img2):\n        img1 = super().__call__(img1)\n        img2 = super().__call__(img2)\n        return img1, img2\n\n\nclass RandomCropPair(torchvision.transforms.RandomCrop):\n    # the crop will be intentionally different for the two images with this class\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.NormalizeBoth","uri":"program://Human3R/class/src.croco.datasets.transforms.NormalizeBoth#L19-L23","kind":"class","name":"NormalizeBoth","path":"src/croco/datasets/transforms.py","language":"python","start_line":19,"end_line":23,"context_start_line":1,"context_end_line":43,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport torch\nimport torchvision.transforms\nimport torchvision.transforms.functional as F\n\n# \"Pair\": apply a transform on a pair\n# \"Both\": apply the exact same transform to both images\n\n\nclass ComposePair(torchvision.transforms.Compose):\n    def __call__(self, img1, img2):\n        for t in self.transforms:\n            img1, img2 = t(img1, img2)\n        return img1, img2\n\n\nclass NormalizeBoth(torchvision.transforms.Normalize):\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ToTensorBoth(torchvision.transforms.ToTensor):\n    def __call__(self, img1, img2):\n        img1 = super().__call__(img1)\n        img2 = super().__call__(img2)\n        return img1, img2\n\n\nclass RandomCropPair(torchvision.transforms.RandomCrop):\n    # the crop will be intentionally different for the two images with this class\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ColorJitterPair(torchvision.transforms.ColorJitter):\n    # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob\n    def __init__(self, assymetric_prob, **kwargs):","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.ToTensorBoth","uri":"program://Human3R/class/src.croco.datasets.transforms.ToTensorBoth#L26-L30","kind":"class","name":"ToTensorBoth","path":"src/croco/datasets/transforms.py","language":"python","start_line":26,"end_line":30,"context_start_line":6,"context_end_line":50,"code":"import torchvision.transforms.functional as F\n\n# \"Pair\": apply a transform on a pair\n# \"Both\": apply the exact same transform to both images\n\n\nclass ComposePair(torchvision.transforms.Compose):\n    def __call__(self, img1, img2):\n        for t in self.transforms:\n            img1, img2 = t(img1, img2)\n        return img1, img2\n\n\nclass NormalizeBoth(torchvision.transforms.Normalize):\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ToTensorBoth(torchvision.transforms.ToTensor):\n    def __call__(self, img1, img2):\n        img1 = super().__call__(img1)\n        img2 = super().__call__(img2)\n        return img1, img2\n\n\nclass RandomCropPair(torchvision.transforms.RandomCrop):\n    # the crop will be intentionally different for the two images with this class\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ColorJitterPair(torchvision.transforms.ColorJitter):\n    # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob\n    def __init__(self, assymetric_prob, **kwargs):\n        super().__init__(**kwargs)\n        self.assymetric_prob = assymetric_prob\n\n    def jitter_one(\n        self,\n        img,\n        fn_idx,","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.RandomCropPair","uri":"program://Human3R/class/src.croco.datasets.transforms.RandomCropPair#L33-L38","kind":"class","name":"RandomCropPair","path":"src/croco/datasets/transforms.py","language":"python","start_line":33,"end_line":38,"context_start_line":13,"context_end_line":58,"code":"    def __call__(self, img1, img2):\n        for t in self.transforms:\n            img1, img2 = t(img1, img2)\n        return img1, img2\n\n\nclass NormalizeBoth(torchvision.transforms.Normalize):\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ToTensorBoth(torchvision.transforms.ToTensor):\n    def __call__(self, img1, img2):\n        img1 = super().__call__(img1)\n        img2 = super().__call__(img2)\n        return img1, img2\n\n\nclass RandomCropPair(torchvision.transforms.RandomCrop):\n    # the crop will be intentionally different for the two images with this class\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ColorJitterPair(torchvision.transforms.ColorJitter):\n    # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob\n    def __init__(self, assymetric_prob, **kwargs):\n        super().__init__(**kwargs)\n        self.assymetric_prob = assymetric_prob\n\n    def jitter_one(\n        self,\n        img,\n        fn_idx,\n        brightness_factor,\n        contrast_factor,\n        saturation_factor,\n        hue_factor,\n    ):\n        for fn_id in fn_idx:\n            if fn_id == 0 and brightness_factor is not None:\n                img = F.adjust_brightness(img, brightness_factor)","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.ColorJitterPair","uri":"program://Human3R/class/src.croco.datasets.transforms.ColorJitterPair#L41-L98","kind":"class","name":"ColorJitterPair","path":"src/croco/datasets/transforms.py","language":"python","start_line":41,"end_line":98,"context_start_line":21,"context_end_line":118,"code":"        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ToTensorBoth(torchvision.transforms.ToTensor):\n    def __call__(self, img1, img2):\n        img1 = super().__call__(img1)\n        img2 = super().__call__(img2)\n        return img1, img2\n\n\nclass RandomCropPair(torchvision.transforms.RandomCrop):\n    # the crop will be intentionally different for the two images with this class\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ColorJitterPair(torchvision.transforms.ColorJitter):\n    # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob\n    def __init__(self, assymetric_prob, **kwargs):\n        super().__init__(**kwargs)\n        self.assymetric_prob = assymetric_prob\n\n    def jitter_one(\n        self,\n        img,\n        fn_idx,\n        brightness_factor,\n        contrast_factor,\n        saturation_factor,\n        hue_factor,\n    ):\n        for fn_id in fn_idx:\n            if fn_id == 0 and brightness_factor is not None:\n                img = F.adjust_brightness(img, brightness_factor)\n            elif fn_id == 1 and contrast_factor is not None:\n                img = F.adjust_contrast(img, contrast_factor)\n            elif fn_id == 2 and saturation_factor is not None:\n                img = F.adjust_saturation(img, saturation_factor)\n            elif fn_id == 3 and hue_factor is not None:\n                img = F.adjust_hue(img, hue_factor)\n        return img\n\n    def forward(self, img1, img2):\n\n        fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = (\n            self.get_params(self.brightness, self.contrast, self.saturation, self.hue)\n        )\n        img1 = self.jitter_one(\n            img1,\n            fn_idx,\n            brightness_factor,\n            contrast_factor,\n            saturation_factor,\n            hue_factor,\n        )\n        if torch.rand(1) < self.assymetric_prob:  # assymetric:\n            (\n                fn_idx,\n                brightness_factor,\n                contrast_factor,\n                saturation_factor,\n                hue_factor,\n            ) = self.get_params(\n                self.brightness, self.contrast, self.saturation, self.hue\n            )\n        img2 = self.jitter_one(\n            img2,\n            fn_idx,\n            brightness_factor,\n            contrast_factor,\n            saturation_factor,\n            hue_factor,\n        )\n        return img1, img2\n\n\ndef get_pair_transforms(transform_str, totensor=True, normalize=True):\n    # transform_str is eg    crop224+color\n    trfs = []\n    for s in transform_str.split(\"+\"):\n        if s.startswith(\"crop\"):\n            size = int(s[len(\"crop\") :])\n            trfs.append(RandomCropPair(size))\n        elif s == \"acolor\":\n            trfs.append(\n                ColorJitterPair(\n                    assymetric_prob=1.0,\n                    brightness=(0.6, 1.4),\n                    contrast=(0.6, 1.4),\n                    saturation=(0.6, 1.4),\n                    hue=0.0,\n                )\n            )\n        elif s == \"\":  # if transform_str was \"\"","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.get_pair_transforms","uri":"program://Human3R/function/src.croco.datasets.transforms.get_pair_transforms#L101-L135","kind":"function","name":"get_pair_transforms","path":"src/croco/datasets/transforms.py","language":"python","start_line":101,"end_line":135,"context_start_line":81,"context_end_line":135,"code":"            (\n                fn_idx,\n                brightness_factor,\n                contrast_factor,\n                saturation_factor,\n                hue_factor,\n            ) = self.get_params(\n                self.brightness, self.contrast, self.saturation, self.hue\n            )\n        img2 = self.jitter_one(\n            img2,\n            fn_idx,\n            brightness_factor,\n            contrast_factor,\n            saturation_factor,\n            hue_factor,\n        )\n        return img1, img2\n\n\ndef get_pair_transforms(transform_str, totensor=True, normalize=True):\n    # transform_str is eg    crop224+color\n    trfs = []\n    for s in transform_str.split(\"+\"):\n        if s.startswith(\"crop\"):\n            size = int(s[len(\"crop\") :])\n            trfs.append(RandomCropPair(size))\n        elif s == \"acolor\":\n            trfs.append(\n                ColorJitterPair(\n                    assymetric_prob=1.0,\n                    brightness=(0.6, 1.4),\n                    contrast=(0.6, 1.4),\n                    saturation=(0.6, 1.4),\n                    hue=0.0,\n                )\n            )\n        elif s == \"\":  # if transform_str was \"\"\n            pass\n        else:\n            raise NotImplementedError(\"Unknown augmentation: \" + s)\n\n    if totensor:\n        trfs.append(ToTensorBoth())\n    if normalize:\n        trfs.append(\n            NormalizeBoth(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])\n        )\n\n    if len(trfs) == 0:\n        return None\n    elif len(trfs) == 1:\n        return trfs\n    else:\n        return ComposePair(trfs)","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.__call__","uri":"program://Human3R/function/src.croco.datasets.transforms.__call__#L27-L30","kind":"function","name":"__call__","path":"src/croco/datasets/transforms.py","language":"python","start_line":27,"end_line":30,"context_start_line":7,"context_end_line":50,"code":"\n# \"Pair\": apply a transform on a pair\n# \"Both\": apply the exact same transform to both images\n\n\nclass ComposePair(torchvision.transforms.Compose):\n    def __call__(self, img1, img2):\n        for t in self.transforms:\n            img1, img2 = t(img1, img2)\n        return img1, img2\n\n\nclass NormalizeBoth(torchvision.transforms.Normalize):\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ToTensorBoth(torchvision.transforms.ToTensor):\n    def __call__(self, img1, img2):\n        img1 = super().__call__(img1)\n        img2 = super().__call__(img2)\n        return img1, img2\n\n\nclass RandomCropPair(torchvision.transforms.RandomCrop):\n    # the crop will be intentionally different for the two images with this class\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ColorJitterPair(torchvision.transforms.ColorJitter):\n    # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob\n    def __init__(self, assymetric_prob, **kwargs):\n        super().__init__(**kwargs)\n        self.assymetric_prob = assymetric_prob\n\n    def jitter_one(\n        self,\n        img,\n        fn_idx,","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.forward","uri":"program://Human3R/function/src.croco.datasets.transforms.forward#L67-L98","kind":"function","name":"forward","path":"src/croco/datasets/transforms.py","language":"python","start_line":67,"end_line":98,"context_start_line":47,"context_end_line":118,"code":"    def jitter_one(\n        self,\n        img,\n        fn_idx,\n        brightness_factor,\n        contrast_factor,\n        saturation_factor,\n        hue_factor,\n    ):\n        for fn_id in fn_idx:\n            if fn_id == 0 and brightness_factor is not None:\n                img = F.adjust_brightness(img, brightness_factor)\n            elif fn_id == 1 and contrast_factor is not None:\n                img = F.adjust_contrast(img, contrast_factor)\n            elif fn_id == 2 and saturation_factor is not None:\n                img = F.adjust_saturation(img, saturation_factor)\n            elif fn_id == 3 and hue_factor is not None:\n                img = F.adjust_hue(img, hue_factor)\n        return img\n\n    def forward(self, img1, img2):\n\n        fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = (\n            self.get_params(self.brightness, self.contrast, self.saturation, self.hue)\n        )\n        img1 = self.jitter_one(\n            img1,\n            fn_idx,\n            brightness_factor,\n            contrast_factor,\n            saturation_factor,\n            hue_factor,\n        )\n        if torch.rand(1) < self.assymetric_prob:  # assymetric:\n            (\n                fn_idx,\n                brightness_factor,\n                contrast_factor,\n                saturation_factor,\n                hue_factor,\n            ) = self.get_params(\n                self.brightness, self.contrast, self.saturation, self.hue\n            )\n        img2 = self.jitter_one(\n            img2,\n            fn_idx,\n            brightness_factor,\n            contrast_factor,\n            saturation_factor,\n            hue_factor,\n        )\n        return img1, img2\n\n\ndef get_pair_transforms(transform_str, totensor=True, normalize=True):\n    # transform_str is eg    crop224+color\n    trfs = []\n    for s in transform_str.split(\"+\"):\n        if s.startswith(\"crop\"):\n            size = int(s[len(\"crop\") :])\n            trfs.append(RandomCropPair(size))\n        elif s == \"acolor\":\n            trfs.append(\n                ColorJitterPair(\n                    assymetric_prob=1.0,\n                    brightness=(0.6, 1.4),\n                    contrast=(0.6, 1.4),\n                    saturation=(0.6, 1.4),\n                    hue=0.0,\n                )\n            )\n        elif s == \"\":  # if transform_str was \"\"","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.__init__","uri":"program://Human3R/function/src.croco.datasets.transforms.__init__#L43-L45","kind":"function","name":"__init__","path":"src/croco/datasets/transforms.py","language":"python","start_line":43,"end_line":45,"context_start_line":23,"context_end_line":65,"code":"        return img1, img2\n\n\nclass ToTensorBoth(torchvision.transforms.ToTensor):\n    def __call__(self, img1, img2):\n        img1 = super().__call__(img1)\n        img2 = super().__call__(img2)\n        return img1, img2\n\n\nclass RandomCropPair(torchvision.transforms.RandomCrop):\n    # the crop will be intentionally different for the two images with this class\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ColorJitterPair(torchvision.transforms.ColorJitter):\n    # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob\n    def __init__(self, assymetric_prob, **kwargs):\n        super().__init__(**kwargs)\n        self.assymetric_prob = assymetric_prob\n\n    def jitter_one(\n        self,\n        img,\n        fn_idx,\n        brightness_factor,\n        contrast_factor,\n        saturation_factor,\n        hue_factor,\n    ):\n        for fn_id in fn_idx:\n            if fn_id == 0 and brightness_factor is not None:\n                img = F.adjust_brightness(img, brightness_factor)\n            elif fn_id == 1 and contrast_factor is not None:\n                img = F.adjust_contrast(img, contrast_factor)\n            elif fn_id == 2 and saturation_factor is not None:\n                img = F.adjust_saturation(img, saturation_factor)\n            elif fn_id == 3 and hue_factor is not None:\n                img = F.adjust_hue(img, hue_factor)\n        return img","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.transforms.jitter_one","uri":"program://Human3R/function/src.croco.datasets.transforms.jitter_one#L47-L65","kind":"function","name":"jitter_one","path":"src/croco/datasets/transforms.py","language":"python","start_line":47,"end_line":65,"context_start_line":27,"context_end_line":85,"code":"    def __call__(self, img1, img2):\n        img1 = super().__call__(img1)\n        img2 = super().__call__(img2)\n        return img1, img2\n\n\nclass RandomCropPair(torchvision.transforms.RandomCrop):\n    # the crop will be intentionally different for the two images with this class\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)\n        img2 = super().forward(img2)\n        return img1, img2\n\n\nclass ColorJitterPair(torchvision.transforms.ColorJitter):\n    # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob\n    def __init__(self, assymetric_prob, **kwargs):\n        super().__init__(**kwargs)\n        self.assymetric_prob = assymetric_prob\n\n    def jitter_one(\n        self,\n        img,\n        fn_idx,\n        brightness_factor,\n        contrast_factor,\n        saturation_factor,\n        hue_factor,\n    ):\n        for fn_id in fn_idx:\n            if fn_id == 0 and brightness_factor is not None:\n                img = F.adjust_brightness(img, brightness_factor)\n            elif fn_id == 1 and contrast_factor is not None:\n                img = F.adjust_contrast(img, contrast_factor)\n            elif fn_id == 2 and saturation_factor is not None:\n                img = F.adjust_saturation(img, saturation_factor)\n            elif fn_id == 3 and hue_factor is not None:\n                img = F.adjust_hue(img, hue_factor)\n        return img\n\n    def forward(self, img1, img2):\n\n        fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = (\n            self.get_params(self.brightness, self.contrast, self.saturation, self.hue)\n        )\n        img1 = self.jitter_one(\n            img1,\n            fn_idx,\n            brightness_factor,\n            contrast_factor,\n            saturation_factor,\n            hue_factor,\n        )\n        if torch.rand(1) < self.assymetric_prob:  # assymetric:\n            (\n                fn_idx,\n                brightness_factor,\n                contrast_factor,\n                saturation_factor,","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images","uri":"program://Human3R/module/src.croco.datasets.crops.extract_crops_from_images#L1-L183","kind":"module","name":"src.croco.datasets.crops.extract_crops_from_images","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":1,"end_line":183,"context_start_line":1,"context_end_line":183,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# Extracting crops for pre-training\n# --------------------------------------------------------\n\nimport os\nimport argparse\nfrom tqdm import tqdm\nfrom PIL import Image\nimport functools\nfrom multiprocessing import Pool\nimport math\n\n\ndef arg_parser():\n    parser = argparse.ArgumentParser(\n        \"Generate cropped image pairs from image crop list\"\n    )\n\n    parser.add_argument(\"--crops\", type=str, required=True, help=\"crop file\")\n    parser.add_argument(\"--root-dir\", type=str, required=True, help=\"root directory\")\n    parser.add_argument(\n        \"--output-dir\", type=str, required=True, help=\"output directory\"\n    )\n    parser.add_argument(\"--imsize\", type=int, default=256, help=\"size of the crops\")\n    parser.add_argument(\n        \"--nthread\", type=int, required=True, help=\"number of simultaneous threads\"\n    )\n    parser.add_argument(\n        \"--max-subdir-levels\",\n        type=int,\n        default=5,\n        help=\"maximum number of subdirectories\",\n    )\n    parser.add_argument(\n        \"--ideal-number-pairs-in-dir\",\n        type=int,\n        default=500,\n        help=\"number of pairs stored in a dir\",\n    )\n    return parser\n\n\ndef main(args):\n    listing_path = os.path.join(args.output_dir, \"listing.txt\")\n\n    print(f\"Loading list of crops ... ({args.nthread} threads)\")\n    crops, num_crops_to_generate = load_crop_file(args.crops)\n\n    print(f\"Preparing jobs ({len(crops)} candidate image pairs)...\")\n    num_levels = min(\n        math.ceil(math.log(num_crops_to_generate, args.ideal_number_pairs_in_dir)),\n        args.max_subdir_levels,\n    )\n    num_pairs_in_dir = math.ceil(num_crops_to_generate ** (1 / num_levels))\n\n    jobs = prepare_jobs(crops, num_levels, num_pairs_in_dir)\n    del crops\n\n    os.makedirs(args.output_dir, exist_ok=True)\n    mmap = Pool(args.nthread).imap_unordered if args.nthread > 1 else map\n    call = functools.partial(save_image_crops, args)\n\n    print(f\"Generating cropped images to {args.output_dir} ...\")\n    with open(listing_path, \"w\") as listing:\n        listing.write(\"# pair_path\\n\")\n        for results in tqdm(mmap(call, jobs), total=len(jobs)):\n            for path in results:\n                listing.write(f\"{path}\\n\")\n    print(\"Finished writing listing to\", listing_path)\n\n\ndef load_crop_file(path):\n    data = open(path).read().splitlines()\n    pairs = []\n    num_crops_to_generate = 0\n    for line in tqdm(data):\n        if line.startswith(\"#\"):\n            continue\n        line = line.split(\", \")\n        if len(line) < 8:\n            img1, img2, rotation = line\n            pairs.append((img1, img2, int(rotation), []))\n        else:\n            l1, r1, t1, b1, l2, r2, t2, b2 = map(int, line)\n            rect1, rect2 = (l1, t1, r1, b1), (l2, t2, r2, b2)\n            pairs[-1][-1].append((rect1, rect2))\n            num_crops_to_generate += 1\n    return pairs, num_crops_to_generate\n\n\ndef prepare_jobs(pairs, num_levels, num_pairs_in_dir):\n    jobs = []\n    powers = [num_pairs_in_dir**level for level in reversed(range(num_levels))]\n\n    def get_path(idx):\n        idx_array = []\n        d = idx\n        for level in range(num_levels - 1):\n            idx_array.append(idx // powers[level])\n            idx = idx % powers[level]\n        idx_array.append(d)\n        return \"/\".join(map(lambda x: hex(x)[2:], idx_array))\n\n    idx = 0\n    for pair_data in tqdm(pairs):\n        img1, img2, rotation, crops = pair_data\n        if -60 <= rotation and rotation <= 60:\n            rotation = 0  # most likely not a true rotation\n        paths = [get_path(idx + k) for k in range(len(crops))]\n        idx += len(crops)\n        jobs.append(((img1, img2), rotation, crops, paths))\n    return jobs\n\n\ndef load_image(path):\n    try:\n        return Image.open(path).convert(\"RGB\")\n    except Exception as e:\n        print(\"skipping\", path, e)\n        raise OSError()\n\n\ndef save_image_crops(args, data):\n    # load images\n    img_pair, rot, crops, paths = data\n    try:\n        img1, img2 = [\n            load_image(os.path.join(args.root_dir, impath)) for impath in img_pair\n        ]\n    except OSError as e:\n        return []\n\n    def area(sz):\n        return sz[0] * sz[1]\n\n    tgt_size = (args.imsize, args.imsize)\n\n    def prepare_crop(img, rect, rot=0):\n        # actual crop\n        img = img.crop(rect)\n\n        # resize to desired size\n        interp = (\n            Image.Resampling.LANCZOS\n            if area(img.size) > 4 * area(tgt_size)\n            else Image.Resampling.BICUBIC\n        )\n        img = img.resize(tgt_size, resample=interp)\n\n        # rotate the image\n        rot90 = (round(rot / 90) % 4) * 90\n        if rot90 == 90:\n            img = img.transpose(Image.Transpose.ROTATE_90)\n        elif rot90 == 180:\n            img = img.transpose(Image.Transpose.ROTATE_180)\n        elif rot90 == 270:\n            img = img.transpose(Image.Transpose.ROTATE_270)\n        return img\n\n    results = []\n    for (rect1, rect2), path in zip(crops, paths):\n        crop1 = prepare_crop(img1, rect1)\n        crop2 = prepare_crop(img2, rect2, rot)\n\n        fullpath1 = os.path.join(args.output_dir, path + \"_1.jpg\")\n        fullpath2 = os.path.join(args.output_dir, path + \"_2.jpg\")\n        os.makedirs(os.path.dirname(fullpath1), exist_ok=True)\n\n        assert not os.path.isfile(fullpath1), fullpath1\n        assert not os.path.isfile(fullpath2), fullpath2\n        crop1.save(fullpath1)\n        crop2.save(fullpath2)\n        results.append(path)\n\n    return results\n\n\nif __name__ == \"__main__\":\n    args = arg_parser().parse_args()\n    main(args)","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images.arg_parser","uri":"program://Human3R/function/src.croco.datasets.crops.extract_crops_from_images.arg_parser#L17-L43","kind":"function","name":"arg_parser","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":17,"end_line":43,"context_start_line":1,"context_end_line":63,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# Extracting crops for pre-training\n# --------------------------------------------------------\n\nimport os\nimport argparse\nfrom tqdm import tqdm\nfrom PIL import Image\nimport functools\nfrom multiprocessing import Pool\nimport math\n\n\ndef arg_parser():\n    parser = argparse.ArgumentParser(\n        \"Generate cropped image pairs from image crop list\"\n    )\n\n    parser.add_argument(\"--crops\", type=str, required=True, help=\"crop file\")\n    parser.add_argument(\"--root-dir\", type=str, required=True, help=\"root directory\")\n    parser.add_argument(\n        \"--output-dir\", type=str, required=True, help=\"output directory\"\n    )\n    parser.add_argument(\"--imsize\", type=int, default=256, help=\"size of the crops\")\n    parser.add_argument(\n        \"--nthread\", type=int, required=True, help=\"number of simultaneous threads\"\n    )\n    parser.add_argument(\n        \"--max-subdir-levels\",\n        type=int,\n        default=5,\n        help=\"maximum number of subdirectories\",\n    )\n    parser.add_argument(\n        \"--ideal-number-pairs-in-dir\",\n        type=int,\n        default=500,\n        help=\"number of pairs stored in a dir\",\n    )\n    return parser\n\n\ndef main(args):\n    listing_path = os.path.join(args.output_dir, \"listing.txt\")\n\n    print(f\"Loading list of crops ... ({args.nthread} threads)\")\n    crops, num_crops_to_generate = load_crop_file(args.crops)\n\n    print(f\"Preparing jobs ({len(crops)} candidate image pairs)...\")\n    num_levels = min(\n        math.ceil(math.log(num_crops_to_generate, args.ideal_number_pairs_in_dir)),\n        args.max_subdir_levels,\n    )\n    num_pairs_in_dir = math.ceil(num_crops_to_generate ** (1 / num_levels))\n\n    jobs = prepare_jobs(crops, num_levels, num_pairs_in_dir)\n    del crops\n\n    os.makedirs(args.output_dir, exist_ok=True)\n    mmap = Pool(args.nthread).imap_unordered if args.nthread > 1 else map","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images.main","uri":"program://Human3R/function/src.croco.datasets.crops.extract_crops_from_images.main#L46-L72","kind":"function","name":"main","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":46,"end_line":72,"context_start_line":26,"context_end_line":92,"code":"    )\n    parser.add_argument(\"--imsize\", type=int, default=256, help=\"size of the crops\")\n    parser.add_argument(\n        \"--nthread\", type=int, required=True, help=\"number of simultaneous threads\"\n    )\n    parser.add_argument(\n        \"--max-subdir-levels\",\n        type=int,\n        default=5,\n        help=\"maximum number of subdirectories\",\n    )\n    parser.add_argument(\n        \"--ideal-number-pairs-in-dir\",\n        type=int,\n        default=500,\n        help=\"number of pairs stored in a dir\",\n    )\n    return parser\n\n\ndef main(args):\n    listing_path = os.path.join(args.output_dir, \"listing.txt\")\n\n    print(f\"Loading list of crops ... ({args.nthread} threads)\")\n    crops, num_crops_to_generate = load_crop_file(args.crops)\n\n    print(f\"Preparing jobs ({len(crops)} candidate image pairs)...\")\n    num_levels = min(\n        math.ceil(math.log(num_crops_to_generate, args.ideal_number_pairs_in_dir)),\n        args.max_subdir_levels,\n    )\n    num_pairs_in_dir = math.ceil(num_crops_to_generate ** (1 / num_levels))\n\n    jobs = prepare_jobs(crops, num_levels, num_pairs_in_dir)\n    del crops\n\n    os.makedirs(args.output_dir, exist_ok=True)\n    mmap = Pool(args.nthread).imap_unordered if args.nthread > 1 else map\n    call = functools.partial(save_image_crops, args)\n\n    print(f\"Generating cropped images to {args.output_dir} ...\")\n    with open(listing_path, \"w\") as listing:\n        listing.write(\"# pair_path\\n\")\n        for results in tqdm(mmap(call, jobs), total=len(jobs)):\n            for path in results:\n                listing.write(f\"{path}\\n\")\n    print(\"Finished writing listing to\", listing_path)\n\n\ndef load_crop_file(path):\n    data = open(path).read().splitlines()\n    pairs = []\n    num_crops_to_generate = 0\n    for line in tqdm(data):\n        if line.startswith(\"#\"):\n            continue\n        line = line.split(\", \")\n        if len(line) < 8:\n            img1, img2, rotation = line\n            pairs.append((img1, img2, int(rotation), []))\n        else:\n            l1, r1, t1, b1, l2, r2, t2, b2 = map(int, line)\n            rect1, rect2 = (l1, t1, r1, b1), (l2, t2, r2, b2)\n            pairs[-1][-1].append((rect1, rect2))\n            num_crops_to_generate += 1\n    return pairs, num_crops_to_generate\n","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images.load_crop_file","uri":"program://Human3R/function/src.croco.datasets.crops.extract_crops_from_images.load_crop_file#L75-L91","kind":"function","name":"load_crop_file","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":75,"end_line":91,"context_start_line":55,"context_end_line":111,"code":"        args.max_subdir_levels,\n    )\n    num_pairs_in_dir = math.ceil(num_crops_to_generate ** (1 / num_levels))\n\n    jobs = prepare_jobs(crops, num_levels, num_pairs_in_dir)\n    del crops\n\n    os.makedirs(args.output_dir, exist_ok=True)\n    mmap = Pool(args.nthread).imap_unordered if args.nthread > 1 else map\n    call = functools.partial(save_image_crops, args)\n\n    print(f\"Generating cropped images to {args.output_dir} ...\")\n    with open(listing_path, \"w\") as listing:\n        listing.write(\"# pair_path\\n\")\n        for results in tqdm(mmap(call, jobs), total=len(jobs)):\n            for path in results:\n                listing.write(f\"{path}\\n\")\n    print(\"Finished writing listing to\", listing_path)\n\n\ndef load_crop_file(path):\n    data = open(path).read().splitlines()\n    pairs = []\n    num_crops_to_generate = 0\n    for line in tqdm(data):\n        if line.startswith(\"#\"):\n            continue\n        line = line.split(\", \")\n        if len(line) < 8:\n            img1, img2, rotation = line\n            pairs.append((img1, img2, int(rotation), []))\n        else:\n            l1, r1, t1, b1, l2, r2, t2, b2 = map(int, line)\n            rect1, rect2 = (l1, t1, r1, b1), (l2, t2, r2, b2)\n            pairs[-1][-1].append((rect1, rect2))\n            num_crops_to_generate += 1\n    return pairs, num_crops_to_generate\n\n\ndef prepare_jobs(pairs, num_levels, num_pairs_in_dir):\n    jobs = []\n    powers = [num_pairs_in_dir**level for level in reversed(range(num_levels))]\n\n    def get_path(idx):\n        idx_array = []\n        d = idx\n        for level in range(num_levels - 1):\n            idx_array.append(idx // powers[level])\n            idx = idx % powers[level]\n        idx_array.append(d)\n        return \"/\".join(map(lambda x: hex(x)[2:], idx_array))\n\n    idx = 0\n    for pair_data in tqdm(pairs):\n        img1, img2, rotation, crops = pair_data\n        if -60 <= rotation and rotation <= 60:\n            rotation = 0  # most likely not a true rotation","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images.prepare_jobs","uri":"program://Human3R/function/src.croco.datasets.crops.extract_crops_from_images.prepare_jobs#L94-L115","kind":"function","name":"prepare_jobs","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":94,"end_line":115,"context_start_line":74,"context_end_line":135,"code":"\ndef load_crop_file(path):\n    data = open(path).read().splitlines()\n    pairs = []\n    num_crops_to_generate = 0\n    for line in tqdm(data):\n        if line.startswith(\"#\"):\n            continue\n        line = line.split(\", \")\n        if len(line) < 8:\n            img1, img2, rotation = line\n            pairs.append((img1, img2, int(rotation), []))\n        else:\n            l1, r1, t1, b1, l2, r2, t2, b2 = map(int, line)\n            rect1, rect2 = (l1, t1, r1, b1), (l2, t2, r2, b2)\n            pairs[-1][-1].append((rect1, rect2))\n            num_crops_to_generate += 1\n    return pairs, num_crops_to_generate\n\n\ndef prepare_jobs(pairs, num_levels, num_pairs_in_dir):\n    jobs = []\n    powers = [num_pairs_in_dir**level for level in reversed(range(num_levels))]\n\n    def get_path(idx):\n        idx_array = []\n        d = idx\n        for level in range(num_levels - 1):\n            idx_array.append(idx // powers[level])\n            idx = idx % powers[level]\n        idx_array.append(d)\n        return \"/\".join(map(lambda x: hex(x)[2:], idx_array))\n\n    idx = 0\n    for pair_data in tqdm(pairs):\n        img1, img2, rotation, crops = pair_data\n        if -60 <= rotation and rotation <= 60:\n            rotation = 0  # most likely not a true rotation\n        paths = [get_path(idx + k) for k in range(len(crops))]\n        idx += len(crops)\n        jobs.append(((img1, img2), rotation, crops, paths))\n    return jobs\n\n\ndef load_image(path):\n    try:\n        return Image.open(path).convert(\"RGB\")\n    except Exception as e:\n        print(\"skipping\", path, e)\n        raise OSError()\n\n\ndef save_image_crops(args, data):\n    # load images\n    img_pair, rot, crops, paths = data\n    try:\n        img1, img2 = [\n            load_image(os.path.join(args.root_dir, impath)) for impath in img_pair\n        ]\n    except OSError as e:\n        return []\n","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images.load_image","uri":"program://Human3R/function/src.croco.datasets.crops.extract_crops_from_images.load_image#L118-L123","kind":"function","name":"load_image","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":118,"end_line":123,"context_start_line":98,"context_end_line":143,"code":"    def get_path(idx):\n        idx_array = []\n        d = idx\n        for level in range(num_levels - 1):\n            idx_array.append(idx // powers[level])\n            idx = idx % powers[level]\n        idx_array.append(d)\n        return \"/\".join(map(lambda x: hex(x)[2:], idx_array))\n\n    idx = 0\n    for pair_data in tqdm(pairs):\n        img1, img2, rotation, crops = pair_data\n        if -60 <= rotation and rotation <= 60:\n            rotation = 0  # most likely not a true rotation\n        paths = [get_path(idx + k) for k in range(len(crops))]\n        idx += len(crops)\n        jobs.append(((img1, img2), rotation, crops, paths))\n    return jobs\n\n\ndef load_image(path):\n    try:\n        return Image.open(path).convert(\"RGB\")\n    except Exception as e:\n        print(\"skipping\", path, e)\n        raise OSError()\n\n\ndef save_image_crops(args, data):\n    # load images\n    img_pair, rot, crops, paths = data\n    try:\n        img1, img2 = [\n            load_image(os.path.join(args.root_dir, impath)) for impath in img_pair\n        ]\n    except OSError as e:\n        return []\n\n    def area(sz):\n        return sz[0] * sz[1]\n\n    tgt_size = (args.imsize, args.imsize)\n\n    def prepare_crop(img, rect, rot=0):\n        # actual crop\n        img = img.crop(rect)","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images.save_image_crops","uri":"program://Human3R/function/src.croco.datasets.crops.extract_crops_from_images.save_image_crops#L126-L178","kind":"function","name":"save_image_crops","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":126,"end_line":178,"context_start_line":106,"context_end_line":183,"code":"\n    idx = 0\n    for pair_data in tqdm(pairs):\n        img1, img2, rotation, crops = pair_data\n        if -60 <= rotation and rotation <= 60:\n            rotation = 0  # most likely not a true rotation\n        paths = [get_path(idx + k) for k in range(len(crops))]\n        idx += len(crops)\n        jobs.append(((img1, img2), rotation, crops, paths))\n    return jobs\n\n\ndef load_image(path):\n    try:\n        return Image.open(path).convert(\"RGB\")\n    except Exception as e:\n        print(\"skipping\", path, e)\n        raise OSError()\n\n\ndef save_image_crops(args, data):\n    # load images\n    img_pair, rot, crops, paths = data\n    try:\n        img1, img2 = [\n            load_image(os.path.join(args.root_dir, impath)) for impath in img_pair\n        ]\n    except OSError as e:\n        return []\n\n    def area(sz):\n        return sz[0] * sz[1]\n\n    tgt_size = (args.imsize, args.imsize)\n\n    def prepare_crop(img, rect, rot=0):\n        # actual crop\n        img = img.crop(rect)\n\n        # resize to desired size\n        interp = (\n            Image.Resampling.LANCZOS\n            if area(img.size) > 4 * area(tgt_size)\n            else Image.Resampling.BICUBIC\n        )\n        img = img.resize(tgt_size, resample=interp)\n\n        # rotate the image\n        rot90 = (round(rot / 90) % 4) * 90\n        if rot90 == 90:\n            img = img.transpose(Image.Transpose.ROTATE_90)\n        elif rot90 == 180:\n            img = img.transpose(Image.Transpose.ROTATE_180)\n        elif rot90 == 270:\n            img = img.transpose(Image.Transpose.ROTATE_270)\n        return img\n\n    results = []\n    for (rect1, rect2), path in zip(crops, paths):\n        crop1 = prepare_crop(img1, rect1)\n        crop2 = prepare_crop(img2, rect2, rot)\n\n        fullpath1 = os.path.join(args.output_dir, path + \"_1.jpg\")\n        fullpath2 = os.path.join(args.output_dir, path + \"_2.jpg\")\n        os.makedirs(os.path.dirname(fullpath1), exist_ok=True)\n\n        assert not os.path.isfile(fullpath1), fullpath1\n        assert not os.path.isfile(fullpath2), fullpath2\n        crop1.save(fullpath1)\n        crop2.save(fullpath2)\n        results.append(path)\n\n    return results\n\n\nif __name__ == \"__main__\":\n    args = arg_parser().parse_args()\n    main(args)","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images.get_path","uri":"program://Human3R/function/src.croco.datasets.crops.extract_crops_from_images.get_path#L98-L105","kind":"function","name":"get_path","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":98,"end_line":105,"context_start_line":78,"context_end_line":125,"code":"    num_crops_to_generate = 0\n    for line in tqdm(data):\n        if line.startswith(\"#\"):\n            continue\n        line = line.split(\", \")\n        if len(line) < 8:\n            img1, img2, rotation = line\n            pairs.append((img1, img2, int(rotation), []))\n        else:\n            l1, r1, t1, b1, l2, r2, t2, b2 = map(int, line)\n            rect1, rect2 = (l1, t1, r1, b1), (l2, t2, r2, b2)\n            pairs[-1][-1].append((rect1, rect2))\n            num_crops_to_generate += 1\n    return pairs, num_crops_to_generate\n\n\ndef prepare_jobs(pairs, num_levels, num_pairs_in_dir):\n    jobs = []\n    powers = [num_pairs_in_dir**level for level in reversed(range(num_levels))]\n\n    def get_path(idx):\n        idx_array = []\n        d = idx\n        for level in range(num_levels - 1):\n            idx_array.append(idx // powers[level])\n            idx = idx % powers[level]\n        idx_array.append(d)\n        return \"/\".join(map(lambda x: hex(x)[2:], idx_array))\n\n    idx = 0\n    for pair_data in tqdm(pairs):\n        img1, img2, rotation, crops = pair_data\n        if -60 <= rotation and rotation <= 60:\n            rotation = 0  # most likely not a true rotation\n        paths = [get_path(idx + k) for k in range(len(crops))]\n        idx += len(crops)\n        jobs.append(((img1, img2), rotation, crops, paths))\n    return jobs\n\n\ndef load_image(path):\n    try:\n        return Image.open(path).convert(\"RGB\")\n    except Exception as e:\n        print(\"skipping\", path, e)\n        raise OSError()\n\n","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images.area","uri":"program://Human3R/function/src.croco.datasets.crops.extract_crops_from_images.area#L136-L137","kind":"function","name":"area","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":136,"end_line":137,"context_start_line":116,"context_end_line":157,"code":"\n\ndef load_image(path):\n    try:\n        return Image.open(path).convert(\"RGB\")\n    except Exception as e:\n        print(\"skipping\", path, e)\n        raise OSError()\n\n\ndef save_image_crops(args, data):\n    # load images\n    img_pair, rot, crops, paths = data\n    try:\n        img1, img2 = [\n            load_image(os.path.join(args.root_dir, impath)) for impath in img_pair\n        ]\n    except OSError as e:\n        return []\n\n    def area(sz):\n        return sz[0] * sz[1]\n\n    tgt_size = (args.imsize, args.imsize)\n\n    def prepare_crop(img, rect, rot=0):\n        # actual crop\n        img = img.crop(rect)\n\n        # resize to desired size\n        interp = (\n            Image.Resampling.LANCZOS\n            if area(img.size) > 4 * area(tgt_size)\n            else Image.Resampling.BICUBIC\n        )\n        img = img.resize(tgt_size, resample=interp)\n\n        # rotate the image\n        rot90 = (round(rot / 90) % 4) * 90\n        if rot90 == 90:\n            img = img.transpose(Image.Transpose.ROTATE_90)\n        elif rot90 == 180:","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.crops.extract_crops_from_images.prepare_crop","uri":"program://Human3R/function/src.croco.datasets.crops.extract_crops_from_images.prepare_crop#L141-L161","kind":"function","name":"prepare_crop","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":141,"end_line":161,"context_start_line":121,"context_end_line":181,"code":"    except Exception as e:\n        print(\"skipping\", path, e)\n        raise OSError()\n\n\ndef save_image_crops(args, data):\n    # load images\n    img_pair, rot, crops, paths = data\n    try:\n        img1, img2 = [\n            load_image(os.path.join(args.root_dir, impath)) for impath in img_pair\n        ]\n    except OSError as e:\n        return []\n\n    def area(sz):\n        return sz[0] * sz[1]\n\n    tgt_size = (args.imsize, args.imsize)\n\n    def prepare_crop(img, rect, rot=0):\n        # actual crop\n        img = img.crop(rect)\n\n        # resize to desired size\n        interp = (\n            Image.Resampling.LANCZOS\n            if area(img.size) > 4 * area(tgt_size)\n            else Image.Resampling.BICUBIC\n        )\n        img = img.resize(tgt_size, resample=interp)\n\n        # rotate the image\n        rot90 = (round(rot / 90) % 4) * 90\n        if rot90 == 90:\n            img = img.transpose(Image.Transpose.ROTATE_90)\n        elif rot90 == 180:\n            img = img.transpose(Image.Transpose.ROTATE_180)\n        elif rot90 == 270:\n            img = img.transpose(Image.Transpose.ROTATE_270)\n        return img\n\n    results = []\n    for (rect1, rect2), path in zip(crops, paths):\n        crop1 = prepare_crop(img1, rect1)\n        crop2 = prepare_crop(img2, rect2, rot)\n\n        fullpath1 = os.path.join(args.output_dir, path + \"_1.jpg\")\n        fullpath2 = os.path.join(args.output_dir, path + \"_2.jpg\")\n        os.makedirs(os.path.dirname(fullpath1), exist_ok=True)\n\n        assert not os.path.isfile(fullpath1), fullpath1\n        assert not os.path.isfile(fullpath2), fullpath2\n        crop1.save(fullpath1)\n        crop2.save(fullpath2)\n        results.append(path)\n\n    return results\n\n\nif __name__ == \"__main__\":","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.paths","uri":"program://Human3R/module/src.croco.datasets.habitat_sim.paths#L1-L179","kind":"module","name":"src.croco.datasets.habitat_sim.paths","path":"src/croco/datasets/habitat_sim/paths.py","language":"python","start_line":1,"end_line":179,"context_start_line":1,"context_end_line":179,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\"\"\"\nPaths to Habitat-Sim scenes\n\"\"\"\n\nimport os\nimport json\nimport collections\nfrom tqdm import tqdm\n\n\n# Hardcoded path to the different scene datasets\nSCENES_DATASET = {\n    \"hm3d\": \"./data/habitat-sim-data/scene_datasets/hm3d/\",\n    \"gibson\": \"./data/habitat-sim-data/scene_datasets/gibson/\",\n    \"habitat-test-scenes\": \"./data/habitat-sim/scene_datasets/habitat-test-scenes/\",\n    \"replica_cad_baked_lighting\": \"./data/habitat-sim/scene_datasets/replica_cad_baked_lighting/\",\n    \"replica_cad\": \"./data/habitat-sim/scene_datasets/replica_cad/\",\n    \"replica\": \"./data/habitat-sim/scene_datasets/ReplicaDataset/\",\n    \"scannet\": \"./data/habitat-sim/scene_datasets/scannet/\",\n}\n\nSceneData = collections.namedtuple(\n    \"SceneData\", [\"scene_dataset_config_file\", \"scene\", \"navmesh\", \"output_dir\"]\n)\n\n\ndef list_replicacad_scenes(base_output_dir, base_path=SCENES_DATASET[\"replica_cad\"]):\n    scene_dataset_config_file = os.path.join(\n        base_path, \"replicaCAD.scene_dataset_config.json\"\n    )\n    scenes = [f\"apt_{i}\" for i in range(6)] + [\"empty_stage\"]\n    navmeshes = [f\"navmeshes/apt_{i}_static_furniture.navmesh\" for i in range(6)] + [\n        \"empty_stage.navmesh\"\n    ]\n    scenes_data = []\n    for idx in range(len(scenes)):\n        output_dir = os.path.join(base_output_dir, \"ReplicaCAD\", scenes[idx])\n        # Add scene\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scenes[idx] + \".scene_instance.json\",\n            navmesh=os.path.join(base_path, navmeshes[idx]),\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data\n\n\ndef list_replica_cad_baked_lighting_scenes(\n    base_output_dir, base_path=SCENES_DATASET[\"replica_cad_baked_lighting\"]\n):\n    scene_dataset_config_file = os.path.join(\n        base_path, \"replicaCAD_baked.scene_dataset_config.json\"\n    )\n    scenes = sum(\n        [[f\"Baked_sc{i}_staging_{j:02}\" for i in range(5)] for j in range(21)], []\n    )\n    navmeshes = \"\"  # [f\"navmeshes/apt_{i}_static_furniture.navmesh\" for i in range(6)] + [\"empty_stage.navmesh\"]\n    scenes_data = []\n    for idx in range(len(scenes)):\n        output_dir = os.path.join(\n            base_output_dir, \"replica_cad_baked_lighting\", scenes[idx]\n        )\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scenes[idx],\n            navmesh=\"\",\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data\n\n\ndef list_replica_scenes(base_output_dir, base_path):\n    scenes_data = []\n    for scene_id in os.listdir(base_path):\n        scene = os.path.join(base_path, scene_id, \"mesh.ply\")\n        navmesh = os.path.join(\n            base_path, scene_id, \"habitat/mesh_preseg_semantic.navmesh\"\n        )  # Not sure if I should use it\n        scene_dataset_config_file = \"\"\n        output_dir = os.path.join(base_output_dir, scene_id)\n        # Add scene only if it does not exist already, or if exist_ok\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scene,\n            navmesh=navmesh,\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data\n\n\ndef list_scenes(base_output_dir, base_path):\n    \"\"\"\n    Generic method iterating through a base_path folder to find scenes.\n    \"\"\"\n    scenes_data = []\n    for root, dirs, files in os.walk(base_path, followlinks=True):\n        folder_scenes_data = []\n        for file in files:\n            name, ext = os.path.splitext(file)\n            if ext == \".glb\":\n                scene = os.path.join(root, name + \".glb\")\n                navmesh = os.path.join(root, name + \".navmesh\")\n                if not os.path.exists(navmesh):\n                    navmesh = \"\"\n                relpath = os.path.relpath(root, base_path)\n                output_dir = os.path.abspath(\n                    os.path.join(base_output_dir, relpath, name)\n                )\n                data = SceneData(\n                    scene_dataset_config_file=\"\",\n                    scene=scene,\n                    navmesh=navmesh,\n                    output_dir=output_dir,\n                )\n                folder_scenes_data.append(data)\n\n        # Specific check for HM3D:\n        # When two meshesxxxx.basis.glb and xxxx.glb are present, use the 'basis' version.\n        basis_scenes = [\n            data.scene[: -len(\".basis.glb\")]\n            for data in folder_scenes_data\n            if data.scene.endswith(\".basis.glb\")\n        ]\n        if len(basis_scenes) != 0:\n            folder_scenes_data = [\n                data\n                for data in folder_scenes_data\n                if not (data.scene[: -len(\".glb\")] in basis_scenes)\n            ]\n\n        scenes_data.extend(folder_scenes_data)\n    return scenes_data\n\n\ndef list_scenes_available(base_output_dir, scenes_dataset_paths=SCENES_DATASET):\n    scenes_data = []\n\n    # HM3D\n    for split in (\"minival\", \"train\", \"val\", \"examples\"):\n        scenes_data += list_scenes(\n            base_output_dir=os.path.join(base_output_dir, f\"hm3d/{split}/\"),\n            base_path=f\"{scenes_dataset_paths['hm3d']}/{split}\",\n        )\n\n    # Gibson\n    scenes_data += list_scenes(\n        base_output_dir=os.path.join(base_output_dir, \"gibson\"),\n        base_path=scenes_dataset_paths[\"gibson\"],\n    )\n\n    # Habitat test scenes (just a few)\n    scenes_data += list_scenes(\n        base_output_dir=os.path.join(base_output_dir, \"habitat-test-scenes\"),\n        base_path=scenes_dataset_paths[\"habitat-test-scenes\"],\n    )\n\n    # ReplicaCAD (baked lightning)\n    scenes_data += list_replica_cad_baked_lighting_scenes(\n        base_output_dir=base_output_dir\n    )\n\n    # ScanNet\n    scenes_data += list_scenes(\n        base_output_dir=os.path.join(base_output_dir, \"scannet\"),\n        base_path=scenes_dataset_paths[\"scannet\"],\n    )\n\n    # Replica\n    list_replica_scenes(\n        base_output_dir=os.path.join(base_output_dir, \"replica\"),\n        base_path=scenes_dataset_paths[\"replica\"],\n    )\n    return scenes_data","source_hash":"2c46a5b329717acf7032810cc9ed6a556007a2ba566ff7374849e5614a3ff37a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.paths.list_replicacad_scenes","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.paths.list_replicacad_scenes#L30-L49","kind":"function","name":"list_replicacad_scenes","path":"src/croco/datasets/habitat_sim/paths.py","language":"python","start_line":30,"end_line":49,"context_start_line":10,"context_end_line":69,"code":"import collections\nfrom tqdm import tqdm\n\n\n# Hardcoded path to the different scene datasets\nSCENES_DATASET = {\n    \"hm3d\": \"./data/habitat-sim-data/scene_datasets/hm3d/\",\n    \"gibson\": \"./data/habitat-sim-data/scene_datasets/gibson/\",\n    \"habitat-test-scenes\": \"./data/habitat-sim/scene_datasets/habitat-test-scenes/\",\n    \"replica_cad_baked_lighting\": \"./data/habitat-sim/scene_datasets/replica_cad_baked_lighting/\",\n    \"replica_cad\": \"./data/habitat-sim/scene_datasets/replica_cad/\",\n    \"replica\": \"./data/habitat-sim/scene_datasets/ReplicaDataset/\",\n    \"scannet\": \"./data/habitat-sim/scene_datasets/scannet/\",\n}\n\nSceneData = collections.namedtuple(\n    \"SceneData\", [\"scene_dataset_config_file\", \"scene\", \"navmesh\", \"output_dir\"]\n)\n\n\ndef list_replicacad_scenes(base_output_dir, base_path=SCENES_DATASET[\"replica_cad\"]):\n    scene_dataset_config_file = os.path.join(\n        base_path, \"replicaCAD.scene_dataset_config.json\"\n    )\n    scenes = [f\"apt_{i}\" for i in range(6)] + [\"empty_stage\"]\n    navmeshes = [f\"navmeshes/apt_{i}_static_furniture.navmesh\" for i in range(6)] + [\n        \"empty_stage.navmesh\"\n    ]\n    scenes_data = []\n    for idx in range(len(scenes)):\n        output_dir = os.path.join(base_output_dir, \"ReplicaCAD\", scenes[idx])\n        # Add scene\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scenes[idx] + \".scene_instance.json\",\n            navmesh=os.path.join(base_path, navmeshes[idx]),\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data\n\n\ndef list_replica_cad_baked_lighting_scenes(\n    base_output_dir, base_path=SCENES_DATASET[\"replica_cad_baked_lighting\"]\n):\n    scene_dataset_config_file = os.path.join(\n        base_path, \"replicaCAD_baked.scene_dataset_config.json\"\n    )\n    scenes = sum(\n        [[f\"Baked_sc{i}_staging_{j:02}\" for i in range(5)] for j in range(21)], []\n    )\n    navmeshes = \"\"  # [f\"navmeshes/apt_{i}_static_furniture.navmesh\" for i in range(6)] + [\"empty_stage.navmesh\"]\n    scenes_data = []\n    for idx in range(len(scenes)):\n        output_dir = os.path.join(\n            base_output_dir, \"replica_cad_baked_lighting\", scenes[idx]\n        )\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scenes[idx],","source_hash":"2c46a5b329717acf7032810cc9ed6a556007a2ba566ff7374849e5614a3ff37a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.paths.list_replica_cad_baked_lighting_scenes","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.paths.list_replica_cad_baked_lighting_scenes#L52-L74","kind":"function","name":"list_replica_cad_baked_lighting_scenes","path":"src/croco/datasets/habitat_sim/paths.py","language":"python","start_line":52,"end_line":74,"context_start_line":32,"context_end_line":94,"code":"        base_path, \"replicaCAD.scene_dataset_config.json\"\n    )\n    scenes = [f\"apt_{i}\" for i in range(6)] + [\"empty_stage\"]\n    navmeshes = [f\"navmeshes/apt_{i}_static_furniture.navmesh\" for i in range(6)] + [\n        \"empty_stage.navmesh\"\n    ]\n    scenes_data = []\n    for idx in range(len(scenes)):\n        output_dir = os.path.join(base_output_dir, \"ReplicaCAD\", scenes[idx])\n        # Add scene\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scenes[idx] + \".scene_instance.json\",\n            navmesh=os.path.join(base_path, navmeshes[idx]),\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data\n\n\ndef list_replica_cad_baked_lighting_scenes(\n    base_output_dir, base_path=SCENES_DATASET[\"replica_cad_baked_lighting\"]\n):\n    scene_dataset_config_file = os.path.join(\n        base_path, \"replicaCAD_baked.scene_dataset_config.json\"\n    )\n    scenes = sum(\n        [[f\"Baked_sc{i}_staging_{j:02}\" for i in range(5)] for j in range(21)], []\n    )\n    navmeshes = \"\"  # [f\"navmeshes/apt_{i}_static_furniture.navmesh\" for i in range(6)] + [\"empty_stage.navmesh\"]\n    scenes_data = []\n    for idx in range(len(scenes)):\n        output_dir = os.path.join(\n            base_output_dir, \"replica_cad_baked_lighting\", scenes[idx]\n        )\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scenes[idx],\n            navmesh=\"\",\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data\n\n\ndef list_replica_scenes(base_output_dir, base_path):\n    scenes_data = []\n    for scene_id in os.listdir(base_path):\n        scene = os.path.join(base_path, scene_id, \"mesh.ply\")\n        navmesh = os.path.join(\n            base_path, scene_id, \"habitat/mesh_preseg_semantic.navmesh\"\n        )  # Not sure if I should use it\n        scene_dataset_config_file = \"\"\n        output_dir = os.path.join(base_output_dir, scene_id)\n        # Add scene only if it does not exist already, or if exist_ok\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scene,\n            navmesh=navmesh,\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data","source_hash":"2c46a5b329717acf7032810cc9ed6a556007a2ba566ff7374849e5614a3ff37a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.paths.list_replica_scenes","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.paths.list_replica_scenes#L77-L94","kind":"function","name":"list_replica_scenes","path":"src/croco/datasets/habitat_sim/paths.py","language":"python","start_line":77,"end_line":94,"context_start_line":57,"context_end_line":114,"code":"    )\n    scenes = sum(\n        [[f\"Baked_sc{i}_staging_{j:02}\" for i in range(5)] for j in range(21)], []\n    )\n    navmeshes = \"\"  # [f\"navmeshes/apt_{i}_static_furniture.navmesh\" for i in range(6)] + [\"empty_stage.navmesh\"]\n    scenes_data = []\n    for idx in range(len(scenes)):\n        output_dir = os.path.join(\n            base_output_dir, \"replica_cad_baked_lighting\", scenes[idx]\n        )\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scenes[idx],\n            navmesh=\"\",\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data\n\n\ndef list_replica_scenes(base_output_dir, base_path):\n    scenes_data = []\n    for scene_id in os.listdir(base_path):\n        scene = os.path.join(base_path, scene_id, \"mesh.ply\")\n        navmesh = os.path.join(\n            base_path, scene_id, \"habitat/mesh_preseg_semantic.navmesh\"\n        )  # Not sure if I should use it\n        scene_dataset_config_file = \"\"\n        output_dir = os.path.join(base_output_dir, scene_id)\n        # Add scene only if it does not exist already, or if exist_ok\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scene,\n            navmesh=navmesh,\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data\n\n\ndef list_scenes(base_output_dir, base_path):\n    \"\"\"\n    Generic method iterating through a base_path folder to find scenes.\n    \"\"\"\n    scenes_data = []\n    for root, dirs, files in os.walk(base_path, followlinks=True):\n        folder_scenes_data = []\n        for file in files:\n            name, ext = os.path.splitext(file)\n            if ext == \".glb\":\n                scene = os.path.join(root, name + \".glb\")\n                navmesh = os.path.join(root, name + \".navmesh\")\n                if not os.path.exists(navmesh):\n                    navmesh = \"\"\n                relpath = os.path.relpath(root, base_path)\n                output_dir = os.path.abspath(\n                    os.path.join(base_output_dir, relpath, name)\n                )","source_hash":"2c46a5b329717acf7032810cc9ed6a556007a2ba566ff7374849e5614a3ff37a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.paths.list_scenes","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.paths.list_scenes#L97-L138","kind":"function","name":"list_scenes","path":"src/croco/datasets/habitat_sim/paths.py","language":"python","start_line":97,"end_line":138,"context_start_line":77,"context_end_line":158,"code":"def list_replica_scenes(base_output_dir, base_path):\n    scenes_data = []\n    for scene_id in os.listdir(base_path):\n        scene = os.path.join(base_path, scene_id, \"mesh.ply\")\n        navmesh = os.path.join(\n            base_path, scene_id, \"habitat/mesh_preseg_semantic.navmesh\"\n        )  # Not sure if I should use it\n        scene_dataset_config_file = \"\"\n        output_dir = os.path.join(base_output_dir, scene_id)\n        # Add scene only if it does not exist already, or if exist_ok\n        data = SceneData(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scene,\n            navmesh=navmesh,\n            output_dir=output_dir,\n        )\n        scenes_data.append(data)\n    return scenes_data\n\n\ndef list_scenes(base_output_dir, base_path):\n    \"\"\"\n    Generic method iterating through a base_path folder to find scenes.\n    \"\"\"\n    scenes_data = []\n    for root, dirs, files in os.walk(base_path, followlinks=True):\n        folder_scenes_data = []\n        for file in files:\n            name, ext = os.path.splitext(file)\n            if ext == \".glb\":\n                scene = os.path.join(root, name + \".glb\")\n                navmesh = os.path.join(root, name + \".navmesh\")\n                if not os.path.exists(navmesh):\n                    navmesh = \"\"\n                relpath = os.path.relpath(root, base_path)\n                output_dir = os.path.abspath(\n                    os.path.join(base_output_dir, relpath, name)\n                )\n                data = SceneData(\n                    scene_dataset_config_file=\"\",\n                    scene=scene,\n                    navmesh=navmesh,\n                    output_dir=output_dir,\n                )\n                folder_scenes_data.append(data)\n\n        # Specific check for HM3D:\n        # When two meshesxxxx.basis.glb and xxxx.glb are present, use the 'basis' version.\n        basis_scenes = [\n            data.scene[: -len(\".basis.glb\")]\n            for data in folder_scenes_data\n            if data.scene.endswith(\".basis.glb\")\n        ]\n        if len(basis_scenes) != 0:\n            folder_scenes_data = [\n                data\n                for data in folder_scenes_data\n                if not (data.scene[: -len(\".glb\")] in basis_scenes)\n            ]\n\n        scenes_data.extend(folder_scenes_data)\n    return scenes_data\n\n\ndef list_scenes_available(base_output_dir, scenes_dataset_paths=SCENES_DATASET):\n    scenes_data = []\n\n    # HM3D\n    for split in (\"minival\", \"train\", \"val\", \"examples\"):\n        scenes_data += list_scenes(\n            base_output_dir=os.path.join(base_output_dir, f\"hm3d/{split}/\"),\n            base_path=f\"{scenes_dataset_paths['hm3d']}/{split}\",\n        )\n\n    # Gibson\n    scenes_data += list_scenes(\n        base_output_dir=os.path.join(base_output_dir, \"gibson\"),\n        base_path=scenes_dataset_paths[\"gibson\"],\n    )\n\n    # Habitat test scenes (just a few)\n    scenes_data += list_scenes(","source_hash":"2c46a5b329717acf7032810cc9ed6a556007a2ba566ff7374849e5614a3ff37a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.paths.list_scenes_available","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.paths.list_scenes_available#L141-L179","kind":"function","name":"list_scenes_available","path":"src/croco/datasets/habitat_sim/paths.py","language":"python","start_line":141,"end_line":179,"context_start_line":121,"context_end_line":179,"code":"                folder_scenes_data.append(data)\n\n        # Specific check for HM3D:\n        # When two meshesxxxx.basis.glb and xxxx.glb are present, use the 'basis' version.\n        basis_scenes = [\n            data.scene[: -len(\".basis.glb\")]\n            for data in folder_scenes_data\n            if data.scene.endswith(\".basis.glb\")\n        ]\n        if len(basis_scenes) != 0:\n            folder_scenes_data = [\n                data\n                for data in folder_scenes_data\n                if not (data.scene[: -len(\".glb\")] in basis_scenes)\n            ]\n\n        scenes_data.extend(folder_scenes_data)\n    return scenes_data\n\n\ndef list_scenes_available(base_output_dir, scenes_dataset_paths=SCENES_DATASET):\n    scenes_data = []\n\n    # HM3D\n    for split in (\"minival\", \"train\", \"val\", \"examples\"):\n        scenes_data += list_scenes(\n            base_output_dir=os.path.join(base_output_dir, f\"hm3d/{split}/\"),\n            base_path=f\"{scenes_dataset_paths['hm3d']}/{split}\",\n        )\n\n    # Gibson\n    scenes_data += list_scenes(\n        base_output_dir=os.path.join(base_output_dir, \"gibson\"),\n        base_path=scenes_dataset_paths[\"gibson\"],\n    )\n\n    # Habitat test scenes (just a few)\n    scenes_data += list_scenes(\n        base_output_dir=os.path.join(base_output_dir, \"habitat-test-scenes\"),\n        base_path=scenes_dataset_paths[\"habitat-test-scenes\"],\n    )\n\n    # ReplicaCAD (baked lightning)\n    scenes_data += list_replica_cad_baked_lighting_scenes(\n        base_output_dir=base_output_dir\n    )\n\n    # ScanNet\n    scenes_data += list_scenes(\n        base_output_dir=os.path.join(base_output_dir, \"scannet\"),\n        base_path=scenes_dataset_paths[\"scannet\"],\n    )\n\n    # Replica\n    list_replica_scenes(\n        base_output_dir=os.path.join(base_output_dir, \"replica\"),\n        base_path=scenes_dataset_paths[\"replica\"],\n    )\n    return scenes_data","source_hash":"2c46a5b329717acf7032810cc9ed6a556007a2ba566ff7374849e5614a3ff37a","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.generate_multiview_images","uri":"program://Human3R/module/src.croco.datasets.habitat_sim.generate_multiview_images#L1-L231","kind":"module","name":"src.croco.datasets.habitat_sim.generate_multiview_images","path":"src/croco/datasets/habitat_sim/generate_multiview_images.py","language":"python","start_line":1,"end_line":231,"context_start_line":1,"context_end_line":231,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nfrom tqdm import tqdm\nimport argparse\nimport PIL.Image\nimport numpy as np\nimport json\nfrom datasets.habitat_sim.multiview_habitat_sim_generator import (\n    MultiviewHabitatSimGenerator,\n    NoNaviguableSpaceError,\n)\nfrom datasets.habitat_sim.paths import list_scenes_available\nimport cv2\nimport quaternion\nimport shutil\n\n\ndef generate_multiview_images_for_scene(\n    scene_dataset_config_file,\n    scene,\n    navmesh,\n    output_dir,\n    views_count,\n    size,\n    exist_ok=False,\n    generate_depth=False,\n    **kwargs,\n):\n    \"\"\"\n    Generate tuples of overlapping views for a given scene.\n    generate_depth: generate depth images and camera parameters.\n    \"\"\"\n    if os.path.exists(output_dir) and not exist_ok:\n        print(f\"Scene {scene}: data already generated. Ignoring generation.\")\n        return\n    try:\n        print(f\"Scene {scene}: {size} multiview acquisitions to generate...\")\n        os.makedirs(output_dir, exist_ok=exist_ok)\n\n        metadata_filename = os.path.join(output_dir, \"metadata.json\")\n\n        metadata_template = dict(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scene,\n            navmesh=navmesh,\n            views_count=views_count,\n            size=size,\n            generate_depth=generate_depth,\n            **kwargs,\n        )\n        metadata_template[\"multiviews\"] = dict()\n\n        if os.path.exists(metadata_filename):\n            print(\"Metadata file already exists:\", metadata_filename)\n            print(\"Loading already generated metadata file...\")\n            with open(metadata_filename, \"r\") as f:\n                metadata = json.load(f)\n\n            for key in metadata_template.keys():\n                if key != \"multiviews\":\n                    assert (\n                        metadata_template[key] == metadata[key]\n                    ), f\"existing file is inconsistent with the input parameters:\\nKey: {key}\\nmetadata: {metadata[key]}\\ntemplate: {metadata_template[key]}.\"\n        else:\n            print(\"No temporary file found. Starting generation from scratch...\")\n            metadata = metadata_template\n\n        starting_id = len(metadata[\"multiviews\"])\n        print(f\"Starting generation from index {starting_id}/{size}...\")\n        if starting_id >= size:\n            print(\"Generation already done.\")\n            return\n\n        generator = MultiviewHabitatSimGenerator(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scene,\n            navmesh=navmesh,\n            views_count=views_count,\n            size=size,\n            **kwargs,\n        )\n\n        for idx in tqdm(range(starting_id, size)):\n            # Generate / re-generate the observations\n            try:\n                data = generator[idx]\n                observations = data[\"observations\"]\n                positions = data[\"positions\"]\n                orientations = data[\"orientations\"]\n\n                idx_label = f\"{idx:08}\"\n                for oidx, observation in enumerate(observations):\n                    observation_label = (\n                        f\"{oidx + 1}\"  # Leonid is indexing starting from 1\n                    )\n                    # Color image saved using PIL\n                    img = PIL.Image.fromarray(observation[\"color\"][:, :, :3])\n                    filename = os.path.join(\n                        output_dir, f\"{idx_label}_{observation_label}.jpeg\"\n                    )\n                    img.save(filename)\n                    if generate_depth:\n                        # Depth image as EXR file\n                        filename = os.path.join(\n                            output_dir, f\"{idx_label}_{observation_label}_depth.exr\"\n                        )\n                        cv2.imwrite(\n                            filename,\n                            observation[\"depth\"],\n                            [cv2.IMWRITE_EXR_TYPE, cv2.IMWRITE_EXR_TYPE_HALF],\n                        )\n                        # Camera parameters\n                        camera_params = dict(\n                            [\n                                (key, observation[key].tolist())\n                                for key in (\n                                    \"camera_intrinsics\",\n                                    \"R_cam2world\",\n                                    \"t_cam2world\",\n                                )\n                            ]\n                        )\n                        filename = os.path.join(\n                            output_dir,\n                            f\"{idx_label}_{observation_label}_camera_params.json\",\n                        )\n                        with open(filename, \"w\") as f:\n                            json.dump(camera_params, f)\n                metadata[\"multiviews\"][idx_label] = {\n                    \"positions\": positions.tolist(),\n                    \"orientations\": orientations.tolist(),\n                    \"covisibility_ratios\": data[\"covisibility_ratios\"].tolist(),\n                    \"valid_fractions\": data[\"valid_fractions\"].tolist(),\n                    \"pairwise_visibility_ratios\": data[\n                        \"pairwise_visibility_ratios\"\n                    ].tolist(),\n                }\n            except RecursionError:\n                print(\n                    \"Recursion error: unable to sample observations for this scene. We will stop there.\"\n                )\n                break\n\n            # Regularly save a temporary metadata file, in case we need to restart the generation\n            if idx % 10 == 0:\n                with open(metadata_filename, \"w\") as f:\n                    json.dump(metadata, f)\n\n        # Save metadata\n        with open(metadata_filename, \"w\") as f:\n            json.dump(metadata, f)\n\n        generator.close()\n    except NoNaviguableSpaceError:\n        pass\n\n\ndef create_commandline(scene_data, generate_depth, exist_ok=False):\n    \"\"\"\n    Create a commandline string to generate a scene.\n    \"\"\"\n\n    def my_formatting(val):\n        if val is None or val == \"\":\n            return '\"\"'\n        else:\n            return val\n\n    commandline = f\"\"\"python {__file__} --scene {my_formatting(scene_data.scene)} \n    --scene_dataset_config_file {my_formatting(scene_data.scene_dataset_config_file)} \n    --navmesh {my_formatting(scene_data.navmesh)} \n    --output_dir {my_formatting(scene_data.output_dir)} \n    --generate_depth {int(generate_depth)} \n    --exist_ok {int(exist_ok)}\n    \"\"\"\n    commandline = \" \".join(commandline.split())\n    return commandline\n\n\nif __name__ == \"__main__\":\n    os.umask(2)\n\n    parser = argparse.ArgumentParser(\n        description=\"\"\"Example of use -- listing commands to generate data for scenes available:\n    > python datasets/habitat_sim/generate_multiview_habitat_images.py --list_commands\n    \"\"\"\n    )\n\n    parser.add_argument(\"--output_dir\", type=str, required=True)\n    parser.add_argument(\n        \"--list_commands\", action=\"store_true\", help=\"list commandlines to run if true\"\n    )\n    parser.add_argument(\"--scene\", type=str, default=\"\")\n    parser.add_argument(\"--scene_dataset_config_file\", type=str, default=\"\")\n    parser.add_argument(\"--navmesh\", type=str, default=\"\")\n\n    parser.add_argument(\"--generate_depth\", type=int, default=1)\n    parser.add_argument(\"--exist_ok\", type=int, default=0)\n\n    kwargs = dict(resolution=(256, 256), hfov=60, views_count=2, size=1000)\n\n    args = parser.parse_args()\n    generate_depth = bool(args.generate_depth)\n    exist_ok = bool(args.exist_ok)\n\n    if args.list_commands:\n        # Listing scenes available...\n        scenes_data = list_scenes_available(base_output_dir=args.output_dir)\n\n        for scene_data in scenes_data:\n            print(\n                create_commandline(\n                    scene_data, generate_depth=generate_depth, exist_ok=exist_ok\n                )\n            )\n    else:\n        if args.scene == \"\" or args.output_dir == \"\":\n            print(\"Missing scene or output dir argument!\")\n            print(parser.format_help())\n        else:\n            generate_multiview_images_for_scene(\n                scene=args.scene,\n                scene_dataset_config_file=args.scene_dataset_config_file,\n                navmesh=args.navmesh,\n                output_dir=args.output_dir,\n                exist_ok=exist_ok,\n                generate_depth=generate_depth,\n                **kwargs,\n            )","source_hash":"f1f29c975905b469f700a147e5acbd9462535dcf6798a041df74d8f91d6498ea","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.generate_multiview_images.generate_multiview_images_for_scene","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.generate_multiview_images.generate_multiview_images_for_scene#L20-L157","kind":"function","name":"generate_multiview_images_for_scene","path":"src/croco/datasets/habitat_sim/generate_multiview_images.py","language":"python","start_line":20,"end_line":157,"context_start_line":1,"context_end_line":177,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nfrom tqdm import tqdm\nimport argparse\nimport PIL.Image\nimport numpy as np\nimport json\nfrom datasets.habitat_sim.multiview_habitat_sim_generator import (\n    MultiviewHabitatSimGenerator,\n    NoNaviguableSpaceError,\n)\nfrom datasets.habitat_sim.paths import list_scenes_available\nimport cv2\nimport quaternion\nimport shutil\n\n\ndef generate_multiview_images_for_scene(\n    scene_dataset_config_file,\n    scene,\n    navmesh,\n    output_dir,\n    views_count,\n    size,\n    exist_ok=False,\n    generate_depth=False,\n    **kwargs,\n):\n    \"\"\"\n    Generate tuples of overlapping views for a given scene.\n    generate_depth: generate depth images and camera parameters.\n    \"\"\"\n    if os.path.exists(output_dir) and not exist_ok:\n        print(f\"Scene {scene}: data already generated. Ignoring generation.\")\n        return\n    try:\n        print(f\"Scene {scene}: {size} multiview acquisitions to generate...\")\n        os.makedirs(output_dir, exist_ok=exist_ok)\n\n        metadata_filename = os.path.join(output_dir, \"metadata.json\")\n\n        metadata_template = dict(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scene,\n            navmesh=navmesh,\n            views_count=views_count,\n            size=size,\n            generate_depth=generate_depth,\n            **kwargs,\n        )\n        metadata_template[\"multiviews\"] = dict()\n\n        if os.path.exists(metadata_filename):\n            print(\"Metadata file already exists:\", metadata_filename)\n            print(\"Loading already generated metadata file...\")\n            with open(metadata_filename, \"r\") as f:\n                metadata = json.load(f)\n\n            for key in metadata_template.keys():\n                if key != \"multiviews\":\n                    assert (\n                        metadata_template[key] == metadata[key]\n                    ), f\"existing file is inconsistent with the input parameters:\\nKey: {key}\\nmetadata: {metadata[key]}\\ntemplate: {metadata_template[key]}.\"\n        else:\n            print(\"No temporary file found. Starting generation from scratch...\")\n            metadata = metadata_template\n\n        starting_id = len(metadata[\"multiviews\"])\n        print(f\"Starting generation from index {starting_id}/{size}...\")\n        if starting_id >= size:\n            print(\"Generation already done.\")\n            return\n\n        generator = MultiviewHabitatSimGenerator(\n            scene_dataset_config_file=scene_dataset_config_file,\n            scene=scene,\n            navmesh=navmesh,\n            views_count=views_count,\n            size=size,\n            **kwargs,\n        )\n\n        for idx in tqdm(range(starting_id, size)):\n            # Generate / re-generate the observations\n            try:\n                data = generator[idx]\n                observations = data[\"observations\"]\n                positions = data[\"positions\"]\n                orientations = data[\"orientations\"]\n\n                idx_label = f\"{idx:08}\"\n                for oidx, observation in enumerate(observations):\n                    observation_label = (\n                        f\"{oidx + 1}\"  # Leonid is indexing starting from 1\n                    )\n                    # Color image saved using PIL\n                    img = PIL.Image.fromarray(observation[\"color\"][:, :, :3])\n                    filename = os.path.join(\n                        output_dir, f\"{idx_label}_{observation_label}.jpeg\"\n                    )\n                    img.save(filename)\n                    if generate_depth:\n                        # Depth image as EXR file\n                        filename = os.path.join(\n                            output_dir, f\"{idx_label}_{observation_label}_depth.exr\"\n                        )\n                        cv2.imwrite(\n                            filename,\n                            observation[\"depth\"],\n                            [cv2.IMWRITE_EXR_TYPE, cv2.IMWRITE_EXR_TYPE_HALF],\n                        )\n                        # Camera parameters\n                        camera_params = dict(\n                            [\n                                (key, observation[key].tolist())\n                                for key in (\n                                    \"camera_intrinsics\",\n                                    \"R_cam2world\",\n                                    \"t_cam2world\",\n                                )\n                            ]\n                        )\n                        filename = os.path.join(\n                            output_dir,\n                            f\"{idx_label}_{observation_label}_camera_params.json\",\n                        )\n                        with open(filename, \"w\") as f:\n                            json.dump(camera_params, f)\n                metadata[\"multiviews\"][idx_label] = {\n                    \"positions\": positions.tolist(),\n                    \"orientations\": orientations.tolist(),\n                    \"covisibility_ratios\": data[\"covisibility_ratios\"].tolist(),\n                    \"valid_fractions\": data[\"valid_fractions\"].tolist(),\n                    \"pairwise_visibility_ratios\": data[\n                        \"pairwise_visibility_ratios\"\n                    ].tolist(),\n                }\n            except RecursionError:\n                print(\n                    \"Recursion error: unable to sample observations for this scene. We will stop there.\"\n                )\n                break\n\n            # Regularly save a temporary metadata file, in case we need to restart the generation\n            if idx % 10 == 0:\n                with open(metadata_filename, \"w\") as f:\n                    json.dump(metadata, f)\n\n        # Save metadata\n        with open(metadata_filename, \"w\") as f:\n            json.dump(metadata, f)\n\n        generator.close()\n    except NoNaviguableSpaceError:\n        pass\n\n\ndef create_commandline(scene_data, generate_depth, exist_ok=False):\n    \"\"\"\n    Create a commandline string to generate a scene.\n    \"\"\"\n\n    def my_formatting(val):\n        if val is None or val == \"\":\n            return '\"\"'\n        else:\n            return val\n\n    commandline = f\"\"\"python {__file__} --scene {my_formatting(scene_data.scene)} \n    --scene_dataset_config_file {my_formatting(scene_data.scene_dataset_config_file)} \n    --navmesh {my_formatting(scene_data.navmesh)} \n    --output_dir {my_formatting(scene_data.output_dir)} \n    --generate_depth {int(generate_depth)} \n    --exist_ok {int(exist_ok)}\n    \"\"\"","source_hash":"f1f29c975905b469f700a147e5acbd9462535dcf6798a041df74d8f91d6498ea","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.generate_multiview_images.create_commandline","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.generate_multiview_images.create_commandline#L160-L179","kind":"function","name":"create_commandline","path":"src/croco/datasets/habitat_sim/generate_multiview_images.py","language":"python","start_line":160,"end_line":179,"context_start_line":140,"context_end_line":199,"code":"            except RecursionError:\n                print(\n                    \"Recursion error: unable to sample observations for this scene. We will stop there.\"\n                )\n                break\n\n            # Regularly save a temporary metadata file, in case we need to restart the generation\n            if idx % 10 == 0:\n                with open(metadata_filename, \"w\") as f:\n                    json.dump(metadata, f)\n\n        # Save metadata\n        with open(metadata_filename, \"w\") as f:\n            json.dump(metadata, f)\n\n        generator.close()\n    except NoNaviguableSpaceError:\n        pass\n\n\ndef create_commandline(scene_data, generate_depth, exist_ok=False):\n    \"\"\"\n    Create a commandline string to generate a scene.\n    \"\"\"\n\n    def my_formatting(val):\n        if val is None or val == \"\":\n            return '\"\"'\n        else:\n            return val\n\n    commandline = f\"\"\"python {__file__} --scene {my_formatting(scene_data.scene)} \n    --scene_dataset_config_file {my_formatting(scene_data.scene_dataset_config_file)} \n    --navmesh {my_formatting(scene_data.navmesh)} \n    --output_dir {my_formatting(scene_data.output_dir)} \n    --generate_depth {int(generate_depth)} \n    --exist_ok {int(exist_ok)}\n    \"\"\"\n    commandline = \" \".join(commandline.split())\n    return commandline\n\n\nif __name__ == \"__main__\":\n    os.umask(2)\n\n    parser = argparse.ArgumentParser(\n        description=\"\"\"Example of use -- listing commands to generate data for scenes available:\n    > python datasets/habitat_sim/generate_multiview_habitat_images.py --list_commands\n    \"\"\"\n    )\n\n    parser.add_argument(\"--output_dir\", type=str, required=True)\n    parser.add_argument(\n        \"--list_commands\", action=\"store_true\", help=\"list commandlines to run if true\"\n    )\n    parser.add_argument(\"--scene\", type=str, default=\"\")\n    parser.add_argument(\"--scene_dataset_config_file\", type=str, default=\"\")\n    parser.add_argument(\"--navmesh\", type=str, default=\"\")\n\n    parser.add_argument(\"--generate_depth\", type=int, default=1)","source_hash":"f1f29c975905b469f700a147e5acbd9462535dcf6798a041df74d8f91d6498ea","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.generate_multiview_images.my_formatting","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.generate_multiview_images.my_formatting#L165-L169","kind":"function","name":"my_formatting","path":"src/croco/datasets/habitat_sim/generate_multiview_images.py","language":"python","start_line":165,"end_line":169,"context_start_line":145,"context_end_line":189,"code":"\n            # Regularly save a temporary metadata file, in case we need to restart the generation\n            if idx % 10 == 0:\n                with open(metadata_filename, \"w\") as f:\n                    json.dump(metadata, f)\n\n        # Save metadata\n        with open(metadata_filename, \"w\") as f:\n            json.dump(metadata, f)\n\n        generator.close()\n    except NoNaviguableSpaceError:\n        pass\n\n\ndef create_commandline(scene_data, generate_depth, exist_ok=False):\n    \"\"\"\n    Create a commandline string to generate a scene.\n    \"\"\"\n\n    def my_formatting(val):\n        if val is None or val == \"\":\n            return '\"\"'\n        else:\n            return val\n\n    commandline = f\"\"\"python {__file__} --scene {my_formatting(scene_data.scene)} \n    --scene_dataset_config_file {my_formatting(scene_data.scene_dataset_config_file)} \n    --navmesh {my_formatting(scene_data.navmesh)} \n    --output_dir {my_formatting(scene_data.output_dir)} \n    --generate_depth {int(generate_depth)} \n    --exist_ok {int(exist_ok)}\n    \"\"\"\n    commandline = \" \".join(commandline.split())\n    return commandline\n\n\nif __name__ == \"__main__\":\n    os.umask(2)\n\n    parser = argparse.ArgumentParser(\n        description=\"\"\"Example of use -- listing commands to generate data for scenes available:\n    > python datasets/habitat_sim/generate_multiview_habitat_images.py --list_commands\n    \"\"\"\n    )","source_hash":"f1f29c975905b469f700a147e5acbd9462535dcf6798a041df74d8f91d6498ea","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator","uri":"program://Human3R/module/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator#L1-L501","kind":"module","name":"src.croco.datasets.habitat_sim.multiview_habitat_sim_generator","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":1,"end_line":501,"context_start_line":1,"context_end_line":501,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nimport numpy as np\nimport quaternion\nimport habitat_sim\nimport json\nfrom sklearn.neighbors import NearestNeighbors\nimport cv2\n\n# OpenCV to habitat camera convention transformation\nR_OPENCV2HABITAT = np.stack(\n    (habitat_sim.geo.RIGHT, -habitat_sim.geo.UP, habitat_sim.geo.FRONT), axis=0\n)\nR_HABITAT2OPENCV = R_OPENCV2HABITAT.T\nDEG2RAD = np.pi / 180\n\n\ndef compute_camera_intrinsics(height, width, hfov):\n    f = width / 2 / np.tan(hfov / 2 * np.pi / 180)\n    cu, cv = width / 2, height / 2\n    return f, cu, cv\n\n\ndef compute_camera_pose_opencv_convention(camera_position, camera_orientation):\n    R_cam2world = quaternion.as_rotation_matrix(camera_orientation) @ R_OPENCV2HABITAT\n    t_cam2world = np.asarray(camera_position)\n    return R_cam2world, t_cam2world\n\n\ndef compute_pointmap(depthmap, hfov):\n    \"\"\"Compute a HxWx3 pointmap in camera frame from a HxW depth map.\"\"\"\n    height, width = depthmap.shape\n    f, cu, cv = compute_camera_intrinsics(height, width, hfov)\n    # Cast depth map to point\n    z_cam = depthmap\n    u, v = np.meshgrid(range(width), range(height))\n    x_cam = (u - cu) / f * z_cam\n    y_cam = (v - cv) / f * z_cam\n    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1)\n    return X_cam\n\n\ndef compute_pointcloud(depthmap, hfov, camera_position, camera_rotation):\n    \"\"\"Return a 3D point cloud corresponding to valid pixels of the depth map\"\"\"\n    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(\n        camera_position, camera_rotation\n    )\n\n    X_cam = compute_pointmap(depthmap=depthmap, hfov=hfov)\n    valid_mask = X_cam[:, :, 2] != 0.0\n\n    X_cam = X_cam.reshape(-1, 3)[valid_mask.flatten()]\n    X_world = X_cam @ R_cam2world.T + t_cam2world.reshape(1, 3)\n    return X_world\n\n\ndef compute_pointcloud_overlaps_scikit(\n    pointcloud1, pointcloud2, distance_threshold, compute_symmetric=False\n):\n    \"\"\"\n    Compute 'overlapping' metrics based on a distance threshold between two point clouds.\n    \"\"\"\n    nbrs = NearestNeighbors(n_neighbors=1, algorithm=\"kd_tree\").fit(pointcloud2)\n    distances, indices = nbrs.kneighbors(pointcloud1)\n    intersection1 = np.count_nonzero(distances.flatten() < distance_threshold)\n\n    data = {\"intersection1\": intersection1, \"size1\": len(pointcloud1)}\n    if compute_symmetric:\n        nbrs = NearestNeighbors(n_neighbors=1, algorithm=\"kd_tree\").fit(pointcloud1)\n        distances, indices = nbrs.kneighbors(pointcloud2)\n        intersection2 = np.count_nonzero(distances.flatten() < distance_threshold)\n        data[\"intersection2\"] = intersection2\n        data[\"size2\"] = len(pointcloud2)\n\n    return data\n\n\ndef _append_camera_parameters(observation, hfov, camera_location, camera_rotation):\n    \"\"\"\n    Add camera parameters to the observation dictionnary produced by Habitat-Sim\n    In-place modifications.\n    \"\"\"\n    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(\n        camera_location, camera_rotation\n    )\n    height, width = observation[\"depth\"].shape\n    f, cu, cv = compute_camera_intrinsics(height, width, hfov)\n    K = np.asarray([[f, 0, cu], [0, f, cv], [0, 0, 1.0]])\n    observation[\"camera_intrinsics\"] = K\n    observation[\"t_cam2world\"] = t_cam2world\n    observation[\"R_cam2world\"] = R_cam2world\n\n\ndef look_at(eye, center, up, return_cam2world=True):\n    \"\"\"\n    Return camera pose looking at a given center point.\n    Analogous of gluLookAt function, using OpenCV camera convention.\n    \"\"\"\n    z = center - eye\n    z /= np.linalg.norm(z, axis=-1, keepdims=True)\n    y = -up\n    y = y - np.sum(y * z, axis=-1, keepdims=True) * z\n    y /= np.linalg.norm(y, axis=-1, keepdims=True)\n    x = np.cross(y, z, axis=-1)\n\n    if return_cam2world:\n        R = np.stack((x, y, z), axis=-1)\n        t = eye\n    else:\n        # World to camera transformation\n        # Transposed matrix\n        R = np.stack((x, y, z), axis=-2)\n        t = -np.einsum(\"...ij, ...j\", R, eye)\n    return R, t\n\n\ndef look_at_for_habitat(eye, center, up, return_cam2world=True):\n    R, t = look_at(eye, center, up)\n    orientation = quaternion.from_rotation_matrix(R @ R_OPENCV2HABITAT.T)\n    return orientation, t\n\n\ndef generate_orientation_noise(pan_range, tilt_range, roll_range):\n    return (\n        quaternion.from_rotation_vector(\n            np.random.uniform(*pan_range) * DEG2RAD * habitat_sim.geo.UP\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*tilt_range) * DEG2RAD * habitat_sim.geo.RIGHT\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*roll_range) * DEG2RAD * habitat_sim.geo.FRONT\n        )\n    )\n\n\nclass NoNaviguableSpaceError(RuntimeError):\n    def __init__(self, *args):\n        super().__init__(*args)\n\n\nclass MultiviewHabitatSimGenerator:\n    def __init__(\n        self,\n        scene,\n        navmesh,\n        scene_dataset_config_file,\n        resolution=(240, 320),\n        views_count=2,\n        hfov=60,\n        gpu_id=0,\n        size=10000,\n        minimum_covisibility=0.5,\n        transform=None,\n    ):\n        self.scene = scene\n        self.navmesh = navmesh\n        self.scene_dataset_config_file = scene_dataset_config_file\n        self.resolution = resolution\n        self.views_count = views_count\n        assert self.views_count >= 1\n        self.hfov = hfov\n        self.gpu_id = gpu_id\n        self.size = size\n        self.transform = transform\n\n        # Noise added to camera orientation\n        self.pan_range = (-3, 3)\n        self.tilt_range = (-10, 10)\n        self.roll_range = (-5, 5)\n\n        # Height range to sample cameras\n        self.height_range = (1.2, 1.8)\n\n        # Random steps between the camera views\n        self.random_steps_count = 5\n        self.random_step_variance = 2.0\n\n        # Minimum fraction of the scene which should be valid (well defined depth)\n        self.minimum_valid_fraction = 0.7\n\n        # Distance threshold to see  to select pairs\n        self.distance_threshold = 0.05\n        # Minimum IoU of a view point cloud with respect to the reference view to be kept.\n        self.minimum_covisibility = minimum_covisibility\n\n        # Maximum number of retries.\n        self.max_attempts_count = 100\n\n        self.seed = None\n        self._lazy_initialization()\n\n    def _lazy_initialization(self):\n        # Lazy random seeding and instantiation of the simulator to deal with multiprocessing properly\n        if self.seed == None:\n            # Re-seed numpy generator\n            np.random.seed()\n            self.seed = np.random.randint(2**32 - 1)\n            sim_cfg = habitat_sim.SimulatorConfiguration()\n            sim_cfg.scene_id = self.scene\n            if (\n                self.scene_dataset_config_file is not None\n                and self.scene_dataset_config_file != \"\"\n            ):\n                sim_cfg.scene_dataset_config_file = self.scene_dataset_config_file\n            sim_cfg.random_seed = self.seed\n            sim_cfg.load_semantic_mesh = False\n            sim_cfg.gpu_device_id = self.gpu_id\n\n            depth_sensor_spec = habitat_sim.CameraSensorSpec()\n            depth_sensor_spec.uuid = \"depth\"\n            depth_sensor_spec.sensor_type = habitat_sim.SensorType.DEPTH\n            depth_sensor_spec.resolution = self.resolution\n            depth_sensor_spec.hfov = self.hfov\n            depth_sensor_spec.position = [0.0, 0.0, 0]\n            depth_sensor_spec.orientation\n\n            rgb_sensor_spec = habitat_sim.CameraSensorSpec()\n            rgb_sensor_spec.uuid = \"color\"\n            rgb_sensor_spec.sensor_type = habitat_sim.SensorType.COLOR\n            rgb_sensor_spec.resolution = self.resolution\n            rgb_sensor_spec.hfov = self.hfov\n            rgb_sensor_spec.position = [0.0, 0.0, 0]\n            agent_cfg = habitat_sim.agent.AgentConfiguration(\n                sensor_specifications=[rgb_sensor_spec, depth_sensor_spec]\n            )\n\n            cfg = habitat_sim.Configuration(sim_cfg, [agent_cfg])\n            self.sim = habitat_sim.Simulator(cfg)\n            if self.navmesh is not None and self.navmesh != \"\":\n                # Use pre-computed navmesh when available (usually better than those generated automatically)\n                self.sim.pathfinder.load_nav_mesh(self.navmesh)\n\n            if not self.sim.pathfinder.is_loaded:\n                # Try to compute a navmesh\n                navmesh_settings = habitat_sim.NavMeshSettings()\n                navmesh_settings.set_defaults()\n                self.sim.recompute_navmesh(self.sim.pathfinder, navmesh_settings, True)\n\n            # Ensure that the navmesh is not empty\n            if not self.sim.pathfinder.is_loaded:\n                raise NoNaviguableSpaceError(\n                    f\"No naviguable location (scene: {self.scene} -- navmesh: {self.navmesh})\"\n                )\n\n            self.agent = self.sim.initialize_agent(agent_id=0)\n\n    def close(self):\n        self.sim.close()\n\n    def __del__(self):\n        self.sim.close()\n\n    def __len__(self):\n        return self.size\n\n    def sample_random_viewpoint(self):\n        \"\"\"Sample a random viewpoint using the navmesh\"\"\"\n        nav_point = self.sim.pathfinder.get_random_navigable_point()\n\n        # Sample a random viewpoint height\n        viewpoint_height = np.random.uniform(*self.height_range)\n        viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP\n        viewpoint_orientation = quaternion.from_rotation_vector(\n            np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP\n        ) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)\n        return viewpoint_position, viewpoint_orientation, nav_point\n\n    def sample_other_random_viewpoint(self, observed_point, nav_point):\n        \"\"\"Sample a random viewpoint close to an existing one, using the navmesh and a reference observed point.\"\"\"\n        other_nav_point = nav_point\n\n        walk_directions = self.random_step_variance * np.asarray([1, 0, 1])\n        for i in range(self.random_steps_count):\n            temp = self.sim.pathfinder.snap_point(\n                other_nav_point + walk_directions * np.random.normal(size=3)\n            )\n            # Snapping may return nan when it fails\n            if not np.isnan(temp[0]):\n                other_nav_point = temp\n\n        other_viewpoint_height = np.random.uniform(*self.height_range)\n        other_viewpoint_position = (\n            other_nav_point + other_viewpoint_height * habitat_sim.geo.UP\n        )\n\n        # Set viewing direction towards the central point\n        rotation, position = look_at_for_habitat(\n            eye=other_viewpoint_position,\n            center=observed_point,\n            up=habitat_sim.geo.UP,\n            return_cam2world=True,\n        )\n        rotation = rotation * generate_orientation_noise(\n            self.pan_range, self.tilt_range, self.roll_range\n        )\n        return position, rotation, other_nav_point\n\n    def is_other_pointcloud_overlapping(self, ref_pointcloud, other_pointcloud):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        pixels_count = self.resolution[0] * self.resolution[1]\n        valid_fraction = len(other_pointcloud) / pixels_count\n        assert valid_fraction <= 1.0 and valid_fraction >= 0.0\n        overlap = compute_pointcloud_overlaps_scikit(\n            ref_pointcloud,\n            other_pointcloud,\n            self.distance_threshold,\n            compute_symmetric=True,\n        )\n        covisibility = min(\n            overlap[\"intersection1\"] / pixels_count,\n            overlap[\"intersection2\"] / pixels_count,\n        )\n        is_valid = (valid_fraction >= self.minimum_valid_fraction) and (\n            covisibility >= self.minimum_covisibility\n        )\n        return is_valid, valid_fraction, covisibility\n\n    def is_other_viewpoint_overlapping(\n        self, ref_pointcloud, observation, position, rotation\n    ):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        other_pointcloud = compute_pointcloud(\n            observation[\"depth\"], self.hfov, position, rotation\n        )\n        return self.is_other_pointcloud_overlapping(ref_pointcloud, other_pointcloud)\n\n    def render_viewpoint(self, viewpoint_position, viewpoint_orientation):\n        agent_state = habitat_sim.AgentState()\n        agent_state.position = viewpoint_position\n        agent_state.rotation = viewpoint_orientation\n        self.agent.set_state(agent_state)\n        viewpoint_observations = self.sim.get_sensor_observations(agent_ids=0)\n        _append_camera_parameters(\n            viewpoint_observations, self.hfov, viewpoint_position, viewpoint_orientation\n        )\n        return viewpoint_observations\n\n    def __getitem__(self, useless_idx):\n        ref_position, ref_orientation, nav_point = self.sample_random_viewpoint()\n        ref_observations = self.render_viewpoint(ref_position, ref_orientation)\n        # Extract point cloud\n        ref_pointcloud = compute_pointcloud(\n            depthmap=ref_observations[\"depth\"],\n            hfov=self.hfov,\n            camera_position=ref_position,\n            camera_rotation=ref_orientation,\n        )\n\n        pixels_count = self.resolution[0] * self.resolution[1]\n        ref_valid_fraction = len(ref_pointcloud) / pixels_count\n        assert ref_valid_fraction <= 1.0 and ref_valid_fraction >= 0.0\n        if ref_valid_fraction < self.minimum_valid_fraction:\n            # This should produce a recursion error at some point when something is very wrong.\n            return self[0]\n        # Pick an reference observed point in the point cloud\n        observed_point = np.mean(ref_pointcloud, axis=0)\n\n        # Add the first image as reference\n        viewpoints_observations = [ref_observations]\n        viewpoints_covisibility = [ref_valid_fraction]\n        viewpoints_positions = [ref_position]\n        viewpoints_orientations = [quaternion.as_float_array(ref_orientation)]\n        viewpoints_clouds = [ref_pointcloud]\n        viewpoints_valid_fractions = [ref_valid_fraction]\n\n        for _ in range(self.views_count - 1):\n            # Generate an other viewpoint using some dummy random walk\n            successful_sampling = False\n            for sampling_attempt in range(self.max_attempts_count):\n                position, rotation, _ = self.sample_other_random_viewpoint(\n                    observed_point, nav_point\n                )\n                # Observation\n                other_viewpoint_observations = self.render_viewpoint(position, rotation)\n                other_pointcloud = compute_pointcloud(\n                    other_viewpoint_observations[\"depth\"], self.hfov, position, rotation\n                )\n\n                is_valid, valid_fraction, covisibility = (\n                    self.is_other_pointcloud_overlapping(\n                        ref_pointcloud, other_pointcloud\n                    )\n                )\n                if is_valid:\n                    successful_sampling = True\n                    break\n            if not successful_sampling:\n                print(\"WARNING: Maximum number of attempts reached.\")\n                # Dirty hack, try using a novel original viewpoint\n                return self[0]\n            viewpoints_observations.append(other_viewpoint_observations)\n            viewpoints_covisibility.append(covisibility)\n            viewpoints_positions.append(position)\n            viewpoints_orientations.append(\n                quaternion.as_float_array(rotation)\n            )  # WXYZ convention for the quaternion encoding.\n            viewpoints_clouds.append(other_pointcloud)\n            viewpoints_valid_fractions.append(valid_fraction)\n\n        # Estimate relations between all pairs of images\n        pairwise_visibility_ratios = np.ones(\n            (len(viewpoints_observations), len(viewpoints_observations))\n        )\n        for i in range(len(viewpoints_observations)):\n            pairwise_visibility_ratios[i, i] = viewpoints_valid_fractions[i]\n            for j in range(i + 1, len(viewpoints_observations)):\n                overlap = compute_pointcloud_overlaps_scikit(\n                    viewpoints_clouds[i],\n                    viewpoints_clouds[j],\n                    self.distance_threshold,\n                    compute_symmetric=True,\n                )\n                pairwise_visibility_ratios[i, j] = (\n                    overlap[\"intersection1\"] / pixels_count\n                )\n                pairwise_visibility_ratios[j, i] = (\n                    overlap[\"intersection2\"] / pixels_count\n                )\n\n        # IoU is relative to the image 0\n        data = {\n            \"observations\": viewpoints_observations,\n            \"positions\": np.asarray(viewpoints_positions),\n            \"orientations\": np.asarray(viewpoints_orientations),\n            \"covisibility_ratios\": np.asarray(viewpoints_covisibility),\n            \"valid_fractions\": np.asarray(viewpoints_valid_fractions, dtype=float),\n            \"pairwise_visibility_ratios\": np.asarray(\n                pairwise_visibility_ratios, dtype=float\n            ),\n        }\n\n        if self.transform is not None:\n            data = self.transform(data)\n        return data\n\n    def generate_random_spiral_trajectory(\n        self,\n        images_count=100,\n        max_radius=0.5,\n        half_turns=5,\n        use_constant_orientation=False,\n    ):\n        \"\"\"\n        Return a list of images corresponding to a spiral trajectory from a random starting point.\n        Useful to generate nice visualisations.\n        Use an even number of half turns to get a nice \"C1-continuous\" loop effect\n        \"\"\"\n        ref_position, ref_orientation, navpoint = self.sample_random_viewpoint()\n        ref_observations = self.render_viewpoint(ref_position, ref_orientation)\n        ref_pointcloud = compute_pointcloud(\n            depthmap=ref_observations[\"depth\"],\n            hfov=self.hfov,\n            camera_position=ref_position,\n            camera_rotation=ref_orientation,\n        )\n        pixels_count = self.resolution[0] * self.resolution[1]\n        if len(ref_pointcloud) / pixels_count < self.minimum_valid_fraction:\n            # Dirty hack: ensure that the valid part of the image is significant\n            return self.generate_random_spiral_trajectory(\n                images_count, max_radius, half_turns, use_constant_orientation\n            )\n\n        # Pick an observed point in the point cloud\n        observed_point = np.mean(ref_pointcloud, axis=0)\n        ref_R, ref_t = compute_camera_pose_opencv_convention(\n            ref_position, ref_orientation\n        )\n\n        images = []\n        is_valid = []\n        # Spiral trajectory, use_constant orientation\n        for i, alpha in enumerate(np.linspace(0, 1, images_count)):\n            r = max_radius * np.abs(\n                np.sin(alpha * np.pi)\n            )  # Increase then decrease the radius\n            theta = alpha * half_turns * np.pi\n            x = r * np.cos(theta)\n            y = r * np.sin(theta)\n            z = 0.0\n            position = (\n                ref_position + (ref_R @ np.asarray([x, y, z]).reshape(3, 1)).flatten()\n            )\n            if use_constant_orientation:\n                orientation = ref_orientation\n            else:\n                # trajectory looking at a mean point in front of the ref observation\n                orientation, position = look_at_for_habitat(\n                    eye=position, center=observed_point, up=habitat_sim.geo.UP\n                )\n            observations = self.render_viewpoint(position, orientation)\n            images.append(observations[\"color\"][..., :3])\n            _is_valid, valid_fraction, iou = self.is_other_viewpoint_overlapping(\n                ref_pointcloud, observations, position, orientation\n            )\n            is_valid.append(_is_valid)\n        ret\n# ... truncated ...","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_camera_intrinsics","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_camera_intrinsics#L20-L23","kind":"function","name":"compute_camera_intrinsics","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":20,"end_line":23,"context_start_line":1,"context_end_line":43,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nimport numpy as np\nimport quaternion\nimport habitat_sim\nimport json\nfrom sklearn.neighbors import NearestNeighbors\nimport cv2\n\n# OpenCV to habitat camera convention transformation\nR_OPENCV2HABITAT = np.stack(\n    (habitat_sim.geo.RIGHT, -habitat_sim.geo.UP, habitat_sim.geo.FRONT), axis=0\n)\nR_HABITAT2OPENCV = R_OPENCV2HABITAT.T\nDEG2RAD = np.pi / 180\n\n\ndef compute_camera_intrinsics(height, width, hfov):\n    f = width / 2 / np.tan(hfov / 2 * np.pi / 180)\n    cu, cv = width / 2, height / 2\n    return f, cu, cv\n\n\ndef compute_camera_pose_opencv_convention(camera_position, camera_orientation):\n    R_cam2world = quaternion.as_rotation_matrix(camera_orientation) @ R_OPENCV2HABITAT\n    t_cam2world = np.asarray(camera_position)\n    return R_cam2world, t_cam2world\n\n\ndef compute_pointmap(depthmap, hfov):\n    \"\"\"Compute a HxWx3 pointmap in camera frame from a HxW depth map.\"\"\"\n    height, width = depthmap.shape\n    f, cu, cv = compute_camera_intrinsics(height, width, hfov)\n    # Cast depth map to point\n    z_cam = depthmap\n    u, v = np.meshgrid(range(width), range(height))\n    x_cam = (u - cu) / f * z_cam\n    y_cam = (v - cv) / f * z_cam\n    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1)\n    return X_cam\n","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_camera_pose_opencv_convention","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_camera_pose_opencv_convention#L26-L29","kind":"function","name":"compute_camera_pose_opencv_convention","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":26,"end_line":29,"context_start_line":6,"context_end_line":49,"code":"import quaternion\nimport habitat_sim\nimport json\nfrom sklearn.neighbors import NearestNeighbors\nimport cv2\n\n# OpenCV to habitat camera convention transformation\nR_OPENCV2HABITAT = np.stack(\n    (habitat_sim.geo.RIGHT, -habitat_sim.geo.UP, habitat_sim.geo.FRONT), axis=0\n)\nR_HABITAT2OPENCV = R_OPENCV2HABITAT.T\nDEG2RAD = np.pi / 180\n\n\ndef compute_camera_intrinsics(height, width, hfov):\n    f = width / 2 / np.tan(hfov / 2 * np.pi / 180)\n    cu, cv = width / 2, height / 2\n    return f, cu, cv\n\n\ndef compute_camera_pose_opencv_convention(camera_position, camera_orientation):\n    R_cam2world = quaternion.as_rotation_matrix(camera_orientation) @ R_OPENCV2HABITAT\n    t_cam2world = np.asarray(camera_position)\n    return R_cam2world, t_cam2world\n\n\ndef compute_pointmap(depthmap, hfov):\n    \"\"\"Compute a HxWx3 pointmap in camera frame from a HxW depth map.\"\"\"\n    height, width = depthmap.shape\n    f, cu, cv = compute_camera_intrinsics(height, width, hfov)\n    # Cast depth map to point\n    z_cam = depthmap\n    u, v = np.meshgrid(range(width), range(height))\n    x_cam = (u - cu) / f * z_cam\n    y_cam = (v - cv) / f * z_cam\n    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1)\n    return X_cam\n\n\ndef compute_pointcloud(depthmap, hfov, camera_position, camera_rotation):\n    \"\"\"Return a 3D point cloud corresponding to valid pixels of the depth map\"\"\"\n    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(\n        camera_position, camera_rotation\n    )","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_pointmap","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_pointmap#L32-L42","kind":"function","name":"compute_pointmap","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":32,"end_line":42,"context_start_line":12,"context_end_line":62,"code":"# OpenCV to habitat camera convention transformation\nR_OPENCV2HABITAT = np.stack(\n    (habitat_sim.geo.RIGHT, -habitat_sim.geo.UP, habitat_sim.geo.FRONT), axis=0\n)\nR_HABITAT2OPENCV = R_OPENCV2HABITAT.T\nDEG2RAD = np.pi / 180\n\n\ndef compute_camera_intrinsics(height, width, hfov):\n    f = width / 2 / np.tan(hfov / 2 * np.pi / 180)\n    cu, cv = width / 2, height / 2\n    return f, cu, cv\n\n\ndef compute_camera_pose_opencv_convention(camera_position, camera_orientation):\n    R_cam2world = quaternion.as_rotation_matrix(camera_orientation) @ R_OPENCV2HABITAT\n    t_cam2world = np.asarray(camera_position)\n    return R_cam2world, t_cam2world\n\n\ndef compute_pointmap(depthmap, hfov):\n    \"\"\"Compute a HxWx3 pointmap in camera frame from a HxW depth map.\"\"\"\n    height, width = depthmap.shape\n    f, cu, cv = compute_camera_intrinsics(height, width, hfov)\n    # Cast depth map to point\n    z_cam = depthmap\n    u, v = np.meshgrid(range(width), range(height))\n    x_cam = (u - cu) / f * z_cam\n    y_cam = (v - cv) / f * z_cam\n    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1)\n    return X_cam\n\n\ndef compute_pointcloud(depthmap, hfov, camera_position, camera_rotation):\n    \"\"\"Return a 3D point cloud corresponding to valid pixels of the depth map\"\"\"\n    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(\n        camera_position, camera_rotation\n    )\n\n    X_cam = compute_pointmap(depthmap=depthmap, hfov=hfov)\n    valid_mask = X_cam[:, :, 2] != 0.0\n\n    X_cam = X_cam.reshape(-1, 3)[valid_mask.flatten()]\n    X_world = X_cam @ R_cam2world.T + t_cam2world.reshape(1, 3)\n    return X_world\n\n\ndef compute_pointcloud_overlaps_scikit(\n    pointcloud1, pointcloud2, distance_threshold, compute_symmetric=False\n):\n    \"\"\"","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_pointcloud","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_pointcloud#L45-L56","kind":"function","name":"compute_pointcloud","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":45,"end_line":56,"context_start_line":25,"context_end_line":76,"code":"\ndef compute_camera_pose_opencv_convention(camera_position, camera_orientation):\n    R_cam2world = quaternion.as_rotation_matrix(camera_orientation) @ R_OPENCV2HABITAT\n    t_cam2world = np.asarray(camera_position)\n    return R_cam2world, t_cam2world\n\n\ndef compute_pointmap(depthmap, hfov):\n    \"\"\"Compute a HxWx3 pointmap in camera frame from a HxW depth map.\"\"\"\n    height, width = depthmap.shape\n    f, cu, cv = compute_camera_intrinsics(height, width, hfov)\n    # Cast depth map to point\n    z_cam = depthmap\n    u, v = np.meshgrid(range(width), range(height))\n    x_cam = (u - cu) / f * z_cam\n    y_cam = (v - cv) / f * z_cam\n    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1)\n    return X_cam\n\n\ndef compute_pointcloud(depthmap, hfov, camera_position, camera_rotation):\n    \"\"\"Return a 3D point cloud corresponding to valid pixels of the depth map\"\"\"\n    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(\n        camera_position, camera_rotation\n    )\n\n    X_cam = compute_pointmap(depthmap=depthmap, hfov=hfov)\n    valid_mask = X_cam[:, :, 2] != 0.0\n\n    X_cam = X_cam.reshape(-1, 3)[valid_mask.flatten()]\n    X_world = X_cam @ R_cam2world.T + t_cam2world.reshape(1, 3)\n    return X_world\n\n\ndef compute_pointcloud_overlaps_scikit(\n    pointcloud1, pointcloud2, distance_threshold, compute_symmetric=False\n):\n    \"\"\"\n    Compute 'overlapping' metrics based on a distance threshold between two point clouds.\n    \"\"\"\n    nbrs = NearestNeighbors(n_neighbors=1, algorithm=\"kd_tree\").fit(pointcloud2)\n    distances, indices = nbrs.kneighbors(pointcloud1)\n    intersection1 = np.count_nonzero(distances.flatten() < distance_threshold)\n\n    data = {\"intersection1\": intersection1, \"size1\": len(pointcloud1)}\n    if compute_symmetric:\n        nbrs = NearestNeighbors(n_neighbors=1, algorithm=\"kd_tree\").fit(pointcloud1)\n        distances, indices = nbrs.kneighbors(pointcloud2)\n        intersection2 = np.count_nonzero(distances.flatten() < distance_threshold)\n        data[\"intersection2\"] = intersection2\n        data[\"size2\"] = len(pointcloud2)\n","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_pointcloud_overlaps_scikit","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.compute_pointcloud_overlaps_scikit#L59-L77","kind":"function","name":"compute_pointcloud_overlaps_scikit","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":59,"end_line":77,"context_start_line":39,"context_end_line":97,"code":"    x_cam = (u - cu) / f * z_cam\n    y_cam = (v - cv) / f * z_cam\n    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1)\n    return X_cam\n\n\ndef compute_pointcloud(depthmap, hfov, camera_position, camera_rotation):\n    \"\"\"Return a 3D point cloud corresponding to valid pixels of the depth map\"\"\"\n    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(\n        camera_position, camera_rotation\n    )\n\n    X_cam = compute_pointmap(depthmap=depthmap, hfov=hfov)\n    valid_mask = X_cam[:, :, 2] != 0.0\n\n    X_cam = X_cam.reshape(-1, 3)[valid_mask.flatten()]\n    X_world = X_cam @ R_cam2world.T + t_cam2world.reshape(1, 3)\n    return X_world\n\n\ndef compute_pointcloud_overlaps_scikit(\n    pointcloud1, pointcloud2, distance_threshold, compute_symmetric=False\n):\n    \"\"\"\n    Compute 'overlapping' metrics based on a distance threshold between two point clouds.\n    \"\"\"\n    nbrs = NearestNeighbors(n_neighbors=1, algorithm=\"kd_tree\").fit(pointcloud2)\n    distances, indices = nbrs.kneighbors(pointcloud1)\n    intersection1 = np.count_nonzero(distances.flatten() < distance_threshold)\n\n    data = {\"intersection1\": intersection1, \"size1\": len(pointcloud1)}\n    if compute_symmetric:\n        nbrs = NearestNeighbors(n_neighbors=1, algorithm=\"kd_tree\").fit(pointcloud1)\n        distances, indices = nbrs.kneighbors(pointcloud2)\n        intersection2 = np.count_nonzero(distances.flatten() < distance_threshold)\n        data[\"intersection2\"] = intersection2\n        data[\"size2\"] = len(pointcloud2)\n\n    return data\n\n\ndef _append_camera_parameters(observation, hfov, camera_location, camera_rotation):\n    \"\"\"\n    Add camera parameters to the observation dictionnary produced by Habitat-Sim\n    In-place modifications.\n    \"\"\"\n    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(\n        camera_location, camera_rotation\n    )\n    height, width = observation[\"depth\"].shape\n    f, cu, cv = compute_camera_intrinsics(height, width, hfov)\n    K = np.asarray([[f, 0, cu], [0, f, cv], [0, 0, 1.0]])\n    observation[\"camera_intrinsics\"] = K\n    observation[\"t_cam2world\"] = t_cam2world\n    observation[\"R_cam2world\"] = R_cam2world\n\n\ndef look_at(eye, center, up, return_cam2world=True):\n    \"\"\"","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator._append_camera_parameters","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator._append_camera_parameters#L80-L93","kind":"function","name":"_append_camera_parameters","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":80,"end_line":93,"context_start_line":60,"context_end_line":113,"code":"    pointcloud1, pointcloud2, distance_threshold, compute_symmetric=False\n):\n    \"\"\"\n    Compute 'overlapping' metrics based on a distance threshold between two point clouds.\n    \"\"\"\n    nbrs = NearestNeighbors(n_neighbors=1, algorithm=\"kd_tree\").fit(pointcloud2)\n    distances, indices = nbrs.kneighbors(pointcloud1)\n    intersection1 = np.count_nonzero(distances.flatten() < distance_threshold)\n\n    data = {\"intersection1\": intersection1, \"size1\": len(pointcloud1)}\n    if compute_symmetric:\n        nbrs = NearestNeighbors(n_neighbors=1, algorithm=\"kd_tree\").fit(pointcloud1)\n        distances, indices = nbrs.kneighbors(pointcloud2)\n        intersection2 = np.count_nonzero(distances.flatten() < distance_threshold)\n        data[\"intersection2\"] = intersection2\n        data[\"size2\"] = len(pointcloud2)\n\n    return data\n\n\ndef _append_camera_parameters(observation, hfov, camera_location, camera_rotation):\n    \"\"\"\n    Add camera parameters to the observation dictionnary produced by Habitat-Sim\n    In-place modifications.\n    \"\"\"\n    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(\n        camera_location, camera_rotation\n    )\n    height, width = observation[\"depth\"].shape\n    f, cu, cv = compute_camera_intrinsics(height, width, hfov)\n    K = np.asarray([[f, 0, cu], [0, f, cv], [0, 0, 1.0]])\n    observation[\"camera_intrinsics\"] = K\n    observation[\"t_cam2world\"] = t_cam2world\n    observation[\"R_cam2world\"] = R_cam2world\n\n\ndef look_at(eye, center, up, return_cam2world=True):\n    \"\"\"\n    Return camera pose looking at a given center point.\n    Analogous of gluLookAt function, using OpenCV camera convention.\n    \"\"\"\n    z = center - eye\n    z /= np.linalg.norm(z, axis=-1, keepdims=True)\n    y = -up\n    y = y - np.sum(y * z, axis=-1, keepdims=True) * z\n    y /= np.linalg.norm(y, axis=-1, keepdims=True)\n    x = np.cross(y, z, axis=-1)\n\n    if return_cam2world:\n        R = np.stack((x, y, z), axis=-1)\n        t = eye\n    else:\n        # World to camera transformation\n        # Transposed matrix","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.look_at","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.look_at#L96-L116","kind":"function","name":"look_at","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":96,"end_line":116,"context_start_line":76,"context_end_line":136,"code":"\n    return data\n\n\ndef _append_camera_parameters(observation, hfov, camera_location, camera_rotation):\n    \"\"\"\n    Add camera parameters to the observation dictionnary produced by Habitat-Sim\n    In-place modifications.\n    \"\"\"\n    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(\n        camera_location, camera_rotation\n    )\n    height, width = observation[\"depth\"].shape\n    f, cu, cv = compute_camera_intrinsics(height, width, hfov)\n    K = np.asarray([[f, 0, cu], [0, f, cv], [0, 0, 1.0]])\n    observation[\"camera_intrinsics\"] = K\n    observation[\"t_cam2world\"] = t_cam2world\n    observation[\"R_cam2world\"] = R_cam2world\n\n\ndef look_at(eye, center, up, return_cam2world=True):\n    \"\"\"\n    Return camera pose looking at a given center point.\n    Analogous of gluLookAt function, using OpenCV camera convention.\n    \"\"\"\n    z = center - eye\n    z /= np.linalg.norm(z, axis=-1, keepdims=True)\n    y = -up\n    y = y - np.sum(y * z, axis=-1, keepdims=True) * z\n    y /= np.linalg.norm(y, axis=-1, keepdims=True)\n    x = np.cross(y, z, axis=-1)\n\n    if return_cam2world:\n        R = np.stack((x, y, z), axis=-1)\n        t = eye\n    else:\n        # World to camera transformation\n        # Transposed matrix\n        R = np.stack((x, y, z), axis=-2)\n        t = -np.einsum(\"...ij, ...j\", R, eye)\n    return R, t\n\n\ndef look_at_for_habitat(eye, center, up, return_cam2world=True):\n    R, t = look_at(eye, center, up)\n    orientation = quaternion.from_rotation_matrix(R @ R_OPENCV2HABITAT.T)\n    return orientation, t\n\n\ndef generate_orientation_noise(pan_range, tilt_range, roll_range):\n    return (\n        quaternion.from_rotation_vector(\n            np.random.uniform(*pan_range) * DEG2RAD * habitat_sim.geo.UP\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*tilt_range) * DEG2RAD * habitat_sim.geo.RIGHT\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*roll_range) * DEG2RAD * habitat_sim.geo.FRONT\n        )\n    )","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.look_at_for_habitat","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.look_at_for_habitat#L119-L122","kind":"function","name":"look_at_for_habitat","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":119,"end_line":122,"context_start_line":99,"context_end_line":142,"code":"    Analogous of gluLookAt function, using OpenCV camera convention.\n    \"\"\"\n    z = center - eye\n    z /= np.linalg.norm(z, axis=-1, keepdims=True)\n    y = -up\n    y = y - np.sum(y * z, axis=-1, keepdims=True) * z\n    y /= np.linalg.norm(y, axis=-1, keepdims=True)\n    x = np.cross(y, z, axis=-1)\n\n    if return_cam2world:\n        R = np.stack((x, y, z), axis=-1)\n        t = eye\n    else:\n        # World to camera transformation\n        # Transposed matrix\n        R = np.stack((x, y, z), axis=-2)\n        t = -np.einsum(\"...ij, ...j\", R, eye)\n    return R, t\n\n\ndef look_at_for_habitat(eye, center, up, return_cam2world=True):\n    R, t = look_at(eye, center, up)\n    orientation = quaternion.from_rotation_matrix(R @ R_OPENCV2HABITAT.T)\n    return orientation, t\n\n\ndef generate_orientation_noise(pan_range, tilt_range, roll_range):\n    return (\n        quaternion.from_rotation_vector(\n            np.random.uniform(*pan_range) * DEG2RAD * habitat_sim.geo.UP\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*tilt_range) * DEG2RAD * habitat_sim.geo.RIGHT\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*roll_range) * DEG2RAD * habitat_sim.geo.FRONT\n        )\n    )\n\n\nclass NoNaviguableSpaceError(RuntimeError):\n    def __init__(self, *args):\n        super().__init__(*args)\n","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.generate_orientation_noise","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.generate_orientation_noise#L125-L136","kind":"function","name":"generate_orientation_noise","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":125,"end_line":136,"context_start_line":105,"context_end_line":156,"code":"    y /= np.linalg.norm(y, axis=-1, keepdims=True)\n    x = np.cross(y, z, axis=-1)\n\n    if return_cam2world:\n        R = np.stack((x, y, z), axis=-1)\n        t = eye\n    else:\n        # World to camera transformation\n        # Transposed matrix\n        R = np.stack((x, y, z), axis=-2)\n        t = -np.einsum(\"...ij, ...j\", R, eye)\n    return R, t\n\n\ndef look_at_for_habitat(eye, center, up, return_cam2world=True):\n    R, t = look_at(eye, center, up)\n    orientation = quaternion.from_rotation_matrix(R @ R_OPENCV2HABITAT.T)\n    return orientation, t\n\n\ndef generate_orientation_noise(pan_range, tilt_range, roll_range):\n    return (\n        quaternion.from_rotation_vector(\n            np.random.uniform(*pan_range) * DEG2RAD * habitat_sim.geo.UP\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*tilt_range) * DEG2RAD * habitat_sim.geo.RIGHT\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*roll_range) * DEG2RAD * habitat_sim.geo.FRONT\n        )\n    )\n\n\nclass NoNaviguableSpaceError(RuntimeError):\n    def __init__(self, *args):\n        super().__init__(*args)\n\n\nclass MultiviewHabitatSimGenerator:\n    def __init__(\n        self,\n        scene,\n        navmesh,\n        scene_dataset_config_file,\n        resolution=(240, 320),\n        views_count=2,\n        hfov=60,\n        gpu_id=0,\n        size=10000,\n        minimum_covisibility=0.5,\n        transform=None,","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.NoNaviguableSpaceError","uri":"program://Human3R/class/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.NoNaviguableSpaceError#L139-L141","kind":"class","name":"NoNaviguableSpaceError","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":139,"end_line":141,"context_start_line":119,"context_end_line":161,"code":"def look_at_for_habitat(eye, center, up, return_cam2world=True):\n    R, t = look_at(eye, center, up)\n    orientation = quaternion.from_rotation_matrix(R @ R_OPENCV2HABITAT.T)\n    return orientation, t\n\n\ndef generate_orientation_noise(pan_range, tilt_range, roll_range):\n    return (\n        quaternion.from_rotation_vector(\n            np.random.uniform(*pan_range) * DEG2RAD * habitat_sim.geo.UP\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*tilt_range) * DEG2RAD * habitat_sim.geo.RIGHT\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*roll_range) * DEG2RAD * habitat_sim.geo.FRONT\n        )\n    )\n\n\nclass NoNaviguableSpaceError(RuntimeError):\n    def __init__(self, *args):\n        super().__init__(*args)\n\n\nclass MultiviewHabitatSimGenerator:\n    def __init__(\n        self,\n        scene,\n        navmesh,\n        scene_dataset_config_file,\n        resolution=(240, 320),\n        views_count=2,\n        hfov=60,\n        gpu_id=0,\n        size=10000,\n        minimum_covisibility=0.5,\n        transform=None,\n    ):\n        self.scene = scene\n        self.navmesh = navmesh\n        self.scene_dataset_config_file = scene_dataset_config_file\n        self.resolution = resolution","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.MultiviewHabitatSimGenerator","uri":"program://Human3R/class/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.MultiviewHabitatSimGenerator#L144-L501","kind":"class","name":"MultiviewHabitatSimGenerator","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":144,"end_line":501,"context_start_line":124,"context_end_line":501,"code":"\ndef generate_orientation_noise(pan_range, tilt_range, roll_range):\n    return (\n        quaternion.from_rotation_vector(\n            np.random.uniform(*pan_range) * DEG2RAD * habitat_sim.geo.UP\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*tilt_range) * DEG2RAD * habitat_sim.geo.RIGHT\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*roll_range) * DEG2RAD * habitat_sim.geo.FRONT\n        )\n    )\n\n\nclass NoNaviguableSpaceError(RuntimeError):\n    def __init__(self, *args):\n        super().__init__(*args)\n\n\nclass MultiviewHabitatSimGenerator:\n    def __init__(\n        self,\n        scene,\n        navmesh,\n        scene_dataset_config_file,\n        resolution=(240, 320),\n        views_count=2,\n        hfov=60,\n        gpu_id=0,\n        size=10000,\n        minimum_covisibility=0.5,\n        transform=None,\n    ):\n        self.scene = scene\n        self.navmesh = navmesh\n        self.scene_dataset_config_file = scene_dataset_config_file\n        self.resolution = resolution\n        self.views_count = views_count\n        assert self.views_count >= 1\n        self.hfov = hfov\n        self.gpu_id = gpu_id\n        self.size = size\n        self.transform = transform\n\n        # Noise added to camera orientation\n        self.pan_range = (-3, 3)\n        self.tilt_range = (-10, 10)\n        self.roll_range = (-5, 5)\n\n        # Height range to sample cameras\n        self.height_range = (1.2, 1.8)\n\n        # Random steps between the camera views\n        self.random_steps_count = 5\n        self.random_step_variance = 2.0\n\n        # Minimum fraction of the scene which should be valid (well defined depth)\n        self.minimum_valid_fraction = 0.7\n\n        # Distance threshold to see  to select pairs\n        self.distance_threshold = 0.05\n        # Minimum IoU of a view point cloud with respect to the reference view to be kept.\n        self.minimum_covisibility = minimum_covisibility\n\n        # Maximum number of retries.\n        self.max_attempts_count = 100\n\n        self.seed = None\n        self._lazy_initialization()\n\n    def _lazy_initialization(self):\n        # Lazy random seeding and instantiation of the simulator to deal with multiprocessing properly\n        if self.seed == None:\n            # Re-seed numpy generator\n            np.random.seed()\n            self.seed = np.random.randint(2**32 - 1)\n            sim_cfg = habitat_sim.SimulatorConfiguration()\n            sim_cfg.scene_id = self.scene\n            if (\n                self.scene_dataset_config_file is not None\n                and self.scene_dataset_config_file != \"\"\n            ):\n                sim_cfg.scene_dataset_config_file = self.scene_dataset_config_file\n            sim_cfg.random_seed = self.seed\n            sim_cfg.load_semantic_mesh = False\n            sim_cfg.gpu_device_id = self.gpu_id\n\n            depth_sensor_spec = habitat_sim.CameraSensorSpec()\n            depth_sensor_spec.uuid = \"depth\"\n            depth_sensor_spec.sensor_type = habitat_sim.SensorType.DEPTH\n            depth_sensor_spec.resolution = self.resolution\n            depth_sensor_spec.hfov = self.hfov\n            depth_sensor_spec.position = [0.0, 0.0, 0]\n            depth_sensor_spec.orientation\n\n            rgb_sensor_spec = habitat_sim.CameraSensorSpec()\n            rgb_sensor_spec.uuid = \"color\"\n            rgb_sensor_spec.sensor_type = habitat_sim.SensorType.COLOR\n            rgb_sensor_spec.resolution = self.resolution\n            rgb_sensor_spec.hfov = self.hfov\n            rgb_sensor_spec.position = [0.0, 0.0, 0]\n            agent_cfg = habitat_sim.agent.AgentConfiguration(\n                sensor_specifications=[rgb_sensor_spec, depth_sensor_spec]\n            )\n\n            cfg = habitat_sim.Configuration(sim_cfg, [agent_cfg])\n            self.sim = habitat_sim.Simulator(cfg)\n            if self.navmesh is not None and self.navmesh != \"\":\n                # Use pre-computed navmesh when available (usually better than those generated automatically)\n                self.sim.pathfinder.load_nav_mesh(self.navmesh)\n\n            if not self.sim.pathfinder.is_loaded:\n                # Try to compute a navmesh\n                navmesh_settings = habitat_sim.NavMeshSettings()\n                navmesh_settings.set_defaults()\n                self.sim.recompute_navmesh(self.sim.pathfinder, navmesh_settings, True)\n\n            # Ensure that the navmesh is not empty\n            if not self.sim.pathfinder.is_loaded:\n                raise NoNaviguableSpaceError(\n                    f\"No naviguable location (scene: {self.scene} -- navmesh: {self.navmesh})\"\n                )\n\n            self.agent = self.sim.initialize_agent(agent_id=0)\n\n    def close(self):\n        self.sim.close()\n\n    def __del__(self):\n        self.sim.close()\n\n    def __len__(self):\n        return self.size\n\n    def sample_random_viewpoint(self):\n        \"\"\"Sample a random viewpoint using the navmesh\"\"\"\n        nav_point = self.sim.pathfinder.get_random_navigable_point()\n\n        # Sample a random viewpoint height\n        viewpoint_height = np.random.uniform(*self.height_range)\n        viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP\n        viewpoint_orientation = quaternion.from_rotation_vector(\n            np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP\n        ) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)\n        return viewpoint_position, viewpoint_orientation, nav_point\n\n    def sample_other_random_viewpoint(self, observed_point, nav_point):\n        \"\"\"Sample a random viewpoint close to an existing one, using the navmesh and a reference observed point.\"\"\"\n        other_nav_point = nav_point\n\n        walk_directions = self.random_step_variance * np.asarray([1, 0, 1])\n        for i in range(self.random_steps_count):\n            temp = self.sim.pathfinder.snap_point(\n                other_nav_point + walk_directions * np.random.normal(size=3)\n            )\n            # Snapping may return nan when it fails\n            if not np.isnan(temp[0]):\n                other_nav_point = temp\n\n        other_viewpoint_height = np.random.uniform(*self.height_range)\n        other_viewpoint_position = (\n            other_nav_point + other_viewpoint_height * habitat_sim.geo.UP\n        )\n\n        # Set viewing direction towards the central point\n        rotation, position = look_at_for_habitat(\n            eye=other_viewpoint_position,\n            center=observed_point,\n            up=habitat_sim.geo.UP,\n            return_cam2world=True,\n        )\n        rotation = rotation * generate_orientation_noise(\n            self.pan_range, self.tilt_range, self.roll_range\n        )\n        return position, rotation, other_nav_point\n\n    def is_other_pointcloud_overlapping(self, ref_pointcloud, other_pointcloud):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        pixels_count = self.resolution[0] * self.resolution[1]\n        valid_fraction = len(other_pointcloud) / pixels_count\n        assert valid_fraction <= 1.0 and valid_fraction >= 0.0\n        overlap = compute_pointcloud_overlaps_scikit(\n            ref_pointcloud,\n            other_pointcloud,\n            self.distance_threshold,\n            compute_symmetric=True,\n        )\n        covisibility = min(\n            overlap[\"intersection1\"] / pixels_count,\n            overlap[\"intersection2\"] / pixels_count,\n        )\n        is_valid = (valid_fraction >= self.minimum_valid_fraction) and (\n            covisibility >= self.minimum_covisibility\n        )\n        return is_valid, valid_fraction, covisibility\n\n    def is_other_viewpoint_overlapping(\n        self, ref_pointcloud, observation, position, rotation\n    ):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        other_pointcloud = compute_pointcloud(\n            observation[\"depth\"], self.hfov, position, rotation\n        )\n        return self.is_other_pointcloud_overlapping(ref_pointcloud, other_pointcloud)\n\n    def render_viewpoint(self, viewpoint_position, viewpoint_orientation):\n        agent_state = habitat_sim.AgentState()\n        agent_state.position = viewpoint_position\n        agent_state.rotation = viewpoint_orientation\n        self.agent.set_state(agent_state)\n        viewpoint_observations = self.sim.get_sensor_observations(agent_ids=0)\n        _append_camera_parameters(\n            viewpoint_observations, self.hfov, viewpoint_position, viewpoint_orientation\n        )\n        return viewpoint_observations\n\n    def __getitem__(self, useless_idx):\n        ref_position, ref_orientation, nav_point = self.sample_random_viewpoint()\n        ref_observations = self.render_viewpoint(ref_position, ref_orientation)\n        # Extract point cloud\n        ref_pointcloud = compute_pointcloud(\n            depthmap=ref_observations[\"depth\"],\n            hfov=self.hfov,\n            camera_position=ref_position,\n            camera_rotation=ref_orientation,\n        )\n\n        pixels_count = self.resolution[0] * self.resolution[1]\n        ref_valid_fraction = len(ref_pointcloud) / pixels_count\n        assert ref_valid_fraction <= 1.0 and ref_valid_fraction >= 0.0\n        if ref_valid_fraction < self.minimum_valid_fraction:\n            # This should produce a recursion error at some point when something is very wrong.\n            return self[0]\n        # Pick an reference observed point in the point cloud\n        observed_point = np.mean(ref_pointcloud, axis=0)\n\n        # Add the first image as reference\n        viewpoints_observations = [ref_observations]\n        viewpoints_covisibility = [ref_valid_fraction]\n        viewpoints_positions = [ref_position]\n        viewpoints_orientations = [quaternion.as_float_array(ref_orientation)]\n        viewpoints_clouds = [ref_pointcloud]\n        viewpoints_valid_fractions = [ref_valid_fraction]\n\n        for _ in range(self.views_count - 1):\n            # Generate an other viewpoint using some dummy random walk\n            successful_sampling = False\n            for sampling_attempt in range(self.max_attempts_count):\n                position, rotation, _ = self.sample_other_random_viewpoint(\n                    observed_point, nav_point\n                )\n                # Observation\n                other_viewpoint_observations = self.render_viewpoint(position, rotation)\n                other_pointcloud = compute_pointcloud(\n                    other_viewpoint_observations[\"depth\"], self.hfov, position, rotation\n                )\n\n                is_valid, valid_fraction, covisibility = (\n                    self.is_other_pointcloud_overlapping(\n                        ref_pointcloud, other_pointcloud\n                    )\n                )\n                if is_valid:\n                    successful_sampling = True\n                    break\n            if not successful_sampling:\n                print(\"WARNING: Maximum number of attempts reached.\")\n                # Dirty hack, try using a novel original viewpoint\n                return self[0]\n            viewpoints_observations.append(other_viewpoint_observations)\n            viewpoints_covisibility.append(covisibility)\n            viewpoints_positions.append(position)\n            viewpoints_orientations.append(\n                quaternion.as_float_array(rotation)\n            )  # WXYZ convention for the quaternion encoding.\n            viewpoints_clouds.append(other_pointcloud)\n            viewpoints_valid_fractions.append(valid_fraction)\n\n        # Estimate relations between all pairs of images\n        pairwise_visibility_ratios = np.ones(\n            (len(viewpoints_observations), len(viewpoints_observations))\n        )\n        for i in range(len(viewpoints_observations)):\n            pairwise_visibility_ratios[i, i] = viewpoints_valid_fractions[i]\n            for j in range(i + 1, len(viewpoints_observations)):\n                overlap = compute_pointcloud_overlaps_scikit(\n                    viewpoints_clouds[i],\n                    viewpoints_clouds[j],\n                    self.distance_threshold,\n                    compute_symmetric=True,\n                )\n                pairwise_visibility_ratios[i, j] = (\n                    overlap[\"intersection1\"] / pixels_count\n                )\n                pairwise_visibility_ratios[j, i] = (\n                    overlap[\"intersection2\"] / pixels_count\n                )\n\n        # IoU is relative to the image 0\n        data = {\n            \"observations\": viewpoints_observations,\n            \"positions\": np.asarray(viewpoints_positions),\n            \"orientations\": np.asarray(viewpoints_orientations),\n            \"covisibility_ratios\": np.asarray(viewpoints_covisibility),\n            \"valid_fractions\": np.asarray(viewpoints_valid_fractions, dtype=float),\n            \"pairwise_visibility_ratios\": np.asarray(\n                pairwise_visibility_ratios, dtype=float\n            ),\n        }\n\n        if self.transform is not None:\n            data = self.transform(data)\n        return data\n\n    def generate_random_spiral_trajectory(\n        self,\n        images_count=100,\n        max_radius=0.5,\n        half_turns=5,\n        use_constant_orientation=False,\n    ):\n        \"\"\"\n        Return a list of images corresponding to a spiral trajectory from a random starting point.\n        Useful to generate nice visualisations.\n        Use an even number of half turns to get a nice \"C1-continuous\" loop effect\n        \"\"\"\n        ref_position, ref_orientation, navpoint = self.sample_random_viewpoint()\n        ref_observations = self.render_viewpoint(ref_position, ref_orientation)\n        ref_pointcloud = compute_pointcloud(\n            depthmap=ref_observations[\"depth\"],\n            hfov=self.hfov,\n            camera_position=ref_position,\n            camera_rotation=ref_orientation,\n        )\n        pixels_count = self.resolution[0] * self.resolution[1]\n        if len(ref_pointcloud) / pixels_count < self.minimum_valid_fraction:\n            # Dirty hack: ensure that the valid part of the image is significant\n            return self.generate_random_spiral_trajectory(\n                images_count, max_radius, half_turns, use_constant_orientation\n            )\n\n        # Pick an observed point in the point cloud\n        observed_point = np.mean(ref_pointcloud, axis=0)\n        ref_R, ref_t = compute_camera_pose_opencv_convention(\n            ref_position, ref_orientation\n        )\n\n        images = []\n        is_valid = []\n        # Spiral trajectory, use_constant orientation\n        for i, alpha in enumerate(np.linspace(0, 1, images_count)):\n            r = max_radius * np.abs(\n                np.sin(alpha * np.pi)\n            )  # Increase then decrease the radius\n            theta = alpha * half_turns * np.pi\n            x = r * np.cos(theta)\n            y = r * np.sin(theta)\n            z = 0.0\n            position = (\n                ref_position + (ref_R @ np.asarray([x, y, z]).reshape(3, 1)).flatten()\n            )\n            if use_constant_orientation:\n                orientation = ref_orientation\n            else:\n                # trajectory looking at a mean point in front of the ref observation\n                orientation, position = look_at_for_habitat(\n                    eye=position, center=observed_point, up=habitat_sim.geo.UP\n                )\n            observations = self.render_viewpoint(position, orientation)\n            images.append(observations[\"color\"][..., :3])\n            _is_valid, valid_fraction, iou = self.is_other_viewpoint_overlapping(\n                ref_pointcloud, observations, position, orientation\n            )\n            is_valid.append(_is_valid)\n        return images, np.all(is_valid)","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.__init__","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.__init__#L145-L193","kind":"function","name":"__init__","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":145,"end_line":193,"context_start_line":125,"context_end_line":213,"code":"def generate_orientation_noise(pan_range, tilt_range, roll_range):\n    return (\n        quaternion.from_rotation_vector(\n            np.random.uniform(*pan_range) * DEG2RAD * habitat_sim.geo.UP\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*tilt_range) * DEG2RAD * habitat_sim.geo.RIGHT\n        )\n        * quaternion.from_rotation_vector(\n            np.random.uniform(*roll_range) * DEG2RAD * habitat_sim.geo.FRONT\n        )\n    )\n\n\nclass NoNaviguableSpaceError(RuntimeError):\n    def __init__(self, *args):\n        super().__init__(*args)\n\n\nclass MultiviewHabitatSimGenerator:\n    def __init__(\n        self,\n        scene,\n        navmesh,\n        scene_dataset_config_file,\n        resolution=(240, 320),\n        views_count=2,\n        hfov=60,\n        gpu_id=0,\n        size=10000,\n        minimum_covisibility=0.5,\n        transform=None,\n    ):\n        self.scene = scene\n        self.navmesh = navmesh\n        self.scene_dataset_config_file = scene_dataset_config_file\n        self.resolution = resolution\n        self.views_count = views_count\n        assert self.views_count >= 1\n        self.hfov = hfov\n        self.gpu_id = gpu_id\n        self.size = size\n        self.transform = transform\n\n        # Noise added to camera orientation\n        self.pan_range = (-3, 3)\n        self.tilt_range = (-10, 10)\n        self.roll_range = (-5, 5)\n\n        # Height range to sample cameras\n        self.height_range = (1.2, 1.8)\n\n        # Random steps between the camera views\n        self.random_steps_count = 5\n        self.random_step_variance = 2.0\n\n        # Minimum fraction of the scene which should be valid (well defined depth)\n        self.minimum_valid_fraction = 0.7\n\n        # Distance threshold to see  to select pairs\n        self.distance_threshold = 0.05\n        # Minimum IoU of a view point cloud with respect to the reference view to be kept.\n        self.minimum_covisibility = minimum_covisibility\n\n        # Maximum number of retries.\n        self.max_attempts_count = 100\n\n        self.seed = None\n        self._lazy_initialization()\n\n    def _lazy_initialization(self):\n        # Lazy random seeding and instantiation of the simulator to deal with multiprocessing properly\n        if self.seed == None:\n            # Re-seed numpy generator\n            np.random.seed()\n            self.seed = np.random.randint(2**32 - 1)\n            sim_cfg = habitat_sim.SimulatorConfiguration()\n            sim_cfg.scene_id = self.scene\n            if (\n                self.scene_dataset_config_file is not None\n                and self.scene_dataset_config_file != \"\"\n            ):\n                sim_cfg.scene_dataset_config_file = self.scene_dataset_config_file\n            sim_cfg.random_seed = self.seed\n            sim_cfg.load_semantic_mesh = False\n            sim_cfg.gpu_device_id = self.gpu_id\n\n            depth_sensor_spec = habitat_sim.CameraSensorSpec()\n            depth_sensor_spec.uuid = \"depth\"","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator._lazy_initialization","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator._lazy_initialization#L195-L248","kind":"function","name":"_lazy_initialization","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":195,"end_line":248,"context_start_line":175,"context_end_line":268,"code":"        self.height_range = (1.2, 1.8)\n\n        # Random steps between the camera views\n        self.random_steps_count = 5\n        self.random_step_variance = 2.0\n\n        # Minimum fraction of the scene which should be valid (well defined depth)\n        self.minimum_valid_fraction = 0.7\n\n        # Distance threshold to see  to select pairs\n        self.distance_threshold = 0.05\n        # Minimum IoU of a view point cloud with respect to the reference view to be kept.\n        self.minimum_covisibility = minimum_covisibility\n\n        # Maximum number of retries.\n        self.max_attempts_count = 100\n\n        self.seed = None\n        self._lazy_initialization()\n\n    def _lazy_initialization(self):\n        # Lazy random seeding and instantiation of the simulator to deal with multiprocessing properly\n        if self.seed == None:\n            # Re-seed numpy generator\n            np.random.seed()\n            self.seed = np.random.randint(2**32 - 1)\n            sim_cfg = habitat_sim.SimulatorConfiguration()\n            sim_cfg.scene_id = self.scene\n            if (\n                self.scene_dataset_config_file is not None\n                and self.scene_dataset_config_file != \"\"\n            ):\n                sim_cfg.scene_dataset_config_file = self.scene_dataset_config_file\n            sim_cfg.random_seed = self.seed\n            sim_cfg.load_semantic_mesh = False\n            sim_cfg.gpu_device_id = self.gpu_id\n\n            depth_sensor_spec = habitat_sim.CameraSensorSpec()\n            depth_sensor_spec.uuid = \"depth\"\n            depth_sensor_spec.sensor_type = habitat_sim.SensorType.DEPTH\n            depth_sensor_spec.resolution = self.resolution\n            depth_sensor_spec.hfov = self.hfov\n            depth_sensor_spec.position = [0.0, 0.0, 0]\n            depth_sensor_spec.orientation\n\n            rgb_sensor_spec = habitat_sim.CameraSensorSpec()\n            rgb_sensor_spec.uuid = \"color\"\n            rgb_sensor_spec.sensor_type = habitat_sim.SensorType.COLOR\n            rgb_sensor_spec.resolution = self.resolution\n            rgb_sensor_spec.hfov = self.hfov\n            rgb_sensor_spec.position = [0.0, 0.0, 0]\n            agent_cfg = habitat_sim.agent.AgentConfiguration(\n                sensor_specifications=[rgb_sensor_spec, depth_sensor_spec]\n            )\n\n            cfg = habitat_sim.Configuration(sim_cfg, [agent_cfg])\n            self.sim = habitat_sim.Simulator(cfg)\n            if self.navmesh is not None and self.navmesh != \"\":\n                # Use pre-computed navmesh when available (usually better than those generated automatically)\n                self.sim.pathfinder.load_nav_mesh(self.navmesh)\n\n            if not self.sim.pathfinder.is_loaded:\n                # Try to compute a navmesh\n                navmesh_settings = habitat_sim.NavMeshSettings()\n                navmesh_settings.set_defaults()\n                self.sim.recompute_navmesh(self.sim.pathfinder, navmesh_settings, True)\n\n            # Ensure that the navmesh is not empty\n            if not self.sim.pathfinder.is_loaded:\n                raise NoNaviguableSpaceError(\n                    f\"No naviguable location (scene: {self.scene} -- navmesh: {self.navmesh})\"\n                )\n\n            self.agent = self.sim.initialize_agent(agent_id=0)\n\n    def close(self):\n        self.sim.close()\n\n    def __del__(self):\n        self.sim.close()\n\n    def __len__(self):\n        return self.size\n\n    def sample_random_viewpoint(self):\n        \"\"\"Sample a random viewpoint using the navmesh\"\"\"\n        nav_point = self.sim.pathfinder.get_random_navigable_point()\n\n        # Sample a random viewpoint height\n        viewpoint_height = np.random.uniform(*self.height_range)\n        viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP\n        viewpoint_orientation = quaternion.from_rotation_vector(\n            np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP\n        ) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.close","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.close#L250-L251","kind":"function","name":"close","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":250,"end_line":251,"context_start_line":230,"context_end_line":271,"code":"            cfg = habitat_sim.Configuration(sim_cfg, [agent_cfg])\n            self.sim = habitat_sim.Simulator(cfg)\n            if self.navmesh is not None and self.navmesh != \"\":\n                # Use pre-computed navmesh when available (usually better than those generated automatically)\n                self.sim.pathfinder.load_nav_mesh(self.navmesh)\n\n            if not self.sim.pathfinder.is_loaded:\n                # Try to compute a navmesh\n                navmesh_settings = habitat_sim.NavMeshSettings()\n                navmesh_settings.set_defaults()\n                self.sim.recompute_navmesh(self.sim.pathfinder, navmesh_settings, True)\n\n            # Ensure that the navmesh is not empty\n            if not self.sim.pathfinder.is_loaded:\n                raise NoNaviguableSpaceError(\n                    f\"No naviguable location (scene: {self.scene} -- navmesh: {self.navmesh})\"\n                )\n\n            self.agent = self.sim.initialize_agent(agent_id=0)\n\n    def close(self):\n        self.sim.close()\n\n    def __del__(self):\n        self.sim.close()\n\n    def __len__(self):\n        return self.size\n\n    def sample_random_viewpoint(self):\n        \"\"\"Sample a random viewpoint using the navmesh\"\"\"\n        nav_point = self.sim.pathfinder.get_random_navigable_point()\n\n        # Sample a random viewpoint height\n        viewpoint_height = np.random.uniform(*self.height_range)\n        viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP\n        viewpoint_orientation = quaternion.from_rotation_vector(\n            np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP\n        ) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)\n        return viewpoint_position, viewpoint_orientation, nav_point\n\n    def sample_other_random_viewpoint(self, observed_point, nav_point):","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.__del__","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.__del__#L253-L254","kind":"function","name":"__del__","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":253,"end_line":254,"context_start_line":233,"context_end_line":274,"code":"                # Use pre-computed navmesh when available (usually better than those generated automatically)\n                self.sim.pathfinder.load_nav_mesh(self.navmesh)\n\n            if not self.sim.pathfinder.is_loaded:\n                # Try to compute a navmesh\n                navmesh_settings = habitat_sim.NavMeshSettings()\n                navmesh_settings.set_defaults()\n                self.sim.recompute_navmesh(self.sim.pathfinder, navmesh_settings, True)\n\n            # Ensure that the navmesh is not empty\n            if not self.sim.pathfinder.is_loaded:\n                raise NoNaviguableSpaceError(\n                    f\"No naviguable location (scene: {self.scene} -- navmesh: {self.navmesh})\"\n                )\n\n            self.agent = self.sim.initialize_agent(agent_id=0)\n\n    def close(self):\n        self.sim.close()\n\n    def __del__(self):\n        self.sim.close()\n\n    def __len__(self):\n        return self.size\n\n    def sample_random_viewpoint(self):\n        \"\"\"Sample a random viewpoint using the navmesh\"\"\"\n        nav_point = self.sim.pathfinder.get_random_navigable_point()\n\n        # Sample a random viewpoint height\n        viewpoint_height = np.random.uniform(*self.height_range)\n        viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP\n        viewpoint_orientation = quaternion.from_rotation_vector(\n            np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP\n        ) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)\n        return viewpoint_position, viewpoint_orientation, nav_point\n\n    def sample_other_random_viewpoint(self, observed_point, nav_point):\n        \"\"\"Sample a random viewpoint close to an existing one, using the navmesh and a reference observed point.\"\"\"\n        other_nav_point = nav_point\n","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.__len__","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.__len__#L256-L257","kind":"function","name":"__len__","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":256,"end_line":257,"context_start_line":236,"context_end_line":277,"code":"            if not self.sim.pathfinder.is_loaded:\n                # Try to compute a navmesh\n                navmesh_settings = habitat_sim.NavMeshSettings()\n                navmesh_settings.set_defaults()\n                self.sim.recompute_navmesh(self.sim.pathfinder, navmesh_settings, True)\n\n            # Ensure that the navmesh is not empty\n            if not self.sim.pathfinder.is_loaded:\n                raise NoNaviguableSpaceError(\n                    f\"No naviguable location (scene: {self.scene} -- navmesh: {self.navmesh})\"\n                )\n\n            self.agent = self.sim.initialize_agent(agent_id=0)\n\n    def close(self):\n        self.sim.close()\n\n    def __del__(self):\n        self.sim.close()\n\n    def __len__(self):\n        return self.size\n\n    def sample_random_viewpoint(self):\n        \"\"\"Sample a random viewpoint using the navmesh\"\"\"\n        nav_point = self.sim.pathfinder.get_random_navigable_point()\n\n        # Sample a random viewpoint height\n        viewpoint_height = np.random.uniform(*self.height_range)\n        viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP\n        viewpoint_orientation = quaternion.from_rotation_vector(\n            np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP\n        ) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)\n        return viewpoint_position, viewpoint_orientation, nav_point\n\n    def sample_other_random_viewpoint(self, observed_point, nav_point):\n        \"\"\"Sample a random viewpoint close to an existing one, using the navmesh and a reference observed point.\"\"\"\n        other_nav_point = nav_point\n\n        walk_directions = self.random_step_variance * np.asarray([1, 0, 1])\n        for i in range(self.random_steps_count):\n            temp = self.sim.pathfinder.snap_point(","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.sample_random_viewpoint","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.sample_random_viewpoint#L259-L269","kind":"function","name":"sample_random_viewpoint","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":259,"end_line":269,"context_start_line":239,"context_end_line":289,"code":"                navmesh_settings.set_defaults()\n                self.sim.recompute_navmesh(self.sim.pathfinder, navmesh_settings, True)\n\n            # Ensure that the navmesh is not empty\n            if not self.sim.pathfinder.is_loaded:\n                raise NoNaviguableSpaceError(\n                    f\"No naviguable location (scene: {self.scene} -- navmesh: {self.navmesh})\"\n                )\n\n            self.agent = self.sim.initialize_agent(agent_id=0)\n\n    def close(self):\n        self.sim.close()\n\n    def __del__(self):\n        self.sim.close()\n\n    def __len__(self):\n        return self.size\n\n    def sample_random_viewpoint(self):\n        \"\"\"Sample a random viewpoint using the navmesh\"\"\"\n        nav_point = self.sim.pathfinder.get_random_navigable_point()\n\n        # Sample a random viewpoint height\n        viewpoint_height = np.random.uniform(*self.height_range)\n        viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP\n        viewpoint_orientation = quaternion.from_rotation_vector(\n            np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP\n        ) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)\n        return viewpoint_position, viewpoint_orientation, nav_point\n\n    def sample_other_random_viewpoint(self, observed_point, nav_point):\n        \"\"\"Sample a random viewpoint close to an existing one, using the navmesh and a reference observed point.\"\"\"\n        other_nav_point = nav_point\n\n        walk_directions = self.random_step_variance * np.asarray([1, 0, 1])\n        for i in range(self.random_steps_count):\n            temp = self.sim.pathfinder.snap_point(\n                other_nav_point + walk_directions * np.random.normal(size=3)\n            )\n            # Snapping may return nan when it fails\n            if not np.isnan(temp[0]):\n                other_nav_point = temp\n\n        other_viewpoint_height = np.random.uniform(*self.height_range)\n        other_viewpoint_position = (\n            other_nav_point + other_viewpoint_height * habitat_sim.geo.UP\n        )\n\n        # Set viewing direction towards the central point","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.sample_other_random_viewpoint","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.sample_other_random_viewpoint#L271-L299","kind":"function","name":"sample_other_random_viewpoint","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":271,"end_line":299,"context_start_line":251,"context_end_line":319,"code":"        self.sim.close()\n\n    def __del__(self):\n        self.sim.close()\n\n    def __len__(self):\n        return self.size\n\n    def sample_random_viewpoint(self):\n        \"\"\"Sample a random viewpoint using the navmesh\"\"\"\n        nav_point = self.sim.pathfinder.get_random_navigable_point()\n\n        # Sample a random viewpoint height\n        viewpoint_height = np.random.uniform(*self.height_range)\n        viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP\n        viewpoint_orientation = quaternion.from_rotation_vector(\n            np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP\n        ) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)\n        return viewpoint_position, viewpoint_orientation, nav_point\n\n    def sample_other_random_viewpoint(self, observed_point, nav_point):\n        \"\"\"Sample a random viewpoint close to an existing one, using the navmesh and a reference observed point.\"\"\"\n        other_nav_point = nav_point\n\n        walk_directions = self.random_step_variance * np.asarray([1, 0, 1])\n        for i in range(self.random_steps_count):\n            temp = self.sim.pathfinder.snap_point(\n                other_nav_point + walk_directions * np.random.normal(size=3)\n            )\n            # Snapping may return nan when it fails\n            if not np.isnan(temp[0]):\n                other_nav_point = temp\n\n        other_viewpoint_height = np.random.uniform(*self.height_range)\n        other_viewpoint_position = (\n            other_nav_point + other_viewpoint_height * habitat_sim.geo.UP\n        )\n\n        # Set viewing direction towards the central point\n        rotation, position = look_at_for_habitat(\n            eye=other_viewpoint_position,\n            center=observed_point,\n            up=habitat_sim.geo.UP,\n            return_cam2world=True,\n        )\n        rotation = rotation * generate_orientation_noise(\n            self.pan_range, self.tilt_range, self.roll_range\n        )\n        return position, rotation, other_nav_point\n\n    def is_other_pointcloud_overlapping(self, ref_pointcloud, other_pointcloud):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        pixels_count = self.resolution[0] * self.resolution[1]\n        valid_fraction = len(other_pointcloud) / pixels_count\n        assert valid_fraction <= 1.0 and valid_fraction >= 0.0\n        overlap = compute_pointcloud_overlaps_scikit(\n            ref_pointcloud,\n            other_pointcloud,\n            self.distance_threshold,\n            compute_symmetric=True,\n        )\n        covisibility = min(\n            overlap[\"intersection1\"] / pixels_count,\n            overlap[\"intersection2\"] / pixels_count,\n        )\n        is_valid = (valid_fraction >= self.minimum_valid_fraction) and (\n            covisibility >= self.minimum_covisibility\n        )","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.is_other_pointcloud_overlapping","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.is_other_pointcloud_overlapping#L301-L320","kind":"function","name":"is_other_pointcloud_overlapping","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":301,"end_line":320,"context_start_line":281,"context_end_line":340,"code":"            if not np.isnan(temp[0]):\n                other_nav_point = temp\n\n        other_viewpoint_height = np.random.uniform(*self.height_range)\n        other_viewpoint_position = (\n            other_nav_point + other_viewpoint_height * habitat_sim.geo.UP\n        )\n\n        # Set viewing direction towards the central point\n        rotation, position = look_at_for_habitat(\n            eye=other_viewpoint_position,\n            center=observed_point,\n            up=habitat_sim.geo.UP,\n            return_cam2world=True,\n        )\n        rotation = rotation * generate_orientation_noise(\n            self.pan_range, self.tilt_range, self.roll_range\n        )\n        return position, rotation, other_nav_point\n\n    def is_other_pointcloud_overlapping(self, ref_pointcloud, other_pointcloud):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        pixels_count = self.resolution[0] * self.resolution[1]\n        valid_fraction = len(other_pointcloud) / pixels_count\n        assert valid_fraction <= 1.0 and valid_fraction >= 0.0\n        overlap = compute_pointcloud_overlaps_scikit(\n            ref_pointcloud,\n            other_pointcloud,\n            self.distance_threshold,\n            compute_symmetric=True,\n        )\n        covisibility = min(\n            overlap[\"intersection1\"] / pixels_count,\n            overlap[\"intersection2\"] / pixels_count,\n        )\n        is_valid = (valid_fraction >= self.minimum_valid_fraction) and (\n            covisibility >= self.minimum_covisibility\n        )\n        return is_valid, valid_fraction, covisibility\n\n    def is_other_viewpoint_overlapping(\n        self, ref_pointcloud, observation, position, rotation\n    ):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        other_pointcloud = compute_pointcloud(\n            observation[\"depth\"], self.hfov, position, rotation\n        )\n        return self.is_other_pointcloud_overlapping(ref_pointcloud, other_pointcloud)\n\n    def render_viewpoint(self, viewpoint_position, viewpoint_orientation):\n        agent_state = habitat_sim.AgentState()\n        agent_state.position = viewpoint_position\n        agent_state.rotation = viewpoint_orientation\n        self.agent.set_state(agent_state)\n        viewpoint_observations = self.sim.get_sensor_observations(agent_ids=0)\n        _append_camera_parameters(\n            viewpoint_observations, self.hfov, viewpoint_position, viewpoint_orientation\n        )","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.is_other_viewpoint_overlapping","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.is_other_viewpoint_overlapping#L322-L330","kind":"function","name":"is_other_viewpoint_overlapping","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":322,"end_line":330,"context_start_line":302,"context_end_line":350,"code":"        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        pixels_count = self.resolution[0] * self.resolution[1]\n        valid_fraction = len(other_pointcloud) / pixels_count\n        assert valid_fraction <= 1.0 and valid_fraction >= 0.0\n        overlap = compute_pointcloud_overlaps_scikit(\n            ref_pointcloud,\n            other_pointcloud,\n            self.distance_threshold,\n            compute_symmetric=True,\n        )\n        covisibility = min(\n            overlap[\"intersection1\"] / pixels_count,\n            overlap[\"intersection2\"] / pixels_count,\n        )\n        is_valid = (valid_fraction >= self.minimum_valid_fraction) and (\n            covisibility >= self.minimum_covisibility\n        )\n        return is_valid, valid_fraction, covisibility\n\n    def is_other_viewpoint_overlapping(\n        self, ref_pointcloud, observation, position, rotation\n    ):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        other_pointcloud = compute_pointcloud(\n            observation[\"depth\"], self.hfov, position, rotation\n        )\n        return self.is_other_pointcloud_overlapping(ref_pointcloud, other_pointcloud)\n\n    def render_viewpoint(self, viewpoint_position, viewpoint_orientation):\n        agent_state = habitat_sim.AgentState()\n        agent_state.position = viewpoint_position\n        agent_state.rotation = viewpoint_orientation\n        self.agent.set_state(agent_state)\n        viewpoint_observations = self.sim.get_sensor_observations(agent_ids=0)\n        _append_camera_parameters(\n            viewpoint_observations, self.hfov, viewpoint_position, viewpoint_orientation\n        )\n        return viewpoint_observations\n\n    def __getitem__(self, useless_idx):\n        ref_position, ref_orientation, nav_point = self.sample_random_viewpoint()\n        ref_observations = self.render_viewpoint(ref_position, ref_orientation)\n        # Extract point cloud\n        ref_pointcloud = compute_pointcloud(\n            depthmap=ref_observations[\"depth\"],\n            hfov=self.hfov,\n            camera_position=ref_position,","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.render_viewpoint","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.render_viewpoint#L332-L341","kind":"function","name":"render_viewpoint","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":332,"end_line":341,"context_start_line":312,"context_end_line":361,"code":"        )\n        covisibility = min(\n            overlap[\"intersection1\"] / pixels_count,\n            overlap[\"intersection2\"] / pixels_count,\n        )\n        is_valid = (valid_fraction >= self.minimum_valid_fraction) and (\n            covisibility >= self.minimum_covisibility\n        )\n        return is_valid, valid_fraction, covisibility\n\n    def is_other_viewpoint_overlapping(\n        self, ref_pointcloud, observation, position, rotation\n    ):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        other_pointcloud = compute_pointcloud(\n            observation[\"depth\"], self.hfov, position, rotation\n        )\n        return self.is_other_pointcloud_overlapping(ref_pointcloud, other_pointcloud)\n\n    def render_viewpoint(self, viewpoint_position, viewpoint_orientation):\n        agent_state = habitat_sim.AgentState()\n        agent_state.position = viewpoint_position\n        agent_state.rotation = viewpoint_orientation\n        self.agent.set_state(agent_state)\n        viewpoint_observations = self.sim.get_sensor_observations(agent_ids=0)\n        _append_camera_parameters(\n            viewpoint_observations, self.hfov, viewpoint_position, viewpoint_orientation\n        )\n        return viewpoint_observations\n\n    def __getitem__(self, useless_idx):\n        ref_position, ref_orientation, nav_point = self.sample_random_viewpoint()\n        ref_observations = self.render_viewpoint(ref_position, ref_orientation)\n        # Extract point cloud\n        ref_pointcloud = compute_pointcloud(\n            depthmap=ref_observations[\"depth\"],\n            hfov=self.hfov,\n            camera_position=ref_position,\n            camera_rotation=ref_orientation,\n        )\n\n        pixels_count = self.resolution[0] * self.resolution[1]\n        ref_valid_fraction = len(ref_pointcloud) / pixels_count\n        assert ref_valid_fraction <= 1.0 and ref_valid_fraction >= 0.0\n        if ref_valid_fraction < self.minimum_valid_fraction:\n            # This should produce a recursion error at some point when something is very wrong.\n            return self[0]\n        # Pick an reference observed point in the point cloud\n        observed_point = np.mean(ref_pointcloud, axis=0)","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.__getitem__","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.__getitem__#L343-L439","kind":"function","name":"__getitem__","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":343,"end_line":439,"context_start_line":323,"context_end_line":459,"code":"        self, ref_pointcloud, observation, position, rotation\n    ):\n        \"\"\"Check if a viewpoint is valid and overlaps significantly with a reference one.\"\"\"\n        # Observation\n        other_pointcloud = compute_pointcloud(\n            observation[\"depth\"], self.hfov, position, rotation\n        )\n        return self.is_other_pointcloud_overlapping(ref_pointcloud, other_pointcloud)\n\n    def render_viewpoint(self, viewpoint_position, viewpoint_orientation):\n        agent_state = habitat_sim.AgentState()\n        agent_state.position = viewpoint_position\n        agent_state.rotation = viewpoint_orientation\n        self.agent.set_state(agent_state)\n        viewpoint_observations = self.sim.get_sensor_observations(agent_ids=0)\n        _append_camera_parameters(\n            viewpoint_observations, self.hfov, viewpoint_position, viewpoint_orientation\n        )\n        return viewpoint_observations\n\n    def __getitem__(self, useless_idx):\n        ref_position, ref_orientation, nav_point = self.sample_random_viewpoint()\n        ref_observations = self.render_viewpoint(ref_position, ref_orientation)\n        # Extract point cloud\n        ref_pointcloud = compute_pointcloud(\n            depthmap=ref_observations[\"depth\"],\n            hfov=self.hfov,\n            camera_position=ref_position,\n            camera_rotation=ref_orientation,\n        )\n\n        pixels_count = self.resolution[0] * self.resolution[1]\n        ref_valid_fraction = len(ref_pointcloud) / pixels_count\n        assert ref_valid_fraction <= 1.0 and ref_valid_fraction >= 0.0\n        if ref_valid_fraction < self.minimum_valid_fraction:\n            # This should produce a recursion error at some point when something is very wrong.\n            return self[0]\n        # Pick an reference observed point in the point cloud\n        observed_point = np.mean(ref_pointcloud, axis=0)\n\n        # Add the first image as reference\n        viewpoints_observations = [ref_observations]\n        viewpoints_covisibility = [ref_valid_fraction]\n        viewpoints_positions = [ref_position]\n        viewpoints_orientations = [quaternion.as_float_array(ref_orientation)]\n        viewpoints_clouds = [ref_pointcloud]\n        viewpoints_valid_fractions = [ref_valid_fraction]\n\n        for _ in range(self.views_count - 1):\n            # Generate an other viewpoint using some dummy random walk\n            successful_sampling = False\n            for sampling_attempt in range(self.max_attempts_count):\n                position, rotation, _ = self.sample_other_random_viewpoint(\n                    observed_point, nav_point\n                )\n                # Observation\n                other_viewpoint_observations = self.render_viewpoint(position, rotation)\n                other_pointcloud = compute_pointcloud(\n                    other_viewpoint_observations[\"depth\"], self.hfov, position, rotation\n                )\n\n                is_valid, valid_fraction, covisibility = (\n                    self.is_other_pointcloud_overlapping(\n                        ref_pointcloud, other_pointcloud\n                    )\n                )\n                if is_valid:\n                    successful_sampling = True\n                    break\n            if not successful_sampling:\n                print(\"WARNING: Maximum number of attempts reached.\")\n                # Dirty hack, try using a novel original viewpoint\n                return self[0]\n            viewpoints_observations.append(other_viewpoint_observations)\n            viewpoints_covisibility.append(covisibility)\n            viewpoints_positions.append(position)\n            viewpoints_orientations.append(\n                quaternion.as_float_array(rotation)\n            )  # WXYZ convention for the quaternion encoding.\n            viewpoints_clouds.append(other_pointcloud)\n            viewpoints_valid_fractions.append(valid_fraction)\n\n        # Estimate relations between all pairs of images\n        pairwise_visibility_ratios = np.ones(\n            (len(viewpoints_observations), len(viewpoints_observations))\n        )\n        for i in range(len(viewpoints_observations)):\n            pairwise_visibility_ratios[i, i] = viewpoints_valid_fractions[i]\n            for j in range(i + 1, len(viewpoints_observations)):\n                overlap = compute_pointcloud_overlaps_scikit(\n                    viewpoints_clouds[i],\n                    viewpoints_clouds[j],\n                    self.distance_threshold,\n                    compute_symmetric=True,\n                )\n                pairwise_visibility_ratios[i, j] = (\n                    overlap[\"intersection1\"] / pixels_count\n                )\n                pairwise_visibility_ratios[j, i] = (\n                    overlap[\"intersection2\"] / pixels_count\n                )\n\n        # IoU is relative to the image 0\n        data = {\n            \"observations\": viewpoints_observations,\n            \"positions\": np.asarray(viewpoints_positions),\n            \"orientations\": np.asarray(viewpoints_orientations),\n            \"covisibility_ratios\": np.asarray(viewpoints_covisibility),\n            \"valid_fractions\": np.asarray(viewpoints_valid_fractions, dtype=float),\n            \"pairwise_visibility_ratios\": np.asarray(\n                pairwise_visibility_ratios, dtype=float\n            ),\n        }\n\n        if self.transform is not None:\n            data = self.transform(data)\n        return data\n\n    def generate_random_spiral_trajectory(\n        self,\n        images_count=100,\n        max_radius=0.5,\n        half_turns=5,\n        use_constant_orientation=False,\n    ):\n        \"\"\"\n        Return a list of images corresponding to a spiral trajectory from a random starting point.\n        Useful to generate nice visualisations.\n        Use an even number of half turns to get a nice \"C1-continuous\" loop effect\n        \"\"\"\n        ref_position, ref_orientation, navpoint = self.sample_random_viewpoint()\n        ref_observations = self.render_viewpoint(ref_position, ref_orientation)\n        ref_pointcloud = compute_pointcloud(\n            depthmap=ref_observations[\"depth\"],\n            hfov=self.hfov,\n            camera_position=ref_position,\n            camera_rotation=ref_orientation,","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.generate_random_spiral_trajectory","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.multiview_habitat_sim_generator.generate_random_spiral_trajectory#L441-L501","kind":"function","name":"generate_random_spiral_trajectory","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":441,"end_line":501,"context_start_line":421,"context_end_line":501,"code":"                pairwise_visibility_ratios[j, i] = (\n                    overlap[\"intersection2\"] / pixels_count\n                )\n\n        # IoU is relative to the image 0\n        data = {\n            \"observations\": viewpoints_observations,\n            \"positions\": np.asarray(viewpoints_positions),\n            \"orientations\": np.asarray(viewpoints_orientations),\n            \"covisibility_ratios\": np.asarray(viewpoints_covisibility),\n            \"valid_fractions\": np.asarray(viewpoints_valid_fractions, dtype=float),\n            \"pairwise_visibility_ratios\": np.asarray(\n                pairwise_visibility_ratios, dtype=float\n            ),\n        }\n\n        if self.transform is not None:\n            data = self.transform(data)\n        return data\n\n    def generate_random_spiral_trajectory(\n        self,\n        images_count=100,\n        max_radius=0.5,\n        half_turns=5,\n        use_constant_orientation=False,\n    ):\n        \"\"\"\n        Return a list of images corresponding to a spiral trajectory from a random starting point.\n        Useful to generate nice visualisations.\n        Use an even number of half turns to get a nice \"C1-continuous\" loop effect\n        \"\"\"\n        ref_position, ref_orientation, navpoint = self.sample_random_viewpoint()\n        ref_observations = self.render_viewpoint(ref_position, ref_orientation)\n        ref_pointcloud = compute_pointcloud(\n            depthmap=ref_observations[\"depth\"],\n            hfov=self.hfov,\n            camera_position=ref_position,\n            camera_rotation=ref_orientation,\n        )\n        pixels_count = self.resolution[0] * self.resolution[1]\n        if len(ref_pointcloud) / pixels_count < self.minimum_valid_fraction:\n            # Dirty hack: ensure that the valid part of the image is significant\n            return self.generate_random_spiral_trajectory(\n                images_count, max_radius, half_turns, use_constant_orientation\n            )\n\n        # Pick an observed point in the point cloud\n        observed_point = np.mean(ref_pointcloud, axis=0)\n        ref_R, ref_t = compute_camera_pose_opencv_convention(\n            ref_position, ref_orientation\n        )\n\n        images = []\n        is_valid = []\n        # Spiral trajectory, use_constant orientation\n        for i, alpha in enumerate(np.linspace(0, 1, images_count)):\n            r = max_radius * np.abs(\n                np.sin(alpha * np.pi)\n            )  # Increase then decrease the radius\n            theta = alpha * half_turns * np.pi\n            x = r * np.cos(theta)\n            y = r * np.sin(theta)\n            z = 0.0\n            position = (\n                ref_position + (ref_R @ np.asarray([x, y, z]).reshape(3, 1)).flatten()\n            )\n            if use_constant_orientation:\n                orientation = ref_orientation\n            else:\n                # trajectory looking at a mean point in front of the ref observation\n                orientation, position = look_at_for_habitat(\n                    eye=position, center=observed_point, up=habitat_sim.geo.UP\n                )\n            observations = self.render_viewpoint(position, orientation)\n            images.append(observations[\"color\"][..., :3])\n            _is_valid, valid_fraction, iou = self.is_other_viewpoint_overlapping(\n                ref_pointcloud, observations, position, orientation\n            )\n            is_valid.append(_is_valid)\n        return images, np.all(is_valid)","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.pack_metadata_files","uri":"program://Human3R/module/src.croco.datasets.habitat_sim.pack_metadata_files#L1-L80","kind":"module","name":"src.croco.datasets.habitat_sim.pack_metadata_files","path":"src/croco/datasets/habitat_sim/pack_metadata_files.py","language":"python","start_line":1,"end_line":80,"context_start_line":1,"context_end_line":80,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\"\"\"\nUtility script to pack metadata files of the dataset in order to be able to re-generate it elsewhere.\n\"\"\"\nimport os\nimport glob\nfrom tqdm import tqdm\nimport shutil\nimport json\nfrom datasets.habitat_sim.paths import *\nimport argparse\nimport collections\n\nif __name__ == \"__main__\":\n    parser = argparse.ArgumentParser()\n    parser.add_argument(\"input_dir\")\n    parser.add_argument(\"output_dir\")\n    args = parser.parse_args()\n\n    input_dirname = args.input_dir\n    output_dirname = args.output_dir\n\n    input_metadata_filenames = glob.iglob(\n        f\"{input_dirname}/**/metadata.json\", recursive=True\n    )\n\n    images_count = collections.defaultdict(lambda: 0)\n\n    os.makedirs(output_dirname)\n    for input_filename in tqdm(input_metadata_filenames):\n        # Ignore empty files\n        with open(input_filename, \"r\") as f:\n            original_metadata = json.load(f)\n            if (\n                \"multiviews\" not in original_metadata\n                or len(original_metadata[\"multiviews\"]) == 0\n            ):\n                print(\"No views in\", input_filename)\n                continue\n\n        relpath = os.path.relpath(input_filename, input_dirname)\n        print(relpath)\n\n        # Copy metadata, while replacing scene paths by generic keys depending on the dataset, for portability.\n        # Data paths are sorted by decreasing length to avoid potential bugs due to paths starting by the same string pattern.\n        scenes_dataset_paths = dict(\n            sorted(SCENES_DATASET.items(), key=lambda x: len(x[1]), reverse=True)\n        )\n        metadata = dict()\n        for key, value in original_metadata.items():\n            if key in (\"scene_dataset_config_file\", \"scene\", \"navmesh\") and value != \"\":\n                known_path = False\n                for dataset, dataset_path in scenes_dataset_paths.items():\n                    if value.startswith(dataset_path):\n                        value = os.path.join(\n                            dataset, os.path.relpath(value, dataset_path)\n                        )\n                        known_path = True\n                        break\n                if not known_path:\n                    raise KeyError(\"Unknown path:\" + value)\n            metadata[key] = value\n\n        # Compile some general statistics while packing data\n        scene_split = metadata[\"scene\"].split(\"/\")\n        upper_level = (\n            \"/\".join(scene_split[:2]) if scene_split[0] == \"hm3d\" else scene_split[0]\n        )\n        images_count[upper_level] += len(metadata[\"multiviews\"])\n\n        output_filename = os.path.join(output_dirname, relpath)\n        os.makedirs(os.path.dirname(output_filename), exist_ok=True)\n        with open(output_filename, \"w\") as f:\n            json.dump(metadata, f)\n\n    # Print statistics\n    print(\"Images count:\")\n    for upper_level, count in images_count.items():\n        print(f\"- {upper_level}: {count}\")","source_hash":"229db302fb8332cbe6635e5f2f05976fe5610332b4d27e24898290f4b5f85c33","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.generate_from_metadata_files","uri":"program://Human3R/module/src.croco.datasets.habitat_sim.generate_from_metadata_files#L1-L36","kind":"module","name":"src.croco.datasets.habitat_sim.generate_from_metadata_files","path":"src/croco/datasets/habitat_sim/generate_from_metadata_files.py","language":"python","start_line":1,"end_line":36,"context_start_line":1,"context_end_line":36,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\"\"\"\nScript generating commandlines to generate image pairs from metadata files.\n\"\"\"\nimport os\nimport glob\nfrom tqdm import tqdm\nimport argparse\n\nif __name__ == \"__main__\":\n    parser = argparse.ArgumentParser()\n    parser.add_argument(\"--input_dir\", required=True)\n    parser.add_argument(\"--output_dir\", required=True)\n    parser.add_argument(\n        \"--prefix\",\n        default=\"\",\n        help=\"Commanline prefix, useful e.g. to setup environment.\",\n    )\n    args = parser.parse_args()\n\n    input_metadata_filenames = glob.iglob(\n        f\"{args.input_dir}/**/metadata.json\", recursive=True\n    )\n\n    for metadata_filename in tqdm(input_metadata_filenames):\n        output_dir = os.path.join(\n            args.output_dir,\n            os.path.relpath(os.path.dirname(metadata_filename), args.input_dir),\n        )\n        # Do not process the scene if the metadata file already exists\n        if os.path.exists(os.path.join(output_dir, \"metadata.json\")):\n            continue\n        commandline = f\"{args.prefix}python datasets/habitat_sim/generate_from_metadata.py --metadata_filename={metadata_filename} --output_dir={output_dir}\"\n        print(commandline)","source_hash":"256b884779bccbbf23b8f17af2003a93da03cbf4288ab8361399ec3e0b0cb944","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.generate_from_metadata","uri":"program://Human3R/module/src.croco.datasets.habitat_sim.generate_from_metadata#L1-L125","kind":"module","name":"src.croco.datasets.habitat_sim.generate_from_metadata","path":"src/croco/datasets/habitat_sim/generate_from_metadata.py","language":"python","start_line":1,"end_line":125,"context_start_line":1,"context_end_line":125,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\"\"\"\nScript to generate image pairs for a given scene reproducing poses provided in a metadata file.\n\"\"\"\nimport os\nfrom datasets.habitat_sim.multiview_habitat_sim_generator import (\n    MultiviewHabitatSimGenerator,\n)\nfrom datasets.habitat_sim.paths import SCENES_DATASET\nimport argparse\nimport quaternion\nimport PIL.Image\nimport cv2\nimport json\nfrom tqdm import tqdm\n\n\ndef generate_multiview_images_from_metadata(\n    metadata_filename,\n    output_dir,\n    overload_params=dict(),\n    scene_datasets_paths=None,\n    exist_ok=False,\n):\n    \"\"\"\n    Generate images from a metadata file for reproducibility purposes.\n    \"\"\"\n    # Reorder paths by decreasing label length, to avoid collisions when testing if a string by such label\n    if scene_datasets_paths is not None:\n        scene_datasets_paths = dict(\n            sorted(scene_datasets_paths.items(), key=lambda x: len(x[0]), reverse=True)\n        )\n\n    with open(metadata_filename, \"r\") as f:\n        input_metadata = json.load(f)\n    metadata = dict()\n    for key, value in input_metadata.items():\n        # Optionally replace some paths\n        if key in (\"scene_dataset_config_file\", \"scene\", \"navmesh\") and value != \"\":\n            if scene_datasets_paths is not None:\n                for dataset_label, dataset_path in scene_datasets_paths.items():\n                    if value.startswith(dataset_label):\n                        value = os.path.normpath(\n                            os.path.join(\n                                dataset_path, os.path.relpath(value, dataset_label)\n                            )\n                        )\n                        break\n        metadata[key] = value\n\n    # Overload some parameters\n    for key, value in overload_params.items():\n        metadata[key] = value\n\n    generation_entries = dict(\n        [\n            (key, value)\n            for key, value in metadata.items()\n            if not (key in (\"multiviews\", \"output_dir\", \"generate_depth\"))\n        ]\n    )\n    generate_depth = metadata[\"generate_depth\"]\n\n    os.makedirs(output_dir, exist_ok=exist_ok)\n\n    generator = MultiviewHabitatSimGenerator(**generation_entries)\n\n    # Generate views\n    for idx_label, data in tqdm(metadata[\"multiviews\"].items()):\n        positions = data[\"positions\"]\n        orientations = data[\"orientations\"]\n        n = len(positions)\n        for oidx in range(n):\n            observation = generator.render_viewpoint(\n                positions[oidx], quaternion.from_float_array(orientations[oidx])\n            )\n            observation_label = f\"{oidx + 1}\"  # Leonid is indexing starting from 1\n            # Color image saved using PIL\n            img = PIL.Image.fromarray(observation[\"color\"][:, :, :3])\n            filename = os.path.join(output_dir, f\"{idx_label}_{observation_label}.jpeg\")\n            img.save(filename)\n            if generate_depth:\n                # Depth image as EXR file\n                filename = os.path.join(\n                    output_dir, f\"{idx_label}_{observation_label}_depth.exr\"\n                )\n                cv2.imwrite(\n                    filename,\n                    observation[\"depth\"],\n                    [cv2.IMWRITE_EXR_TYPE, cv2.IMWRITE_EXR_TYPE_HALF],\n                )\n                # Camera parameters\n                camera_params = dict(\n                    [\n                        (key, observation[key].tolist())\n                        for key in (\"camera_intrinsics\", \"R_cam2world\", \"t_cam2world\")\n                    ]\n                )\n                filename = os.path.join(\n                    output_dir, f\"{idx_label}_{observation_label}_camera_params.json\"\n                )\n                with open(filename, \"w\") as f:\n                    json.dump(camera_params, f)\n                # Save metadata\n    with open(os.path.join(output_dir, \"metadata.json\"), \"w\") as f:\n        json.dump(metadata, f)\n\n    generator.close()\n\n\nif __name__ == \"__main__\":\n    parser = argparse.ArgumentParser()\n    parser.add_argument(\"--metadata_filename\", required=True)\n    parser.add_argument(\"--output_dir\", required=True)\n    args = parser.parse_args()\n\n    generate_multiview_images_from_metadata(\n        metadata_filename=args.metadata_filename,\n        output_dir=args.output_dir,\n        scene_datasets_paths=SCENES_DATASET,\n        overload_params=dict(),\n        exist_ok=True,\n    )","source_hash":"e1f03764fc02c17b46a0194c5751d5d582ebdb1f4e00c8e89c6e84d55ffb27c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.croco.datasets.habitat_sim.generate_from_metadata.generate_multiview_images_from_metadata","uri":"program://Human3R/function/src.croco.datasets.habitat_sim.generate_from_metadata.generate_multiview_images_from_metadata#L20-L110","kind":"function","name":"generate_multiview_images_from_metadata","path":"src/croco/datasets/habitat_sim/generate_from_metadata.py","language":"python","start_line":20,"end_line":110,"context_start_line":1,"context_end_line":125,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\"\"\"\nScript to generate image pairs for a given scene reproducing poses provided in a metadata file.\n\"\"\"\nimport os\nfrom datasets.habitat_sim.multiview_habitat_sim_generator import (\n    MultiviewHabitatSimGenerator,\n)\nfrom datasets.habitat_sim.paths import SCENES_DATASET\nimport argparse\nimport quaternion\nimport PIL.Image\nimport cv2\nimport json\nfrom tqdm import tqdm\n\n\ndef generate_multiview_images_from_metadata(\n    metadata_filename,\n    output_dir,\n    overload_params=dict(),\n    scene_datasets_paths=None,\n    exist_ok=False,\n):\n    \"\"\"\n    Generate images from a metadata file for reproducibility purposes.\n    \"\"\"\n    # Reorder paths by decreasing label length, to avoid collisions when testing if a string by such label\n    if scene_datasets_paths is not None:\n        scene_datasets_paths = dict(\n            sorted(scene_datasets_paths.items(), key=lambda x: len(x[0]), reverse=True)\n        )\n\n    with open(metadata_filename, \"r\") as f:\n        input_metadata = json.load(f)\n    metadata = dict()\n    for key, value in input_metadata.items():\n        # Optionally replace some paths\n        if key in (\"scene_dataset_config_file\", \"scene\", \"navmesh\") and value != \"\":\n            if scene_datasets_paths is not None:\n                for dataset_label, dataset_path in scene_datasets_paths.items():\n                    if value.startswith(dataset_label):\n                        value = os.path.normpath(\n                            os.path.join(\n                                dataset_path, os.path.relpath(value, dataset_label)\n                            )\n                        )\n                        break\n        metadata[key] = value\n\n    # Overload some parameters\n    for key, value in overload_params.items():\n        metadata[key] = value\n\n    generation_entries = dict(\n        [\n            (key, value)\n            for key, value in metadata.items()\n            if not (key in (\"multiviews\", \"output_dir\", \"generate_depth\"))\n        ]\n    )\n    generate_depth = metadata[\"generate_depth\"]\n\n    os.makedirs(output_dir, exist_ok=exist_ok)\n\n    generator = MultiviewHabitatSimGenerator(**generation_entries)\n\n    # Generate views\n    for idx_label, data in tqdm(metadata[\"multiviews\"].items()):\n        positions = data[\"positions\"]\n        orientations = data[\"orientations\"]\n        n = len(positions)\n        for oidx in range(n):\n            observation = generator.render_viewpoint(\n                positions[oidx], quaternion.from_float_array(orientations[oidx])\n            )\n            observation_label = f\"{oidx + 1}\"  # Leonid is indexing starting from 1\n            # Color image saved using PIL\n            img = PIL.Image.fromarray(observation[\"color\"][:, :, :3])\n            filename = os.path.join(output_dir, f\"{idx_label}_{observation_label}.jpeg\")\n            img.save(filename)\n            if generate_depth:\n                # Depth image as EXR file\n                filename = os.path.join(\n                    output_dir, f\"{idx_label}_{observation_label}_depth.exr\"\n                )\n                cv2.imwrite(\n                    filename,\n                    observation[\"depth\"],\n                    [cv2.IMWRITE_EXR_TYPE, cv2.IMWRITE_EXR_TYPE_HALF],\n                )\n                # Camera parameters\n                camera_params = dict(\n                    [\n                        (key, observation[key].tolist())\n                        for key in (\"camera_intrinsics\", \"R_cam2world\", \"t_cam2world\")\n                    ]\n                )\n                filename = os.path.join(\n                    output_dir, f\"{idx_label}_{observation_label}_camera_params.json\"\n                )\n                with open(filename, \"w\") as f:\n                    json.dump(camera_params, f)\n                # Save metadata\n    with open(os.path.join(output_dir, \"metadata.json\"), \"w\") as f:\n        json.dump(metadata, f)\n\n    generator.close()\n\n\nif __name__ == \"__main__\":\n    parser = argparse.ArgumentParser()\n    parser.add_argument(\"--metadata_filename\", required=True)\n    parser.add_argument(\"--output_dir\", required=True)\n    args = parser.parse_args()\n\n    generate_multiview_images_from_metadata(\n        metadata_filename=args.metadata_filename,\n        output_dir=args.output_dir,\n        scene_datasets_paths=SCENES_DATASET,\n        overload_params=dict(),\n        exist_ok=True,\n    )","source_hash":"e1f03764fc02c17b46a0194c5751d5d582ebdb1f4e00c8e89c6e84d55ffb27c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks","uri":"program://Human3R/module/src.dust3r.blocks#L1-L546","kind":"module","name":"src.dust3r.blocks","path":"src/dust3r/blocks.py","language":"python","start_line":1,"end_line":546,"context_start_line":1,"context_end_line":546,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc\nfrom torch.nn.functional import scaled_dot_product_attention\nfrom functools import partial\n\n\ndef _ntuple(n):\n    def parse(x):\n        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):\n            return x\n        return tuple(repeat(x, n))\n\n    return parse\n\n\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n\n        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])\n        self.act = act_layer()\n        self.drop1 = nn.Dropout(drop_probs[0])\n        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])\n        self.drop2 = nn.Dropout(drop_probs[1])\n\n    def forward(self, x):\n        return self.drop2(self.fc2(self.drop1(self.act(self.fc1(x)))))\n\n\nclass Attention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, x, xpos):\n        B, N, C = x.shape\n\n        qkv = (\n            self.qkv(x)\n            .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n            .transpose(1, 3)\n        )\n        q, k, v = [qkv[:, :, i] for i in range(3)]\n\n        q_type = q.dtype\n        k_type = k.dtype\n        if self.rope is not None:\n            q = q.float()\n            k = k.float()\n            with torch.autocast(device_type=\"cuda\", enabled=False):\n                q = self.rope(q, xpos)\n                k = self.rope(k, xpos)\n            q = q.to(q_type)\n            k = k.to(k_type)\n\n        x = (\n            scaled_dot_product_attention(\n                query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n            )\n            .transpose(1, 2)\n            .reshape(B, N, C)\n        )\n\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        return x\n\n\nclass Block(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x)))\n        return x\n\n\nclass CrossAttention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n\n        self.projq = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projk = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projv = nn.Linear(dim, dim, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, query, key, value, qpos, kpos, use_ttt3r=False):\n        B, Nq, C = query.shape\n        Nk = key.shape[1]\n        Nv = value.shape[1]\n\n        q = (\n            self.projq(query)\n            .reshape(B, Nq, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n        k = (\n            self.projk(key)\n            .reshape(B, Nk, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n        v = (\n            self.projv(value)\n            .reshape(B, Nv, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n\n        q_type = q.dtype\n        k_type = k.dtype\n        if self.rope is not None:\n            if qpos is not None:\n                q = q.float()\n                with torch.autocast(device_type=\"cuda\", enabled=False):\n                    q = self.rope(q, qpos)\n                q = q.to(q_type)\n\n            if kpos is not None:\n                k = k.float()\n                with torch.autocast(device_type=\"cuda\", enabled=False):\n                    k = self.rope(k, kpos)\n                k = k.to(k_type)\n\n        if use_ttt3r:\n            # original attention\n            attn = (q @ k.transpose(-2, -1)) * self.scale  # [B, num_heads, Nq, Nk] [1, 16, 768, 1 + 576]\n            attn_before_softmax = attn.detach().clone()\n\n            attn = attn.softmax(dim=-1)\n            attn = self.attn_drop(attn)\n            x = (attn @ v).transpose(1, 2).reshape(B, Nq, C) # [B, Nq, C] [1, 768, 768]\n\n            x = self.proj(x)\n            x = self.proj_drop(x)\n\n            return x, attn_before_softmax\n        else:\n            x = (\n                scaled_dot_product_attention(\n                    query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n                )\n                .transpose(1, 2)\n                .reshape(B, Nq, C)\n            )\n\n            x = self.proj(x)\n            x = self.proj_drop(x)\n            return x, None\n\n\nclass DecoderBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.cross_attn = CrossAttention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        self.norm3 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, xpos, ypos, use_ttt3r=False):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        cross_attn_output, cross_attn = self.cross_attn(self.norm2(x), y_, y_, xpos, ypos, use_ttt3r=use_ttt3r)\n        x = x + self.drop_path(cross_attn_output)\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y, cross_attn\n\n\nclass CustomDecoderBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.cross_attn = CrossAttention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        self.norm3 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n        self.norm_z = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, z, xpos, ypos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        z_ = self.norm_z(z)\n        x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, z_, xpos, ypos))\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y\n\n\nclass ModLN(nn.Module):\n    \"\"\"\n    Modulation with adaLN.\n\n    References:\n    DiT: https://github.com/facebookresearch/DiT/blob/main/models.py#L101\n    \"\"\"\n\n    def __init__(self, inner_dim: int, mod_dim: int, eps: float):\n        super().__init__()\n        self.norm = nn.LayerNorm(inner_dim, eps=eps)\n        self.mlp = nn.Sequential(\n            nn.SiLU(),\n            nn.Linear(mod_dim, inner_dim * 2),\n        )\n\n    @staticmethod\n    def modulate(x, shift, scale):\n\n        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)\n\n    def forward(self, x: torch.Tensor, mod: torch.Tensor) -> torch.Tensor:\n        shift, scale = self.mlp(mod).chunk(2, dim=-1)  # [N, D]\n        return self.modulate(self.norm(x), shift, scale)  # [N, L, D]\n\n\nclass ConditionModulationBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=partial(ModLN, eps=1e-6),\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim, dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim, dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, mod, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x, mod), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x, mod)))\n        return x\n\n\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n        self.num_patches = self.grid_size[0] * self.grid_size[1]\n        self.flatten = flatten\n\n        self.proj = nn.Conv2d(\n            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n        )\n        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()\n\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data\n        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))\n\n\nif __name__ == \"__main__\":\n    import os\n    import sys\n\n    sys.path.append(os.path.dirname(os.path.dirname(__file__)))\n    import dust3r.utils.path_to_croco\n    from models.pos_embed import get_2d_sincos_pos_embed, RoPE2D\n    from functools import partial\n    from torch.utils.checkpoint import checkpoint\n\n    torch.manual_seed(0)\n\n    enc_blocks_ray_map = (\n        nn.ModuleList(\n            [\n                Block(\n                    768,\n                    16,\n                    4,\n                    qkv_bias=True,\n                    norm_layer=partial(nn.LayerNorm, eps=1e-6),\n                    rope=RoPE2D(100),\n                )\n                for _ in range(2)\n            ]\n        )\n        .cuda()\n        .train()\n    )\n\n    x = torch.randn(2, 196, 768, requires_grad=True).cuda()\n    xpos = torch.arange(0, 196).unsqueeze(0).unsqueeze(-1).repeat(2, 1, 2).cuda().long()\n    enc_blocks_ray_map.zero_grad()\n    for blk in enc_blocks_ray_map:\n\n        x = checkpoint(blk, x, xpos)\n    enc_blocks_ray_map.zero_grad()\n    x.sum().backward()\n\n    grad_not_checkpointed = {}\n    for name, param in enc_blocks_ray_map.named_parameters():\n        grad_not_checkpointed[name] = param.grad.data.clone()\n        print(name, grad_not_checkpointed[name])\n        break","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks._ntuple","uri":"program://Human3R/function/src.dust3r.blocks._ntuple#L16-L22","kind":"function","name":"_ntuple","path":"src/dust3r/blocks.py","language":"python","start_line":16,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc\nfrom torch.nn.functional import scaled_dot_product_attention\nfrom functools import partial\n\n\ndef _ntuple(n):\n    def parse(x):\n        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):\n            return x\n        return tuple(repeat(x, n))\n\n    return parse\n\n\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.drop_path","uri":"program://Human3R/function/src.dust3r.blocks.drop_path#L28-L41","kind":"function","name":"drop_path","path":"src/dust3r/blocks.py","language":"python","start_line":28,"end_line":41,"context_start_line":8,"context_end_line":61,"code":"import torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc\nfrom torch.nn.functional import scaled_dot_product_attention\nfrom functools import partial\n\n\ndef _ntuple(n):\n    def parse(x):\n        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):\n            return x\n        return tuple(repeat(x, n))\n\n    return parse\n\n\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.DropPath","uri":"program://Human3R/class/src.dust3r.blocks.DropPath#L44-L56","kind":"class","name":"DropPath","path":"src/dust3r/blocks.py","language":"python","start_line":44,"end_line":56,"context_start_line":24,"context_end_line":76,"code":"\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.Mlp","uri":"program://Human3R/class/src.dust3r.blocks.Mlp#L59-L84","kind":"class","name":"Mlp","path":"src/dust3r/blocks.py","language":"python","start_line":59,"end_line":84,"context_start_line":39,"context_end_line":104,"code":"    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n\n        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])\n        self.act = act_layer()\n        self.drop1 = nn.Dropout(drop_probs[0])\n        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])\n        self.drop2 = nn.Dropout(drop_probs[1])\n\n    def forward(self, x):\n        return self.drop2(self.fc2(self.drop1(self.act(self.fc1(x)))))\n\n\nclass Attention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, x, xpos):\n        B, N, C = x.shape\n","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.Attention","uri":"program://Human3R/class/src.dust3r.blocks.Attention#L87-L133","kind":"class","name":"Attention","path":"src/dust3r/blocks.py","language":"python","start_line":87,"end_line":133,"context_start_line":67,"context_end_line":153,"code":"        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n\n        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])\n        self.act = act_layer()\n        self.drop1 = nn.Dropout(drop_probs[0])\n        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])\n        self.drop2 = nn.Dropout(drop_probs[1])\n\n    def forward(self, x):\n        return self.drop2(self.fc2(self.drop1(self.act(self.fc1(x)))))\n\n\nclass Attention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, x, xpos):\n        B, N, C = x.shape\n\n        qkv = (\n            self.qkv(x)\n            .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n            .transpose(1, 3)\n        )\n        q, k, v = [qkv[:, :, i] for i in range(3)]\n\n        q_type = q.dtype\n        k_type = k.dtype\n        if self.rope is not None:\n            q = q.float()\n            k = k.float()\n            with torch.autocast(device_type=\"cuda\", enabled=False):\n                q = self.rope(q, xpos)\n                k = self.rope(k, xpos)\n            q = q.to(q_type)\n            k = k.to(k_type)\n\n        x = (\n            scaled_dot_product_attention(\n                query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n            )\n            .transpose(1, 2)\n            .reshape(B, N, C)\n        )\n\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        return x\n\n\nclass Block(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.Block","uri":"program://Human3R/class/src.dust3r.blocks.Block#L136-L175","kind":"class","name":"Block","path":"src/dust3r/blocks.py","language":"python","start_line":136,"end_line":175,"context_start_line":116,"context_end_line":195,"code":"            k = k.float()\n            with torch.autocast(device_type=\"cuda\", enabled=False):\n                q = self.rope(q, xpos)\n                k = self.rope(k, xpos)\n            q = q.to(q_type)\n            k = k.to(k_type)\n\n        x = (\n            scaled_dot_product_attention(\n                query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n            )\n            .transpose(1, 2)\n            .reshape(B, N, C)\n        )\n\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        return x\n\n\nclass Block(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x)))\n        return x\n\n\nclass CrossAttention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n\n        self.projq = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projk = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projv = nn.Linear(dim, dim, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n\n        self.rope = rope.float() if rope is not None else None","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.CrossAttention","uri":"program://Human3R/class/src.dust3r.blocks.CrossAttention#L178-L257","kind":"class","name":"CrossAttention","path":"src/dust3r/blocks.py","language":"python","start_line":178,"end_line":257,"context_start_line":158,"context_end_line":277,"code":"            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x)))\n        return x\n\n\nclass CrossAttention(nn.Module):\n\n    def __init__(\n        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0\n    ):\n        super().__init__()\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = head_dim**-0.5\n\n        self.projq = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projk = nn.Linear(dim, dim, bias=qkv_bias)\n        self.projv = nn.Linear(dim, dim, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n\n        self.rope = rope.float() if rope is not None else None\n\n    def forward(self, query, key, value, qpos, kpos, use_ttt3r=False):\n        B, Nq, C = query.shape\n        Nk = key.shape[1]\n        Nv = value.shape[1]\n\n        q = (\n            self.projq(query)\n            .reshape(B, Nq, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n        k = (\n            self.projk(key)\n            .reshape(B, Nk, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n        v = (\n            self.projv(value)\n            .reshape(B, Nv, self.num_heads, C // self.num_heads)\n            .permute(0, 2, 1, 3)\n        )\n\n        q_type = q.dtype\n        k_type = k.dtype\n        if self.rope is not None:\n            if qpos is not None:\n                q = q.float()\n                with torch.autocast(device_type=\"cuda\", enabled=False):\n                    q = self.rope(q, qpos)\n                q = q.to(q_type)\n\n            if kpos is not None:\n                k = k.float()\n                with torch.autocast(device_type=\"cuda\", enabled=False):\n                    k = self.rope(k, kpos)\n                k = k.to(k_type)\n\n        if use_ttt3r:\n            # original attention\n            attn = (q @ k.transpose(-2, -1)) * self.scale  # [B, num_heads, Nq, Nk] [1, 16, 768, 1 + 576]\n            attn_before_softmax = attn.detach().clone()\n\n            attn = attn.softmax(dim=-1)\n            attn = self.attn_drop(attn)\n            x = (attn @ v).transpose(1, 2).reshape(B, Nq, C) # [B, Nq, C] [1, 768, 768]\n\n            x = self.proj(x)\n            x = self.proj_drop(x)\n\n            return x, attn_before_softmax\n        else:\n            x = (\n                scaled_dot_product_attention(\n                    query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n                )\n                .transpose(1, 2)\n                .reshape(B, Nq, C)\n            )\n\n            x = self.proj(x)\n            x = self.proj_drop(x)\n            return x, None\n\n\nclass DecoderBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.DecoderBlock","uri":"program://Human3R/class/src.dust3r.blocks.DecoderBlock#L260-L312","kind":"class","name":"DecoderBlock","path":"src/dust3r/blocks.py","language":"python","start_line":260,"end_line":312,"context_start_line":240,"context_end_line":332,"code":"            x = (attn @ v).transpose(1, 2).reshape(B, Nq, C) # [B, Nq, C] [1, 768, 768]\n\n            x = self.proj(x)\n            x = self.proj_drop(x)\n\n            return x, attn_before_softmax\n        else:\n            x = (\n                scaled_dot_product_attention(\n                    query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale\n                )\n                .transpose(1, 2)\n                .reshape(B, Nq, C)\n            )\n\n            x = self.proj(x)\n            x = self.proj_drop(x)\n            return x, None\n\n\nclass DecoderBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.cross_attn = CrossAttention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        self.norm3 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, xpos, ypos, use_ttt3r=False):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        cross_attn_output, cross_attn = self.cross_attn(self.norm2(x), y_, y_, xpos, ypos, use_ttt3r=use_ttt3r)\n        x = x + self.drop_path(cross_attn_output)\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y, cross_attn\n\n\nclass CustomDecoderBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.CustomDecoderBlock","uri":"program://Human3R/class/src.dust3r.blocks.CustomDecoderBlock#L315-L368","kind":"class","name":"CustomDecoderBlock","path":"src/dust3r/blocks.py","language":"python","start_line":315,"end_line":368,"context_start_line":295,"context_end_line":388,"code":"        self.norm2 = norm_layer(dim)\n        self.norm3 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, xpos, ypos, use_ttt3r=False):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        cross_attn_output, cross_attn = self.cross_attn(self.norm2(x), y_, y_, xpos, ypos, use_ttt3r=use_ttt3r)\n        x = x + self.drop_path(cross_attn_output)\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y, cross_attn\n\n\nclass CustomDecoderBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.cross_attn = CrossAttention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        self.norm3 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n        self.norm_z = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, z, xpos, ypos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        z_ = self.norm_z(z)\n        x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, z_, xpos, ypos))\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y\n\n\nclass ModLN(nn.Module):\n    \"\"\"\n    Modulation with adaLN.\n\n    References:\n    DiT: https://github.com/facebookresearch/DiT/blob/main/models.py#L101\n    \"\"\"\n\n    def __init__(self, inner_dim: int, mod_dim: int, eps: float):\n        super().__init__()\n        self.norm = nn.LayerNorm(inner_dim, eps=eps)\n        self.mlp = nn.Sequential(\n            nn.SiLU(),\n            nn.Linear(mod_dim, inner_dim * 2),\n        )\n\n    @staticmethod\n    def modulate(x, shift, scale):","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.ModLN","uri":"program://Human3R/class/src.dust3r.blocks.ModLN#L371-L394","kind":"class","name":"ModLN","path":"src/dust3r/blocks.py","language":"python","start_line":371,"end_line":394,"context_start_line":351,"context_end_line":414,"code":"        self.norm3 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()\n        self.norm_z = norm_layer(dim) if norm_mem else nn.Identity()\n\n    def forward(self, x, y, z, xpos, ypos):\n        x = x + self.drop_path(self.attn(self.norm1(x), xpos))\n        y_ = self.norm_y(y)\n        z_ = self.norm_z(z)\n        x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, z_, xpos, ypos))\n        x = x + self.drop_path(self.mlp(self.norm3(x)))\n        return x, y\n\n\nclass ModLN(nn.Module):\n    \"\"\"\n    Modulation with adaLN.\n\n    References:\n    DiT: https://github.com/facebookresearch/DiT/blob/main/models.py#L101\n    \"\"\"\n\n    def __init__(self, inner_dim: int, mod_dim: int, eps: float):\n        super().__init__()\n        self.norm = nn.LayerNorm(inner_dim, eps=eps)\n        self.mlp = nn.Sequential(\n            nn.SiLU(),\n            nn.Linear(mod_dim, inner_dim * 2),\n        )\n\n    @staticmethod\n    def modulate(x, shift, scale):\n\n        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)\n\n    def forward(self, x: torch.Tensor, mod: torch.Tensor) -> torch.Tensor:\n        shift, scale = self.mlp(mod).chunk(2, dim=-1)  # [N, D]\n        return self.modulate(self.norm(x), shift, scale)  # [N, L, D]\n\n\nclass ConditionModulationBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=partial(ModLN, eps=1e-6),\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim, dim)\n        self.attn = Attention(","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.ConditionModulationBlock","uri":"program://Human3R/class/src.dust3r.blocks.ConditionModulationBlock#L397-L435","kind":"class","name":"ConditionModulationBlock","path":"src/dust3r/blocks.py","language":"python","start_line":397,"end_line":435,"context_start_line":377,"context_end_line":455,"code":"    \"\"\"\n\n    def __init__(self, inner_dim: int, mod_dim: int, eps: float):\n        super().__init__()\n        self.norm = nn.LayerNorm(inner_dim, eps=eps)\n        self.mlp = nn.Sequential(\n            nn.SiLU(),\n            nn.Linear(mod_dim, inner_dim * 2),\n        )\n\n    @staticmethod\n    def modulate(x, shift, scale):\n\n        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)\n\n    def forward(self, x: torch.Tensor, mod: torch.Tensor) -> torch.Tensor:\n        shift, scale = self.mlp(mod).chunk(2, dim=-1)  # [N, D]\n        return self.modulate(self.norm(x), shift, scale)  # [N, L, D]\n\n\nclass ConditionModulationBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=partial(ModLN, eps=1e-6),\n        rope=None,\n    ):\n        super().__init__()\n        self.norm1 = norm_layer(dim, dim)\n        self.attn = Attention(\n            dim,\n            rope=rope,\n            num_heads=num_heads,\n            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim, dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, mod, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x, mod), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x, mod)))\n        return x\n\n\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.PositionGetter","uri":"program://Human3R/class/src.dust3r.blocks.PositionGetter#L438-L450","kind":"class","name":"PositionGetter","path":"src/dust3r/blocks.py","language":"python","start_line":438,"end_line":450,"context_start_line":418,"context_end_line":470,"code":"            qkv_bias=qkv_bias,\n            attn_drop=attn_drop,\n            proj_drop=drop,\n        )\n        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n        self.norm2 = norm_layer(dim, dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, mod, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x, mod), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x, mod)))\n        return x\n\n\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.PatchEmbed","uri":"program://Human3R/class/src.dust3r.blocks.PatchEmbed#L453-L500","kind":"class","name":"PatchEmbed","path":"src/dust3r/blocks.py","language":"python","start_line":453,"end_line":500,"context_start_line":433,"context_end_line":520,"code":"        x = x + self.drop_path(self.attn(self.norm1(x, mod), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x, mod)))\n        return x\n\n\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n        self.num_patches = self.grid_size[0] * self.grid_size[1]\n        self.flatten = flatten\n\n        self.proj = nn.Conv2d(\n            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n        )\n        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()\n\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data\n        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))\n\n\nif __name__ == \"__main__\":\n    import os\n    import sys\n\n    sys.path.append(os.path.dirname(os.path.dirname(__file__)))\n    import dust3r.utils.path_to_croco\n    from models.pos_embed import get_2d_sincos_pos_embed, RoPE2D\n    from functools import partial\n    from torch.utils.checkpoint import checkpoint\n\n    torch.manual_seed(0)\n\n    enc_blocks_ray_map = (\n        nn.ModuleList(\n            [\n                Block(\n                    768,\n                    16,","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.parse","uri":"program://Human3R/function/src.dust3r.blocks.parse#L17-L20","kind":"function","name":"parse","path":"src/dust3r/blocks.py","language":"python","start_line":17,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc\nfrom torch.nn.functional import scaled_dot_product_attention\nfrom functools import partial\n\n\ndef _ntuple(n):\n    def parse(x):\n        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):\n            return x\n        return tuple(repeat(x, n))\n\n    return parse\n\n\nto_2tuple = _ntuple(2)\n\n\ndef drop_path(\n    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True\n):\n    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n    if drop_prob == 0.0 or not training:\n        return x\n    keep_prob = 1 - drop_prob\n    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.__init__","uri":"program://Human3R/function/src.dust3r.blocks.__init__#L456-L479","kind":"function","name":"__init__","path":"src/dust3r/blocks.py","language":"python","start_line":456,"end_line":479,"context_start_line":436,"context_end_line":499,"code":"\n\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n        self.num_patches = self.grid_size[0] * self.grid_size[1]\n        self.flatten = flatten\n\n        self.proj = nn.Conv2d(\n            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n        )\n        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()\n\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.forward","uri":"program://Human3R/function/src.dust3r.blocks.forward#L481-L496","kind":"function","name":"forward","path":"src/dust3r/blocks.py","language":"python","start_line":481,"end_line":496,"context_start_line":461,"context_end_line":516,"code":"        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n        self.num_patches = self.grid_size[0] * self.grid_size[1]\n        self.flatten = flatten\n\n        self.proj = nn.Conv2d(\n            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n        )\n        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()\n\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data\n        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))\n\n\nif __name__ == \"__main__\":\n    import os\n    import sys\n\n    sys.path.append(os.path.dirname(os.path.dirname(__file__)))\n    import dust3r.utils.path_to_croco\n    from models.pos_embed import get_2d_sincos_pos_embed, RoPE2D\n    from functools import partial\n    from torch.utils.checkpoint import checkpoint\n\n    torch.manual_seed(0)\n\n    enc_blocks_ray_map = (\n        nn.ModuleList(","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.extra_repr","uri":"program://Human3R/function/src.dust3r.blocks.extra_repr#L55-L56","kind":"function","name":"extra_repr","path":"src/dust3r/blocks.py","language":"python","start_line":55,"end_line":56,"context_start_line":35,"context_end_line":76,"code":"    shape = (x.shape[0],) + (1,) * (\n        x.ndim - 1\n    )  # work with diff dim tensors, not just 2D ConvNets\n    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)\n    if keep_prob > 0.0 and scale_by_keep:\n        random_tensor.div_(keep_prob)\n    return x * random_tensor\n\n\nclass DropPath(nn.Module):\n    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\"\"\"\n\n    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):\n        super(DropPath, self).__init__()\n        self.drop_prob = drop_prob\n        self.scale_by_keep = scale_by_keep\n\n    def forward(self, x):\n        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)\n\n    def extra_repr(self):\n        return f\"drop_prob={round(self.drop_prob,3):0.3f}\"\n\n\nclass Mlp(nn.Module):\n    \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n    def __init__(\n        self,\n        in_features,\n        hidden_features=None,\n        out_features=None,\n        act_layer=nn.GELU,\n        bias=True,\n        drop=0.0,\n    ):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        bias = to_2tuple(bias)\n        drop_probs = to_2tuple(drop)\n","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.modulate","uri":"program://Human3R/function/src.dust3r.blocks.modulate#L388-L390","kind":"function","name":"modulate","path":"src/dust3r/blocks.py","language":"python","start_line":388,"end_line":390,"context_start_line":368,"context_end_line":410,"code":"        return x, y\n\n\nclass ModLN(nn.Module):\n    \"\"\"\n    Modulation with adaLN.\n\n    References:\n    DiT: https://github.com/facebookresearch/DiT/blob/main/models.py#L101\n    \"\"\"\n\n    def __init__(self, inner_dim: int, mod_dim: int, eps: float):\n        super().__init__()\n        self.norm = nn.LayerNorm(inner_dim, eps=eps)\n        self.mlp = nn.Sequential(\n            nn.SiLU(),\n            nn.Linear(mod_dim, inner_dim * 2),\n        )\n\n    @staticmethod\n    def modulate(x, shift, scale):\n\n        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)\n\n    def forward(self, x: torch.Tensor, mod: torch.Tensor) -> torch.Tensor:\n        shift, scale = self.mlp(mod).chunk(2, dim=-1)  # [N, D]\n        return self.modulate(self.norm(x), shift, scale)  # [N, L, D]\n\n\nclass ConditionModulationBlock(nn.Module):\n\n    def __init__(\n        self,\n        dim,\n        num_heads,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=partial(ModLN, eps=1e-6),\n        rope=None,","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks.__call__","uri":"program://Human3R/function/src.dust3r.blocks.__call__#L444-L450","kind":"function","name":"__call__","path":"src/dust3r/blocks.py","language":"python","start_line":444,"end_line":450,"context_start_line":424,"context_end_line":470,"code":"        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(\n            in_features=dim,\n            hidden_features=mlp_hidden_dim,\n            act_layer=act_layer,\n            drop=drop,\n        )\n\n    def forward(self, x, mod, xpos):\n        x = x + self.drop_path(self.attn(self.norm1(x, mod), xpos))\n        x = x + self.drop_path(self.mlp(self.norm2(x, mod)))\n        return x\n\n\nclass PositionGetter(object):\n    \"\"\"return positions of patches\"\"\"\n\n    def __init__(self):\n        self.cache_positions = {}\n\n    def __call__(self, b, h, w, device):\n        if not (h, w) in self.cache_positions:\n            x = torch.arange(w, device=device)\n            y = torch.arange(h, device=device)\n            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)\n        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()\n        return pos\n\n\nclass PatchEmbed(nn.Module):\n    \"\"\"just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed\"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.blocks._init_weights","uri":"program://Human3R/function/src.dust3r.blocks._init_weights#L498-L500","kind":"function","name":"_init_weights","path":"src/dust3r/blocks.py","language":"python","start_line":498,"end_line":500,"context_start_line":478,"context_end_line":520,"code":"\n        self.position_getter = PositionGetter()\n\n    def forward(self, x):\n        B, C, H, W = x.shape\n        torch._assert(\n            H == self.img_size[0],\n            f\"Input image height ({H}) doesn't match model ({self.img_size[0]}).\",\n        )\n        torch._assert(\n            W == self.img_size[1],\n            f\"Input image width ({W}) doesn't match model ({self.img_size[1]}).\",\n        )\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n    def _init_weights(self):\n        w = self.proj.weight.data\n        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))\n\n\nif __name__ == \"__main__\":\n    import os\n    import sys\n\n    sys.path.append(os.path.dirname(os.path.dirname(__file__)))\n    import dust3r.utils.path_to_croco\n    from models.pos_embed import get_2d_sincos_pos_embed, RoPE2D\n    from functools import partial\n    from torch.utils.checkpoint import checkpoint\n\n    torch.manual_seed(0)\n\n    enc_blocks_ray_map = (\n        nn.ModuleList(\n            [\n                Block(\n                    768,\n                    16,","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses","uri":"program://Human3R/module/src.dust3r.losses#L1-L1649","kind":"module","name":"src.dust3r.losses","path":"src/dust3r/losses.py","language":"python","start_line":1,"end_line":1649,"context_start_line":1,"context_end_line":1649,"code":"from copy import copy, deepcopy\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom dust3r.utils.geometry import (\n    inv,\n    geotrf,\n    normalize_pointcloud_group,\n    get_group_pointcloud_center_scale,\n    to_euclidean_dist,\n)\nimport numpy as np\nfrom dust3r.utils.camera import (\n    pose_encoding_to_camera,\n    camera_to_pose_encoding,\n    relative_pose_absT_quatR,\n)\nfrom dust3r.utils.image import unpad_image\nfrom dust3r.utils import SMPL_Layer\n\nimport roma\nfrom tqdm import tqdm\n\ndef Sum(*losses_and_masks):\n    loss, mask = losses_and_masks[0]\n    if loss.ndim > 0:\n        # we are actually returning the loss for every pixels\n        return losses_and_masks\n    else:\n        # we are returning the global loss\n        for loss2, mask2 in losses_and_masks[1:]:\n            loss = loss + loss2\n        return loss\n\n\ndef stack_view(ls, k=None):\n    if isinstance(ls[0], dict):\n        v = torch.stack([g[k] for g in ls], dim=0)\n    else:\n        v = torch.stack(ls, dim=0)\n    return v.view(-1, *v.shape[2:])\n\n\ndef _neg_loss(pred, gt):\n  '''\n  Code modified from: https://github.com/xingyizhou/CenterNet/blob/4c50fd3a46bdf63dbf2082c5cbb3458d39579e6c/src/lib/models/losses.py#L42\n    Modified focal loss. Exactly the same as CornerNet.\n      Runs faster and costs a little bit more memory\n    Arguments:\n      pred (batch x c x h x w)\n      gt_regr (batch x c x h x w)\n    \n  '''\n  assert pred.shape == gt.shape\n\n  pos_inds = gt.eq(1).float()\n  neg_inds = gt.lt(1).float()\n\n  neg_weights = torch.pow(1 - gt, 4)\n\n  loss = 0\n\n  eps = 1e-7\n\n  pos_loss = torch.log(pred + eps) * torch.pow(1 - pred, 2) * pos_inds\n  neg_loss = torch.log(1 - pred + eps) * torch.pow(pred, 2) * neg_weights * neg_inds\n\n  num_pos  = pos_inds.float().sum()\n  pos_loss = pos_loss.sum()\n  neg_loss = neg_loss.sum()\n\n  if num_pos == 0:\n    loss = loss - neg_loss\n  else:\n    loss = loss - (pos_loss + neg_loss) / num_pos\n  return loss\n\n\nclass BaseCriterion(nn.Module):\n    def __init__(self, reduction=\"mean\"):\n        super().__init__()\n        self.reduction = reduction\n\n\nclass LLoss(BaseCriterion):\n    \"\"\"L-norm loss\"\"\"\n\n    def forward(self, a, b):\n        # assert (\n        #     a.shape == b.shape and a.ndim >= 2 and 1 <= a.shape[-1] <= 3\n        # ), f\"Bad shape = {a.shape}\"\n        dist = self.distance(a, b)\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def distance(self, a, b):\n        raise NotImplementedError()\n\n\nclass L21Loss(LLoss):\n    \"\"\"Euclidean distance between 3d points\"\"\"\n\n    def distance(self, a, b):\n        return torch.norm(a - b, dim=-1)  # normalized L2 distance\n\n\nL21 = L21Loss()\n\n\nclass MSELoss(LLoss):\n    def distance(self, a, b):\n        return (a - b) ** 2\n\n\nMSE = MSELoss()\n\n\nclass L1Loss(LLoss):\n    def distance(self, a, b):\n        return (a - b).abs()\n\nL1 = L1Loss()\n\n\nclass Criterion(nn.Module):\n    def __init__(self, criterion=None):\n        super().__init__()\n        assert isinstance(\n            criterion, BaseCriterion\n        ), f\"{criterion} is not a proper criterion!\"\n        self.criterion = copy(criterion)\n\n    def get_name(self):\n        return f\"{type(self).__name__}({self.criterion})\"\n\n    def with_reduction(self, mode=\"none\"):\n        res = loss = deepcopy(self)\n        while loss is not None:\n            assert isinstance(loss, Criterion)\n            loss.criterion.reduction = mode  # make it return the loss for each sample\n            loss = loss._loss2  # we assume loss is a Multiloss\n        return res\n\n\nclass MultiLoss(nn.Module):\n    \"\"\"Easily combinable losses (also keep track of individual loss values):\n        loss = MyLoss1() + 0.1*MyLoss2()\n    Usage:\n        Inherit from this class and override get_name() and compute_loss()\n    \"\"\"\n\n    def __init__(self):\n        super().__init__()\n        self._alpha = 1\n        self._loss2 = None\n\n    def compute_loss(self, *args, **kwargs):\n        raise NotImplementedError()\n\n    def get_name(self):\n        raise NotImplementedError()\n\n    def __mul__(self, alpha):\n        assert isinstance(alpha, (int, float))\n        res = copy(self)\n        res._alpha = alpha\n        return res\n\n    __rmul__ = __mul__  # same\n\n    def __add__(self, loss2):\n        assert isinstance(loss2, MultiLoss)\n        res = cur = copy(self)\n        # find the end of the chain\n        while cur._loss2 is not None:\n            cur = cur._loss2\n        cur._loss2 = loss2\n        return res\n\n    def __repr__(self):\n        name = self.get_name()\n        if self._alpha != 1:\n            name = f\"{self._alpha:g}*{name}\"\n        if self._loss2:\n            name = f\"{name} + {self._loss2}\"\n        return name\n\n    def forward(self, *args, **kwargs):\n        loss = self.compute_loss(*args, **kwargs)\n        if isinstance(loss, tuple):\n            loss, details = loss\n        elif loss.ndim == 0:\n            details = {self.get_name(): float(loss)}\n        else:\n            details = {}\n        loss = loss * self._alpha\n\n        if self._loss2:\n            loss2, details2 = self._loss2(*args, **kwargs)\n            loss = loss + loss2\n            details |= details2\n\n        return loss, details\n\n\nclass SSIM(nn.Module):\n    \"\"\"Layer to compute the SSIM loss between a pair of images\"\"\"\n\n    def __init__(self):\n        super(SSIM, self).__init__()\n        self.mu_x_pool = nn.AvgPool2d(3, 1)\n        self.mu_y_pool = nn.AvgPool2d(3, 1)\n        self.sig_x_pool = nn.AvgPool2d(3, 1)\n        self.sig_y_pool = nn.AvgPool2d(3, 1)\n        self.sig_xy_pool = nn.AvgPool2d(3, 1)\n\n        self.refl = nn.ReflectionPad2d(1)\n\n        self.C1 = 0.01**2\n        self.C2 = 0.03**2\n\n    def forward(self, x, y):\n        x = self.refl(x)\n        y = self.refl(y)\n\n        mu_x = self.mu_x_pool(x)\n        mu_y = self.mu_y_pool(y)\n\n        sigma_x = self.sig_x_pool(x**2) - mu_x**2\n        sigma_y = self.sig_y_pool(y**2) - mu_y**2\n        sigma_xy = self.sig_xy_pool(x * y) - mu_x * mu_y\n\n        SSIM_n = (2 * mu_x * mu_y + self.C1) * (2 * sigma_xy + self.C2)\n        SSIM_d = (mu_x**2 + mu_y**2 + self.C1) * (sigma_x + sigma_y + self.C2)\n\n        return torch.clamp((1 - SSIM_n / SSIM_d) / 2, 0, 1)\n\n\nclass RGBLoss(Criterion, MultiLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        self.ssim = SSIM()\n\n    def img_loss(self, a, b):\n        return self.criterion(a, b)\n\n    def compute_loss(self, gts, preds, **kw):\n        gt_rgbs = [gt[\"img\"].permute(0, 2, 3, 1) for gt in gts]\n        pred_rgbs = [pred[\"rgb\"] for pred in preds]\n        ls = [\n            self.img_loss(pred_rgb, gt_rgb)\n            for pred_rgb, gt_rgb in zip(pred_rgbs, gt_rgbs)\n        ]\n        details = {}\n        self_name = type(self).__name__\n        for i, l in enumerate(ls):\n            details[self_name + f\"_rgb/{i+1}\"] = float(l)\n            details[f\"pred_rgb_{i+1}\"] = pred_rgbs[i]\n        rgb_loss = sum(ls) / len(ls)\n        return rgb_loss, details\n\n\nclass SMPLLoss(Criterion, MultiLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        scale = 0.01\n        self.alpha_msk = 100.0 * scale\n        self.alpha_bce = 10.0 * scale\n        self.alpha_rotmat = 100.0 * scale\n        self.alpha_shape = 10.0 * scale\n        self.alpha_transl = 100.0 * scale\n        self.alpha_j3d = 100.0 * scale\n        self.alpha_v3d = 100.0 * scale\n        self.alpha_j2d = 1.0 * scale\n        self.alpha_v2d = 1.0 * scale\n\n        # SMPL layer\n        person_center = 'head'\n        dict_smpl_layer = {\n            'neutral': {\n                10: SMPL_Layer(type='smplx', gender='neutral', num_betas=10, kid=False, person_center=person_center),\n                11: SMPL_Layer(type='smplx', gender='neutral', num_betas=11, kid=False, person_center=person_center),\n                }\n            }\n        _moduleDict = []\n        for k, _smpl_layer in dict_smpl_layer.items():\n            for x, y in _smpl_layer.items():\n                _moduleDict.append([f\"{k}_{x}\", deepcopy(y)])\n        self.smpl_layer = nn.ModuleDict(_moduleDict)\n\n    def get_name(self):\n        return \"SMPLLoss\"\n\n    def mask_loss(self, gts, preds, masks, ret_pred=False):\n        gt_msks = [gt[\"msk_mhmr\"].unsqueeze(-1) for gt in gts]\n        pred_msks = [pred[\"msk\"] for pred in preds]\n        ls = [\n            F.binary_cross_entropy(p[m], g[m])\n            for p, g, m in zip(pred_msks, gt_msks, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        for i, l in enumerate(ls):\n            details[self_name + f\"_msk/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_msk_{i+1}\"] = pred_msks[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n        \n    def bce(self, gts, preds, masks, ret_pred=False):\n        gt_scores = [(gt[\"smpl_scores\"] >= 1).to(int).unsqueeze(-1) for gt in gts]\n        pred_scores = [pred[\"smpl_scores\"] for pred in preds]\n        ls = [\n            _neg_loss(p[m], g[m])\n            for p, g, m in zip(pred_scores, gt_scores, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        for i, l in enumerate(ls):\n            details[self_name + f\"_scores/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_smpl_scores_{i+1}\"] = pred_scores[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n    \n    def smpl_param_loss(self, gts, preds, masks, k, ret_pred=False):\n        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        ls = [\n            self.criterion(p[m], g[m])\n            for p, g, m in zip(preds, gts, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def point3d_loss(self, gts, preds, gt_t_p, pr_t_ps, masks, k, ret_pred=False):\n        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        if gt_t_p is None or pr_t_ps is None:\n            ls = [\n                self.criterion(p[m], g[m])\n                for p, g, m in zip(preds, gts, masks)\n            ]\n        else:\n            ls = [\n                self.criterion(p[m]-pr_p[m], g[m]-gt_p[m])\n                for p, pr_p, g, gt_p, m in zip(preds, pr_t_ps, gts, gt_t_p, masks)\n            ]\n\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        if gt_t_p is None or pr_t_ps is None:\n            k_name = \"c\" + k_name\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def point2d_loss(self, gts, preds, masks, k, shape=None, ret_pred=False):\n        if isinstance(gts[0], dict):\n            shape = gts[0]['true_shape'][0]\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        valid_mask = [\n            ((gt[..., 0] > 0) & (gt[..., 0] < shape[1]) & (gt[..., 1] > 0) & (gt[..., 1] < shape[0]\n            )) for gt in gts]\n\n        ls = [\n            self.criterion(p[m1.unsqueeze(-1) & m2], g[m1.unsqueeze(-1) & m2])\n            for p, g, m1, m2 in zip(preds, gts, masks, valid_mask)\n        ]\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def compute_loss(self, gts, preds, **kw):\n        img_mask_list = [gt[\"img_mask\"] for gt in gts]\n        smpl_mask_list = [gt[\"smpl_mask\"] for gt in gts]\n        masks_list = [a.unsqueeze(1) & b for a, b in zip(img_mask_list, smpl_mask_list)]\n\n        # Detection loss\n        score_loss, score_details = self.bce(gts, preds, img_mask_list, ret_pred=True)\n\n        has_msk = \"msk\" in preds[0]\n        if has_msk:\n            msk_loss, msk_details = self.mask_loss(gts, preds, img_mask_list, ret_pred=True)\n\n        K = stack_view(gts, 'camera_intrinsics')\n        img_mask = stack_view(img_mask_list,'img_mask').unsqueeze(1)\n        smpl_mask = stack_view(smpl_mask_list, 'smpl_mask') * img_mask\n        idx_h = torch.where(smpl_mask)\n        if int(smpl_mask.sum()) == 0:\n            total_loss = self.alpha_bce * score_loss\n            details = {\n                **score_details,\n            }\n            if has_msk:\n                total_loss += self.alpha_msk * msk_loss\n                details.update(msk_details)\n            return total_loss, details\n    \n        # Prediction\n        pred_rotmat = stack_view(preds, 'smpl_rotmat')\n        pred_rotvec = roma.rotmat_to_rotvec(pred_rotmat[smpl_mask])\n        pred_shape = stack_view(preds, 'smpl_shape')\n        pred_transl = [pred.pop(\"smpl_transl\") for pred in preds]\n        pred_transl = stack_view(pred_transl, 'smpl_transl')\n        pred_expression = stack_view(preds, 'smpl_expression')\n        \n        smpl_out = self.smpl_layer[f\"neutral_{pred_shape.shape[-1]}\"](\n            pred_rotvec, \n            pred_shape[smpl_mask], \n            pred_transl[smpl_mask], \n            None, None, \n            K=K[idx_h[0]], \n            expression=pred_expression[smpl_mask])\n        \n        pred_smpl = {}\n        batch_size = img_mask_list[0].shape[0]\n        num_view = len(gts)\n        max_humans = smpl_mask.shape[1]\n\n        for k, v in smpl_out.items():\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, *v.shape[1:], \n                device=v.device, dtype=v.dtype,\n            )\n            full_out[smpl_mask] = v\n            pred_smpl[k] = full_out.chunk(num_view, dim=0)\n\n        # SMPL-X params\n        rotmat_loss, rotmat_details = self.smpl_param_loss(gts, preds, masks_list, \"smpl_rotmat\")\n        transl_loss, transl_details = self.smpl_param_loss(gts, pred_smpl['smpl_transl'], masks_list, \"smpl_transl\")\n        shape_dim = min([gts[0]['smpl_shape'].shape[-1], preds[0]['smpl_shape'].shape[-1]])\n        gt_shape = [gt['smpl_shape'][...,:shape_dim] for gt in gts]\n        pred_shape = [pred['smpl_shape'][...,:shape_dim] for pred in preds]\n        shape_loss, shape_details = self.smpl_param_loss(gt_shape, pred_shape, masks_list, \"smpl_shape\")\n\n        # 3D points\n        gt_transl_pelvis= [gt['smpl_transl_pelvis'][..., None, :] for gt in gts]\n        pred_transl_pelvis = pred_smpl['smpl_transl_pelvis']\n        j3d_loss, j3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_j3d'], \n                                    gt_transl_pelvis, \n                                    pred_transl_pelvis, \n                                    masks_list, \"smpl_j3d\")\n        v3d_loss, v3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_v3d'], \n                                    gt_transl_pelvis, \n                                    pred_transl_pelvis, \n                                    masks_list, \"smpl_v3d\",\n                                    ret_pred=True)\n        cj3d_loss, cj3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_j3d'], \n                                    None, \n                                    None, \n                                    masks_list, \"smpl_j3d\")\n        cv3d_loss, cv3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_v3d'], \n                                    None, \n                                    None, \n                                    masks_list, \"smpl_v3d\")\n        \n        # total loss\n        total_loss = self.alpha_bce * score_loss +\\\n                     self.alpha_rotmat * rotmat_loss +\\\n                     self.alpha_shape * shape_loss +\\\n                     self.alpha_transl * transl_loss +\\\n                     self.alpha_j3d * j3d_loss +\\\n                     self.alpha_v3d * v3d_loss +\\\n                     self.alpha_j3d * cj3d_loss +\\\n                     self.alpha_v3d * cv3d_loss\n\n        details = {\n            **score_details,\n            **rotmat_details,\n            **transl_details,\n            **shape_details,\n            **j3d_details,\n            **v3d_details,\n            **cj3d_details,\n            **cv3d_details,\n        }\n\n        if has_msk:\n            total_loss += self.alpha_msk * msk_loss\n            details.update(msk_details)\n\n        if cv3d_loss < 1.0:\n            # 2D reprojection\n            j2d_loss, j2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_j2d'], \n                                        masks_list, \"smpl_j2d\")\n            v2d_loss, v2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_v2d'], \n                                        masks_list, \"smpl_v2d\")\n            total_loss += self.alpha_j2d * j2d_loss +\\\n                          self.alpha_v2d * v2d_loss\n            details.update({\n                **j2d_details,\n                **v2d_details\n            })\n\n        return total_loss, details\n\nclass NaiveSMPLLoss(SMPLLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        \n    def compute_loss(self, gts, preds, **kw):\n        img_mask_list = [gt[\"img_mask\"] for gt in gts]\n        smpl_mask_list = [gt[\"smpl_mask\"] for gt in gts]\n        masks_list = [a.unsqueeze(1) & b for a, b in zip(img_mask_list, smpl_mask_list)]\n\n        # Detection loss\n        score_loss, score_details = self.bce(gts, preds, img_mask_list, ret_pred=True)\n\n        # only for inference SMPL model\n        K = stack_view(gts, 'camera_intrinsics')\n        img_mask = stack_view(img_mask_list,'img_mask').unsqueeze(1)\n        smpl_mask = stack_view(smpl_mask_list, 'smpl_mask') * img_mask\n        idx_h = torch.where(smpl_mask)\n        if int(smpl_mask.sum()) == 0:\n            total_loss = self.alpha_bce * score_loss\n            details = {\n                **score_details,\n            }\n            return total_loss, details\n    \n        # Prediction\n        pred_rotmat = stack_view(preds, 'smpl_rotmat')\n        pred_rotvec = roma.rotmat_to_rotvec(pred_rotmat[smpl_mask])\n        pred_shape = stack_view(preds, 'smpl_shape')\n        pred_transl = [pred.pop(\"smpl_transl\") for pred in preds]\n        pred_transl = stack_view(pred_transl, 'smpl_transl')\n        pred_expression = stack_view(preds, 'smpl_expression')\n        \n        # Neutral for MHMR\n        K_mhmr = stack_view(gts, 'K_mhmr')\n        mhmr_img_res = gts[0][\"img_mhmr\"].shape[-1]\n        # fine head uv\n        pred_loc = stack_view(preds, 'smpl_loc')\n        # Distance \n        dist = pred_transl[smpl_mask][:, 0].unsqueeze(-1)\n        dist = to_euclidean_dist(mhmr_img_res, dist, K_mhmr[idx_h[0]])  # use K GT\n        smpl_out = self.smpl_layer[f\"neutral_{pred_shape.shape[-1]}\"](\n            pred_rotvec, \n            pred_shape[smpl_mask], \n            None, \n            pred_loc[smpl_mask], \n            dist,\n# ... truncated ...","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.Sum","uri":"program://Human3R/function/src.dust3r.losses.Sum#L25-L34","kind":"function","name":"Sum","path":"src/dust3r/losses.py","language":"python","start_line":25,"end_line":34,"context_start_line":5,"context_end_line":54,"code":"\nfrom dust3r.utils.geometry import (\n    inv,\n    geotrf,\n    normalize_pointcloud_group,\n    get_group_pointcloud_center_scale,\n    to_euclidean_dist,\n)\nimport numpy as np\nfrom dust3r.utils.camera import (\n    pose_encoding_to_camera,\n    camera_to_pose_encoding,\n    relative_pose_absT_quatR,\n)\nfrom dust3r.utils.image import unpad_image\nfrom dust3r.utils import SMPL_Layer\n\nimport roma\nfrom tqdm import tqdm\n\ndef Sum(*losses_and_masks):\n    loss, mask = losses_and_masks[0]\n    if loss.ndim > 0:\n        # we are actually returning the loss for every pixels\n        return losses_and_masks\n    else:\n        # we are returning the global loss\n        for loss2, mask2 in losses_and_masks[1:]:\n            loss = loss + loss2\n        return loss\n\n\ndef stack_view(ls, k=None):\n    if isinstance(ls[0], dict):\n        v = torch.stack([g[k] for g in ls], dim=0)\n    else:\n        v = torch.stack(ls, dim=0)\n    return v.view(-1, *v.shape[2:])\n\n\ndef _neg_loss(pred, gt):\n  '''\n  Code modified from: https://github.com/xingyizhou/CenterNet/blob/4c50fd3a46bdf63dbf2082c5cbb3458d39579e6c/src/lib/models/losses.py#L42\n    Modified focal loss. Exactly the same as CornerNet.\n      Runs faster and costs a little bit more memory\n    Arguments:\n      pred (batch x c x h x w)\n      gt_regr (batch x c x h x w)\n    \n  '''","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.stack_view","uri":"program://Human3R/function/src.dust3r.losses.stack_view#L37-L42","kind":"function","name":"stack_view","path":"src/dust3r/losses.py","language":"python","start_line":37,"end_line":42,"context_start_line":17,"context_end_line":62,"code":"    relative_pose_absT_quatR,\n)\nfrom dust3r.utils.image import unpad_image\nfrom dust3r.utils import SMPL_Layer\n\nimport roma\nfrom tqdm import tqdm\n\ndef Sum(*losses_and_masks):\n    loss, mask = losses_and_masks[0]\n    if loss.ndim > 0:\n        # we are actually returning the loss for every pixels\n        return losses_and_masks\n    else:\n        # we are returning the global loss\n        for loss2, mask2 in losses_and_masks[1:]:\n            loss = loss + loss2\n        return loss\n\n\ndef stack_view(ls, k=None):\n    if isinstance(ls[0], dict):\n        v = torch.stack([g[k] for g in ls], dim=0)\n    else:\n        v = torch.stack(ls, dim=0)\n    return v.view(-1, *v.shape[2:])\n\n\ndef _neg_loss(pred, gt):\n  '''\n  Code modified from: https://github.com/xingyizhou/CenterNet/blob/4c50fd3a46bdf63dbf2082c5cbb3458d39579e6c/src/lib/models/losses.py#L42\n    Modified focal loss. Exactly the same as CornerNet.\n      Runs faster and costs a little bit more memory\n    Arguments:\n      pred (batch x c x h x w)\n      gt_regr (batch x c x h x w)\n    \n  '''\n  assert pred.shape == gt.shape\n\n  pos_inds = gt.eq(1).float()\n  neg_inds = gt.lt(1).float()\n\n  neg_weights = torch.pow(1 - gt, 4)\n\n  loss = 0","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses._neg_loss","uri":"program://Human3R/function/src.dust3r.losses._neg_loss#L45-L77","kind":"function","name":"_neg_loss","path":"src/dust3r/losses.py","language":"python","start_line":45,"end_line":77,"context_start_line":25,"context_end_line":97,"code":"def Sum(*losses_and_masks):\n    loss, mask = losses_and_masks[0]\n    if loss.ndim > 0:\n        # we are actually returning the loss for every pixels\n        return losses_and_masks\n    else:\n        # we are returning the global loss\n        for loss2, mask2 in losses_and_masks[1:]:\n            loss = loss + loss2\n        return loss\n\n\ndef stack_view(ls, k=None):\n    if isinstance(ls[0], dict):\n        v = torch.stack([g[k] for g in ls], dim=0)\n    else:\n        v = torch.stack(ls, dim=0)\n    return v.view(-1, *v.shape[2:])\n\n\ndef _neg_loss(pred, gt):\n  '''\n  Code modified from: https://github.com/xingyizhou/CenterNet/blob/4c50fd3a46bdf63dbf2082c5cbb3458d39579e6c/src/lib/models/losses.py#L42\n    Modified focal loss. Exactly the same as CornerNet.\n      Runs faster and costs a little bit more memory\n    Arguments:\n      pred (batch x c x h x w)\n      gt_regr (batch x c x h x w)\n    \n  '''\n  assert pred.shape == gt.shape\n\n  pos_inds = gt.eq(1).float()\n  neg_inds = gt.lt(1).float()\n\n  neg_weights = torch.pow(1 - gt, 4)\n\n  loss = 0\n\n  eps = 1e-7\n\n  pos_loss = torch.log(pred + eps) * torch.pow(1 - pred, 2) * pos_inds\n  neg_loss = torch.log(1 - pred + eps) * torch.pow(pred, 2) * neg_weights * neg_inds\n\n  num_pos  = pos_inds.float().sum()\n  pos_loss = pos_loss.sum()\n  neg_loss = neg_loss.sum()\n\n  if num_pos == 0:\n    loss = loss - neg_loss\n  else:\n    loss = loss - (pos_loss + neg_loss) / num_pos\n  return loss\n\n\nclass BaseCriterion(nn.Module):\n    def __init__(self, reduction=\"mean\"):\n        super().__init__()\n        self.reduction = reduction\n\n\nclass LLoss(BaseCriterion):\n    \"\"\"L-norm loss\"\"\"\n\n    def forward(self, a, b):\n        # assert (\n        #     a.shape == b.shape and a.ndim >= 2 and 1 <= a.shape[-1] <= 3\n        # ), f\"Bad shape = {a.shape}\"\n        dist = self.distance(a, b)\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.BaseCriterion","uri":"program://Human3R/class/src.dust3r.losses.BaseCriterion#L80-L83","kind":"class","name":"BaseCriterion","path":"src/dust3r/losses.py","language":"python","start_line":80,"end_line":83,"context_start_line":60,"context_end_line":103,"code":"  neg_weights = torch.pow(1 - gt, 4)\n\n  loss = 0\n\n  eps = 1e-7\n\n  pos_loss = torch.log(pred + eps) * torch.pow(1 - pred, 2) * pos_inds\n  neg_loss = torch.log(1 - pred + eps) * torch.pow(pred, 2) * neg_weights * neg_inds\n\n  num_pos  = pos_inds.float().sum()\n  pos_loss = pos_loss.sum()\n  neg_loss = neg_loss.sum()\n\n  if num_pos == 0:\n    loss = loss - neg_loss\n  else:\n    loss = loss - (pos_loss + neg_loss) / num_pos\n  return loss\n\n\nclass BaseCriterion(nn.Module):\n    def __init__(self, reduction=\"mean\"):\n        super().__init__()\n        self.reduction = reduction\n\n\nclass LLoss(BaseCriterion):\n    \"\"\"L-norm loss\"\"\"\n\n    def forward(self, a, b):\n        # assert (\n        #     a.shape == b.shape and a.ndim >= 2 and 1 <= a.shape[-1] <= 3\n        # ), f\"Bad shape = {a.shape}\"\n        dist = self.distance(a, b)\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def distance(self, a, b):\n        raise NotImplementedError()","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.LLoss","uri":"program://Human3R/class/src.dust3r.losses.LLoss#L86-L103","kind":"class","name":"LLoss","path":"src/dust3r/losses.py","language":"python","start_line":86,"end_line":103,"context_start_line":66,"context_end_line":123,"code":"  pos_loss = torch.log(pred + eps) * torch.pow(1 - pred, 2) * pos_inds\n  neg_loss = torch.log(1 - pred + eps) * torch.pow(pred, 2) * neg_weights * neg_inds\n\n  num_pos  = pos_inds.float().sum()\n  pos_loss = pos_loss.sum()\n  neg_loss = neg_loss.sum()\n\n  if num_pos == 0:\n    loss = loss - neg_loss\n  else:\n    loss = loss - (pos_loss + neg_loss) / num_pos\n  return loss\n\n\nclass BaseCriterion(nn.Module):\n    def __init__(self, reduction=\"mean\"):\n        super().__init__()\n        self.reduction = reduction\n\n\nclass LLoss(BaseCriterion):\n    \"\"\"L-norm loss\"\"\"\n\n    def forward(self, a, b):\n        # assert (\n        #     a.shape == b.shape and a.ndim >= 2 and 1 <= a.shape[-1] <= 3\n        # ), f\"Bad shape = {a.shape}\"\n        dist = self.distance(a, b)\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def distance(self, a, b):\n        raise NotImplementedError()\n\n\nclass L21Loss(LLoss):\n    \"\"\"Euclidean distance between 3d points\"\"\"\n\n    def distance(self, a, b):\n        return torch.norm(a - b, dim=-1)  # normalized L2 distance\n\n\nL21 = L21Loss()\n\n\nclass MSELoss(LLoss):\n    def distance(self, a, b):\n        return (a - b) ** 2\n\n\nMSE = MSELoss()\n\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.L21Loss","uri":"program://Human3R/class/src.dust3r.losses.L21Loss#L106-L110","kind":"class","name":"L21Loss","path":"src/dust3r/losses.py","language":"python","start_line":106,"end_line":110,"context_start_line":86,"context_end_line":130,"code":"class LLoss(BaseCriterion):\n    \"\"\"L-norm loss\"\"\"\n\n    def forward(self, a, b):\n        # assert (\n        #     a.shape == b.shape and a.ndim >= 2 and 1 <= a.shape[-1] <= 3\n        # ), f\"Bad shape = {a.shape}\"\n        dist = self.distance(a, b)\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def distance(self, a, b):\n        raise NotImplementedError()\n\n\nclass L21Loss(LLoss):\n    \"\"\"Euclidean distance between 3d points\"\"\"\n\n    def distance(self, a, b):\n        return torch.norm(a - b, dim=-1)  # normalized L2 distance\n\n\nL21 = L21Loss()\n\n\nclass MSELoss(LLoss):\n    def distance(self, a, b):\n        return (a - b) ** 2\n\n\nMSE = MSELoss()\n\n\nclass L1Loss(LLoss):\n    def distance(self, a, b):\n        return (a - b).abs()\n\nL1 = L1Loss()\n\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.MSELoss","uri":"program://Human3R/class/src.dust3r.losses.MSELoss#L116-L118","kind":"class","name":"MSELoss","path":"src/dust3r/losses.py","language":"python","start_line":116,"end_line":118,"context_start_line":96,"context_end_line":138,"code":"        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def distance(self, a, b):\n        raise NotImplementedError()\n\n\nclass L21Loss(LLoss):\n    \"\"\"Euclidean distance between 3d points\"\"\"\n\n    def distance(self, a, b):\n        return torch.norm(a - b, dim=-1)  # normalized L2 distance\n\n\nL21 = L21Loss()\n\n\nclass MSELoss(LLoss):\n    def distance(self, a, b):\n        return (a - b) ** 2\n\n\nMSE = MSELoss()\n\n\nclass L1Loss(LLoss):\n    def distance(self, a, b):\n        return (a - b).abs()\n\nL1 = L1Loss()\n\n\nclass Criterion(nn.Module):\n    def __init__(self, criterion=None):\n        super().__init__()\n        assert isinstance(\n            criterion, BaseCriterion\n        ), f\"{criterion} is not a proper criterion!\"\n        self.criterion = copy(criterion)\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.L1Loss","uri":"program://Human3R/class/src.dust3r.losses.L1Loss#L124-L126","kind":"class","name":"L1Loss","path":"src/dust3r/losses.py","language":"python","start_line":124,"end_line":126,"context_start_line":104,"context_end_line":146,"code":"\n\nclass L21Loss(LLoss):\n    \"\"\"Euclidean distance between 3d points\"\"\"\n\n    def distance(self, a, b):\n        return torch.norm(a - b, dim=-1)  # normalized L2 distance\n\n\nL21 = L21Loss()\n\n\nclass MSELoss(LLoss):\n    def distance(self, a, b):\n        return (a - b) ** 2\n\n\nMSE = MSELoss()\n\n\nclass L1Loss(LLoss):\n    def distance(self, a, b):\n        return (a - b).abs()\n\nL1 = L1Loss()\n\n\nclass Criterion(nn.Module):\n    def __init__(self, criterion=None):\n        super().__init__()\n        assert isinstance(\n            criterion, BaseCriterion\n        ), f\"{criterion} is not a proper criterion!\"\n        self.criterion = copy(criterion)\n\n    def get_name(self):\n        return f\"{type(self).__name__}({self.criterion})\"\n\n    def with_reduction(self, mode=\"none\"):\n        res = loss = deepcopy(self)\n        while loss is not None:\n            assert isinstance(loss, Criterion)\n            loss.criterion.reduction = mode  # make it return the loss for each sample","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.Criterion","uri":"program://Human3R/class/src.dust3r.losses.Criterion#L131-L148","kind":"class","name":"Criterion","path":"src/dust3r/losses.py","language":"python","start_line":131,"end_line":148,"context_start_line":111,"context_end_line":168,"code":"\n\nL21 = L21Loss()\n\n\nclass MSELoss(LLoss):\n    def distance(self, a, b):\n        return (a - b) ** 2\n\n\nMSE = MSELoss()\n\n\nclass L1Loss(LLoss):\n    def distance(self, a, b):\n        return (a - b).abs()\n\nL1 = L1Loss()\n\n\nclass Criterion(nn.Module):\n    def __init__(self, criterion=None):\n        super().__init__()\n        assert isinstance(\n            criterion, BaseCriterion\n        ), f\"{criterion} is not a proper criterion!\"\n        self.criterion = copy(criterion)\n\n    def get_name(self):\n        return f\"{type(self).__name__}({self.criterion})\"\n\n    def with_reduction(self, mode=\"none\"):\n        res = loss = deepcopy(self)\n        while loss is not None:\n            assert isinstance(loss, Criterion)\n            loss.criterion.reduction = mode  # make it return the loss for each sample\n            loss = loss._loss2  # we assume loss is a Multiloss\n        return res\n\n\nclass MultiLoss(nn.Module):\n    \"\"\"Easily combinable losses (also keep track of individual loss values):\n        loss = MyLoss1() + 0.1*MyLoss2()\n    Usage:\n        Inherit from this class and override get_name() and compute_loss()\n    \"\"\"\n\n    def __init__(self):\n        super().__init__()\n        self._alpha = 1\n        self._loss2 = None\n\n    def compute_loss(self, *args, **kwargs):\n        raise NotImplementedError()\n\n    def get_name(self):\n        raise NotImplementedError()\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.MultiLoss","uri":"program://Human3R/class/src.dust3r.losses.MultiLoss#L151-L209","kind":"class","name":"MultiLoss","path":"src/dust3r/losses.py","language":"python","start_line":151,"end_line":209,"context_start_line":131,"context_end_line":229,"code":"class Criterion(nn.Module):\n    def __init__(self, criterion=None):\n        super().__init__()\n        assert isinstance(\n            criterion, BaseCriterion\n        ), f\"{criterion} is not a proper criterion!\"\n        self.criterion = copy(criterion)\n\n    def get_name(self):\n        return f\"{type(self).__name__}({self.criterion})\"\n\n    def with_reduction(self, mode=\"none\"):\n        res = loss = deepcopy(self)\n        while loss is not None:\n            assert isinstance(loss, Criterion)\n            loss.criterion.reduction = mode  # make it return the loss for each sample\n            loss = loss._loss2  # we assume loss is a Multiloss\n        return res\n\n\nclass MultiLoss(nn.Module):\n    \"\"\"Easily combinable losses (also keep track of individual loss values):\n        loss = MyLoss1() + 0.1*MyLoss2()\n    Usage:\n        Inherit from this class and override get_name() and compute_loss()\n    \"\"\"\n\n    def __init__(self):\n        super().__init__()\n        self._alpha = 1\n        self._loss2 = None\n\n    def compute_loss(self, *args, **kwargs):\n        raise NotImplementedError()\n\n    def get_name(self):\n        raise NotImplementedError()\n\n    def __mul__(self, alpha):\n        assert isinstance(alpha, (int, float))\n        res = copy(self)\n        res._alpha = alpha\n        return res\n\n    __rmul__ = __mul__  # same\n\n    def __add__(self, loss2):\n        assert isinstance(loss2, MultiLoss)\n        res = cur = copy(self)\n        # find the end of the chain\n        while cur._loss2 is not None:\n            cur = cur._loss2\n        cur._loss2 = loss2\n        return res\n\n    def __repr__(self):\n        name = self.get_name()\n        if self._alpha != 1:\n            name = f\"{self._alpha:g}*{name}\"\n        if self._loss2:\n            name = f\"{name} + {self._loss2}\"\n        return name\n\n    def forward(self, *args, **kwargs):\n        loss = self.compute_loss(*args, **kwargs)\n        if isinstance(loss, tuple):\n            loss, details = loss\n        elif loss.ndim == 0:\n            details = {self.get_name(): float(loss)}\n        else:\n            details = {}\n        loss = loss * self._alpha\n\n        if self._loss2:\n            loss2, details2 = self._loss2(*args, **kwargs)\n            loss = loss + loss2\n            details |= details2\n\n        return loss, details\n\n\nclass SSIM(nn.Module):\n    \"\"\"Layer to compute the SSIM loss between a pair of images\"\"\"\n\n    def __init__(self):\n        super(SSIM, self).__init__()\n        self.mu_x_pool = nn.AvgPool2d(3, 1)\n        self.mu_y_pool = nn.AvgPool2d(3, 1)\n        self.sig_x_pool = nn.AvgPool2d(3, 1)\n        self.sig_y_pool = nn.AvgPool2d(3, 1)\n        self.sig_xy_pool = nn.AvgPool2d(3, 1)\n\n        self.refl = nn.ReflectionPad2d(1)\n\n        self.C1 = 0.01**2\n        self.C2 = 0.03**2\n\n    def forward(self, x, y):\n        x = self.refl(x)","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.SSIM","uri":"program://Human3R/class/src.dust3r.losses.SSIM#L212-L242","kind":"class","name":"SSIM","path":"src/dust3r/losses.py","language":"python","start_line":212,"end_line":242,"context_start_line":192,"context_end_line":262,"code":"        return name\n\n    def forward(self, *args, **kwargs):\n        loss = self.compute_loss(*args, **kwargs)\n        if isinstance(loss, tuple):\n            loss, details = loss\n        elif loss.ndim == 0:\n            details = {self.get_name(): float(loss)}\n        else:\n            details = {}\n        loss = loss * self._alpha\n\n        if self._loss2:\n            loss2, details2 = self._loss2(*args, **kwargs)\n            loss = loss + loss2\n            details |= details2\n\n        return loss, details\n\n\nclass SSIM(nn.Module):\n    \"\"\"Layer to compute the SSIM loss between a pair of images\"\"\"\n\n    def __init__(self):\n        super(SSIM, self).__init__()\n        self.mu_x_pool = nn.AvgPool2d(3, 1)\n        self.mu_y_pool = nn.AvgPool2d(3, 1)\n        self.sig_x_pool = nn.AvgPool2d(3, 1)\n        self.sig_y_pool = nn.AvgPool2d(3, 1)\n        self.sig_xy_pool = nn.AvgPool2d(3, 1)\n\n        self.refl = nn.ReflectionPad2d(1)\n\n        self.C1 = 0.01**2\n        self.C2 = 0.03**2\n\n    def forward(self, x, y):\n        x = self.refl(x)\n        y = self.refl(y)\n\n        mu_x = self.mu_x_pool(x)\n        mu_y = self.mu_y_pool(y)\n\n        sigma_x = self.sig_x_pool(x**2) - mu_x**2\n        sigma_y = self.sig_y_pool(y**2) - mu_y**2\n        sigma_xy = self.sig_xy_pool(x * y) - mu_x * mu_y\n\n        SSIM_n = (2 * mu_x * mu_y + self.C1) * (2 * sigma_xy + self.C2)\n        SSIM_d = (mu_x**2 + mu_y**2 + self.C1) * (sigma_x + sigma_y + self.C2)\n\n        return torch.clamp((1 - SSIM_n / SSIM_d) / 2, 0, 1)\n\n\nclass RGBLoss(Criterion, MultiLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        self.ssim = SSIM()\n\n    def img_loss(self, a, b):\n        return self.criterion(a, b)\n\n    def compute_loss(self, gts, preds, **kw):\n        gt_rgbs = [gt[\"img\"].permute(0, 2, 3, 1) for gt in gts]\n        pred_rgbs = [pred[\"rgb\"] for pred in preds]\n        ls = [\n            self.img_loss(pred_rgb, gt_rgb)\n            for pred_rgb, gt_rgb in zip(pred_rgbs, gt_rgbs)\n        ]\n        details = {}\n        self_name = type(self).__name__\n        for i, l in enumerate(ls):","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.RGBLoss","uri":"program://Human3R/class/src.dust3r.losses.RGBLoss#L245-L266","kind":"class","name":"RGBLoss","path":"src/dust3r/losses.py","language":"python","start_line":245,"end_line":266,"context_start_line":225,"context_end_line":286,"code":"        self.C1 = 0.01**2\n        self.C2 = 0.03**2\n\n    def forward(self, x, y):\n        x = self.refl(x)\n        y = self.refl(y)\n\n        mu_x = self.mu_x_pool(x)\n        mu_y = self.mu_y_pool(y)\n\n        sigma_x = self.sig_x_pool(x**2) - mu_x**2\n        sigma_y = self.sig_y_pool(y**2) - mu_y**2\n        sigma_xy = self.sig_xy_pool(x * y) - mu_x * mu_y\n\n        SSIM_n = (2 * mu_x * mu_y + self.C1) * (2 * sigma_xy + self.C2)\n        SSIM_d = (mu_x**2 + mu_y**2 + self.C1) * (sigma_x + sigma_y + self.C2)\n\n        return torch.clamp((1 - SSIM_n / SSIM_d) / 2, 0, 1)\n\n\nclass RGBLoss(Criterion, MultiLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        self.ssim = SSIM()\n\n    def img_loss(self, a, b):\n        return self.criterion(a, b)\n\n    def compute_loss(self, gts, preds, **kw):\n        gt_rgbs = [gt[\"img\"].permute(0, 2, 3, 1) for gt in gts]\n        pred_rgbs = [pred[\"rgb\"] for pred in preds]\n        ls = [\n            self.img_loss(pred_rgb, gt_rgb)\n            for pred_rgb, gt_rgb in zip(pred_rgbs, gt_rgbs)\n        ]\n        details = {}\n        self_name = type(self).__name__\n        for i, l in enumerate(ls):\n            details[self_name + f\"_rgb/{i+1}\"] = float(l)\n            details[f\"pred_rgb_{i+1}\"] = pred_rgbs[i]\n        rgb_loss = sum(ls) / len(ls)\n        return rgb_loss, details\n\n\nclass SMPLLoss(Criterion, MultiLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        scale = 0.01\n        self.alpha_msk = 100.0 * scale\n        self.alpha_bce = 10.0 * scale\n        self.alpha_rotmat = 100.0 * scale\n        self.alpha_shape = 10.0 * scale\n        self.alpha_transl = 100.0 * scale\n        self.alpha_j3d = 100.0 * scale\n        self.alpha_v3d = 100.0 * scale\n        self.alpha_j2d = 1.0 * scale\n        self.alpha_v2d = 1.0 * scale\n\n        # SMPL layer\n        person_center = 'head'\n        dict_smpl_layer = {\n            'neutral': {","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.SMPLLoss","uri":"program://Human3R/class/src.dust3r.losses.SMPLLoss#L269-L543","kind":"class","name":"SMPLLoss","path":"src/dust3r/losses.py","language":"python","start_line":269,"end_line":543,"context_start_line":249,"context_end_line":563,"code":"\n    def img_loss(self, a, b):\n        return self.criterion(a, b)\n\n    def compute_loss(self, gts, preds, **kw):\n        gt_rgbs = [gt[\"img\"].permute(0, 2, 3, 1) for gt in gts]\n        pred_rgbs = [pred[\"rgb\"] for pred in preds]\n        ls = [\n            self.img_loss(pred_rgb, gt_rgb)\n            for pred_rgb, gt_rgb in zip(pred_rgbs, gt_rgbs)\n        ]\n        details = {}\n        self_name = type(self).__name__\n        for i, l in enumerate(ls):\n            details[self_name + f\"_rgb/{i+1}\"] = float(l)\n            details[f\"pred_rgb_{i+1}\"] = pred_rgbs[i]\n        rgb_loss = sum(ls) / len(ls)\n        return rgb_loss, details\n\n\nclass SMPLLoss(Criterion, MultiLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        scale = 0.01\n        self.alpha_msk = 100.0 * scale\n        self.alpha_bce = 10.0 * scale\n        self.alpha_rotmat = 100.0 * scale\n        self.alpha_shape = 10.0 * scale\n        self.alpha_transl = 100.0 * scale\n        self.alpha_j3d = 100.0 * scale\n        self.alpha_v3d = 100.0 * scale\n        self.alpha_j2d = 1.0 * scale\n        self.alpha_v2d = 1.0 * scale\n\n        # SMPL layer\n        person_center = 'head'\n        dict_smpl_layer = {\n            'neutral': {\n                10: SMPL_Layer(type='smplx', gender='neutral', num_betas=10, kid=False, person_center=person_center),\n                11: SMPL_Layer(type='smplx', gender='neutral', num_betas=11, kid=False, person_center=person_center),\n                }\n            }\n        _moduleDict = []\n        for k, _smpl_layer in dict_smpl_layer.items():\n            for x, y in _smpl_layer.items():\n                _moduleDict.append([f\"{k}_{x}\", deepcopy(y)])\n        self.smpl_layer = nn.ModuleDict(_moduleDict)\n\n    def get_name(self):\n        return \"SMPLLoss\"\n\n    def mask_loss(self, gts, preds, masks, ret_pred=False):\n        gt_msks = [gt[\"msk_mhmr\"].unsqueeze(-1) for gt in gts]\n        pred_msks = [pred[\"msk\"] for pred in preds]\n        ls = [\n            F.binary_cross_entropy(p[m], g[m])\n            for p, g, m in zip(pred_msks, gt_msks, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        for i, l in enumerate(ls):\n            details[self_name + f\"_msk/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_msk_{i+1}\"] = pred_msks[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n        \n    def bce(self, gts, preds, masks, ret_pred=False):\n        gt_scores = [(gt[\"smpl_scores\"] >= 1).to(int).unsqueeze(-1) for gt in gts]\n        pred_scores = [pred[\"smpl_scores\"] for pred in preds]\n        ls = [\n            _neg_loss(p[m], g[m])\n            for p, g, m in zip(pred_scores, gt_scores, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        for i, l in enumerate(ls):\n            details[self_name + f\"_scores/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_smpl_scores_{i+1}\"] = pred_scores[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n    \n    def smpl_param_loss(self, gts, preds, masks, k, ret_pred=False):\n        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        ls = [\n            self.criterion(p[m], g[m])\n            for p, g, m in zip(preds, gts, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def point3d_loss(self, gts, preds, gt_t_p, pr_t_ps, masks, k, ret_pred=False):\n        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        if gt_t_p is None or pr_t_ps is None:\n            ls = [\n                self.criterion(p[m], g[m])\n                for p, g, m in zip(preds, gts, masks)\n            ]\n        else:\n            ls = [\n                self.criterion(p[m]-pr_p[m], g[m]-gt_p[m])\n                for p, pr_p, g, gt_p, m in zip(preds, pr_t_ps, gts, gt_t_p, masks)\n            ]\n\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        if gt_t_p is None or pr_t_ps is None:\n            k_name = \"c\" + k_name\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def point2d_loss(self, gts, preds, masks, k, shape=None, ret_pred=False):\n        if isinstance(gts[0], dict):\n            shape = gts[0]['true_shape'][0]\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        valid_mask = [\n            ((gt[..., 0] > 0) & (gt[..., 0] < shape[1]) & (gt[..., 1] > 0) & (gt[..., 1] < shape[0]\n            )) for gt in gts]\n\n        ls = [\n            self.criterion(p[m1.unsqueeze(-1) & m2], g[m1.unsqueeze(-1) & m2])\n            for p, g, m1, m2 in zip(preds, gts, masks, valid_mask)\n        ]\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def compute_loss(self, gts, preds, **kw):\n        img_mask_list = [gt[\"img_mask\"] for gt in gts]\n        smpl_mask_list = [gt[\"smpl_mask\"] for gt in gts]\n        masks_list = [a.unsqueeze(1) & b for a, b in zip(img_mask_list, smpl_mask_list)]\n\n        # Detection loss\n        score_loss, score_details = self.bce(gts, preds, img_mask_list, ret_pred=True)\n\n        has_msk = \"msk\" in preds[0]\n        if has_msk:\n            msk_loss, msk_details = self.mask_loss(gts, preds, img_mask_list, ret_pred=True)\n\n        K = stack_view(gts, 'camera_intrinsics')\n        img_mask = stack_view(img_mask_list,'img_mask').unsqueeze(1)\n        smpl_mask = stack_view(smpl_mask_list, 'smpl_mask') * img_mask\n        idx_h = torch.where(smpl_mask)\n        if int(smpl_mask.sum()) == 0:\n            total_loss = self.alpha_bce * score_loss\n            details = {\n                **score_details,\n            }\n            if has_msk:\n                total_loss += self.alpha_msk * msk_loss\n                details.update(msk_details)\n            return total_loss, details\n    \n        # Prediction\n        pred_rotmat = stack_view(preds, 'smpl_rotmat')\n        pred_rotvec = roma.rotmat_to_rotvec(pred_rotmat[smpl_mask])\n        pred_shape = stack_view(preds, 'smpl_shape')\n        pred_transl = [pred.pop(\"smpl_transl\") for pred in preds]\n        pred_transl = stack_view(pred_transl, 'smpl_transl')\n        pred_expression = stack_view(preds, 'smpl_expression')\n        \n        smpl_out = self.smpl_layer[f\"neutral_{pred_shape.shape[-1]}\"](\n            pred_rotvec, \n            pred_shape[smpl_mask], \n            pred_transl[smpl_mask], \n            None, None, \n            K=K[idx_h[0]], \n            expression=pred_expression[smpl_mask])\n        \n        pred_smpl = {}\n        batch_size = img_mask_list[0].shape[0]\n        num_view = len(gts)\n        max_humans = smpl_mask.shape[1]\n\n        for k, v in smpl_out.items():\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, *v.shape[1:], \n                device=v.device, dtype=v.dtype,\n            )\n            full_out[smpl_mask] = v\n            pred_smpl[k] = full_out.chunk(num_view, dim=0)\n\n        # SMPL-X params\n        rotmat_loss, rotmat_details = self.smpl_param_loss(gts, preds, masks_list, \"smpl_rotmat\")\n        transl_loss, transl_details = self.smpl_param_loss(gts, pred_smpl['smpl_transl'], masks_list, \"smpl_transl\")\n        shape_dim = min([gts[0]['smpl_shape'].shape[-1], preds[0]['smpl_shape'].shape[-1]])\n        gt_shape = [gt['smpl_shape'][...,:shape_dim] for gt in gts]\n        pred_shape = [pred['smpl_shape'][...,:shape_dim] for pred in preds]\n        shape_loss, shape_details = self.smpl_param_loss(gt_shape, pred_shape, masks_list, \"smpl_shape\")\n\n        # 3D points\n        gt_transl_pelvis= [gt['smpl_transl_pelvis'][..., None, :] for gt in gts]\n        pred_transl_pelvis = pred_smpl['smpl_transl_pelvis']\n        j3d_loss, j3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_j3d'], \n                                    gt_transl_pelvis, \n                                    pred_transl_pelvis, \n                                    masks_list, \"smpl_j3d\")\n        v3d_loss, v3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_v3d'], \n                                    gt_transl_pelvis, \n                                    pred_transl_pelvis, \n                                    masks_list, \"smpl_v3d\",\n                                    ret_pred=True)\n        cj3d_loss, cj3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_j3d'], \n                                    None, \n                                    None, \n                                    masks_list, \"smpl_j3d\")\n        cv3d_loss, cv3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_v3d'], \n                                    None, \n                                    None, \n                                    masks_list, \"smpl_v3d\")\n        \n        # total loss\n        total_loss = self.alpha_bce * score_loss +\\\n                     self.alpha_rotmat * rotmat_loss +\\\n                     self.alpha_shape * shape_loss +\\\n                     self.alpha_transl * transl_loss +\\\n                     self.alpha_j3d * j3d_loss +\\\n                     self.alpha_v3d * v3d_loss +\\\n                     self.alpha_j3d * cj3d_loss +\\\n                     self.alpha_v3d * cv3d_loss\n\n        details = {\n            **score_details,\n            **rotmat_details,\n            **transl_details,\n            **shape_details,\n            **j3d_details,\n            **v3d_details,\n            **cj3d_details,\n            **cv3d_details,\n        }\n\n        if has_msk:\n            total_loss += self.alpha_msk * msk_loss\n            details.update(msk_details)\n\n        if cv3d_loss < 1.0:\n            # 2D reprojection\n            j2d_loss, j2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_j2d'], \n                                        masks_list, \"smpl_j2d\")\n            v2d_loss, v2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_v2d'], \n                                        masks_list, \"smpl_v2d\")\n            total_loss += self.alpha_j2d * j2d_loss +\\\n                          self.alpha_v2d * v2d_loss\n            details.update({\n                **j2d_details,\n                **v2d_details\n            })\n\n        return total_loss, details\n\nclass NaiveSMPLLoss(SMPLLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        \n    def compute_loss(self, gts, preds, **kw):\n        img_mask_list = [gt[\"img_mask\"] for gt in gts]\n        smpl_mask_list = [gt[\"smpl_mask\"] for gt in gts]\n        masks_list = [a.unsqueeze(1) & b for a, b in zip(img_mask_list, smpl_mask_list)]\n\n        # Detection loss\n        score_loss, score_details = self.bce(gts, preds, img_mask_list, ret_pred=True)\n\n        # only for inference SMPL model\n        K = stack_view(gts, 'camera_intrinsics')\n        img_mask = stack_view(img_mask_list,'img_mask').unsqueeze(1)\n        smpl_mask = stack_view(smpl_mask_list, 'smpl_mask') * img_mask\n        idx_h = torch.where(smpl_mask)\n        if int(smpl_mask.sum()) == 0:\n            total_loss = self.alpha_bce * score_loss","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.NaiveSMPLLoss","uri":"program://Human3R/class/src.dust3r.losses.NaiveSMPLLoss#L545-L684","kind":"class","name":"NaiveSMPLLoss","path":"src/dust3r/losses.py","language":"python","start_line":545,"end_line":684,"context_start_line":525,"context_end_line":704,"code":"\n        if cv3d_loss < 1.0:\n            # 2D reprojection\n            j2d_loss, j2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_j2d'], \n                                        masks_list, \"smpl_j2d\")\n            v2d_loss, v2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_v2d'], \n                                        masks_list, \"smpl_v2d\")\n            total_loss += self.alpha_j2d * j2d_loss +\\\n                          self.alpha_v2d * v2d_loss\n            details.update({\n                **j2d_details,\n                **v2d_details\n            })\n\n        return total_loss, details\n\nclass NaiveSMPLLoss(SMPLLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        \n    def compute_loss(self, gts, preds, **kw):\n        img_mask_list = [gt[\"img_mask\"] for gt in gts]\n        smpl_mask_list = [gt[\"smpl_mask\"] for gt in gts]\n        masks_list = [a.unsqueeze(1) & b for a, b in zip(img_mask_list, smpl_mask_list)]\n\n        # Detection loss\n        score_loss, score_details = self.bce(gts, preds, img_mask_list, ret_pred=True)\n\n        # only for inference SMPL model\n        K = stack_view(gts, 'camera_intrinsics')\n        img_mask = stack_view(img_mask_list,'img_mask').unsqueeze(1)\n        smpl_mask = stack_view(smpl_mask_list, 'smpl_mask') * img_mask\n        idx_h = torch.where(smpl_mask)\n        if int(smpl_mask.sum()) == 0:\n            total_loss = self.alpha_bce * score_loss\n            details = {\n                **score_details,\n            }\n            return total_loss, details\n    \n        # Prediction\n        pred_rotmat = stack_view(preds, 'smpl_rotmat')\n        pred_rotvec = roma.rotmat_to_rotvec(pred_rotmat[smpl_mask])\n        pred_shape = stack_view(preds, 'smpl_shape')\n        pred_transl = [pred.pop(\"smpl_transl\") for pred in preds]\n        pred_transl = stack_view(pred_transl, 'smpl_transl')\n        pred_expression = stack_view(preds, 'smpl_expression')\n        \n        # Neutral for MHMR\n        K_mhmr = stack_view(gts, 'K_mhmr')\n        mhmr_img_res = gts[0][\"img_mhmr\"].shape[-1]\n        # fine head uv\n        pred_loc = stack_view(preds, 'smpl_loc')\n        # Distance \n        dist = pred_transl[smpl_mask][:, 0].unsqueeze(-1)\n        dist = to_euclidean_dist(mhmr_img_res, dist, K_mhmr[idx_h[0]])  # use K GT\n        smpl_out = self.smpl_layer[f\"neutral_{pred_shape.shape[-1]}\"](\n            pred_rotvec, \n            pred_shape[smpl_mask], \n            None, \n            pred_loc[smpl_mask], \n            dist, \n            K=K_mhmr[idx_h[0]],\n            expression=pred_expression[smpl_mask],\n            K_to_proj=K[idx_h[0]], # if use K of CUT3R for projection\n            )\n        \n        pred_smpl = {}\n        batch_size = img_mask_list[0].shape[0]\n        num_view = len(gts)\n        max_humans = smpl_mask.shape[1]\n\n        for k, v in smpl_out.items():\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, *v.shape[1:], \n                device=v.device, dtype=v.dtype,\n            )\n            full_out[smpl_mask] = v\n            pred_smpl[k] = full_out.chunk(num_view, dim=0)\n\n        # SMPL-X params\n        rotmat_loss, rotmat_details = self.smpl_param_loss(gts, preds, masks_list, \"smpl_rotmat\")\n        transl_loss, transl_details = self.smpl_param_loss(gts, pred_smpl['smpl_transl'], masks_list, \"smpl_transl\")\n        shape_dim = min([gts[0]['smpl_shape'].shape[-1], preds[0]['smpl_shape'].shape[-1]])\n        gt_shape = [gt['smpl_shape'][...,:shape_dim] for gt in gts]\n        pred_shape = [pred['smpl_shape'][...,:shape_dim] for pred in preds]\n        shape_loss, shape_details = self.smpl_param_loss(gt_shape, pred_shape, masks_list, \"smpl_shape\")\n\n        # 3D points\n        gt_transl_pelvis= [gt['smpl_transl_pelvis'][..., None, :] for gt in gts]\n        pred_transl_pelvis = pred_smpl['smpl_transl_pelvis']\n        j3d_loss, j3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_j3d'], \n                                    gt_transl_pelvis, \n                                    pred_transl_pelvis, \n                                    masks_list, \"smpl_j3d\")\n        v3d_loss, v3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_v3d'], \n                                    gt_transl_pelvis, \n                                    pred_transl_pelvis, \n                                    masks_list, \"smpl_v3d\",\n                                    ret_pred=True)\n        cj3d_loss, cj3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_j3d'], \n                                    None, \n                                    None, \n                                    masks_list, \"smpl_j3d\")\n        cv3d_loss, cv3d_details = self.point3d_loss(\n                                    gts, \n                                    pred_smpl['smpl_v3d'], \n                                    None, \n                                    None, \n                                    masks_list, \"smpl_v3d\")\n        \n        # total loss\n        total_loss = self.alpha_bce * score_loss +\\\n                     self.alpha_rotmat * rotmat_loss +\\\n                     self.alpha_shape * shape_loss +\\\n                     self.alpha_transl * transl_loss +\\\n                     self.alpha_j3d * j3d_loss +\\\n                     self.alpha_v3d * v3d_loss +\\\n                     self.alpha_j3d * cj3d_loss +\\\n                     self.alpha_v3d * cv3d_loss\n\n        details = {\n            **score_details,\n            **rotmat_details,\n            **transl_details,\n            **shape_details,\n            **j3d_details,\n            **v3d_details,\n            **cj3d_details,\n            **cv3d_details,\n        }\n\n        if cv3d_loss < 1.0:\n            # 2D reprojection\n            j2d_loss, j2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_j2d'], \n                                        masks_list, \"smpl_j2d\")\n            v2d_loss, v2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_v2d'], \n                                        masks_list, \"smpl_v2d\")\n            total_loss += self.alpha_j2d * j2d_loss +\\\n                          self.alpha_v2d * v2d_loss\n            details.update({\n                **j2d_details,\n                **v2d_details\n            })\n\n        return total_loss, details\n\nclass DepthScaleShiftInvLoss(BaseCriterion):\n    \"\"\"scale and shift invariant loss\"\"\"\n\n    def __init__(self, reduction=\"none\"):\n        super().__init__(reduction)\n\n    def forward(self, pred, gt, mask):\n        assert pred.shape == gt.shape and pred.ndim == 3, f\"Bad shape = {pred.shape}\"\n        dist = self.distance(pred, gt, mask)\n        # assert dist.ndim == a.ndim - 1  # one dimension less\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def normalize(self, x, mask):","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.DepthScaleShiftInvLoss","uri":"program://Human3R/class/src.dust3r.losses.DepthScaleShiftInvLoss#L686-L719","kind":"class","name":"DepthScaleShiftInvLoss","path":"src/dust3r/losses.py","language":"python","start_line":686,"end_line":719,"context_start_line":666,"context_end_line":739,"code":"\n        if cv3d_loss < 1.0:\n            # 2D reprojection\n            j2d_loss, j2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_j2d'], \n                                        masks_list, \"smpl_j2d\")\n            v2d_loss, v2d_details = self.point2d_loss(\n                                        gts, \n                                        pred_smpl['smpl_v2d'], \n                                        masks_list, \"smpl_v2d\")\n            total_loss += self.alpha_j2d * j2d_loss +\\\n                          self.alpha_v2d * v2d_loss\n            details.update({\n                **j2d_details,\n                **v2d_details\n            })\n\n        return total_loss, details\n\nclass DepthScaleShiftInvLoss(BaseCriterion):\n    \"\"\"scale and shift invariant loss\"\"\"\n\n    def __init__(self, reduction=\"none\"):\n        super().__init__(reduction)\n\n    def forward(self, pred, gt, mask):\n        assert pred.shape == gt.shape and pred.ndim == 3, f\"Bad shape = {pred.shape}\"\n        dist = self.distance(pred, gt, mask)\n        # assert dist.ndim == a.ndim - 1  # one dimension less\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def normalize(self, x, mask):\n        x_valid = x[mask]\n        splits = mask.sum(dim=(1, 2)).tolist()\n        x_valid_list = torch.split(x_valid, splits)\n        shift = [x.mean() for x in x_valid_list]\n        x_valid_centered = [x - m for x, m in zip(x_valid_list, shift)]\n        scale = [x.abs().mean() for x in x_valid_centered]\n        scale = torch.stack(scale)\n        shift = torch.stack(shift)\n        x = (x - shift.view(-1, 1, 1)) / scale.view(-1, 1, 1).clamp(min=1e-6)\n        return x\n\n    def distance(self, pred, gt, mask):\n        pred = self.normalize(pred, mask)\n        gt = self.normalize(gt, mask)\n        return torch.abs((pred - gt)[mask])\n\n\nclass ScaleInvLoss(BaseCriterion):\n    \"\"\"scale invariant loss\"\"\"\n\n    def __init__(self, reduction=\"none\"):\n        super().__init__(reduction)\n\n    def forward(self, pred, gt, mask):\n        assert pred.shape == gt.shape and pred.ndim == 4, f\"Bad shape = {pred.shape}\"\n        dist = self.distance(pred, gt, mask)\n        # assert dist.ndim == a.ndim - 1  # one dimension less\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.ScaleInvLoss","uri":"program://Human3R/class/src.dust3r.losses.ScaleInvLoss#L722-L749","kind":"class","name":"ScaleInvLoss","path":"src/dust3r/losses.py","language":"python","start_line":722,"end_line":749,"context_start_line":702,"context_end_line":769,"code":"        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def normalize(self, x, mask):\n        x_valid = x[mask]\n        splits = mask.sum(dim=(1, 2)).tolist()\n        x_valid_list = torch.split(x_valid, splits)\n        shift = [x.mean() for x in x_valid_list]\n        x_valid_centered = [x - m for x, m in zip(x_valid_list, shift)]\n        scale = [x.abs().mean() for x in x_valid_centered]\n        scale = torch.stack(scale)\n        shift = torch.stack(shift)\n        x = (x - shift.view(-1, 1, 1)) / scale.view(-1, 1, 1).clamp(min=1e-6)\n        return x\n\n    def distance(self, pred, gt, mask):\n        pred = self.normalize(pred, mask)\n        gt = self.normalize(gt, mask)\n        return torch.abs((pred - gt)[mask])\n\n\nclass ScaleInvLoss(BaseCriterion):\n    \"\"\"scale invariant loss\"\"\"\n\n    def __init__(self, reduction=\"none\"):\n        super().__init__(reduction)\n\n    def forward(self, pred, gt, mask):\n        assert pred.shape == gt.shape and pred.ndim == 4, f\"Bad shape = {pred.shape}\"\n        dist = self.distance(pred, gt, mask)\n        # assert dist.ndim == a.ndim - 1  # one dimension less\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def distance(self, pred, gt, mask):\n        pred_norm_factor = (torch.norm(pred, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum(\n            dim=(1, 2)\n        ).clamp(min=1e-6)\n        gt_norm_factor = (torch.norm(gt, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum(\n            dim=(1, 2)\n        ).clamp(min=1e-6)\n        pred = pred / pred_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6)\n        gt = gt / gt_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6)\n        return torch.norm(pred - gt, dim=-1)[mask]\n\n\nclass Regr3DPose(Criterion, MultiLoss):\n    \"\"\"Ensure that all 3D points are correct.\n    Asymmetric loss: view1 is supposed to be the anchor.\n\n    P1 = RT1 @ D1\n    P2 = RT2 @ D2\n    loss1 = (I @ pred_D1) - (RT1^-1 @ RT1 @ D1)\n    loss2 = (RT21 @ pred_D2) - (RT1^-1 @ P2)\n          = (RT21 @ pred_D2) - (RT1^-1 @ RT2 @ D2)\n    \"\"\"\n\n    def __init__(\n        self,\n        criterion,\n        norm_mode=\"?avg_dis\",\n        gt_scale=False,\n        sky_loss_value=2,\n        max_metric_scale=False,","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.Regr3DPose","uri":"program://Human3R/class/src.dust3r.losses.Regr3DPose#L752-L1325","kind":"class","name":"Regr3DPose","path":"src/dust3r/losses.py","language":"python","start_line":752,"end_line":1325,"context_start_line":732,"context_end_line":1345,"code":"        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def distance(self, pred, gt, mask):\n        pred_norm_factor = (torch.norm(pred, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum(\n            dim=(1, 2)\n        ).clamp(min=1e-6)\n        gt_norm_factor = (torch.norm(gt, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum(\n            dim=(1, 2)\n        ).clamp(min=1e-6)\n        pred = pred / pred_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6)\n        gt = gt / gt_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6)\n        return torch.norm(pred - gt, dim=-1)[mask]\n\n\nclass Regr3DPose(Criterion, MultiLoss):\n    \"\"\"Ensure that all 3D points are correct.\n    Asymmetric loss: view1 is supposed to be the anchor.\n\n    P1 = RT1 @ D1\n    P2 = RT2 @ D2\n    loss1 = (I @ pred_D1) - (RT1^-1 @ RT1 @ D1)\n    loss2 = (RT21 @ pred_D2) - (RT1^-1 @ P2)\n          = (RT21 @ pred_D2) - (RT1^-1 @ RT2 @ D2)\n    \"\"\"\n\n    def __init__(\n        self,\n        criterion,\n        norm_mode=\"?avg_dis\",\n        gt_scale=False,\n        sky_loss_value=2,\n        max_metric_scale=False,\n    ):\n        super().__init__(criterion)\n        if norm_mode.startswith(\"?\"):\n            # do no norm pts from metric scale datasets\n            self.norm_all = False\n            self.norm_mode = norm_mode[1:]\n        else:\n            self.norm_all = True\n            self.norm_mode = norm_mode\n        self.gt_scale = gt_scale\n\n        self.sky_loss_value = sky_loss_value\n        self.max_metric_scale = max_metric_scale\n\n    def get_norm_factor_point_cloud(\n        self, pts_self, pts_cross, valids, conf_self, conf_cross, norm_self_only=False\n    ):\n        if norm_self_only:\n            norm_factor = normalize_pointcloud_group(\n                pts_self, self.norm_mode, valids, conf_self, ret_factor_only=True\n            )\n        else:\n            pts = [torch.cat([x, y], dim=2) for x, y in zip(pts_self, pts_cross)]\n            valids = [torch.cat([x, x], dim=2) for x in valids]\n            confs = [torch.cat([x, y], dim=2) for x, y in zip(conf_self, conf_cross)]\n            norm_factor = normalize_pointcloud_group(\n                pts, self.norm_mode, valids, confs, ret_factor_only=True\n            )\n        return norm_factor\n\n    def get_norm_factor_poses(self, gt_trans, pr_trans, not_metric_mask):\n\n        if self.norm_mode and not self.gt_scale:\n            gt_trans = [x[:, None, None, :].clone() for x in gt_trans]\n            valids = [torch.ones_like(x[..., 0], dtype=torch.bool) for x in gt_trans]\n            norm_factor_gt = (\n                normalize_pointcloud_group(\n                    gt_trans,\n                    self.norm_mode,\n                    valids,\n                    ret_factor_only=True,\n                )\n                .squeeze(-1)\n                .squeeze(-1)\n            )\n        else:\n            norm_factor_gt = torch.ones(\n                len(gt_trans), dtype=gt_trans[0].dtype, device=gt_trans[0].device\n            )\n\n        norm_factor_pr = norm_factor_gt.clone()\n        if self.norm_mode and not_metric_mask.sum() > 0 and not self.gt_scale:\n            pr_trans_not_metric = [\n                x[not_metric_mask][:, None, None, :].clone() for x in pr_trans\n            ]\n            valids = [\n                torch.ones_like(x[..., 0], dtype=torch.bool)\n                for x in pr_trans_not_metric\n            ]\n            norm_factor_pr_not_metric = (\n                normalize_pointcloud_group(\n                    pr_trans_not_metric,\n                    self.norm_mode,\n                    valids,\n                    ret_factor_only=True,\n                )\n                .squeeze(-1)\n                .squeeze(-1)\n            )\n            norm_factor_pr[not_metric_mask] = norm_factor_pr_not_metric\n        return norm_factor_gt, norm_factor_pr\n\n    def get_all_pts3d(\n        self,\n        gts,\n        preds,\n        dist_clip=None,\n        norm_self_only=False,\n        norm_pose_separately=False,\n        eps=1e-3,\n        camera1=None,\n    ):\n        # everything is normalized w.r.t. camera of view1\n        in_camera1 = inv(gts[0][\"camera_pose\"]) if camera1 is None else inv(camera1)\n        gt_pts_self = [geotrf(inv(gt[\"camera_pose\"]), gt[\"pts3d\"]) for gt in gts]\n        gt_pts_cross = [geotrf(in_camera1, gt[\"pts3d\"]) for gt in gts]\n        valids = [gt[\"valid_mask\"].clone() for gt in gts]\n        camera_only = gts[0][\"camera_only\"]\n\n        if dist_clip is not None:\n            # points that are too far-away == invalid\n            dis = [gt_pt.norm(dim=-1) for gt_pt in gt_pts_cross]\n            valids = [valid & (dis <= dist_clip) for valid, dis in zip(valids, dis)]\n\n        pr_pts_self = [pred[\"pts3d_in_self_view\"] for pred in preds]\n        pr_pts_cross = [pred[\"pts3d_in_other_view\"] for pred in preds]\n        conf_self = [torch.log(pred[\"conf_self\"]).detach().clip(eps) for pred in preds]\n        conf_cross = [torch.log(pred[\"conf\"]).detach().clip(eps) for pred in preds]\n\n        if not self.norm_all:\n            if self.max_metric_scale:\n                B = valids[0].shape[0]\n                dist = [\n                    torch.where(valid, torch.linalg.norm(gt_pt_cross, dim=-1), 0).view(\n                        B, -1\n                    )\n                    for valid, gt_pt_cross in zip(valids, gt_pts_cross)\n                ]\n                for d in dist:\n                    gts[0][\"is_metric\"] = gts[0][\"is_metric_scale\"] & (\n                        d.max(dim=-1).values < self.max_metric_scale\n                    )\n            not_metric_mask = ~gts[0][\"is_metric\"]\n        else:\n            not_metric_mask = torch.ones_like(gts[0][\"is_metric\"])\n\n        # normalize 3d points\n        # compute the scale using only the self view point maps\n        if self.norm_mode and not self.gt_scale:\n            norm_factor_gt = self.get_norm_factor_point_cloud(\n                gt_pts_self,\n                gt_pts_cross,\n                valids,\n                conf_self,\n                conf_cross,\n                norm_self_only=norm_self_only,\n            )\n        else:\n            norm_factor_gt = torch.ones_like(\n                preds[0][\"pts3d_in_other_view\"][:, :1, :1, :1]\n            )\n\n        norm_factor_pr = norm_factor_gt.clone()\n        if self.norm_mode and not_metric_mask.sum() > 0 and not self.gt_scale:\n            norm_factor_pr_not_metric = self.get_norm_factor_point_cloud(\n                [pr_pt_self[not_metric_mask] for pr_pt_self in pr_pts_self],\n                [pr_pt_cross[not_metric_mask] for pr_pt_cross in pr_pts_cross],\n                [valid[not_metric_mask] for valid in valids],\n                [conf[not_metric_mask] for conf in conf_self],\n                [conf[not_metric_mask] for conf in conf_cross],\n                norm_self_only=norm_self_only,\n            )\n            norm_factor_pr[not_metric_mask] = norm_factor_pr_not_metric\n\n        norm_factor_gt = norm_factor_gt.clip(eps)\n        norm_factor_pr = norm_factor_pr.clip(eps)\n\n        gt_pts_self = [pts / norm_factor_gt for pts in gt_pts_self]\n        gt_pts_cross = [pts / norm_factor_gt for pts in gt_pts_cross]\n        pr_pts_self = [pts / norm_factor_pr for pts in pr_pts_self]\n        pr_pts_cross = [pts / norm_factor_pr for pts in pr_pts_cross]\n\n        # [(Bx3, BX4), (BX3, BX4), ...], 3 for translation, 4 for quaternion\n        gt_poses = [\n            camera_to_pose_encoding(in_camera1 @ gt[\"camera_pose\"]).clone()\n            for gt in gts\n        ]\n        pr_poses = [pred[\"camera_pose\"].clone() for pred in preds]\n        pose_norm_factor_gt = norm_factor_gt.clone().squeeze(2, 3)\n        pose_norm_factor_pr = norm_factor_pr.clone().squeeze(2, 3)\n\n        if norm_pose_separately:\n            gt_trans = [gt[:, :3] for gt in gt_poses]\n            pr_trans = [pr[:, :3] for pr in pr_poses]\n            pose_norm_factor_gt, pose_norm_factor_pr = self.get_norm_factor_poses(\n                gt_trans, pr_trans, not_metric_mask\n            )\n        elif any(camera_only):\n            gt_trans = [gt[:, :3] for gt in gt_poses]\n            pr_trans = [pr[:, :3] for pr in pr_poses]\n            pose_only_norm_factor_gt, pose_only_norm_factor_pr = (\n                self.get_norm_factor_poses(gt_trans, pr_trans, not_metric_mask)\n            )\n            pose_norm_factor_gt = torch.where(\n                camera_only[:, None], pose_only_norm_factor_gt, pose_norm_factor_gt\n            )\n            pose_norm_factor_pr = torch.where(\n                camera_only[:, None], pose_only_norm_factor_pr, pose_norm_factor_pr\n            )\n\n        gt_poses = [\n            (gt[:, :3] / pose_norm_factor_gt.clip(eps), gt[:, 3:]) for gt in gt_poses\n        ]\n        pr_poses = [\n            (pr[:, :3] / pose_norm_factor_pr.clip(eps), pr[:, 3:]) for pr in pr_poses\n        ]\n        pose_masks = (pose_norm_factor_gt.squeeze() > eps) & (\n            pose_norm_factor_pr.squeeze() > eps\n        )\n\n        if any(camera_only):\n            # this is equal to a loss for camera intrinsics\n            gt_pts_self = [\n                torch.where(\n                    camera_only[:, None, None, None],\n                    (gt / gt[..., -1:].clip(1e-6)).clip(-2, 2),\n                    gt,\n                )\n                for gt in gt_pts_self\n            ]\n            pr_pts_self = [\n                torch.where(\n                    camera_only[:, None, None, None],\n                    (pr / pr[..., -1:].clip(1e-6)).clip(-2, 2),\n                    pr,\n                )\n                for pr in pr_pts_self\n            ]\n            # # do not add cross view loss when there is only camera supervision\n\n        skys = [gt[\"sky_mask\"] & ~valid for gt, valid in zip(gts, valids)]\n        return (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            valids,\n            skys,\n            pose_masks,\n            {},\n        )\n\n    def get_all_pts3d_with_scale_loss(\n        self,\n        gts,\n        preds,\n        dist_clip=None,\n        norm_self_only=False,\n        norm_pose_separately=False,\n        eps=1e-3,\n    ):\n        # everything is normalized w.r.t. camera of view1\n        in_camera1 = inv(gts[0][\"camera_pose\"])\n        gt_pts_self = [geotrf(inv(gt[\"camera_pose\"]), gt[\"pts3d\"]) for gt in gts]\n        gt_pts_cross = [geotrf(in_camera1, gt[\"pts3d\"]) for gt in gts]\n        valids = [gt[\"valid_mask\"].clone() for gt in gts]\n        camera_only = gts[0][\"camera_only\"]\n\n        if dist_clip is not None:\n            # points that are too far-away == invalid\n            dis = [gt_pt.norm(dim=-1) for gt_pt in gt_pts_cross]\n            valids = [valid & (dis <= dist_clip) for valid, dis in zip(valids, dis)]\n\n        pr_pts_self = [pred[\"pts3d_in_self_view\"] for pred in preds]\n        pr_pts_cross = [pred[\"pts3d_in_other_view\"] for pred in preds]\n        conf_self = [torch.log(pred[\"conf_self\"]).detach().clip(eps) for pred in preds]\n        conf_cross = [torch.log(pred[\"conf\"]).detach().clip(eps) for pred in preds]\n\n        if not self.norm_all:\n            if self.max_metric_scale:\n                B = valids[0].shape[0]\n                dist = [\n                    torch.where(valid, torch.linalg.norm(gt_pt_cross, dim=-1), 0).view(\n                        B, -1\n                    )\n                    for valid, gt_pt_cross in zip(valids, gt_pts_cross)\n                ]\n                for d in dist:\n                    gts[0][\"is_metric\"] = gts[0][\"is_metric_scale\"] & (\n                        d.max(dim=-1).values < self.max_metric_scale\n                    )\n            not_metric_mask = ~gts[0][\"is_metric\"]\n        else:\n            not_metric_mask = torch.ones_like(gts[0][\"is_metric\"])\n\n        # normalize 3d points\n        # compute the scale using only the self view point maps\n        if self.norm_mode and not self.gt_scale:\n            norm_factor_gt = self.get_norm_factor_point_cloud(\n                gt_pts_self[:1],\n                gt_pts_cross[:1],\n                valids[:1],\n                conf_self[:1],\n                conf_cross[:1],\n                norm_self_only=norm_self_only,\n            )\n        else:\n            norm_factor_gt = torch.ones_like(\n                preds[0][\"pts3d_in_other_view\"][:, :1, :1, :1]\n            )\n\n        if self.norm_mode:\n            norm_factor_pr = self.get_norm_factor_point_cloud(\n                pr_pts_self[:1],\n                pr_pts_cross[:1],\n                valids[:1],\n                conf_self[:1],\n                conf_cross[:1],\n                norm_self_only=norm_self_only,\n            )\n        else:\n            raise NotImplementedError\n        # only add loss to metric scale norm factor\n        if (~not_metric_mask).sum() > 0:\n            pts_scale_loss = torch.abs(\n                norm_factor_pr[~not_metric_mask] - norm_factor_gt[~not_metric_mask]\n            ).mean()\n        else:\n            pts_scale_loss = 0.0\n\n        norm_factor_gt = norm_factor_gt.clip(eps)\n        norm_factor_pr = norm_factor_pr.clip(eps)\n\n        gt_pts_self = [pts / norm_factor_gt for pts in gt_pts_self]\n        gt_pts_cross = [pts / norm_factor_gt for pts in gt_pts_cross]\n        pr_pts_self = [pts / norm_factor_pr for pts in pr_pts_self]\n        pr_pts_cross = [pts / norm_factor_pr for pts in pr_pts_cross]\n\n        # [(Bx3, BX4), (BX3, BX4), ...], 3 for translation, 4 for quaternion\n        gt_poses = [\n            camera_to_pose_encoding(in_camera1 @ gt[\"camera_pose\"]).clone()\n            for gt in gts\n        ]\n        pr_poses = [pred[\"camera_pose\"].clone() for pred in preds]\n        pose_norm_factor_gt = norm_factor_gt.clone().squeeze(2, 3)\n        pose_norm_factor_pr = norm_factor_pr.clone().squeeze(2, 3)\n\n        if norm_pose_separately:\n            gt_trans = [gt[:, :3] for gt in gt_poses][:1]\n            pr_trans = [pr[:, :3] for pr in pr_poses][:1]\n            pose_norm_factor_gt, pose_norm_factor_pr = self.get_norm_factor_poses(\n                gt_trans, pr_trans, torch.ones_like(not_metric_mask)\n            )\n        elif any(camera_only):\n            gt_trans = [gt[:, :3] for gt in gt_poses][:1]\n            pr_trans = [pr[:, :3] for pr in pr_poses][:1]\n            pose_only_norm_factor_gt, pose_only_norm_factor_pr = (\n                self.get_norm_factor_poses(\n                    gt_trans, pr_trans, torch.ones_like(not_metric_mask)\n                )\n            )\n            pose_norm_factor_gt = torch.where(\n                camera_only[:, None], pose_only_norm_factor_gt, pose_norm_factor_gt\n            )\n            pose_norm_factor_pr = torch.where(\n                camera_only[:, None], pose_only_norm_factor_pr, pose_norm_factor_pr\n            )\n        # only add loss to metric scale norm factor\n        if (~not_metric_mask).sum() > 0:\n            pose_scale_loss = torch.abs(\n                pose_norm_factor_pr[~not_metric_mask]\n                - pose_norm_factor_gt[~not_metric_mask]\n            ).mean()\n        else:\n            pose_scale_loss = 0.0\n        gt_poses = [\n            (gt[:, :3] / pose_norm_factor_gt.clip(eps), gt[:, 3:]) for gt in gt_poses\n        ]\n        pr_poses = [\n            (pr[:, :3] / pose_norm_factor_pr.clip(eps), pr[:, 3:]) for pr in pr_poses\n        ]\n\n        pose_masks = (pose_norm_factor_gt.squeeze() > eps) & (\n            pose_norm_factor_pr.squeeze() > eps\n        )\n\n        if any(camera_only):\n            # this is equal to a loss for camera intrinsics\n            gt_pts_self = [\n                torch.where(\n                    camera_only[:, None, None, None],\n                    (gt / gt[..., -1:].clip(1e-6)).clip(-2, 2),\n                    gt,\n                )\n                for gt in gt_pts_self\n            ]\n            pr_pts_self = [\n                torch.where(\n                    camera_only[:, None, None, None],\n                    (pr / pr[..., -1:].clip(1e-6)).clip(-2, 2),\n                    pr,\n                )\n                for pr in pr_pts_self\n            ]\n            # # do not add cross view loss when there is only camera supervision\n\n        skys = [gt[\"sky_mask\"] & ~valid for gt, valid in zip(gts, valids)]\n        return (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            valids,\n            skys,\n            pose_masks,\n            {\"scale_loss\": pose_scale_loss + pts_scale_loss},\n        )\n\n    def compute_relative_pose_loss(\n        self, gt_trans, gt_quats, pr_trans, pr_quats, masks=None\n    ):\n        if masks is None:\n            masks = torch.ones(len(gt_trans), dtype=torch.bool, device=gt_trans.device)\n        gt_trans_matrix1 = gt_trans[:, :, None, :].repeat(1, 1, gt_trans.shape[1], 1)[\n            masks\n        ]\n        gt_trans_matrix2 = gt_trans[:, None, :, :].repeat(1, gt_trans.shape[1], 1, 1)[\n            masks\n        ]\n        gt_quats_matrix1 = gt_quats[:, :, None, :].repeat(1, 1, gt_quats.shape[1], 1)[\n            masks\n        ]\n        gt_quats_matrix2 = gt_quats[:, None, :, :].repeat(1, gt_quats.shape[1], 1, 1)[\n            masks\n        ]\n        pr_trans_matrix1 = pr_trans[:, :, None, :].repeat(1, 1, pr_trans.shape[1], 1)[\n            masks\n        ]\n        pr_trans_matrix2 = pr_trans[:, None, :, :].repeat(1, pr_trans.shape[1], 1, 1)[\n            masks\n        ]\n        pr_quats_matrix1 = pr_quats[:, :, None, :].repeat(1, 1, pr_quats.shape[1], 1)[\n            masks\n        ]\n        pr_quats_matrix2 = pr_quats[:, None, :, :].repeat(1, pr_quats.shape[1], 1, 1)[\n            masks\n        ]\n\n        gt_rel_trans, gt_rel_quats = relative_pose_absT_quatR(\n            gt_trans_matrix1, gt_quats_matrix1, gt_trans_matrix2, gt_quats_matrix2\n        )\n        pr_rel_trans, pr_rel_quats = relative_pose_absT_quatR(\n            pr_trans_matrix1, pr_quats_matrix1, pr_trans_matrix2, pr_quats_matrix2\n        )\n        rel_trans_err = torch.norm(gt_rel_trans - pr_rel_trans, dim=-1)\n        rel_quats_err = torch.norm(gt_rel_quats - pr_rel_quats, dim=-1)\n        return rel_trans_err.mean() + rel_quats_err.mean()\n\n    def compute_pose_loss(self, gt_poses, pred_poses, masks=None):\n        \"\"\"\n        gt_pose: list of (Bx3, Bx4)\n        pred_pose: list of (Bx3, Bx4)\n        masks: None, or B\n        \"\"\"\n        gt_trans = torch.stack([gt[0] for gt in gt_poses], dim=1)  # BxNx3\n        gt_quats = torch.stack([gt[1] for gt in gt_poses], dim=1)  # BXNX4\n        pred_trans = torch.stack([pr[0] for pr in pred_poses], dim=1)  # BxNx3\n        pred_quats = torch.stack([pr[1] for pr in pred_poses], dim=1)  # BxNx4\n        if masks == None:\n            pose_loss = (\n                torch.norm(pred_trans - gt_trans, dim=-1).mean()\n                + torch.norm(pred_quats - gt_quats, dim=-1).mean()\n            )\n        else:\n            if not any(masks):\n                return torch.tensor(0.0)\n            pose_loss = (\n                torch.norm(pred_trans - gt_trans, dim=-1)[masks].mean()\n                + torch.norm(pred_quats - gt_quats, dim=-1)[masks].mean()\n            )\n\n        return pose_loss\n\n    def compute_loss(self, gts, preds, **kw):\n        (\n            gt_pts_self,\n            gt_pts_cross,\n            pred_pts_self,\n            pred_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,\n        ) = self.get_all_pts3d(gts, preds, **kw)\n\n        if self.sky_loss_value > 0:\n            assert (\n                self.criterion.reduction == \"none\"\n            ), \"sky_loss_value should be 0 if no conf loss\"\n            masks = [mask | sky for mask, sky in zip(masks, skys)]\n\n        # self view loss and details\n        if \"Quantile\" in self.criterion.__class__.__name__:\n# ... truncated ...","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.Regr3DPoseBatchList","uri":"program://Human3R/class/src.dust3r.losses.Regr3DPoseBatchList#L1328-L1509","kind":"class","name":"Regr3DPoseBatchList","path":"src/dust3r/losses.py","language":"python","start_line":1328,"end_line":1509,"context_start_line":1308,"context_end_line":1529,"code":"                    skys_cross[i][masks_cross[i]], self.sky_loss_value, l\n                )\n\n        for i in range(len(ls_cross)):\n            details[self_name + f\"_pts3d/{i+1}\"] = float(\n                ls_cross[i].mean() if ls_cross[i].numel() > 0 else 0\n            )\n            details[f\"conf_{i+1}\"] = preds[i][\"conf\"].detach()\n\n        ls = ls_self + ls_cross\n        masks = masks + masks_cross\n        details[\"is_self\"] = [True] * len(ls_self) + [False] * len(ls_cross)\n        details[\"img_ids\"] = (\n            np.arange(len(ls_self)).tolist() + np.arange(len(ls_cross)).tolist()\n        )\n        details[\"pose_loss\"] = self.compute_pose_loss(gt_poses, pr_poses, pose_masks)\n\n        return Sum(*list(zip(ls, masks))), (details | monitoring)\n\n\nclass Regr3DPoseBatchList(Regr3DPose):\n    \"\"\"Ensure that all 3D points are correct.\n    Asymmetric loss: view1 is supposed to be the anchor.\n\n    P1 = RT1 @ D1\n    P2 = RT2 @ D2\n    loss1 = (I @ pred_D1) - (RT1^-1 @ RT1 @ D1)\n    loss2 = (RT21 @ pred_D2) - (RT1^-1 @ P2)\n          = (RT21 @ pred_D2) - (RT1^-1 @ RT2 @ D2)\n    \"\"\"\n\n    def __init__(\n        self,\n        criterion,\n        norm_mode=\"?avg_dis\",\n        gt_scale=False,\n        sky_loss_value=2,\n        max_metric_scale=False,\n    ):\n        super().__init__(\n            criterion, norm_mode, gt_scale, sky_loss_value, max_metric_scale\n        )\n        self.depth_only_criterion = DepthScaleShiftInvLoss()\n        self.single_view_criterion = ScaleInvLoss()\n\n    def reorg(self, ls_b, masks_b):\n        ids_split = [mask.sum(dim=(1, 2)) for mask in masks_b]\n        ls = [[] for _ in range(len(masks_b[0]))]\n        for i in range(len(ls_b)):\n            ls_splitted_i = torch.split(ls_b[i], ids_split[i].tolist())\n            for j in range(len(masks_b[0])):\n                ls[j].append(ls_splitted_i[j])\n        ls = [torch.cat(l) for l in ls]\n        return ls\n\n    def compute_loss(self, gts, preds, **kw):\n        (\n            gt_pts_self,\n            gt_pts_cross,\n            pred_pts_self,\n            pred_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,\n        ) = self.get_all_pts3d(gts, preds, **kw)\n\n        if self.sky_loss_value > 0:\n            assert (\n                self.criterion.reduction == \"none\"\n            ), \"sky_loss_value should be 0 if no conf loss\"\n            masks = [mask | sky for mask, sky in zip(masks, skys)]\n\n        camera_only = gts[0][\"camera_only\"]\n        depth_only = gts[0][\"depth_only\"]\n        single_view = gts[0][\"single_view\"]\n        is_metric = gts[0][\"is_metric\"]\n\n        # self view loss and details\n        if \"Quantile\" in self.criterion.__class__.__name__:\n            raise NotImplementedError\n        else:\n            # list [(B, h, w, 3)] x num_views -> list [num_views, h, w, 3] x B\n            gt_pts_self_b = torch.unbind(torch.stack(gt_pts_self, dim=1), dim=0)\n            pred_pts_self_b = torch.unbind(torch.stack(pred_pts_self, dim=1), dim=0)\n            masks_b = torch.unbind(torch.stack(masks, dim=1), dim=0)\n            ls_self_b = []\n            for i in range(len(gt_pts_self_b)):\n                if depth_only[\n                    i\n                ]:  # if only have relative depth, no intrinsics or anything\n                    ls_self_b.append(\n                        self.depth_only_criterion(\n                            pred_pts_self_b[i][..., -1],\n                            gt_pts_self_b[i][..., -1],\n                            masks_b[i],\n                        )\n                    )\n                elif (\n                    single_view[i] and not is_metric[i]\n                ):  # if single view, with intrinsics and not metric\n                    ls_self_b.append(\n                        self.single_view_criterion(\n                            pred_pts_self_b[i], gt_pts_self_b[i], masks_b[i]\n                        )\n                    )\n                else:  # if multiple view, or metric single view\n                    ls_self_b.append(\n                        self.criterion(\n                            pred_pts_self_b[i][masks_b[i]], gt_pts_self_b[i][masks_b[i]]\n                        )\n                    )\n            ls_self = self.reorg(ls_self_b, masks_b)\n\n        if self.sky_loss_value > 0:\n            assert (\n                self.criterion.reduction == \"none\"\n            ), \"sky_loss_value should be 0 if no conf loss\"\n            for i, l in enumerate(ls_self):\n                ls_self[i] = torch.where(skys[i][masks[i]], self.sky_loss_value, l)\n\n        self_name = type(self).__name__\n\n        details = {}\n        for i in range(len(ls_self)):\n            details[self_name + f\"_self_pts3d/{i+1}\"] = float(ls_self[i].mean())\n            details[f\"self_conf_{i+1}\"] = preds[i][\"conf_self\"].detach()\n            details[f\"gt_img{i+1}\"] = gts[i][\"img\"].permute(0, 2, 3, 1).detach()\n            details[f\"valid_mask_{i+1}\"] = masks[i].detach()\n\n            if \"img_mask\" in gts[i] and \"ray_mask\" in gts[i]:\n                details[f\"img_mask_{i+1}\"] = gts[i][\"img_mask\"].detach()\n                details[f\"ray_mask_{i+1}\"] = gts[i][\"ray_mask\"].detach()\n\n            if \"desc\" in preds[i]:\n                details[f\"desc_{i+1}\"] = preds[i][\"desc\"].detach()\n\n        if \"Quantile\" in self.criterion.__class__.__name__:\n            # quantile masks have already been determined by self view losses, here pass in None as quantile\n            raise NotImplementedError\n        else:\n            gt_pts_cross_b = torch.unbind(\n                torch.stack(gt_pts_cross, dim=1)[~camera_only], dim=0\n            )\n            pred_pts_cross_b = torch.unbind(\n                torch.stack(pred_pts_cross, dim=1)[~camera_only], dim=0\n            )\n            masks_cross_b = torch.unbind(torch.stack(masks, dim=1)[~camera_only], dim=0)\n            ls_cross_b = []\n            for i in range(len(gt_pts_cross_b)):\n                if depth_only[~camera_only][i]:\n                    ls_cross_b.append(\n                        self.depth_only_criterion(\n                            pred_pts_cross_b[i][..., -1],\n                            gt_pts_cross_b[i][..., -1],\n                            masks_cross_b[i],\n                        )\n                    )\n                elif single_view[~camera_only][i] and not is_metric[~camera_only][i]:\n                    ls_cross_b.append(\n                        self.single_view_criterion(\n                            pred_pts_cross_b[i], gt_pts_cross_b[i], masks_cross_b[i]\n                        )\n                    )\n                else:\n                    ls_cross_b.append(\n                        self.criterion(\n                            pred_pts_cross_b[i][masks_cross_b[i]],\n                            gt_pts_cross_b[i][masks_cross_b[i]],\n                        )\n                    )\n            ls_cross = self.reorg(ls_cross_b, masks_cross_b)\n\n        if self.sky_loss_value > 0:\n            assert (\n                self.criterion.reduction == \"none\"\n            ), \"sky_loss_value should be 0 if no conf loss\"\n            masks_cross = [mask[~camera_only] for mask in masks]\n            skys_cross = [sky[~camera_only] for sky in skys]\n            for i, l in enumerate(ls_cross):\n                ls_cross[i] = torch.where(\n                    skys_cross[i][masks_cross[i]], self.sky_loss_value, l\n                )\n\n        for i in range(len(ls_cross)):\n            details[self_name + f\"_pts3d/{i+1}\"] = float(\n                ls_cross[i].mean() if ls_cross[i].numel() > 0 else 0\n            )\n            details[f\"conf_{i+1}\"] = preds[i][\"conf\"].detach()\n\n        ls = ls_self + ls_cross\n        masks = masks + masks_cross\n        details[\"is_self\"] = [True] * len(ls_self) + [False] * len(ls_cross)\n        details[\"img_ids\"] = (\n            np.arange(len(ls_self)).tolist() + np.arange(len(ls_cross)).tolist()\n        )\n        pose_masks = pose_masks * gts[i][\"img_mask\"]\n        details[\"pose_loss\"] = self.compute_pose_loss(gt_poses, pr_poses, pose_masks)\n\n        return Sum(*list(zip(ls, masks))), (details | monitoring)\n\n\nclass ConfLoss(MultiLoss):\n    \"\"\"Weighted regression by learned confidence.\n        Assuming the input pixel_loss is a pixel-level regression loss.\n\n    Principle:\n        high-confidence means high conf = 0.1 ==> conf_loss = x / 10 + alpha*log(10)\n        low  confidence means low  conf = 10  ==> conf_loss = x * 10 - alpha*log(10)\n\n        alpha: hyperparameter\n    \"\"\"\n\n    def __init__(self, pixel_loss, alpha=1):\n        super().__init__()\n        assert alpha > 0\n        self.alpha = alpha\n        self.pixel_loss = pixel_loss.with_reduction(\"none\")\n\n    def get_name(self):","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.ConfLoss","uri":"program://Human3R/class/src.dust3r.losses.ConfLoss#L1512-L1584","kind":"class","name":"ConfLoss","path":"src/dust3r/losses.py","language":"python","start_line":1512,"end_line":1584,"context_start_line":1492,"context_end_line":1604,"code":"                )\n\n        for i in range(len(ls_cross)):\n            details[self_name + f\"_pts3d/{i+1}\"] = float(\n                ls_cross[i].mean() if ls_cross[i].numel() > 0 else 0\n            )\n            details[f\"conf_{i+1}\"] = preds[i][\"conf\"].detach()\n\n        ls = ls_self + ls_cross\n        masks = masks + masks_cross\n        details[\"is_self\"] = [True] * len(ls_self) + [False] * len(ls_cross)\n        details[\"img_ids\"] = (\n            np.arange(len(ls_self)).tolist() + np.arange(len(ls_cross)).tolist()\n        )\n        pose_masks = pose_masks * gts[i][\"img_mask\"]\n        details[\"pose_loss\"] = self.compute_pose_loss(gt_poses, pr_poses, pose_masks)\n\n        return Sum(*list(zip(ls, masks))), (details | monitoring)\n\n\nclass ConfLoss(MultiLoss):\n    \"\"\"Weighted regression by learned confidence.\n        Assuming the input pixel_loss is a pixel-level regression loss.\n\n    Principle:\n        high-confidence means high conf = 0.1 ==> conf_loss = x / 10 + alpha*log(10)\n        low  confidence means low  conf = 10  ==> conf_loss = x * 10 - alpha*log(10)\n\n        alpha: hyperparameter\n    \"\"\"\n\n    def __init__(self, pixel_loss, alpha=1):\n        super().__init__()\n        assert alpha > 0\n        self.alpha = alpha\n        self.pixel_loss = pixel_loss.with_reduction(\"none\")\n\n    def get_name(self):\n        return f\"ConfLoss({self.pixel_loss})\"\n\n    def get_conf_log(self, x):\n        return x, torch.log(x)\n\n    def compute_loss(self, gts, preds, **kw):\n        # compute per-pixel loss\n        losses_and_masks, details = self.pixel_loss(gts, preds, **kw)\n        if \"is_self\" in details and \"img_ids\" in details:\n            is_self = details[\"is_self\"]\n            img_ids = details[\"img_ids\"]\n        else:\n            is_self = [False] * len(losses_and_masks)\n            img_ids = list(range(len(losses_and_masks)))\n\n        # weight by confidence\n        conf_losses = []\n\n        for i in range(len(losses_and_masks)):\n            pred = preds[img_ids[i]]\n            conf_key = \"conf_self\" if is_self[i] else \"conf\"\n            if not is_self[i]:\n                camera_only = gts[0][\"camera_only\"]\n                conf, log_conf = self.get_conf_log(\n                    pred[conf_key][~camera_only][losses_and_masks[i][1]]\n                )\n            else:\n                conf, log_conf = self.get_conf_log(\n                    pred[conf_key][losses_and_masks[i][1]]\n                )\n\n            conf_loss = losses_and_masks[i][0] * conf - self.alpha * log_conf\n            conf_loss = conf_loss.mean() if conf_loss.numel() > 0 else 0\n            conf_losses.append(conf_loss)\n\n            if is_self[i]:\n                details[self.get_name() + f\"_conf_loss_self/{img_ids[i]+1}\"] = float(\n                    conf_loss\n                )\n            else:\n                details[self.get_name() + f\"_conf_loss/{img_ids[i]+1}\"] = float(\n                    conf_loss\n                )\n\n        details.pop(\"is_self\", None)\n        details.pop(\"img_ids\", None)\n\n        final_loss = sum(conf_losses) / len(conf_losses) * 2.0\n        if \"pose_loss\" in details:\n            final_loss = (\n                final_loss + details[\"pose_loss\"]\n            )  # , details\n        if \"scale_loss\" in details:\n            final_loss = final_loss + details[\"scale_loss\"]\n        return final_loss, details\n\n\nclass Regr3DPose_ScaleInv(Regr3DPose):\n    \"\"\"Same than Regr3D but invariant to depth shift.\n    if gt_scale == True: enforce the prediction to take the same scale than GT\n    \"\"\"\n\n    def get_all_pts3d(self, gts, preds):\n        # compute depth-normalized points\n        (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.Regr3DPose_ScaleInv","uri":"program://Human3R/class/src.dust3r.losses.Regr3DPose_ScaleInv#L1587-L1649","kind":"class","name":"Regr3DPose_ScaleInv","path":"src/dust3r/losses.py","language":"python","start_line":1587,"end_line":1649,"context_start_line":1567,"context_end_line":1649,"code":"                    conf_loss\n                )\n            else:\n                details[self.get_name() + f\"_conf_loss/{img_ids[i]+1}\"] = float(\n                    conf_loss\n                )\n\n        details.pop(\"is_self\", None)\n        details.pop(\"img_ids\", None)\n\n        final_loss = sum(conf_losses) / len(conf_losses) * 2.0\n        if \"pose_loss\" in details:\n            final_loss = (\n                final_loss + details[\"pose_loss\"]\n            )  # , details\n        if \"scale_loss\" in details:\n            final_loss = final_loss + details[\"scale_loss\"]\n        return final_loss, details\n\n\nclass Regr3DPose_ScaleInv(Regr3DPose):\n    \"\"\"Same than Regr3D but invariant to depth shift.\n    if gt_scale == True: enforce the prediction to take the same scale than GT\n    \"\"\"\n\n    def get_all_pts3d(self, gts, preds):\n        # compute depth-normalized points\n        (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,\n        ) = super().get_all_pts3d(gts, preds)\n\n        # measure scene scale\n        _, gt_scale_self = get_group_pointcloud_center_scale(gt_pts_self, masks)\n        _, pred_scale_self = get_group_pointcloud_center_scale(pr_pts_self, masks)\n\n        _, gt_scale_cross = get_group_pointcloud_center_scale(gt_pts_cross, masks)\n        _, pred_scale_cross = get_group_pointcloud_center_scale(pr_pts_cross, masks)\n\n        # prevent predictions to be in a ridiculous range\n        pred_scale_self = pred_scale_self.clip(min=1e-3, max=1e3)\n        pred_scale_cross = pred_scale_cross.clip(min=1e-3, max=1e3)\n\n        # subtract the median depth\n        if self.gt_scale:\n            pr_pts_self = [\n                pr_pt_self * gt_scale_self / pred_scale_self\n                for pr_pt_self in pr_pts_self\n            ]\n            pr_pts_cross = [\n                pr_pt_cross * gt_scale_cross / pred_scale_cross\n                for pr_pt_cross in pr_pts_cross\n            ]\n        else:\n            gt_pts_self = [gt_pt_self / gt_scale_self for gt_pt_self in gt_pts_self]\n            gt_pts_cross = [\n                gt_pt_cross / gt_scale_cross for gt_pt_cross in gt_pts_cross\n            ]\n            pr_pts_self = [pr_pt_self / pred_scale_self for pr_pt_self in pr_pts_self]\n            pr_pts_cross = [\n                pr_pt_cross / pred_scale_cross for pr_pt_cross in pr_pts_cross\n            ]\n\n        return (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,\n        )","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.__init__","uri":"program://Human3R/function/src.dust3r.losses.__init__#L1523-L1527","kind":"function","name":"__init__","path":"src/dust3r/losses.py","language":"python","start_line":1523,"end_line":1527,"context_start_line":1503,"context_end_line":1547,"code":"        details[\"img_ids\"] = (\n            np.arange(len(ls_self)).tolist() + np.arange(len(ls_cross)).tolist()\n        )\n        pose_masks = pose_masks * gts[i][\"img_mask\"]\n        details[\"pose_loss\"] = self.compute_pose_loss(gt_poses, pr_poses, pose_masks)\n\n        return Sum(*list(zip(ls, masks))), (details | monitoring)\n\n\nclass ConfLoss(MultiLoss):\n    \"\"\"Weighted regression by learned confidence.\n        Assuming the input pixel_loss is a pixel-level regression loss.\n\n    Principle:\n        high-confidence means high conf = 0.1 ==> conf_loss = x / 10 + alpha*log(10)\n        low  confidence means low  conf = 10  ==> conf_loss = x * 10 - alpha*log(10)\n\n        alpha: hyperparameter\n    \"\"\"\n\n    def __init__(self, pixel_loss, alpha=1):\n        super().__init__()\n        assert alpha > 0\n        self.alpha = alpha\n        self.pixel_loss = pixel_loss.with_reduction(\"none\")\n\n    def get_name(self):\n        return f\"ConfLoss({self.pixel_loss})\"\n\n    def get_conf_log(self, x):\n        return x, torch.log(x)\n\n    def compute_loss(self, gts, preds, **kw):\n        # compute per-pixel loss\n        losses_and_masks, details = self.pixel_loss(gts, preds, **kw)\n        if \"is_self\" in details and \"img_ids\" in details:\n            is_self = details[\"is_self\"]\n            img_ids = details[\"img_ids\"]\n        else:\n            is_self = [False] * len(losses_and_masks)\n            img_ids = list(range(len(losses_and_masks)))\n\n        # weight by confidence\n        conf_losses = []\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.forward","uri":"program://Human3R/function/src.dust3r.losses.forward#L728-L738","kind":"function","name":"forward","path":"src/dust3r/losses.py","language":"python","start_line":728,"end_line":738,"context_start_line":708,"context_end_line":758,"code":"        shift = [x.mean() for x in x_valid_list]\n        x_valid_centered = [x - m for x, m in zip(x_valid_list, shift)]\n        scale = [x.abs().mean() for x in x_valid_centered]\n        scale = torch.stack(scale)\n        shift = torch.stack(shift)\n        x = (x - shift.view(-1, 1, 1)) / scale.view(-1, 1, 1).clamp(min=1e-6)\n        return x\n\n    def distance(self, pred, gt, mask):\n        pred = self.normalize(pred, mask)\n        gt = self.normalize(gt, mask)\n        return torch.abs((pred - gt)[mask])\n\n\nclass ScaleInvLoss(BaseCriterion):\n    \"\"\"scale invariant loss\"\"\"\n\n    def __init__(self, reduction=\"none\"):\n        super().__init__(reduction)\n\n    def forward(self, pred, gt, mask):\n        assert pred.shape == gt.shape and pred.ndim == 4, f\"Bad shape = {pred.shape}\"\n        dist = self.distance(pred, gt, mask)\n        # assert dist.ndim == a.ndim - 1  # one dimension less\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def distance(self, pred, gt, mask):\n        pred_norm_factor = (torch.norm(pred, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum(\n            dim=(1, 2)\n        ).clamp(min=1e-6)\n        gt_norm_factor = (torch.norm(gt, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum(\n            dim=(1, 2)\n        ).clamp(min=1e-6)\n        pred = pred / pred_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6)\n        gt = gt / gt_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6)\n        return torch.norm(pred - gt, dim=-1)[mask]\n\n\nclass Regr3DPose(Criterion, MultiLoss):\n    \"\"\"Ensure that all 3D points are correct.\n    Asymmetric loss: view1 is supposed to be the anchor.\n\n    P1 = RT1 @ D1\n    P2 = RT2 @ D2\n    loss1 = (I @ pred_D1) - (RT1^-1 @ RT1 @ D1)","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.distance","uri":"program://Human3R/function/src.dust3r.losses.distance#L740-L749","kind":"function","name":"distance","path":"src/dust3r/losses.py","language":"python","start_line":740,"end_line":749,"context_start_line":720,"context_end_line":769,"code":"\n\nclass ScaleInvLoss(BaseCriterion):\n    \"\"\"scale invariant loss\"\"\"\n\n    def __init__(self, reduction=\"none\"):\n        super().__init__(reduction)\n\n    def forward(self, pred, gt, mask):\n        assert pred.shape == gt.shape and pred.ndim == 4, f\"Bad shape = {pred.shape}\"\n        dist = self.distance(pred, gt, mask)\n        # assert dist.ndim == a.ndim - 1  # one dimension less\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def distance(self, pred, gt, mask):\n        pred_norm_factor = (torch.norm(pred, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum(\n            dim=(1, 2)\n        ).clamp(min=1e-6)\n        gt_norm_factor = (torch.norm(gt, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum(\n            dim=(1, 2)\n        ).clamp(min=1e-6)\n        pred = pred / pred_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6)\n        gt = gt / gt_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6)\n        return torch.norm(pred - gt, dim=-1)[mask]\n\n\nclass Regr3DPose(Criterion, MultiLoss):\n    \"\"\"Ensure that all 3D points are correct.\n    Asymmetric loss: view1 is supposed to be the anchor.\n\n    P1 = RT1 @ D1\n    P2 = RT2 @ D2\n    loss1 = (I @ pred_D1) - (RT1^-1 @ RT1 @ D1)\n    loss2 = (RT21 @ pred_D2) - (RT1^-1 @ P2)\n          = (RT21 @ pred_D2) - (RT1^-1 @ RT2 @ D2)\n    \"\"\"\n\n    def __init__(\n        self,\n        criterion,\n        norm_mode=\"?avg_dis\",\n        gt_scale=False,\n        sky_loss_value=2,\n        max_metric_scale=False,","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.get_name","uri":"program://Human3R/function/src.dust3r.losses.get_name#L1529-L1530","kind":"function","name":"get_name","path":"src/dust3r/losses.py","language":"python","start_line":1529,"end_line":1530,"context_start_line":1509,"context_end_line":1550,"code":"        return Sum(*list(zip(ls, masks))), (details | monitoring)\n\n\nclass ConfLoss(MultiLoss):\n    \"\"\"Weighted regression by learned confidence.\n        Assuming the input pixel_loss is a pixel-level regression loss.\n\n    Principle:\n        high-confidence means high conf = 0.1 ==> conf_loss = x / 10 + alpha*log(10)\n        low  confidence means low  conf = 10  ==> conf_loss = x * 10 - alpha*log(10)\n\n        alpha: hyperparameter\n    \"\"\"\n\n    def __init__(self, pixel_loss, alpha=1):\n        super().__init__()\n        assert alpha > 0\n        self.alpha = alpha\n        self.pixel_loss = pixel_loss.with_reduction(\"none\")\n\n    def get_name(self):\n        return f\"ConfLoss({self.pixel_loss})\"\n\n    def get_conf_log(self, x):\n        return x, torch.log(x)\n\n    def compute_loss(self, gts, preds, **kw):\n        # compute per-pixel loss\n        losses_and_masks, details = self.pixel_loss(gts, preds, **kw)\n        if \"is_self\" in details and \"img_ids\" in details:\n            is_self = details[\"is_self\"]\n            img_ids = details[\"img_ids\"]\n        else:\n            is_self = [False] * len(losses_and_masks)\n            img_ids = list(range(len(losses_and_masks)))\n\n        # weight by confidence\n        conf_losses = []\n\n        for i in range(len(losses_and_masks)):\n            pred = preds[img_ids[i]]\n            conf_key = \"conf_self\" if is_self[i] else \"conf\"","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.with_reduction","uri":"program://Human3R/function/src.dust3r.losses.with_reduction#L142-L148","kind":"function","name":"with_reduction","path":"src/dust3r/losses.py","language":"python","start_line":142,"end_line":148,"context_start_line":122,"context_end_line":168,"code":"\n\nclass L1Loss(LLoss):\n    def distance(self, a, b):\n        return (a - b).abs()\n\nL1 = L1Loss()\n\n\nclass Criterion(nn.Module):\n    def __init__(self, criterion=None):\n        super().__init__()\n        assert isinstance(\n            criterion, BaseCriterion\n        ), f\"{criterion} is not a proper criterion!\"\n        self.criterion = copy(criterion)\n\n    def get_name(self):\n        return f\"{type(self).__name__}({self.criterion})\"\n\n    def with_reduction(self, mode=\"none\"):\n        res = loss = deepcopy(self)\n        while loss is not None:\n            assert isinstance(loss, Criterion)\n            loss.criterion.reduction = mode  # make it return the loss for each sample\n            loss = loss._loss2  # we assume loss is a Multiloss\n        return res\n\n\nclass MultiLoss(nn.Module):\n    \"\"\"Easily combinable losses (also keep track of individual loss values):\n        loss = MyLoss1() + 0.1*MyLoss2()\n    Usage:\n        Inherit from this class and override get_name() and compute_loss()\n    \"\"\"\n\n    def __init__(self):\n        super().__init__()\n        self._alpha = 1\n        self._loss2 = None\n\n    def compute_loss(self, *args, **kwargs):\n        raise NotImplementedError()\n\n    def get_name(self):\n        raise NotImplementedError()\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.compute_loss","uri":"program://Human3R/function/src.dust3r.losses.compute_loss#L1535-L1584","kind":"function","name":"compute_loss","path":"src/dust3r/losses.py","language":"python","start_line":1535,"end_line":1584,"context_start_line":1515,"context_end_line":1604,"code":"\n    Principle:\n        high-confidence means high conf = 0.1 ==> conf_loss = x / 10 + alpha*log(10)\n        low  confidence means low  conf = 10  ==> conf_loss = x * 10 - alpha*log(10)\n\n        alpha: hyperparameter\n    \"\"\"\n\n    def __init__(self, pixel_loss, alpha=1):\n        super().__init__()\n        assert alpha > 0\n        self.alpha = alpha\n        self.pixel_loss = pixel_loss.with_reduction(\"none\")\n\n    def get_name(self):\n        return f\"ConfLoss({self.pixel_loss})\"\n\n    def get_conf_log(self, x):\n        return x, torch.log(x)\n\n    def compute_loss(self, gts, preds, **kw):\n        # compute per-pixel loss\n        losses_and_masks, details = self.pixel_loss(gts, preds, **kw)\n        if \"is_self\" in details and \"img_ids\" in details:\n            is_self = details[\"is_self\"]\n            img_ids = details[\"img_ids\"]\n        else:\n            is_self = [False] * len(losses_and_masks)\n            img_ids = list(range(len(losses_and_masks)))\n\n        # weight by confidence\n        conf_losses = []\n\n        for i in range(len(losses_and_masks)):\n            pred = preds[img_ids[i]]\n            conf_key = \"conf_self\" if is_self[i] else \"conf\"\n            if not is_self[i]:\n                camera_only = gts[0][\"camera_only\"]\n                conf, log_conf = self.get_conf_log(\n                    pred[conf_key][~camera_only][losses_and_masks[i][1]]\n                )\n            else:\n                conf, log_conf = self.get_conf_log(\n                    pred[conf_key][losses_and_masks[i][1]]\n                )\n\n            conf_loss = losses_and_masks[i][0] * conf - self.alpha * log_conf\n            conf_loss = conf_loss.mean() if conf_loss.numel() > 0 else 0\n            conf_losses.append(conf_loss)\n\n            if is_self[i]:\n                details[self.get_name() + f\"_conf_loss_self/{img_ids[i]+1}\"] = float(\n                    conf_loss\n                )\n            else:\n                details[self.get_name() + f\"_conf_loss/{img_ids[i]+1}\"] = float(\n                    conf_loss\n                )\n\n        details.pop(\"is_self\", None)\n        details.pop(\"img_ids\", None)\n\n        final_loss = sum(conf_losses) / len(conf_losses) * 2.0\n        if \"pose_loss\" in details:\n            final_loss = (\n                final_loss + details[\"pose_loss\"]\n            )  # , details\n        if \"scale_loss\" in details:\n            final_loss = final_loss + details[\"scale_loss\"]\n        return final_loss, details\n\n\nclass Regr3DPose_ScaleInv(Regr3DPose):\n    \"\"\"Same than Regr3D but invariant to depth shift.\n    if gt_scale == True: enforce the prediction to take the same scale than GT\n    \"\"\"\n\n    def get_all_pts3d(self, gts, preds):\n        # compute depth-normalized points\n        (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.__mul__","uri":"program://Human3R/function/src.dust3r.losses.__mul__#L169-L173","kind":"function","name":"__mul__","path":"src/dust3r/losses.py","language":"python","start_line":169,"end_line":173,"context_start_line":149,"context_end_line":193,"code":"\n\nclass MultiLoss(nn.Module):\n    \"\"\"Easily combinable losses (also keep track of individual loss values):\n        loss = MyLoss1() + 0.1*MyLoss2()\n    Usage:\n        Inherit from this class and override get_name() and compute_loss()\n    \"\"\"\n\n    def __init__(self):\n        super().__init__()\n        self._alpha = 1\n        self._loss2 = None\n\n    def compute_loss(self, *args, **kwargs):\n        raise NotImplementedError()\n\n    def get_name(self):\n        raise NotImplementedError()\n\n    def __mul__(self, alpha):\n        assert isinstance(alpha, (int, float))\n        res = copy(self)\n        res._alpha = alpha\n        return res\n\n    __rmul__ = __mul__  # same\n\n    def __add__(self, loss2):\n        assert isinstance(loss2, MultiLoss)\n        res = cur = copy(self)\n        # find the end of the chain\n        while cur._loss2 is not None:\n            cur = cur._loss2\n        cur._loss2 = loss2\n        return res\n\n    def __repr__(self):\n        name = self.get_name()\n        if self._alpha != 1:\n            name = f\"{self._alpha:g}*{name}\"\n        if self._loss2:\n            name = f\"{name} + {self._loss2}\"\n        return name\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.__add__","uri":"program://Human3R/function/src.dust3r.losses.__add__#L177-L184","kind":"function","name":"__add__","path":"src/dust3r/losses.py","language":"python","start_line":177,"end_line":184,"context_start_line":157,"context_end_line":204,"code":"\n    def __init__(self):\n        super().__init__()\n        self._alpha = 1\n        self._loss2 = None\n\n    def compute_loss(self, *args, **kwargs):\n        raise NotImplementedError()\n\n    def get_name(self):\n        raise NotImplementedError()\n\n    def __mul__(self, alpha):\n        assert isinstance(alpha, (int, float))\n        res = copy(self)\n        res._alpha = alpha\n        return res\n\n    __rmul__ = __mul__  # same\n\n    def __add__(self, loss2):\n        assert isinstance(loss2, MultiLoss)\n        res = cur = copy(self)\n        # find the end of the chain\n        while cur._loss2 is not None:\n            cur = cur._loss2\n        cur._loss2 = loss2\n        return res\n\n    def __repr__(self):\n        name = self.get_name()\n        if self._alpha != 1:\n            name = f\"{self._alpha:g}*{name}\"\n        if self._loss2:\n            name = f\"{name} + {self._loss2}\"\n        return name\n\n    def forward(self, *args, **kwargs):\n        loss = self.compute_loss(*args, **kwargs)\n        if isinstance(loss, tuple):\n            loss, details = loss\n        elif loss.ndim == 0:\n            details = {self.get_name(): float(loss)}\n        else:\n            details = {}\n        loss = loss * self._alpha\n\n        if self._loss2:","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.__repr__","uri":"program://Human3R/function/src.dust3r.losses.__repr__#L186-L192","kind":"function","name":"__repr__","path":"src/dust3r/losses.py","language":"python","start_line":186,"end_line":192,"context_start_line":166,"context_end_line":212,"code":"    def get_name(self):\n        raise NotImplementedError()\n\n    def __mul__(self, alpha):\n        assert isinstance(alpha, (int, float))\n        res = copy(self)\n        res._alpha = alpha\n        return res\n\n    __rmul__ = __mul__  # same\n\n    def __add__(self, loss2):\n        assert isinstance(loss2, MultiLoss)\n        res = cur = copy(self)\n        # find the end of the chain\n        while cur._loss2 is not None:\n            cur = cur._loss2\n        cur._loss2 = loss2\n        return res\n\n    def __repr__(self):\n        name = self.get_name()\n        if self._alpha != 1:\n            name = f\"{self._alpha:g}*{name}\"\n        if self._loss2:\n            name = f\"{name} + {self._loss2}\"\n        return name\n\n    def forward(self, *args, **kwargs):\n        loss = self.compute_loss(*args, **kwargs)\n        if isinstance(loss, tuple):\n            loss, details = loss\n        elif loss.ndim == 0:\n            details = {self.get_name(): float(loss)}\n        else:\n            details = {}\n        loss = loss * self._alpha\n\n        if self._loss2:\n            loss2, details2 = self._loss2(*args, **kwargs)\n            loss = loss + loss2\n            details |= details2\n\n        return loss, details\n\n\nclass SSIM(nn.Module):","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.img_loss","uri":"program://Human3R/function/src.dust3r.losses.img_loss#L250-L251","kind":"function","name":"img_loss","path":"src/dust3r/losses.py","language":"python","start_line":250,"end_line":251,"context_start_line":230,"context_end_line":271,"code":"        y = self.refl(y)\n\n        mu_x = self.mu_x_pool(x)\n        mu_y = self.mu_y_pool(y)\n\n        sigma_x = self.sig_x_pool(x**2) - mu_x**2\n        sigma_y = self.sig_y_pool(y**2) - mu_y**2\n        sigma_xy = self.sig_xy_pool(x * y) - mu_x * mu_y\n\n        SSIM_n = (2 * mu_x * mu_y + self.C1) * (2 * sigma_xy + self.C2)\n        SSIM_d = (mu_x**2 + mu_y**2 + self.C1) * (sigma_x + sigma_y + self.C2)\n\n        return torch.clamp((1 - SSIM_n / SSIM_d) / 2, 0, 1)\n\n\nclass RGBLoss(Criterion, MultiLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)\n        self.ssim = SSIM()\n\n    def img_loss(self, a, b):\n        return self.criterion(a, b)\n\n    def compute_loss(self, gts, preds, **kw):\n        gt_rgbs = [gt[\"img\"].permute(0, 2, 3, 1) for gt in gts]\n        pred_rgbs = [pred[\"rgb\"] for pred in preds]\n        ls = [\n            self.img_loss(pred_rgb, gt_rgb)\n            for pred_rgb, gt_rgb in zip(pred_rgbs, gt_rgbs)\n        ]\n        details = {}\n        self_name = type(self).__name__\n        for i, l in enumerate(ls):\n            details[self_name + f\"_rgb/{i+1}\"] = float(l)\n            details[f\"pred_rgb_{i+1}\"] = pred_rgbs[i]\n        rgb_loss = sum(ls) / len(ls)\n        return rgb_loss, details\n\n\nclass SMPLLoss(Criterion, MultiLoss):\n    def __init__(self, criterion):\n        super().__init__(criterion)","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.mask_loss","uri":"program://Human3R/function/src.dust3r.losses.mask_loss#L300-L314","kind":"function","name":"mask_loss","path":"src/dust3r/losses.py","language":"python","start_line":300,"end_line":314,"context_start_line":280,"context_end_line":334,"code":"        self.alpha_j2d = 1.0 * scale\n        self.alpha_v2d = 1.0 * scale\n\n        # SMPL layer\n        person_center = 'head'\n        dict_smpl_layer = {\n            'neutral': {\n                10: SMPL_Layer(type='smplx', gender='neutral', num_betas=10, kid=False, person_center=person_center),\n                11: SMPL_Layer(type='smplx', gender='neutral', num_betas=11, kid=False, person_center=person_center),\n                }\n            }\n        _moduleDict = []\n        for k, _smpl_layer in dict_smpl_layer.items():\n            for x, y in _smpl_layer.items():\n                _moduleDict.append([f\"{k}_{x}\", deepcopy(y)])\n        self.smpl_layer = nn.ModuleDict(_moduleDict)\n\n    def get_name(self):\n        return \"SMPLLoss\"\n\n    def mask_loss(self, gts, preds, masks, ret_pred=False):\n        gt_msks = [gt[\"msk_mhmr\"].unsqueeze(-1) for gt in gts]\n        pred_msks = [pred[\"msk\"] for pred in preds]\n        ls = [\n            F.binary_cross_entropy(p[m], g[m])\n            for p, g, m in zip(pred_msks, gt_msks, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        for i, l in enumerate(ls):\n            details[self_name + f\"_msk/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_msk_{i+1}\"] = pred_msks[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n        \n    def bce(self, gts, preds, masks, ret_pred=False):\n        gt_scores = [(gt[\"smpl_scores\"] >= 1).to(int).unsqueeze(-1) for gt in gts]\n        pred_scores = [pred[\"smpl_scores\"] for pred in preds]\n        ls = [\n            _neg_loss(p[m], g[m])\n            for p, g, m in zip(pred_scores, gt_scores, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        for i, l in enumerate(ls):\n            details[self_name + f\"_scores/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_smpl_scores_{i+1}\"] = pred_scores[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n    \n    def smpl_param_loss(self, gts, preds, masks, k, ret_pred=False):\n        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.bce","uri":"program://Human3R/function/src.dust3r.losses.bce#L316-L330","kind":"function","name":"bce","path":"src/dust3r/losses.py","language":"python","start_line":316,"end_line":330,"context_start_line":296,"context_end_line":350,"code":"\n    def get_name(self):\n        return \"SMPLLoss\"\n\n    def mask_loss(self, gts, preds, masks, ret_pred=False):\n        gt_msks = [gt[\"msk_mhmr\"].unsqueeze(-1) for gt in gts]\n        pred_msks = [pred[\"msk\"] for pred in preds]\n        ls = [\n            F.binary_cross_entropy(p[m], g[m])\n            for p, g, m in zip(pred_msks, gt_msks, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        for i, l in enumerate(ls):\n            details[self_name + f\"_msk/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_msk_{i+1}\"] = pred_msks[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n        \n    def bce(self, gts, preds, masks, ret_pred=False):\n        gt_scores = [(gt[\"smpl_scores\"] >= 1).to(int).unsqueeze(-1) for gt in gts]\n        pred_scores = [pred[\"smpl_scores\"] for pred in preds]\n        ls = [\n            _neg_loss(p[m], g[m])\n            for p, g, m in zip(pred_scores, gt_scores, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        for i, l in enumerate(ls):\n            details[self_name + f\"_scores/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_smpl_scores_{i+1}\"] = pred_scores[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n    \n    def smpl_param_loss(self, gts, preds, masks, k, ret_pred=False):\n        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        ls = [\n            self.criterion(p[m], g[m])\n            for p, g, m in zip(preds, gts, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.smpl_param_loss","uri":"program://Human3R/function/src.dust3r.losses.smpl_param_loss#L332-L351","kind":"function","name":"smpl_param_loss","path":"src/dust3r/losses.py","language":"python","start_line":332,"end_line":351,"context_start_line":312,"context_end_line":371,"code":"                details[f\"pred_msk_{i+1}\"] = pred_msks[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n        \n    def bce(self, gts, preds, masks, ret_pred=False):\n        gt_scores = [(gt[\"smpl_scores\"] >= 1).to(int).unsqueeze(-1) for gt in gts]\n        pred_scores = [pred[\"smpl_scores\"] for pred in preds]\n        ls = [\n            _neg_loss(p[m], g[m])\n            for p, g, m in zip(pred_scores, gt_scores, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        for i, l in enumerate(ls):\n            details[self_name + f\"_scores/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_smpl_scores_{i+1}\"] = pred_scores[i]\n        bce = sum(ls) / len(ls)\n        return bce, details\n    \n    def smpl_param_loss(self, gts, preds, masks, k, ret_pred=False):\n        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        ls = [\n            self.criterion(p[m], g[m])\n            for p, g, m in zip(preds, gts, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def point3d_loss(self, gts, preds, gt_t_p, pr_t_ps, masks, k, ret_pred=False):\n        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        if gt_t_p is None or pr_t_ps is None:\n            ls = [\n                self.criterion(p[m], g[m])\n                for p, g, m in zip(preds, gts, masks)\n            ]\n        else:\n            ls = [\n                self.criterion(p[m]-pr_p[m], g[m]-gt_p[m])\n                for p, pr_p, g, gt_p, m in zip(preds, pr_t_ps, gts, gt_t_p, masks)\n            ]\n\n        details = {}\n        self_name = self.get_name()","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.point3d_loss","uri":"program://Human3R/function/src.dust3r.losses.point3d_loss#L353-L381","kind":"function","name":"point3d_loss","path":"src/dust3r/losses.py","language":"python","start_line":353,"end_line":381,"context_start_line":333,"context_end_line":401,"code":"        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        ls = [\n            self.criterion(p[m], g[m])\n            for p, g, m in zip(preds, gts, masks)\n        ]\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def point3d_loss(self, gts, preds, gt_t_p, pr_t_ps, masks, k, ret_pred=False):\n        if isinstance(gts[0], dict):\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        if gt_t_p is None or pr_t_ps is None:\n            ls = [\n                self.criterion(p[m], g[m])\n                for p, g, m in zip(preds, gts, masks)\n            ]\n        else:\n            ls = [\n                self.criterion(p[m]-pr_p[m], g[m]-gt_p[m])\n                for p, pr_p, g, gt_p, m in zip(preds, pr_t_ps, gts, gt_t_p, masks)\n            ]\n\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        if gt_t_p is None or pr_t_ps is None:\n            k_name = \"c\" + k_name\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def point2d_loss(self, gts, preds, masks, k, shape=None, ret_pred=False):\n        if isinstance(gts[0], dict):\n            shape = gts[0]['true_shape'][0]\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        valid_mask = [\n            ((gt[..., 0] > 0) & (gt[..., 0] < shape[1]) & (gt[..., 1] > 0) & (gt[..., 1] < shape[0]\n            )) for gt in gts]\n\n        ls = [\n            self.criterion(p[m1.unsqueeze(-1) & m2], g[m1.unsqueeze(-1) & m2])\n            for p, g, m1, m2 in zip(preds, gts, masks, valid_mask)\n        ]\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        for i, l in enumerate(ls):","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.point2d_loss","uri":"program://Human3R/function/src.dust3r.losses.point2d_loss#L383-L407","kind":"function","name":"point2d_loss","path":"src/dust3r/losses.py","language":"python","start_line":383,"end_line":407,"context_start_line":363,"context_end_line":427,"code":"            ]\n        else:\n            ls = [\n                self.criterion(p[m]-pr_p[m], g[m]-gt_p[m])\n                for p, pr_p, g, gt_p, m in zip(preds, pr_t_ps, gts, gt_t_p, masks)\n            ]\n\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        if gt_t_p is None or pr_t_ps is None:\n            k_name = \"c\" + k_name\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def point2d_loss(self, gts, preds, masks, k, shape=None, ret_pred=False):\n        if isinstance(gts[0], dict):\n            shape = gts[0]['true_shape'][0]\n            gts = [gt[k] for gt in gts]\n        if isinstance(preds[0], dict):\n            preds = [pred[k] for pred in preds]\n        \n        valid_mask = [\n            ((gt[..., 0] > 0) & (gt[..., 0] < shape[1]) & (gt[..., 1] > 0) & (gt[..., 1] < shape[0]\n            )) for gt in gts]\n\n        ls = [\n            self.criterion(p[m1.unsqueeze(-1) & m2], g[m1.unsqueeze(-1) & m2])\n            for p, g, m1, m2 in zip(preds, gts, masks, valid_mask)\n        ]\n        details = {}\n        self_name = self.get_name()\n        k_name = k.split('smpl_')[-1] if 'smpl_' in k else k\n        for i, l in enumerate(ls):\n            details[self_name + f\"_{k_name}/{i+1}\"] = float(l)\n            if ret_pred:\n                details[f\"pred_{k}_{i+1}\"] = preds[i]\n        loss = sum(ls) / len(ls)\n\n        return loss, details\n\n    def compute_loss(self, gts, preds, **kw):\n        img_mask_list = [gt[\"img_mask\"] for gt in gts]\n        smpl_mask_list = [gt[\"smpl_mask\"] for gt in gts]\n        masks_list = [a.unsqueeze(1) & b for a, b in zip(img_mask_list, smpl_mask_list)]\n\n        # Detection loss\n        score_loss, score_details = self.bce(gts, preds, img_mask_list, ret_pred=True)\n\n        has_msk = \"msk\" in preds[0]\n        if has_msk:\n            msk_loss, msk_details = self.mask_loss(gts, preds, img_mask_list, ret_pred=True)\n\n        K = stack_view(gts, 'camera_intrinsics')\n        img_mask = stack_view(img_mask_list,'img_mask').unsqueeze(1)\n        smpl_mask = stack_view(smpl_mask_list, 'smpl_mask') * img_mask\n        idx_h = torch.where(smpl_mask)\n        if int(smpl_mask.sum()) == 0:\n            total_loss = self.alpha_bce * score_loss\n            details = {","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.normalize","uri":"program://Human3R/function/src.dust3r.losses.normalize#L704-L714","kind":"function","name":"normalize","path":"src/dust3r/losses.py","language":"python","start_line":704,"end_line":714,"context_start_line":684,"context_end_line":734,"code":"        return total_loss, details\n\nclass DepthScaleShiftInvLoss(BaseCriterion):\n    \"\"\"scale and shift invariant loss\"\"\"\n\n    def __init__(self, reduction=\"none\"):\n        super().__init__(reduction)\n\n    def forward(self, pred, gt, mask):\n        assert pred.shape == gt.shape and pred.ndim == 3, f\"Bad shape = {pred.shape}\"\n        dist = self.distance(pred, gt, mask)\n        # assert dist.ndim == a.ndim - 1  # one dimension less\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":\n            return dist.sum()\n        if self.reduction == \"mean\":\n            return dist.mean() if dist.numel() > 0 else dist.new_zeros(())\n        raise ValueError(f\"bad {self.reduction=} mode\")\n\n    def normalize(self, x, mask):\n        x_valid = x[mask]\n        splits = mask.sum(dim=(1, 2)).tolist()\n        x_valid_list = torch.split(x_valid, splits)\n        shift = [x.mean() for x in x_valid_list]\n        x_valid_centered = [x - m for x, m in zip(x_valid_list, shift)]\n        scale = [x.abs().mean() for x in x_valid_centered]\n        scale = torch.stack(scale)\n        shift = torch.stack(shift)\n        x = (x - shift.view(-1, 1, 1)) / scale.view(-1, 1, 1).clamp(min=1e-6)\n        return x\n\n    def distance(self, pred, gt, mask):\n        pred = self.normalize(pred, mask)\n        gt = self.normalize(gt, mask)\n        return torch.abs((pred - gt)[mask])\n\n\nclass ScaleInvLoss(BaseCriterion):\n    \"\"\"scale invariant loss\"\"\"\n\n    def __init__(self, reduction=\"none\"):\n        super().__init__(reduction)\n\n    def forward(self, pred, gt, mask):\n        assert pred.shape == gt.shape and pred.ndim == 4, f\"Bad shape = {pred.shape}\"\n        dist = self.distance(pred, gt, mask)\n        # assert dist.ndim == a.ndim - 1  # one dimension less\n        if self.reduction == \"none\":\n            return dist\n        if self.reduction == \"sum\":","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.get_norm_factor_point_cloud","uri":"program://Human3R/function/src.dust3r.losses.get_norm_factor_point_cloud#L784-L798","kind":"function","name":"get_norm_factor_point_cloud","path":"src/dust3r/losses.py","language":"python","start_line":784,"end_line":798,"context_start_line":764,"context_end_line":818,"code":"        self,\n        criterion,\n        norm_mode=\"?avg_dis\",\n        gt_scale=False,\n        sky_loss_value=2,\n        max_metric_scale=False,\n    ):\n        super().__init__(criterion)\n        if norm_mode.startswith(\"?\"):\n            # do no norm pts from metric scale datasets\n            self.norm_all = False\n            self.norm_mode = norm_mode[1:]\n        else:\n            self.norm_all = True\n            self.norm_mode = norm_mode\n        self.gt_scale = gt_scale\n\n        self.sky_loss_value = sky_loss_value\n        self.max_metric_scale = max_metric_scale\n\n    def get_norm_factor_point_cloud(\n        self, pts_self, pts_cross, valids, conf_self, conf_cross, norm_self_only=False\n    ):\n        if norm_self_only:\n            norm_factor = normalize_pointcloud_group(\n                pts_self, self.norm_mode, valids, conf_self, ret_factor_only=True\n            )\n        else:\n            pts = [torch.cat([x, y], dim=2) for x, y in zip(pts_self, pts_cross)]\n            valids = [torch.cat([x, x], dim=2) for x in valids]\n            confs = [torch.cat([x, y], dim=2) for x, y in zip(conf_self, conf_cross)]\n            norm_factor = normalize_pointcloud_group(\n                pts, self.norm_mode, valids, confs, ret_factor_only=True\n            )\n        return norm_factor\n\n    def get_norm_factor_poses(self, gt_trans, pr_trans, not_metric_mask):\n\n        if self.norm_mode and not self.gt_scale:\n            gt_trans = [x[:, None, None, :].clone() for x in gt_trans]\n            valids = [torch.ones_like(x[..., 0], dtype=torch.bool) for x in gt_trans]\n            norm_factor_gt = (\n                normalize_pointcloud_group(\n                    gt_trans,\n                    self.norm_mode,\n                    valids,\n                    ret_factor_only=True,\n                )\n                .squeeze(-1)\n                .squeeze(-1)\n            )\n        else:\n            norm_factor_gt = torch.ones(\n                len(gt_trans), dtype=gt_trans[0].dtype, device=gt_trans[0].device\n            )","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.get_norm_factor_poses","uri":"program://Human3R/function/src.dust3r.losses.get_norm_factor_poses#L800-L840","kind":"function","name":"get_norm_factor_poses","path":"src/dust3r/losses.py","language":"python","start_line":800,"end_line":840,"context_start_line":780,"context_end_line":860,"code":"\n        self.sky_loss_value = sky_loss_value\n        self.max_metric_scale = max_metric_scale\n\n    def get_norm_factor_point_cloud(\n        self, pts_self, pts_cross, valids, conf_self, conf_cross, norm_self_only=False\n    ):\n        if norm_self_only:\n            norm_factor = normalize_pointcloud_group(\n                pts_self, self.norm_mode, valids, conf_self, ret_factor_only=True\n            )\n        else:\n            pts = [torch.cat([x, y], dim=2) for x, y in zip(pts_self, pts_cross)]\n            valids = [torch.cat([x, x], dim=2) for x in valids]\n            confs = [torch.cat([x, y], dim=2) for x, y in zip(conf_self, conf_cross)]\n            norm_factor = normalize_pointcloud_group(\n                pts, self.norm_mode, valids, confs, ret_factor_only=True\n            )\n        return norm_factor\n\n    def get_norm_factor_poses(self, gt_trans, pr_trans, not_metric_mask):\n\n        if self.norm_mode and not self.gt_scale:\n            gt_trans = [x[:, None, None, :].clone() for x in gt_trans]\n            valids = [torch.ones_like(x[..., 0], dtype=torch.bool) for x in gt_trans]\n            norm_factor_gt = (\n                normalize_pointcloud_group(\n                    gt_trans,\n                    self.norm_mode,\n                    valids,\n                    ret_factor_only=True,\n                )\n                .squeeze(-1)\n                .squeeze(-1)\n            )\n        else:\n            norm_factor_gt = torch.ones(\n                len(gt_trans), dtype=gt_trans[0].dtype, device=gt_trans[0].device\n            )\n\n        norm_factor_pr = norm_factor_gt.clone()\n        if self.norm_mode and not_metric_mask.sum() > 0 and not self.gt_scale:\n            pr_trans_not_metric = [\n                x[not_metric_mask][:, None, None, :].clone() for x in pr_trans\n            ]\n            valids = [\n                torch.ones_like(x[..., 0], dtype=torch.bool)\n                for x in pr_trans_not_metric\n            ]\n            norm_factor_pr_not_metric = (\n                normalize_pointcloud_group(\n                    pr_trans_not_metric,\n                    self.norm_mode,\n                    valids,\n                    ret_factor_only=True,\n                )\n                .squeeze(-1)\n                .squeeze(-1)\n            )\n            norm_factor_pr[not_metric_mask] = norm_factor_pr_not_metric\n        return norm_factor_gt, norm_factor_pr\n\n    def get_all_pts3d(\n        self,\n        gts,\n        preds,\n        dist_clip=None,\n        norm_self_only=False,\n        norm_pose_separately=False,\n        eps=1e-3,\n        camera1=None,\n    ):\n        # everything is normalized w.r.t. camera of view1\n        in_camera1 = inv(gts[0][\"camera_pose\"]) if camera1 is None else inv(camera1)\n        gt_pts_self = [geotrf(inv(gt[\"camera_pose\"]), gt[\"pts3d\"]) for gt in gts]\n        gt_pts_cross = [geotrf(in_camera1, gt[\"pts3d\"]) for gt in gts]\n        valids = [gt[\"valid_mask\"].clone() for gt in gts]\n        camera_only = gts[0][\"camera_only\"]\n\n        if dist_clip is not None:\n            # points that are too far-away == invalid","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.get_all_pts3d","uri":"program://Human3R/function/src.dust3r.losses.get_all_pts3d#L1592-L1649","kind":"function","name":"get_all_pts3d","path":"src/dust3r/losses.py","language":"python","start_line":1592,"end_line":1649,"context_start_line":1572,"context_end_line":1649,"code":"                )\n\n        details.pop(\"is_self\", None)\n        details.pop(\"img_ids\", None)\n\n        final_loss = sum(conf_losses) / len(conf_losses) * 2.0\n        if \"pose_loss\" in details:\n            final_loss = (\n                final_loss + details[\"pose_loss\"]\n            )  # , details\n        if \"scale_loss\" in details:\n            final_loss = final_loss + details[\"scale_loss\"]\n        return final_loss, details\n\n\nclass Regr3DPose_ScaleInv(Regr3DPose):\n    \"\"\"Same than Regr3D but invariant to depth shift.\n    if gt_scale == True: enforce the prediction to take the same scale than GT\n    \"\"\"\n\n    def get_all_pts3d(self, gts, preds):\n        # compute depth-normalized points\n        (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,\n        ) = super().get_all_pts3d(gts, preds)\n\n        # measure scene scale\n        _, gt_scale_self = get_group_pointcloud_center_scale(gt_pts_self, masks)\n        _, pred_scale_self = get_group_pointcloud_center_scale(pr_pts_self, masks)\n\n        _, gt_scale_cross = get_group_pointcloud_center_scale(gt_pts_cross, masks)\n        _, pred_scale_cross = get_group_pointcloud_center_scale(pr_pts_cross, masks)\n\n        # prevent predictions to be in a ridiculous range\n        pred_scale_self = pred_scale_self.clip(min=1e-3, max=1e3)\n        pred_scale_cross = pred_scale_cross.clip(min=1e-3, max=1e3)\n\n        # subtract the median depth\n        if self.gt_scale:\n            pr_pts_self = [\n                pr_pt_self * gt_scale_self / pred_scale_self\n                for pr_pt_self in pr_pts_self\n            ]\n            pr_pts_cross = [\n                pr_pt_cross * gt_scale_cross / pred_scale_cross\n                for pr_pt_cross in pr_pts_cross\n            ]\n        else:\n            gt_pts_self = [gt_pt_self / gt_scale_self for gt_pt_self in gt_pts_self]\n            gt_pts_cross = [\n                gt_pt_cross / gt_scale_cross for gt_pt_cross in gt_pts_cross\n            ]\n            pr_pts_self = [pr_pt_self / pred_scale_self for pr_pt_self in pr_pts_self]\n            pr_pts_cross = [\n                pr_pt_cross / pred_scale_cross for pr_pt_cross in pr_pts_cross\n            ]\n\n        return (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,\n        )","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.get_all_pts3d_with_scale_loss","uri":"program://Human3R/function/src.dust3r.losses.get_all_pts3d_with_scale_loss#L994-L1160","kind":"function","name":"get_all_pts3d_with_scale_loss","path":"src/dust3r/losses.py","language":"python","start_line":994,"end_line":1160,"context_start_line":974,"context_end_line":1180,"code":"                    pr,\n                )\n                for pr in pr_pts_self\n            ]\n            # # do not add cross view loss when there is only camera supervision\n\n        skys = [gt[\"sky_mask\"] & ~valid for gt, valid in zip(gts, valids)]\n        return (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            valids,\n            skys,\n            pose_masks,\n            {},\n        )\n\n    def get_all_pts3d_with_scale_loss(\n        self,\n        gts,\n        preds,\n        dist_clip=None,\n        norm_self_only=False,\n        norm_pose_separately=False,\n        eps=1e-3,\n    ):\n        # everything is normalized w.r.t. camera of view1\n        in_camera1 = inv(gts[0][\"camera_pose\"])\n        gt_pts_self = [geotrf(inv(gt[\"camera_pose\"]), gt[\"pts3d\"]) for gt in gts]\n        gt_pts_cross = [geotrf(in_camera1, gt[\"pts3d\"]) for gt in gts]\n        valids = [gt[\"valid_mask\"].clone() for gt in gts]\n        camera_only = gts[0][\"camera_only\"]\n\n        if dist_clip is not None:\n            # points that are too far-away == invalid\n            dis = [gt_pt.norm(dim=-1) for gt_pt in gt_pts_cross]\n            valids = [valid & (dis <= dist_clip) for valid, dis in zip(valids, dis)]\n\n        pr_pts_self = [pred[\"pts3d_in_self_view\"] for pred in preds]\n        pr_pts_cross = [pred[\"pts3d_in_other_view\"] for pred in preds]\n        conf_self = [torch.log(pred[\"conf_self\"]).detach().clip(eps) for pred in preds]\n        conf_cross = [torch.log(pred[\"conf\"]).detach().clip(eps) for pred in preds]\n\n        if not self.norm_all:\n            if self.max_metric_scale:\n                B = valids[0].shape[0]\n                dist = [\n                    torch.where(valid, torch.linalg.norm(gt_pt_cross, dim=-1), 0).view(\n                        B, -1\n                    )\n                    for valid, gt_pt_cross in zip(valids, gt_pts_cross)\n                ]\n                for d in dist:\n                    gts[0][\"is_metric\"] = gts[0][\"is_metric_scale\"] & (\n                        d.max(dim=-1).values < self.max_metric_scale\n                    )\n            not_metric_mask = ~gts[0][\"is_metric\"]\n        else:\n            not_metric_mask = torch.ones_like(gts[0][\"is_metric\"])\n\n        # normalize 3d points\n        # compute the scale using only the self view point maps\n        if self.norm_mode and not self.gt_scale:\n            norm_factor_gt = self.get_norm_factor_point_cloud(\n                gt_pts_self[:1],\n                gt_pts_cross[:1],\n                valids[:1],\n                conf_self[:1],\n                conf_cross[:1],\n                norm_self_only=norm_self_only,\n            )\n        else:\n            norm_factor_gt = torch.ones_like(\n                preds[0][\"pts3d_in_other_view\"][:, :1, :1, :1]\n            )\n\n        if self.norm_mode:\n            norm_factor_pr = self.get_norm_factor_point_cloud(\n                pr_pts_self[:1],\n                pr_pts_cross[:1],\n                valids[:1],\n                conf_self[:1],\n                conf_cross[:1],\n                norm_self_only=norm_self_only,\n            )\n        else:\n            raise NotImplementedError\n        # only add loss to metric scale norm factor\n        if (~not_metric_mask).sum() > 0:\n            pts_scale_loss = torch.abs(\n                norm_factor_pr[~not_metric_mask] - norm_factor_gt[~not_metric_mask]\n            ).mean()\n        else:\n            pts_scale_loss = 0.0\n\n        norm_factor_gt = norm_factor_gt.clip(eps)\n        norm_factor_pr = norm_factor_pr.clip(eps)\n\n        gt_pts_self = [pts / norm_factor_gt for pts in gt_pts_self]\n        gt_pts_cross = [pts / norm_factor_gt for pts in gt_pts_cross]\n        pr_pts_self = [pts / norm_factor_pr for pts in pr_pts_self]\n        pr_pts_cross = [pts / norm_factor_pr for pts in pr_pts_cross]\n\n        # [(Bx3, BX4), (BX3, BX4), ...], 3 for translation, 4 for quaternion\n        gt_poses = [\n            camera_to_pose_encoding(in_camera1 @ gt[\"camera_pose\"]).clone()\n            for gt in gts\n        ]\n        pr_poses = [pred[\"camera_pose\"].clone() for pred in preds]\n        pose_norm_factor_gt = norm_factor_gt.clone().squeeze(2, 3)\n        pose_norm_factor_pr = norm_factor_pr.clone().squeeze(2, 3)\n\n        if norm_pose_separately:\n            gt_trans = [gt[:, :3] for gt in gt_poses][:1]\n            pr_trans = [pr[:, :3] for pr in pr_poses][:1]\n            pose_norm_factor_gt, pose_norm_factor_pr = self.get_norm_factor_poses(\n                gt_trans, pr_trans, torch.ones_like(not_metric_mask)\n            )\n        elif any(camera_only):\n            gt_trans = [gt[:, :3] for gt in gt_poses][:1]\n            pr_trans = [pr[:, :3] for pr in pr_poses][:1]\n            pose_only_norm_factor_gt, pose_only_norm_factor_pr = (\n                self.get_norm_factor_poses(\n                    gt_trans, pr_trans, torch.ones_like(not_metric_mask)\n                )\n            )\n            pose_norm_factor_gt = torch.where(\n                camera_only[:, None], pose_only_norm_factor_gt, pose_norm_factor_gt\n            )\n            pose_norm_factor_pr = torch.where(\n                camera_only[:, None], pose_only_norm_factor_pr, pose_norm_factor_pr\n            )\n        # only add loss to metric scale norm factor\n        if (~not_metric_mask).sum() > 0:\n            pose_scale_loss = torch.abs(\n                pose_norm_factor_pr[~not_metric_mask]\n                - pose_norm_factor_gt[~not_metric_mask]\n            ).mean()\n        else:\n            pose_scale_loss = 0.0\n        gt_poses = [\n            (gt[:, :3] / pose_norm_factor_gt.clip(eps), gt[:, 3:]) for gt in gt_poses\n        ]\n        pr_poses = [\n            (pr[:, :3] / pose_norm_factor_pr.clip(eps), pr[:, 3:]) for pr in pr_poses\n        ]\n\n        pose_masks = (pose_norm_factor_gt.squeeze() > eps) & (\n            pose_norm_factor_pr.squeeze() > eps\n        )\n\n        if any(camera_only):\n            # this is equal to a loss for camera intrinsics\n            gt_pts_self = [\n                torch.where(\n                    camera_only[:, None, None, None],\n                    (gt / gt[..., -1:].clip(1e-6)).clip(-2, 2),\n                    gt,\n                )\n                for gt in gt_pts_self\n            ]\n            pr_pts_self = [\n                torch.where(\n                    camera_only[:, None, None, None],\n                    (pr / pr[..., -1:].clip(1e-6)).clip(-2, 2),\n                    pr,\n                )\n                for pr in pr_pts_self\n            ]\n            # # do not add cross view loss when there is only camera supervision\n\n        skys = [gt[\"sky_mask\"] & ~valid for gt, valid in zip(gts, valids)]\n        return (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            valids,\n            skys,\n            pose_masks,\n            {\"scale_loss\": pose_scale_loss + pts_scale_loss},\n        )\n\n    def compute_relative_pose_loss(\n        self, gt_trans, gt_quats, pr_trans, pr_quats, masks=None\n    ):\n        if masks is None:\n            masks = torch.ones(len(gt_trans), dtype=torch.bool, device=gt_trans.device)\n        gt_trans_matrix1 = gt_trans[:, :, None, :].repeat(1, 1, gt_trans.shape[1], 1)[\n            masks\n        ]\n        gt_trans_matrix2 = gt_trans[:, None, :, :].repeat(1, gt_trans.shape[1], 1, 1)[\n            masks\n        ]\n        gt_quats_matrix1 = gt_quats[:, :, None, :].repeat(1, 1, gt_quats.shape[1], 1)[\n            masks\n        ]\n        gt_quats_matrix2 = gt_quats[:, None, :, :].repeat(1, gt_quats.shape[1], 1, 1)[\n            masks\n        ]\n        pr_trans_matrix1 = pr_trans[:, :, None, :].repeat(1, 1, pr_trans.shape[1], 1)[\n            masks","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.compute_relative_pose_loss","uri":"program://Human3R/function/src.dust3r.losses.compute_relative_pose_loss#L1162-L1200","kind":"function","name":"compute_relative_pose_loss","path":"src/dust3r/losses.py","language":"python","start_line":1162,"end_line":1200,"context_start_line":1142,"context_end_line":1220,"code":"                    pr,\n                )\n                for pr in pr_pts_self\n            ]\n            # # do not add cross view loss when there is only camera supervision\n\n        skys = [gt[\"sky_mask\"] & ~valid for gt, valid in zip(gts, valids)]\n        return (\n            gt_pts_self,\n            gt_pts_cross,\n            pr_pts_self,\n            pr_pts_cross,\n            gt_poses,\n            pr_poses,\n            valids,\n            skys,\n            pose_masks,\n            {\"scale_loss\": pose_scale_loss + pts_scale_loss},\n        )\n\n    def compute_relative_pose_loss(\n        self, gt_trans, gt_quats, pr_trans, pr_quats, masks=None\n    ):\n        if masks is None:\n            masks = torch.ones(len(gt_trans), dtype=torch.bool, device=gt_trans.device)\n        gt_trans_matrix1 = gt_trans[:, :, None, :].repeat(1, 1, gt_trans.shape[1], 1)[\n            masks\n        ]\n        gt_trans_matrix2 = gt_trans[:, None, :, :].repeat(1, gt_trans.shape[1], 1, 1)[\n            masks\n        ]\n        gt_quats_matrix1 = gt_quats[:, :, None, :].repeat(1, 1, gt_quats.shape[1], 1)[\n            masks\n        ]\n        gt_quats_matrix2 = gt_quats[:, None, :, :].repeat(1, gt_quats.shape[1], 1, 1)[\n            masks\n        ]\n        pr_trans_matrix1 = pr_trans[:, :, None, :].repeat(1, 1, pr_trans.shape[1], 1)[\n            masks\n        ]\n        pr_trans_matrix2 = pr_trans[:, None, :, :].repeat(1, pr_trans.shape[1], 1, 1)[\n            masks\n        ]\n        pr_quats_matrix1 = pr_quats[:, :, None, :].repeat(1, 1, pr_quats.shape[1], 1)[\n            masks\n        ]\n        pr_quats_matrix2 = pr_quats[:, None, :, :].repeat(1, pr_quats.shape[1], 1, 1)[\n            masks\n        ]\n\n        gt_rel_trans, gt_rel_quats = relative_pose_absT_quatR(\n            gt_trans_matrix1, gt_quats_matrix1, gt_trans_matrix2, gt_quats_matrix2\n        )\n        pr_rel_trans, pr_rel_quats = relative_pose_absT_quatR(\n            pr_trans_matrix1, pr_quats_matrix1, pr_trans_matrix2, pr_quats_matrix2\n        )\n        rel_trans_err = torch.norm(gt_rel_trans - pr_rel_trans, dim=-1)\n        rel_quats_err = torch.norm(gt_rel_quats - pr_rel_quats, dim=-1)\n        return rel_trans_err.mean() + rel_quats_err.mean()\n\n    def compute_pose_loss(self, gt_poses, pred_poses, masks=None):\n        \"\"\"\n        gt_pose: list of (Bx3, Bx4)\n        pred_pose: list of (Bx3, Bx4)\n        masks: None, or B\n        \"\"\"\n        gt_trans = torch.stack([gt[0] for gt in gt_poses], dim=1)  # BxNx3\n        gt_quats = torch.stack([gt[1] for gt in gt_poses], dim=1)  # BXNX4\n        pred_trans = torch.stack([pr[0] for pr in pred_poses], dim=1)  # BxNx3\n        pred_quats = torch.stack([pr[1] for pr in pred_poses], dim=1)  # BxNx4\n        if masks == None:\n            pose_loss = (\n                torch.norm(pred_trans - gt_trans, dim=-1).mean()\n                + torch.norm(pred_quats - gt_quats, dim=-1).mean()\n            )\n        else:\n            if not any(masks):\n                return torch.tensor(0.0)\n            pose_loss = (","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.compute_pose_loss","uri":"program://Human3R/function/src.dust3r.losses.compute_pose_loss#L1202-L1225","kind":"function","name":"compute_pose_loss","path":"src/dust3r/losses.py","language":"python","start_line":1202,"end_line":1225,"context_start_line":1182,"context_end_line":1245,"code":"        pr_trans_matrix2 = pr_trans[:, None, :, :].repeat(1, pr_trans.shape[1], 1, 1)[\n            masks\n        ]\n        pr_quats_matrix1 = pr_quats[:, :, None, :].repeat(1, 1, pr_quats.shape[1], 1)[\n            masks\n        ]\n        pr_quats_matrix2 = pr_quats[:, None, :, :].repeat(1, pr_quats.shape[1], 1, 1)[\n            masks\n        ]\n\n        gt_rel_trans, gt_rel_quats = relative_pose_absT_quatR(\n            gt_trans_matrix1, gt_quats_matrix1, gt_trans_matrix2, gt_quats_matrix2\n        )\n        pr_rel_trans, pr_rel_quats = relative_pose_absT_quatR(\n            pr_trans_matrix1, pr_quats_matrix1, pr_trans_matrix2, pr_quats_matrix2\n        )\n        rel_trans_err = torch.norm(gt_rel_trans - pr_rel_trans, dim=-1)\n        rel_quats_err = torch.norm(gt_rel_quats - pr_rel_quats, dim=-1)\n        return rel_trans_err.mean() + rel_quats_err.mean()\n\n    def compute_pose_loss(self, gt_poses, pred_poses, masks=None):\n        \"\"\"\n        gt_pose: list of (Bx3, Bx4)\n        pred_pose: list of (Bx3, Bx4)\n        masks: None, or B\n        \"\"\"\n        gt_trans = torch.stack([gt[0] for gt in gt_poses], dim=1)  # BxNx3\n        gt_quats = torch.stack([gt[1] for gt in gt_poses], dim=1)  # BXNX4\n        pred_trans = torch.stack([pr[0] for pr in pred_poses], dim=1)  # BxNx3\n        pred_quats = torch.stack([pr[1] for pr in pred_poses], dim=1)  # BxNx4\n        if masks == None:\n            pose_loss = (\n                torch.norm(pred_trans - gt_trans, dim=-1).mean()\n                + torch.norm(pred_quats - gt_quats, dim=-1).mean()\n            )\n        else:\n            if not any(masks):\n                return torch.tensor(0.0)\n            pose_loss = (\n                torch.norm(pred_trans - gt_trans, dim=-1)[masks].mean()\n                + torch.norm(pred_quats - gt_quats, dim=-1)[masks].mean()\n            )\n\n        return pose_loss\n\n    def compute_loss(self, gts, preds, **kw):\n        (\n            gt_pts_self,\n            gt_pts_cross,\n            pred_pts_self,\n            pred_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,\n        ) = self.get_all_pts3d(gts, preds, **kw)\n\n        if self.sky_loss_value > 0:\n            assert (\n                self.criterion.reduction == \"none\"\n            ), \"sky_loss_value should be 0 if no conf loss\"\n            masks = [mask | sky for mask, sky in zip(masks, skys)]","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.reorg","uri":"program://Human3R/function/src.dust3r.losses.reorg#L1353-L1361","kind":"function","name":"reorg","path":"src/dust3r/losses.py","language":"python","start_line":1353,"end_line":1361,"context_start_line":1333,"context_end_line":1381,"code":"    P2 = RT2 @ D2\n    loss1 = (I @ pred_D1) - (RT1^-1 @ RT1 @ D1)\n    loss2 = (RT21 @ pred_D2) - (RT1^-1 @ P2)\n          = (RT21 @ pred_D2) - (RT1^-1 @ RT2 @ D2)\n    \"\"\"\n\n    def __init__(\n        self,\n        criterion,\n        norm_mode=\"?avg_dis\",\n        gt_scale=False,\n        sky_loss_value=2,\n        max_metric_scale=False,\n    ):\n        super().__init__(\n            criterion, norm_mode, gt_scale, sky_loss_value, max_metric_scale\n        )\n        self.depth_only_criterion = DepthScaleShiftInvLoss()\n        self.single_view_criterion = ScaleInvLoss()\n\n    def reorg(self, ls_b, masks_b):\n        ids_split = [mask.sum(dim=(1, 2)) for mask in masks_b]\n        ls = [[] for _ in range(len(masks_b[0]))]\n        for i in range(len(ls_b)):\n            ls_splitted_i = torch.split(ls_b[i], ids_split[i].tolist())\n            for j in range(len(masks_b[0])):\n                ls[j].append(ls_splitted_i[j])\n        ls = [torch.cat(l) for l in ls]\n        return ls\n\n    def compute_loss(self, gts, preds, **kw):\n        (\n            gt_pts_self,\n            gt_pts_cross,\n            pred_pts_self,\n            pred_pts_cross,\n            gt_poses,\n            pr_poses,\n            masks,\n            skys,\n            pose_masks,\n            monitoring,\n        ) = self.get_all_pts3d(gts, preds, **kw)\n\n        if self.sky_loss_value > 0:\n            assert (\n                self.criterion.reduction == \"none\"\n            ), \"sky_loss_value should be 0 if no conf loss\"\n            masks = [mask | sky for mask, sky in zip(masks, skys)]","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.losses.get_conf_log","uri":"program://Human3R/function/src.dust3r.losses.get_conf_log#L1532-L1533","kind":"function","name":"get_conf_log","path":"src/dust3r/losses.py","language":"python","start_line":1532,"end_line":1533,"context_start_line":1512,"context_end_line":1553,"code":"class ConfLoss(MultiLoss):\n    \"\"\"Weighted regression by learned confidence.\n        Assuming the input pixel_loss is a pixel-level regression loss.\n\n    Principle:\n        high-confidence means high conf = 0.1 ==> conf_loss = x / 10 + alpha*log(10)\n        low  confidence means low  conf = 10  ==> conf_loss = x * 10 - alpha*log(10)\n\n        alpha: hyperparameter\n    \"\"\"\n\n    def __init__(self, pixel_loss, alpha=1):\n        super().__init__()\n        assert alpha > 0\n        self.alpha = alpha\n        self.pixel_loss = pixel_loss.with_reduction(\"none\")\n\n    def get_name(self):\n        return f\"ConfLoss({self.pixel_loss})\"\n\n    def get_conf_log(self, x):\n        return x, torch.log(x)\n\n    def compute_loss(self, gts, preds, **kw):\n        # compute per-pixel loss\n        losses_and_masks, details = self.pixel_loss(gts, preds, **kw)\n        if \"is_self\" in details and \"img_ids\" in details:\n            is_self = details[\"is_self\"]\n            img_ids = details[\"img_ids\"]\n        else:\n            is_self = [False] * len(losses_and_masks)\n            img_ids = list(range(len(losses_and_masks)))\n\n        # weight by confidence\n        conf_losses = []\n\n        for i in range(len(losses_and_masks)):\n            pred = preds[img_ids[i]]\n            conf_key = \"conf_self\" if is_self[i] else \"conf\"\n            if not is_self[i]:\n                camera_only = gts[0][\"camera_only\"]\n                conf, log_conf = self.get_conf_log(","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model","uri":"program://Human3R/module/src.dust3r.smpl_model#L1-L441","kind":"module","name":"src.dust3r.smpl_model","path":"src/dust3r/smpl_model.py","language":"python","start_line":1,"end_line":441,"context_start_line":1,"context_end_line":441,"code":"# modified from Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nimport numpy as np\nimport smplx\nfrom smplx.joint_names import JOINT_NAMES\nfrom dust3r.utils.geometry import (\n    perspective_projection, \n    resize_camera_intrinsics,\n    get_camera_parameters\n)\nfrom dust3r.utils.image import pad_image\nimport roma\nimport pickle\nimport os\n\ncurrent_dir = os.path.dirname(os.path.abspath(__file__))\nsrc_dir = os.path.dirname(current_dir)\nSMPLX_DIR = os.path.join(src_dir, 'models')\nMEAN_PARAMS = os.path.join(src_dir, 'models', 'smpl_mean_params.npz')\nSMPLX2SMPL = os.path.join(src_dir, 'models', 'smplx', 'smplx2smpl.pkl')\n\nclass SMPLModel(object):\n    def __init__(self, device, model_args={}, eval_args={}):\n        self.device = device\n        self.person_center = 'head'\n        \n        self.patch_size = model_args.get('patch_size', 16)\n        self.mhmr_img_res = model_args.get('mhmr_img_res', 896)\n        self.bb_patch_size = model_args.get('bb_patch_size', 14)\n\n        # Parametric 3D human models\n        self.smplx_neutral_11 = smplx.create(\n            SMPLX_DIR, 'smplx', gender='neutral', use_pca=False, flat_hand_mean=True, num_betas=11).to(self.device)\n        self.smplx_neutral_10 = smplx.create(\n            SMPLX_DIR, 'smplx', gender='neutral', use_pca=False, flat_hand_mean=True, num_betas=10).to(self.device)\n        \n        # Evaluation\n        self.use_fake_K = eval_args.get('use_fake_K', False)\n        dataset = eval_args.get('dataset', None)\n        if dataset is not None:\n            self.smpl = [\n                smplx.create(SMPLX_DIR, 'smpl', gender=g).to(self.device) for g in ['neutral', 'male', 'female']]\n            self.smpl_faces = {'smpl': self.smpl[0].faces, 'smplx': self.smplx_neutral_11.faces}\n            with open(SMPLX2SMPL, 'rb') as f:\n                self.smplx2smpl = torch.from_numpy(pickle.load(f)['matrix'].astype(np.float32)).to(self.device)\n\n            if dataset in ['rich']:\n                self.smplx = {\n                    g: smplx.create(SMPLX_DIR, 'smplx', gender=g, num_pca_comps=12\n                                    ).to(self.device) for g in ['male', 'female']}\n            self._setup_dataset_config(dataset)        \n        \n    def _setup_dataset_config(self, dataset):\n        self.j_smpl = self.smpl[0].J_regressor[:24]\n        if dataset in ['3dpw']:\n            h36m_to_14 = [6, 5, 4, 1, 2, 3, 16, 15, 14, 11, 12, 13, 8, 10, 0, 7, 9][:14]\n            self.j_h36m = torch.Tensor(np.load('src/models/smpl/J_regressor_h36m.npy'))\n            self.j_regressor = self.j_h36m[h36m_to_14]\n            self.pelvis_idx = [2, 3]\n            self.params_type = 'smpl'\n        elif dataset in ['bedlam', 'rich']:\n            self.j_regressor = self.j_smpl\n            self.pelvis_idx = [1, 2]\n            self.params_type = 'smplx'\n        else:\n            self.j_regressor = self.j_smpl\n            self.pelvis_idx = [1, 2]\n            self.params_type = 'smpl'\n\n    def forward_smpl(self, dataset, smpl_dict, smpl_mask):\n        nhv = int(smpl_mask.sum())\n\n        if dataset in ['bedlam']:\n            out = self.smplx_neutral_11(\n                global_orient=smpl_dict['smplx_root_pose'][smpl_mask].reshape(-1, 3),\n                body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1, 21*3),\n                jaw_pose=smpl_dict['smplx_jaw_pose'][smpl_mask].reshape(-1, 3),\n                leye_pose=smpl_dict['smplx_leye_pose'][smpl_mask].reshape(-1, 3),\n                reye_pose=smpl_dict['smplx_reye_pose'][smpl_mask].reshape(-1, 3),\n                left_hand_pose=smpl_dict['smplx_left_hand_pose'][smpl_mask].reshape(-1, 15*3),\n                right_hand_pose=smpl_dict['smplx_right_hand_pose'][smpl_mask].reshape(-1, 15*3),\n                betas=smpl_dict['smplx_shape'][smpl_mask].reshape(-1, 11),\n                transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1, 3),\n                expression=self.smplx_neutral_11.expression.repeat(nhv, 1),\n            )\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n        elif dataset in ['3dpw']:\n            smpl_params = {\n                'global_orient': smpl_dict['smpl_root_pose'][smpl_mask].reshape(-1,3),\n                'body_pose': smpl_dict['smpl_body_pose'][smpl_mask].reshape(-1,23*3),\n                'betas': smpl_dict['smpl_shape'][smpl_mask].reshape(-1,10),\n                'transl': smpl_dict['smpl_transl'][smpl_mask].reshape(-1,3),\n                }\n            out = self.smpl[1](**smpl_params)\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n            # update verts/joints if this is not the right gender\n            if int(smpl_dict['smpl_gender_id'].max()) == 2:\n                out_female = self.smpl[2](**smpl_params)\n                idx = torch.where(smpl_dict['smpl_gender_id'] == 2)[1]\n                verts[idx] = out_female.vertices.reshape(nhv, -1, 3)[idx]\n                \n        elif dataset in ['emdb', 'emdb1', 'emdb2']:\n            gender = smpl_dict['smpl_gender_id'].max()\n            out = self.smpl[gender](\n                global_orient=smpl_dict['smpl_root_pose_w'][smpl_mask].reshape(-1,3),\n                body_pose=smpl_dict['smpl_body_pose'][smpl_mask].reshape(-1,23*3),\n                betas=smpl_dict['smpl_shape'][smpl_mask].reshape(-1,10),\n                transl=smpl_dict['smpl_transl_w'][smpl_mask].reshape(-1,3),\n            )\n            verts = out.vertices.reshape(nhv, -1, 3) # world space\n                \n        elif dataset in ['rich']:\n            gender = {1: 'male', 2: 'female'}[int(smpl_dict['smplx_gender_id'].max())]\n            out = self.smplx[gender](\n                global_orient=smpl_dict['smplx_global_orient'][smpl_mask].reshape(-1,3),\n                body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1,21*3),\n                jaw_pose=torch.zeros([nhv, 3]),\n                leye_pose=torch.zeros([nhv, 3]),\n                reye_pose=torch.zeros([nhv, 3]),\n                left_hand_pose=torch.zeros([nhv, 12]),\n                right_hand_pose=torch.zeros([nhv, 12]),\n                betas=smpl_dict['smplx_betas'][smpl_mask].reshape(-1,10),\n                transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1,3),\n                expression=torch.zeros([nhv, 10]),    \n            )\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n        if self.params_type == 'smplx':\n            verts = self.smplx2smpl @ verts\n        jts = self.j_regressor @ verts\n\n        return verts, jts\n\n    def update_smpl_gt(self, views):\n        target = {}\n\n        batch_size = views[0][\"img\"].shape[0]\n\n        smpl_keys = [k for k in views[0].keys() if 'smpl' in k]\n        smpl_dict = {\n            k: (stacked := torch.stack(\n                [view.pop(k) for view in views], dim=0)).view(-1, *stacked.shape[2:])\n            for k in smpl_keys\n        }   # Shape: (num_views * batch_size, 10, ...)\n        smpl_mask = smpl_dict['smpl_mask']\n        idx_h = torch.where(smpl_mask) # frame_idx, batch_idx, human_idx\n        K = torch.stack(\n            [view['camera_intrinsics'] for view in views], dim=0\n        )\n        K = K.view(-1, *K.shape[2:])\n        nhv = int(smpl_mask.sum())\n\n        # Get MHMR input image (high-res, square)\n        imgs = torch.stack([view[\"img\"] for view in views], dim=0)\n        imgs = imgs.view(-1, *imgs.shape[2:])\n        K_mhmr = resize_camera_intrinsics(K, *imgs.shape[2:], self.mhmr_img_res)\n        imgs_mhmr = pad_image(imgs, self.mhmr_img_res)\n\n        # SMPLX forward - BEDLAM\n        has_smplx_params = 1\n        out = self.smplx_neutral_11(\n            global_orient=smpl_dict['smplx_root_pose'][smpl_mask].reshape(-1, 3),\n            body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1, 21*3),\n            jaw_pose=smpl_dict['smplx_jaw_pose'][smpl_mask].reshape(-1, 3),\n            leye_pose=smpl_dict['smplx_leye_pose'][smpl_mask].reshape(-1, 3),\n            reye_pose=smpl_dict['smplx_reye_pose'][smpl_mask].reshape(-1, 3),\n            left_hand_pose=smpl_dict['smplx_left_hand_pose'][smpl_mask].reshape(-1, 15*3),\n            right_hand_pose=smpl_dict['smplx_right_hand_pose'][smpl_mask].reshape(-1, 15*3),\n            betas=smpl_dict['smplx_shape'][smpl_mask].reshape(-1, 11),\n            transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1, 3),\n            expression=self.smplx_neutral_11.expression.repeat(nhv, 1),\n        )\n        verts, jts = out.vertices.reshape(nhv, -1, 3), out.joints.reshape(nhv, -1, 3)\n\n        j2d = perspective_projection(jts, K[idx_h[0]])\n        v2d = perspective_projection(verts, K[idx_h[0]])\n\n        # Translation of the primary keypoint\n        root_joint_idx = JOINT_NAMES.index(self.person_center)\n        target['smpl_transl'] = jts[:,root_joint_idx] # [nhv,3]\n        target['smpl_transl_pelvis'] = jts[:,0] # [nhv,3]\n\n        # Fill in target\n        target['smpl_v3d'] = verts\n        target['smpl_j3d'] = jts\n        target['smpl_j2d'] = j2d\n        target['smpl_v2d'] = v2d\n\n        if has_smplx_params:\n            target['smpl_rotvec'] = torch.cat([smpl_dict['smplx_root_pose'],\n                                        smpl_dict['smplx_body_pose'],\n                                        smpl_dict['smplx_left_hand_pose'],\n                                        smpl_dict['smplx_right_hand_pose'],\n                                        smpl_dict['smplx_jaw_pose']],2)[smpl_mask] # [bs,nhmax]\n            target['smpl_rotmat'] = roma.rotvec_to_rotmat(target['smpl_rotvec'])\n            target['smpl_shape'] = smpl_dict['smplx_shape'][smpl_mask]\n\n        \n        true_shapes = torch.stack([view[\"true_shape\"] for view in views], dim=0)\n        if len(torch.unique(true_shapes, dim=0)) != 1:\n            raise NotImplementedError\n        \n        # Creating the target heatmap for the primary keypoint\n        pk = target['smpl_transl'].unsqueeze(1) # (nhv,3)\n        \n        # For 512 res (CUT3R, patch_size=16)\n        pk_loc = perspective_projection(pk, K[idx_h[0]]).squeeze(1) # original pixel uv coordinates (nhv,2): W, H\n        n_patch_16, pk_idx_16 = get_patch_uv(true_shapes[0][0], self.patch_size, pk_loc)\n        target['smpl_uv_16'] = pk_idx_16[:, [1, 0]]\n\n        # For 896 res (MHMR, patch_size=14)\n        pk_loc_mhmr = perspective_projection(pk, K_mhmr[idx_h[0]]).squeeze(1) # original pixel uv coordinates (nhv,2): W, H\n        n_patch_14, pk_idx_14 = get_patch_uv(self.mhmr_img_res, self.bb_patch_size, pk_loc_mhmr)\n        smpl_mask_14, visible_humans_14, scores_14 = get_score(n_patch_14, pk_idx_14, smpl_mask.clone())\n        target['smpl_uv'] = pk_idx_14[:, [1, 0]]\n\n        # Rebatch and Update with visibility indice\n        _target = {}\n        num_view = len(views)\n        max_humans = smpl_mask_14.shape[1]\n        idx_vis = torch.where(visible_humans_14)[0]\n\n        for k, v in target.items():\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, *v.shape[1:], \n                device=v.device, dtype=v.dtype,\n            )\n            full_out[smpl_mask_14] = v[idx_vis] # discard unvisible humans due to olccusion\n            _target[k] = full_out.chunk(num_view, dim=0) # .view(num_view, batch_size, *full_out.shape[1:])\n\n        _target['smpl_scores'] = scores_14.chunk(num_view, dim=0)\n        _target['smpl_mask'] = smpl_mask_14.chunk(num_view, dim=0)\n        _target['K_mhmr'] = K_mhmr.chunk(num_view, dim=0)\n        _target['img_mhmr'] = imgs_mhmr.chunk(num_view, dim=0)\n\n        if \"msk\" in views[0]:\n            msks = torch.stack([view[\"msk\"] for view in views], dim=0)\n            msks = msks.view(-1, *msks.shape[2:])\n            msks_mhmr = pad_image(msks, self.mhmr_img_res, pad_value=0.0)  # bs,288,512->bs,896,896\n            msks_mhmr = (msks_mhmr > 0.1).float()\n            _target['msk_mhmr'] = msks_mhmr.chunk(num_view, dim=0)\n\n        for i, v in enumerate(zip(*_target.values())):\n            views[i].update(dict(zip(_target.keys(), v)))\n\n        torch.cuda.empty_cache()\n    \n    def update_smpl_gt_eval(self, views, dataset):\n        from dust3r.utils.geometry import geotrf\n\n        target = {}\n        batch_size = views[0][\"img\"].shape[0]\n\n        smpl_keys = [k for k in views[0].keys() if 'smpl' in k]\n        smpl_dict = {\n            k: (stacked := torch.stack(\n                [view.pop(k) for view in views], dim=0)).view(-1, *stacked.shape[2:])\n            for k in smpl_keys\n        }   # Shape: (num_views * batch_size, 10, ...)\n        smpl_mask = smpl_dict['smpl_mask']\n        idx_h = torch.where(smpl_mask) # frame_idx, batch_idx, human_idx\n        K = torch.stack([view['camera_intrinsics'] for view in views], dim=0)\n        K = K.view(-1, *K.shape[2:])\n\n        # Get MHMR input image (high-res, square)\n        imgs = torch.stack([view[\"img\"] for view in views], dim=0)\n        imgs = imgs.view(-1, *imgs.shape[2:])\n        K_mhmr = resize_camera_intrinsics(K, *imgs.shape[2:], self.mhmr_img_res)\n        imgs_mhmr = pad_image(imgs, self.mhmr_img_res)\n\n        verts, jts = self.forward_smpl(dataset, smpl_dict, smpl_mask)\n\n        if dataset in ['emdb', 'emdb1', 'emdb2', 'rich']:\n            target['smpl_v3d_w'] = verts\n            target['smpl_j3d_w'] = jts\n            T_w2c = torch.stack([view['T_w2c'] for view in views], dim=0)\n            T_w2c = T_w2c.view(-1, *T_w2c.shape[2:])\n            target['smpl_v3d_c'] = geotrf(T_w2c[idx_h[0]], verts)\n            target['smpl_j3d_c'] = geotrf(T_w2c[idx_h[0]], jts)\n \n        else:\n            target['smpl_v3d_c'] = verts\n            target['smpl_j3d_c'] = jts\n            T_c2w = torch.stack([view['camera_pose'] for view in views], dim=0)\n            T_c2w = T_c2w.view(-1, *T_c2w.shape[2:])\n            target['smpl_v3d_w'] = geotrf(T_c2w[idx_h[0]], verts)\n            target['smpl_j3d_w'] = geotrf(T_c2w[idx_h[0]], jts)\n\n        target['smpl_j2d'] = perspective_projection(target['smpl_j3d_c'], K[idx_h[0]])\n        target['smpl_v2d'] = perspective_projection(target['smpl_v3d_c'], K[idx_h[0]])\n\n        # Rebatch and Update with visibility indice\n        _target = {}\n        num_view = len(views)\n        max_humans = smpl_mask.shape[1]\n        for k, v in target.items():\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, *v.shape[1:], \n                device=v.device, dtype=v.dtype,\n            )\n            full_out[smpl_mask] = v # discard unvisible humans due to olccusion\n            _target[k] = full_out.chunk(num_view, dim=0) # .view(num_view, batch_size, *full_out.shape[1:])\n\n        if self.use_fake_K:\n            K_mhmr = get_camera_parameters(self.mhmr_img_res, device=K.device) # if use pseudo K\n            K_mhmr = K_mhmr.expand(K.shape[0], -1, -1)\n\n        _target['smpl_mask'] = smpl_mask.chunk(num_view, dim=0)\n        _target['K_mhmr'] = K_mhmr.chunk(num_view, dim=0)\n        _target['img_mhmr'] = imgs_mhmr.chunk(num_view, dim=0)\n\n        if \"msk\" in views[0]:\n            msks = torch.stack([view[\"msk\"] for view in views], dim=0)\n            msks = msks.view(-1, *msks.shape[2:])\n            msks_mhmr = pad_image(msks, self.mhmr_img_res, pad_value=0.0)  # bs,288,512->bs,896,896\n            msks_mhmr = (msks_mhmr > 0.1).float()\n            _target['msk_mhmr'] = msks_mhmr.chunk(num_view, dim=0)\n\n        for i, v in enumerate(zip(*_target.values())):\n            views[i].update(dict(zip(_target.keys(), v)))\n\n        torch.cuda.empty_cache()\n\n\ndef get_patch_uv(imgshape, patch_size, pk_loc):\n    n_patch = imgshape // patch_size  # H, W\n    pk_idx = (pk_loc // patch_size).int()\n    return n_patch, pk_idx\n\ndef get_score(n_patch, pk_idx, smpl_mask):\n    # Scores & updating valid_humans according to occlusion - wap X and Y for scores only\n    idx_h = torch.where(smpl_mask)\n    nhv = int(smpl_mask.sum())\n    bs = smpl_mask.shape[0]\n    device = smpl_mask.device\n\n    if isinstance(n_patch, (int, float)):\n        patch_h, patch_w = int(n_patch), int(n_patch)\n    else:\n        patch_h, patch_w = n_patch[0], n_patch[1]\n\n    scores = torch.zeros((bs, patch_h, patch_w)).to(device)\n    visible_humans = torch.ones(nhv).to(device) # by default no occlusion so all visible\n\n    for k in range(nhv):\n        i = int(idx_h[0][k]) # index of the image\n        j = int(idx_h[1][k]) # index of the human in this image\n        _x = pk_idx[k,1] # patch center H\n        _y = pk_idx[k,0] # patch center W\n        # filter out heads out of cropping bounds\n        if _x >= 0 and _x < patch_h and _y >= 0 and _y < patch_w:\n            if scores[i,_x,_y] == 1:\n                smpl_mask[i,j] = 0\n                visible_humans[k] = 0\n            else:\n                scores[i,_x,_y] = 1\n        else:\n            smpl_mask[i,j] = 0\n            visible_humans[k] = 0\n    \n    return smpl_mask, visible_humans, scores\n\n\nimport torch.nn as nn\nfrom croco.models.blocks import Mlp_flex\n\nclass SMPLDecoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        target_dim=1,\n        mlp_ratio=1,\n        num_layers=2,\n    ):\n        super().__init__()\n        self.mlp = Mlp_flex(\n            in_features=hidden_size,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=target_dim,\n            num_layers=num_layers,\n            drop=0,\n        )\n\n    def forward(\n        self,\n        feat,\n    ):\n        \"\"\"\n        feat: BxC\n        \"\"\"\n\n        pred = self.mlp(feat)\n        return pred\n\n\ndef regression_mlp(layers_sizes):\n    \"\"\"\n    Return a fully connected network.\n    \"\"\"\n    assert len(layers_sizes) >= 2\n    in_features = layers_sizes[0]\n    layers = []\n    for i in range(1, len(layers_sizes)-1):\n        out_features = layers_sizes[i]\n        layers.append(torch.nn.Linear(in_features, out_features))\n        layers.append(torch.nn.ReLU())\n        in_features = out_features\n    layers.append(torch.nn.Linear(in_features, layers_sizes[-1]))\n    return torch.nn.Sequential(*layers)\n\ndef apply_threshold(det_thresh, _scores):\n    \"\"\" Apply thresholding to detection scores; if stack_K is used and det_thresh is a list, apply to each channel separately \"\"\"\n    if isinstance(det_thresh, list):\n        det_thresh = det_thresh[0]\n    idx = torch.where(_scores >= det_thresh)\n    return idx\n\ndef nms(heat, kernel=3):\n    \"\"\" easy non maximal supression (as in CenterNet) \"\"\"\n\n    if kernel not in [2, 4]:\n        pad = (kernel - 1) // 2\n    else:\n        if kernel == 2:\n            pad = 1\n        else:\n            pad = 2\n\n    hmax = nn.functional.max_pool2d( heat, (kernel, kernel), stride=1, padding=pad)\n\n    if hmax.shape[2] > heat.shape[2]:\n        hmax = hmax[:, :, :heat.shape[2], :heat.shape[3]]\n\n    keep = (hmax == heat).float()\n\n    return heat * keep","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.SMPLModel","uri":"program://Human3R/class/src.dust3r.smpl_model.SMPLModel#L25-L327","kind":"class","name":"SMPLModel","path":"src/dust3r/smpl_model.py","language":"python","start_line":25,"end_line":327,"context_start_line":5,"context_end_line":347,"code":"import torch\nimport numpy as np\nimport smplx\nfrom smplx.joint_names import JOINT_NAMES\nfrom dust3r.utils.geometry import (\n    perspective_projection, \n    resize_camera_intrinsics,\n    get_camera_parameters\n)\nfrom dust3r.utils.image import pad_image\nimport roma\nimport pickle\nimport os\n\ncurrent_dir = os.path.dirname(os.path.abspath(__file__))\nsrc_dir = os.path.dirname(current_dir)\nSMPLX_DIR = os.path.join(src_dir, 'models')\nMEAN_PARAMS = os.path.join(src_dir, 'models', 'smpl_mean_params.npz')\nSMPLX2SMPL = os.path.join(src_dir, 'models', 'smplx', 'smplx2smpl.pkl')\n\nclass SMPLModel(object):\n    def __init__(self, device, model_args={}, eval_args={}):\n        self.device = device\n        self.person_center = 'head'\n        \n        self.patch_size = model_args.get('patch_size', 16)\n        self.mhmr_img_res = model_args.get('mhmr_img_res', 896)\n        self.bb_patch_size = model_args.get('bb_patch_size', 14)\n\n        # Parametric 3D human models\n        self.smplx_neutral_11 = smplx.create(\n            SMPLX_DIR, 'smplx', gender='neutral', use_pca=False, flat_hand_mean=True, num_betas=11).to(self.device)\n        self.smplx_neutral_10 = smplx.create(\n            SMPLX_DIR, 'smplx', gender='neutral', use_pca=False, flat_hand_mean=True, num_betas=10).to(self.device)\n        \n        # Evaluation\n        self.use_fake_K = eval_args.get('use_fake_K', False)\n        dataset = eval_args.get('dataset', None)\n        if dataset is not None:\n            self.smpl = [\n                smplx.create(SMPLX_DIR, 'smpl', gender=g).to(self.device) for g in ['neutral', 'male', 'female']]\n            self.smpl_faces = {'smpl': self.smpl[0].faces, 'smplx': self.smplx_neutral_11.faces}\n            with open(SMPLX2SMPL, 'rb') as f:\n                self.smplx2smpl = torch.from_numpy(pickle.load(f)['matrix'].astype(np.float32)).to(self.device)\n\n            if dataset in ['rich']:\n                self.smplx = {\n                    g: smplx.create(SMPLX_DIR, 'smplx', gender=g, num_pca_comps=12\n                                    ).to(self.device) for g in ['male', 'female']}\n            self._setup_dataset_config(dataset)        \n        \n    def _setup_dataset_config(self, dataset):\n        self.j_smpl = self.smpl[0].J_regressor[:24]\n        if dataset in ['3dpw']:\n            h36m_to_14 = [6, 5, 4, 1, 2, 3, 16, 15, 14, 11, 12, 13, 8, 10, 0, 7, 9][:14]\n            self.j_h36m = torch.Tensor(np.load('src/models/smpl/J_regressor_h36m.npy'))\n            self.j_regressor = self.j_h36m[h36m_to_14]\n            self.pelvis_idx = [2, 3]\n            self.params_type = 'smpl'\n        elif dataset in ['bedlam', 'rich']:\n            self.j_regressor = self.j_smpl\n            self.pelvis_idx = [1, 2]\n            self.params_type = 'smplx'\n        else:\n            self.j_regressor = self.j_smpl\n            self.pelvis_idx = [1, 2]\n            self.params_type = 'smpl'\n\n    def forward_smpl(self, dataset, smpl_dict, smpl_mask):\n        nhv = int(smpl_mask.sum())\n\n        if dataset in ['bedlam']:\n            out = self.smplx_neutral_11(\n                global_orient=smpl_dict['smplx_root_pose'][smpl_mask].reshape(-1, 3),\n                body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1, 21*3),\n                jaw_pose=smpl_dict['smplx_jaw_pose'][smpl_mask].reshape(-1, 3),\n                leye_pose=smpl_dict['smplx_leye_pose'][smpl_mask].reshape(-1, 3),\n                reye_pose=smpl_dict['smplx_reye_pose'][smpl_mask].reshape(-1, 3),\n                left_hand_pose=smpl_dict['smplx_left_hand_pose'][smpl_mask].reshape(-1, 15*3),\n                right_hand_pose=smpl_dict['smplx_right_hand_pose'][smpl_mask].reshape(-1, 15*3),\n                betas=smpl_dict['smplx_shape'][smpl_mask].reshape(-1, 11),\n                transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1, 3),\n                expression=self.smplx_neutral_11.expression.repeat(nhv, 1),\n            )\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n        elif dataset in ['3dpw']:\n            smpl_params = {\n                'global_orient': smpl_dict['smpl_root_pose'][smpl_mask].reshape(-1,3),\n                'body_pose': smpl_dict['smpl_body_pose'][smpl_mask].reshape(-1,23*3),\n                'betas': smpl_dict['smpl_shape'][smpl_mask].reshape(-1,10),\n                'transl': smpl_dict['smpl_transl'][smpl_mask].reshape(-1,3),\n                }\n            out = self.smpl[1](**smpl_params)\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n            # update verts/joints if this is not the right gender\n            if int(smpl_dict['smpl_gender_id'].max()) == 2:\n                out_female = self.smpl[2](**smpl_params)\n                idx = torch.where(smpl_dict['smpl_gender_id'] == 2)[1]\n                verts[idx] = out_female.vertices.reshape(nhv, -1, 3)[idx]\n                \n        elif dataset in ['emdb', 'emdb1', 'emdb2']:\n            gender = smpl_dict['smpl_gender_id'].max()\n            out = self.smpl[gender](\n                global_orient=smpl_dict['smpl_root_pose_w'][smpl_mask].reshape(-1,3),\n                body_pose=smpl_dict['smpl_body_pose'][smpl_mask].reshape(-1,23*3),\n                betas=smpl_dict['smpl_shape'][smpl_mask].reshape(-1,10),\n                transl=smpl_dict['smpl_transl_w'][smpl_mask].reshape(-1,3),\n            )\n            verts = out.vertices.reshape(nhv, -1, 3) # world space\n                \n        elif dataset in ['rich']:\n            gender = {1: 'male', 2: 'female'}[int(smpl_dict['smplx_gender_id'].max())]\n            out = self.smplx[gender](\n                global_orient=smpl_dict['smplx_global_orient'][smpl_mask].reshape(-1,3),\n                body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1,21*3),\n                jaw_pose=torch.zeros([nhv, 3]),\n                leye_pose=torch.zeros([nhv, 3]),\n                reye_pose=torch.zeros([nhv, 3]),\n                left_hand_pose=torch.zeros([nhv, 12]),\n                right_hand_pose=torch.zeros([nhv, 12]),\n                betas=smpl_dict['smplx_betas'][smpl_mask].reshape(-1,10),\n                transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1,3),\n                expression=torch.zeros([nhv, 10]),    \n            )\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n        if self.params_type == 'smplx':\n            verts = self.smplx2smpl @ verts\n        jts = self.j_regressor @ verts\n\n        return verts, jts\n\n    def update_smpl_gt(self, views):\n        target = {}\n\n        batch_size = views[0][\"img\"].shape[0]\n\n        smpl_keys = [k for k in views[0].keys() if 'smpl' in k]\n        smpl_dict = {\n            k: (stacked := torch.stack(\n                [view.pop(k) for view in views], dim=0)).view(-1, *stacked.shape[2:])\n            for k in smpl_keys\n        }   # Shape: (num_views * batch_size, 10, ...)\n        smpl_mask = smpl_dict['smpl_mask']\n        idx_h = torch.where(smpl_mask) # frame_idx, batch_idx, human_idx\n        K = torch.stack(\n            [view['camera_intrinsics'] for view in views], dim=0\n        )\n        K = K.view(-1, *K.shape[2:])\n        nhv = int(smpl_mask.sum())\n\n        # Get MHMR input image (high-res, square)\n        imgs = torch.stack([view[\"img\"] for view in views], dim=0)\n        imgs = imgs.view(-1, *imgs.shape[2:])\n        K_mhmr = resize_camera_intrinsics(K, *imgs.shape[2:], self.mhmr_img_res)\n        imgs_mhmr = pad_image(imgs, self.mhmr_img_res)\n\n        # SMPLX forward - BEDLAM\n        has_smplx_params = 1\n        out = self.smplx_neutral_11(\n            global_orient=smpl_dict['smplx_root_pose'][smpl_mask].reshape(-1, 3),\n            body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1, 21*3),\n            jaw_pose=smpl_dict['smplx_jaw_pose'][smpl_mask].reshape(-1, 3),\n            leye_pose=smpl_dict['smplx_leye_pose'][smpl_mask].reshape(-1, 3),\n            reye_pose=smpl_dict['smplx_reye_pose'][smpl_mask].reshape(-1, 3),\n            left_hand_pose=smpl_dict['smplx_left_hand_pose'][smpl_mask].reshape(-1, 15*3),\n            right_hand_pose=smpl_dict['smplx_right_hand_pose'][smpl_mask].reshape(-1, 15*3),\n            betas=smpl_dict['smplx_shape'][smpl_mask].reshape(-1, 11),\n            transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1, 3),\n            expression=self.smplx_neutral_11.expression.repeat(nhv, 1),\n        )\n        verts, jts = out.vertices.reshape(nhv, -1, 3), out.joints.reshape(nhv, -1, 3)\n\n        j2d = perspective_projection(jts, K[idx_h[0]])\n        v2d = perspective_projection(verts, K[idx_h[0]])\n\n        # Translation of the primary keypoint\n        root_joint_idx = JOINT_NAMES.index(self.person_center)\n        target['smpl_transl'] = jts[:,root_joint_idx] # [nhv,3]\n        target['smpl_transl_pelvis'] = jts[:,0] # [nhv,3]\n\n        # Fill in target\n        target['smpl_v3d'] = verts\n        target['smpl_j3d'] = jts\n        target['smpl_j2d'] = j2d\n        target['smpl_v2d'] = v2d\n\n        if has_smplx_params:\n            target['smpl_rotvec'] = torch.cat([smpl_dict['smplx_root_pose'],\n                                        smpl_dict['smplx_body_pose'],\n                                        smpl_dict['smplx_left_hand_pose'],\n                                        smpl_dict['smplx_right_hand_pose'],\n                                        smpl_dict['smplx_jaw_pose']],2)[smpl_mask] # [bs,nhmax]\n            target['smpl_rotmat'] = roma.rotvec_to_rotmat(target['smpl_rotvec'])\n            target['smpl_shape'] = smpl_dict['smplx_shape'][smpl_mask]\n\n        \n        true_shapes = torch.stack([view[\"true_shape\"] for view in views], dim=0)\n        if len(torch.unique(true_shapes, dim=0)) != 1:\n            raise NotImplementedError\n        \n        # Creating the target heatmap for the primary keypoint\n        pk = target['smpl_transl'].unsqueeze(1) # (nhv,3)\n        \n        # For 512 res (CUT3R, patch_size=16)\n        pk_loc = perspective_projection(pk, K[idx_h[0]]).squeeze(1) # original pixel uv coordinates (nhv,2): W, H\n        n_patch_16, pk_idx_16 = get_patch_uv(true_shapes[0][0], self.patch_size, pk_loc)\n        target['smpl_uv_16'] = pk_idx_16[:, [1, 0]]\n\n        # For 896 res (MHMR, patch_size=14)\n        pk_loc_mhmr = perspective_projection(pk, K_mhmr[idx_h[0]]).squeeze(1) # original pixel uv coordinates (nhv,2): W, H\n        n_patch_14, pk_idx_14 = get_patch_uv(self.mhmr_img_res, self.bb_patch_size, pk_loc_mhmr)\n        smpl_mask_14, visible_humans_14, scores_14 = get_score(n_patch_14, pk_idx_14, smpl_mask.clone())\n        target['smpl_uv'] = pk_idx_14[:, [1, 0]]\n\n        # Rebatch and Update with visibility indice\n        _target = {}\n        num_view = len(views)\n        max_humans = smpl_mask_14.shape[1]\n        idx_vis = torch.where(visible_humans_14)[0]\n\n        for k, v in target.items():\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, *v.shape[1:], \n                device=v.device, dtype=v.dtype,\n            )\n            full_out[smpl_mask_14] = v[idx_vis] # discard unvisible humans due to olccusion\n            _target[k] = full_out.chunk(num_view, dim=0) # .view(num_view, batch_size, *full_out.shape[1:])\n\n        _target['smpl_scores'] = scores_14.chunk(num_view, dim=0)\n        _target['smpl_mask'] = smpl_mask_14.chunk(num_view, dim=0)\n        _target['K_mhmr'] = K_mhmr.chunk(num_view, dim=0)\n        _target['img_mhmr'] = imgs_mhmr.chunk(num_view, dim=0)\n\n        if \"msk\" in views[0]:\n            msks = torch.stack([view[\"msk\"] for view in views], dim=0)\n            msks = msks.view(-1, *msks.shape[2:])\n            msks_mhmr = pad_image(msks, self.mhmr_img_res, pad_value=0.0)  # bs,288,512->bs,896,896\n            msks_mhmr = (msks_mhmr > 0.1).float()\n            _target['msk_mhmr'] = msks_mhmr.chunk(num_view, dim=0)\n\n        for i, v in enumerate(zip(*_target.values())):\n            views[i].update(dict(zip(_target.keys(), v)))\n\n        torch.cuda.empty_cache()\n    \n    def update_smpl_gt_eval(self, views, dataset):\n        from dust3r.utils.geometry import geotrf\n\n        target = {}\n        batch_size = views[0][\"img\"].shape[0]\n\n        smpl_keys = [k for k in views[0].keys() if 'smpl' in k]\n        smpl_dict = {\n            k: (stacked := torch.stack(\n                [view.pop(k) for view in views], dim=0)).view(-1, *stacked.shape[2:])\n            for k in smpl_keys\n        }   # Shape: (num_views * batch_size, 10, ...)\n        smpl_mask = smpl_dict['smpl_mask']\n        idx_h = torch.where(smpl_mask) # frame_idx, batch_idx, human_idx\n        K = torch.stack([view['camera_intrinsics'] for view in views], dim=0)\n        K = K.view(-1, *K.shape[2:])\n\n        # Get MHMR input image (high-res, square)\n        imgs = torch.stack([view[\"img\"] for view in views], dim=0)\n        imgs = imgs.view(-1, *imgs.shape[2:])\n        K_mhmr = resize_camera_intrinsics(K, *imgs.shape[2:], self.mhmr_img_res)\n        imgs_mhmr = pad_image(imgs, self.mhmr_img_res)\n\n        verts, jts = self.forward_smpl(dataset, smpl_dict, smpl_mask)\n\n        if dataset in ['emdb', 'emdb1', 'emdb2', 'rich']:\n            target['smpl_v3d_w'] = verts\n            target['smpl_j3d_w'] = jts\n            T_w2c = torch.stack([view['T_w2c'] for view in views], dim=0)\n            T_w2c = T_w2c.view(-1, *T_w2c.shape[2:])\n            target['smpl_v3d_c'] = geotrf(T_w2c[idx_h[0]], verts)\n            target['smpl_j3d_c'] = geotrf(T_w2c[idx_h[0]], jts)\n \n        else:\n            target['smpl_v3d_c'] = verts\n            target['smpl_j3d_c'] = jts\n            T_c2w = torch.stack([view['camera_pose'] for view in views], dim=0)\n            T_c2w = T_c2w.view(-1, *T_c2w.shape[2:])\n            target['smpl_v3d_w'] = geotrf(T_c2w[idx_h[0]], verts)\n            target['smpl_j3d_w'] = geotrf(T_c2w[idx_h[0]], jts)\n\n        target['smpl_j2d'] = perspective_projection(target['smpl_j3d_c'], K[idx_h[0]])\n        target['smpl_v2d'] = perspective_projection(target['smpl_v3d_c'], K[idx_h[0]])\n\n        # Rebatch and Update with visibility indice\n        _target = {}\n        num_view = len(views)\n        max_humans = smpl_mask.shape[1]\n        for k, v in target.items():\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, *v.shape[1:], \n                device=v.device, dtype=v.dtype,\n            )\n            full_out[smpl_mask] = v # discard unvisible humans due to olccusion\n            _target[k] = full_out.chunk(num_view, dim=0) # .view(num_view, batch_size, *full_out.shape[1:])\n\n        if self.use_fake_K:\n            K_mhmr = get_camera_parameters(self.mhmr_img_res, device=K.device) # if use pseudo K\n            K_mhmr = K_mhmr.expand(K.shape[0], -1, -1)\n\n        _target['smpl_mask'] = smpl_mask.chunk(num_view, dim=0)\n        _target['K_mhmr'] = K_mhmr.chunk(num_view, dim=0)\n        _target['img_mhmr'] = imgs_mhmr.chunk(num_view, dim=0)\n\n        if \"msk\" in views[0]:\n            msks = torch.stack([view[\"msk\"] for view in views], dim=0)\n            msks = msks.view(-1, *msks.shape[2:])\n            msks_mhmr = pad_image(msks, self.mhmr_img_res, pad_value=0.0)  # bs,288,512->bs,896,896\n            msks_mhmr = (msks_mhmr > 0.1).float()\n            _target['msk_mhmr'] = msks_mhmr.chunk(num_view, dim=0)\n\n        for i, v in enumerate(zip(*_target.values())):\n            views[i].update(dict(zip(_target.keys(), v)))\n\n        torch.cuda.empty_cache()\n\n\ndef get_patch_uv(imgshape, patch_size, pk_loc):\n    n_patch = imgshape // patch_size  # H, W\n    pk_idx = (pk_loc // patch_size).int()\n    return n_patch, pk_idx\n\ndef get_score(n_patch, pk_idx, smpl_mask):\n    # Scores & updating valid_humans according to occlusion - wap X and Y for scores only\n    idx_h = torch.where(smpl_mask)\n    nhv = int(smpl_mask.sum())\n    bs = smpl_mask.shape[0]\n    device = smpl_mask.device\n\n    if isinstance(n_patch, (int, float)):\n        patch_h, patch_w = int(n_patch), int(n_patch)\n    else:\n        patch_h, patch_w = n_patch[0], n_patch[1]\n\n    scores = torch.zeros((bs, patch_h, patch_w)).to(device)","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.get_patch_uv","uri":"program://Human3R/function/src.dust3r.smpl_model.get_patch_uv#L330-L333","kind":"function","name":"get_patch_uv","path":"src/dust3r/smpl_model.py","language":"python","start_line":330,"end_line":333,"context_start_line":310,"context_end_line":353,"code":"            K_mhmr = get_camera_parameters(self.mhmr_img_res, device=K.device) # if use pseudo K\n            K_mhmr = K_mhmr.expand(K.shape[0], -1, -1)\n\n        _target['smpl_mask'] = smpl_mask.chunk(num_view, dim=0)\n        _target['K_mhmr'] = K_mhmr.chunk(num_view, dim=0)\n        _target['img_mhmr'] = imgs_mhmr.chunk(num_view, dim=0)\n\n        if \"msk\" in views[0]:\n            msks = torch.stack([view[\"msk\"] for view in views], dim=0)\n            msks = msks.view(-1, *msks.shape[2:])\n            msks_mhmr = pad_image(msks, self.mhmr_img_res, pad_value=0.0)  # bs,288,512->bs,896,896\n            msks_mhmr = (msks_mhmr > 0.1).float()\n            _target['msk_mhmr'] = msks_mhmr.chunk(num_view, dim=0)\n\n        for i, v in enumerate(zip(*_target.values())):\n            views[i].update(dict(zip(_target.keys(), v)))\n\n        torch.cuda.empty_cache()\n\n\ndef get_patch_uv(imgshape, patch_size, pk_loc):\n    n_patch = imgshape // patch_size  # H, W\n    pk_idx = (pk_loc // patch_size).int()\n    return n_patch, pk_idx\n\ndef get_score(n_patch, pk_idx, smpl_mask):\n    # Scores & updating valid_humans according to occlusion - wap X and Y for scores only\n    idx_h = torch.where(smpl_mask)\n    nhv = int(smpl_mask.sum())\n    bs = smpl_mask.shape[0]\n    device = smpl_mask.device\n\n    if isinstance(n_patch, (int, float)):\n        patch_h, patch_w = int(n_patch), int(n_patch)\n    else:\n        patch_h, patch_w = n_patch[0], n_patch[1]\n\n    scores = torch.zeros((bs, patch_h, patch_w)).to(device)\n    visible_humans = torch.ones(nhv).to(device) # by default no occlusion so all visible\n\n    for k in range(nhv):\n        i = int(idx_h[0][k]) # index of the image\n        j = int(idx_h[1][k]) # index of the human in this image\n        _x = pk_idx[k,1] # patch center H","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.get_score","uri":"program://Human3R/function/src.dust3r.smpl_model.get_score#L335-L366","kind":"function","name":"get_score","path":"src/dust3r/smpl_model.py","language":"python","start_line":335,"end_line":366,"context_start_line":315,"context_end_line":386,"code":"        _target['img_mhmr'] = imgs_mhmr.chunk(num_view, dim=0)\n\n        if \"msk\" in views[0]:\n            msks = torch.stack([view[\"msk\"] for view in views], dim=0)\n            msks = msks.view(-1, *msks.shape[2:])\n            msks_mhmr = pad_image(msks, self.mhmr_img_res, pad_value=0.0)  # bs,288,512->bs,896,896\n            msks_mhmr = (msks_mhmr > 0.1).float()\n            _target['msk_mhmr'] = msks_mhmr.chunk(num_view, dim=0)\n\n        for i, v in enumerate(zip(*_target.values())):\n            views[i].update(dict(zip(_target.keys(), v)))\n\n        torch.cuda.empty_cache()\n\n\ndef get_patch_uv(imgshape, patch_size, pk_loc):\n    n_patch = imgshape // patch_size  # H, W\n    pk_idx = (pk_loc // patch_size).int()\n    return n_patch, pk_idx\n\ndef get_score(n_patch, pk_idx, smpl_mask):\n    # Scores & updating valid_humans according to occlusion - wap X and Y for scores only\n    idx_h = torch.where(smpl_mask)\n    nhv = int(smpl_mask.sum())\n    bs = smpl_mask.shape[0]\n    device = smpl_mask.device\n\n    if isinstance(n_patch, (int, float)):\n        patch_h, patch_w = int(n_patch), int(n_patch)\n    else:\n        patch_h, patch_w = n_patch[0], n_patch[1]\n\n    scores = torch.zeros((bs, patch_h, patch_w)).to(device)\n    visible_humans = torch.ones(nhv).to(device) # by default no occlusion so all visible\n\n    for k in range(nhv):\n        i = int(idx_h[0][k]) # index of the image\n        j = int(idx_h[1][k]) # index of the human in this image\n        _x = pk_idx[k,1] # patch center H\n        _y = pk_idx[k,0] # patch center W\n        # filter out heads out of cropping bounds\n        if _x >= 0 and _x < patch_h and _y >= 0 and _y < patch_w:\n            if scores[i,_x,_y] == 1:\n                smpl_mask[i,j] = 0\n                visible_humans[k] = 0\n            else:\n                scores[i,_x,_y] = 1\n        else:\n            smpl_mask[i,j] = 0\n            visible_humans[k] = 0\n    \n    return smpl_mask, visible_humans, scores\n\n\nimport torch.nn as nn\nfrom croco.models.blocks import Mlp_flex\n\nclass SMPLDecoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        target_dim=1,\n        mlp_ratio=1,\n        num_layers=2,\n    ):\n        super().__init__()\n        self.mlp = Mlp_flex(\n            in_features=hidden_size,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=target_dim,\n            num_layers=num_layers,\n            drop=0,","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.SMPLDecoder","uri":"program://Human3R/class/src.dust3r.smpl_model.SMPLDecoder#L372-L398","kind":"class","name":"SMPLDecoder","path":"src/dust3r/smpl_model.py","language":"python","start_line":372,"end_line":398,"context_start_line":352,"context_end_line":418,"code":"        j = int(idx_h[1][k]) # index of the human in this image\n        _x = pk_idx[k,1] # patch center H\n        _y = pk_idx[k,0] # patch center W\n        # filter out heads out of cropping bounds\n        if _x >= 0 and _x < patch_h and _y >= 0 and _y < patch_w:\n            if scores[i,_x,_y] == 1:\n                smpl_mask[i,j] = 0\n                visible_humans[k] = 0\n            else:\n                scores[i,_x,_y] = 1\n        else:\n            smpl_mask[i,j] = 0\n            visible_humans[k] = 0\n    \n    return smpl_mask, visible_humans, scores\n\n\nimport torch.nn as nn\nfrom croco.models.blocks import Mlp_flex\n\nclass SMPLDecoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        target_dim=1,\n        mlp_ratio=1,\n        num_layers=2,\n    ):\n        super().__init__()\n        self.mlp = Mlp_flex(\n            in_features=hidden_size,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=target_dim,\n            num_layers=num_layers,\n            drop=0,\n        )\n\n    def forward(\n        self,\n        feat,\n    ):\n        \"\"\"\n        feat: BxC\n        \"\"\"\n\n        pred = self.mlp(feat)\n        return pred\n\n\ndef regression_mlp(layers_sizes):\n    \"\"\"\n    Return a fully connected network.\n    \"\"\"\n    assert len(layers_sizes) >= 2\n    in_features = layers_sizes[0]\n    layers = []\n    for i in range(1, len(layers_sizes)-1):\n        out_features = layers_sizes[i]\n        layers.append(torch.nn.Linear(in_features, out_features))\n        layers.append(torch.nn.ReLU())\n        in_features = out_features\n    layers.append(torch.nn.Linear(in_features, layers_sizes[-1]))\n    return torch.nn.Sequential(*layers)\n\ndef apply_threshold(det_thresh, _scores):\n    \"\"\" Apply thresholding to detection scores; if stack_K is used and det_thresh is a list, apply to each channel separately \"\"\"\n    if isinstance(det_thresh, list):","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.regression_mlp","uri":"program://Human3R/function/src.dust3r.smpl_model.regression_mlp#L401-L414","kind":"function","name":"regression_mlp","path":"src/dust3r/smpl_model.py","language":"python","start_line":401,"end_line":414,"context_start_line":381,"context_end_line":434,"code":"        self.mlp = Mlp_flex(\n            in_features=hidden_size,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=target_dim,\n            num_layers=num_layers,\n            drop=0,\n        )\n\n    def forward(\n        self,\n        feat,\n    ):\n        \"\"\"\n        feat: BxC\n        \"\"\"\n\n        pred = self.mlp(feat)\n        return pred\n\n\ndef regression_mlp(layers_sizes):\n    \"\"\"\n    Return a fully connected network.\n    \"\"\"\n    assert len(layers_sizes) >= 2\n    in_features = layers_sizes[0]\n    layers = []\n    for i in range(1, len(layers_sizes)-1):\n        out_features = layers_sizes[i]\n        layers.append(torch.nn.Linear(in_features, out_features))\n        layers.append(torch.nn.ReLU())\n        in_features = out_features\n    layers.append(torch.nn.Linear(in_features, layers_sizes[-1]))\n    return torch.nn.Sequential(*layers)\n\ndef apply_threshold(det_thresh, _scores):\n    \"\"\" Apply thresholding to detection scores; if stack_K is used and det_thresh is a list, apply to each channel separately \"\"\"\n    if isinstance(det_thresh, list):\n        det_thresh = det_thresh[0]\n    idx = torch.where(_scores >= det_thresh)\n    return idx\n\ndef nms(heat, kernel=3):\n    \"\"\" easy non maximal supression (as in CenterNet) \"\"\"\n\n    if kernel not in [2, 4]:\n        pad = (kernel - 1) // 2\n    else:\n        if kernel == 2:\n            pad = 1\n        else:\n            pad = 2\n\n    hmax = nn.functional.max_pool2d( heat, (kernel, kernel), stride=1, padding=pad)","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.apply_threshold","uri":"program://Human3R/function/src.dust3r.smpl_model.apply_threshold#L416-L421","kind":"function","name":"apply_threshold","path":"src/dust3r/smpl_model.py","language":"python","start_line":416,"end_line":421,"context_start_line":396,"context_end_line":441,"code":"\n        pred = self.mlp(feat)\n        return pred\n\n\ndef regression_mlp(layers_sizes):\n    \"\"\"\n    Return a fully connected network.\n    \"\"\"\n    assert len(layers_sizes) >= 2\n    in_features = layers_sizes[0]\n    layers = []\n    for i in range(1, len(layers_sizes)-1):\n        out_features = layers_sizes[i]\n        layers.append(torch.nn.Linear(in_features, out_features))\n        layers.append(torch.nn.ReLU())\n        in_features = out_features\n    layers.append(torch.nn.Linear(in_features, layers_sizes[-1]))\n    return torch.nn.Sequential(*layers)\n\ndef apply_threshold(det_thresh, _scores):\n    \"\"\" Apply thresholding to detection scores; if stack_K is used and det_thresh is a list, apply to each channel separately \"\"\"\n    if isinstance(det_thresh, list):\n        det_thresh = det_thresh[0]\n    idx = torch.where(_scores >= det_thresh)\n    return idx\n\ndef nms(heat, kernel=3):\n    \"\"\" easy non maximal supression (as in CenterNet) \"\"\"\n\n    if kernel not in [2, 4]:\n        pad = (kernel - 1) // 2\n    else:\n        if kernel == 2:\n            pad = 1\n        else:\n            pad = 2\n\n    hmax = nn.functional.max_pool2d( heat, (kernel, kernel), stride=1, padding=pad)\n\n    if hmax.shape[2] > heat.shape[2]:\n        hmax = hmax[:, :, :heat.shape[2], :heat.shape[3]]\n\n    keep = (hmax == heat).float()\n\n    return heat * keep","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.nms","uri":"program://Human3R/function/src.dust3r.smpl_model.nms#L423-L441","kind":"function","name":"nms","path":"src/dust3r/smpl_model.py","language":"python","start_line":423,"end_line":441,"context_start_line":403,"context_end_line":441,"code":"    Return a fully connected network.\n    \"\"\"\n    assert len(layers_sizes) >= 2\n    in_features = layers_sizes[0]\n    layers = []\n    for i in range(1, len(layers_sizes)-1):\n        out_features = layers_sizes[i]\n        layers.append(torch.nn.Linear(in_features, out_features))\n        layers.append(torch.nn.ReLU())\n        in_features = out_features\n    layers.append(torch.nn.Linear(in_features, layers_sizes[-1]))\n    return torch.nn.Sequential(*layers)\n\ndef apply_threshold(det_thresh, _scores):\n    \"\"\" Apply thresholding to detection scores; if stack_K is used and det_thresh is a list, apply to each channel separately \"\"\"\n    if isinstance(det_thresh, list):\n        det_thresh = det_thresh[0]\n    idx = torch.where(_scores >= det_thresh)\n    return idx\n\ndef nms(heat, kernel=3):\n    \"\"\" easy non maximal supression (as in CenterNet) \"\"\"\n\n    if kernel not in [2, 4]:\n        pad = (kernel - 1) // 2\n    else:\n        if kernel == 2:\n            pad = 1\n        else:\n            pad = 2\n\n    hmax = nn.functional.max_pool2d( heat, (kernel, kernel), stride=1, padding=pad)\n\n    if hmax.shape[2] > heat.shape[2]:\n        hmax = hmax[:, :, :heat.shape[2], :heat.shape[3]]\n\n    keep = (hmax == heat).float()\n\n    return heat * keep","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.__init__","uri":"program://Human3R/function/src.dust3r.smpl_model.__init__#L373-L387","kind":"function","name":"__init__","path":"src/dust3r/smpl_model.py","language":"python","start_line":373,"end_line":387,"context_start_line":353,"context_end_line":407,"code":"        _x = pk_idx[k,1] # patch center H\n        _y = pk_idx[k,0] # patch center W\n        # filter out heads out of cropping bounds\n        if _x >= 0 and _x < patch_h and _y >= 0 and _y < patch_w:\n            if scores[i,_x,_y] == 1:\n                smpl_mask[i,j] = 0\n                visible_humans[k] = 0\n            else:\n                scores[i,_x,_y] = 1\n        else:\n            smpl_mask[i,j] = 0\n            visible_humans[k] = 0\n    \n    return smpl_mask, visible_humans, scores\n\n\nimport torch.nn as nn\nfrom croco.models.blocks import Mlp_flex\n\nclass SMPLDecoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        target_dim=1,\n        mlp_ratio=1,\n        num_layers=2,\n    ):\n        super().__init__()\n        self.mlp = Mlp_flex(\n            in_features=hidden_size,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=target_dim,\n            num_layers=num_layers,\n            drop=0,\n        )\n\n    def forward(\n        self,\n        feat,\n    ):\n        \"\"\"\n        feat: BxC\n        \"\"\"\n\n        pred = self.mlp(feat)\n        return pred\n\n\ndef regression_mlp(layers_sizes):\n    \"\"\"\n    Return a fully connected network.\n    \"\"\"\n    assert len(layers_sizes) >= 2\n    in_features = layers_sizes[0]\n    layers = []","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model._setup_dataset_config","uri":"program://Human3R/function/src.dust3r.smpl_model._setup_dataset_config#L56-L71","kind":"function","name":"_setup_dataset_config","path":"src/dust3r/smpl_model.py","language":"python","start_line":56,"end_line":71,"context_start_line":36,"context_end_line":91,"code":"            SMPLX_DIR, 'smplx', gender='neutral', use_pca=False, flat_hand_mean=True, num_betas=11).to(self.device)\n        self.smplx_neutral_10 = smplx.create(\n            SMPLX_DIR, 'smplx', gender='neutral', use_pca=False, flat_hand_mean=True, num_betas=10).to(self.device)\n        \n        # Evaluation\n        self.use_fake_K = eval_args.get('use_fake_K', False)\n        dataset = eval_args.get('dataset', None)\n        if dataset is not None:\n            self.smpl = [\n                smplx.create(SMPLX_DIR, 'smpl', gender=g).to(self.device) for g in ['neutral', 'male', 'female']]\n            self.smpl_faces = {'smpl': self.smpl[0].faces, 'smplx': self.smplx_neutral_11.faces}\n            with open(SMPLX2SMPL, 'rb') as f:\n                self.smplx2smpl = torch.from_numpy(pickle.load(f)['matrix'].astype(np.float32)).to(self.device)\n\n            if dataset in ['rich']:\n                self.smplx = {\n                    g: smplx.create(SMPLX_DIR, 'smplx', gender=g, num_pca_comps=12\n                                    ).to(self.device) for g in ['male', 'female']}\n            self._setup_dataset_config(dataset)        \n        \n    def _setup_dataset_config(self, dataset):\n        self.j_smpl = self.smpl[0].J_regressor[:24]\n        if dataset in ['3dpw']:\n            h36m_to_14 = [6, 5, 4, 1, 2, 3, 16, 15, 14, 11, 12, 13, 8, 10, 0, 7, 9][:14]\n            self.j_h36m = torch.Tensor(np.load('src/models/smpl/J_regressor_h36m.npy'))\n            self.j_regressor = self.j_h36m[h36m_to_14]\n            self.pelvis_idx = [2, 3]\n            self.params_type = 'smpl'\n        elif dataset in ['bedlam', 'rich']:\n            self.j_regressor = self.j_smpl\n            self.pelvis_idx = [1, 2]\n            self.params_type = 'smplx'\n        else:\n            self.j_regressor = self.j_smpl\n            self.pelvis_idx = [1, 2]\n            self.params_type = 'smpl'\n\n    def forward_smpl(self, dataset, smpl_dict, smpl_mask):\n        nhv = int(smpl_mask.sum())\n\n        if dataset in ['bedlam']:\n            out = self.smplx_neutral_11(\n                global_orient=smpl_dict['smplx_root_pose'][smpl_mask].reshape(-1, 3),\n                body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1, 21*3),\n                jaw_pose=smpl_dict['smplx_jaw_pose'][smpl_mask].reshape(-1, 3),\n                leye_pose=smpl_dict['smplx_leye_pose'][smpl_mask].reshape(-1, 3),\n                reye_pose=smpl_dict['smplx_reye_pose'][smpl_mask].reshape(-1, 3),\n                left_hand_pose=smpl_dict['smplx_left_hand_pose'][smpl_mask].reshape(-1, 15*3),\n                right_hand_pose=smpl_dict['smplx_right_hand_pose'][smpl_mask].reshape(-1, 15*3),\n                betas=smpl_dict['smplx_shape'][smpl_mask].reshape(-1, 11),\n                transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1, 3),\n                expression=self.smplx_neutral_11.expression.repeat(nhv, 1),\n            )\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n        elif dataset in ['3dpw']:","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.forward_smpl","uri":"program://Human3R/function/src.dust3r.smpl_model.forward_smpl#L73-L137","kind":"function","name":"forward_smpl","path":"src/dust3r/smpl_model.py","language":"python","start_line":73,"end_line":137,"context_start_line":53,"context_end_line":157,"code":"                                    ).to(self.device) for g in ['male', 'female']}\n            self._setup_dataset_config(dataset)        \n        \n    def _setup_dataset_config(self, dataset):\n        self.j_smpl = self.smpl[0].J_regressor[:24]\n        if dataset in ['3dpw']:\n            h36m_to_14 = [6, 5, 4, 1, 2, 3, 16, 15, 14, 11, 12, 13, 8, 10, 0, 7, 9][:14]\n            self.j_h36m = torch.Tensor(np.load('src/models/smpl/J_regressor_h36m.npy'))\n            self.j_regressor = self.j_h36m[h36m_to_14]\n            self.pelvis_idx = [2, 3]\n            self.params_type = 'smpl'\n        elif dataset in ['bedlam', 'rich']:\n            self.j_regressor = self.j_smpl\n            self.pelvis_idx = [1, 2]\n            self.params_type = 'smplx'\n        else:\n            self.j_regressor = self.j_smpl\n            self.pelvis_idx = [1, 2]\n            self.params_type = 'smpl'\n\n    def forward_smpl(self, dataset, smpl_dict, smpl_mask):\n        nhv = int(smpl_mask.sum())\n\n        if dataset in ['bedlam']:\n            out = self.smplx_neutral_11(\n                global_orient=smpl_dict['smplx_root_pose'][smpl_mask].reshape(-1, 3),\n                body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1, 21*3),\n                jaw_pose=smpl_dict['smplx_jaw_pose'][smpl_mask].reshape(-1, 3),\n                leye_pose=smpl_dict['smplx_leye_pose'][smpl_mask].reshape(-1, 3),\n                reye_pose=smpl_dict['smplx_reye_pose'][smpl_mask].reshape(-1, 3),\n                left_hand_pose=smpl_dict['smplx_left_hand_pose'][smpl_mask].reshape(-1, 15*3),\n                right_hand_pose=smpl_dict['smplx_right_hand_pose'][smpl_mask].reshape(-1, 15*3),\n                betas=smpl_dict['smplx_shape'][smpl_mask].reshape(-1, 11),\n                transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1, 3),\n                expression=self.smplx_neutral_11.expression.repeat(nhv, 1),\n            )\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n        elif dataset in ['3dpw']:\n            smpl_params = {\n                'global_orient': smpl_dict['smpl_root_pose'][smpl_mask].reshape(-1,3),\n                'body_pose': smpl_dict['smpl_body_pose'][smpl_mask].reshape(-1,23*3),\n                'betas': smpl_dict['smpl_shape'][smpl_mask].reshape(-1,10),\n                'transl': smpl_dict['smpl_transl'][smpl_mask].reshape(-1,3),\n                }\n            out = self.smpl[1](**smpl_params)\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n            # update verts/joints if this is not the right gender\n            if int(smpl_dict['smpl_gender_id'].max()) == 2:\n                out_female = self.smpl[2](**smpl_params)\n                idx = torch.where(smpl_dict['smpl_gender_id'] == 2)[1]\n                verts[idx] = out_female.vertices.reshape(nhv, -1, 3)[idx]\n                \n        elif dataset in ['emdb', 'emdb1', 'emdb2']:\n            gender = smpl_dict['smpl_gender_id'].max()\n            out = self.smpl[gender](\n                global_orient=smpl_dict['smpl_root_pose_w'][smpl_mask].reshape(-1,3),\n                body_pose=smpl_dict['smpl_body_pose'][smpl_mask].reshape(-1,23*3),\n                betas=smpl_dict['smpl_shape'][smpl_mask].reshape(-1,10),\n                transl=smpl_dict['smpl_transl_w'][smpl_mask].reshape(-1,3),\n            )\n            verts = out.vertices.reshape(nhv, -1, 3) # world space\n                \n        elif dataset in ['rich']:\n            gender = {1: 'male', 2: 'female'}[int(smpl_dict['smplx_gender_id'].max())]\n            out = self.smplx[gender](\n                global_orient=smpl_dict['smplx_global_orient'][smpl_mask].reshape(-1,3),\n                body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1,21*3),\n                jaw_pose=torch.zeros([nhv, 3]),\n                leye_pose=torch.zeros([nhv, 3]),\n                reye_pose=torch.zeros([nhv, 3]),\n                left_hand_pose=torch.zeros([nhv, 12]),\n                right_hand_pose=torch.zeros([nhv, 12]),\n                betas=smpl_dict['smplx_betas'][smpl_mask].reshape(-1,10),\n                transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1,3),\n                expression=torch.zeros([nhv, 10]),    \n            )\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n        if self.params_type == 'smplx':\n            verts = self.smplx2smpl @ verts\n        jts = self.j_regressor @ verts\n\n        return verts, jts\n\n    def update_smpl_gt(self, views):\n        target = {}\n\n        batch_size = views[0][\"img\"].shape[0]\n\n        smpl_keys = [k for k in views[0].keys() if 'smpl' in k]\n        smpl_dict = {\n            k: (stacked := torch.stack(\n                [view.pop(k) for view in views], dim=0)).view(-1, *stacked.shape[2:])\n            for k in smpl_keys\n        }   # Shape: (num_views * batch_size, 10, ...)\n        smpl_mask = smpl_dict['smpl_mask']\n        idx_h = torch.where(smpl_mask) # frame_idx, batch_idx, human_idx\n        K = torch.stack(\n            [view['camera_intrinsics'] for view in views], dim=0\n        )\n        K = K.view(-1, *K.shape[2:])\n        nhv = int(smpl_mask.sum())\n","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.update_smpl_gt","uri":"program://Human3R/function/src.dust3r.smpl_model.update_smpl_gt#L139-L251","kind":"function","name":"update_smpl_gt","path":"src/dust3r/smpl_model.py","language":"python","start_line":139,"end_line":251,"context_start_line":119,"context_end_line":271,"code":"            out = self.smplx[gender](\n                global_orient=smpl_dict['smplx_global_orient'][smpl_mask].reshape(-1,3),\n                body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1,21*3),\n                jaw_pose=torch.zeros([nhv, 3]),\n                leye_pose=torch.zeros([nhv, 3]),\n                reye_pose=torch.zeros([nhv, 3]),\n                left_hand_pose=torch.zeros([nhv, 12]),\n                right_hand_pose=torch.zeros([nhv, 12]),\n                betas=smpl_dict['smplx_betas'][smpl_mask].reshape(-1,10),\n                transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1,3),\n                expression=torch.zeros([nhv, 10]),    \n            )\n            verts = out.vertices.reshape(nhv, -1, 3)\n\n        if self.params_type == 'smplx':\n            verts = self.smplx2smpl @ verts\n        jts = self.j_regressor @ verts\n\n        return verts, jts\n\n    def update_smpl_gt(self, views):\n        target = {}\n\n        batch_size = views[0][\"img\"].shape[0]\n\n        smpl_keys = [k for k in views[0].keys() if 'smpl' in k]\n        smpl_dict = {\n            k: (stacked := torch.stack(\n                [view.pop(k) for view in views], dim=0)).view(-1, *stacked.shape[2:])\n            for k in smpl_keys\n        }   # Shape: (num_views * batch_size, 10, ...)\n        smpl_mask = smpl_dict['smpl_mask']\n        idx_h = torch.where(smpl_mask) # frame_idx, batch_idx, human_idx\n        K = torch.stack(\n            [view['camera_intrinsics'] for view in views], dim=0\n        )\n        K = K.view(-1, *K.shape[2:])\n        nhv = int(smpl_mask.sum())\n\n        # Get MHMR input image (high-res, square)\n        imgs = torch.stack([view[\"img\"] for view in views], dim=0)\n        imgs = imgs.view(-1, *imgs.shape[2:])\n        K_mhmr = resize_camera_intrinsics(K, *imgs.shape[2:], self.mhmr_img_res)\n        imgs_mhmr = pad_image(imgs, self.mhmr_img_res)\n\n        # SMPLX forward - BEDLAM\n        has_smplx_params = 1\n        out = self.smplx_neutral_11(\n            global_orient=smpl_dict['smplx_root_pose'][smpl_mask].reshape(-1, 3),\n            body_pose=smpl_dict['smplx_body_pose'][smpl_mask].reshape(-1, 21*3),\n            jaw_pose=smpl_dict['smplx_jaw_pose'][smpl_mask].reshape(-1, 3),\n            leye_pose=smpl_dict['smplx_leye_pose'][smpl_mask].reshape(-1, 3),\n            reye_pose=smpl_dict['smplx_reye_pose'][smpl_mask].reshape(-1, 3),\n            left_hand_pose=smpl_dict['smplx_left_hand_pose'][smpl_mask].reshape(-1, 15*3),\n            right_hand_pose=smpl_dict['smplx_right_hand_pose'][smpl_mask].reshape(-1, 15*3),\n            betas=smpl_dict['smplx_shape'][smpl_mask].reshape(-1, 11),\n            transl=smpl_dict['smplx_transl'][smpl_mask].reshape(-1, 3),\n            expression=self.smplx_neutral_11.expression.repeat(nhv, 1),\n        )\n        verts, jts = out.vertices.reshape(nhv, -1, 3), out.joints.reshape(nhv, -1, 3)\n\n        j2d = perspective_projection(jts, K[idx_h[0]])\n        v2d = perspective_projection(verts, K[idx_h[0]])\n\n        # Translation of the primary keypoint\n        root_joint_idx = JOINT_NAMES.index(self.person_center)\n        target['smpl_transl'] = jts[:,root_joint_idx] # [nhv,3]\n        target['smpl_transl_pelvis'] = jts[:,0] # [nhv,3]\n\n        # Fill in target\n        target['smpl_v3d'] = verts\n        target['smpl_j3d'] = jts\n        target['smpl_j2d'] = j2d\n        target['smpl_v2d'] = v2d\n\n        if has_smplx_params:\n            target['smpl_rotvec'] = torch.cat([smpl_dict['smplx_root_pose'],\n                                        smpl_dict['smplx_body_pose'],\n                                        smpl_dict['smplx_left_hand_pose'],\n                                        smpl_dict['smplx_right_hand_pose'],\n                                        smpl_dict['smplx_jaw_pose']],2)[smpl_mask] # [bs,nhmax]\n            target['smpl_rotmat'] = roma.rotvec_to_rotmat(target['smpl_rotvec'])\n            target['smpl_shape'] = smpl_dict['smplx_shape'][smpl_mask]\n\n        \n        true_shapes = torch.stack([view[\"true_shape\"] for view in views], dim=0)\n        if len(torch.unique(true_shapes, dim=0)) != 1:\n            raise NotImplementedError\n        \n        # Creating the target heatmap for the primary keypoint\n        pk = target['smpl_transl'].unsqueeze(1) # (nhv,3)\n        \n        # For 512 res (CUT3R, patch_size=16)\n        pk_loc = perspective_projection(pk, K[idx_h[0]]).squeeze(1) # original pixel uv coordinates (nhv,2): W, H\n        n_patch_16, pk_idx_16 = get_patch_uv(true_shapes[0][0], self.patch_size, pk_loc)\n        target['smpl_uv_16'] = pk_idx_16[:, [1, 0]]\n\n        # For 896 res (MHMR, patch_size=14)\n        pk_loc_mhmr = perspective_projection(pk, K_mhmr[idx_h[0]]).squeeze(1) # original pixel uv coordinates (nhv,2): W, H\n        n_patch_14, pk_idx_14 = get_patch_uv(self.mhmr_img_res, self.bb_patch_size, pk_loc_mhmr)\n        smpl_mask_14, visible_humans_14, scores_14 = get_score(n_patch_14, pk_idx_14, smpl_mask.clone())\n        target['smpl_uv'] = pk_idx_14[:, [1, 0]]\n\n        # Rebatch and Update with visibility indice\n        _target = {}\n        num_view = len(views)\n        max_humans = smpl_mask_14.shape[1]\n        idx_vis = torch.where(visible_humans_14)[0]\n\n        for k, v in target.items():\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, *v.shape[1:], \n                device=v.device, dtype=v.dtype,\n            )\n            full_out[smpl_mask_14] = v[idx_vis] # discard unvisible humans due to olccusion\n            _target[k] = full_out.chunk(num_view, dim=0) # .view(num_view, batch_size, *full_out.shape[1:])\n\n        _target['smpl_scores'] = scores_14.chunk(num_view, dim=0)\n        _target['smpl_mask'] = smpl_mask_14.chunk(num_view, dim=0)\n        _target['K_mhmr'] = K_mhmr.chunk(num_view, dim=0)\n        _target['img_mhmr'] = imgs_mhmr.chunk(num_view, dim=0)\n\n        if \"msk\" in views[0]:\n            msks = torch.stack([view[\"msk\"] for view in views], dim=0)\n            msks = msks.view(-1, *msks.shape[2:])\n            msks_mhmr = pad_image(msks, self.mhmr_img_res, pad_value=0.0)  # bs,288,512->bs,896,896\n            msks_mhmr = (msks_mhmr > 0.1).float()\n            _target['msk_mhmr'] = msks_mhmr.chunk(num_view, dim=0)\n\n        for i, v in enumerate(zip(*_target.values())):\n            views[i].update(dict(zip(_target.keys(), v)))\n\n        torch.cuda.empty_cache()\n    \n    def update_smpl_gt_eval(self, views, dataset):\n        from dust3r.utils.geometry import geotrf\n\n        target = {}\n        batch_size = views[0][\"img\"].shape[0]\n\n        smpl_keys = [k for k in views[0].keys() if 'smpl' in k]\n        smpl_dict = {\n            k: (stacked := torch.stack(\n                [view.pop(k) for view in views], dim=0)).view(-1, *stacked.shape[2:])\n            for k in smpl_keys\n        }   # Shape: (num_views * batch_size, 10, ...)\n        smpl_mask = smpl_dict['smpl_mask']\n        idx_h = torch.where(smpl_mask) # frame_idx, batch_idx, human_idx\n        K = torch.stack([view['camera_intrinsics'] for view in views], dim=0)\n        K = K.view(-1, *K.shape[2:])\n\n        # Get MHMR input image (high-res, square)\n        imgs = torch.stack([view[\"img\"] for view in views], dim=0)","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.update_smpl_gt_eval","uri":"program://Human3R/function/src.dust3r.smpl_model.update_smpl_gt_eval#L253-L327","kind":"function","name":"update_smpl_gt_eval","path":"src/dust3r/smpl_model.py","language":"python","start_line":253,"end_line":327,"context_start_line":233,"context_end_line":347,"code":"            full_out[smpl_mask_14] = v[idx_vis] # discard unvisible humans due to olccusion\n            _target[k] = full_out.chunk(num_view, dim=0) # .view(num_view, batch_size, *full_out.shape[1:])\n\n        _target['smpl_scores'] = scores_14.chunk(num_view, dim=0)\n        _target['smpl_mask'] = smpl_mask_14.chunk(num_view, dim=0)\n        _target['K_mhmr'] = K_mhmr.chunk(num_view, dim=0)\n        _target['img_mhmr'] = imgs_mhmr.chunk(num_view, dim=0)\n\n        if \"msk\" in views[0]:\n            msks = torch.stack([view[\"msk\"] for view in views], dim=0)\n            msks = msks.view(-1, *msks.shape[2:])\n            msks_mhmr = pad_image(msks, self.mhmr_img_res, pad_value=0.0)  # bs,288,512->bs,896,896\n            msks_mhmr = (msks_mhmr > 0.1).float()\n            _target['msk_mhmr'] = msks_mhmr.chunk(num_view, dim=0)\n\n        for i, v in enumerate(zip(*_target.values())):\n            views[i].update(dict(zip(_target.keys(), v)))\n\n        torch.cuda.empty_cache()\n    \n    def update_smpl_gt_eval(self, views, dataset):\n        from dust3r.utils.geometry import geotrf\n\n        target = {}\n        batch_size = views[0][\"img\"].shape[0]\n\n        smpl_keys = [k for k in views[0].keys() if 'smpl' in k]\n        smpl_dict = {\n            k: (stacked := torch.stack(\n                [view.pop(k) for view in views], dim=0)).view(-1, *stacked.shape[2:])\n            for k in smpl_keys\n        }   # Shape: (num_views * batch_size, 10, ...)\n        smpl_mask = smpl_dict['smpl_mask']\n        idx_h = torch.where(smpl_mask) # frame_idx, batch_idx, human_idx\n        K = torch.stack([view['camera_intrinsics'] for view in views], dim=0)\n        K = K.view(-1, *K.shape[2:])\n\n        # Get MHMR input image (high-res, square)\n        imgs = torch.stack([view[\"img\"] for view in views], dim=0)\n        imgs = imgs.view(-1, *imgs.shape[2:])\n        K_mhmr = resize_camera_intrinsics(K, *imgs.shape[2:], self.mhmr_img_res)\n        imgs_mhmr = pad_image(imgs, self.mhmr_img_res)\n\n        verts, jts = self.forward_smpl(dataset, smpl_dict, smpl_mask)\n\n        if dataset in ['emdb', 'emdb1', 'emdb2', 'rich']:\n            target['smpl_v3d_w'] = verts\n            target['smpl_j3d_w'] = jts\n            T_w2c = torch.stack([view['T_w2c'] for view in views], dim=0)\n            T_w2c = T_w2c.view(-1, *T_w2c.shape[2:])\n            target['smpl_v3d_c'] = geotrf(T_w2c[idx_h[0]], verts)\n            target['smpl_j3d_c'] = geotrf(T_w2c[idx_h[0]], jts)\n \n        else:\n            target['smpl_v3d_c'] = verts\n            target['smpl_j3d_c'] = jts\n            T_c2w = torch.stack([view['camera_pose'] for view in views], dim=0)\n            T_c2w = T_c2w.view(-1, *T_c2w.shape[2:])\n            target['smpl_v3d_w'] = geotrf(T_c2w[idx_h[0]], verts)\n            target['smpl_j3d_w'] = geotrf(T_c2w[idx_h[0]], jts)\n\n        target['smpl_j2d'] = perspective_projection(target['smpl_j3d_c'], K[idx_h[0]])\n        target['smpl_v2d'] = perspective_projection(target['smpl_v3d_c'], K[idx_h[0]])\n\n        # Rebatch and Update with visibility indice\n        _target = {}\n        num_view = len(views)\n        max_humans = smpl_mask.shape[1]\n        for k, v in target.items():\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, *v.shape[1:], \n                device=v.device, dtype=v.dtype,\n            )\n            full_out[smpl_mask] = v # discard unvisible humans due to olccusion\n            _target[k] = full_out.chunk(num_view, dim=0) # .view(num_view, batch_size, *full_out.shape[1:])\n\n        if self.use_fake_K:\n            K_mhmr = get_camera_parameters(self.mhmr_img_res, device=K.device) # if use pseudo K\n            K_mhmr = K_mhmr.expand(K.shape[0], -1, -1)\n\n        _target['smpl_mask'] = smpl_mask.chunk(num_view, dim=0)\n        _target['K_mhmr'] = K_mhmr.chunk(num_view, dim=0)\n        _target['img_mhmr'] = imgs_mhmr.chunk(num_view, dim=0)\n\n        if \"msk\" in views[0]:\n            msks = torch.stack([view[\"msk\"] for view in views], dim=0)\n            msks = msks.view(-1, *msks.shape[2:])\n            msks_mhmr = pad_image(msks, self.mhmr_img_res, pad_value=0.0)  # bs,288,512->bs,896,896\n            msks_mhmr = (msks_mhmr > 0.1).float()\n            _target['msk_mhmr'] = msks_mhmr.chunk(num_view, dim=0)\n\n        for i, v in enumerate(zip(*_target.values())):\n            views[i].update(dict(zip(_target.keys(), v)))\n\n        torch.cuda.empty_cache()\n\n\ndef get_patch_uv(imgshape, patch_size, pk_loc):\n    n_patch = imgshape // patch_size  # H, W\n    pk_idx = (pk_loc // patch_size).int()\n    return n_patch, pk_idx\n\ndef get_score(n_patch, pk_idx, smpl_mask):\n    # Scores & updating valid_humans according to occlusion - wap X and Y for scores only\n    idx_h = torch.where(smpl_mask)\n    nhv = int(smpl_mask.sum())\n    bs = smpl_mask.shape[0]\n    device = smpl_mask.device\n\n    if isinstance(n_patch, (int, float)):\n        patch_h, patch_w = int(n_patch), int(n_patch)\n    else:\n        patch_h, patch_w = n_patch[0], n_patch[1]\n\n    scores = torch.zeros((bs, patch_h, patch_w)).to(device)","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.smpl_model.forward","uri":"program://Human3R/function/src.dust3r.smpl_model.forward#L389-L398","kind":"function","name":"forward","path":"src/dust3r/smpl_model.py","language":"python","start_line":389,"end_line":398,"context_start_line":369,"context_end_line":418,"code":"import torch.nn as nn\nfrom croco.models.blocks import Mlp_flex\n\nclass SMPLDecoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        target_dim=1,\n        mlp_ratio=1,\n        num_layers=2,\n    ):\n        super().__init__()\n        self.mlp = Mlp_flex(\n            in_features=hidden_size,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=target_dim,\n            num_layers=num_layers,\n            drop=0,\n        )\n\n    def forward(\n        self,\n        feat,\n    ):\n        \"\"\"\n        feat: BxC\n        \"\"\"\n\n        pred = self.mlp(feat)\n        return pred\n\n\ndef regression_mlp(layers_sizes):\n    \"\"\"\n    Return a fully connected network.\n    \"\"\"\n    assert len(layers_sizes) >= 2\n    in_features = layers_sizes[0]\n    layers = []\n    for i in range(1, len(layers_sizes)-1):\n        out_features = layers_sizes[i]\n        layers.append(torch.nn.Linear(in_features, out_features))\n        layers.append(torch.nn.ReLU())\n        in_features = out_features\n    layers.append(torch.nn.Linear(in_features, layers_sizes[-1]))\n    return torch.nn.Sequential(*layers)\n\ndef apply_threshold(det_thresh, _scores):\n    \"\"\" Apply thresholding to detection scores; if stack_K is used and det_thresh is a list, apply to each channel separately \"\"\"\n    if isinstance(det_thresh, list):","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference","uri":"program://Human3R/module/src.dust3r.inference#L1-L276","kind":"module","name":"src.dust3r.inference","path":"src/dust3r/inference.py","language":"python","start_line":1,"end_line":276,"context_start_line":1,"context_end_line":276,"code":"import tqdm\nimport torch\nfrom dust3r.utils.device import to_cpu, collate_with_cat\nfrom dust3r.utils.misc import invalid_to_nans\nfrom dust3r.utils.geometry import depthmap_to_pts3d, geotrf\nfrom dust3r.model import ARCroco3DStereo\nfrom dust3r.smpl_model import SMPLModel\nfrom accelerate import Accelerator\nimport re\n\n\ndef custom_sort_key(key):\n    text = key.split(\"/\")\n    if len(text) > 1:\n        text, num = text[0], text[-1]\n        return (text, int(num))\n    else:\n        return (key, -1)\n\n\ndef merge_chunk_dict(old_dict, curr_dict, add_number):\n    new_dict = {}\n    for key, value in curr_dict.items():\n\n        match = re.search(r\"(\\d+)$\", key)\n        if match:\n\n            num_part = int(match.group()) + add_number\n\n            new_key = re.sub(r\"(\\d+)$\", str(num_part), key, 1)\n            new_dict[new_key] = value\n        else:\n            new_dict[key] = value\n    new_dict = old_dict | new_dict\n    return {k: new_dict[k] for k in sorted(new_dict.keys(), key=custom_sort_key)}\n\n\ndef _interleave_imgs(img1, img2):\n    res = {}\n    for key, value1 in img1.items():\n        value2 = img2[key]\n        if isinstance(value1, torch.Tensor):\n            value = torch.stack((value1, value2), dim=1).flatten(0, 1)\n        else:\n            value = [x for pair in zip(value1, value2) for x in pair]\n        res[key] = value\n    return res\n\n\ndef make_batch_symmetric(batch):\n    view1, view2 = batch\n    view1, view2 = (_interleave_imgs(view1, view2), _interleave_imgs(view2, view1))\n    return view1, view2\n\n\ndef loss_of_one_batch(\n    batch,\n    model,\n    criterion,\n    accelerator: Accelerator,\n    symmetrize_batch=False,\n    use_amp=False,\n    ret=None,\n    img_mask=None,\n    inference=False,\n    smpl_model: SMPLModel = None\n):\n    if len(batch) > 2:\n        assert (\n            symmetrize_batch is False\n        ), \"cannot symmetrize batch with more than 2 views\"\n    if symmetrize_batch:\n        batch = make_batch_symmetric(batch)\n\n    with torch.cuda.amp.autocast(enabled=not inference):\n        if inference:\n            output, state_args = model(batch, ret_state=True, inference=True)\n            preds, batch = output.ress, output.views\n            result = dict(views=batch, pred=preds)\n            return result[ret] if ret else result, state_args\n        else:\n            smpl_model.update_smpl_gt(batch)\n            output = model(batch)\n            preds, batch = output.ress, output.views\n\n        with torch.cuda.amp.autocast(enabled=False):\n            loss = criterion(batch, preds) if criterion is not None else None\n\n    result = dict(views=batch, pred=preds, loss=loss)\n    return result[ret] if ret else result\n\n@torch.no_grad()\ndef inference(groups, model, device, verbose=True):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for view in groups:\n        for name in view.keys():  # pseudo_focal\n            if name in ignore_keys:\n                continue\n            if isinstance(view[name], tuple) or isinstance(view[name], list):\n                view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n            else:\n                view[name] = view[name].to(device, non_blocking=True)\n\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    res, state_args = loss_of_one_batch(groups, model, None, None, inference=True)\n    result = to_cpu(res)\n    return result, state_args\n\n\n@torch.no_grad()\ndef inference_step(view, state_args, model, device, verbose=True):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for name in view.keys():  # pseudo_focal\n        if name in ignore_keys:\n            continue\n        if isinstance(view[name], tuple) or isinstance(view[name], list):\n            view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n        else:\n            view[name] = view[name].to(device, non_blocking=True)\n\n    with torch.cuda.amp.autocast(enabled=False):\n        state_feat, state_pos, init_state_feat, mem, init_mem = state_args\n        pred, _ = model.inference_step(\n            view, state_feat, state_pos, init_state_feat, mem, init_mem\n        )\n\n    res = dict(pred=pred)\n    result = to_cpu(res)\n    return result\n\n\n@torch.no_grad()\ndef inference_recurrent(groups, model, device, verbose=True):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for view in groups:\n        for name in view.keys():  # pseudo_focal\n            if name in ignore_keys:\n                continue\n            if isinstance(view[name], tuple) or isinstance(view[name], list):\n                view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n            else:\n                view[name] = view[name].to(device, non_blocking=True)\n\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    with torch.cuda.amp.autocast(enabled=False):\n        preds, batch, state_args = model.forward_recurrent(\n            groups, device, ret_state=True\n        )\n        res = dict(views=batch, pred=preds)\n    result = to_cpu(res)\n    return result, state_args\n\n@torch.no_grad()\ndef inference_recurrent_lighter(groups, model, device, verbose=True, is_naive=False, use_ttt3r=False):\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    with torch.cuda.amp.autocast(enabled=False):\n        if is_naive:\n            preds, batch, state_args = model.forward_recurrent_lighter_naive(\n            groups, device, ret_state=True, use_ttt3r=use_ttt3r\n        )\n        else:\n            preds, batch, state_args = model.forward_recurrent_lighter(\n                groups, device, ret_state=True, use_ttt3r=use_ttt3r\n            )\n        res = dict(views=batch, pred=preds)\n    return res, state_args\n\ndef check_if_same_size(pairs):\n    shapes1 = [img1[\"img\"].shape[-2:] for img1, img2 in pairs]\n    shapes2 = [img2[\"img\"].shape[-2:] for img1, img2 in pairs]\n    return all(shapes1[0] == s for s in shapes1) and all(\n        shapes2[0] == s for s in shapes2\n    )\n\n\ndef get_pred_pts3d(gt, pred, use_pose=False, inplace=False):\n    if \"depth\" in pred and \"pseudo_focal\" in pred:\n        try:\n            pp = gt[\"camera_intrinsics\"][..., :2, 2]\n        except KeyError:\n            pp = None\n        pts3d = depthmap_to_pts3d(**pred, pp=pp)\n\n    elif \"pts3d\" in pred:\n\n        pts3d = pred[\"pts3d\"]\n\n    elif \"pts3d_in_other_view\" in pred:\n\n        assert use_pose is True\n        return (\n            pred[\"pts3d_in_other_view\"]\n            if inplace\n            else pred[\"pts3d_in_other_view\"].clone()\n        )\n\n    if use_pose:\n        camera_pose = pred.get(\"camera_pose\")\n        assert camera_pose is not None\n        pts3d = geotrf(camera_pose, pts3d)\n\n    return pts3d\n\n\ndef find_opt_scaling(\n    gt_pts1,\n    gt_pts2,\n    pr_pts1,\n    pr_pts2=None,\n    fit_mode=\"weiszfeld_stop_grad\",\n    valid1=None,\n    valid2=None,\n):\n    assert gt_pts1.ndim == pr_pts1.ndim == 4\n    assert gt_pts1.shape == pr_pts1.shape\n    if gt_pts2 is not None:\n        assert gt_pts2.ndim == pr_pts2.ndim == 4\n        assert gt_pts2.shape == pr_pts2.shape\n\n    nan_gt_pts1 = invalid_to_nans(gt_pts1, valid1).flatten(1, 2)\n    nan_gt_pts2 = (\n        invalid_to_nans(gt_pts2, valid2).flatten(1, 2) if gt_pts2 is not None else None\n    )\n\n    pr_pts1 = invalid_to_nans(pr_pts1, valid1).flatten(1, 2)\n    pr_pts2 = (\n        invalid_to_nans(pr_pts2, valid2).flatten(1, 2) if pr_pts2 is not None else None\n    )\n\n    all_gt = (\n        torch.cat((nan_gt_pts1, nan_gt_pts2), dim=1)\n        if gt_pts2 is not None\n        else nan_gt_pts1\n    )\n    all_pr = torch.cat((pr_pts1, pr_pts2), dim=1) if pr_pts2 is not None else pr_pts1\n\n    dot_gt_pr = (all_pr * all_gt).sum(dim=-1)\n    dot_gt_gt = all_gt.square().sum(dim=-1)\n\n    if fit_mode.startswith(\"avg\"):\n\n        scaling = dot_gt_pr.nanmean(dim=1) / dot_gt_gt.nanmean(dim=1)\n    elif fit_mode.startswith(\"median\"):\n        scaling = (dot_gt_pr / dot_gt_gt).nanmedian(dim=1).values\n    elif fit_mode.startswith(\"weiszfeld\"):\n\n        scaling = dot_gt_pr.nanmean(dim=1) / dot_gt_gt.nanmean(dim=1)\n\n        for iter in range(10):\n\n            dis = (all_pr - scaling.view(-1, 1, 1) * all_gt).norm(dim=-1)\n\n            w = dis.clip_(min=1e-8).reciprocal()\n\n            scaling = (w * dot_gt_pr).nanmean(dim=1) / (w * dot_gt_gt).nanmean(dim=1)\n    else:\n        raise ValueError(f\"bad {fit_mode=}\")\n\n    if fit_mode.endswith(\"stop_grad\"):\n        scaling = scaling.detach()\n\n    scaling = scaling.clip(min=1e-3)\n\n    return scaling","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.custom_sort_key","uri":"program://Human3R/function/src.dust3r.inference.custom_sort_key#L12-L18","kind":"function","name":"custom_sort_key","path":"src/dust3r/inference.py","language":"python","start_line":12,"end_line":18,"context_start_line":1,"context_end_line":38,"code":"import tqdm\nimport torch\nfrom dust3r.utils.device import to_cpu, collate_with_cat\nfrom dust3r.utils.misc import invalid_to_nans\nfrom dust3r.utils.geometry import depthmap_to_pts3d, geotrf\nfrom dust3r.model import ARCroco3DStereo\nfrom dust3r.smpl_model import SMPLModel\nfrom accelerate import Accelerator\nimport re\n\n\ndef custom_sort_key(key):\n    text = key.split(\"/\")\n    if len(text) > 1:\n        text, num = text[0], text[-1]\n        return (text, int(num))\n    else:\n        return (key, -1)\n\n\ndef merge_chunk_dict(old_dict, curr_dict, add_number):\n    new_dict = {}\n    for key, value in curr_dict.items():\n\n        match = re.search(r\"(\\d+)$\", key)\n        if match:\n\n            num_part = int(match.group()) + add_number\n\n            new_key = re.sub(r\"(\\d+)$\", str(num_part), key, 1)\n            new_dict[new_key] = value\n        else:\n            new_dict[key] = value\n    new_dict = old_dict | new_dict\n    return {k: new_dict[k] for k in sorted(new_dict.keys(), key=custom_sort_key)}\n\n\ndef _interleave_imgs(img1, img2):","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.merge_chunk_dict","uri":"program://Human3R/function/src.dust3r.inference.merge_chunk_dict#L21-L35","kind":"function","name":"merge_chunk_dict","path":"src/dust3r/inference.py","language":"python","start_line":21,"end_line":35,"context_start_line":1,"context_end_line":55,"code":"import tqdm\nimport torch\nfrom dust3r.utils.device import to_cpu, collate_with_cat\nfrom dust3r.utils.misc import invalid_to_nans\nfrom dust3r.utils.geometry import depthmap_to_pts3d, geotrf\nfrom dust3r.model import ARCroco3DStereo\nfrom dust3r.smpl_model import SMPLModel\nfrom accelerate import Accelerator\nimport re\n\n\ndef custom_sort_key(key):\n    text = key.split(\"/\")\n    if len(text) > 1:\n        text, num = text[0], text[-1]\n        return (text, int(num))\n    else:\n        return (key, -1)\n\n\ndef merge_chunk_dict(old_dict, curr_dict, add_number):\n    new_dict = {}\n    for key, value in curr_dict.items():\n\n        match = re.search(r\"(\\d+)$\", key)\n        if match:\n\n            num_part = int(match.group()) + add_number\n\n            new_key = re.sub(r\"(\\d+)$\", str(num_part), key, 1)\n            new_dict[new_key] = value\n        else:\n            new_dict[key] = value\n    new_dict = old_dict | new_dict\n    return {k: new_dict[k] for k in sorted(new_dict.keys(), key=custom_sort_key)}\n\n\ndef _interleave_imgs(img1, img2):\n    res = {}\n    for key, value1 in img1.items():\n        value2 = img2[key]\n        if isinstance(value1, torch.Tensor):\n            value = torch.stack((value1, value2), dim=1).flatten(0, 1)\n        else:\n            value = [x for pair in zip(value1, value2) for x in pair]\n        res[key] = value\n    return res\n\n\ndef make_batch_symmetric(batch):\n    view1, view2 = batch\n    view1, view2 = (_interleave_imgs(view1, view2), _interleave_imgs(view2, view1))\n    return view1, view2\n\n","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference._interleave_imgs","uri":"program://Human3R/function/src.dust3r.inference._interleave_imgs#L38-L47","kind":"function","name":"_interleave_imgs","path":"src/dust3r/inference.py","language":"python","start_line":38,"end_line":47,"context_start_line":18,"context_end_line":67,"code":"        return (key, -1)\n\n\ndef merge_chunk_dict(old_dict, curr_dict, add_number):\n    new_dict = {}\n    for key, value in curr_dict.items():\n\n        match = re.search(r\"(\\d+)$\", key)\n        if match:\n\n            num_part = int(match.group()) + add_number\n\n            new_key = re.sub(r\"(\\d+)$\", str(num_part), key, 1)\n            new_dict[new_key] = value\n        else:\n            new_dict[key] = value\n    new_dict = old_dict | new_dict\n    return {k: new_dict[k] for k in sorted(new_dict.keys(), key=custom_sort_key)}\n\n\ndef _interleave_imgs(img1, img2):\n    res = {}\n    for key, value1 in img1.items():\n        value2 = img2[key]\n        if isinstance(value1, torch.Tensor):\n            value = torch.stack((value1, value2), dim=1).flatten(0, 1)\n        else:\n            value = [x for pair in zip(value1, value2) for x in pair]\n        res[key] = value\n    return res\n\n\ndef make_batch_symmetric(batch):\n    view1, view2 = batch\n    view1, view2 = (_interleave_imgs(view1, view2), _interleave_imgs(view2, view1))\n    return view1, view2\n\n\ndef loss_of_one_batch(\n    batch,\n    model,\n    criterion,\n    accelerator: Accelerator,\n    symmetrize_batch=False,\n    use_amp=False,\n    ret=None,\n    img_mask=None,\n    inference=False,\n    smpl_model: SMPLModel = None\n):","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.make_batch_symmetric","uri":"program://Human3R/function/src.dust3r.inference.make_batch_symmetric#L50-L53","kind":"function","name":"make_batch_symmetric","path":"src/dust3r/inference.py","language":"python","start_line":50,"end_line":53,"context_start_line":30,"context_end_line":73,"code":"            new_key = re.sub(r\"(\\d+)$\", str(num_part), key, 1)\n            new_dict[new_key] = value\n        else:\n            new_dict[key] = value\n    new_dict = old_dict | new_dict\n    return {k: new_dict[k] for k in sorted(new_dict.keys(), key=custom_sort_key)}\n\n\ndef _interleave_imgs(img1, img2):\n    res = {}\n    for key, value1 in img1.items():\n        value2 = img2[key]\n        if isinstance(value1, torch.Tensor):\n            value = torch.stack((value1, value2), dim=1).flatten(0, 1)\n        else:\n            value = [x for pair in zip(value1, value2) for x in pair]\n        res[key] = value\n    return res\n\n\ndef make_batch_symmetric(batch):\n    view1, view2 = batch\n    view1, view2 = (_interleave_imgs(view1, view2), _interleave_imgs(view2, view1))\n    return view1, view2\n\n\ndef loss_of_one_batch(\n    batch,\n    model,\n    criterion,\n    accelerator: Accelerator,\n    symmetrize_batch=False,\n    use_amp=False,\n    ret=None,\n    img_mask=None,\n    inference=False,\n    smpl_model: SMPLModel = None\n):\n    if len(batch) > 2:\n        assert (\n            symmetrize_batch is False\n        ), \"cannot symmetrize batch with more than 2 views\"\n    if symmetrize_batch:\n        batch = make_batch_symmetric(batch)","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.loss_of_one_batch","uri":"program://Human3R/function/src.dust3r.inference.loss_of_one_batch#L56-L90","kind":"function","name":"loss_of_one_batch","path":"src/dust3r/inference.py","language":"python","start_line":56,"end_line":90,"context_start_line":36,"context_end_line":110,"code":"\n\ndef _interleave_imgs(img1, img2):\n    res = {}\n    for key, value1 in img1.items():\n        value2 = img2[key]\n        if isinstance(value1, torch.Tensor):\n            value = torch.stack((value1, value2), dim=1).flatten(0, 1)\n        else:\n            value = [x for pair in zip(value1, value2) for x in pair]\n        res[key] = value\n    return res\n\n\ndef make_batch_symmetric(batch):\n    view1, view2 = batch\n    view1, view2 = (_interleave_imgs(view1, view2), _interleave_imgs(view2, view1))\n    return view1, view2\n\n\ndef loss_of_one_batch(\n    batch,\n    model,\n    criterion,\n    accelerator: Accelerator,\n    symmetrize_batch=False,\n    use_amp=False,\n    ret=None,\n    img_mask=None,\n    inference=False,\n    smpl_model: SMPLModel = None\n):\n    if len(batch) > 2:\n        assert (\n            symmetrize_batch is False\n        ), \"cannot symmetrize batch with more than 2 views\"\n    if symmetrize_batch:\n        batch = make_batch_symmetric(batch)\n\n    with torch.cuda.amp.autocast(enabled=not inference):\n        if inference:\n            output, state_args = model(batch, ret_state=True, inference=True)\n            preds, batch = output.ress, output.views\n            result = dict(views=batch, pred=preds)\n            return result[ret] if ret else result, state_args\n        else:\n            smpl_model.update_smpl_gt(batch)\n            output = model(batch)\n            preds, batch = output.ress, output.views\n\n        with torch.cuda.amp.autocast(enabled=False):\n            loss = criterion(batch, preds) if criterion is not None else None\n\n    result = dict(views=batch, pred=preds, loss=loss)\n    return result[ret] if ret else result\n\n@torch.no_grad()\ndef inference(groups, model, device, verbose=True):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for view in groups:\n        for name in view.keys():  # pseudo_focal\n            if name in ignore_keys:\n                continue\n            if isinstance(view[name], tuple) or isinstance(view[name], list):\n                view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n            else:\n                view[name] = view[name].to(device, non_blocking=True)\n\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    res, state_args = loss_of_one_batch(groups, model, None, None, inference=True)\n    result = to_cpu(res)","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.inference","uri":"program://Human3R/function/src.dust3r.inference.inference#L93-L111","kind":"function","name":"inference","path":"src/dust3r/inference.py","language":"python","start_line":93,"end_line":111,"context_start_line":73,"context_end_line":131,"code":"        batch = make_batch_symmetric(batch)\n\n    with torch.cuda.amp.autocast(enabled=not inference):\n        if inference:\n            output, state_args = model(batch, ret_state=True, inference=True)\n            preds, batch = output.ress, output.views\n            result = dict(views=batch, pred=preds)\n            return result[ret] if ret else result, state_args\n        else:\n            smpl_model.update_smpl_gt(batch)\n            output = model(batch)\n            preds, batch = output.ress, output.views\n\n        with torch.cuda.amp.autocast(enabled=False):\n            loss = criterion(batch, preds) if criterion is not None else None\n\n    result = dict(views=batch, pred=preds, loss=loss)\n    return result[ret] if ret else result\n\n@torch.no_grad()\ndef inference(groups, model, device, verbose=True):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for view in groups:\n        for name in view.keys():  # pseudo_focal\n            if name in ignore_keys:\n                continue\n            if isinstance(view[name], tuple) or isinstance(view[name], list):\n                view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n            else:\n                view[name] = view[name].to(device, non_blocking=True)\n\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    res, state_args = loss_of_one_batch(groups, model, None, None, inference=True)\n    result = to_cpu(res)\n    return result, state_args\n\n\n@torch.no_grad()\ndef inference_step(view, state_args, model, device, verbose=True):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for name in view.keys():  # pseudo_focal\n        if name in ignore_keys:\n            continue\n        if isinstance(view[name], tuple) or isinstance(view[name], list):\n            view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n        else:\n            view[name] = view[name].to(device, non_blocking=True)\n\n    with torch.cuda.amp.autocast(enabled=False):\n        state_feat, state_pos, init_state_feat, mem, init_mem = state_args\n        pred, _ = model.inference_step(\n            view, state_feat, state_pos, init_state_feat, mem, init_mem\n        )","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.inference_step","uri":"program://Human3R/function/src.dust3r.inference.inference_step#L115-L135","kind":"function","name":"inference_step","path":"src/dust3r/inference.py","language":"python","start_line":115,"end_line":135,"context_start_line":95,"context_end_line":155,"code":"        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for view in groups:\n        for name in view.keys():  # pseudo_focal\n            if name in ignore_keys:\n                continue\n            if isinstance(view[name], tuple) or isinstance(view[name], list):\n                view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n            else:\n                view[name] = view[name].to(device, non_blocking=True)\n\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    res, state_args = loss_of_one_batch(groups, model, None, None, inference=True)\n    result = to_cpu(res)\n    return result, state_args\n\n\n@torch.no_grad()\ndef inference_step(view, state_args, model, device, verbose=True):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for name in view.keys():  # pseudo_focal\n        if name in ignore_keys:\n            continue\n        if isinstance(view[name], tuple) or isinstance(view[name], list):\n            view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n        else:\n            view[name] = view[name].to(device, non_blocking=True)\n\n    with torch.cuda.amp.autocast(enabled=False):\n        state_feat, state_pos, init_state_feat, mem, init_mem = state_args\n        pred, _ = model.inference_step(\n            view, state_feat, state_pos, init_state_feat, mem, init_mem\n        )\n\n    res = dict(pred=pred)\n    result = to_cpu(res)\n    return result\n\n\n@torch.no_grad()\ndef inference_recurrent(groups, model, device, verbose=True):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for view in groups:\n        for name in view.keys():  # pseudo_focal\n            if name in ignore_keys:\n                continue\n            if isinstance(view[name], tuple) or isinstance(view[name], list):\n                view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n            else:\n                view[name] = view[name].to(device, non_blocking=True)\n\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    with torch.cuda.amp.autocast(enabled=False):","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.inference_recurrent","uri":"program://Human3R/function/src.dust3r.inference.inference_recurrent#L139-L161","kind":"function","name":"inference_recurrent","path":"src/dust3r/inference.py","language":"python","start_line":139,"end_line":161,"context_start_line":119,"context_end_line":181,"code":"    for name in view.keys():  # pseudo_focal\n        if name in ignore_keys:\n            continue\n        if isinstance(view[name], tuple) or isinstance(view[name], list):\n            view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n        else:\n            view[name] = view[name].to(device, non_blocking=True)\n\n    with torch.cuda.amp.autocast(enabled=False):\n        state_feat, state_pos, init_state_feat, mem, init_mem = state_args\n        pred, _ = model.inference_step(\n            view, state_feat, state_pos, init_state_feat, mem, init_mem\n        )\n\n    res = dict(pred=pred)\n    result = to_cpu(res)\n    return result\n\n\n@torch.no_grad()\ndef inference_recurrent(groups, model, device, verbose=True):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"true_shape\", \"rng\"]\n    )\n    for view in groups:\n        for name in view.keys():  # pseudo_focal\n            if name in ignore_keys:\n                continue\n            if isinstance(view[name], tuple) or isinstance(view[name], list):\n                view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n            else:\n                view[name] = view[name].to(device, non_blocking=True)\n\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    with torch.cuda.amp.autocast(enabled=False):\n        preds, batch, state_args = model.forward_recurrent(\n            groups, device, ret_state=True\n        )\n        res = dict(views=batch, pred=preds)\n    result = to_cpu(res)\n    return result, state_args\n\n@torch.no_grad()\ndef inference_recurrent_lighter(groups, model, device, verbose=True, is_naive=False, use_ttt3r=False):\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    with torch.cuda.amp.autocast(enabled=False):\n        if is_naive:\n            preds, batch, state_args = model.forward_recurrent_lighter_naive(\n            groups, device, ret_state=True, use_ttt3r=use_ttt3r\n        )\n        else:\n            preds, batch, state_args = model.forward_recurrent_lighter(\n                groups, device, ret_state=True, use_ttt3r=use_ttt3r\n            )\n        res = dict(views=batch, pred=preds)\n    return res, state_args\n\ndef check_if_same_size(pairs):\n    shapes1 = [img1[\"img\"].shape[-2:] for img1, img2 in pairs]","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.inference_recurrent_lighter","uri":"program://Human3R/function/src.dust3r.inference.inference_recurrent_lighter#L164-L178","kind":"function","name":"inference_recurrent_lighter","path":"src/dust3r/inference.py","language":"python","start_line":164,"end_line":178,"context_start_line":144,"context_end_line":198,"code":"        for name in view.keys():  # pseudo_focal\n            if name in ignore_keys:\n                continue\n            if isinstance(view[name], tuple) or isinstance(view[name], list):\n                view[name] = [x.to(device, non_blocking=True) for x in view[name]]\n            else:\n                view[name] = view[name].to(device, non_blocking=True)\n\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    with torch.cuda.amp.autocast(enabled=False):\n        preds, batch, state_args = model.forward_recurrent(\n            groups, device, ret_state=True\n        )\n        res = dict(views=batch, pred=preds)\n    result = to_cpu(res)\n    return result, state_args\n\n@torch.no_grad()\ndef inference_recurrent_lighter(groups, model, device, verbose=True, is_naive=False, use_ttt3r=False):\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    with torch.cuda.amp.autocast(enabled=False):\n        if is_naive:\n            preds, batch, state_args = model.forward_recurrent_lighter_naive(\n            groups, device, ret_state=True, use_ttt3r=use_ttt3r\n        )\n        else:\n            preds, batch, state_args = model.forward_recurrent_lighter(\n                groups, device, ret_state=True, use_ttt3r=use_ttt3r\n            )\n        res = dict(views=batch, pred=preds)\n    return res, state_args\n\ndef check_if_same_size(pairs):\n    shapes1 = [img1[\"img\"].shape[-2:] for img1, img2 in pairs]\n    shapes2 = [img2[\"img\"].shape[-2:] for img1, img2 in pairs]\n    return all(shapes1[0] == s for s in shapes1) and all(\n        shapes2[0] == s for s in shapes2\n    )\n\n\ndef get_pred_pts3d(gt, pred, use_pose=False, inplace=False):\n    if \"depth\" in pred and \"pseudo_focal\" in pred:\n        try:\n            pp = gt[\"camera_intrinsics\"][..., :2, 2]\n        except KeyError:\n            pp = None\n        pts3d = depthmap_to_pts3d(**pred, pp=pp)\n\n    elif \"pts3d\" in pred:\n\n        pts3d = pred[\"pts3d\"]","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.check_if_same_size","uri":"program://Human3R/function/src.dust3r.inference.check_if_same_size#L180-L185","kind":"function","name":"check_if_same_size","path":"src/dust3r/inference.py","language":"python","start_line":180,"end_line":185,"context_start_line":160,"context_end_line":205,"code":"    result = to_cpu(res)\n    return result, state_args\n\n@torch.no_grad()\ndef inference_recurrent_lighter(groups, model, device, verbose=True, is_naive=False, use_ttt3r=False):\n    if verbose:\n        print(f\">> Inference with model on {len(groups)} image/raymaps\")\n\n    with torch.cuda.amp.autocast(enabled=False):\n        if is_naive:\n            preds, batch, state_args = model.forward_recurrent_lighter_naive(\n            groups, device, ret_state=True, use_ttt3r=use_ttt3r\n        )\n        else:\n            preds, batch, state_args = model.forward_recurrent_lighter(\n                groups, device, ret_state=True, use_ttt3r=use_ttt3r\n            )\n        res = dict(views=batch, pred=preds)\n    return res, state_args\n\ndef check_if_same_size(pairs):\n    shapes1 = [img1[\"img\"].shape[-2:] for img1, img2 in pairs]\n    shapes2 = [img2[\"img\"].shape[-2:] for img1, img2 in pairs]\n    return all(shapes1[0] == s for s in shapes1) and all(\n        shapes2[0] == s for s in shapes2\n    )\n\n\ndef get_pred_pts3d(gt, pred, use_pose=False, inplace=False):\n    if \"depth\" in pred and \"pseudo_focal\" in pred:\n        try:\n            pp = gt[\"camera_intrinsics\"][..., :2, 2]\n        except KeyError:\n            pp = None\n        pts3d = depthmap_to_pts3d(**pred, pp=pp)\n\n    elif \"pts3d\" in pred:\n\n        pts3d = pred[\"pts3d\"]\n\n    elif \"pts3d_in_other_view\" in pred:\n\n        assert use_pose is True\n        return (\n            pred[\"pts3d_in_other_view\"]\n            if inplace","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.get_pred_pts3d","uri":"program://Human3R/function/src.dust3r.inference.get_pred_pts3d#L188-L214","kind":"function","name":"get_pred_pts3d","path":"src/dust3r/inference.py","language":"python","start_line":188,"end_line":214,"context_start_line":168,"context_end_line":234,"code":"    with torch.cuda.amp.autocast(enabled=False):\n        if is_naive:\n            preds, batch, state_args = model.forward_recurrent_lighter_naive(\n            groups, device, ret_state=True, use_ttt3r=use_ttt3r\n        )\n        else:\n            preds, batch, state_args = model.forward_recurrent_lighter(\n                groups, device, ret_state=True, use_ttt3r=use_ttt3r\n            )\n        res = dict(views=batch, pred=preds)\n    return res, state_args\n\ndef check_if_same_size(pairs):\n    shapes1 = [img1[\"img\"].shape[-2:] for img1, img2 in pairs]\n    shapes2 = [img2[\"img\"].shape[-2:] for img1, img2 in pairs]\n    return all(shapes1[0] == s for s in shapes1) and all(\n        shapes2[0] == s for s in shapes2\n    )\n\n\ndef get_pred_pts3d(gt, pred, use_pose=False, inplace=False):\n    if \"depth\" in pred and \"pseudo_focal\" in pred:\n        try:\n            pp = gt[\"camera_intrinsics\"][..., :2, 2]\n        except KeyError:\n            pp = None\n        pts3d = depthmap_to_pts3d(**pred, pp=pp)\n\n    elif \"pts3d\" in pred:\n\n        pts3d = pred[\"pts3d\"]\n\n    elif \"pts3d_in_other_view\" in pred:\n\n        assert use_pose is True\n        return (\n            pred[\"pts3d_in_other_view\"]\n            if inplace\n            else pred[\"pts3d_in_other_view\"].clone()\n        )\n\n    if use_pose:\n        camera_pose = pred.get(\"camera_pose\")\n        assert camera_pose is not None\n        pts3d = geotrf(camera_pose, pts3d)\n\n    return pts3d\n\n\ndef find_opt_scaling(\n    gt_pts1,\n    gt_pts2,\n    pr_pts1,\n    pr_pts2=None,\n    fit_mode=\"weiszfeld_stop_grad\",\n    valid1=None,\n    valid2=None,\n):\n    assert gt_pts1.ndim == pr_pts1.ndim == 4\n    assert gt_pts1.shape == pr_pts1.shape\n    if gt_pts2 is not None:\n        assert gt_pts2.ndim == pr_pts2.ndim == 4\n        assert gt_pts2.shape == pr_pts2.shape\n\n    nan_gt_pts1 = invalid_to_nans(gt_pts1, valid1).flatten(1, 2)\n    nan_gt_pts2 = (\n        invalid_to_nans(gt_pts2, valid2).flatten(1, 2) if gt_pts2 is not None else None","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.inference.find_opt_scaling","uri":"program://Human3R/function/src.dust3r.inference.find_opt_scaling#L217-L276","kind":"function","name":"find_opt_scaling","path":"src/dust3r/inference.py","language":"python","start_line":217,"end_line":276,"context_start_line":197,"context_end_line":276,"code":"\n        pts3d = pred[\"pts3d\"]\n\n    elif \"pts3d_in_other_view\" in pred:\n\n        assert use_pose is True\n        return (\n            pred[\"pts3d_in_other_view\"]\n            if inplace\n            else pred[\"pts3d_in_other_view\"].clone()\n        )\n\n    if use_pose:\n        camera_pose = pred.get(\"camera_pose\")\n        assert camera_pose is not None\n        pts3d = geotrf(camera_pose, pts3d)\n\n    return pts3d\n\n\ndef find_opt_scaling(\n    gt_pts1,\n    gt_pts2,\n    pr_pts1,\n    pr_pts2=None,\n    fit_mode=\"weiszfeld_stop_grad\",\n    valid1=None,\n    valid2=None,\n):\n    assert gt_pts1.ndim == pr_pts1.ndim == 4\n    assert gt_pts1.shape == pr_pts1.shape\n    if gt_pts2 is not None:\n        assert gt_pts2.ndim == pr_pts2.ndim == 4\n        assert gt_pts2.shape == pr_pts2.shape\n\n    nan_gt_pts1 = invalid_to_nans(gt_pts1, valid1).flatten(1, 2)\n    nan_gt_pts2 = (\n        invalid_to_nans(gt_pts2, valid2).flatten(1, 2) if gt_pts2 is not None else None\n    )\n\n    pr_pts1 = invalid_to_nans(pr_pts1, valid1).flatten(1, 2)\n    pr_pts2 = (\n        invalid_to_nans(pr_pts2, valid2).flatten(1, 2) if pr_pts2 is not None else None\n    )\n\n    all_gt = (\n        torch.cat((nan_gt_pts1, nan_gt_pts2), dim=1)\n        if gt_pts2 is not None\n        else nan_gt_pts1\n    )\n    all_pr = torch.cat((pr_pts1, pr_pts2), dim=1) if pr_pts2 is not None else pr_pts1\n\n    dot_gt_pr = (all_pr * all_gt).sum(dim=-1)\n    dot_gt_gt = all_gt.square().sum(dim=-1)\n\n    if fit_mode.startswith(\"avg\"):\n\n        scaling = dot_gt_pr.nanmean(dim=1) / dot_gt_gt.nanmean(dim=1)\n    elif fit_mode.startswith(\"median\"):\n        scaling = (dot_gt_pr / dot_gt_gt).nanmedian(dim=1).values\n    elif fit_mode.startswith(\"weiszfeld\"):\n\n        scaling = dot_gt_pr.nanmean(dim=1) / dot_gt_gt.nanmean(dim=1)\n\n        for iter in range(10):\n\n            dis = (all_pr - scaling.view(-1, 1, 1) * all_gt).norm(dim=-1)\n\n            w = dis.clip_(min=1e-8).reciprocal()\n\n            scaling = (w * dot_gt_pr).nanmean(dim=1) / (w * dot_gt_gt).nanmean(dim=1)\n    else:\n        raise ValueError(f\"bad {fit_mode=}\")\n\n    if fit_mode.endswith(\"stop_grad\"):\n        scaling = scaling.detach()\n\n    scaling = scaling.clip(min=1e-3)\n\n    return scaling","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.patch_embed","uri":"program://Human3R/module/src.dust3r.patch_embed#L1-L93","kind":"module","name":"src.dust3r.patch_embed","path":"src/dust3r/patch_embed.py","language":"python","start_line":1,"end_line":93,"context_start_line":1,"context_end_line":93,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport dust3r.utils.path_to_croco  # noqa: F401\nfrom models.blocks import PatchEmbed  # noqa\n\n\ndef get_patch_embed(patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3):\n    assert patch_embed_cls in [\"PatchEmbedDust3R\", \"ManyAR_PatchEmbed\"]\n    patch_embed = eval(patch_embed_cls)(img_size, patch_size, in_chans, enc_embed_dim)\n    return patch_embed\n\n\nclass PatchEmbedDust3R(PatchEmbed):\n    def forward(self, x, **kw):\n        B, C, H, W = x.shape\n        assert (\n            H % self.patch_size[0] == 0\n        ), f\"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]}).\"\n        assert (\n            W % self.patch_size[1] == 0\n        ), f\"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]}).\"\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n\nclass ManyAR_PatchEmbed(PatchEmbed):\n    \"\"\"Handle images with non-square aspect ratio.\n    All images in the same batch have the same aspect ratio.\n    true_shape = [(height, width) ...] indicates the actual shape of each image.\n    \"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        self.embed_dim = embed_dim\n        super().__init__(img_size, patch_size, in_chans, embed_dim, norm_layer, flatten)\n\n    def forward(self, img, true_shape):\n        B, C, H, W = img.shape\n\n        assert (\n            H % self.patch_size[0] == 0\n        ), f\"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]}).\"\n        assert (\n            W % self.patch_size[1] == 0\n        ), f\"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]}).\"\n        assert true_shape.shape == (\n            B,\n            2,\n        ), f\"true_shape has the wrong shape={true_shape.shape}\"\n\n        W //= self.patch_size[0]\n        H //= self.patch_size[1]\n        n_tokens = H * W\n\n        height, width = true_shape.T\n\n        is_landscape = torch.ones_like(width, dtype=torch.bool)\n        is_portrait = ~is_landscape\n\n        x = img.new_zeros((B, n_tokens, self.embed_dim))\n        pos = img.new_zeros((B, n_tokens, 2), dtype=torch.int64)\n\n        x[is_landscape] = (\n            self.proj(img[is_landscape]).permute(0, 2, 3, 1).flatten(1, 2).float()\n        )\n        x[is_portrait] = (\n            self.proj(img[is_portrait].swapaxes(-1, -2))\n            .permute(0, 2, 3, 1)\n            .flatten(1, 2)\n            .float()\n        )\n\n        pos[is_landscape] = self.position_getter(1, H, W, pos.device)\n        pos[is_portrait] = self.position_getter(1, W, H, pos.device)\n\n        x = self.norm(x)\n        return x, pos","source_hash":"c0dae36e876d2124f0e31403a3a306db7e6d5bd69d8e9c6f16bbfb81c82eef3d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.patch_embed.get_patch_embed","uri":"program://Human3R/function/src.dust3r.patch_embed.get_patch_embed#L12-L15","kind":"function","name":"get_patch_embed","path":"src/dust3r/patch_embed.py","language":"python","start_line":12,"end_line":15,"context_start_line":1,"context_end_line":35,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport dust3r.utils.path_to_croco  # noqa: F401\nfrom models.blocks import PatchEmbed  # noqa\n\n\ndef get_patch_embed(patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3):\n    assert patch_embed_cls in [\"PatchEmbedDust3R\", \"ManyAR_PatchEmbed\"]\n    patch_embed = eval(patch_embed_cls)(img_size, patch_size, in_chans, enc_embed_dim)\n    return patch_embed\n\n\nclass PatchEmbedDust3R(PatchEmbed):\n    def forward(self, x, **kw):\n        B, C, H, W = x.shape\n        assert (\n            H % self.patch_size[0] == 0\n        ), f\"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]}).\"\n        assert (\n            W % self.patch_size[1] == 0\n        ), f\"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]}).\"\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n\nclass ManyAR_PatchEmbed(PatchEmbed):","source_hash":"c0dae36e876d2124f0e31403a3a306db7e6d5bd69d8e9c6f16bbfb81c82eef3d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.patch_embed.PatchEmbedDust3R","uri":"program://Human3R/class/src.dust3r.patch_embed.PatchEmbedDust3R#L18-L32","kind":"class","name":"PatchEmbedDust3R","path":"src/dust3r/patch_embed.py","language":"python","start_line":18,"end_line":32,"context_start_line":1,"context_end_line":52,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport dust3r.utils.path_to_croco  # noqa: F401\nfrom models.blocks import PatchEmbed  # noqa\n\n\ndef get_patch_embed(patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3):\n    assert patch_embed_cls in [\"PatchEmbedDust3R\", \"ManyAR_PatchEmbed\"]\n    patch_embed = eval(patch_embed_cls)(img_size, patch_size, in_chans, enc_embed_dim)\n    return patch_embed\n\n\nclass PatchEmbedDust3R(PatchEmbed):\n    def forward(self, x, **kw):\n        B, C, H, W = x.shape\n        assert (\n            H % self.patch_size[0] == 0\n        ), f\"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]}).\"\n        assert (\n            W % self.patch_size[1] == 0\n        ), f\"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]}).\"\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n\nclass ManyAR_PatchEmbed(PatchEmbed):\n    \"\"\"Handle images with non-square aspect ratio.\n    All images in the same batch have the same aspect ratio.\n    true_shape = [(height, width) ...] indicates the actual shape of each image.\n    \"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        self.embed_dim = embed_dim\n        super().__init__(img_size, patch_size, in_chans, embed_dim, norm_layer, flatten)\n","source_hash":"c0dae36e876d2124f0e31403a3a306db7e6d5bd69d8e9c6f16bbfb81c82eef3d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.patch_embed.ManyAR_PatchEmbed","uri":"program://Human3R/class/src.dust3r.patch_embed.ManyAR_PatchEmbed#L35-L93","kind":"class","name":"ManyAR_PatchEmbed","path":"src/dust3r/patch_embed.py","language":"python","start_line":35,"end_line":93,"context_start_line":15,"context_end_line":93,"code":"    return patch_embed\n\n\nclass PatchEmbedDust3R(PatchEmbed):\n    def forward(self, x, **kw):\n        B, C, H, W = x.shape\n        assert (\n            H % self.patch_size[0] == 0\n        ), f\"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]}).\"\n        assert (\n            W % self.patch_size[1] == 0\n        ), f\"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]}).\"\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n\nclass ManyAR_PatchEmbed(PatchEmbed):\n    \"\"\"Handle images with non-square aspect ratio.\n    All images in the same batch have the same aspect ratio.\n    true_shape = [(height, width) ...] indicates the actual shape of each image.\n    \"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        self.embed_dim = embed_dim\n        super().__init__(img_size, patch_size, in_chans, embed_dim, norm_layer, flatten)\n\n    def forward(self, img, true_shape):\n        B, C, H, W = img.shape\n\n        assert (\n            H % self.patch_size[0] == 0\n        ), f\"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]}).\"\n        assert (\n            W % self.patch_size[1] == 0\n        ), f\"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]}).\"\n        assert true_shape.shape == (\n            B,\n            2,\n        ), f\"true_shape has the wrong shape={true_shape.shape}\"\n\n        W //= self.patch_size[0]\n        H //= self.patch_size[1]\n        n_tokens = H * W\n\n        height, width = true_shape.T\n\n        is_landscape = torch.ones_like(width, dtype=torch.bool)\n        is_portrait = ~is_landscape\n\n        x = img.new_zeros((B, n_tokens, self.embed_dim))\n        pos = img.new_zeros((B, n_tokens, 2), dtype=torch.int64)\n\n        x[is_landscape] = (\n            self.proj(img[is_landscape]).permute(0, 2, 3, 1).flatten(1, 2).float()\n        )\n        x[is_portrait] = (\n            self.proj(img[is_portrait].swapaxes(-1, -2))\n            .permute(0, 2, 3, 1)\n            .flatten(1, 2)\n            .float()\n        )\n\n        pos[is_landscape] = self.position_getter(1, H, W, pos.device)\n        pos[is_portrait] = self.position_getter(1, W, H, pos.device)\n\n        x = self.norm(x)\n        return x, pos","source_hash":"c0dae36e876d2124f0e31403a3a306db7e6d5bd69d8e9c6f16bbfb81c82eef3d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.patch_embed.forward","uri":"program://Human3R/function/src.dust3r.patch_embed.forward#L53-L93","kind":"function","name":"forward","path":"src/dust3r/patch_embed.py","language":"python","start_line":53,"end_line":93,"context_start_line":33,"context_end_line":93,"code":"\n\nclass ManyAR_PatchEmbed(PatchEmbed):\n    \"\"\"Handle images with non-square aspect ratio.\n    All images in the same batch have the same aspect ratio.\n    true_shape = [(height, width) ...] indicates the actual shape of each image.\n    \"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        self.embed_dim = embed_dim\n        super().__init__(img_size, patch_size, in_chans, embed_dim, norm_layer, flatten)\n\n    def forward(self, img, true_shape):\n        B, C, H, W = img.shape\n\n        assert (\n            H % self.patch_size[0] == 0\n        ), f\"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]}).\"\n        assert (\n            W % self.patch_size[1] == 0\n        ), f\"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]}).\"\n        assert true_shape.shape == (\n            B,\n            2,\n        ), f\"true_shape has the wrong shape={true_shape.shape}\"\n\n        W //= self.patch_size[0]\n        H //= self.patch_size[1]\n        n_tokens = H * W\n\n        height, width = true_shape.T\n\n        is_landscape = torch.ones_like(width, dtype=torch.bool)\n        is_portrait = ~is_landscape\n\n        x = img.new_zeros((B, n_tokens, self.embed_dim))\n        pos = img.new_zeros((B, n_tokens, 2), dtype=torch.int64)\n\n        x[is_landscape] = (\n            self.proj(img[is_landscape]).permute(0, 2, 3, 1).flatten(1, 2).float()\n        )\n        x[is_portrait] = (\n            self.proj(img[is_portrait].swapaxes(-1, -2))\n            .permute(0, 2, 3, 1)\n            .flatten(1, 2)\n            .float()\n        )\n\n        pos[is_landscape] = self.position_getter(1, H, W, pos.device)\n        pos[is_portrait] = self.position_getter(1, W, H, pos.device)\n\n        x = self.norm(x)\n        return x, pos","source_hash":"c0dae36e876d2124f0e31403a3a306db7e6d5bd69d8e9c6f16bbfb81c82eef3d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.patch_embed.__init__","uri":"program://Human3R/function/src.dust3r.patch_embed.__init__#L41-L51","kind":"function","name":"__init__","path":"src/dust3r/patch_embed.py","language":"python","start_line":41,"end_line":51,"context_start_line":21,"context_end_line":71,"code":"        assert (\n            H % self.patch_size[0] == 0\n        ), f\"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]}).\"\n        assert (\n            W % self.patch_size[1] == 0\n        ), f\"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]}).\"\n        x = self.proj(x)\n        pos = self.position_getter(B, x.size(2), x.size(3), x.device)\n        if self.flatten:\n            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n        x = self.norm(x)\n        return x, pos\n\n\nclass ManyAR_PatchEmbed(PatchEmbed):\n    \"\"\"Handle images with non-square aspect ratio.\n    All images in the same batch have the same aspect ratio.\n    true_shape = [(height, width) ...] indicates the actual shape of each image.\n    \"\"\"\n\n    def __init__(\n        self,\n        img_size=224,\n        patch_size=16,\n        in_chans=3,\n        embed_dim=768,\n        norm_layer=None,\n        flatten=True,\n    ):\n        self.embed_dim = embed_dim\n        super().__init__(img_size, patch_size, in_chans, embed_dim, norm_layer, flatten)\n\n    def forward(self, img, true_shape):\n        B, C, H, W = img.shape\n\n        assert (\n            H % self.patch_size[0] == 0\n        ), f\"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]}).\"\n        assert (\n            W % self.patch_size[1] == 0\n        ), f\"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]}).\"\n        assert true_shape.shape == (\n            B,\n            2,\n        ), f\"true_shape has the wrong shape={true_shape.shape}\"\n\n        W //= self.patch_size[0]\n        H //= self.patch_size[1]\n        n_tokens = H * W\n\n        height, width = true_shape.T","source_hash":"c0dae36e876d2124f0e31403a3a306db7e6d5bd69d8e9c6f16bbfb81c82eef3d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model","uri":"program://Human3R/module/src.dust3r.model#L1-L1982","kind":"module","name":"src.dust3r.model","path":"src/dust3r/model.py","language":"python","start_line":1,"end_line":1982,"context_start_line":1,"context_end_line":1982,"code":"import sys\nimport os\n\nsys.path.append(os.path.dirname(os.path.dirname(__file__)))\nfrom collections import OrderedDict\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.checkpoint import checkpoint\nfrom copy import deepcopy\nfrom functools import partial\nfrom typing import Optional, Tuple, List, Any\nfrom dataclasses import dataclass\nfrom transformers import PretrainedConfig\nfrom transformers import PreTrainedModel\nfrom transformers.modeling_outputs import BaseModelOutput\nfrom transformers.file_utils import ModelOutput\nimport time\nfrom dust3r.utils.misc import (\n    fill_default_args,\n    freeze_all_params,\n    fix_all_params,\n    is_symmetrized,\n    interleave,\n    transpose_to_landscape,\n)\nfrom dust3r.heads import head_factory\nfrom dust3r.utils.camera import PoseEncoder\nfrom dust3r.patch_embed import get_patch_embed\nimport dust3r.utils.path_to_croco  # noqa: F401\nfrom models.croco import CroCoNet, CrocoConfig  # noqa\nfrom dust3r.blocks import (\n    Block,\n    DecoderBlock,\n    Mlp,\n    Attention,\n    CrossAttention,\n    DropPath,\n    CustomDecoderBlock,\n)  # noqa\n\ninf = float(\"inf\")\nfrom accelerate.logging import get_logger\n\nfrom dust3r.smpl_model import nms, apply_threshold\nfrom einops import rearrange\n\nfrom dust3r.utils.geometry import inverse_perspective_projection, get_camera_parameters\nfrom dust3r.utils.image import unpad_uv, log_optimal_transport\nfrom mhmr.blocks import Dinov2Backbone, FourierPositionEncoding, TransformerDecoder\nprinter = get_logger(__name__, log_level=\"DEBUG\")\n\nfrom dust3r.utils.device import to_cpu, to_gpu\n\n@dataclass\nclass ARCroco3DStereoOutput(ModelOutput):\n    \"\"\"\n    Custom output class for ARCroco3DStereo.\n    \"\"\"\n\n    ress: Optional[List[Any]] = None\n    views: Optional[List[Any]] = None\n\n\ndef strip_module(state_dict):\n    \"\"\"\n    Removes the 'module.' prefix from the keys of a state_dict.\n    Args:\n        state_dict (dict): The original state_dict with possible 'module.' prefixes.\n    Returns:\n        OrderedDict: A new state_dict with 'module.' prefixes removed.\n    \"\"\"\n    new_state_dict = OrderedDict()\n    for k, v in state_dict.items():\n        name = k[7:] if k.startswith(\"module.\") else k\n        new_state_dict[name] = v\n    return new_state_dict\n\n\ndef strip_module_mhmr(state_dict):\n    \"\"\"\n    Load Multi-HMR pretrained model\n    \"\"\"\n    new_state_dict = OrderedDict()\n    for k, v in state_dict.items():\n        if k.startswith((\"mlp_classif.\", \"mlp_offset.\")):\n            name = f\"downstream_head.{k}\"\n        elif k.startswith((\"x_attention_head.dec\")):\n            name = f\"downstream_head.{k[17:]}\"\n        elif k.startswith((\"x_attention_head.transformer.\", \"x_attention_head.cross_\")):\n            name = k[17:]\n        elif k.startswith((\"backbone\")):\n            name = k\n        else:\n            continue\n        new_state_dict[name] = v\n    return new_state_dict\n\n\ndef load_model(model_path, device, verbose=True):\n    if verbose:\n        print(\"... loading model from\", model_path)\n    ckpt = torch.load(model_path, map_location=\"cpu\")\n    args = ckpt[\"args\"].model.replace(\n        \"ManyAR_PatchEmbed\", \"PatchEmbedDust3R\"\n    )  # ManyAR only for aspect ratio not consistent\n    if \"landscape_only\" not in args:\n        args = args[:-2] + \", landscape_only=False))\"\n    else:\n        args = args.replace(\" \", \"\").replace(\n            \"landscape_only=True\", \"landscape_only=False\"\n        )\n    assert \"landscape_only=False\" in args\n    if verbose:\n        print(f\"instantiating : {args}\")\n    net = eval(args)\n    s = net.load_state_dict(ckpt[\"model\"], strict=False)\n    if verbose:\n        print(s)\n    return net.to(device)\n\n\nclass ARCroco3DStereoConfig(PretrainedConfig):\n    model_type = \"arcroco_3d_stereo\"\n\n    def __init__(\n        self,\n        output_mode=\"pts3d\",\n        head_type=\"linear\",  # or dpt\n        depth_mode=(\"exp\", -float(\"inf\"), float(\"inf\")),\n        conf_mode=(\"exp\", 1, float(\"inf\")),\n        pose_mode=(\"exp\", -float(\"inf\"), float(\"inf\")),\n        freeze=\"none\",\n        landscape_only=True,\n        patch_embed_cls=\"PatchEmbedDust3R\",\n        ray_enc_depth=2,\n        state_size=324,\n        local_mem_size=256,\n        state_pe=\"2d\",\n        state_dec_num_heads=16,\n        depth_head=False,\n        rgb_head=False,\n        pose_conf_head=False,\n        pose_head=False,\n        msk_head=False,\n        use_prompt=False,\n        is_shallow=False,\n        prompt_size=None,\n        backbone='dinov2_vitl14',\n        mhmr_img_res=None,\n        **croco_kwargs,\n    ):\n        super().__init__()\n        self.output_mode = output_mode\n        self.head_type = head_type\n        self.depth_mode = depth_mode\n        self.conf_mode = conf_mode\n        self.pose_mode = pose_mode\n        self.freeze = freeze\n        self.landscape_only = landscape_only\n        self.patch_embed_cls = patch_embed_cls\n        self.ray_enc_depth = ray_enc_depth\n        self.state_size = state_size\n        self.state_pe = state_pe\n        self.state_dec_num_heads = state_dec_num_heads\n        self.local_mem_size = local_mem_size\n        self.depth_head = depth_head\n        self.rgb_head = rgb_head\n        self.pose_conf_head = pose_conf_head\n        self.pose_head = pose_head\n        self.msk_head = msk_head\n        self.backbone = backbone\n        self.mhmr_img_res = mhmr_img_res\n        self.croco_kwargs = croco_kwargs\n\n\nclass LocalMemory(nn.Module):\n    def __init__(\n        self,\n        size,\n        k_dim,\n        v_dim,\n        num_heads,\n        depth=2,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ) -> None:\n        super().__init__()\n        self.v_dim = v_dim\n        self.proj_q = nn.Linear(k_dim, v_dim)\n        self.masked_token = nn.Parameter(\n            torch.randn(1, 1, v_dim) * 0.2, requires_grad=True\n        )\n        self.mem = nn.Parameter(\n            torch.randn(1, size, 2 * v_dim) * 0.2, requires_grad=True\n        )\n        self.write_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    2 * v_dim,\n                    num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=qkv_bias,\n                    norm_layer=norm_layer,\n                    attn_drop=attn_drop,\n                    drop=drop,\n                    drop_path=drop_path,\n                    act_layer=act_layer,\n                    norm_mem=norm_mem,\n                    rope=rope,\n                )\n                for _ in range(depth)\n            ]\n        )\n        self.read_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    2 * v_dim,\n                    num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=qkv_bias,\n                    norm_layer=norm_layer,\n                    attn_drop=attn_drop,\n                    drop=drop,\n                    drop_path=drop_path,\n                    act_layer=act_layer,\n                    norm_mem=norm_mem,\n                    rope=rope,\n                )\n                for _ in range(depth)\n            ]\n        )\n\n    def update_mem(self, mem, feat_k, feat_v):\n        \"\"\"\n        mem_k: [B, size, C]\n        mem_v: [B, size, C]\n        feat_k: [B, 1, C]\n        feat_v: [B, 1, C]\n        \"\"\"\n        feat_k = self.proj_q(feat_k)  # [B, 1, C]\n        feat = torch.cat([feat_k, feat_v], dim=-1)\n        for blk in self.write_blocks:\n            mem, _, _ = blk(mem, feat, None, None)\n        return mem\n\n    def inquire(self, query, mem):\n        x = self.proj_q(query)  # [B, 1, C]\n        x = torch.cat([x, self.masked_token.expand(x.shape[0], -1, -1)], dim=-1)\n        for blk in self.read_blocks:\n            x, _, _ = blk(x, mem, None, None)\n        return x[..., -self.v_dim :]\n\n\nclass ARCroco3DStereo(CroCoNet):\n    config_class = ARCroco3DStereoConfig\n    base_model_prefix = \"arcroco3dstereo\"\n    supports_gradient_checkpointing = True\n\n    def __init__(self, config: ARCroco3DStereoConfig):\n        self.gradient_checkpointing = False\n        self.fixed_input_length = True\n        config.croco_kwargs = fill_default_args(\n            config.croco_kwargs, CrocoConfig.__init__\n        )\n        self.config = config\n        self.patch_embed_cls = config.patch_embed_cls\n        self.croco_args = config.croco_kwargs\n        croco_cfg = CrocoConfig(**self.croco_args)\n        super().__init__(croco_cfg)\n        self.enc_blocks_ray_map = nn.ModuleList(\n            [\n                Block(\n                    self.enc_embed_dim,\n                    16,\n                    4,\n                    qkv_bias=True,\n                    norm_layer=partial(nn.LayerNorm, eps=1e-6),\n                    rope=self.rope,\n                )\n                for _ in range(config.ray_enc_depth)\n            ]\n        )\n        self.enc_norm_ray_map = nn.LayerNorm(self.enc_embed_dim, eps=1e-6)\n        self.dec_num_heads = self.croco_args[\"dec_num_heads\"]\n        self.pose_head_flag = config.pose_head\n        self.msk_head_flag = config.msk_head\n        if self.pose_head_flag:\n            self.pose_token = nn.Parameter(\n                torch.randn(1, 1, self.dec_embed_dim) * 0.02, requires_grad=True\n            )\n            self.pose_retriever = LocalMemory(\n                size=config.local_mem_size,\n                k_dim=self.enc_embed_dim,\n                v_dim=self.dec_embed_dim,\n                num_heads=self.dec_num_heads,\n                mlp_ratio=4,\n                qkv_bias=True,\n                attn_drop=0.0,\n                norm_layer=partial(nn.LayerNorm, eps=1e-6),\n                rope=None,\n            )\n        self.register_tokens = nn.Embedding(config.state_size, self.enc_embed_dim)\n        self.state_size = config.state_size\n        self.state_pe = config.state_pe\n        self.masked_img_token = nn.Parameter(\n            torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True\n        )\n        self.masked_ray_map_token = nn.Parameter(\n            torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True\n        )\n        self.masked_smpl_token = nn.Parameter(\n            torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True\n        )\n\n        # MHMR\n        # 'dinov2_vits14': 384, 'dinov2_vitb14': 768, 'dinov2_vitl14': 1024\n        self.backbone = Dinov2Backbone(config.backbone, pretrained=False)\n        self.bb_patch_size = self.backbone.patch_size\n        self.backbone_dim = self.backbone.embed_dim\n        self.mhmr_img_res = config.mhmr_img_res\n        self.bb_token_res = self.mhmr_img_res // self.bb_patch_size\n\n        if config.output_mode == 'naive':\n            self.fourier_camera = FourierPositionEncoding(n=3, num_bands=16, max_resolution=64)\n            self.camera_embed_dim = self.fourier_camera.channels\n            context_dim = self.backbone_dim + self.camera_embed_dim\n\n            transformer_args = dict(\n                num_tokens=1,\n                token_dim=(318+10+3+context_dim),\n                dim=1024,\n                depth=2,\n                heads=8,\n                mlp_dim=1024,\n                dim_head=32,\n                dropout=0.0,\n                emb_dropout=0.0,\n                context_dim=context_dim,\n            )\n            self.transformer = TransformerDecoder(**transformer_args)\n            # Init learned embeddings for queries\n            self.cross_queries_x = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_queries_x, std=0.2)\n            self.cross_queries_y = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_queries_y, std=0.2)\n            self.cross_values_x = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_values_x, std=0.2)\n            self.cross_values_y = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_values_y, std=0.2)\n\n\n        self.mhmr_masked_smpl_token = nn.Parameter(\n            torch.randn(\n                1, context_dim if config.output_mode == \"naive\" else self.backbone_dim\n                ) * 0.02, requires_grad=True\n        )\n        self.mhmr_masked_img_token = nn.Parameter(\n            torch.randn(1, self.backbone_dim) * 0.02, requires_grad=True\n        )\n\n        self._set_state_decoder(\n            self.enc_embed_dim,\n            self.dec_embed_dim,\n            config.state_dec_num_heads,\n            self.dec_depth,\n            self.croco_args.get(\"mlp_ratio\", None),\n            self.croco_args.get(\"norm_layer\", None),\n            self.croco_args.get(\"norm_im2_in_dec\", None),\n        )\n        self.set_downstream_head(\n            config.output_mode,\n            config.head_type,\n            config.landscape_only,\n            config.depth_mode,\n            config.conf_mode,\n            config.pose_mode,\n            config.depth_head,\n            config.rgb_head,\n            config.pose_conf_head,\n            config.pose_head,\n            config.msk_head,\n            **self.croco_args,\n        )\n        self.set_freeze(config.freeze)\n\n    @classmethod\n    def from_pretrained(cls, pretrained_model_name_or_path, **kw):\n        if os.path.isfile(pretrained_model_name_or_path):\n            return load_model(pretrained_model_name_or_path, device=\"cpu\")\n        else:\n            try:\n                model = super(ARCroco3DStereo, cls).from_pretrained(\n                    pretrained_model_name_or_path, **kw\n                )\n            except TypeError as e:\n                raise Exception(\n                    f\"tried to load {pretrained_model_name_or_path} from huggingface, but failed\"\n                )\n            return model\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3\n        )\n        self.patch_embed_ray_map = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=6\n        )\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_state_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth_state = dec_depth\n        self.dec_embed_dim_state = dec_embed_dim\n        self.decoder_embed_state = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks_state = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        self.dec_norm_state = norm_layer(dec_embed_dim)\n\n    def load_state_dict(self, ckpt, **kw):\n        if all(k.startswith(\"module\") for k in ckpt):\n            ckpt = strip_module(ckpt)\n        new_ckpt = dict(ckpt)\n        if not any(k.startswith(\"dec_blocks_state\") for k in ckpt):\n            for key, value in ckpt.items():\n                if key.startswith(\"dec_blocks\"):\n                    new_ckpt[key.replace(\"dec_blocks\", \"dec_blocks_state\")] = value\n        try:\n            return super().load_state_dict(new_ckpt, **kw)\n        except:\n            try:\n                new_new_ckpt = {\n                    k: v\n                    for k, v in new_ckpt.items()\n                    if not k.startswith(\"dec_blocks\")\n                    and not k.startswith(\"dec_norm\")\n                    and not k.startswith(\"decoder_embed\")\n                }\n                return super().load_state_dict(new_new_ckpt, **kw)\n            except:\n                new_new_ckpt = {}\n                for key in new_ckpt:\n                    if key in self.state_dict():\n                        if new_ckpt[key].size() == self.state_dict()[key].size():\n                            new_new_ckpt[key] = new_ckpt[key]\n                        else:\n                            printer.info(\n                                f\"Skipping '{key}': size mismatch (ckpt: {new_ckpt[key].size()}, model: {self.state_dict()[key].size()})\"\n                            )\n                    else:\n                        printer.info(f\"Skipping '{key}': not found in model\")\n                return super().load_state_dict(new_new_ckpt, **kw)\n\n    def set_freeze(self, freeze):  # this is for use by downstream models\n        self.freeze = freeze\n        to_be_frozen = {\n            \"none\": [],\n            \"mask\": [self.mask_token] if hasattr(self, \"mask_token\") else [],\n            \"encoder\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n            ],\n            \"encoder_and_head\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.downstream_head,\n            ],\n            \"encoder_and_decoder\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n                self.register_tokens,\n                self.decoder_embed_state,\n                self.decoder_embed,\n                self.dec_norm,\n                self.dec_norm_state,\n            ],\n            \"decoder\": [\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n            ],\n            \"encoder_and_decoder_and_head\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n                self.register_tokens,\n                self.decoder_embed_state,\n                self.decoder_embed,\n                self.dec_norm,\n                self.dec_norm_state,\n                self.downstre\n# ... truncated ...","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.ARCroco3DStereoOutput","uri":"program://Human3R/class/src.dust3r.model.ARCroco3DStereoOutput#L56-L62","kind":"class","name":"ARCroco3DStereoOutput","path":"src/dust3r/model.py","language":"python","start_line":56,"end_line":62,"context_start_line":36,"context_end_line":82,"code":"    Attention,\n    CrossAttention,\n    DropPath,\n    CustomDecoderBlock,\n)  # noqa\n\ninf = float(\"inf\")\nfrom accelerate.logging import get_logger\n\nfrom dust3r.smpl_model import nms, apply_threshold\nfrom einops import rearrange\n\nfrom dust3r.utils.geometry import inverse_perspective_projection, get_camera_parameters\nfrom dust3r.utils.image import unpad_uv, log_optimal_transport\nfrom mhmr.blocks import Dinov2Backbone, FourierPositionEncoding, TransformerDecoder\nprinter = get_logger(__name__, log_level=\"DEBUG\")\n\nfrom dust3r.utils.device import to_cpu, to_gpu\n\n@dataclass\nclass ARCroco3DStereoOutput(ModelOutput):\n    \"\"\"\n    Custom output class for ARCroco3DStereo.\n    \"\"\"\n\n    ress: Optional[List[Any]] = None\n    views: Optional[List[Any]] = None\n\n\ndef strip_module(state_dict):\n    \"\"\"\n    Removes the 'module.' prefix from the keys of a state_dict.\n    Args:\n        state_dict (dict): The original state_dict with possible 'module.' prefixes.\n    Returns:\n        OrderedDict: A new state_dict with 'module.' prefixes removed.\n    \"\"\"\n    new_state_dict = OrderedDict()\n    for k, v in state_dict.items():\n        name = k[7:] if k.startswith(\"module.\") else k\n        new_state_dict[name] = v\n    return new_state_dict\n\n\ndef strip_module_mhmr(state_dict):\n    \"\"\"\n    Load Multi-HMR pretrained model","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.strip_module","uri":"program://Human3R/function/src.dust3r.model.strip_module#L65-L77","kind":"function","name":"strip_module","path":"src/dust3r/model.py","language":"python","start_line":65,"end_line":77,"context_start_line":45,"context_end_line":97,"code":"from dust3r.smpl_model import nms, apply_threshold\nfrom einops import rearrange\n\nfrom dust3r.utils.geometry import inverse_perspective_projection, get_camera_parameters\nfrom dust3r.utils.image import unpad_uv, log_optimal_transport\nfrom mhmr.blocks import Dinov2Backbone, FourierPositionEncoding, TransformerDecoder\nprinter = get_logger(__name__, log_level=\"DEBUG\")\n\nfrom dust3r.utils.device import to_cpu, to_gpu\n\n@dataclass\nclass ARCroco3DStereoOutput(ModelOutput):\n    \"\"\"\n    Custom output class for ARCroco3DStereo.\n    \"\"\"\n\n    ress: Optional[List[Any]] = None\n    views: Optional[List[Any]] = None\n\n\ndef strip_module(state_dict):\n    \"\"\"\n    Removes the 'module.' prefix from the keys of a state_dict.\n    Args:\n        state_dict (dict): The original state_dict with possible 'module.' prefixes.\n    Returns:\n        OrderedDict: A new state_dict with 'module.' prefixes removed.\n    \"\"\"\n    new_state_dict = OrderedDict()\n    for k, v in state_dict.items():\n        name = k[7:] if k.startswith(\"module.\") else k\n        new_state_dict[name] = v\n    return new_state_dict\n\n\ndef strip_module_mhmr(state_dict):\n    \"\"\"\n    Load Multi-HMR pretrained model\n    \"\"\"\n    new_state_dict = OrderedDict()\n    for k, v in state_dict.items():\n        if k.startswith((\"mlp_classif.\", \"mlp_offset.\")):\n            name = f\"downstream_head.{k}\"\n        elif k.startswith((\"x_attention_head.dec\")):\n            name = f\"downstream_head.{k[17:]}\"\n        elif k.startswith((\"x_attention_head.transformer.\", \"x_attention_head.cross_\")):\n            name = k[17:]\n        elif k.startswith((\"backbone\")):\n            name = k\n        else:\n            continue\n        new_state_dict[name] = v\n    return new_state_dict","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.strip_module_mhmr","uri":"program://Human3R/function/src.dust3r.model.strip_module_mhmr#L80-L97","kind":"function","name":"strip_module_mhmr","path":"src/dust3r/model.py","language":"python","start_line":80,"end_line":97,"context_start_line":60,"context_end_line":117,"code":"\n    ress: Optional[List[Any]] = None\n    views: Optional[List[Any]] = None\n\n\ndef strip_module(state_dict):\n    \"\"\"\n    Removes the 'module.' prefix from the keys of a state_dict.\n    Args:\n        state_dict (dict): The original state_dict with possible 'module.' prefixes.\n    Returns:\n        OrderedDict: A new state_dict with 'module.' prefixes removed.\n    \"\"\"\n    new_state_dict = OrderedDict()\n    for k, v in state_dict.items():\n        name = k[7:] if k.startswith(\"module.\") else k\n        new_state_dict[name] = v\n    return new_state_dict\n\n\ndef strip_module_mhmr(state_dict):\n    \"\"\"\n    Load Multi-HMR pretrained model\n    \"\"\"\n    new_state_dict = OrderedDict()\n    for k, v in state_dict.items():\n        if k.startswith((\"mlp_classif.\", \"mlp_offset.\")):\n            name = f\"downstream_head.{k}\"\n        elif k.startswith((\"x_attention_head.dec\")):\n            name = f\"downstream_head.{k[17:]}\"\n        elif k.startswith((\"x_attention_head.transformer.\", \"x_attention_head.cross_\")):\n            name = k[17:]\n        elif k.startswith((\"backbone\")):\n            name = k\n        else:\n            continue\n        new_state_dict[name] = v\n    return new_state_dict\n\n\ndef load_model(model_path, device, verbose=True):\n    if verbose:\n        print(\"... loading model from\", model_path)\n    ckpt = torch.load(model_path, map_location=\"cpu\")\n    args = ckpt[\"args\"].model.replace(\n        \"ManyAR_PatchEmbed\", \"PatchEmbedDust3R\"\n    )  # ManyAR only for aspect ratio not consistent\n    if \"landscape_only\" not in args:\n        args = args[:-2] + \", landscape_only=False))\"\n    else:\n        args = args.replace(\" \", \"\").replace(\n            \"landscape_only=True\", \"landscape_only=False\"\n        )\n    assert \"landscape_only=False\" in args\n    if verbose:\n        print(f\"instantiating : {args}\")\n    net = eval(args)\n    s = net.load_state_dict(ckpt[\"model\"], strict=False)","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.load_model","uri":"program://Human3R/function/src.dust3r.model.load_model#L100-L120","kind":"function","name":"load_model","path":"src/dust3r/model.py","language":"python","start_line":100,"end_line":120,"context_start_line":80,"context_end_line":140,"code":"def strip_module_mhmr(state_dict):\n    \"\"\"\n    Load Multi-HMR pretrained model\n    \"\"\"\n    new_state_dict = OrderedDict()\n    for k, v in state_dict.items():\n        if k.startswith((\"mlp_classif.\", \"mlp_offset.\")):\n            name = f\"downstream_head.{k}\"\n        elif k.startswith((\"x_attention_head.dec\")):\n            name = f\"downstream_head.{k[17:]}\"\n        elif k.startswith((\"x_attention_head.transformer.\", \"x_attention_head.cross_\")):\n            name = k[17:]\n        elif k.startswith((\"backbone\")):\n            name = k\n        else:\n            continue\n        new_state_dict[name] = v\n    return new_state_dict\n\n\ndef load_model(model_path, device, verbose=True):\n    if verbose:\n        print(\"... loading model from\", model_path)\n    ckpt = torch.load(model_path, map_location=\"cpu\")\n    args = ckpt[\"args\"].model.replace(\n        \"ManyAR_PatchEmbed\", \"PatchEmbedDust3R\"\n    )  # ManyAR only for aspect ratio not consistent\n    if \"landscape_only\" not in args:\n        args = args[:-2] + \", landscape_only=False))\"\n    else:\n        args = args.replace(\" \", \"\").replace(\n            \"landscape_only=True\", \"landscape_only=False\"\n        )\n    assert \"landscape_only=False\" in args\n    if verbose:\n        print(f\"instantiating : {args}\")\n    net = eval(args)\n    s = net.load_state_dict(ckpt[\"model\"], strict=False)\n    if verbose:\n        print(s)\n    return net.to(device)\n\n\nclass ARCroco3DStereoConfig(PretrainedConfig):\n    model_type = \"arcroco_3d_stereo\"\n\n    def __init__(\n        self,\n        output_mode=\"pts3d\",\n        head_type=\"linear\",  # or dpt\n        depth_mode=(\"exp\", -float(\"inf\"), float(\"inf\")),\n        conf_mode=(\"exp\", 1, float(\"inf\")),\n        pose_mode=(\"exp\", -float(\"inf\"), float(\"inf\")),\n        freeze=\"none\",\n        landscape_only=True,\n        patch_embed_cls=\"PatchEmbedDust3R\",\n        ray_enc_depth=2,\n        state_size=324,\n        local_mem_size=256,\n        state_pe=\"2d\",\n        state_dec_num_heads=16,","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.ARCroco3DStereoConfig","uri":"program://Human3R/class/src.dust3r.model.ARCroco3DStereoConfig#L123-L174","kind":"class","name":"ARCroco3DStereoConfig","path":"src/dust3r/model.py","language":"python","start_line":123,"end_line":174,"context_start_line":103,"context_end_line":194,"code":"    ckpt = torch.load(model_path, map_location=\"cpu\")\n    args = ckpt[\"args\"].model.replace(\n        \"ManyAR_PatchEmbed\", \"PatchEmbedDust3R\"\n    )  # ManyAR only for aspect ratio not consistent\n    if \"landscape_only\" not in args:\n        args = args[:-2] + \", landscape_only=False))\"\n    else:\n        args = args.replace(\" \", \"\").replace(\n            \"landscape_only=True\", \"landscape_only=False\"\n        )\n    assert \"landscape_only=False\" in args\n    if verbose:\n        print(f\"instantiating : {args}\")\n    net = eval(args)\n    s = net.load_state_dict(ckpt[\"model\"], strict=False)\n    if verbose:\n        print(s)\n    return net.to(device)\n\n\nclass ARCroco3DStereoConfig(PretrainedConfig):\n    model_type = \"arcroco_3d_stereo\"\n\n    def __init__(\n        self,\n        output_mode=\"pts3d\",\n        head_type=\"linear\",  # or dpt\n        depth_mode=(\"exp\", -float(\"inf\"), float(\"inf\")),\n        conf_mode=(\"exp\", 1, float(\"inf\")),\n        pose_mode=(\"exp\", -float(\"inf\"), float(\"inf\")),\n        freeze=\"none\",\n        landscape_only=True,\n        patch_embed_cls=\"PatchEmbedDust3R\",\n        ray_enc_depth=2,\n        state_size=324,\n        local_mem_size=256,\n        state_pe=\"2d\",\n        state_dec_num_heads=16,\n        depth_head=False,\n        rgb_head=False,\n        pose_conf_head=False,\n        pose_head=False,\n        msk_head=False,\n        use_prompt=False,\n        is_shallow=False,\n        prompt_size=None,\n        backbone='dinov2_vitl14',\n        mhmr_img_res=None,\n        **croco_kwargs,\n    ):\n        super().__init__()\n        self.output_mode = output_mode\n        self.head_type = head_type\n        self.depth_mode = depth_mode\n        self.conf_mode = conf_mode\n        self.pose_mode = pose_mode\n        self.freeze = freeze\n        self.landscape_only = landscape_only\n        self.patch_embed_cls = patch_embed_cls\n        self.ray_enc_depth = ray_enc_depth\n        self.state_size = state_size\n        self.state_pe = state_pe\n        self.state_dec_num_heads = state_dec_num_heads\n        self.local_mem_size = local_mem_size\n        self.depth_head = depth_head\n        self.rgb_head = rgb_head\n        self.pose_conf_head = pose_conf_head\n        self.pose_head = pose_head\n        self.msk_head = msk_head\n        self.backbone = backbone\n        self.mhmr_img_res = mhmr_img_res\n        self.croco_kwargs = croco_kwargs\n\n\nclass LocalMemory(nn.Module):\n    def __init__(\n        self,\n        size,\n        k_dim,\n        v_dim,\n        num_heads,\n        depth=2,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ) -> None:","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.LocalMemory","uri":"program://Human3R/class/src.dust3r.model.LocalMemory#L177-L259","kind":"class","name":"LocalMemory","path":"src/dust3r/model.py","language":"python","start_line":177,"end_line":259,"context_start_line":157,"context_end_line":279,"code":"        self.conf_mode = conf_mode\n        self.pose_mode = pose_mode\n        self.freeze = freeze\n        self.landscape_only = landscape_only\n        self.patch_embed_cls = patch_embed_cls\n        self.ray_enc_depth = ray_enc_depth\n        self.state_size = state_size\n        self.state_pe = state_pe\n        self.state_dec_num_heads = state_dec_num_heads\n        self.local_mem_size = local_mem_size\n        self.depth_head = depth_head\n        self.rgb_head = rgb_head\n        self.pose_conf_head = pose_conf_head\n        self.pose_head = pose_head\n        self.msk_head = msk_head\n        self.backbone = backbone\n        self.mhmr_img_res = mhmr_img_res\n        self.croco_kwargs = croco_kwargs\n\n\nclass LocalMemory(nn.Module):\n    def __init__(\n        self,\n        size,\n        k_dim,\n        v_dim,\n        num_heads,\n        depth=2,\n        mlp_ratio=4.0,\n        qkv_bias=False,\n        drop=0.0,\n        attn_drop=0.0,\n        drop_path=0.0,\n        act_layer=nn.GELU,\n        norm_layer=nn.LayerNorm,\n        norm_mem=True,\n        rope=None,\n    ) -> None:\n        super().__init__()\n        self.v_dim = v_dim\n        self.proj_q = nn.Linear(k_dim, v_dim)\n        self.masked_token = nn.Parameter(\n            torch.randn(1, 1, v_dim) * 0.2, requires_grad=True\n        )\n        self.mem = nn.Parameter(\n            torch.randn(1, size, 2 * v_dim) * 0.2, requires_grad=True\n        )\n        self.write_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    2 * v_dim,\n                    num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=qkv_bias,\n                    norm_layer=norm_layer,\n                    attn_drop=attn_drop,\n                    drop=drop,\n                    drop_path=drop_path,\n                    act_layer=act_layer,\n                    norm_mem=norm_mem,\n                    rope=rope,\n                )\n                for _ in range(depth)\n            ]\n        )\n        self.read_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    2 * v_dim,\n                    num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=qkv_bias,\n                    norm_layer=norm_layer,\n                    attn_drop=attn_drop,\n                    drop=drop,\n                    drop_path=drop_path,\n                    act_layer=act_layer,\n                    norm_mem=norm_mem,\n                    rope=rope,\n                )\n                for _ in range(depth)\n            ]\n        )\n\n    def update_mem(self, mem, feat_k, feat_v):\n        \"\"\"\n        mem_k: [B, size, C]\n        mem_v: [B, size, C]\n        feat_k: [B, 1, C]\n        feat_v: [B, 1, C]\n        \"\"\"\n        feat_k = self.proj_q(feat_k)  # [B, 1, C]\n        feat = torch.cat([feat_k, feat_v], dim=-1)\n        for blk in self.write_blocks:\n            mem, _, _ = blk(mem, feat, None, None)\n        return mem\n\n    def inquire(self, query, mem):\n        x = self.proj_q(query)  # [B, 1, C]\n        x = torch.cat([x, self.masked_token.expand(x.shape[0], -1, -1)], dim=-1)\n        for blk in self.read_blocks:\n            x, _, _ = blk(x, mem, None, None)\n        return x[..., -self.v_dim :]\n\n\nclass ARCroco3DStereo(CroCoNet):\n    config_class = ARCroco3DStereoConfig\n    base_model_prefix = \"arcroco3dstereo\"\n    supports_gradient_checkpointing = True\n\n    def __init__(self, config: ARCroco3DStereoConfig):\n        self.gradient_checkpointing = False\n        self.fixed_input_length = True\n        config.croco_kwargs = fill_default_args(\n            config.croco_kwargs, CrocoConfig.__init__\n        )\n        self.config = config\n        self.patch_embed_cls = config.patch_embed_cls\n        self.croco_args = config.croco_kwargs\n        croco_cfg = CrocoConfig(**self.croco_args)\n        super().__init__(croco_cfg)\n        self.enc_blocks_ray_map = nn.ModuleList(\n            [","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.ARCroco3DStereo","uri":"program://Human3R/class/src.dust3r.model.ARCroco3DStereo#L262-L1958","kind":"class","name":"ARCroco3DStereo","path":"src/dust3r/model.py","language":"python","start_line":262,"end_line":1958,"context_start_line":242,"context_end_line":1978,"code":"        \"\"\"\n        mem_k: [B, size, C]\n        mem_v: [B, size, C]\n        feat_k: [B, 1, C]\n        feat_v: [B, 1, C]\n        \"\"\"\n        feat_k = self.proj_q(feat_k)  # [B, 1, C]\n        feat = torch.cat([feat_k, feat_v], dim=-1)\n        for blk in self.write_blocks:\n            mem, _, _ = blk(mem, feat, None, None)\n        return mem\n\n    def inquire(self, query, mem):\n        x = self.proj_q(query)  # [B, 1, C]\n        x = torch.cat([x, self.masked_token.expand(x.shape[0], -1, -1)], dim=-1)\n        for blk in self.read_blocks:\n            x, _, _ = blk(x, mem, None, None)\n        return x[..., -self.v_dim :]\n\n\nclass ARCroco3DStereo(CroCoNet):\n    config_class = ARCroco3DStereoConfig\n    base_model_prefix = \"arcroco3dstereo\"\n    supports_gradient_checkpointing = True\n\n    def __init__(self, config: ARCroco3DStereoConfig):\n        self.gradient_checkpointing = False\n        self.fixed_input_length = True\n        config.croco_kwargs = fill_default_args(\n            config.croco_kwargs, CrocoConfig.__init__\n        )\n        self.config = config\n        self.patch_embed_cls = config.patch_embed_cls\n        self.croco_args = config.croco_kwargs\n        croco_cfg = CrocoConfig(**self.croco_args)\n        super().__init__(croco_cfg)\n        self.enc_blocks_ray_map = nn.ModuleList(\n            [\n                Block(\n                    self.enc_embed_dim,\n                    16,\n                    4,\n                    qkv_bias=True,\n                    norm_layer=partial(nn.LayerNorm, eps=1e-6),\n                    rope=self.rope,\n                )\n                for _ in range(config.ray_enc_depth)\n            ]\n        )\n        self.enc_norm_ray_map = nn.LayerNorm(self.enc_embed_dim, eps=1e-6)\n        self.dec_num_heads = self.croco_args[\"dec_num_heads\"]\n        self.pose_head_flag = config.pose_head\n        self.msk_head_flag = config.msk_head\n        if self.pose_head_flag:\n            self.pose_token = nn.Parameter(\n                torch.randn(1, 1, self.dec_embed_dim) * 0.02, requires_grad=True\n            )\n            self.pose_retriever = LocalMemory(\n                size=config.local_mem_size,\n                k_dim=self.enc_embed_dim,\n                v_dim=self.dec_embed_dim,\n                num_heads=self.dec_num_heads,\n                mlp_ratio=4,\n                qkv_bias=True,\n                attn_drop=0.0,\n                norm_layer=partial(nn.LayerNorm, eps=1e-6),\n                rope=None,\n            )\n        self.register_tokens = nn.Embedding(config.state_size, self.enc_embed_dim)\n        self.state_size = config.state_size\n        self.state_pe = config.state_pe\n        self.masked_img_token = nn.Parameter(\n            torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True\n        )\n        self.masked_ray_map_token = nn.Parameter(\n            torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True\n        )\n        self.masked_smpl_token = nn.Parameter(\n            torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True\n        )\n\n        # MHMR\n        # 'dinov2_vits14': 384, 'dinov2_vitb14': 768, 'dinov2_vitl14': 1024\n        self.backbone = Dinov2Backbone(config.backbone, pretrained=False)\n        self.bb_patch_size = self.backbone.patch_size\n        self.backbone_dim = self.backbone.embed_dim\n        self.mhmr_img_res = config.mhmr_img_res\n        self.bb_token_res = self.mhmr_img_res // self.bb_patch_size\n\n        if config.output_mode == 'naive':\n            self.fourier_camera = FourierPositionEncoding(n=3, num_bands=16, max_resolution=64)\n            self.camera_embed_dim = self.fourier_camera.channels\n            context_dim = self.backbone_dim + self.camera_embed_dim\n\n            transformer_args = dict(\n                num_tokens=1,\n                token_dim=(318+10+3+context_dim),\n                dim=1024,\n                depth=2,\n                heads=8,\n                mlp_dim=1024,\n                dim_head=32,\n                dropout=0.0,\n                emb_dropout=0.0,\n                context_dim=context_dim,\n            )\n            self.transformer = TransformerDecoder(**transformer_args)\n            # Init learned embeddings for queries\n            self.cross_queries_x = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_queries_x, std=0.2)\n            self.cross_queries_y = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_queries_y, std=0.2)\n            self.cross_values_x = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_values_x, std=0.2)\n            self.cross_values_y = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_values_y, std=0.2)\n\n\n        self.mhmr_masked_smpl_token = nn.Parameter(\n            torch.randn(\n                1, context_dim if config.output_mode == \"naive\" else self.backbone_dim\n                ) * 0.02, requires_grad=True\n        )\n        self.mhmr_masked_img_token = nn.Parameter(\n            torch.randn(1, self.backbone_dim) * 0.02, requires_grad=True\n        )\n\n        self._set_state_decoder(\n            self.enc_embed_dim,\n            self.dec_embed_dim,\n            config.state_dec_num_heads,\n            self.dec_depth,\n            self.croco_args.get(\"mlp_ratio\", None),\n            self.croco_args.get(\"norm_layer\", None),\n            self.croco_args.get(\"norm_im2_in_dec\", None),\n        )\n        self.set_downstream_head(\n            config.output_mode,\n            config.head_type,\n            config.landscape_only,\n            config.depth_mode,\n            config.conf_mode,\n            config.pose_mode,\n            config.depth_head,\n            config.rgb_head,\n            config.pose_conf_head,\n            config.pose_head,\n            config.msk_head,\n            **self.croco_args,\n        )\n        self.set_freeze(config.freeze)\n\n    @classmethod\n    def from_pretrained(cls, pretrained_model_name_or_path, **kw):\n        if os.path.isfile(pretrained_model_name_or_path):\n            return load_model(pretrained_model_name_or_path, device=\"cpu\")\n        else:\n            try:\n                model = super(ARCroco3DStereo, cls).from_pretrained(\n                    pretrained_model_name_or_path, **kw\n                )\n            except TypeError as e:\n                raise Exception(\n                    f\"tried to load {pretrained_model_name_or_path} from huggingface, but failed\"\n                )\n            return model\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3\n        )\n        self.patch_embed_ray_map = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=6\n        )\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_state_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth_state = dec_depth\n        self.dec_embed_dim_state = dec_embed_dim\n        self.decoder_embed_state = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks_state = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        self.dec_norm_state = norm_layer(dec_embed_dim)\n\n    def load_state_dict(self, ckpt, **kw):\n        if all(k.startswith(\"module\") for k in ckpt):\n            ckpt = strip_module(ckpt)\n        new_ckpt = dict(ckpt)\n        if not any(k.startswith(\"dec_blocks_state\") for k in ckpt):\n            for key, value in ckpt.items():\n                if key.startswith(\"dec_blocks\"):\n                    new_ckpt[key.replace(\"dec_blocks\", \"dec_blocks_state\")] = value\n        try:\n            return super().load_state_dict(new_ckpt, **kw)\n        except:\n            try:\n                new_new_ckpt = {\n                    k: v\n                    for k, v in new_ckpt.items()\n                    if not k.startswith(\"dec_blocks\")\n                    and not k.startswith(\"dec_norm\")\n                    and not k.startswith(\"decoder_embed\")\n                }\n                return super().load_state_dict(new_new_ckpt, **kw)\n            except:\n                new_new_ckpt = {}\n                for key in new_ckpt:\n                    if key in self.state_dict():\n                        if new_ckpt[key].size() == self.state_dict()[key].size():\n                            new_new_ckpt[key] = new_ckpt[key]\n                        else:\n                            printer.info(\n                                f\"Skipping '{key}': size mismatch (ckpt: {new_ckpt[key].size()}, model: {self.state_dict()[key].size()})\"\n                            )\n                    else:\n                        printer.info(f\"Skipping '{key}': not found in model\")\n                return super().load_state_dict(new_new_ckpt, **kw)\n\n    def set_freeze(self, freeze):  # this is for use by downstream models\n        self.freeze = freeze\n        to_be_frozen = {\n            \"none\": [],\n            \"mask\": [self.mask_token] if hasattr(self, \"mask_token\") else [],\n            \"encoder\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n            ],\n            \"encoder_and_head\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.downstream_head,\n            ],\n            \"encoder_and_decoder\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n                self.register_tokens,\n                self.decoder_embed_state,\n                self.decoder_embed,\n                self.dec_norm,\n                self.dec_norm_state,\n            ],\n            \"decoder\": [\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n            ],\n            \"encoder_and_decoder_and_head\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n                self.register_tokens,\n                self.decoder_embed_state,\n                self.decoder_embed,\n                self.dec_norm,\n                self.dec_norm_state,\n                self.downstream_head.dpt_self,\n                self.downstream_head.final_transform,\n                self.downstream_head.dpt_cross,\n                self.downstream_head.dpt_rgb,\n                self.downstream_head.pose_head,\n            ],\n            \"mhmr\": [\n                self.backbone,\n                self.mhmr_masked_img_token,\n                self.downstream_head.mlp_classif,\n                self.downstream_head.mlp_offset,\n            ],\n        }\n        if self.output_mode == \"naive\":\n            to_be_frozen[\"all\"] = [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n                self.register_tokens,\n                self.decoder_embed_state,\n                self.decoder_embed,\n                self.dec_norm,\n                self.dec_norm_state,\n                self.downstream_head.dpt_self,\n                self.downstream_head.final_transform,\n                self.downstream_head.dpt_cross,\n                self.downstream_head.dpt_rgb,\n                self.downstream_head.pose_head,\n                self.backbone,\n                self.mhmr_masked_smpl_token,\n                self.mhmr_masked_img_token,\n                self.transformer,\n                self.cross_queries_x,\n                self.cross_queries_y,\n                self.cross_values_x,\n                self.cross_values_y,\n                self.downstream_head.mlp_classif,\n                self.downstream_head.mlp_offset,\n                self.downstream_head.decpose, \n                self.downstream_head.decshape, \n                self.downstream_head.deccam, \n                self.downstream_head.decexpression,\n            ]\n\n        if freeze == \"encoder_and_decoder_and_head\":\n            fix_all_params(to_be_frozen[\"encoder_and_decoder_and_head\"]) # will not be updated\n            freeze_all_params(to_be_frozen[\"encoder\"]) # requires_grad = False\n            freeze_all_params(to_be_frozen[\"mhmr\"])\n        else:\n            freeze_all_params(to_be_frozen[freeze])\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head\"\"\"\n        return\n\n    def set_downstream_head(\n        self,\n        output_mode,\n        head_type,\n        landscape_only,\n        depth_mode,\n        conf_mode,\n        pose_mode,\n        depth_head,\n        rgb_head,\n        pose_conf_head,\n        pose_head,\n        msk_head,\n        patch_size,\n        img_size,\n        **kw,\n    ):\n        assert (\n            img_size[0] % patch_size == 0 and img_size[1] % patch_size == 0\n        ), f\"{img_size=} must be multiple of {patch_size=}\"\n        self.output_mode = output_mode\n        self.head_type = head_type\n        self.depth_mode = depth_mode\n        self.conf_mode = conf_mode\n        self.pose_mode = pose_mode\n        self.downstream_head = head_factory(\n            head_type,\n            output_mode,\n            self,\n            has_conf=bool(conf_mode),\n            has_depth=bool(depth_head),\n            has_rgb=bool(rgb_head),\n            has_pose_conf=bool(pose_conf_head),\n            has_pose=bool(pose_head),\n            has_msk=bool(msk_head),\n        )\n        self.head = transpose_to_landscape(\n            self.downstream_head, activate=landscape_only\n        )\n\n    def _encode_image(self, image, true_shape):\n        x, pos = self.patch_embed(image, true_shape=true_shape)\n        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)\n        x = self.enc_norm(x)\n        return [x], pos, None\n\n    def _encode_ray_map(self, ray_map, true_shape):\n        x, pos = self.patch_embed_ray_map(ray_map, true_shape=true_shape)\n        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks_ray_map:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)\n        x = self.enc_norm_ray_map(x)\n        return [x], pos, None\n\n    def _encode_state(self, image_tokens, image_pos):\n        batch_size = image_tokens.shape[0]\n        state_feat = self.register_tokens(\n            torch.arange(self.state_size, device=image_pos.device)\n        )\n        if self.state_pe == \"1d\":\n            state_pos = (\n                torch.tensor(\n                    [[i, i] for i in range(self.state_size)],\n                    dtype=image_pos.dtype,\n                    device=image_pos.device,\n                )[None]\n                .expand(batch_size, -1, -1)\n                .contiguous()\n            )  # .long()\n        elif self.state_pe == \"2d\":\n            width = int(self.state_size**0.5)\n            width = width + 1 if width % 2 == 1 else width\n            state_pos = (\n                torch.tensor(\n                    [[i // width, i % width] for i in range(self.state_size)],\n                    dtype=image_pos.dtype,\n                    device=image_pos.device,\n                )[None]\n                .expand(batch_size, -1, -1)\n                .contiguous()\n            )\n        elif self.state_pe == \"none\":\n            state_pos = None\n        state_feat = state_feat[None].expand(batch_size, -1, -1)\n        return state_feat, state_pos, None\n\n    def _encode_views_mhmr(self, views, img_mask=None, ray_mask=None):\n        device = views[0][\"img\"].device\n        batch_size = views[0][\"img\"].shape[0]\n        given = True\n        if img_mask is None and ray_mask is None:\n            given = False\n        if not given:\n            img_mask = torch.stack(\n                [view[\"img_mask\"] for view in views], dim=0\n            )  # Shape: (num_views, batch_size)\n            ray_mask = torch.stack(\n                [view[\"ray_mask\"] for view in views], dim=0\n            )  # Shape: (num_views, batch_size)\n        imgs = torch.stack(\n            [view[\"img\"] for view in views], dim=0\n        )  # Shape: (num_views, batch_size, C, H, W)\n        ray_maps = torch.stack(\n            [view[\"ray_map\"] for view in views], dim=0\n        )  # Shape: (num_views, batch_size, H, W, C)\n        shapes = []\n        for view in views:\n            if \"true_shape\" in view:\n                shapes.append(view[\"true_shape\"])\n            else:\n                shape = torch.tensor(view[\"img\"].shape[-2:], device=device)\n                shapes.append(shape.unsqueeze(0).repeat(batch_size, 1))\n        shapes = torch.stack(shapes, dim=0).to(\n            imgs.device\n        )  # Shape: (num_views, batch_size, 2)\n        imgs = imgs.view(\n            -1, *imgs.shape[2:]\n        )  # Shape: (num_views * batch_size, C, H, W)\n        ray_maps = ray_maps.view(\n            -1, *ray_maps.shape[2:]\n        )  # Shape: (num_views * batch_size, H, W, C)\n        shapes = shapes.view(-1, 2)  # Shape: (num_views * batch_size, 2)\n        img_masks_flat = img_mask.vie\n# ... truncated ...","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.__init__","uri":"program://Human3R/function/src.dust3r.model.__init__#L267-L392","kind":"function","name":"__init__","path":"src/dust3r/model.py","language":"python","start_line":267,"end_line":392,"context_start_line":247,"context_end_line":412,"code":"        \"\"\"\n        feat_k = self.proj_q(feat_k)  # [B, 1, C]\n        feat = torch.cat([feat_k, feat_v], dim=-1)\n        for blk in self.write_blocks:\n            mem, _, _ = blk(mem, feat, None, None)\n        return mem\n\n    def inquire(self, query, mem):\n        x = self.proj_q(query)  # [B, 1, C]\n        x = torch.cat([x, self.masked_token.expand(x.shape[0], -1, -1)], dim=-1)\n        for blk in self.read_blocks:\n            x, _, _ = blk(x, mem, None, None)\n        return x[..., -self.v_dim :]\n\n\nclass ARCroco3DStereo(CroCoNet):\n    config_class = ARCroco3DStereoConfig\n    base_model_prefix = \"arcroco3dstereo\"\n    supports_gradient_checkpointing = True\n\n    def __init__(self, config: ARCroco3DStereoConfig):\n        self.gradient_checkpointing = False\n        self.fixed_input_length = True\n        config.croco_kwargs = fill_default_args(\n            config.croco_kwargs, CrocoConfig.__init__\n        )\n        self.config = config\n        self.patch_embed_cls = config.patch_embed_cls\n        self.croco_args = config.croco_kwargs\n        croco_cfg = CrocoConfig(**self.croco_args)\n        super().__init__(croco_cfg)\n        self.enc_blocks_ray_map = nn.ModuleList(\n            [\n                Block(\n                    self.enc_embed_dim,\n                    16,\n                    4,\n                    qkv_bias=True,\n                    norm_layer=partial(nn.LayerNorm, eps=1e-6),\n                    rope=self.rope,\n                )\n                for _ in range(config.ray_enc_depth)\n            ]\n        )\n        self.enc_norm_ray_map = nn.LayerNorm(self.enc_embed_dim, eps=1e-6)\n        self.dec_num_heads = self.croco_args[\"dec_num_heads\"]\n        self.pose_head_flag = config.pose_head\n        self.msk_head_flag = config.msk_head\n        if self.pose_head_flag:\n            self.pose_token = nn.Parameter(\n                torch.randn(1, 1, self.dec_embed_dim) * 0.02, requires_grad=True\n            )\n            self.pose_retriever = LocalMemory(\n                size=config.local_mem_size,\n                k_dim=self.enc_embed_dim,\n                v_dim=self.dec_embed_dim,\n                num_heads=self.dec_num_heads,\n                mlp_ratio=4,\n                qkv_bias=True,\n                attn_drop=0.0,\n                norm_layer=partial(nn.LayerNorm, eps=1e-6),\n                rope=None,\n            )\n        self.register_tokens = nn.Embedding(config.state_size, self.enc_embed_dim)\n        self.state_size = config.state_size\n        self.state_pe = config.state_pe\n        self.masked_img_token = nn.Parameter(\n            torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True\n        )\n        self.masked_ray_map_token = nn.Parameter(\n            torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True\n        )\n        self.masked_smpl_token = nn.Parameter(\n            torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True\n        )\n\n        # MHMR\n        # 'dinov2_vits14': 384, 'dinov2_vitb14': 768, 'dinov2_vitl14': 1024\n        self.backbone = Dinov2Backbone(config.backbone, pretrained=False)\n        self.bb_patch_size = self.backbone.patch_size\n        self.backbone_dim = self.backbone.embed_dim\n        self.mhmr_img_res = config.mhmr_img_res\n        self.bb_token_res = self.mhmr_img_res // self.bb_patch_size\n\n        if config.output_mode == 'naive':\n            self.fourier_camera = FourierPositionEncoding(n=3, num_bands=16, max_resolution=64)\n            self.camera_embed_dim = self.fourier_camera.channels\n            context_dim = self.backbone_dim + self.camera_embed_dim\n\n            transformer_args = dict(\n                num_tokens=1,\n                token_dim=(318+10+3+context_dim),\n                dim=1024,\n                depth=2,\n                heads=8,\n                mlp_dim=1024,\n                dim_head=32,\n                dropout=0.0,\n                emb_dropout=0.0,\n                context_dim=context_dim,\n            )\n            self.transformer = TransformerDecoder(**transformer_args)\n            # Init learned embeddings for queries\n            self.cross_queries_x = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_queries_x, std=0.2)\n            self.cross_queries_y = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_queries_y, std=0.2)\n            self.cross_values_x = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_values_x, std=0.2)\n            self.cross_values_y = nn.Parameter(torch.zeros(self.bb_token_res, context_dim))\n            torch.nn.init.normal_(self.cross_values_y, std=0.2)\n\n\n        self.mhmr_masked_smpl_token = nn.Parameter(\n            torch.randn(\n                1, context_dim if config.output_mode == \"naive\" else self.backbone_dim\n                ) * 0.02, requires_grad=True\n        )\n        self.mhmr_masked_img_token = nn.Parameter(\n            torch.randn(1, self.backbone_dim) * 0.02, requires_grad=True\n        )\n\n        self._set_state_decoder(\n            self.enc_embed_dim,\n            self.dec_embed_dim,\n            config.state_dec_num_heads,\n            self.dec_depth,\n            self.croco_args.get(\"mlp_ratio\", None),\n            self.croco_args.get(\"norm_layer\", None),\n            self.croco_args.get(\"norm_im2_in_dec\", None),\n        )\n        self.set_downstream_head(\n            config.output_mode,\n            config.head_type,\n            config.landscape_only,\n            config.depth_mode,\n            config.conf_mode,\n            config.pose_mode,\n            config.depth_head,\n            config.rgb_head,\n            config.pose_conf_head,\n            config.pose_head,\n            config.msk_head,\n            **self.croco_args,\n        )\n        self.set_freeze(config.freeze)\n\n    @classmethod\n    def from_pretrained(cls, pretrained_model_name_or_path, **kw):\n        if os.path.isfile(pretrained_model_name_or_path):\n            return load_model(pretrained_model_name_or_path, device=\"cpu\")\n        else:\n            try:\n                model = super(ARCroco3DStereo, cls).from_pretrained(\n                    pretrained_model_name_or_path, **kw\n                )\n            except TypeError as e:\n                raise Exception(\n                    f\"tried to load {pretrained_model_name_or_path} from huggingface, but failed\"\n                )\n            return model\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3\n        )","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.update_mem","uri":"program://Human3R/function/src.dust3r.model.update_mem#L241-L252","kind":"function","name":"update_mem","path":"src/dust3r/model.py","language":"python","start_line":241,"end_line":252,"context_start_line":221,"context_end_line":272,"code":"        )\n        self.read_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    2 * v_dim,\n                    num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=qkv_bias,\n                    norm_layer=norm_layer,\n                    attn_drop=attn_drop,\n                    drop=drop,\n                    drop_path=drop_path,\n                    act_layer=act_layer,\n                    norm_mem=norm_mem,\n                    rope=rope,\n                )\n                for _ in range(depth)\n            ]\n        )\n\n    def update_mem(self, mem, feat_k, feat_v):\n        \"\"\"\n        mem_k: [B, size, C]\n        mem_v: [B, size, C]\n        feat_k: [B, 1, C]\n        feat_v: [B, 1, C]\n        \"\"\"\n        feat_k = self.proj_q(feat_k)  # [B, 1, C]\n        feat = torch.cat([feat_k, feat_v], dim=-1)\n        for blk in self.write_blocks:\n            mem, _, _ = blk(mem, feat, None, None)\n        return mem\n\n    def inquire(self, query, mem):\n        x = self.proj_q(query)  # [B, 1, C]\n        x = torch.cat([x, self.masked_token.expand(x.shape[0], -1, -1)], dim=-1)\n        for blk in self.read_blocks:\n            x, _, _ = blk(x, mem, None, None)\n        return x[..., -self.v_dim :]\n\n\nclass ARCroco3DStereo(CroCoNet):\n    config_class = ARCroco3DStereoConfig\n    base_model_prefix = \"arcroco3dstereo\"\n    supports_gradient_checkpointing = True\n\n    def __init__(self, config: ARCroco3DStereoConfig):\n        self.gradient_checkpointing = False\n        self.fixed_input_length = True\n        config.croco_kwargs = fill_default_args(\n            config.croco_kwargs, CrocoConfig.__init__\n        )","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.inquire","uri":"program://Human3R/function/src.dust3r.model.inquire#L254-L259","kind":"function","name":"inquire","path":"src/dust3r/model.py","language":"python","start_line":254,"end_line":259,"context_start_line":234,"context_end_line":279,"code":"                    norm_mem=norm_mem,\n                    rope=rope,\n                )\n                for _ in range(depth)\n            ]\n        )\n\n    def update_mem(self, mem, feat_k, feat_v):\n        \"\"\"\n        mem_k: [B, size, C]\n        mem_v: [B, size, C]\n        feat_k: [B, 1, C]\n        feat_v: [B, 1, C]\n        \"\"\"\n        feat_k = self.proj_q(feat_k)  # [B, 1, C]\n        feat = torch.cat([feat_k, feat_v], dim=-1)\n        for blk in self.write_blocks:\n            mem, _, _ = blk(mem, feat, None, None)\n        return mem\n\n    def inquire(self, query, mem):\n        x = self.proj_q(query)  # [B, 1, C]\n        x = torch.cat([x, self.masked_token.expand(x.shape[0], -1, -1)], dim=-1)\n        for blk in self.read_blocks:\n            x, _, _ = blk(x, mem, None, None)\n        return x[..., -self.v_dim :]\n\n\nclass ARCroco3DStereo(CroCoNet):\n    config_class = ARCroco3DStereoConfig\n    base_model_prefix = \"arcroco3dstereo\"\n    supports_gradient_checkpointing = True\n\n    def __init__(self, config: ARCroco3DStereoConfig):\n        self.gradient_checkpointing = False\n        self.fixed_input_length = True\n        config.croco_kwargs = fill_default_args(\n            config.croco_kwargs, CrocoConfig.__init__\n        )\n        self.config = config\n        self.patch_embed_cls = config.patch_embed_cls\n        self.croco_args = config.croco_kwargs\n        croco_cfg = CrocoConfig(**self.croco_args)\n        super().__init__(croco_cfg)\n        self.enc_blocks_ray_map = nn.ModuleList(\n            [","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.from_pretrained","uri":"program://Human3R/function/src.dust3r.model.from_pretrained#L395-L407","kind":"function","name":"from_pretrained","path":"src/dust3r/model.py","language":"python","start_line":395,"end_line":407,"context_start_line":375,"context_end_line":427,"code":"            self.croco_args.get(\"norm_layer\", None),\n            self.croco_args.get(\"norm_im2_in_dec\", None),\n        )\n        self.set_downstream_head(\n            config.output_mode,\n            config.head_type,\n            config.landscape_only,\n            config.depth_mode,\n            config.conf_mode,\n            config.pose_mode,\n            config.depth_head,\n            config.rgb_head,\n            config.pose_conf_head,\n            config.pose_head,\n            config.msk_head,\n            **self.croco_args,\n        )\n        self.set_freeze(config.freeze)\n\n    @classmethod\n    def from_pretrained(cls, pretrained_model_name_or_path, **kw):\n        if os.path.isfile(pretrained_model_name_or_path):\n            return load_model(pretrained_model_name_or_path, device=\"cpu\")\n        else:\n            try:\n                model = super(ARCroco3DStereo, cls).from_pretrained(\n                    pretrained_model_name_or_path, **kw\n                )\n            except TypeError as e:\n                raise Exception(\n                    f\"tried to load {pretrained_model_name_or_path} from huggingface, but failed\"\n                )\n            return model\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3\n        )\n        self.patch_embed_ray_map = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=6\n        )\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._set_patch_embed","uri":"program://Human3R/function/src.dust3r.model._set_patch_embed#L409-L415","kind":"function","name":"_set_patch_embed","path":"src/dust3r/model.py","language":"python","start_line":409,"end_line":415,"context_start_line":389,"context_end_line":435,"code":"            config.msk_head,\n            **self.croco_args,\n        )\n        self.set_freeze(config.freeze)\n\n    @classmethod\n    def from_pretrained(cls, pretrained_model_name_or_path, **kw):\n        if os.path.isfile(pretrained_model_name_or_path):\n            return load_model(pretrained_model_name_or_path, device=\"cpu\")\n        else:\n            try:\n                model = super(ARCroco3DStereo, cls).from_pretrained(\n                    pretrained_model_name_or_path, **kw\n                )\n            except TypeError as e:\n                raise Exception(\n                    f\"tried to load {pretrained_model_name_or_path} from huggingface, but failed\"\n                )\n            return model\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3\n        )\n        self.patch_embed_ray_map = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=6\n        )\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._set_decoder","uri":"program://Human3R/function/src.dust3r.model._set_decoder#L417-L444","kind":"function","name":"_set_decoder","path":"src/dust3r/model.py","language":"python","start_line":417,"end_line":444,"context_start_line":397,"context_end_line":464,"code":"            return load_model(pretrained_model_name_or_path, device=\"cpu\")\n        else:\n            try:\n                model = super(ARCroco3DStereo, cls).from_pretrained(\n                    pretrained_model_name_or_path, **kw\n                )\n            except TypeError as e:\n                raise Exception(\n                    f\"tried to load {pretrained_model_name_or_path} from huggingface, but failed\"\n                )\n            return model\n\n    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):\n        self.patch_embed = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3\n        )\n        self.patch_embed_ray_map = get_patch_embed(\n            self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=6\n        )\n\n    def _set_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_state_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth_state = dec_depth\n        self.dec_embed_dim_state = dec_embed_dim\n        self.decoder_embed_state = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks_state = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._set_state_decoder","uri":"program://Human3R/function/src.dust3r.model._set_state_decoder#L446-L473","kind":"function","name":"_set_state_decoder","path":"src/dust3r/model.py","language":"python","start_line":446,"end_line":473,"context_start_line":426,"context_end_line":493,"code":"    ):\n        self.dec_depth = dec_depth\n        self.dec_embed_dim = dec_embed_dim\n        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        self.dec_norm = norm_layer(dec_embed_dim)\n\n    def _set_state_decoder(\n        self,\n        enc_embed_dim,\n        dec_embed_dim,\n        dec_num_heads,\n        dec_depth,\n        mlp_ratio,\n        norm_layer,\n        norm_im2_in_dec,\n    ):\n        self.dec_depth_state = dec_depth\n        self.dec_embed_dim_state = dec_embed_dim\n        self.decoder_embed_state = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks_state = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        self.dec_norm_state = norm_layer(dec_embed_dim)\n\n    def load_state_dict(self, ckpt, **kw):\n        if all(k.startswith(\"module\") for k in ckpt):\n            ckpt = strip_module(ckpt)\n        new_ckpt = dict(ckpt)\n        if not any(k.startswith(\"dec_blocks_state\") for k in ckpt):\n            for key, value in ckpt.items():\n                if key.startswith(\"dec_blocks\"):\n                    new_ckpt[key.replace(\"dec_blocks\", \"dec_blocks_state\")] = value\n        try:\n            return super().load_state_dict(new_ckpt, **kw)\n        except:\n            try:\n                new_new_ckpt = {\n                    k: v\n                    for k, v in new_ckpt.items()\n                    if not k.startswith(\"dec_blocks\")\n                    and not k.startswith(\"dec_norm\")\n                    and not k.startswith(\"decoder_embed\")\n                }","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.load_state_dict","uri":"program://Human3R/function/src.dust3r.model.load_state_dict#L475-L507","kind":"function","name":"load_state_dict","path":"src/dust3r/model.py","language":"python","start_line":475,"end_line":507,"context_start_line":455,"context_end_line":527,"code":"    ):\n        self.dec_depth_state = dec_depth\n        self.dec_embed_dim_state = dec_embed_dim\n        self.decoder_embed_state = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)\n        self.dec_blocks_state = nn.ModuleList(\n            [\n                DecoderBlock(\n                    dec_embed_dim,\n                    dec_num_heads,\n                    mlp_ratio=mlp_ratio,\n                    qkv_bias=True,\n                    norm_layer=norm_layer,\n                    norm_mem=norm_im2_in_dec,\n                    rope=self.rope,\n                )\n                for i in range(dec_depth)\n            ]\n        )\n        self.dec_norm_state = norm_layer(dec_embed_dim)\n\n    def load_state_dict(self, ckpt, **kw):\n        if all(k.startswith(\"module\") for k in ckpt):\n            ckpt = strip_module(ckpt)\n        new_ckpt = dict(ckpt)\n        if not any(k.startswith(\"dec_blocks_state\") for k in ckpt):\n            for key, value in ckpt.items():\n                if key.startswith(\"dec_blocks\"):\n                    new_ckpt[key.replace(\"dec_blocks\", \"dec_blocks_state\")] = value\n        try:\n            return super().load_state_dict(new_ckpt, **kw)\n        except:\n            try:\n                new_new_ckpt = {\n                    k: v\n                    for k, v in new_ckpt.items()\n                    if not k.startswith(\"dec_blocks\")\n                    and not k.startswith(\"dec_norm\")\n                    and not k.startswith(\"decoder_embed\")\n                }\n                return super().load_state_dict(new_new_ckpt, **kw)\n            except:\n                new_new_ckpt = {}\n                for key in new_ckpt:\n                    if key in self.state_dict():\n                        if new_ckpt[key].size() == self.state_dict()[key].size():\n                            new_new_ckpt[key] = new_ckpt[key]\n                        else:\n                            printer.info(\n                                f\"Skipping '{key}': size mismatch (ckpt: {new_ckpt[key].size()}, model: {self.state_dict()[key].size()})\"\n                            )\n                    else:\n                        printer.info(f\"Skipping '{key}': not found in model\")\n                return super().load_state_dict(new_new_ckpt, **kw)\n\n    def set_freeze(self, freeze):  # this is for use by downstream models\n        self.freeze = freeze\n        to_be_frozen = {\n            \"none\": [],\n            \"mask\": [self.mask_token] if hasattr(self, \"mask_token\") else [],\n            \"encoder\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n            ],\n            \"encoder_and_head\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.set_freeze","uri":"program://Human3R/function/src.dust3r.model.set_freeze#L509-L635","kind":"function","name":"set_freeze","path":"src/dust3r/model.py","language":"python","start_line":509,"end_line":635,"context_start_line":489,"context_end_line":655,"code":"                    for k, v in new_ckpt.items()\n                    if not k.startswith(\"dec_blocks\")\n                    and not k.startswith(\"dec_norm\")\n                    and not k.startswith(\"decoder_embed\")\n                }\n                return super().load_state_dict(new_new_ckpt, **kw)\n            except:\n                new_new_ckpt = {}\n                for key in new_ckpt:\n                    if key in self.state_dict():\n                        if new_ckpt[key].size() == self.state_dict()[key].size():\n                            new_new_ckpt[key] = new_ckpt[key]\n                        else:\n                            printer.info(\n                                f\"Skipping '{key}': size mismatch (ckpt: {new_ckpt[key].size()}, model: {self.state_dict()[key].size()})\"\n                            )\n                    else:\n                        printer.info(f\"Skipping '{key}': not found in model\")\n                return super().load_state_dict(new_new_ckpt, **kw)\n\n    def set_freeze(self, freeze):  # this is for use by downstream models\n        self.freeze = freeze\n        to_be_frozen = {\n            \"none\": [],\n            \"mask\": [self.mask_token] if hasattr(self, \"mask_token\") else [],\n            \"encoder\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n            ],\n            \"encoder_and_head\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.downstream_head,\n            ],\n            \"encoder_and_decoder\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n                self.register_tokens,\n                self.decoder_embed_state,\n                self.decoder_embed,\n                self.dec_norm,\n                self.dec_norm_state,\n            ],\n            \"decoder\": [\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n            ],\n            \"encoder_and_decoder_and_head\": [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_img_token,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n                self.register_tokens,\n                self.decoder_embed_state,\n                self.decoder_embed,\n                self.dec_norm,\n                self.dec_norm_state,\n                self.downstream_head.dpt_self,\n                self.downstream_head.final_transform,\n                self.downstream_head.dpt_cross,\n                self.downstream_head.dpt_rgb,\n                self.downstream_head.pose_head,\n            ],\n            \"mhmr\": [\n                self.backbone,\n                self.mhmr_masked_img_token,\n                self.downstream_head.mlp_classif,\n                self.downstream_head.mlp_offset,\n            ],\n        }\n        if self.output_mode == \"naive\":\n            to_be_frozen[\"all\"] = [\n                self.patch_embed,\n                self.patch_embed_ray_map,\n                self.masked_ray_map_token,\n                self.enc_blocks,\n                self.enc_blocks_ray_map,\n                self.enc_norm,\n                self.enc_norm_ray_map,\n                self.dec_blocks,\n                self.dec_blocks_state,\n                self.pose_retriever,\n                self.pose_token,\n                self.register_tokens,\n                self.decoder_embed_state,\n                self.decoder_embed,\n                self.dec_norm,\n                self.dec_norm_state,\n                self.downstream_head.dpt_self,\n                self.downstream_head.final_transform,\n                self.downstream_head.dpt_cross,\n                self.downstream_head.dpt_rgb,\n                self.downstream_head.pose_head,\n                self.backbone,\n                self.mhmr_masked_smpl_token,\n                self.mhmr_masked_img_token,\n                self.transformer,\n                self.cross_queries_x,\n                self.cross_queries_y,\n                self.cross_values_x,\n                self.cross_values_y,\n                self.downstream_head.mlp_classif,\n                self.downstream_head.mlp_offset,\n                self.downstream_head.decpose, \n                self.downstream_head.decshape, \n                self.downstream_head.deccam, \n                self.downstream_head.decexpression,\n            ]\n\n        if freeze == \"encoder_and_decoder_and_head\":\n            fix_all_params(to_be_frozen[\"encoder_and_decoder_and_head\"]) # will not be updated\n            freeze_all_params(to_be_frozen[\"encoder\"]) # requires_grad = False\n            freeze_all_params(to_be_frozen[\"mhmr\"])\n        else:\n            freeze_all_params(to_be_frozen[freeze])\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head\"\"\"\n        return\n\n    def set_downstream_head(\n        self,\n        output_mode,\n        head_type,\n        landscape_only,\n        depth_mode,\n        conf_mode,\n        pose_mode,\n        depth_head,\n        rgb_head,\n        pose_conf_head,\n        pose_head,\n        msk_head,\n        patch_size,\n        img_size,","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._set_prediction_head","uri":"program://Human3R/function/src.dust3r.model._set_prediction_head#L637-L639","kind":"function","name":"_set_prediction_head","path":"src/dust3r/model.py","language":"python","start_line":637,"end_line":639,"context_start_line":617,"context_end_line":659,"code":"                self.transformer,\n                self.cross_queries_x,\n                self.cross_queries_y,\n                self.cross_values_x,\n                self.cross_values_y,\n                self.downstream_head.mlp_classif,\n                self.downstream_head.mlp_offset,\n                self.downstream_head.decpose, \n                self.downstream_head.decshape, \n                self.downstream_head.deccam, \n                self.downstream_head.decexpression,\n            ]\n\n        if freeze == \"encoder_and_decoder_and_head\":\n            fix_all_params(to_be_frozen[\"encoder_and_decoder_and_head\"]) # will not be updated\n            freeze_all_params(to_be_frozen[\"encoder\"]) # requires_grad = False\n            freeze_all_params(to_be_frozen[\"mhmr\"])\n        else:\n            freeze_all_params(to_be_frozen[freeze])\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head\"\"\"\n        return\n\n    def set_downstream_head(\n        self,\n        output_mode,\n        head_type,\n        landscape_only,\n        depth_mode,\n        conf_mode,\n        pose_mode,\n        depth_head,\n        rgb_head,\n        pose_conf_head,\n        pose_head,\n        msk_head,\n        patch_size,\n        img_size,\n        **kw,\n    ):\n        assert (\n            img_size[0] % patch_size == 0 and img_size[1] % patch_size == 0","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.set_downstream_head","uri":"program://Human3R/function/src.dust3r.model.set_downstream_head#L641-L679","kind":"function","name":"set_downstream_head","path":"src/dust3r/model.py","language":"python","start_line":641,"end_line":679,"context_start_line":621,"context_end_line":699,"code":"                self.cross_values_y,\n                self.downstream_head.mlp_classif,\n                self.downstream_head.mlp_offset,\n                self.downstream_head.decpose, \n                self.downstream_head.decshape, \n                self.downstream_head.deccam, \n                self.downstream_head.decexpression,\n            ]\n\n        if freeze == \"encoder_and_decoder_and_head\":\n            fix_all_params(to_be_frozen[\"encoder_and_decoder_and_head\"]) # will not be updated\n            freeze_all_params(to_be_frozen[\"encoder\"]) # requires_grad = False\n            freeze_all_params(to_be_frozen[\"mhmr\"])\n        else:\n            freeze_all_params(to_be_frozen[freeze])\n\n    def _set_prediction_head(self, *args, **kwargs):\n        \"\"\"No prediction head\"\"\"\n        return\n\n    def set_downstream_head(\n        self,\n        output_mode,\n        head_type,\n        landscape_only,\n        depth_mode,\n        conf_mode,\n        pose_mode,\n        depth_head,\n        rgb_head,\n        pose_conf_head,\n        pose_head,\n        msk_head,\n        patch_size,\n        img_size,\n        **kw,\n    ):\n        assert (\n            img_size[0] % patch_size == 0 and img_size[1] % patch_size == 0\n        ), f\"{img_size=} must be multiple of {patch_size=}\"\n        self.output_mode = output_mode\n        self.head_type = head_type\n        self.depth_mode = depth_mode\n        self.conf_mode = conf_mode\n        self.pose_mode = pose_mode\n        self.downstream_head = head_factory(\n            head_type,\n            output_mode,\n            self,\n            has_conf=bool(conf_mode),\n            has_depth=bool(depth_head),\n            has_rgb=bool(rgb_head),\n            has_pose_conf=bool(pose_conf_head),\n            has_pose=bool(pose_head),\n            has_msk=bool(msk_head),\n        )\n        self.head = transpose_to_landscape(\n            self.downstream_head, activate=landscape_only\n        )\n\n    def _encode_image(self, image, true_shape):\n        x, pos = self.patch_embed(image, true_shape=true_shape)\n        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)\n        x = self.enc_norm(x)\n        return [x], pos, None\n\n    def _encode_ray_map(self, ray_map, true_shape):\n        x, pos = self.patch_embed_ray_map(ray_map, true_shape=true_shape)\n        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks_ray_map:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._encode_image","uri":"program://Human3R/function/src.dust3r.model._encode_image#L681-L690","kind":"function","name":"_encode_image","path":"src/dust3r/model.py","language":"python","start_line":681,"end_line":690,"context_start_line":661,"context_end_line":710,"code":"        self.output_mode = output_mode\n        self.head_type = head_type\n        self.depth_mode = depth_mode\n        self.conf_mode = conf_mode\n        self.pose_mode = pose_mode\n        self.downstream_head = head_factory(\n            head_type,\n            output_mode,\n            self,\n            has_conf=bool(conf_mode),\n            has_depth=bool(depth_head),\n            has_rgb=bool(rgb_head),\n            has_pose_conf=bool(pose_conf_head),\n            has_pose=bool(pose_head),\n            has_msk=bool(msk_head),\n        )\n        self.head = transpose_to_landscape(\n            self.downstream_head, activate=landscape_only\n        )\n\n    def _encode_image(self, image, true_shape):\n        x, pos = self.patch_embed(image, true_shape=true_shape)\n        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)\n        x = self.enc_norm(x)\n        return [x], pos, None\n\n    def _encode_ray_map(self, ray_map, true_shape):\n        x, pos = self.patch_embed_ray_map(ray_map, true_shape=true_shape)\n        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks_ray_map:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)\n        x = self.enc_norm_ray_map(x)\n        return [x], pos, None\n\n    def _encode_state(self, image_tokens, image_pos):\n        batch_size = image_tokens.shape[0]\n        state_feat = self.register_tokens(\n            torch.arange(self.state_size, device=image_pos.device)\n        )\n        if self.state_pe == \"1d\":\n            state_pos = (\n                torch.tensor(","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._encode_ray_map","uri":"program://Human3R/function/src.dust3r.model._encode_ray_map#L692-L701","kind":"function","name":"_encode_ray_map","path":"src/dust3r/model.py","language":"python","start_line":692,"end_line":701,"context_start_line":672,"context_end_line":721,"code":"            has_rgb=bool(rgb_head),\n            has_pose_conf=bool(pose_conf_head),\n            has_pose=bool(pose_head),\n            has_msk=bool(msk_head),\n        )\n        self.head = transpose_to_landscape(\n            self.downstream_head, activate=landscape_only\n        )\n\n    def _encode_image(self, image, true_shape):\n        x, pos = self.patch_embed(image, true_shape=true_shape)\n        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)\n        x = self.enc_norm(x)\n        return [x], pos, None\n\n    def _encode_ray_map(self, ray_map, true_shape):\n        x, pos = self.patch_embed_ray_map(ray_map, true_shape=true_shape)\n        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks_ray_map:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)\n        x = self.enc_norm_ray_map(x)\n        return [x], pos, None\n\n    def _encode_state(self, image_tokens, image_pos):\n        batch_size = image_tokens.shape[0]\n        state_feat = self.register_tokens(\n            torch.arange(self.state_size, device=image_pos.device)\n        )\n        if self.state_pe == \"1d\":\n            state_pos = (\n                torch.tensor(\n                    [[i, i] for i in range(self.state_size)],\n                    dtype=image_pos.dtype,\n                    device=image_pos.device,\n                )[None]\n                .expand(batch_size, -1, -1)\n                .contiguous()\n            )  # .long()\n        elif self.state_pe == \"2d\":\n            width = int(self.state_size**0.5)\n            width = width + 1 if width % 2 == 1 else width\n            state_pos = (","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._encode_state","uri":"program://Human3R/function/src.dust3r.model._encode_state#L703-L733","kind":"function","name":"_encode_state","path":"src/dust3r/model.py","language":"python","start_line":703,"end_line":733,"context_start_line":683,"context_end_line":753,"code":"        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)\n        x = self.enc_norm(x)\n        return [x], pos, None\n\n    def _encode_ray_map(self, ray_map, true_shape):\n        x, pos = self.patch_embed_ray_map(ray_map, true_shape=true_shape)\n        assert self.enc_pos_embed is None\n        for blk in self.enc_blocks_ray_map:\n            if self.gradient_checkpointing and self.training:\n                x = checkpoint(blk, x, pos, use_reentrant=False)\n            else:\n                x = blk(x, pos)\n        x = self.enc_norm_ray_map(x)\n        return [x], pos, None\n\n    def _encode_state(self, image_tokens, image_pos):\n        batch_size = image_tokens.shape[0]\n        state_feat = self.register_tokens(\n            torch.arange(self.state_size, device=image_pos.device)\n        )\n        if self.state_pe == \"1d\":\n            state_pos = (\n                torch.tensor(\n                    [[i, i] for i in range(self.state_size)],\n                    dtype=image_pos.dtype,\n                    device=image_pos.device,\n                )[None]\n                .expand(batch_size, -1, -1)\n                .contiguous()\n            )  # .long()\n        elif self.state_pe == \"2d\":\n            width = int(self.state_size**0.5)\n            width = width + 1 if width % 2 == 1 else width\n            state_pos = (\n                torch.tensor(\n                    [[i // width, i % width] for i in range(self.state_size)],\n                    dtype=image_pos.dtype,\n                    device=image_pos.device,\n                )[None]\n                .expand(batch_size, -1, -1)\n                .contiguous()\n            )\n        elif self.state_pe == \"none\":\n            state_pos = None\n        state_feat = state_feat[None].expand(batch_size, -1, -1)\n        return state_feat, state_pos, None\n\n    def _encode_views_mhmr(self, views, img_mask=None, ray_mask=None):\n        device = views[0][\"img\"].device\n        batch_size = views[0][\"img\"].shape[0]\n        given = True\n        if img_mask is None and ray_mask is None:\n            given = False\n        if not given:\n            img_mask = torch.stack(\n                [view[\"img_mask\"] for view in views], dim=0\n            )  # Shape: (num_views, batch_size)\n            ray_mask = torch.stack(\n                [view[\"ray_mask\"] for view in views], dim=0\n            )  # Shape: (num_views, batch_size)\n        imgs = torch.stack(\n            [view[\"img\"] for view in views], dim=0\n        )  # Shape: (num_views, batch_size, C, H, W)\n        ray_maps = torch.stack(\n            [view[\"ray_map\"] for view in views], dim=0\n        )  # Shape: (num_views, batch_size, H, W, C)","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._encode_views_mhmr","uri":"program://Human3R/function/src.dust3r.model._encode_views_mhmr#L735-L851","kind":"function","name":"_encode_views_mhmr","path":"src/dust3r/model.py","language":"python","start_line":735,"end_line":851,"context_start_line":715,"context_end_line":871,"code":"                .expand(batch_size, -1, -1)\n                .contiguous()\n            )  # .long()\n        elif self.state_pe == \"2d\":\n            width = int(self.state_size**0.5)\n            width = width + 1 if width % 2 == 1 else width\n            state_pos = (\n                torch.tensor(\n                    [[i // width, i % width] for i in range(self.state_size)],\n                    dtype=image_pos.dtype,\n                    device=image_pos.device,\n                )[None]\n                .expand(batch_size, -1, -1)\n                .contiguous()\n            )\n        elif self.state_pe == \"none\":\n            state_pos = None\n        state_feat = state_feat[None].expand(batch_size, -1, -1)\n        return state_feat, state_pos, None\n\n    def _encode_views_mhmr(self, views, img_mask=None, ray_mask=None):\n        device = views[0][\"img\"].device\n        batch_size = views[0][\"img\"].shape[0]\n        given = True\n        if img_mask is None and ray_mask is None:\n            given = False\n        if not given:\n            img_mask = torch.stack(\n                [view[\"img_mask\"] for view in views], dim=0\n            )  # Shape: (num_views, batch_size)\n            ray_mask = torch.stack(\n                [view[\"ray_mask\"] for view in views], dim=0\n            )  # Shape: (num_views, batch_size)\n        imgs = torch.stack(\n            [view[\"img\"] for view in views], dim=0\n        )  # Shape: (num_views, batch_size, C, H, W)\n        ray_maps = torch.stack(\n            [view[\"ray_map\"] for view in views], dim=0\n        )  # Shape: (num_views, batch_size, H, W, C)\n        shapes = []\n        for view in views:\n            if \"true_shape\" in view:\n                shapes.append(view[\"true_shape\"])\n            else:\n                shape = torch.tensor(view[\"img\"].shape[-2:], device=device)\n                shapes.append(shape.unsqueeze(0).repeat(batch_size, 1))\n        shapes = torch.stack(shapes, dim=0).to(\n            imgs.device\n        )  # Shape: (num_views, batch_size, 2)\n        imgs = imgs.view(\n            -1, *imgs.shape[2:]\n        )  # Shape: (num_views * batch_size, C, H, W)\n        ray_maps = ray_maps.view(\n            -1, *ray_maps.shape[2:]\n        )  # Shape: (num_views * batch_size, H, W, C)\n        shapes = shapes.view(-1, 2)  # Shape: (num_views * batch_size, 2)\n        img_masks_flat = img_mask.view(-1)  # Shape: (num_views * batch_size)\n        ray_masks_flat = ray_mask.view(-1)\n        selected_imgs = imgs[img_masks_flat]\n        selected_shapes = shapes[img_masks_flat]\n        if selected_imgs.size(0) > 0:\n            img_out, img_pos, _ = self._encode_image(selected_imgs, selected_shapes)\n        else:\n            raise NotImplementedError\n        full_out = [\n            torch.zeros(\n                len(views) * batch_size, *img_out[0].shape[1:], device=img_out[0].device\n            )\n            for _ in range(len(img_out))\n        ]\n        full_pos = torch.zeros(\n            len(views) * batch_size,\n            *img_pos.shape[1:],\n            device=img_pos.device,\n            dtype=img_pos.dtype,\n        )\n        for i in range(len(img_out)):\n            full_out[i][img_masks_flat] += img_out[i]\n            full_out[i][~img_masks_flat] += self.masked_img_token\n        full_pos[img_masks_flat] += img_pos\n\n        # MHMR\n        imgs_mhmr = torch.stack(\n            [view[\"img_mhmr\"] for view in views], dim=0\n        )  # Shape: (num_views, batch_size, C, H, W)\n        imgs_mhmr = imgs_mhmr.view(\n            -1, *imgs_mhmr.shape[2:]\n        )  # Shape: (num_views * batch_size, C, H, W)\n        selected_imgs_mhmr = imgs_mhmr[img_masks_flat]\n        if selected_imgs_mhmr.size(0) > 0:\n            mean = torch.tensor([0.485, 0.456, 0.406], device=device)[None, :, None, None]\n            std = torch.tensor([0.229, 0.224, 0.225], device=device)[None, :, None, None]\n            selected_imgs_mhmr = (selected_imgs_mhmr * 0.5 + 0.5 - mean) / std\n            mhmr_img_out = [self.backbone(selected_imgs_mhmr)] # image[bs, 3, h, w] -> image feature [bs, h_nb_patches * w_nb_patches, D]\n        else:\n            raise NotImplementedError\n        \n        mhmr_full_out = [\n            torch.zeros(\n                len(views) * batch_size, *mhmr_img_out[0].shape[1:], device=mhmr_img_out[0].device\n            )\n            for _ in range(len(mhmr_img_out))\n        ]\n        for i in range(len(mhmr_img_out)):\n            mhmr_full_out[i][img_masks_flat] += mhmr_img_out[i]\n            mhmr_full_out[i][~img_masks_flat] += self.mhmr_masked_img_token\n\n        ray_maps = ray_maps.permute(0, 3, 1, 2)  # Change shape to (N, C, H, W)\n        selected_ray_maps = ray_maps[ray_masks_flat]\n        selected_shapes_ray = shapes[ray_masks_flat]\n        if selected_ray_maps.size(0) > 0:\n            ray_out, ray_pos, _ = self._encode_ray_map(\n                selected_ray_maps, selected_shapes_ray\n            )\n            assert len(ray_out) == len(full_out), f\"{len(ray_out)}, {len(full_out)}\"\n            for i in range(len(ray_out)):\n                full_out[i][ray_masks_flat] += ray_out[i]\n                full_out[i][~ray_masks_flat] += self.masked_ray_map_token\n            full_pos[ray_masks_flat] += (\n                ray_pos * (~img_masks_flat[ray_masks_flat][:, None, None]).long()\n            )\n        else:\n            raymaps = torch.zeros(\n                1, 6, imgs[0].shape[-2], imgs[0].shape[-1], device=img_out[0].device\n            )\n            ray_mask_flat = torch.zeros_like(img_masks_flat)\n            ray_mask_flat[:1] = True\n            ray_out, ray_pos, _ = self._encode_ray_map(raymaps, shapes[ray_mask_flat])\n            for i in range(len(ray_out)):\n                full_out[i][ray_mask_flat] += ray_out[i] * 0.0\n                full_out[i][~ray_mask_flat] += self.masked_ray_map_token * 0.0\n        return (\n            shapes.chunk(len(views), dim=0),\n            [out.chunk(len(views), dim=0) for out in full_out],\n            full_pos.chunk(len(views), dim=0),\n            [mhmr_out.chunk(len(views), dim=0) for mhmr_out in mhmr_full_out],\n        )\n\n    def _decoder(self, f_state, pos_state, f_img, pos_img, f_pose, pos_pose, f_smpl, pos_smpl, use_ttt3r=False):\n        final_output = [(f_state, f_img)]  # before projection\n        assert f_state.shape[-1] == self.dec_embed_dim\n        f_img = self.decoder_embed(f_img)\n        if self.pose_head_flag:\n            assert f_pose is not None and pos_pose is not None\n            if f_smpl is not None:\n                f_img = torch.cat([f_pose, f_img, f_smpl], dim=1)\n                pos_img = torch.cat([pos_pose, pos_img, pos_smpl], dim=1)\n            else:\n                f_img = torch.cat([f_pose, f_img], dim=1) # used for naive CUT3R+MHMR\n                pos_img = torch.cat([pos_pose, pos_img], dim=1) # used for naive CUT3R+MHMR\n        final_output.append((f_state, f_img))\n        cross_attn_states = []\n        for blk_state, blk_img in zip(self.dec_blocks_state, self.dec_blocks):\n            if (\n                self.gradient_checkpointing\n                and self.training\n                and torch.is_grad_enabled()","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._decoder","uri":"program://Human3R/function/src.dust3r.model._decoder#L853-L902","kind":"function","name":"_decoder","path":"src/dust3r/model.py","language":"python","start_line":853,"end_line":902,"context_start_line":833,"context_end_line":922,"code":"            full_pos[ray_masks_flat] += (\n                ray_pos * (~img_masks_flat[ray_masks_flat][:, None, None]).long()\n            )\n        else:\n            raymaps = torch.zeros(\n                1, 6, imgs[0].shape[-2], imgs[0].shape[-1], device=img_out[0].device\n            )\n            ray_mask_flat = torch.zeros_like(img_masks_flat)\n            ray_mask_flat[:1] = True\n            ray_out, ray_pos, _ = self._encode_ray_map(raymaps, shapes[ray_mask_flat])\n            for i in range(len(ray_out)):\n                full_out[i][ray_mask_flat] += ray_out[i] * 0.0\n                full_out[i][~ray_mask_flat] += self.masked_ray_map_token * 0.0\n        return (\n            shapes.chunk(len(views), dim=0),\n            [out.chunk(len(views), dim=0) for out in full_out],\n            full_pos.chunk(len(views), dim=0),\n            [mhmr_out.chunk(len(views), dim=0) for mhmr_out in mhmr_full_out],\n        )\n\n    def _decoder(self, f_state, pos_state, f_img, pos_img, f_pose, pos_pose, f_smpl, pos_smpl, use_ttt3r=False):\n        final_output = [(f_state, f_img)]  # before projection\n        assert f_state.shape[-1] == self.dec_embed_dim\n        f_img = self.decoder_embed(f_img)\n        if self.pose_head_flag:\n            assert f_pose is not None and pos_pose is not None\n            if f_smpl is not None:\n                f_img = torch.cat([f_pose, f_img, f_smpl], dim=1)\n                pos_img = torch.cat([pos_pose, pos_img, pos_smpl], dim=1)\n            else:\n                f_img = torch.cat([f_pose, f_img], dim=1) # used for naive CUT3R+MHMR\n                pos_img = torch.cat([pos_pose, pos_img], dim=1) # used for naive CUT3R+MHMR\n        final_output.append((f_state, f_img))\n        cross_attn_states = []\n        for blk_state, blk_img in zip(self.dec_blocks_state, self.dec_blocks):\n            if (\n                self.gradient_checkpointing\n                and self.training\n                and torch.is_grad_enabled()\n            ):\n                f_state, _, cross_attn_state = checkpoint(\n                    blk_state,\n                    *final_output[-1][::+1],\n                    pos_state,\n                    pos_img,\n                    use_ttt3r=use_ttt3r,\n                    use_reentrant=not self.fixed_input_length,\n                )\n                f_img, _, _ = checkpoint(\n                    blk_img,\n                    *final_output[-1][::-1],\n                    pos_img,\n                    pos_state,\n                    use_ttt3r=False,\n                    use_reentrant=not self.fixed_input_length,\n                )\n            else:\n                f_state, _, cross_attn_state = blk_state(*final_output[-1][::+1], pos_state, pos_img, use_ttt3r=use_ttt3r)\n                f_img, _, _ = blk_img(*final_output[-1][::-1], pos_img, pos_state, use_ttt3r=False)\n\n            final_output.append((f_state, f_img))\n            cross_attn_states.append(cross_attn_state)\n\n        del final_output[1]  # duplicate with final_output[0]\n        final_output[-1] = (\n            self.dec_norm_state(final_output[-1][0]),\n            self.dec_norm(final_output[-1][1]),\n        )\n\n        return zip(*final_output), cross_attn_states\n\n    def _downstream_head(self, decout, img_shape, **kwargs):\n        B, S, D = decout[-1].shape\n        head = getattr(self, f\"head\")\n        return head(decout, img_shape, **kwargs)\n\n    def _init_state(self, image_tokens, image_pos):\n        \"\"\"\n        Current Version: input the first frame img feature and pose to initialize the state feature and pose\n        \"\"\"\n        state_feat, state_pos, _ = self._encode_state(image_tokens, image_pos)\n        state_feat = self.decoder_embed_state(state_feat)\n        return state_feat, state_pos\n\n    def _recurrent_rollout(\n        self,\n        state_feat,\n        state_pos,\n        current_feat,\n        current_pos,","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._downstream_head","uri":"program://Human3R/function/src.dust3r.model._downstream_head#L904-L907","kind":"function","name":"_downstream_head","path":"src/dust3r/model.py","language":"python","start_line":904,"end_line":907,"context_start_line":884,"context_end_line":927,"code":"                    pos_img,\n                    pos_state,\n                    use_ttt3r=False,\n                    use_reentrant=not self.fixed_input_length,\n                )\n            else:\n                f_state, _, cross_attn_state = blk_state(*final_output[-1][::+1], pos_state, pos_img, use_ttt3r=use_ttt3r)\n                f_img, _, _ = blk_img(*final_output[-1][::-1], pos_img, pos_state, use_ttt3r=False)\n\n            final_output.append((f_state, f_img))\n            cross_attn_states.append(cross_attn_state)\n\n        del final_output[1]  # duplicate with final_output[0]\n        final_output[-1] = (\n            self.dec_norm_state(final_output[-1][0]),\n            self.dec_norm(final_output[-1][1]),\n        )\n\n        return zip(*final_output), cross_attn_states\n\n    def _downstream_head(self, decout, img_shape, **kwargs):\n        B, S, D = decout[-1].shape\n        head = getattr(self, f\"head\")\n        return head(decout, img_shape, **kwargs)\n\n    def _init_state(self, image_tokens, image_pos):\n        \"\"\"\n        Current Version: input the first frame img feature and pose to initialize the state feature and pose\n        \"\"\"\n        state_feat, state_pos, _ = self._encode_state(image_tokens, image_pos)\n        state_feat = self.decoder_embed_state(state_feat)\n        return state_feat, state_pos\n\n    def _recurrent_rollout(\n        self,\n        state_feat,\n        state_pos,\n        current_feat,\n        current_pos,\n        pose_feat,\n        pose_pos,\n        smpl_feat,\n        smpl_pos,\n        init_state_feat,","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._init_state","uri":"program://Human3R/function/src.dust3r.model._init_state#L909-L915","kind":"function","name":"_init_state","path":"src/dust3r/model.py","language":"python","start_line":909,"end_line":915,"context_start_line":889,"context_end_line":935,"code":"            else:\n                f_state, _, cross_attn_state = blk_state(*final_output[-1][::+1], pos_state, pos_img, use_ttt3r=use_ttt3r)\n                f_img, _, _ = blk_img(*final_output[-1][::-1], pos_img, pos_state, use_ttt3r=False)\n\n            final_output.append((f_state, f_img))\n            cross_attn_states.append(cross_attn_state)\n\n        del final_output[1]  # duplicate with final_output[0]\n        final_output[-1] = (\n            self.dec_norm_state(final_output[-1][0]),\n            self.dec_norm(final_output[-1][1]),\n        )\n\n        return zip(*final_output), cross_attn_states\n\n    def _downstream_head(self, decout, img_shape, **kwargs):\n        B, S, D = decout[-1].shape\n        head = getattr(self, f\"head\")\n        return head(decout, img_shape, **kwargs)\n\n    def _init_state(self, image_tokens, image_pos):\n        \"\"\"\n        Current Version: input the first frame img feature and pose to initialize the state feature and pose\n        \"\"\"\n        state_feat, state_pos, _ = self._encode_state(image_tokens, image_pos)\n        state_feat = self.decoder_embed_state(state_feat)\n        return state_feat, state_pos\n\n    def _recurrent_rollout(\n        self,\n        state_feat,\n        state_pos,\n        current_feat,\n        current_pos,\n        pose_feat,\n        pose_pos,\n        smpl_feat,\n        smpl_pos,\n        init_state_feat,\n        img_mask=None,\n        reset_mask=None,\n        update=None,\n        use_ttt3r=False,\n    ):\n        (new_state_feat, dec), cross_attn_states = self._decoder(\n            state_feat, state_pos, current_feat, current_pos, pose_feat, pose_pos, smpl_feat, smpl_pos, use_ttt3r\n        )","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._recurrent_rollout","uri":"program://Human3R/function/src.dust3r.model._recurrent_rollout#L917-L937","kind":"function","name":"_recurrent_rollout","path":"src/dust3r/model.py","language":"python","start_line":917,"end_line":937,"context_start_line":897,"context_end_line":957,"code":"        final_output[-1] = (\n            self.dec_norm_state(final_output[-1][0]),\n            self.dec_norm(final_output[-1][1]),\n        )\n\n        return zip(*final_output), cross_attn_states\n\n    def _downstream_head(self, decout, img_shape, **kwargs):\n        B, S, D = decout[-1].shape\n        head = getattr(self, f\"head\")\n        return head(decout, img_shape, **kwargs)\n\n    def _init_state(self, image_tokens, image_pos):\n        \"\"\"\n        Current Version: input the first frame img feature and pose to initialize the state feature and pose\n        \"\"\"\n        state_feat, state_pos, _ = self._encode_state(image_tokens, image_pos)\n        state_feat = self.decoder_embed_state(state_feat)\n        return state_feat, state_pos\n\n    def _recurrent_rollout(\n        self,\n        state_feat,\n        state_pos,\n        current_feat,\n        current_pos,\n        pose_feat,\n        pose_pos,\n        smpl_feat,\n        smpl_pos,\n        init_state_feat,\n        img_mask=None,\n        reset_mask=None,\n        update=None,\n        use_ttt3r=False,\n    ):\n        (new_state_feat, dec), cross_attn_states = self._decoder(\n            state_feat, state_pos, current_feat, current_pos, pose_feat, pose_pos, smpl_feat, smpl_pos, use_ttt3r\n        )\n        new_state_feat = new_state_feat[-1]\n        return new_state_feat, dec, cross_attn_states\n\n    def _get_img_level_feat(self, feat):\n        return torch.mean(feat, dim=1, keepdim=True)\n\n    def embedd_camera(self, K, n_patch):\n        \"\"\" Embed viewing directions using fourrier encoding.\"\"\"\n        bs = K.shape[0]\n        _h, _w = n_patch\n        points = torch.stack([\n            torch.arange(0,_h,1).reshape(-1,1).repeat(1,_w), \n            torch.arange(0,_w,1).reshape(1,-1).repeat(_h,1)],\n            -1).to(K.device).float() # [h,w,2]\n        points = points * self.bb_patch_size + self.bb_patch_size // 2 # move to pixel space - we give the pixel center of each token\n        points = points.reshape(1,-1,2).repeat(bs,1,1) # (bs, hw, 2): 2D points\n        distance = torch.ones(bs,points.shape[1],1).to(K.device) # (bs, N, 1): distance in the 3D world\n        rays = inverse_perspective_projection(points, K, distance) # (bs, N, 3)\n        rays_embeddings = self.fourier_camera(pos=rays)\n\n        # Repeat for each element of the batch\n        z_K = rays_embeddings.reshape(bs,_h,_w,self.camera_embed_dim) # [bs,h,w,99]","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._get_img_level_feat","uri":"program://Human3R/function/src.dust3r.model._get_img_level_feat#L939-L940","kind":"function","name":"_get_img_level_feat","path":"src/dust3r/model.py","language":"python","start_line":939,"end_line":940,"context_start_line":919,"context_end_line":960,"code":"        state_feat,\n        state_pos,\n        current_feat,\n        current_pos,\n        pose_feat,\n        pose_pos,\n        smpl_feat,\n        smpl_pos,\n        init_state_feat,\n        img_mask=None,\n        reset_mask=None,\n        update=None,\n        use_ttt3r=False,\n    ):\n        (new_state_feat, dec), cross_attn_states = self._decoder(\n            state_feat, state_pos, current_feat, current_pos, pose_feat, pose_pos, smpl_feat, smpl_pos, use_ttt3r\n        )\n        new_state_feat = new_state_feat[-1]\n        return new_state_feat, dec, cross_attn_states\n\n    def _get_img_level_feat(self, feat):\n        return torch.mean(feat, dim=1, keepdim=True)\n\n    def embedd_camera(self, K, n_patch):\n        \"\"\" Embed viewing directions using fourrier encoding.\"\"\"\n        bs = K.shape[0]\n        _h, _w = n_patch\n        points = torch.stack([\n            torch.arange(0,_h,1).reshape(-1,1).repeat(1,_w), \n            torch.arange(0,_w,1).reshape(1,-1).repeat(_h,1)],\n            -1).to(K.device).float() # [h,w,2]\n        points = points * self.bb_patch_size + self.bb_patch_size // 2 # move to pixel space - we give the pixel center of each token\n        points = points.reshape(1,-1,2).repeat(bs,1,1) # (bs, hw, 2): 2D points\n        distance = torch.ones(bs,points.shape[1],1).to(K.device) # (bs, N, 1): distance in the 3D world\n        rays = inverse_perspective_projection(points, K, distance) # (bs, N, 3)\n        rays_embeddings = self.fourier_camera(pos=rays)\n\n        # Repeat for each element of the batch\n        z_K = rays_embeddings.reshape(bs,_h,_w,self.camera_embed_dim) # [bs,h,w,99]\n        return z_K \n\n    def smpl_tokenizer_mhmr(self, feat, pos, views, inference=False):","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.embedd_camera","uri":"program://Human3R/function/src.dust3r.model.embedd_camera#L942-L958","kind":"function","name":"embedd_camera","path":"src/dust3r/model.py","language":"python","start_line":942,"end_line":958,"context_start_line":922,"context_end_line":978,"code":"        current_pos,\n        pose_feat,\n        pose_pos,\n        smpl_feat,\n        smpl_pos,\n        init_state_feat,\n        img_mask=None,\n        reset_mask=None,\n        update=None,\n        use_ttt3r=False,\n    ):\n        (new_state_feat, dec), cross_attn_states = self._decoder(\n            state_feat, state_pos, current_feat, current_pos, pose_feat, pose_pos, smpl_feat, smpl_pos, use_ttt3r\n        )\n        new_state_feat = new_state_feat[-1]\n        return new_state_feat, dec, cross_attn_states\n\n    def _get_img_level_feat(self, feat):\n        return torch.mean(feat, dim=1, keepdim=True)\n\n    def embedd_camera(self, K, n_patch):\n        \"\"\" Embed viewing directions using fourrier encoding.\"\"\"\n        bs = K.shape[0]\n        _h, _w = n_patch\n        points = torch.stack([\n            torch.arange(0,_h,1).reshape(-1,1).repeat(1,_w), \n            torch.arange(0,_w,1).reshape(1,-1).repeat(_h,1)],\n            -1).to(K.device).float() # [h,w,2]\n        points = points * self.bb_patch_size + self.bb_patch_size // 2 # move to pixel space - we give the pixel center of each token\n        points = points.reshape(1,-1,2).repeat(bs,1,1) # (bs, hw, 2): 2D points\n        distance = torch.ones(bs,points.shape[1],1).to(K.device) # (bs, N, 1): distance in the 3D world\n        rays = inverse_perspective_projection(points, K, distance) # (bs, N, 3)\n        rays_embeddings = self.fourier_camera(pos=rays)\n\n        # Repeat for each element of the batch\n        z_K = rays_embeddings.reshape(bs,_h,_w,self.camera_embed_dim) # [bs,h,w,99]\n        return z_K \n\n    def smpl_tokenizer_mhmr(self, feat, pos, views, inference=False):\n        feat = torch.stack([f.detach() for f in feat], dim=0) #(num_view, bs, 576, 1024)\n        num_view, batch_size = feat.shape[:2]\n\n        feat = feat.view(-1, *feat.shape[2:]) #(num_view * bs, 576, 1024)\n        scores = self.downstream_head.detect_mhmr(feat) #(num_view * bs, 576, 1)\n        if self.msk_head_flag:\n            msks = self.downstream_head.segment(feat.detach()) #(num_view * bs, 576, 14*14)\n\n        # Restore Height and Width dimensions.\n        n_patch = self.bb_token_res # H,W\n        scores = rearrange(scores, \"b (nh nw) c -> b c nh nw\", nh=n_patch, nw=n_patch)\n        feat = rearrange(feat, \"b (nh nw) c -> b nh nw c\", nh=n_patch, nw=n_patch) # head token extraction: (num_view * bs, h, w, 1024)\n        if self.msk_head_flag:\n            msks = rearrange(msks, \"b (nh nw) c -> b c nh nw\", nh=n_patch, nw=n_patch)\n            msks = F.pixel_shuffle(msks, self.bb_patch_size)  # (num_view * bs, 1, h, w)\n\n        if self.output_mode == \"naive\":\n            # use GT K","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.smpl_tokenizer_mhmr","uri":"program://Human3R/function/src.dust3r.model.smpl_tokenizer_mhmr#L960-L1099","kind":"function","name":"smpl_tokenizer_mhmr","path":"src/dust3r/model.py","language":"python","start_line":960,"end_line":1099,"context_start_line":940,"context_end_line":1119,"code":"        return torch.mean(feat, dim=1, keepdim=True)\n\n    def embedd_camera(self, K, n_patch):\n        \"\"\" Embed viewing directions using fourrier encoding.\"\"\"\n        bs = K.shape[0]\n        _h, _w = n_patch\n        points = torch.stack([\n            torch.arange(0,_h,1).reshape(-1,1).repeat(1,_w), \n            torch.arange(0,_w,1).reshape(1,-1).repeat(_h,1)],\n            -1).to(K.device).float() # [h,w,2]\n        points = points * self.bb_patch_size + self.bb_patch_size // 2 # move to pixel space - we give the pixel center of each token\n        points = points.reshape(1,-1,2).repeat(bs,1,1) # (bs, hw, 2): 2D points\n        distance = torch.ones(bs,points.shape[1],1).to(K.device) # (bs, N, 1): distance in the 3D world\n        rays = inverse_perspective_projection(points, K, distance) # (bs, N, 3)\n        rays_embeddings = self.fourier_camera(pos=rays)\n\n        # Repeat for each element of the batch\n        z_K = rays_embeddings.reshape(bs,_h,_w,self.camera_embed_dim) # [bs,h,w,99]\n        return z_K \n\n    def smpl_tokenizer_mhmr(self, feat, pos, views, inference=False):\n        feat = torch.stack([f.detach() for f in feat], dim=0) #(num_view, bs, 576, 1024)\n        num_view, batch_size = feat.shape[:2]\n\n        feat = feat.view(-1, *feat.shape[2:]) #(num_view * bs, 576, 1024)\n        scores = self.downstream_head.detect_mhmr(feat) #(num_view * bs, 576, 1)\n        if self.msk_head_flag:\n            msks = self.downstream_head.segment(feat.detach()) #(num_view * bs, 576, 14*14)\n\n        # Restore Height and Width dimensions.\n        n_patch = self.bb_token_res # H,W\n        scores = rearrange(scores, \"b (nh nw) c -> b c nh nw\", nh=n_patch, nw=n_patch)\n        feat = rearrange(feat, \"b (nh nw) c -> b nh nw c\", nh=n_patch, nw=n_patch) # head token extraction: (num_view * bs, h, w, 1024)\n        if self.msk_head_flag:\n            msks = rearrange(msks, \"b (nh nw) c -> b c nh nw\", nh=n_patch, nw=n_patch)\n            msks = F.pixel_shuffle(msks, self.bb_patch_size)  # (num_view * bs, 1, h, w)\n\n        if self.output_mode == \"naive\":\n            # use GT K\n            K = torch.stack([v['K_mhmr'] for v in views], dim=0)\n            K = K.view(-1, *K.shape[2:])\n            # # use pseudo K\n            # K = get_camera_parameters(self.mhmr_img_res, device=feat.device)\n            # K = K.expand(feat.shape[0], -1, -1)\n            feat_K = self.embedd_camera(K, [n_patch, n_patch]) # Embed viewing directions. [num_view * bs,h,w,99]\n        \n        if inference:\n            scores = nms(scores, kernel=3) # (num_view * bs, 1, h, w)\n            _scores = scores.permute((0, 2, 3, 1)) # (num_view * bs, h, w, 1)\n            # Binary decision (keep confident detections)\n            idx = apply_threshold(0.3, _scores)\n            img_id, h_id, w_id = idx[0], idx[1], idx[2]\n        else:\n            smpl_mask = torch.stack([view[\"smpl_mask\"] for view in views], dim=0)\n            smpl_mask = smpl_mask.view(-1, *smpl_mask.shape[2:])\n            max_humans = smpl_mask.shape[1]\n            smpl_uv = torch.stack([view[\"smpl_uv\"] for view in views], dim=0)\n            smpl_uv = smpl_uv.view(-1, *smpl_uv.shape[2:])[smpl_mask]\n            img_id = torch.where(smpl_mask)[0]\n            h_id, w_id = smpl_uv.T\n\n        # Scores  \n        scores = scores.permute((0, 2, 3, 1)) # (num_view * bs, h, w, 1)\n        if self.msk_head_flag:\n            msks = msks.permute((0, 2, 3, 1)) # (num_view * bs, h, w, 1)\n        \n        # Head token and offset\n        feat_central = feat[img_id, h_id, w_id] # (nvh, 1024)\n        offset = self.downstream_head.mlp_offset(feat_central)# [nhv,2]\n        # Distance for estimating the 3D location in 3D space\n        loc = torch.stack([w_id, h_id]).permute(1,0) # x,y\n        loc = (loc + 0.5 + offset) * self.bb_patch_size # Moving to higher res the location of the pelvis\n\n        if self.output_mode == \"naive\":\n            # Concat with camera embedding\n            feat_K_central = feat_K[img_id, h_id, w_id] # (nvh, 99)\n            feat_central = torch.cat([feat_central, feat_K_central], 1) # feature + camera embedding for heads only to query tokens [nhv, 1123]\n            feat_all = torch.cat([feat, feat_K], -1).permute(0,3,1,2) # feature + camera embedding for full image for the cross-attention only. [bs,1123,nh,nw]\n\n            # Get learned embeddings for queries, at positions with detected people.\n            queries_xy = self.cross_queries_x[h_id] + self.cross_queries_y[w_id]\n            # Add the embedding to the central features.\n            feat_central = feat_central + queries_xy # [nhv, 1123]\n            # Inject leared embeddings for key/values at detected locations. \n            values_xy = self.cross_values_x[h_id] + self.cross_values_y[w_id]\n            feat_all[img_id, :, h_id, w_id] += values_xy  # [bs, 1123, nh, nw]\n            feat_all = rearrange(feat_all, \"b c h w -> b (h w) c\") # (num_view * bs, nh*nw, 1024)\n\n        if inference:\n            head_token = feat_central\n            head_loc = loc\n            expand = lambda x: x.expand(*feat_central.shape[:-1] , -1)\n        else:\n            # concat with mask token\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, feat_central.shape[1], \n                device=feat_central.device\n            )\n            full_out[smpl_mask] += feat_central\n            full_out[~smpl_mask] += self.mhmr_masked_smpl_token\n\n            loc_full_out = torch.zeros(\n                num_view * batch_size, max_humans, loc.shape[1], \n                device=loc.device\n            )\n            loc_full_out[smpl_mask] += loc\n        \n            head_token = full_out\n            head_loc = loc_full_out\n            expand = lambda x: x.expand(num_view * batch_size, max_humans , -1)\n\n        if self.output_mode == \"naive\":\n            # Get initial smpl token from MHMR\n            pred_body_pose, pred_betas, pred_cam, pred_expression = [expand(x) for x in\n                    [self.downstream_head.init_body_pose, \n                    self.downstream_head.init_betas, \n                    self.downstream_head.init_cam, \n                    self.downstream_head.init_expression,\n                    ]]\n            head_token = torch.cat([\n                head_token, pred_body_pose, pred_betas, pred_cam, \n                ], dim=-1)  # training: [bs, 10, 1454]; inference: [nhv, 1454]\n\n        if inference:\n            smpl_query_list, smpl_pos_list = [], []\n            for i in range(num_view * batch_size):\n                if self.output_mode == \"naive\":\n                    smpl_query = self.transformer(\n                        head_token[img_id == i].unsqueeze(0), \n                        context=feat_all[i].unsqueeze(0), \n                        mask=None) # train:[bs, 10, 1024]), inference:[1, nhv, 1024]\n                else:\n                    smpl_query = head_token[img_id == i].unsqueeze(0)   # use mhmr vit token\n                smpl_query_list.append(smpl_query)\n                smpl_pos = torch.zeros(\n                    *smpl_query.shape[:2], 2).to(smpl_query.device).to(pos[0].dtype)\n                smpl_pos_list.append(smpl_pos)\n            loc_list = [\n                head_loc[img_id == i].unsqueeze(0) for i in range(num_view * batch_size)]\n        else:\n            if self.output_mode == \"naive\":\n                smpl_query = self.transformer(\n                    head_token, \n                    context=feat_all, \n                    mask=smpl_mask.type(torch.float32)) # [bs, 10, 1024])\n            else:\n                smpl_query = head_token   # use mhmr vit token\n            smpl_query_list = smpl_query.chunk(num_view, dim=0)\n            loc_list = head_loc.chunk(num_view, dim=0)\n            full_pos = torch.zeros(\n                *smpl_query.shape[:2], 2).to(smpl_query.device).to(pos[0].dtype)\n            smpl_pos_list = full_pos.chunk(num_view, dim=0)\n\n        return (\n            scores.chunk(num_view, dim=0), \n            smpl_query_list,\n            smpl_pos_list,\n            loc_list,\n            msks.chunk(num_view, dim=0) if self.msk_head_flag else None,\n        )\n\n    def smpl_tokenizer_cut3r(self, feat, pos, views, loc, inference=False):\n        feat = torch.stack([f.detach() for f in feat], dim=0) #(num_view, bs, 576, 1024)\n        num_view, batch_size = feat.shape[:2]\n\n        feat = feat.view(-1, *feat.shape[2:]) #(num_view * bs, 576, 1024)\n        pos = torch.stack([p.detach() for p in pos], dim=0) #(num_view, bs, 576, 2)\n        pos = pos.view(-1, *pos.shape[2:]) #(num_view * bs, 576, 2)\n\n        # Restore Height and Width dimensions.\n        n_patch = views[0][\"true_shape\"][0] // self.croco_args['patch_size'] # H,W\n        feat = rearrange(feat, \"b (nh nw) c -> b nh nw c\", nh=n_patch[0], nw=n_patch[1]) # (num_view * bs, h, w, 1024)\n        pos = rearrange(pos, \"b (nh nw) c -> b nh nw c\", nh=n_patch[0], nw=n_patch[1]) # (num_view * bs, h, w, 2)\n\n        if inference:\n            num_humans = [l.shape[1] for l in loc]\n            img_id = torch.repeat_interleave(\n                torch.arange(len(loc), device=loc[0].device), \n                torch.tensor(num_humans, device=loc[0].device)\n            )","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.smpl_tokenizer_cut3r","uri":"program://Human3R/function/src.dust3r.model.smpl_tokenizer_cut3r#L1101-L1181","kind":"function","name":"smpl_tokenizer_cut3r","path":"src/dust3r/model.py","language":"python","start_line":1101,"end_line":1181,"context_start_line":1081,"context_end_line":1201,"code":"                smpl_query = self.transformer(\n                    head_token, \n                    context=feat_all, \n                    mask=smpl_mask.type(torch.float32)) # [bs, 10, 1024])\n            else:\n                smpl_query = head_token   # use mhmr vit token\n            smpl_query_list = smpl_query.chunk(num_view, dim=0)\n            loc_list = head_loc.chunk(num_view, dim=0)\n            full_pos = torch.zeros(\n                *smpl_query.shape[:2], 2).to(smpl_query.device).to(pos[0].dtype)\n            smpl_pos_list = full_pos.chunk(num_view, dim=0)\n\n        return (\n            scores.chunk(num_view, dim=0), \n            smpl_query_list,\n            smpl_pos_list,\n            loc_list,\n            msks.chunk(num_view, dim=0) if self.msk_head_flag else None,\n        )\n\n    def smpl_tokenizer_cut3r(self, feat, pos, views, loc, inference=False):\n        feat = torch.stack([f.detach() for f in feat], dim=0) #(num_view, bs, 576, 1024)\n        num_view, batch_size = feat.shape[:2]\n\n        feat = feat.view(-1, *feat.shape[2:]) #(num_view * bs, 576, 1024)\n        pos = torch.stack([p.detach() for p in pos], dim=0) #(num_view, bs, 576, 2)\n        pos = pos.view(-1, *pos.shape[2:]) #(num_view * bs, 576, 2)\n\n        # Restore Height and Width dimensions.\n        n_patch = views[0][\"true_shape\"][0] // self.croco_args['patch_size'] # H,W\n        feat = rearrange(feat, \"b (nh nw) c -> b nh nw c\", nh=n_patch[0], nw=n_patch[1]) # (num_view * bs, h, w, 1024)\n        pos = rearrange(pos, \"b (nh nw) c -> b nh nw c\", nh=n_patch[0], nw=n_patch[1]) # (num_view * bs, h, w, 2)\n\n        if inference:\n            num_humans = [l.shape[1] for l in loc]\n            img_id = torch.repeat_interleave(\n                torch.arange(len(loc), device=loc[0].device), \n                torch.tensor(num_humans, device=loc[0].device)\n            )\n            loc = torch.cat([l.squeeze(0) for l in loc], dim=0) # (nvh, 2)\n            loc_cut3r = unpad_uv(loc, self.mhmr_img_res, *views[0][\"true_shape\"][0])\n            smpl_uv = (loc_cut3r // self.croco_args['patch_size']).int()\n            w_id, h_id = smpl_uv.T\n        else:\n            smpl_mask = torch.stack([view[\"smpl_mask\"] for view in views], dim=0)\n            smpl_mask = smpl_mask.view(-1, *smpl_mask.shape[2:])\n            max_humans = smpl_mask.shape[1]\n            loc = torch.stack([l.detach() for l in loc], dim=0) # high-res head uv in mhmr: (num_view, bs, 10, 2)\n            loc = loc.view(-1, *loc.shape[2:]) #(num_view * bs, 10, 2)\n            loc_cut3r = unpad_uv(loc[smpl_mask], self.mhmr_img_res, *views[0][\"true_shape\"][0]) # high-res head uv in cut3r\n            smpl_uv = (loc_cut3r // self.croco_args['patch_size']).int() # low-res head uv in cut3r\n            img_id = torch.where(smpl_mask)[0]\n            w_id, h_id = smpl_uv.T\n\n        # Head token\n        feat_central = feat[img_id, h_id, w_id] # (nvh, 1024)\n        pos_central = pos[img_id, h_id, w_id] # (nvh, 2)\n\n        if inference:\n            smpl_query = feat_central\n            head_uv = smpl_uv\n        else:\n            # concat with mask token and mean SMPL params\n            full_out = torch.zeros(\n                num_view * batch_size, max_humans, feat_central.shape[1], \n                device=feat_central.device\n            )\n            full_pos = torch.zeros(\n                num_view * batch_size, max_humans, pos_central.shape[1], \n                device=pos_central.device, dtype=pos_central.dtype,\n            )\n            full_out[smpl_mask] += feat_central\n            full_out[~smpl_mask] += self.masked_smpl_token\n            full_pos[smpl_mask] += pos_central\n            smpl_query = full_out\n\n            uv_full_out = torch.zeros(\n                num_view * batch_size, max_humans, smpl_uv.shape[1], \n                device=loc.device,\n                dtype=smpl_uv.dtype\n            )\n            uv_full_out[smpl_mask] += smpl_uv\n            head_uv = uv_full_out\n\n        if inference:\n            smpl_query_list = [\n                smpl_query[img_id == i].unsqueeze(0) for i in range(num_view * batch_size)]\n            smpl_pos_list = [\n                pos_central[img_id == i].unsqueeze(0) for i in range(num_view * batch_size)]\n            smpl_uv_list = [\n                head_uv[img_id == i].unsqueeze(0) for i in range(num_view * batch_size)]\n        else:\n            smpl_query_list = smpl_query.chunk(num_view, dim=0)\n            smpl_pos_list = full_pos.chunk(num_view, dim=0)\n            smpl_uv_list = head_uv.chunk(num_view, dim=0)\n\n        return (\n            smpl_query_list,\n            smpl_pos_list,\n            smpl_uv_list,\n        )\n\n    def token_fuse(self, tk_mhmr, tk_cut3r, inference):\n        if inference:\n            num_humans = [t.shape[1] for t in tk_mhmr]\n            num_view = len(tk_mhmr)\n            img_id = torch.repeat_interleave(\n                torch.arange(num_view, device=tk_mhmr[0].device), \n                torch.tensor(num_humans, device=tk_mhmr[0].device)\n            )\n            tk_mhmr = torch.cat([t.squeeze(0) for t in tk_mhmr], dim=0) # (nvh, 1024)\n            tk_cut3r = torch.cat([t.squeeze(0) for t in tk_cut3r], dim=0) # (nvh, 1024)\n            tk = torch.cat([tk_mhmr, tk_cut3r], dim=-1) #(nvh, 2048)\n            fused_tk = self.downstream_head.mlp_fuse(tk)\n            fused_tk_list = [\n                fused_tk[img_id == i].unsqueeze(0) for i in range(num_view)]\n        else:\n            tk_mhmr = torch.stack([t.detach() for t in tk_mhmr], dim=0) #(num_view, bs, 10, 1024)\n            tk_cut3r = torch.stack([t.detach() for t in tk_cut3r], dim=0) #(num_view, bs, 10, 1024)\n            num_view, batch_size = tk_mhmr.shape[:2]\n        ","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.token_fuse","uri":"program://Human3R/function/src.dust3r.model.token_fuse#L1183-L1208","kind":"function","name":"token_fuse","path":"src/dust3r/model.py","language":"python","start_line":1183,"end_line":1208,"context_start_line":1163,"context_end_line":1228,"code":"            head_uv = uv_full_out\n\n        if inference:\n            smpl_query_list = [\n                smpl_query[img_id == i].unsqueeze(0) for i in range(num_view * batch_size)]\n            smpl_pos_list = [\n                pos_central[img_id == i].unsqueeze(0) for i in range(num_view * batch_size)]\n            smpl_uv_list = [\n                head_uv[img_id == i].unsqueeze(0) for i in range(num_view * batch_size)]\n        else:\n            smpl_query_list = smpl_query.chunk(num_view, dim=0)\n            smpl_pos_list = full_pos.chunk(num_view, dim=0)\n            smpl_uv_list = head_uv.chunk(num_view, dim=0)\n\n        return (\n            smpl_query_list,\n            smpl_pos_list,\n            smpl_uv_list,\n        )\n\n    def token_fuse(self, tk_mhmr, tk_cut3r, inference):\n        if inference:\n            num_humans = [t.shape[1] for t in tk_mhmr]\n            num_view = len(tk_mhmr)\n            img_id = torch.repeat_interleave(\n                torch.arange(num_view, device=tk_mhmr[0].device), \n                torch.tensor(num_humans, device=tk_mhmr[0].device)\n            )\n            tk_mhmr = torch.cat([t.squeeze(0) for t in tk_mhmr], dim=0) # (nvh, 1024)\n            tk_cut3r = torch.cat([t.squeeze(0) for t in tk_cut3r], dim=0) # (nvh, 1024)\n            tk = torch.cat([tk_mhmr, tk_cut3r], dim=-1) #(nvh, 2048)\n            fused_tk = self.downstream_head.mlp_fuse(tk)\n            fused_tk_list = [\n                fused_tk[img_id == i].unsqueeze(0) for i in range(num_view)]\n        else:\n            tk_mhmr = torch.stack([t.detach() for t in tk_mhmr], dim=0) #(num_view, bs, 10, 1024)\n            tk_cut3r = torch.stack([t.detach() for t in tk_cut3r], dim=0) #(num_view, bs, 10, 1024)\n            num_view, batch_size = tk_mhmr.shape[:2]\n        \n            tk_mhmr = tk_mhmr.view(-1, *tk_mhmr.shape[2:]) #(num_view * bs, 10, 1024)\n            tk_cut3r = tk_cut3r.view(-1, *tk_cut3r.shape[2:]) #(num_view * bs, 10, 1024)\n            tk = torch.cat([tk_mhmr, tk_cut3r], dim=-1) #(num_view * bs, 10, 2048)\n\n            fused_tk = self.downstream_head.mlp_fuse(tk) #(num_view * bs, 10, 768)\n            fused_tk_list = fused_tk.chunk(num_view, dim=0)\n        return fused_tk_list\n\n    def _forward_impl(self, views, ret_state=False, inference=False):\n        shape, feat_ls, pos, mhmr_feat_ls = self._encode_views_mhmr(views)\n        feat = feat_ls[-1]\n        mhmr_feat = mhmr_feat_ls[-1]\n\n        scores, smpl_tk_mhmr, pos_mhmr, smpl_loc, msks = self.smpl_tokenizer_mhmr(\n            mhmr_feat, pos, views, inference)\n        smpl_tk_cut3r, pos_cut3r, smpl_uv_cut3r = self.smpl_tokenizer_cut3r(\n            feat, pos, views, smpl_loc, inference)\n        \n        # fuse CUT3R and MHMR smpl tokens\n        smpl_query = self.token_fuse(smpl_tk_mhmr, smpl_tk_cut3r, inference)\n        pos_central = pos_cut3r\n\n        state_feat, state_pos = self._init_state(feat[0], pos[0])\n        mem = self.pose_retriever.mem.expand(feat[0].shape[0], -1, -1)  # [b, 256, 1536]\n        init_state_feat = state_feat.clone()\n        init_mem = mem.clone()\n        all_state_args = [(state_feat, state_pos, init_state_feat, mem, init_mem)]","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._forward_impl","uri":"program://Human3R/function/src.dust3r.model._forward_impl#L1210-L1320","kind":"function","name":"_forward_impl","path":"src/dust3r/model.py","language":"python","start_line":1210,"end_line":1320,"context_start_line":1190,"context_end_line":1340,"code":"            )\n            tk_mhmr = torch.cat([t.squeeze(0) for t in tk_mhmr], dim=0) # (nvh, 1024)\n            tk_cut3r = torch.cat([t.squeeze(0) for t in tk_cut3r], dim=0) # (nvh, 1024)\n            tk = torch.cat([tk_mhmr, tk_cut3r], dim=-1) #(nvh, 2048)\n            fused_tk = self.downstream_head.mlp_fuse(tk)\n            fused_tk_list = [\n                fused_tk[img_id == i].unsqueeze(0) for i in range(num_view)]\n        else:\n            tk_mhmr = torch.stack([t.detach() for t in tk_mhmr], dim=0) #(num_view, bs, 10, 1024)\n            tk_cut3r = torch.stack([t.detach() for t in tk_cut3r], dim=0) #(num_view, bs, 10, 1024)\n            num_view, batch_size = tk_mhmr.shape[:2]\n        \n            tk_mhmr = tk_mhmr.view(-1, *tk_mhmr.shape[2:]) #(num_view * bs, 10, 1024)\n            tk_cut3r = tk_cut3r.view(-1, *tk_cut3r.shape[2:]) #(num_view * bs, 10, 1024)\n            tk = torch.cat([tk_mhmr, tk_cut3r], dim=-1) #(num_view * bs, 10, 2048)\n\n            fused_tk = self.downstream_head.mlp_fuse(tk) #(num_view * bs, 10, 768)\n            fused_tk_list = fused_tk.chunk(num_view, dim=0)\n        return fused_tk_list\n\n    def _forward_impl(self, views, ret_state=False, inference=False):\n        shape, feat_ls, pos, mhmr_feat_ls = self._encode_views_mhmr(views)\n        feat = feat_ls[-1]\n        mhmr_feat = mhmr_feat_ls[-1]\n\n        scores, smpl_tk_mhmr, pos_mhmr, smpl_loc, msks = self.smpl_tokenizer_mhmr(\n            mhmr_feat, pos, views, inference)\n        smpl_tk_cut3r, pos_cut3r, smpl_uv_cut3r = self.smpl_tokenizer_cut3r(\n            feat, pos, views, smpl_loc, inference)\n        \n        # fuse CUT3R and MHMR smpl tokens\n        smpl_query = self.token_fuse(smpl_tk_mhmr, smpl_tk_cut3r, inference)\n        pos_central = pos_cut3r\n\n        state_feat, state_pos = self._init_state(feat[0], pos[0])\n        mem = self.pose_retriever.mem.expand(feat[0].shape[0], -1, -1)  # [b, 256, 1536]\n        init_state_feat = state_feat.clone()\n        init_mem = mem.clone()\n        all_state_args = [(state_feat, state_pos, init_state_feat, mem, init_mem)]\n        ress = []\n        for i in range(len(views)):\n            feat_i = feat[i]\n            pos_i = pos[i]\n            smpl_feat_i = smpl_query[i]\n            smpl_pos_i = pos_central[i]\n            n_humans_i = smpl_feat_i.shape[1]\n\n            if self.pose_head_flag:\n                global_img_feat_i = self._get_img_level_feat(feat_i)    # [b, 1, 1024]\n                if i == 0:\n                    pose_feat_i = self.pose_token.expand(feat_i.shape[0], -1, -1)   # coarse pose feat: [b, 1, 768]\n                else:\n                    pose_feat_i = self.pose_retriever.inquire(global_img_feat_i, mem)   # coarse pose feat: [b, 1, 768]\n                pose_pos_i = -torch.ones(\n                    feat_i.shape[0], 1, 2, device=feat_i.device, dtype=pos_i.dtype\n                )\n            else:\n                pose_feat_i = None\n                pose_pos_i = None\n\n            new_state_feat, dec, _ = self._recurrent_rollout(\n                state_feat,\n                state_pos,\n                feat_i,\n                pos_i,\n                pose_feat_i,\n                pose_pos_i,\n                smpl_feat_i,\n                smpl_pos_i,\n                init_state_feat,\n                img_mask=views[i][\"img_mask\"],\n                reset_mask=views[i][\"reset\"],\n                update=views[i].get(\"update\", None),\n            )\n            out_pose_feat_i = dec[-1][:, 0:1]   # After Cross-Attention, refined pose feat: [b, 1, 768]\n            new_mem = self.pose_retriever.update_mem(\n                mem, global_img_feat_i, out_pose_feat_i\n            )   # [b, 256, 1536]\n\n            assert len(dec) == self.dec_depth + 1\n            if n_humans_i > 0:\n                head_input = [\n                    dec[0].float(),\n                    dec[self.dec_depth * 2 // 4][:, 1:-n_humans_i].float(),\n                    dec[self.dec_depth * 3 // 4][:, 1:-n_humans_i].float(),\n                    dec[self.dec_depth][:, :-n_humans_i].float(),\n                ]\n                smpl_token = dec[self.dec_depth][:, -n_humans_i:].float()\n                smpl_token = torch.cat([smpl_token, smpl_tk_mhmr[i]], dim=-1)\n            else:\n                head_input = [\n                    dec[0].float(),\n                    dec[self.dec_depth * 2 // 4][:, 1:].float(),\n                    dec[self.dec_depth * 3 // 4][:, 1:].float(),\n                    dec[self.dec_depth].float(),\n                ]\n                smpl_token = None\n            res = self._downstream_head(\n                head_input, shape[i], pos=pos_i, n_humans=n_humans_i, smpl_token=smpl_token)\n            if self.msk_head_flag:\n                res['msk'] = msks[i]\n            ress.append({\n                **res, 'smpl_scores': scores[i], 'smpl_loc': smpl_loc[i]})\n            img_mask = views[i][\"img_mask\"]\n            update = views[i].get(\"update\", None)\n            if update is not None:\n                update_mask = (\n                    img_mask & update\n                )  # if don't update, then whatever img_mask\n            else:\n                update_mask = img_mask\n            update_mask = update_mask[:, None, None].float()\n            state_feat = new_state_feat * update_mask + state_feat * (\n                1 - update_mask\n            )  # update global state\n            mem = new_mem * update_mask + mem * (\n                1 - update_mask\n            )  # then update local state\n            reset_mask = views[i][\"reset\"]\n            if reset_mask is not None:\n                reset_mask = reset_mask[:, None, None].float()\n                state_feat = init_state_feat * reset_mask + state_feat * (\n                    1 - reset_mask\n                )\n                mem = init_mem * reset_mask + mem * (1 - reset_mask)\n            all_state_args.append(\n                (state_feat, state_pos, init_state_feat, mem, init_mem)\n            )\n        if ret_state:\n            return ress, views, all_state_args\n        return ress, views\n\n    def _forward_impl_naive(self, views, ret_state=False, inference=False):\n        shape, feat_ls, pos, mhmr_feat_ls = self._encode_views_mhmr(views)\n        feat = feat_ls[-1]\n        mhmr_feat = mhmr_feat_ls[-1]\n\n        scores, smpl_tk_mhmr, pos_mhmr, smpl_loc, msks = self.smpl_tokenizer_mhmr(\n            mhmr_feat, pos, views, inference)\n\n        # naive CUT3R+MHMR\n        smpl_query = smpl_tk_mhmr\n        pos_central = pos_mhmr\n\n        state_feat, state_pos = self._init_state(feat[0], pos[0])\n        mem = self.pose_retriever.mem.expand(feat[0].shape[0], -1, -1)  # [b, 256, 1536]\n        init_state_feat = state_feat.clone()\n        init_mem = mem.clone()\n        all_state_args = [(state_feat, state_pos, init_state_feat, mem, init_mem)]\n        ress = []\n        for i in range(len(views)):","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model._forward_impl_naive","uri":"program://Human3R/function/src.dust3r.model._forward_impl_naive#L1322-L1422","kind":"function","name":"_forward_impl_naive","path":"src/dust3r/model.py","language":"python","start_line":1322,"end_line":1422,"context_start_line":1302,"context_end_line":1442,"code":"            state_feat = new_state_feat * update_mask + state_feat * (\n                1 - update_mask\n            )  # update global state\n            mem = new_mem * update_mask + mem * (\n                1 - update_mask\n            )  # then update local state\n            reset_mask = views[i][\"reset\"]\n            if reset_mask is not None:\n                reset_mask = reset_mask[:, None, None].float()\n                state_feat = init_state_feat * reset_mask + state_feat * (\n                    1 - reset_mask\n                )\n                mem = init_mem * reset_mask + mem * (1 - reset_mask)\n            all_state_args.append(\n                (state_feat, state_pos, init_state_feat, mem, init_mem)\n            )\n        if ret_state:\n            return ress, views, all_state_args\n        return ress, views\n\n    def _forward_impl_naive(self, views, ret_state=False, inference=False):\n        shape, feat_ls, pos, mhmr_feat_ls = self._encode_views_mhmr(views)\n        feat = feat_ls[-1]\n        mhmr_feat = mhmr_feat_ls[-1]\n\n        scores, smpl_tk_mhmr, pos_mhmr, smpl_loc, msks = self.smpl_tokenizer_mhmr(\n            mhmr_feat, pos, views, inference)\n\n        # naive CUT3R+MHMR\n        smpl_query = smpl_tk_mhmr\n        pos_central = pos_mhmr\n\n        state_feat, state_pos = self._init_state(feat[0], pos[0])\n        mem = self.pose_retriever.mem.expand(feat[0].shape[0], -1, -1)  # [b, 256, 1536]\n        init_state_feat = state_feat.clone()\n        init_mem = mem.clone()\n        all_state_args = [(state_feat, state_pos, init_state_feat, mem, init_mem)]\n        ress = []\n        for i in range(len(views)):\n            feat_i = feat[i]\n            pos_i = pos[i]\n            smpl_feat_i = smpl_query[i]\n            smpl_pos_i = pos_central[i]\n            n_humans_i = smpl_feat_i.shape[1]\n\n            if self.pose_head_flag:\n                global_img_feat_i = self._get_img_level_feat(feat_i)    # [b, 1, 1024]\n                if i == 0:\n                    pose_feat_i = self.pose_token.expand(feat_i.shape[0], -1, -1)   # coarse pose feat: [b, 1, 768]\n                else:\n                    pose_feat_i = self.pose_retriever.inquire(global_img_feat_i, mem)   # coarse pose feat: [b, 1, 768]\n                pose_pos_i = -torch.ones(\n                    feat_i.shape[0], 1, 2, device=feat_i.device, dtype=pos_i.dtype\n                )\n            else:\n                pose_feat_i = None\n                pose_pos_i = None\n\n            new_state_feat, dec, _ = self._recurrent_rollout(\n                state_feat,\n                state_pos,\n                feat_i,\n                pos_i,\n                pose_feat_i,\n                pose_pos_i,\n                None,\n                None,\n                init_state_feat,\n                img_mask=views[i][\"img_mask\"],\n                reset_mask=views[i][\"reset\"],\n                update=views[i].get(\"update\", None),\n            )\n            out_pose_feat_i = dec[-1][:, 0:1]\n            new_mem = self.pose_retriever.update_mem(\n                mem, global_img_feat_i, out_pose_feat_i\n            )   # [b, 256, 1536]\n\n            assert len(dec) == self.dec_depth + 1\n            head_input = [\n                dec[0].float(),\n                dec[self.dec_depth * 2 // 4][:, 1:].float(),\n                dec[self.dec_depth * 3 // 4][:, 1:].float(),\n                dec[self.dec_depth].float(),\n            ]\n            if n_humans_i > 0:\n                smpl_token = smpl_feat_i # used for naive CUT3R+MHMR\n            else:\n                smpl_token = None\n            res = self._downstream_head(\n                head_input, shape[i], pos=pos_i, n_humans=n_humans_i, smpl_token=smpl_token)\n\n            ress.append({\n                **res, 'smpl_scores': scores[i], 'smpl_loc': smpl_loc[i]})\n            img_mask = views[i][\"img_mask\"]\n            update = views[i].get(\"update\", None)\n            if update is not None:\n                update_mask = (\n                    img_mask & update\n                )  # if don't update, then whatever img_mask\n            else:\n                update_mask = img_mask\n            update_mask = update_mask[:, None, None].float()\n            state_feat = new_state_feat * update_mask + state_feat * (\n                1 - update_mask\n            )  # update global state\n            mem = new_mem * update_mask + mem * (\n                1 - update_mask\n            )  # then update local state\n            reset_mask = views[i][\"reset\"]\n            if reset_mask is not None:\n                reset_mask = reset_mask[:, None, None].float()\n                state_feat = init_state_feat * reset_mask + state_feat * (\n                    1 - reset_mask\n                )\n                mem = init_mem * reset_mask + mem * (1 - reset_mask)\n            all_state_args.append(\n                (state_feat, state_pos, init_state_feat, mem, init_mem)\n            )\n        if ret_state:\n            return ress, views, all_state_args\n        return ress, views\n\n    def forward(self, views, ret_state=False, inference=False):\n        if self.output_mode == \"naive\":\n            if ret_state:\n                ress, views, state_args = self._forward_impl_naive(views, ret_state=ret_state, inference=inference)\n                return ARCroco3DStereoOutput(ress=ress, views=views), state_args\n            else:\n                ress, views = self._forward_impl_naive(views, ret_state=ret_state)\n                return ARCroco3DStereoOutput(ress=ress, views=views)\n        else:\n            if ret_state:\n                ress, views, state_args = self._forward_impl(views, ret_state=ret_state, inference=inference)\n                return ARCroco3DStereoOutput(ress=ress, views=views), state_args\n            else:\n                ress, views = self._forward_impl(views, ret_state=ret_state)\n                return ARCroco3DStereoOutput(ress=ress, views=views)\n\n\n    def forward_recurrent_lighter(self, views, device, ret_state=False, use_ttt3r=False):\n        ress = []","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.forward","uri":"program://Human3R/function/src.dust3r.model.forward#L1424-L1438","kind":"function","name":"forward","path":"src/dust3r/model.py","language":"python","start_line":1424,"end_line":1438,"context_start_line":1404,"context_end_line":1458,"code":"            state_feat = new_state_feat * update_mask + state_feat * (\n                1 - update_mask\n            )  # update global state\n            mem = new_mem * update_mask + mem * (\n                1 - update_mask\n            )  # then update local state\n            reset_mask = views[i][\"reset\"]\n            if reset_mask is not None:\n                reset_mask = reset_mask[:, None, None].float()\n                state_feat = init_state_feat * reset_mask + state_feat * (\n                    1 - reset_mask\n                )\n                mem = init_mem * reset_mask + mem * (1 - reset_mask)\n            all_state_args.append(\n                (state_feat, state_pos, init_state_feat, mem, init_mem)\n            )\n        if ret_state:\n            return ress, views, all_state_args\n        return ress, views\n\n    def forward(self, views, ret_state=False, inference=False):\n        if self.output_mode == \"naive\":\n            if ret_state:\n                ress, views, state_args = self._forward_impl_naive(views, ret_state=ret_state, inference=inference)\n                return ARCroco3DStereoOutput(ress=ress, views=views), state_args\n            else:\n                ress, views = self._forward_impl_naive(views, ret_state=ret_state)\n                return ARCroco3DStereoOutput(ress=ress, views=views)\n        else:\n            if ret_state:\n                ress, views, state_args = self._forward_impl(views, ret_state=ret_state, inference=inference)\n                return ARCroco3DStereoOutput(ress=ress, views=views), state_args\n            else:\n                ress, views = self._forward_impl(views, ret_state=ret_state)\n                return ARCroco3DStereoOutput(ress=ress, views=views)\n\n\n    def forward_recurrent_lighter(self, views, device, ret_state=False, use_ttt3r=False):\n        ress = []\n        all_state_args = []\n        last_smpl_tk = None\n        last_smpl_id = None\n        max_smpl_id = -1\n        reset_mask = False\n        for i, _view in enumerate(views):\n            view = to_gpu(_view, device)\n            batch_size = view[\"img\"].shape[0]\n            img_mask = view[\"img_mask\"].reshape(\n                -1, batch_size\n            )  # Shape: (1, batch_size)\n            imgs = view[\"img\"].unsqueeze(0)  # Shape: (1, batch_size, C, H, W)\n            shapes = (\n                view[\"true_shape\"].unsqueeze(0)\n                if \"true_shape\" in view\n                else torch.tensor(view[\"img\"].shape[-2:], device=device)","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.forward_recurrent_lighter","uri":"program://Human3R/function/src.dust3r.model.forward_recurrent_lighter#L1441-L1701","kind":"function","name":"forward_recurrent_lighter","path":"src/dust3r/model.py","language":"python","start_line":1441,"end_line":1701,"context_start_line":1421,"context_end_line":1721,"code":"            return ress, views, all_state_args\n        return ress, views\n\n    def forward(self, views, ret_state=False, inference=False):\n        if self.output_mode == \"naive\":\n            if ret_state:\n                ress, views, state_args = self._forward_impl_naive(views, ret_state=ret_state, inference=inference)\n                return ARCroco3DStereoOutput(ress=ress, views=views), state_args\n            else:\n                ress, views = self._forward_impl_naive(views, ret_state=ret_state)\n                return ARCroco3DStereoOutput(ress=ress, views=views)\n        else:\n            if ret_state:\n                ress, views, state_args = self._forward_impl(views, ret_state=ret_state, inference=inference)\n                return ARCroco3DStereoOutput(ress=ress, views=views), state_args\n            else:\n                ress, views = self._forward_impl(views, ret_state=ret_state)\n                return ARCroco3DStereoOutput(ress=ress, views=views)\n\n\n    def forward_recurrent_lighter(self, views, device, ret_state=False, use_ttt3r=False):\n        ress = []\n        all_state_args = []\n        last_smpl_tk = None\n        last_smpl_id = None\n        max_smpl_id = -1\n        reset_mask = False\n        for i, _view in enumerate(views):\n            view = to_gpu(_view, device)\n            batch_size = view[\"img\"].shape[0]\n            img_mask = view[\"img_mask\"].reshape(\n                -1, batch_size\n            )  # Shape: (1, batch_size)\n            imgs = view[\"img\"].unsqueeze(0)  # Shape: (1, batch_size, C, H, W)\n            shapes = (\n                view[\"true_shape\"].unsqueeze(0)\n                if \"true_shape\" in view\n                else torch.tensor(view[\"img\"].shape[-2:], device=device)\n                .unsqueeze(0)\n                .repeat(batch_size, 1)\n                .unsqueeze(0)\n            )  # Shape: (num_views, batch_size, 2)\n            imgs = imgs.view(\n                -1, *imgs.shape[2:]\n            )  # Shape: (num_views * batch_size, C, H, W)\n            shapes = shapes.view(-1, 2)  # Shape: (num_views * batch_size, 2)\n            img_masks_flat = img_mask.view(-1)  # Shape: (num_views * batch_size)\n            selected_imgs = imgs[img_masks_flat]\n            selected_shapes = shapes[img_masks_flat]\n            if selected_imgs.size(0) > 0:\n                img_out, img_pos, _ = self._encode_image(selected_imgs, selected_shapes)\n            else:\n                img_out, img_pos = None, None\n\n            shape = shapes\n            feat_i = img_out[-1]\n            pos_i = img_pos\n            \n            # MHMR vit\n            imgs_mhmr = view[\"img_mhmr\"].unsqueeze(0)  # Shape: (1, batch_size, C, H, W)\n            imgs_mhmr = imgs_mhmr.view(\n                -1, *imgs_mhmr.shape[2:]\n            )  # Shape: (num_views * batch_size, C, H, W)\n            selected_imgs_mhmr = imgs_mhmr[img_masks_flat]\n            if selected_imgs_mhmr.size(0) > 0:\n                mean = torch.tensor([0.485, 0.456, 0.406], device=device)[None, :, None, None]\n                std = torch.tensor([0.229, 0.224, 0.225], device=device)[None, :, None, None]\n                selected_imgs_mhmr = (selected_imgs_mhmr * 0.5 + 0.5 - mean) / std\n                mhmr_img_out = [self.backbone(selected_imgs_mhmr)] # image[bs, 3, h, w] -> image feature [bs, h_patches*w_patches, D]\n            feat_mhmr_i = mhmr_img_out[-1]\n\n            # MHMR smpl tokenizer\n            n_patch_mhmr = self.bb_token_res\n            scores = self.downstream_head.detect_mhmr(feat_mhmr_i) #(num_view * bs, h_patches*w_patches, 1)\n            scores = rearrange(scores, \"b (nh nw) c -> b c nh nw\", nh=n_patch_mhmr, nw=n_patch_mhmr) # [num_view * bs, h_nb_patches * w_nb_patches, 1] -> [num_view * bs, 1, h, w]\n            if self.msk_head_flag:\n                msks = self.downstream_head.segment(feat_mhmr_i) # low-res mask\n                msks = rearrange(msks, \"b (nh nw) c -> b c nh nw\", nh=n_patch_mhmr, nw=n_patch_mhmr)\n                msks = F.pixel_shuffle(msks, self.bb_patch_size)  # (num_view * bs, 1, h, w)\n                msks = msks.permute((0, 2, 3, 1))\n            feat_mhmr_i = rearrange(feat_mhmr_i, \"b (nh nw) c -> b nh nw c\", nh=n_patch_mhmr, nw=n_patch_mhmr) # head token extraction: (num_view * bs, h, w, 1024)\n\n            scores = nms(scores, kernel=3) # (num_view * bs, 1, h, w)\n            scores = scores.permute((0, 2, 3, 1)) # (num_view * bs, h, w, 1)\n            idx = apply_threshold(0.3, scores)\n            img_id, h_id, w_id = idx[0], idx[1], idx[2]\n\n            # Head token and offset\n            feat_central_mhmr = feat_mhmr_i[img_id, h_id, w_id] # (nvh, 1024)\n            offset = self.downstream_head.mlp_offset(feat_central_mhmr)# [nhv,2]\n            # Distance for estimating the 3D location in 3D space\n            loc = torch.stack([w_id, h_id]).permute(1,0) # x,y\n            loc = (loc + 0.5 + offset) * self.bb_patch_size # Moving to higher res the location of the pelvis\n\n            smpl_tk_mhmr = feat_central_mhmr.unsqueeze(0)   # use mhmr vit token\n\n            # CUT3R smpl tokenizer\n            n_patch_cut3r = shape[0] // self.croco_args['patch_size'] # H,W\n            feat_cut3r_i = rearrange(\n                feat_i, \"b (nh nw) c -> b nh nw c\", nh=n_patch_cut3r[0], nw=n_patch_cut3r[1]) # (num_view * bs, h, w, 1024)\n            pos_cut3r_i = rearrange(\n                pos_i, \"b (nh nw) c -> b nh nw c\", nh=n_patch_cut3r[0], nw=n_patch_cut3r[1]) # (num_view * bs, h, w, 2)\n            \n            loc_cut3r = unpad_uv(loc, self.mhmr_img_res, *shape[0])\n            smpl_uv_cut3r = (loc_cut3r // self.croco_args['patch_size']).int()\n            w_id_cut3r, h_id_cut3r = smpl_uv_cut3r.T\n            feat_central_cut3r = feat_cut3r_i[img_id, h_id_cut3r, w_id_cut3r] # (nvh, 1024)\n            pos_central_cut3r = pos_cut3r_i[img_id, h_id_cut3r, w_id_cut3r] # (nvh, 2)\n\n            smpl_tk_cut3r = feat_central_cut3r.unsqueeze(0)\n            smpl_pos_cut3r = pos_central_cut3r.unsqueeze(0)\n\n            # fuse CUT3R and MHMR smpl tokens\n            fused_tk = torch.cat([smpl_tk_mhmr, smpl_tk_cut3r], dim=-1) #(1, nvh, 2048)\n            fused_tk = self.downstream_head.mlp_fuse(fused_tk) # (1, nvh, 768)\n\n            smpl_feat_i = fused_tk # (1,nvh, 768)\n            smpl_pos_i = smpl_pos_cut3r # (1,nvh, 2)\n     \n            n_humans_i = smpl_feat_i.shape[1]\n            if i == 0:\n                state_feat, state_pos = self._init_state(feat_i, pos_i)\n                mem = self.pose_retriever.mem.expand(feat_i.shape[0], -1, -1)\n                init_state_feat = state_feat.clone()\n                init_mem = mem.clone()\n\n            if self.pose_head_flag:\n                global_img_feat_i = self._get_img_level_feat(feat_i)\n                if i == 0 or reset_mask:\n                    pose_feat_i = self.pose_token.expand(feat_i.shape[0], -1, -1)\n                else:\n                    pose_feat_i = self.pose_retriever.inquire(global_img_feat_i, mem)\n                pose_pos_i = -torch.ones(\n                    feat_i.shape[0], 1, 2, device=device, dtype=pos_i.dtype\n                )\n            else:\n                pose_feat_i = None\n                pose_pos_i = None\n            new_state_feat, dec, cross_attn_states = self._recurrent_rollout(\n                state_feat,\n                state_pos,\n                feat_i,\n                pos_i,\n                pose_feat_i,\n                pose_pos_i,\n                smpl_feat_i,\n                smpl_pos_i,\n                init_state_feat,\n                img_mask=view[\"img_mask\"],\n                reset_mask=view[\"reset\"],\n                update=view.get(\"update\", None),\n                use_ttt3r=use_ttt3r,\n            )\n            out_pose_feat_i = dec[-1][:, 0:1]\n            new_mem = self.pose_retriever.update_mem(\n                mem, global_img_feat_i, out_pose_feat_i\n            )\n            assert len(dec) == self.dec_depth + 1\n            if n_humans_i > 0:\n                head_input = [\n                    dec[0].float(),\n                    dec[self.dec_depth * 2 // 4][:, 1:-n_humans_i].float(),\n                    dec[self.dec_depth * 3 // 4][:, 1:-n_humans_i].float(),\n                    dec[self.dec_depth][:, :-n_humans_i].float(),\n                ]\n                smpl_token = dec[self.dec_depth][:, -n_humans_i:].float()\n                smpl_token_cat = torch.cat([smpl_token, smpl_tk_mhmr], dim=-1)\n            else:\n                head_input = [\n                    dec[0].float(),\n                    dec[self.dec_depth * 2 // 4][:, 1:].float(),\n                    dec[self.dec_depth * 3 // 4][:, 1:].float(),\n                    dec[self.dec_depth].float(),\n                ]\n                smpl_token = None\n                smpl_token_cat = None\n            res = self._downstream_head(\n                head_input, shape, pos=pos_i, n_humans=n_humans_i, smpl_token=smpl_token_cat)\n\n            # tracking\n            num_miss_match0 = 0\n            if last_smpl_tk is not None and smpl_token is not None:\n                cost_mat = -torch.cdist(last_smpl_tk, smpl_token, p=2)\n                cost_mat = log_optimal_transport(\n                    cost_mat, alpha=torch.tensor(-10.0, device=device), iters=20)\n                matches = cost_mat[:, :-1, :-1]\n                max0, max1 = matches.max(2), matches.max(1)\n                indices0, indices1 = max0.indices, max1.indices\n                mutual0 = torch.arange(\n                    indices0.shape[1], device=device\n                    )[None] == indices1.gather(1, indices0)\n                mutual1 = torch.arange(\n                    indices1.shape[1], device=device\n                    )[None] == indices0.gather(1, indices1)\n                zero = matches.new_tensor(0)\n                mscores0 = torch.where(mutual0, max0.values.exp(), zero)\n                mscores1 = torch.where(mutual1, mscores0.gather(1, indices1), zero)\n\n                match_threshold = 0.2\n                valid0 = mutual0 & (mscores0 > match_threshold) # 1,n\n                valid1 = mutual1 & valid0.gather(1, indices1) # 1,m\n                # get the final matching indices, invalid matches set to -1\n                indices0 = torch.where(valid0, indices0, indices0.new_tensor(-1)) # [1, n] current frame matches for last frame\n                indices1 = torch.where(valid1, indices1, indices1.new_tensor(-1)) # [1, m] last frame matches for current frame\n\n                smpl_id = indices1.new_full(indices1.shape, -1) # 1,m\n                valid_match1 = indices1[valid1]\n                if valid_match1.numel() > 0:\n                    smpl_id[valid1] = last_smpl_id.gather(1, valid_match1[None]).flatten()\n\n                num_miss_match0 = int((~valid0).sum())\n                num_new_persons = len(smpl_id[~valid1])\n                if num_new_persons > 0:\n                    new_ids = torch.arange(\n                        max_smpl_id + 1,\n                        max_smpl_id + 1 + num_new_persons,\n                        device=device\n                    )\n                    smpl_id[~valid1] = new_ids\n                    max_smpl_id += num_new_persons\n            else:\n                # first frame with humans\n                if smpl_token is not None:\n                    smpl_id = torch.arange(n_humans_i, device=device)[None]  # (1, nvh)\n                    max_smpl_id = n_humans_i - 1\n                else:\n                    smpl_id = None\n\n            if smpl_token is not None:\n                if num_miss_match0 > 0:\n                    miss_match_id0 = last_smpl_id[~valid0][None]\n                    miss_match_tk0 = last_smpl_tk[~valid0][None]\n                    last_smpl_id = torch.cat([smpl_id, miss_match_id0], dim=1)\n                    last_smpl_tk = torch.cat([smpl_token, miss_match_tk0], dim=1)\n                else:\n                    last_smpl_tk = smpl_token.clone()\n                    last_smpl_id = smpl_id.clone()\n\n            if smpl_id is not None:\n                res['smpl_id'] = smpl_id\n\n            if self.msk_head_flag:\n                res['msk'] = msks\n            res_cpu = to_cpu({**res, 'smpl_scores': scores, 'smpl_loc': loc[None]})\n            ress.append(res_cpu)\n            # ress.append(res)\n\n            # updating the state and memory\n            img_mask = view[\"img_mask\"]\n            update = view.get(\"update\", None)\n            if update is not None:\n                update_mask = (\n                    img_mask & update\n                )  # if don't update, then whatever img_mask\n            else:\n                update_mask = img_mask\n            update_mask = update_mask[:, None, None].float()\n\n            if use_ttt3r and i != 0 and not reset_mask:\n                cross_attn_states = rearrange(torch.cat(cross_attn_states, dim=0), 'l h nstate nimg -> 1 nstate nimg (l h)').mean(dim=(-1, -2))\n                update_mask_state = update_mask * torch.sigmoid(cross_attn_states)[..., None]\n            else:\n                update_mask_state = update_mask\n\n            state_feat = new_state_feat * update_mask_state + state_feat * (\n                1 - update_mask_state\n            )  # update global state\n            mem = new_mem * update_mask + mem * (\n                1 - update_mask\n            )  # then update local state\n            reset_mask = view[\"reset\"]\n            if reset_mask is not None:\n                reset_mask = reset_mask[:, None, None].float()\n                state_feat = init_state_feat * reset_mask + state_feat * (\n                    1 - reset_mask\n                )\n                mem = init_mem * reset_mask + mem * (1 - reset_mask)\n           \n        if ret_state:\n            return ress, views, all_state_args\n        return ress, views\n\n    def forward_recurrent_lighter_naive(self, views, device, ret_state=False, use_ttt3r=False):\n        ress = []\n        all_state_args = []\n        last_smpl_tk = None\n        last_smpl_id = None\n        max_smpl_id = -1\n        reset_mask = False\n        for i, _view in enumerate(views):\n            view = to_gpu(_view, device)\n            batch_size = view[\"img\"].shape[0]\n            img_mask = view[\"img_mask\"].reshape(\n                -1, batch_size\n            )  # Shape: (1, batch_size)\n            imgs = view[\"img\"].unsqueeze(0)  # Shape: (1, batch_size, C, H, W)\n            shapes = (\n                view[\"true_shape\"].unsqueeze(0)\n                if \"true_shape\" in view\n                else torch.tensor(view[\"img\"].shape[-2:], device=device)\n                .unsqueeze(0)","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.model.forward_recurrent_lighter_naive","uri":"program://Human3R/function/src.dust3r.model.forward_recurrent_lighter_naive#L1703-L1958","kind":"function","name":"forward_recurrent_lighter_naive","path":"src/dust3r/model.py","language":"python","start_line":1703,"end_line":1958,"context_start_line":1683,"context_end_line":1978,"code":"                update_mask_state = update_mask\n\n            state_feat = new_state_feat * update_mask_state + state_feat * (\n                1 - update_mask_state\n            )  # update global state\n            mem = new_mem * update_mask + mem * (\n                1 - update_mask\n            )  # then update local state\n            reset_mask = view[\"reset\"]\n            if reset_mask is not None:\n                reset_mask = reset_mask[:, None, None].float()\n                state_feat = init_state_feat * reset_mask + state_feat * (\n                    1 - reset_mask\n                )\n                mem = init_mem * reset_mask + mem * (1 - reset_mask)\n           \n        if ret_state:\n            return ress, views, all_state_args\n        return ress, views\n\n    def forward_recurrent_lighter_naive(self, views, device, ret_state=False, use_ttt3r=False):\n        ress = []\n        all_state_args = []\n        last_smpl_tk = None\n        last_smpl_id = None\n        max_smpl_id = -1\n        reset_mask = False\n        for i, _view in enumerate(views):\n            view = to_gpu(_view, device)\n            batch_size = view[\"img\"].shape[0]\n            img_mask = view[\"img_mask\"].reshape(\n                -1, batch_size\n            )  # Shape: (1, batch_size)\n            imgs = view[\"img\"].unsqueeze(0)  # Shape: (1, batch_size, C, H, W)\n            shapes = (\n                view[\"true_shape\"].unsqueeze(0)\n                if \"true_shape\" in view\n                else torch.tensor(view[\"img\"].shape[-2:], device=device)\n                .unsqueeze(0)\n                .repeat(batch_size, 1)\n                .unsqueeze(0)\n            )  # Shape: (num_views, batch_size, 2)\n            imgs = imgs.view(\n                -1, *imgs.shape[2:]\n            )  # Shape: (num_views * batch_size, C, H, W)\n            shapes = shapes.view(-1, 2)  # Shape: (num_views * batch_size, 2)\n            img_masks_flat = img_mask.view(-1)  # Shape: (num_views * batch_size)\n            # ray_masks_flat = ray_mask.view(-1)\n            selected_imgs = imgs[img_masks_flat]\n            selected_shapes = shapes[img_masks_flat]\n            if selected_imgs.size(0) > 0:\n                img_out, img_pos, _ = self._encode_image(selected_imgs, selected_shapes)\n            else:\n                img_out, img_pos = None, None\n\n            shape = shapes\n            feat_i = img_out[-1]\n            pos_i = img_pos\n            \n            # MHMR vit\n            imgs_mhmr = view[\"img_mhmr\"].unsqueeze(0)  # Shape: (1, batch_size, C, H, W)\n            imgs_mhmr = imgs_mhmr.view(\n                -1, *imgs_mhmr.shape[2:]\n            )  # Shape: (num_views * batch_size, C, H, W)\n            selected_imgs_mhmr = imgs_mhmr[img_masks_flat]\n            if selected_imgs_mhmr.size(0) > 0:\n                mean = torch.tensor([0.485, 0.456, 0.406], device=device)[None, :, None, None]\n                std = torch.tensor([0.229, 0.224, 0.225], device=device)[None, :, None, None]\n                selected_imgs_mhmr = (selected_imgs_mhmr * 0.5 + 0.5 - mean) / std\n                mhmr_img_out = [self.backbone(selected_imgs_mhmr)] # image[bs, 3, h, w] -> image feature [bs, h_patches*w_patches, D]\n            feat_mhmr_i = mhmr_img_out[-1]\n\n\n            # MHMR smpl tokenizer\n            n_patch_mhmr = self.bb_token_res\n            scores = self.downstream_head.detect_mhmr(feat_mhmr_i) #(num_view * bs, h_patches*w_patches, 1)\n            scores = rearrange(scores, \"b (nh nw) c -> b c nh nw\", nh=n_patch_mhmr, nw=n_patch_mhmr) # [num_view * bs, h_nb_patches * w_nb_patches, 1] -> [num_view * bs, 1, h, w]\n            feat_mhmr_i = rearrange(feat_mhmr_i, \"b (nh nw) c -> b nh nw c\", nh=n_patch_mhmr, nw=n_patch_mhmr) # head token extraction: (num_view * bs, h, w, 1024)\n\n            scores = nms(scores, kernel=3) # (num_view * bs, 1, h, w)\n            scores = scores.permute((0, 2, 3, 1)) # (num_view * bs, h, w, 1)\n            idx = apply_threshold(0.3, scores)\n            img_id, h_id, w_id = idx[0], idx[1], idx[2]\n\n            # Head token and offset\n            feat_central_mhmr = feat_mhmr_i[img_id, h_id, w_id] # (nvh, 1024)\n            offset = self.downstream_head.mlp_offset(feat_central_mhmr)# [nhv,2]\n            # Distance for estimating the 3D location in 3D space\n            loc = torch.stack([w_id, h_id]).permute(1,0) # x,y\n            loc = (loc + 0.5 + offset) * self.bb_patch_size # Moving to higher res the location of the pelvis\n\n            # Concat with camera embedding\n            # K = get_camera_parameters(self.mhmr_img_res, device=device) # use pseudo K\n            K = view[\"K_mhmr\"] # use GT K\n            feat_K = self.embedd_camera(K, [n_patch_mhmr, n_patch_mhmr]) # Embed viewing directions. [num_view * bs,h,w,99]\n            feat_K_central = feat_K[img_id, h_id, w_id] # (nvh, 99)\n            feat_central_mhmr = torch.cat([feat_central_mhmr, feat_K_central], 1) # feature + camera embedding for heads only to query tokens [nhv, 1123]\n            feat_all = torch.cat([feat_mhmr_i, feat_K], -1).permute(0,3,1,2) # feature + camera embedding for full image for the cross-attention only. [bs,1123,nh,nw]\n\n            # Get learned embeddings for queries, at positions with detected people.\n            queries_xy = self.cross_queries_x[h_id] + self.cross_queries_y[w_id]\n            # Add the embedding to the central features.\n            feat_central_mhmr = feat_central_mhmr + queries_xy # [nhv, 1123]\n            # Inject leared embeddings for key/values at detected locations. \n            values_xy = self.cross_values_x[h_id] + self.cross_values_y[w_id]\n            feat_all[img_id, :, h_id, w_id] += values_xy  # [bs, 1123, nh, nw]\n            feat_all = rearrange(feat_all, \"b c h w -> b (h w) c\") # (num_view * bs, nh*nw, 1024)\n\n            # Get initial smpl token from MHMR\n            expand = lambda x: x.expand(*feat_central_mhmr.shape[:-1] , -1)\n            pred_body_pose, pred_betas, pred_cam, pred_expression = [expand(x) for x in\n                    [self.downstream_head.init_body_pose, \n                    self.downstream_head.init_betas, \n                    self.downstream_head.init_cam, \n                    self.downstream_head.init_expression,\n                    ]]\n            feat_central_mhmr = torch.cat([\n                feat_central_mhmr, pred_body_pose, pred_betas, pred_cam, \n                ], dim=-1)  # training: [bs, 10, 1454]; inference: [nhv, 1454]\n\n            smpl_tk_mhmr = self.transformer(\n                feat_central_mhmr.unsqueeze(0), \n                context=feat_all, \n                mask=None) # inference:[1, nhv, 1024]\n            \n            smpl_feat_i = smpl_tk_mhmr # (1,nvh, 768)\n\n            n_humans_i = smpl_feat_i.shape[1]\n            if i == 0 :\n                state_feat, state_pos = self._init_state(feat_i, pos_i)\n                mem = self.pose_retriever.mem.expand(feat_i.shape[0], -1, -1)\n                init_state_feat = state_feat.clone()\n                init_mem = mem.clone()\n\n            if self.pose_head_flag:\n                global_img_feat_i = self._get_img_level_feat(feat_i)\n                if i == 0 or reset_mask:\n                    pose_feat_i = self.pose_token.expand(feat_i.shape[0], -1, -1)\n                else:\n                    pose_feat_i = self.pose_retriever.inquire(global_img_feat_i, mem)\n                pose_pos_i = -torch.ones(\n                    feat_i.shape[0], 1, 2, device=device, dtype=pos_i.dtype\n                )\n            else:\n                pose_feat_i = None\n                pose_pos_i = None\n            new_state_feat, dec, cross_attn_states = self._recurrent_rollout(\n                state_feat,\n                state_pos,\n                feat_i,\n                pos_i,\n                pose_feat_i,\n                pose_pos_i,\n                None,\n                None,\n                init_state_feat,\n                img_mask=view[\"img_mask\"],\n                reset_mask=view[\"reset\"],\n                update=view.get(\"update\", None),\n                use_ttt3r=use_ttt3r,\n            )\n            out_pose_feat_i = dec[-1][:, 0:1]\n            new_mem = self.pose_retriever.update_mem(\n                mem, global_img_feat_i, out_pose_feat_i\n            )\n            assert len(dec) == self.dec_depth + 1\n            head_input = [\n                dec[0].float(),\n                dec[self.dec_depth * 2 // 4][:, 1:].float(),\n                dec[self.dec_depth * 3 // 4][:, 1:].float(),\n                dec[self.dec_depth].float(),\n            ]\n            if n_humans_i > 0:\n                smpl_token = smpl_feat_i\n            else:\n                smpl_token = None\n            res = self._downstream_head(\n                head_input, shape, pos=pos_i, n_humans=n_humans_i, smpl_token=smpl_token)\n\n            # tracking\n            num_miss_match0 = 0\n            if last_smpl_tk is not None and smpl_token is not None:\n                cost_mat = -torch.cdist(last_smpl_tk, smpl_token, p=2)\n                cost_mat = log_optimal_transport(\n                    cost_mat, alpha=torch.tensor(-10.0, device=device), iters=20)\n                matches = cost_mat[:, :-1, :-1]\n                max0, max1 = matches.max(2), matches.max(1)\n                indices0, indices1 = max0.indices, max1.indices\n                mutual0 = torch.arange(\n                    indices0.shape[1], device=device\n                    )[None] == indices1.gather(1, indices0)\n                mutual1 = torch.arange(\n                    indices1.shape[1], device=device\n                    )[None] == indices0.gather(1, indices1)\n                zero = matches.new_tensor(0)\n                mscores0 = torch.where(mutual0, max0.values.exp(), zero)\n                mscores1 = torch.where(mutual1, mscores0.gather(1, indices1), zero)\n\n                match_threshold = 0.2\n                valid0 = mutual0 & (mscores0 > match_threshold) # 1,n\n                valid1 = mutual1 & valid0.gather(1, indices1) # 1,m\n                # get the final matching indices, invalid matches set to -1\n                indices0 = torch.where(valid0, indices0, indices0.new_tensor(-1)) # [1, n] current frame matches for last frame\n                indices1 = torch.where(valid1, indices1, indices1.new_tensor(-1)) # [1, m] last frame matches for current frame\n\n                smpl_id = indices1.new_full(indices1.shape, -1) # 1,m\n                valid_match1 = indices1[valid1]\n                if valid_match1.numel() > 0:\n                    smpl_id[valid1] = last_smpl_id.gather(1, valid_match1[None]).flatten()\n\n                num_miss_match0 = int((~valid0).sum())\n                num_new_persons = len(smpl_id[~valid1])\n                if num_new_persons > 0:\n                    new_ids = torch.arange(\n                        max_smpl_id + 1,\n                        max_smpl_id + 1 + num_new_persons,\n                        device=device\n                    )\n                    smpl_id[~valid1] = new_ids\n                    max_smpl_id += num_new_persons\n            else:\n                # first frame with humans\n                if smpl_token is not None:\n                    smpl_id = torch.arange(n_humans_i, device=device)[None]  # (1, nvh)\n                    max_smpl_id = n_humans_i - 1\n                else:\n                    smpl_id = None\n\n            if smpl_token is not None:\n                if num_miss_match0 > 0:\n                    miss_match_id0 = last_smpl_id[~valid0][None]\n                    miss_match_tk0 = last_smpl_tk[~valid0][None]\n                    last_smpl_id = torch.cat([smpl_id, miss_match_id0], dim=1)\n                    last_smpl_tk = torch.cat([smpl_token, miss_match_tk0], dim=1)\n                else:\n                    last_smpl_tk = smpl_token.clone()\n                    last_smpl_id = smpl_id.clone()\n\n            if smpl_id is not None:\n                res['smpl_id'] = smpl_id\n\n            res_cpu = to_cpu({**res, 'smpl_scores': scores, 'smpl_loc': loc[None]})\n            ress.append(res_cpu)\n            img_mask = view[\"img_mask\"]\n            update = view.get(\"update\", None)\n            if update is not None:\n                update_mask = (\n                    img_mask & update\n                )  # if don't update, then whatever img_mask\n            else:\n                update_mask = img_mask\n            update_mask = update_mask[:, None, None].float()\n\n\n            if use_ttt3r and i != 0 and not reset_mask:\n                cross_attn_states = rearrange(torch.cat(cross_attn_states, dim=0), 'l h nstate nimg -> 1 nstate nimg (l h)').mean(dim=(-1, -2))\n                update_mask_state = update_mask * torch.sigmoid(cross_attn_states)[..., None]\n            else:\n                update_mask_state = update_mask\n\n            state_feat = new_state_feat * update_mask_state + state_feat * (\n                1 - update_mask_state\n            )  # update global state\n            mem = new_mem * update_mask + mem * (\n                1 - update_mask\n            )  # then update local state\n            reset_mask = view[\"reset\"]\n            if reset_mask is not None:\n                reset_mask = reset_mask[:, None, None].float()\n                state_feat = init_state_feat * reset_mask + state_feat * (\n                    1 - reset_mask\n                )\n                mem = init_mem * reset_mask + mem * (1 - reset_mask)\n        if ret_state:\n            return ress, views, all_state_args\n        return ress, views\n\n\nif __name__ == \"__main__\":\n    print(ARCroco3DStereo.mro())\n    cfg = ARCroco3DStereoConfig(\n        state_size=256,\n        pos_embed=\"RoPE100\",\n        rgb_head=True,\n        pose_head=True,\n        msk_head=False,\n        img_size=(224, 224),\n        head_type=\"linear\",\n        output_mode=\"pts3d+pose\",\n        depth_mode=(\"exp\", -inf, inf),\n        conf_mode=(\"exp\", 1, inf),\n        pose_mode=(\"exp\", -inf, inf),\n        enc_embed_dim=1024,\n        enc_depth=24,\n        enc_num_heads=16,\n        dec_embed_dim=768,","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz","uri":"program://Human3R/module/src.dust3r.viz#L1-L1089","kind":"module","name":"src.dust3r.viz","path":"src/dust3r/viz.py","language":"python","start_line":1,"end_line":1089,"context_start_line":1,"context_end_line":1089,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport PIL.Image\nimport numpy as np\nfrom scipy.spatial.transform import Rotation\nimport torch\nimport cv2\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom dust3r.utils.geometry import (\n    geotrf,\n    get_med_dist_between_poses,\n    depthmap_to_absolute_camera_coordinates,\n)\nfrom dust3r.utils.device import to_numpy\nfrom dust3r.utils.image import rgb, img_to_arr\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom matplotlib.figure import Figure\n\ntry:\n    import trimesh\nexcept ImportError:\n    print(\"/!\\\\ module trimesh is not installed, cannot visualize results /!\\\\\")\n\n\ndef float2uint8(x):\n    return (255.0 * x).astype(np.uint8)\n\n\ndef uint82float(img):\n    return np.ascontiguousarray(img) / 255.0\n\n\ndef cat_3d(vecs):\n    if isinstance(vecs, (np.ndarray, torch.Tensor)):\n        vecs = [vecs]\n    return np.concatenate([p.reshape(-1, 3) for p in to_numpy(vecs)])\n\n\ndef show_raw_pointcloud(pts3d, colors, point_size=2):\n    scene = trimesh.Scene()\n\n    pct = trimesh.PointCloud(cat_3d(pts3d), colors=cat_3d(colors))\n    scene.add_geometry(pct)\n\n    scene.show(line_settings={\"point_size\": point_size})\n\n\ndef pts3d_to_trimesh(img, pts3d, valid=None):\n    H, W, THREE = img.shape\n    assert THREE == 3\n    assert img.shape == pts3d.shape\n\n    vertices = pts3d.reshape(-1, 3)\n\n    idx = np.arange(len(vertices)).reshape(H, W)\n    idx1 = idx[:-1, :-1].ravel()  # top-left corner\n    idx2 = idx[:-1, +1:].ravel()  # right-left corner\n    idx3 = idx[+1:, :-1].ravel()  # bottom-left corner\n    idx4 = idx[+1:, +1:].ravel()  # bottom-right corner\n    faces = np.concatenate(\n        (\n            np.c_[idx1, idx2, idx3],\n            np.c_[\n                idx3, idx2, idx1\n            ],  # same triangle, but backward (cheap solution to cancel face culling)\n            np.c_[idx2, idx3, idx4],\n            np.c_[\n                idx4, idx3, idx2\n            ],  # same triangle, but backward (cheap solution to cancel face culling)\n        ),\n        axis=0,\n    )\n\n    face_colors = np.concatenate(\n        (\n            img[:-1, :-1].reshape(-1, 3),\n            img[:-1, :-1].reshape(-1, 3),\n            img[+1:, +1:].reshape(-1, 3),\n            img[+1:, +1:].reshape(-1, 3),\n        ),\n        axis=0,\n    )\n\n    if valid is not None:\n        assert valid.shape == (H, W)\n        valid_idxs = valid.ravel()\n        valid_faces = valid_idxs[faces].all(axis=-1)\n        faces = faces[valid_faces]\n        face_colors = face_colors[valid_faces]\n\n    assert len(faces) == len(face_colors)\n    return dict(vertices=vertices, face_colors=face_colors, faces=faces)\n\n\ndef cat_meshes(meshes):\n    vertices, faces, colors = zip(\n        *[(m[\"vertices\"], m[\"faces\"], m[\"face_colors\"]) for m in meshes]\n    )\n    n_vertices = np.cumsum([0] + [len(v) for v in vertices])\n    for i in range(len(faces)):\n        faces[i][:] += n_vertices[i]\n\n    vertices = np.concatenate(vertices)\n    colors = np.concatenate(colors)\n    faces = np.concatenate(faces)\n    return dict(vertices=vertices, face_colors=colors, faces=faces)\n\n\ndef show_duster_pairs(view1, view2, pred1, pred2):\n    import matplotlib.pyplot as pl\n\n    pl.ion()\n\n    for e in range(len(view1[\"instance\"])):\n        i = view1[\"idx\"][e]\n        j = view2[\"idx\"][e]\n        img1 = rgb(view1[\"img\"][e])\n        img2 = rgb(view2[\"img\"][e])\n        conf1 = pred1[\"conf\"][e].squeeze()\n        conf2 = pred2[\"conf\"][e].squeeze()\n        score = conf1.mean() * conf2.mean()\n        print(f\">> Showing pair #{e} {i}-{j} {score=:g}\")\n        pl.clf()\n        pl.subplot(221).imshow(img1)\n        pl.subplot(223).imshow(img2)\n        pl.subplot(222).imshow(conf1, vmin=1, vmax=30)\n        pl.subplot(224).imshow(conf2, vmin=1, vmax=30)\n        pts1 = pred1[\"pts3d\"][e]\n        pts2 = pred2[\"pts3d_in_other_view\"][e]\n        pl.subplots_adjust(0, 0, 1, 1, 0, 0)\n        if input(\"show pointcloud? (y/n) \") == \"y\":\n            show_raw_pointcloud(cat(pts1, pts2), cat(img1, img2), point_size=5)\n\n\ndef auto_cam_size(im_poses):\n    return 0.1 * get_med_dist_between_poses(im_poses)\n\n\nclass SceneViz:\n    def __init__(self):\n        self.scene = trimesh.Scene()\n\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n        image = img_to_arr(image)\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n\n        pts3d = depthmap_to_pts3d(depth, intrinsics, cam2world=cam2world)\n\n        return self.add_pointcloud(\n            pts3d, image, mask=(depth < zfar) if mask is None else mask\n        )\n\n    def add_pointcloud(self, pts3d, color=(0, 0, 0), mask=None, denoise=False):\n        pts3d = to_numpy(pts3d)\n        mask = to_numpy(mask)\n        if not isinstance(pts3d, list):\n            pts3d = [pts3d.reshape(-1, 3)]\n            if mask is not None:\n                mask = [mask.ravel()]\n        if not isinstance(color, (tuple, list)):\n            color = [color.reshape(-1, 3)]\n        if mask is None:\n            mask = [slice(None)] * len(pts3d)\n\n        pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])\n        pct = trimesh.PointCloud(pts)\n\n        if isinstance(color, (list, np.ndarray, torch.Tensor)):\n            color = to_numpy(color)\n            col = np.concatenate([p[m] for p, m in zip(color, mask)])\n            assert col.shape == pts.shape, bb()\n            pct.visual.vertex_colors = uint8(col.reshape(-1, 3))\n        else:\n            assert len(color) == 3\n            pct.visual.vertex_colors = np.broadcast_to(uint8(color), pts.shape)\n\n        if denoise:\n\n            centroid = np.median(pct.vertices, axis=0)\n            dist_to_centroid = np.linalg.norm(pct.vertices - centroid, axis=-1)\n            dist_thr = np.quantile(dist_to_centroid, 0.99)\n            valid = dist_to_centroid < dist_thr\n\n            pct = trimesh.PointCloud(\n                pct.vertices[valid], color=pct.visual.vertex_colors[valid]\n            )\n\n        self.scene.add_geometry(pct)\n        return self\n\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n\n        pts3d, mask2 = depthmap_to_absolute_camera_coordinates(\n            depth, intrinsics, cam2world\n        )\n        mask2 &= depth < zfar\n\n        if mask is not None:\n            mask2 &= mask\n\n        return self.add_pointcloud(pts3d, image, mask=mask2)\n\n    def add_camera(\n        self,\n        pose_c2w,\n        focal=None,\n        color=(0, 0, 0),\n        image=None,\n        imsize=None,\n        cam_size=0.03,\n    ):\n        pose_c2w, focal, color, image = to_numpy((pose_c2w, focal, color, image))\n        image = img_to_arr(image)\n        if isinstance(focal, np.ndarray) and focal.shape == (3, 3):\n            intrinsics = focal\n            focal = (intrinsics[0, 0] * intrinsics[1, 1]) ** 0.5\n            if imsize is None:\n                imsize = (2 * intrinsics[0, 2], 2 * intrinsics[1, 2])\n\n        add_scene_cam(\n            self.scene,\n            pose_c2w,\n            color,\n            image,\n            focal,\n            imsize=imsize,\n            screen_width=cam_size,\n            marker=None,\n        )\n        return self\n\n    def add_cameras(\n        self, poses, focals=None, images=None, imsizes=None, colors=None, **kw\n    ):\n        get = lambda arr, idx: None if arr is None else arr[idx]\n        for i, pose_c2w in enumerate(poses):\n            self.add_camera(\n                pose_c2w,\n                get(focals, i),\n                image=get(images, i),\n                color=get(colors, i),\n                imsize=get(imsizes, i),\n                **kw,\n            )\n        return self\n\n    def show(self, point_size=2):\n        self.scene.show(line_settings={\"point_size\": point_size})\n\n\ndef show_raw_pointcloud_with_cams(\n    imgs, pts3d, mask, focals, cams2world, point_size=2, cam_size=0.05, cam_color=None\n):\n    \"\"\"Visualization of a pointcloud with cameras\n    imgs = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    pts3d = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    focals = (N,) or N-size list of [focal, ...]\n    cams2world = (N,4,4) or N-size list of [(4,4), ...]\n    \"\"\"\n    assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)\n    pts3d = to_numpy(pts3d)\n    imgs = to_numpy(imgs)\n    focals = to_numpy(focals)\n    cams2world = to_numpy(cams2world)\n\n    scene = trimesh.Scene()\n\n    pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])\n    col = np.concatenate([p[m] for p, m in zip(imgs, mask)])\n    pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3))\n    scene.add_geometry(pct)\n\n    for i, pose_c2w in enumerate(cams2world):\n        if isinstance(cam_color, list):\n            camera_edge_color = cam_color[i]\n        else:\n            camera_edge_color = cam_color or CAM_COLORS[i % len(CAM_COLORS)]\n        add_scene_cam(\n            scene,\n            pose_c2w,\n            camera_edge_color,\n            imgs[i] if i < len(imgs) else None,\n            focals[i],\n            screen_width=cam_size,\n        )\n\n    scene.show(line_settings={\"point_size\": point_size})\n\n\ndef add_scene_cam(\n    scene,\n    pose_c2w,\n    edge_color,\n    image=None,\n    focal=None,\n    imsize=None,\n    screen_width=0.03,\n    marker=None,\n):\n    if image is not None:\n        image = np.asarray(image)\n        H, W, THREE = image.shape\n        assert THREE == 3\n        if image.dtype != np.uint8:\n            image = np.uint8(255 * image)\n    elif imsize is not None:\n        W, H = imsize\n    elif focal is not None:\n        H = W = focal / 1.1\n    else:\n        H = W = 1\n\n    if isinstance(focal, np.ndarray):\n        focal = focal[0]\n    if not focal:\n        focal = min(H, W) * 1.1  # default value\n\n    height = max(screen_width / 10, focal * screen_width / H)\n    width = screen_width * 0.5**0.5\n    rot45 = np.eye(4)\n    rot45[:3, :3] = Rotation.from_euler(\"z\", np.deg2rad(45)).as_matrix()\n    rot45[2, 3] = -height  # set the tip of the cone = optical center\n    aspect_ratio = np.eye(4)\n    aspect_ratio[0, 0] = W / H\n    transform = pose_c2w @ OPENGL @ aspect_ratio @ rot45\n    cam = trimesh.creation.cone(width, height, sections=4)  # , transform=transform)\n\n    if image is not None:\n        vertices = geotrf(transform, cam.vertices[[4, 5, 1, 3]])\n        faces = np.array([[0, 1, 2], [0, 2, 3], [2, 1, 0], [3, 2, 0]])\n        img = trimesh.Trimesh(vertices=vertices, faces=faces)\n        uv_coords = np.float32([[0, 0], [1, 0], [1, 1], [0, 1]])\n        img.visual = trimesh.visual.TextureVisuals(\n            uv_coords, image=PIL.Image.fromarray(image)\n        )\n        scene.add_geometry(img)\n\n    rot2 = np.eye(4)\n    rot2[:3, :3] = Rotation.from_euler(\"z\", np.deg2rad(2)).as_matrix()\n    vertices = np.r_[cam.vertices, 0.95 * cam.vertices, geotrf(rot2, cam.vertices)]\n    vertices = geotrf(transform, vertices)\n    faces = []\n    for face in cam.faces:\n        if 0 in face:\n            continue\n        a, b, c = face\n        a2, b2, c2 = face + len(cam.vertices)\n        a3, b3, c3 = face + 2 * len(cam.vertices)\n\n        faces.append((a, b, b2))\n        faces.append((a, a2, c))\n        faces.append((c2, b, c))\n\n        faces.append((a, b, b3))\n        faces.append((a, a3, c))\n        faces.append((c3, b, c))\n\n    faces += [(c, b, a) for a, b, c in faces]\n\n    cam = trimesh.Trimesh(vertices=vertices, faces=faces)\n    cam.visual.face_colors[:, :3] = edge_color\n    scene.add_geometry(cam)\n\n    if marker == \"o\":\n        marker = trimesh.creation.icosphere(3, radius=screen_width / 4)\n        marker.vertices += pose_c2w[:3, 3]\n        marker.visual.face_colors[:, :3] = edge_color\n        scene.add_geometry(marker)\n\n\ndef cat(a, b):\n    return np.concatenate((a.reshape(-1, 3), b.reshape(-1, 3)))\n\n\nOPENGL = np.array([[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]])\n\n\nCAM_COLORS = [\n    (255, 0, 0),\n    (0, 0, 255),\n    (0, 255, 0),\n    (255, 0, 255),\n    (255, 204, 0),\n    (0, 204, 204),\n    (128, 255, 255),\n    (255, 128, 255),\n    (255, 255, 128),\n    (0, 0, 0),\n    (128, 128, 128),\n]\n\n\ndef uint8(colors):\n    if not isinstance(colors, np.ndarray):\n        colors = np.array(colors)\n    if np.issubdtype(colors.dtype, np.floating):\n        colors *= 255\n    assert 0 <= colors.min() and colors.max() < 256\n    return np.uint8(colors)\n\n\ndef segment_sky(image):\n    import cv2\n    from scipy import ndimage\n\n    image = to_numpy(image)\n    if np.issubdtype(image.dtype, np.floating):\n        image = np.uint8(255 * image.clip(min=0, max=1))\n    hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\n\n    lower_blue = np.array([0, 0, 100])\n    upper_blue = np.array([30, 255, 255])\n    mask = cv2.inRange(hsv, lower_blue, upper_blue).view(bool)\n\n    mask |= (hsv[:, :, 1] < 10) & (hsv[:, :, 2] > 150)\n    mask |= (hsv[:, :, 1] < 30) & (hsv[:, :, 2] > 180)\n    mask |= (hsv[:, :, 1] < 50) & (hsv[:, :, 2] > 220)\n\n    kernel = np.ones((5, 5), np.uint8)\n    mask2 = ndimage.binary_opening(mask, structure=kernel)\n\n    _, labels, stats, _ = cv2.connectedComponentsWithStats(\n        mask2.view(np.uint8), connectivity=8\n    )\n    cc_sizes = stats[1:, cv2.CC_STAT_AREA]\n    order = cc_sizes.argsort()[::-1]  # bigger first\n    i = 0\n    selection = []\n    while i < len(order) and cc_sizes[order[i]] > cc_sizes[order[0]] / 2:\n        selection.append(1 + order[i])\n        i += 1\n    mask3 = np.in1d(labels, selection).reshape(labels.shape)\n\n    return torch.from_numpy(mask3)\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n\n    tick_cnt = 6\n    tick_loc = np.linspace(vmin, vmax, tick_cnt)\n    cb1 = mpl.colorbar.ColorbarBase(\n        ax, cmap=cmap, norm=norm, ticks=tick_loc, orientation=\"vertical\"\n    )\n\n    tick_label = [str(np.round(x, cbar_precision)) for x in tick_loc]\n    if cbar_precision == 0:\n        tick_label = [x[:-2] for x in tick_label]\n\n    cb1.set_ticklabels(tick_label)\n\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    fig.tight_layout()\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]\n    \"\"\"\n    if range is not None:\n        vmin, vmax = range\n    elif mask is not None:\n\n        vmin = np.min(x[mask][np.nonzero(x[mask])])\n        vmax = np.max(x[mask])\n\n        x[np.logical_not(mask)] = vmin\n\n    else:\n        vmin, vmax = np.percentile(x, (1, 100))\n        vmax += 1e-6\n\n    x = np.clip(x, vmin, vmax)\n    x = (x - vmin) / (vmax - vmin)\n\n    cmap = cm.get_cmap(cmap_name)\n    x_new = cmap(x)[:, :, :3]\n\n    if mask is not None:\n        mask = np.float32(mask[:, :, np.newaxis])\n        x_new = x_new * mask + np.ones_like(x_new) * (1.0 - mask)\n\n    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out\n\n\ndef draw_correspondences(\n    imgs1, imgs2, coords1, coords2, interval=10, color_by=0, radius=2\n):\n    \"\"\"\n    draw correspondences between two images\n    :param img1: tensor [B, H, W, 3]\n    :param img2: tensor [B, H, W, 3]\n    :param coord1: tensor [B, N, 2]\n    :param coord2: tensor [B, N, 2]\n    :param interval: int the interval between two points\n    :param color_by: specify the color based on image 1 or image 2, 0 or 1\n    :return: [B, 2*H, W, 3]\n    \"\"\"\n    batch_size = len(imgs1)\n    out = []\n    for i in range(batch_size):\n        img1 = imgs1[i].detach().cpu().numpy()\n        img2 = imgs2[i].detach().cpu().numpy()\n        coord1 = (\n            coords1[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        coord2 = (\n            coords2[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        img = drawMatches(\n            img1, img2, coord1, coord2, radius=radius, color_by=color_by, row_cat=True\n        )\n        out.append(img)\n    out = np.stack(out)\n    return out\n\n\ndef draw_correspondences_lines(\n    imgs1, imgs2, coords1, coords2, interval=10, color_by=0, radius=2\n):\n    \"\"\"\n    draw correspondences between two images\n    :param img1: tensor [B, H, W, 3]\n    :param img2: tensor [B, H, W, 3]\n    :param coord1: tensor [B, N, 2]\n    :param coord2: tensor [B, N, 2]\n    :param interval: int the interval between two points\n    :param color_by: specify the color based on image 1 or image 2, 0 or 1\n    :return: [B, 2*H, W, 3]\n    \"\"\"\n    batch_size = len(imgs1)\n    out = []\n    for i in range(batch_size):\n        img1 = imgs1[i].detach().cpu().numpy()\n        img2 = imgs2[i].detach().cpu().numpy()\n        coord1 = (\n            coords1[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        coord2 = (\n            coords2[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        img = drawMatches_lines(\n            img1, img2, coord1, coord2, radius=radius, color_by=color_by, row_cat=True\n        )\n        out.append(img)\n    out = np.stack(out)\n    return out\n\n\ndef drawMatches(img1, img2, kp1, kp2, radius=2, mask=None, color_by=0, row_cat=False):\n\n    h1, w1 = img1.shape[:\n# ... truncated ...","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.float2uint8","uri":"program://Human3R/function/src.dust3r.viz.float2uint8#L31-L32","kind":"function","name":"float2uint8","path":"src/dust3r/viz.py","language":"python","start_line":31,"end_line":32,"context_start_line":11,"context_end_line":52,"code":"import cv2\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom dust3r.utils.geometry import (\n    geotrf,\n    get_med_dist_between_poses,\n    depthmap_to_absolute_camera_coordinates,\n)\nfrom dust3r.utils.device import to_numpy\nfrom dust3r.utils.image import rgb, img_to_arr\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom matplotlib.figure import Figure\n\ntry:\n    import trimesh\nexcept ImportError:\n    print(\"/!\\\\ module trimesh is not installed, cannot visualize results /!\\\\\")\n\n\ndef float2uint8(x):\n    return (255.0 * x).astype(np.uint8)\n\n\ndef uint82float(img):\n    return np.ascontiguousarray(img) / 255.0\n\n\ndef cat_3d(vecs):\n    if isinstance(vecs, (np.ndarray, torch.Tensor)):\n        vecs = [vecs]\n    return np.concatenate([p.reshape(-1, 3) for p in to_numpy(vecs)])\n\n\ndef show_raw_pointcloud(pts3d, colors, point_size=2):\n    scene = trimesh.Scene()\n\n    pct = trimesh.PointCloud(cat_3d(pts3d), colors=cat_3d(colors))\n    scene.add_geometry(pct)\n\n    scene.show(line_settings={\"point_size\": point_size})\n","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.uint82float","uri":"program://Human3R/function/src.dust3r.viz.uint82float#L35-L36","kind":"function","name":"uint82float","path":"src/dust3r/viz.py","language":"python","start_line":35,"end_line":36,"context_start_line":15,"context_end_line":56,"code":"from dust3r.utils.geometry import (\n    geotrf,\n    get_med_dist_between_poses,\n    depthmap_to_absolute_camera_coordinates,\n)\nfrom dust3r.utils.device import to_numpy\nfrom dust3r.utils.image import rgb, img_to_arr\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom matplotlib.figure import Figure\n\ntry:\n    import trimesh\nexcept ImportError:\n    print(\"/!\\\\ module trimesh is not installed, cannot visualize results /!\\\\\")\n\n\ndef float2uint8(x):\n    return (255.0 * x).astype(np.uint8)\n\n\ndef uint82float(img):\n    return np.ascontiguousarray(img) / 255.0\n\n\ndef cat_3d(vecs):\n    if isinstance(vecs, (np.ndarray, torch.Tensor)):\n        vecs = [vecs]\n    return np.concatenate([p.reshape(-1, 3) for p in to_numpy(vecs)])\n\n\ndef show_raw_pointcloud(pts3d, colors, point_size=2):\n    scene = trimesh.Scene()\n\n    pct = trimesh.PointCloud(cat_3d(pts3d), colors=cat_3d(colors))\n    scene.add_geometry(pct)\n\n    scene.show(line_settings={\"point_size\": point_size})\n\n\ndef pts3d_to_trimesh(img, pts3d, valid=None):\n    H, W, THREE = img.shape\n    assert THREE == 3","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.cat_3d","uri":"program://Human3R/function/src.dust3r.viz.cat_3d#L39-L42","kind":"function","name":"cat_3d","path":"src/dust3r/viz.py","language":"python","start_line":39,"end_line":42,"context_start_line":19,"context_end_line":62,"code":")\nfrom dust3r.utils.device import to_numpy\nfrom dust3r.utils.image import rgb, img_to_arr\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom matplotlib.figure import Figure\n\ntry:\n    import trimesh\nexcept ImportError:\n    print(\"/!\\\\ module trimesh is not installed, cannot visualize results /!\\\\\")\n\n\ndef float2uint8(x):\n    return (255.0 * x).astype(np.uint8)\n\n\ndef uint82float(img):\n    return np.ascontiguousarray(img) / 255.0\n\n\ndef cat_3d(vecs):\n    if isinstance(vecs, (np.ndarray, torch.Tensor)):\n        vecs = [vecs]\n    return np.concatenate([p.reshape(-1, 3) for p in to_numpy(vecs)])\n\n\ndef show_raw_pointcloud(pts3d, colors, point_size=2):\n    scene = trimesh.Scene()\n\n    pct = trimesh.PointCloud(cat_3d(pts3d), colors=cat_3d(colors))\n    scene.add_geometry(pct)\n\n    scene.show(line_settings={\"point_size\": point_size})\n\n\ndef pts3d_to_trimesh(img, pts3d, valid=None):\n    H, W, THREE = img.shape\n    assert THREE == 3\n    assert img.shape == pts3d.shape\n\n    vertices = pts3d.reshape(-1, 3)\n\n    idx = np.arange(len(vertices)).reshape(H, W)\n    idx1 = idx[:-1, :-1].ravel()  # top-left corner","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.show_raw_pointcloud","uri":"program://Human3R/function/src.dust3r.viz.show_raw_pointcloud#L45-L51","kind":"function","name":"show_raw_pointcloud","path":"src/dust3r/viz.py","language":"python","start_line":45,"end_line":51,"context_start_line":25,"context_end_line":71,"code":"try:\n    import trimesh\nexcept ImportError:\n    print(\"/!\\\\ module trimesh is not installed, cannot visualize results /!\\\\\")\n\n\ndef float2uint8(x):\n    return (255.0 * x).astype(np.uint8)\n\n\ndef uint82float(img):\n    return np.ascontiguousarray(img) / 255.0\n\n\ndef cat_3d(vecs):\n    if isinstance(vecs, (np.ndarray, torch.Tensor)):\n        vecs = [vecs]\n    return np.concatenate([p.reshape(-1, 3) for p in to_numpy(vecs)])\n\n\ndef show_raw_pointcloud(pts3d, colors, point_size=2):\n    scene = trimesh.Scene()\n\n    pct = trimesh.PointCloud(cat_3d(pts3d), colors=cat_3d(colors))\n    scene.add_geometry(pct)\n\n    scene.show(line_settings={\"point_size\": point_size})\n\n\ndef pts3d_to_trimesh(img, pts3d, valid=None):\n    H, W, THREE = img.shape\n    assert THREE == 3\n    assert img.shape == pts3d.shape\n\n    vertices = pts3d.reshape(-1, 3)\n\n    idx = np.arange(len(vertices)).reshape(H, W)\n    idx1 = idx[:-1, :-1].ravel()  # top-left corner\n    idx2 = idx[:-1, +1:].ravel()  # right-left corner\n    idx3 = idx[+1:, :-1].ravel()  # bottom-left corner\n    idx4 = idx[+1:, +1:].ravel()  # bottom-right corner\n    faces = np.concatenate(\n        (\n            np.c_[idx1, idx2, idx3],\n            np.c_[\n                idx3, idx2, idx1\n            ],  # same triangle, but backward (cheap solution to cancel face culling)","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.pts3d_to_trimesh","uri":"program://Human3R/function/src.dust3r.viz.pts3d_to_trimesh#L54-L98","kind":"function","name":"pts3d_to_trimesh","path":"src/dust3r/viz.py","language":"python","start_line":54,"end_line":98,"context_start_line":34,"context_end_line":118,"code":"\ndef uint82float(img):\n    return np.ascontiguousarray(img) / 255.0\n\n\ndef cat_3d(vecs):\n    if isinstance(vecs, (np.ndarray, torch.Tensor)):\n        vecs = [vecs]\n    return np.concatenate([p.reshape(-1, 3) for p in to_numpy(vecs)])\n\n\ndef show_raw_pointcloud(pts3d, colors, point_size=2):\n    scene = trimesh.Scene()\n\n    pct = trimesh.PointCloud(cat_3d(pts3d), colors=cat_3d(colors))\n    scene.add_geometry(pct)\n\n    scene.show(line_settings={\"point_size\": point_size})\n\n\ndef pts3d_to_trimesh(img, pts3d, valid=None):\n    H, W, THREE = img.shape\n    assert THREE == 3\n    assert img.shape == pts3d.shape\n\n    vertices = pts3d.reshape(-1, 3)\n\n    idx = np.arange(len(vertices)).reshape(H, W)\n    idx1 = idx[:-1, :-1].ravel()  # top-left corner\n    idx2 = idx[:-1, +1:].ravel()  # right-left corner\n    idx3 = idx[+1:, :-1].ravel()  # bottom-left corner\n    idx4 = idx[+1:, +1:].ravel()  # bottom-right corner\n    faces = np.concatenate(\n        (\n            np.c_[idx1, idx2, idx3],\n            np.c_[\n                idx3, idx2, idx1\n            ],  # same triangle, but backward (cheap solution to cancel face culling)\n            np.c_[idx2, idx3, idx4],\n            np.c_[\n                idx4, idx3, idx2\n            ],  # same triangle, but backward (cheap solution to cancel face culling)\n        ),\n        axis=0,\n    )\n\n    face_colors = np.concatenate(\n        (\n            img[:-1, :-1].reshape(-1, 3),\n            img[:-1, :-1].reshape(-1, 3),\n            img[+1:, +1:].reshape(-1, 3),\n            img[+1:, +1:].reshape(-1, 3),\n        ),\n        axis=0,\n    )\n\n    if valid is not None:\n        assert valid.shape == (H, W)\n        valid_idxs = valid.ravel()\n        valid_faces = valid_idxs[faces].all(axis=-1)\n        faces = faces[valid_faces]\n        face_colors = face_colors[valid_faces]\n\n    assert len(faces) == len(face_colors)\n    return dict(vertices=vertices, face_colors=face_colors, faces=faces)\n\n\ndef cat_meshes(meshes):\n    vertices, faces, colors = zip(\n        *[(m[\"vertices\"], m[\"faces\"], m[\"face_colors\"]) for m in meshes]\n    )\n    n_vertices = np.cumsum([0] + [len(v) for v in vertices])\n    for i in range(len(faces)):\n        faces[i][:] += n_vertices[i]\n\n    vertices = np.concatenate(vertices)\n    colors = np.concatenate(colors)\n    faces = np.concatenate(faces)\n    return dict(vertices=vertices, face_colors=colors, faces=faces)\n\n\ndef show_duster_pairs(view1, view2, pred1, pred2):\n    import matplotlib.pyplot as pl\n\n    pl.ion()","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.cat_meshes","uri":"program://Human3R/function/src.dust3r.viz.cat_meshes#L101-L112","kind":"function","name":"cat_meshes","path":"src/dust3r/viz.py","language":"python","start_line":101,"end_line":112,"context_start_line":81,"context_end_line":132,"code":"        (\n            img[:-1, :-1].reshape(-1, 3),\n            img[:-1, :-1].reshape(-1, 3),\n            img[+1:, +1:].reshape(-1, 3),\n            img[+1:, +1:].reshape(-1, 3),\n        ),\n        axis=0,\n    )\n\n    if valid is not None:\n        assert valid.shape == (H, W)\n        valid_idxs = valid.ravel()\n        valid_faces = valid_idxs[faces].all(axis=-1)\n        faces = faces[valid_faces]\n        face_colors = face_colors[valid_faces]\n\n    assert len(faces) == len(face_colors)\n    return dict(vertices=vertices, face_colors=face_colors, faces=faces)\n\n\ndef cat_meshes(meshes):\n    vertices, faces, colors = zip(\n        *[(m[\"vertices\"], m[\"faces\"], m[\"face_colors\"]) for m in meshes]\n    )\n    n_vertices = np.cumsum([0] + [len(v) for v in vertices])\n    for i in range(len(faces)):\n        faces[i][:] += n_vertices[i]\n\n    vertices = np.concatenate(vertices)\n    colors = np.concatenate(colors)\n    faces = np.concatenate(faces)\n    return dict(vertices=vertices, face_colors=colors, faces=faces)\n\n\ndef show_duster_pairs(view1, view2, pred1, pred2):\n    import matplotlib.pyplot as pl\n\n    pl.ion()\n\n    for e in range(len(view1[\"instance\"])):\n        i = view1[\"idx\"][e]\n        j = view2[\"idx\"][e]\n        img1 = rgb(view1[\"img\"][e])\n        img2 = rgb(view2[\"img\"][e])\n        conf1 = pred1[\"conf\"][e].squeeze()\n        conf2 = pred2[\"conf\"][e].squeeze()\n        score = conf1.mean() * conf2.mean()\n        print(f\">> Showing pair #{e} {i}-{j} {score=:g}\")\n        pl.clf()\n        pl.subplot(221).imshow(img1)\n        pl.subplot(223).imshow(img2)\n        pl.subplot(222).imshow(conf1, vmin=1, vmax=30)","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.show_duster_pairs","uri":"program://Human3R/function/src.dust3r.viz.show_duster_pairs#L115-L138","kind":"function","name":"show_duster_pairs","path":"src/dust3r/viz.py","language":"python","start_line":115,"end_line":138,"context_start_line":95,"context_end_line":158,"code":"        face_colors = face_colors[valid_faces]\n\n    assert len(faces) == len(face_colors)\n    return dict(vertices=vertices, face_colors=face_colors, faces=faces)\n\n\ndef cat_meshes(meshes):\n    vertices, faces, colors = zip(\n        *[(m[\"vertices\"], m[\"faces\"], m[\"face_colors\"]) for m in meshes]\n    )\n    n_vertices = np.cumsum([0] + [len(v) for v in vertices])\n    for i in range(len(faces)):\n        faces[i][:] += n_vertices[i]\n\n    vertices = np.concatenate(vertices)\n    colors = np.concatenate(colors)\n    faces = np.concatenate(faces)\n    return dict(vertices=vertices, face_colors=colors, faces=faces)\n\n\ndef show_duster_pairs(view1, view2, pred1, pred2):\n    import matplotlib.pyplot as pl\n\n    pl.ion()\n\n    for e in range(len(view1[\"instance\"])):\n        i = view1[\"idx\"][e]\n        j = view2[\"idx\"][e]\n        img1 = rgb(view1[\"img\"][e])\n        img2 = rgb(view2[\"img\"][e])\n        conf1 = pred1[\"conf\"][e].squeeze()\n        conf2 = pred2[\"conf\"][e].squeeze()\n        score = conf1.mean() * conf2.mean()\n        print(f\">> Showing pair #{e} {i}-{j} {score=:g}\")\n        pl.clf()\n        pl.subplot(221).imshow(img1)\n        pl.subplot(223).imshow(img2)\n        pl.subplot(222).imshow(conf1, vmin=1, vmax=30)\n        pl.subplot(224).imshow(conf2, vmin=1, vmax=30)\n        pts1 = pred1[\"pts3d\"][e]\n        pts2 = pred2[\"pts3d_in_other_view\"][e]\n        pl.subplots_adjust(0, 0, 1, 1, 0, 0)\n        if input(\"show pointcloud? (y/n) \") == \"y\":\n            show_raw_pointcloud(cat(pts1, pts2), cat(img1, img2), point_size=5)\n\n\ndef auto_cam_size(im_poses):\n    return 0.1 * get_med_dist_between_poses(im_poses)\n\n\nclass SceneViz:\n    def __init__(self):\n        self.scene = trimesh.Scene()\n\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n        image = img_to_arr(image)\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.auto_cam_size","uri":"program://Human3R/function/src.dust3r.viz.auto_cam_size#L141-L142","kind":"function","name":"auto_cam_size","path":"src/dust3r/viz.py","language":"python","start_line":141,"end_line":142,"context_start_line":121,"context_end_line":162,"code":"        i = view1[\"idx\"][e]\n        j = view2[\"idx\"][e]\n        img1 = rgb(view1[\"img\"][e])\n        img2 = rgb(view2[\"img\"][e])\n        conf1 = pred1[\"conf\"][e].squeeze()\n        conf2 = pred2[\"conf\"][e].squeeze()\n        score = conf1.mean() * conf2.mean()\n        print(f\">> Showing pair #{e} {i}-{j} {score=:g}\")\n        pl.clf()\n        pl.subplot(221).imshow(img1)\n        pl.subplot(223).imshow(img2)\n        pl.subplot(222).imshow(conf1, vmin=1, vmax=30)\n        pl.subplot(224).imshow(conf2, vmin=1, vmax=30)\n        pts1 = pred1[\"pts3d\"][e]\n        pts2 = pred2[\"pts3d_in_other_view\"][e]\n        pl.subplots_adjust(0, 0, 1, 1, 0, 0)\n        if input(\"show pointcloud? (y/n) \") == \"y\":\n            show_raw_pointcloud(cat(pts1, pts2), cat(img1, img2), point_size=5)\n\n\ndef auto_cam_size(im_poses):\n    return 0.1 * get_med_dist_between_poses(im_poses)\n\n\nclass SceneViz:\n    def __init__(self):\n        self.scene = trimesh.Scene()\n\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n        image = img_to_arr(image)\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n\n        pts3d = depthmap_to_pts3d(depth, intrinsics, cam2world=cam2world)\n\n        return self.add_pointcloud(\n            pts3d, image, mask=(depth < zfar) if mask is None else mask","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.SceneViz","uri":"program://Human3R/class/src.dust3r.viz.SceneViz#L145-L267","kind":"class","name":"SceneViz","path":"src/dust3r/viz.py","language":"python","start_line":145,"end_line":267,"context_start_line":125,"context_end_line":287,"code":"        conf1 = pred1[\"conf\"][e].squeeze()\n        conf2 = pred2[\"conf\"][e].squeeze()\n        score = conf1.mean() * conf2.mean()\n        print(f\">> Showing pair #{e} {i}-{j} {score=:g}\")\n        pl.clf()\n        pl.subplot(221).imshow(img1)\n        pl.subplot(223).imshow(img2)\n        pl.subplot(222).imshow(conf1, vmin=1, vmax=30)\n        pl.subplot(224).imshow(conf2, vmin=1, vmax=30)\n        pts1 = pred1[\"pts3d\"][e]\n        pts2 = pred2[\"pts3d_in_other_view\"][e]\n        pl.subplots_adjust(0, 0, 1, 1, 0, 0)\n        if input(\"show pointcloud? (y/n) \") == \"y\":\n            show_raw_pointcloud(cat(pts1, pts2), cat(img1, img2), point_size=5)\n\n\ndef auto_cam_size(im_poses):\n    return 0.1 * get_med_dist_between_poses(im_poses)\n\n\nclass SceneViz:\n    def __init__(self):\n        self.scene = trimesh.Scene()\n\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n        image = img_to_arr(image)\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n\n        pts3d = depthmap_to_pts3d(depth, intrinsics, cam2world=cam2world)\n\n        return self.add_pointcloud(\n            pts3d, image, mask=(depth < zfar) if mask is None else mask\n        )\n\n    def add_pointcloud(self, pts3d, color=(0, 0, 0), mask=None, denoise=False):\n        pts3d = to_numpy(pts3d)\n        mask = to_numpy(mask)\n        if not isinstance(pts3d, list):\n            pts3d = [pts3d.reshape(-1, 3)]\n            if mask is not None:\n                mask = [mask.ravel()]\n        if not isinstance(color, (tuple, list)):\n            color = [color.reshape(-1, 3)]\n        if mask is None:\n            mask = [slice(None)] * len(pts3d)\n\n        pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])\n        pct = trimesh.PointCloud(pts)\n\n        if isinstance(color, (list, np.ndarray, torch.Tensor)):\n            color = to_numpy(color)\n            col = np.concatenate([p[m] for p, m in zip(color, mask)])\n            assert col.shape == pts.shape, bb()\n            pct.visual.vertex_colors = uint8(col.reshape(-1, 3))\n        else:\n            assert len(color) == 3\n            pct.visual.vertex_colors = np.broadcast_to(uint8(color), pts.shape)\n\n        if denoise:\n\n            centroid = np.median(pct.vertices, axis=0)\n            dist_to_centroid = np.linalg.norm(pct.vertices - centroid, axis=-1)\n            dist_thr = np.quantile(dist_to_centroid, 0.99)\n            valid = dist_to_centroid < dist_thr\n\n            pct = trimesh.PointCloud(\n                pct.vertices[valid], color=pct.visual.vertex_colors[valid]\n            )\n\n        self.scene.add_geometry(pct)\n        return self\n\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n\n        pts3d, mask2 = depthmap_to_absolute_camera_coordinates(\n            depth, intrinsics, cam2world\n        )\n        mask2 &= depth < zfar\n\n        if mask is not None:\n            mask2 &= mask\n\n        return self.add_pointcloud(pts3d, image, mask=mask2)\n\n    def add_camera(\n        self,\n        pose_c2w,\n        focal=None,\n        color=(0, 0, 0),\n        image=None,\n        imsize=None,\n        cam_size=0.03,\n    ):\n        pose_c2w, focal, color, image = to_numpy((pose_c2w, focal, color, image))\n        image = img_to_arr(image)\n        if isinstance(focal, np.ndarray) and focal.shape == (3, 3):\n            intrinsics = focal\n            focal = (intrinsics[0, 0] * intrinsics[1, 1]) ** 0.5\n            if imsize is None:\n                imsize = (2 * intrinsics[0, 2], 2 * intrinsics[1, 2])\n\n        add_scene_cam(\n            self.scene,\n            pose_c2w,\n            color,\n            image,\n            focal,\n            imsize=imsize,\n            screen_width=cam_size,\n            marker=None,\n        )\n        return self\n\n    def add_cameras(\n        self, poses, focals=None, images=None, imsizes=None, colors=None, **kw\n    ):\n        get = lambda arr, idx: None if arr is None else arr[idx]\n        for i, pose_c2w in enumerate(poses):\n            self.add_camera(\n                pose_c2w,\n                get(focals, i),\n                image=get(images, i),\n                color=get(colors, i),\n                imsize=get(imsizes, i),\n                **kw,\n            )\n        return self\n\n    def show(self, point_size=2):\n        self.scene.show(line_settings={\"point_size\": point_size})\n\n\ndef show_raw_pointcloud_with_cams(\n    imgs, pts3d, mask, focals, cams2world, point_size=2, cam_size=0.05, cam_color=None\n):\n    \"\"\"Visualization of a pointcloud with cameras\n    imgs = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    pts3d = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    focals = (N,) or N-size list of [focal, ...]\n    cams2world = (N,4,4) or N-size list of [(4,4), ...]\n    \"\"\"\n    assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)\n    pts3d = to_numpy(pts3d)\n    imgs = to_numpy(imgs)\n    focals = to_numpy(focals)\n    cams2world = to_numpy(cams2world)\n\n    scene = trimesh.Scene()\n\n    pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.show_raw_pointcloud_with_cams","uri":"program://Human3R/function/src.dust3r.viz.show_raw_pointcloud_with_cams#L270-L306","kind":"function","name":"show_raw_pointcloud_with_cams","path":"src/dust3r/viz.py","language":"python","start_line":270,"end_line":306,"context_start_line":250,"context_end_line":326,"code":"\n    def add_cameras(\n        self, poses, focals=None, images=None, imsizes=None, colors=None, **kw\n    ):\n        get = lambda arr, idx: None if arr is None else arr[idx]\n        for i, pose_c2w in enumerate(poses):\n            self.add_camera(\n                pose_c2w,\n                get(focals, i),\n                image=get(images, i),\n                color=get(colors, i),\n                imsize=get(imsizes, i),\n                **kw,\n            )\n        return self\n\n    def show(self, point_size=2):\n        self.scene.show(line_settings={\"point_size\": point_size})\n\n\ndef show_raw_pointcloud_with_cams(\n    imgs, pts3d, mask, focals, cams2world, point_size=2, cam_size=0.05, cam_color=None\n):\n    \"\"\"Visualization of a pointcloud with cameras\n    imgs = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    pts3d = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    focals = (N,) or N-size list of [focal, ...]\n    cams2world = (N,4,4) or N-size list of [(4,4), ...]\n    \"\"\"\n    assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)\n    pts3d = to_numpy(pts3d)\n    imgs = to_numpy(imgs)\n    focals = to_numpy(focals)\n    cams2world = to_numpy(cams2world)\n\n    scene = trimesh.Scene()\n\n    pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])\n    col = np.concatenate([p[m] for p, m in zip(imgs, mask)])\n    pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3))\n    scene.add_geometry(pct)\n\n    for i, pose_c2w in enumerate(cams2world):\n        if isinstance(cam_color, list):\n            camera_edge_color = cam_color[i]\n        else:\n            camera_edge_color = cam_color or CAM_COLORS[i % len(CAM_COLORS)]\n        add_scene_cam(\n            scene,\n            pose_c2w,\n            camera_edge_color,\n            imgs[i] if i < len(imgs) else None,\n            focals[i],\n            screen_width=cam_size,\n        )\n\n    scene.show(line_settings={\"point_size\": point_size})\n\n\ndef add_scene_cam(\n    scene,\n    pose_c2w,\n    edge_color,\n    image=None,\n    focal=None,\n    imsize=None,\n    screen_width=0.03,\n    marker=None,\n):\n    if image is not None:\n        image = np.asarray(image)\n        H, W, THREE = image.shape\n        assert THREE == 3\n        if image.dtype != np.uint8:\n            image = np.uint8(255 * image)\n    elif imsize is not None:\n        W, H = imsize","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.add_scene_cam","uri":"program://Human3R/function/src.dust3r.viz.add_scene_cam#L309-L387","kind":"function","name":"add_scene_cam","path":"src/dust3r/viz.py","language":"python","start_line":309,"end_line":387,"context_start_line":289,"context_end_line":407,"code":"    pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3))\n    scene.add_geometry(pct)\n\n    for i, pose_c2w in enumerate(cams2world):\n        if isinstance(cam_color, list):\n            camera_edge_color = cam_color[i]\n        else:\n            camera_edge_color = cam_color or CAM_COLORS[i % len(CAM_COLORS)]\n        add_scene_cam(\n            scene,\n            pose_c2w,\n            camera_edge_color,\n            imgs[i] if i < len(imgs) else None,\n            focals[i],\n            screen_width=cam_size,\n        )\n\n    scene.show(line_settings={\"point_size\": point_size})\n\n\ndef add_scene_cam(\n    scene,\n    pose_c2w,\n    edge_color,\n    image=None,\n    focal=None,\n    imsize=None,\n    screen_width=0.03,\n    marker=None,\n):\n    if image is not None:\n        image = np.asarray(image)\n        H, W, THREE = image.shape\n        assert THREE == 3\n        if image.dtype != np.uint8:\n            image = np.uint8(255 * image)\n    elif imsize is not None:\n        W, H = imsize\n    elif focal is not None:\n        H = W = focal / 1.1\n    else:\n        H = W = 1\n\n    if isinstance(focal, np.ndarray):\n        focal = focal[0]\n    if not focal:\n        focal = min(H, W) * 1.1  # default value\n\n    height = max(screen_width / 10, focal * screen_width / H)\n    width = screen_width * 0.5**0.5\n    rot45 = np.eye(4)\n    rot45[:3, :3] = Rotation.from_euler(\"z\", np.deg2rad(45)).as_matrix()\n    rot45[2, 3] = -height  # set the tip of the cone = optical center\n    aspect_ratio = np.eye(4)\n    aspect_ratio[0, 0] = W / H\n    transform = pose_c2w @ OPENGL @ aspect_ratio @ rot45\n    cam = trimesh.creation.cone(width, height, sections=4)  # , transform=transform)\n\n    if image is not None:\n        vertices = geotrf(transform, cam.vertices[[4, 5, 1, 3]])\n        faces = np.array([[0, 1, 2], [0, 2, 3], [2, 1, 0], [3, 2, 0]])\n        img = trimesh.Trimesh(vertices=vertices, faces=faces)\n        uv_coords = np.float32([[0, 0], [1, 0], [1, 1], [0, 1]])\n        img.visual = trimesh.visual.TextureVisuals(\n            uv_coords, image=PIL.Image.fromarray(image)\n        )\n        scene.add_geometry(img)\n\n    rot2 = np.eye(4)\n    rot2[:3, :3] = Rotation.from_euler(\"z\", np.deg2rad(2)).as_matrix()\n    vertices = np.r_[cam.vertices, 0.95 * cam.vertices, geotrf(rot2, cam.vertices)]\n    vertices = geotrf(transform, vertices)\n    faces = []\n    for face in cam.faces:\n        if 0 in face:\n            continue\n        a, b, c = face\n        a2, b2, c2 = face + len(cam.vertices)\n        a3, b3, c3 = face + 2 * len(cam.vertices)\n\n        faces.append((a, b, b2))\n        faces.append((a, a2, c))\n        faces.append((c2, b, c))\n\n        faces.append((a, b, b3))\n        faces.append((a, a3, c))\n        faces.append((c3, b, c))\n\n    faces += [(c, b, a) for a, b, c in faces]\n\n    cam = trimesh.Trimesh(vertices=vertices, faces=faces)\n    cam.visual.face_colors[:, :3] = edge_color\n    scene.add_geometry(cam)\n\n    if marker == \"o\":\n        marker = trimesh.creation.icosphere(3, radius=screen_width / 4)\n        marker.vertices += pose_c2w[:3, 3]\n        marker.visual.face_colors[:, :3] = edge_color\n        scene.add_geometry(marker)\n\n\ndef cat(a, b):\n    return np.concatenate((a.reshape(-1, 3), b.reshape(-1, 3)))\n\n\nOPENGL = np.array([[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]])\n\n\nCAM_COLORS = [\n    (255, 0, 0),\n    (0, 0, 255),\n    (0, 255, 0),\n    (255, 0, 255),\n    (255, 204, 0),\n    (0, 204, 204),\n    (128, 255, 255),\n    (255, 128, 255),\n    (255, 255, 128),\n    (0, 0, 0),","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.cat","uri":"program://Human3R/function/src.dust3r.viz.cat#L390-L391","kind":"function","name":"cat","path":"src/dust3r/viz.py","language":"python","start_line":390,"end_line":391,"context_start_line":370,"context_end_line":411,"code":"        faces.append((a, a2, c))\n        faces.append((c2, b, c))\n\n        faces.append((a, b, b3))\n        faces.append((a, a3, c))\n        faces.append((c3, b, c))\n\n    faces += [(c, b, a) for a, b, c in faces]\n\n    cam = trimesh.Trimesh(vertices=vertices, faces=faces)\n    cam.visual.face_colors[:, :3] = edge_color\n    scene.add_geometry(cam)\n\n    if marker == \"o\":\n        marker = trimesh.creation.icosphere(3, radius=screen_width / 4)\n        marker.vertices += pose_c2w[:3, 3]\n        marker.visual.face_colors[:, :3] = edge_color\n        scene.add_geometry(marker)\n\n\ndef cat(a, b):\n    return np.concatenate((a.reshape(-1, 3), b.reshape(-1, 3)))\n\n\nOPENGL = np.array([[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]])\n\n\nCAM_COLORS = [\n    (255, 0, 0),\n    (0, 0, 255),\n    (0, 255, 0),\n    (255, 0, 255),\n    (255, 204, 0),\n    (0, 204, 204),\n    (128, 255, 255),\n    (255, 128, 255),\n    (255, 255, 128),\n    (0, 0, 0),\n    (128, 128, 128),\n]\n\n","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.uint8","uri":"program://Human3R/function/src.dust3r.viz.uint8#L412-L418","kind":"function","name":"uint8","path":"src/dust3r/viz.py","language":"python","start_line":412,"end_line":418,"context_start_line":392,"context_end_line":438,"code":"\n\nOPENGL = np.array([[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]])\n\n\nCAM_COLORS = [\n    (255, 0, 0),\n    (0, 0, 255),\n    (0, 255, 0),\n    (255, 0, 255),\n    (255, 204, 0),\n    (0, 204, 204),\n    (128, 255, 255),\n    (255, 128, 255),\n    (255, 255, 128),\n    (0, 0, 0),\n    (128, 128, 128),\n]\n\n\ndef uint8(colors):\n    if not isinstance(colors, np.ndarray):\n        colors = np.array(colors)\n    if np.issubdtype(colors.dtype, np.floating):\n        colors *= 255\n    assert 0 <= colors.min() and colors.max() < 256\n    return np.uint8(colors)\n\n\ndef segment_sky(image):\n    import cv2\n    from scipy import ndimage\n\n    image = to_numpy(image)\n    if np.issubdtype(image.dtype, np.floating):\n        image = np.uint8(255 * image.clip(min=0, max=1))\n    hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\n\n    lower_blue = np.array([0, 0, 100])\n    upper_blue = np.array([30, 255, 255])\n    mask = cv2.inRange(hsv, lower_blue, upper_blue).view(bool)\n\n    mask |= (hsv[:, :, 1] < 10) & (hsv[:, :, 2] > 150)\n    mask |= (hsv[:, :, 1] < 30) & (hsv[:, :, 2] > 180)\n    mask |= (hsv[:, :, 1] < 50) & (hsv[:, :, 2] > 220)\n\n    kernel = np.ones((5, 5), np.uint8)","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.segment_sky","uri":"program://Human3R/function/src.dust3r.viz.segment_sky#L421-L453","kind":"function","name":"segment_sky","path":"src/dust3r/viz.py","language":"python","start_line":421,"end_line":453,"context_start_line":401,"context_end_line":473,"code":"    (255, 0, 255),\n    (255, 204, 0),\n    (0, 204, 204),\n    (128, 255, 255),\n    (255, 128, 255),\n    (255, 255, 128),\n    (0, 0, 0),\n    (128, 128, 128),\n]\n\n\ndef uint8(colors):\n    if not isinstance(colors, np.ndarray):\n        colors = np.array(colors)\n    if np.issubdtype(colors.dtype, np.floating):\n        colors *= 255\n    assert 0 <= colors.min() and colors.max() < 256\n    return np.uint8(colors)\n\n\ndef segment_sky(image):\n    import cv2\n    from scipy import ndimage\n\n    image = to_numpy(image)\n    if np.issubdtype(image.dtype, np.floating):\n        image = np.uint8(255 * image.clip(min=0, max=1))\n    hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\n\n    lower_blue = np.array([0, 0, 100])\n    upper_blue = np.array([30, 255, 255])\n    mask = cv2.inRange(hsv, lower_blue, upper_blue).view(bool)\n\n    mask |= (hsv[:, :, 1] < 10) & (hsv[:, :, 2] > 150)\n    mask |= (hsv[:, :, 1] < 30) & (hsv[:, :, 2] > 180)\n    mask |= (hsv[:, :, 1] < 50) & (hsv[:, :, 2] > 220)\n\n    kernel = np.ones((5, 5), np.uint8)\n    mask2 = ndimage.binary_opening(mask, structure=kernel)\n\n    _, labels, stats, _ = cv2.connectedComponentsWithStats(\n        mask2.view(np.uint8), connectivity=8\n    )\n    cc_sizes = stats[1:, cv2.CC_STAT_AREA]\n    order = cc_sizes.argsort()[::-1]  # bigger first\n    i = 0\n    selection = []\n    while i < len(order) and cc_sizes[order[i]] > cc_sizes[order[0]] / 2:\n        selection.append(1 + order[i])\n        i += 1\n    mask3 = np.in1d(labels, selection).reshape(labels.shape)\n\n    return torch.from_numpy(mask3)\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.get_vertical_colorbar","uri":"program://Human3R/function/src.dust3r.viz.get_vertical_colorbar#L456-L502","kind":"function","name":"get_vertical_colorbar","path":"src/dust3r/viz.py","language":"python","start_line":456,"end_line":502,"context_start_line":436,"context_end_line":522,"code":"    mask |= (hsv[:, :, 1] < 50) & (hsv[:, :, 2] > 220)\n\n    kernel = np.ones((5, 5), np.uint8)\n    mask2 = ndimage.binary_opening(mask, structure=kernel)\n\n    _, labels, stats, _ = cv2.connectedComponentsWithStats(\n        mask2.view(np.uint8), connectivity=8\n    )\n    cc_sizes = stats[1:, cv2.CC_STAT_AREA]\n    order = cc_sizes.argsort()[::-1]  # bigger first\n    i = 0\n    selection = []\n    while i < len(order) and cc_sizes[order[i]] > cc_sizes[order[0]] / 2:\n        selection.append(1 + order[i])\n        i += 1\n    mask3 = np.in1d(labels, selection).reshape(labels.shape)\n\n    return torch.from_numpy(mask3)\n\n\ndef get_vertical_colorbar(h, vmin, vmax, cmap_name=\"jet\", label=None, cbar_precision=2):\n    \"\"\"\n    :param w: pixels\n    :param h: pixels\n    :param vmin: min value\n    :param vmax: max value\n    :param cmap_name:\n    :param label\n    :return:\n    \"\"\"\n    fig = Figure(figsize=(2, 8), dpi=100)\n    fig.subplots_adjust(right=1.5)\n    canvas = FigureCanvasAgg(fig)\n\n    ax = fig.add_subplot(111)\n    cmap = cm.get_cmap(cmap_name)\n    norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)\n\n    tick_cnt = 6\n    tick_loc = np.linspace(vmin, vmax, tick_cnt)\n    cb1 = mpl.colorbar.ColorbarBase(\n        ax, cmap=cmap, norm=norm, ticks=tick_loc, orientation=\"vertical\"\n    )\n\n    tick_label = [str(np.round(x, cbar_precision)) for x in tick_loc]\n    if cbar_precision == 0:\n        tick_label = [x[:-2] for x in tick_label]\n\n    cb1.set_ticklabels(tick_label)\n\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    fig.tight_layout()\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.colorize_np","uri":"program://Human3R/function/src.dust3r.viz.colorize_np#L505-L564","kind":"function","name":"colorize_np","path":"src/dust3r/viz.py","language":"python","start_line":505,"end_line":564,"context_start_line":485,"context_end_line":584,"code":"\n    cb1.ax.tick_params(labelsize=18, rotation=0)\n    if label is not None:\n        cb1.set_label(label)\n\n    fig.tight_layout()\n\n    canvas.draw()\n    s, (width, height) = canvas.print_to_buffer()\n\n    im = np.frombuffer(s, np.uint8).reshape((height, width, 4))\n\n    im = im[:, :, :3].astype(np.float32) / 255.0\n    if h != im.shape[0]:\n        w = int(im.shape[1] / im.shape[0] * h)\n        im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA)\n\n    return im\n\n\ndef colorize_np(\n    x,\n    cmap_name=\"jet\",\n    mask=None,\n    range=None,\n    append_cbar=False,\n    cbar_in_image=False,\n    cbar_precision=2,\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: input grayscale, [H, W]\n    :param cmap_name: the colorization method\n    :param mask: the mask image, [H, W]\n    :param range: the range for scaling, automatic if None, [min, max]\n    :param append_cbar: if append the color bar\n    :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image\n    :return: colorized image, [H, W]\n    \"\"\"\n    if range is not None:\n        vmin, vmax = range\n    elif mask is not None:\n\n        vmin = np.min(x[mask][np.nonzero(x[mask])])\n        vmax = np.max(x[mask])\n\n        x[np.logical_not(mask)] = vmin\n\n    else:\n        vmin, vmax = np.percentile(x, (1, 100))\n        vmax += 1e-6\n\n    x = np.clip(x, vmin, vmax)\n    x = (x - vmin) / (vmax - vmin)\n\n    cmap = cm.get_cmap(cmap_name)\n    x_new = cmap(x)[:, :, :3]\n\n    if mask is not None:\n        mask = np.float32(mask[:, :, np.newaxis])\n        x_new = x_new * mask + np.ones_like(x_new) * (1.0 - mask)\n\n    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.colorize","uri":"program://Human3R/function/src.dust3r.viz.colorize#L567-L595","kind":"function","name":"colorize","path":"src/dust3r/viz.py","language":"python","start_line":567,"end_line":595,"context_start_line":547,"context_end_line":615,"code":"    cbar = get_vertical_colorbar(\n        h=x.shape[0],\n        vmin=vmin,\n        vmax=vmax,\n        cmap_name=cmap_name,\n        cbar_precision=cbar_precision,\n    )\n\n    if append_cbar:\n        if cbar_in_image:\n            x_new[:, -cbar.shape[1] :, :] = cbar\n        else:\n            x_new = np.concatenate(\n                (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1\n            )\n        return x_new\n    else:\n        return x_new\n\n\ndef colorize(\n    x, cmap_name=\"jet\", mask=None, range=None, append_cbar=False, cbar_in_image=False\n):\n    \"\"\"\n    turn a grayscale image into a color image\n    :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W]\n    :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None\n    \"\"\"\n\n    device = x.device\n    x = x.cpu().numpy()\n    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out\n\n\ndef draw_correspondences(\n    imgs1, imgs2, coords1, coords2, interval=10, color_by=0, radius=2\n):\n    \"\"\"\n    draw correspondences between two images\n    :param img1: tensor [B, H, W, 3]\n    :param img2: tensor [B, H, W, 3]\n    :param coord1: tensor [B, N, 2]\n    :param coord2: tensor [B, N, 2]\n    :param interval: int the interval between two points\n    :param color_by: specify the color based on image 1 or image 2, 0 or 1\n    :return: [B, 2*H, W, 3]\n    \"\"\"\n    batch_size = len(imgs1)\n    out = []\n    for i in range(batch_size):\n        img1 = imgs1[i].detach().cpu().numpy()\n        img2 = imgs2[i].detach().cpu().numpy()","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.draw_correspondences","uri":"program://Human3R/function/src.dust3r.viz.draw_correspondences#L598-L627","kind":"function","name":"draw_correspondences","path":"src/dust3r/viz.py","language":"python","start_line":598,"end_line":627,"context_start_line":578,"context_end_line":647,"code":"    if mask is not None:\n        mask = mask.cpu().numpy() > 0.99\n        kernel = np.ones((3, 3), np.uint8)\n\n    if x.ndim == 2:\n        x = x[None]\n        if mask is not None:\n            mask = mask[None]\n\n    out = []\n    for x_ in x:\n        if mask is not None:\n            mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool)\n\n        x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image)\n        out.append(torch.from_numpy(x_).to(device).float())\n    out = torch.stack(out).squeeze(0)\n    return out\n\n\ndef draw_correspondences(\n    imgs1, imgs2, coords1, coords2, interval=10, color_by=0, radius=2\n):\n    \"\"\"\n    draw correspondences between two images\n    :param img1: tensor [B, H, W, 3]\n    :param img2: tensor [B, H, W, 3]\n    :param coord1: tensor [B, N, 2]\n    :param coord2: tensor [B, N, 2]\n    :param interval: int the interval between two points\n    :param color_by: specify the color based on image 1 or image 2, 0 or 1\n    :return: [B, 2*H, W, 3]\n    \"\"\"\n    batch_size = len(imgs1)\n    out = []\n    for i in range(batch_size):\n        img1 = imgs1[i].detach().cpu().numpy()\n        img2 = imgs2[i].detach().cpu().numpy()\n        coord1 = (\n            coords1[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        coord2 = (\n            coords2[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        img = drawMatches(\n            img1, img2, coord1, coord2, radius=radius, color_by=color_by, row_cat=True\n        )\n        out.append(img)\n    out = np.stack(out)\n    return out\n\n\ndef draw_correspondences_lines(\n    imgs1, imgs2, coords1, coords2, interval=10, color_by=0, radius=2\n):\n    \"\"\"\n    draw correspondences between two images\n    :param img1: tensor [B, H, W, 3]\n    :param img2: tensor [B, H, W, 3]\n    :param coord1: tensor [B, N, 2]\n    :param coord2: tensor [B, N, 2]\n    :param interval: int the interval between two points\n    :param color_by: specify the color based on image 1 or image 2, 0 or 1\n    :return: [B, 2*H, W, 3]\n    \"\"\"\n    batch_size = len(imgs1)\n    out = []\n    for i in range(batch_size):\n        img1 = imgs1[i].detach().cpu().numpy()\n        img2 = imgs2[i].detach().cpu().numpy()","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.draw_correspondences_lines","uri":"program://Human3R/function/src.dust3r.viz.draw_correspondences_lines#L630-L659","kind":"function","name":"draw_correspondences_lines","path":"src/dust3r/viz.py","language":"python","start_line":630,"end_line":659,"context_start_line":610,"context_end_line":679,"code":"    \"\"\"\n    batch_size = len(imgs1)\n    out = []\n    for i in range(batch_size):\n        img1 = imgs1[i].detach().cpu().numpy()\n        img2 = imgs2[i].detach().cpu().numpy()\n        coord1 = (\n            coords1[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        coord2 = (\n            coords2[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        img = drawMatches(\n            img1, img2, coord1, coord2, radius=radius, color_by=color_by, row_cat=True\n        )\n        out.append(img)\n    out = np.stack(out)\n    return out\n\n\ndef draw_correspondences_lines(\n    imgs1, imgs2, coords1, coords2, interval=10, color_by=0, radius=2\n):\n    \"\"\"\n    draw correspondences between two images\n    :param img1: tensor [B, H, W, 3]\n    :param img2: tensor [B, H, W, 3]\n    :param coord1: tensor [B, N, 2]\n    :param coord2: tensor [B, N, 2]\n    :param interval: int the interval between two points\n    :param color_by: specify the color based on image 1 or image 2, 0 or 1\n    :return: [B, 2*H, W, 3]\n    \"\"\"\n    batch_size = len(imgs1)\n    out = []\n    for i in range(batch_size):\n        img1 = imgs1[i].detach().cpu().numpy()\n        img2 = imgs2[i].detach().cpu().numpy()\n        coord1 = (\n            coords1[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        coord2 = (\n            coords2[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        img = drawMatches_lines(\n            img1, img2, coord1, coord2, radius=radius, color_by=color_by, row_cat=True\n        )\n        out.append(img)\n    out = np.stack(out)\n    return out\n\n\ndef drawMatches(img1, img2, kp1, kp2, radius=2, mask=None, color_by=0, row_cat=False):\n\n    h1, w1 = img1.shape[:2]\n    h2, w2 = img2.shape[:2]\n\n    img1 = np.ascontiguousarray(float2uint8(img1))\n    img2 = np.ascontiguousarray(float2uint8(img2))\n\n    center1 = np.median(kp1, axis=0)\n    center2 = np.median(kp2, axis=0)\n\n    set_max = range(128)\n    colors = {m: i for i, m in enumerate(set_max)}\n    colors = {\n        m: (255 * np.array(plt.cm.hsv(i / float(len(colors))))[:3][::-1]).astype(\n            np.int32\n        )\n        for m, i in colors.items()","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.drawMatches","uri":"program://Human3R/function/src.dust3r.viz.drawMatches#L662-L735","kind":"function","name":"drawMatches","path":"src/dust3r/viz.py","language":"python","start_line":662,"end_line":735,"context_start_line":642,"context_end_line":755,"code":"    \"\"\"\n    batch_size = len(imgs1)\n    out = []\n    for i in range(batch_size):\n        img1 = imgs1[i].detach().cpu().numpy()\n        img2 = imgs2[i].detach().cpu().numpy()\n        coord1 = (\n            coords1[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        coord2 = (\n            coords2[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2)\n        )\n        img = drawMatches_lines(\n            img1, img2, coord1, coord2, radius=radius, color_by=color_by, row_cat=True\n        )\n        out.append(img)\n    out = np.stack(out)\n    return out\n\n\ndef drawMatches(img1, img2, kp1, kp2, radius=2, mask=None, color_by=0, row_cat=False):\n\n    h1, w1 = img1.shape[:2]\n    h2, w2 = img2.shape[:2]\n\n    img1 = np.ascontiguousarray(float2uint8(img1))\n    img2 = np.ascontiguousarray(float2uint8(img2))\n\n    center1 = np.median(kp1, axis=0)\n    center2 = np.median(kp2, axis=0)\n\n    set_max = range(128)\n    colors = {m: i for i, m in enumerate(set_max)}\n    colors = {\n        m: (255 * np.array(plt.cm.hsv(i / float(len(colors))))[:3][::-1]).astype(\n            np.int32\n        )\n        for m, i in colors.items()\n    }\n\n    if mask is not None:\n        ind = np.argsort(mask)[::-1]\n        kp1 = kp1[ind]\n        kp2 = kp2[ind]\n        mask = mask[ind]\n\n    for i, (pt1, pt2) in enumerate(zip(kp1, kp2)):\n\n        if color_by == 0:\n            coord_angle = np.arctan2(pt1[1] - center1[1], pt1[0] - center1[0])\n        elif color_by == 1:\n            coord_angle = np.arctan2(pt2[1] - center2[1], pt2[0] - center2[0])\n\n        corr_color = np.int32(64 * coord_angle / np.pi) % 128\n        color = tuple(colors[corr_color].tolist())\n\n        if (\n            (pt1[0] <= w1 - 1)\n            and (pt1[0] >= 0)\n            and (pt1[1] <= h1 - 1)\n            and (pt1[1] >= 0)\n        ):\n            img1 = cv2.circle(\n                img1, (int(pt1[0]), int(pt1[1])), radius, color, -1, cv2.LINE_AA\n            )\n\n        if (\n            (pt2[0] <= w2 - 1)\n            and (pt2[0] >= 0)\n            and (pt2[1] <= h2 - 1)\n            and (pt2[1] >= 0)\n        ):\n            if mask is not None and mask[i]:\n                img2 = cv2.drawMarker(\n                    img2,\n                    (int(pt2[0]), int(pt2[1])),\n                    color,\n                    markerType=cv2.MARKER_CROSS,\n                    markerSize=int(5 * radius),\n                    thickness=int(radius / 2),\n                    line_type=cv2.LINE_AA,\n                )\n            else:\n                img2 = cv2.circle(\n                    img2, (int(pt2[0]), int(pt2[1])), radius, color, -1, cv2.LINE_AA\n                )\n    if row_cat:\n        whole_img = np.concatenate([img1, img2], axis=0)\n    else:\n        whole_img = np.concatenate([img1, img2], axis=1)\n    return whole_img\n    if row_cat:\n        return np.concatenate([img1, img2], axis=0)\n    return np.concatenate([img1, img2], axis=1)\n\n\ndef drawMatches_lines(\n    img1, img2, kp1, kp2, radius=2, mask=None, color_by=0, row_cat=False\n):\n\n    h1, w1 = img1.shape[:2]\n    h2, w2 = img2.shape[:2]\n\n    img1 = np.ascontiguousarray(float2uint8(img1))\n    img2 = np.ascontiguousarray(float2uint8(img2))\n\n    center1 = np.median(kp1, axis=0)\n    center2 = np.median(kp2, axis=0)\n\n    set_max = range(128)\n    colors = {m: i for i, m in enumerate(set_max)}\n    colors = {\n        m: (255 * np.array(plt.cm.hsv(i / float(len(colors))))[:3][::-1]).astype(\n            np.int32","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.drawMatches_lines","uri":"program://Human3R/function/src.dust3r.viz.drawMatches_lines#L738-L838","kind":"function","name":"drawMatches_lines","path":"src/dust3r/viz.py","language":"python","start_line":738,"end_line":838,"context_start_line":718,"context_end_line":858,"code":"                    color,\n                    markerType=cv2.MARKER_CROSS,\n                    markerSize=int(5 * radius),\n                    thickness=int(radius / 2),\n                    line_type=cv2.LINE_AA,\n                )\n            else:\n                img2 = cv2.circle(\n                    img2, (int(pt2[0]), int(pt2[1])), radius, color, -1, cv2.LINE_AA\n                )\n    if row_cat:\n        whole_img = np.concatenate([img1, img2], axis=0)\n    else:\n        whole_img = np.concatenate([img1, img2], axis=1)\n    return whole_img\n    if row_cat:\n        return np.concatenate([img1, img2], axis=0)\n    return np.concatenate([img1, img2], axis=1)\n\n\ndef drawMatches_lines(\n    img1, img2, kp1, kp2, radius=2, mask=None, color_by=0, row_cat=False\n):\n\n    h1, w1 = img1.shape[:2]\n    h2, w2 = img2.shape[:2]\n\n    img1 = np.ascontiguousarray(float2uint8(img1))\n    img2 = np.ascontiguousarray(float2uint8(img2))\n\n    center1 = np.median(kp1, axis=0)\n    center2 = np.median(kp2, axis=0)\n\n    set_max = range(128)\n    colors = {m: i for i, m in enumerate(set_max)}\n    colors = {\n        m: (255 * np.array(plt.cm.hsv(i / float(len(colors))))[:3][::-1]).astype(\n            np.int32\n        )\n        for m, i in colors.items()\n    }\n\n    if mask is not None:\n        ind = np.argsort(mask)[::-1]\n        kp1 = kp1[ind]\n        kp2 = kp2[ind]\n        mask = mask[ind]\n\n    if row_cat:\n        whole_img = np.concatenate([img1, img2], axis=0)\n    else:\n        whole_img = np.concatenate([img1, img2], axis=1)\n    for i, (pt1, pt2) in enumerate(zip(kp1, kp2)):\n        if color_by == 0:\n            coord_angle = np.arctan2(pt1[1] - center1[1], pt1[0] - center1[0])\n        elif color_by == 1:\n            coord_angle = np.arctan2(pt2[1] - center2[1], pt2[0] - center2[0])\n\n        corr_color = np.int32(64 * coord_angle / np.pi) % 128\n        color = tuple(colors[corr_color].tolist())\n        rand_val = np.random.rand()\n        if rand_val < 0.1:\n            if (\n                (pt1[0] <= w1 - 1)\n                and (pt1[0] >= 0)\n                and (pt1[1] <= h1 - 1)\n                and (pt1[1] >= 0)\n            ) and (\n                (pt2[0] <= w2 - 1)\n                and (pt2[0] >= 0)\n                and (pt2[1] <= h2 - 1)\n                and (pt2[1] >= 0)\n            ):\n\n                whole_img = cv2.circle(\n                    whole_img,\n                    (int(pt1[0]), int(pt1[1])),\n                    radius,\n                    color,\n                    -1,\n                    cv2.LINE_AA,\n                )\n\n                if row_cat:\n                    whole_img = cv2.circle(\n                        whole_img,\n                        (int(pt2[0]), int(pt2[1] + h1)),\n                        radius,\n                        color,\n                        -1,\n                        cv2.LINE_AA,\n                    )\n                    cv2.line(\n                        whole_img,\n                        (int(pt1[0]), int(pt1[1])),\n                        (int(pt2[0]), int(pt2[1] + h1)),\n                        color,\n                        1,\n                        cv2.LINE_AA,\n                    )\n                else:\n                    whole_img = cv2.circle(\n                        whole_img,\n                        (int(pt2[0] + w1), int(pt2[1])),\n                        radius,\n                        color,\n                        -1,\n                        cv2.LINE_AA,\n                    )\n                    cv2.line(\n                        whole_img,\n                        (int(pt1[0]), int(pt1[1])),\n                        (int(pt2[0] + w1), int(pt2[1])),\n                        color,\n                        1,\n                        cv2.LINE_AA,\n                    )\n    return whole_img\n    if row_cat:\n        return np.concatenate([img1, img2], axis=0)\n    return np.concatenate([img1, img2], axis=1)\n\n\nimport torch\nimport os\nimport time\nimport viser\n\n\ndef rotation_matrix_to_quaternion(R):\n    \"\"\"\n    :param R: [3, 3]\n    :return: [4]\n    \"\"\"\n    tr = np.trace(R)\n    Rxx = R[0, 0]\n    Ryy = R[1, 1]\n    Rzz = R[2, 2]\n    q = np.zeros(4)\n    q[0] = 0.5 * np.sqrt(1 + tr)\n    q[1] = (R[2, 1] - R[1, 2]) / (4 * q[0])","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.rotation_matrix_to_quaternion","uri":"program://Human3R/function/src.dust3r.viz.rotation_matrix_to_quaternion#L847-L861","kind":"function","name":"rotation_matrix_to_quaternion","path":"src/dust3r/viz.py","language":"python","start_line":847,"end_line":861,"context_start_line":827,"context_end_line":881,"code":"                    cv2.line(\n                        whole_img,\n                        (int(pt1[0]), int(pt1[1])),\n                        (int(pt2[0] + w1), int(pt2[1])),\n                        color,\n                        1,\n                        cv2.LINE_AA,\n                    )\n    return whole_img\n    if row_cat:\n        return np.concatenate([img1, img2], axis=0)\n    return np.concatenate([img1, img2], axis=1)\n\n\nimport torch\nimport os\nimport time\nimport viser\n\n\ndef rotation_matrix_to_quaternion(R):\n    \"\"\"\n    :param R: [3, 3]\n    :return: [4]\n    \"\"\"\n    tr = np.trace(R)\n    Rxx = R[0, 0]\n    Ryy = R[1, 1]\n    Rzz = R[2, 2]\n    q = np.zeros(4)\n    q[0] = 0.5 * np.sqrt(1 + tr)\n    q[1] = (R[2, 1] - R[1, 2]) / (4 * q[0])\n    q[2] = (R[0, 2] - R[2, 0]) / (4 * q[0])\n    q[3] = (R[1, 0] - R[0, 1]) / (4 * q[0])\n    return q\n\n\nclass PointCloudViewer:\n    def __init__(self, pc_dir, device=\"cpu\"):\n        self.server = viser.ViserServer()\n        self.server.set_up_direction(\"-y\")\n        self.device = device\n        self.tt = lambda x: torch.from_numpy(x).float().to(device)\n        self.pc_dir = pc_dir\n        self.pcs, self.all_steps = self.read_data()\n        self.num_frames = len(self.all_steps)\n\n        self.fix_camera = False\n        self.camera_scale = self.server.add_gui_slider(\n            \"camera_scale\",\n            min=0.01,\n            max=1.0,\n            step=0.01,\n            initial_value=0.1,\n        )","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.PointCloudViewer","uri":"program://Human3R/class/src.dust3r.viz.PointCloudViewer#L864-L1053","kind":"class","name":"PointCloudViewer","path":"src/dust3r/viz.py","language":"python","start_line":864,"end_line":1053,"context_start_line":844,"context_end_line":1073,"code":"import viser\n\n\ndef rotation_matrix_to_quaternion(R):\n    \"\"\"\n    :param R: [3, 3]\n    :return: [4]\n    \"\"\"\n    tr = np.trace(R)\n    Rxx = R[0, 0]\n    Ryy = R[1, 1]\n    Rzz = R[2, 2]\n    q = np.zeros(4)\n    q[0] = 0.5 * np.sqrt(1 + tr)\n    q[1] = (R[2, 1] - R[1, 2]) / (4 * q[0])\n    q[2] = (R[0, 2] - R[2, 0]) / (4 * q[0])\n    q[3] = (R[1, 0] - R[0, 1]) / (4 * q[0])\n    return q\n\n\nclass PointCloudViewer:\n    def __init__(self, pc_dir, device=\"cpu\"):\n        self.server = viser.ViserServer()\n        self.server.set_up_direction(\"-y\")\n        self.device = device\n        self.tt = lambda x: torch.from_numpy(x).float().to(device)\n        self.pc_dir = pc_dir\n        self.pcs, self.all_steps = self.read_data()\n        self.num_frames = len(self.all_steps)\n\n        self.fix_camera = False\n        self.camera_scale = self.server.add_gui_slider(\n            \"camera_scale\",\n            min=0.01,\n            max=1.0,\n            step=0.01,\n            initial_value=0.1,\n        )\n\n        self.camera_handles = []\n\n    def read_data(self):\n        pc_list = os.listdir(self.pc_dir)\n        pc_list.sort(key=lambda x: int(x.split(\".\")[0].split(\"_\")[-1]))\n        pcs = {}\n        step_list = []\n        for pc_name in pc_list:\n            pc = np.load(os.path.join(self.pc_dir, pc_name))\n            step = int(pc_name.split(\".\")[0].split(\"_\")[-1])\n            pcs.update({step: {\"pc\": pc}})\n            step_list.append(step)\n        return pcs, step_list\n\n    def parse_pc_data(self, pc, batch_idx=-1):\n        idx = batch_idx\n        ret_dict = {}\n        for i in range(len(pc.keys()) // 2):\n            pred_pts = pc[f\"pts3d_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            color = pc[f\"colors_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            ret_dict.update({f\"pred_pts_{i+1}\": pred_pts, f\"color_{i+1}\": color})\n        return ret_dict\n\n    def add_pc(self, step):\n        pc = self.pcs[step][\"pc\"]\n        pc_dict = self.parse_pc_data(pc)\n\n        for i in range(len(pc_dict.keys()) // 2):\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts_{i+1}_{step}\",\n                points=pc_dict[f\"pred_pts_{i+1}\"],\n                colors=pc_dict[f\"color_{i+1}\"],\n                point_size=0.002,\n            )\n\n        if not self.fix_camera:\n            raise NotImplementedError\n\n            R21, T21 = find_rigid_alignment_batched(\n                torch.from_numpy(pc_dict[\"pred_pts1_2\"][None]),\n                torch.from_numpy(pc_dict[\"pred_pts1_1\"][None]),\n            )\n            R12, T12 = find_rigid_alignment_batched(\n                torch.from_numpy(pc_dict[\"pred_pts2_1\"][None]),\n                torch.from_numpy(pc_dict[\"pred_pts2_2\"][None]),\n            )\n            R21 = R21[0].numpy()\n            T21 = T21.numpy()\n            R12 = R12[0].numpy()\n            T12 = T12.numpy()\n            pred_pts1_2 = pc_dict[\"pred_pts1_2\"] @ R21.T + T21\n            pred_pts2_1 = pc_dict[\"pred_pts2_1\"] @ R12.T + T12\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts1_2_{step}\",\n                points=pred_pts1_2,\n                colors=pc_dict[\"color1_2\"],\n                point_size=0.002,\n            )\n\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts2_1_{step}\",\n                points=pred_pts2_1,\n                colors=pc_dict[\"color2_1\"],\n                point_size=0.002,\n            )\n            img1 = pc_dict[\"color1_1\"].reshape(224, 224, 3)\n            img2 = pc_dict[\"color2_2\"].reshape(224, 224, 3)\n            self.camera_handles.append(\n                self.server.add_camera_frustum(\n                    name=f\"/frames/{step}/camera1_{step}\",\n                    fov=2.0 * np.arctan(224.0 / 490.0),\n                    aspect=1.0,\n                    scale=self.camera_scale.value,\n                    color=(1.0, 0, 0),\n                    image=img1,\n                )\n            )\n            self.camera_handles.append(\n                self.server.add_camera_frustum(\n                    name=f\"/frames/{step}/camera2_{step}\",\n                    fov=2.0 * np.arctan(224.0 / 490.0),\n                    aspect=1.0,\n                    scale=self.camera_scale.value,\n                    color=(0, 0, 1.0),\n                    wxyz=rotation_matrix_to_quaternion(R21),\n                    position=T21,\n                    image=img2,\n                )\n            )\n\n    def animate(self):\n        with self.server.add_gui_folder(\"Playback\"):\n            gui_timestep = self.server.add_gui_slider(\n                \"Train Step\",\n                min=0,\n                max=self.num_frames - 1,\n                step=1,\n                initial_value=0,\n                disabled=True,\n            )\n            gui_next_frame = self.server.add_gui_button(\"Next Step\", disabled=True)\n            gui_prev_frame = self.server.add_gui_button(\"Prev Step\", disabled=True)\n            gui_playing = self.server.add_gui_checkbox(\"Playing\", False)\n            gui_framerate = self.server.add_gui_slider(\n                \"FPS\", min=1, max=60, step=0.1, initial_value=1\n            )\n            gui_framerate_options = self.server.add_gui_button_group(\n                \"FPS options\", (\"10\", \"20\", \"30\", \"60\")\n            )\n\n        @gui_next_frame.on_click\n        def _(_) -> None:\n            gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n\n        @gui_prev_frame.on_click\n        def _(_) -> None:\n            gui_timestep.value = (gui_timestep.value - 1) % self.num_frames\n\n        @gui_playing.on_update\n        def _(_) -> None:\n            gui_timestep.disabled = gui_playing.value\n            gui_next_frame.disabled = gui_playing.value\n            gui_prev_frame.disabled = gui_playing.value\n\n        @gui_framerate_options.on_click\n        def _(_) -> None:\n            gui_framerate.value = int(gui_framerate_options.value)\n\n        prev_timestep = gui_timestep.value\n\n        @gui_timestep.on_update\n        def _(_) -> None:\n            nonlocal prev_timestep\n            current_timestep = gui_timestep.value\n            with self.server.atomic():\n                frame_nodes[current_timestep].visible = True\n                frame_nodes[prev_timestep].visible = False\n            prev_timestep = current_timestep\n            self.server.flush()  # Optional!\n\n        self.server.add_frame(\n            \"/frames\",\n            show_axes=False,\n        )\n        frame_nodes = []\n        for i in range(self.num_frames):\n            step = self.all_steps[i]\n            frame_nodes.append(\n                self.server.add_frame(\n                    f\"/frames/{step}\",\n                    show_axes=False,\n                )\n            )\n            self.add_pc(step)\n\n        for i, frame_node in enumerate(frame_nodes):\n\n            frame_node.visible = i == gui_timestep.value\n\n        prev_timestep = gui_timestep.value\n        while True:\n            if gui_playing.value:\n                gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n            for handle in self.camera_handles:\n                handle.scale = self.camera_scale.value\n            time.sleep(1.0 / gui_framerate.value)\n\n    def run(self):\n        self.animate()\n        while True:\n            time.sleep(10.0)\n\n\nfrom sklearn.decomposition import PCA\n\n\ndef colorize_feature_map(x):\n    \"\"\"\n    Args:\n        x: torch.Tensor, [B, H, W, D]\n    Returns:\n        torch.Tensor, [B, H, W, 3]\n    \"\"\"\n    device = x.device\n    x = x.cpu().numpy()\n\n    out = []\n    for x_ in x:\n        x_ = colorize_feature_map_np(x_)\n        out.append(torch.from_numpy(x_).to(device))\n    out = torch.stack(out).squeeze(0)","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.colorize_feature_map","uri":"program://Human3R/function/src.dust3r.viz.colorize_feature_map#L1059-L1074","kind":"function","name":"colorize_feature_map","path":"src/dust3r/viz.py","language":"python","start_line":1059,"end_line":1074,"context_start_line":1039,"context_end_line":1089,"code":"\n            frame_node.visible = i == gui_timestep.value\n\n        prev_timestep = gui_timestep.value\n        while True:\n            if gui_playing.value:\n                gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n            for handle in self.camera_handles:\n                handle.scale = self.camera_scale.value\n            time.sleep(1.0 / gui_framerate.value)\n\n    def run(self):\n        self.animate()\n        while True:\n            time.sleep(10.0)\n\n\nfrom sklearn.decomposition import PCA\n\n\ndef colorize_feature_map(x):\n    \"\"\"\n    Args:\n        x: torch.Tensor, [B, H, W, D]\n    Returns:\n        torch.Tensor, [B, H, W, 3]\n    \"\"\"\n    device = x.device\n    x = x.cpu().numpy()\n\n    out = []\n    for x_ in x:\n        x_ = colorize_feature_map_np(x_)\n        out.append(torch.from_numpy(x_).to(device))\n    out = torch.stack(out).squeeze(0)\n    return out\n\n\ndef colorize_feature_map_np(x):\n    \"\"\"\n    Args:\n        x: np.ndarray, [H, W, D]\n    \"\"\"\n    pca = PCA(n_components=3)\n    pca_features = pca.fit_transform(x.reshape(-1, x.shape[-1]))\n\n    pca_features = (pca_features - pca_features.min()) / (\n        pca_features.max() - pca_features.min()\n    )\n    pca_features = pca_features.reshape(x.shape[0], x.shape[1], 3)\n    return pca_features","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.colorize_feature_map_np","uri":"program://Human3R/function/src.dust3r.viz.colorize_feature_map_np#L1077-L1089","kind":"function","name":"colorize_feature_map_np","path":"src/dust3r/viz.py","language":"python","start_line":1077,"end_line":1089,"context_start_line":1057,"context_end_line":1089,"code":"\n\ndef colorize_feature_map(x):\n    \"\"\"\n    Args:\n        x: torch.Tensor, [B, H, W, D]\n    Returns:\n        torch.Tensor, [B, H, W, 3]\n    \"\"\"\n    device = x.device\n    x = x.cpu().numpy()\n\n    out = []\n    for x_ in x:\n        x_ = colorize_feature_map_np(x_)\n        out.append(torch.from_numpy(x_).to(device))\n    out = torch.stack(out).squeeze(0)\n    return out\n\n\ndef colorize_feature_map_np(x):\n    \"\"\"\n    Args:\n        x: np.ndarray, [H, W, D]\n    \"\"\"\n    pca = PCA(n_components=3)\n    pca_features = pca.fit_transform(x.reshape(-1, x.shape[-1]))\n\n    pca_features = (pca_features - pca_features.min()) / (\n        pca_features.max() - pca_features.min()\n    )\n    pca_features = pca_features.reshape(x.shape[0], x.shape[1], 3)\n    return pca_features","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.__init__","uri":"program://Human3R/function/src.dust3r.viz.__init__#L865-L883","kind":"function","name":"__init__","path":"src/dust3r/viz.py","language":"python","start_line":865,"end_line":883,"context_start_line":845,"context_end_line":903,"code":"\n\ndef rotation_matrix_to_quaternion(R):\n    \"\"\"\n    :param R: [3, 3]\n    :return: [4]\n    \"\"\"\n    tr = np.trace(R)\n    Rxx = R[0, 0]\n    Ryy = R[1, 1]\n    Rzz = R[2, 2]\n    q = np.zeros(4)\n    q[0] = 0.5 * np.sqrt(1 + tr)\n    q[1] = (R[2, 1] - R[1, 2]) / (4 * q[0])\n    q[2] = (R[0, 2] - R[2, 0]) / (4 * q[0])\n    q[3] = (R[1, 0] - R[0, 1]) / (4 * q[0])\n    return q\n\n\nclass PointCloudViewer:\n    def __init__(self, pc_dir, device=\"cpu\"):\n        self.server = viser.ViserServer()\n        self.server.set_up_direction(\"-y\")\n        self.device = device\n        self.tt = lambda x: torch.from_numpy(x).float().to(device)\n        self.pc_dir = pc_dir\n        self.pcs, self.all_steps = self.read_data()\n        self.num_frames = len(self.all_steps)\n\n        self.fix_camera = False\n        self.camera_scale = self.server.add_gui_slider(\n            \"camera_scale\",\n            min=0.01,\n            max=1.0,\n            step=0.01,\n            initial_value=0.1,\n        )\n\n        self.camera_handles = []\n\n    def read_data(self):\n        pc_list = os.listdir(self.pc_dir)\n        pc_list.sort(key=lambda x: int(x.split(\".\")[0].split(\"_\")[-1]))\n        pcs = {}\n        step_list = []\n        for pc_name in pc_list:\n            pc = np.load(os.path.join(self.pc_dir, pc_name))\n            step = int(pc_name.split(\".\")[0].split(\"_\")[-1])\n            pcs.update({step: {\"pc\": pc}})\n            step_list.append(step)\n        return pcs, step_list\n\n    def parse_pc_data(self, pc, batch_idx=-1):\n        idx = batch_idx\n        ret_dict = {}\n        for i in range(len(pc.keys()) // 2):\n            pred_pts = pc[f\"pts3d_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            color = pc[f\"colors_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            ret_dict.update({f\"pred_pts_{i+1}\": pred_pts, f\"color_{i+1}\": color})","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.add_rgbd","uri":"program://Human3R/function/src.dust3r.viz.add_rgbd#L203-L220","kind":"function","name":"add_rgbd","path":"src/dust3r/viz.py","language":"python","start_line":203,"end_line":220,"context_start_line":183,"context_end_line":240,"code":"            assert col.shape == pts.shape, bb()\n            pct.visual.vertex_colors = uint8(col.reshape(-1, 3))\n        else:\n            assert len(color) == 3\n            pct.visual.vertex_colors = np.broadcast_to(uint8(color), pts.shape)\n\n        if denoise:\n\n            centroid = np.median(pct.vertices, axis=0)\n            dist_to_centroid = np.linalg.norm(pct.vertices - centroid, axis=-1)\n            dist_thr = np.quantile(dist_to_centroid, 0.99)\n            valid = dist_to_centroid < dist_thr\n\n            pct = trimesh.PointCloud(\n                pct.vertices[valid], color=pct.visual.vertex_colors[valid]\n            )\n\n        self.scene.add_geometry(pct)\n        return self\n\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n\n        pts3d, mask2 = depthmap_to_absolute_camera_coordinates(\n            depth, intrinsics, cam2world\n        )\n        mask2 &= depth < zfar\n\n        if mask is not None:\n            mask2 &= mask\n\n        return self.add_pointcloud(pts3d, image, mask=mask2)\n\n    def add_camera(\n        self,\n        pose_c2w,\n        focal=None,\n        color=(0, 0, 0),\n        image=None,\n        imsize=None,\n        cam_size=0.03,\n    ):\n        pose_c2w, focal, color, image = to_numpy((pose_c2w, focal, color, image))\n        image = img_to_arr(image)\n        if isinstance(focal, np.ndarray) and focal.shape == (3, 3):\n            intrinsics = focal\n            focal = (intrinsics[0, 0] * intrinsics[1, 1]) ** 0.5\n            if imsize is None:\n                imsize = (2 * intrinsics[0, 2], 2 * intrinsics[1, 2])\n\n        add_scene_cam(\n            self.scene,","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.add_pointcloud","uri":"program://Human3R/function/src.dust3r.viz.add_pointcloud#L165-L201","kind":"function","name":"add_pointcloud","path":"src/dust3r/viz.py","language":"python","start_line":165,"end_line":201,"context_start_line":145,"context_end_line":221,"code":"class SceneViz:\n    def __init__(self):\n        self.scene = trimesh.Scene()\n\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n        image = img_to_arr(image)\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n\n        pts3d = depthmap_to_pts3d(depth, intrinsics, cam2world=cam2world)\n\n        return self.add_pointcloud(\n            pts3d, image, mask=(depth < zfar) if mask is None else mask\n        )\n\n    def add_pointcloud(self, pts3d, color=(0, 0, 0), mask=None, denoise=False):\n        pts3d = to_numpy(pts3d)\n        mask = to_numpy(mask)\n        if not isinstance(pts3d, list):\n            pts3d = [pts3d.reshape(-1, 3)]\n            if mask is not None:\n                mask = [mask.ravel()]\n        if not isinstance(color, (tuple, list)):\n            color = [color.reshape(-1, 3)]\n        if mask is None:\n            mask = [slice(None)] * len(pts3d)\n\n        pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])\n        pct = trimesh.PointCloud(pts)\n\n        if isinstance(color, (list, np.ndarray, torch.Tensor)):\n            color = to_numpy(color)\n            col = np.concatenate([p[m] for p, m in zip(color, mask)])\n            assert col.shape == pts.shape, bb()\n            pct.visual.vertex_colors = uint8(col.reshape(-1, 3))\n        else:\n            assert len(color) == 3\n            pct.visual.vertex_colors = np.broadcast_to(uint8(color), pts.shape)\n\n        if denoise:\n\n            centroid = np.median(pct.vertices, axis=0)\n            dist_to_centroid = np.linalg.norm(pct.vertices - centroid, axis=-1)\n            dist_thr = np.quantile(dist_to_centroid, 0.99)\n            valid = dist_to_centroid < dist_thr\n\n            pct = trimesh.PointCloud(\n                pct.vertices[valid], color=pct.visual.vertex_colors[valid]\n            )\n\n        self.scene.add_geometry(pct)\n        return self\n\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n\n        pts3d, mask2 = depthmap_to_absolute_camera_coordinates(\n            depth, intrinsics, cam2world\n        )\n        mask2 &= depth < zfar\n\n        if mask is not None:\n            mask2 &= mask\n\n        return self.add_pointcloud(pts3d, image, mask=mask2)\n","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.add_camera","uri":"program://Human3R/function/src.dust3r.viz.add_camera#L222-L249","kind":"function","name":"add_camera","path":"src/dust3r/viz.py","language":"python","start_line":222,"end_line":249,"context_start_line":202,"context_end_line":269,"code":"\n    def add_rgbd(\n        self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None\n    ):\n\n        if intrinsics is None:\n            H, W, THREE = image.shape\n            focal = max(H, W)\n            intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]])\n\n        pts3d, mask2 = depthmap_to_absolute_camera_coordinates(\n            depth, intrinsics, cam2world\n        )\n        mask2 &= depth < zfar\n\n        if mask is not None:\n            mask2 &= mask\n\n        return self.add_pointcloud(pts3d, image, mask=mask2)\n\n    def add_camera(\n        self,\n        pose_c2w,\n        focal=None,\n        color=(0, 0, 0),\n        image=None,\n        imsize=None,\n        cam_size=0.03,\n    ):\n        pose_c2w, focal, color, image = to_numpy((pose_c2w, focal, color, image))\n        image = img_to_arr(image)\n        if isinstance(focal, np.ndarray) and focal.shape == (3, 3):\n            intrinsics = focal\n            focal = (intrinsics[0, 0] * intrinsics[1, 1]) ** 0.5\n            if imsize is None:\n                imsize = (2 * intrinsics[0, 2], 2 * intrinsics[1, 2])\n\n        add_scene_cam(\n            self.scene,\n            pose_c2w,\n            color,\n            image,\n            focal,\n            imsize=imsize,\n            screen_width=cam_size,\n            marker=None,\n        )\n        return self\n\n    def add_cameras(\n        self, poses, focals=None, images=None, imsizes=None, colors=None, **kw\n    ):\n        get = lambda arr, idx: None if arr is None else arr[idx]\n        for i, pose_c2w in enumerate(poses):\n            self.add_camera(\n                pose_c2w,\n                get(focals, i),\n                image=get(images, i),\n                color=get(colors, i),\n                imsize=get(imsizes, i),\n                **kw,\n            )\n        return self\n\n    def show(self, point_size=2):\n        self.scene.show(line_settings={\"point_size\": point_size})\n\n","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.add_cameras","uri":"program://Human3R/function/src.dust3r.viz.add_cameras#L251-L264","kind":"function","name":"add_cameras","path":"src/dust3r/viz.py","language":"python","start_line":251,"end_line":264,"context_start_line":231,"context_end_line":284,"code":"        pose_c2w, focal, color, image = to_numpy((pose_c2w, focal, color, image))\n        image = img_to_arr(image)\n        if isinstance(focal, np.ndarray) and focal.shape == (3, 3):\n            intrinsics = focal\n            focal = (intrinsics[0, 0] * intrinsics[1, 1]) ** 0.5\n            if imsize is None:\n                imsize = (2 * intrinsics[0, 2], 2 * intrinsics[1, 2])\n\n        add_scene_cam(\n            self.scene,\n            pose_c2w,\n            color,\n            image,\n            focal,\n            imsize=imsize,\n            screen_width=cam_size,\n            marker=None,\n        )\n        return self\n\n    def add_cameras(\n        self, poses, focals=None, images=None, imsizes=None, colors=None, **kw\n    ):\n        get = lambda arr, idx: None if arr is None else arr[idx]\n        for i, pose_c2w in enumerate(poses):\n            self.add_camera(\n                pose_c2w,\n                get(focals, i),\n                image=get(images, i),\n                color=get(colors, i),\n                imsize=get(imsizes, i),\n                **kw,\n            )\n        return self\n\n    def show(self, point_size=2):\n        self.scene.show(line_settings={\"point_size\": point_size})\n\n\ndef show_raw_pointcloud_with_cams(\n    imgs, pts3d, mask, focals, cams2world, point_size=2, cam_size=0.05, cam_color=None\n):\n    \"\"\"Visualization of a pointcloud with cameras\n    imgs = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    pts3d = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    focals = (N,) or N-size list of [focal, ...]\n    cams2world = (N,4,4) or N-size list of [(4,4), ...]\n    \"\"\"\n    assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)\n    pts3d = to_numpy(pts3d)\n    imgs = to_numpy(imgs)\n    focals = to_numpy(focals)\n    cams2world = to_numpy(cams2world)\n","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.show","uri":"program://Human3R/function/src.dust3r.viz.show#L266-L267","kind":"function","name":"show","path":"src/dust3r/viz.py","language":"python","start_line":266,"end_line":267,"context_start_line":246,"context_end_line":287,"code":"            screen_width=cam_size,\n            marker=None,\n        )\n        return self\n\n    def add_cameras(\n        self, poses, focals=None, images=None, imsizes=None, colors=None, **kw\n    ):\n        get = lambda arr, idx: None if arr is None else arr[idx]\n        for i, pose_c2w in enumerate(poses):\n            self.add_camera(\n                pose_c2w,\n                get(focals, i),\n                image=get(images, i),\n                color=get(colors, i),\n                imsize=get(imsizes, i),\n                **kw,\n            )\n        return self\n\n    def show(self, point_size=2):\n        self.scene.show(line_settings={\"point_size\": point_size})\n\n\ndef show_raw_pointcloud_with_cams(\n    imgs, pts3d, mask, focals, cams2world, point_size=2, cam_size=0.05, cam_color=None\n):\n    \"\"\"Visualization of a pointcloud with cameras\n    imgs = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    pts3d = (N, H, W, 3) or N-size list of [(H,W,3), ...]\n    focals = (N,) or N-size list of [focal, ...]\n    cams2world = (N,4,4) or N-size list of [(4,4), ...]\n    \"\"\"\n    assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)\n    pts3d = to_numpy(pts3d)\n    imgs = to_numpy(imgs)\n    focals = to_numpy(focals)\n    cams2world = to_numpy(cams2world)\n\n    scene = trimesh.Scene()\n\n    pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.read_data","uri":"program://Human3R/function/src.dust3r.viz.read_data#L885-L895","kind":"function","name":"read_data","path":"src/dust3r/viz.py","language":"python","start_line":885,"end_line":895,"context_start_line":865,"context_end_line":915,"code":"    def __init__(self, pc_dir, device=\"cpu\"):\n        self.server = viser.ViserServer()\n        self.server.set_up_direction(\"-y\")\n        self.device = device\n        self.tt = lambda x: torch.from_numpy(x).float().to(device)\n        self.pc_dir = pc_dir\n        self.pcs, self.all_steps = self.read_data()\n        self.num_frames = len(self.all_steps)\n\n        self.fix_camera = False\n        self.camera_scale = self.server.add_gui_slider(\n            \"camera_scale\",\n            min=0.01,\n            max=1.0,\n            step=0.01,\n            initial_value=0.1,\n        )\n\n        self.camera_handles = []\n\n    def read_data(self):\n        pc_list = os.listdir(self.pc_dir)\n        pc_list.sort(key=lambda x: int(x.split(\".\")[0].split(\"_\")[-1]))\n        pcs = {}\n        step_list = []\n        for pc_name in pc_list:\n            pc = np.load(os.path.join(self.pc_dir, pc_name))\n            step = int(pc_name.split(\".\")[0].split(\"_\")[-1])\n            pcs.update({step: {\"pc\": pc}})\n            step_list.append(step)\n        return pcs, step_list\n\n    def parse_pc_data(self, pc, batch_idx=-1):\n        idx = batch_idx\n        ret_dict = {}\n        for i in range(len(pc.keys()) // 2):\n            pred_pts = pc[f\"pts3d_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            color = pc[f\"colors_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            ret_dict.update({f\"pred_pts_{i+1}\": pred_pts, f\"color_{i+1}\": color})\n        return ret_dict\n\n    def add_pc(self, step):\n        pc = self.pcs[step][\"pc\"]\n        pc_dict = self.parse_pc_data(pc)\n\n        for i in range(len(pc_dict.keys()) // 2):\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts_{i+1}_{step}\",\n                points=pc_dict[f\"pred_pts_{i+1}\"],\n                colors=pc_dict[f\"color_{i+1}\"],\n                point_size=0.002,","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.parse_pc_data","uri":"program://Human3R/function/src.dust3r.viz.parse_pc_data#L897-L904","kind":"function","name":"parse_pc_data","path":"src/dust3r/viz.py","language":"python","start_line":897,"end_line":904,"context_start_line":877,"context_end_line":924,"code":"            min=0.01,\n            max=1.0,\n            step=0.01,\n            initial_value=0.1,\n        )\n\n        self.camera_handles = []\n\n    def read_data(self):\n        pc_list = os.listdir(self.pc_dir)\n        pc_list.sort(key=lambda x: int(x.split(\".\")[0].split(\"_\")[-1]))\n        pcs = {}\n        step_list = []\n        for pc_name in pc_list:\n            pc = np.load(os.path.join(self.pc_dir, pc_name))\n            step = int(pc_name.split(\".\")[0].split(\"_\")[-1])\n            pcs.update({step: {\"pc\": pc}})\n            step_list.append(step)\n        return pcs, step_list\n\n    def parse_pc_data(self, pc, batch_idx=-1):\n        idx = batch_idx\n        ret_dict = {}\n        for i in range(len(pc.keys()) // 2):\n            pred_pts = pc[f\"pts3d_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            color = pc[f\"colors_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            ret_dict.update({f\"pred_pts_{i+1}\": pred_pts, f\"color_{i+1}\": color})\n        return ret_dict\n\n    def add_pc(self, step):\n        pc = self.pcs[step][\"pc\"]\n        pc_dict = self.parse_pc_data(pc)\n\n        for i in range(len(pc_dict.keys()) // 2):\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts_{i+1}_{step}\",\n                points=pc_dict[f\"pred_pts_{i+1}\"],\n                colors=pc_dict[f\"color_{i+1}\"],\n                point_size=0.002,\n            )\n\n        if not self.fix_camera:\n            raise NotImplementedError\n\n            R21, T21 = find_rigid_alignment_batched(\n                torch.from_numpy(pc_dict[\"pred_pts1_2\"][None]),\n                torch.from_numpy(pc_dict[\"pred_pts1_1\"][None]),\n            )","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.add_pc","uri":"program://Human3R/function/src.dust3r.viz.add_pc#L906-L971","kind":"function","name":"add_pc","path":"src/dust3r/viz.py","language":"python","start_line":906,"end_line":971,"context_start_line":886,"context_end_line":991,"code":"        pc_list = os.listdir(self.pc_dir)\n        pc_list.sort(key=lambda x: int(x.split(\".\")[0].split(\"_\")[-1]))\n        pcs = {}\n        step_list = []\n        for pc_name in pc_list:\n            pc = np.load(os.path.join(self.pc_dir, pc_name))\n            step = int(pc_name.split(\".\")[0].split(\"_\")[-1])\n            pcs.update({step: {\"pc\": pc}})\n            step_list.append(step)\n        return pcs, step_list\n\n    def parse_pc_data(self, pc, batch_idx=-1):\n        idx = batch_idx\n        ret_dict = {}\n        for i in range(len(pc.keys()) // 2):\n            pred_pts = pc[f\"pts3d_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            color = pc[f\"colors_{i+1}\"][idx].reshape(-1, 3)  # [N, 3]\n            ret_dict.update({f\"pred_pts_{i+1}\": pred_pts, f\"color_{i+1}\": color})\n        return ret_dict\n\n    def add_pc(self, step):\n        pc = self.pcs[step][\"pc\"]\n        pc_dict = self.parse_pc_data(pc)\n\n        for i in range(len(pc_dict.keys()) // 2):\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts_{i+1}_{step}\",\n                points=pc_dict[f\"pred_pts_{i+1}\"],\n                colors=pc_dict[f\"color_{i+1}\"],\n                point_size=0.002,\n            )\n\n        if not self.fix_camera:\n            raise NotImplementedError\n\n            R21, T21 = find_rigid_alignment_batched(\n                torch.from_numpy(pc_dict[\"pred_pts1_2\"][None]),\n                torch.from_numpy(pc_dict[\"pred_pts1_1\"][None]),\n            )\n            R12, T12 = find_rigid_alignment_batched(\n                torch.from_numpy(pc_dict[\"pred_pts2_1\"][None]),\n                torch.from_numpy(pc_dict[\"pred_pts2_2\"][None]),\n            )\n            R21 = R21[0].numpy()\n            T21 = T21.numpy()\n            R12 = R12[0].numpy()\n            T12 = T12.numpy()\n            pred_pts1_2 = pc_dict[\"pred_pts1_2\"] @ R21.T + T21\n            pred_pts2_1 = pc_dict[\"pred_pts2_1\"] @ R12.T + T12\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts1_2_{step}\",\n                points=pred_pts1_2,\n                colors=pc_dict[\"color1_2\"],\n                point_size=0.002,\n            )\n\n            self.server.add_point_cloud(\n                name=f\"/frames/{step}/pred_pts2_1_{step}\",\n                points=pred_pts2_1,\n                colors=pc_dict[\"color2_1\"],\n                point_size=0.002,\n            )\n            img1 = pc_dict[\"color1_1\"].reshape(224, 224, 3)\n            img2 = pc_dict[\"color2_2\"].reshape(224, 224, 3)\n            self.camera_handles.append(\n                self.server.add_camera_frustum(\n                    name=f\"/frames/{step}/camera1_{step}\",\n                    fov=2.0 * np.arctan(224.0 / 490.0),\n                    aspect=1.0,\n                    scale=self.camera_scale.value,\n                    color=(1.0, 0, 0),\n                    image=img1,\n                )\n            )\n            self.camera_handles.append(\n                self.server.add_camera_frustum(\n                    name=f\"/frames/{step}/camera2_{step}\",\n                    fov=2.0 * np.arctan(224.0 / 490.0),\n                    aspect=1.0,\n                    scale=self.camera_scale.value,\n                    color=(0, 0, 1.0),\n                    wxyz=rotation_matrix_to_quaternion(R21),\n                    position=T21,\n                    image=img2,\n                )\n            )\n\n    def animate(self):\n        with self.server.add_gui_folder(\"Playback\"):\n            gui_timestep = self.server.add_gui_slider(\n                \"Train Step\",\n                min=0,\n                max=self.num_frames - 1,\n                step=1,\n                initial_value=0,\n                disabled=True,\n            )\n            gui_next_frame = self.server.add_gui_button(\"Next Step\", disabled=True)\n            gui_prev_frame = self.server.add_gui_button(\"Prev Step\", disabled=True)\n            gui_playing = self.server.add_gui_checkbox(\"Playing\", False)\n            gui_framerate = self.server.add_gui_slider(\n                \"FPS\", min=1, max=60, step=0.1, initial_value=1\n            )\n            gui_framerate_options = self.server.add_gui_button_group(\n                \"FPS options\", (\"10\", \"20\", \"30\", \"60\")\n            )","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.animate","uri":"program://Human3R/function/src.dust3r.viz.animate#L973-L1048","kind":"function","name":"animate","path":"src/dust3r/viz.py","language":"python","start_line":973,"end_line":1048,"context_start_line":953,"context_end_line":1068,"code":"                    fov=2.0 * np.arctan(224.0 / 490.0),\n                    aspect=1.0,\n                    scale=self.camera_scale.value,\n                    color=(1.0, 0, 0),\n                    image=img1,\n                )\n            )\n            self.camera_handles.append(\n                self.server.add_camera_frustum(\n                    name=f\"/frames/{step}/camera2_{step}\",\n                    fov=2.0 * np.arctan(224.0 / 490.0),\n                    aspect=1.0,\n                    scale=self.camera_scale.value,\n                    color=(0, 0, 1.0),\n                    wxyz=rotation_matrix_to_quaternion(R21),\n                    position=T21,\n                    image=img2,\n                )\n            )\n\n    def animate(self):\n        with self.server.add_gui_folder(\"Playback\"):\n            gui_timestep = self.server.add_gui_slider(\n                \"Train Step\",\n                min=0,\n                max=self.num_frames - 1,\n                step=1,\n                initial_value=0,\n                disabled=True,\n            )\n            gui_next_frame = self.server.add_gui_button(\"Next Step\", disabled=True)\n            gui_prev_frame = self.server.add_gui_button(\"Prev Step\", disabled=True)\n            gui_playing = self.server.add_gui_checkbox(\"Playing\", False)\n            gui_framerate = self.server.add_gui_slider(\n                \"FPS\", min=1, max=60, step=0.1, initial_value=1\n            )\n            gui_framerate_options = self.server.add_gui_button_group(\n                \"FPS options\", (\"10\", \"20\", \"30\", \"60\")\n            )\n\n        @gui_next_frame.on_click\n        def _(_) -> None:\n            gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n\n        @gui_prev_frame.on_click\n        def _(_) -> None:\n            gui_timestep.value = (gui_timestep.value - 1) % self.num_frames\n\n        @gui_playing.on_update\n        def _(_) -> None:\n            gui_timestep.disabled = gui_playing.value\n            gui_next_frame.disabled = gui_playing.value\n            gui_prev_frame.disabled = gui_playing.value\n\n        @gui_framerate_options.on_click\n        def _(_) -> None:\n            gui_framerate.value = int(gui_framerate_options.value)\n\n        prev_timestep = gui_timestep.value\n\n        @gui_timestep.on_update\n        def _(_) -> None:\n            nonlocal prev_timestep\n            current_timestep = gui_timestep.value\n            with self.server.atomic():\n                frame_nodes[current_timestep].visible = True\n                frame_nodes[prev_timestep].visible = False\n            prev_timestep = current_timestep\n            self.server.flush()  # Optional!\n\n        self.server.add_frame(\n            \"/frames\",\n            show_axes=False,\n        )\n        frame_nodes = []\n        for i in range(self.num_frames):\n            step = self.all_steps[i]\n            frame_nodes.append(\n                self.server.add_frame(\n                    f\"/frames/{step}\",\n                    show_axes=False,\n                )\n            )\n            self.add_pc(step)\n\n        for i, frame_node in enumerate(frame_nodes):\n\n            frame_node.visible = i == gui_timestep.value\n\n        prev_timestep = gui_timestep.value\n        while True:\n            if gui_playing.value:\n                gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n            for handle in self.camera_handles:\n                handle.scale = self.camera_scale.value\n            time.sleep(1.0 / gui_framerate.value)\n\n    def run(self):\n        self.animate()\n        while True:\n            time.sleep(10.0)\n\n\nfrom sklearn.decomposition import PCA\n\n\ndef colorize_feature_map(x):\n    \"\"\"\n    Args:\n        x: torch.Tensor, [B, H, W, D]\n    Returns:\n        torch.Tensor, [B, H, W, 3]\n    \"\"\"\n    device = x.device\n    x = x.cpu().numpy()\n","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz.run","uri":"program://Human3R/function/src.dust3r.viz.run#L1050-L1053","kind":"function","name":"run","path":"src/dust3r/viz.py","language":"python","start_line":1050,"end_line":1053,"context_start_line":1030,"context_end_line":1073,"code":"            frame_nodes.append(\n                self.server.add_frame(\n                    f\"/frames/{step}\",\n                    show_axes=False,\n                )\n            )\n            self.add_pc(step)\n\n        for i, frame_node in enumerate(frame_nodes):\n\n            frame_node.visible = i == gui_timestep.value\n\n        prev_timestep = gui_timestep.value\n        while True:\n            if gui_playing.value:\n                gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n            for handle in self.camera_handles:\n                handle.scale = self.camera_scale.value\n            time.sleep(1.0 / gui_framerate.value)\n\n    def run(self):\n        self.animate()\n        while True:\n            time.sleep(10.0)\n\n\nfrom sklearn.decomposition import PCA\n\n\ndef colorize_feature_map(x):\n    \"\"\"\n    Args:\n        x: torch.Tensor, [B, H, W, D]\n    Returns:\n        torch.Tensor, [B, H, W, 3]\n    \"\"\"\n    device = x.device\n    x = x.cpu().numpy()\n\n    out = []\n    for x_ in x:\n        x_ = colorize_feature_map_np(x_)\n        out.append(torch.from_numpy(x_).to(device))\n    out = torch.stack(out).squeeze(0)","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.viz._","uri":"program://Human3R/function/src.dust3r.viz._#L1014-L1021","kind":"function","name":"_","path":"src/dust3r/viz.py","language":"python","start_line":1014,"end_line":1021,"context_start_line":994,"context_end_line":1041,"code":"        def _(_) -> None:\n            gui_timestep.value = (gui_timestep.value + 1) % self.num_frames\n\n        @gui_prev_frame.on_click\n        def _(_) -> None:\n            gui_timestep.value = (gui_timestep.value - 1) % self.num_frames\n\n        @gui_playing.on_update\n        def _(_) -> None:\n            gui_timestep.disabled = gui_playing.value\n            gui_next_frame.disabled = gui_playing.value\n            gui_prev_frame.disabled = gui_playing.value\n\n        @gui_framerate_options.on_click\n        def _(_) -> None:\n            gui_framerate.value = int(gui_framerate_options.value)\n\n        prev_timestep = gui_timestep.value\n\n        @gui_timestep.on_update\n        def _(_) -> None:\n            nonlocal prev_timestep\n            current_timestep = gui_timestep.value\n            with self.server.atomic():\n                frame_nodes[current_timestep].visible = True\n                frame_nodes[prev_timestep].visible = False\n            prev_timestep = current_timestep\n            self.server.flush()  # Optional!\n\n        self.server.add_frame(\n            \"/frames\",\n            show_axes=False,\n        )\n        frame_nodes = []\n        for i in range(self.num_frames):\n            step = self.all_steps[i]\n            frame_nodes.append(\n                self.server.add_frame(\n                    f\"/frames/{step}\",\n                    show_axes=False,\n                )\n            )\n            self.add_pc(step)\n\n        for i, frame_node in enumerate(frame_nodes):\n\n            frame_node.visible = i == gui_timestep.value\n","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.post_process","uri":"program://Human3R/module/src.dust3r.post_process#L1-L64","kind":"module","name":"src.dust3r.post_process","path":"src/dust3r/post_process.py","language":"python","start_line":1,"end_line":64,"context_start_line":1,"context_end_line":64,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nimport torch\nfrom dust3r.utils.geometry import xy_grid\n\n\ndef estimate_focal_knowing_depth(\n    pts3d, pp, focal_mode=\"median\", min_focal=0.0, max_focal=np.inf\n):\n    \"\"\"Reprojection method, for when the absolute depth is known:\n    1) estimate the camera focal using a robust estimator\n    2) reproject points onto true rays, minimizing a certain error\n    \"\"\"\n    B, H, W, THREE = pts3d.shape\n    assert THREE == 3\n\n    pixels = xy_grid(W, H, device=pts3d.device).view(1, -1, 2) - pp.view(\n        -1, 1, 2\n    )  # B,HW,2\n    pts3d = pts3d.flatten(1, 2)  # (B, HW, 3)\n\n    if focal_mode == \"median\":\n        with torch.no_grad():\n\n            u, v = pixels.unbind(dim=-1)\n            x, y, z = pts3d.unbind(dim=-1)\n            fx_votes = (u * z) / x\n            fy_votes = (v * z) / y\n\n            f_votes = torch.cat((fx_votes.view(B, -1), fy_votes.view(B, -1)), dim=-1)\n            focal = torch.nanmedian(f_votes, dim=-1).values\n\n    elif focal_mode == \"weiszfeld\":\n\n        xy_over_z = (pts3d[..., :2] / pts3d[..., 2:3]).nan_to_num(\n            posinf=0, neginf=0\n        )  # homogeneous (x,y,1)\n\n        dot_xy_px = (xy_over_z * pixels).sum(dim=-1)\n        dot_xy_xy = xy_over_z.square().sum(dim=-1)\n\n        focal = dot_xy_px.mean(dim=1) / dot_xy_xy.mean(dim=1)\n\n        for iter in range(10):\n\n            dis = (pixels - focal.view(-1, 1, 1) * xy_over_z).norm(dim=-1)\n\n            w = dis.clip(min=1e-8).reciprocal()\n\n            focal = (w * dot_xy_px).mean(dim=1) / (w * dot_xy_xy).mean(dim=1)\n    else:\n        raise ValueError(f\"bad {focal_mode=}\")\n\n    focal_base = max(H, W) / (\n        2 * np.tan(np.deg2rad(60) / 2)\n    )  # size / 1.1547005383792515\n    focal = focal.clip(min=min_focal * focal_base, max=max_focal * focal_base)\n\n    return focal","source_hash":"e2ef308ee09ae547f64c3e8e5314cc6dd60a5ff37b2e548812af53f8f1ad1f37","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.post_process.estimate_focal_knowing_depth","uri":"program://Human3R/function/src.dust3r.post_process.estimate_focal_knowing_depth#L12-L64","kind":"function","name":"estimate_focal_knowing_depth","path":"src/dust3r/post_process.py","language":"python","start_line":12,"end_line":64,"context_start_line":1,"context_end_line":64,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nimport torch\nfrom dust3r.utils.geometry import xy_grid\n\n\ndef estimate_focal_knowing_depth(\n    pts3d, pp, focal_mode=\"median\", min_focal=0.0, max_focal=np.inf\n):\n    \"\"\"Reprojection method, for when the absolute depth is known:\n    1) estimate the camera focal using a robust estimator\n    2) reproject points onto true rays, minimizing a certain error\n    \"\"\"\n    B, H, W, THREE = pts3d.shape\n    assert THREE == 3\n\n    pixels = xy_grid(W, H, device=pts3d.device).view(1, -1, 2) - pp.view(\n        -1, 1, 2\n    )  # B,HW,2\n    pts3d = pts3d.flatten(1, 2)  # (B, HW, 3)\n\n    if focal_mode == \"median\":\n        with torch.no_grad():\n\n            u, v = pixels.unbind(dim=-1)\n            x, y, z = pts3d.unbind(dim=-1)\n            fx_votes = (u * z) / x\n            fy_votes = (v * z) / y\n\n            f_votes = torch.cat((fx_votes.view(B, -1), fy_votes.view(B, -1)), dim=-1)\n            focal = torch.nanmedian(f_votes, dim=-1).values\n\n    elif focal_mode == \"weiszfeld\":\n\n        xy_over_z = (pts3d[..., :2] / pts3d[..., 2:3]).nan_to_num(\n            posinf=0, neginf=0\n        )  # homogeneous (x,y,1)\n\n        dot_xy_px = (xy_over_z * pixels).sum(dim=-1)\n        dot_xy_xy = xy_over_z.square().sum(dim=-1)\n\n        focal = dot_xy_px.mean(dim=1) / dot_xy_xy.mean(dim=1)\n\n        for iter in range(10):\n\n            dis = (pixels - focal.view(-1, 1, 1) * xy_over_z).norm(dim=-1)\n\n            w = dis.clip(min=1e-8).reciprocal()\n\n            focal = (w * dot_xy_px).mean(dim=1) / (w * dot_xy_xy).mean(dim=1)\n    else:\n        raise ValueError(f\"bad {focal_mode=}\")\n\n    focal_base = max(H, W) / (\n        2 * np.tan(np.deg2rad(60) / 2)\n    )  # size / 1.1547005383792515\n    focal = focal.clip(min=min_focal * focal_base, max=max_focal * focal_base)\n\n    return focal","source_hash":"e2ef308ee09ae547f64c3e8e5314cc6dd60a5ff37b2e548812af53f8f1ad1f37","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.device","uri":"program://Human3R/module/src.dust3r.utils.device#L1-L108","kind":"module","name":"src.dust3r.utils.device","path":"src/dust3r/utils/device.py","language":"python","start_line":1,"end_line":108,"context_start_line":1,"context_end_line":108,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nimport torch\n\n\ndef todevice(batch, device, callback=None, non_blocking=False):\n    \"\"\"Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy).\n\n    batch: list, tuple, dict of tensors or other things\n    device: pytorch device or 'numpy'\n    callback: function that would be called on every sub-elements.\n    \"\"\"\n    if callback:\n        batch = callback(batch)\n\n    if isinstance(batch, dict):\n        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef to_cpu(x):\n    return todevice(x, \"cpu\")\n\n\ndef to_cuda(x):\n    return todevice(x, \"cuda\")\n\n\ndef collate_with_cat(whatever, lists=False):\n    if isinstance(whatever, dict):\n        return {k: collate_with_cat(vals, lists=lists) for k, vals in whatever.items()}\n\n    elif isinstance(whatever, (tuple, list)):\n        if len(whatever) == 0:\n            return whatever\n        elem = whatever[0]\n        T = type(whatever)\n\n        if elem is None:\n            return None\n        if isinstance(elem, (bool, float, int, str)):\n            return whatever\n        if isinstance(elem, tuple):\n            return T(collate_with_cat(x, lists=lists) for x in zip(*whatever))\n        if isinstance(elem, dict):\n            return {\n                k: collate_with_cat([e[k] for e in whatever], lists=lists) for k in elem\n            }\n\n        if isinstance(elem, torch.Tensor):\n            return listify(whatever) if lists else torch.cat(whatever)\n        if isinstance(elem, np.ndarray):\n            return (\n                listify(whatever)\n                if lists\n                else torch.cat([torch.from_numpy(x) for x in whatever])\n            )\n\n        return sum(whatever, T())\n\n\ndef listify(elems):\n    return [x for e in elems for x in e]\n\n\ndef to_gpu(_view, device):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"rng\", \n         \"ray_map\", \"camera_pose\", \"camera_intrinsics\", \"ray_mask\", \"fov_x\",\n         \"fov_y\", \"T_w2c\", \"smpl_v3d_w\", \"smpl_j3d_w\", \"smpl_v3d_c\", \"smpl_j3d_c\",\n         \"smpl_j2d\", \"smpl_v2d\", \"smpl_mask\", \"msk\",\n         ]\n    )\n    view = {}\n    for name in _view.keys():  # pseudo_focal\n        if name in ignore_keys:\n            continue\n        if isinstance(_view[name], tuple) or isinstance(_view[name], list):\n            view[name] = [x.clone().to(device, non_blocking=True) for x in _view[name]]\n        else:\n            view[name] = _view[name].clone().to(device, non_blocking=True)\n    \n    return view","source_hash":"d6eed957d20ec85bcb142727736b58f5c034322d6fc610964924f0b95d3e977c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.device.todevice","uri":"program://Human3R/function/src.dust3r.utils.device.todevice#L11-L36","kind":"function","name":"todevice","path":"src/dust3r/utils/device.py","language":"python","start_line":11,"end_line":36,"context_start_line":1,"context_end_line":56,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nimport torch\n\n\ndef todevice(batch, device, callback=None, non_blocking=False):\n    \"\"\"Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy).\n\n    batch: list, tuple, dict of tensors or other things\n    device: pytorch device or 'numpy'\n    callback: function that would be called on every sub-elements.\n    \"\"\"\n    if callback:\n        batch = callback(batch)\n\n    if isinstance(batch, dict):\n        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef to_cpu(x):\n    return todevice(x, \"cpu\")\n\n\ndef to_cuda(x):\n    return todevice(x, \"cuda\")\n\n\ndef collate_with_cat(whatever, lists=False):\n    if isinstance(whatever, dict):\n        return {k: collate_with_cat(vals, lists=lists) for k, vals in whatever.items()}","source_hash":"d6eed957d20ec85bcb142727736b58f5c034322d6fc610964924f0b95d3e977c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.device.to_numpy","uri":"program://Human3R/function/src.dust3r.utils.device.to_numpy#L42-L43","kind":"function","name":"to_numpy","path":"src/dust3r/utils/device.py","language":"python","start_line":42,"end_line":43,"context_start_line":22,"context_end_line":63,"code":"        return {k: todevice(v, device) for k, v in batch.items()}\n\n    if isinstance(batch, (tuple, list)):\n        return type(batch)(todevice(x, device) for x in batch)\n\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef to_cpu(x):\n    return todevice(x, \"cpu\")\n\n\ndef to_cuda(x):\n    return todevice(x, \"cuda\")\n\n\ndef collate_with_cat(whatever, lists=False):\n    if isinstance(whatever, dict):\n        return {k: collate_with_cat(vals, lists=lists) for k, vals in whatever.items()}\n\n    elif isinstance(whatever, (tuple, list)):\n        if len(whatever) == 0:\n            return whatever\n        elem = whatever[0]\n        T = type(whatever)\n","source_hash":"d6eed957d20ec85bcb142727736b58f5c034322d6fc610964924f0b95d3e977c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.device.to_cpu","uri":"program://Human3R/function/src.dust3r.utils.device.to_cpu#L46-L47","kind":"function","name":"to_cpu","path":"src/dust3r/utils/device.py","language":"python","start_line":46,"end_line":47,"context_start_line":26,"context_end_line":67,"code":"\n    x = batch\n    if device == \"numpy\":\n        if isinstance(x, torch.Tensor):\n            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef to_cpu(x):\n    return todevice(x, \"cpu\")\n\n\ndef to_cuda(x):\n    return todevice(x, \"cuda\")\n\n\ndef collate_with_cat(whatever, lists=False):\n    if isinstance(whatever, dict):\n        return {k: collate_with_cat(vals, lists=lists) for k, vals in whatever.items()}\n\n    elif isinstance(whatever, (tuple, list)):\n        if len(whatever) == 0:\n            return whatever\n        elem = whatever[0]\n        T = type(whatever)\n\n        if elem is None:\n            return None\n        if isinstance(elem, (bool, float, int, str)):\n            return whatever","source_hash":"d6eed957d20ec85bcb142727736b58f5c034322d6fc610964924f0b95d3e977c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.device.to_cuda","uri":"program://Human3R/function/src.dust3r.utils.device.to_cuda#L50-L51","kind":"function","name":"to_cuda","path":"src/dust3r/utils/device.py","language":"python","start_line":50,"end_line":51,"context_start_line":30,"context_end_line":71,"code":"            x = x.detach().cpu().numpy()\n    elif x is not None:\n        if isinstance(x, np.ndarray):\n            x = torch.from_numpy(x)\n        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef to_cpu(x):\n    return todevice(x, \"cpu\")\n\n\ndef to_cuda(x):\n    return todevice(x, \"cuda\")\n\n\ndef collate_with_cat(whatever, lists=False):\n    if isinstance(whatever, dict):\n        return {k: collate_with_cat(vals, lists=lists) for k, vals in whatever.items()}\n\n    elif isinstance(whatever, (tuple, list)):\n        if len(whatever) == 0:\n            return whatever\n        elem = whatever[0]\n        T = type(whatever)\n\n        if elem is None:\n            return None\n        if isinstance(elem, (bool, float, int, str)):\n            return whatever\n        if isinstance(elem, tuple):\n            return T(collate_with_cat(x, lists=lists) for x in zip(*whatever))\n        if isinstance(elem, dict):\n            return {","source_hash":"d6eed957d20ec85bcb142727736b58f5c034322d6fc610964924f0b95d3e977c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.device.collate_with_cat","uri":"program://Human3R/function/src.dust3r.utils.device.collate_with_cat#L54-L84","kind":"function","name":"collate_with_cat","path":"src/dust3r/utils/device.py","language":"python","start_line":54,"end_line":84,"context_start_line":34,"context_end_line":104,"code":"        if torch.is_tensor(x):\n            x = x.to(device, non_blocking=non_blocking)\n    return x\n\n\nto_device = todevice  # alias\n\n\ndef to_numpy(x):\n    return todevice(x, \"numpy\")\n\n\ndef to_cpu(x):\n    return todevice(x, \"cpu\")\n\n\ndef to_cuda(x):\n    return todevice(x, \"cuda\")\n\n\ndef collate_with_cat(whatever, lists=False):\n    if isinstance(whatever, dict):\n        return {k: collate_with_cat(vals, lists=lists) for k, vals in whatever.items()}\n\n    elif isinstance(whatever, (tuple, list)):\n        if len(whatever) == 0:\n            return whatever\n        elem = whatever[0]\n        T = type(whatever)\n\n        if elem is None:\n            return None\n        if isinstance(elem, (bool, float, int, str)):\n            return whatever\n        if isinstance(elem, tuple):\n            return T(collate_with_cat(x, lists=lists) for x in zip(*whatever))\n        if isinstance(elem, dict):\n            return {\n                k: collate_with_cat([e[k] for e in whatever], lists=lists) for k in elem\n            }\n\n        if isinstance(elem, torch.Tensor):\n            return listify(whatever) if lists else torch.cat(whatever)\n        if isinstance(elem, np.ndarray):\n            return (\n                listify(whatever)\n                if lists\n                else torch.cat([torch.from_numpy(x) for x in whatever])\n            )\n\n        return sum(whatever, T())\n\n\ndef listify(elems):\n    return [x for e in elems for x in e]\n\n\ndef to_gpu(_view, device):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"rng\", \n         \"ray_map\", \"camera_pose\", \"camera_intrinsics\", \"ray_mask\", \"fov_x\",\n         \"fov_y\", \"T_w2c\", \"smpl_v3d_w\", \"smpl_j3d_w\", \"smpl_v3d_c\", \"smpl_j3d_c\",\n         \"smpl_j2d\", \"smpl_v2d\", \"smpl_mask\", \"msk\",\n         ]\n    )\n    view = {}\n    for name in _view.keys():  # pseudo_focal\n        if name in ignore_keys:\n            continue\n        if isinstance(_view[name], tuple) or isinstance(_view[name], list):\n            view[name] = [x.clone().to(device, non_blocking=True) for x in _view[name]]","source_hash":"d6eed957d20ec85bcb142727736b58f5c034322d6fc610964924f0b95d3e977c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.device.listify","uri":"program://Human3R/function/src.dust3r.utils.device.listify#L87-L88","kind":"function","name":"listify","path":"src/dust3r/utils/device.py","language":"python","start_line":87,"end_line":88,"context_start_line":67,"context_end_line":108,"code":"            return whatever\n        if isinstance(elem, tuple):\n            return T(collate_with_cat(x, lists=lists) for x in zip(*whatever))\n        if isinstance(elem, dict):\n            return {\n                k: collate_with_cat([e[k] for e in whatever], lists=lists) for k in elem\n            }\n\n        if isinstance(elem, torch.Tensor):\n            return listify(whatever) if lists else torch.cat(whatever)\n        if isinstance(elem, np.ndarray):\n            return (\n                listify(whatever)\n                if lists\n                else torch.cat([torch.from_numpy(x) for x in whatever])\n            )\n\n        return sum(whatever, T())\n\n\ndef listify(elems):\n    return [x for e in elems for x in e]\n\n\ndef to_gpu(_view, device):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"rng\", \n         \"ray_map\", \"camera_pose\", \"camera_intrinsics\", \"ray_mask\", \"fov_x\",\n         \"fov_y\", \"T_w2c\", \"smpl_v3d_w\", \"smpl_j3d_w\", \"smpl_v3d_c\", \"smpl_j3d_c\",\n         \"smpl_j2d\", \"smpl_v2d\", \"smpl_mask\", \"msk\",\n         ]\n    )\n    view = {}\n    for name in _view.keys():  # pseudo_focal\n        if name in ignore_keys:\n            continue\n        if isinstance(_view[name], tuple) or isinstance(_view[name], list):\n            view[name] = [x.clone().to(device, non_blocking=True) for x in _view[name]]\n        else:\n            view[name] = _view[name].clone().to(device, non_blocking=True)\n    \n    return view","source_hash":"d6eed957d20ec85bcb142727736b58f5c034322d6fc610964924f0b95d3e977c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.device.to_gpu","uri":"program://Human3R/function/src.dust3r.utils.device.to_gpu#L91-L108","kind":"function","name":"to_gpu","path":"src/dust3r/utils/device.py","language":"python","start_line":91,"end_line":108,"context_start_line":71,"context_end_line":108,"code":"            return {\n                k: collate_with_cat([e[k] for e in whatever], lists=lists) for k in elem\n            }\n\n        if isinstance(elem, torch.Tensor):\n            return listify(whatever) if lists else torch.cat(whatever)\n        if isinstance(elem, np.ndarray):\n            return (\n                listify(whatever)\n                if lists\n                else torch.cat([torch.from_numpy(x) for x in whatever])\n            )\n\n        return sum(whatever, T())\n\n\ndef listify(elems):\n    return [x for e in elems for x in e]\n\n\ndef to_gpu(_view, device):\n    ignore_keys = set(\n        [\"depthmap\", \"dataset\", \"label\", \"instance\", \"idx\", \"rng\", \n         \"ray_map\", \"camera_pose\", \"camera_intrinsics\", \"ray_mask\", \"fov_x\",\n         \"fov_y\", \"T_w2c\", \"smpl_v3d_w\", \"smpl_j3d_w\", \"smpl_v3d_c\", \"smpl_j3d_c\",\n         \"smpl_j2d\", \"smpl_v2d\", \"smpl_mask\", \"msk\",\n         ]\n    )\n    view = {}\n    for name in _view.keys():  # pseudo_focal\n        if name in ignore_keys:\n            continue\n        if isinstance(_view[name], tuple) or isinstance(_view[name], list):\n            view[name] = [x.clone().to(device, non_blocking=True) for x in _view[name]]\n        else:\n            view[name] = _view[name].clone().to(device, non_blocking=True)\n    \n    return view","source_hash":"d6eed957d20ec85bcb142727736b58f5c034322d6fc610964924f0b95d3e977c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc","uri":"program://Human3R/module/src.dust3r.utils.misc#L1-L135","kind":"module","name":"src.dust3r.utils.misc","path":"src/dust3r/utils/misc.py","language":"python","start_line":1,"end_line":135,"context_start_line":1,"context_end_line":135,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\n\n\ndef fill_default_args(kwargs, func):\n    import inspect  # a bit hacky but it works reliably\n\n    signature = inspect.signature(func)\n\n    for k, v in signature.parameters.items():\n        if v.default is inspect.Parameter.empty:\n            continue\n        kwargs.setdefault(k, v.default)\n\n    return kwargs\n\n\ndef freeze_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param.requires_grad = False\n        except AttributeError:\n\n            module.requires_grad = False\n\ndef fix_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param._is_frozen = True\n        except AttributeError:\n            module._is_frozen = True\n\n\ndef is_symmetrized(gt1, gt2):\n    x = gt1[\"instance\"]\n    y = gt2[\"instance\"]\n    if len(x) == len(y) and len(x) == 1:\n        return False  # special case of batchsize 1\n    ok = True\n    for i in range(0, len(x), 2):\n        ok = ok and (x[i] == y[i + 1]) and (x[i + 1] == y[i])\n    return ok\n\n\ndef flip(tensor):\n    \"\"\"flip so that tensor[0::2] <=> tensor[1::2]\"\"\"\n    return torch.stack((tensor[1::2], tensor[0::2]), dim=1).flatten(0, 1)\n\n\ndef interleave(tensor1, tensor2):\n    res1 = torch.stack((tensor1, tensor2), dim=1).flatten(0, 1)\n    res2 = torch.stack((tensor2, tensor1), dim=1).flatten(0, 1)\n    return res1, res2\n\n\ndef transpose_to_landscape(head, activate=True):\n    \"\"\"Predict in the correct aspect-ratio,\n    then transpose the result in landscape\n    and stack everything back together.\n    \"\"\"\n\n    def wrapper_no(decout, true_shape, **kwargs):\n        B = len(true_shape)\n        assert true_shape[0:1].allclose(true_shape), \"true_shape must be all identical\"\n        H, W = true_shape[0].cpu().tolist()\n        res = head(decout, (H, W), **kwargs)\n        return res\n\n    def wrapper_yes(decout, true_shape, **kwargs):\n        B = len(true_shape)\n\n        H, W = int(true_shape.min()), int(true_shape.max())\n\n        height, width = true_shape.T\n        is_landscape = width >= height\n        is_portrait = ~is_landscape\n\n        if is_landscape.all():\n            return head(decout, (H, W), **kwargs)\n        if is_portrait.all():\n            return transposed(head(decout, (W, H), **kwargs))\n\n        def selout(ar):\n            return [d[ar] for d in decout]\n\n        if \"pos\" in kwargs:\n            kwargs_landscape = kwargs.copy()\n            kwargs_landscape[\"pos\"] = kwargs[\"pos\"][is_landscape]\n            kwargs_portrait = kwargs.copy()\n            kwargs_portrait[\"pos\"] = kwargs[\"pos\"][is_portrait]\n        l_result = head(selout(is_landscape), (H, W), **kwargs_landscape)\n        p_result = transposed(head(selout(is_portrait), (W, H), **kwargs_portrait))\n\n        result = {}\n        for k in l_result | p_result:\n            x = l_result[k].new(B, *l_result[k].shape[1:])\n            x[is_landscape] = l_result[k]\n            x[is_portrait] = p_result[k]\n            result[k] = x\n\n        return result\n\n    return wrapper_yes if activate else wrapper_no\n\n\ndef transposed(dic):\n    return {k: v.swapaxes(1, 2) if v.ndim > 2 else v for k, v in dic.items()}\n\n\ndef invalid_to_nans(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = float(\"nan\")\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)\n    return arr\n\n\ndef invalid_to_zeros(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = 0\n        nnz = valid_mask.view(len(valid_mask), -1).sum(1)\n    else:\n        nnz = arr.numel() // len(arr) if len(arr) else 0  # number of point per image\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)\n    return arr, nnz","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.fill_default_args","uri":"program://Human3R/function/src.dust3r.utils.misc.fill_default_args#L10-L20","kind":"function","name":"fill_default_args","path":"src/dust3r/utils/misc.py","language":"python","start_line":10,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\n\n\ndef fill_default_args(kwargs, func):\n    import inspect  # a bit hacky but it works reliably\n\n    signature = inspect.signature(func)\n\n    for k, v in signature.parameters.items():\n        if v.default is inspect.Parameter.empty:\n            continue\n        kwargs.setdefault(k, v.default)\n\n    return kwargs\n\n\ndef freeze_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param.requires_grad = False\n        except AttributeError:\n\n            module.requires_grad = False\n\ndef fix_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param._is_frozen = True\n        except AttributeError:\n            module._is_frozen = True\n\n","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.freeze_all_params","uri":"program://Human3R/function/src.dust3r.utils.misc.freeze_all_params#L23-L30","kind":"function","name":"freeze_all_params","path":"src/dust3r/utils/misc.py","language":"python","start_line":23,"end_line":30,"context_start_line":3,"context_end_line":50,"code":"#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\n\n\ndef fill_default_args(kwargs, func):\n    import inspect  # a bit hacky but it works reliably\n\n    signature = inspect.signature(func)\n\n    for k, v in signature.parameters.items():\n        if v.default is inspect.Parameter.empty:\n            continue\n        kwargs.setdefault(k, v.default)\n\n    return kwargs\n\n\ndef freeze_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param.requires_grad = False\n        except AttributeError:\n\n            module.requires_grad = False\n\ndef fix_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param._is_frozen = True\n        except AttributeError:\n            module._is_frozen = True\n\n\ndef is_symmetrized(gt1, gt2):\n    x = gt1[\"instance\"]\n    y = gt2[\"instance\"]\n    if len(x) == len(y) and len(x) == 1:\n        return False  # special case of batchsize 1\n    ok = True\n    for i in range(0, len(x), 2):\n        ok = ok and (x[i] == y[i + 1]) and (x[i + 1] == y[i])\n    return ok\n","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.fix_all_params","uri":"program://Human3R/function/src.dust3r.utils.misc.fix_all_params#L32-L38","kind":"function","name":"fix_all_params","path":"src/dust3r/utils/misc.py","language":"python","start_line":32,"end_line":38,"context_start_line":12,"context_end_line":58,"code":"\n    signature = inspect.signature(func)\n\n    for k, v in signature.parameters.items():\n        if v.default is inspect.Parameter.empty:\n            continue\n        kwargs.setdefault(k, v.default)\n\n    return kwargs\n\n\ndef freeze_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param.requires_grad = False\n        except AttributeError:\n\n            module.requires_grad = False\n\ndef fix_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param._is_frozen = True\n        except AttributeError:\n            module._is_frozen = True\n\n\ndef is_symmetrized(gt1, gt2):\n    x = gt1[\"instance\"]\n    y = gt2[\"instance\"]\n    if len(x) == len(y) and len(x) == 1:\n        return False  # special case of batchsize 1\n    ok = True\n    for i in range(0, len(x), 2):\n        ok = ok and (x[i] == y[i + 1]) and (x[i + 1] == y[i])\n    return ok\n\n\ndef flip(tensor):\n    \"\"\"flip so that tensor[0::2] <=> tensor[1::2]\"\"\"\n    return torch.stack((tensor[1::2], tensor[0::2]), dim=1).flatten(0, 1)\n\n\ndef interleave(tensor1, tensor2):\n    res1 = torch.stack((tensor1, tensor2), dim=1).flatten(0, 1)","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.is_symmetrized","uri":"program://Human3R/function/src.dust3r.utils.misc.is_symmetrized#L41-L49","kind":"function","name":"is_symmetrized","path":"src/dust3r/utils/misc.py","language":"python","start_line":41,"end_line":49,"context_start_line":21,"context_end_line":69,"code":"\n\ndef freeze_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param.requires_grad = False\n        except AttributeError:\n\n            module.requires_grad = False\n\ndef fix_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param._is_frozen = True\n        except AttributeError:\n            module._is_frozen = True\n\n\ndef is_symmetrized(gt1, gt2):\n    x = gt1[\"instance\"]\n    y = gt2[\"instance\"]\n    if len(x) == len(y) and len(x) == 1:\n        return False  # special case of batchsize 1\n    ok = True\n    for i in range(0, len(x), 2):\n        ok = ok and (x[i] == y[i + 1]) and (x[i + 1] == y[i])\n    return ok\n\n\ndef flip(tensor):\n    \"\"\"flip so that tensor[0::2] <=> tensor[1::2]\"\"\"\n    return torch.stack((tensor[1::2], tensor[0::2]), dim=1).flatten(0, 1)\n\n\ndef interleave(tensor1, tensor2):\n    res1 = torch.stack((tensor1, tensor2), dim=1).flatten(0, 1)\n    res2 = torch.stack((tensor2, tensor1), dim=1).flatten(0, 1)\n    return res1, res2\n\n\ndef transpose_to_landscape(head, activate=True):\n    \"\"\"Predict in the correct aspect-ratio,\n    then transpose the result in landscape\n    and stack everything back together.\n    \"\"\"\n\n    def wrapper_no(decout, true_shape, **kwargs):","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.flip","uri":"program://Human3R/function/src.dust3r.utils.misc.flip#L52-L54","kind":"function","name":"flip","path":"src/dust3r/utils/misc.py","language":"python","start_line":52,"end_line":54,"context_start_line":32,"context_end_line":74,"code":"def fix_all_params(modules):\n    for module in modules:\n        try:\n            for n, param in module.named_parameters():\n                param._is_frozen = True\n        except AttributeError:\n            module._is_frozen = True\n\n\ndef is_symmetrized(gt1, gt2):\n    x = gt1[\"instance\"]\n    y = gt2[\"instance\"]\n    if len(x) == len(y) and len(x) == 1:\n        return False  # special case of batchsize 1\n    ok = True\n    for i in range(0, len(x), 2):\n        ok = ok and (x[i] == y[i + 1]) and (x[i + 1] == y[i])\n    return ok\n\n\ndef flip(tensor):\n    \"\"\"flip so that tensor[0::2] <=> tensor[1::2]\"\"\"\n    return torch.stack((tensor[1::2], tensor[0::2]), dim=1).flatten(0, 1)\n\n\ndef interleave(tensor1, tensor2):\n    res1 = torch.stack((tensor1, tensor2), dim=1).flatten(0, 1)\n    res2 = torch.stack((tensor2, tensor1), dim=1).flatten(0, 1)\n    return res1, res2\n\n\ndef transpose_to_landscape(head, activate=True):\n    \"\"\"Predict in the correct aspect-ratio,\n    then transpose the result in landscape\n    and stack everything back together.\n    \"\"\"\n\n    def wrapper_no(decout, true_shape, **kwargs):\n        B = len(true_shape)\n        assert true_shape[0:1].allclose(true_shape), \"true_shape must be all identical\"\n        H, W = true_shape[0].cpu().tolist()\n        res = head(decout, (H, W), **kwargs)\n        return res","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.interleave","uri":"program://Human3R/function/src.dust3r.utils.misc.interleave#L57-L60","kind":"function","name":"interleave","path":"src/dust3r/utils/misc.py","language":"python","start_line":57,"end_line":60,"context_start_line":37,"context_end_line":80,"code":"        except AttributeError:\n            module._is_frozen = True\n\n\ndef is_symmetrized(gt1, gt2):\n    x = gt1[\"instance\"]\n    y = gt2[\"instance\"]\n    if len(x) == len(y) and len(x) == 1:\n        return False  # special case of batchsize 1\n    ok = True\n    for i in range(0, len(x), 2):\n        ok = ok and (x[i] == y[i + 1]) and (x[i + 1] == y[i])\n    return ok\n\n\ndef flip(tensor):\n    \"\"\"flip so that tensor[0::2] <=> tensor[1::2]\"\"\"\n    return torch.stack((tensor[1::2], tensor[0::2]), dim=1).flatten(0, 1)\n\n\ndef interleave(tensor1, tensor2):\n    res1 = torch.stack((tensor1, tensor2), dim=1).flatten(0, 1)\n    res2 = torch.stack((tensor2, tensor1), dim=1).flatten(0, 1)\n    return res1, res2\n\n\ndef transpose_to_landscape(head, activate=True):\n    \"\"\"Predict in the correct aspect-ratio,\n    then transpose the result in landscape\n    and stack everything back together.\n    \"\"\"\n\n    def wrapper_no(decout, true_shape, **kwargs):\n        B = len(true_shape)\n        assert true_shape[0:1].allclose(true_shape), \"true_shape must be all identical\"\n        H, W = true_shape[0].cpu().tolist()\n        res = head(decout, (H, W), **kwargs)\n        return res\n\n    def wrapper_yes(decout, true_shape, **kwargs):\n        B = len(true_shape)\n\n        H, W = int(true_shape.min()), int(true_shape.max())\n","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.transpose_to_landscape","uri":"program://Human3R/function/src.dust3r.utils.misc.transpose_to_landscape#L63-L110","kind":"function","name":"transpose_to_landscape","path":"src/dust3r/utils/misc.py","language":"python","start_line":63,"end_line":110,"context_start_line":43,"context_end_line":130,"code":"    y = gt2[\"instance\"]\n    if len(x) == len(y) and len(x) == 1:\n        return False  # special case of batchsize 1\n    ok = True\n    for i in range(0, len(x), 2):\n        ok = ok and (x[i] == y[i + 1]) and (x[i + 1] == y[i])\n    return ok\n\n\ndef flip(tensor):\n    \"\"\"flip so that tensor[0::2] <=> tensor[1::2]\"\"\"\n    return torch.stack((tensor[1::2], tensor[0::2]), dim=1).flatten(0, 1)\n\n\ndef interleave(tensor1, tensor2):\n    res1 = torch.stack((tensor1, tensor2), dim=1).flatten(0, 1)\n    res2 = torch.stack((tensor2, tensor1), dim=1).flatten(0, 1)\n    return res1, res2\n\n\ndef transpose_to_landscape(head, activate=True):\n    \"\"\"Predict in the correct aspect-ratio,\n    then transpose the result in landscape\n    and stack everything back together.\n    \"\"\"\n\n    def wrapper_no(decout, true_shape, **kwargs):\n        B = len(true_shape)\n        assert true_shape[0:1].allclose(true_shape), \"true_shape must be all identical\"\n        H, W = true_shape[0].cpu().tolist()\n        res = head(decout, (H, W), **kwargs)\n        return res\n\n    def wrapper_yes(decout, true_shape, **kwargs):\n        B = len(true_shape)\n\n        H, W = int(true_shape.min()), int(true_shape.max())\n\n        height, width = true_shape.T\n        is_landscape = width >= height\n        is_portrait = ~is_landscape\n\n        if is_landscape.all():\n            return head(decout, (H, W), **kwargs)\n        if is_portrait.all():\n            return transposed(head(decout, (W, H), **kwargs))\n\n        def selout(ar):\n            return [d[ar] for d in decout]\n\n        if \"pos\" in kwargs:\n            kwargs_landscape = kwargs.copy()\n            kwargs_landscape[\"pos\"] = kwargs[\"pos\"][is_landscape]\n            kwargs_portrait = kwargs.copy()\n            kwargs_portrait[\"pos\"] = kwargs[\"pos\"][is_portrait]\n        l_result = head(selout(is_landscape), (H, W), **kwargs_landscape)\n        p_result = transposed(head(selout(is_portrait), (W, H), **kwargs_portrait))\n\n        result = {}\n        for k in l_result | p_result:\n            x = l_result[k].new(B, *l_result[k].shape[1:])\n            x[is_landscape] = l_result[k]\n            x[is_portrait] = p_result[k]\n            result[k] = x\n\n        return result\n\n    return wrapper_yes if activate else wrapper_no\n\n\ndef transposed(dic):\n    return {k: v.swapaxes(1, 2) if v.ndim > 2 else v for k, v in dic.items()}\n\n\ndef invalid_to_nans(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = float(\"nan\")\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)\n    return arr\n\n\ndef invalid_to_zeros(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = 0\n        nnz = valid_mask.view(len(valid_mask), -1).sum(1)","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.transposed","uri":"program://Human3R/function/src.dust3r.utils.misc.transposed#L113-L114","kind":"function","name":"transposed","path":"src/dust3r/utils/misc.py","language":"python","start_line":113,"end_line":114,"context_start_line":93,"context_end_line":134,"code":"        if \"pos\" in kwargs:\n            kwargs_landscape = kwargs.copy()\n            kwargs_landscape[\"pos\"] = kwargs[\"pos\"][is_landscape]\n            kwargs_portrait = kwargs.copy()\n            kwargs_portrait[\"pos\"] = kwargs[\"pos\"][is_portrait]\n        l_result = head(selout(is_landscape), (H, W), **kwargs_landscape)\n        p_result = transposed(head(selout(is_portrait), (W, H), **kwargs_portrait))\n\n        result = {}\n        for k in l_result | p_result:\n            x = l_result[k].new(B, *l_result[k].shape[1:])\n            x[is_landscape] = l_result[k]\n            x[is_portrait] = p_result[k]\n            result[k] = x\n\n        return result\n\n    return wrapper_yes if activate else wrapper_no\n\n\ndef transposed(dic):\n    return {k: v.swapaxes(1, 2) if v.ndim > 2 else v for k, v in dic.items()}\n\n\ndef invalid_to_nans(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = float(\"nan\")\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)\n    return arr\n\n\ndef invalid_to_zeros(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = 0\n        nnz = valid_mask.view(len(valid_mask), -1).sum(1)\n    else:\n        nnz = arr.numel() // len(arr) if len(arr) else 0  # number of point per image\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.invalid_to_nans","uri":"program://Human3R/function/src.dust3r.utils.misc.invalid_to_nans#L117-L123","kind":"function","name":"invalid_to_nans","path":"src/dust3r/utils/misc.py","language":"python","start_line":117,"end_line":123,"context_start_line":97,"context_end_line":135,"code":"            kwargs_portrait[\"pos\"] = kwargs[\"pos\"][is_portrait]\n        l_result = head(selout(is_landscape), (H, W), **kwargs_landscape)\n        p_result = transposed(head(selout(is_portrait), (W, H), **kwargs_portrait))\n\n        result = {}\n        for k in l_result | p_result:\n            x = l_result[k].new(B, *l_result[k].shape[1:])\n            x[is_landscape] = l_result[k]\n            x[is_portrait] = p_result[k]\n            result[k] = x\n\n        return result\n\n    return wrapper_yes if activate else wrapper_no\n\n\ndef transposed(dic):\n    return {k: v.swapaxes(1, 2) if v.ndim > 2 else v for k, v in dic.items()}\n\n\ndef invalid_to_nans(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = float(\"nan\")\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)\n    return arr\n\n\ndef invalid_to_zeros(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = 0\n        nnz = valid_mask.view(len(valid_mask), -1).sum(1)\n    else:\n        nnz = arr.numel() // len(arr) if len(arr) else 0  # number of point per image\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)\n    return arr, nnz","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.invalid_to_zeros","uri":"program://Human3R/function/src.dust3r.utils.misc.invalid_to_zeros#L126-L135","kind":"function","name":"invalid_to_zeros","path":"src/dust3r/utils/misc.py","language":"python","start_line":126,"end_line":135,"context_start_line":106,"context_end_line":135,"code":"            result[k] = x\n\n        return result\n\n    return wrapper_yes if activate else wrapper_no\n\n\ndef transposed(dic):\n    return {k: v.swapaxes(1, 2) if v.ndim > 2 else v for k, v in dic.items()}\n\n\ndef invalid_to_nans(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = float(\"nan\")\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)\n    return arr\n\n\ndef invalid_to_zeros(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = 0\n        nnz = valid_mask.view(len(valid_mask), -1).sum(1)\n    else:\n        nnz = arr.numel() // len(arr) if len(arr) else 0  # number of point per image\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)\n    return arr, nnz","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.wrapper_no","uri":"program://Human3R/function/src.dust3r.utils.misc.wrapper_no#L69-L74","kind":"function","name":"wrapper_no","path":"src/dust3r/utils/misc.py","language":"python","start_line":69,"end_line":74,"context_start_line":49,"context_end_line":94,"code":"    return ok\n\n\ndef flip(tensor):\n    \"\"\"flip so that tensor[0::2] <=> tensor[1::2]\"\"\"\n    return torch.stack((tensor[1::2], tensor[0::2]), dim=1).flatten(0, 1)\n\n\ndef interleave(tensor1, tensor2):\n    res1 = torch.stack((tensor1, tensor2), dim=1).flatten(0, 1)\n    res2 = torch.stack((tensor2, tensor1), dim=1).flatten(0, 1)\n    return res1, res2\n\n\ndef transpose_to_landscape(head, activate=True):\n    \"\"\"Predict in the correct aspect-ratio,\n    then transpose the result in landscape\n    and stack everything back together.\n    \"\"\"\n\n    def wrapper_no(decout, true_shape, **kwargs):\n        B = len(true_shape)\n        assert true_shape[0:1].allclose(true_shape), \"true_shape must be all identical\"\n        H, W = true_shape[0].cpu().tolist()\n        res = head(decout, (H, W), **kwargs)\n        return res\n\n    def wrapper_yes(decout, true_shape, **kwargs):\n        B = len(true_shape)\n\n        H, W = int(true_shape.min()), int(true_shape.max())\n\n        height, width = true_shape.T\n        is_landscape = width >= height\n        is_portrait = ~is_landscape\n\n        if is_landscape.all():\n            return head(decout, (H, W), **kwargs)\n        if is_portrait.all():\n            return transposed(head(decout, (W, H), **kwargs))\n\n        def selout(ar):\n            return [d[ar] for d in decout]\n\n        if \"pos\" in kwargs:\n            kwargs_landscape = kwargs.copy()","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.wrapper_yes","uri":"program://Human3R/function/src.dust3r.utils.misc.wrapper_yes#L76-L108","kind":"function","name":"wrapper_yes","path":"src/dust3r/utils/misc.py","language":"python","start_line":76,"end_line":108,"context_start_line":56,"context_end_line":128,"code":"\ndef interleave(tensor1, tensor2):\n    res1 = torch.stack((tensor1, tensor2), dim=1).flatten(0, 1)\n    res2 = torch.stack((tensor2, tensor1), dim=1).flatten(0, 1)\n    return res1, res2\n\n\ndef transpose_to_landscape(head, activate=True):\n    \"\"\"Predict in the correct aspect-ratio,\n    then transpose the result in landscape\n    and stack everything back together.\n    \"\"\"\n\n    def wrapper_no(decout, true_shape, **kwargs):\n        B = len(true_shape)\n        assert true_shape[0:1].allclose(true_shape), \"true_shape must be all identical\"\n        H, W = true_shape[0].cpu().tolist()\n        res = head(decout, (H, W), **kwargs)\n        return res\n\n    def wrapper_yes(decout, true_shape, **kwargs):\n        B = len(true_shape)\n\n        H, W = int(true_shape.min()), int(true_shape.max())\n\n        height, width = true_shape.T\n        is_landscape = width >= height\n        is_portrait = ~is_landscape\n\n        if is_landscape.all():\n            return head(decout, (H, W), **kwargs)\n        if is_portrait.all():\n            return transposed(head(decout, (W, H), **kwargs))\n\n        def selout(ar):\n            return [d[ar] for d in decout]\n\n        if \"pos\" in kwargs:\n            kwargs_landscape = kwargs.copy()\n            kwargs_landscape[\"pos\"] = kwargs[\"pos\"][is_landscape]\n            kwargs_portrait = kwargs.copy()\n            kwargs_portrait[\"pos\"] = kwargs[\"pos\"][is_portrait]\n        l_result = head(selout(is_landscape), (H, W), **kwargs_landscape)\n        p_result = transposed(head(selout(is_portrait), (W, H), **kwargs_portrait))\n\n        result = {}\n        for k in l_result | p_result:\n            x = l_result[k].new(B, *l_result[k].shape[1:])\n            x[is_landscape] = l_result[k]\n            x[is_portrait] = p_result[k]\n            result[k] = x\n\n        return result\n\n    return wrapper_yes if activate else wrapper_no\n\n\ndef transposed(dic):\n    return {k: v.swapaxes(1, 2) if v.ndim > 2 else v for k, v in dic.items()}\n\n\ndef invalid_to_nans(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()\n        arr[~valid_mask] = float(\"nan\")\n    if arr.ndim > ndim:\n        arr = arr.flatten(-2 - (arr.ndim - ndim), -2)\n    return arr\n\n\ndef invalid_to_zeros(arr, valid_mask, ndim=999):\n    if valid_mask is not None:\n        arr = arr.clone()","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.misc.selout","uri":"program://Human3R/function/src.dust3r.utils.misc.selout#L90-L91","kind":"function","name":"selout","path":"src/dust3r/utils/misc.py","language":"python","start_line":90,"end_line":91,"context_start_line":70,"context_end_line":111,"code":"        B = len(true_shape)\n        assert true_shape[0:1].allclose(true_shape), \"true_shape must be all identical\"\n        H, W = true_shape[0].cpu().tolist()\n        res = head(decout, (H, W), **kwargs)\n        return res\n\n    def wrapper_yes(decout, true_shape, **kwargs):\n        B = len(true_shape)\n\n        H, W = int(true_shape.min()), int(true_shape.max())\n\n        height, width = true_shape.T\n        is_landscape = width >= height\n        is_portrait = ~is_landscape\n\n        if is_landscape.all():\n            return head(decout, (H, W), **kwargs)\n        if is_portrait.all():\n            return transposed(head(decout, (W, H), **kwargs))\n\n        def selout(ar):\n            return [d[ar] for d in decout]\n\n        if \"pos\" in kwargs:\n            kwargs_landscape = kwargs.copy()\n            kwargs_landscape[\"pos\"] = kwargs[\"pos\"][is_landscape]\n            kwargs_portrait = kwargs.copy()\n            kwargs_portrait[\"pos\"] = kwargs[\"pos\"][is_portrait]\n        l_result = head(selout(is_landscape), (H, W), **kwargs_landscape)\n        p_result = transposed(head(selout(is_portrait), (W, H), **kwargs_portrait))\n\n        result = {}\n        for k in l_result | p_result:\n            x = l_result[k].new(B, *l_result[k].shape[1:])\n            x[is_landscape] = l_result[k]\n            x[is_portrait] = p_result[k]\n            result[k] = x\n\n        return result\n\n    return wrapper_yes if activate else wrapper_no\n","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera","uri":"program://Human3R/module/src.dust3r.utils.camera#L1-L462","kind":"module","name":"src.dust3r.utils.camera","path":"src/dust3r/utils/camera.py","language":"python","start_line":1,"end_line":462,"context_start_line":1,"context_end_line":462,"code":"from typing import Optional\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom croco.models.blocks import Mlp\nfrom dust3r.heads.postprocess import postprocess_pose\n\ninf = float(\"inf\")\n\n\nclass PoseDecoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        mlp_ratio=4,\n        pose_encoding_type=\"absT_quaR\",\n    ):\n        super().__init__()\n\n        self.pose_encoding_type = pose_encoding_type\n        if self.pose_encoding_type == \"absT_quaR\":\n            self.target_dim = 7\n\n        self.mlp = Mlp(\n            in_features=hidden_size,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=self.target_dim,\n            drop=0,\n        )\n\n    def forward(\n        self,\n        pose_feat,\n    ):\n        \"\"\"\n        pose_feat: BxC\n        preliminary_cameras: cameras in opencv coordinate.\n        \"\"\"\n\n        pred_cameras = self.mlp(pose_feat)  # Bx7, 3 for absT, 4 for quaR\n        return pred_cameras\n\n\nclass PoseEncoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        mlp_ratio=4,\n        pose_mode=(\"exp\", -inf, inf),\n        pose_encoding_type=\"absT_quaR\",\n    ):\n        super().__init__()\n        self.pose_encoding_type = pose_encoding_type\n        self.pose_mode = pose_mode\n\n        if self.pose_encoding_type == \"absT_quaR\":\n            self.target_dim = 7\n\n        self.embed_pose = PoseEmbedding(\n            target_dim=self.target_dim,\n            out_dim=hidden_size,\n            n_harmonic_functions=10,\n            append_input=True,\n        )\n        self.pose_encoder = Mlp(\n            in_features=self.embed_pose.out_dim,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=hidden_size,\n            drop=0,\n        )\n\n    def forward(self, camera):\n        pose_enc = camera_to_pose_encoding(\n            camera,\n            pose_encoding_type=self.pose_encoding_type,\n        ).to(camera.dtype)\n        pose_enc = postprocess_pose(pose_enc, self.pose_mode, inverse=True)\n        pose_feat = self.embed_pose(pose_enc)\n        pose_feat = self.pose_encoder(pose_feat)\n        return pose_feat\n\n\nclass HarmonicEmbedding(torch.nn.Module):\n    def __init__(\n        self,\n        n_harmonic_functions: int = 6,\n        omega_0: float = 1.0,\n        logspace: bool = True,\n        append_input: bool = True,\n    ) -> None:\n        \"\"\"\n        The harmonic embedding layer supports the classical\n        Nerf positional encoding described in\n        `NeRF <https://arxiv.org/abs/2003.08934>`_\n        and the integrated position encoding in\n        `MIP-NeRF <https://arxiv.org/abs/2103.13415>`_.\n\n        During the inference you can provide the extra argument `diag_cov`.\n\n        If `diag_cov is None`, it converts\n        rays parametrized with a `ray_bundle` to 3D points by\n        extending each ray according to the corresponding length.\n        Then it converts each feature\n        (i.e. vector along the last dimension) in `x`\n        into a series of harmonic features `embedding`,\n        where for each i in range(dim) the following are present\n        in embedding[...]::\n\n            [\n                sin(f_1*x[..., i]),\n                sin(f_2*x[..., i]),\n                ...\n                sin(f_N * x[..., i]),\n                cos(f_1*x[..., i]),\n                cos(f_2*x[..., i]),\n                ...\n                cos(f_N * x[..., i]),\n                x[..., i],              # only present if append_input is True.\n            ]\n\n        where N corresponds to `n_harmonic_functions-1`, and f_i is a scalar\n        denoting the i-th frequency of the harmonic embedding.\n\n\n        If `diag_cov is not None`, it approximates\n        conical frustums following a ray bundle as gaussians,\n        defined by x, the means of the gaussians and diag_cov,\n        the diagonal covariances.\n        Then it converts each gaussian\n        into a series of harmonic features `embedding`,\n        where for each i in range(dim) the following are present\n        in embedding[...]::\n\n            [\n                sin(f_1*x[..., i]) * exp(0.5 * f_1**2 * diag_cov[..., i,]),\n                sin(f_2*x[..., i]) * exp(0.5 * f_2**2 * diag_cov[..., i,]),\n                ...\n                sin(f_N * x[..., i]) * exp(0.5 * f_N**2 * diag_cov[..., i,]),\n                cos(f_1*x[..., i]) * exp(0.5 * f_1**2 * diag_cov[..., i,]),\n                cos(f_2*x[..., i]) * exp(0.5 * f_2**2 * diag_cov[..., i,]),,\n                ...\n                cos(f_N * x[..., i]) * exp(0.5 * f_N**2 * diag_cov[..., i,]),\n                x[..., i],              # only present if append_input is True.\n            ]\n\n        where N equals `n_harmonic_functions-1`, and f_i is a scalar\n        denoting the i-th frequency of the harmonic embedding.\n\n        If `logspace==True`, the frequencies `[f_1, ..., f_N]` are\n        powers of 2:\n            `f_1, ..., f_N = 2**torch.arange(n_harmonic_functions)`\n\n        If `logspace==False`, frequencies are linearly spaced between\n        `1.0` and `2**(n_harmonic_functions-1)`:\n            `f_1, ..., f_N = torch.linspace(\n                1.0, 2**(n_harmonic_functions-1), n_harmonic_functions\n            )`\n\n        Note that `x` is also premultiplied by the base frequency `omega_0`\n        before evaluating the harmonic functions.\n\n        Args:\n            n_harmonic_functions: int, number of harmonic\n                features\n            omega_0: float, base frequency\n            logspace: bool, Whether to space the frequencies in\n                logspace or linear space\n            append_input: bool, whether to concat the original\n                input to the harmonic embedding. If true the\n                output is of the form (embed.sin(), embed.cos(), x)\n        \"\"\"\n        super().__init__()\n\n        if logspace:\n            frequencies = 2.0 ** torch.arange(n_harmonic_functions, dtype=torch.float32)\n        else:\n            frequencies = torch.linspace(\n                1.0,\n                2.0 ** (n_harmonic_functions - 1),\n                n_harmonic_functions,\n                dtype=torch.float32,\n            )\n\n        self.register_buffer(\"_frequencies\", frequencies * omega_0, persistent=False)\n        self.register_buffer(\n            \"_zero_half_pi\",\n            torch.tensor([0.0, 0.5 * torch.pi]),\n            persistent=False,\n        )\n        self.append_input = append_input\n\n    def forward(\n        self, x: torch.Tensor, diag_cov: Optional[torch.Tensor] = None, **kwargs\n    ) -> torch.Tensor:\n        \"\"\"\n        Args:\n            x: tensor of shape [..., dim]\n            diag_cov: An optional tensor of shape `(..., dim)`\n                representing the diagonal covariance matrices of our Gaussians, joined with x\n                as means of the Gaussians.\n\n        Returns:\n            embedding: a harmonic embedding of `x` of shape\n            [..., (n_harmonic_functions * 2 + int(append_input)) * num_points_per_ray]\n        \"\"\"\n\n        embed = x[..., None] * self._frequencies\n\n        embed = embed[..., None, :, :] + self._zero_half_pi[..., None, None]\n\n        embed = embed.sin()\n        if diag_cov is not None:\n            x_var = diag_cov[..., None] * torch.pow(self._frequencies, 2)\n            exp_var = torch.exp(-0.5 * x_var)\n\n            embed = embed * exp_var[..., None, :, :]\n\n        embed = embed.reshape(*x.shape[:-1], -1)\n\n        if self.append_input:\n            return torch.cat([embed, x], dim=-1)\n        return embed\n\n    @staticmethod\n    def get_output_dim_static(\n        input_dims: int, n_harmonic_functions: int, append_input: bool\n    ) -> int:\n        \"\"\"\n        Utility to help predict the shape of the output of `forward`.\n\n        Args:\n            input_dims: length of the last dimension of the input tensor\n            n_harmonic_functions: number of embedding frequencies\n            append_input: whether or not to concat the original\n                input to the harmonic embedding\n        Returns:\n            int: the length of the last dimension of the output tensor\n        \"\"\"\n        return input_dims * (2 * n_harmonic_functions + int(append_input))\n\n    def get_output_dim(self, input_dims: int = 3) -> int:\n        \"\"\"\n        Same as above. The default for input_dims is 3 for 3D applications\n        which use harmonic embedding for positional encoding,\n        so the input might be xyz.\n        \"\"\"\n        return self.get_output_dim_static(\n            input_dims, len(self._frequencies), self.append_input\n        )\n\n\nclass PoseEmbedding(nn.Module):\n    def __init__(self, target_dim, out_dim, n_harmonic_functions=10, append_input=True):\n        super().__init__()\n\n        self._emb_pose = HarmonicEmbedding(\n            n_harmonic_functions=n_harmonic_functions, append_input=append_input\n        )\n\n        self.out_dim = self._emb_pose.get_output_dim(target_dim)\n\n    def forward(self, pose_encoding):\n        e_pose_encoding = self._emb_pose(pose_encoding)\n        return e_pose_encoding\n\n\ndef _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Returns torch.sqrt(torch.max(0, x))\n    but with a zero subgradient where x is 0.\n    \"\"\"\n    ret = torch.zeros_like(x)\n    positive_mask = x > 0\n    ret[positive_mask] = torch.sqrt(x[positive_mask])\n    return ret\n\n\ndef matrix_to_quaternion(matrix: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert rotations given as rotation matrices to quaternions.\n\n    Args:\n        matrix: Rotation matrices as tensor of shape (..., 3, 3).\n\n    Returns:\n        quaternions with real part first, as tensor of shape (..., 4).\n    \"\"\"\n    if matrix.size(-1) != 3 or matrix.size(-2) != 3:\n        raise ValueError(f\"Invalid rotation matrix shape {matrix.shape}.\")\n\n    batch_dim = matrix.shape[:-2]\n    m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind(\n        matrix.reshape(batch_dim + (9,)), dim=-1\n    )\n\n    q_abs = _sqrt_positive_part(\n        torch.stack(\n            [\n                1.0 + m00 + m11 + m22,\n                1.0 + m00 - m11 - m22,\n                1.0 - m00 + m11 - m22,\n                1.0 - m00 - m11 + m22,\n            ],\n            dim=-1,\n        )\n    )\n\n    quat_by_rijk = torch.stack(\n        [\n            torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1),\n            torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1),\n            torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1),\n            torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1),\n        ],\n        dim=-2,\n    )\n\n    flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device)\n    quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr))\n\n    out = quat_candidates[\n        F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, :\n    ].reshape(batch_dim + (4,))\n    return standardize_quaternion(out)\n\n\ndef standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert a unit quaternion to a standard form: one in which the real\n    part is non negative.\n\n    Args:\n        quaternions: Quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Standardized quaternions as tensor of shape (..., 4).\n    \"\"\"\n    quaternions = F.normalize(quaternions, p=2, dim=-1)\n    return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions)\n\n\ndef camera_to_pose_encoding(\n    camera,\n    pose_encoding_type=\"absT_quaR\",\n):\n    \"\"\"\n    Inverse to pose_encoding_to_camera\n    camera: opencv, cam2world\n    \"\"\"\n    if pose_encoding_type == \"absT_quaR\":\n\n        quaternion_R = matrix_to_quaternion(camera[:, :3, :3])\n\n        pose_encoding = torch.cat([camera[:, :3, 3], quaternion_R], dim=-1)\n    else:\n        raise ValueError(f\"Unknown pose encoding {pose_encoding_type}\")\n\n    return pose_encoding\n\n\ndef quaternion_to_matrix(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert rotations given as quaternions to rotation matrices.\n\n    Args:\n        quaternions: quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Rotation matrices as tensor of shape (..., 3, 3).\n    \"\"\"\n    r, i, j, k = torch.unbind(quaternions, -1)\n\n    two_s = 2.0 / (quaternions * quaternions).sum(-1)\n\n    o = torch.stack(\n        (\n            1 - two_s * (j * j + k * k),\n            two_s * (i * j - k * r),\n            two_s * (i * k + j * r),\n            two_s * (i * j + k * r),\n            1 - two_s * (i * i + k * k),\n            two_s * (j * k - i * r),\n            two_s * (i * k - j * r),\n            two_s * (j * k + i * r),\n            1 - two_s * (i * i + j * j),\n        ),\n        -1,\n    )\n    return o.reshape(quaternions.shape[:-1] + (3, 3))\n\n\ndef pose_encoding_to_camera(\n    pose_encoding,\n    pose_encoding_type=\"absT_quaR\",\n):\n    \"\"\"\n    Args:\n        pose_encoding: A tensor of shape `BxC`, containing a batch of\n                        `B` `C`-dimensional pose encodings.\n        pose_encoding_type: The type of pose encoding,\n    \"\"\"\n\n    if pose_encoding_type == \"absT_quaR\":\n\n        abs_T = pose_encoding[:, :3]\n        quaternion_R = pose_encoding[:, 3:7]\n        R = quaternion_to_matrix(quaternion_R)\n    else:\n        raise ValueError(f\"Unknown pose encoding {pose_encoding_type}\")\n\n    c2w_mats = torch.eye(4, 4).to(R.dtype).to(R.device)\n    c2w_mats = c2w_mats[None].repeat(len(R), 1, 1)\n    c2w_mats[:, :3, :3] = R\n    c2w_mats[:, :3, 3] = abs_T\n\n    return c2w_mats\n\n\ndef quaternion_conjugate(q):\n    \"\"\"Compute the conjugate of quaternion q (w, x, y, z).\"\"\"\n\n    q_conj = torch.cat([q[..., :1], -q[..., 1:]], dim=-1)\n    return q_conj\n\n\ndef quaternion_multiply(q1, q2):\n    \"\"\"Multiply two quaternions q1 and q2.\"\"\"\n    w1, x1, y1, z1 = q1.unbind(dim=-1)\n    w2, x2, y2, z2 = q2.unbind(dim=-1)\n\n    w = w1 * w2 - x1 * x2 - y1 * y2 - z1 * z2\n    x = w1 * x2 + x1 * w2 + y1 * z2 - z1 * y2\n    y = w1 * y2 - x1 * z2 + y1 * w2 + z1 * x2\n    z = w1 * z2 + x1 * y2 - y1 * x2 + z1 * w2\n\n    return torch.stack((w, x, y, z), dim=-1)\n\n\ndef rotate_vector(q, v):\n    \"\"\"Rotate vector v by quaternion q.\"\"\"\n    q_vec = q[..., 1:]\n    q_w = q[..., :1]\n\n    t = 2.0 * torch.cross(q_vec, v, dim=-1)\n    v_rot = v + q_w * t + torch.cross(q_vec, t, dim=-1)\n    return v_rot\n\n\ndef relative_pose_absT_quatR(t1, q1, t2, q2):\n    \"\"\"Compute the relative translation and quaternion between two poses.\"\"\"\n\n    q1_inv = quaternion_conjugate(q1)\n\n    q_rel = quaternion_multiply(q1_inv, q2)\n\n    delta_t = t2 - t1\n    t_rel = rotate_vector(q1_inv, delta_t)\n    return t_rel, q_rel","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.PoseDecoder","uri":"program://Human3R/class/src.dust3r.utils.camera.PoseDecoder#L13-L43","kind":"class","name":"PoseDecoder","path":"src/dust3r/utils/camera.py","language":"python","start_line":13,"end_line":43,"context_start_line":1,"context_end_line":63,"code":"from typing import Optional\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom croco.models.blocks import Mlp\nfrom dust3r.heads.postprocess import postprocess_pose\n\ninf = float(\"inf\")\n\n\nclass PoseDecoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        mlp_ratio=4,\n        pose_encoding_type=\"absT_quaR\",\n    ):\n        super().__init__()\n\n        self.pose_encoding_type = pose_encoding_type\n        if self.pose_encoding_type == \"absT_quaR\":\n            self.target_dim = 7\n\n        self.mlp = Mlp(\n            in_features=hidden_size,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=self.target_dim,\n            drop=0,\n        )\n\n    def forward(\n        self,\n        pose_feat,\n    ):\n        \"\"\"\n        pose_feat: BxC\n        preliminary_cameras: cameras in opencv coordinate.\n        \"\"\"\n\n        pred_cameras = self.mlp(pose_feat)  # Bx7, 3 for absT, 4 for quaR\n        return pred_cameras\n\n\nclass PoseEncoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        mlp_ratio=4,\n        pose_mode=(\"exp\", -inf, inf),\n        pose_encoding_type=\"absT_quaR\",\n    ):\n        super().__init__()\n        self.pose_encoding_type = pose_encoding_type\n        self.pose_mode = pose_mode\n\n        if self.pose_encoding_type == \"absT_quaR\":\n            self.target_dim = 7\n\n        self.embed_pose = PoseEmbedding(\n            target_dim=self.target_dim,\n            out_dim=hidden_size,","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.PoseEncoder","uri":"program://Human3R/class/src.dust3r.utils.camera.PoseEncoder#L46-L82","kind":"class","name":"PoseEncoder","path":"src/dust3r/utils/camera.py","language":"python","start_line":46,"end_line":82,"context_start_line":26,"context_end_line":102,"code":"        self.mlp = Mlp(\n            in_features=hidden_size,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=self.target_dim,\n            drop=0,\n        )\n\n    def forward(\n        self,\n        pose_feat,\n    ):\n        \"\"\"\n        pose_feat: BxC\n        preliminary_cameras: cameras in opencv coordinate.\n        \"\"\"\n\n        pred_cameras = self.mlp(pose_feat)  # Bx7, 3 for absT, 4 for quaR\n        return pred_cameras\n\n\nclass PoseEncoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        mlp_ratio=4,\n        pose_mode=(\"exp\", -inf, inf),\n        pose_encoding_type=\"absT_quaR\",\n    ):\n        super().__init__()\n        self.pose_encoding_type = pose_encoding_type\n        self.pose_mode = pose_mode\n\n        if self.pose_encoding_type == \"absT_quaR\":\n            self.target_dim = 7\n\n        self.embed_pose = PoseEmbedding(\n            target_dim=self.target_dim,\n            out_dim=hidden_size,\n            n_harmonic_functions=10,\n            append_input=True,\n        )\n        self.pose_encoder = Mlp(\n            in_features=self.embed_pose.out_dim,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=hidden_size,\n            drop=0,\n        )\n\n    def forward(self, camera):\n        pose_enc = camera_to_pose_encoding(\n            camera,\n            pose_encoding_type=self.pose_encoding_type,\n        ).to(camera.dtype)\n        pose_enc = postprocess_pose(pose_enc, self.pose_mode, inverse=True)\n        pose_feat = self.embed_pose(pose_enc)\n        pose_feat = self.pose_encoder(pose_feat)\n        return pose_feat\n\n\nclass HarmonicEmbedding(torch.nn.Module):\n    def __init__(\n        self,\n        n_harmonic_functions: int = 6,\n        omega_0: float = 1.0,\n        logspace: bool = True,\n        append_input: bool = True,\n    ) -> None:\n        \"\"\"\n        The harmonic embedding layer supports the classical\n        Nerf positional encoding described in\n        `NeRF <https://arxiv.org/abs/2003.08934>`_\n        and the integrated position encoding in\n        `MIP-NeRF <https://arxiv.org/abs/2103.13415>`_.\n\n        During the inference you can provide the extra argument `diag_cov`.\n\n        If `diag_cov is None`, it converts","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.HarmonicEmbedding","uri":"program://Human3R/class/src.dust3r.utils.camera.HarmonicEmbedding#L85-L251","kind":"class","name":"HarmonicEmbedding","path":"src/dust3r/utils/camera.py","language":"python","start_line":85,"end_line":251,"context_start_line":65,"context_end_line":271,"code":"            append_input=True,\n        )\n        self.pose_encoder = Mlp(\n            in_features=self.embed_pose.out_dim,\n            hidden_features=int(hidden_size * mlp_ratio),\n            out_features=hidden_size,\n            drop=0,\n        )\n\n    def forward(self, camera):\n        pose_enc = camera_to_pose_encoding(\n            camera,\n            pose_encoding_type=self.pose_encoding_type,\n        ).to(camera.dtype)\n        pose_enc = postprocess_pose(pose_enc, self.pose_mode, inverse=True)\n        pose_feat = self.embed_pose(pose_enc)\n        pose_feat = self.pose_encoder(pose_feat)\n        return pose_feat\n\n\nclass HarmonicEmbedding(torch.nn.Module):\n    def __init__(\n        self,\n        n_harmonic_functions: int = 6,\n        omega_0: float = 1.0,\n        logspace: bool = True,\n        append_input: bool = True,\n    ) -> None:\n        \"\"\"\n        The harmonic embedding layer supports the classical\n        Nerf positional encoding described in\n        `NeRF <https://arxiv.org/abs/2003.08934>`_\n        and the integrated position encoding in\n        `MIP-NeRF <https://arxiv.org/abs/2103.13415>`_.\n\n        During the inference you can provide the extra argument `diag_cov`.\n\n        If `diag_cov is None`, it converts\n        rays parametrized with a `ray_bundle` to 3D points by\n        extending each ray according to the corresponding length.\n        Then it converts each feature\n        (i.e. vector along the last dimension) in `x`\n        into a series of harmonic features `embedding`,\n        where for each i in range(dim) the following are present\n        in embedding[...]::\n\n            [\n                sin(f_1*x[..., i]),\n                sin(f_2*x[..., i]),\n                ...\n                sin(f_N * x[..., i]),\n                cos(f_1*x[..., i]),\n                cos(f_2*x[..., i]),\n                ...\n                cos(f_N * x[..., i]),\n                x[..., i],              # only present if append_input is True.\n            ]\n\n        where N corresponds to `n_harmonic_functions-1`, and f_i is a scalar\n        denoting the i-th frequency of the harmonic embedding.\n\n\n        If `diag_cov is not None`, it approximates\n        conical frustums following a ray bundle as gaussians,\n        defined by x, the means of the gaussians and diag_cov,\n        the diagonal covariances.\n        Then it converts each gaussian\n        into a series of harmonic features `embedding`,\n        where for each i in range(dim) the following are present\n        in embedding[...]::\n\n            [\n                sin(f_1*x[..., i]) * exp(0.5 * f_1**2 * diag_cov[..., i,]),\n                sin(f_2*x[..., i]) * exp(0.5 * f_2**2 * diag_cov[..., i,]),\n                ...\n                sin(f_N * x[..., i]) * exp(0.5 * f_N**2 * diag_cov[..., i,]),\n                cos(f_1*x[..., i]) * exp(0.5 * f_1**2 * diag_cov[..., i,]),\n                cos(f_2*x[..., i]) * exp(0.5 * f_2**2 * diag_cov[..., i,]),,\n                ...\n                cos(f_N * x[..., i]) * exp(0.5 * f_N**2 * diag_cov[..., i,]),\n                x[..., i],              # only present if append_input is True.\n            ]\n\n        where N equals `n_harmonic_functions-1`, and f_i is a scalar\n        denoting the i-th frequency of the harmonic embedding.\n\n        If `logspace==True`, the frequencies `[f_1, ..., f_N]` are\n        powers of 2:\n            `f_1, ..., f_N = 2**torch.arange(n_harmonic_functions)`\n\n        If `logspace==False`, frequencies are linearly spaced between\n        `1.0` and `2**(n_harmonic_functions-1)`:\n            `f_1, ..., f_N = torch.linspace(\n                1.0, 2**(n_harmonic_functions-1), n_harmonic_functions\n            )`\n\n        Note that `x` is also premultiplied by the base frequency `omega_0`\n        before evaluating the harmonic functions.\n\n        Args:\n            n_harmonic_functions: int, number of harmonic\n                features\n            omega_0: float, base frequency\n            logspace: bool, Whether to space the frequencies in\n                logspace or linear space\n            append_input: bool, whether to concat the original\n                input to the harmonic embedding. If true the\n                output is of the form (embed.sin(), embed.cos(), x)\n        \"\"\"\n        super().__init__()\n\n        if logspace:\n            frequencies = 2.0 ** torch.arange(n_harmonic_functions, dtype=torch.float32)\n        else:\n            frequencies = torch.linspace(\n                1.0,\n                2.0 ** (n_harmonic_functions - 1),\n                n_harmonic_functions,\n                dtype=torch.float32,\n            )\n\n        self.register_buffer(\"_frequencies\", frequencies * omega_0, persistent=False)\n        self.register_buffer(\n            \"_zero_half_pi\",\n            torch.tensor([0.0, 0.5 * torch.pi]),\n            persistent=False,\n        )\n        self.append_input = append_input\n\n    def forward(\n        self, x: torch.Tensor, diag_cov: Optional[torch.Tensor] = None, **kwargs\n    ) -> torch.Tensor:\n        \"\"\"\n        Args:\n            x: tensor of shape [..., dim]\n            diag_cov: An optional tensor of shape `(..., dim)`\n                representing the diagonal covariance matrices of our Gaussians, joined with x\n                as means of the Gaussians.\n\n        Returns:\n            embedding: a harmonic embedding of `x` of shape\n            [..., (n_harmonic_functions * 2 + int(append_input)) * num_points_per_ray]\n        \"\"\"\n\n        embed = x[..., None] * self._frequencies\n\n        embed = embed[..., None, :, :] + self._zero_half_pi[..., None, None]\n\n        embed = embed.sin()\n        if diag_cov is not None:\n            x_var = diag_cov[..., None] * torch.pow(self._frequencies, 2)\n            exp_var = torch.exp(-0.5 * x_var)\n\n            embed = embed * exp_var[..., None, :, :]\n\n        embed = embed.reshape(*x.shape[:-1], -1)\n\n        if self.append_input:\n            return torch.cat([embed, x], dim=-1)\n        return embed\n\n    @staticmethod\n    def get_output_dim_static(\n        input_dims: int, n_harmonic_functions: int, append_input: bool\n    ) -> int:\n        \"\"\"\n        Utility to help predict the shape of the output of `forward`.\n\n        Args:\n            input_dims: length of the last dimension of the input tensor\n            n_harmonic_functions: number of embedding frequencies\n            append_input: whether or not to concat the original\n                input to the harmonic embedding\n        Returns:\n            int: the length of the last dimension of the output tensor\n        \"\"\"\n        return input_dims * (2 * n_harmonic_functions + int(append_input))\n\n    def get_output_dim(self, input_dims: int = 3) -> int:\n        \"\"\"\n        Same as above. The default for input_dims is 3 for 3D applications\n        which use harmonic embedding for positional encoding,\n        so the input might be xyz.\n        \"\"\"\n        return self.get_output_dim_static(\n            input_dims, len(self._frequencies), self.append_input\n        )\n\n\nclass PoseEmbedding(nn.Module):\n    def __init__(self, target_dim, out_dim, n_harmonic_functions=10, append_input=True):\n        super().__init__()\n\n        self._emb_pose = HarmonicEmbedding(\n            n_harmonic_functions=n_harmonic_functions, append_input=append_input\n        )\n\n        self.out_dim = self._emb_pose.get_output_dim(target_dim)\n\n    def forward(self, pose_encoding):\n        e_pose_encoding = self._emb_pose(pose_encoding)\n        return e_pose_encoding\n\n\ndef _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Returns torch.sqrt(torch.max(0, x))","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.PoseEmbedding","uri":"program://Human3R/class/src.dust3r.utils.camera.PoseEmbedding#L254-L266","kind":"class","name":"PoseEmbedding","path":"src/dust3r/utils/camera.py","language":"python","start_line":254,"end_line":266,"context_start_line":234,"context_end_line":286,"code":"            input_dims: length of the last dimension of the input tensor\n            n_harmonic_functions: number of embedding frequencies\n            append_input: whether or not to concat the original\n                input to the harmonic embedding\n        Returns:\n            int: the length of the last dimension of the output tensor\n        \"\"\"\n        return input_dims * (2 * n_harmonic_functions + int(append_input))\n\n    def get_output_dim(self, input_dims: int = 3) -> int:\n        \"\"\"\n        Same as above. The default for input_dims is 3 for 3D applications\n        which use harmonic embedding for positional encoding,\n        so the input might be xyz.\n        \"\"\"\n        return self.get_output_dim_static(\n            input_dims, len(self._frequencies), self.append_input\n        )\n\n\nclass PoseEmbedding(nn.Module):\n    def __init__(self, target_dim, out_dim, n_harmonic_functions=10, append_input=True):\n        super().__init__()\n\n        self._emb_pose = HarmonicEmbedding(\n            n_harmonic_functions=n_harmonic_functions, append_input=append_input\n        )\n\n        self.out_dim = self._emb_pose.get_output_dim(target_dim)\n\n    def forward(self, pose_encoding):\n        e_pose_encoding = self._emb_pose(pose_encoding)\n        return e_pose_encoding\n\n\ndef _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Returns torch.sqrt(torch.max(0, x))\n    but with a zero subgradient where x is 0.\n    \"\"\"\n    ret = torch.zeros_like(x)\n    positive_mask = x > 0\n    ret[positive_mask] = torch.sqrt(x[positive_mask])\n    return ret\n\n\ndef matrix_to_quaternion(matrix: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert rotations given as rotation matrices to quaternions.\n\n    Args:\n        matrix: Rotation matrices as tensor of shape (..., 3, 3).\n","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera._sqrt_positive_part","uri":"program://Human3R/function/src.dust3r.utils.camera._sqrt_positive_part#L269-L277","kind":"function","name":"_sqrt_positive_part","path":"src/dust3r/utils/camera.py","language":"python","start_line":269,"end_line":277,"context_start_line":249,"context_end_line":297,"code":"        return self.get_output_dim_static(\n            input_dims, len(self._frequencies), self.append_input\n        )\n\n\nclass PoseEmbedding(nn.Module):\n    def __init__(self, target_dim, out_dim, n_harmonic_functions=10, append_input=True):\n        super().__init__()\n\n        self._emb_pose = HarmonicEmbedding(\n            n_harmonic_functions=n_harmonic_functions, append_input=append_input\n        )\n\n        self.out_dim = self._emb_pose.get_output_dim(target_dim)\n\n    def forward(self, pose_encoding):\n        e_pose_encoding = self._emb_pose(pose_encoding)\n        return e_pose_encoding\n\n\ndef _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Returns torch.sqrt(torch.max(0, x))\n    but with a zero subgradient where x is 0.\n    \"\"\"\n    ret = torch.zeros_like(x)\n    positive_mask = x > 0\n    ret[positive_mask] = torch.sqrt(x[positive_mask])\n    return ret\n\n\ndef matrix_to_quaternion(matrix: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert rotations given as rotation matrices to quaternions.\n\n    Args:\n        matrix: Rotation matrices as tensor of shape (..., 3, 3).\n\n    Returns:\n        quaternions with real part first, as tensor of shape (..., 4).\n    \"\"\"\n    if matrix.size(-1) != 3 or matrix.size(-2) != 3:\n        raise ValueError(f\"Invalid rotation matrix shape {matrix.shape}.\")\n\n    batch_dim = matrix.shape[:-2]\n    m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind(\n        matrix.reshape(batch_dim + (9,)), dim=-1\n    )\n","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.matrix_to_quaternion","uri":"program://Human3R/function/src.dust3r.utils.camera.matrix_to_quaternion#L280-L326","kind":"function","name":"matrix_to_quaternion","path":"src/dust3r/utils/camera.py","language":"python","start_line":280,"end_line":326,"context_start_line":260,"context_end_line":346,"code":"        )\n\n        self.out_dim = self._emb_pose.get_output_dim(target_dim)\n\n    def forward(self, pose_encoding):\n        e_pose_encoding = self._emb_pose(pose_encoding)\n        return e_pose_encoding\n\n\ndef _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Returns torch.sqrt(torch.max(0, x))\n    but with a zero subgradient where x is 0.\n    \"\"\"\n    ret = torch.zeros_like(x)\n    positive_mask = x > 0\n    ret[positive_mask] = torch.sqrt(x[positive_mask])\n    return ret\n\n\ndef matrix_to_quaternion(matrix: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert rotations given as rotation matrices to quaternions.\n\n    Args:\n        matrix: Rotation matrices as tensor of shape (..., 3, 3).\n\n    Returns:\n        quaternions with real part first, as tensor of shape (..., 4).\n    \"\"\"\n    if matrix.size(-1) != 3 or matrix.size(-2) != 3:\n        raise ValueError(f\"Invalid rotation matrix shape {matrix.shape}.\")\n\n    batch_dim = matrix.shape[:-2]\n    m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind(\n        matrix.reshape(batch_dim + (9,)), dim=-1\n    )\n\n    q_abs = _sqrt_positive_part(\n        torch.stack(\n            [\n                1.0 + m00 + m11 + m22,\n                1.0 + m00 - m11 - m22,\n                1.0 - m00 + m11 - m22,\n                1.0 - m00 - m11 + m22,\n            ],\n            dim=-1,\n        )\n    )\n\n    quat_by_rijk = torch.stack(\n        [\n            torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1),\n            torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1),\n            torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1),\n            torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1),\n        ],\n        dim=-2,\n    )\n\n    flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device)\n    quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr))\n\n    out = quat_candidates[\n        F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, :\n    ].reshape(batch_dim + (4,))\n    return standardize_quaternion(out)\n\n\ndef standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert a unit quaternion to a standard form: one in which the real\n    part is non negative.\n\n    Args:\n        quaternions: Quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Standardized quaternions as tensor of shape (..., 4).\n    \"\"\"\n    quaternions = F.normalize(quaternions, p=2, dim=-1)\n    return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions)\n\n\ndef camera_to_pose_encoding(\n    camera,","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.standardize_quaternion","uri":"program://Human3R/function/src.dust3r.utils.camera.standardize_quaternion#L329-L342","kind":"function","name":"standardize_quaternion","path":"src/dust3r/utils/camera.py","language":"python","start_line":329,"end_line":342,"context_start_line":309,"context_end_line":362,"code":"\n    quat_by_rijk = torch.stack(\n        [\n            torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1),\n            torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1),\n            torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1),\n            torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1),\n        ],\n        dim=-2,\n    )\n\n    flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device)\n    quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr))\n\n    out = quat_candidates[\n        F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, :\n    ].reshape(batch_dim + (4,))\n    return standardize_quaternion(out)\n\n\ndef standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert a unit quaternion to a standard form: one in which the real\n    part is non negative.\n\n    Args:\n        quaternions: Quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Standardized quaternions as tensor of shape (..., 4).\n    \"\"\"\n    quaternions = F.normalize(quaternions, p=2, dim=-1)\n    return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions)\n\n\ndef camera_to_pose_encoding(\n    camera,\n    pose_encoding_type=\"absT_quaR\",\n):\n    \"\"\"\n    Inverse to pose_encoding_to_camera\n    camera: opencv, cam2world\n    \"\"\"\n    if pose_encoding_type == \"absT_quaR\":\n\n        quaternion_R = matrix_to_quaternion(camera[:, :3, :3])\n\n        pose_encoding = torch.cat([camera[:, :3, 3], quaternion_R], dim=-1)\n    else:\n        raise ValueError(f\"Unknown pose encoding {pose_encoding_type}\")\n\n    return pose_encoding\n","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.camera_to_pose_encoding","uri":"program://Human3R/function/src.dust3r.utils.camera.camera_to_pose_encoding#L345-L361","kind":"function","name":"camera_to_pose_encoding","path":"src/dust3r/utils/camera.py","language":"python","start_line":345,"end_line":361,"context_start_line":325,"context_end_line":381,"code":"    ].reshape(batch_dim + (4,))\n    return standardize_quaternion(out)\n\n\ndef standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert a unit quaternion to a standard form: one in which the real\n    part is non negative.\n\n    Args:\n        quaternions: Quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Standardized quaternions as tensor of shape (..., 4).\n    \"\"\"\n    quaternions = F.normalize(quaternions, p=2, dim=-1)\n    return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions)\n\n\ndef camera_to_pose_encoding(\n    camera,\n    pose_encoding_type=\"absT_quaR\",\n):\n    \"\"\"\n    Inverse to pose_encoding_to_camera\n    camera: opencv, cam2world\n    \"\"\"\n    if pose_encoding_type == \"absT_quaR\":\n\n        quaternion_R = matrix_to_quaternion(camera[:, :3, :3])\n\n        pose_encoding = torch.cat([camera[:, :3, 3], quaternion_R], dim=-1)\n    else:\n        raise ValueError(f\"Unknown pose encoding {pose_encoding_type}\")\n\n    return pose_encoding\n\n\ndef quaternion_to_matrix(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert rotations given as quaternions to rotation matrices.\n\n    Args:\n        quaternions: quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Rotation matrices as tensor of shape (..., 3, 3).\n    \"\"\"\n    r, i, j, k = torch.unbind(quaternions, -1)\n\n    two_s = 2.0 / (quaternions * quaternions).sum(-1)\n\n    o = torch.stack(\n        (\n            1 - two_s * (j * j + k * k),","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.quaternion_to_matrix","uri":"program://Human3R/function/src.dust3r.utils.camera.quaternion_to_matrix#L364-L393","kind":"function","name":"quaternion_to_matrix","path":"src/dust3r/utils/camera.py","language":"python","start_line":364,"end_line":393,"context_start_line":344,"context_end_line":413,"code":"\ndef camera_to_pose_encoding(\n    camera,\n    pose_encoding_type=\"absT_quaR\",\n):\n    \"\"\"\n    Inverse to pose_encoding_to_camera\n    camera: opencv, cam2world\n    \"\"\"\n    if pose_encoding_type == \"absT_quaR\":\n\n        quaternion_R = matrix_to_quaternion(camera[:, :3, :3])\n\n        pose_encoding = torch.cat([camera[:, :3, 3], quaternion_R], dim=-1)\n    else:\n        raise ValueError(f\"Unknown pose encoding {pose_encoding_type}\")\n\n    return pose_encoding\n\n\ndef quaternion_to_matrix(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert rotations given as quaternions to rotation matrices.\n\n    Args:\n        quaternions: quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Rotation matrices as tensor of shape (..., 3, 3).\n    \"\"\"\n    r, i, j, k = torch.unbind(quaternions, -1)\n\n    two_s = 2.0 / (quaternions * quaternions).sum(-1)\n\n    o = torch.stack(\n        (\n            1 - two_s * (j * j + k * k),\n            two_s * (i * j - k * r),\n            two_s * (i * k + j * r),\n            two_s * (i * j + k * r),\n            1 - two_s * (i * i + k * k),\n            two_s * (j * k - i * r),\n            two_s * (i * k - j * r),\n            two_s * (j * k + i * r),\n            1 - two_s * (i * i + j * j),\n        ),\n        -1,\n    )\n    return o.reshape(quaternions.shape[:-1] + (3, 3))\n\n\ndef pose_encoding_to_camera(\n    pose_encoding,\n    pose_encoding_type=\"absT_quaR\",\n):\n    \"\"\"\n    Args:\n        pose_encoding: A tensor of shape `BxC`, containing a batch of\n                        `B` `C`-dimensional pose encodings.\n        pose_encoding_type: The type of pose encoding,\n    \"\"\"\n\n    if pose_encoding_type == \"absT_quaR\":\n\n        abs_T = pose_encoding[:, :3]\n        quaternion_R = pose_encoding[:, 3:7]\n        R = quaternion_to_matrix(quaternion_R)\n    else:\n        raise ValueError(f\"Unknown pose encoding {pose_encoding_type}\")","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.pose_encoding_to_camera","uri":"program://Human3R/function/src.dust3r.utils.camera.pose_encoding_to_camera#L396-L420","kind":"function","name":"pose_encoding_to_camera","path":"src/dust3r/utils/camera.py","language":"python","start_line":396,"end_line":420,"context_start_line":376,"context_end_line":440,"code":"\n    two_s = 2.0 / (quaternions * quaternions).sum(-1)\n\n    o = torch.stack(\n        (\n            1 - two_s * (j * j + k * k),\n            two_s * (i * j - k * r),\n            two_s * (i * k + j * r),\n            two_s * (i * j + k * r),\n            1 - two_s * (i * i + k * k),\n            two_s * (j * k - i * r),\n            two_s * (i * k - j * r),\n            two_s * (j * k + i * r),\n            1 - two_s * (i * i + j * j),\n        ),\n        -1,\n    )\n    return o.reshape(quaternions.shape[:-1] + (3, 3))\n\n\ndef pose_encoding_to_camera(\n    pose_encoding,\n    pose_encoding_type=\"absT_quaR\",\n):\n    \"\"\"\n    Args:\n        pose_encoding: A tensor of shape `BxC`, containing a batch of\n                        `B` `C`-dimensional pose encodings.\n        pose_encoding_type: The type of pose encoding,\n    \"\"\"\n\n    if pose_encoding_type == \"absT_quaR\":\n\n        abs_T = pose_encoding[:, :3]\n        quaternion_R = pose_encoding[:, 3:7]\n        R = quaternion_to_matrix(quaternion_R)\n    else:\n        raise ValueError(f\"Unknown pose encoding {pose_encoding_type}\")\n\n    c2w_mats = torch.eye(4, 4).to(R.dtype).to(R.device)\n    c2w_mats = c2w_mats[None].repeat(len(R), 1, 1)\n    c2w_mats[:, :3, :3] = R\n    c2w_mats[:, :3, 3] = abs_T\n\n    return c2w_mats\n\n\ndef quaternion_conjugate(q):\n    \"\"\"Compute the conjugate of quaternion q (w, x, y, z).\"\"\"\n\n    q_conj = torch.cat([q[..., :1], -q[..., 1:]], dim=-1)\n    return q_conj\n\n\ndef quaternion_multiply(q1, q2):\n    \"\"\"Multiply two quaternions q1 and q2.\"\"\"\n    w1, x1, y1, z1 = q1.unbind(dim=-1)\n    w2, x2, y2, z2 = q2.unbind(dim=-1)\n\n    w = w1 * w2 - x1 * x2 - y1 * y2 - z1 * z2\n    x = w1 * x2 + x1 * w2 + y1 * z2 - z1 * y2\n    y = w1 * y2 - x1 * z2 + y1 * w2 + z1 * x2\n    z = w1 * z2 + x1 * y2 - y1 * x2 + z1 * w2\n\n    return torch.stack((w, x, y, z), dim=-1)","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.quaternion_conjugate","uri":"program://Human3R/function/src.dust3r.utils.camera.quaternion_conjugate#L423-L427","kind":"function","name":"quaternion_conjugate","path":"src/dust3r/utils/camera.py","language":"python","start_line":423,"end_line":427,"context_start_line":403,"context_end_line":447,"code":"                        `B` `C`-dimensional pose encodings.\n        pose_encoding_type: The type of pose encoding,\n    \"\"\"\n\n    if pose_encoding_type == \"absT_quaR\":\n\n        abs_T = pose_encoding[:, :3]\n        quaternion_R = pose_encoding[:, 3:7]\n        R = quaternion_to_matrix(quaternion_R)\n    else:\n        raise ValueError(f\"Unknown pose encoding {pose_encoding_type}\")\n\n    c2w_mats = torch.eye(4, 4).to(R.dtype).to(R.device)\n    c2w_mats = c2w_mats[None].repeat(len(R), 1, 1)\n    c2w_mats[:, :3, :3] = R\n    c2w_mats[:, :3, 3] = abs_T\n\n    return c2w_mats\n\n\ndef quaternion_conjugate(q):\n    \"\"\"Compute the conjugate of quaternion q (w, x, y, z).\"\"\"\n\n    q_conj = torch.cat([q[..., :1], -q[..., 1:]], dim=-1)\n    return q_conj\n\n\ndef quaternion_multiply(q1, q2):\n    \"\"\"Multiply two quaternions q1 and q2.\"\"\"\n    w1, x1, y1, z1 = q1.unbind(dim=-1)\n    w2, x2, y2, z2 = q2.unbind(dim=-1)\n\n    w = w1 * w2 - x1 * x2 - y1 * y2 - z1 * z2\n    x = w1 * x2 + x1 * w2 + y1 * z2 - z1 * y2\n    y = w1 * y2 - x1 * z2 + y1 * w2 + z1 * x2\n    z = w1 * z2 + x1 * y2 - y1 * x2 + z1 * w2\n\n    return torch.stack((w, x, y, z), dim=-1)\n\n\ndef rotate_vector(q, v):\n    \"\"\"Rotate vector v by quaternion q.\"\"\"\n    q_vec = q[..., 1:]\n    q_w = q[..., :1]\n","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.quaternion_multiply","uri":"program://Human3R/function/src.dust3r.utils.camera.quaternion_multiply#L430-L440","kind":"function","name":"quaternion_multiply","path":"src/dust3r/utils/camera.py","language":"python","start_line":430,"end_line":440,"context_start_line":410,"context_end_line":460,"code":"        quaternion_R = pose_encoding[:, 3:7]\n        R = quaternion_to_matrix(quaternion_R)\n    else:\n        raise ValueError(f\"Unknown pose encoding {pose_encoding_type}\")\n\n    c2w_mats = torch.eye(4, 4).to(R.dtype).to(R.device)\n    c2w_mats = c2w_mats[None].repeat(len(R), 1, 1)\n    c2w_mats[:, :3, :3] = R\n    c2w_mats[:, :3, 3] = abs_T\n\n    return c2w_mats\n\n\ndef quaternion_conjugate(q):\n    \"\"\"Compute the conjugate of quaternion q (w, x, y, z).\"\"\"\n\n    q_conj = torch.cat([q[..., :1], -q[..., 1:]], dim=-1)\n    return q_conj\n\n\ndef quaternion_multiply(q1, q2):\n    \"\"\"Multiply two quaternions q1 and q2.\"\"\"\n    w1, x1, y1, z1 = q1.unbind(dim=-1)\n    w2, x2, y2, z2 = q2.unbind(dim=-1)\n\n    w = w1 * w2 - x1 * x2 - y1 * y2 - z1 * z2\n    x = w1 * x2 + x1 * w2 + y1 * z2 - z1 * y2\n    y = w1 * y2 - x1 * z2 + y1 * w2 + z1 * x2\n    z = w1 * z2 + x1 * y2 - y1 * x2 + z1 * w2\n\n    return torch.stack((w, x, y, z), dim=-1)\n\n\ndef rotate_vector(q, v):\n    \"\"\"Rotate vector v by quaternion q.\"\"\"\n    q_vec = q[..., 1:]\n    q_w = q[..., :1]\n\n    t = 2.0 * torch.cross(q_vec, v, dim=-1)\n    v_rot = v + q_w * t + torch.cross(q_vec, t, dim=-1)\n    return v_rot\n\n\ndef relative_pose_absT_quatR(t1, q1, t2, q2):\n    \"\"\"Compute the relative translation and quaternion between two poses.\"\"\"\n\n    q1_inv = quaternion_conjugate(q1)\n\n    q_rel = quaternion_multiply(q1_inv, q2)\n\n    delta_t = t2 - t1","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.rotate_vector","uri":"program://Human3R/function/src.dust3r.utils.camera.rotate_vector#L443-L450","kind":"function","name":"rotate_vector","path":"src/dust3r/utils/camera.py","language":"python","start_line":443,"end_line":450,"context_start_line":423,"context_end_line":462,"code":"def quaternion_conjugate(q):\n    \"\"\"Compute the conjugate of quaternion q (w, x, y, z).\"\"\"\n\n    q_conj = torch.cat([q[..., :1], -q[..., 1:]], dim=-1)\n    return q_conj\n\n\ndef quaternion_multiply(q1, q2):\n    \"\"\"Multiply two quaternions q1 and q2.\"\"\"\n    w1, x1, y1, z1 = q1.unbind(dim=-1)\n    w2, x2, y2, z2 = q2.unbind(dim=-1)\n\n    w = w1 * w2 - x1 * x2 - y1 * y2 - z1 * z2\n    x = w1 * x2 + x1 * w2 + y1 * z2 - z1 * y2\n    y = w1 * y2 - x1 * z2 + y1 * w2 + z1 * x2\n    z = w1 * z2 + x1 * y2 - y1 * x2 + z1 * w2\n\n    return torch.stack((w, x, y, z), dim=-1)\n\n\ndef rotate_vector(q, v):\n    \"\"\"Rotate vector v by quaternion q.\"\"\"\n    q_vec = q[..., 1:]\n    q_w = q[..., :1]\n\n    t = 2.0 * torch.cross(q_vec, v, dim=-1)\n    v_rot = v + q_w * t + torch.cross(q_vec, t, dim=-1)\n    return v_rot\n\n\ndef relative_pose_absT_quatR(t1, q1, t2, q2):\n    \"\"\"Compute the relative translation and quaternion between two poses.\"\"\"\n\n    q1_inv = quaternion_conjugate(q1)\n\n    q_rel = quaternion_multiply(q1_inv, q2)\n\n    delta_t = t2 - t1\n    t_rel = rotate_vector(q1_inv, delta_t)\n    return t_rel, q_rel","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.relative_pose_absT_quatR","uri":"program://Human3R/function/src.dust3r.utils.camera.relative_pose_absT_quatR#L453-L462","kind":"function","name":"relative_pose_absT_quatR","path":"src/dust3r/utils/camera.py","language":"python","start_line":453,"end_line":462,"context_start_line":433,"context_end_line":462,"code":"    w2, x2, y2, z2 = q2.unbind(dim=-1)\n\n    w = w1 * w2 - x1 * x2 - y1 * y2 - z1 * z2\n    x = w1 * x2 + x1 * w2 + y1 * z2 - z1 * y2\n    y = w1 * y2 - x1 * z2 + y1 * w2 + z1 * x2\n    z = w1 * z2 + x1 * y2 - y1 * x2 + z1 * w2\n\n    return torch.stack((w, x, y, z), dim=-1)\n\n\ndef rotate_vector(q, v):\n    \"\"\"Rotate vector v by quaternion q.\"\"\"\n    q_vec = q[..., 1:]\n    q_w = q[..., :1]\n\n    t = 2.0 * torch.cross(q_vec, v, dim=-1)\n    v_rot = v + q_w * t + torch.cross(q_vec, t, dim=-1)\n    return v_rot\n\n\ndef relative_pose_absT_quatR(t1, q1, t2, q2):\n    \"\"\"Compute the relative translation and quaternion between two poses.\"\"\"\n\n    q1_inv = quaternion_conjugate(q1)\n\n    q_rel = quaternion_multiply(q1_inv, q2)\n\n    delta_t = t2 - t1\n    t_rel = rotate_vector(q1_inv, delta_t)\n    return t_rel, q_rel","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.__init__","uri":"program://Human3R/function/src.dust3r.utils.camera.__init__#L255-L262","kind":"function","name":"__init__","path":"src/dust3r/utils/camera.py","language":"python","start_line":255,"end_line":262,"context_start_line":235,"context_end_line":282,"code":"            n_harmonic_functions: number of embedding frequencies\n            append_input: whether or not to concat the original\n                input to the harmonic embedding\n        Returns:\n            int: the length of the last dimension of the output tensor\n        \"\"\"\n        return input_dims * (2 * n_harmonic_functions + int(append_input))\n\n    def get_output_dim(self, input_dims: int = 3) -> int:\n        \"\"\"\n        Same as above. The default for input_dims is 3 for 3D applications\n        which use harmonic embedding for positional encoding,\n        so the input might be xyz.\n        \"\"\"\n        return self.get_output_dim_static(\n            input_dims, len(self._frequencies), self.append_input\n        )\n\n\nclass PoseEmbedding(nn.Module):\n    def __init__(self, target_dim, out_dim, n_harmonic_functions=10, append_input=True):\n        super().__init__()\n\n        self._emb_pose = HarmonicEmbedding(\n            n_harmonic_functions=n_harmonic_functions, append_input=append_input\n        )\n\n        self.out_dim = self._emb_pose.get_output_dim(target_dim)\n\n    def forward(self, pose_encoding):\n        e_pose_encoding = self._emb_pose(pose_encoding)\n        return e_pose_encoding\n\n\ndef _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Returns torch.sqrt(torch.max(0, x))\n    but with a zero subgradient where x is 0.\n    \"\"\"\n    ret = torch.zeros_like(x)\n    positive_mask = x > 0\n    ret[positive_mask] = torch.sqrt(x[positive_mask])\n    return ret\n\n\ndef matrix_to_quaternion(matrix: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert rotations given as rotation matrices to quaternions.","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.forward","uri":"program://Human3R/function/src.dust3r.utils.camera.forward#L264-L266","kind":"function","name":"forward","path":"src/dust3r/utils/camera.py","language":"python","start_line":264,"end_line":266,"context_start_line":244,"context_end_line":286,"code":"        \"\"\"\n        Same as above. The default for input_dims is 3 for 3D applications\n        which use harmonic embedding for positional encoding,\n        so the input might be xyz.\n        \"\"\"\n        return self.get_output_dim_static(\n            input_dims, len(self._frequencies), self.append_input\n        )\n\n\nclass PoseEmbedding(nn.Module):\n    def __init__(self, target_dim, out_dim, n_harmonic_functions=10, append_input=True):\n        super().__init__()\n\n        self._emb_pose = HarmonicEmbedding(\n            n_harmonic_functions=n_harmonic_functions, append_input=append_input\n        )\n\n        self.out_dim = self._emb_pose.get_output_dim(target_dim)\n\n    def forward(self, pose_encoding):\n        e_pose_encoding = self._emb_pose(pose_encoding)\n        return e_pose_encoding\n\n\ndef _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Returns torch.sqrt(torch.max(0, x))\n    but with a zero subgradient where x is 0.\n    \"\"\"\n    ret = torch.zeros_like(x)\n    positive_mask = x > 0\n    ret[positive_mask] = torch.sqrt(x[positive_mask])\n    return ret\n\n\ndef matrix_to_quaternion(matrix: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert rotations given as rotation matrices to quaternions.\n\n    Args:\n        matrix: Rotation matrices as tensor of shape (..., 3, 3).\n","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.get_output_dim_static","uri":"program://Human3R/function/src.dust3r.utils.camera.get_output_dim_static#L227-L241","kind":"function","name":"get_output_dim_static","path":"src/dust3r/utils/camera.py","language":"python","start_line":227,"end_line":241,"context_start_line":207,"context_end_line":261,"code":"        \"\"\"\n\n        embed = x[..., None] * self._frequencies\n\n        embed = embed[..., None, :, :] + self._zero_half_pi[..., None, None]\n\n        embed = embed.sin()\n        if diag_cov is not None:\n            x_var = diag_cov[..., None] * torch.pow(self._frequencies, 2)\n            exp_var = torch.exp(-0.5 * x_var)\n\n            embed = embed * exp_var[..., None, :, :]\n\n        embed = embed.reshape(*x.shape[:-1], -1)\n\n        if self.append_input:\n            return torch.cat([embed, x], dim=-1)\n        return embed\n\n    @staticmethod\n    def get_output_dim_static(\n        input_dims: int, n_harmonic_functions: int, append_input: bool\n    ) -> int:\n        \"\"\"\n        Utility to help predict the shape of the output of `forward`.\n\n        Args:\n            input_dims: length of the last dimension of the input tensor\n            n_harmonic_functions: number of embedding frequencies\n            append_input: whether or not to concat the original\n                input to the harmonic embedding\n        Returns:\n            int: the length of the last dimension of the output tensor\n        \"\"\"\n        return input_dims * (2 * n_harmonic_functions + int(append_input))\n\n    def get_output_dim(self, input_dims: int = 3) -> int:\n        \"\"\"\n        Same as above. The default for input_dims is 3 for 3D applications\n        which use harmonic embedding for positional encoding,\n        so the input might be xyz.\n        \"\"\"\n        return self.get_output_dim_static(\n            input_dims, len(self._frequencies), self.append_input\n        )\n\n\nclass PoseEmbedding(nn.Module):\n    def __init__(self, target_dim, out_dim, n_harmonic_functions=10, append_input=True):\n        super().__init__()\n\n        self._emb_pose = HarmonicEmbedding(\n            n_harmonic_functions=n_harmonic_functions, append_input=append_input\n        )\n","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.camera.get_output_dim","uri":"program://Human3R/function/src.dust3r.utils.camera.get_output_dim#L243-L251","kind":"function","name":"get_output_dim","path":"src/dust3r/utils/camera.py","language":"python","start_line":243,"end_line":251,"context_start_line":223,"context_end_line":271,"code":"            return torch.cat([embed, x], dim=-1)\n        return embed\n\n    @staticmethod\n    def get_output_dim_static(\n        input_dims: int, n_harmonic_functions: int, append_input: bool\n    ) -> int:\n        \"\"\"\n        Utility to help predict the shape of the output of `forward`.\n\n        Args:\n            input_dims: length of the last dimension of the input tensor\n            n_harmonic_functions: number of embedding frequencies\n            append_input: whether or not to concat the original\n                input to the harmonic embedding\n        Returns:\n            int: the length of the last dimension of the output tensor\n        \"\"\"\n        return input_dims * (2 * n_harmonic_functions + int(append_input))\n\n    def get_output_dim(self, input_dims: int = 3) -> int:\n        \"\"\"\n        Same as above. The default for input_dims is 3 for 3D applications\n        which use harmonic embedding for positional encoding,\n        so the input might be xyz.\n        \"\"\"\n        return self.get_output_dim_static(\n            input_dims, len(self._frequencies), self.append_input\n        )\n\n\nclass PoseEmbedding(nn.Module):\n    def __init__(self, target_dim, out_dim, n_harmonic_functions=10, append_input=True):\n        super().__init__()\n\n        self._emb_pose = HarmonicEmbedding(\n            n_harmonic_functions=n_harmonic_functions, append_input=append_input\n        )\n\n        self.out_dim = self._emb_pose.get_output_dim(target_dim)\n\n    def forward(self, pose_encoding):\n        e_pose_encoding = self._emb_pose(pose_encoding)\n        return e_pose_encoding\n\n\ndef _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Returns torch.sqrt(torch.max(0, x))","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.render","uri":"program://Human3R/module/src.dust3r.utils.render#L1-L314","kind":"module","name":"src.dust3r.utils.render","path":"src/dust3r/utils/render.py","language":"python","start_line":1,"end_line":314,"context_start_line":1,"context_end_line":314,"code":"import os\nos.environ['PYOPENGL_PLATFORM'] = 'egl'\n\nimport torch\nfrom gsplat import rasterization\nfrom dust3r.utils.geometry import inv, geotrf\nfrom dust3r.utils.image import unpad_image\nimport numpy as np\ntry:\n    import pyrender\nexcept:\n    import pyrender\n\nimport trimesh\nfrom PIL import Image\n\ndef render(\n    intrinsics: torch.Tensor,\n    pts3d: torch.Tensor,\n    rgbs: torch.Tensor | None = None,\n    scale: float = 0.002,\n    opacity: float = 0.95,\n):\n\n    device = pts3d.device\n    batch_size = len(intrinsics)\n    img_size = pts3d.shape[1:3]\n    pts3d = pts3d.reshape(batch_size, -1, 3)\n    num_pts = pts3d.shape[1]\n    quats = torch.randn((num_pts, 4), device=device)\n    quats = quats / quats.norm(dim=-1, keepdim=True)\n    scales = scale * torch.ones((num_pts, 3), device=device)\n    opacities = opacity * torch.ones((num_pts), device=device)\n    if rgbs is not None:\n        assert rgbs.shape[1] == 3\n        rgbs = rgbs.reshape(batch_size, 3, -1).transpose(1, 2)\n    else:\n        rgbs = torch.ones_like(pts3d[:, :, :3])\n\n    rendered_rgbs = []\n    rendered_depths = []\n    accs = []\n    for i in range(batch_size):\n        rgbd, acc, _ = rasterization(\n            pts3d[i],\n            quats,\n            scales,\n            opacities,\n            rgbs[i],\n            torch.eye(4, device=device)[None],\n            intrinsics[[i]],\n            width=img_size[1],\n            height=img_size[0],\n            packed=False,\n            render_mode=\"RGB+D\",\n        )\n\n        rendered_depths.append(rgbd[..., 3])\n\n    rendered_depths = torch.cat(rendered_depths, dim=0)\n\n    return rendered_rgbs, rendered_depths, accs\n\n\ndef get_render_results(gts, preds, self_view=False):\n    device = preds[0][\"pts3d_in_self_view\"].device\n    with torch.no_grad():\n        depths = []\n        gt_depths = []\n        for i, (gt, pred) in enumerate(zip(gts, preds)):\n            if self_view:\n                camera = inv(gt[\"camera_pose\"]).to(device)\n                intrinsics = gt[\"camera_intrinsics\"].to(device)\n                pred = pred[\"pts3d_in_self_view\"]\n            else:\n                camera = inv(gts[0][\"camera_pose\"]).to(device)\n                intrinsics = gts[0][\"camera_intrinsics\"].to(device)\n                pred = pred[\"pts3d_in_other_view\"]\n            gt_img = gt[\"img\"].to(device)\n            gt_pts3d = gt[\"pts3d\"].to(device)\n\n            _, depth, _ = render(intrinsics, pred, gt_img)\n            _, gt_depth, _ = render(intrinsics, geotrf(camera, gt_pts3d), gt_img)\n            depths.append(depth)\n            gt_depths.append(gt_depth)\n    return depths, gt_depths\n\n\ndef vis_heatmap(img, scores):\n    hm = scores.clone()\n    hm = torch.clamp(hm + 0.1, 0, 1)  # for visu purpose only\n    hm = hm.unsqueeze(0).unsqueeze(0)\n    hm = torch.nn.functional.interpolate(\n        hm, \n        size=(img.shape[0], img.shape[1]),\n        mode='nearest'\n    ).squeeze(0).squeeze(0)\n    \n    hm = hm.unsqueeze(-1)\n    \n    return img * hm\n\ndef get_render_smpl(gts, preds, smpl_model, loss_details, has_msk=False):\n    with torch.no_grad():\n        smpl_faces = {\n            'neutral': {\n                10: smpl_model.smplx_neutral_10.faces,\n                11: smpl_model.smplx_neutral_11.faces,\n            }\n        }\n    \n        gt_hms_list, pr_hms_list, gt_smpls_list, pr_smpls_list= [], [], [], []\n        gt_msks_list, pr_msks_list = [], []\n        for i, (gt, pred) in enumerate(zip(gts, preds)):\n            gt_img = gt[\"img\"]\n            smpl_mask = gt[\"smpl_mask\"]\n            K = gt[\"camera_intrinsics\"]\n\n            idx_h = torch.where(smpl_mask)\n            if has_msk:\n                gt_msk, pr_msk = gt[\"msk_mhmr\"], pred[\"msk\"][...,0]\n                gt_msk = unpad_image(gt_msk[:,None], gt_img.shape[2:])[:, 0]\n                pr_msk = unpad_image(pr_msk[:,None], gt_img.shape[2:])[:, 0]\n            gt_scores, pr_scores = gt[\"smpl_scores\"], pred[\"smpl_scores\"][...,0]\n            gt_scores = unpad_image(gt_scores[:,None], gt_img.shape[2:])[:, 0] # if use K of CUT3R, unpad the scores\n            pr_scores = unpad_image(pr_scores[:,None], gt_img.shape[2:])[:, 0]\n            \n            gt_shape, pr_shape = gt[\"smpl_shape\"].shape[-1], pred[\"smpl_shape\"].shape[-1]\n            gt_v3d = gt[\"smpl_v3d\"][smpl_mask]\n            if int(smpl_mask.sum()) == 0:\n                pr_v3d = torch.zeros_like(gt_v3d)\n            else:\n                pr_v3d = loss_details[f\"pred_smpl_v3d_{i+1}\"][smpl_mask]\n            \n            gt_hms, pr_hms, gt_msks, pr_msks, gt_smpls, pr_smpls= [], [], [], [], [], []\n            for k in range(len(gt_img)):\n                img_array = (0.5 * (gt_img[k] + 1.0)).permute(1, 2, 0)\n\n                if has_msk:\n                    gt_msk_array = vis_heatmap(img_array, gt_msk[k])\n                    pr_msk_array = vis_heatmap(img_array, pr_msk[k])\n\n                gt_hm_array = vis_heatmap(img_array, gt_scores[k])\n                pr_hm_array = vis_heatmap(img_array, pr_scores[k])\n\n                img_array_np = (img_array * 255).cpu().numpy().astype(np.uint8)\n                focal = K[k,[0,1],[0,1]].cpu().numpy()\n                princpt = K[k,[0,1],[-1,-1]].cpu().numpy()\n                gt_verts, gt_faces, pr_verts, pr_faces = [], [], [], []\n                for j in range(len(idx_h[0])):\n                    if idx_h[0][j] == k:\n                        gt_verts.append(gt_v3d[j].detach().cpu().numpy().reshape(-1,3))\n                        gt_faces.append(smpl_faces[\"neutral\"][gt_shape])\n                        pr_verts.append(pr_v3d[j].detach().cpu().numpy().reshape(-1,3))\n                        pr_faces.append(smpl_faces[\"neutral\"][pr_shape])\n                gt_rend_array = torch.as_tensor(\n                    render_meshes(img_array_np.copy(), \n                    gt_verts, gt_faces,\n                    {'focal': focal, 'princpt': princpt}),\n                    ) / 255.0\n                pr_rend_array = torch.as_tensor(\n                    render_meshes(img_array_np.copy(), \n                    pr_verts, pr_faces,\n                    {'focal': focal, 'princpt': princpt}),\n                    ) / 255.0\n\n                gt_hms.append(gt_hm_array)\n                pr_hms.append(pr_hm_array)\n                gt_smpls.append(gt_rend_array)\n                pr_smpls.append(pr_rend_array)\n                if has_msk:\n                    gt_msks.append(gt_msk_array)\n                    pr_msks.append(pr_msk_array)\n       \n            gt_hms_list.append(torch.stack(gt_hms, 0))\n            pr_hms_list.append(torch.stack(pr_hms, 0))\n            gt_smpls_list.append(torch.stack(gt_smpls, 0))\n            pr_smpls_list.append(torch.stack(pr_smpls, 0))\n            if has_msk:\n                gt_msks_list.append(torch.stack(gt_msks, 0))\n                pr_msks_list.append(torch.stack(pr_msks, 0)) \n    return (\n        gt_msks_list, pr_msks_list, \n        gt_hms_list, pr_hms_list, \n        gt_smpls_list, pr_smpls_list\n        )\n\n\nOPENCV_TO_OPENGL_CAMERA_CONVENTION = np.array([[1, 0, 0, 0],\n                                               [0, -1, 0, 0],\n                                               [0, 0, -1, 0],\n                                               [0, 0, 0, 1]])\n\ndef render_meshes(img, l_mesh, l_face, cam_param, color=None, alpha=1.0, \n                  show_camera=False,\n                  intensity=3.0,\n                  metallicFactor=0., roughnessFactor=0.5, smooth=True,\n                  ):\n    \"\"\"\n    Rendering multiple mesh and project then in the initial image.\n    Args:\n        - img: np.array [w,h,3]\n        - l_mesh: np.array list of [v,3]\n        - l_face: np.array list of [f,3]\n        - cam_param: info about the camera intrinsics (focal, princpt) and (R,t) is possible\n    Return:\n        - img: np.array [w,h,3]\n    \"\"\"\n    # scene\n    scene = pyrender.Scene(ambient_light=(0.3, 0.3, 0.3))\n\n    # mesh\n    for i, mesh in enumerate(l_mesh):\n        if color is None:\n            _color = (np.random.choice(range(1,225))/255, np.random.choice(range(1,225))/255, np.random.choice(range(1,225))/255)\n        else:\n            if isinstance(color,list):\n                _color = color[i]\n            elif isinstance(color,tuple):\n                _color = color\n            else:\n                raise NotImplementedError\n        mesh = trimesh.Trimesh(mesh, l_face[i])\n        \n        material = pyrender.MetallicRoughnessMaterial(\n            metallicFactor=metallicFactor,\n            roughnessFactor=roughnessFactor,\n            alphaMode='OPAQUE',\n            baseColorFactor=(_color[0], _color[1], _color[2], 1.0))\n        mesh = pyrender.Mesh.from_trimesh(mesh, material=material, smooth=smooth)\n        scene.add(mesh, f\"mesh_{i}\")\n\n    # Adding coordinate system at (0,0,2) for the moment\n    # Using lines defined in pyramid https://docs.pyvista.org/version/stable/api/utilities/_autosummary/pyvista.Pyramid.html\n    if show_camera:\n        import pyvista\n\n        def get_faces(x):\n            return x.faces.astype(np.uint32).reshape((x.n_faces, 4))[:, 1:]\n        \n        # Camera = Box + Cone (or Cylinder?)\n        material_cam = pyrender.MetallicRoughnessMaterial(metallicFactor=metallicFactor, roughnessFactor=roughnessFactor, alphaMode='OPAQUE', baseColorFactor=(0.5,0.5,0.5))\n        height = 0.2\n        radius = 0.1\n        cone = pyvista.Cone(center=(0.0, 0.0, -height/2), direction=(0.0, 0.0, -1.0), height=height, radius=radius).extract_surface().triangulate()\n        verts = cone.points\n        mesh = pyrender.Mesh.from_trimesh(trimesh.Trimesh(verts, get_faces(cone)), material=material_cam, smooth=smooth)\n        scene.add(mesh, f\"cone\")\n\n        size = 0.1\n        box = pyvista.Box(bounds=(-size, size, \n                                  -size, size, \n                                  verts[:,-1].min() - 3*size, verts[:,-1].min())).extract_surface().triangulate()\n        verts = box.points\n        mesh = pyrender.Mesh.from_trimesh(trimesh.Trimesh(verts, get_faces(box)), material=material_cam, smooth=smooth)\n        scene.add(mesh, f\"box\")\n        \n\n        # Coordinate system\n        # https://docs.pyvista.org/version/stable/api/utilities/_autosummary/pyvista.Arrow.html\n        l_color = [(1,0,0,1.0), (0,1,0,1.0), (0,0,1,1.0)]\n        l_direction = [(1,0,0), (0,1,0), (0,0,1)]\n        scale = 0.2\n        pose3d = [2*scale, 0.0, -scale]\n        for i in range(len(l_color)):\n            arrow = pyvista.Arrow(direction=l_direction[i], scale=scale)\n            arrow = arrow.extract_surface().triangulate()\n            verts = arrow.points + np.asarray([pose3d])\n            faces = arrow.faces.astype(np.uint32).reshape((arrow.n_faces, 4))[:, 1:]\n            mesh = trimesh.Trimesh(verts, faces)\n            material = pyrender.MetallicRoughnessMaterial(metallicFactor=metallicFactor, roughnessFactor=roughnessFactor, alphaMode='OPAQUE', baseColorFactor=l_color[i])\n            mesh = pyrender.Mesh.from_trimesh(mesh, material=material, smooth=smooth)\n            scene.add(mesh, f\"arrow_{i}\")\n    \n    focal, princpt = cam_param['focal'], cam_param['princpt']\n    camera_pose = np.eye(4)\n    if 'R' in cam_param.keys():\n        camera_pose[:3, :3] = cam_param['R']\n    if 't' in cam_param.keys():\n        camera_pose[:3, 3] = cam_param['t']\n    camera = pyrender.IntrinsicsCamera(fx=focal[0], fy=focal[1], cx=princpt[0], cy=princpt[1])\n    \n    # camera\n    camera_pose = OPENCV_TO_OPENGL_CAMERA_CONVENTION @ camera_pose\n    camera_pose = np.linalg.inv(camera_pose)\n    scene.add(camera, pose=camera_pose)\n \n    # renderer\n    renderer = pyrender.OffscreenRenderer(viewport_width=img.shape[1], viewport_height=img.shape[0], point_size=1.0)\n   \n    # light\n    light = pyrender.DirectionalLight(intensity=intensity)\n    scene.add(light, pose=camera_pose)\n\n    # render\n    rgb, depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)\n    rgb = rgb[:,:,:3].astype(np.float32)\n    fg = (depth > 0)[:,:,None].astype(np.float32)\n\n    # Simple smoothing of the mask\n    bg_blending_radius = 1\n    bg_blending_kernel = 2.0 * torch.ones((1, 1, 2 * bg_blending_radius + 1, 2 * bg_blending_radius + 1)) / (2 * bg_blending_radius + 1) ** 2\n    bg_blending_bias =  -torch.ones(1)\n    fg = fg.reshape((fg.shape[0],fg.shape[1]))\n    fg = torch.from_numpy(fg).unsqueeze(0)\n    fg = torch.clamp_min(torch.nn.functional.conv2d(fg, weight=bg_blending_kernel, bias=bg_blending_bias, padding=bg_blending_radius) * fg, 0.0)\n    fg = fg.permute(1,2,0).numpy()\n\n    # Alpha-blending\n    img = (fg * (alpha * rgb + (1.0-alpha) * img) + (1-fg) * img).astype(np.uint8)\n\n    renderer.delete()\n\n    return img.astype(np.uint8)","source_hash":"90a8f9d4510f97514fa7c7a7883b7d20b06567f8d036b66bba6dd74d06532bbe","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.render.render","uri":"program://Human3R/function/src.dust3r.utils.render.render#L17-L62","kind":"function","name":"render","path":"src/dust3r/utils/render.py","language":"python","start_line":17,"end_line":62,"context_start_line":1,"context_end_line":82,"code":"import os\nos.environ['PYOPENGL_PLATFORM'] = 'egl'\n\nimport torch\nfrom gsplat import rasterization\nfrom dust3r.utils.geometry import inv, geotrf\nfrom dust3r.utils.image import unpad_image\nimport numpy as np\ntry:\n    import pyrender\nexcept:\n    import pyrender\n\nimport trimesh\nfrom PIL import Image\n\ndef render(\n    intrinsics: torch.Tensor,\n    pts3d: torch.Tensor,\n    rgbs: torch.Tensor | None = None,\n    scale: float = 0.002,\n    opacity: float = 0.95,\n):\n\n    device = pts3d.device\n    batch_size = len(intrinsics)\n    img_size = pts3d.shape[1:3]\n    pts3d = pts3d.reshape(batch_size, -1, 3)\n    num_pts = pts3d.shape[1]\n    quats = torch.randn((num_pts, 4), device=device)\n    quats = quats / quats.norm(dim=-1, keepdim=True)\n    scales = scale * torch.ones((num_pts, 3), device=device)\n    opacities = opacity * torch.ones((num_pts), device=device)\n    if rgbs is not None:\n        assert rgbs.shape[1] == 3\n        rgbs = rgbs.reshape(batch_size, 3, -1).transpose(1, 2)\n    else:\n        rgbs = torch.ones_like(pts3d[:, :, :3])\n\n    rendered_rgbs = []\n    rendered_depths = []\n    accs = []\n    for i in range(batch_size):\n        rgbd, acc, _ = rasterization(\n            pts3d[i],\n            quats,\n            scales,\n            opacities,\n            rgbs[i],\n            torch.eye(4, device=device)[None],\n            intrinsics[[i]],\n            width=img_size[1],\n            height=img_size[0],\n            packed=False,\n            render_mode=\"RGB+D\",\n        )\n\n        rendered_depths.append(rgbd[..., 3])\n\n    rendered_depths = torch.cat(rendered_depths, dim=0)\n\n    return rendered_rgbs, rendered_depths, accs\n\n\ndef get_render_results(gts, preds, self_view=False):\n    device = preds[0][\"pts3d_in_self_view\"].device\n    with torch.no_grad():\n        depths = []\n        gt_depths = []\n        for i, (gt, pred) in enumerate(zip(gts, preds)):\n            if self_view:\n                camera = inv(gt[\"camera_pose\"]).to(device)\n                intrinsics = gt[\"camera_intrinsics\"].to(device)\n                pred = pred[\"pts3d_in_self_view\"]\n            else:\n                camera = inv(gts[0][\"camera_pose\"]).to(device)\n                intrinsics = gts[0][\"camera_intrinsics\"].to(device)\n                pred = pred[\"pts3d_in_other_view\"]\n            gt_img = gt[\"img\"].to(device)\n            gt_pts3d = gt[\"pts3d\"].to(device)\n\n            _, depth, _ = render(intrinsics, pred, gt_img)","source_hash":"90a8f9d4510f97514fa7c7a7883b7d20b06567f8d036b66bba6dd74d06532bbe","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.render.get_render_results","uri":"program://Human3R/function/src.dust3r.utils.render.get_render_results#L65-L86","kind":"function","name":"get_render_results","path":"src/dust3r/utils/render.py","language":"python","start_line":65,"end_line":86,"context_start_line":45,"context_end_line":106,"code":"            pts3d[i],\n            quats,\n            scales,\n            opacities,\n            rgbs[i],\n            torch.eye(4, device=device)[None],\n            intrinsics[[i]],\n            width=img_size[1],\n            height=img_size[0],\n            packed=False,\n            render_mode=\"RGB+D\",\n        )\n\n        rendered_depths.append(rgbd[..., 3])\n\n    rendered_depths = torch.cat(rendered_depths, dim=0)\n\n    return rendered_rgbs, rendered_depths, accs\n\n\ndef get_render_results(gts, preds, self_view=False):\n    device = preds[0][\"pts3d_in_self_view\"].device\n    with torch.no_grad():\n        depths = []\n        gt_depths = []\n        for i, (gt, pred) in enumerate(zip(gts, preds)):\n            if self_view:\n                camera = inv(gt[\"camera_pose\"]).to(device)\n                intrinsics = gt[\"camera_intrinsics\"].to(device)\n                pred = pred[\"pts3d_in_self_view\"]\n            else:\n                camera = inv(gts[0][\"camera_pose\"]).to(device)\n                intrinsics = gts[0][\"camera_intrinsics\"].to(device)\n                pred = pred[\"pts3d_in_other_view\"]\n            gt_img = gt[\"img\"].to(device)\n            gt_pts3d = gt[\"pts3d\"].to(device)\n\n            _, depth, _ = render(intrinsics, pred, gt_img)\n            _, gt_depth, _ = render(intrinsics, geotrf(camera, gt_pts3d), gt_img)\n            depths.append(depth)\n            gt_depths.append(gt_depth)\n    return depths, gt_depths\n\n\ndef vis_heatmap(img, scores):\n    hm = scores.clone()\n    hm = torch.clamp(hm + 0.1, 0, 1)  # for visu purpose only\n    hm = hm.unsqueeze(0).unsqueeze(0)\n    hm = torch.nn.functional.interpolate(\n        hm, \n        size=(img.shape[0], img.shape[1]),\n        mode='nearest'\n    ).squeeze(0).squeeze(0)\n    \n    hm = hm.unsqueeze(-1)\n    \n    return img * hm\n\ndef get_render_smpl(gts, preds, smpl_model, loss_details, has_msk=False):\n    with torch.no_grad():\n        smpl_faces = {\n            'neutral': {","source_hash":"90a8f9d4510f97514fa7c7a7883b7d20b06567f8d036b66bba6dd74d06532bbe","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.render.vis_heatmap","uri":"program://Human3R/function/src.dust3r.utils.render.vis_heatmap#L89-L101","kind":"function","name":"vis_heatmap","path":"src/dust3r/utils/render.py","language":"python","start_line":89,"end_line":101,"context_start_line":69,"context_end_line":121,"code":"        gt_depths = []\n        for i, (gt, pred) in enumerate(zip(gts, preds)):\n            if self_view:\n                camera = inv(gt[\"camera_pose\"]).to(device)\n                intrinsics = gt[\"camera_intrinsics\"].to(device)\n                pred = pred[\"pts3d_in_self_view\"]\n            else:\n                camera = inv(gts[0][\"camera_pose\"]).to(device)\n                intrinsics = gts[0][\"camera_intrinsics\"].to(device)\n                pred = pred[\"pts3d_in_other_view\"]\n            gt_img = gt[\"img\"].to(device)\n            gt_pts3d = gt[\"pts3d\"].to(device)\n\n            _, depth, _ = render(intrinsics, pred, gt_img)\n            _, gt_depth, _ = render(intrinsics, geotrf(camera, gt_pts3d), gt_img)\n            depths.append(depth)\n            gt_depths.append(gt_depth)\n    return depths, gt_depths\n\n\ndef vis_heatmap(img, scores):\n    hm = scores.clone()\n    hm = torch.clamp(hm + 0.1, 0, 1)  # for visu purpose only\n    hm = hm.unsqueeze(0).unsqueeze(0)\n    hm = torch.nn.functional.interpolate(\n        hm, \n        size=(img.shape[0], img.shape[1]),\n        mode='nearest'\n    ).squeeze(0).squeeze(0)\n    \n    hm = hm.unsqueeze(-1)\n    \n    return img * hm\n\ndef get_render_smpl(gts, preds, smpl_model, loss_details, has_msk=False):\n    with torch.no_grad():\n        smpl_faces = {\n            'neutral': {\n                10: smpl_model.smplx_neutral_10.faces,\n                11: smpl_model.smplx_neutral_11.faces,\n            }\n        }\n    \n        gt_hms_list, pr_hms_list, gt_smpls_list, pr_smpls_list= [], [], [], []\n        gt_msks_list, pr_msks_list = [], []\n        for i, (gt, pred) in enumerate(zip(gts, preds)):\n            gt_img = gt[\"img\"]\n            smpl_mask = gt[\"smpl_mask\"]\n            K = gt[\"camera_intrinsics\"]\n\n            idx_h = torch.where(smpl_mask)\n            if has_msk:\n                gt_msk, pr_msk = gt[\"msk_mhmr\"], pred[\"msk\"][...,0]","source_hash":"90a8f9d4510f97514fa7c7a7883b7d20b06567f8d036b66bba6dd74d06532bbe","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.render.get_render_smpl","uri":"program://Human3R/function/src.dust3r.utils.render.get_render_smpl#L103-L186","kind":"function","name":"get_render_smpl","path":"src/dust3r/utils/render.py","language":"python","start_line":103,"end_line":186,"context_start_line":83,"context_end_line":206,"code":"            _, gt_depth, _ = render(intrinsics, geotrf(camera, gt_pts3d), gt_img)\n            depths.append(depth)\n            gt_depths.append(gt_depth)\n    return depths, gt_depths\n\n\ndef vis_heatmap(img, scores):\n    hm = scores.clone()\n    hm = torch.clamp(hm + 0.1, 0, 1)  # for visu purpose only\n    hm = hm.unsqueeze(0).unsqueeze(0)\n    hm = torch.nn.functional.interpolate(\n        hm, \n        size=(img.shape[0], img.shape[1]),\n        mode='nearest'\n    ).squeeze(0).squeeze(0)\n    \n    hm = hm.unsqueeze(-1)\n    \n    return img * hm\n\ndef get_render_smpl(gts, preds, smpl_model, loss_details, has_msk=False):\n    with torch.no_grad():\n        smpl_faces = {\n            'neutral': {\n                10: smpl_model.smplx_neutral_10.faces,\n                11: smpl_model.smplx_neutral_11.faces,\n            }\n        }\n    \n        gt_hms_list, pr_hms_list, gt_smpls_list, pr_smpls_list= [], [], [], []\n        gt_msks_list, pr_msks_list = [], []\n        for i, (gt, pred) in enumerate(zip(gts, preds)):\n            gt_img = gt[\"img\"]\n            smpl_mask = gt[\"smpl_mask\"]\n            K = gt[\"camera_intrinsics\"]\n\n            idx_h = torch.where(smpl_mask)\n            if has_msk:\n                gt_msk, pr_msk = gt[\"msk_mhmr\"], pred[\"msk\"][...,0]\n                gt_msk = unpad_image(gt_msk[:,None], gt_img.shape[2:])[:, 0]\n                pr_msk = unpad_image(pr_msk[:,None], gt_img.shape[2:])[:, 0]\n            gt_scores, pr_scores = gt[\"smpl_scores\"], pred[\"smpl_scores\"][...,0]\n            gt_scores = unpad_image(gt_scores[:,None], gt_img.shape[2:])[:, 0] # if use K of CUT3R, unpad the scores\n            pr_scores = unpad_image(pr_scores[:,None], gt_img.shape[2:])[:, 0]\n            \n            gt_shape, pr_shape = gt[\"smpl_shape\"].shape[-1], pred[\"smpl_shape\"].shape[-1]\n            gt_v3d = gt[\"smpl_v3d\"][smpl_mask]\n            if int(smpl_mask.sum()) == 0:\n                pr_v3d = torch.zeros_like(gt_v3d)\n            else:\n                pr_v3d = loss_details[f\"pred_smpl_v3d_{i+1}\"][smpl_mask]\n            \n            gt_hms, pr_hms, gt_msks, pr_msks, gt_smpls, pr_smpls= [], [], [], [], [], []\n            for k in range(len(gt_img)):\n                img_array = (0.5 * (gt_img[k] + 1.0)).permute(1, 2, 0)\n\n                if has_msk:\n                    gt_msk_array = vis_heatmap(img_array, gt_msk[k])\n                    pr_msk_array = vis_heatmap(img_array, pr_msk[k])\n\n                gt_hm_array = vis_heatmap(img_array, gt_scores[k])\n                pr_hm_array = vis_heatmap(img_array, pr_scores[k])\n\n                img_array_np = (img_array * 255).cpu().numpy().astype(np.uint8)\n                focal = K[k,[0,1],[0,1]].cpu().numpy()\n                princpt = K[k,[0,1],[-1,-1]].cpu().numpy()\n                gt_verts, gt_faces, pr_verts, pr_faces = [], [], [], []\n                for j in range(len(idx_h[0])):\n                    if idx_h[0][j] == k:\n                        gt_verts.append(gt_v3d[j].detach().cpu().numpy().reshape(-1,3))\n                        gt_faces.append(smpl_faces[\"neutral\"][gt_shape])\n                        pr_verts.append(pr_v3d[j].detach().cpu().numpy().reshape(-1,3))\n                        pr_faces.append(smpl_faces[\"neutral\"][pr_shape])\n                gt_rend_array = torch.as_tensor(\n                    render_meshes(img_array_np.copy(), \n                    gt_verts, gt_faces,\n                    {'focal': focal, 'princpt': princpt}),\n                    ) / 255.0\n                pr_rend_array = torch.as_tensor(\n                    render_meshes(img_array_np.copy(), \n                    pr_verts, pr_faces,\n                    {'focal': focal, 'princpt': princpt}),\n                    ) / 255.0\n\n                gt_hms.append(gt_hm_array)\n                pr_hms.append(pr_hm_array)\n                gt_smpls.append(gt_rend_array)\n                pr_smpls.append(pr_rend_array)\n                if has_msk:\n                    gt_msks.append(gt_msk_array)\n                    pr_msks.append(pr_msk_array)\n       \n            gt_hms_list.append(torch.stack(gt_hms, 0))\n            pr_hms_list.append(torch.stack(pr_hms, 0))\n            gt_smpls_list.append(torch.stack(gt_smpls, 0))\n            pr_smpls_list.append(torch.stack(pr_smpls, 0))\n            if has_msk:\n                gt_msks_list.append(torch.stack(gt_msks, 0))\n                pr_msks_list.append(torch.stack(pr_msks, 0)) \n    return (\n        gt_msks_list, pr_msks_list, \n        gt_hms_list, pr_hms_list, \n        gt_smpls_list, pr_smpls_list\n        )\n\n\nOPENCV_TO_OPENGL_CAMERA_CONVENTION = np.array([[1, 0, 0, 0],\n                                               [0, -1, 0, 0],\n                                               [0, 0, -1, 0],\n                                               [0, 0, 0, 1]])\n\ndef render_meshes(img, l_mesh, l_face, cam_param, color=None, alpha=1.0, \n                  show_camera=False,\n                  intensity=3.0,\n                  metallicFactor=0., roughnessFactor=0.5, smooth=True,\n                  ):\n    \"\"\"\n    Rendering multiple mesh and project then in the initial image.\n    Args:\n        - img: np.array [w,h,3]\n        - l_mesh: np.array list of [v,3]\n        - l_face: np.array list of [f,3]\n        - cam_param: info about the camera intrinsics (focal, princpt) and (R,t) is possible\n    Return:","source_hash":"90a8f9d4510f97514fa7c7a7883b7d20b06567f8d036b66bba6dd74d06532bbe","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.render.render_meshes","uri":"program://Human3R/function/src.dust3r.utils.render.render_meshes#L194-L314","kind":"function","name":"render_meshes","path":"src/dust3r/utils/render.py","language":"python","start_line":194,"end_line":314,"context_start_line":174,"context_end_line":314,"code":"       \n            gt_hms_list.append(torch.stack(gt_hms, 0))\n            pr_hms_list.append(torch.stack(pr_hms, 0))\n            gt_smpls_list.append(torch.stack(gt_smpls, 0))\n            pr_smpls_list.append(torch.stack(pr_smpls, 0))\n            if has_msk:\n                gt_msks_list.append(torch.stack(gt_msks, 0))\n                pr_msks_list.append(torch.stack(pr_msks, 0)) \n    return (\n        gt_msks_list, pr_msks_list, \n        gt_hms_list, pr_hms_list, \n        gt_smpls_list, pr_smpls_list\n        )\n\n\nOPENCV_TO_OPENGL_CAMERA_CONVENTION = np.array([[1, 0, 0, 0],\n                                               [0, -1, 0, 0],\n                                               [0, 0, -1, 0],\n                                               [0, 0, 0, 1]])\n\ndef render_meshes(img, l_mesh, l_face, cam_param, color=None, alpha=1.0, \n                  show_camera=False,\n                  intensity=3.0,\n                  metallicFactor=0., roughnessFactor=0.5, smooth=True,\n                  ):\n    \"\"\"\n    Rendering multiple mesh and project then in the initial image.\n    Args:\n        - img: np.array [w,h,3]\n        - l_mesh: np.array list of [v,3]\n        - l_face: np.array list of [f,3]\n        - cam_param: info about the camera intrinsics (focal, princpt) and (R,t) is possible\n    Return:\n        - img: np.array [w,h,3]\n    \"\"\"\n    # scene\n    scene = pyrender.Scene(ambient_light=(0.3, 0.3, 0.3))\n\n    # mesh\n    for i, mesh in enumerate(l_mesh):\n        if color is None:\n            _color = (np.random.choice(range(1,225))/255, np.random.choice(range(1,225))/255, np.random.choice(range(1,225))/255)\n        else:\n            if isinstance(color,list):\n                _color = color[i]\n            elif isinstance(color,tuple):\n                _color = color\n            else:\n                raise NotImplementedError\n        mesh = trimesh.Trimesh(mesh, l_face[i])\n        \n        material = pyrender.MetallicRoughnessMaterial(\n            metallicFactor=metallicFactor,\n            roughnessFactor=roughnessFactor,\n            alphaMode='OPAQUE',\n            baseColorFactor=(_color[0], _color[1], _color[2], 1.0))\n        mesh = pyrender.Mesh.from_trimesh(mesh, material=material, smooth=smooth)\n        scene.add(mesh, f\"mesh_{i}\")\n\n    # Adding coordinate system at (0,0,2) for the moment\n    # Using lines defined in pyramid https://docs.pyvista.org/version/stable/api/utilities/_autosummary/pyvista.Pyramid.html\n    if show_camera:\n        import pyvista\n\n        def get_faces(x):\n            return x.faces.astype(np.uint32).reshape((x.n_faces, 4))[:, 1:]\n        \n        # Camera = Box + Cone (or Cylinder?)\n        material_cam = pyrender.MetallicRoughnessMaterial(metallicFactor=metallicFactor, roughnessFactor=roughnessFactor, alphaMode='OPAQUE', baseColorFactor=(0.5,0.5,0.5))\n        height = 0.2\n        radius = 0.1\n        cone = pyvista.Cone(center=(0.0, 0.0, -height/2), direction=(0.0, 0.0, -1.0), height=height, radius=radius).extract_surface().triangulate()\n        verts = cone.points\n        mesh = pyrender.Mesh.from_trimesh(trimesh.Trimesh(verts, get_faces(cone)), material=material_cam, smooth=smooth)\n        scene.add(mesh, f\"cone\")\n\n        size = 0.1\n        box = pyvista.Box(bounds=(-size, size, \n                                  -size, size, \n                                  verts[:,-1].min() - 3*size, verts[:,-1].min())).extract_surface().triangulate()\n        verts = box.points\n        mesh = pyrender.Mesh.from_trimesh(trimesh.Trimesh(verts, get_faces(box)), material=material_cam, smooth=smooth)\n        scene.add(mesh, f\"box\")\n        \n\n        # Coordinate system\n        # https://docs.pyvista.org/version/stable/api/utilities/_autosummary/pyvista.Arrow.html\n        l_color = [(1,0,0,1.0), (0,1,0,1.0), (0,0,1,1.0)]\n        l_direction = [(1,0,0), (0,1,0), (0,0,1)]\n        scale = 0.2\n        pose3d = [2*scale, 0.0, -scale]\n        for i in range(len(l_color)):\n            arrow = pyvista.Arrow(direction=l_direction[i], scale=scale)\n            arrow = arrow.extract_surface().triangulate()\n            verts = arrow.points + np.asarray([pose3d])\n            faces = arrow.faces.astype(np.uint32).reshape((arrow.n_faces, 4))[:, 1:]\n            mesh = trimesh.Trimesh(verts, faces)\n            material = pyrender.MetallicRoughnessMaterial(metallicFactor=metallicFactor, roughnessFactor=roughnessFactor, alphaMode='OPAQUE', baseColorFactor=l_color[i])\n            mesh = pyrender.Mesh.from_trimesh(mesh, material=material, smooth=smooth)\n            scene.add(mesh, f\"arrow_{i}\")\n    \n    focal, princpt = cam_param['focal'], cam_param['princpt']\n    camera_pose = np.eye(4)\n    if 'R' in cam_param.keys():\n        camera_pose[:3, :3] = cam_param['R']\n    if 't' in cam_param.keys():\n        camera_pose[:3, 3] = cam_param['t']\n    camera = pyrender.IntrinsicsCamera(fx=focal[0], fy=focal[1], cx=princpt[0], cy=princpt[1])\n    \n    # camera\n    camera_pose = OPENCV_TO_OPENGL_CAMERA_CONVENTION @ camera_pose\n    camera_pose = np.linalg.inv(camera_pose)\n    scene.add(camera, pose=camera_pose)\n \n    # renderer\n    renderer = pyrender.OffscreenRenderer(viewport_width=img.shape[1], viewport_height=img.shape[0], point_size=1.0)\n   \n    # light\n    light = pyrender.DirectionalLight(intensity=intensity)\n    scene.add(light, pose=camera_pose)\n\n    # render\n    rgb, depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)\n    rgb = rgb[:,:,:3].astype(np.float32)\n    fg = (depth > 0)[:,:,None].astype(np.float32)\n\n    # Simple smoothing of the mask\n    bg_blending_radius = 1\n    bg_blending_kernel = 2.0 * torch.ones((1, 1, 2 * bg_blending_radius + 1, 2 * bg_blending_radius + 1)) / (2 * bg_blending_radius + 1) ** 2\n    bg_blending_bias =  -torch.ones(1)\n    fg = fg.reshape((fg.shape[0],fg.shape[1]))\n    fg = torch.from_numpy(fg).unsqueeze(0)\n    fg = torch.clamp_min(torch.nn.functional.conv2d(fg, weight=bg_blending_kernel, bias=bg_blending_bias, padding=bg_blending_radius) * fg, 0.0)\n    fg = fg.permute(1,2,0).numpy()\n\n    # Alpha-blending\n    img = (fg * (alpha * rgb + (1.0-alpha) * img) + (1-fg) * img).astype(np.uint8)\n\n    renderer.delete()\n\n    return img.astype(np.uint8)","source_hash":"90a8f9d4510f97514fa7c7a7883b7d20b06567f8d036b66bba6dd74d06532bbe","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.render.get_faces","uri":"program://Human3R/function/src.dust3r.utils.render.get_faces#L238-L239","kind":"function","name":"get_faces","path":"src/dust3r/utils/render.py","language":"python","start_line":238,"end_line":239,"context_start_line":218,"context_end_line":259,"code":"                _color = color[i]\n            elif isinstance(color,tuple):\n                _color = color\n            else:\n                raise NotImplementedError\n        mesh = trimesh.Trimesh(mesh, l_face[i])\n        \n        material = pyrender.MetallicRoughnessMaterial(\n            metallicFactor=metallicFactor,\n            roughnessFactor=roughnessFactor,\n            alphaMode='OPAQUE',\n            baseColorFactor=(_color[0], _color[1], _color[2], 1.0))\n        mesh = pyrender.Mesh.from_trimesh(mesh, material=material, smooth=smooth)\n        scene.add(mesh, f\"mesh_{i}\")\n\n    # Adding coordinate system at (0,0,2) for the moment\n    # Using lines defined in pyramid https://docs.pyvista.org/version/stable/api/utilities/_autosummary/pyvista.Pyramid.html\n    if show_camera:\n        import pyvista\n\n        def get_faces(x):\n            return x.faces.astype(np.uint32).reshape((x.n_faces, 4))[:, 1:]\n        \n        # Camera = Box + Cone (or Cylinder?)\n        material_cam = pyrender.MetallicRoughnessMaterial(metallicFactor=metallicFactor, roughnessFactor=roughnessFactor, alphaMode='OPAQUE', baseColorFactor=(0.5,0.5,0.5))\n        height = 0.2\n        radius = 0.1\n        cone = pyvista.Cone(center=(0.0, 0.0, -height/2), direction=(0.0, 0.0, -1.0), height=height, radius=radius).extract_surface().triangulate()\n        verts = cone.points\n        mesh = pyrender.Mesh.from_trimesh(trimesh.Trimesh(verts, get_faces(cone)), material=material_cam, smooth=smooth)\n        scene.add(mesh, f\"cone\")\n\n        size = 0.1\n        box = pyvista.Box(bounds=(-size, size, \n                                  -size, size, \n                                  verts[:,-1].min() - 3*size, verts[:,-1].min())).extract_surface().triangulate()\n        verts = box.points\n        mesh = pyrender.Mesh.from_trimesh(trimesh.Trimesh(verts, get_faces(box)), material=material_cam, smooth=smooth)\n        scene.add(mesh, f\"box\")\n        \n\n        # Coordinate system","source_hash":"90a8f9d4510f97514fa7c7a7883b7d20b06567f8d036b66bba6dd74d06532bbe","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.smpl_layer","uri":"program://Human3R/module/src.dust3r.utils.smpl_layer#L1-L154","kind":"module","name":"src.dust3r.utils.smpl_layer","path":"src/dust3r/utils/smpl_layer.py","language":"python","start_line":1,"end_line":154,"context_start_line":1,"context_end_line":154,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\nimport smplx\nimport torch\nfrom dust3r.utils.geometry import inverse_perspective_projection, perspective_projection\nimport roma\nfrom dust3r.smpl_model import SMPLX_DIR\nfrom smplx.joint_names import JOINT_NAMES\n\nclass SMPL_Layer(nn.Module):\n    \"\"\"\n    Extension of the SMPL Layer with information about the camera for (inverse) projection the camera plane.\n    \"\"\"\n    def __init__(self, \n                 type='smplx', \n                 gender='neutral', \n                 num_betas=10,\n                 kid=False,\n                 person_center=None,\n                 *args, \n                 **kwargs,\n                 ):\n        super().__init__()\n\n        # Args\n        assert type == 'smplx'\n        self.type = type\n        self.kid = kid\n        self.num_betas = num_betas\n        self.bm_x = smplx.create(SMPLX_DIR, 'smplx', gender=gender, use_pca=False, flat_hand_mean=True, num_betas=num_betas)\n\n        # Primary keypoint - root\n        self.joint_names = JOINT_NAMES[:127]\n        self.person_center = person_center\n        self.person_center_idx = None\n        if self.person_center is not None:\n            self.person_center_idx = self.joint_names.index(self.person_center)\n\n    def forward(self,\n                pose, shape, transl,\n                loc, dist,\n                K,\n                expression=None, # facial expression\n                K_to_proj=None,\n                ):\n        \"\"\"\n        Args:\n            - pose: pose of the person in axis-angle - torch.Tensor [bs,24,3]\n            - shape: torch.Tensor [bs,10]\n            - loc: 2D location of the pelvis in pixel space - torch.Tensor [bs,2]\n            - dist: distance of the pelvis from the camera in m - torch.Tensor [bs,1]\n        Return:\n            - dict containing a bunch of useful information about each person\n        \"\"\"\n        \n        if loc is not None and dist is not None:\n            assert pose.shape[0] == shape.shape[0] == loc.shape[0] == dist.shape[0]\n            assert len(loc.shape) == 2 and list(loc.shape[1:]) == [2]\n            assert len(dist.shape) == 2 and list(dist.shape[1:]) == [1]\n        if self.type == 'smpl':\n            assert len(pose.shape) == 3 and list(pose.shape[1:]) == [24,3]\n        elif self.type == 'smplx':\n            assert len(pose.shape) == 3 and list(pose.shape[1:]) == [53,3] # taking root_orient, body_pose, lhand, rhan and jaw for the moment\n        else:\n            raise NameError\n        assert len(shape.shape) == 2 and (list(shape.shape[1:]) == [self.num_betas] or list(shape.shape[1:]) == [self.num_betas+1])\n        assert (transl is not None) or (loc is not None and dist is not None)\n\n        bs = pose.shape[0]\n\n        out = {}\n\n        # No humans\n        if bs == 0:\n            return {}\n        \n        # Low dimensional parameters        \n        kwargs_pose = {\n            'betas': shape,\n        }\n        kwargs_pose['global_orient'] = self.bm_x.global_orient.repeat(bs,1) # 0，0，0\n        kwargs_pose['body_pose'] = pose[:,1:22].flatten(1)\n        kwargs_pose['left_hand_pose'] = pose[:,22:37].flatten(1)\n        kwargs_pose['right_hand_pose'] = pose[:,37:52].flatten(1)\n        kwargs_pose['jaw_pose'] = pose[:,52:53].flatten(1)\n\n        if expression is not None:\n            kwargs_pose['expression'] = expression.flatten(1) # [bs,10]\n        else:\n            kwargs_pose['expression'] = self.bm_x.expression.repeat(bs,1)\n\n        # default - to be generalized\n        kwargs_pose['leye_pose'] = self.bm_x.leye_pose.repeat(bs,1)\n        kwargs_pose['reye_pose'] = self.bm_x.reye_pose.repeat(bs,1)        \n        \n        # Forward using the parametric 3d model SMPL-X layer\n        output = self.bm_x(**kwargs_pose)\n        verts = output.vertices\n        j3d = output.joints # 45 joints\n        R = roma.rotvec_to_rotmat(pose[:,0])\n\n        # Apply global orientation on 3D points\n        pelvis = j3d[:,[0]]\n        j3d = (R.unsqueeze(1) @ (j3d - pelvis).unsqueeze(-1)).squeeze(-1)\n        \n        # Apply global orientation on 3D points - bis\n        verts = (R.unsqueeze(1) @ (verts - pelvis).unsqueeze(-1)).squeeze(-1) # to pelvis-center coordinates: R(v-p)\n\n        # Location of the person in 3D\n        if transl is None:\n            if K.dtype == torch.float16:\n                # because of torch.inverse - not working with float16 at the moment\n                transl = inverse_perspective_projection(loc.unsqueeze(1).float(), K.float(), dist.unsqueeze(1).float())[:,0]    \n                transl = transl.half()\n            else:\n                transl = inverse_perspective_projection(loc.unsqueeze(1), K, dist.unsqueeze(1))[:,0] # head center in camera coordinates\n\n        # Updating transl if we choose a certain person center\n        transl_up = transl.clone()\n\n        # Definition of the translation depend on the args: 1) vanilla SMPL - 2) computed from a given joint\n        if self.person_center_idx is None:\n            # Add pelvis to transl - standard way for SMPLX layer\n            transl_up = transl_up + pelvis[:,0] # back to smpl original coordinates\n        else:\n            # Center around the joint because teh translation is computed from this joint\n            person_center = j3d[:, [self.person_center_idx]] # head center in pelvis-center coordinates\n            verts = verts - person_center # nomalize to head-center coordinates\n            j3d = j3d - person_center\n\n        # Moving into the camera coordinate system\n        j3d_cam = j3d + transl_up.unsqueeze(1) # move to camera coordinates\n        verts_cam = verts + transl_up.unsqueeze(1)\n\n        if K_to_proj is None:\n            K_to_proj = K\n        # Projection in camera plane\n        j2d = perspective_projection(j3d_cam, K_to_proj)\n        v2d = perspective_projection(verts_cam, K_to_proj)\n\n        out.update({\n            'smpl_v3d': verts_cam, # in 3d camera space\n            'smpl_j3d': j3d_cam, # in 3d camera space\n            'smpl_j2d': j2d, \n            'smpl_v2d': v2d, \n            'smpl_transl': transl, # translation of the primary keypoint in camera coordinates\n            'smpl_transl_pelvis': j3d_cam[:,[0]], # root=pelvis in camera coordinates\n        })\n            \n        return out","source_hash":"60e448f46c122d3463d5e52753ff8f19c232dff613351f87cb7d5fbdf603ffeb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.smpl_layer.SMPL_Layer","uri":"program://Human3R/class/src.dust3r.utils.smpl_layer.SMPL_Layer#L14-L154","kind":"class","name":"SMPL_Layer","path":"src/dust3r/utils/smpl_layer.py","language":"python","start_line":14,"end_line":154,"context_start_line":1,"context_end_line":154,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\nimport smplx\nimport torch\nfrom dust3r.utils.geometry import inverse_perspective_projection, perspective_projection\nimport roma\nfrom dust3r.smpl_model import SMPLX_DIR\nfrom smplx.joint_names import JOINT_NAMES\n\nclass SMPL_Layer(nn.Module):\n    \"\"\"\n    Extension of the SMPL Layer with information about the camera for (inverse) projection the camera plane.\n    \"\"\"\n    def __init__(self, \n                 type='smplx', \n                 gender='neutral', \n                 num_betas=10,\n                 kid=False,\n                 person_center=None,\n                 *args, \n                 **kwargs,\n                 ):\n        super().__init__()\n\n        # Args\n        assert type == 'smplx'\n        self.type = type\n        self.kid = kid\n        self.num_betas = num_betas\n        self.bm_x = smplx.create(SMPLX_DIR, 'smplx', gender=gender, use_pca=False, flat_hand_mean=True, num_betas=num_betas)\n\n        # Primary keypoint - root\n        self.joint_names = JOINT_NAMES[:127]\n        self.person_center = person_center\n        self.person_center_idx = None\n        if self.person_center is not None:\n            self.person_center_idx = self.joint_names.index(self.person_center)\n\n    def forward(self,\n                pose, shape, transl,\n                loc, dist,\n                K,\n                expression=None, # facial expression\n                K_to_proj=None,\n                ):\n        \"\"\"\n        Args:\n            - pose: pose of the person in axis-angle - torch.Tensor [bs,24,3]\n            - shape: torch.Tensor [bs,10]\n            - loc: 2D location of the pelvis in pixel space - torch.Tensor [bs,2]\n            - dist: distance of the pelvis from the camera in m - torch.Tensor [bs,1]\n        Return:\n            - dict containing a bunch of useful information about each person\n        \"\"\"\n        \n        if loc is not None and dist is not None:\n            assert pose.shape[0] == shape.shape[0] == loc.shape[0] == dist.shape[0]\n            assert len(loc.shape) == 2 and list(loc.shape[1:]) == [2]\n            assert len(dist.shape) == 2 and list(dist.shape[1:]) == [1]\n        if self.type == 'smpl':\n            assert len(pose.shape) == 3 and list(pose.shape[1:]) == [24,3]\n        elif self.type == 'smplx':\n            assert len(pose.shape) == 3 and list(pose.shape[1:]) == [53,3] # taking root_orient, body_pose, lhand, rhan and jaw for the moment\n        else:\n            raise NameError\n        assert len(shape.shape) == 2 and (list(shape.shape[1:]) == [self.num_betas] or list(shape.shape[1:]) == [self.num_betas+1])\n        assert (transl is not None) or (loc is not None and dist is not None)\n\n        bs = pose.shape[0]\n\n        out = {}\n\n        # No humans\n        if bs == 0:\n            return {}\n        \n        # Low dimensional parameters        \n        kwargs_pose = {\n            'betas': shape,\n        }\n        kwargs_pose['global_orient'] = self.bm_x.global_orient.repeat(bs,1) # 0，0，0\n        kwargs_pose['body_pose'] = pose[:,1:22].flatten(1)\n        kwargs_pose['left_hand_pose'] = pose[:,22:37].flatten(1)\n        kwargs_pose['right_hand_pose'] = pose[:,37:52].flatten(1)\n        kwargs_pose['jaw_pose'] = pose[:,52:53].flatten(1)\n\n        if expression is not None:\n            kwargs_pose['expression'] = expression.flatten(1) # [bs,10]\n        else:\n            kwargs_pose['expression'] = self.bm_x.expression.repeat(bs,1)\n\n        # default - to be generalized\n        kwargs_pose['leye_pose'] = self.bm_x.leye_pose.repeat(bs,1)\n        kwargs_pose['reye_pose'] = self.bm_x.reye_pose.repeat(bs,1)        \n        \n        # Forward using the parametric 3d model SMPL-X layer\n        output = self.bm_x(**kwargs_pose)\n        verts = output.vertices\n        j3d = output.joints # 45 joints\n        R = roma.rotvec_to_rotmat(pose[:,0])\n\n        # Apply global orientation on 3D points\n        pelvis = j3d[:,[0]]\n        j3d = (R.unsqueeze(1) @ (j3d - pelvis).unsqueeze(-1)).squeeze(-1)\n        \n        # Apply global orientation on 3D points - bis\n        verts = (R.unsqueeze(1) @ (verts - pelvis).unsqueeze(-1)).squeeze(-1) # to pelvis-center coordinates: R(v-p)\n\n        # Location of the person in 3D\n        if transl is None:\n            if K.dtype == torch.float16:\n                # because of torch.inverse - not working with float16 at the moment\n                transl = inverse_perspective_projection(loc.unsqueeze(1).float(), K.float(), dist.unsqueeze(1).float())[:,0]    \n                transl = transl.half()\n            else:\n                transl = inverse_perspective_projection(loc.unsqueeze(1), K, dist.unsqueeze(1))[:,0] # head center in camera coordinates\n\n        # Updating transl if we choose a certain person center\n        transl_up = transl.clone()\n\n        # Definition of the translation depend on the args: 1) vanilla SMPL - 2) computed from a given joint\n        if self.person_center_idx is None:\n            # Add pelvis to transl - standard way for SMPLX layer\n            transl_up = transl_up + pelvis[:,0] # back to smpl original coordinates\n        else:\n            # Center around the joint because teh translation is computed from this joint\n            person_center = j3d[:, [self.person_center_idx]] # head center in pelvis-center coordinates\n            verts = verts - person_center # nomalize to head-center coordinates\n            j3d = j3d - person_center\n\n        # Moving into the camera coordinate system\n        j3d_cam = j3d + transl_up.unsqueeze(1) # move to camera coordinates\n        verts_cam = verts + transl_up.unsqueeze(1)\n\n        if K_to_proj is None:\n            K_to_proj = K\n        # Projection in camera plane\n        j2d = perspective_projection(j3d_cam, K_to_proj)\n        v2d = perspective_projection(verts_cam, K_to_proj)\n\n        out.update({\n            'smpl_v3d': verts_cam, # in 3d camera space\n            'smpl_j3d': j3d_cam, # in 3d camera space\n            'smpl_j2d': j2d, \n            'smpl_v2d': v2d, \n            'smpl_transl': transl, # translation of the primary keypoint in camera coordinates\n            'smpl_transl_pelvis': j3d_cam[:,[0]], # root=pelvis in camera coordinates\n        })\n            \n        return out","source_hash":"60e448f46c122d3463d5e52753ff8f19c232dff613351f87cb7d5fbdf603ffeb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.smpl_layer.__init__","uri":"program://Human3R/function/src.dust3r.utils.smpl_layer.__init__#L18-L41","kind":"function","name":"__init__","path":"src/dust3r/utils/smpl_layer.py","language":"python","start_line":18,"end_line":41,"context_start_line":1,"context_end_line":61,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\nimport smplx\nimport torch\nfrom dust3r.utils.geometry import inverse_perspective_projection, perspective_projection\nimport roma\nfrom dust3r.smpl_model import SMPLX_DIR\nfrom smplx.joint_names import JOINT_NAMES\n\nclass SMPL_Layer(nn.Module):\n    \"\"\"\n    Extension of the SMPL Layer with information about the camera for (inverse) projection the camera plane.\n    \"\"\"\n    def __init__(self, \n                 type='smplx', \n                 gender='neutral', \n                 num_betas=10,\n                 kid=False,\n                 person_center=None,\n                 *args, \n                 **kwargs,\n                 ):\n        super().__init__()\n\n        # Args\n        assert type == 'smplx'\n        self.type = type\n        self.kid = kid\n        self.num_betas = num_betas\n        self.bm_x = smplx.create(SMPLX_DIR, 'smplx', gender=gender, use_pca=False, flat_hand_mean=True, num_betas=num_betas)\n\n        # Primary keypoint - root\n        self.joint_names = JOINT_NAMES[:127]\n        self.person_center = person_center\n        self.person_center_idx = None\n        if self.person_center is not None:\n            self.person_center_idx = self.joint_names.index(self.person_center)\n\n    def forward(self,\n                pose, shape, transl,\n                loc, dist,\n                K,\n                expression=None, # facial expression\n                K_to_proj=None,\n                ):\n        \"\"\"\n        Args:\n            - pose: pose of the person in axis-angle - torch.Tensor [bs,24,3]\n            - shape: torch.Tensor [bs,10]\n            - loc: 2D location of the pelvis in pixel space - torch.Tensor [bs,2]\n            - dist: distance of the pelvis from the camera in m - torch.Tensor [bs,1]\n        Return:\n            - dict containing a bunch of useful information about each person\n        \"\"\"\n        \n        if loc is not None and dist is not None:\n            assert pose.shape[0] == shape.shape[0] == loc.shape[0] == dist.shape[0]","source_hash":"60e448f46c122d3463d5e52753ff8f19c232dff613351f87cb7d5fbdf603ffeb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.smpl_layer.forward","uri":"program://Human3R/function/src.dust3r.utils.smpl_layer.forward#L43-L154","kind":"function","name":"forward","path":"src/dust3r/utils/smpl_layer.py","language":"python","start_line":43,"end_line":154,"context_start_line":23,"context_end_line":154,"code":"                 person_center=None,\n                 *args, \n                 **kwargs,\n                 ):\n        super().__init__()\n\n        # Args\n        assert type == 'smplx'\n        self.type = type\n        self.kid = kid\n        self.num_betas = num_betas\n        self.bm_x = smplx.create(SMPLX_DIR, 'smplx', gender=gender, use_pca=False, flat_hand_mean=True, num_betas=num_betas)\n\n        # Primary keypoint - root\n        self.joint_names = JOINT_NAMES[:127]\n        self.person_center = person_center\n        self.person_center_idx = None\n        if self.person_center is not None:\n            self.person_center_idx = self.joint_names.index(self.person_center)\n\n    def forward(self,\n                pose, shape, transl,\n                loc, dist,\n                K,\n                expression=None, # facial expression\n                K_to_proj=None,\n                ):\n        \"\"\"\n        Args:\n            - pose: pose of the person in axis-angle - torch.Tensor [bs,24,3]\n            - shape: torch.Tensor [bs,10]\n            - loc: 2D location of the pelvis in pixel space - torch.Tensor [bs,2]\n            - dist: distance of the pelvis from the camera in m - torch.Tensor [bs,1]\n        Return:\n            - dict containing a bunch of useful information about each person\n        \"\"\"\n        \n        if loc is not None and dist is not None:\n            assert pose.shape[0] == shape.shape[0] == loc.shape[0] == dist.shape[0]\n            assert len(loc.shape) == 2 and list(loc.shape[1:]) == [2]\n            assert len(dist.shape) == 2 and list(dist.shape[1:]) == [1]\n        if self.type == 'smpl':\n            assert len(pose.shape) == 3 and list(pose.shape[1:]) == [24,3]\n        elif self.type == 'smplx':\n            assert len(pose.shape) == 3 and list(pose.shape[1:]) == [53,3] # taking root_orient, body_pose, lhand, rhan and jaw for the moment\n        else:\n            raise NameError\n        assert len(shape.shape) == 2 and (list(shape.shape[1:]) == [self.num_betas] or list(shape.shape[1:]) == [self.num_betas+1])\n        assert (transl is not None) or (loc is not None and dist is not None)\n\n        bs = pose.shape[0]\n\n        out = {}\n\n        # No humans\n        if bs == 0:\n            return {}\n        \n        # Low dimensional parameters        \n        kwargs_pose = {\n            'betas': shape,\n        }\n        kwargs_pose['global_orient'] = self.bm_x.global_orient.repeat(bs,1) # 0，0，0\n        kwargs_pose['body_pose'] = pose[:,1:22].flatten(1)\n        kwargs_pose['left_hand_pose'] = pose[:,22:37].flatten(1)\n        kwargs_pose['right_hand_pose'] = pose[:,37:52].flatten(1)\n        kwargs_pose['jaw_pose'] = pose[:,52:53].flatten(1)\n\n        if expression is not None:\n            kwargs_pose['expression'] = expression.flatten(1) # [bs,10]\n        else:\n            kwargs_pose['expression'] = self.bm_x.expression.repeat(bs,1)\n\n        # default - to be generalized\n        kwargs_pose['leye_pose'] = self.bm_x.leye_pose.repeat(bs,1)\n        kwargs_pose['reye_pose'] = self.bm_x.reye_pose.repeat(bs,1)        \n        \n        # Forward using the parametric 3d model SMPL-X layer\n        output = self.bm_x(**kwargs_pose)\n        verts = output.vertices\n        j3d = output.joints # 45 joints\n        R = roma.rotvec_to_rotmat(pose[:,0])\n\n        # Apply global orientation on 3D points\n        pelvis = j3d[:,[0]]\n        j3d = (R.unsqueeze(1) @ (j3d - pelvis).unsqueeze(-1)).squeeze(-1)\n        \n        # Apply global orientation on 3D points - bis\n        verts = (R.unsqueeze(1) @ (verts - pelvis).unsqueeze(-1)).squeeze(-1) # to pelvis-center coordinates: R(v-p)\n\n        # Location of the person in 3D\n        if transl is None:\n            if K.dtype == torch.float16:\n                # because of torch.inverse - not working with float16 at the moment\n                transl = inverse_perspective_projection(loc.unsqueeze(1).float(), K.float(), dist.unsqueeze(1).float())[:,0]    \n                transl = transl.half()\n            else:\n                transl = inverse_perspective_projection(loc.unsqueeze(1), K, dist.unsqueeze(1))[:,0] # head center in camera coordinates\n\n        # Updating transl if we choose a certain person center\n        transl_up = transl.clone()\n\n        # Definition of the translation depend on the args: 1) vanilla SMPL - 2) computed from a given joint\n        if self.person_center_idx is None:\n            # Add pelvis to transl - standard way for SMPLX layer\n            transl_up = transl_up + pelvis[:,0] # back to smpl original coordinates\n        else:\n            # Center around the joint because teh translation is computed from this joint\n            person_center = j3d[:, [self.person_center_idx]] # head center in pelvis-center coordinates\n            verts = verts - person_center # nomalize to head-center coordinates\n            j3d = j3d - person_center\n\n        # Moving into the camera coordinate system\n        j3d_cam = j3d + transl_up.unsqueeze(1) # move to camera coordinates\n        verts_cam = verts + transl_up.unsqueeze(1)\n\n        if K_to_proj is None:\n            K_to_proj = K\n        # Projection in camera plane\n        j2d = perspective_projection(j3d_cam, K_to_proj)\n        v2d = perspective_projection(verts_cam, K_to_proj)\n\n        out.update({\n            'smpl_v3d': verts_cam, # in 3d camera space\n            'smpl_j3d': j3d_cam, # in 3d camera space\n            'smpl_j2d': j2d, \n            'smpl_v2d': v2d, \n            'smpl_transl': transl, # translation of the primary keypoint in camera coordinates\n            'smpl_transl_pelvis': j3d_cam[:,[0]], # root=pelvis in camera coordinates\n        })\n            \n        return out","source_hash":"60e448f46c122d3463d5e52753ff8f19c232dff613351f87cb7d5fbdf603ffeb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.path_to_croco","uri":"program://Human3R/module/src.dust3r.utils.path_to_croco#L1-L21","kind":"module","name":"src.dust3r.utils.path_to_croco","path":"src/dust3r/utils/path_to_croco.py","language":"python","start_line":1,"end_line":21,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport sys\nimport os.path as path\n\nHERE_PATH = path.normpath(path.dirname(__file__))\nCROCO_REPO_PATH = path.normpath(path.join(HERE_PATH, \"../../croco\"))\nCROCO_MODELS_PATH = path.join(CROCO_REPO_PATH, \"models\")\n\nif path.isdir(CROCO_MODELS_PATH):\n\n    sys.path.insert(0, CROCO_REPO_PATH)\nelse:\n    raise ImportError(\n        f\"croco is not initialized, could not find: {CROCO_MODELS_PATH}.\\n \"\n        \"Did you forget to run 'git submodule update --init --recursive' ?\"\n    )","source_hash":"3d8c68320432fd7e3527f6bb94390f2a44195957ee5460539c1d807faf35367e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry","uri":"program://Human3R/module/src.dust3r.utils.geometry#L1-L729","kind":"module","name":"src.dust3r.utils.geometry","path":"src/dust3r/utils/geometry.py","language":"python","start_line":1,"end_line":729,"context_start_line":1,"context_end_line":729,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport numpy as np\nfrom scipy.spatial import cKDTree as KDTree\n\nfrom dust3r.utils.misc import invalid_to_zeros, invalid_to_nans\nfrom dust3r.utils.device import to_numpy\n\n\ndef xy_grid(\n    W,\n    H,\n    device=None,\n    origin=(0, 0),\n    unsqueeze=None,\n    cat_dim=-1,\n    homogeneous=False,\n    **arange_kw,\n):\n    \"\"\"Output a (H,W,2) array of int32\n    with output[j,i,0] = i + origin[0]\n         output[j,i,1] = j + origin[1]\n    \"\"\"\n    if device is None:\n\n        arange, meshgrid, stack, ones = np.arange, np.meshgrid, np.stack, np.ones\n    else:\n\n        arange = lambda *a, **kw: torch.arange(*a, device=device, **kw)\n        meshgrid, stack = torch.meshgrid, torch.stack\n        ones = lambda *a: torch.ones(*a, device=device)\n\n    tw, th = [arange(o, o + s, **arange_kw) for s, o in zip((W, H), origin)]\n    grid = meshgrid(tw, th, indexing=\"xy\")\n    if homogeneous:\n        grid = grid + (ones((H, W)),)\n    if unsqueeze is not None:\n        grid = (grid[0].unsqueeze(unsqueeze), grid[1].unsqueeze(unsqueeze))\n    if cat_dim is not None:\n        grid = stack(grid, cat_dim)\n    return grid\n\n\ndef geotrf(Trf, pts, ncol=None, norm=False):\n    \"\"\"Apply a geometric transformation to a list of 3-D points.\n\n    H: 3x3 or 4x4 projection matrix (typically a Homography)\n    p: numpy/torch/tuple of coordinates. Shape must be (...,2) or (...,3)\n\n    ncol: int. number of columns of the result (2 or 3)\n    norm: float. if != 0, the resut is projected on the z=norm plane.\n\n    Returns an array of projected 2d points.\n    \"\"\"\n    assert Trf.ndim >= 2\n    if isinstance(Trf, np.ndarray):\n        pts = np.asarray(pts)\n    elif isinstance(Trf, torch.Tensor):\n        pts = torch.as_tensor(pts, dtype=Trf.dtype)\n\n    output_reshape = pts.shape[:-1]\n    ncol = ncol or pts.shape[-1]\n\n    if (\n        isinstance(Trf, torch.Tensor)\n        and isinstance(pts, torch.Tensor)\n        and Trf.ndim == 3\n        and pts.ndim == 4\n    ):\n        d = pts.shape[3]\n        if Trf.shape[-1] == d:\n            pts = torch.einsum(\"bij, bhwj -> bhwi\", Trf, pts)\n        elif Trf.shape[-1] == d + 1:\n            pts = (\n                torch.einsum(\"bij, bhwj -> bhwi\", Trf[:, :d, :d], pts)\n                + Trf[:, None, None, :d, d]\n            )\n        else:\n            raise ValueError(f\"bad shape, not ending with 3 or 4, for {pts.shape=}\")\n    else:\n        if Trf.ndim >= 3:\n            n = Trf.ndim - 2\n            assert Trf.shape[:n] == pts.shape[:n], \"batch size does not match\"\n            Trf = Trf.reshape(-1, Trf.shape[-2], Trf.shape[-1])\n\n            if pts.ndim > Trf.ndim:\n\n                pts = pts.reshape(Trf.shape[0], -1, pts.shape[-1])\n            elif pts.ndim == 2:\n\n                pts = pts[:, None, :]\n\n        if pts.shape[-1] + 1 == Trf.shape[-1]:\n            Trf = Trf.swapaxes(-1, -2)  # transpose Trf\n            pts = pts @ Trf[..., :-1, :] + Trf[..., -1:, :]\n        elif pts.shape[-1] == Trf.shape[-1]:\n            Trf = Trf.swapaxes(-1, -2)  # transpose Trf\n            pts = pts @ Trf\n        else:\n            pts = Trf @ pts.T\n            if pts.ndim >= 2:\n                pts = pts.swapaxes(-1, -2)\n\n    if norm:\n        pts = pts / pts[..., -1:]  # DONT DO /= BECAUSE OF WEIRD PYTORCH BUG\n        if norm != 1:\n            pts *= norm\n\n    res = pts[..., :ncol].reshape(*output_reshape, ncol)\n    return res\n\n\ndef inv(mat):\n    \"\"\"Invert a torch or numpy matrix\"\"\"\n    if isinstance(mat, torch.Tensor):\n        return torch.linalg.inv(mat)\n    if isinstance(mat, np.ndarray):\n        return np.linalg.inv(mat)\n    raise ValueError(f\"bad matrix type = {type(mat)}\")\n\n\ndef depthmap_to_pts3d(depth, pseudo_focal, pp=None, **_):\n    \"\"\"\n    Args:\n        - depthmap (BxHxW array):\n        - pseudo_focal: [B,H,W] ; [B,2,H,W] or [B,1,H,W]\n    Returns:\n        pointmap of absolute coordinates (BxHxWx3 array)\n    \"\"\"\n\n    if len(depth.shape) == 4:\n        B, H, W, n = depth.shape\n    else:\n        B, H, W = depth.shape\n        n = None\n\n    if len(pseudo_focal.shape) == 3:  # [B,H,W]\n        pseudo_focalx = pseudo_focaly = pseudo_focal\n    elif len(pseudo_focal.shape) == 4:  # [B,2,H,W] or [B,1,H,W]\n        pseudo_focalx = pseudo_focal[:, 0]\n        if pseudo_focal.shape[1] == 2:\n            pseudo_focaly = pseudo_focal[:, 1]\n        else:\n            pseudo_focaly = pseudo_focalx\n    else:\n        raise NotImplementedError(\"Error, unknown input focal shape format.\")\n\n    assert pseudo_focalx.shape == depth.shape[:3]\n    assert pseudo_focaly.shape == depth.shape[:3]\n    grid_x, grid_y = xy_grid(W, H, cat_dim=0, device=depth.device)[:, None]\n\n    if pp is None:\n        grid_x = grid_x - (W - 1) / 2\n        grid_y = grid_y - (H - 1) / 2\n    else:\n        grid_x = grid_x.expand(B, -1, -1) - pp[:, 0, None, None]\n        grid_y = grid_y.expand(B, -1, -1) - pp[:, 1, None, None]\n\n    if n is None:\n        pts3d = torch.empty((B, H, W, 3), device=depth.device)\n        pts3d[..., 0] = depth * grid_x / pseudo_focalx\n        pts3d[..., 1] = depth * grid_y / pseudo_focaly\n        pts3d[..., 2] = depth\n    else:\n        pts3d = torch.empty((B, H, W, 3, n), device=depth.device)\n        pts3d[..., 0, :] = depth * (grid_x / pseudo_focalx)[..., None]\n        pts3d[..., 1, :] = depth * (grid_y / pseudo_focaly)[..., None]\n        pts3d[..., 2, :] = depth\n    return pts3d\n\n\ndef depthmap_to_camera_coordinates(depthmap, camera_intrinsics, pseudo_focal=None):\n    \"\"\"\n    Args:\n        - depthmap (HxW array):\n        - camera_intrinsics: a 3x3 matrix\n    Returns:\n        pointmap of absolute coordinates (HxWx3 array), and a mask specifying valid pixels.\n    \"\"\"\n    camera_intrinsics = np.float32(camera_intrinsics)\n    H, W = depthmap.shape\n\n    assert camera_intrinsics[0, 1] == 0.0\n    assert camera_intrinsics[1, 0] == 0.0\n    if pseudo_focal is None:\n        fu = camera_intrinsics[0, 0]\n        fv = camera_intrinsics[1, 1]\n    else:\n        assert pseudo_focal.shape == (H, W)\n        fu = fv = pseudo_focal\n    cu = camera_intrinsics[0, 2]\n    cv = camera_intrinsics[1, 2]\n\n    u, v = np.meshgrid(np.arange(W), np.arange(H))\n    z_cam = depthmap\n    x_cam = (u - cu) * z_cam / fu\n    y_cam = (v - cv) * z_cam / fv\n    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1).astype(np.float32)\n\n    valid_mask = depthmap > 0.0\n    return X_cam, valid_mask\n\n\ndef depthmap_to_absolute_camera_coordinates(\n    depthmap, camera_intrinsics, camera_pose, **kw\n):\n    \"\"\"\n    Args:\n        - depthmap (HxW array):\n        - camera_intrinsics: a 3x3 matrix\n        - camera_pose: a 4x3 or 4x4 cam2world matrix\n    Returns:\n        pointmap of absolute coordinates (HxWx3 array), and a mask specifying valid pixels.\n    \"\"\"\n    X_cam, valid_mask = depthmap_to_camera_coordinates(depthmap, camera_intrinsics)\n\n    X_world = X_cam  # default\n    if camera_pose is not None:\n\n        R_cam2world = camera_pose[:3, :3]\n        t_cam2world = camera_pose[:3, 3]\n\n        X_world = (\n            np.einsum(\"ik, vuk -> vui\", R_cam2world, X_cam) + t_cam2world[None, None, :]\n        )\n\n    return X_world, valid_mask\n\n\ndef colmap_to_opencv_intrinsics(K):\n    \"\"\"\n    Modify camera intrinsics to follow a different convention.\n    Coordinates of the center of the top-left pixels are by default:\n    - (0.5, 0.5) in Colmap\n    - (0,0) in OpenCV\n    \"\"\"\n    K = K.copy()\n    K[0, 2] -= 0.5\n    K[1, 2] -= 0.5\n    return K\n\n\ndef opencv_to_colmap_intrinsics(K):\n    \"\"\"\n    Modify camera intrinsics to follow a different convention.\n    Coordinates of the center of the top-left pixels are by default:\n    - (0.5, 0.5) in Colmap\n    - (0,0) in OpenCV\n    \"\"\"\n    K = K.copy()\n    K[0, 2] += 0.5\n    K[1, 2] += 0.5\n    return K\n\n\ndef normalize_pointcloud(\n    pts1, pts2, norm_mode=\"avg_dis\", valid1=None, valid2=None, ret_factor=False\n):\n    \"\"\"renorm pointmaps pts1, pts2 with norm_mode\"\"\"\n    assert pts1.ndim >= 3 and pts1.shape[-1] == 3\n    assert pts2 is None or (pts2.ndim >= 3 and pts2.shape[-1] == 3)\n    norm_mode, dis_mode = norm_mode.split(\"_\")\n\n    if norm_mode == \"avg\":\n\n        nan_pts1, nnz1 = invalid_to_zeros(pts1, valid1, ndim=3)\n        nan_pts2, nnz2 = (\n            invalid_to_zeros(pts2, valid2, ndim=3) if pts2 is not None else (None, 0)\n        )\n        all_pts = (\n            torch.cat((nan_pts1, nan_pts2), dim=1) if pts2 is not None else nan_pts1\n        )\n\n        all_dis = all_pts.norm(dim=-1)\n        if dis_mode == \"dis\":\n            pass  # do nothing\n        elif dis_mode == \"log1p\":\n            all_dis = torch.log1p(all_dis)\n        elif dis_mode == \"warp-log1p\":\n\n            log_dis = torch.log1p(all_dis)\n            warp_factor = log_dis / all_dis.clip(min=1e-8)\n            H1, W1 = pts1.shape[1:-1]\n            pts1 = pts1 * warp_factor[:, : W1 * H1].view(-1, H1, W1, 1)\n            if pts2 is not None:\n                H2, W2 = pts2.shape[1:-1]\n                pts2 = pts2 * warp_factor[:, W1 * H1 :].view(-1, H2, W2, 1)\n            all_dis = log_dis  # this is their true distance afterwards\n        else:\n            raise ValueError(f\"bad {dis_mode=}\")\n\n        norm_factor = all_dis.sum(dim=1) / (nnz1 + nnz2 + 1e-8)\n    else:\n\n        nan_pts1 = invalid_to_nans(pts1, valid1, ndim=3)\n        nan_pts2 = invalid_to_nans(pts2, valid2, ndim=3) if pts2 is not None else None\n        all_pts = (\n            torch.cat((nan_pts1, nan_pts2), dim=1) if pts2 is not None else nan_pts1\n        )\n\n        all_dis = all_pts.norm(dim=-1)\n\n        if norm_mode == \"avg\":\n            norm_factor = all_dis.nanmean(dim=1)\n        elif norm_mode == \"median\":\n            norm_factor = all_dis.nanmedian(dim=1).values.detach()\n        elif norm_mode == \"sqrt\":\n            norm_factor = all_dis.sqrt().nanmean(dim=1) ** 2\n        else:\n            raise ValueError(f\"bad {norm_mode=}\")\n\n    norm_factor = norm_factor.clip(min=1e-8)\n    while norm_factor.ndim < pts1.ndim:\n        norm_factor.unsqueeze_(-1)\n\n    res = pts1 / norm_factor\n    if pts2 is not None:\n        res = (res, pts2 / norm_factor)\n    if ret_factor:\n        res = res + (norm_factor,)\n    return res\n\n\ndef normalize_pointcloud_group(\n    pts_list,\n    norm_mode=\"avg_dis\",\n    valid_list=None,\n    conf_list=None,\n    ret_factor=False,\n    ret_factor_only=False,\n):\n    \"\"\"renorm pointmaps pts1, pts2 with norm_mode\"\"\"\n    for pts in pts_list:\n        assert pts.ndim >= 3 and pts.shape[-1] == 3\n\n    norm_mode, dis_mode = norm_mode.split(\"_\")\n\n    if norm_mode == \"avg\":\n\n        nan_pts_list, nnz_list = zip(\n            *[\n                invalid_to_zeros(pts1, valid1, ndim=3)\n                for pts1, valid1 in zip(pts_list, valid_list)\n            ]\n        )\n        all_pts = torch.cat(nan_pts_list, dim=1)\n        if conf_list is not None:\n            nan_conf_list = [\n                invalid_to_zeros(conf1[..., None], valid1, ndim=3)[0]\n                for conf1, valid1 in zip(conf_list, valid_list)\n            ]\n            all_conf = torch.cat(nan_conf_list, dim=1)[..., 0]\n        else:\n            all_conf = torch.ones_like(all_pts[..., 0])\n\n        all_dis = all_pts.norm(dim=-1)\n        if dis_mode == \"dis\":\n            pass  # do nothing\n        elif dis_mode == \"log1p\":\n            all_dis = torch.log1p(all_dis)\n        elif dis_mode == \"warp-log1p\":\n\n            log_dis = torch.log1p(all_dis)\n            warp_factor = log_dis / all_dis.clip(min=1e-8)\n            H_W_list = [pts.shape[1:-1] for pts in pts_list]\n            pts_list = [\n                pts\n                * warp_factor[:, sum(H_W_list[:i]) : sum(H_W_list[: i + 1])].view(\n                    -1, H, W, 1\n                )\n                for i, (pts, (H, W)) in enumerate(zip(pts_list, H_W_list))\n            ]\n            all_dis = log_dis  # this is their true distance afterwards\n        else:\n            raise ValueError(f\"bad {dis_mode=}\")\n\n        norm_factor = (all_conf * all_dis).sum(dim=1) / (all_conf.sum(dim=1) + 1e-8)\n    else:\n\n        nan_pts_list = [\n            invalid_to_nans(pts1, valid1, ndim=3)\n            for pts1, valid1 in zip(pts_list, valid_list)\n        ]\n\n        all_pts = torch.cat(nan_pts_list, dim=1)\n\n        all_dis = all_pts.norm(dim=-1)\n\n        if norm_mode == \"avg\":\n            norm_factor = all_dis.nanmean(dim=1)\n        elif norm_mode == \"median\":\n            norm_factor = all_dis.nanmedian(dim=1).values.detach()\n        elif norm_mode == \"sqrt\":\n            norm_factor = all_dis.sqrt().nanmean(dim=1) ** 2\n        else:\n            raise ValueError(f\"bad {norm_mode=}\")\n\n    norm_factor = norm_factor.clip(min=1e-8)\n    while norm_factor.ndim < pts_list[0].ndim:\n        norm_factor.unsqueeze_(-1)\n\n    if ret_factor_only:\n\n        return norm_factor\n\n    res = [pts / norm_factor for pts in pts_list]\n    if ret_factor:\n        return res, norm_factor\n    return res\n\n\n@torch.no_grad()\ndef get_joint_pointcloud_depth(z1, z2, valid_mask1, valid_mask2=None, quantile=0.5):\n\n    _z1 = invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1)\n    _z2 = (\n        invalid_to_nans(z2, valid_mask2).reshape(len(z2), -1)\n        if z2 is not None\n        else None\n    )\n    _z = torch.cat((_z1, _z2), dim=-1) if z2 is not None else _z1\n\n    if quantile == 0.5:\n        shift_z = torch.nanmedian(_z, dim=-1).values\n    else:\n        shift_z = torch.nanquantile(_z, quantile, dim=-1)\n    return shift_z  # (B,)\n\n\n@torch.no_grad()\ndef get_group_pointcloud_depth(zs, valid_masks, quantile=0.5):\n\n    _zs = [\n        invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1)\n        for z1, valid_mask1 in zip(zs, valid_masks)\n    ]\n    _z = torch.cat(_zs, dim=-1)\n\n    if quantile == 0.5:\n        shift_z = torch.nanmedian(_z, dim=-1).values\n    else:\n        shift_z = torch.nanquantile(_z, quantile, dim=-1)\n    return shift_z  # (B,)\n\n\n@torch.no_grad()\ndef get_joint_pointcloud_center_scale(\n    pts1, pts2, valid_mask1=None, valid_mask2=None, z_only=False, center=True\n):\n\n    _pts1 = invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3)\n    _pts2 = (\n        invalid_to_nans(pts2, valid_mask2).reshape(len(pts2), -1, 3)\n        if pts2 is not None\n        else None\n    )\n    _pts = torch.cat((_pts1, _pts2), dim=1) if pts2 is not None else _pts1\n\n    _center = torch.nanmedian(_pts, dim=1, keepdim=True).values  # (B,1,3)\n    if z_only:\n        _center[..., :2] = 0  # do not center X and Y\n\n    _norm = ((_pts - _center) if center else _pts).norm(dim=-1)\n    scale = torch.nanmedian(_norm, dim=1).values\n    return _center[:, None, :, :], scale[:, None, None, None]\n\n\n@torch.no_grad()\ndef get_group_pointcloud_center_scale(pts, valid_masks=None, z_only=False, center=True):\n\n    _pts = [\n        invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3)\n        for pts1, valid_mask1 in zip(pts, valid_masks)\n    ]\n    _pts = torch.cat(_pts, dim=1)\n\n    _center = torch.nanmedian(_pts, dim=1, keepdim=True).values  # (B,1,3)\n    if z_only:\n        _center[..., :2] = 0  # do not center X and Y\n\n    _norm = ((_pts - _center) if center else _pts).norm(dim=-1)\n    scale = torch.nanmedian(_norm, dim=1).values\n    return _center[:, None, :, :], scale[:, None, None, None]\n\n\ndef find_reciprocal_matches(P1, P2):\n    \"\"\"\n    returns 3 values:\n    1 - reciprocal_in_P2: a boolean array of size P2.shape[0], a \"True\" value indicates a match\n    2 - nn2_in_P1: a int array of size P2.shape[0], it contains the indexes of the closest points in P1\n    3 - reciprocal_in_P2.sum(): the number of matches\n    \"\"\"\n    tree1 = KDTree(P1)\n    tree2 = KDTree(P2)\n\n    _, nn1_in_P2 = tree2.query(P1, workers=8)\n    _, nn2_in_P1 = tree1.query(P2, workers=8)\n\n    reciprocal_in_P1 = nn2_in_P1[nn1_in_P2] == np.arange(len(nn1_in_P2))\n    reciprocal_in_P2 = nn1_in_P2[nn2_in_P1] == np.arange(len(nn2_in_P1))\n    assert reciprocal_in_P1.sum() == reciprocal_in_P2.sum()\n    return reciprocal_in_P2, nn2_in_P1, reciprocal_in_P2.sum()\n\n\ndef get_med_dist_between_poses(poses):\n    from scipy.spatial.distance import pdist\n\n    return np.median(pdist([to_numpy(p[:3, 3]) for p in poses]))\n\n\ndef weighted_procrustes(A, B, w, use_weights=True, eps=1e-16, return_T=False):\n    \"\"\"\n    X: torch tensor B x N x 3\n    Y: torch tensor B x N x 3\n    w: torch tensor B x N\n    \"\"\"\n    assert len(A) == len(B)\n    if use_weights:\n        W1 = torch.abs(w).sum(1, keepdim=True)\n        w_norm = (w / (W1 + eps)).unsqueeze(-1)\n        a_mean = (w_norm * A).sum(dim=1, keepdim=True)\n        b_mean = (w_norm * B).sum(dim=1, keepdim=True)\n\n        A_c = A - a_mean\n        B_c = B - b_mean\n\n        H = torch.einsum(\"bni,bnj->bij\", A_c, w_norm * B_c)\n\n    else:\n        a_mean = A.mean(axis=1, keepdim=True)\n        b_mean = B.mean(axis=1, keepdim=True)\n\n        A_c = A - a_mean\n        B_c = B - b_mean\n\n        H = torch.einsum(\"bij,bik->bjk\", A_c, B_c)\n\n    U, S, V = torch.svd(H)  # U: B x 3 x 3, S: B x 3, V: B x 3 x 3\n    Z = torch.eye(3).unsqueeze(0).repeat(A.shape[0], 1, 1).to(A.device)\n    Z[:, -1, -1] = torch.sign(torch.linalg.det(U @ V.transpose(1, 2)))  # B x 3 x 3\n    R = V @ Z @ U.transpose(1, 2)  # B x 3 x 3\n    t = b_mean - torch.einsum(\"bij,bjk->bik\", R, a_mean.transpose(-2, -1)).transpose(\n        -2, -1\n    )\n    if return_T:\n        T = torch.eye(4).unsqueeze(0).repeat(A.shape[0], 1, 1).to(A.device)\n        T[:, :3, :3] = R\n        T[:, :3, 3] = t.squeeze()\n        return T\n    return R, t.squeeze()\n\n\ndef perspective_projection(x, K):\n    \"\"\"\n    This function computes the perspective projection of a set of points assuming the extrinsinc params have already been applied\n    Args:\n        - x [bs,N,3]: 3D points\n        - K [bs,3,3]: Camera instrincs params\n    \"\"\"\n    # Apply perspective distortion\n    y = x / x[:, :, -1].unsqueeze(-1)  # (bs, N, 3)\n\n    # Apply camera intrinsics\n    y = torch.einsum('bij,bkj->bki', K, y)  # (bs, N, 3)\n\n    return y[:, :, :2]\n\n\ndef inverse_perspective_projection(points, K, distance):\n    \"\"\"\n    This function computes the inverse perspective projection of a set of points given an estimated distance.\n    Input:\n        points (bs, N, 2): 2D points\n        K (bs,3,3): camera intrinsics params\n        distance (bs, N, 1): distance in the 3D world\n    Similar to:\n        - pts_l_norm = cv2.undistortPoints(np.expand_dims(pts_l, axis=1), cameraMatrix=K_l, distCoeffs=None)\n    \"\"\"\n    # Apply camera intrinsics\n    points = torch.cat([points, torch.ones_like(points[..., :1])], -1)\n    points = torch.einsum('bij,bkj->bki', torch.inverse(K), points)\n\n    # Apply perspective distortion\n    if distance == None:\n        return points\n    points = points * distance\n    return points\n\ndef get_focalLength_from_fieldOfView(fov=60, img_size=512):\n    \"\"\"\n    Compute the focal length of the camera lens by assuming a certain FOV for the entire image\n    Args:\n        - fov: float, expressed in degree\n        - img_size: int\n    Return:\n        focal: float\n    \"\"\"\n    focal = img_size / (2 * np.tan(np.radians(fov) /2))\n    return focal\n\ndef focal_length_normalization(x, f, fovn=60, img_size=448):\n    \"\"\"\n    Section 3.1 of https://arxiv.org/pdf/1904.02028.pdf\n    E = (fn/f) * E' where E is 1/d\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    y = x * (fn/f)\n    return y\n\ndef undo_focal_length_normalization(y, f, fovn=60, img_size=448):\n    \"\"\"\n    Undo focal_length_normalization()\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    x = y * (f/fn)\n    return x\n\nEPS_LOG = 1e-10\ndef log_depth(x, eps=EPS_LOG):\n    \"\"\"\n    Move depth to log space\n    \"\"\"\n    return torch.log(x + eps)\n\ndef undo_log_depth(y, eps=EPS_LOG):\n    \"\"\"\n    Undo log_depth()\n    \"\"\"\n    return torch.exp(y) - eps\n\ndef get_camera_parameters(img_size, fov=60, p_x=None, p_y=None, device=torch.device('cuda')):\n    \"\"\" Given image size, fov and principal point coordinates, return K the camera parameter matrix\"\"\"\n    K = torch.\n# ... truncated ...","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.xy_grid","uri":"program://Human3R/function/src.dust3r.utils.geometry.xy_grid#L15-L46","kind":"function","name":"xy_grid","path":"src/dust3r/utils/geometry.py","language":"python","start_line":15,"end_line":46,"context_start_line":1,"context_end_line":66,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport numpy as np\nfrom scipy.spatial import cKDTree as KDTree\n\nfrom dust3r.utils.misc import invalid_to_zeros, invalid_to_nans\nfrom dust3r.utils.device import to_numpy\n\n\ndef xy_grid(\n    W,\n    H,\n    device=None,\n    origin=(0, 0),\n    unsqueeze=None,\n    cat_dim=-1,\n    homogeneous=False,\n    **arange_kw,\n):\n    \"\"\"Output a (H,W,2) array of int32\n    with output[j,i,0] = i + origin[0]\n         output[j,i,1] = j + origin[1]\n    \"\"\"\n    if device is None:\n\n        arange, meshgrid, stack, ones = np.arange, np.meshgrid, np.stack, np.ones\n    else:\n\n        arange = lambda *a, **kw: torch.arange(*a, device=device, **kw)\n        meshgrid, stack = torch.meshgrid, torch.stack\n        ones = lambda *a: torch.ones(*a, device=device)\n\n    tw, th = [arange(o, o + s, **arange_kw) for s, o in zip((W, H), origin)]\n    grid = meshgrid(tw, th, indexing=\"xy\")\n    if homogeneous:\n        grid = grid + (ones((H, W)),)\n    if unsqueeze is not None:\n        grid = (grid[0].unsqueeze(unsqueeze), grid[1].unsqueeze(unsqueeze))\n    if cat_dim is not None:\n        grid = stack(grid, cat_dim)\n    return grid\n\n\ndef geotrf(Trf, pts, ncol=None, norm=False):\n    \"\"\"Apply a geometric transformation to a list of 3-D points.\n\n    H: 3x3 or 4x4 projection matrix (typically a Homography)\n    p: numpy/torch/tuple of coordinates. Shape must be (...,2) or (...,3)\n\n    ncol: int. number of columns of the result (2 or 3)\n    norm: float. if != 0, the resut is projected on the z=norm plane.\n\n    Returns an array of projected 2d points.\n    \"\"\"\n    assert Trf.ndim >= 2\n    if isinstance(Trf, np.ndarray):\n        pts = np.asarray(pts)\n    elif isinstance(Trf, torch.Tensor):\n        pts = torch.as_tensor(pts, dtype=Trf.dtype)\n\n    output_reshape = pts.shape[:-1]","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.geotrf","uri":"program://Human3R/function/src.dust3r.utils.geometry.geotrf#L49-L115","kind":"function","name":"geotrf","path":"src/dust3r/utils/geometry.py","language":"python","start_line":49,"end_line":115,"context_start_line":29,"context_end_line":135,"code":"    if device is None:\n\n        arange, meshgrid, stack, ones = np.arange, np.meshgrid, np.stack, np.ones\n    else:\n\n        arange = lambda *a, **kw: torch.arange(*a, device=device, **kw)\n        meshgrid, stack = torch.meshgrid, torch.stack\n        ones = lambda *a: torch.ones(*a, device=device)\n\n    tw, th = [arange(o, o + s, **arange_kw) for s, o in zip((W, H), origin)]\n    grid = meshgrid(tw, th, indexing=\"xy\")\n    if homogeneous:\n        grid = grid + (ones((H, W)),)\n    if unsqueeze is not None:\n        grid = (grid[0].unsqueeze(unsqueeze), grid[1].unsqueeze(unsqueeze))\n    if cat_dim is not None:\n        grid = stack(grid, cat_dim)\n    return grid\n\n\ndef geotrf(Trf, pts, ncol=None, norm=False):\n    \"\"\"Apply a geometric transformation to a list of 3-D points.\n\n    H: 3x3 or 4x4 projection matrix (typically a Homography)\n    p: numpy/torch/tuple of coordinates. Shape must be (...,2) or (...,3)\n\n    ncol: int. number of columns of the result (2 or 3)\n    norm: float. if != 0, the resut is projected on the z=norm plane.\n\n    Returns an array of projected 2d points.\n    \"\"\"\n    assert Trf.ndim >= 2\n    if isinstance(Trf, np.ndarray):\n        pts = np.asarray(pts)\n    elif isinstance(Trf, torch.Tensor):\n        pts = torch.as_tensor(pts, dtype=Trf.dtype)\n\n    output_reshape = pts.shape[:-1]\n    ncol = ncol or pts.shape[-1]\n\n    if (\n        isinstance(Trf, torch.Tensor)\n        and isinstance(pts, torch.Tensor)\n        and Trf.ndim == 3\n        and pts.ndim == 4\n    ):\n        d = pts.shape[3]\n        if Trf.shape[-1] == d:\n            pts = torch.einsum(\"bij, bhwj -> bhwi\", Trf, pts)\n        elif Trf.shape[-1] == d + 1:\n            pts = (\n                torch.einsum(\"bij, bhwj -> bhwi\", Trf[:, :d, :d], pts)\n                + Trf[:, None, None, :d, d]\n            )\n        else:\n            raise ValueError(f\"bad shape, not ending with 3 or 4, for {pts.shape=}\")\n    else:\n        if Trf.ndim >= 3:\n            n = Trf.ndim - 2\n            assert Trf.shape[:n] == pts.shape[:n], \"batch size does not match\"\n            Trf = Trf.reshape(-1, Trf.shape[-2], Trf.shape[-1])\n\n            if pts.ndim > Trf.ndim:\n\n                pts = pts.reshape(Trf.shape[0], -1, pts.shape[-1])\n            elif pts.ndim == 2:\n\n                pts = pts[:, None, :]\n\n        if pts.shape[-1] + 1 == Trf.shape[-1]:\n            Trf = Trf.swapaxes(-1, -2)  # transpose Trf\n            pts = pts @ Trf[..., :-1, :] + Trf[..., -1:, :]\n        elif pts.shape[-1] == Trf.shape[-1]:\n            Trf = Trf.swapaxes(-1, -2)  # transpose Trf\n            pts = pts @ Trf\n        else:\n            pts = Trf @ pts.T\n            if pts.ndim >= 2:\n                pts = pts.swapaxes(-1, -2)\n\n    if norm:\n        pts = pts / pts[..., -1:]  # DONT DO /= BECAUSE OF WEIRD PYTORCH BUG\n        if norm != 1:\n            pts *= norm\n\n    res = pts[..., :ncol].reshape(*output_reshape, ncol)\n    return res\n\n\ndef inv(mat):\n    \"\"\"Invert a torch or numpy matrix\"\"\"\n    if isinstance(mat, torch.Tensor):\n        return torch.linalg.inv(mat)\n    if isinstance(mat, np.ndarray):\n        return np.linalg.inv(mat)\n    raise ValueError(f\"bad matrix type = {type(mat)}\")\n\n\ndef depthmap_to_pts3d(depth, pseudo_focal, pp=None, **_):\n    \"\"\"\n    Args:\n        - depthmap (BxHxW array):\n        - pseudo_focal: [B,H,W] ; [B,2,H,W] or [B,1,H,W]\n    Returns:\n        pointmap of absolute coordinates (BxHxWx3 array)\n    \"\"\"\n","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.inv","uri":"program://Human3R/function/src.dust3r.utils.geometry.inv#L118-L124","kind":"function","name":"inv","path":"src/dust3r/utils/geometry.py","language":"python","start_line":118,"end_line":124,"context_start_line":98,"context_end_line":144,"code":"        if pts.shape[-1] + 1 == Trf.shape[-1]:\n            Trf = Trf.swapaxes(-1, -2)  # transpose Trf\n            pts = pts @ Trf[..., :-1, :] + Trf[..., -1:, :]\n        elif pts.shape[-1] == Trf.shape[-1]:\n            Trf = Trf.swapaxes(-1, -2)  # transpose Trf\n            pts = pts @ Trf\n        else:\n            pts = Trf @ pts.T\n            if pts.ndim >= 2:\n                pts = pts.swapaxes(-1, -2)\n\n    if norm:\n        pts = pts / pts[..., -1:]  # DONT DO /= BECAUSE OF WEIRD PYTORCH BUG\n        if norm != 1:\n            pts *= norm\n\n    res = pts[..., :ncol].reshape(*output_reshape, ncol)\n    return res\n\n\ndef inv(mat):\n    \"\"\"Invert a torch or numpy matrix\"\"\"\n    if isinstance(mat, torch.Tensor):\n        return torch.linalg.inv(mat)\n    if isinstance(mat, np.ndarray):\n        return np.linalg.inv(mat)\n    raise ValueError(f\"bad matrix type = {type(mat)}\")\n\n\ndef depthmap_to_pts3d(depth, pseudo_focal, pp=None, **_):\n    \"\"\"\n    Args:\n        - depthmap (BxHxW array):\n        - pseudo_focal: [B,H,W] ; [B,2,H,W] or [B,1,H,W]\n    Returns:\n        pointmap of absolute coordinates (BxHxWx3 array)\n    \"\"\"\n\n    if len(depth.shape) == 4:\n        B, H, W, n = depth.shape\n    else:\n        B, H, W = depth.shape\n        n = None\n\n    if len(pseudo_focal.shape) == 3:  # [B,H,W]\n        pseudo_focalx = pseudo_focaly = pseudo_focal\n    elif len(pseudo_focal.shape) == 4:  # [B,2,H,W] or [B,1,H,W]","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.depthmap_to_pts3d","uri":"program://Human3R/function/src.dust3r.utils.geometry.depthmap_to_pts3d#L127-L174","kind":"function","name":"depthmap_to_pts3d","path":"src/dust3r/utils/geometry.py","language":"python","start_line":127,"end_line":174,"context_start_line":107,"context_end_line":194,"code":"                pts = pts.swapaxes(-1, -2)\n\n    if norm:\n        pts = pts / pts[..., -1:]  # DONT DO /= BECAUSE OF WEIRD PYTORCH BUG\n        if norm != 1:\n            pts *= norm\n\n    res = pts[..., :ncol].reshape(*output_reshape, ncol)\n    return res\n\n\ndef inv(mat):\n    \"\"\"Invert a torch or numpy matrix\"\"\"\n    if isinstance(mat, torch.Tensor):\n        return torch.linalg.inv(mat)\n    if isinstance(mat, np.ndarray):\n        return np.linalg.inv(mat)\n    raise ValueError(f\"bad matrix type = {type(mat)}\")\n\n\ndef depthmap_to_pts3d(depth, pseudo_focal, pp=None, **_):\n    \"\"\"\n    Args:\n        - depthmap (BxHxW array):\n        - pseudo_focal: [B,H,W] ; [B,2,H,W] or [B,1,H,W]\n    Returns:\n        pointmap of absolute coordinates (BxHxWx3 array)\n    \"\"\"\n\n    if len(depth.shape) == 4:\n        B, H, W, n = depth.shape\n    else:\n        B, H, W = depth.shape\n        n = None\n\n    if len(pseudo_focal.shape) == 3:  # [B,H,W]\n        pseudo_focalx = pseudo_focaly = pseudo_focal\n    elif len(pseudo_focal.shape) == 4:  # [B,2,H,W] or [B,1,H,W]\n        pseudo_focalx = pseudo_focal[:, 0]\n        if pseudo_focal.shape[1] == 2:\n            pseudo_focaly = pseudo_focal[:, 1]\n        else:\n            pseudo_focaly = pseudo_focalx\n    else:\n        raise NotImplementedError(\"Error, unknown input focal shape format.\")\n\n    assert pseudo_focalx.shape == depth.shape[:3]\n    assert pseudo_focaly.shape == depth.shape[:3]\n    grid_x, grid_y = xy_grid(W, H, cat_dim=0, device=depth.device)[:, None]\n\n    if pp is None:\n        grid_x = grid_x - (W - 1) / 2\n        grid_y = grid_y - (H - 1) / 2\n    else:\n        grid_x = grid_x.expand(B, -1, -1) - pp[:, 0, None, None]\n        grid_y = grid_y.expand(B, -1, -1) - pp[:, 1, None, None]\n\n    if n is None:\n        pts3d = torch.empty((B, H, W, 3), device=depth.device)\n        pts3d[..., 0] = depth * grid_x / pseudo_focalx\n        pts3d[..., 1] = depth * grid_y / pseudo_focaly\n        pts3d[..., 2] = depth\n    else:\n        pts3d = torch.empty((B, H, W, 3, n), device=depth.device)\n        pts3d[..., 0, :] = depth * (grid_x / pseudo_focalx)[..., None]\n        pts3d[..., 1, :] = depth * (grid_y / pseudo_focaly)[..., None]\n        pts3d[..., 2, :] = depth\n    return pts3d\n\n\ndef depthmap_to_camera_coordinates(depthmap, camera_intrinsics, pseudo_focal=None):\n    \"\"\"\n    Args:\n        - depthmap (HxW array):\n        - camera_intrinsics: a 3x3 matrix\n    Returns:\n        pointmap of absolute coordinates (HxWx3 array), and a mask specifying valid pixels.\n    \"\"\"\n    camera_intrinsics = np.float32(camera_intrinsics)\n    H, W = depthmap.shape\n\n    assert camera_intrinsics[0, 1] == 0.0\n    assert camera_intrinsics[1, 0] == 0.0\n    if pseudo_focal is None:\n        fu = camera_intrinsics[0, 0]\n        fv = camera_intrinsics[1, 1]\n    else:\n        assert pseudo_focal.shape == (H, W)","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.depthmap_to_camera_coordinates","uri":"program://Human3R/function/src.dust3r.utils.geometry.depthmap_to_camera_coordinates#L177-L206","kind":"function","name":"depthmap_to_camera_coordinates","path":"src/dust3r/utils/geometry.py","language":"python","start_line":177,"end_line":206,"context_start_line":157,"context_end_line":226,"code":"    if pp is None:\n        grid_x = grid_x - (W - 1) / 2\n        grid_y = grid_y - (H - 1) / 2\n    else:\n        grid_x = grid_x.expand(B, -1, -1) - pp[:, 0, None, None]\n        grid_y = grid_y.expand(B, -1, -1) - pp[:, 1, None, None]\n\n    if n is None:\n        pts3d = torch.empty((B, H, W, 3), device=depth.device)\n        pts3d[..., 0] = depth * grid_x / pseudo_focalx\n        pts3d[..., 1] = depth * grid_y / pseudo_focaly\n        pts3d[..., 2] = depth\n    else:\n        pts3d = torch.empty((B, H, W, 3, n), device=depth.device)\n        pts3d[..., 0, :] = depth * (grid_x / pseudo_focalx)[..., None]\n        pts3d[..., 1, :] = depth * (grid_y / pseudo_focaly)[..., None]\n        pts3d[..., 2, :] = depth\n    return pts3d\n\n\ndef depthmap_to_camera_coordinates(depthmap, camera_intrinsics, pseudo_focal=None):\n    \"\"\"\n    Args:\n        - depthmap (HxW array):\n        - camera_intrinsics: a 3x3 matrix\n    Returns:\n        pointmap of absolute coordinates (HxWx3 array), and a mask specifying valid pixels.\n    \"\"\"\n    camera_intrinsics = np.float32(camera_intrinsics)\n    H, W = depthmap.shape\n\n    assert camera_intrinsics[0, 1] == 0.0\n    assert camera_intrinsics[1, 0] == 0.0\n    if pseudo_focal is None:\n        fu = camera_intrinsics[0, 0]\n        fv = camera_intrinsics[1, 1]\n    else:\n        assert pseudo_focal.shape == (H, W)\n        fu = fv = pseudo_focal\n    cu = camera_intrinsics[0, 2]\n    cv = camera_intrinsics[1, 2]\n\n    u, v = np.meshgrid(np.arange(W), np.arange(H))\n    z_cam = depthmap\n    x_cam = (u - cu) * z_cam / fu\n    y_cam = (v - cv) * z_cam / fv\n    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1).astype(np.float32)\n\n    valid_mask = depthmap > 0.0\n    return X_cam, valid_mask\n\n\ndef depthmap_to_absolute_camera_coordinates(\n    depthmap, camera_intrinsics, camera_pose, **kw\n):\n    \"\"\"\n    Args:\n        - depthmap (HxW array):\n        - camera_intrinsics: a 3x3 matrix\n        - camera_pose: a 4x3 or 4x4 cam2world matrix\n    Returns:\n        pointmap of absolute coordinates (HxWx3 array), and a mask specifying valid pixels.\n    \"\"\"\n    X_cam, valid_mask = depthmap_to_camera_coordinates(depthmap, camera_intrinsics)\n\n    X_world = X_cam  # default\n    if camera_pose is not None:\n\n        R_cam2world = camera_pose[:3, :3]\n        t_cam2world = camera_pose[:3, 3]","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.depthmap_to_absolute_camera_coordinates","uri":"program://Human3R/function/src.dust3r.utils.geometry.depthmap_to_absolute_camera_coordinates#L209-L232","kind":"function","name":"depthmap_to_absolute_camera_coordinates","path":"src/dust3r/utils/geometry.py","language":"python","start_line":209,"end_line":232,"context_start_line":189,"context_end_line":252,"code":"    assert camera_intrinsics[1, 0] == 0.0\n    if pseudo_focal is None:\n        fu = camera_intrinsics[0, 0]\n        fv = camera_intrinsics[1, 1]\n    else:\n        assert pseudo_focal.shape == (H, W)\n        fu = fv = pseudo_focal\n    cu = camera_intrinsics[0, 2]\n    cv = camera_intrinsics[1, 2]\n\n    u, v = np.meshgrid(np.arange(W), np.arange(H))\n    z_cam = depthmap\n    x_cam = (u - cu) * z_cam / fu\n    y_cam = (v - cv) * z_cam / fv\n    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1).astype(np.float32)\n\n    valid_mask = depthmap > 0.0\n    return X_cam, valid_mask\n\n\ndef depthmap_to_absolute_camera_coordinates(\n    depthmap, camera_intrinsics, camera_pose, **kw\n):\n    \"\"\"\n    Args:\n        - depthmap (HxW array):\n        - camera_intrinsics: a 3x3 matrix\n        - camera_pose: a 4x3 or 4x4 cam2world matrix\n    Returns:\n        pointmap of absolute coordinates (HxWx3 array), and a mask specifying valid pixels.\n    \"\"\"\n    X_cam, valid_mask = depthmap_to_camera_coordinates(depthmap, camera_intrinsics)\n\n    X_world = X_cam  # default\n    if camera_pose is not None:\n\n        R_cam2world = camera_pose[:3, :3]\n        t_cam2world = camera_pose[:3, 3]\n\n        X_world = (\n            np.einsum(\"ik, vuk -> vui\", R_cam2world, X_cam) + t_cam2world[None, None, :]\n        )\n\n    return X_world, valid_mask\n\n\ndef colmap_to_opencv_intrinsics(K):\n    \"\"\"\n    Modify camera intrinsics to follow a different convention.\n    Coordinates of the center of the top-left pixels are by default:\n    - (0.5, 0.5) in Colmap\n    - (0,0) in OpenCV\n    \"\"\"\n    K = K.copy()\n    K[0, 2] -= 0.5\n    K[1, 2] -= 0.5\n    return K\n\n\ndef opencv_to_colmap_intrinsics(K):\n    \"\"\"\n    Modify camera intrinsics to follow a different convention.\n    Coordinates of the center of the top-left pixels are by default:\n    - (0.5, 0.5) in Colmap","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.colmap_to_opencv_intrinsics","uri":"program://Human3R/function/src.dust3r.utils.geometry.colmap_to_opencv_intrinsics#L235-L245","kind":"function","name":"colmap_to_opencv_intrinsics","path":"src/dust3r/utils/geometry.py","language":"python","start_line":235,"end_line":245,"context_start_line":215,"context_end_line":265,"code":"        - camera_intrinsics: a 3x3 matrix\n        - camera_pose: a 4x3 or 4x4 cam2world matrix\n    Returns:\n        pointmap of absolute coordinates (HxWx3 array), and a mask specifying valid pixels.\n    \"\"\"\n    X_cam, valid_mask = depthmap_to_camera_coordinates(depthmap, camera_intrinsics)\n\n    X_world = X_cam  # default\n    if camera_pose is not None:\n\n        R_cam2world = camera_pose[:3, :3]\n        t_cam2world = camera_pose[:3, 3]\n\n        X_world = (\n            np.einsum(\"ik, vuk -> vui\", R_cam2world, X_cam) + t_cam2world[None, None, :]\n        )\n\n    return X_world, valid_mask\n\n\ndef colmap_to_opencv_intrinsics(K):\n    \"\"\"\n    Modify camera intrinsics to follow a different convention.\n    Coordinates of the center of the top-left pixels are by default:\n    - (0.5, 0.5) in Colmap\n    - (0,0) in OpenCV\n    \"\"\"\n    K = K.copy()\n    K[0, 2] -= 0.5\n    K[1, 2] -= 0.5\n    return K\n\n\ndef opencv_to_colmap_intrinsics(K):\n    \"\"\"\n    Modify camera intrinsics to follow a different convention.\n    Coordinates of the center of the top-left pixels are by default:\n    - (0.5, 0.5) in Colmap\n    - (0,0) in OpenCV\n    \"\"\"\n    K = K.copy()\n    K[0, 2] += 0.5\n    K[1, 2] += 0.5\n    return K\n\n\ndef normalize_pointcloud(\n    pts1, pts2, norm_mode=\"avg_dis\", valid1=None, valid2=None, ret_factor=False\n):\n    \"\"\"renorm pointmaps pts1, pts2 with norm_mode\"\"\"\n    assert pts1.ndim >= 3 and pts1.shape[-1] == 3","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.opencv_to_colmap_intrinsics","uri":"program://Human3R/function/src.dust3r.utils.geometry.opencv_to_colmap_intrinsics#L248-L258","kind":"function","name":"opencv_to_colmap_intrinsics","path":"src/dust3r/utils/geometry.py","language":"python","start_line":248,"end_line":258,"context_start_line":228,"context_end_line":278,"code":"        X_world = (\n            np.einsum(\"ik, vuk -> vui\", R_cam2world, X_cam) + t_cam2world[None, None, :]\n        )\n\n    return X_world, valid_mask\n\n\ndef colmap_to_opencv_intrinsics(K):\n    \"\"\"\n    Modify camera intrinsics to follow a different convention.\n    Coordinates of the center of the top-left pixels are by default:\n    - (0.5, 0.5) in Colmap\n    - (0,0) in OpenCV\n    \"\"\"\n    K = K.copy()\n    K[0, 2] -= 0.5\n    K[1, 2] -= 0.5\n    return K\n\n\ndef opencv_to_colmap_intrinsics(K):\n    \"\"\"\n    Modify camera intrinsics to follow a different convention.\n    Coordinates of the center of the top-left pixels are by default:\n    - (0.5, 0.5) in Colmap\n    - (0,0) in OpenCV\n    \"\"\"\n    K = K.copy()\n    K[0, 2] += 0.5\n    K[1, 2] += 0.5\n    return K\n\n\ndef normalize_pointcloud(\n    pts1, pts2, norm_mode=\"avg_dis\", valid1=None, valid2=None, ret_factor=False\n):\n    \"\"\"renorm pointmaps pts1, pts2 with norm_mode\"\"\"\n    assert pts1.ndim >= 3 and pts1.shape[-1] == 3\n    assert pts2 is None or (pts2.ndim >= 3 and pts2.shape[-1] == 3)\n    norm_mode, dis_mode = norm_mode.split(\"_\")\n\n    if norm_mode == \"avg\":\n\n        nan_pts1, nnz1 = invalid_to_zeros(pts1, valid1, ndim=3)\n        nan_pts2, nnz2 = (\n            invalid_to_zeros(pts2, valid2, ndim=3) if pts2 is not None else (None, 0)\n        )\n        all_pts = (\n            torch.cat((nan_pts1, nan_pts2), dim=1) if pts2 is not None else nan_pts1\n        )\n","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.normalize_pointcloud","uri":"program://Human3R/function/src.dust3r.utils.geometry.normalize_pointcloud#L261-L326","kind":"function","name":"normalize_pointcloud","path":"src/dust3r/utils/geometry.py","language":"python","start_line":261,"end_line":326,"context_start_line":241,"context_end_line":346,"code":"    \"\"\"\n    K = K.copy()\n    K[0, 2] -= 0.5\n    K[1, 2] -= 0.5\n    return K\n\n\ndef opencv_to_colmap_intrinsics(K):\n    \"\"\"\n    Modify camera intrinsics to follow a different convention.\n    Coordinates of the center of the top-left pixels are by default:\n    - (0.5, 0.5) in Colmap\n    - (0,0) in OpenCV\n    \"\"\"\n    K = K.copy()\n    K[0, 2] += 0.5\n    K[1, 2] += 0.5\n    return K\n\n\ndef normalize_pointcloud(\n    pts1, pts2, norm_mode=\"avg_dis\", valid1=None, valid2=None, ret_factor=False\n):\n    \"\"\"renorm pointmaps pts1, pts2 with norm_mode\"\"\"\n    assert pts1.ndim >= 3 and pts1.shape[-1] == 3\n    assert pts2 is None or (pts2.ndim >= 3 and pts2.shape[-1] == 3)\n    norm_mode, dis_mode = norm_mode.split(\"_\")\n\n    if norm_mode == \"avg\":\n\n        nan_pts1, nnz1 = invalid_to_zeros(pts1, valid1, ndim=3)\n        nan_pts2, nnz2 = (\n            invalid_to_zeros(pts2, valid2, ndim=3) if pts2 is not None else (None, 0)\n        )\n        all_pts = (\n            torch.cat((nan_pts1, nan_pts2), dim=1) if pts2 is not None else nan_pts1\n        )\n\n        all_dis = all_pts.norm(dim=-1)\n        if dis_mode == \"dis\":\n            pass  # do nothing\n        elif dis_mode == \"log1p\":\n            all_dis = torch.log1p(all_dis)\n        elif dis_mode == \"warp-log1p\":\n\n            log_dis = torch.log1p(all_dis)\n            warp_factor = log_dis / all_dis.clip(min=1e-8)\n            H1, W1 = pts1.shape[1:-1]\n            pts1 = pts1 * warp_factor[:, : W1 * H1].view(-1, H1, W1, 1)\n            if pts2 is not None:\n                H2, W2 = pts2.shape[1:-1]\n                pts2 = pts2 * warp_factor[:, W1 * H1 :].view(-1, H2, W2, 1)\n            all_dis = log_dis  # this is their true distance afterwards\n        else:\n            raise ValueError(f\"bad {dis_mode=}\")\n\n        norm_factor = all_dis.sum(dim=1) / (nnz1 + nnz2 + 1e-8)\n    else:\n\n        nan_pts1 = invalid_to_nans(pts1, valid1, ndim=3)\n        nan_pts2 = invalid_to_nans(pts2, valid2, ndim=3) if pts2 is not None else None\n        all_pts = (\n            torch.cat((nan_pts1, nan_pts2), dim=1) if pts2 is not None else nan_pts1\n        )\n\n        all_dis = all_pts.norm(dim=-1)\n\n        if norm_mode == \"avg\":\n            norm_factor = all_dis.nanmean(dim=1)\n        elif norm_mode == \"median\":\n            norm_factor = all_dis.nanmedian(dim=1).values.detach()\n        elif norm_mode == \"sqrt\":\n            norm_factor = all_dis.sqrt().nanmean(dim=1) ** 2\n        else:\n            raise ValueError(f\"bad {norm_mode=}\")\n\n    norm_factor = norm_factor.clip(min=1e-8)\n    while norm_factor.ndim < pts1.ndim:\n        norm_factor.unsqueeze_(-1)\n\n    res = pts1 / norm_factor\n    if pts2 is not None:\n        res = (res, pts2 / norm_factor)\n    if ret_factor:\n        res = res + (norm_factor,)\n    return res\n\n\ndef normalize_pointcloud_group(\n    pts_list,\n    norm_mode=\"avg_dis\",\n    valid_list=None,\n    conf_list=None,\n    ret_factor=False,\n    ret_factor_only=False,\n):\n    \"\"\"renorm pointmaps pts1, pts2 with norm_mode\"\"\"\n    for pts in pts_list:\n        assert pts.ndim >= 3 and pts.shape[-1] == 3\n\n    norm_mode, dis_mode = norm_mode.split(\"_\")\n\n    if norm_mode == \"avg\":\n\n        nan_pts_list, nnz_list = zip(\n            *[","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.normalize_pointcloud_group","uri":"program://Human3R/function/src.dust3r.utils.geometry.normalize_pointcloud_group#L329-L414","kind":"function","name":"normalize_pointcloud_group","path":"src/dust3r/utils/geometry.py","language":"python","start_line":329,"end_line":414,"context_start_line":309,"context_end_line":434,"code":"            norm_factor = all_dis.nanmean(dim=1)\n        elif norm_mode == \"median\":\n            norm_factor = all_dis.nanmedian(dim=1).values.detach()\n        elif norm_mode == \"sqrt\":\n            norm_factor = all_dis.sqrt().nanmean(dim=1) ** 2\n        else:\n            raise ValueError(f\"bad {norm_mode=}\")\n\n    norm_factor = norm_factor.clip(min=1e-8)\n    while norm_factor.ndim < pts1.ndim:\n        norm_factor.unsqueeze_(-1)\n\n    res = pts1 / norm_factor\n    if pts2 is not None:\n        res = (res, pts2 / norm_factor)\n    if ret_factor:\n        res = res + (norm_factor,)\n    return res\n\n\ndef normalize_pointcloud_group(\n    pts_list,\n    norm_mode=\"avg_dis\",\n    valid_list=None,\n    conf_list=None,\n    ret_factor=False,\n    ret_factor_only=False,\n):\n    \"\"\"renorm pointmaps pts1, pts2 with norm_mode\"\"\"\n    for pts in pts_list:\n        assert pts.ndim >= 3 and pts.shape[-1] == 3\n\n    norm_mode, dis_mode = norm_mode.split(\"_\")\n\n    if norm_mode == \"avg\":\n\n        nan_pts_list, nnz_list = zip(\n            *[\n                invalid_to_zeros(pts1, valid1, ndim=3)\n                for pts1, valid1 in zip(pts_list, valid_list)\n            ]\n        )\n        all_pts = torch.cat(nan_pts_list, dim=1)\n        if conf_list is not None:\n            nan_conf_list = [\n                invalid_to_zeros(conf1[..., None], valid1, ndim=3)[0]\n                for conf1, valid1 in zip(conf_list, valid_list)\n            ]\n            all_conf = torch.cat(nan_conf_list, dim=1)[..., 0]\n        else:\n            all_conf = torch.ones_like(all_pts[..., 0])\n\n        all_dis = all_pts.norm(dim=-1)\n        if dis_mode == \"dis\":\n            pass  # do nothing\n        elif dis_mode == \"log1p\":\n            all_dis = torch.log1p(all_dis)\n        elif dis_mode == \"warp-log1p\":\n\n            log_dis = torch.log1p(all_dis)\n            warp_factor = log_dis / all_dis.clip(min=1e-8)\n            H_W_list = [pts.shape[1:-1] for pts in pts_list]\n            pts_list = [\n                pts\n                * warp_factor[:, sum(H_W_list[:i]) : sum(H_W_list[: i + 1])].view(\n                    -1, H, W, 1\n                )\n                for i, (pts, (H, W)) in enumerate(zip(pts_list, H_W_list))\n            ]\n            all_dis = log_dis  # this is their true distance afterwards\n        else:\n            raise ValueError(f\"bad {dis_mode=}\")\n\n        norm_factor = (all_conf * all_dis).sum(dim=1) / (all_conf.sum(dim=1) + 1e-8)\n    else:\n\n        nan_pts_list = [\n            invalid_to_nans(pts1, valid1, ndim=3)\n            for pts1, valid1 in zip(pts_list, valid_list)\n        ]\n\n        all_pts = torch.cat(nan_pts_list, dim=1)\n\n        all_dis = all_pts.norm(dim=-1)\n\n        if norm_mode == \"avg\":\n            norm_factor = all_dis.nanmean(dim=1)\n        elif norm_mode == \"median\":\n            norm_factor = all_dis.nanmedian(dim=1).values.detach()\n        elif norm_mode == \"sqrt\":\n            norm_factor = all_dis.sqrt().nanmean(dim=1) ** 2\n        else:\n            raise ValueError(f\"bad {norm_mode=}\")\n\n    norm_factor = norm_factor.clip(min=1e-8)\n    while norm_factor.ndim < pts_list[0].ndim:\n        norm_factor.unsqueeze_(-1)\n\n    if ret_factor_only:\n\n        return norm_factor\n\n    res = [pts / norm_factor for pts in pts_list]\n    if ret_factor:\n        return res, norm_factor\n    return res\n\n\n@torch.no_grad()\ndef get_joint_pointcloud_depth(z1, z2, valid_mask1, valid_mask2=None, quantile=0.5):\n\n    _z1 = invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1)\n    _z2 = (\n        invalid_to_nans(z2, valid_mask2).reshape(len(z2), -1)\n        if z2 is not None\n        else None\n    )\n    _z = torch.cat((_z1, _z2), dim=-1) if z2 is not None else _z1\n\n    if quantile == 0.5:\n        shift_z = torch.nanmedian(_z, dim=-1).values\n    else:\n        shift_z = torch.nanquantile(_z, quantile, dim=-1)\n    return shift_z  # (B,)\n\n","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.get_joint_pointcloud_depth","uri":"program://Human3R/function/src.dust3r.utils.geometry.get_joint_pointcloud_depth#L418-L432","kind":"function","name":"get_joint_pointcloud_depth","path":"src/dust3r/utils/geometry.py","language":"python","start_line":418,"end_line":432,"context_start_line":398,"context_end_line":452,"code":"        elif norm_mode == \"sqrt\":\n            norm_factor = all_dis.sqrt().nanmean(dim=1) ** 2\n        else:\n            raise ValueError(f\"bad {norm_mode=}\")\n\n    norm_factor = norm_factor.clip(min=1e-8)\n    while norm_factor.ndim < pts_list[0].ndim:\n        norm_factor.unsqueeze_(-1)\n\n    if ret_factor_only:\n\n        return norm_factor\n\n    res = [pts / norm_factor for pts in pts_list]\n    if ret_factor:\n        return res, norm_factor\n    return res\n\n\n@torch.no_grad()\ndef get_joint_pointcloud_depth(z1, z2, valid_mask1, valid_mask2=None, quantile=0.5):\n\n    _z1 = invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1)\n    _z2 = (\n        invalid_to_nans(z2, valid_mask2).reshape(len(z2), -1)\n        if z2 is not None\n        else None\n    )\n    _z = torch.cat((_z1, _z2), dim=-1) if z2 is not None else _z1\n\n    if quantile == 0.5:\n        shift_z = torch.nanmedian(_z, dim=-1).values\n    else:\n        shift_z = torch.nanquantile(_z, quantile, dim=-1)\n    return shift_z  # (B,)\n\n\n@torch.no_grad()\ndef get_group_pointcloud_depth(zs, valid_masks, quantile=0.5):\n\n    _zs = [\n        invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1)\n        for z1, valid_mask1 in zip(zs, valid_masks)\n    ]\n    _z = torch.cat(_zs, dim=-1)\n\n    if quantile == 0.5:\n        shift_z = torch.nanmedian(_z, dim=-1).values\n    else:\n        shift_z = torch.nanquantile(_z, quantile, dim=-1)\n    return shift_z  # (B,)\n\n\n@torch.no_grad()\ndef get_joint_pointcloud_center_scale(","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.get_group_pointcloud_depth","uri":"program://Human3R/function/src.dust3r.utils.geometry.get_group_pointcloud_depth#L436-L448","kind":"function","name":"get_group_pointcloud_depth","path":"src/dust3r/utils/geometry.py","language":"python","start_line":436,"end_line":448,"context_start_line":416,"context_end_line":468,"code":"\n@torch.no_grad()\ndef get_joint_pointcloud_depth(z1, z2, valid_mask1, valid_mask2=None, quantile=0.5):\n\n    _z1 = invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1)\n    _z2 = (\n        invalid_to_nans(z2, valid_mask2).reshape(len(z2), -1)\n        if z2 is not None\n        else None\n    )\n    _z = torch.cat((_z1, _z2), dim=-1) if z2 is not None else _z1\n\n    if quantile == 0.5:\n        shift_z = torch.nanmedian(_z, dim=-1).values\n    else:\n        shift_z = torch.nanquantile(_z, quantile, dim=-1)\n    return shift_z  # (B,)\n\n\n@torch.no_grad()\ndef get_group_pointcloud_depth(zs, valid_masks, quantile=0.5):\n\n    _zs = [\n        invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1)\n        for z1, valid_mask1 in zip(zs, valid_masks)\n    ]\n    _z = torch.cat(_zs, dim=-1)\n\n    if quantile == 0.5:\n        shift_z = torch.nanmedian(_z, dim=-1).values\n    else:\n        shift_z = torch.nanquantile(_z, quantile, dim=-1)\n    return shift_z  # (B,)\n\n\n@torch.no_grad()\ndef get_joint_pointcloud_center_scale(\n    pts1, pts2, valid_mask1=None, valid_mask2=None, z_only=False, center=True\n):\n\n    _pts1 = invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3)\n    _pts2 = (\n        invalid_to_nans(pts2, valid_mask2).reshape(len(pts2), -1, 3)\n        if pts2 is not None\n        else None\n    )\n    _pts = torch.cat((_pts1, _pts2), dim=1) if pts2 is not None else _pts1\n\n    _center = torch.nanmedian(_pts, dim=1, keepdim=True).values  # (B,1,3)\n    if z_only:\n        _center[..., :2] = 0  # do not center X and Y\n\n    _norm = ((_pts - _center) if center else _pts).norm(dim=-1)","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.get_joint_pointcloud_center_scale","uri":"program://Human3R/function/src.dust3r.utils.geometry.get_joint_pointcloud_center_scale#L452-L470","kind":"function","name":"get_joint_pointcloud_center_scale","path":"src/dust3r/utils/geometry.py","language":"python","start_line":452,"end_line":470,"context_start_line":432,"context_end_line":490,"code":"    return shift_z  # (B,)\n\n\n@torch.no_grad()\ndef get_group_pointcloud_depth(zs, valid_masks, quantile=0.5):\n\n    _zs = [\n        invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1)\n        for z1, valid_mask1 in zip(zs, valid_masks)\n    ]\n    _z = torch.cat(_zs, dim=-1)\n\n    if quantile == 0.5:\n        shift_z = torch.nanmedian(_z, dim=-1).values\n    else:\n        shift_z = torch.nanquantile(_z, quantile, dim=-1)\n    return shift_z  # (B,)\n\n\n@torch.no_grad()\ndef get_joint_pointcloud_center_scale(\n    pts1, pts2, valid_mask1=None, valid_mask2=None, z_only=False, center=True\n):\n\n    _pts1 = invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3)\n    _pts2 = (\n        invalid_to_nans(pts2, valid_mask2).reshape(len(pts2), -1, 3)\n        if pts2 is not None\n        else None\n    )\n    _pts = torch.cat((_pts1, _pts2), dim=1) if pts2 is not None else _pts1\n\n    _center = torch.nanmedian(_pts, dim=1, keepdim=True).values  # (B,1,3)\n    if z_only:\n        _center[..., :2] = 0  # do not center X and Y\n\n    _norm = ((_pts - _center) if center else _pts).norm(dim=-1)\n    scale = torch.nanmedian(_norm, dim=1).values\n    return _center[:, None, :, :], scale[:, None, None, None]\n\n\n@torch.no_grad()\ndef get_group_pointcloud_center_scale(pts, valid_masks=None, z_only=False, center=True):\n\n    _pts = [\n        invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3)\n        for pts1, valid_mask1 in zip(pts, valid_masks)\n    ]\n    _pts = torch.cat(_pts, dim=1)\n\n    _center = torch.nanmedian(_pts, dim=1, keepdim=True).values  # (B,1,3)\n    if z_only:\n        _center[..., :2] = 0  # do not center X and Y\n\n    _norm = ((_pts - _center) if center else _pts).norm(dim=-1)\n    scale = torch.nanmedian(_norm, dim=1).values\n    return _center[:, None, :, :], scale[:, None, None, None]\n\n","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.get_group_pointcloud_center_scale","uri":"program://Human3R/function/src.dust3r.utils.geometry.get_group_pointcloud_center_scale#L474-L488","kind":"function","name":"get_group_pointcloud_center_scale","path":"src/dust3r/utils/geometry.py","language":"python","start_line":474,"end_line":488,"context_start_line":454,"context_end_line":508,"code":"):\n\n    _pts1 = invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3)\n    _pts2 = (\n        invalid_to_nans(pts2, valid_mask2).reshape(len(pts2), -1, 3)\n        if pts2 is not None\n        else None\n    )\n    _pts = torch.cat((_pts1, _pts2), dim=1) if pts2 is not None else _pts1\n\n    _center = torch.nanmedian(_pts, dim=1, keepdim=True).values  # (B,1,3)\n    if z_only:\n        _center[..., :2] = 0  # do not center X and Y\n\n    _norm = ((_pts - _center) if center else _pts).norm(dim=-1)\n    scale = torch.nanmedian(_norm, dim=1).values\n    return _center[:, None, :, :], scale[:, None, None, None]\n\n\n@torch.no_grad()\ndef get_group_pointcloud_center_scale(pts, valid_masks=None, z_only=False, center=True):\n\n    _pts = [\n        invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3)\n        for pts1, valid_mask1 in zip(pts, valid_masks)\n    ]\n    _pts = torch.cat(_pts, dim=1)\n\n    _center = torch.nanmedian(_pts, dim=1, keepdim=True).values  # (B,1,3)\n    if z_only:\n        _center[..., :2] = 0  # do not center X and Y\n\n    _norm = ((_pts - _center) if center else _pts).norm(dim=-1)\n    scale = torch.nanmedian(_norm, dim=1).values\n    return _center[:, None, :, :], scale[:, None, None, None]\n\n\ndef find_reciprocal_matches(P1, P2):\n    \"\"\"\n    returns 3 values:\n    1 - reciprocal_in_P2: a boolean array of size P2.shape[0], a \"True\" value indicates a match\n    2 - nn2_in_P1: a int array of size P2.shape[0], it contains the indexes of the closest points in P1\n    3 - reciprocal_in_P2.sum(): the number of matches\n    \"\"\"\n    tree1 = KDTree(P1)\n    tree2 = KDTree(P2)\n\n    _, nn1_in_P2 = tree2.query(P1, workers=8)\n    _, nn2_in_P1 = tree1.query(P2, workers=8)\n\n    reciprocal_in_P1 = nn2_in_P1[nn1_in_P2] == np.arange(len(nn1_in_P2))\n    reciprocal_in_P2 = nn1_in_P2[nn2_in_P1] == np.arange(len(nn2_in_P1))\n    assert reciprocal_in_P1.sum() == reciprocal_in_P2.sum()\n    return reciprocal_in_P2, nn2_in_P1, reciprocal_in_P2.sum()\n","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.find_reciprocal_matches","uri":"program://Human3R/function/src.dust3r.utils.geometry.find_reciprocal_matches#L491-L507","kind":"function","name":"find_reciprocal_matches","path":"src/dust3r/utils/geometry.py","language":"python","start_line":491,"end_line":507,"context_start_line":471,"context_end_line":527,"code":"\n\n@torch.no_grad()\ndef get_group_pointcloud_center_scale(pts, valid_masks=None, z_only=False, center=True):\n\n    _pts = [\n        invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3)\n        for pts1, valid_mask1 in zip(pts, valid_masks)\n    ]\n    _pts = torch.cat(_pts, dim=1)\n\n    _center = torch.nanmedian(_pts, dim=1, keepdim=True).values  # (B,1,3)\n    if z_only:\n        _center[..., :2] = 0  # do not center X and Y\n\n    _norm = ((_pts - _center) if center else _pts).norm(dim=-1)\n    scale = torch.nanmedian(_norm, dim=1).values\n    return _center[:, None, :, :], scale[:, None, None, None]\n\n\ndef find_reciprocal_matches(P1, P2):\n    \"\"\"\n    returns 3 values:\n    1 - reciprocal_in_P2: a boolean array of size P2.shape[0], a \"True\" value indicates a match\n    2 - nn2_in_P1: a int array of size P2.shape[0], it contains the indexes of the closest points in P1\n    3 - reciprocal_in_P2.sum(): the number of matches\n    \"\"\"\n    tree1 = KDTree(P1)\n    tree2 = KDTree(P2)\n\n    _, nn1_in_P2 = tree2.query(P1, workers=8)\n    _, nn2_in_P1 = tree1.query(P2, workers=8)\n\n    reciprocal_in_P1 = nn2_in_P1[nn1_in_P2] == np.arange(len(nn1_in_P2))\n    reciprocal_in_P2 = nn1_in_P2[nn2_in_P1] == np.arange(len(nn2_in_P1))\n    assert reciprocal_in_P1.sum() == reciprocal_in_P2.sum()\n    return reciprocal_in_P2, nn2_in_P1, reciprocal_in_P2.sum()\n\n\ndef get_med_dist_between_poses(poses):\n    from scipy.spatial.distance import pdist\n\n    return np.median(pdist([to_numpy(p[:3, 3]) for p in poses]))\n\n\ndef weighted_procrustes(A, B, w, use_weights=True, eps=1e-16, return_T=False):\n    \"\"\"\n    X: torch tensor B x N x 3\n    Y: torch tensor B x N x 3\n    w: torch tensor B x N\n    \"\"\"\n    assert len(A) == len(B)\n    if use_weights:\n        W1 = torch.abs(w).sum(1, keepdim=True)\n        w_norm = (w / (W1 + eps)).unsqueeze(-1)\n        a_mean = (w_norm * A).sum(dim=1, keepdim=True)\n        b_mean = (w_norm * B).sum(dim=1, keepdim=True)","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.get_med_dist_between_poses","uri":"program://Human3R/function/src.dust3r.utils.geometry.get_med_dist_between_poses#L510-L513","kind":"function","name":"get_med_dist_between_poses","path":"src/dust3r/utils/geometry.py","language":"python","start_line":510,"end_line":513,"context_start_line":490,"context_end_line":533,"code":"\ndef find_reciprocal_matches(P1, P2):\n    \"\"\"\n    returns 3 values:\n    1 - reciprocal_in_P2: a boolean array of size P2.shape[0], a \"True\" value indicates a match\n    2 - nn2_in_P1: a int array of size P2.shape[0], it contains the indexes of the closest points in P1\n    3 - reciprocal_in_P2.sum(): the number of matches\n    \"\"\"\n    tree1 = KDTree(P1)\n    tree2 = KDTree(P2)\n\n    _, nn1_in_P2 = tree2.query(P1, workers=8)\n    _, nn2_in_P1 = tree1.query(P2, workers=8)\n\n    reciprocal_in_P1 = nn2_in_P1[nn1_in_P2] == np.arange(len(nn1_in_P2))\n    reciprocal_in_P2 = nn1_in_P2[nn2_in_P1] == np.arange(len(nn2_in_P1))\n    assert reciprocal_in_P1.sum() == reciprocal_in_P2.sum()\n    return reciprocal_in_P2, nn2_in_P1, reciprocal_in_P2.sum()\n\n\ndef get_med_dist_between_poses(poses):\n    from scipy.spatial.distance import pdist\n\n    return np.median(pdist([to_numpy(p[:3, 3]) for p in poses]))\n\n\ndef weighted_procrustes(A, B, w, use_weights=True, eps=1e-16, return_T=False):\n    \"\"\"\n    X: torch tensor B x N x 3\n    Y: torch tensor B x N x 3\n    w: torch tensor B x N\n    \"\"\"\n    assert len(A) == len(B)\n    if use_weights:\n        W1 = torch.abs(w).sum(1, keepdim=True)\n        w_norm = (w / (W1 + eps)).unsqueeze(-1)\n        a_mean = (w_norm * A).sum(dim=1, keepdim=True)\n        b_mean = (w_norm * B).sum(dim=1, keepdim=True)\n\n        A_c = A - a_mean\n        B_c = B - b_mean\n\n        H = torch.einsum(\"bni,bnj->bij\", A_c, w_norm * B_c)\n","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.weighted_procrustes","uri":"program://Human3R/function/src.dust3r.utils.geometry.weighted_procrustes#L516-L555","kind":"function","name":"weighted_procrustes","path":"src/dust3r/utils/geometry.py","language":"python","start_line":516,"end_line":555,"context_start_line":496,"context_end_line":575,"code":"    3 - reciprocal_in_P2.sum(): the number of matches\n    \"\"\"\n    tree1 = KDTree(P1)\n    tree2 = KDTree(P2)\n\n    _, nn1_in_P2 = tree2.query(P1, workers=8)\n    _, nn2_in_P1 = tree1.query(P2, workers=8)\n\n    reciprocal_in_P1 = nn2_in_P1[nn1_in_P2] == np.arange(len(nn1_in_P2))\n    reciprocal_in_P2 = nn1_in_P2[nn2_in_P1] == np.arange(len(nn2_in_P1))\n    assert reciprocal_in_P1.sum() == reciprocal_in_P2.sum()\n    return reciprocal_in_P2, nn2_in_P1, reciprocal_in_P2.sum()\n\n\ndef get_med_dist_between_poses(poses):\n    from scipy.spatial.distance import pdist\n\n    return np.median(pdist([to_numpy(p[:3, 3]) for p in poses]))\n\n\ndef weighted_procrustes(A, B, w, use_weights=True, eps=1e-16, return_T=False):\n    \"\"\"\n    X: torch tensor B x N x 3\n    Y: torch tensor B x N x 3\n    w: torch tensor B x N\n    \"\"\"\n    assert len(A) == len(B)\n    if use_weights:\n        W1 = torch.abs(w).sum(1, keepdim=True)\n        w_norm = (w / (W1 + eps)).unsqueeze(-1)\n        a_mean = (w_norm * A).sum(dim=1, keepdim=True)\n        b_mean = (w_norm * B).sum(dim=1, keepdim=True)\n\n        A_c = A - a_mean\n        B_c = B - b_mean\n\n        H = torch.einsum(\"bni,bnj->bij\", A_c, w_norm * B_c)\n\n    else:\n        a_mean = A.mean(axis=1, keepdim=True)\n        b_mean = B.mean(axis=1, keepdim=True)\n\n        A_c = A - a_mean\n        B_c = B - b_mean\n\n        H = torch.einsum(\"bij,bik->bjk\", A_c, B_c)\n\n    U, S, V = torch.svd(H)  # U: B x 3 x 3, S: B x 3, V: B x 3 x 3\n    Z = torch.eye(3).unsqueeze(0).repeat(A.shape[0], 1, 1).to(A.device)\n    Z[:, -1, -1] = torch.sign(torch.linalg.det(U @ V.transpose(1, 2)))  # B x 3 x 3\n    R = V @ Z @ U.transpose(1, 2)  # B x 3 x 3\n    t = b_mean - torch.einsum(\"bij,bjk->bik\", R, a_mean.transpose(-2, -1)).transpose(\n        -2, -1\n    )\n    if return_T:\n        T = torch.eye(4).unsqueeze(0).repeat(A.shape[0], 1, 1).to(A.device)\n        T[:, :3, :3] = R\n        T[:, :3, 3] = t.squeeze()\n        return T\n    return R, t.squeeze()\n\n\ndef perspective_projection(x, K):\n    \"\"\"\n    This function computes the perspective projection of a set of points assuming the extrinsinc params have already been applied\n    Args:\n        - x [bs,N,3]: 3D points\n        - K [bs,3,3]: Camera instrincs params\n    \"\"\"\n    # Apply perspective distortion\n    y = x / x[:, :, -1].unsqueeze(-1)  # (bs, N, 3)\n\n    # Apply camera intrinsics\n    y = torch.einsum('bij,bkj->bki', K, y)  # (bs, N, 3)\n\n    return y[:, :, :2]\n\n\ndef inverse_perspective_projection(points, K, distance):\n    \"\"\"","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.perspective_projection","uri":"program://Human3R/function/src.dust3r.utils.geometry.perspective_projection#L558-L571","kind":"function","name":"perspective_projection","path":"src/dust3r/utils/geometry.py","language":"python","start_line":558,"end_line":571,"context_start_line":538,"context_end_line":591,"code":"        A_c = A - a_mean\n        B_c = B - b_mean\n\n        H = torch.einsum(\"bij,bik->bjk\", A_c, B_c)\n\n    U, S, V = torch.svd(H)  # U: B x 3 x 3, S: B x 3, V: B x 3 x 3\n    Z = torch.eye(3).unsqueeze(0).repeat(A.shape[0], 1, 1).to(A.device)\n    Z[:, -1, -1] = torch.sign(torch.linalg.det(U @ V.transpose(1, 2)))  # B x 3 x 3\n    R = V @ Z @ U.transpose(1, 2)  # B x 3 x 3\n    t = b_mean - torch.einsum(\"bij,bjk->bik\", R, a_mean.transpose(-2, -1)).transpose(\n        -2, -1\n    )\n    if return_T:\n        T = torch.eye(4).unsqueeze(0).repeat(A.shape[0], 1, 1).to(A.device)\n        T[:, :3, :3] = R\n        T[:, :3, 3] = t.squeeze()\n        return T\n    return R, t.squeeze()\n\n\ndef perspective_projection(x, K):\n    \"\"\"\n    This function computes the perspective projection of a set of points assuming the extrinsinc params have already been applied\n    Args:\n        - x [bs,N,3]: 3D points\n        - K [bs,3,3]: Camera instrincs params\n    \"\"\"\n    # Apply perspective distortion\n    y = x / x[:, :, -1].unsqueeze(-1)  # (bs, N, 3)\n\n    # Apply camera intrinsics\n    y = torch.einsum('bij,bkj->bki', K, y)  # (bs, N, 3)\n\n    return y[:, :, :2]\n\n\ndef inverse_perspective_projection(points, K, distance):\n    \"\"\"\n    This function computes the inverse perspective projection of a set of points given an estimated distance.\n    Input:\n        points (bs, N, 2): 2D points\n        K (bs,3,3): camera intrinsics params\n        distance (bs, N, 1): distance in the 3D world\n    Similar to:\n        - pts_l_norm = cv2.undistortPoints(np.expand_dims(pts_l, axis=1), cameraMatrix=K_l, distCoeffs=None)\n    \"\"\"\n    # Apply camera intrinsics\n    points = torch.cat([points, torch.ones_like(points[..., :1])], -1)\n    points = torch.einsum('bij,bkj->bki', torch.inverse(K), points)\n\n    # Apply perspective distortion\n    if distance == None:\n        return points\n    points = points * distance","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.inverse_perspective_projection","uri":"program://Human3R/function/src.dust3r.utils.geometry.inverse_perspective_projection#L574-L592","kind":"function","name":"inverse_perspective_projection","path":"src/dust3r/utils/geometry.py","language":"python","start_line":574,"end_line":592,"context_start_line":554,"context_end_line":612,"code":"        return T\n    return R, t.squeeze()\n\n\ndef perspective_projection(x, K):\n    \"\"\"\n    This function computes the perspective projection of a set of points assuming the extrinsinc params have already been applied\n    Args:\n        - x [bs,N,3]: 3D points\n        - K [bs,3,3]: Camera instrincs params\n    \"\"\"\n    # Apply perspective distortion\n    y = x / x[:, :, -1].unsqueeze(-1)  # (bs, N, 3)\n\n    # Apply camera intrinsics\n    y = torch.einsum('bij,bkj->bki', K, y)  # (bs, N, 3)\n\n    return y[:, :, :2]\n\n\ndef inverse_perspective_projection(points, K, distance):\n    \"\"\"\n    This function computes the inverse perspective projection of a set of points given an estimated distance.\n    Input:\n        points (bs, N, 2): 2D points\n        K (bs,3,3): camera intrinsics params\n        distance (bs, N, 1): distance in the 3D world\n    Similar to:\n        - pts_l_norm = cv2.undistortPoints(np.expand_dims(pts_l, axis=1), cameraMatrix=K_l, distCoeffs=None)\n    \"\"\"\n    # Apply camera intrinsics\n    points = torch.cat([points, torch.ones_like(points[..., :1])], -1)\n    points = torch.einsum('bij,bkj->bki', torch.inverse(K), points)\n\n    # Apply perspective distortion\n    if distance == None:\n        return points\n    points = points * distance\n    return points\n\ndef get_focalLength_from_fieldOfView(fov=60, img_size=512):\n    \"\"\"\n    Compute the focal length of the camera lens by assuming a certain FOV for the entire image\n    Args:\n        - fov: float, expressed in degree\n        - img_size: int\n    Return:\n        focal: float\n    \"\"\"\n    focal = img_size / (2 * np.tan(np.radians(fov) /2))\n    return focal\n\ndef focal_length_normalization(x, f, fovn=60, img_size=448):\n    \"\"\"\n    Section 3.1 of https://arxiv.org/pdf/1904.02028.pdf\n    E = (fn/f) * E' where E is 1/d\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    y = x * (fn/f)","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.get_focalLength_from_fieldOfView","uri":"program://Human3R/function/src.dust3r.utils.geometry.get_focalLength_from_fieldOfView#L594-L604","kind":"function","name":"get_focalLength_from_fieldOfView","path":"src/dust3r/utils/geometry.py","language":"python","start_line":594,"end_line":604,"context_start_line":574,"context_end_line":624,"code":"def inverse_perspective_projection(points, K, distance):\n    \"\"\"\n    This function computes the inverse perspective projection of a set of points given an estimated distance.\n    Input:\n        points (bs, N, 2): 2D points\n        K (bs,3,3): camera intrinsics params\n        distance (bs, N, 1): distance in the 3D world\n    Similar to:\n        - pts_l_norm = cv2.undistortPoints(np.expand_dims(pts_l, axis=1), cameraMatrix=K_l, distCoeffs=None)\n    \"\"\"\n    # Apply camera intrinsics\n    points = torch.cat([points, torch.ones_like(points[..., :1])], -1)\n    points = torch.einsum('bij,bkj->bki', torch.inverse(K), points)\n\n    # Apply perspective distortion\n    if distance == None:\n        return points\n    points = points * distance\n    return points\n\ndef get_focalLength_from_fieldOfView(fov=60, img_size=512):\n    \"\"\"\n    Compute the focal length of the camera lens by assuming a certain FOV for the entire image\n    Args:\n        - fov: float, expressed in degree\n        - img_size: int\n    Return:\n        focal: float\n    \"\"\"\n    focal = img_size / (2 * np.tan(np.radians(fov) /2))\n    return focal\n\ndef focal_length_normalization(x, f, fovn=60, img_size=448):\n    \"\"\"\n    Section 3.1 of https://arxiv.org/pdf/1904.02028.pdf\n    E = (fn/f) * E' where E is 1/d\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    y = x * (fn/f)\n    return y\n\ndef undo_focal_length_normalization(y, f, fovn=60, img_size=448):\n    \"\"\"\n    Undo focal_length_normalization()\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    x = y * (f/fn)\n    return x\n\nEPS_LOG = 1e-10\ndef log_depth(x, eps=EPS_LOG):","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.focal_length_normalization","uri":"program://Human3R/function/src.dust3r.utils.geometry.focal_length_normalization#L606-L613","kind":"function","name":"focal_length_normalization","path":"src/dust3r/utils/geometry.py","language":"python","start_line":606,"end_line":613,"context_start_line":586,"context_end_line":633,"code":"    points = torch.einsum('bij,bkj->bki', torch.inverse(K), points)\n\n    # Apply perspective distortion\n    if distance == None:\n        return points\n    points = points * distance\n    return points\n\ndef get_focalLength_from_fieldOfView(fov=60, img_size=512):\n    \"\"\"\n    Compute the focal length of the camera lens by assuming a certain FOV for the entire image\n    Args:\n        - fov: float, expressed in degree\n        - img_size: int\n    Return:\n        focal: float\n    \"\"\"\n    focal = img_size / (2 * np.tan(np.radians(fov) /2))\n    return focal\n\ndef focal_length_normalization(x, f, fovn=60, img_size=448):\n    \"\"\"\n    Section 3.1 of https://arxiv.org/pdf/1904.02028.pdf\n    E = (fn/f) * E' where E is 1/d\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    y = x * (fn/f)\n    return y\n\ndef undo_focal_length_normalization(y, f, fovn=60, img_size=448):\n    \"\"\"\n    Undo focal_length_normalization()\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    x = y * (f/fn)\n    return x\n\nEPS_LOG = 1e-10\ndef log_depth(x, eps=EPS_LOG):\n    \"\"\"\n    Move depth to log space\n    \"\"\"\n    return torch.log(x + eps)\n\ndef undo_log_depth(y, eps=EPS_LOG):\n    \"\"\"\n    Undo log_depth()\n    \"\"\"","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.undo_focal_length_normalization","uri":"program://Human3R/function/src.dust3r.utils.geometry.undo_focal_length_normalization#L615-L621","kind":"function","name":"undo_focal_length_normalization","path":"src/dust3r/utils/geometry.py","language":"python","start_line":615,"end_line":621,"context_start_line":595,"context_end_line":641,"code":"    \"\"\"\n    Compute the focal length of the camera lens by assuming a certain FOV for the entire image\n    Args:\n        - fov: float, expressed in degree\n        - img_size: int\n    Return:\n        focal: float\n    \"\"\"\n    focal = img_size / (2 * np.tan(np.radians(fov) /2))\n    return focal\n\ndef focal_length_normalization(x, f, fovn=60, img_size=448):\n    \"\"\"\n    Section 3.1 of https://arxiv.org/pdf/1904.02028.pdf\n    E = (fn/f) * E' where E is 1/d\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    y = x * (fn/f)\n    return y\n\ndef undo_focal_length_normalization(y, f, fovn=60, img_size=448):\n    \"\"\"\n    Undo focal_length_normalization()\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    x = y * (f/fn)\n    return x\n\nEPS_LOG = 1e-10\ndef log_depth(x, eps=EPS_LOG):\n    \"\"\"\n    Move depth to log space\n    \"\"\"\n    return torch.log(x + eps)\n\ndef undo_log_depth(y, eps=EPS_LOG):\n    \"\"\"\n    Undo log_depth()\n    \"\"\"\n    return torch.exp(y) - eps\n\ndef get_camera_parameters(img_size, fov=60, p_x=None, p_y=None, device=torch.device('cuda')):\n    \"\"\" Given image size, fov and principal point coordinates, return K the camera parameter matrix\"\"\"\n    K = torch.eye(3)\n    # Get focal length.\n    focal = get_focalLength_from_fieldOfView(fov=fov, img_size=img_size)\n    K[0,0], K[1,1] = focal, focal","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.log_depth","uri":"program://Human3R/function/src.dust3r.utils.geometry.log_depth#L624-L628","kind":"function","name":"log_depth","path":"src/dust3r/utils/geometry.py","language":"python","start_line":624,"end_line":628,"context_start_line":604,"context_end_line":648,"code":"    return focal\n\ndef focal_length_normalization(x, f, fovn=60, img_size=448):\n    \"\"\"\n    Section 3.1 of https://arxiv.org/pdf/1904.02028.pdf\n    E = (fn/f) * E' where E is 1/d\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    y = x * (fn/f)\n    return y\n\ndef undo_focal_length_normalization(y, f, fovn=60, img_size=448):\n    \"\"\"\n    Undo focal_length_normalization()\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    x = y * (f/fn)\n    return x\n\nEPS_LOG = 1e-10\ndef log_depth(x, eps=EPS_LOG):\n    \"\"\"\n    Move depth to log space\n    \"\"\"\n    return torch.log(x + eps)\n\ndef undo_log_depth(y, eps=EPS_LOG):\n    \"\"\"\n    Undo log_depth()\n    \"\"\"\n    return torch.exp(y) - eps\n\ndef get_camera_parameters(img_size, fov=60, p_x=None, p_y=None, device=torch.device('cuda')):\n    \"\"\" Given image size, fov and principal point coordinates, return K the camera parameter matrix\"\"\"\n    K = torch.eye(3)\n    # Get focal length.\n    focal = get_focalLength_from_fieldOfView(fov=fov, img_size=img_size)\n    K[0,0], K[1,1] = focal, focal\n\n    # Set principal point\n    if p_x is not None and p_y is not None:\n            K[0,-1], K[1,-1] = p_x * img_size, p_y * img_size\n    else:\n            K[0,-1], K[1,-1] = img_size//2, img_size//2\n","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.undo_log_depth","uri":"program://Human3R/function/src.dust3r.utils.geometry.undo_log_depth#L630-L634","kind":"function","name":"undo_log_depth","path":"src/dust3r/utils/geometry.py","language":"python","start_line":630,"end_line":634,"context_start_line":610,"context_end_line":654,"code":"    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    y = x * (fn/f)\n    return y\n\ndef undo_focal_length_normalization(y, f, fovn=60, img_size=448):\n    \"\"\"\n    Undo focal_length_normalization()\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    x = y * (f/fn)\n    return x\n\nEPS_LOG = 1e-10\ndef log_depth(x, eps=EPS_LOG):\n    \"\"\"\n    Move depth to log space\n    \"\"\"\n    return torch.log(x + eps)\n\ndef undo_log_depth(y, eps=EPS_LOG):\n    \"\"\"\n    Undo log_depth()\n    \"\"\"\n    return torch.exp(y) - eps\n\ndef get_camera_parameters(img_size, fov=60, p_x=None, p_y=None, device=torch.device('cuda')):\n    \"\"\" Given image size, fov and principal point coordinates, return K the camera parameter matrix\"\"\"\n    K = torch.eye(3)\n    # Get focal length.\n    focal = get_focalLength_from_fieldOfView(fov=fov, img_size=img_size)\n    K[0,0], K[1,1] = focal, focal\n\n    # Set principal point\n    if p_x is not None and p_y is not None:\n            K[0,-1], K[1,-1] = p_x * img_size, p_y * img_size\n    else:\n            K[0,-1], K[1,-1] = img_size//2, img_size//2\n\n    # Add batch dimension\n    K = K.unsqueeze(0).to(device)\n    return K\n\ndef to_euclidean_dist(img_size, dist, _K, fovn=60):\n    # Focal length normalization","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.get_camera_parameters","uri":"program://Human3R/function/src.dust3r.utils.geometry.get_camera_parameters#L636-L651","kind":"function","name":"get_camera_parameters","path":"src/dust3r/utils/geometry.py","language":"python","start_line":636,"end_line":651,"context_start_line":616,"context_end_line":671,"code":"    \"\"\"\n    Undo focal_length_normalization()\n    \"\"\"\n    fn = get_focalLength_from_fieldOfView(fov=fovn, img_size=img_size)\n    x = y * (f/fn)\n    return x\n\nEPS_LOG = 1e-10\ndef log_depth(x, eps=EPS_LOG):\n    \"\"\"\n    Move depth to log space\n    \"\"\"\n    return torch.log(x + eps)\n\ndef undo_log_depth(y, eps=EPS_LOG):\n    \"\"\"\n    Undo log_depth()\n    \"\"\"\n    return torch.exp(y) - eps\n\ndef get_camera_parameters(img_size, fov=60, p_x=None, p_y=None, device=torch.device('cuda')):\n    \"\"\" Given image size, fov and principal point coordinates, return K the camera parameter matrix\"\"\"\n    K = torch.eye(3)\n    # Get focal length.\n    focal = get_focalLength_from_fieldOfView(fov=fov, img_size=img_size)\n    K[0,0], K[1,1] = focal, focal\n\n    # Set principal point\n    if p_x is not None and p_y is not None:\n            K[0,-1], K[1,-1] = p_x * img_size, p_y * img_size\n    else:\n            K[0,-1], K[1,-1] = img_size//2, img_size//2\n\n    # Add batch dimension\n    K = K.unsqueeze(0).to(device)\n    return K\n\ndef to_euclidean_dist(img_size, dist, _K, fovn=60):\n    # Focal length normalization\n    focal = _K[:,[0],[0]]\n    dist = undo_focal_length_normalization(dist, focal, fovn=fovn, img_size=img_size)\n\n    # log space\n    dist = undo_log_depth(dist)\n\n    # Clamping\n    dist = torch.clamp(dist, 0, 50)\n\n    return dist\n\n\ndef resize_camera_intrinsics(K, original_height, original_width, target_size):\n    \"\"\"\n    Resize camera intrinsics matrix to target image size\n    \n    Args:","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.to_euclidean_dist","uri":"program://Human3R/function/src.dust3r.utils.geometry.to_euclidean_dist#L653-L664","kind":"function","name":"to_euclidean_dist","path":"src/dust3r/utils/geometry.py","language":"python","start_line":653,"end_line":664,"context_start_line":633,"context_end_line":684,"code":"    \"\"\"\n    return torch.exp(y) - eps\n\ndef get_camera_parameters(img_size, fov=60, p_x=None, p_y=None, device=torch.device('cuda')):\n    \"\"\" Given image size, fov and principal point coordinates, return K the camera parameter matrix\"\"\"\n    K = torch.eye(3)\n    # Get focal length.\n    focal = get_focalLength_from_fieldOfView(fov=fov, img_size=img_size)\n    K[0,0], K[1,1] = focal, focal\n\n    # Set principal point\n    if p_x is not None and p_y is not None:\n            K[0,-1], K[1,-1] = p_x * img_size, p_y * img_size\n    else:\n            K[0,-1], K[1,-1] = img_size//2, img_size//2\n\n    # Add batch dimension\n    K = K.unsqueeze(0).to(device)\n    return K\n\ndef to_euclidean_dist(img_size, dist, _K, fovn=60):\n    # Focal length normalization\n    focal = _K[:,[0],[0]]\n    dist = undo_focal_length_normalization(dist, focal, fovn=fovn, img_size=img_size)\n\n    # log space\n    dist = undo_log_depth(dist)\n\n    # Clamping\n    dist = torch.clamp(dist, 0, 50)\n\n    return dist\n\n\ndef resize_camera_intrinsics(K, original_height, original_width, target_size):\n    \"\"\"\n    Resize camera intrinsics matrix to target image size\n    \n    Args:\n        K: Camera intrinsics matrix (N, 3, 3)\n        original_width: Original image width\n        original_height: Original image height  \n        target_size: Target image size (assumed square)\n    \n    Returns:\n        K_resized: Resized camera intrinsics matrix\n    \"\"\"\n    K_resized = K.clone()\n    \n    # normalize between 0 and 1\n    princpt_width = K[:, 0, 2] / original_width\n    princpt_height = K[:, 1, 2] / original_height","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.resize_camera_intrinsics","uri":"program://Human3R/function/src.dust3r.utils.geometry.resize_camera_intrinsics#L667-L688","kind":"function","name":"resize_camera_intrinsics","path":"src/dust3r/utils/geometry.py","language":"python","start_line":667,"end_line":688,"context_start_line":647,"context_end_line":708,"code":"            K[0,-1], K[1,-1] = img_size//2, img_size//2\n\n    # Add batch dimension\n    K = K.unsqueeze(0).to(device)\n    return K\n\ndef to_euclidean_dist(img_size, dist, _K, fovn=60):\n    # Focal length normalization\n    focal = _K[:,[0],[0]]\n    dist = undo_focal_length_normalization(dist, focal, fovn=fovn, img_size=img_size)\n\n    # log space\n    dist = undo_log_depth(dist)\n\n    # Clamping\n    dist = torch.clamp(dist, 0, 50)\n\n    return dist\n\n\ndef resize_camera_intrinsics(K, original_height, original_width, target_size):\n    \"\"\"\n    Resize camera intrinsics matrix to target image size\n    \n    Args:\n        K: Camera intrinsics matrix (N, 3, 3)\n        original_width: Original image width\n        original_height: Original image height  \n        target_size: Target image size (assumed square)\n    \n    Returns:\n        K_resized: Resized camera intrinsics matrix\n    \"\"\"\n    K_resized = K.clone()\n    \n    # normalize between 0 and 1\n    princpt_width = K[:, 0, 2] / original_width\n    princpt_height = K[:, 1, 2] / original_height\n    K_resized[:, [0,1], 2] = target_size * torch.stack([princpt_width, princpt_height], dim=1)\n    K_resized[:, [0,1], [0,1]] /= (max(original_width, original_height) / target_size)\n    \n    return K_resized\n\ndef unresize_camera_intrinsics(K, original_size, target_height, target_width):\n    \"\"\"\n    Resize camera intrinsics matrix from square size back to target rectangular size\n    \n    Args:\n        K: Camera intrinsics matrix (N, 3, 3) - from square image\n        original_size: Original square image size \n        target_height: Target image height\n        target_width: Target image width\n    \n    Returns:\n        K_resized: Resized camera intrinsics matrix\n    \"\"\"\n    K_resized = K\n    \n    # normalize between 0 and 1 (from square coordinates)\n    princpt_width = K[:, 0, 2] / original_size\n    princpt_height = K[:, 1, 2] / original_size\n    ","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.unresize_camera_intrinsics","uri":"program://Human3R/function/src.dust3r.utils.geometry.unresize_camera_intrinsics#L690-L717","kind":"function","name":"unresize_camera_intrinsics","path":"src/dust3r/utils/geometry.py","language":"python","start_line":690,"end_line":717,"context_start_line":670,"context_end_line":729,"code":"    \n    Args:\n        K: Camera intrinsics matrix (N, 3, 3)\n        original_width: Original image width\n        original_height: Original image height  \n        target_size: Target image size (assumed square)\n    \n    Returns:\n        K_resized: Resized camera intrinsics matrix\n    \"\"\"\n    K_resized = K.clone()\n    \n    # normalize between 0 and 1\n    princpt_width = K[:, 0, 2] / original_width\n    princpt_height = K[:, 1, 2] / original_height\n    K_resized[:, [0,1], 2] = target_size * torch.stack([princpt_width, princpt_height], dim=1)\n    K_resized[:, [0,1], [0,1]] /= (max(original_width, original_height) / target_size)\n    \n    return K_resized\n\ndef unresize_camera_intrinsics(K, original_size, target_height, target_width):\n    \"\"\"\n    Resize camera intrinsics matrix from square size back to target rectangular size\n    \n    Args:\n        K: Camera intrinsics matrix (N, 3, 3) - from square image\n        original_size: Original square image size \n        target_height: Target image height\n        target_width: Target image width\n    \n    Returns:\n        K_resized: Resized camera intrinsics matrix\n    \"\"\"\n    K_resized = K\n    \n    # normalize between 0 and 1 (from square coordinates)\n    princpt_width = K[:, 0, 2] / original_size\n    princpt_height = K[:, 1, 2] / original_size\n    \n    # convert to target rectangular coordinates\n    K_resized[:, 0, 2] = princpt_width * target_width\n    K_resized[:, 1, 2] = princpt_height * target_height\n    \n    # adjust focal lengths\n    scale_factor = max(target_height, target_width) / original_size\n    K_resized[:, [0,1], [0,1]] *= scale_factor\n    \n    return K_resized\n\n\ndef matrix_cumprod(matrices):\n    if len(matrices) == 0:\n        return matrices\n    \n    result = torch.empty_like(matrices)\n    result[0] = matrices[0]\n    \n    for i in range(1, len(matrices)):\n        torch.matmul(result[i-1], matrices[i], out=result[i])\n    return result","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.geometry.matrix_cumprod","uri":"program://Human3R/function/src.dust3r.utils.geometry.matrix_cumprod#L720-L729","kind":"function","name":"matrix_cumprod","path":"src/dust3r/utils/geometry.py","language":"python","start_line":720,"end_line":729,"context_start_line":700,"context_end_line":729,"code":"    Returns:\n        K_resized: Resized camera intrinsics matrix\n    \"\"\"\n    K_resized = K\n    \n    # normalize between 0 and 1 (from square coordinates)\n    princpt_width = K[:, 0, 2] / original_size\n    princpt_height = K[:, 1, 2] / original_size\n    \n    # convert to target rectangular coordinates\n    K_resized[:, 0, 2] = princpt_width * target_width\n    K_resized[:, 1, 2] = princpt_height * target_height\n    \n    # adjust focal lengths\n    scale_factor = max(target_height, target_width) / original_size\n    K_resized[:, [0,1], [0,1]] *= scale_factor\n    \n    return K_resized\n\n\ndef matrix_cumprod(matrices):\n    if len(matrices) == 0:\n        return matrices\n    \n    result = torch.empty_like(matrices)\n    result[0] = matrices[0]\n    \n    for i in range(1, len(matrices)):\n        torch.matmul(result[i-1], matrices[i], out=result[i])\n    return result","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.parallel","uri":"program://Human3R/module/src.dust3r.utils.parallel#L1-L87","kind":"module","name":"src.dust3r.utils.parallel","path":"src/dust3r/utils/parallel.py","language":"python","start_line":1,"end_line":87,"context_start_line":1,"context_end_line":87,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nfrom tqdm import tqdm\nfrom multiprocessing.dummy import Pool as ThreadPool\nfrom multiprocessing import cpu_count\n\n\ndef parallel_threads(\n    function,\n    args,\n    workers=0,\n    star_args=False,\n    kw_args=False,\n    front_num=1,\n    Pool=ThreadPool,\n    **tqdm_kw\n):\n    \"\"\"tqdm but with parallel execution.\n\n    Will essentially return\n      res = [ function(arg) # default\n              function(*arg) # if star_args is True\n              function(**arg) # if kw_args is True\n              for arg in args]\n\n    Note:\n        the <front_num> first elements of args will not be parallelized.\n        This can be useful for debugging.\n    \"\"\"\n    while workers <= 0:\n        workers += cpu_count()\n    if workers == 1:\n        front_num = float(\"inf\")\n\n    try:\n        n_args_parallel = len(args) - front_num\n    except TypeError:\n        n_args_parallel = None\n    args = iter(args)\n\n    front = []\n    while len(front) < front_num:\n        try:\n            a = next(args)\n        except StopIteration:\n            return front  # end of the iterable\n        front.append(\n            function(*a) if star_args else function(**a) if kw_args else function(a)\n        )\n\n    out = []\n    with Pool(workers) as pool:\n\n        if star_args:\n            futures = pool.imap(starcall, [(function, a) for a in args])\n        elif kw_args:\n            futures = pool.imap(starstarcall, [(function, a) for a in args])\n        else:\n            futures = pool.imap(function, args)\n\n        for f in tqdm(futures, total=n_args_parallel, **tqdm_kw):\n            out.append(f)\n    return front + out\n\n\ndef parallel_processes(*args, **kwargs):\n    \"\"\"Same as parallel_threads, with processes\"\"\"\n    import multiprocessing as mp\n\n    kwargs[\"Pool\"] = mp.Pool\n    return parallel_threads(*args, **kwargs)\n\n\ndef starcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(*args)\n\n\ndef starstarcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(**args)","source_hash":"8af1ba7e2e52a7daf1af51a955b98f7b5de3e24e70e66ab5dd3555c8df48aefb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.parallel.parallel_threads","uri":"program://Human3R/function/src.dust3r.utils.parallel.parallel_threads#L12-L67","kind":"function","name":"parallel_threads","path":"src/dust3r/utils/parallel.py","language":"python","start_line":12,"end_line":67,"context_start_line":1,"context_end_line":87,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nfrom tqdm import tqdm\nfrom multiprocessing.dummy import Pool as ThreadPool\nfrom multiprocessing import cpu_count\n\n\ndef parallel_threads(\n    function,\n    args,\n    workers=0,\n    star_args=False,\n    kw_args=False,\n    front_num=1,\n    Pool=ThreadPool,\n    **tqdm_kw\n):\n    \"\"\"tqdm but with parallel execution.\n\n    Will essentially return\n      res = [ function(arg) # default\n              function(*arg) # if star_args is True\n              function(**arg) # if kw_args is True\n              for arg in args]\n\n    Note:\n        the <front_num> first elements of args will not be parallelized.\n        This can be useful for debugging.\n    \"\"\"\n    while workers <= 0:\n        workers += cpu_count()\n    if workers == 1:\n        front_num = float(\"inf\")\n\n    try:\n        n_args_parallel = len(args) - front_num\n    except TypeError:\n        n_args_parallel = None\n    args = iter(args)\n\n    front = []\n    while len(front) < front_num:\n        try:\n            a = next(args)\n        except StopIteration:\n            return front  # end of the iterable\n        front.append(\n            function(*a) if star_args else function(**a) if kw_args else function(a)\n        )\n\n    out = []\n    with Pool(workers) as pool:\n\n        if star_args:\n            futures = pool.imap(starcall, [(function, a) for a in args])\n        elif kw_args:\n            futures = pool.imap(starstarcall, [(function, a) for a in args])\n        else:\n            futures = pool.imap(function, args)\n\n        for f in tqdm(futures, total=n_args_parallel, **tqdm_kw):\n            out.append(f)\n    return front + out\n\n\ndef parallel_processes(*args, **kwargs):\n    \"\"\"Same as parallel_threads, with processes\"\"\"\n    import multiprocessing as mp\n\n    kwargs[\"Pool\"] = mp.Pool\n    return parallel_threads(*args, **kwargs)\n\n\ndef starcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(*args)\n\n\ndef starstarcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(**args)","source_hash":"8af1ba7e2e52a7daf1af51a955b98f7b5de3e24e70e66ab5dd3555c8df48aefb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.parallel.parallel_processes","uri":"program://Human3R/function/src.dust3r.utils.parallel.parallel_processes#L70-L75","kind":"function","name":"parallel_processes","path":"src/dust3r/utils/parallel.py","language":"python","start_line":70,"end_line":75,"context_start_line":50,"context_end_line":87,"code":"            return front  # end of the iterable\n        front.append(\n            function(*a) if star_args else function(**a) if kw_args else function(a)\n        )\n\n    out = []\n    with Pool(workers) as pool:\n\n        if star_args:\n            futures = pool.imap(starcall, [(function, a) for a in args])\n        elif kw_args:\n            futures = pool.imap(starstarcall, [(function, a) for a in args])\n        else:\n            futures = pool.imap(function, args)\n\n        for f in tqdm(futures, total=n_args_parallel, **tqdm_kw):\n            out.append(f)\n    return front + out\n\n\ndef parallel_processes(*args, **kwargs):\n    \"\"\"Same as parallel_threads, with processes\"\"\"\n    import multiprocessing as mp\n\n    kwargs[\"Pool\"] = mp.Pool\n    return parallel_threads(*args, **kwargs)\n\n\ndef starcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(*args)\n\n\ndef starstarcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(**args)","source_hash":"8af1ba7e2e52a7daf1af51a955b98f7b5de3e24e70e66ab5dd3555c8df48aefb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.parallel.starcall","uri":"program://Human3R/function/src.dust3r.utils.parallel.starcall#L78-L81","kind":"function","name":"starcall","path":"src/dust3r/utils/parallel.py","language":"python","start_line":78,"end_line":81,"context_start_line":58,"context_end_line":87,"code":"        if star_args:\n            futures = pool.imap(starcall, [(function, a) for a in args])\n        elif kw_args:\n            futures = pool.imap(starstarcall, [(function, a) for a in args])\n        else:\n            futures = pool.imap(function, args)\n\n        for f in tqdm(futures, total=n_args_parallel, **tqdm_kw):\n            out.append(f)\n    return front + out\n\n\ndef parallel_processes(*args, **kwargs):\n    \"\"\"Same as parallel_threads, with processes\"\"\"\n    import multiprocessing as mp\n\n    kwargs[\"Pool\"] = mp.Pool\n    return parallel_threads(*args, **kwargs)\n\n\ndef starcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(*args)\n\n\ndef starstarcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(**args)","source_hash":"8af1ba7e2e52a7daf1af51a955b98f7b5de3e24e70e66ab5dd3555c8df48aefb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.parallel.starstarcall","uri":"program://Human3R/function/src.dust3r.utils.parallel.starstarcall#L84-L87","kind":"function","name":"starstarcall","path":"src/dust3r/utils/parallel.py","language":"python","start_line":84,"end_line":87,"context_start_line":64,"context_end_line":87,"code":"\n        for f in tqdm(futures, total=n_args_parallel, **tqdm_kw):\n            out.append(f)\n    return front + out\n\n\ndef parallel_processes(*args, **kwargs):\n    \"\"\"Same as parallel_threads, with processes\"\"\"\n    import multiprocessing as mp\n\n    kwargs[\"Pool\"] = mp.Pool\n    return parallel_threads(*args, **kwargs)\n\n\ndef starcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(*args)\n\n\ndef starstarcall(args):\n    \"\"\"convenient wrapper for Process.Pool\"\"\"\n    function, args = args\n    return function(**args)","source_hash":"8af1ba7e2e52a7daf1af51a955b98f7b5de3e24e70e66ab5dd3555c8df48aefb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image","uri":"program://Human3R/module/src.dust3r.utils.image#L1-L527","kind":"module","name":"src.dust3r.utils.image","path":"src/dust3r/utils/image.py","language":"python","start_line":1,"end_line":527,"context_start_line":1,"context_end_line":527,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport os\nimport torch\nimport numpy as np\nimport PIL.Image\nfrom PIL.ImageOps import exif_transpose\nimport torchvision.transforms as tvf\n\nos.environ[\"OPENCV_IO_ENABLE_OPENEXR\"] = \"1\"\nimport cv2  # noqa\n\ntry:\n    from pillow_heif import register_heif_opener  # noqa\n\n    register_heif_opener()\n    heif_support_enabled = True\nexcept ImportError:\n    heif_support_enabled = False\n\nImgNorm = tvf.Compose([tvf.ToTensor(), tvf.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])\n\n\ndef img_to_arr(img):\n    if isinstance(img, str):\n        img = imread_cv2(img)\n    return img\n\n\ndef imread_cv2(path, options=cv2.IMREAD_COLOR):\n    \"\"\"Open an image or a depthmap with opencv-python.\"\"\"\n    if path.endswith((\".exr\", \"EXR\")):\n        options = cv2.IMREAD_ANYDEPTH\n    img = cv2.imread(path, options)\n    if img is None:\n        raise IOError(f\"Could not load image={path} with {options=}\")\n    if img.ndim == 3:\n        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n    return img\n\n\ndef rgb(ftensor, true_shape=None):\n    if isinstance(ftensor, list):\n        return [rgb(x, true_shape=true_shape) for x in ftensor]\n    if isinstance(ftensor, torch.Tensor):\n        ftensor = ftensor.detach().cpu().numpy()  # H,W,3\n    if ftensor.ndim == 3 and ftensor.shape[0] == 3:\n        ftensor = ftensor.transpose(1, 2, 0)\n    elif ftensor.ndim == 4 and ftensor.shape[1] == 3:\n        ftensor = ftensor.transpose(0, 2, 3, 1)\n    if true_shape is not None:\n        H, W = true_shape\n        ftensor = ftensor[:H, :W]\n    if ftensor.dtype == np.uint8:\n        img = np.float32(ftensor) / 255\n    else:\n        img = (ftensor * 0.5) + 0.5\n    return img.clip(min=0, max=1)\n\n\ndef _resize_pil_image(img, long_edge_size):\n    S = max(img.size)\n    if S > long_edge_size:\n        interp = PIL.Image.LANCZOS\n    elif S <= long_edge_size:\n        interp = PIL.Image.BICUBIC\n    new_size = tuple(int(round(x * long_edge_size / S)) for x in img.size)\n    return img.resize(new_size, interp)\n\n\ndef load_images(folder_or_list, size, square_ok=False, verbose=True):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        if verbose:\n            print(f\">> Loading a list of {len(folder_or_list)} images\")\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\", \".bmp\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]\n    supported_images_extensions = tuple(supported_images_extensions)\n\n    imgs = []\n    for path in folder_content:\n        if not path.lower().endswith(supported_images_extensions):\n            continue\n        img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert(\"RGB\")\n        W1, H1 = img.size\n        if size == 224:\n\n            img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1)))\n        else:\n\n            img = _resize_pil_image(img, size)\n        W, H = img.size\n        cx, cy = W // 2, H // 2\n        if size == 224:\n            half = min(cx, cy)\n            img = img.crop((cx - half, cy - half, cx + half, cy + half))\n        else:\n            halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8\n            if not (square_ok) and W == H:\n                halfh = 3 * halfw / 4\n            img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n\n        W2, H2 = img.size\n        if verbose:\n            print(f\" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}\")\n        imgs.append(\n            dict(\n                img=ImgNorm(img)[None],\n                true_shape=np.int32([img.size[::-1]]),\n                idx=len(imgs),\n                instance=str(len(imgs)),\n            )\n        )\n\n    assert imgs, \"no images foud at \" + root\n    if verbose:\n        print(f\" (Found {len(imgs)} images)\")\n    return imgs\n\n\ndef load_images_for_eval(\n    folder_or_list, size, square_ok=False, verbose=True, crop=True\n):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        if verbose:\n            print(f\">> Loading a list of {len(folder_or_list)} images\")\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]\n    supported_images_extensions = tuple(supported_images_extensions)\n\n    imgs = []\n    for path in folder_content:\n        if not path.lower().endswith(supported_images_extensions):\n            continue\n        img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert(\"RGB\")\n        W1, H1 = img.size\n        if size == 224:\n            # resize short side to 224 (then crop)\n            img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1)))\n        else:\n            # resize long side to 512\n            img = _resize_pil_image(img, size)\n        W, H = img.size\n        cx, cy = W // 2, H // 2\n        if size == 224:\n            half = min(cx, cy)\n            if crop:\n                img = img.crop((cx - half, cy - half, cx + half, cy + half))\n            else:  # resize\n                img = img.resize((2 * half, 2 * half), PIL.Image.LANCZOS)\n        else:\n            halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8\n            if not (square_ok) and W == H:\n                halfh = 3 * halfw / 4\n            if crop:\n                img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n            else:  # resize\n                img = img.resize((2 * halfw, 2 * halfh), PIL.Image.LANCZOS)\n        W2, H2 = img.size\n        if verbose:\n            print(f\" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}\")\n        imgs.append(\n            dict(\n                img=ImgNorm(img)[None],\n                true_shape=np.int32([img.size[::-1]]),\n                idx=len(imgs),\n                instance=str(len(imgs)),\n                ori_shape=np.int32([[H1, W1]])\n            )\n        )\n\n    assert imgs, \"no images foud at \" + root\n    if verbose:\n        print(f\" (Found {len(imgs)} images)\")\n    return imgs\n\n\n\ndef load_masks_for_eval(\n    folder_or_list, size, square_ok=False, verbose=True, crop=True\n):\n    if isinstance(folder_or_list, str):\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]\n    supported_images_extensions = tuple(supported_images_extensions)\n\n    imgs = []\n    for path in folder_content:\n        if not path.lower().endswith(supported_images_extensions):\n            continue\n        img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert(\"RGB\")\n        W1, H1 = img.size\n        if size == 224:\n            # resize short side to 224 (then crop)\n            img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1)))\n        else:\n            # resize long side to 512\n            img = _resize_pil_image(img, size)\n        W, H = img.size\n        cx, cy = W // 2, H // 2\n        if size == 224:\n            half = min(cx, cy)\n            if crop:\n                img = img.crop((cx - half, cy - half, cx + half, cy + half))\n            else:  # resize\n                img = img.resize((2 * half, 2 * half), PIL.Image.LANCZOS)\n        else:\n            halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8\n            if not (square_ok) and W == H:\n                halfh = 3 * halfw / 4\n            if crop:\n                img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n            else:  # resize\n                img = img.resize((2 * halfw, 2 * halfh), PIL.Image.LANCZOS)\n\n        img = tvf.ToTensor()(img)\n        img = (img[0] > 0.5).float()\n        imgs.append(img[None])\n\n    assert imgs, \"no images foud at \" + root\n    return imgs\n\n\ndef load_images_512(folder_or_list, size, square_ok=False, verbose=True):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        if verbose:\n            print(f\">> Loading a list of {len(folder_or_list)} images\")\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\", \".bmp\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]\n    supported_images_extensions = tuple(supported_images_extensions)\n\n    imgs = []\n    for path in folder_content:\n        if not path.lower().endswith(supported_images_extensions):\n            continue\n        img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert(\"RGB\")\n        img = img.resize((512, 384))\n        W1, H1 = img.size\n        if size == 224:\n\n            img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1)))\n        else:\n\n            img = _resize_pil_image(img, size)\n        W, H = img.size\n        cx, cy = W // 2, H // 2\n        if size == 224:\n            half = min(cx, cy)\n            img = img.crop((cx - half, cy - half, cx + half, cy + half))\n        else:\n            halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8\n            if not (square_ok) and W == H:\n                halfh = 3 * halfw / 4\n            img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n\n        W2, H2 = img.size\n        if verbose:\n            print(f\" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}\")\n        imgs.append(\n            dict(\n                img=ImgNorm(img)[None],\n                true_shape=np.int32([img.size[::-1]]),\n                idx=len(imgs),\n                instance=str(len(imgs)),\n            )\n        )\n\n    assert imgs, \"no images foud at \" + root\n    if verbose:\n        print(f\" (Found {len(imgs)} images)\")\n    return imgs\n\n\nIMG_NORM_MEAN = [0.5, 0.5, 0.5]\nIMG_NORM_STD = [0.5, 0.5, 0.5]\n\n\ndef normalize_rgb(img, imagenet_normalization=True):\n    \"\"\"\n    Args:\n        - img: np.array - (W,H,3) - np.uint8 - 0/255\n    Return:\n        - img: np.array - (3,W,H) - np.float - -3/3\n    \"\"\"\n    img = img.astype(np.float32) / 255.\n    img = np.transpose(img, (2,0,1))\n    if imagenet_normalization:\n        img = (img - np.asarray(IMG_NORM_MEAN).reshape(3,1,1)) / np.asarray(IMG_NORM_STD).reshape(3,1,1)\n    img = img.astype(np.float32)\n    return img\n\n\ndef denormalize_rgb(img, imagenet_normalization=True):\n    \"\"\"\n    Args:\n        - img: np.array - (3,W,H) - np.float - -3/3\n    Return:\n        - img: np.array - (W,H,3) - np.uint8 - 0/255\n    \"\"\"\n    if imagenet_normalization:\n        img = (img * np.asarray(IMG_NORM_STD).reshape(3,1,1)) + np.asarray(IMG_NORM_MEAN).reshape(3,1,1)\n    img = np.transpose(img, (1,2,0)) * 255.\n    img = img.astype(np.uint8)\n    return img\n\n\nimport torch.nn.functional as F\n\ndef pad_image(img_tensor, target_size, pad_value=-1.0):\n    \"\"\"\n    torch version of ImageOps.pad, equivalent to the combination of contain and pad\n    \n    Args:\n        img_tensor: torch tensor, shape [C, H, W] or [B, C, H, W]\n        target_size: int, target size (square)\n    \n    Returns:\n        torch tensor, shape [C, target_size, target_size] or [B, C, target_size, target_size]\n    \"\"\"\n    \n    # process input dimension\n    if img_tensor.dim() == 3:\n        img_tensor = img_tensor.unsqueeze(0)\n        squeeze_output = True\n    else:\n        squeeze_output = False\n    \n    batch_size, channels, height, width = img_tensor.shape\n    \n    # calculate scale (contain function)\n    scale = min(target_size / height, target_size / width)\n    \n    # resize image\n    new_height = int(height * scale)\n    new_width = int(width * scale)\n    \n    img_resized = F.interpolate(\n        img_tensor, \n        size=(new_height, new_width), \n        mode='bilinear',     # bicubic\n        align_corners=False\n    )\n    \n    # calculate padding (pad function)\n    pad_height = target_size - new_height\n    pad_width = target_size - new_width\n    \n    # center padding\n    pad_top = pad_height // 2\n    pad_bottom = pad_height - pad_top\n    pad_left = pad_width // 2\n    pad_right = pad_width - pad_left\n    \n    # apply padding (left, right, top, bottom)\n    img_padded = F.pad(\n        img_resized, \n        (pad_left, pad_right, pad_top, pad_bottom), \n        mode='constant', \n        value=pad_value\n    )\n    \n    if squeeze_output:\n        img_padded = img_padded.squeeze(0)\n    \n    return img_padded\n\n\ndef unpad_image(img_tensor, target_size):\n    \"\"\"\n    torch version of unpad, reverse operation of pad_image\n    \n    Args:\n        img_tensor: torch tensor, shape [C, H, W] or [B, C, H, W], assumed to be square and padded\n        target_size: tuple/list [H, W], target height and width\n    \n    Returns:\n        torch tensor, shape [C, H, W] or [B, C, H, W] with target_size dimensions\n    \"\"\"\n    \n    # process input dimension\n    if img_tensor.dim() == 3:\n        img_tensor = img_tensor.unsqueeze(0)\n        squeeze_output = True\n    else:\n        squeeze_output = False\n    \n    target_height, target_width = target_size\n    max_target = max(target_height, target_width)\n    \n    # first resize to the larger dimension size (square)\n    img_resized = F.interpolate(\n        img_tensor, \n        size=(max_target, max_target), \n        mode='nearest',\n        # align_corners=False\n    )\n    \n    # then crop to target size (center crop)\n    pad_height = max_target - target_height\n    pad_width = max_target - target_width\n    pad_top = pad_height // 2\n    pad_left = pad_width // 2\n    \n    img_cropped = img_resized[\n        :, :,\n        pad_top:pad_top + target_height,\n        pad_left:pad_left + target_width\n    ]\n    \n    if squeeze_output:\n        img_cropped = img_cropped.squeeze(0)\n    \n    return img_cropped\n\n\ndef unpad_uv(uv, original_size, target_height, target_width):\n    \"\"\"\n    transform uv from padded image to unpadded image\n    \n    Args:\n        uv: uv coordinates tensor, shape [batch_size, num_points, 2] or [num_points, 2]\n        original_size: original size of the image (int)\n        target_height: target height of the image (int)\n        target_width: target width of the image (int)\n    \n    Returns:\n        uv_transformed: transformed uv coordinates tensor, shape [batch_size, num_points, 2] or [num_points, 2]\n    \"\"\"\n    # calculate the maximum size of the target\n    max_target = max(target_height, target_width)\n    \n    # first, scale the uv from original_size to max_target\n    scale_factor = max_target / original_size\n    uv_scaled = uv * scale_factor\n    \n    # then, subtract the padding offset\n    pad_left = (max_target - target_width) // 2\n    pad_top = (max_target - target_height) // 2\n    \n    # create the offset tensor, shape [2]\n    offset = torch.tensor([pad_left, pad_top], dtype=uv.dtype, device=uv.device)\n    \n    # broadcast subtraction\n    uv_transformed = uv_scaled - offset\n    uv_transformed[..., 0] = torch.clamp(uv_transformed[..., 0], 0, target_width - 1)   # u\n    uv_transformed[..., 1] = torch.clamp(uv_transformed[..., 1], 0, target_height - 1)  # v\n    return uv_transformed\n\n\ndef log_sinkhorn_iterations(Z: torch.Tensor, log_mu: torch.Tensor, log_nu: torch.Tensor, iters: int) -> torch.Tensor:\n    \"\"\" Perform Sinkhorn Normalization in Log-space for stability\"\"\"\n    u, v = torch.zeros_like(log_mu), torch.zeros_like(log_nu)\n    for _ in range(iters):\n        u = log_mu - torch.logsumexp(Z + v.unsqueeze(1), dim=2)\n        v = log_nu - torch.logsumexp(Z + u.unsqueeze(2), dim=1)\n    return Z + u.unsqueeze(2) + v.unsqueeze(1)\n\n\ndef log_optimal_transport(scores: torch.Tensor, alpha: torch.Tensor, iters: int) -> torch.Tensor:\n    \"\"\" Perform Differentiable Optimal Transport in Log-space for stability\"\"\"\n    b, m, n = scores.shape\n    one = scores.new_tensor(1)\n    ms, ns = (m*one).to(scores), (n*one).to(scores)\n\n    bins0 = alpha.expand(b, m, 1)\n    bins1 = alpha.expand(b, 1, n)\n    alpha = alpha.expand(b, 1, 1)\n\n    couplings = torch.cat([torch.cat([scores, bins0], -1),\n                           torch.cat([bins1, alpha], -1)], 1)\n\n    norm = - (ms + ns).log()\n    log_mu = torch.cat([norm.expand(m), ns.log()[None] + norm])\n    log_nu = torch.cat([norm.expand(n), ms.log()[None] + norm])\n    log_mu, log_nu = log_mu[None].expand(b, -1), log_nu[None].expand(b, -1)\n\n    Z = log_sinkhorn_iterations(couplings, log_mu, log_nu, iters)\n    Z = Z - norm  # multiply probabilities by M+N\n    return Z","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.img_to_arr","uri":"program://Human3R/function/src.dust3r.utils.image.img_to_arr#L28-L31","kind":"function","name":"img_to_arr","path":"src/dust3r/utils/image.py","language":"python","start_line":28,"end_line":31,"context_start_line":8,"context_end_line":51,"code":"import torch\nimport numpy as np\nimport PIL.Image\nfrom PIL.ImageOps import exif_transpose\nimport torchvision.transforms as tvf\n\nos.environ[\"OPENCV_IO_ENABLE_OPENEXR\"] = \"1\"\nimport cv2  # noqa\n\ntry:\n    from pillow_heif import register_heif_opener  # noqa\n\n    register_heif_opener()\n    heif_support_enabled = True\nexcept ImportError:\n    heif_support_enabled = False\n\nImgNorm = tvf.Compose([tvf.ToTensor(), tvf.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])\n\n\ndef img_to_arr(img):\n    if isinstance(img, str):\n        img = imread_cv2(img)\n    return img\n\n\ndef imread_cv2(path, options=cv2.IMREAD_COLOR):\n    \"\"\"Open an image or a depthmap with opencv-python.\"\"\"\n    if path.endswith((\".exr\", \"EXR\")):\n        options = cv2.IMREAD_ANYDEPTH\n    img = cv2.imread(path, options)\n    if img is None:\n        raise IOError(f\"Could not load image={path} with {options=}\")\n    if img.ndim == 3:\n        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n    return img\n\n\ndef rgb(ftensor, true_shape=None):\n    if isinstance(ftensor, list):\n        return [rgb(x, true_shape=true_shape) for x in ftensor]\n    if isinstance(ftensor, torch.Tensor):\n        ftensor = ftensor.detach().cpu().numpy()  # H,W,3\n    if ftensor.ndim == 3 and ftensor.shape[0] == 3:","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.imread_cv2","uri":"program://Human3R/function/src.dust3r.utils.image.imread_cv2#L34-L43","kind":"function","name":"imread_cv2","path":"src/dust3r/utils/image.py","language":"python","start_line":34,"end_line":43,"context_start_line":14,"context_end_line":63,"code":"os.environ[\"OPENCV_IO_ENABLE_OPENEXR\"] = \"1\"\nimport cv2  # noqa\n\ntry:\n    from pillow_heif import register_heif_opener  # noqa\n\n    register_heif_opener()\n    heif_support_enabled = True\nexcept ImportError:\n    heif_support_enabled = False\n\nImgNorm = tvf.Compose([tvf.ToTensor(), tvf.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])\n\n\ndef img_to_arr(img):\n    if isinstance(img, str):\n        img = imread_cv2(img)\n    return img\n\n\ndef imread_cv2(path, options=cv2.IMREAD_COLOR):\n    \"\"\"Open an image or a depthmap with opencv-python.\"\"\"\n    if path.endswith((\".exr\", \"EXR\")):\n        options = cv2.IMREAD_ANYDEPTH\n    img = cv2.imread(path, options)\n    if img is None:\n        raise IOError(f\"Could not load image={path} with {options=}\")\n    if img.ndim == 3:\n        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n    return img\n\n\ndef rgb(ftensor, true_shape=None):\n    if isinstance(ftensor, list):\n        return [rgb(x, true_shape=true_shape) for x in ftensor]\n    if isinstance(ftensor, torch.Tensor):\n        ftensor = ftensor.detach().cpu().numpy()  # H,W,3\n    if ftensor.ndim == 3 and ftensor.shape[0] == 3:\n        ftensor = ftensor.transpose(1, 2, 0)\n    elif ftensor.ndim == 4 and ftensor.shape[1] == 3:\n        ftensor = ftensor.transpose(0, 2, 3, 1)\n    if true_shape is not None:\n        H, W = true_shape\n        ftensor = ftensor[:H, :W]\n    if ftensor.dtype == np.uint8:\n        img = np.float32(ftensor) / 255\n    else:\n        img = (ftensor * 0.5) + 0.5\n    return img.clip(min=0, max=1)\n","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.rgb","uri":"program://Human3R/function/src.dust3r.utils.image.rgb#L46-L62","kind":"function","name":"rgb","path":"src/dust3r/utils/image.py","language":"python","start_line":46,"end_line":62,"context_start_line":26,"context_end_line":82,"code":"\n\ndef img_to_arr(img):\n    if isinstance(img, str):\n        img = imread_cv2(img)\n    return img\n\n\ndef imread_cv2(path, options=cv2.IMREAD_COLOR):\n    \"\"\"Open an image or a depthmap with opencv-python.\"\"\"\n    if path.endswith((\".exr\", \"EXR\")):\n        options = cv2.IMREAD_ANYDEPTH\n    img = cv2.imread(path, options)\n    if img is None:\n        raise IOError(f\"Could not load image={path} with {options=}\")\n    if img.ndim == 3:\n        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n    return img\n\n\ndef rgb(ftensor, true_shape=None):\n    if isinstance(ftensor, list):\n        return [rgb(x, true_shape=true_shape) for x in ftensor]\n    if isinstance(ftensor, torch.Tensor):\n        ftensor = ftensor.detach().cpu().numpy()  # H,W,3\n    if ftensor.ndim == 3 and ftensor.shape[0] == 3:\n        ftensor = ftensor.transpose(1, 2, 0)\n    elif ftensor.ndim == 4 and ftensor.shape[1] == 3:\n        ftensor = ftensor.transpose(0, 2, 3, 1)\n    if true_shape is not None:\n        H, W = true_shape\n        ftensor = ftensor[:H, :W]\n    if ftensor.dtype == np.uint8:\n        img = np.float32(ftensor) / 255\n    else:\n        img = (ftensor * 0.5) + 0.5\n    return img.clip(min=0, max=1)\n\n\ndef _resize_pil_image(img, long_edge_size):\n    S = max(img.size)\n    if S > long_edge_size:\n        interp = PIL.Image.LANCZOS\n    elif S <= long_edge_size:\n        interp = PIL.Image.BICUBIC\n    new_size = tuple(int(round(x * long_edge_size / S)) for x in img.size)\n    return img.resize(new_size, interp)\n\n\ndef load_images(folder_or_list, size, square_ok=False, verbose=True):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image._resize_pil_image","uri":"program://Human3R/function/src.dust3r.utils.image._resize_pil_image#L65-L72","kind":"function","name":"_resize_pil_image","path":"src/dust3r/utils/image.py","language":"python","start_line":65,"end_line":72,"context_start_line":45,"context_end_line":92,"code":"\ndef rgb(ftensor, true_shape=None):\n    if isinstance(ftensor, list):\n        return [rgb(x, true_shape=true_shape) for x in ftensor]\n    if isinstance(ftensor, torch.Tensor):\n        ftensor = ftensor.detach().cpu().numpy()  # H,W,3\n    if ftensor.ndim == 3 and ftensor.shape[0] == 3:\n        ftensor = ftensor.transpose(1, 2, 0)\n    elif ftensor.ndim == 4 and ftensor.shape[1] == 3:\n        ftensor = ftensor.transpose(0, 2, 3, 1)\n    if true_shape is not None:\n        H, W = true_shape\n        ftensor = ftensor[:H, :W]\n    if ftensor.dtype == np.uint8:\n        img = np.float32(ftensor) / 255\n    else:\n        img = (ftensor * 0.5) + 0.5\n    return img.clip(min=0, max=1)\n\n\ndef _resize_pil_image(img, long_edge_size):\n    S = max(img.size)\n    if S > long_edge_size:\n        interp = PIL.Image.LANCZOS\n    elif S <= long_edge_size:\n        interp = PIL.Image.BICUBIC\n    new_size = tuple(int(round(x * long_edge_size / S)) for x in img.size)\n    return img.resize(new_size, interp)\n\n\ndef load_images(folder_or_list, size, square_ok=False, verbose=True):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        if verbose:\n            print(f\">> Loading a list of {len(folder_or_list)} images\")\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\", \".bmp\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.load_images","uri":"program://Human3R/function/src.dust3r.utils.image.load_images#L75-L133","kind":"function","name":"load_images","path":"src/dust3r/utils/image.py","language":"python","start_line":75,"end_line":133,"context_start_line":55,"context_end_line":153,"code":"    if true_shape is not None:\n        H, W = true_shape\n        ftensor = ftensor[:H, :W]\n    if ftensor.dtype == np.uint8:\n        img = np.float32(ftensor) / 255\n    else:\n        img = (ftensor * 0.5) + 0.5\n    return img.clip(min=0, max=1)\n\n\ndef _resize_pil_image(img, long_edge_size):\n    S = max(img.size)\n    if S > long_edge_size:\n        interp = PIL.Image.LANCZOS\n    elif S <= long_edge_size:\n        interp = PIL.Image.BICUBIC\n    new_size = tuple(int(round(x * long_edge_size / S)) for x in img.size)\n    return img.resize(new_size, interp)\n\n\ndef load_images(folder_or_list, size, square_ok=False, verbose=True):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        if verbose:\n            print(f\">> Loading a list of {len(folder_or_list)} images\")\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\", \".bmp\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]\n    supported_images_extensions = tuple(supported_images_extensions)\n\n    imgs = []\n    for path in folder_content:\n        if not path.lower().endswith(supported_images_extensions):\n            continue\n        img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert(\"RGB\")\n        W1, H1 = img.size\n        if size == 224:\n\n            img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1)))\n        else:\n\n            img = _resize_pil_image(img, size)\n        W, H = img.size\n        cx, cy = W // 2, H // 2\n        if size == 224:\n            half = min(cx, cy)\n            img = img.crop((cx - half, cy - half, cx + half, cy + half))\n        else:\n            halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8\n            if not (square_ok) and W == H:\n                halfh = 3 * halfw / 4\n            img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n\n        W2, H2 = img.size\n        if verbose:\n            print(f\" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}\")\n        imgs.append(\n            dict(\n                img=ImgNorm(img)[None],\n                true_shape=np.int32([img.size[::-1]]),\n                idx=len(imgs),\n                instance=str(len(imgs)),\n            )\n        )\n\n    assert imgs, \"no images foud at \" + root\n    if verbose:\n        print(f\" (Found {len(imgs)} images)\")\n    return imgs\n\n\ndef load_images_for_eval(\n    folder_or_list, size, square_ok=False, verbose=True, crop=True\n):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        if verbose:\n            print(f\">> Loading a list of {len(folder_or_list)} images\")\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\"]","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.load_images_for_eval","uri":"program://Human3R/function/src.dust3r.utils.image.load_images_for_eval#L136-L202","kind":"function","name":"load_images_for_eval","path":"src/dust3r/utils/image.py","language":"python","start_line":136,"end_line":202,"context_start_line":116,"context_end_line":222,"code":"            img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n\n        W2, H2 = img.size\n        if verbose:\n            print(f\" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}\")\n        imgs.append(\n            dict(\n                img=ImgNorm(img)[None],\n                true_shape=np.int32([img.size[::-1]]),\n                idx=len(imgs),\n                instance=str(len(imgs)),\n            )\n        )\n\n    assert imgs, \"no images foud at \" + root\n    if verbose:\n        print(f\" (Found {len(imgs)} images)\")\n    return imgs\n\n\ndef load_images_for_eval(\n    folder_or_list, size, square_ok=False, verbose=True, crop=True\n):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        if verbose:\n            print(f\">> Loading a list of {len(folder_or_list)} images\")\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]\n    supported_images_extensions = tuple(supported_images_extensions)\n\n    imgs = []\n    for path in folder_content:\n        if not path.lower().endswith(supported_images_extensions):\n            continue\n        img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert(\"RGB\")\n        W1, H1 = img.size\n        if size == 224:\n            # resize short side to 224 (then crop)\n            img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1)))\n        else:\n            # resize long side to 512\n            img = _resize_pil_image(img, size)\n        W, H = img.size\n        cx, cy = W // 2, H // 2\n        if size == 224:\n            half = min(cx, cy)\n            if crop:\n                img = img.crop((cx - half, cy - half, cx + half, cy + half))\n            else:  # resize\n                img = img.resize((2 * half, 2 * half), PIL.Image.LANCZOS)\n        else:\n            halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8\n            if not (square_ok) and W == H:\n                halfh = 3 * halfw / 4\n            if crop:\n                img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n            else:  # resize\n                img = img.resize((2 * halfw, 2 * halfh), PIL.Image.LANCZOS)\n        W2, H2 = img.size\n        if verbose:\n            print(f\" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}\")\n        imgs.append(\n            dict(\n                img=ImgNorm(img)[None],\n                true_shape=np.int32([img.size[::-1]]),\n                idx=len(imgs),\n                instance=str(len(imgs)),\n                ori_shape=np.int32([[H1, W1]])\n            )\n        )\n\n    assert imgs, \"no images foud at \" + root\n    if verbose:\n        print(f\" (Found {len(imgs)} images)\")\n    return imgs\n\n\n\ndef load_masks_for_eval(\n    folder_or_list, size, square_ok=False, verbose=True, crop=True\n):\n    if isinstance(folder_or_list, str):\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]\n    supported_images_extensions = tuple(supported_images_extensions)\n","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.load_masks_for_eval","uri":"program://Human3R/function/src.dust3r.utils.image.load_masks_for_eval#L206-L257","kind":"function","name":"load_masks_for_eval","path":"src/dust3r/utils/image.py","language":"python","start_line":206,"end_line":257,"context_start_line":186,"context_end_line":277,"code":"        W2, H2 = img.size\n        if verbose:\n            print(f\" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}\")\n        imgs.append(\n            dict(\n                img=ImgNorm(img)[None],\n                true_shape=np.int32([img.size[::-1]]),\n                idx=len(imgs),\n                instance=str(len(imgs)),\n                ori_shape=np.int32([[H1, W1]])\n            )\n        )\n\n    assert imgs, \"no images foud at \" + root\n    if verbose:\n        print(f\" (Found {len(imgs)} images)\")\n    return imgs\n\n\n\ndef load_masks_for_eval(\n    folder_or_list, size, square_ok=False, verbose=True, crop=True\n):\n    if isinstance(folder_or_list, str):\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]\n    supported_images_extensions = tuple(supported_images_extensions)\n\n    imgs = []\n    for path in folder_content:\n        if not path.lower().endswith(supported_images_extensions):\n            continue\n        img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert(\"RGB\")\n        W1, H1 = img.size\n        if size == 224:\n            # resize short side to 224 (then crop)\n            img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1)))\n        else:\n            # resize long side to 512\n            img = _resize_pil_image(img, size)\n        W, H = img.size\n        cx, cy = W // 2, H // 2\n        if size == 224:\n            half = min(cx, cy)\n            if crop:\n                img = img.crop((cx - half, cy - half, cx + half, cy + half))\n            else:  # resize\n                img = img.resize((2 * half, 2 * half), PIL.Image.LANCZOS)\n        else:\n            halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8\n            if not (square_ok) and W == H:\n                halfh = 3 * halfw / 4\n            if crop:\n                img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n            else:  # resize\n                img = img.resize((2 * halfw, 2 * halfh), PIL.Image.LANCZOS)\n\n        img = tvf.ToTensor()(img)\n        img = (img[0] > 0.5).float()\n        imgs.append(img[None])\n\n    assert imgs, \"no images foud at \" + root\n    return imgs\n\n\ndef load_images_512(folder_or_list, size, square_ok=False, verbose=True):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        if verbose:\n            print(f\">> Loading a list of {len(folder_or_list)} images\")\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\", \".bmp\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.load_images_512","uri":"program://Human3R/function/src.dust3r.utils.image.load_images_512#L260-L319","kind":"function","name":"load_images_512","path":"src/dust3r/utils/image.py","language":"python","start_line":260,"end_line":319,"context_start_line":240,"context_end_line":339,"code":"                img = img.crop((cx - half, cy - half, cx + half, cy + half))\n            else:  # resize\n                img = img.resize((2 * half, 2 * half), PIL.Image.LANCZOS)\n        else:\n            halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8\n            if not (square_ok) and W == H:\n                halfh = 3 * halfw / 4\n            if crop:\n                img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n            else:  # resize\n                img = img.resize((2 * halfw, 2 * halfh), PIL.Image.LANCZOS)\n\n        img = tvf.ToTensor()(img)\n        img = (img[0] > 0.5).float()\n        imgs.append(img[None])\n\n    assert imgs, \"no images foud at \" + root\n    return imgs\n\n\ndef load_images_512(folder_or_list, size, square_ok=False, verbose=True):\n    \"\"\"open and convert all images in a list or folder to proper input format for DUSt3R\"\"\"\n    if isinstance(folder_or_list, str):\n        if verbose:\n            print(f\">> Loading images from {folder_or_list}\")\n        root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list))\n\n    elif isinstance(folder_or_list, list):\n        if verbose:\n            print(f\">> Loading a list of {len(folder_or_list)} images\")\n        root, folder_content = \"\", folder_or_list\n\n    else:\n        raise ValueError(f\"bad {folder_or_list=} ({type(folder_or_list)})\")\n\n    supported_images_extensions = [\".jpg\", \".jpeg\", \".png\", \".bmp\"]\n    if heif_support_enabled:\n        supported_images_extensions += [\".heic\", \".heif\"]\n    supported_images_extensions = tuple(supported_images_extensions)\n\n    imgs = []\n    for path in folder_content:\n        if not path.lower().endswith(supported_images_extensions):\n            continue\n        img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert(\"RGB\")\n        img = img.resize((512, 384))\n        W1, H1 = img.size\n        if size == 224:\n\n            img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1)))\n        else:\n\n            img = _resize_pil_image(img, size)\n        W, H = img.size\n        cx, cy = W // 2, H // 2\n        if size == 224:\n            half = min(cx, cy)\n            img = img.crop((cx - half, cy - half, cx + half, cy + half))\n        else:\n            halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8\n            if not (square_ok) and W == H:\n                halfh = 3 * halfw / 4\n            img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh))\n\n        W2, H2 = img.size\n        if verbose:\n            print(f\" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}\")\n        imgs.append(\n            dict(\n                img=ImgNorm(img)[None],\n                true_shape=np.int32([img.size[::-1]]),\n                idx=len(imgs),\n                instance=str(len(imgs)),\n            )\n        )\n\n    assert imgs, \"no images foud at \" + root\n    if verbose:\n        print(f\" (Found {len(imgs)} images)\")\n    return imgs\n\n\nIMG_NORM_MEAN = [0.5, 0.5, 0.5]\nIMG_NORM_STD = [0.5, 0.5, 0.5]\n\n\ndef normalize_rgb(img, imagenet_normalization=True):\n    \"\"\"\n    Args:\n        - img: np.array - (W,H,3) - np.uint8 - 0/255\n    Return:\n        - img: np.array - (3,W,H) - np.float - -3/3\n    \"\"\"\n    img = img.astype(np.float32) / 255.\n    img = np.transpose(img, (2,0,1))\n    if imagenet_normalization:\n        img = (img - np.asarray(IMG_NORM_MEAN).reshape(3,1,1)) / np.asarray(IMG_NORM_STD).reshape(3,1,1)\n    img = img.astype(np.float32)\n    return img\n","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.normalize_rgb","uri":"program://Human3R/function/src.dust3r.utils.image.normalize_rgb#L326-L338","kind":"function","name":"normalize_rgb","path":"src/dust3r/utils/image.py","language":"python","start_line":326,"end_line":338,"context_start_line":306,"context_end_line":358,"code":"            print(f\" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}\")\n        imgs.append(\n            dict(\n                img=ImgNorm(img)[None],\n                true_shape=np.int32([img.size[::-1]]),\n                idx=len(imgs),\n                instance=str(len(imgs)),\n            )\n        )\n\n    assert imgs, \"no images foud at \" + root\n    if verbose:\n        print(f\" (Found {len(imgs)} images)\")\n    return imgs\n\n\nIMG_NORM_MEAN = [0.5, 0.5, 0.5]\nIMG_NORM_STD = [0.5, 0.5, 0.5]\n\n\ndef normalize_rgb(img, imagenet_normalization=True):\n    \"\"\"\n    Args:\n        - img: np.array - (W,H,3) - np.uint8 - 0/255\n    Return:\n        - img: np.array - (3,W,H) - np.float - -3/3\n    \"\"\"\n    img = img.astype(np.float32) / 255.\n    img = np.transpose(img, (2,0,1))\n    if imagenet_normalization:\n        img = (img - np.asarray(IMG_NORM_MEAN).reshape(3,1,1)) / np.asarray(IMG_NORM_STD).reshape(3,1,1)\n    img = img.astype(np.float32)\n    return img\n\n\ndef denormalize_rgb(img, imagenet_normalization=True):\n    \"\"\"\n    Args:\n        - img: np.array - (3,W,H) - np.float - -3/3\n    Return:\n        - img: np.array - (W,H,3) - np.uint8 - 0/255\n    \"\"\"\n    if imagenet_normalization:\n        img = (img * np.asarray(IMG_NORM_STD).reshape(3,1,1)) + np.asarray(IMG_NORM_MEAN).reshape(3,1,1)\n    img = np.transpose(img, (1,2,0)) * 255.\n    img = img.astype(np.uint8)\n    return img\n\n\nimport torch.nn.functional as F\n\ndef pad_image(img_tensor, target_size, pad_value=-1.0):\n    \"\"\"","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.denormalize_rgb","uri":"program://Human3R/function/src.dust3r.utils.image.denormalize_rgb#L341-L352","kind":"function","name":"denormalize_rgb","path":"src/dust3r/utils/image.py","language":"python","start_line":341,"end_line":352,"context_start_line":321,"context_end_line":372,"code":"\nIMG_NORM_MEAN = [0.5, 0.5, 0.5]\nIMG_NORM_STD = [0.5, 0.5, 0.5]\n\n\ndef normalize_rgb(img, imagenet_normalization=True):\n    \"\"\"\n    Args:\n        - img: np.array - (W,H,3) - np.uint8 - 0/255\n    Return:\n        - img: np.array - (3,W,H) - np.float - -3/3\n    \"\"\"\n    img = img.astype(np.float32) / 255.\n    img = np.transpose(img, (2,0,1))\n    if imagenet_normalization:\n        img = (img - np.asarray(IMG_NORM_MEAN).reshape(3,1,1)) / np.asarray(IMG_NORM_STD).reshape(3,1,1)\n    img = img.astype(np.float32)\n    return img\n\n\ndef denormalize_rgb(img, imagenet_normalization=True):\n    \"\"\"\n    Args:\n        - img: np.array - (3,W,H) - np.float - -3/3\n    Return:\n        - img: np.array - (W,H,3) - np.uint8 - 0/255\n    \"\"\"\n    if imagenet_normalization:\n        img = (img * np.asarray(IMG_NORM_STD).reshape(3,1,1)) + np.asarray(IMG_NORM_MEAN).reshape(3,1,1)\n    img = np.transpose(img, (1,2,0)) * 255.\n    img = img.astype(np.uint8)\n    return img\n\n\nimport torch.nn.functional as F\n\ndef pad_image(img_tensor, target_size, pad_value=-1.0):\n    \"\"\"\n    torch version of ImageOps.pad, equivalent to the combination of contain and pad\n    \n    Args:\n        img_tensor: torch tensor, shape [C, H, W] or [B, C, H, W]\n        target_size: int, target size (square)\n    \n    Returns:\n        torch tensor, shape [C, target_size, target_size] or [B, C, target_size, target_size]\n    \"\"\"\n    \n    # process input dimension\n    if img_tensor.dim() == 3:\n        img_tensor = img_tensor.unsqueeze(0)\n        squeeze_output = True","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.pad_image","uri":"program://Human3R/function/src.dust3r.utils.image.pad_image#L357-L413","kind":"function","name":"pad_image","path":"src/dust3r/utils/image.py","language":"python","start_line":357,"end_line":413,"context_start_line":337,"context_end_line":433,"code":"    img = img.astype(np.float32)\n    return img\n\n\ndef denormalize_rgb(img, imagenet_normalization=True):\n    \"\"\"\n    Args:\n        - img: np.array - (3,W,H) - np.float - -3/3\n    Return:\n        - img: np.array - (W,H,3) - np.uint8 - 0/255\n    \"\"\"\n    if imagenet_normalization:\n        img = (img * np.asarray(IMG_NORM_STD).reshape(3,1,1)) + np.asarray(IMG_NORM_MEAN).reshape(3,1,1)\n    img = np.transpose(img, (1,2,0)) * 255.\n    img = img.astype(np.uint8)\n    return img\n\n\nimport torch.nn.functional as F\n\ndef pad_image(img_tensor, target_size, pad_value=-1.0):\n    \"\"\"\n    torch version of ImageOps.pad, equivalent to the combination of contain and pad\n    \n    Args:\n        img_tensor: torch tensor, shape [C, H, W] or [B, C, H, W]\n        target_size: int, target size (square)\n    \n    Returns:\n        torch tensor, shape [C, target_size, target_size] or [B, C, target_size, target_size]\n    \"\"\"\n    \n    # process input dimension\n    if img_tensor.dim() == 3:\n        img_tensor = img_tensor.unsqueeze(0)\n        squeeze_output = True\n    else:\n        squeeze_output = False\n    \n    batch_size, channels, height, width = img_tensor.shape\n    \n    # calculate scale (contain function)\n    scale = min(target_size / height, target_size / width)\n    \n    # resize image\n    new_height = int(height * scale)\n    new_width = int(width * scale)\n    \n    img_resized = F.interpolate(\n        img_tensor, \n        size=(new_height, new_width), \n        mode='bilinear',     # bicubic\n        align_corners=False\n    )\n    \n    # calculate padding (pad function)\n    pad_height = target_size - new_height\n    pad_width = target_size - new_width\n    \n    # center padding\n    pad_top = pad_height // 2\n    pad_bottom = pad_height - pad_top\n    pad_left = pad_width // 2\n    pad_right = pad_width - pad_left\n    \n    # apply padding (left, right, top, bottom)\n    img_padded = F.pad(\n        img_resized, \n        (pad_left, pad_right, pad_top, pad_bottom), \n        mode='constant', \n        value=pad_value\n    )\n    \n    if squeeze_output:\n        img_padded = img_padded.squeeze(0)\n    \n    return img_padded\n\n\ndef unpad_image(img_tensor, target_size):\n    \"\"\"\n    torch version of unpad, reverse operation of pad_image\n    \n    Args:\n        img_tensor: torch tensor, shape [C, H, W] or [B, C, H, W], assumed to be square and padded\n        target_size: tuple/list [H, W], target height and width\n    \n    Returns:\n        torch tensor, shape [C, H, W] or [B, C, H, W] with target_size dimensions\n    \"\"\"\n    \n    # process input dimension\n    if img_tensor.dim() == 3:\n        img_tensor = img_tensor.unsqueeze(0)\n        squeeze_output = True\n    else:\n        squeeze_output = False","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.unpad_image","uri":"program://Human3R/function/src.dust3r.utils.image.unpad_image#L416-L461","kind":"function","name":"unpad_image","path":"src/dust3r/utils/image.py","language":"python","start_line":416,"end_line":461,"context_start_line":396,"context_end_line":481,"code":"    # center padding\n    pad_top = pad_height // 2\n    pad_bottom = pad_height - pad_top\n    pad_left = pad_width // 2\n    pad_right = pad_width - pad_left\n    \n    # apply padding (left, right, top, bottom)\n    img_padded = F.pad(\n        img_resized, \n        (pad_left, pad_right, pad_top, pad_bottom), \n        mode='constant', \n        value=pad_value\n    )\n    \n    if squeeze_output:\n        img_padded = img_padded.squeeze(0)\n    \n    return img_padded\n\n\ndef unpad_image(img_tensor, target_size):\n    \"\"\"\n    torch version of unpad, reverse operation of pad_image\n    \n    Args:\n        img_tensor: torch tensor, shape [C, H, W] or [B, C, H, W], assumed to be square and padded\n        target_size: tuple/list [H, W], target height and width\n    \n    Returns:\n        torch tensor, shape [C, H, W] or [B, C, H, W] with target_size dimensions\n    \"\"\"\n    \n    # process input dimension\n    if img_tensor.dim() == 3:\n        img_tensor = img_tensor.unsqueeze(0)\n        squeeze_output = True\n    else:\n        squeeze_output = False\n    \n    target_height, target_width = target_size\n    max_target = max(target_height, target_width)\n    \n    # first resize to the larger dimension size (square)\n    img_resized = F.interpolate(\n        img_tensor, \n        size=(max_target, max_target), \n        mode='nearest',\n        # align_corners=False\n    )\n    \n    # then crop to target size (center crop)\n    pad_height = max_target - target_height\n    pad_width = max_target - target_width\n    pad_top = pad_height // 2\n    pad_left = pad_width // 2\n    \n    img_cropped = img_resized[\n        :, :,\n        pad_top:pad_top + target_height,\n        pad_left:pad_left + target_width\n    ]\n    \n    if squeeze_output:\n        img_cropped = img_cropped.squeeze(0)\n    \n    return img_cropped\n\n\ndef unpad_uv(uv, original_size, target_height, target_width):\n    \"\"\"\n    transform uv from padded image to unpadded image\n    \n    Args:\n        uv: uv coordinates tensor, shape [batch_size, num_points, 2] or [num_points, 2]\n        original_size: original size of the image (int)\n        target_height: target height of the image (int)\n        target_width: target width of the image (int)\n    \n    Returns:\n        uv_transformed: transformed uv coordinates tensor, shape [batch_size, num_points, 2] or [num_points, 2]\n    \"\"\"\n    # calculate the maximum size of the target\n    max_target = max(target_height, target_width)\n    \n    # first, scale the uv from original_size to max_target\n    scale_factor = max_target / original_size","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.unpad_uv","uri":"program://Human3R/function/src.dust3r.utils.image.unpad_uv#L464-L495","kind":"function","name":"unpad_uv","path":"src/dust3r/utils/image.py","language":"python","start_line":464,"end_line":495,"context_start_line":444,"context_end_line":515,"code":"    )\n    \n    # then crop to target size (center crop)\n    pad_height = max_target - target_height\n    pad_width = max_target - target_width\n    pad_top = pad_height // 2\n    pad_left = pad_width // 2\n    \n    img_cropped = img_resized[\n        :, :,\n        pad_top:pad_top + target_height,\n        pad_left:pad_left + target_width\n    ]\n    \n    if squeeze_output:\n        img_cropped = img_cropped.squeeze(0)\n    \n    return img_cropped\n\n\ndef unpad_uv(uv, original_size, target_height, target_width):\n    \"\"\"\n    transform uv from padded image to unpadded image\n    \n    Args:\n        uv: uv coordinates tensor, shape [batch_size, num_points, 2] or [num_points, 2]\n        original_size: original size of the image (int)\n        target_height: target height of the image (int)\n        target_width: target width of the image (int)\n    \n    Returns:\n        uv_transformed: transformed uv coordinates tensor, shape [batch_size, num_points, 2] or [num_points, 2]\n    \"\"\"\n    # calculate the maximum size of the target\n    max_target = max(target_height, target_width)\n    \n    # first, scale the uv from original_size to max_target\n    scale_factor = max_target / original_size\n    uv_scaled = uv * scale_factor\n    \n    # then, subtract the padding offset\n    pad_left = (max_target - target_width) // 2\n    pad_top = (max_target - target_height) // 2\n    \n    # create the offset tensor, shape [2]\n    offset = torch.tensor([pad_left, pad_top], dtype=uv.dtype, device=uv.device)\n    \n    # broadcast subtraction\n    uv_transformed = uv_scaled - offset\n    uv_transformed[..., 0] = torch.clamp(uv_transformed[..., 0], 0, target_width - 1)   # u\n    uv_transformed[..., 1] = torch.clamp(uv_transformed[..., 1], 0, target_height - 1)  # v\n    return uv_transformed\n\n\ndef log_sinkhorn_iterations(Z: torch.Tensor, log_mu: torch.Tensor, log_nu: torch.Tensor, iters: int) -> torch.Tensor:\n    \"\"\" Perform Sinkhorn Normalization in Log-space for stability\"\"\"\n    u, v = torch.zeros_like(log_mu), torch.zeros_like(log_nu)\n    for _ in range(iters):\n        u = log_mu - torch.logsumexp(Z + v.unsqueeze(1), dim=2)\n        v = log_nu - torch.logsumexp(Z + u.unsqueeze(2), dim=1)\n    return Z + u.unsqueeze(2) + v.unsqueeze(1)\n\n\ndef log_optimal_transport(scores: torch.Tensor, alpha: torch.Tensor, iters: int) -> torch.Tensor:\n    \"\"\" Perform Differentiable Optimal Transport in Log-space for stability\"\"\"\n    b, m, n = scores.shape\n    one = scores.new_tensor(1)\n    ms, ns = (m*one).to(scores), (n*one).to(scores)\n\n    bins0 = alpha.expand(b, m, 1)\n    bins1 = alpha.expand(b, 1, n)\n    alpha = alpha.expand(b, 1, 1)","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.log_sinkhorn_iterations","uri":"program://Human3R/function/src.dust3r.utils.image.log_sinkhorn_iterations#L498-L504","kind":"function","name":"log_sinkhorn_iterations","path":"src/dust3r/utils/image.py","language":"python","start_line":498,"end_line":504,"context_start_line":478,"context_end_line":524,"code":"    max_target = max(target_height, target_width)\n    \n    # first, scale the uv from original_size to max_target\n    scale_factor = max_target / original_size\n    uv_scaled = uv * scale_factor\n    \n    # then, subtract the padding offset\n    pad_left = (max_target - target_width) // 2\n    pad_top = (max_target - target_height) // 2\n    \n    # create the offset tensor, shape [2]\n    offset = torch.tensor([pad_left, pad_top], dtype=uv.dtype, device=uv.device)\n    \n    # broadcast subtraction\n    uv_transformed = uv_scaled - offset\n    uv_transformed[..., 0] = torch.clamp(uv_transformed[..., 0], 0, target_width - 1)   # u\n    uv_transformed[..., 1] = torch.clamp(uv_transformed[..., 1], 0, target_height - 1)  # v\n    return uv_transformed\n\n\ndef log_sinkhorn_iterations(Z: torch.Tensor, log_mu: torch.Tensor, log_nu: torch.Tensor, iters: int) -> torch.Tensor:\n    \"\"\" Perform Sinkhorn Normalization in Log-space for stability\"\"\"\n    u, v = torch.zeros_like(log_mu), torch.zeros_like(log_nu)\n    for _ in range(iters):\n        u = log_mu - torch.logsumexp(Z + v.unsqueeze(1), dim=2)\n        v = log_nu - torch.logsumexp(Z + u.unsqueeze(2), dim=1)\n    return Z + u.unsqueeze(2) + v.unsqueeze(1)\n\n\ndef log_optimal_transport(scores: torch.Tensor, alpha: torch.Tensor, iters: int) -> torch.Tensor:\n    \"\"\" Perform Differentiable Optimal Transport in Log-space for stability\"\"\"\n    b, m, n = scores.shape\n    one = scores.new_tensor(1)\n    ms, ns = (m*one).to(scores), (n*one).to(scores)\n\n    bins0 = alpha.expand(b, m, 1)\n    bins1 = alpha.expand(b, 1, n)\n    alpha = alpha.expand(b, 1, 1)\n\n    couplings = torch.cat([torch.cat([scores, bins0], -1),\n                           torch.cat([bins1, alpha], -1)], 1)\n\n    norm = - (ms + ns).log()\n    log_mu = torch.cat([norm.expand(m), ns.log()[None] + norm])\n    log_nu = torch.cat([norm.expand(n), ms.log()[None] + norm])\n    log_mu, log_nu = log_mu[None].expand(b, -1), log_nu[None].expand(b, -1)\n","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.utils.image.log_optimal_transport","uri":"program://Human3R/function/src.dust3r.utils.image.log_optimal_transport#L507-L527","kind":"function","name":"log_optimal_transport","path":"src/dust3r/utils/image.py","language":"python","start_line":507,"end_line":527,"context_start_line":487,"context_end_line":527,"code":"    \n    # create the offset tensor, shape [2]\n    offset = torch.tensor([pad_left, pad_top], dtype=uv.dtype, device=uv.device)\n    \n    # broadcast subtraction\n    uv_transformed = uv_scaled - offset\n    uv_transformed[..., 0] = torch.clamp(uv_transformed[..., 0], 0, target_width - 1)   # u\n    uv_transformed[..., 1] = torch.clamp(uv_transformed[..., 1], 0, target_height - 1)  # v\n    return uv_transformed\n\n\ndef log_sinkhorn_iterations(Z: torch.Tensor, log_mu: torch.Tensor, log_nu: torch.Tensor, iters: int) -> torch.Tensor:\n    \"\"\" Perform Sinkhorn Normalization in Log-space for stability\"\"\"\n    u, v = torch.zeros_like(log_mu), torch.zeros_like(log_nu)\n    for _ in range(iters):\n        u = log_mu - torch.logsumexp(Z + v.unsqueeze(1), dim=2)\n        v = log_nu - torch.logsumexp(Z + u.unsqueeze(2), dim=1)\n    return Z + u.unsqueeze(2) + v.unsqueeze(1)\n\n\ndef log_optimal_transport(scores: torch.Tensor, alpha: torch.Tensor, iters: int) -> torch.Tensor:\n    \"\"\" Perform Differentiable Optimal Transport in Log-space for stability\"\"\"\n    b, m, n = scores.shape\n    one = scores.new_tensor(1)\n    ms, ns = (m*one).to(scores), (n*one).to(scores)\n\n    bins0 = alpha.expand(b, m, 1)\n    bins1 = alpha.expand(b, 1, n)\n    alpha = alpha.expand(b, 1, 1)\n\n    couplings = torch.cat([torch.cat([scores, bins0], -1),\n                           torch.cat([bins1, alpha], -1)], 1)\n\n    norm = - (ms + ns).log()\n    log_mu = torch.cat([norm.expand(m), ns.log()[None] + norm])\n    log_nu = torch.cat([norm.expand(n), ms.log()[None] + norm])\n    log_mu, log_nu = log_mu[None].expand(b, -1), log_nu[None].expand(b, -1)\n\n    Z = log_sinkhorn_iterations(couplings, log_mu, log_nu, iters)\n    Z = Z - norm  # multiply probabilities by M+N\n    return Z","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvimgnet","uri":"program://Human3R/module/src.dust3r.datasets.mvimgnet#L1-L145","kind":"module","name":"src.dust3r.datasets.mvimgnet","path":"src/dust3r/datasets/mvimgnet.py","language":"python","start_line":1,"end_line":145,"context_start_line":1,"context_end_line":145,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVImgNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 32\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n\n            num_imgs = len(basenames)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []\n            for view_idx in image_idxs:\n                scene_id = self.sceneids[view_idx]\n                scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n                rgb_dir = osp.join(scene_dir, \"rgb\")\n                cam_dir = osp.join(scene_dir, \"cam\")\n\n                basename = self.images[view_idx]\n\n                try:\n                    # Load RGB image\n                    rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n                    # Load depthmap, no depth, set to all ones\n                    depthmap = np.ones_like(rgb_image[..., 0], dtype=np.float32)\n                    cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n                    camera_pose = cam[\"pose\"]\n                    intrinsics = np.eye(3)\n                    intrinsics[0, 0] = cam[\"intrinsics\"][0, 0]\n                    intrinsics[1, 1] = cam[\"intrinsics\"][0, 0]\n                    intrinsics[0, 2] = cam[\"intrinsics\"][1, 1]\n                    intrinsics[1, 2] = cam[\"intrinsics\"][0, 2]\n                except:\n                    print(f\"Error loading {scene} {basename}, skipping\")\n                    self.invalid_scenes[scene] = True\n                    break\n\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap.astype(np.float32),\n                        camera_pose=camera_pose.astype(np.float32),\n                        camera_intrinsics=intrinsics.astype(np.float32),\n                        dataset=\"MVImgnet\",\n                        label=self.scenes[scene_id] + \"_\" + basename,\n                        instance=f\"{str(idx)}_{str(view_idx)}\",\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.98, dtype=np.float32),\n                        img_mask=True,\n                        ray_mask=False,\n                        camera_only=True,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n            if len(views) == num_views:\n                invalid_seq = False\n        return views","source_hash":"a75252d8106f05fe95cf409028aba31d93c3a7b1dce68b0b3d4bbdd6c0833c95","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvimgnet.MVImgNet_Multi","uri":"program://Human3R/class/src.dust3r.datasets.mvimgnet.MVImgNet_Multi#L14-L145","kind":"class","name":"MVImgNet_Multi","path":"src/dust3r/datasets/mvimgnet.py","language":"python","start_line":14,"end_line":145,"context_start_line":1,"context_end_line":145,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVImgNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 32\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n\n            num_imgs = len(basenames)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []\n            for view_idx in image_idxs:\n                scene_id = self.sceneids[view_idx]\n                scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n                rgb_dir = osp.join(scene_dir, \"rgb\")\n                cam_dir = osp.join(scene_dir, \"cam\")\n\n                basename = self.images[view_idx]\n\n                try:\n                    # Load RGB image\n                    rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n                    # Load depthmap, no depth, set to all ones\n                    depthmap = np.ones_like(rgb_image[..., 0], dtype=np.float32)\n                    cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n                    camera_pose = cam[\"pose\"]\n                    intrinsics = np.eye(3)\n                    intrinsics[0, 0] = cam[\"intrinsics\"][0, 0]\n                    intrinsics[1, 1] = cam[\"intrinsics\"][0, 0]\n                    intrinsics[0, 2] = cam[\"intrinsics\"][1, 1]\n                    intrinsics[1, 2] = cam[\"intrinsics\"][0, 2]\n                except:\n                    print(f\"Error loading {scene} {basename}, skipping\")\n                    self.invalid_scenes[scene] = True\n                    break\n\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap.astype(np.float32),\n                        camera_pose=camera_pose.astype(np.float32),\n                        camera_intrinsics=intrinsics.astype(np.float32),\n                        dataset=\"MVImgnet\",\n                        label=self.scenes[scene_id] + \"_\" + basename,\n                        instance=f\"{str(idx)}_{str(view_idx)}\",\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.98, dtype=np.float32),\n                        img_mask=True,\n                        ray_mask=False,\n                        camera_only=True,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n            if len(views) == num_views:\n                invalid_seq = False\n        return views","source_hash":"a75252d8106f05fe95cf409028aba31d93c3a7b1dce68b0b3d4bbdd6c0833c95","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvimgnet.__init__","uri":"program://Human3R/function/src.dust3r.datasets.mvimgnet.__init__#L15-L22","kind":"function","name":"__init__","path":"src/dust3r/datasets/mvimgnet.py","language":"python","start_line":15,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVImgNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 32\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n\n            num_imgs = len(basenames)","source_hash":"a75252d8106f05fe95cf409028aba31d93c3a7b1dce68b0b3d4bbdd6c0833c95","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvimgnet._load_data","uri":"program://Human3R/function/src.dust3r.datasets.mvimgnet._load_data#L24-L70","kind":"function","name":"_load_data","path":"src/dust3r/datasets/mvimgnet.py","language":"python","start_line":24,"end_line":70,"context_start_line":4,"context_end_line":90,"code":"import itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVImgNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 32\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n\n            num_imgs = len(basenames)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )","source_hash":"a75252d8106f05fe95cf409028aba31d93c3a7b1dce68b0b3d4bbdd6c0833c95","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvimgnet.__len__","uri":"program://Human3R/function/src.dust3r.datasets.mvimgnet.__len__#L72-L73","kind":"function","name":"__len__","path":"src/dust3r/datasets/mvimgnet.py","language":"python","start_line":72,"end_line":73,"context_start_line":52,"context_end_line":93,"code":"            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []","source_hash":"a75252d8106f05fe95cf409028aba31d93c3a7b1dce68b0b3d4bbdd6c0833c95","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvimgnet.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.mvimgnet.get_image_num#L75-L76","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/mvimgnet.py","language":"python","start_line":75,"end_line":76,"context_start_line":55,"context_end_line":96,"code":"            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []\n            for view_idx in image_idxs:\n                scene_id = self.sceneids[view_idx]\n                scene_dir = osp.join(self.ROOT, self.scenes[scene_id])","source_hash":"a75252d8106f05fe95cf409028aba31d93c3a7b1dce68b0b3d4bbdd6c0833c95","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvimgnet._get_views","uri":"program://Human3R/function/src.dust3r.datasets.mvimgnet._get_views#L78-L145","kind":"function","name":"_get_views","path":"src/dust3r/datasets/mvimgnet.py","language":"python","start_line":78,"end_line":145,"context_start_line":58,"context_end_line":145,"code":"            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []\n            for view_idx in image_idxs:\n                scene_id = self.sceneids[view_idx]\n                scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n                rgb_dir = osp.join(scene_dir, \"rgb\")\n                cam_dir = osp.join(scene_dir, \"cam\")\n\n                basename = self.images[view_idx]\n\n                try:\n                    # Load RGB image\n                    rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n                    # Load depthmap, no depth, set to all ones\n                    depthmap = np.ones_like(rgb_image[..., 0], dtype=np.float32)\n                    cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n                    camera_pose = cam[\"pose\"]\n                    intrinsics = np.eye(3)\n                    intrinsics[0, 0] = cam[\"intrinsics\"][0, 0]\n                    intrinsics[1, 1] = cam[\"intrinsics\"][0, 0]\n                    intrinsics[0, 2] = cam[\"intrinsics\"][1, 1]\n                    intrinsics[1, 2] = cam[\"intrinsics\"][0, 2]\n                except:\n                    print(f\"Error loading {scene} {basename}, skipping\")\n                    self.invalid_scenes[scene] = True\n                    break\n\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap.astype(np.float32),\n                        camera_pose=camera_pose.astype(np.float32),\n                        camera_intrinsics=intrinsics.astype(np.float32),\n                        dataset=\"MVImgnet\",\n                        label=self.scenes[scene_id] + \"_\" + basename,\n                        instance=f\"{str(idx)}_{str(view_idx)}\",\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.98, dtype=np.float32),\n                        img_mask=True,\n                        ray_mask=False,\n                        camera_only=True,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n            if len(views) == num_views:\n                invalid_seq = False\n        return views","source_hash":"a75252d8106f05fe95cf409028aba31d93c3a7b1dce68b0b3d4bbdd6c0833c95","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.urbansyn","uri":"program://Human3R/module/src.dust3r.datasets.urbansyn#L1-L82","kind":"module","name":"src.dust3r.datasets.urbansyn","path":"src/dust3r/datasets/urbansyn.py","language":"python","start_line":1,"end_line":82,"context_start_line":1,"context_end_line":82,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass UrbanSyn(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")])\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for img_name in img_names:\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            depthmap[depthmap > 200] = 0.0\n\n            intrinsics = np.load(osp.join(self.ROOT, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"urbansyn\",\n                    label=img_name,\n                    instance=f\"{str(idx)}_{img_name}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"1f5f40b60354c47212c541c812dd9e041a5c442d92bd1c61fd85262a60702b22","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.urbansyn.UrbanSyn","uri":"program://Human3R/class/src.dust3r.datasets.urbansyn.UrbanSyn#L14-L82","kind":"class","name":"UrbanSyn","path":"src/dust3r/datasets/urbansyn.py","language":"python","start_line":14,"end_line":82,"context_start_line":1,"context_end_line":82,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass UrbanSyn(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")])\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for img_name in img_names:\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            depthmap[depthmap > 200] = 0.0\n\n            intrinsics = np.load(osp.join(self.ROOT, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"urbansyn\",\n                    label=img_name,\n                    instance=f\"{str(idx)}_{img_name}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"1f5f40b60354c47212c541c812dd9e041a5c442d92bd1c61fd85262a60702b22","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.urbansyn.__init__","uri":"program://Human3R/function/src.dust3r.datasets.urbansyn.__init__#L15-L20","kind":"function","name":"__init__","path":"src/dust3r/datasets/urbansyn.py","language":"python","start_line":15,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass UrbanSyn(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")])\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for img_name in img_names:\n            # Load RGB image","source_hash":"1f5f40b60354c47212c541c812dd9e041a5c442d92bd1c61fd85262a60702b22","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.urbansyn._load_data","uri":"program://Human3R/function/src.dust3r.datasets.urbansyn._load_data#L22-L25","kind":"function","name":"_load_data","path":"src/dust3r/datasets/urbansyn.py","language":"python","start_line":22,"end_line":25,"context_start_line":2,"context_end_line":45,"code":"import cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass UrbanSyn(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")])\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for img_name in img_names:\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127","source_hash":"1f5f40b60354c47212c541c812dd9e041a5c442d92bd1c61fd85262a60702b22","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.urbansyn.__len__","uri":"program://Human3R/function/src.dust3r.datasets.urbansyn.__len__#L27-L28","kind":"function","name":"__len__","path":"src/dust3r/datasets/urbansyn.py","language":"python","start_line":27,"end_line":28,"context_start_line":7,"context_end_line":48,"code":"\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass UrbanSyn(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")])\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for img_name in img_names:\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)","source_hash":"1f5f40b60354c47212c541c812dd9e041a5c442d92bd1c61fd85262a60702b22","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.urbansyn.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.urbansyn.get_image_num#L30-L31","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/urbansyn.py","language":"python","start_line":30,"end_line":31,"context_start_line":10,"context_end_line":51,"code":"from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass UrbanSyn(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")])\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for img_name in img_names:\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            depthmap[depthmap > 200] = 0.0\n\n            intrinsics = np.load(osp.join(self.ROOT, \"cam\", f\"{img_name}.npz\"))[","source_hash":"1f5f40b60354c47212c541c812dd9e041a5c442d92bd1c61fd85262a60702b22","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.urbansyn._get_views","uri":"program://Human3R/function/src.dust3r.datasets.urbansyn._get_views#L33-L82","kind":"function","name":"_get_views","path":"src/dust3r/datasets/urbansyn.py","language":"python","start_line":33,"end_line":82,"context_start_line":13,"context_end_line":82,"code":"\nclass UrbanSyn(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")])\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for img_name in img_names:\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            depthmap[depthmap > 200] = 0.0\n\n            intrinsics = np.load(osp.join(self.ROOT, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"urbansyn\",\n                    label=img_name,\n                    instance=f\"{str(idx)}_{img_name}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"1f5f40b60354c47212c541c812dd9e041a5c442d92bd1c61fd85262a60702b22","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.spring","uri":"program://Human3R/module/src.dust3r.datasets.spring#L1-L137","kind":"module","name":"src.dust3r.datasets.spring","path":"src/dust3r/datasets/spring.py","language":"python","start_line":1,"end_line":137,"context_start_line":1,"context_end_line":137,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Spring(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.10, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"spring\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"98f3c2bb4565d2eff0cf26b79f50759cfe2e138117b4dae20ff675297ef4a088","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.spring.Spring","uri":"program://Human3R/class/src.dust3r.datasets.spring.Spring#L14-L137","kind":"class","name":"Spring","path":"src/dust3r/datasets/spring.py","language":"python","start_line":14,"end_line":137,"context_start_line":1,"context_end_line":137,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Spring(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.10, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"spring\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"98f3c2bb4565d2eff0cf26b79f50759cfe2e138117b4dae20ff675297ef4a088","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.spring.__init__","uri":"program://Human3R/function/src.dust3r.datasets.spring.__init__#L15-L22","kind":"function","name":"__init__","path":"src/dust3r/datasets/spring.py","language":"python","start_line":15,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Spring(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)","source_hash":"98f3c2bb4565d2eff0cf26b79f50759cfe2e138117b4dae20ff675297ef4a088","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.spring._load_data","uri":"program://Human3R/function/src.dust3r.datasets.spring._load_data#L24-L66","kind":"function","name":"_load_data","path":"src/dust3r/datasets/spring.py","language":"python","start_line":24,"end_line":66,"context_start_line":4,"context_end_line":86,"code":"import itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Spring(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]","source_hash":"98f3c2bb4565d2eff0cf26b79f50759cfe2e138117b4dae20ff675297ef4a088","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.spring.__len__","uri":"program://Human3R/function/src.dust3r.datasets.spring.__len__#L68-L69","kind":"function","name":"__len__","path":"src/dust3r/datasets/spring.py","language":"python","start_line":68,"end_line":69,"context_start_line":48,"context_end_line":89,"code":"            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):","source_hash":"98f3c2bb4565d2eff0cf26b79f50759cfe2e138117b4dae20ff675297ef4a088","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.spring.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.spring.get_image_num#L71-L72","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/spring.py","language":"python","start_line":71,"end_line":72,"context_start_line":51,"context_end_line":92,"code":"\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")","source_hash":"98f3c2bb4565d2eff0cf26b79f50759cfe2e138117b4dae20ff675297ef4a088","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.spring._get_views","uri":"program://Human3R/function/src.dust3r.datasets.spring._get_views#L74-L137","kind":"function","name":"_get_views","path":"src/dust3r/datasets/spring.py","language":"python","start_line":74,"end_line":137,"context_start_line":54,"context_end_line":137,"code":"            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.10, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"spring\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"98f3c2bb4565d2eff0cf26b79f50759cfe2e138117b4dae20ff675297ef4a088","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.uasol","uri":"program://Human3R/module/src.dust3r.datasets.uasol#L1-L148","kind":"module","name":"src.dust3r.datasets.uasol","path":"src/dust3r/datasets/uasol.py","language":"python","start_line":1,"end_line":148,"context_start_line":1,"context_end_line":148,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass UASOL_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 40\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=extract_number,\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.75,\n            fix_interval_prob=0.75,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            depthmap[depthmap >= 20] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"UASOL\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".png\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.9, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"f729162f30ae3c3895e0f11af5cbd767d7a2236a5554455d31fd914706a1b17d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.uasol.extract_number","uri":"program://Human3R/function/src.dust3r.datasets.uasol.extract_number#L16-L20","kind":"function","name":"extract_number","path":"src/dust3r/datasets/uasol.py","language":"python","start_line":16,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass UASOL_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 40\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []","source_hash":"f729162f30ae3c3895e0f11af5cbd767d7a2236a5554455d31fd914706a1b17d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.uasol.UASOL_Multi","uri":"program://Human3R/class/src.dust3r.datasets.uasol.UASOL_Multi#L23-L148","kind":"class","name":"UASOL_Multi","path":"src/dust3r/datasets/uasol.py","language":"python","start_line":23,"end_line":148,"context_start_line":3,"context_end_line":148,"code":"import numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass UASOL_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 40\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=extract_number,\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.75,\n            fix_interval_prob=0.75,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            depthmap[depthmap >= 20] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"UASOL\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".png\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.9, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"f729162f30ae3c3895e0f11af5cbd767d7a2236a5554455d31fd914706a1b17d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.uasol.__init__","uri":"program://Human3R/function/src.dust3r.datasets.uasol.__init__#L24-L30","kind":"function","name":"__init__","path":"src/dust3r/datasets/uasol.py","language":"python","start_line":24,"end_line":30,"context_start_line":4,"context_end_line":50,"code":"import itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass UASOL_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 40\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=extract_number,\n            )\n            num_imgs = len(basenames)","source_hash":"f729162f30ae3c3895e0f11af5cbd767d7a2236a5554455d31fd914706a1b17d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.uasol._load_data","uri":"program://Human3R/function/src.dust3r.datasets.uasol._load_data#L32-L76","kind":"function","name":"_load_data","path":"src/dust3r/datasets/uasol.py","language":"python","start_line":32,"end_line":76,"context_start_line":12,"context_end_line":96,"code":"\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass UASOL_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 40\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=extract_number,\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.75,\n            fix_interval_prob=0.75,\n        )\n        image_idxs = np.array(all_image_ids)[pos]","source_hash":"f729162f30ae3c3895e0f11af5cbd767d7a2236a5554455d31fd914706a1b17d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.uasol.__len__","uri":"program://Human3R/function/src.dust3r.datasets.uasol.__len__#L78-L79","kind":"function","name":"__len__","path":"src/dust3r/datasets/uasol.py","language":"python","start_line":78,"end_line":79,"context_start_line":58,"context_end_line":99,"code":"            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.75,\n            fix_interval_prob=0.75,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):","source_hash":"f729162f30ae3c3895e0f11af5cbd767d7a2236a5554455d31fd914706a1b17d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.uasol.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.uasol.get_image_num#L81-L82","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/uasol.py","language":"python","start_line":81,"end_line":82,"context_start_line":61,"context_end_line":102,"code":"\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.75,\n            fix_interval_prob=0.75,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")","source_hash":"f729162f30ae3c3895e0f11af5cbd767d7a2236a5554455d31fd914706a1b17d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.uasol._get_views","uri":"program://Human3R/function/src.dust3r.datasets.uasol._get_views#L84-L148","kind":"function","name":"_get_views","path":"src/dust3r/datasets/uasol.py","language":"python","start_line":84,"end_line":148,"context_start_line":64,"context_end_line":148,"code":"            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.75,\n            fix_interval_prob=0.75,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            depthmap[depthmap >= 20] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"UASOL\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".png\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.9, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"f729162f30ae3c3895e0f11af5cbd767d7a2236a5554455d31fd914706a1b17d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.bedlam","uri":"program://Human3R/module/src.dust3r.datasets.bedlam#L1-L356","kind":"module","name":"src.dust3r.datasets.bedlam","path":"src/dust3r/datasets/bedlam.py","language":"python","start_line":1,"end_line":356,"context_start_line":1,"context_end_line":356,"code":"import os.path as osp\nimport numpy as np\nimport os\nimport sys\nimport pickle\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\ninvalid_seqs = [\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000042\",\n    \"20221024_10_100_batch01handhair_zoom_suburb_d_seq_000059\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000079\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000978\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000081\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000268\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000089\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000189\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000034\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000889\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000293\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000067\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000904\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000434\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000044\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000013\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000396\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000012\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000082\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000120\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000324\",\n    \"20221013_3_250_batch01hand_static_bigOffice_seq_000038\",\n    \"20221012_3-10_500_batch01hand_zoom_highSchoolGym_seq_000486\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000421\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000226\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000012\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000149\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000311\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000080\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000122\",\n    \"20221012_3-10_500_batch01hand_zoom_highSchoolGym_seq_000079\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000077\",\n    \"20221014_3_250_batch01hand_orbit_archVizUI3_time15_seq_000095\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000062\",\n    \"20221013_3_250_batch01hand_static_bigOffice_seq_000015\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000095\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000119\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000297\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000011\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000196\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000316\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000283\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000085\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000287\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000163\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000804\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000842\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000027\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000182\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000982\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000029\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000031\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000025\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000250\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000785\",\n    \"20221024_10_100_batch01handhair_zoom_suburb_d_seq_000069\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000122\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000246\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000352\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000425\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000192\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000900\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000043\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000063\",\n    \"20221014_3_250_batch01hand_orbit_archVizUI3_time15_seq_000096\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000091\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000013\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000309\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000114\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000969\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000361\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000267\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000083\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000383\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000890\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000003\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000045\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000317\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000076\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000082\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000907\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000279\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000076\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000004\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000061\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000811\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000800\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000841\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000794\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000308\",\n    \"20221024_10_100_batch01handhair_zoom_suburb_d_seq_000064\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000284\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000752\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000269\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000036\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000419\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000290\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000322\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000818\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000327\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000326\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000002\",\n    \"20221024_10_100_batch01handhair_zoom_suburb_d_seq_000060\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000348\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000059\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000016\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000817\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000332\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000094\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000193\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000779\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000177\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000368\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000023\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000024\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000310\",\n    \"20221014_3_250_batch01hand_orbit_archVizUI3_time15_seq_000086\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000038\",\n    \"20221024_10_100_batch01handhair_zoom_suburb_d_seq_000071\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000768\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000017\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000053\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000097\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000856\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000827\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000161\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000084\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000106\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000207\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000007\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000013\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000251\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000796\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000105\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000251\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000046\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000334\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000453\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000373\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000283\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000249\",\n]\nhdri_scenes = [\n    \"20221010_3_1000_batch01hand\",\n    \"20221017_3_1000_batch01hand\",\n    \"20221018_3-8_250_batch01hand\",\n    \"20221019_3_250_highbmihand\",\n]\n\n\nclass BEDLAM_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        self.max_humans = 10\n        self.smpl_key2shape= {\n            'smplx_root_pose': (1, 3), \n            'smplx_body_pose': (21, 3), \n            'smplx_jaw_pose': (1, 3), \n            'smplx_leye_pose': (1, 3), \n            'smplx_reye_pose': (1, 3), \n            'smplx_left_hand_pose': (15, 3), \n            'smplx_right_hand_pose': (15, 3), \n            'smplx_shape': (11,), \n            'smplx_transl': (3,), \n            'smplx_gender_id': (),\n            }\n\n        super().__init__(*args, **kwargs)\n\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Test\"\n        else:\n            raise ValueError(\"\")\n        \n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scenes = os.listdir(osp.join(self.ROOT, split))\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            if scene in invalid_seqs:\n                continue\n            if any([scene.startswith(x) for x in hdri_scenes]):\n                continue\n            if \"closeup\" in scene:\n                continue\n            scene_dir = osp.join(self.ROOT, split, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            mask_dir = osp.join(scene_dir, \"mask\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n            smpl_dir = osp.join(scene_dir, \"smpl\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load mask image\n            if os.path.exists(mask_dir):\n                mask_image = imread_cv2(osp.join(mask_dir, basename + \".png\"))\n            else:\n                mask_image = None\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            depthmap[depthmap > 200.0] = 0.0\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n\n            annot_file = osp.join(smpl_dir, f\"{basename}.pkl\")\n            annots = []\n            smpl_mask = np.zeros(self.max_humans, dtype=np.bool_)\n\n            if os.path.isfile(annot_file):\n                with open(annot_file, 'rb') as f:\n                    annots = pickle.load(f)\n                humans = [hum for hum in annots if hum['smplx_transl'][-1] > 0.01] # the person should be in front of the camera\n                if len(humans) > 0:\n                    smpl_mask[:len(humans)] = 1.\n                    l_dist = [hum['smplx_transl'][-1] for hum in humans]\n                    indexed_lst = list(enumerate(l_dist))\n                    sorted_indexed = sorted(indexed_lst, key=lambda x: x[1], reverse=False)\n                    sorted_indices = [index for index, _ in sorted_indexed]\n                    annots = [humans[h_idx] for h_idx in sorted_indices]\n\n                    # Update smplx_gender - 0=neutral - 1=male - 2=female - kids?\n                    for hum in annots:\n                        hum['smplx_gender_id'] = np.asarray({'neutral': 0}[hum['smplx_gender']])\n\n            if mask_image is not None:\n                rgb_image, depthmap, mask_image, intrinsics = self._crop_resize_if_necessary_mask(\n                    rgb_image, depthmap, mask_image, intrinsics, resolution, rng=rng, info=view_idx\n                )\n            else:\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n                )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.00, 0.15]\n            )\n            # Reorganize the smpl annotations\n            smpl_dict = {}\n            for k in self.smpl_key2shape.keys():\n                smpl_dict[k] = np.zeros((self.max_humans, *self.smpl_key2shape[k]), dtype=np.float32)\n                if len(humans) > 0:\n                    for h in range(len(humans)):\n                        smpl_dict[k][h] = annots[h][k].astype(np.float32)\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    msk=False if mask_image is None else mask_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"BEDLAM\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".png\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                    smpl_mask=smpl_mask,\n                    **smpl_dict,\n                )\n            )\n\n        assert len(views) == num_views\n        return views","source_hash":"813bf1fcea87d92249abdcd53500ae1fd82ea9b86bf2a9bc98d3a9c1ca4b29f0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.bedlam.BEDLAM_Multi","uri":"program://Human3R/class/src.dust3r.datasets.bedlam.BEDLAM_Multi#L162-L356","kind":"class","name":"BEDLAM_Multi","path":"src/dust3r/datasets/bedlam.py","language":"python","start_line":162,"end_line":356,"context_start_line":142,"context_end_line":356,"code":"    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000013\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000251\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000796\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000105\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000251\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000046\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000334\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000453\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000373\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000283\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000249\",\n]\nhdri_scenes = [\n    \"20221010_3_1000_batch01hand\",\n    \"20221017_3_1000_batch01hand\",\n    \"20221018_3-8_250_batch01hand\",\n    \"20221019_3_250_highbmihand\",\n]\n\n\nclass BEDLAM_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        self.max_humans = 10\n        self.smpl_key2shape= {\n            'smplx_root_pose': (1, 3), \n            'smplx_body_pose': (21, 3), \n            'smplx_jaw_pose': (1, 3), \n            'smplx_leye_pose': (1, 3), \n            'smplx_reye_pose': (1, 3), \n            'smplx_left_hand_pose': (15, 3), \n            'smplx_right_hand_pose': (15, 3), \n            'smplx_shape': (11,), \n            'smplx_transl': (3,), \n            'smplx_gender_id': (),\n            }\n\n        super().__init__(*args, **kwargs)\n\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Test\"\n        else:\n            raise ValueError(\"\")\n        \n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scenes = os.listdir(osp.join(self.ROOT, split))\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            if scene in invalid_seqs:\n                continue\n            if any([scene.startswith(x) for x in hdri_scenes]):\n                continue\n            if \"closeup\" in scene:\n                continue\n            scene_dir = osp.join(self.ROOT, split, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            mask_dir = osp.join(scene_dir, \"mask\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n            smpl_dir = osp.join(scene_dir, \"smpl\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load mask image\n            if os.path.exists(mask_dir):\n                mask_image = imread_cv2(osp.join(mask_dir, basename + \".png\"))\n            else:\n                mask_image = None\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            depthmap[depthmap > 200.0] = 0.0\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n\n            annot_file = osp.join(smpl_dir, f\"{basename}.pkl\")\n            annots = []\n            smpl_mask = np.zeros(self.max_humans, dtype=np.bool_)\n\n            if os.path.isfile(annot_file):\n                with open(annot_file, 'rb') as f:\n                    annots = pickle.load(f)\n                humans = [hum for hum in annots if hum['smplx_transl'][-1] > 0.01] # the person should be in front of the camera\n                if len(humans) > 0:\n                    smpl_mask[:len(humans)] = 1.\n                    l_dist = [hum['smplx_transl'][-1] for hum in humans]\n                    indexed_lst = list(enumerate(l_dist))\n                    sorted_indexed = sorted(indexed_lst, key=lambda x: x[1], reverse=False)\n                    sorted_indices = [index for index, _ in sorted_indexed]\n                    annots = [humans[h_idx] for h_idx in sorted_indices]\n\n                    # Update smplx_gender - 0=neutral - 1=male - 2=female - kids?\n                    for hum in annots:\n                        hum['smplx_gender_id'] = np.asarray({'neutral': 0}[hum['smplx_gender']])\n\n            if mask_image is not None:\n                rgb_image, depthmap, mask_image, intrinsics = self._crop_resize_if_necessary_mask(\n                    rgb_image, depthmap, mask_image, intrinsics, resolution, rng=rng, info=view_idx\n                )\n            else:\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n                )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.00, 0.15]\n            )\n            # Reorganize the smpl annotations\n            smpl_dict = {}\n            for k in self.smpl_key2shape.keys():\n                smpl_dict[k] = np.zeros((self.max_humans, *self.smpl_key2shape[k]), dtype=np.float32)\n                if len(humans) > 0:\n                    for h in range(len(humans)):\n                        smpl_dict[k][h] = annots[h][k].astype(np.float32)\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    msk=False if mask_image is None else mask_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"BEDLAM\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".png\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                    smpl_mask=smpl_mask,\n                    **smpl_dict,\n                )\n            )\n\n        assert len(views) == num_views\n        return views","source_hash":"813bf1fcea87d92249abdcd53500ae1fd82ea9b86bf2a9bc98d3a9c1ca4b29f0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.bedlam.__init__","uri":"program://Human3R/function/src.dust3r.datasets.bedlam.__init__#L163-L191","kind":"function","name":"__init__","path":"src/dust3r/datasets/bedlam.py","language":"python","start_line":163,"end_line":191,"context_start_line":143,"context_end_line":211,"code":"    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000251\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000796\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000105\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000251\",\n    \"20221019_3-8_250_highbmihand_orbit_stadium_seq_000046\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000334\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000453\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000373\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000283\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000249\",\n]\nhdri_scenes = [\n    \"20221010_3_1000_batch01hand\",\n    \"20221017_3_1000_batch01hand\",\n    \"20221018_3-8_250_batch01hand\",\n    \"20221019_3_250_highbmihand\",\n]\n\n\nclass BEDLAM_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        self.max_humans = 10\n        self.smpl_key2shape= {\n            'smplx_root_pose': (1, 3), \n            'smplx_body_pose': (21, 3), \n            'smplx_jaw_pose': (1, 3), \n            'smplx_leye_pose': (1, 3), \n            'smplx_reye_pose': (1, 3), \n            'smplx_left_hand_pose': (15, 3), \n            'smplx_right_hand_pose': (15, 3), \n            'smplx_shape': (11,), \n            'smplx_transl': (3,), \n            'smplx_gender_id': (),\n            }\n\n        super().__init__(*args, **kwargs)\n\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Test\"\n        else:\n            raise ValueError(\"\")\n        \n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scenes = os.listdir(osp.join(self.ROOT, split))\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            if scene in invalid_seqs:\n                continue\n            if any([scene.startswith(x) for x in hdri_scenes]):\n                continue\n            if \"closeup\" in scene:\n                continue\n            scene_dir = osp.join(self.ROOT, split, scene)","source_hash":"813bf1fcea87d92249abdcd53500ae1fd82ea9b86bf2a9bc98d3a9c1ca4b29f0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.bedlam._load_data","uri":"program://Human3R/function/src.dust3r.datasets.bedlam._load_data#L193-L240","kind":"function","name":"_load_data","path":"src/dust3r/datasets/bedlam.py","language":"python","start_line":193,"end_line":240,"context_start_line":173,"context_end_line":260,"code":"            'smplx_leye_pose': (1, 3), \n            'smplx_reye_pose': (1, 3), \n            'smplx_left_hand_pose': (15, 3), \n            'smplx_right_hand_pose': (15, 3), \n            'smplx_shape': (11,), \n            'smplx_transl': (3,), \n            'smplx_gender_id': (),\n            }\n\n        super().__init__(*args, **kwargs)\n\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Test\"\n        else:\n            raise ValueError(\"\")\n        \n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scenes = os.listdir(osp.join(self.ROOT, split))\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            if scene in invalid_seqs:\n                continue\n            if any([scene.startswith(x) for x in hdri_scenes]):\n                continue\n            if \"closeup\" in scene:\n                continue\n            scene_dir = osp.join(self.ROOT, split, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]","source_hash":"813bf1fcea87d92249abdcd53500ae1fd82ea9b86bf2a9bc98d3a9c1ca4b29f0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.bedlam.__len__","uri":"program://Human3R/function/src.dust3r.datasets.bedlam.__len__#L242-L243","kind":"function","name":"__len__","path":"src/dust3r/datasets/bedlam.py","language":"python","start_line":242,"end_line":243,"context_start_line":222,"context_end_line":263,"code":"                print(f\"Skipping {scene}\")\n                continue\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):","source_hash":"813bf1fcea87d92249abdcd53500ae1fd82ea9b86bf2a9bc98d3a9c1ca4b29f0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.bedlam.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.bedlam.get_image_num#L245-L246","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/bedlam.py","language":"python","start_line":245,"end_line":246,"context_start_line":225,"context_end_line":266,"code":"\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")","source_hash":"813bf1fcea87d92249abdcd53500ae1fd82ea9b86bf2a9bc98d3a9c1ca4b29f0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.bedlam._get_views","uri":"program://Human3R/function/src.dust3r.datasets.bedlam._get_views#L248-L356","kind":"function","name":"_get_views","path":"src/dust3r/datasets/bedlam.py","language":"python","start_line":248,"end_line":356,"context_start_line":228,"context_end_line":356,"code":"            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            mask_dir = osp.join(scene_dir, \"mask\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n            smpl_dir = osp.join(scene_dir, \"smpl\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load mask image\n            if os.path.exists(mask_dir):\n                mask_image = imread_cv2(osp.join(mask_dir, basename + \".png\"))\n            else:\n                mask_image = None\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            depthmap[depthmap > 200.0] = 0.0\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n\n            annot_file = osp.join(smpl_dir, f\"{basename}.pkl\")\n            annots = []\n            smpl_mask = np.zeros(self.max_humans, dtype=np.bool_)\n\n            if os.path.isfile(annot_file):\n                with open(annot_file, 'rb') as f:\n                    annots = pickle.load(f)\n                humans = [hum for hum in annots if hum['smplx_transl'][-1] > 0.01] # the person should be in front of the camera\n                if len(humans) > 0:\n                    smpl_mask[:len(humans)] = 1.\n                    l_dist = [hum['smplx_transl'][-1] for hum in humans]\n                    indexed_lst = list(enumerate(l_dist))\n                    sorted_indexed = sorted(indexed_lst, key=lambda x: x[1], reverse=False)\n                    sorted_indices = [index for index, _ in sorted_indexed]\n                    annots = [humans[h_idx] for h_idx in sorted_indices]\n\n                    # Update smplx_gender - 0=neutral - 1=male - 2=female - kids?\n                    for hum in annots:\n                        hum['smplx_gender_id'] = np.asarray({'neutral': 0}[hum['smplx_gender']])\n\n            if mask_image is not None:\n                rgb_image, depthmap, mask_image, intrinsics = self._crop_resize_if_necessary_mask(\n                    rgb_image, depthmap, mask_image, intrinsics, resolution, rng=rng, info=view_idx\n                )\n            else:\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n                )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.00, 0.15]\n            )\n            # Reorganize the smpl annotations\n            smpl_dict = {}\n            for k in self.smpl_key2shape.keys():\n                smpl_dict[k] = np.zeros((self.max_humans, *self.smpl_key2shape[k]), dtype=np.float32)\n                if len(humans) > 0:\n                    for h in range(len(humans)):\n                        smpl_dict[k][h] = annots[h][k].astype(np.float32)\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    msk=False if mask_image is None else mask_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"BEDLAM\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".png\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                    smpl_mask=smpl_mask,\n                    **smpl_dict,\n                )\n            )\n\n        assert len(views) == num_views\n        return views","source_hash":"813bf1fcea87d92249abdcd53500ae1fd82ea9b86bf2a9bc98d3a9c1ca4b29f0","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.realestate10k","uri":"program://Human3R/module/src.dust3r.datasets.realestate10k#L1-L139","kind":"module","name":"src.dust3r.datasets.realestate10k","path":"src/dust3r/datasets/realestate10k.py","language":"python","start_line":1,"end_line":139,"context_start_line":1,"context_end_line":139,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass RE10K_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 128\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=lambda x: int(x),\n            )\n\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []\n            for view_idx in image_idxs:\n                scene_id = self.sceneids[view_idx]\n                scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n                rgb_dir = osp.join(scene_dir, \"rgb\")\n                cam_dir = osp.join(scene_dir, \"cam\")\n\n                basename = self.images[view_idx]\n\n                try:\n                    # Load RGB image\n                    rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n                    # Load depthmap, no depth, set to all ones\n                    depthmap = np.ones_like(rgb_image[..., 0], dtype=np.float32)\n                    cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n                    intrinsics = cam[\"intrinsics\"]\n                    camera_pose = cam[\"pose\"]\n                except:\n                    print(f\"Error loading {scene} {basename}, skipping\")\n                    self.invalid_scenes[scene] = True\n                    break\n\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap.astype(np.float32),\n                        camera_pose=camera_pose.astype(np.float32),\n                        camera_intrinsics=intrinsics.astype(np.float32),\n                        dataset=\"realestate10k\",\n                        label=self.scenes[scene_id] + \"_\" + basename,\n                        instance=f\"{str(idx)}_{str(view_idx)}\",\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.98, dtype=np.float32),\n                        img_mask=True,\n                        ray_mask=False,\n                        camera_only=True,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n            if len(views) == num_views:\n                invalid_seq = False\n        return views","source_hash":"d56877dac5303ffef371832f370bdc4e24c13d8fe45cec26c645e7707d470747","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.realestate10k.RE10K_Multi","uri":"program://Human3R/class/src.dust3r.datasets.realestate10k.RE10K_Multi#L14-L139","kind":"class","name":"RE10K_Multi","path":"src/dust3r/datasets/realestate10k.py","language":"python","start_line":14,"end_line":139,"context_start_line":1,"context_end_line":139,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass RE10K_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 128\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=lambda x: int(x),\n            )\n\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []\n            for view_idx in image_idxs:\n                scene_id = self.sceneids[view_idx]\n                scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n                rgb_dir = osp.join(scene_dir, \"rgb\")\n                cam_dir = osp.join(scene_dir, \"cam\")\n\n                basename = self.images[view_idx]\n\n                try:\n                    # Load RGB image\n                    rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n                    # Load depthmap, no depth, set to all ones\n                    depthmap = np.ones_like(rgb_image[..., 0], dtype=np.float32)\n                    cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n                    intrinsics = cam[\"intrinsics\"]\n                    camera_pose = cam[\"pose\"]\n                except:\n                    print(f\"Error loading {scene} {basename}, skipping\")\n                    self.invalid_scenes[scene] = True\n                    break\n\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap.astype(np.float32),\n                        camera_pose=camera_pose.astype(np.float32),\n                        camera_intrinsics=intrinsics.astype(np.float32),\n                        dataset=\"realestate10k\",\n                        label=self.scenes[scene_id] + \"_\" + basename,\n                        instance=f\"{str(idx)}_{str(view_idx)}\",\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.98, dtype=np.float32),\n                        img_mask=True,\n                        ray_mask=False,\n                        camera_only=True,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n            if len(views) == num_views:\n                invalid_seq = False\n        return views","source_hash":"d56877dac5303ffef371832f370bdc4e24c13d8fe45cec26c645e7707d470747","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.realestate10k.__init__","uri":"program://Human3R/function/src.dust3r.datasets.realestate10k.__init__#L15-L21","kind":"function","name":"__init__","path":"src/dust3r/datasets/realestate10k.py","language":"python","start_line":15,"end_line":21,"context_start_line":1,"context_end_line":41,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass RE10K_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 128\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=lambda x: int(x),\n            )\n","source_hash":"d56877dac5303ffef371832f370bdc4e24c13d8fe45cec26c645e7707d470747","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.realestate10k._load_data","uri":"program://Human3R/function/src.dust3r.datasets.realestate10k._load_data#L23-L68","kind":"function","name":"_load_data","path":"src/dust3r/datasets/realestate10k.py","language":"python","start_line":23,"end_line":68,"context_start_line":3,"context_end_line":88,"code":"import numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass RE10K_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 128\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=lambda x: int(x),\n            )\n\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )","source_hash":"d56877dac5303ffef371832f370bdc4e24c13d8fe45cec26c645e7707d470747","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.realestate10k.__len__","uri":"program://Human3R/function/src.dust3r.datasets.realestate10k.__len__#L70-L71","kind":"function","name":"__len__","path":"src/dust3r/datasets/realestate10k.py","language":"python","start_line":70,"end_line":71,"context_start_line":50,"context_end_line":91,"code":"            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []","source_hash":"d56877dac5303ffef371832f370bdc4e24c13d8fe45cec26c645e7707d470747","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.realestate10k.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.realestate10k.get_image_num#L73-L74","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/realestate10k.py","language":"python","start_line":73,"end_line":74,"context_start_line":53,"context_end_line":94,"code":"            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []\n            for view_idx in image_idxs:\n                scene_id = self.sceneids[view_idx]\n                scene_dir = osp.join(self.ROOT, self.scenes[scene_id])","source_hash":"d56877dac5303ffef371832f370bdc4e24c13d8fe45cec26c645e7707d470747","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.realestate10k._get_views","uri":"program://Human3R/function/src.dust3r.datasets.realestate10k._get_views#L76-L139","kind":"function","name":"_get_views","path":"src/dust3r/datasets/realestate10k.py","language":"python","start_line":76,"end_line":139,"context_start_line":56,"context_end_line":139,"code":"            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n        self.invalid_scenes = {scene: False for scene in self.scenes}\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene, start_id = self.start_img_ids[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene]:\n                idx = rng.integers(low=0, high=len(self.start_img_ids))\n                scene, start_id = self.start_img_ids[idx]\n\n            all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n            )\n            image_idxs = np.array(all_image_ids)[pos]\n\n            views = []\n            for view_idx in image_idxs:\n                scene_id = self.sceneids[view_idx]\n                scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n                rgb_dir = osp.join(scene_dir, \"rgb\")\n                cam_dir = osp.join(scene_dir, \"cam\")\n\n                basename = self.images[view_idx]\n\n                try:\n                    # Load RGB image\n                    rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n                    # Load depthmap, no depth, set to all ones\n                    depthmap = np.ones_like(rgb_image[..., 0], dtype=np.float32)\n                    cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n                    intrinsics = cam[\"intrinsics\"]\n                    camera_pose = cam[\"pose\"]\n                except:\n                    print(f\"Error loading {scene} {basename}, skipping\")\n                    self.invalid_scenes[scene] = True\n                    break\n\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap.astype(np.float32),\n                        camera_pose=camera_pose.astype(np.float32),\n                        camera_intrinsics=intrinsics.astype(np.float32),\n                        dataset=\"realestate10k\",\n                        label=self.scenes[scene_id] + \"_\" + basename,\n                        instance=f\"{str(idx)}_{str(view_idx)}\",\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.98, dtype=np.float32),\n                        img_mask=True,\n                        ray_mask=False,\n                        camera_only=True,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n            if len(views) == num_views:\n                invalid_seq = False\n        return views","source_hash":"d56877dac5303ffef371832f370bdc4e24c13d8fe45cec26c645e7707d470747","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.threedkb","uri":"program://Human3R/module/src.dust3r.datasets.threedkb#L1-L111","kind":"module","name":"src.dust3r.datasets.threedkb","path":"src/dust3r/datasets/threedkb.py","language":"python","start_line":1,"end_line":111,"context_start_line":1,"context_end_line":111,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ThreeDKenBurns(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n\n            num_imgs = len(basenames)\n            img_ids_ = list(np.arange(num_imgs) + offset)\n\n            img_ids.extend(img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.img_ids = img_ids\n\n    def __len__(self):\n        return len(self.img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        image_idxs = new_rng.choice(self.img_ids, num_views, replace=False)\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            depthmap = imread_cv2(osp.join(depth_dir, basename + \".exr\"))\n            depthmap[depthmap > 20000] = 0.0\n            depthmap = depthmap / 1000.0\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            intrinsics = cam[\"intrinsics\"]\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"3DKenBurns\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"5e16195ffff7cf350365c24fb986de0c2f6c73050243f3a5c131ae9c27b4b0eb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.threedkb.ThreeDKenBurns","uri":"program://Human3R/class/src.dust3r.datasets.threedkb.ThreeDKenBurns#L14-L111","kind":"class","name":"ThreeDKenBurns","path":"src/dust3r/datasets/threedkb.py","language":"python","start_line":14,"end_line":111,"context_start_line":1,"context_end_line":111,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ThreeDKenBurns(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n\n            num_imgs = len(basenames)\n            img_ids_ = list(np.arange(num_imgs) + offset)\n\n            img_ids.extend(img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.img_ids = img_ids\n\n    def __len__(self):\n        return len(self.img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        image_idxs = new_rng.choice(self.img_ids, num_views, replace=False)\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            depthmap = imread_cv2(osp.join(depth_dir, basename + \".exr\"))\n            depthmap[depthmap > 20000] = 0.0\n            depthmap = depthmap / 1000.0\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            intrinsics = cam[\"intrinsics\"]\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"3DKenBurns\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"5e16195ffff7cf350365c24fb986de0c2f6c73050243f3a5c131ae9c27b4b0eb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.threedkb.__init__","uri":"program://Human3R/function/src.dust3r.datasets.threedkb.__init__#L15-L20","kind":"function","name":"__init__","path":"src/dust3r/datasets/threedkb.py","language":"python","start_line":15,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ThreeDKenBurns(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n\n            num_imgs = len(basenames)\n            img_ids_ = list(np.arange(num_imgs) + offset)","source_hash":"5e16195ffff7cf350365c24fb986de0c2f6c73050243f3a5c131ae9c27b4b0eb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.threedkb._load_data","uri":"program://Human3R/function/src.dust3r.datasets.threedkb._load_data#L22-L54","kind":"function","name":"_load_data","path":"src/dust3r/datasets/threedkb.py","language":"python","start_line":22,"end_line":54,"context_start_line":2,"context_end_line":74,"code":"import cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ThreeDKenBurns(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n\n            num_imgs = len(basenames)\n            img_ids_ = list(np.arange(num_imgs) + offset)\n\n            img_ids.extend(img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.img_ids = img_ids\n\n    def __len__(self):\n        return len(self.img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        image_idxs = new_rng.choice(self.img_ids, num_views, replace=False)\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n","source_hash":"5e16195ffff7cf350365c24fb986de0c2f6c73050243f3a5c131ae9c27b4b0eb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.threedkb.__len__","uri":"program://Human3R/function/src.dust3r.datasets.threedkb.__len__#L56-L57","kind":"function","name":"__len__","path":"src/dust3r/datasets/threedkb.py","language":"python","start_line":56,"end_line":57,"context_start_line":36,"context_end_line":77,"code":"                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n\n            num_imgs = len(basenames)\n            img_ids_ = list(np.arange(num_imgs) + offset)\n\n            img_ids.extend(img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.img_ids = img_ids\n\n    def __len__(self):\n        return len(self.img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        image_idxs = new_rng.choice(self.img_ids, num_views, replace=False)\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image","source_hash":"5e16195ffff7cf350365c24fb986de0c2f6c73050243f3a5c131ae9c27b4b0eb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.threedkb.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.threedkb.get_image_num#L59-L60","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/threedkb.py","language":"python","start_line":59,"end_line":60,"context_start_line":39,"context_end_line":80,"code":"            num_imgs = len(basenames)\n            img_ids_ = list(np.arange(num_imgs) + offset)\n\n            img_ids.extend(img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.img_ids = img_ids\n\n    def __len__(self):\n        return len(self.img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        image_idxs = new_rng.choice(self.img_ids, num_views, replace=False)\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            depthmap = imread_cv2(osp.join(depth_dir, basename + \".exr\"))\n            depthmap[depthmap > 20000] = 0.0","source_hash":"5e16195ffff7cf350365c24fb986de0c2f6c73050243f3a5c131ae9c27b4b0eb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.threedkb._get_views","uri":"program://Human3R/function/src.dust3r.datasets.threedkb._get_views#L62-L111","kind":"function","name":"_get_views","path":"src/dust3r/datasets/threedkb.py","language":"python","start_line":62,"end_line":111,"context_start_line":42,"context_end_line":111,"code":"            img_ids.extend(img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.img_ids = img_ids\n\n    def __len__(self):\n        return len(self.img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        image_idxs = new_rng.choice(self.img_ids, num_views, replace=False)\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            depthmap = imread_cv2(osp.join(depth_dir, basename + \".exr\"))\n            depthmap[depthmap > 20000] = 0.0\n            depthmap = depthmap / 1000.0\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            intrinsics = cam[\"intrinsics\"]\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"3DKenBurns\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"5e16195ffff7cf350365c24fb986de0c2f6c73050243f3a5c131ae9c27b4b0eb","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d","uri":"program://Human3R/module/src.dust3r.datasets.co3d#L1-L190","kind":"module","name":"src.dust3r.datasets.co3d","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":1,"end_line":190,"context_start_line":1,"context_end_line":190,"code":"import os.path as osp\nimport json\nimport itertools\nfrom collections import deque\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\nimport time\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Co3d_Multi(BaseMultiViewDataset):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        super().__init__(*args, **kwargs)\n        assert mask_bg in (True, False, \"rand\")\n        self.mask_bg = mask_bg\n        self.is_metric = False\n        self.dataset_label = \"Co3d_v2\"\n\n        # load all scenes\n        with open(osp.join(self.ROOT, f\"selected_seqs_{self.split}.json\"), \"r\") as f:\n            self.scenes = json.load(f)\n            self.scenes = {k: v for k, v in self.scenes.items() if len(v) > 0}\n            self.scenes = {\n                (k, k2): v2 for k, v in self.scenes.items() for k2, v2 in v.items()\n            }\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num(\n            input_metadata[\"maximum_depth\"]\n        )\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]\n            if len(image_pool) < self.cut_off:\n                print(\"Invalid scene!\")\n                self.invalid_scenes[scene_info] = True\n                continue\n\n            imgs_idxs, ordered_video = self.get_seq_from_start_id(\n                num_views, ref_img_idx, image_pool, rng\n            )\n\n            if resolution not in self.invalidate[obj, instance]:  # flag invalid images\n                self.invalidate[obj, instance][resolution] = [\n                    False for _ in range(len(image_pool))\n                ]\n            # decide now if we mask the bg\n            mask_bg = (self.mask_bg == True) or (\n                self.mask_bg == \"rand\" and rng.choice(2, p=[0.9, 0.1])\n            )\n            views = []\n\n            imgs_idxs = deque(imgs_idxs)\n\n            while len(imgs_idxs) > 0:  # some images (few) have zero depth\n                if (\n                    len(image_pool) - sum(self.invalidate[obj, instance][resolution])\n                    < self.cut_off\n                ):\n                    print(\"Invalid scene!\")\n                    invalid_seq = True\n                    self.invalid_scenes[scene_info] = True\n                    break\n\n                im_idx = imgs_idxs.pop()\n                if self.invalidate[obj, instance][resolution][im_idx]:\n                    # search for a valid image\n                    ordered_video = False\n                    random_direction = 2 * rng.choice(2) - 1\n                    for offset in range(1, len(image_pool)):\n                        tentative_im_idx = (im_idx + (random_direction * offset)) % len(\n                            image_pool\n                        )\n                        if not self.invalidate[obj, instance][resolution][\n                            tentative_im_idx\n                        ]:\n                            im_idx = tentative_im_idx\n                            break\n                view_idx = image_pool[im_idx]\n                impath = self._get_impath(obj, instance, view_idx)\n                depthpath = self._get_depthpath(obj, instance, view_idx)\n\n                # load camera params\n                metadata_path = self._get_metadatapath(obj, instance, view_idx)\n                input_metadata = np.load(metadata_path)\n                camera_pose = input_metadata[\"camera_pose\"].astype(np.float32)\n                intrinsics = input_metadata[\"camera_intrinsics\"].astype(np.float32)\n\n                # load image and depth\n                rgb_image = imread_cv2(impath)\n                depthmap = self._read_depthmap(depthpath, input_metadata)\n\n                if mask_bg:\n                    # load object mask\n                    maskpath = self._get_maskpath(obj, instance, view_idx)\n                    maskmap = imread_cv2(maskpath, cv2.IMREAD_UNCHANGED).astype(\n                        np.float32\n                    )\n                    maskmap = (maskmap / 255.0) > 0.1\n\n                    # update the depthmap with mask\n                    depthmap *= maskmap\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=impath\n                )\n                num_valid = (depthmap > 0.0).sum()\n                if num_valid == 0:\n                    # problem, invalidate image and retry\n                    self.invalidate[obj, instance][resolution][im_idx] = True\n                    imgs_idxs.append(im_idx)\n                    continue\n\n                # generate img mask and raymap mask\n                img_mask, ray_mask = self.get_img_and_ray_masks(\n                    self.is_metric, len(views), rng\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap,\n                        camera_pose=camera_pose,\n                        camera_intrinsics=intrinsics,\n                        dataset=self.dataset_label,\n                        label=osp.join(obj, instance),\n                        instance=osp.split(impath)[1],\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.9, dtype=np.float32),\n                        img_mask=img_mask,\n                        ray_mask=ray_mask,\n                        camera_only=False,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n\n            if len(views) == num_views and not all(\n                [view[\"instance\"] == views[0][\"instance\"] for view in views]\n            ):\n                invalid_seq = False\n        return views","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d.Co3d_Multi","uri":"program://Human3R/class/src.dust3r.datasets.co3d.Co3d_Multi#L16-L190","kind":"class","name":"Co3d_Multi","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":16,"end_line":190,"context_start_line":1,"context_end_line":190,"code":"import os.path as osp\nimport json\nimport itertools\nfrom collections import deque\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\nimport time\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Co3d_Multi(BaseMultiViewDataset):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        super().__init__(*args, **kwargs)\n        assert mask_bg in (True, False, \"rand\")\n        self.mask_bg = mask_bg\n        self.is_metric = False\n        self.dataset_label = \"Co3d_v2\"\n\n        # load all scenes\n        with open(osp.join(self.ROOT, f\"selected_seqs_{self.split}.json\"), \"r\") as f:\n            self.scenes = json.load(f)\n            self.scenes = {k: v for k, v in self.scenes.items() if len(v) > 0}\n            self.scenes = {\n                (k, k2): v2 for k, v in self.scenes.items() for k2, v2 in v.items()\n            }\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num(\n            input_metadata[\"maximum_depth\"]\n        )\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]\n            if len(image_pool) < self.cut_off:\n                print(\"Invalid scene!\")\n                self.invalid_scenes[scene_info] = True\n                continue\n\n            imgs_idxs, ordered_video = self.get_seq_from_start_id(\n                num_views, ref_img_idx, image_pool, rng\n            )\n\n            if resolution not in self.invalidate[obj, instance]:  # flag invalid images\n                self.invalidate[obj, instance][resolution] = [\n                    False for _ in range(len(image_pool))\n                ]\n            # decide now if we mask the bg\n            mask_bg = (self.mask_bg == True) or (\n                self.mask_bg == \"rand\" and rng.choice(2, p=[0.9, 0.1])\n            )\n            views = []\n\n            imgs_idxs = deque(imgs_idxs)\n\n            while len(imgs_idxs) > 0:  # some images (few) have zero depth\n                if (\n                    len(image_pool) - sum(self.invalidate[obj, instance][resolution])\n                    < self.cut_off\n                ):\n                    print(\"Invalid scene!\")\n                    invalid_seq = True\n                    self.invalid_scenes[scene_info] = True\n                    break\n\n                im_idx = imgs_idxs.pop()\n                if self.invalidate[obj, instance][resolution][im_idx]:\n                    # search for a valid image\n                    ordered_video = False\n                    random_direction = 2 * rng.choice(2) - 1\n                    for offset in range(1, len(image_pool)):\n                        tentative_im_idx = (im_idx + (random_direction * offset)) % len(\n                            image_pool\n                        )\n                        if not self.invalidate[obj, instance][resolution][\n                            tentative_im_idx\n                        ]:\n                            im_idx = tentative_im_idx\n                            break\n                view_idx = image_pool[im_idx]\n                impath = self._get_impath(obj, instance, view_idx)\n                depthpath = self._get_depthpath(obj, instance, view_idx)\n\n                # load camera params\n                metadata_path = self._get_metadatapath(obj, instance, view_idx)\n                input_metadata = np.load(metadata_path)\n                camera_pose = input_metadata[\"camera_pose\"].astype(np.float32)\n                intrinsics = input_metadata[\"camera_intrinsics\"].astype(np.float32)\n\n                # load image and depth\n                rgb_image = imread_cv2(impath)\n                depthmap = self._read_depthmap(depthpath, input_metadata)\n\n                if mask_bg:\n                    # load object mask\n                    maskpath = self._get_maskpath(obj, instance, view_idx)\n                    maskmap = imread_cv2(maskpath, cv2.IMREAD_UNCHANGED).astype(\n                        np.float32\n                    )\n                    maskmap = (maskmap / 255.0) > 0.1\n\n                    # update the depthmap with mask\n                    depthmap *= maskmap\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=impath\n                )\n                num_valid = (depthmap > 0.0).sum()\n                if num_valid == 0:\n                    # problem, invalidate image and retry\n                    self.invalidate[obj, instance][resolution][im_idx] = True\n                    imgs_idxs.append(im_idx)\n                    continue\n\n                # generate img mask and raymap mask\n                img_mask, ray_mask = self.get_img_and_ray_masks(\n                    self.is_metric, len(views), rng\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap,\n                        camera_pose=camera_pose,\n                        camera_intrinsics=intrinsics,\n                        dataset=self.dataset_label,\n                        label=osp.join(obj, instance),\n                        instance=osp.split(impath)[1],\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.9, dtype=np.float32),\n                        img_mask=img_mask,\n                        ray_mask=ray_mask,\n                        camera_only=False,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n\n            if len(views) == num_views and not all(\n                [view[\"instance\"] == views[0][\"instance\"] for view in views]\n            ):\n                invalid_seq = False\n        return views","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d.__init__","uri":"program://Human3R/function/src.dust3r.datasets.co3d.__init__#L17-L43","kind":"function","name":"__init__","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":17,"end_line":43,"context_start_line":1,"context_end_line":63,"code":"import os.path as osp\nimport json\nimport itertools\nfrom collections import deque\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\nimport time\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Co3d_Multi(BaseMultiViewDataset):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        super().__init__(*args, **kwargs)\n        assert mask_bg in (True, False, \"rand\")\n        self.mask_bg = mask_bg\n        self.is_metric = False\n        self.dataset_label = \"Co3d_v2\"\n\n        # load all scenes\n        with open(osp.join(self.ROOT, f\"selected_seqs_{self.split}.json\"), \"r\") as f:\n            self.scenes = json.load(f)\n            self.scenes = {k: v for k, v in self.scenes.items() if len(v) > 0}\n            self.scenes = {\n                (k, k2): v2 for k, v in self.scenes.items() for k2, v2 in v.items()\n            }\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d.__len__","uri":"program://Human3R/function/src.dust3r.datasets.co3d.__len__#L45-L46","kind":"function","name":"__len__","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":45,"end_line":46,"context_start_line":25,"context_end_line":66,"code":"        # load all scenes\n        with open(osp.join(self.ROOT, f\"selected_seqs_{self.split}.json\"), \"r\") as f:\n            self.scenes = json.load(f)\n            self.scenes = {k: v for k, v in self.scenes.items() if len(v) > 0}\n            self.scenes = {\n                (k, k2): v2 for k, v in self.scenes.items() for k2, v2 in v.items()\n            }\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num(\n            input_metadata[\"maximum_depth\"]\n        )","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d._get_metadatapath","uri":"program://Human3R/function/src.dust3r.datasets.co3d._get_metadatapath#L48-L49","kind":"function","name":"_get_metadatapath","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":48,"end_line":49,"context_start_line":28,"context_end_line":69,"code":"            self.scenes = {k: v for k, v in self.scenes.items() if len(v) > 0}\n            self.scenes = {\n                (k, k2): v2 for k, v in self.scenes.items() for k2, v2 in v.items()\n            }\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num(\n            input_metadata[\"maximum_depth\"]\n        )\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d._get_impath","uri":"program://Human3R/function/src.dust3r.datasets.co3d._get_impath#L51-L52","kind":"function","name":"_get_impath","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":51,"end_line":52,"context_start_line":31,"context_end_line":72,"code":"            }\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num(\n            input_metadata[\"maximum_depth\"]\n        )\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d._get_depthpath","uri":"program://Human3R/function/src.dust3r.datasets.co3d._get_depthpath#L54-L57","kind":"function","name":"_get_depthpath","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":54,"end_line":57,"context_start_line":34,"context_end_line":77,"code":"            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num(\n            input_metadata[\"maximum_depth\"]\n        )\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d._get_maskpath","uri":"program://Human3R/function/src.dust3r.datasets.co3d._get_maskpath#L59-L60","kind":"function","name":"_get_maskpath","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":59,"end_line":60,"context_start_line":39,"context_end_line":80,"code":"            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num(\n            input_metadata[\"maximum_depth\"]\n        )\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d._read_depthmap","uri":"program://Human3R/function/src.dust3r.datasets.co3d._read_depthmap#L62-L67","kind":"function","name":"_read_depthmap","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":62,"end_line":67,"context_start_line":42,"context_end_line":87,"code":"        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num(\n            input_metadata[\"maximum_depth\"]\n        )\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]\n            if len(image_pool) < self.cut_off:\n                print(\"Invalid scene!\")\n                self.invalid_scenes[scene_info] = True\n                continue\n\n            imgs_idxs, ordered_video = self.get_seq_from_start_id(\n                num_views, ref_img_idx, image_pool, rng","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.co3d._get_views","uri":"program://Human3R/function/src.dust3r.datasets.co3d._get_views#L69-L190","kind":"function","name":"_get_views","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":69,"end_line":190,"context_start_line":49,"context_end_line":190,"code":"        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(\n            self.ROOT, obj, instance, \"depths\", f\"frame{view_idx:06n}.jpg.geometric.png\"\n        )\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num(\n            input_metadata[\"maximum_depth\"]\n        )\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]\n            if len(image_pool) < self.cut_off:\n                print(\"Invalid scene!\")\n                self.invalid_scenes[scene_info] = True\n                continue\n\n            imgs_idxs, ordered_video = self.get_seq_from_start_id(\n                num_views, ref_img_idx, image_pool, rng\n            )\n\n            if resolution not in self.invalidate[obj, instance]:  # flag invalid images\n                self.invalidate[obj, instance][resolution] = [\n                    False for _ in range(len(image_pool))\n                ]\n            # decide now if we mask the bg\n            mask_bg = (self.mask_bg == True) or (\n                self.mask_bg == \"rand\" and rng.choice(2, p=[0.9, 0.1])\n            )\n            views = []\n\n            imgs_idxs = deque(imgs_idxs)\n\n            while len(imgs_idxs) > 0:  # some images (few) have zero depth\n                if (\n                    len(image_pool) - sum(self.invalidate[obj, instance][resolution])\n                    < self.cut_off\n                ):\n                    print(\"Invalid scene!\")\n                    invalid_seq = True\n                    self.invalid_scenes[scene_info] = True\n                    break\n\n                im_idx = imgs_idxs.pop()\n                if self.invalidate[obj, instance][resolution][im_idx]:\n                    # search for a valid image\n                    ordered_video = False\n                    random_direction = 2 * rng.choice(2) - 1\n                    for offset in range(1, len(image_pool)):\n                        tentative_im_idx = (im_idx + (random_direction * offset)) % len(\n                            image_pool\n                        )\n                        if not self.invalidate[obj, instance][resolution][\n                            tentative_im_idx\n                        ]:\n                            im_idx = tentative_im_idx\n                            break\n                view_idx = image_pool[im_idx]\n                impath = self._get_impath(obj, instance, view_idx)\n                depthpath = self._get_depthpath(obj, instance, view_idx)\n\n                # load camera params\n                metadata_path = self._get_metadatapath(obj, instance, view_idx)\n                input_metadata = np.load(metadata_path)\n                camera_pose = input_metadata[\"camera_pose\"].astype(np.float32)\n                intrinsics = input_metadata[\"camera_intrinsics\"].astype(np.float32)\n\n                # load image and depth\n                rgb_image = imread_cv2(impath)\n                depthmap = self._read_depthmap(depthpath, input_metadata)\n\n                if mask_bg:\n                    # load object mask\n                    maskpath = self._get_maskpath(obj, instance, view_idx)\n                    maskmap = imread_cv2(maskpath, cv2.IMREAD_UNCHANGED).astype(\n                        np.float32\n                    )\n                    maskmap = (maskmap / 255.0) > 0.1\n\n                    # update the depthmap with mask\n                    depthmap *= maskmap\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=impath\n                )\n                num_valid = (depthmap > 0.0).sum()\n                if num_valid == 0:\n                    # problem, invalidate image and retry\n                    self.invalidate[obj, instance][resolution][im_idx] = True\n                    imgs_idxs.append(im_idx)\n                    continue\n\n                # generate img mask and raymap mask\n                img_mask, ray_mask = self.get_img_and_ray_masks(\n                    self.is_metric, len(views), rng\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap,\n                        camera_pose=camera_pose,\n                        camera_intrinsics=intrinsics,\n                        dataset=self.dataset_label,\n                        label=osp.join(obj, instance),\n                        instance=osp.split(impath)[1],\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.9, dtype=np.float32),\n                        img_mask=img_mask,\n                        ray_mask=ray_mask,\n                        camera_only=False,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n\n            if len(views) == num_views and not all(\n                [view[\"instance\"] == views[0][\"instance\"] for view in views]\n            ):\n                invalid_seq = False\n        return views","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hypersim","uri":"program://Human3R/module/src.dust3r.datasets.hypersim#L1-L141","kind":"module","name":"src.dust3r.datasets.hypersim","path":"src/dust3r/datasets/hypersim.py","language":"python","start_line":1,"end_line":141,"context_start_line":1,"context_end_line":141,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HyperSim_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.all_scenes = sorted(\n            [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))]\n        )\n        subscenes = []\n        for scene in self.all_scenes:\n            # not empty\n            subscenes.extend(\n                [\n                    osp.join(scene, f)\n                    for f in os.listdir(osp.join(self.ROOT, scene))\n                    if os.path.isdir(osp.join(self.ROOT, scene, f))\n                    and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0\n                ]\n            )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        j = 0\n        for scene_idx, scene in enumerate(subscenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_paths = sorted([f for f in os.listdir(scene_dir) if f.endswith(\".png\")])\n            assert len(rgb_paths) > 0, f\"{scene_dir} is empty.\"\n            num_imgs = len(rgb_paths)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.start_img_ids = start_img_ids\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n\n            rgb_path = self.images[view_idx]\n            depth_path = rgb_path.replace(\"rgb.png\", \"depth.npy\")\n            cam_path = rgb_path.replace(\"rgb.png\", \"cam.npz\")\n\n            rgb_image = imread_cv2(osp.join(scene_dir, rgb_path), cv2.IMREAD_COLOR)\n            depthmap = np.load(osp.join(scene_dir, depth_path)).astype(np.float32)\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            cam_file = np.load(osp.join(scene_dir, cam_path))\n            intrinsics = cam_file[\"intrinsics\"].astype(np.float32)\n            camera_pose = cam_file[\"pose\"].astype(np.float32)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"hypersim\",\n                    label=self.scenes[scene_id] + \"_\" + rgb_path,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"43d4851eab94fb053c0d919adca3d534f7ac04532b35396ddb3adbf974cfda74","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hypersim.HyperSim_Multi","uri":"program://Human3R/class/src.dust3r.datasets.hypersim.HyperSim_Multi#L14-L141","kind":"class","name":"HyperSim_Multi","path":"src/dust3r/datasets/hypersim.py","language":"python","start_line":14,"end_line":141,"context_start_line":1,"context_end_line":141,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HyperSim_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.all_scenes = sorted(\n            [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))]\n        )\n        subscenes = []\n        for scene in self.all_scenes:\n            # not empty\n            subscenes.extend(\n                [\n                    osp.join(scene, f)\n                    for f in os.listdir(osp.join(self.ROOT, scene))\n                    if os.path.isdir(osp.join(self.ROOT, scene, f))\n                    and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0\n                ]\n            )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        j = 0\n        for scene_idx, scene in enumerate(subscenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_paths = sorted([f for f in os.listdir(scene_dir) if f.endswith(\".png\")])\n            assert len(rgb_paths) > 0, f\"{scene_dir} is empty.\"\n            num_imgs = len(rgb_paths)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.start_img_ids = start_img_ids\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n\n            rgb_path = self.images[view_idx]\n            depth_path = rgb_path.replace(\"rgb.png\", \"depth.npy\")\n            cam_path = rgb_path.replace(\"rgb.png\", \"cam.npz\")\n\n            rgb_image = imread_cv2(osp.join(scene_dir, rgb_path), cv2.IMREAD_COLOR)\n            depthmap = np.load(osp.join(scene_dir, depth_path)).astype(np.float32)\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            cam_file = np.load(osp.join(scene_dir, cam_path))\n            intrinsics = cam_file[\"intrinsics\"].astype(np.float32)\n            camera_pose = cam_file[\"pose\"].astype(np.float32)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"hypersim\",\n                    label=self.scenes[scene_id] + \"_\" + rgb_path,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"43d4851eab94fb053c0d919adca3d534f7ac04532b35396ddb3adbf974cfda74","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hypersim.__init__","uri":"program://Human3R/function/src.dust3r.datasets.hypersim.__init__#L15-L22","kind":"function","name":"__init__","path":"src/dust3r/datasets/hypersim.py","language":"python","start_line":15,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HyperSim_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.all_scenes = sorted(\n            [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))]\n        )\n        subscenes = []\n        for scene in self.all_scenes:\n            # not empty\n            subscenes.extend(\n                [\n                    osp.join(scene, f)\n                    for f in os.listdir(osp.join(self.ROOT, scene))\n                    if os.path.isdir(osp.join(self.ROOT, scene, f))\n                    and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0\n                ]\n            )\n\n        offset = 0\n        scenes = []\n        sceneids = []","source_hash":"43d4851eab94fb053c0d919adca3d534f7ac04532b35396ddb3adbf974cfda74","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hypersim._load_data","uri":"program://Human3R/function/src.dust3r.datasets.hypersim._load_data#L24-L73","kind":"function","name":"_load_data","path":"src/dust3r/datasets/hypersim.py","language":"python","start_line":24,"end_line":73,"context_start_line":4,"context_end_line":93,"code":"import itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HyperSim_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.all_scenes = sorted(\n            [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))]\n        )\n        subscenes = []\n        for scene in self.all_scenes:\n            # not empty\n            subscenes.extend(\n                [\n                    osp.join(scene, f)\n                    for f in os.listdir(osp.join(self.ROOT, scene))\n                    if os.path.isdir(osp.join(self.ROOT, scene, f))\n                    and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0\n                ]\n            )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        j = 0\n        for scene_idx, scene in enumerate(subscenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_paths = sorted([f for f in os.listdir(scene_dir) if f.endswith(\".png\")])\n            assert len(rgb_paths) > 0, f\"{scene_dir} is empty.\"\n            num_imgs = len(rgb_paths)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.start_img_ids = start_img_ids\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )","source_hash":"43d4851eab94fb053c0d919adca3d534f7ac04532b35396ddb3adbf974cfda74","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hypersim.__len__","uri":"program://Human3R/function/src.dust3r.datasets.hypersim.__len__#L75-L76","kind":"function","name":"__len__","path":"src/dust3r/datasets/hypersim.py","language":"python","start_line":75,"end_line":76,"context_start_line":55,"context_end_line":96,"code":"            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.start_img_ids = start_img_ids\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n        views = []\n        for v, view_idx in enumerate(image_idxs):","source_hash":"43d4851eab94fb053c0d919adca3d534f7ac04532b35396ddb3adbf974cfda74","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hypersim.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.hypersim.get_image_num#L78-L79","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/hypersim.py","language":"python","start_line":78,"end_line":79,"context_start_line":58,"context_end_line":99,"code":"            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.start_img_ids = start_img_ids\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n","source_hash":"43d4851eab94fb053c0d919adca3d534f7ac04532b35396ddb3adbf974cfda74","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hypersim._get_views","uri":"program://Human3R/function/src.dust3r.datasets.hypersim._get_views#L81-L141","kind":"function","name":"_get_views","path":"src/dust3r/datasets/hypersim.py","language":"python","start_line":81,"end_line":141,"context_start_line":61,"context_end_line":141,"code":"            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.start_img_ids = start_img_ids\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n\n            rgb_path = self.images[view_idx]\n            depth_path = rgb_path.replace(\"rgb.png\", \"depth.npy\")\n            cam_path = rgb_path.replace(\"rgb.png\", \"cam.npz\")\n\n            rgb_image = imread_cv2(osp.join(scene_dir, rgb_path), cv2.IMREAD_COLOR)\n            depthmap = np.load(osp.join(scene_dir, depth_path)).astype(np.float32)\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            cam_file = np.load(osp.join(scene_dir, cam_path))\n            intrinsics = cam_file[\"intrinsics\"].astype(np.float32)\n            camera_pose = cam_file[\"pose\"].astype(np.float32)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"hypersim\",\n                    label=self.scenes[scene_id] + \"_\" + rgb_path,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"43d4851eab94fb053c0d919adca3d534f7ac04532b35396ddb3adbf974cfda74","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes_highres","uri":"program://Human3R/module/src.dust3r.datasets.arkitscenes_highres#L1-L175","kind":"module","name":"src.dust3r.datasets.arkitscenes_highres","path":"src/dust3r/datasets/arkitscenes_highres.py","language":"python","start_line":1,"end_line":175,"context_start_line":1,"context_end_line":175,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\nimport h5py\nimport math\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ARKitScenesHighRes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 8\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Validation\"\n        else:\n            raise ValueError(\"\")\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        all_scenes = sorted(\n            [\n                d\n                for d in os.listdir(osp.join(self.ROOT, split))\n                if osp.isdir(osp.join(self.ROOT, split, d))\n            ]\n        )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        timestamps = []\n        intrinsics = []\n        trajectories = []\n        scene_id = 0\n        for scene in all_scenes:\n            scene_dir = osp.join(self.ROOT, self.split, scene)\n            with np.load(osp.join(scene_dir, \"scene_metadata.npz\")) as data:\n                imgs_with_indices = sorted(\n                    enumerate(data[\"images\"]), key=lambda x: x[1]\n                )\n                imgs = [x[1] for x in imgs_with_indices]\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                if len(imgs) < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n                indices = [x[0] for x in imgs_with_indices]\n                tsps = np.array(\n                    [float(img_name.split(\"_\")[1][:-4]) for img_name in imgs]\n                )\n                assert [img[:8] == scene for img in imgs], f\"{scene}, {imgs}\"\n                num_imgs = data[\"images\"].shape[0]\n                img_ids = list(np.arange(num_imgs) + offset)\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                scenes.append(scene)\n                scene_img_list.append(img_ids)\n                sceneids.extend([scene_id] * num_imgs)\n                images.extend(imgs)\n                start_img_ids.extend(start_img_ids_)\n                timestamps.extend(tsps)\n\n                K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0)\n                intrins = data[\"intrinsics\"][indices]\n                K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins]\n                K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins]\n                K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins]\n                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]\n                intrinsics.extend(list(K))\n                trajectories.extend(list(data[\"trajectories\"][indices]))\n\n                # offset groups\n                offset += num_imgs\n                scene_id += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.start_img_ids = start_img_ids\n        assert len(self.images) == len(self.intrinsics) == len(self.trajectories)\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n\n            intrinsics = self.intrinsics[view_idx]\n            camera_pose = self.trajectories[view_idx]\n            basename = self.images[view_idx]\n            assert (\n                basename[:8] == self.scenes[scene_id]\n            ), f\"{basename}, {self.scenes[scene_id]}\"\n            # print(scene_dir, basename)\n            # Load RGB image\n            rgb_image = imread_cv2(\n                osp.join(scene_dir, \"vga_wide\", basename.replace(\".png\", \".jpg\"))\n            )\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(scene_dir, \"highres_depth\", basename), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000.0\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.7, 0.25, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"arkitscenes_highres\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"f79e8f322b84b677ad3ac4f2c2a30de694db17110ad601a2da641bfd1592b14c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes_highres.ARKitScenesHighRes_Multi","uri":"program://Human3R/class/src.dust3r.datasets.arkitscenes_highres.ARKitScenesHighRes_Multi#L15-L175","kind":"class","name":"ARKitScenesHighRes_Multi","path":"src/dust3r/datasets/arkitscenes_highres.py","language":"python","start_line":15,"end_line":175,"context_start_line":1,"context_end_line":175,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\nimport h5py\nimport math\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ARKitScenesHighRes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 8\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Validation\"\n        else:\n            raise ValueError(\"\")\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        all_scenes = sorted(\n            [\n                d\n                for d in os.listdir(osp.join(self.ROOT, split))\n                if osp.isdir(osp.join(self.ROOT, split, d))\n            ]\n        )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        timestamps = []\n        intrinsics = []\n        trajectories = []\n        scene_id = 0\n        for scene in all_scenes:\n            scene_dir = osp.join(self.ROOT, self.split, scene)\n            with np.load(osp.join(scene_dir, \"scene_metadata.npz\")) as data:\n                imgs_with_indices = sorted(\n                    enumerate(data[\"images\"]), key=lambda x: x[1]\n                )\n                imgs = [x[1] for x in imgs_with_indices]\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                if len(imgs) < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n                indices = [x[0] for x in imgs_with_indices]\n                tsps = np.array(\n                    [float(img_name.split(\"_\")[1][:-4]) for img_name in imgs]\n                )\n                assert [img[:8] == scene for img in imgs], f\"{scene}, {imgs}\"\n                num_imgs = data[\"images\"].shape[0]\n                img_ids = list(np.arange(num_imgs) + offset)\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                scenes.append(scene)\n                scene_img_list.append(img_ids)\n                sceneids.extend([scene_id] * num_imgs)\n                images.extend(imgs)\n                start_img_ids.extend(start_img_ids_)\n                timestamps.extend(tsps)\n\n                K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0)\n                intrins = data[\"intrinsics\"][indices]\n                K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins]\n                K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins]\n                K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins]\n                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]\n                intrinsics.extend(list(K))\n                trajectories.extend(list(data[\"trajectories\"][indices]))\n\n                # offset groups\n                offset += num_imgs\n                scene_id += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.start_img_ids = start_img_ids\n        assert len(self.images) == len(self.intrinsics) == len(self.trajectories)\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n\n            intrinsics = self.intrinsics[view_idx]\n            camera_pose = self.trajectories[view_idx]\n            basename = self.images[view_idx]\n            assert (\n                basename[:8] == self.scenes[scene_id]\n            ), f\"{basename}, {self.scenes[scene_id]}\"\n            # print(scene_dir, basename)\n            # Load RGB image\n            rgb_image = imread_cv2(\n                osp.join(scene_dir, \"vga_wide\", basename.replace(\".png\", \".jpg\"))\n            )\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(scene_dir, \"highres_depth\", basename), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000.0\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.7, 0.25, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"arkitscenes_highres\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"f79e8f322b84b677ad3ac4f2c2a30de694db17110ad601a2da641bfd1592b14c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes_highres.__init__","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes_highres.__init__#L16-L29","kind":"function","name":"__init__","path":"src/dust3r/datasets/arkitscenes_highres.py","language":"python","start_line":16,"end_line":29,"context_start_line":1,"context_end_line":49,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\nimport h5py\nimport math\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ARKitScenesHighRes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 8\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Validation\"\n        else:\n            raise ValueError(\"\")\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        all_scenes = sorted(\n            [\n                d\n                for d in os.listdir(osp.join(self.ROOT, split))\n                if osp.isdir(osp.join(self.ROOT, split, d))\n            ]\n        )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        timestamps = []\n        intrinsics = []\n        trajectories = []\n        scene_id = 0\n        for scene in all_scenes:","source_hash":"f79e8f322b84b677ad3ac4f2c2a30de694db17110ad601a2da641bfd1592b14c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes_highres._load_data","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes_highres._load_data#L31-L100","kind":"function","name":"_load_data","path":"src/dust3r/datasets/arkitscenes_highres.py","language":"python","start_line":31,"end_line":100,"context_start_line":11,"context_end_line":120,"code":"from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ARKitScenesHighRes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 8\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Validation\"\n        else:\n            raise ValueError(\"\")\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        all_scenes = sorted(\n            [\n                d\n                for d in os.listdir(osp.join(self.ROOT, split))\n                if osp.isdir(osp.join(self.ROOT, split, d))\n            ]\n        )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        timestamps = []\n        intrinsics = []\n        trajectories = []\n        scene_id = 0\n        for scene in all_scenes:\n            scene_dir = osp.join(self.ROOT, self.split, scene)\n            with np.load(osp.join(scene_dir, \"scene_metadata.npz\")) as data:\n                imgs_with_indices = sorted(\n                    enumerate(data[\"images\"]), key=lambda x: x[1]\n                )\n                imgs = [x[1] for x in imgs_with_indices]\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                if len(imgs) < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n                indices = [x[0] for x in imgs_with_indices]\n                tsps = np.array(\n                    [float(img_name.split(\"_\")[1][:-4]) for img_name in imgs]\n                )\n                assert [img[:8] == scene for img in imgs], f\"{scene}, {imgs}\"\n                num_imgs = data[\"images\"].shape[0]\n                img_ids = list(np.arange(num_imgs) + offset)\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                scenes.append(scene)\n                scene_img_list.append(img_ids)\n                sceneids.extend([scene_id] * num_imgs)\n                images.extend(imgs)\n                start_img_ids.extend(start_img_ids_)\n                timestamps.extend(tsps)\n\n                K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0)\n                intrins = data[\"intrinsics\"][indices]\n                K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins]\n                K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins]\n                K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins]\n                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]\n                intrinsics.extend(list(K))\n                trajectories.extend(list(data[\"trajectories\"][indices]))\n\n                # offset groups\n                offset += num_imgs\n                scene_id += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.start_img_ids = start_img_ids\n        assert len(self.images) == len(self.intrinsics) == len(self.trajectories)\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n","source_hash":"f79e8f322b84b677ad3ac4f2c2a30de694db17110ad601a2da641bfd1592b14c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes_highres.__len__","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes_highres.__len__#L102-L103","kind":"function","name":"__len__","path":"src/dust3r/datasets/arkitscenes_highres.py","language":"python","start_line":102,"end_line":103,"context_start_line":82,"context_end_line":123,"code":"                K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins]\n                K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins]\n                K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins]\n                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]\n                intrinsics.extend(list(K))\n                trajectories.extend(list(data[\"trajectories\"][indices]))\n\n                # offset groups\n                offset += num_imgs\n                scene_id += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.start_img_ids = start_img_ids\n        assert len(self.images) == len(self.intrinsics) == len(self.trajectories)\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):","source_hash":"f79e8f322b84b677ad3ac4f2c2a30de694db17110ad601a2da641bfd1592b14c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes_highres.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes_highres.get_image_num#L105-L106","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/arkitscenes_highres.py","language":"python","start_line":105,"end_line":106,"context_start_line":85,"context_end_line":126,"code":"                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]\n                intrinsics.extend(list(K))\n                trajectories.extend(list(data[\"trajectories\"][indices]))\n\n                # offset groups\n                offset += num_imgs\n                scene_id += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.start_img_ids = start_img_ids\n        assert len(self.images) == len(self.intrinsics) == len(self.trajectories)\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n","source_hash":"f79e8f322b84b677ad3ac4f2c2a30de694db17110ad601a2da641bfd1592b14c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes_highres._get_views","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes_highres._get_views#L108-L175","kind":"function","name":"_get_views","path":"src/dust3r/datasets/arkitscenes_highres.py","language":"python","start_line":108,"end_line":175,"context_start_line":88,"context_end_line":175,"code":"\n                # offset groups\n                offset += num_imgs\n                scene_id += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.scene_img_list = scene_img_list\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.start_img_ids = start_img_ids\n        assert len(self.images) == len(self.intrinsics) == len(self.trajectories)\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n\n            intrinsics = self.intrinsics[view_idx]\n            camera_pose = self.trajectories[view_idx]\n            basename = self.images[view_idx]\n            assert (\n                basename[:8] == self.scenes[scene_id]\n            ), f\"{basename}, {self.scenes[scene_id]}\"\n            # print(scene_dir, basename)\n            # Load RGB image\n            rgb_image = imread_cv2(\n                osp.join(scene_dir, \"vga_wide\", basename.replace(\".png\", \".jpg\"))\n            )\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(scene_dir, \"highres_depth\", basename), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000.0\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.7, 0.25, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"arkitscenes_highres\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"f79e8f322b84b677ad3ac4f2c2a30de694db17110ad601a2da641bfd1592b14c","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree","uri":"program://Human3R/module/src.dust3r.datasets.mapfree#L1-L282","kind":"module","name":"src.dust3r.datasets.mapfree","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":1,"end_line":282,"context_start_line":1,"context_end_line":282,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nimport pickle\nimport h5py\nfrom tqdm import tqdm\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MapFree_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self._load_data()\n\n    def imgid2path(self, img_id, scene):\n        first_seq_id, first_frame_id = img_id\n        return os.path.join(\n            self.ROOT,\n            scene,\n            f\"dense{first_seq_id}\",\n            \"rgb\",\n            f\"frame_{first_frame_id:05d}.jpg\",\n        )\n\n    def path2imgid(self, subscene, filename):\n        first_seq_id = int(subscene[5:])\n        first_frame_id = int(filename[6:-4])\n        return [first_seq_id, first_frame_id]\n\n    def _load_data(self):\n        cache_file = f\"{self.ROOT}/cached_metadata_50_col_only.h5\"\n        if os.path.exists(cache_file):\n            print(f\"Loading cached metadata from {cache_file}\")\n            with h5py.File(cache_file, \"r\") as hf:\n                self.scenes = list(map(lambda x: x.decode(\"utf-8\"), hf[\"scenes\"][:]))\n                self.sceneids = hf[\"sceneids\"][:]\n                self.scope = hf[\"scope\"][:]\n                self.video_flags = hf[\"video_flags\"][:]\n                self.groups = hf[\"groups\"][:]\n                self.id_ranges = hf[\"id_ranges\"][:]\n                self.images = hf[\"images\"][:]\n        else:\n            scene_dirs = sorted(\n                [\n                    d\n                    for d in os.listdir(self.ROOT)\n                    if os.path.isdir(os.path.join(self.ROOT, d))\n                ]\n            )\n            scenes = []\n            sceneids = []\n            groups = []\n            scope = []\n            images = []\n            id_ranges = []\n            is_video = []\n            start = 0\n            j = 0\n            offset = 0\n\n            for scene in tqdm(scene_dirs):\n                scenes.append(scene)\n                # video sequences\n                subscenes = sorted(\n                    [\n                        d\n                        for d in os.listdir(os.path.join(self.ROOT, scene))\n                        if d.startswith(\"dense\")\n                    ]\n                )\n                id_range_subscenes = []\n                for subscene in subscenes:\n                    rgb_paths = sorted(\n                        [\n                            d\n                            for d in os.listdir(\n                                os.path.join(self.ROOT, scene, subscene, \"rgb\")\n                            )\n                            if d.endswith(\".jpg\")\n                        ]\n                    )\n                    assert (\n                        len(rgb_paths) > 0\n                    ), f\"{os.path.join(self.ROOT, scene, subscene)} is empty.\"\n                    num_imgs = len(rgb_paths)\n                    images.extend(\n                        [self.path2imgid(subscene, rgb_path) for rgb_path in rgb_paths]\n                    )\n                    id_range_subscenes.append((offset, offset + num_imgs))\n                    offset += num_imgs\n\n                # image collections\n                metadata = pickle.load(\n                    open(os.path.join(self.ROOT, scene, \"metadata.pkl\"), \"rb\")\n                )\n                ref_imgs = list(metadata.keys())\n                img_groups = []\n                for ref_img in ref_imgs:\n                    other_imgs = metadata[ref_img]\n                    if len(other_imgs) + 1 < self.num_views:\n                        continue\n                    group = [(*other_img[0], other_img[1]) for other_img in other_imgs]\n                    group.insert(0, (*ref_img, 1))\n                    img_groups.append(np.array(group))\n                    id_ranges.append(id_range_subscenes[ref_img[0]])\n                    scope.append(start)\n                    start = start + len(group)\n\n                num_groups = len(img_groups)\n                sceneids.extend([j] * num_groups)\n                groups.extend(img_groups)\n                is_video.extend([False] * num_groups)\n                j += 1\n\n            self.scenes = np.array(scenes)\n            self.sceneids = np.array(sceneids)\n            self.scope = np.array(scope)\n            self.video_flags = np.array(is_video)\n            self.groups = np.concatenate(groups, 0)\n            self.id_ranges = np.array(id_ranges)\n            self.images = np.array(images)\n\n            data = dict(\n                scenes=self.scenes,\n                sceneids=self.sceneids,\n                scope=self.scope,\n                video_flags=self.video_flags,\n                groups=self.groups,\n                id_ranges=self.id_ranges,\n                images=self.images,\n            )\n\n            with h5py.File(cache_file, \"w\") as h5f:\n                h5f.create_dataset(\n                    \"scenes\",\n                    data=data[\"scenes\"].astype(object),\n                    dtype=h5py.string_dtype(encoding=\"utf-8\"),\n                    compression=\"lzf\",\n                    chunks=True,\n                )\n                h5f.create_dataset(\n                    \"sceneids\", data=data[\"sceneids\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"scope\", data=data[\"scope\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"video_flags\",\n                    data=data[\"video_flags\"],\n                    compression=\"lzf\",\n                    chunks=True,\n                )\n                h5f.create_dataset(\n                    \"groups\", data=data[\"groups\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"id_ranges\", data=data[\"id_ranges\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"images\", data=data[\"images\"], compression=\"lzf\", chunks=True\n                )\n\n    def __len__(self):\n        return len(self.scope)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene = self.scenes[self.sceneids[idx]]\n        if rng.random() < 0.6:\n            ids = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_ids = ids[: len(ids) - cut_off + 1]\n            start_id = rng.choice(start_ids)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                ids.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            ids = np.array(ids)[pos]\n            image_idxs = self.images[ids]\n        else:\n            ordered_video = False\n            seq_start_index = self.scope[idx]\n            seq_end_index = self.scope[idx + 1] if idx < len(self.scope) - 1 else None\n            image_idxs = (\n                self.groups[seq_start_index:seq_end_index]\n                if seq_end_index is not None\n                else self.groups[seq_start_index:]\n            )\n            image_idxs, overlap_scores = image_idxs[:, :2], image_idxs[:, 2]\n            replace = (\n                True\n                if self.allow_repeat\n                or len(overlap_scores[overlap_scores > 0]) < num_views\n                else False\n            )\n            image_idxs = rng.choice(\n                image_idxs,\n                num_views,\n                replace=replace,\n                p=overlap_scores / np.sum(overlap_scores),\n            )\n            image_idxs = image_idxs.astype(np.int64)\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            img_path = self.imgid2path(view_idx, scene)\n            depth_path = img_path.replace(\"rgb\", \"depth\").replace(\".jpg\", \".npy\")\n            cam_path = img_path.replace(\"rgb\", \"cam\").replace(\".jpg\", \".npz\")\n            sky_mask_path = img_path.replace(\"rgb\", \"sky_mask\")\n            image = imread_cv2(img_path)\n            depthmap = np.load(depth_path)\n            camera_params = np.load(cam_path)\n            sky_mask = cv2.imread(sky_mask_path, cv2.IMREAD_UNCHANGED) >= 127\n\n            intrinsics = camera_params[\"intrinsic\"].astype(np.float32)\n            camera_pose = camera_params[\"pose\"].astype(np.float32)\n\n            depthmap[sky_mask] = -1.0\n            depthmap[depthmap > 400.0] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(img_path)\n            )\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"MapFree\",\n                    label=img_path,\n                    is_metric=self.is_metric,\n                    instance=img_path,\n                    is_video=ordered_video,\n                    quantile=np.array(0.96, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree.MapFree_Multi","uri":"program://Human3R/class/src.dust3r.datasets.mapfree.MapFree_Multi#L18-L282","kind":"class","name":"MapFree_Multi","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":18,"end_line":282,"context_start_line":1,"context_end_line":282,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nimport pickle\nimport h5py\nfrom tqdm import tqdm\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MapFree_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self._load_data()\n\n    def imgid2path(self, img_id, scene):\n        first_seq_id, first_frame_id = img_id\n        return os.path.join(\n            self.ROOT,\n            scene,\n            f\"dense{first_seq_id}\",\n            \"rgb\",\n            f\"frame_{first_frame_id:05d}.jpg\",\n        )\n\n    def path2imgid(self, subscene, filename):\n        first_seq_id = int(subscene[5:])\n        first_frame_id = int(filename[6:-4])\n        return [first_seq_id, first_frame_id]\n\n    def _load_data(self):\n        cache_file = f\"{self.ROOT}/cached_metadata_50_col_only.h5\"\n        if os.path.exists(cache_file):\n            print(f\"Loading cached metadata from {cache_file}\")\n            with h5py.File(cache_file, \"r\") as hf:\n                self.scenes = list(map(lambda x: x.decode(\"utf-8\"), hf[\"scenes\"][:]))\n                self.sceneids = hf[\"sceneids\"][:]\n                self.scope = hf[\"scope\"][:]\n                self.video_flags = hf[\"video_flags\"][:]\n                self.groups = hf[\"groups\"][:]\n                self.id_ranges = hf[\"id_ranges\"][:]\n                self.images = hf[\"images\"][:]\n        else:\n            scene_dirs = sorted(\n                [\n                    d\n                    for d in os.listdir(self.ROOT)\n                    if os.path.isdir(os.path.join(self.ROOT, d))\n                ]\n            )\n            scenes = []\n            sceneids = []\n            groups = []\n            scope = []\n            images = []\n            id_ranges = []\n            is_video = []\n            start = 0\n            j = 0\n            offset = 0\n\n            for scene in tqdm(scene_dirs):\n                scenes.append(scene)\n                # video sequences\n                subscenes = sorted(\n                    [\n                        d\n                        for d in os.listdir(os.path.join(self.ROOT, scene))\n                        if d.startswith(\"dense\")\n                    ]\n                )\n                id_range_subscenes = []\n                for subscene in subscenes:\n                    rgb_paths = sorted(\n                        [\n                            d\n                            for d in os.listdir(\n                                os.path.join(self.ROOT, scene, subscene, \"rgb\")\n                            )\n                            if d.endswith(\".jpg\")\n                        ]\n                    )\n                    assert (\n                        len(rgb_paths) > 0\n                    ), f\"{os.path.join(self.ROOT, scene, subscene)} is empty.\"\n                    num_imgs = len(rgb_paths)\n                    images.extend(\n                        [self.path2imgid(subscene, rgb_path) for rgb_path in rgb_paths]\n                    )\n                    id_range_subscenes.append((offset, offset + num_imgs))\n                    offset += num_imgs\n\n                # image collections\n                metadata = pickle.load(\n                    open(os.path.join(self.ROOT, scene, \"metadata.pkl\"), \"rb\")\n                )\n                ref_imgs = list(metadata.keys())\n                img_groups = []\n                for ref_img in ref_imgs:\n                    other_imgs = metadata[ref_img]\n                    if len(other_imgs) + 1 < self.num_views:\n                        continue\n                    group = [(*other_img[0], other_img[1]) for other_img in other_imgs]\n                    group.insert(0, (*ref_img, 1))\n                    img_groups.append(np.array(group))\n                    id_ranges.append(id_range_subscenes[ref_img[0]])\n                    scope.append(start)\n                    start = start + len(group)\n\n                num_groups = len(img_groups)\n                sceneids.extend([j] * num_groups)\n                groups.extend(img_groups)\n                is_video.extend([False] * num_groups)\n                j += 1\n\n            self.scenes = np.array(scenes)\n            self.sceneids = np.array(sceneids)\n            self.scope = np.array(scope)\n            self.video_flags = np.array(is_video)\n            self.groups = np.concatenate(groups, 0)\n            self.id_ranges = np.array(id_ranges)\n            self.images = np.array(images)\n\n            data = dict(\n                scenes=self.scenes,\n                sceneids=self.sceneids,\n                scope=self.scope,\n                video_flags=self.video_flags,\n                groups=self.groups,\n                id_ranges=self.id_ranges,\n                images=self.images,\n            )\n\n            with h5py.File(cache_file, \"w\") as h5f:\n                h5f.create_dataset(\n                    \"scenes\",\n                    data=data[\"scenes\"].astype(object),\n                    dtype=h5py.string_dtype(encoding=\"utf-8\"),\n                    compression=\"lzf\",\n                    chunks=True,\n                )\n                h5f.create_dataset(\n                    \"sceneids\", data=data[\"sceneids\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"scope\", data=data[\"scope\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"video_flags\",\n                    data=data[\"video_flags\"],\n                    compression=\"lzf\",\n                    chunks=True,\n                )\n                h5f.create_dataset(\n                    \"groups\", data=data[\"groups\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"id_ranges\", data=data[\"id_ranges\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"images\", data=data[\"images\"], compression=\"lzf\", chunks=True\n                )\n\n    def __len__(self):\n        return len(self.scope)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene = self.scenes[self.sceneids[idx]]\n        if rng.random() < 0.6:\n            ids = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_ids = ids[: len(ids) - cut_off + 1]\n            start_id = rng.choice(start_ids)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                ids.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            ids = np.array(ids)[pos]\n            image_idxs = self.images[ids]\n        else:\n            ordered_video = False\n            seq_start_index = self.scope[idx]\n            seq_end_index = self.scope[idx + 1] if idx < len(self.scope) - 1 else None\n            image_idxs = (\n                self.groups[seq_start_index:seq_end_index]\n                if seq_end_index is not None\n                else self.groups[seq_start_index:]\n            )\n            image_idxs, overlap_scores = image_idxs[:, :2], image_idxs[:, 2]\n            replace = (\n                True\n                if self.allow_repeat\n                or len(overlap_scores[overlap_scores > 0]) < num_views\n                else False\n            )\n            image_idxs = rng.choice(\n                image_idxs,\n                num_views,\n                replace=replace,\n                p=overlap_scores / np.sum(overlap_scores),\n            )\n            image_idxs = image_idxs.astype(np.int64)\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            img_path = self.imgid2path(view_idx, scene)\n            depth_path = img_path.replace(\"rgb\", \"depth\").replace(\".jpg\", \".npy\")\n            cam_path = img_path.replace(\"rgb\", \"cam\").replace(\".jpg\", \".npz\")\n            sky_mask_path = img_path.replace(\"rgb\", \"sky_mask\")\n            image = imread_cv2(img_path)\n            depthmap = np.load(depth_path)\n            camera_params = np.load(cam_path)\n            sky_mask = cv2.imread(sky_mask_path, cv2.IMREAD_UNCHANGED) >= 127\n\n            intrinsics = camera_params[\"intrinsic\"].astype(np.float32)\n            camera_pose = camera_params[\"pose\"].astype(np.float32)\n\n            depthmap[sky_mask] = -1.0\n            depthmap[depthmap > 400.0] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(img_path)\n            )\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"MapFree\",\n                    label=img_path,\n                    is_metric=self.is_metric,\n                    instance=img_path,\n                    is_video=ordered_video,\n                    quantile=np.array(0.96, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree.__init__","uri":"program://Human3R/function/src.dust3r.datasets.mapfree.__init__#L20-L27","kind":"function","name":"__init__","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":20,"end_line":27,"context_start_line":1,"context_end_line":47,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nimport pickle\nimport h5py\nfrom tqdm import tqdm\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MapFree_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self._load_data()\n\n    def imgid2path(self, img_id, scene):\n        first_seq_id, first_frame_id = img_id\n        return os.path.join(\n            self.ROOT,\n            scene,\n            f\"dense{first_seq_id}\",\n            \"rgb\",\n            f\"frame_{first_frame_id:05d}.jpg\",\n        )\n\n    def path2imgid(self, subscene, filename):\n        first_seq_id = int(subscene[5:])\n        first_frame_id = int(filename[6:-4])\n        return [first_seq_id, first_frame_id]\n\n    def _load_data(self):\n        cache_file = f\"{self.ROOT}/cached_metadata_50_col_only.h5\"\n        if os.path.exists(cache_file):\n            print(f\"Loading cached metadata from {cache_file}\")","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree.imgid2path","uri":"program://Human3R/function/src.dust3r.datasets.mapfree.imgid2path#L29-L37","kind":"function","name":"imgid2path","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":29,"end_line":37,"context_start_line":9,"context_end_line":57,"code":"import h5py\nfrom tqdm import tqdm\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MapFree_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self._load_data()\n\n    def imgid2path(self, img_id, scene):\n        first_seq_id, first_frame_id = img_id\n        return os.path.join(\n            self.ROOT,\n            scene,\n            f\"dense{first_seq_id}\",\n            \"rgb\",\n            f\"frame_{first_frame_id:05d}.jpg\",\n        )\n\n    def path2imgid(self, subscene, filename):\n        first_seq_id = int(subscene[5:])\n        first_frame_id = int(filename[6:-4])\n        return [first_seq_id, first_frame_id]\n\n    def _load_data(self):\n        cache_file = f\"{self.ROOT}/cached_metadata_50_col_only.h5\"\n        if os.path.exists(cache_file):\n            print(f\"Loading cached metadata from {cache_file}\")\n            with h5py.File(cache_file, \"r\") as hf:\n                self.scenes = list(map(lambda x: x.decode(\"utf-8\"), hf[\"scenes\"][:]))\n                self.sceneids = hf[\"sceneids\"][:]\n                self.scope = hf[\"scope\"][:]\n                self.video_flags = hf[\"video_flags\"][:]\n                self.groups = hf[\"groups\"][:]\n                self.id_ranges = hf[\"id_ranges\"][:]\n                self.images = hf[\"images\"][:]\n        else:\n            scene_dirs = sorted(","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree.path2imgid","uri":"program://Human3R/function/src.dust3r.datasets.mapfree.path2imgid#L39-L42","kind":"function","name":"path2imgid","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":39,"end_line":42,"context_start_line":19,"context_end_line":62,"code":"\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self._load_data()\n\n    def imgid2path(self, img_id, scene):\n        first_seq_id, first_frame_id = img_id\n        return os.path.join(\n            self.ROOT,\n            scene,\n            f\"dense{first_seq_id}\",\n            \"rgb\",\n            f\"frame_{first_frame_id:05d}.jpg\",\n        )\n\n    def path2imgid(self, subscene, filename):\n        first_seq_id = int(subscene[5:])\n        first_frame_id = int(filename[6:-4])\n        return [first_seq_id, first_frame_id]\n\n    def _load_data(self):\n        cache_file = f\"{self.ROOT}/cached_metadata_50_col_only.h5\"\n        if os.path.exists(cache_file):\n            print(f\"Loading cached metadata from {cache_file}\")\n            with h5py.File(cache_file, \"r\") as hf:\n                self.scenes = list(map(lambda x: x.decode(\"utf-8\"), hf[\"scenes\"][:]))\n                self.sceneids = hf[\"sceneids\"][:]\n                self.scope = hf[\"scope\"][:]\n                self.video_flags = hf[\"video_flags\"][:]\n                self.groups = hf[\"groups\"][:]\n                self.id_ranges = hf[\"id_ranges\"][:]\n                self.images = hf[\"images\"][:]\n        else:\n            scene_dirs = sorted(\n                [\n                    d\n                    for d in os.listdir(self.ROOT)\n                    if os.path.isdir(os.path.join(self.ROOT, d))\n                ]","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree._load_data","uri":"program://Human3R/function/src.dust3r.datasets.mapfree._load_data#L44-L175","kind":"function","name":"_load_data","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":44,"end_line":175,"context_start_line":24,"context_end_line":195,"code":"        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self._load_data()\n\n    def imgid2path(self, img_id, scene):\n        first_seq_id, first_frame_id = img_id\n        return os.path.join(\n            self.ROOT,\n            scene,\n            f\"dense{first_seq_id}\",\n            \"rgb\",\n            f\"frame_{first_frame_id:05d}.jpg\",\n        )\n\n    def path2imgid(self, subscene, filename):\n        first_seq_id = int(subscene[5:])\n        first_frame_id = int(filename[6:-4])\n        return [first_seq_id, first_frame_id]\n\n    def _load_data(self):\n        cache_file = f\"{self.ROOT}/cached_metadata_50_col_only.h5\"\n        if os.path.exists(cache_file):\n            print(f\"Loading cached metadata from {cache_file}\")\n            with h5py.File(cache_file, \"r\") as hf:\n                self.scenes = list(map(lambda x: x.decode(\"utf-8\"), hf[\"scenes\"][:]))\n                self.sceneids = hf[\"sceneids\"][:]\n                self.scope = hf[\"scope\"][:]\n                self.video_flags = hf[\"video_flags\"][:]\n                self.groups = hf[\"groups\"][:]\n                self.id_ranges = hf[\"id_ranges\"][:]\n                self.images = hf[\"images\"][:]\n        else:\n            scene_dirs = sorted(\n                [\n                    d\n                    for d in os.listdir(self.ROOT)\n                    if os.path.isdir(os.path.join(self.ROOT, d))\n                ]\n            )\n            scenes = []\n            sceneids = []\n            groups = []\n            scope = []\n            images = []\n            id_ranges = []\n            is_video = []\n            start = 0\n            j = 0\n            offset = 0\n\n            for scene in tqdm(scene_dirs):\n                scenes.append(scene)\n                # video sequences\n                subscenes = sorted(\n                    [\n                        d\n                        for d in os.listdir(os.path.join(self.ROOT, scene))\n                        if d.startswith(\"dense\")\n                    ]\n                )\n                id_range_subscenes = []\n                for subscene in subscenes:\n                    rgb_paths = sorted(\n                        [\n                            d\n                            for d in os.listdir(\n                                os.path.join(self.ROOT, scene, subscene, \"rgb\")\n                            )\n                            if d.endswith(\".jpg\")\n                        ]\n                    )\n                    assert (\n                        len(rgb_paths) > 0\n                    ), f\"{os.path.join(self.ROOT, scene, subscene)} is empty.\"\n                    num_imgs = len(rgb_paths)\n                    images.extend(\n                        [self.path2imgid(subscene, rgb_path) for rgb_path in rgb_paths]\n                    )\n                    id_range_subscenes.append((offset, offset + num_imgs))\n                    offset += num_imgs\n\n                # image collections\n                metadata = pickle.load(\n                    open(os.path.join(self.ROOT, scene, \"metadata.pkl\"), \"rb\")\n                )\n                ref_imgs = list(metadata.keys())\n                img_groups = []\n                for ref_img in ref_imgs:\n                    other_imgs = metadata[ref_img]\n                    if len(other_imgs) + 1 < self.num_views:\n                        continue\n                    group = [(*other_img[0], other_img[1]) for other_img in other_imgs]\n                    group.insert(0, (*ref_img, 1))\n                    img_groups.append(np.array(group))\n                    id_ranges.append(id_range_subscenes[ref_img[0]])\n                    scope.append(start)\n                    start = start + len(group)\n\n                num_groups = len(img_groups)\n                sceneids.extend([j] * num_groups)\n                groups.extend(img_groups)\n                is_video.extend([False] * num_groups)\n                j += 1\n\n            self.scenes = np.array(scenes)\n            self.sceneids = np.array(sceneids)\n            self.scope = np.array(scope)\n            self.video_flags = np.array(is_video)\n            self.groups = np.concatenate(groups, 0)\n            self.id_ranges = np.array(id_ranges)\n            self.images = np.array(images)\n\n            data = dict(\n                scenes=self.scenes,\n                sceneids=self.sceneids,\n                scope=self.scope,\n                video_flags=self.video_flags,\n                groups=self.groups,\n                id_ranges=self.id_ranges,\n                images=self.images,\n            )\n\n            with h5py.File(cache_file, \"w\") as h5f:\n                h5f.create_dataset(\n                    \"scenes\",\n                    data=data[\"scenes\"].astype(object),\n                    dtype=h5py.string_dtype(encoding=\"utf-8\"),\n                    compression=\"lzf\",\n                    chunks=True,\n                )\n                h5f.create_dataset(\n                    \"sceneids\", data=data[\"sceneids\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"scope\", data=data[\"scope\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"video_flags\",\n                    data=data[\"video_flags\"],\n                    compression=\"lzf\",\n                    chunks=True,\n                )\n                h5f.create_dataset(\n                    \"groups\", data=data[\"groups\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"id_ranges\", data=data[\"id_ranges\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"images\", data=data[\"images\"], compression=\"lzf\", chunks=True\n                )\n\n    def __len__(self):\n        return len(self.scope)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene = self.scenes[self.sceneids[idx]]\n        if rng.random() < 0.6:\n            ids = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_ids = ids[: len(ids) - cut_off + 1]\n            start_id = rng.choice(start_ids)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree.__len__","uri":"program://Human3R/function/src.dust3r.datasets.mapfree.__len__#L177-L178","kind":"function","name":"__len__","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":177,"end_line":178,"context_start_line":157,"context_end_line":198,"code":"                )\n                h5f.create_dataset(\n                    \"scope\", data=data[\"scope\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"video_flags\",\n                    data=data[\"video_flags\"],\n                    compression=\"lzf\",\n                    chunks=True,\n                )\n                h5f.create_dataset(\n                    \"groups\", data=data[\"groups\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"id_ranges\", data=data[\"id_ranges\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"images\", data=data[\"images\"], compression=\"lzf\", chunks=True\n                )\n\n    def __len__(self):\n        return len(self.scope)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene = self.scenes[self.sceneids[idx]]\n        if rng.random() < 0.6:\n            ids = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_ids = ids[: len(ids) - cut_off + 1]\n            start_id = rng.choice(start_ids)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                ids.tolist(),\n                rng,\n                max_interval=self.max_interval,","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.mapfree.get_image_num#L180-L181","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":180,"end_line":181,"context_start_line":160,"context_end_line":201,"code":"                )\n                h5f.create_dataset(\n                    \"video_flags\",\n                    data=data[\"video_flags\"],\n                    compression=\"lzf\",\n                    chunks=True,\n                )\n                h5f.create_dataset(\n                    \"groups\", data=data[\"groups\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"id_ranges\", data=data[\"id_ranges\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"images\", data=data[\"images\"], compression=\"lzf\", chunks=True\n                )\n\n    def __len__(self):\n        return len(self.scope)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene = self.scenes[self.sceneids[idx]]\n        if rng.random() < 0.6:\n            ids = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_ids = ids[: len(ids) - cut_off + 1]\n            start_id = rng.choice(start_ids)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                ids.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree.get_stats","uri":"program://Human3R/function/src.dust3r.datasets.mapfree.get_stats#L183-L184","kind":"function","name":"get_stats","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":183,"end_line":184,"context_start_line":163,"context_end_line":204,"code":"                    data=data[\"video_flags\"],\n                    compression=\"lzf\",\n                    chunks=True,\n                )\n                h5f.create_dataset(\n                    \"groups\", data=data[\"groups\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"id_ranges\", data=data[\"id_ranges\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"images\", data=data[\"images\"], compression=\"lzf\", chunks=True\n                )\n\n    def __len__(self):\n        return len(self.scope)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene = self.scenes[self.sceneids[idx]]\n        if rng.random() < 0.6:\n            ids = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_ids = ids[: len(ids) - cut_off + 1]\n            start_id = rng.choice(start_ids)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                ids.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            ids = np.array(ids)[pos]\n            image_idxs = self.images[ids]","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mapfree._get_views","uri":"program://Human3R/function/src.dust3r.datasets.mapfree._get_views#L186-L282","kind":"function","name":"_get_views","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":186,"end_line":282,"context_start_line":166,"context_end_line":282,"code":"                )\n                h5f.create_dataset(\n                    \"groups\", data=data[\"groups\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"id_ranges\", data=data[\"id_ranges\"], compression=\"lzf\", chunks=True\n                )\n                h5f.create_dataset(\n                    \"images\", data=data[\"images\"], compression=\"lzf\", chunks=True\n                )\n\n    def __len__(self):\n        return len(self.scope)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene = self.scenes[self.sceneids[idx]]\n        if rng.random() < 0.6:\n            ids = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_ids = ids[: len(ids) - cut_off + 1]\n            start_id = rng.choice(start_ids)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                ids.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            ids = np.array(ids)[pos]\n            image_idxs = self.images[ids]\n        else:\n            ordered_video = False\n            seq_start_index = self.scope[idx]\n            seq_end_index = self.scope[idx + 1] if idx < len(self.scope) - 1 else None\n            image_idxs = (\n                self.groups[seq_start_index:seq_end_index]\n                if seq_end_index is not None\n                else self.groups[seq_start_index:]\n            )\n            image_idxs, overlap_scores = image_idxs[:, :2], image_idxs[:, 2]\n            replace = (\n                True\n                if self.allow_repeat\n                or len(overlap_scores[overlap_scores > 0]) < num_views\n                else False\n            )\n            image_idxs = rng.choice(\n                image_idxs,\n                num_views,\n                replace=replace,\n                p=overlap_scores / np.sum(overlap_scores),\n            )\n            image_idxs = image_idxs.astype(np.int64)\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            img_path = self.imgid2path(view_idx, scene)\n            depth_path = img_path.replace(\"rgb\", \"depth\").replace(\".jpg\", \".npy\")\n            cam_path = img_path.replace(\"rgb\", \"cam\").replace(\".jpg\", \".npz\")\n            sky_mask_path = img_path.replace(\"rgb\", \"sky_mask\")\n            image = imread_cv2(img_path)\n            depthmap = np.load(depth_path)\n            camera_params = np.load(cam_path)\n            sky_mask = cv2.imread(sky_mask_path, cv2.IMREAD_UNCHANGED) >= 127\n\n            intrinsics = camera_params[\"intrinsic\"].astype(np.float32)\n            camera_pose = camera_params[\"pose\"].astype(np.float32)\n\n            depthmap[sky_mask] = -1.0\n            depthmap[depthmap > 400.0] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(img_path)\n            )\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"MapFree\",\n                    label=img_path,\n                    is_metric=self.is_metric,\n                    instance=img_path,\n                    is_video=ordered_video,\n                    quantile=np.array(0.96, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.synscapes","uri":"program://Human3R/module/src.dust3r.datasets.synscapes#L1-L85","kind":"module","name":"src.dust3r.datasets.synscapes","path":"src/dust3r/datasets/synscapes.py","language":"python","start_line":1,"end_line":85,"context_start_line":1,"context_end_line":85,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SynScapes(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted(\n            [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n            key=lambda x: int(x),\n        )\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            depthmap[depthmap > 200] = 0.0\n\n            intrinsics = np.load(osp.join(self.ROOT, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"synscapes\",\n                    label=img_name,\n                    instance=f\"{str(idx)}_{img_name}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"2c2e047055922150a529b6b5b8b41ad47bd66810fbd6d7806051aa3e05b0f03b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.synscapes.SynScapes","uri":"program://Human3R/class/src.dust3r.datasets.synscapes.SynScapes#L14-L85","kind":"class","name":"SynScapes","path":"src/dust3r/datasets/synscapes.py","language":"python","start_line":14,"end_line":85,"context_start_line":1,"context_end_line":85,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SynScapes(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted(\n            [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n            key=lambda x: int(x),\n        )\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            depthmap[depthmap > 200] = 0.0\n\n            intrinsics = np.load(osp.join(self.ROOT, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"synscapes\",\n                    label=img_name,\n                    instance=f\"{str(idx)}_{img_name}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"2c2e047055922150a529b6b5b8b41ad47bd66810fbd6d7806051aa3e05b0f03b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.synscapes.__init__","uri":"program://Human3R/function/src.dust3r.datasets.synscapes.__init__#L15-L20","kind":"function","name":"__init__","path":"src/dust3r/datasets/synscapes.py","language":"python","start_line":15,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SynScapes(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted(\n            [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n            key=lambda x: int(x),\n        )\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n","source_hash":"2c2e047055922150a529b6b5b8b41ad47bd66810fbd6d7806051aa3e05b0f03b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.synscapes._load_data","uri":"program://Human3R/function/src.dust3r.datasets.synscapes._load_data#L22-L28","kind":"function","name":"_load_data","path":"src/dust3r/datasets/synscapes.py","language":"python","start_line":22,"end_line":28,"context_start_line":2,"context_end_line":48,"code":"import cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SynScapes(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted(\n            [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n            key=lambda x: int(x),\n        )\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127","source_hash":"2c2e047055922150a529b6b5b8b41ad47bd66810fbd6d7806051aa3e05b0f03b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.synscapes.__len__","uri":"program://Human3R/function/src.dust3r.datasets.synscapes.__len__#L30-L31","kind":"function","name":"__len__","path":"src/dust3r/datasets/synscapes.py","language":"python","start_line":30,"end_line":31,"context_start_line":10,"context_end_line":51,"code":"from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SynScapes(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted(\n            [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n            key=lambda x: int(x),\n        )\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)","source_hash":"2c2e047055922150a529b6b5b8b41ad47bd66810fbd6d7806051aa3e05b0f03b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.synscapes.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.synscapes.get_image_num#L33-L34","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/synscapes.py","language":"python","start_line":33,"end_line":34,"context_start_line":13,"context_end_line":54,"code":"\nclass SynScapes(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted(\n            [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n            key=lambda x: int(x),\n        )\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            depthmap[depthmap > 200] = 0.0\n\n            intrinsics = np.load(osp.join(self.ROOT, \"cam\", f\"{img_name}.npz\"))[","source_hash":"2c2e047055922150a529b6b5b8b41ad47bd66810fbd6d7806051aa3e05b0f03b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.synscapes._get_views","uri":"program://Human3R/function/src.dust3r.datasets.synscapes._get_views#L36-L85","kind":"function","name":"_get_views","path":"src/dust3r/datasets/synscapes.py","language":"python","start_line":36,"end_line":85,"context_start_line":16,"context_end_line":85,"code":"        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        rgb_dir = osp.join(self.ROOT, \"rgb\")\n        basenames = sorted(\n            [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n            key=lambda x: int(x),\n        )\n        self.img_names = basenames\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(self.ROOT, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, \"depth\", f\"{img_name}.npy\"))\n            sky_mask = (\n                imread_cv2(osp.join(self.ROOT, \"sky_mask\", f\"{img_name}.png\"))[..., 0]\n                >= 127\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            depthmap[depthmap > 200] = 0.0\n\n            intrinsics = np.load(osp.join(self.ROOT, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"synscapes\",\n                    label=img_name,\n                    instance=f\"{str(idx)}_{img_name}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"2c2e047055922150a529b6b5b8b41ad47bd66810fbd6d7806051aa3e05b0f03b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.pointodyssey","uri":"program://Human3R/module/src.dust3r.datasets.pointodyssey#L1-L178","kind":"module","name":"src.dust3r.datasets.pointodyssey","path":"src/dust3r/datasets/pointodyssey.py","language":"python","start_line":1,"end_line":178,"context_start_line":1,"context_end_line":178,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass PointOdyssey_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n        assert self.split in [\"train\", \"test\", \"val\"]\n        self.scenes_to_use = [\n            # 'cab_h_bench_3rd', 'cab_h_bench_ego1', 'cab_h_bench_ego2',\n            \"cnb_dlab_0215_3rd\",\n            \"cnb_dlab_0215_ego1\",\n            \"cnb_dlab_0225_3rd\",\n            \"cnb_dlab_0225_ego1\",\n            \"dancing\",\n            \"dancingroom0_3rd\",\n            \"footlab_3rd\",\n            \"footlab_ego1\",\n            \"footlab_ego2\",\n            \"girl\",\n            \"girl_egocentric\",\n            \"human_egocentric\",\n            \"human_in_scene\",\n            \"human_in_scene1\",\n            \"kg\",\n            \"kg_ego1\",\n            \"kg_ego2\",\n            \"kitchen_gfloor\",\n            \"kitchen_gfloor_ego1\",\n            \"kitchen_gfloor_ego2\",\n            \"scene_carb_h_tables\",\n            \"scene_carb_h_tables_ego1\",\n            \"scene_carb_h_tables_ego2\",\n            \"scene_j716_3rd\",\n            \"scene_j716_ego1\",\n            \"scene_j716_ego2\",\n            \"scene_recording_20210910_S05_S06_0_3rd\",\n            \"scene_recording_20210910_S05_S06_0_ego2\",\n            \"scene1_0129\",\n            \"scene1_0129_ego\",\n            \"seminar_h52_3rd\",\n            \"seminar_h52_ego1\",\n            \"seminar_h52_ego2\",\n        ]\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        root = os.path.join(self.ROOT, split)\n        self.scenes = []\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(os.listdir(root)):\n            if scene not in self.scenes_to_use:\n                continue\n            scene_dir = osp.join(root, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            depthmap[depthmap > 1000] = 0.0\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.9, 0.05, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"PointOdyssey\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".jpg\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"5ead728f8a1a89568d38f8a7ca8c1423c9cb41194ebbf01a5271ddfa9854dd9d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.pointodyssey.PointOdyssey_Multi","uri":"program://Human3R/class/src.dust3r.datasets.pointodyssey.PointOdyssey_Multi#L14-L178","kind":"class","name":"PointOdyssey_Multi","path":"src/dust3r/datasets/pointodyssey.py","language":"python","start_line":14,"end_line":178,"context_start_line":1,"context_end_line":178,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass PointOdyssey_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n        assert self.split in [\"train\", \"test\", \"val\"]\n        self.scenes_to_use = [\n            # 'cab_h_bench_3rd', 'cab_h_bench_ego1', 'cab_h_bench_ego2',\n            \"cnb_dlab_0215_3rd\",\n            \"cnb_dlab_0215_ego1\",\n            \"cnb_dlab_0225_3rd\",\n            \"cnb_dlab_0225_ego1\",\n            \"dancing\",\n            \"dancingroom0_3rd\",\n            \"footlab_3rd\",\n            \"footlab_ego1\",\n            \"footlab_ego2\",\n            \"girl\",\n            \"girl_egocentric\",\n            \"human_egocentric\",\n            \"human_in_scene\",\n            \"human_in_scene1\",\n            \"kg\",\n            \"kg_ego1\",\n            \"kg_ego2\",\n            \"kitchen_gfloor\",\n            \"kitchen_gfloor_ego1\",\n            \"kitchen_gfloor_ego2\",\n            \"scene_carb_h_tables\",\n            \"scene_carb_h_tables_ego1\",\n            \"scene_carb_h_tables_ego2\",\n            \"scene_j716_3rd\",\n            \"scene_j716_ego1\",\n            \"scene_j716_ego2\",\n            \"scene_recording_20210910_S05_S06_0_3rd\",\n            \"scene_recording_20210910_S05_S06_0_ego2\",\n            \"scene1_0129\",\n            \"scene1_0129_ego\",\n            \"seminar_h52_3rd\",\n            \"seminar_h52_ego1\",\n            \"seminar_h52_ego2\",\n        ]\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        root = os.path.join(self.ROOT, split)\n        self.scenes = []\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(os.listdir(root)):\n            if scene not in self.scenes_to_use:\n                continue\n            scene_dir = osp.join(root, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            depthmap[depthmap > 1000] = 0.0\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.9, 0.05, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"PointOdyssey\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".jpg\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"5ead728f8a1a89568d38f8a7ca8c1423c9cb41194ebbf01a5271ddfa9854dd9d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.pointodyssey.__init__","uri":"program://Human3R/function/src.dust3r.datasets.pointodyssey.__init__#L15-L58","kind":"function","name":"__init__","path":"src/dust3r/datasets/pointodyssey.py","language":"python","start_line":15,"end_line":58,"context_start_line":1,"context_end_line":78,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass PointOdyssey_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n        assert self.split in [\"train\", \"test\", \"val\"]\n        self.scenes_to_use = [\n            # 'cab_h_bench_3rd', 'cab_h_bench_ego1', 'cab_h_bench_ego2',\n            \"cnb_dlab_0215_3rd\",\n            \"cnb_dlab_0215_ego1\",\n            \"cnb_dlab_0225_3rd\",\n            \"cnb_dlab_0225_ego1\",\n            \"dancing\",\n            \"dancingroom0_3rd\",\n            \"footlab_3rd\",\n            \"footlab_ego1\",\n            \"footlab_ego2\",\n            \"girl\",\n            \"girl_egocentric\",\n            \"human_egocentric\",\n            \"human_in_scene\",\n            \"human_in_scene1\",\n            \"kg\",\n            \"kg_ego1\",\n            \"kg_ego2\",\n            \"kitchen_gfloor\",\n            \"kitchen_gfloor_ego1\",\n            \"kitchen_gfloor_ego2\",\n            \"scene_carb_h_tables\",\n            \"scene_carb_h_tables_ego1\",\n            \"scene_carb_h_tables_ego2\",\n            \"scene_j716_3rd\",\n            \"scene_j716_ego1\",\n            \"scene_j716_ego2\",\n            \"scene_recording_20210910_S05_S06_0_3rd\",\n            \"scene_recording_20210910_S05_S06_0_ego2\",\n            \"scene1_0129\",\n            \"scene1_0129_ego\",\n            \"seminar_h52_3rd\",\n            \"seminar_h52_ego1\",\n            \"seminar_h52_ego2\",\n        ]\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        root = os.path.join(self.ROOT, split)\n        self.scenes = []\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(os.listdir(root)):\n            if scene not in self.scenes_to_use:\n                continue\n            scene_dir = osp.join(root, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]","source_hash":"5ead728f8a1a89568d38f8a7ca8c1423c9cb41194ebbf01a5271ddfa9854dd9d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.pointodyssey._load_data","uri":"program://Human3R/function/src.dust3r.datasets.pointodyssey._load_data#L60-L106","kind":"function","name":"_load_data","path":"src/dust3r/datasets/pointodyssey.py","language":"python","start_line":60,"end_line":106,"context_start_line":40,"context_end_line":126,"code":"            \"kg_ego2\",\n            \"kitchen_gfloor\",\n            \"kitchen_gfloor_ego1\",\n            \"kitchen_gfloor_ego2\",\n            \"scene_carb_h_tables\",\n            \"scene_carb_h_tables_ego1\",\n            \"scene_carb_h_tables_ego2\",\n            \"scene_j716_3rd\",\n            \"scene_j716_ego1\",\n            \"scene_j716_ego2\",\n            \"scene_recording_20210910_S05_S06_0_3rd\",\n            \"scene_recording_20210910_S05_S06_0_ego2\",\n            \"scene1_0129\",\n            \"scene1_0129_ego\",\n            \"seminar_h52_3rd\",\n            \"seminar_h52_ego1\",\n            \"seminar_h52_ego2\",\n        ]\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        root = os.path.join(self.ROOT, split)\n        self.scenes = []\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(os.listdir(root)):\n            if scene not in self.scenes_to_use:\n                continue\n            scene_dir = osp.join(root, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]","source_hash":"5ead728f8a1a89568d38f8a7ca8c1423c9cb41194ebbf01a5271ddfa9854dd9d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.pointodyssey.__len__","uri":"program://Human3R/function/src.dust3r.datasets.pointodyssey.__len__#L108-L109","kind":"function","name":"__len__","path":"src/dust3r/datasets/pointodyssey.py","language":"python","start_line":108,"end_line":109,"context_start_line":88,"context_end_line":129,"code":"            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):","source_hash":"5ead728f8a1a89568d38f8a7ca8c1423c9cb41194ebbf01a5271ddfa9854dd9d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.pointodyssey.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.pointodyssey.get_image_num#L111-L112","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/pointodyssey.py","language":"python","start_line":111,"end_line":112,"context_start_line":91,"context_end_line":132,"code":"\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")","source_hash":"5ead728f8a1a89568d38f8a7ca8c1423c9cb41194ebbf01a5271ddfa9854dd9d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.pointodyssey._get_views","uri":"program://Human3R/function/src.dust3r.datasets.pointodyssey._get_views#L114-L178","kind":"function","name":"_get_views","path":"src/dust3r/datasets/pointodyssey.py","language":"python","start_line":114,"end_line":178,"context_start_line":94,"context_end_line":178,"code":"            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            depthmap[depthmap > 1000] = 0.0\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.9, 0.05, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"PointOdyssey\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".jpg\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"5ead728f8a1a89568d38f8a7ca8c1423c9cb41194ebbf01a5271ddfa9854dd9d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.tartanair","uri":"program://Human3R/module/src.dust3r.datasets.tartanair#L1-L164","kind":"module","name":"src.dust3r.datasets.tartanair","path":"src/dust3r/datasets/tartanair.py","language":"python","start_line":1,"end_line":164,"context_start_line":1,"context_end_line":164,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass TartanAir_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 20\n        super().__init__(*args, **kwargs)\n        # loading all\n        assert self.split is None\n        self._load_data()\n\n    def _load_data(self):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for scene in scene_dirs:\n            for mode in [\"Easy\", \"Hard\"]:\n                seq_dirs = sorted(\n                    [\n                        os.path.join(self.ROOT, scene, mode, d)\n                        for d in os.listdir(os.path.join(self.ROOT, scene, mode))\n                        if os.path.isdir(os.path.join(self.ROOT, scene, mode, d))\n                    ]\n                )\n                for seq_dir in seq_dirs:\n                    basenames = sorted(\n                        [f[:-8] for f in os.listdir(seq_dir) if f.endswith(\".png\")]\n                    )\n                    num_imgs = len(basenames)\n                    cut_off = (\n                        self.num_views\n                        if not self.allow_repeat\n                        else max(self.num_views // 3, 3)\n                    )\n\n                    if num_imgs < cut_off:\n                        print(f\"Skipping {scene}\")\n                        continue\n                    img_ids = list(np.arange(num_imgs) + offset)\n                    start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                    scenes.append(seq_dir)\n                    scene_img_list.append(img_ids)\n                    sceneids.extend([j] * num_imgs)\n                    images.extend(basenames)\n                    start_img_ids.extend(start_img_ids_)\n                    offset += num_imgs\n                    j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.8,\n            fix_interval_prob=0.8,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = self.scenes[scene_id]\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.png\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = np.load(osp.join(scene_dir, basename + \"_depth.npy\"))\n            camera_params = np.load(osp.join(scene_dir, basename + \"_cam.npz\"))\n\n            intrinsics = camera_params[\"camera_intrinsics\"]\n            camera_pose = camera_params[\"camera_pose\"]\n\n            sky_mask = depthmap >= 1000\n            depthmap[sky_mask] = -1.0  # sky\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"TartanAir\",\n                    label=scene_dir,\n                    is_metric=self.is_metric,\n                    instance=scene_dir + \"_\" + img,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"351cf02bfb408b6fe9c9aa1481dcdfb0f31c260bfe11dd796c8d028cb42ad56e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.tartanair.TartanAir_Multi","uri":"program://Human3R/class/src.dust3r.datasets.tartanair.TartanAir_Multi#L15-L164","kind":"class","name":"TartanAir_Multi","path":"src/dust3r/datasets/tartanair.py","language":"python","start_line":15,"end_line":164,"context_start_line":1,"context_end_line":164,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass TartanAir_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 20\n        super().__init__(*args, **kwargs)\n        # loading all\n        assert self.split is None\n        self._load_data()\n\n    def _load_data(self):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for scene in scene_dirs:\n            for mode in [\"Easy\", \"Hard\"]:\n                seq_dirs = sorted(\n                    [\n                        os.path.join(self.ROOT, scene, mode, d)\n                        for d in os.listdir(os.path.join(self.ROOT, scene, mode))\n                        if os.path.isdir(os.path.join(self.ROOT, scene, mode, d))\n                    ]\n                )\n                for seq_dir in seq_dirs:\n                    basenames = sorted(\n                        [f[:-8] for f in os.listdir(seq_dir) if f.endswith(\".png\")]\n                    )\n                    num_imgs = len(basenames)\n                    cut_off = (\n                        self.num_views\n                        if not self.allow_repeat\n                        else max(self.num_views // 3, 3)\n                    )\n\n                    if num_imgs < cut_off:\n                        print(f\"Skipping {scene}\")\n                        continue\n                    img_ids = list(np.arange(num_imgs) + offset)\n                    start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                    scenes.append(seq_dir)\n                    scene_img_list.append(img_ids)\n                    sceneids.extend([j] * num_imgs)\n                    images.extend(basenames)\n                    start_img_ids.extend(start_img_ids_)\n                    offset += num_imgs\n                    j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.8,\n            fix_interval_prob=0.8,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = self.scenes[scene_id]\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.png\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = np.load(osp.join(scene_dir, basename + \"_depth.npy\"))\n            camera_params = np.load(osp.join(scene_dir, basename + \"_cam.npz\"))\n\n            intrinsics = camera_params[\"camera_intrinsics\"]\n            camera_pose = camera_params[\"camera_pose\"]\n\n            sky_mask = depthmap >= 1000\n            depthmap[sky_mask] = -1.0  # sky\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"TartanAir\",\n                    label=scene_dir,\n                    is_metric=self.is_metric,\n                    instance=scene_dir + \"_\" + img,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"351cf02bfb408b6fe9c9aa1481dcdfb0f31c260bfe11dd796c8d028cb42ad56e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.tartanair.__init__","uri":"program://Human3R/function/src.dust3r.datasets.tartanair.__init__#L17-L25","kind":"function","name":"__init__","path":"src/dust3r/datasets/tartanair.py","language":"python","start_line":17,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass TartanAir_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 20\n        super().__init__(*args, **kwargs)\n        # loading all\n        assert self.split is None\n        self._load_data()\n\n    def _load_data(self):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for scene in scene_dirs:\n            for mode in [\"Easy\", \"Hard\"]:","source_hash":"351cf02bfb408b6fe9c9aa1481dcdfb0f31c260bfe11dd796c8d028cb42ad56e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.tartanair._load_data","uri":"program://Human3R/function/src.dust3r.datasets.tartanair._load_data#L27-L82","kind":"function","name":"_load_data","path":"src/dust3r/datasets/tartanair.py","language":"python","start_line":27,"end_line":82,"context_start_line":7,"context_end_line":102,"code":"import sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass TartanAir_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 20\n        super().__init__(*args, **kwargs)\n        # loading all\n        assert self.split is None\n        self._load_data()\n\n    def _load_data(self):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for scene in scene_dirs:\n            for mode in [\"Easy\", \"Hard\"]:\n                seq_dirs = sorted(\n                    [\n                        os.path.join(self.ROOT, scene, mode, d)\n                        for d in os.listdir(os.path.join(self.ROOT, scene, mode))\n                        if os.path.isdir(os.path.join(self.ROOT, scene, mode, d))\n                    ]\n                )\n                for seq_dir in seq_dirs:\n                    basenames = sorted(\n                        [f[:-8] for f in os.listdir(seq_dir) if f.endswith(\".png\")]\n                    )\n                    num_imgs = len(basenames)\n                    cut_off = (\n                        self.num_views\n                        if not self.allow_repeat\n                        else max(self.num_views // 3, 3)\n                    )\n\n                    if num_imgs < cut_off:\n                        print(f\"Skipping {scene}\")\n                        continue\n                    img_ids = list(np.arange(num_imgs) + offset)\n                    start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                    scenes.append(seq_dir)\n                    scene_img_list.append(img_ids)\n                    sceneids.extend([j] * num_imgs)\n                    images.extend(basenames)\n                    start_img_ids.extend(start_img_ids_)\n                    offset += num_imgs\n                    j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,","source_hash":"351cf02bfb408b6fe9c9aa1481dcdfb0f31c260bfe11dd796c8d028cb42ad56e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.tartanair.__len__","uri":"program://Human3R/function/src.dust3r.datasets.tartanair.__len__#L84-L85","kind":"function","name":"__len__","path":"src/dust3r/datasets/tartanair.py","language":"python","start_line":84,"end_line":85,"context_start_line":64,"context_end_line":105,"code":"                    if num_imgs < cut_off:\n                        print(f\"Skipping {scene}\")\n                        continue\n                    img_ids = list(np.arange(num_imgs) + offset)\n                    start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                    scenes.append(seq_dir)\n                    scene_img_list.append(img_ids)\n                    sceneids.extend([j] * num_imgs)\n                    images.extend(basenames)\n                    start_img_ids.extend(start_img_ids_)\n                    offset += num_imgs\n                    j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.8,\n            fix_interval_prob=0.8,\n            block_shuffle=16,","source_hash":"351cf02bfb408b6fe9c9aa1481dcdfb0f31c260bfe11dd796c8d028cb42ad56e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.tartanair.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.tartanair.get_image_num#L87-L88","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/tartanair.py","language":"python","start_line":87,"end_line":88,"context_start_line":67,"context_end_line":108,"code":"                    img_ids = list(np.arange(num_imgs) + offset)\n                    start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                    scenes.append(seq_dir)\n                    scene_img_list.append(img_ids)\n                    sceneids.extend([j] * num_imgs)\n                    images.extend(basenames)\n                    start_img_ids.extend(start_img_ids_)\n                    offset += num_imgs\n                    j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.8,\n            fix_interval_prob=0.8,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n","source_hash":"351cf02bfb408b6fe9c9aa1481dcdfb0f31c260bfe11dd796c8d028cb42ad56e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.tartanair.get_stats","uri":"program://Human3R/function/src.dust3r.datasets.tartanair.get_stats#L90-L91","kind":"function","name":"get_stats","path":"src/dust3r/datasets/tartanair.py","language":"python","start_line":90,"end_line":91,"context_start_line":70,"context_end_line":111,"code":"                    scenes.append(seq_dir)\n                    scene_img_list.append(img_ids)\n                    sceneids.extend([j] * num_imgs)\n                    images.extend(basenames)\n                    start_img_ids.extend(start_img_ids_)\n                    offset += num_imgs\n                    j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.8,\n            fix_interval_prob=0.8,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):","source_hash":"351cf02bfb408b6fe9c9aa1481dcdfb0f31c260bfe11dd796c8d028cb42ad56e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.tartanair._get_views","uri":"program://Human3R/function/src.dust3r.datasets.tartanair._get_views#L93-L164","kind":"function","name":"_get_views","path":"src/dust3r/datasets/tartanair.py","language":"python","start_line":93,"end_line":164,"context_start_line":73,"context_end_line":164,"code":"                    images.extend(basenames)\n                    start_img_ids.extend(start_img_ids_)\n                    offset += num_imgs\n                    j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.8,\n            fix_interval_prob=0.8,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = self.scenes[scene_id]\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.png\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = np.load(osp.join(scene_dir, basename + \"_depth.npy\"))\n            camera_params = np.load(osp.join(scene_dir, basename + \"_cam.npz\"))\n\n            intrinsics = camera_params[\"camera_intrinsics\"]\n            camera_pose = camera_params[\"camera_pose\"]\n\n            sky_mask = depthmap >= 1000\n            depthmap[sky_mask] = -1.0  # sky\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"TartanAir\",\n                    label=scene_dir,\n                    is_metric=self.is_metric,\n                    instance=scene_dir + \"_\" + img,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"351cf02bfb408b6fe9c9aa1481dcdfb0f31c260bfe11dd796c8d028cb42ad56e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.unreal4k","uri":"program://Human3R/module/src.dust3r.datasets.unreal4k#L1-L159","kind":"module","name":"src.dust3r.datasets.unreal4k","path":"src/dust3r/datasets/unreal4k.py","language":"python","start_line":1,"end_line":159,"context_start_line":1,"context_end_line":159,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nR_conv = np.array([[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]).astype(\n    np.float32\n)\n\n\nclass UnReal4K_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 2\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        # loading all\n        assert self.split is None\n        self._load_data()\n\n    def _load_data(self):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        j = 0\n\n        seq_dirs = sorted(\n            [\n                os.path.join(self.ROOT, scene, mode)\n                for scene in scene_dirs\n                for mode in [\"0\", \"1\"]\n            ]\n        )\n        for seq_dir in seq_dirs:\n            basenames = sorted(\n                [f[:-8] for f in os.listdir(seq_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {seq_dir}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(seq_dir)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)//10} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = self.scenes[scene_id]\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.png\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = np.load(osp.join(scene_dir, basename + \"_depth.npy\"))\n            camera_params = np.load(osp.join(scene_dir, basename + \".npz\"))\n\n            intrinsics = camera_params[\"intrinsics\"].astype(np.float32)\n            camera_pose = camera_params[\"cam2world\"].astype(np.float32)\n\n            camera_pose = R_conv @ camera_pose\n\n            sky_mask = depthmap >= 1000\n            depthmap[sky_mask] = -1.0  # sky\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"UnReal4K\",\n                    label=scene_dir,\n                    is_metric=self.is_metric,\n                    instance=scene_dir + \"_\" + img,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"1b5528a7239f01051125b7378898ff95fc0b7b7813894d237c9124e38e69f0ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.unreal4k.UnReal4K_Multi","uri":"program://Human3R/class/src.dust3r.datasets.unreal4k.UnReal4K_Multi#L19-L159","kind":"class","name":"UnReal4K_Multi","path":"src/dust3r/datasets/unreal4k.py","language":"python","start_line":19,"end_line":159,"context_start_line":1,"context_end_line":159,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nR_conv = np.array([[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]).astype(\n    np.float32\n)\n\n\nclass UnReal4K_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 2\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        # loading all\n        assert self.split is None\n        self._load_data()\n\n    def _load_data(self):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        j = 0\n\n        seq_dirs = sorted(\n            [\n                os.path.join(self.ROOT, scene, mode)\n                for scene in scene_dirs\n                for mode in [\"0\", \"1\"]\n            ]\n        )\n        for seq_dir in seq_dirs:\n            basenames = sorted(\n                [f[:-8] for f in os.listdir(seq_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {seq_dir}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(seq_dir)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)//10} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = self.scenes[scene_id]\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.png\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = np.load(osp.join(scene_dir, basename + \"_depth.npy\"))\n            camera_params = np.load(osp.join(scene_dir, basename + \".npz\"))\n\n            intrinsics = camera_params[\"intrinsics\"].astype(np.float32)\n            camera_pose = camera_params[\"cam2world\"].astype(np.float32)\n\n            camera_pose = R_conv @ camera_pose\n\n            sky_mask = depthmap >= 1000\n            depthmap[sky_mask] = -1.0  # sky\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"UnReal4K\",\n                    label=scene_dir,\n                    is_metric=self.is_metric,\n                    instance=scene_dir + \"_\" + img,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"1b5528a7239f01051125b7378898ff95fc0b7b7813894d237c9124e38e69f0ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.unreal4k.__init__","uri":"program://Human3R/function/src.dust3r.datasets.unreal4k.__init__#L21-L28","kind":"function","name":"__init__","path":"src/dust3r/datasets/unreal4k.py","language":"python","start_line":21,"end_line":28,"context_start_line":1,"context_end_line":48,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nR_conv = np.array([[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]).astype(\n    np.float32\n)\n\n\nclass UnReal4K_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 2\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        # loading all\n        assert self.split is None\n        self._load_data()\n\n    def _load_data(self):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        j = 0\n\n        seq_dirs = sorted(\n            [","source_hash":"1b5528a7239f01051125b7378898ff95fc0b7b7813894d237c9124e38e69f0ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.unreal4k._load_data","uri":"program://Human3R/function/src.dust3r.datasets.unreal4k._load_data#L30-L84","kind":"function","name":"_load_data","path":"src/dust3r/datasets/unreal4k.py","language":"python","start_line":30,"end_line":84,"context_start_line":10,"context_end_line":104,"code":"\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nR_conv = np.array([[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]).astype(\n    np.float32\n)\n\n\nclass UnReal4K_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 2\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        # loading all\n        assert self.split is None\n        self._load_data()\n\n    def _load_data(self):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        j = 0\n\n        seq_dirs = sorted(\n            [\n                os.path.join(self.ROOT, scene, mode)\n                for scene in scene_dirs\n                for mode in [\"0\", \"1\"]\n            ]\n        )\n        for seq_dir in seq_dirs:\n            basenames = sorted(\n                [f[:-8] for f in os.listdir(seq_dir) if f.endswith(\".png\")]\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            # start_img_ids_ = img_ids[:-self.num_views+1]\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {seq_dir}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(seq_dir)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)//10} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []","source_hash":"1b5528a7239f01051125b7378898ff95fc0b7b7813894d237c9124e38e69f0ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.unreal4k.__len__","uri":"program://Human3R/function/src.dust3r.datasets.unreal4k.__len__#L86-L87","kind":"function","name":"__len__","path":"src/dust3r/datasets/unreal4k.py","language":"python","start_line":86,"end_line":87,"context_start_line":66,"context_end_line":107,"code":"            if num_imgs < cut_off:\n                print(f\"Skipping {seq_dir}\")\n                continue\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(seq_dir)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)//10} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]","source_hash":"1b5528a7239f01051125b7378898ff95fc0b7b7813894d237c9124e38e69f0ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.unreal4k.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.unreal4k.get_image_num#L89-L90","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/unreal4k.py","language":"python","start_line":89,"end_line":90,"context_start_line":69,"context_end_line":110,"code":"\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(seq_dir)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)//10} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = self.scenes[scene_id]\n            basename = self.images[view_idx]\n","source_hash":"1b5528a7239f01051125b7378898ff95fc0b7b7813894d237c9124e38e69f0ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.unreal4k.get_stats","uri":"program://Human3R/function/src.dust3r.datasets.unreal4k.get_stats#L92-L93","kind":"function","name":"get_stats","path":"src/dust3r/datasets/unreal4k.py","language":"python","start_line":92,"end_line":93,"context_start_line":72,"context_end_line":113,"code":"            images.extend(basenames)\n            scenes.append(seq_dir)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)//10} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = self.scenes[scene_id]\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.png\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = np.load(osp.join(scene_dir, basename + \"_depth.npy\"))","source_hash":"1b5528a7239f01051125b7378898ff95fc0b7b7813894d237c9124e38e69f0ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.unreal4k._get_views","uri":"program://Human3R/function/src.dust3r.datasets.unreal4k._get_views#L95-L159","kind":"function","name":"_get_views","path":"src/dust3r/datasets/unreal4k.py","language":"python","start_line":95,"end_line":159,"context_start_line":75,"context_end_line":159,"code":"\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids) * 10\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)//10} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=self.max_interval\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = self.scenes[scene_id]\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.png\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = np.load(osp.join(scene_dir, basename + \"_depth.npy\"))\n            camera_params = np.load(osp.join(scene_dir, basename + \".npz\"))\n\n            intrinsics = camera_params[\"intrinsics\"].astype(np.float32)\n            camera_pose = camera_params[\"cam2world\"].astype(np.float32)\n\n            camera_pose = R_conv @ camera_pose\n\n            sky_mask = depthmap >= 1000\n            depthmap[sky_mask] = -1.0  # sky\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"UnReal4K\",\n                    label=scene_dir,\n                    is_metric=self.is_metric,\n                    instance=scene_dir + \"_\" + img,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"1b5528a7239f01051125b7378898ff95fc0b7b7813894d237c9124e38e69f0ce","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.smartportraits","uri":"program://Human3R/module/src.dust3r.datasets.smartportraits#L1-L85","kind":"module","name":"src.dust3r.datasets.smartportraits","path":"src/dust3r/datasets/smartportraits.py","language":"python","start_line":1,"end_line":85,"context_start_line":1,"context_end_line":85,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SmartPortraits_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"SmartPortraits\",\n                    label=img_name,\n                    instance=osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"),\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"6acc2076c1e4b6866dfada9c3fb9848333e2f83012c98fd955253d8dac19092b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.smartportraits.SmartPortraits_Multi","uri":"program://Human3R/class/src.dust3r.datasets.smartportraits.SmartPortraits_Multi#L14-L85","kind":"class","name":"SmartPortraits_Multi","path":"src/dust3r/datasets/smartportraits.py","language":"python","start_line":14,"end_line":85,"context_start_line":1,"context_end_line":85,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SmartPortraits_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"SmartPortraits\",\n                    label=img_name,\n                    instance=osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"),\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"6acc2076c1e4b6866dfada9c3fb9848333e2f83012c98fd955253d8dac19092b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.smartportraits.__init__","uri":"program://Human3R/function/src.dust3r.datasets.smartportraits.__init__#L15-L20","kind":"function","name":"__init__","path":"src/dust3r/datasets/smartportraits.py","language":"python","start_line":15,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SmartPortraits_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n","source_hash":"6acc2076c1e4b6866dfada9c3fb9848333e2f83012c98fd955253d8dac19092b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.smartportraits._load_data","uri":"program://Human3R/function/src.dust3r.datasets.smartportraits._load_data#L22-L33","kind":"function","name":"_load_data","path":"src/dust3r/datasets/smartportraits.py","language":"python","start_line":22,"end_line":33,"context_start_line":2,"context_end_line":53,"code":"import cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SmartPortraits_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n","source_hash":"6acc2076c1e4b6866dfada9c3fb9848333e2f83012c98fd955253d8dac19092b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.smartportraits.__len__","uri":"program://Human3R/function/src.dust3r.datasets.smartportraits.__len__#L35-L36","kind":"function","name":"__len__","path":"src/dust3r/datasets/smartportraits.py","language":"python","start_line":35,"end_line":36,"context_start_line":15,"context_end_line":56,"code":"    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]","source_hash":"6acc2076c1e4b6866dfada9c3fb9848333e2f83012c98fd955253d8dac19092b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.smartportraits.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.smartportraits.get_image_num#L38-L39","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/smartportraits.py","language":"python","start_line":38,"end_line":39,"context_start_line":18,"context_end_line":59,"code":"        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n","source_hash":"6acc2076c1e4b6866dfada9c3fb9848333e2f83012c98fd955253d8dac19092b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.smartportraits._get_views","uri":"program://Human3R/function/src.dust3r.datasets.smartportraits._get_views#L41-L85","kind":"function","name":"_get_views","path":"src/dust3r/datasets/smartportraits.py","language":"python","start_line":41,"end_line":85,"context_start_line":21,"context_end_line":85,"code":"\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"SmartPortraits\",\n                    label=img_name,\n                    instance=osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"),\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"6acc2076c1e4b6866dfada9c3fb9848333e2f83012c98fd955253d8dac19092b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hoi4d","uri":"program://Human3R/module/src.dust3r.datasets.hoi4d#L1-L84","kind":"module","name":"src.dust3r.datasets.hoi4d","path":"src/dust3r/datasets/hoi4d.py","language":"python","start_line":1,"end_line":84,"context_start_line":1,"context_end_line":84,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nsys.path.append(osp.join(osp.dirname(__file__), '..','..'))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HOI4D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, 'rgb')\n            basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith('.png')])\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n         \n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        invalid_seq = True\n        while invalid_seq:\n            img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n            views = []\n            for v, img_name in enumerate(img_names):\n                # Load RGB image\n                scene, img_name = img_name\n                try:\n                    rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n                    depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n                    depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n                    intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\"intrinsics\"]\n                except:\n                    print(f\"Error loading {scene} {img_name}, skipping\")\n                    break\n                # camera pose is not provided, placeholder\n                camera_pose = np.eye(4) \n\n                rgb_image, depthmap, intrinsics= self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name)\n\n                views.append(dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset='HOI4D',\n                    label=img_name,\n                    instance=osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"),\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                ))\n            if len(views) == num_views:\n                invalid_seq = False\n        return views","source_hash":"471296779fbd64093c0aee3d7f279d1e02e1366abb07afef990621489b72e3f5","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hoi4d.HOI4D_Multi","uri":"program://Human3R/class/src.dust3r.datasets.hoi4d.HOI4D_Multi#L13-L84","kind":"class","name":"HOI4D_Multi","path":"src/dust3r/datasets/hoi4d.py","language":"python","start_line":13,"end_line":84,"context_start_line":1,"context_end_line":84,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nsys.path.append(osp.join(osp.dirname(__file__), '..','..'))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HOI4D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, 'rgb')\n            basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith('.png')])\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n         \n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        invalid_seq = True\n        while invalid_seq:\n            img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n            views = []\n            for v, img_name in enumerate(img_names):\n                # Load RGB image\n                scene, img_name = img_name\n                try:\n                    rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n                    depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n                    depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n                    intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\"intrinsics\"]\n                except:\n                    print(f\"Error loading {scene} {img_name}, skipping\")\n                    break\n                # camera pose is not provided, placeholder\n                camera_pose = np.eye(4) \n\n                rgb_image, depthmap, intrinsics= self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name)\n\n                views.append(dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset='HOI4D',\n                    label=img_name,\n                    instance=osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"),\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                ))\n            if len(views) == num_views:\n                invalid_seq = False\n        return views","source_hash":"471296779fbd64093c0aee3d7f279d1e02e1366abb07afef990621489b72e3f5","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hoi4d.__init__","uri":"program://Human3R/function/src.dust3r.datasets.hoi4d.__init__#L14-L19","kind":"function","name":"__init__","path":"src/dust3r/datasets/hoi4d.py","language":"python","start_line":14,"end_line":19,"context_start_line":1,"context_end_line":39,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nsys.path.append(osp.join(osp.dirname(__file__), '..','..'))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HOI4D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, 'rgb')\n            basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith('.png')])\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n         \n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx","source_hash":"471296779fbd64093c0aee3d7f279d1e02e1366abb07afef990621489b72e3f5","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hoi4d._load_data","uri":"program://Human3R/function/src.dust3r.datasets.hoi4d._load_data#L21-L30","kind":"function","name":"_load_data","path":"src/dust3r/datasets/hoi4d.py","language":"python","start_line":21,"end_line":30,"context_start_line":1,"context_end_line":50,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nsys.path.append(osp.join(osp.dirname(__file__), '..','..'))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HOI4D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, 'rgb')\n            basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith('.png')])\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n         \n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        invalid_seq = True\n        while invalid_seq:\n            img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n            views = []\n            for v, img_name in enumerate(img_names):\n                # Load RGB image\n                scene, img_name = img_name\n                try:\n                    rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))","source_hash":"471296779fbd64093c0aee3d7f279d1e02e1366abb07afef990621489b72e3f5","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hoi4d.__len__","uri":"program://Human3R/function/src.dust3r.datasets.hoi4d.__len__#L32-L33","kind":"function","name":"__len__","path":"src/dust3r/datasets/hoi4d.py","language":"python","start_line":32,"end_line":33,"context_start_line":12,"context_end_line":53,"code":"\nclass HOI4D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, 'rgb')\n            basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith('.png')])\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n         \n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        invalid_seq = True\n        while invalid_seq:\n            img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n            views = []\n            for v, img_name in enumerate(img_names):\n                # Load RGB image\n                scene, img_name = img_name\n                try:\n                    rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n                    depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n                    depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n","source_hash":"471296779fbd64093c0aee3d7f279d1e02e1366abb07afef990621489b72e3f5","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hoi4d.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.hoi4d.get_image_num#L35-L36","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/hoi4d.py","language":"python","start_line":35,"end_line":36,"context_start_line":15,"context_end_line":56,"code":"        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, 'rgb')\n            basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith('.png')])\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n         \n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        invalid_seq = True\n        while invalid_seq:\n            img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n            views = []\n            for v, img_name in enumerate(img_names):\n                # Load RGB image\n                scene, img_name = img_name\n                try:\n                    rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n                    depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n                    depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n                    intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\"intrinsics\"]\n                except:\n                    print(f\"Error loading {scene} {img_name}, skipping\")","source_hash":"471296779fbd64093c0aee3d7f279d1e02e1366abb07afef990621489b72e3f5","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.hoi4d._get_views","uri":"program://Human3R/function/src.dust3r.datasets.hoi4d._get_views#L38-L84","kind":"function","name":"_get_views","path":"src/dust3r/datasets/hoi4d.py","language":"python","start_line":38,"end_line":84,"context_start_line":18,"context_end_line":84,"code":"        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, 'rgb')\n            basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith('.png')])\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n         \n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        invalid_seq = True\n        while invalid_seq:\n            img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n            views = []\n            for v, img_name in enumerate(img_names):\n                # Load RGB image\n                scene, img_name = img_name\n                try:\n                    rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n                    depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n                    depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n                    intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\"intrinsics\"]\n                except:\n                    print(f\"Error loading {scene} {img_name}, skipping\")\n                    break\n                # camera pose is not provided, placeholder\n                camera_pose = np.eye(4) \n\n                rgb_image, depthmap, intrinsics= self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name)\n\n                views.append(dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset='HOI4D',\n                    label=img_name,\n                    instance=osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"),\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                ))\n            if len(views) == num_views:\n                invalid_seq = False\n        return views","source_hash":"471296779fbd64093c0aee3d7f279d1e02e1366abb07afef990621489b72e3f5","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.wildrgbd","uri":"program://Human3R/module/src.dust3r.datasets.wildrgbd#L1-L56","kind":"module","name":"src.dust3r.datasets.wildrgbd","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":1,"end_line":56,"context_start_line":1,"context_end_line":56,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass WildRGBD_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"WildRGBD\"\n        self.is_metric = True\n        # load all scenes\n        self.scenes.pop((\"box\", \"scenes/scene_257\"), None)\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"metadata\", f\"{view_idx:0>5d}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"rgb\", f\"{view_idx:0>5d}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"depth\", f\"{view_idx:0>5d}.png\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"{view_idx:0>5d}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        # We store depths in the depth scale of 1000.\n        # That is, when we load depth image and divide by 1000, we could get depth in meters.\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = depthmap.astype(np.float32) / 1000.0\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        views = super()._get_views(idx, resolution, rng, num_views)\n        for view in views:\n            assert view[\"is_metric\"]\n            view[\"quantile\"] = np.array(0.96, dtype=np.float32)\n        return views","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.wildrgbd.WildRGBD_Multi","uri":"program://Human3R/class/src.dust3r.datasets.wildrgbd.WildRGBD_Multi#L12-L56","kind":"class","name":"WildRGBD_Multi","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":12,"end_line":56,"context_start_line":1,"context_end_line":56,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass WildRGBD_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"WildRGBD\"\n        self.is_metric = True\n        # load all scenes\n        self.scenes.pop((\"box\", \"scenes/scene_257\"), None)\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"metadata\", f\"{view_idx:0>5d}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"rgb\", f\"{view_idx:0>5d}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"depth\", f\"{view_idx:0>5d}.png\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"{view_idx:0>5d}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        # We store depths in the depth scale of 1000.\n        # That is, when we load depth image and divide by 1000, we could get depth in meters.\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = depthmap.astype(np.float32) / 1000.0\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        views = super()._get_views(idx, resolution, rng, num_views)\n        for view in views:\n            assert view[\"is_metric\"]\n            view[\"quantile\"] = np.array(0.96, dtype=np.float32)\n        return views","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.wildrgbd.__init__","uri":"program://Human3R/function/src.dust3r.datasets.wildrgbd.__init__#L13-L30","kind":"function","name":"__init__","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":13,"end_line":30,"context_start_line":1,"context_end_line":50,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass WildRGBD_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"WildRGBD\"\n        self.is_metric = True\n        # load all scenes\n        self.scenes.pop((\"box\", \"scenes/scene_257\"), None)\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"metadata\", f\"{view_idx:0>5d}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"rgb\", f\"{view_idx:0>5d}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"depth\", f\"{view_idx:0>5d}.png\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"{view_idx:0>5d}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        # We store depths in the depth scale of 1000.\n        # That is, when we load depth image and divide by 1000, we could get depth in meters.\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = depthmap.astype(np.float32) / 1000.0\n        return depthmap\n","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.wildrgbd._get_metadatapath","uri":"program://Human3R/function/src.dust3r.datasets.wildrgbd._get_metadatapath#L32-L33","kind":"function","name":"_get_metadatapath","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":32,"end_line":33,"context_start_line":12,"context_end_line":53,"code":"class WildRGBD_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"WildRGBD\"\n        self.is_metric = True\n        # load all scenes\n        self.scenes.pop((\"box\", \"scenes/scene_257\"), None)\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"metadata\", f\"{view_idx:0>5d}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"rgb\", f\"{view_idx:0>5d}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"depth\", f\"{view_idx:0>5d}.png\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"{view_idx:0>5d}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        # We store depths in the depth scale of 1000.\n        # That is, when we load depth image and divide by 1000, we could get depth in meters.\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = depthmap.astype(np.float32) / 1000.0\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        views = super()._get_views(idx, resolution, rng, num_views)\n        for view in views:","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.wildrgbd._get_impath","uri":"program://Human3R/function/src.dust3r.datasets.wildrgbd._get_impath#L35-L36","kind":"function","name":"_get_impath","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":35,"end_line":36,"context_start_line":15,"context_end_line":56,"code":"        self.dataset_label = \"WildRGBD\"\n        self.is_metric = True\n        # load all scenes\n        self.scenes.pop((\"box\", \"scenes/scene_257\"), None)\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"metadata\", f\"{view_idx:0>5d}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"rgb\", f\"{view_idx:0>5d}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"depth\", f\"{view_idx:0>5d}.png\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"{view_idx:0>5d}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        # We store depths in the depth scale of 1000.\n        # That is, when we load depth image and divide by 1000, we could get depth in meters.\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = depthmap.astype(np.float32) / 1000.0\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        views = super()._get_views(idx, resolution, rng, num_views)\n        for view in views:\n            assert view[\"is_metric\"]\n            view[\"quantile\"] = np.array(0.96, dtype=np.float32)\n        return views","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.wildrgbd._get_depthpath","uri":"program://Human3R/function/src.dust3r.datasets.wildrgbd._get_depthpath#L38-L39","kind":"function","name":"_get_depthpath","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":38,"end_line":39,"context_start_line":18,"context_end_line":56,"code":"        self.scenes.pop((\"box\", \"scenes/scene_257\"), None)\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"metadata\", f\"{view_idx:0>5d}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"rgb\", f\"{view_idx:0>5d}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"depth\", f\"{view_idx:0>5d}.png\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"{view_idx:0>5d}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        # We store depths in the depth scale of 1000.\n        # That is, when we load depth image and divide by 1000, we could get depth in meters.\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = depthmap.astype(np.float32) / 1000.0\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        views = super()._get_views(idx, resolution, rng, num_views)\n        for view in views:\n            assert view[\"is_metric\"]\n            view[\"quantile\"] = np.array(0.96, dtype=np.float32)\n        return views","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.wildrgbd._get_maskpath","uri":"program://Human3R/function/src.dust3r.datasets.wildrgbd._get_maskpath#L41-L42","kind":"function","name":"_get_maskpath","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":41,"end_line":42,"context_start_line":21,"context_end_line":56,"code":"            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n        )\n        self.cut_off = cut_off\n        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"metadata\", f\"{view_idx:0>5d}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"rgb\", f\"{view_idx:0>5d}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"depth\", f\"{view_idx:0>5d}.png\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"{view_idx:0>5d}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        # We store depths in the depth scale of 1000.\n        # That is, when we load depth image and divide by 1000, we could get depth in meters.\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = depthmap.astype(np.float32) / 1000.0\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        views = super()._get_views(idx, resolution, rng, num_views)\n        for view in views:\n            assert view[\"is_metric\"]\n            view[\"quantile\"] = np.array(0.96, dtype=np.float32)\n        return views","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.wildrgbd._read_depthmap","uri":"program://Human3R/function/src.dust3r.datasets.wildrgbd._read_depthmap#L44-L49","kind":"function","name":"_read_depthmap","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":44,"end_line":49,"context_start_line":24,"context_end_line":56,"code":"        self.all_ref_imgs = [\n            (key, value)\n            for key, values in self.scenes.items()\n            for value in values[: len(values) - cut_off + 1]\n        ]\n        self.invalidate = {scene: {} for scene in self.scene_list}\n        self.invalid_scenes = {scene: False for scene in self.scene_list}\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"metadata\", f\"{view_idx:0>5d}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"rgb\", f\"{view_idx:0>5d}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"depth\", f\"{view_idx:0>5d}.png\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"{view_idx:0>5d}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        # We store depths in the depth scale of 1000.\n        # That is, when we load depth image and divide by 1000, we could get depth in meters.\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = depthmap.astype(np.float32) / 1000.0\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        views = super()._get_views(idx, resolution, rng, num_views)\n        for view in views:\n            assert view[\"is_metric\"]\n            view[\"quantile\"] = np.array(0.96, dtype=np.float32)\n        return views","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.wildrgbd._get_views","uri":"program://Human3R/function/src.dust3r.datasets.wildrgbd._get_views#L51-L56","kind":"function","name":"_get_views","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":51,"end_line":56,"context_start_line":31,"context_end_line":56,"code":"\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"metadata\", f\"{view_idx:0>5d}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"rgb\", f\"{view_idx:0>5d}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"depth\", f\"{view_idx:0>5d}.png\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"{view_idx:0>5d}.png\")\n\n    def _read_depthmap(self, depthpath, input_metadata):\n        # We store depths in the depth scale of 1000.\n        # That is, when we load depth image and divide by 1000, we could get depth in meters.\n        depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED)\n        depthmap = depthmap.astype(np.float32) / 1000.0\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        views = super()._get_views(idx, resolution, rng, num_views)\n        for view in views:\n            assert view[\"is_metric\"]\n            view[\"quantile\"] = np.array(0.96, dtype=np.float32)\n        return views","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.vkitti2","uri":"program://Human3R/module/src.dust3r.datasets.vkitti2#L1-L169","kind":"module","name":"src.dust3r.datasets.vkitti2","path":"src/dust3r/datasets/vkitti2.py","language":"python","start_line":1,"end_line":169,"context_start_line":1,"context_end_line":169,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass VirtualKITTI2_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 5\n        super().__init__(*args, **kwargs)\n        # loading all\n        self._load_data(self.split)\n\n    def _load_data(self, split=None):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n        if split == \"train\":\n            scene_dirs = scene_dirs[:-1]\n        elif split == \"test\":\n            scene_dirs = scene_dirs[-1:]\n        seq_dirs = []\n        for scene in scene_dirs:\n            seq_dirs += sorted(\n                [\n                    os.path.join(scene, d)\n                    for d in os.listdir(os.path.join(self.ROOT, scene))\n                    if os.path.isdir(os.path.join(self.ROOT, scene, d))\n                ]\n            )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for seq_idx, seq in enumerate(seq_dirs):\n            seq_path = osp.join(self.ROOT, seq)\n            for cam in [\"Camera_0\", \"Camera_1\"]:\n                basenames = sorted(\n                    [\n                        f[:5]\n                        for f in os.listdir(seq_path + \"/\" + cam)\n                        if f.endswith(\".jpg\")\n                    ]\n                )\n                num_imgs = len(basenames)\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n                img_ids = list(np.arange(num_imgs) + offset)\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                scenes.append(seq + \"/\" + cam)\n                scene_img_list.append(img_ids)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                start_img_ids.extend(start_img_ids_)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=0.9,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.jpg\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = (\n                cv2.imread(\n                    osp.join(scene_dir, basename + \"_depth.png\"),\n                    cv2.IMREAD_ANYCOLOR | cv2.IMREAD_ANYDEPTH,\n                ).astype(np.float32)\n                / 100.0\n            )\n            camera_params = np.load(osp.join(scene_dir, basename + \"_cam.npz\"))\n\n            intrinsics = camera_params[\"camera_intrinsics\"]\n            camera_pose = camera_params[\"camera_pose\"]\n\n            sky_mask = depthmap >= 655\n            depthmap[sky_mask] = -1.0  # sky\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.1, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"VirtualKITTI2\",\n                    label=scene_dir,\n                    is_metric=self.is_metric,\n                    instance=scene_dir + \"_\" + img,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"1a1d9ca9ee4628332411dadbb333eb4b6f131d681ade6f4aa15100a8e2b58706","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.vkitti2.VirtualKITTI2_Multi","uri":"program://Human3R/class/src.dust3r.datasets.vkitti2.VirtualKITTI2_Multi#L15-L169","kind":"class","name":"VirtualKITTI2_Multi","path":"src/dust3r/datasets/vkitti2.py","language":"python","start_line":15,"end_line":169,"context_start_line":1,"context_end_line":169,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass VirtualKITTI2_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 5\n        super().__init__(*args, **kwargs)\n        # loading all\n        self._load_data(self.split)\n\n    def _load_data(self, split=None):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n        if split == \"train\":\n            scene_dirs = scene_dirs[:-1]\n        elif split == \"test\":\n            scene_dirs = scene_dirs[-1:]\n        seq_dirs = []\n        for scene in scene_dirs:\n            seq_dirs += sorted(\n                [\n                    os.path.join(scene, d)\n                    for d in os.listdir(os.path.join(self.ROOT, scene))\n                    if os.path.isdir(os.path.join(self.ROOT, scene, d))\n                ]\n            )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for seq_idx, seq in enumerate(seq_dirs):\n            seq_path = osp.join(self.ROOT, seq)\n            for cam in [\"Camera_0\", \"Camera_1\"]:\n                basenames = sorted(\n                    [\n                        f[:5]\n                        for f in os.listdir(seq_path + \"/\" + cam)\n                        if f.endswith(\".jpg\")\n                    ]\n                )\n                num_imgs = len(basenames)\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n                img_ids = list(np.arange(num_imgs) + offset)\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                scenes.append(seq + \"/\" + cam)\n                scene_img_list.append(img_ids)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                start_img_ids.extend(start_img_ids_)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=0.9,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.jpg\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = (\n                cv2.imread(\n                    osp.join(scene_dir, basename + \"_depth.png\"),\n                    cv2.IMREAD_ANYCOLOR | cv2.IMREAD_ANYDEPTH,\n                ).astype(np.float32)\n                / 100.0\n            )\n            camera_params = np.load(osp.join(scene_dir, basename + \"_cam.npz\"))\n\n            intrinsics = camera_params[\"camera_intrinsics\"]\n            camera_pose = camera_params[\"camera_pose\"]\n\n            sky_mask = depthmap >= 655\n            depthmap[sky_mask] = -1.0  # sky\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.1, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"VirtualKITTI2\",\n                    label=scene_dir,\n                    is_metric=self.is_metric,\n                    instance=scene_dir + \"_\" + img,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"1a1d9ca9ee4628332411dadbb333eb4b6f131d681ade6f4aa15100a8e2b58706","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.vkitti2.__init__","uri":"program://Human3R/function/src.dust3r.datasets.vkitti2.__init__#L17-L24","kind":"function","name":"__init__","path":"src/dust3r/datasets/vkitti2.py","language":"python","start_line":17,"end_line":24,"context_start_line":1,"context_end_line":44,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass VirtualKITTI2_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 5\n        super().__init__(*args, **kwargs)\n        # loading all\n        self._load_data(self.split)\n\n    def _load_data(self, split=None):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n        if split == \"train\":\n            scene_dirs = scene_dirs[:-1]\n        elif split == \"test\":\n            scene_dirs = scene_dirs[-1:]\n        seq_dirs = []\n        for scene in scene_dirs:\n            seq_dirs += sorted(\n                [\n                    os.path.join(scene, d)\n                    for d in os.listdir(os.path.join(self.ROOT, scene))\n                    if os.path.isdir(os.path.join(self.ROOT, scene, d))","source_hash":"1a1d9ca9ee4628332411dadbb333eb4b6f131d681ade6f4aa15100a8e2b58706","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.vkitti2._load_data","uri":"program://Human3R/function/src.dust3r.datasets.vkitti2._load_data#L26-L89","kind":"function","name":"_load_data","path":"src/dust3r/datasets/vkitti2.py","language":"python","start_line":26,"end_line":89,"context_start_line":6,"context_end_line":109,"code":"import os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass VirtualKITTI2_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 5\n        super().__init__(*args, **kwargs)\n        # loading all\n        self._load_data(self.split)\n\n    def _load_data(self, split=None):\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n        if split == \"train\":\n            scene_dirs = scene_dirs[:-1]\n        elif split == \"test\":\n            scene_dirs = scene_dirs[-1:]\n        seq_dirs = []\n        for scene in scene_dirs:\n            seq_dirs += sorted(\n                [\n                    os.path.join(scene, d)\n                    for d in os.listdir(os.path.join(self.ROOT, scene))\n                    if os.path.isdir(os.path.join(self.ROOT, scene, d))\n                ]\n            )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for seq_idx, seq in enumerate(seq_dirs):\n            seq_path = osp.join(self.ROOT, seq)\n            for cam in [\"Camera_0\", \"Camera_1\"]:\n                basenames = sorted(\n                    [\n                        f[:5]\n                        for f in os.listdir(seq_path + \"/\" + cam)\n                        if f.endswith(\".jpg\")\n                    ]\n                )\n                num_imgs = len(basenames)\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n                img_ids = list(np.arange(num_imgs) + offset)\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                scenes.append(seq + \"/\" + cam)\n                scene_img_list.append(img_ids)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                start_img_ids.extend(start_img_ids_)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,","source_hash":"1a1d9ca9ee4628332411dadbb333eb4b6f131d681ade6f4aa15100a8e2b58706","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.vkitti2.__len__","uri":"program://Human3R/function/src.dust3r.datasets.vkitti2.__len__#L91-L92","kind":"function","name":"__len__","path":"src/dust3r/datasets/vkitti2.py","language":"python","start_line":91,"end_line":92,"context_start_line":71,"context_end_line":112,"code":"                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n                img_ids = list(np.arange(num_imgs) + offset)\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                scenes.append(seq + \"/\" + cam)\n                scene_img_list.append(img_ids)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                start_img_ids.extend(start_img_ids_)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=0.9,\n        )","source_hash":"1a1d9ca9ee4628332411dadbb333eb4b6f131d681ade6f4aa15100a8e2b58706","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.vkitti2.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.vkitti2.get_image_num#L94-L95","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/vkitti2.py","language":"python","start_line":94,"end_line":95,"context_start_line":74,"context_end_line":115,"code":"                img_ids = list(np.arange(num_imgs) + offset)\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                scenes.append(seq + \"/\" + cam)\n                scene_img_list.append(img_ids)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                start_img_ids.extend(start_img_ids_)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=0.9,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []","source_hash":"1a1d9ca9ee4628332411dadbb333eb4b6f131d681ade6f4aa15100a8e2b58706","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.vkitti2.get_stats","uri":"program://Human3R/function/src.dust3r.datasets.vkitti2.get_stats#L97-L98","kind":"function","name":"get_stats","path":"src/dust3r/datasets/vkitti2.py","language":"python","start_line":97,"end_line":98,"context_start_line":77,"context_end_line":118,"code":"                scenes.append(seq + \"/\" + cam)\n                scene_img_list.append(img_ids)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                start_img_ids.extend(start_img_ids_)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=0.9,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]","source_hash":"1a1d9ca9ee4628332411dadbb333eb4b6f131d681ade6f4aa15100a8e2b58706","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.vkitti2._get_views","uri":"program://Human3R/function/src.dust3r.datasets.vkitti2._get_views#L100-L169","kind":"function","name":"_get_views","path":"src/dust3r/datasets/vkitti2.py","language":"python","start_line":100,"end_line":169,"context_start_line":80,"context_end_line":169,"code":"                images.extend(basenames)\n                start_img_ids.extend(start_img_ids_)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=0.9,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            basename = self.images[view_idx]\n\n            img = basename + \"_rgb.jpg\"\n            image = imread_cv2(osp.join(scene_dir, img))\n            depthmap = (\n                cv2.imread(\n                    osp.join(scene_dir, basename + \"_depth.png\"),\n                    cv2.IMREAD_ANYCOLOR | cv2.IMREAD_ANYDEPTH,\n                ).astype(np.float32)\n                / 100.0\n            )\n            camera_params = np.load(osp.join(scene_dir, basename + \"_cam.npz\"))\n\n            intrinsics = camera_params[\"camera_intrinsics\"]\n            camera_pose = camera_params[\"camera_pose\"]\n\n            sky_mask = depthmap >= 655\n            depthmap[sky_mask] = -1.0  # sky\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.1, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"VirtualKITTI2\",\n                    label=scene_dir,\n                    is_metric=self.is_metric,\n                    instance=scene_dir + \"_\" + img,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"1a1d9ca9ee4628332411dadbb333eb4b6f131d681ade6f4aa15100a8e2b58706","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.waymo","uri":"program://Human3R/module/src.dust3r.datasets.waymo#L1-L178","kind":"module","name":"src.dust3r.datasets.waymo","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":1,"end_line":178,"context_start_line":1,"context_end_line":178,"code":"import os.path as osp\nimport os\nimport numpy as np\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport h5py\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Waymo_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 8\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        assert self.split is None\n        self._load_data()\n\n    def load_invalid_dict(self, h5_file_path):\n        invalid_dict = {}\n        with h5py.File(h5_file_path, \"r\") as h5f:\n            for scene in h5f:\n                data = h5f[scene][\"invalid_pairs\"][:]\n                invalid_pairs = set(\n                    tuple(pair.decode(\"utf-8\").split(\"_\")) for pair in data\n                )\n                invalid_dict[scene] = invalid_pairs\n        return invalid_dict\n\n    def _load_data(self):\n        invalid_dict = self.load_invalid_dict(\n            os.path.join(self.ROOT, \"invalid_files.h5\")\n        )\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        is_video = []\n        j = 0\n\n        for scene in scene_dirs:\n            scene_dir = osp.join(self.ROOT, scene)\n            invalid_pairs = invalid_dict.get(scene, set())\n            seq2frames = {}\n            for f in os.listdir(scene_dir):\n                if not f.endswith(\".jpg\"):\n                    continue\n                basename = f[:-4]\n                frame_id = basename.split(\"_\")[0]\n                seq_id = basename.split(\"_\")[1]\n                if seq_id == \"5\":\n                    continue\n                if (seq_id, frame_id) in invalid_pairs:\n                    continue  # Skip invalid files\n                if seq_id not in seq2frames:\n                    seq2frames[seq_id] = []\n                seq2frames[seq_id].append(frame_id)\n\n            for seq_id, frame_ids in seq2frames.items():\n                frame_ids = sorted(frame_ids)\n                num_imgs = len(frame_ids)\n                img_ids = list(np.arange(num_imgs) + offset)\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}_{seq_id}\")\n                    continue\n\n                scenes.append((scene, seq_id))\n                sceneids.extend([j] * num_imgs)\n                images.extend(frame_ids)\n                start_img_ids.extend(start_img_ids_)\n                scene_img_list.append(img_ids)\n\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n        self.is_video = is_video\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        _, seq_id = self.scenes[self.sceneids[start_id]]\n        max_interval = self.max_interval // 2 if seq_id == \"4\" else self.max_interval\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=max_interval,\n            video_prob=0.9,\n            fix_interval_prob=0.9,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n        views = []\n        ordered_video = True\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir, seq_id = self.scenes[scene_id]\n            scene_dir = osp.join(self.ROOT, scene_dir)\n            frame_id = self.images[view_idx]\n\n            impath = f\"{frame_id}_{seq_id}\"\n            image = imread_cv2(osp.join(scene_dir, impath + \".jpg\"))\n            depthmap = imread_cv2(osp.join(scene_dir, impath + \".exr\"))\n            camera_params = np.load(osp.join(scene_dir, impath + \".npz\"))\n\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.float32(camera_params[\"cam2world\"])\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, impath)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.10, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"Waymo\",\n                    label=osp.relpath(scene_dir, self.ROOT),\n                    is_metric=self.is_metric,\n                    instance=osp.join(scene_dir, impath + \".jpg\"),\n                    is_video=ordered_video,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n\n        return views","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.waymo.Waymo_Multi","uri":"program://Human3R/class/src.dust3r.datasets.waymo.Waymo_Multi#L12-L178","kind":"class","name":"Waymo_Multi","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":12,"end_line":178,"context_start_line":1,"context_end_line":178,"code":"import os.path as osp\nimport os\nimport numpy as np\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport h5py\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Waymo_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 8\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        assert self.split is None\n        self._load_data()\n\n    def load_invalid_dict(self, h5_file_path):\n        invalid_dict = {}\n        with h5py.File(h5_file_path, \"r\") as h5f:\n            for scene in h5f:\n                data = h5f[scene][\"invalid_pairs\"][:]\n                invalid_pairs = set(\n                    tuple(pair.decode(\"utf-8\").split(\"_\")) for pair in data\n                )\n                invalid_dict[scene] = invalid_pairs\n        return invalid_dict\n\n    def _load_data(self):\n        invalid_dict = self.load_invalid_dict(\n            os.path.join(self.ROOT, \"invalid_files.h5\")\n        )\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        is_video = []\n        j = 0\n\n        for scene in scene_dirs:\n            scene_dir = osp.join(self.ROOT, scene)\n            invalid_pairs = invalid_dict.get(scene, set())\n            seq2frames = {}\n            for f in os.listdir(scene_dir):\n                if not f.endswith(\".jpg\"):\n                    continue\n                basename = f[:-4]\n                frame_id = basename.split(\"_\")[0]\n                seq_id = basename.split(\"_\")[1]\n                if seq_id == \"5\":\n                    continue\n                if (seq_id, frame_id) in invalid_pairs:\n                    continue  # Skip invalid files\n                if seq_id not in seq2frames:\n                    seq2frames[seq_id] = []\n                seq2frames[seq_id].append(frame_id)\n\n            for seq_id, frame_ids in seq2frames.items():\n                frame_ids = sorted(frame_ids)\n                num_imgs = len(frame_ids)\n                img_ids = list(np.arange(num_imgs) + offset)\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}_{seq_id}\")\n                    continue\n\n                scenes.append((scene, seq_id))\n                sceneids.extend([j] * num_imgs)\n                images.extend(frame_ids)\n                start_img_ids.extend(start_img_ids_)\n                scene_img_list.append(img_ids)\n\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n        self.is_video = is_video\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        _, seq_id = self.scenes[self.sceneids[start_id]]\n        max_interval = self.max_interval // 2 if seq_id == \"4\" else self.max_interval\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=max_interval,\n            video_prob=0.9,\n            fix_interval_prob=0.9,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n        views = []\n        ordered_video = True\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir, seq_id = self.scenes[scene_id]\n            scene_dir = osp.join(self.ROOT, scene_dir)\n            frame_id = self.images[view_idx]\n\n            impath = f\"{frame_id}_{seq_id}\"\n            image = imread_cv2(osp.join(scene_dir, impath + \".jpg\"))\n            depthmap = imread_cv2(osp.join(scene_dir, impath + \".exr\"))\n            camera_params = np.load(osp.join(scene_dir, impath + \".npz\"))\n\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.float32(camera_params[\"cam2world\"])\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, impath)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.10, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"Waymo\",\n                    label=osp.relpath(scene_dir, self.ROOT),\n                    is_metric=self.is_metric,\n                    instance=osp.join(scene_dir, impath + \".jpg\"),\n                    is_video=ordered_video,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n\n        return views","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.waymo.__init__","uri":"program://Human3R/function/src.dust3r.datasets.waymo.__init__#L15-L22","kind":"function","name":"__init__","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":15,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport os\nimport numpy as np\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport h5py\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Waymo_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 8\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        assert self.split is None\n        self._load_data()\n\n    def load_invalid_dict(self, h5_file_path):\n        invalid_dict = {}\n        with h5py.File(h5_file_path, \"r\") as h5f:\n            for scene in h5f:\n                data = h5f[scene][\"invalid_pairs\"][:]\n                invalid_pairs = set(\n                    tuple(pair.decode(\"utf-8\").split(\"_\")) for pair in data\n                )\n                invalid_dict[scene] = invalid_pairs\n        return invalid_dict\n\n    def _load_data(self):\n        invalid_dict = self.load_invalid_dict(\n            os.path.join(self.ROOT, \"invalid_files.h5\")\n        )\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.waymo.load_invalid_dict","uri":"program://Human3R/function/src.dust3r.datasets.waymo.load_invalid_dict#L24-L33","kind":"function","name":"load_invalid_dict","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":24,"end_line":33,"context_start_line":4,"context_end_line":53,"code":"import sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport h5py\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Waymo_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 8\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        assert self.split is None\n        self._load_data()\n\n    def load_invalid_dict(self, h5_file_path):\n        invalid_dict = {}\n        with h5py.File(h5_file_path, \"r\") as h5f:\n            for scene in h5f:\n                data = h5f[scene][\"invalid_pairs\"][:]\n                invalid_pairs = set(\n                    tuple(pair.decode(\"utf-8\").split(\"_\")) for pair in data\n                )\n                invalid_dict[scene] = invalid_pairs\n        return invalid_dict\n\n    def _load_data(self):\n        invalid_dict = self.load_invalid_dict(\n            os.path.join(self.ROOT, \"invalid_files.h5\")\n        )\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        is_video = []\n        j = 0","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.waymo._load_data","uri":"program://Human3R/function/src.dust3r.datasets.waymo._load_data#L35-L102","kind":"function","name":"_load_data","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":35,"end_line":102,"context_start_line":15,"context_end_line":122,"code":"    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 8\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        assert self.split is None\n        self._load_data()\n\n    def load_invalid_dict(self, h5_file_path):\n        invalid_dict = {}\n        with h5py.File(h5_file_path, \"r\") as h5f:\n            for scene in h5f:\n                data = h5f[scene][\"invalid_pairs\"][:]\n                invalid_pairs = set(\n                    tuple(pair.decode(\"utf-8\").split(\"_\")) for pair in data\n                )\n                invalid_dict[scene] = invalid_pairs\n        return invalid_dict\n\n    def _load_data(self):\n        invalid_dict = self.load_invalid_dict(\n            os.path.join(self.ROOT, \"invalid_files.h5\")\n        )\n        scene_dirs = sorted(\n            [\n                d\n                for d in os.listdir(self.ROOT)\n                if os.path.isdir(os.path.join(self.ROOT, d))\n            ]\n        )\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        start_img_ids = []\n        scene_img_list = []\n        is_video = []\n        j = 0\n\n        for scene in scene_dirs:\n            scene_dir = osp.join(self.ROOT, scene)\n            invalid_pairs = invalid_dict.get(scene, set())\n            seq2frames = {}\n            for f in os.listdir(scene_dir):\n                if not f.endswith(\".jpg\"):\n                    continue\n                basename = f[:-4]\n                frame_id = basename.split(\"_\")[0]\n                seq_id = basename.split(\"_\")[1]\n                if seq_id == \"5\":\n                    continue\n                if (seq_id, frame_id) in invalid_pairs:\n                    continue  # Skip invalid files\n                if seq_id not in seq2frames:\n                    seq2frames[seq_id] = []\n                seq2frames[seq_id].append(frame_id)\n\n            for seq_id, frame_ids in seq2frames.items():\n                frame_ids = sorted(frame_ids)\n                num_imgs = len(frame_ids)\n                img_ids = list(np.arange(num_imgs) + offset)\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}_{seq_id}\")\n                    continue\n\n                scenes.append((scene, seq_id))\n                sceneids.extend([j] * num_imgs)\n                images.extend(frame_ids)\n                start_img_ids.extend(start_img_ids_)\n                scene_img_list.append(img_ids)\n\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n        self.is_video = is_video\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        _, seq_id = self.scenes[self.sceneids[start_id]]\n        max_interval = self.max_interval // 2 if seq_id == \"4\" else self.max_interval\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.waymo.__len__","uri":"program://Human3R/function/src.dust3r.datasets.waymo.__len__#L104-L105","kind":"function","name":"__len__","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":104,"end_line":105,"context_start_line":84,"context_end_line":125,"code":"                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}_{seq_id}\")\n                    continue\n\n                scenes.append((scene, seq_id))\n                sceneids.extend([j] * num_imgs)\n                images.extend(frame_ids)\n                start_img_ids.extend(start_img_ids_)\n                scene_img_list.append(img_ids)\n\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n        self.is_video = is_video\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        _, seq_id = self.scenes[self.sceneids[start_id]]\n        max_interval = self.max_interval // 2 if seq_id == \"4\" else self.max_interval\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=max_interval,\n            video_prob=0.9,\n            fix_interval_prob=0.9,","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.waymo.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.waymo.get_image_num#L107-L108","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":107,"end_line":108,"context_start_line":87,"context_end_line":128,"code":"\n                scenes.append((scene, seq_id))\n                sceneids.extend([j] * num_imgs)\n                images.extend(frame_ids)\n                start_img_ids.extend(start_img_ids_)\n                scene_img_list.append(img_ids)\n\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n        self.is_video = is_video\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        _, seq_id = self.scenes[self.sceneids[start_id]]\n        max_interval = self.max_interval // 2 if seq_id == \"4\" else self.max_interval\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=max_interval,\n            video_prob=0.9,\n            fix_interval_prob=0.9,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.waymo.get_stats","uri":"program://Human3R/function/src.dust3r.datasets.waymo.get_stats#L110-L111","kind":"function","name":"get_stats","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":110,"end_line":111,"context_start_line":90,"context_end_line":131,"code":"                images.extend(frame_ids)\n                start_img_ids.extend(start_img_ids_)\n                scene_img_list.append(img_ids)\n\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n        self.is_video = is_video\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        _, seq_id = self.scenes[self.sceneids[start_id]]\n        max_interval = self.max_interval // 2 if seq_id == \"4\" else self.max_interval\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=max_interval,\n            video_prob=0.9,\n            fix_interval_prob=0.9,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n        views = []\n        ordered_video = True\n","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.waymo._get_views","uri":"program://Human3R/function/src.dust3r.datasets.waymo._get_views#L113-L178","kind":"function","name":"_get_views","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":113,"end_line":178,"context_start_line":93,"context_end_line":178,"code":"\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n        self.is_video = is_video\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        _, seq_id = self.scenes[self.sceneids[start_id]]\n        max_interval = self.max_interval // 2 if seq_id == \"4\" else self.max_interval\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=max_interval,\n            video_prob=0.9,\n            fix_interval_prob=0.9,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n        views = []\n        ordered_video = True\n\n        views = []\n\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir, seq_id = self.scenes[scene_id]\n            scene_dir = osp.join(self.ROOT, scene_dir)\n            frame_id = self.images[view_idx]\n\n            impath = f\"{frame_id}_{seq_id}\"\n            image = imread_cv2(osp.join(scene_dir, impath + \".jpg\"))\n            depthmap = imread_cv2(osp.join(scene_dir, impath + \".exr\"))\n            camera_params = np.load(osp.join(scene_dir, impath + \".npz\"))\n\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.float32(camera_params[\"cam2world\"])\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, impath)\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.10, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"Waymo\",\n                    label=osp.relpath(scene_dir, self.ROOT),\n                    is_metric=self.is_metric,\n                    instance=osp.join(scene_dir, impath + \".jpg\"),\n                    is_video=ordered_video,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n\n        return views","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mp3d","uri":"program://Human3R/module/src.dust3r.datasets.mp3d#L1-L132","kind":"module","name":"src.dust3r.datasets.mp3d","path":"src/dust3r/datasets/mp3d.py","language":"python","start_line":1,"end_line":132,"context_start_line":1,"context_end_line":132,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MP3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        offset = 0\n        overlaps = {scene: [] for scene in scenes}\n        scene_img_list = {scene: [] for scene in scenes}\n        images = []\n\n        j = 0\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            overlap = np.load(osp.join(scene_dir, \"overlap.npy\"))\n            overlaps[scene] = overlap\n            num_imgs = len(basenames)\n\n            images.extend(\n                [(scene, i, basename) for i, basename in enumerate(basenames)]\n            )\n            scene_img_list[scene] = np.arange(num_imgs) + offset\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.scene_img_list = scene_img_list\n        self.images = images\n        self.overlaps = overlaps\n\n    def __len__(self):\n        return len(self.images)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        num_views_posible = 0\n        num_unique = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        while num_views_posible < num_unique - 1:\n            scene, img_idx, _ = self.images[idx]\n            overlap = self.overlaps[scene]\n            sel_img_idx = np.where(overlap[:, 0] == img_idx)[0]\n            overlap_sel = overlap[sel_img_idx]\n            overlap_sel = overlap_sel[\n                (overlap_sel[:, 2] > 0.01) * (overlap_sel[:, 2] < 1)\n            ]\n            num_views_posible = len(overlap_sel)\n            if num_views_posible >= num_unique - 1:\n                break\n            idx = rng.choice(len(self.images))\n\n        ref_id = self.scene_img_list[scene][img_idx]\n        ids = self.scene_img_list[scene][overlap_sel[:, 1].astype(np.int64)]\n        replace = False if not self.allow_repeat else True\n        image_idxs = rng.choice(\n            ids,\n            num_views - 1,\n            replace=replace,\n            p=overlap_sel[:, 2] / np.sum(overlap_sel[:, 2]),\n        )\n        image_idxs = np.concatenate([[ref_id], image_idxs])\n\n        ordered_video = False\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene, _, basename = self.images[view_idx]\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_path = osp.join(scene_dir, \"rgb\", basename + \".png\")\n            depth_path = osp.join(scene_dir, \"depth\", basename + \".npy\")\n            cam_path = osp.join(scene_dir, \"cam\", basename + \".npz\")\n\n            rgb_image = imread_cv2(rgb_path, cv2.IMREAD_COLOR)\n            depthmap = np.load(depth_path).astype(np.float32)\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            cam_file = np.load(cam_path)\n            intrinsics = cam_file[\"intrinsics\"]\n            camera_pose = cam_file[\"pose\"]\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.1, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"mp3d\",\n                    label=scene + \"_\" + rgb_path,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"3e20428e776f42c93423aa2aabfdbd174febbe6a92c9470d83afd3688050258b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mp3d.MP3D_Multi","uri":"program://Human3R/class/src.dust3r.datasets.mp3d.MP3D_Multi#L14-L132","kind":"class","name":"MP3D_Multi","path":"src/dust3r/datasets/mp3d.py","language":"python","start_line":14,"end_line":132,"context_start_line":1,"context_end_line":132,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MP3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        offset = 0\n        overlaps = {scene: [] for scene in scenes}\n        scene_img_list = {scene: [] for scene in scenes}\n        images = []\n\n        j = 0\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            overlap = np.load(osp.join(scene_dir, \"overlap.npy\"))\n            overlaps[scene] = overlap\n            num_imgs = len(basenames)\n\n            images.extend(\n                [(scene, i, basename) for i, basename in enumerate(basenames)]\n            )\n            scene_img_list[scene] = np.arange(num_imgs) + offset\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.scene_img_list = scene_img_list\n        self.images = images\n        self.overlaps = overlaps\n\n    def __len__(self):\n        return len(self.images)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        num_views_posible = 0\n        num_unique = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        while num_views_posible < num_unique - 1:\n            scene, img_idx, _ = self.images[idx]\n            overlap = self.overlaps[scene]\n            sel_img_idx = np.where(overlap[:, 0] == img_idx)[0]\n            overlap_sel = overlap[sel_img_idx]\n            overlap_sel = overlap_sel[\n                (overlap_sel[:, 2] > 0.01) * (overlap_sel[:, 2] < 1)\n            ]\n            num_views_posible = len(overlap_sel)\n            if num_views_posible >= num_unique - 1:\n                break\n            idx = rng.choice(len(self.images))\n\n        ref_id = self.scene_img_list[scene][img_idx]\n        ids = self.scene_img_list[scene][overlap_sel[:, 1].astype(np.int64)]\n        replace = False if not self.allow_repeat else True\n        image_idxs = rng.choice(\n            ids,\n            num_views - 1,\n            replace=replace,\n            p=overlap_sel[:, 2] / np.sum(overlap_sel[:, 2]),\n        )\n        image_idxs = np.concatenate([[ref_id], image_idxs])\n\n        ordered_video = False\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene, _, basename = self.images[view_idx]\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_path = osp.join(scene_dir, \"rgb\", basename + \".png\")\n            depth_path = osp.join(scene_dir, \"depth\", basename + \".npy\")\n            cam_path = osp.join(scene_dir, \"cam\", basename + \".npz\")\n\n            rgb_image = imread_cv2(rgb_path, cv2.IMREAD_COLOR)\n            depthmap = np.load(depth_path).astype(np.float32)\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            cam_file = np.load(cam_path)\n            intrinsics = cam_file[\"intrinsics\"]\n            camera_pose = cam_file[\"pose\"]\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.1, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"mp3d\",\n                    label=scene + \"_\" + rgb_path,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"3e20428e776f42c93423aa2aabfdbd174febbe6a92c9470d83afd3688050258b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mp3d.__init__","uri":"program://Human3R/function/src.dust3r.datasets.mp3d.__init__#L15-L21","kind":"function","name":"__init__","path":"src/dust3r/datasets/mp3d.py","language":"python","start_line":15,"end_line":21,"context_start_line":1,"context_end_line":41,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MP3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        offset = 0\n        overlaps = {scene: [] for scene in scenes}\n        scene_img_list = {scene: [] for scene in scenes}\n        images = []\n\n        j = 0\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            overlap = np.load(osp.join(scene_dir, \"overlap.npy\"))\n            overlaps[scene] = overlap\n            num_imgs = len(basenames)\n\n            images.extend(","source_hash":"3e20428e776f42c93423aa2aabfdbd174febbe6a92c9470d83afd3688050258b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mp3d._load_data","uri":"program://Human3R/function/src.dust3r.datasets.mp3d._load_data#L23-L51","kind":"function","name":"_load_data","path":"src/dust3r/datasets/mp3d.py","language":"python","start_line":23,"end_line":51,"context_start_line":3,"context_end_line":71,"code":"import sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MP3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        offset = 0\n        overlaps = {scene: [] for scene in scenes}\n        scene_img_list = {scene: [] for scene in scenes}\n        images = []\n\n        j = 0\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            overlap = np.load(osp.join(scene_dir, \"overlap.npy\"))\n            overlaps[scene] = overlap\n            num_imgs = len(basenames)\n\n            images.extend(\n                [(scene, i, basename) for i, basename in enumerate(basenames)]\n            )\n            scene_img_list[scene] = np.arange(num_imgs) + offset\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.scene_img_list = scene_img_list\n        self.images = images\n        self.overlaps = overlaps\n\n    def __len__(self):\n        return len(self.images)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        num_views_posible = 0\n        num_unique = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        while num_views_posible < num_unique - 1:\n            scene, img_idx, _ = self.images[idx]\n            overlap = self.overlaps[scene]\n            sel_img_idx = np.where(overlap[:, 0] == img_idx)[0]\n            overlap_sel = overlap[sel_img_idx]\n            overlap_sel = overlap_sel[\n                (overlap_sel[:, 2] > 0.01) * (overlap_sel[:, 2] < 1)\n            ]\n            num_views_posible = len(overlap_sel)\n            if num_views_posible >= num_unique - 1:","source_hash":"3e20428e776f42c93423aa2aabfdbd174febbe6a92c9470d83afd3688050258b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mp3d.__len__","uri":"program://Human3R/function/src.dust3r.datasets.mp3d.__len__#L53-L54","kind":"function","name":"__len__","path":"src/dust3r/datasets/mp3d.py","language":"python","start_line":53,"end_line":54,"context_start_line":33,"context_end_line":74,"code":"            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            overlap = np.load(osp.join(scene_dir, \"overlap.npy\"))\n            overlaps[scene] = overlap\n            num_imgs = len(basenames)\n\n            images.extend(\n                [(scene, i, basename) for i, basename in enumerate(basenames)]\n            )\n            scene_img_list[scene] = np.arange(num_imgs) + offset\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.scene_img_list = scene_img_list\n        self.images = images\n        self.overlaps = overlaps\n\n    def __len__(self):\n        return len(self.images)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        num_views_posible = 0\n        num_unique = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        while num_views_posible < num_unique - 1:\n            scene, img_idx, _ = self.images[idx]\n            overlap = self.overlaps[scene]\n            sel_img_idx = np.where(overlap[:, 0] == img_idx)[0]\n            overlap_sel = overlap[sel_img_idx]\n            overlap_sel = overlap_sel[\n                (overlap_sel[:, 2] > 0.01) * (overlap_sel[:, 2] < 1)\n            ]\n            num_views_posible = len(overlap_sel)\n            if num_views_posible >= num_unique - 1:\n                break\n            idx = rng.choice(len(self.images))\n","source_hash":"3e20428e776f42c93423aa2aabfdbd174febbe6a92c9470d83afd3688050258b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mp3d.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.mp3d.get_image_num#L56-L57","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/mp3d.py","language":"python","start_line":56,"end_line":57,"context_start_line":36,"context_end_line":77,"code":"            )\n            overlap = np.load(osp.join(scene_dir, \"overlap.npy\"))\n            overlaps[scene] = overlap\n            num_imgs = len(basenames)\n\n            images.extend(\n                [(scene, i, basename) for i, basename in enumerate(basenames)]\n            )\n            scene_img_list[scene] = np.arange(num_imgs) + offset\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.scene_img_list = scene_img_list\n        self.images = images\n        self.overlaps = overlaps\n\n    def __len__(self):\n        return len(self.images)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        num_views_posible = 0\n        num_unique = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        while num_views_posible < num_unique - 1:\n            scene, img_idx, _ = self.images[idx]\n            overlap = self.overlaps[scene]\n            sel_img_idx = np.where(overlap[:, 0] == img_idx)[0]\n            overlap_sel = overlap[sel_img_idx]\n            overlap_sel = overlap_sel[\n                (overlap_sel[:, 2] > 0.01) * (overlap_sel[:, 2] < 1)\n            ]\n            num_views_posible = len(overlap_sel)\n            if num_views_posible >= num_unique - 1:\n                break\n            idx = rng.choice(len(self.images))\n\n        ref_id = self.scene_img_list[scene][img_idx]\n        ids = self.scene_img_list[scene][overlap_sel[:, 1].astype(np.int64)]\n        replace = False if not self.allow_repeat else True","source_hash":"3e20428e776f42c93423aa2aabfdbd174febbe6a92c9470d83afd3688050258b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mp3d._get_views","uri":"program://Human3R/function/src.dust3r.datasets.mp3d._get_views#L59-L132","kind":"function","name":"_get_views","path":"src/dust3r/datasets/mp3d.py","language":"python","start_line":59,"end_line":132,"context_start_line":39,"context_end_line":132,"code":"            num_imgs = len(basenames)\n\n            images.extend(\n                [(scene, i, basename) for i, basename in enumerate(basenames)]\n            )\n            scene_img_list[scene] = np.arange(num_imgs) + offset\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.scene_img_list = scene_img_list\n        self.images = images\n        self.overlaps = overlaps\n\n    def __len__(self):\n        return len(self.images)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        num_views_posible = 0\n        num_unique = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        while num_views_posible < num_unique - 1:\n            scene, img_idx, _ = self.images[idx]\n            overlap = self.overlaps[scene]\n            sel_img_idx = np.where(overlap[:, 0] == img_idx)[0]\n            overlap_sel = overlap[sel_img_idx]\n            overlap_sel = overlap_sel[\n                (overlap_sel[:, 2] > 0.01) * (overlap_sel[:, 2] < 1)\n            ]\n            num_views_posible = len(overlap_sel)\n            if num_views_posible >= num_unique - 1:\n                break\n            idx = rng.choice(len(self.images))\n\n        ref_id = self.scene_img_list[scene][img_idx]\n        ids = self.scene_img_list[scene][overlap_sel[:, 1].astype(np.int64)]\n        replace = False if not self.allow_repeat else True\n        image_idxs = rng.choice(\n            ids,\n            num_views - 1,\n            replace=replace,\n            p=overlap_sel[:, 2] / np.sum(overlap_sel[:, 2]),\n        )\n        image_idxs = np.concatenate([[ref_id], image_idxs])\n\n        ordered_video = False\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene, _, basename = self.images[view_idx]\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_path = osp.join(scene_dir, \"rgb\", basename + \".png\")\n            depth_path = osp.join(scene_dir, \"depth\", basename + \".npy\")\n            cam_path = osp.join(scene_dir, \"cam\", basename + \".npz\")\n\n            rgb_image = imread_cv2(rgb_path, cv2.IMREAD_COLOR)\n            depthmap = np.load(depth_path).astype(np.float32)\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            cam_file = np.load(cam_path)\n            intrinsics = cam_file[\"intrinsics\"]\n            camera_pose = cam_file[\"pose\"]\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.1, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"mp3d\",\n                    label=scene + \"_\" + rgb_path,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"3e20428e776f42c93423aa2aabfdbd174febbe6a92c9470d83afd3688050258b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.megadepth","uri":"program://Human3R/module/src.dust3r.datasets.megadepth#L1-L98","kind":"module","name":"src.dust3r.datasets.megadepth","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":1,"end_line":98,"context_start_line":1,"context_end_line":98,"code":"import os.path as osp\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MegaDepth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        super().__init__(*args, **kwargs)\n        self._load_data(self.split)\n        self.is_metric = False\n        if self.split is None:\n            pass\n        elif self.split == \"train\":\n            self.select_scene((\"0015\", \"0022\"), opposite=True)\n        elif self.split == \"val\":\n            self.select_scene((\"0015\", \"0022\"))\n        else:\n            raise ValueError(f\"bad {self.split=}\")\n\n    def _load_data(self, split):\n        with np.load(\n            osp.join(self.ROOT, \"megadepth_sets_64.npz\"), allow_pickle=True\n        ) as data:\n            self.all_scenes = data[\"scenes\"]\n            self.all_images = data[\"images\"]\n            self.sets = data[\"sets\"]\n\n    def __len__(self):\n        return len(self.sets)\n\n    def get_image_num(self):\n        return len(self.all_images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups from {len(self.all_scenes)} scenes\"\n\n    def select_scene(self, scene, *instances, opposite=False):\n        scenes = (scene,) if isinstance(scene, str) else tuple(scene)\n        scene_id = [s.startswith(scenes) for s in self.all_scenes]\n        assert any(scene_id), \"no scene found\"\n        valid = np.in1d(self.sets[:, 0], np.nonzero(scene_id)[0])\n        if instances:\n            raise NotImplementedError(\"selecting instances not implemented\")\n        if opposite:\n            valid = ~valid\n        assert valid.any()\n        self.sets = self.sets[valid]\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene_id = self.sets[idx][0]\n        image_idxs = self.sets[idx][1:65]\n        replace = False if not self.allow_repeat else True\n        image_idxs = rng.choice(image_idxs, num_views, replace=replace)\n        scene, subscene = self.all_scenes[scene_id].split()\n        seq_path = osp.join(self.ROOT, scene, subscene)\n        views = []\n        for im_id in image_idxs:\n            img = self.all_images[im_id]\n            try:\n                image = imread_cv2(osp.join(seq_path, img + \".jpg\"))\n                depthmap = imread_cv2(osp.join(seq_path, img + \".exr\"))\n                camera_params = np.load(osp.join(seq_path, img + \".npz\"))\n            except Exception as e:\n                raise OSError(f\"cannot load {img}, got exception {e}\")\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.float32(camera_params[\"cam2world\"])\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(seq_path, img)\n            )\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"MegaDepth\",\n                    label=osp.relpath(seq_path, self.ROOT),\n                    is_metric=self.is_metric,\n                    instance=img,\n                    is_video=False,\n                    quantile=np.array(0.96, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.megadepth.MegaDepth_Multi","uri":"program://Human3R/class/src.dust3r.datasets.megadepth.MegaDepth_Multi#L12-L98","kind":"class","name":"MegaDepth_Multi","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":12,"end_line":98,"context_start_line":1,"context_end_line":98,"code":"import os.path as osp\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MegaDepth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        super().__init__(*args, **kwargs)\n        self._load_data(self.split)\n        self.is_metric = False\n        if self.split is None:\n            pass\n        elif self.split == \"train\":\n            self.select_scene((\"0015\", \"0022\"), opposite=True)\n        elif self.split == \"val\":\n            self.select_scene((\"0015\", \"0022\"))\n        else:\n            raise ValueError(f\"bad {self.split=}\")\n\n    def _load_data(self, split):\n        with np.load(\n            osp.join(self.ROOT, \"megadepth_sets_64.npz\"), allow_pickle=True\n        ) as data:\n            self.all_scenes = data[\"scenes\"]\n            self.all_images = data[\"images\"]\n            self.sets = data[\"sets\"]\n\n    def __len__(self):\n        return len(self.sets)\n\n    def get_image_num(self):\n        return len(self.all_images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups from {len(self.all_scenes)} scenes\"\n\n    def select_scene(self, scene, *instances, opposite=False):\n        scenes = (scene,) if isinstance(scene, str) else tuple(scene)\n        scene_id = [s.startswith(scenes) for s in self.all_scenes]\n        assert any(scene_id), \"no scene found\"\n        valid = np.in1d(self.sets[:, 0], np.nonzero(scene_id)[0])\n        if instances:\n            raise NotImplementedError(\"selecting instances not implemented\")\n        if opposite:\n            valid = ~valid\n        assert valid.any()\n        self.sets = self.sets[valid]\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene_id = self.sets[idx][0]\n        image_idxs = self.sets[idx][1:65]\n        replace = False if not self.allow_repeat else True\n        image_idxs = rng.choice(image_idxs, num_views, replace=replace)\n        scene, subscene = self.all_scenes[scene_id].split()\n        seq_path = osp.join(self.ROOT, scene, subscene)\n        views = []\n        for im_id in image_idxs:\n            img = self.all_images[im_id]\n            try:\n                image = imread_cv2(osp.join(seq_path, img + \".jpg\"))\n                depthmap = imread_cv2(osp.join(seq_path, img + \".exr\"))\n                camera_params = np.load(osp.join(seq_path, img + \".npz\"))\n            except Exception as e:\n                raise OSError(f\"cannot load {img}, got exception {e}\")\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.float32(camera_params[\"cam2world\"])\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(seq_path, img)\n            )\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"MegaDepth\",\n                    label=osp.relpath(seq_path, self.ROOT),\n                    is_metric=self.is_metric,\n                    instance=img,\n                    is_video=False,\n                    quantile=np.array(0.96, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.megadepth.__init__","uri":"program://Human3R/function/src.dust3r.datasets.megadepth.__init__#L13-L25","kind":"function","name":"__init__","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":13,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"import os.path as osp\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MegaDepth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        super().__init__(*args, **kwargs)\n        self._load_data(self.split)\n        self.is_metric = False\n        if self.split is None:\n            pass\n        elif self.split == \"train\":\n            self.select_scene((\"0015\", \"0022\"), opposite=True)\n        elif self.split == \"val\":\n            self.select_scene((\"0015\", \"0022\"))\n        else:\n            raise ValueError(f\"bad {self.split=}\")\n\n    def _load_data(self, split):\n        with np.load(\n            osp.join(self.ROOT, \"megadepth_sets_64.npz\"), allow_pickle=True\n        ) as data:\n            self.all_scenes = data[\"scenes\"]\n            self.all_images = data[\"images\"]\n            self.sets = data[\"sets\"]\n\n    def __len__(self):\n        return len(self.sets)\n\n    def get_image_num(self):\n        return len(self.all_images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups from {len(self.all_scenes)} scenes\"\n\n    def select_scene(self, scene, *instances, opposite=False):\n        scenes = (scene,) if isinstance(scene, str) else tuple(scene)","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.megadepth._load_data","uri":"program://Human3R/function/src.dust3r.datasets.megadepth._load_data#L27-L33","kind":"function","name":"_load_data","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":27,"end_line":33,"context_start_line":7,"context_end_line":53,"code":"sys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MegaDepth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        super().__init__(*args, **kwargs)\n        self._load_data(self.split)\n        self.is_metric = False\n        if self.split is None:\n            pass\n        elif self.split == \"train\":\n            self.select_scene((\"0015\", \"0022\"), opposite=True)\n        elif self.split == \"val\":\n            self.select_scene((\"0015\", \"0022\"))\n        else:\n            raise ValueError(f\"bad {self.split=}\")\n\n    def _load_data(self, split):\n        with np.load(\n            osp.join(self.ROOT, \"megadepth_sets_64.npz\"), allow_pickle=True\n        ) as data:\n            self.all_scenes = data[\"scenes\"]\n            self.all_images = data[\"images\"]\n            self.sets = data[\"sets\"]\n\n    def __len__(self):\n        return len(self.sets)\n\n    def get_image_num(self):\n        return len(self.all_images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups from {len(self.all_scenes)} scenes\"\n\n    def select_scene(self, scene, *instances, opposite=False):\n        scenes = (scene,) if isinstance(scene, str) else tuple(scene)\n        scene_id = [s.startswith(scenes) for s in self.all_scenes]\n        assert any(scene_id), \"no scene found\"\n        valid = np.in1d(self.sets[:, 0], np.nonzero(scene_id)[0])\n        if instances:\n            raise NotImplementedError(\"selecting instances not implemented\")\n        if opposite:\n            valid = ~valid\n        assert valid.any()","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.megadepth.__len__","uri":"program://Human3R/function/src.dust3r.datasets.megadepth.__len__#L35-L36","kind":"function","name":"__len__","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":35,"end_line":36,"context_start_line":15,"context_end_line":56,"code":"        super().__init__(*args, **kwargs)\n        self._load_data(self.split)\n        self.is_metric = False\n        if self.split is None:\n            pass\n        elif self.split == \"train\":\n            self.select_scene((\"0015\", \"0022\"), opposite=True)\n        elif self.split == \"val\":\n            self.select_scene((\"0015\", \"0022\"))\n        else:\n            raise ValueError(f\"bad {self.split=}\")\n\n    def _load_data(self, split):\n        with np.load(\n            osp.join(self.ROOT, \"megadepth_sets_64.npz\"), allow_pickle=True\n        ) as data:\n            self.all_scenes = data[\"scenes\"]\n            self.all_images = data[\"images\"]\n            self.sets = data[\"sets\"]\n\n    def __len__(self):\n        return len(self.sets)\n\n    def get_image_num(self):\n        return len(self.all_images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups from {len(self.all_scenes)} scenes\"\n\n    def select_scene(self, scene, *instances, opposite=False):\n        scenes = (scene,) if isinstance(scene, str) else tuple(scene)\n        scene_id = [s.startswith(scenes) for s in self.all_scenes]\n        assert any(scene_id), \"no scene found\"\n        valid = np.in1d(self.sets[:, 0], np.nonzero(scene_id)[0])\n        if instances:\n            raise NotImplementedError(\"selecting instances not implemented\")\n        if opposite:\n            valid = ~valid\n        assert valid.any()\n        self.sets = self.sets[valid]\n\n    def _get_views(self, idx, resolution, rng, num_views):","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.megadepth.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.megadepth.get_image_num#L38-L39","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":38,"end_line":39,"context_start_line":18,"context_end_line":59,"code":"        if self.split is None:\n            pass\n        elif self.split == \"train\":\n            self.select_scene((\"0015\", \"0022\"), opposite=True)\n        elif self.split == \"val\":\n            self.select_scene((\"0015\", \"0022\"))\n        else:\n            raise ValueError(f\"bad {self.split=}\")\n\n    def _load_data(self, split):\n        with np.load(\n            osp.join(self.ROOT, \"megadepth_sets_64.npz\"), allow_pickle=True\n        ) as data:\n            self.all_scenes = data[\"scenes\"]\n            self.all_images = data[\"images\"]\n            self.sets = data[\"sets\"]\n\n    def __len__(self):\n        return len(self.sets)\n\n    def get_image_num(self):\n        return len(self.all_images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups from {len(self.all_scenes)} scenes\"\n\n    def select_scene(self, scene, *instances, opposite=False):\n        scenes = (scene,) if isinstance(scene, str) else tuple(scene)\n        scene_id = [s.startswith(scenes) for s in self.all_scenes]\n        assert any(scene_id), \"no scene found\"\n        valid = np.in1d(self.sets[:, 0], np.nonzero(scene_id)[0])\n        if instances:\n            raise NotImplementedError(\"selecting instances not implemented\")\n        if opposite:\n            valid = ~valid\n        assert valid.any()\n        self.sets = self.sets[valid]\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene_id = self.sets[idx][0]\n        image_idxs = self.sets[idx][1:65]\n        replace = False if not self.allow_repeat else True","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.megadepth.get_stats","uri":"program://Human3R/function/src.dust3r.datasets.megadepth.get_stats#L41-L42","kind":"function","name":"get_stats","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":41,"end_line":42,"context_start_line":21,"context_end_line":62,"code":"            self.select_scene((\"0015\", \"0022\"), opposite=True)\n        elif self.split == \"val\":\n            self.select_scene((\"0015\", \"0022\"))\n        else:\n            raise ValueError(f\"bad {self.split=}\")\n\n    def _load_data(self, split):\n        with np.load(\n            osp.join(self.ROOT, \"megadepth_sets_64.npz\"), allow_pickle=True\n        ) as data:\n            self.all_scenes = data[\"scenes\"]\n            self.all_images = data[\"images\"]\n            self.sets = data[\"sets\"]\n\n    def __len__(self):\n        return len(self.sets)\n\n    def get_image_num(self):\n        return len(self.all_images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups from {len(self.all_scenes)} scenes\"\n\n    def select_scene(self, scene, *instances, opposite=False):\n        scenes = (scene,) if isinstance(scene, str) else tuple(scene)\n        scene_id = [s.startswith(scenes) for s in self.all_scenes]\n        assert any(scene_id), \"no scene found\"\n        valid = np.in1d(self.sets[:, 0], np.nonzero(scene_id)[0])\n        if instances:\n            raise NotImplementedError(\"selecting instances not implemented\")\n        if opposite:\n            valid = ~valid\n        assert valid.any()\n        self.sets = self.sets[valid]\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene_id = self.sets[idx][0]\n        image_idxs = self.sets[idx][1:65]\n        replace = False if not self.allow_repeat else True\n        image_idxs = rng.choice(image_idxs, num_views, replace=replace)\n        scene, subscene = self.all_scenes[scene_id].split()\n        seq_path = osp.join(self.ROOT, scene, subscene)","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.megadepth.select_scene","uri":"program://Human3R/function/src.dust3r.datasets.megadepth.select_scene#L44-L54","kind":"function","name":"select_scene","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":44,"end_line":54,"context_start_line":24,"context_end_line":74,"code":"        else:\n            raise ValueError(f\"bad {self.split=}\")\n\n    def _load_data(self, split):\n        with np.load(\n            osp.join(self.ROOT, \"megadepth_sets_64.npz\"), allow_pickle=True\n        ) as data:\n            self.all_scenes = data[\"scenes\"]\n            self.all_images = data[\"images\"]\n            self.sets = data[\"sets\"]\n\n    def __len__(self):\n        return len(self.sets)\n\n    def get_image_num(self):\n        return len(self.all_images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups from {len(self.all_scenes)} scenes\"\n\n    def select_scene(self, scene, *instances, opposite=False):\n        scenes = (scene,) if isinstance(scene, str) else tuple(scene)\n        scene_id = [s.startswith(scenes) for s in self.all_scenes]\n        assert any(scene_id), \"no scene found\"\n        valid = np.in1d(self.sets[:, 0], np.nonzero(scene_id)[0])\n        if instances:\n            raise NotImplementedError(\"selecting instances not implemented\")\n        if opposite:\n            valid = ~valid\n        assert valid.any()\n        self.sets = self.sets[valid]\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene_id = self.sets[idx][0]\n        image_idxs = self.sets[idx][1:65]\n        replace = False if not self.allow_repeat else True\n        image_idxs = rng.choice(image_idxs, num_views, replace=replace)\n        scene, subscene = self.all_scenes[scene_id].split()\n        seq_path = osp.join(self.ROOT, scene, subscene)\n        views = []\n        for im_id in image_idxs:\n            img = self.all_images[im_id]\n            try:\n                image = imread_cv2(osp.join(seq_path, img + \".jpg\"))\n                depthmap = imread_cv2(osp.join(seq_path, img + \".exr\"))\n                camera_params = np.load(osp.join(seq_path, img + \".npz\"))\n            except Exception as e:\n                raise OSError(f\"cannot load {img}, got exception {e}\")\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.float32(camera_params[\"cam2world\"])\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.megadepth._get_views","uri":"program://Human3R/function/src.dust3r.datasets.megadepth._get_views#L56-L98","kind":"function","name":"_get_views","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":56,"end_line":98,"context_start_line":36,"context_end_line":98,"code":"        return len(self.sets)\n\n    def get_image_num(self):\n        return len(self.all_images)\n\n    def get_stats(self):\n        return f\"{len(self)} groups from {len(self.all_scenes)} scenes\"\n\n    def select_scene(self, scene, *instances, opposite=False):\n        scenes = (scene,) if isinstance(scene, str) else tuple(scene)\n        scene_id = [s.startswith(scenes) for s in self.all_scenes]\n        assert any(scene_id), \"no scene found\"\n        valid = np.in1d(self.sets[:, 0], np.nonzero(scene_id)[0])\n        if instances:\n            raise NotImplementedError(\"selecting instances not implemented\")\n        if opposite:\n            valid = ~valid\n        assert valid.any()\n        self.sets = self.sets[valid]\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene_id = self.sets[idx][0]\n        image_idxs = self.sets[idx][1:65]\n        replace = False if not self.allow_repeat else True\n        image_idxs = rng.choice(image_idxs, num_views, replace=replace)\n        scene, subscene = self.all_scenes[scene_id].split()\n        seq_path = osp.join(self.ROOT, scene, subscene)\n        views = []\n        for im_id in image_idxs:\n            img = self.all_images[im_id]\n            try:\n                image = imread_cv2(osp.join(seq_path, img + \".jpg\"))\n                depthmap = imread_cv2(osp.join(seq_path, img + \".exr\"))\n                camera_params = np.load(osp.join(seq_path, img + \".npz\"))\n            except Exception as e:\n                raise OSError(f\"cannot load {img}, got exception {e}\")\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.float32(camera_params[\"cam2world\"])\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(seq_path, img)\n            )\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"MegaDepth\",\n                    label=osp.relpath(seq_path, self.ROOT),\n                    is_metric=self.is_metric,\n                    instance=img,\n                    is_video=False,\n                    quantile=np.array(0.96, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dl3dv","uri":"program://Human3R/module/src.dust3r.datasets.dl3dv#L1-L166","kind":"module","name":"src.dust3r.datasets.dl3dv","path":"src/dust3r/datasets/dl3dv.py","language":"python","start_line":1,"end_line":166,"context_start_line":1,"context_end_line":166,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DL3DV_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 20\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.all_scenes = sorted(\n            [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))]\n        )\n        subscenes = []\n        for scene in self.all_scenes:\n            # not empty\n            subscenes.extend(\n                [\n                    osp.join(scene, f)\n                    for f in os.listdir(osp.join(self.ROOT, scene))\n                    if os.path.isdir(osp.join(self.ROOT, scene, f))\n                    and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0\n                ]\n            )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for scene_idx, scene in enumerate(subscenes):\n            scene_dir = osp.join(self.ROOT, scene, \"dense\")\n            rgb_paths = sorted(\n                [\n                    f\n                    for f in os.listdir(os.path.join(scene_dir, \"rgb\"))\n                    if f.endswith(\".png\")\n                ]\n            )\n            assert len(rgb_paths) > 0, f\"{scene_dir} is empty.\"\n            num_imgs = len(rgb_paths)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=25,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id], \"dense\")\n\n            rgb_path = self.images[view_idx]\n            basename = rgb_path[:-4]\n\n            rgb_image = imread_cv2(\n                osp.join(scene_dir, \"rgb\", rgb_path), cv2.IMREAD_COLOR\n            )\n            depthmap = np.load(osp.join(scene_dir, \"depth\", basename + \".npy\")).astype(\n                np.float32\n            )\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            cam_file = np.load(osp.join(scene_dir, \"cam\", basename + \".npz\"))\n            sky_mask = (\n                cv2.imread(\n                    osp.join(scene_dir, \"sky_mask\", rgb_path), cv2.IMREAD_UNCHANGED\n                )\n                >= 127\n            )\n            outlier_mask = cv2.imread(\n                osp.join(scene_dir, \"outlier_mask\", rgb_path), cv2.IMREAD_UNCHANGED\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap[outlier_mask >= 127] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n\n            intrinsics = cam_file[\"intrinsic\"].astype(np.float32)\n            camera_pose = cam_file[\"pose\"].astype(np.float32)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"dl3dv\",\n                    label=self.scenes[scene_id] + \"_\" + rgb_path,\n                    instance=osp.join(scene_dir, \"rgb\", rgb_path),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.9, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        return views","source_hash":"a6df42342e9cf8ec382750f73d79bb007161e1bf5a7dbd4845f22291feafb2e1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dl3dv.DL3DV_Multi","uri":"program://Human3R/class/src.dust3r.datasets.dl3dv.DL3DV_Multi#L14-L166","kind":"class","name":"DL3DV_Multi","path":"src/dust3r/datasets/dl3dv.py","language":"python","start_line":14,"end_line":166,"context_start_line":1,"context_end_line":166,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DL3DV_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 20\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.all_scenes = sorted(\n            [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))]\n        )\n        subscenes = []\n        for scene in self.all_scenes:\n            # not empty\n            subscenes.extend(\n                [\n                    osp.join(scene, f)\n                    for f in os.listdir(osp.join(self.ROOT, scene))\n                    if os.path.isdir(osp.join(self.ROOT, scene, f))\n                    and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0\n                ]\n            )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for scene_idx, scene in enumerate(subscenes):\n            scene_dir = osp.join(self.ROOT, scene, \"dense\")\n            rgb_paths = sorted(\n                [\n                    f\n                    for f in os.listdir(os.path.join(scene_dir, \"rgb\"))\n                    if f.endswith(\".png\")\n                ]\n            )\n            assert len(rgb_paths) > 0, f\"{scene_dir} is empty.\"\n            num_imgs = len(rgb_paths)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=25,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id], \"dense\")\n\n            rgb_path = self.images[view_idx]\n            basename = rgb_path[:-4]\n\n            rgb_image = imread_cv2(\n                osp.join(scene_dir, \"rgb\", rgb_path), cv2.IMREAD_COLOR\n            )\n            depthmap = np.load(osp.join(scene_dir, \"depth\", basename + \".npy\")).astype(\n                np.float32\n            )\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            cam_file = np.load(osp.join(scene_dir, \"cam\", basename + \".npz\"))\n            sky_mask = (\n                cv2.imread(\n                    osp.join(scene_dir, \"sky_mask\", rgb_path), cv2.IMREAD_UNCHANGED\n                )\n                >= 127\n            )\n            outlier_mask = cv2.imread(\n                osp.join(scene_dir, \"outlier_mask\", rgb_path), cv2.IMREAD_UNCHANGED\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap[outlier_mask >= 127] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n\n            intrinsics = cam_file[\"intrinsic\"].astype(np.float32)\n            camera_pose = cam_file[\"pose\"].astype(np.float32)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"dl3dv\",\n                    label=self.scenes[scene_id] + \"_\" + rgb_path,\n                    instance=osp.join(scene_dir, \"rgb\", rgb_path),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.9, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        return views","source_hash":"a6df42342e9cf8ec382750f73d79bb007161e1bf5a7dbd4845f22291feafb2e1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dl3dv.__init__","uri":"program://Human3R/function/src.dust3r.datasets.dl3dv.__init__#L15-L22","kind":"function","name":"__init__","path":"src/dust3r/datasets/dl3dv.py","language":"python","start_line":15,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DL3DV_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 20\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.all_scenes = sorted(\n            [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))]\n        )\n        subscenes = []\n        for scene in self.all_scenes:\n            # not empty\n            subscenes.extend(\n                [\n                    osp.join(scene, f)\n                    for f in os.listdir(osp.join(self.ROOT, scene))\n                    if os.path.isdir(osp.join(self.ROOT, scene, f))\n                    and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0\n                ]\n            )\n\n        offset = 0\n        scenes = []\n        sceneids = []","source_hash":"a6df42342e9cf8ec382750f73d79bb007161e1bf5a7dbd4845f22291feafb2e1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dl3dv._load_data","uri":"program://Human3R/function/src.dust3r.datasets.dl3dv._load_data#L24-L82","kind":"function","name":"_load_data","path":"src/dust3r/datasets/dl3dv.py","language":"python","start_line":24,"end_line":82,"context_start_line":4,"context_end_line":102,"code":"import itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DL3DV_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 20\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.all_scenes = sorted(\n            [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))]\n        )\n        subscenes = []\n        for scene in self.all_scenes:\n            # not empty\n            subscenes.extend(\n                [\n                    osp.join(scene, f)\n                    for f in os.listdir(osp.join(self.ROOT, scene))\n                    if os.path.isdir(osp.join(self.ROOT, scene, f))\n                    and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0\n                ]\n            )\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        scene_img_list = []\n        start_img_ids = []\n        j = 0\n\n        for scene_idx, scene in enumerate(subscenes):\n            scene_dir = osp.join(self.ROOT, scene, \"dense\")\n            rgb_paths = sorted(\n                [\n                    f\n                    for f in os.listdir(os.path.join(scene_dir, \"rgb\"))\n                    if f.endswith(\".png\")\n                ]\n            )\n            assert len(rgb_paths) > 0, f\"{scene_dir} is empty.\"\n            num_imgs = len(rgb_paths)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=25,\n        )\n        image_idxs = np.array(all_image_ids)[pos]","source_hash":"a6df42342e9cf8ec382750f73d79bb007161e1bf5a7dbd4845f22291feafb2e1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dl3dv.__len__","uri":"program://Human3R/function/src.dust3r.datasets.dl3dv.__len__#L84-L85","kind":"function","name":"__len__","path":"src/dust3r/datasets/dl3dv.py","language":"python","start_line":84,"end_line":85,"context_start_line":64,"context_end_line":105,"code":"                print(f\"Skipping {scene}\")\n                continue\n\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=25,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for view_idx in image_idxs:","source_hash":"a6df42342e9cf8ec382750f73d79bb007161e1bf5a7dbd4845f22291feafb2e1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dl3dv.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.dl3dv.get_image_num#L87-L88","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/dl3dv.py","language":"python","start_line":87,"end_line":88,"context_start_line":67,"context_end_line":108,"code":"            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=25,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id], \"dense\")\n","source_hash":"a6df42342e9cf8ec382750f73d79bb007161e1bf5a7dbd4845f22291feafb2e1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dl3dv._get_views","uri":"program://Human3R/function/src.dust3r.datasets.dl3dv._get_views#L90-L166","kind":"function","name":"_get_views","path":"src/dust3r/datasets/dl3dv.py","language":"python","start_line":90,"end_line":166,"context_start_line":70,"context_end_line":166,"code":"            scenes.append(scene)\n            scene_img_list.append(img_ids)\n            sceneids.extend([j] * num_imgs)\n            images.extend(rgb_paths)\n            start_img_ids.extend(start_img_ids_)\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        scene_id = self.sceneids[start_id]\n        all_image_ids = self.scene_img_list[scene_id]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            block_shuffle=25,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for view_idx in image_idxs:\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id], \"dense\")\n\n            rgb_path = self.images[view_idx]\n            basename = rgb_path[:-4]\n\n            rgb_image = imread_cv2(\n                osp.join(scene_dir, \"rgb\", rgb_path), cv2.IMREAD_COLOR\n            )\n            depthmap = np.load(osp.join(scene_dir, \"depth\", basename + \".npy\")).astype(\n                np.float32\n            )\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            cam_file = np.load(osp.join(scene_dir, \"cam\", basename + \".npz\"))\n            sky_mask = (\n                cv2.imread(\n                    osp.join(scene_dir, \"sky_mask\", rgb_path), cv2.IMREAD_UNCHANGED\n                )\n                >= 127\n            )\n            outlier_mask = cv2.imread(\n                osp.join(scene_dir, \"outlier_mask\", rgb_path), cv2.IMREAD_UNCHANGED\n            )\n            depthmap[sky_mask] = -1.0\n            depthmap[outlier_mask >= 127] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n\n            intrinsics = cam_file[\"intrinsic\"].astype(np.float32)\n            camera_pose = cam_file[\"pose\"].astype(np.float32)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"dl3dv\",\n                    label=self.scenes[scene_id] + \"_\" + rgb_path,\n                    instance=osp.join(scene_dir, \"rgb\", rgb_path),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.9, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        return views","source_hash":"a6df42342e9cf8ec382750f73d79bb007161e1bf5a7dbd4845f22291feafb2e1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvs_synth","uri":"program://Human3R/module/src.dust3r.datasets.mvs_synth#L1-L143","kind":"module","name":"src.dust3r.datasets.mvs_synth","path":"src/dust3r/datasets/mvs_synth.py","language":"python","start_line":1,"end_line":143,"context_start_line":1,"context_end_line":143,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVS_Synth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n            num_imgs = len(basenames)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n            depthmap[depthmap > 1000] = 0.0\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.8, 0.15, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"MVS_Synth\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".jpg\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"c87b10ff5331934399cb4c883a051b6432c67c316edab551daee3998ab8d36f3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvs_synth.MVS_Synth_Multi","uri":"program://Human3R/class/src.dust3r.datasets.mvs_synth.MVS_Synth_Multi#L14-L143","kind":"class","name":"MVS_Synth_Multi","path":"src/dust3r/datasets/mvs_synth.py","language":"python","start_line":14,"end_line":143,"context_start_line":1,"context_end_line":143,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVS_Synth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n            num_imgs = len(basenames)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n            depthmap[depthmap > 1000] = 0.0\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.8, 0.15, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"MVS_Synth\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".jpg\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"c87b10ff5331934399cb4c883a051b6432c67c316edab551daee3998ab8d36f3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvs_synth.__init__","uri":"program://Human3R/function/src.dust3r.datasets.mvs_synth.__init__#L15-L21","kind":"function","name":"__init__","path":"src/dust3r/datasets/mvs_synth.py","language":"python","start_line":15,"end_line":21,"context_start_line":1,"context_end_line":41,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVS_Synth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n            num_imgs = len(basenames)\n            cut_off = (","source_hash":"c87b10ff5331934399cb4c883a051b6432c67c316edab551daee3998ab8d36f3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvs_synth._load_data","uri":"program://Human3R/function/src.dust3r.datasets.mvs_synth._load_data#L23-L65","kind":"function","name":"_load_data","path":"src/dust3r/datasets/mvs_synth.py","language":"python","start_line":23,"end_line":65,"context_start_line":3,"context_end_line":85,"code":"import numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVS_Synth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = os.listdir(self.ROOT)\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".jpg\")]\n            )\n            num_imgs = len(basenames)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]","source_hash":"c87b10ff5331934399cb4c883a051b6432c67c316edab551daee3998ab8d36f3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvs_synth.__len__","uri":"program://Human3R/function/src.dust3r.datasets.mvs_synth.__len__#L67-L68","kind":"function","name":"__len__","path":"src/dust3r/datasets/mvs_synth.py","language":"python","start_line":67,"end_line":68,"context_start_line":47,"context_end_line":88,"code":"                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):","source_hash":"c87b10ff5331934399cb4c883a051b6432c67c316edab551daee3998ab8d36f3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvs_synth.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.mvs_synth.get_image_num#L70-L71","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/mvs_synth.py","language":"python","start_line":70,"end_line":71,"context_start_line":50,"context_end_line":91,"code":"\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")","source_hash":"c87b10ff5331934399cb4c883a051b6432c67c316edab551daee3998ab8d36f3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.mvs_synth._get_views","uri":"program://Human3R/function/src.dust3r.datasets.mvs_synth._get_views#L73-L143","kind":"function","name":"_get_views","path":"src/dust3r/datasets/mvs_synth.py","language":"python","start_line":73,"end_line":143,"context_start_line":53,"context_end_line":143,"code":"            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n            threshold = (\n                np.percentile(depthmap[depthmap > 0], 98)\n                if depthmap[depthmap > 0].size > 0\n                else 0\n            )\n            depthmap[depthmap > threshold] = 0.0\n            depthmap[depthmap > 1000] = 0.0\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.8, 0.15, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"MVS_Synth\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=osp.join(rgb_dir, basename + \".jpg\"),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"c87b10ff5331934399cb4c883a051b6432c67c316edab551daee3998ab8d36f3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.eden","uri":"program://Human3R/module/src.dust3r.datasets.eden#L1-L94","kind":"module","name":"src.dust3r.datasets.eden","path":"src/dust3r/datasets/eden.py","language":"python","start_line":1,"end_line":94,"context_start_line":1,"context_end_line":94,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass EDEN_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.permutation(self.img_names)\n\n        views = []\n        i = 0\n        while len(views) < num_views:\n            # Load RGB image\n            scene, img_name = img_names[i]\n            try:\n                rgb_image = imread_cv2(\n                    osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\")\n                )\n                depthmap = np.load(\n                    osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\")\n                )\n                depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n                intrinsics = np.load(\n                    osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\")\n                )[\"intrinsics\"]\n                # camera pose is not provided, placeholder\n                camera_pose = np.eye(4)\n            except:\n                i += 1\n                continue\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"EDEN\",\n                    label=img_name,\n                    instance=osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"),\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n            i += 1\n        return views","source_hash":"67146a2ea020d46842392a85258e5848bc14f0821b47a7b6b5ba083db06cb4ff","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.eden.EDEN_Multi","uri":"program://Human3R/class/src.dust3r.datasets.eden.EDEN_Multi#L14-L94","kind":"class","name":"EDEN_Multi","path":"src/dust3r/datasets/eden.py","language":"python","start_line":14,"end_line":94,"context_start_line":1,"context_end_line":94,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass EDEN_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.permutation(self.img_names)\n\n        views = []\n        i = 0\n        while len(views) < num_views:\n            # Load RGB image\n            scene, img_name = img_names[i]\n            try:\n                rgb_image = imread_cv2(\n                    osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\")\n                )\n                depthmap = np.load(\n                    osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\")\n                )\n                depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n                intrinsics = np.load(\n                    osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\")\n                )[\"intrinsics\"]\n                # camera pose is not provided, placeholder\n                camera_pose = np.eye(4)\n            except:\n                i += 1\n                continue\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"EDEN\",\n                    label=img_name,\n                    instance=osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"),\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n            i += 1\n        return views","source_hash":"67146a2ea020d46842392a85258e5848bc14f0821b47a7b6b5ba083db06cb4ff","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.eden.__init__","uri":"program://Human3R/function/src.dust3r.datasets.eden.__init__#L15-L20","kind":"function","name":"__init__","path":"src/dust3r/datasets/eden.py","language":"python","start_line":15,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass EDEN_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n","source_hash":"67146a2ea020d46842392a85258e5848bc14f0821b47a7b6b5ba083db06cb4ff","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.eden._load_data","uri":"program://Human3R/function/src.dust3r.datasets.eden._load_data#L22-L33","kind":"function","name":"_load_data","path":"src/dust3r/datasets/eden.py","language":"python","start_line":22,"end_line":33,"context_start_line":2,"context_end_line":53,"code":"import cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass EDEN_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.permutation(self.img_names)\n\n        views = []\n        i = 0\n        while len(views) < num_views:\n            # Load RGB image\n            scene, img_name = img_names[i]\n            try:\n                rgb_image = imread_cv2(\n                    osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\")","source_hash":"67146a2ea020d46842392a85258e5848bc14f0821b47a7b6b5ba083db06cb4ff","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.eden.__len__","uri":"program://Human3R/function/src.dust3r.datasets.eden.__len__#L35-L36","kind":"function","name":"__len__","path":"src/dust3r/datasets/eden.py","language":"python","start_line":35,"end_line":36,"context_start_line":15,"context_end_line":56,"code":"    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.permutation(self.img_names)\n\n        views = []\n        i = 0\n        while len(views) < num_views:\n            # Load RGB image\n            scene, img_name = img_names[i]\n            try:\n                rgb_image = imread_cv2(\n                    osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\")\n                )\n                depthmap = np.load(\n                    osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\")","source_hash":"67146a2ea020d46842392a85258e5848bc14f0821b47a7b6b5ba083db06cb4ff","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.eden.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.eden.get_image_num#L38-L39","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/eden.py","language":"python","start_line":38,"end_line":39,"context_start_line":18,"context_end_line":59,"code":"        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.permutation(self.img_names)\n\n        views = []\n        i = 0\n        while len(views) < num_views:\n            # Load RGB image\n            scene, img_name = img_names[i]\n            try:\n                rgb_image = imread_cv2(\n                    osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\")\n                )\n                depthmap = np.load(\n                    osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\")\n                )\n                depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n","source_hash":"67146a2ea020d46842392a85258e5848bc14f0821b47a7b6b5ba083db06cb4ff","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.eden._get_views","uri":"program://Human3R/function/src.dust3r.datasets.eden._get_views#L41-L94","kind":"function","name":"_get_views","path":"src/dust3r/datasets/eden.py","language":"python","start_line":41,"end_line":94,"context_start_line":21,"context_end_line":94,"code":"\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.permutation(self.img_names)\n\n        views = []\n        i = 0\n        while len(views) < num_views:\n            # Load RGB image\n            scene, img_name = img_names[i]\n            try:\n                rgb_image = imread_cv2(\n                    osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\")\n                )\n                depthmap = np.load(\n                    osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\")\n                )\n                depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n                intrinsics = np.load(\n                    osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\")\n                )[\"intrinsics\"]\n                # camera pose is not provided, placeholder\n                camera_pose = np.eye(4)\n            except:\n                i += 1\n                continue\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"EDEN\",\n                    label=img_name,\n                    instance=osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"),\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n            i += 1\n        return views","source_hash":"67146a2ea020d46842392a85258e5848bc14f0821b47a7b6b5ba083db06cb4ff","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes","uri":"program://Human3R/module/src.dust3r.datasets.arkitscenes#L1-L247","kind":"module","name":"src.dust3r.datasets.arkitscenes","path":"src/dust3r/datasets/arkitscenes.py","language":"python","start_line":1,"end_line":247,"context_start_line":1,"context_end_line":247,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\ndef stratified_sampling(indices, num_samples, rng=None):\n    if num_samples > len(indices):\n        raise ValueError(\"num_samples cannot exceed the number of available indices.\")\n    elif num_samples == len(indices):\n        return indices\n\n    sorted_indices = sorted(indices)\n    stride = len(sorted_indices) / num_samples\n    sampled_indices = []\n    if rng is None:\n        rng = np.random.default_rng()\n\n    for i in range(num_samples):\n        start = int(i * stride)\n        end = int((i + 1) * stride)\n        # Ensure end does not exceed the list\n        end = min(end, len(sorted_indices))\n        if start < end:\n            # Randomly select within the current stratum\n            rand_idx = rng.integers(start, end)\n            sampled_indices.append(sorted_indices[rand_idx])\n        else:\n            # In case of any rounding issues, select the last index\n            sampled_indices.append(sorted_indices[-1])\n\n    return rng.permutation(sampled_indices)\n\n\nclass ARKitScenes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 8\n        super().__init__(*args, **kwargs)\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Test\"\n        else:\n            raise ValueError(\"\")\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        with np.load(osp.join(self.ROOT, split, \"all_metadata.npz\")) as data:\n            self.scenes: np.ndarray = data[\"scenes\"]\n           # check if highres data exists\n            highres_path = os.path.join(\n                self.ROOT.rstrip(\"/\") + \"_highres\",\n                split if split == \"Training\" else \"Validation\"\n            )\n            \n            if os.path.exists(highres_path):\n                # if highres data exists, use the highres data (cut3r logic)\n                high_res_list = np.array(\n                    [\n                        d\n                        for d in os.listdir(highres_path)\n                        if os.path.join(self.ROOT + \"_highres\", split, d)\n                    ]\n                )\n                self.scenes = np.setdiff1d(self.scenes, high_res_list)\n            # if highres data does not exist, use the normal data list\n\n        offset = 0\n        counts = []\n        scenes = []\n        sceneids = []\n        images = []\n        intrinsics = []\n        trajectories = []\n        groups = []\n        id_ranges = []\n        j = 0\n        for scene_idx, scene in enumerate(self.scenes):\n            scene_dir = osp.join(self.ROOT, self.split, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                imgs = data[\"images\"]\n                intrins = data[\"intrinsics\"]\n                traj = data[\"trajectories\"]\n                min_seq_len = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                if len(imgs) < min_seq_len:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                collections = {}\n                assert \"image_collection\" in data, \"Image collection not found\"\n                collections[\"image\"] = data[\"image_collection\"]\n\n                num_imgs = imgs.shape[0]\n                img_groups = []\n                min_group_len = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                for ref_id, group in collections[\"image\"].item().items():\n                    if len(group) + 1 < min_group_len:\n                        continue\n\n                    # groups are (idx, score)s\n                    group.insert(0, (ref_id, 1.0))\n                    group = [int(x[0] + offset) for x in group]\n                    img_groups.append(sorted(group))\n\n                if len(img_groups) == 0:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                scenes.append(scene)\n                sceneids.extend([j] * num_imgs)\n                id_ranges.extend([(offset, offset + num_imgs) for _ in range(num_imgs)])\n                images.extend(imgs)\n                K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0)\n\n                K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins]\n                K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins]\n                K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins]\n                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]\n                intrinsics.extend(list(K))\n                trajectories.extend(list(traj))\n\n                # offset groups\n                groups.extend(img_groups)\n                counts.append(offset)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.id_ranges = id_ranges\n        self.images = images\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n\n        if rng.choice([True, False]):\n            image_idxs = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_image_idxs = image_idxs[: len(image_idxs) - cut_off + 1]\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            image_idxs = np.array(image_idxs)[pos]\n        else:\n            ordered_video = False\n            image_idxs = self.groups[idx]\n            image_idxs = rng.permutation(image_idxs)\n            if len(image_idxs) > num_views:\n                image_idxs = image_idxs[:num_views]\n            else:\n                if rng.random() < 0.8:\n                    image_idxs = rng.choice(image_idxs, size=num_views, replace=True)\n                else:\n                    repeat_num = num_views // len(image_idxs) + 1\n                    image_idxs = np.tile(image_idxs, repeat_num)[:num_views]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n\n            intrinsics = self.intrinsics[view_idx]\n            camera_pose = self.trajectories[view_idx]\n            basename = self.images[view_idx]\n            assert (\n                basename[:8] == self.scenes[scene_id]\n            ), f\"{basename}, {self.scenes[scene_id]}\"\n            # print(scene_dir, basename)\n            # Load RGB image\n            rgb_image = imread_cv2(\n                osp.join(scene_dir, \"vga_wide\", basename.replace(\".png\", \".jpg\"))\n            )\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(scene_dir, \"lowres_depth\", basename), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000.0\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"arkitscenes\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"190b16bd1d3f8f27e36812fdbecfba6b979028b589cc3a5dab5afe430f181efd","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes.stratified_sampling","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes.stratified_sampling#L14-L39","kind":"function","name":"stratified_sampling","path":"src/dust3r/datasets/arkitscenes.py","language":"python","start_line":14,"end_line":39,"context_start_line":1,"context_end_line":59,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\ndef stratified_sampling(indices, num_samples, rng=None):\n    if num_samples > len(indices):\n        raise ValueError(\"num_samples cannot exceed the number of available indices.\")\n    elif num_samples == len(indices):\n        return indices\n\n    sorted_indices = sorted(indices)\n    stride = len(sorted_indices) / num_samples\n    sampled_indices = []\n    if rng is None:\n        rng = np.random.default_rng()\n\n    for i in range(num_samples):\n        start = int(i * stride)\n        end = int((i + 1) * stride)\n        # Ensure end does not exceed the list\n        end = min(end, len(sorted_indices))\n        if start < end:\n            # Randomly select within the current stratum\n            rand_idx = rng.integers(start, end)\n            sampled_indices.append(sorted_indices[rand_idx])\n        else:\n            # In case of any rounding issues, select the last index\n            sampled_indices.append(sorted_indices[-1])\n\n    return rng.permutation(sampled_indices)\n\n\nclass ARKitScenes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 8\n        super().__init__(*args, **kwargs)\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Test\"\n        else:\n            raise ValueError(\"\")\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        with np.load(osp.join(self.ROOT, split, \"all_metadata.npz\")) as data:","source_hash":"190b16bd1d3f8f27e36812fdbecfba6b979028b589cc3a5dab5afe430f181efd","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes.ARKitScenes_Multi","uri":"program://Human3R/class/src.dust3r.datasets.arkitscenes.ARKitScenes_Multi#L42-L247","kind":"class","name":"ARKitScenes_Multi","path":"src/dust3r/datasets/arkitscenes.py","language":"python","start_line":42,"end_line":247,"context_start_line":22,"context_end_line":247,"code":"    sampled_indices = []\n    if rng is None:\n        rng = np.random.default_rng()\n\n    for i in range(num_samples):\n        start = int(i * stride)\n        end = int((i + 1) * stride)\n        # Ensure end does not exceed the list\n        end = min(end, len(sorted_indices))\n        if start < end:\n            # Randomly select within the current stratum\n            rand_idx = rng.integers(start, end)\n            sampled_indices.append(sorted_indices[rand_idx])\n        else:\n            # In case of any rounding issues, select the last index\n            sampled_indices.append(sorted_indices[-1])\n\n    return rng.permutation(sampled_indices)\n\n\nclass ARKitScenes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 8\n        super().__init__(*args, **kwargs)\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Test\"\n        else:\n            raise ValueError(\"\")\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        with np.load(osp.join(self.ROOT, split, \"all_metadata.npz\")) as data:\n            self.scenes: np.ndarray = data[\"scenes\"]\n           # check if highres data exists\n            highres_path = os.path.join(\n                self.ROOT.rstrip(\"/\") + \"_highres\",\n                split if split == \"Training\" else \"Validation\"\n            )\n            \n            if os.path.exists(highres_path):\n                # if highres data exists, use the highres data (cut3r logic)\n                high_res_list = np.array(\n                    [\n                        d\n                        for d in os.listdir(highres_path)\n                        if os.path.join(self.ROOT + \"_highres\", split, d)\n                    ]\n                )\n                self.scenes = np.setdiff1d(self.scenes, high_res_list)\n            # if highres data does not exist, use the normal data list\n\n        offset = 0\n        counts = []\n        scenes = []\n        sceneids = []\n        images = []\n        intrinsics = []\n        trajectories = []\n        groups = []\n        id_ranges = []\n        j = 0\n        for scene_idx, scene in enumerate(self.scenes):\n            scene_dir = osp.join(self.ROOT, self.split, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                imgs = data[\"images\"]\n                intrins = data[\"intrinsics\"]\n                traj = data[\"trajectories\"]\n                min_seq_len = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                if len(imgs) < min_seq_len:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                collections = {}\n                assert \"image_collection\" in data, \"Image collection not found\"\n                collections[\"image\"] = data[\"image_collection\"]\n\n                num_imgs = imgs.shape[0]\n                img_groups = []\n                min_group_len = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                for ref_id, group in collections[\"image\"].item().items():\n                    if len(group) + 1 < min_group_len:\n                        continue\n\n                    # groups are (idx, score)s\n                    group.insert(0, (ref_id, 1.0))\n                    group = [int(x[0] + offset) for x in group]\n                    img_groups.append(sorted(group))\n\n                if len(img_groups) == 0:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                scenes.append(scene)\n                sceneids.extend([j] * num_imgs)\n                id_ranges.extend([(offset, offset + num_imgs) for _ in range(num_imgs)])\n                images.extend(imgs)\n                K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0)\n\n                K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins]\n                K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins]\n                K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins]\n                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]\n                intrinsics.extend(list(K))\n                trajectories.extend(list(traj))\n\n                # offset groups\n                groups.extend(img_groups)\n                counts.append(offset)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.id_ranges = id_ranges\n        self.images = images\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n\n        if rng.choice([True, False]):\n            image_idxs = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_image_idxs = image_idxs[: len(image_idxs) - cut_off + 1]\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            image_idxs = np.array(image_idxs)[pos]\n        else:\n            ordered_video = False\n            image_idxs = self.groups[idx]\n            image_idxs = rng.permutation(image_idxs)\n            if len(image_idxs) > num_views:\n                image_idxs = image_idxs[:num_views]\n            else:\n                if rng.random() < 0.8:\n                    image_idxs = rng.choice(image_idxs, size=num_views, replace=True)\n                else:\n                    repeat_num = num_views // len(image_idxs) + 1\n                    image_idxs = np.tile(image_idxs, repeat_num)[:num_views]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n\n            intrinsics = self.intrinsics[view_idx]\n            camera_pose = self.trajectories[view_idx]\n            basename = self.images[view_idx]\n            assert (\n                basename[:8] == self.scenes[scene_id]\n            ), f\"{basename}, {self.scenes[scene_id]}\"\n            # print(scene_dir, basename)\n            # Load RGB image\n            rgb_image = imread_cv2(\n                osp.join(scene_dir, \"vga_wide\", basename.replace(\".png\", \".jpg\"))\n            )\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(scene_dir, \"lowres_depth\", basename), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000.0\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"arkitscenes\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"190b16bd1d3f8f27e36812fdbecfba6b979028b589cc3a5dab5afe430f181efd","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes.__init__","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes.__init__#L43-L56","kind":"function","name":"__init__","path":"src/dust3r/datasets/arkitscenes.py","language":"python","start_line":43,"end_line":56,"context_start_line":23,"context_end_line":76,"code":"    if rng is None:\n        rng = np.random.default_rng()\n\n    for i in range(num_samples):\n        start = int(i * stride)\n        end = int((i + 1) * stride)\n        # Ensure end does not exceed the list\n        end = min(end, len(sorted_indices))\n        if start < end:\n            # Randomly select within the current stratum\n            rand_idx = rng.integers(start, end)\n            sampled_indices.append(sorted_indices[rand_idx])\n        else:\n            # In case of any rounding issues, select the last index\n            sampled_indices.append(sorted_indices[-1])\n\n    return rng.permutation(sampled_indices)\n\n\nclass ARKitScenes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 8\n        super().__init__(*args, **kwargs)\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Test\"\n        else:\n            raise ValueError(\"\")\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        with np.load(osp.join(self.ROOT, split, \"all_metadata.npz\")) as data:\n            self.scenes: np.ndarray = data[\"scenes\"]\n           # check if highres data exists\n            highres_path = os.path.join(\n                self.ROOT.rstrip(\"/\") + \"_highres\",\n                split if split == \"Training\" else \"Validation\"\n            )\n            \n            if os.path.exists(highres_path):\n                # if highres data exists, use the highres data (cut3r logic)\n                high_res_list = np.array(\n                    [\n                        d\n                        for d in os.listdir(highres_path)\n                        if os.path.join(self.ROOT + \"_highres\", split, d)\n                    ]\n                )\n                self.scenes = np.setdiff1d(self.scenes, high_res_list)","source_hash":"190b16bd1d3f8f27e36812fdbecfba6b979028b589cc3a5dab5afe430f181efd","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes._load_data","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes._load_data#L58-L155","kind":"function","name":"_load_data","path":"src/dust3r/datasets/arkitscenes.py","language":"python","start_line":58,"end_line":155,"context_start_line":38,"context_end_line":175,"code":"\n    return rng.permutation(sampled_indices)\n\n\nclass ARKitScenes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 8\n        super().__init__(*args, **kwargs)\n        if split == \"train\":\n            self.split = \"Training\"\n        elif split == \"test\":\n            self.split = \"Test\"\n        else:\n            raise ValueError(\"\")\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        with np.load(osp.join(self.ROOT, split, \"all_metadata.npz\")) as data:\n            self.scenes: np.ndarray = data[\"scenes\"]\n           # check if highres data exists\n            highres_path = os.path.join(\n                self.ROOT.rstrip(\"/\") + \"_highres\",\n                split if split == \"Training\" else \"Validation\"\n            )\n            \n            if os.path.exists(highres_path):\n                # if highres data exists, use the highres data (cut3r logic)\n                high_res_list = np.array(\n                    [\n                        d\n                        for d in os.listdir(highres_path)\n                        if os.path.join(self.ROOT + \"_highres\", split, d)\n                    ]\n                )\n                self.scenes = np.setdiff1d(self.scenes, high_res_list)\n            # if highres data does not exist, use the normal data list\n\n        offset = 0\n        counts = []\n        scenes = []\n        sceneids = []\n        images = []\n        intrinsics = []\n        trajectories = []\n        groups = []\n        id_ranges = []\n        j = 0\n        for scene_idx, scene in enumerate(self.scenes):\n            scene_dir = osp.join(self.ROOT, self.split, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                imgs = data[\"images\"]\n                intrins = data[\"intrinsics\"]\n                traj = data[\"trajectories\"]\n                min_seq_len = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                if len(imgs) < min_seq_len:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                collections = {}\n                assert \"image_collection\" in data, \"Image collection not found\"\n                collections[\"image\"] = data[\"image_collection\"]\n\n                num_imgs = imgs.shape[0]\n                img_groups = []\n                min_group_len = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                for ref_id, group in collections[\"image\"].item().items():\n                    if len(group) + 1 < min_group_len:\n                        continue\n\n                    # groups are (idx, score)s\n                    group.insert(0, (ref_id, 1.0))\n                    group = [int(x[0] + offset) for x in group]\n                    img_groups.append(sorted(group))\n\n                if len(img_groups) == 0:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                scenes.append(scene)\n                sceneids.extend([j] * num_imgs)\n                id_ranges.extend([(offset, offset + num_imgs) for _ in range(num_imgs)])\n                images.extend(imgs)\n                K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0)\n\n                K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins]\n                K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins]\n                K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins]\n                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]\n                intrinsics.extend(list(K))\n                trajectories.extend(list(traj))\n\n                # offset groups\n                groups.extend(img_groups)\n                counts.append(offset)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.id_ranges = id_ranges\n        self.images = images\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n\n        if rng.choice([True, False]):\n            image_idxs = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_image_idxs = image_idxs[: len(image_idxs) - cut_off + 1]\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs.tolist(),\n                rng,\n                max_interval=self.max_interval,","source_hash":"190b16bd1d3f8f27e36812fdbecfba6b979028b589cc3a5dab5afe430f181efd","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes.__len__","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes.__len__#L157-L158","kind":"function","name":"__len__","path":"src/dust3r/datasets/arkitscenes.py","language":"python","start_line":157,"end_line":158,"context_start_line":137,"context_end_line":178,"code":"                K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins]\n                K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins]\n                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]\n                intrinsics.extend(list(K))\n                trajectories.extend(list(traj))\n\n                # offset groups\n                groups.extend(img_groups)\n                counts.append(offset)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.id_ranges = id_ranges\n        self.images = images\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n\n        if rng.choice([True, False]):\n            image_idxs = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_image_idxs = image_idxs[: len(image_idxs) - cut_off + 1]\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,","source_hash":"190b16bd1d3f8f27e36812fdbecfba6b979028b589cc3a5dab5afe430f181efd","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes.get_image_num#L160-L161","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/arkitscenes.py","language":"python","start_line":160,"end_line":161,"context_start_line":140,"context_end_line":181,"code":"                intrinsics.extend(list(K))\n                trajectories.extend(list(traj))\n\n                # offset groups\n                groups.extend(img_groups)\n                counts.append(offset)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.id_ranges = id_ranges\n        self.images = images\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n\n        if rng.choice([True, False]):\n            image_idxs = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_image_idxs = image_idxs[: len(image_idxs) - cut_off + 1]\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            image_idxs = np.array(image_idxs)[pos]\n        else:","source_hash":"190b16bd1d3f8f27e36812fdbecfba6b979028b589cc3a5dab5afe430f181efd","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.arkitscenes._get_views","uri":"program://Human3R/function/src.dust3r.datasets.arkitscenes._get_views#L163-L247","kind":"function","name":"_get_views","path":"src/dust3r/datasets/arkitscenes.py","language":"python","start_line":163,"end_line":247,"context_start_line":143,"context_end_line":247,"code":"                # offset groups\n                groups.extend(img_groups)\n                counts.append(offset)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.id_ranges = id_ranges\n        self.images = images\n        self.intrinsics = intrinsics\n        self.trajectories = trajectories\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n\n        if rng.choice([True, False]):\n            image_idxs = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1])\n            cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n            start_image_idxs = image_idxs[: len(image_idxs) - cut_off + 1]\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs.tolist(),\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            image_idxs = np.array(image_idxs)[pos]\n        else:\n            ordered_video = False\n            image_idxs = self.groups[idx]\n            image_idxs = rng.permutation(image_idxs)\n            if len(image_idxs) > num_views:\n                image_idxs = image_idxs[:num_views]\n            else:\n                if rng.random() < 0.8:\n                    image_idxs = rng.choice(image_idxs, size=num_views, replace=True)\n                else:\n                    repeat_num = num_views // len(image_idxs) + 1\n                    image_idxs = np.tile(image_idxs, repeat_num)[:num_views]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])\n\n            intrinsics = self.intrinsics[view_idx]\n            camera_pose = self.trajectories[view_idx]\n            basename = self.images[view_idx]\n            assert (\n                basename[:8] == self.scenes[scene_id]\n            ), f\"{basename}, {self.scenes[scene_id]}\"\n            # print(scene_dir, basename)\n            # Load RGB image\n            rgb_image = imread_cv2(\n                osp.join(scene_dir, \"vga_wide\", basename.replace(\".png\", \".jpg\"))\n            )\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(scene_dir, \"lowres_depth\", basename), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000.0\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"arkitscenes\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"190b16bd1d3f8f27e36812fdbecfba6b979028b589cc3a5dab5afe430f181efd","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannetpp","uri":"program://Human3R/module/src.dust3r.datasets.scannetpp#L1-L191","kind":"module","name":"src.dust3r.datasets.scannetpp","path":"src/dust3r/datasets/scannetpp.py","language":"python","start_line":1,"end_line":191,"context_start_line":1,"context_end_line":191,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNetpp_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 3\n        super().__init__(*args, **kwargs)\n        assert self.split == \"train\"\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        with np.load(osp.join(self.ROOT, \"all_metadata.npz\")) as data:\n            self.scenes = data[\"scenes\"]\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        intrinsics = []\n        trajectories = []\n        groups = []\n        id_ranges = []\n        j = 0\n        self.image_num = 0\n        for scene in self.scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                imgs = data[\"images\"]\n                self.image_num += len(imgs)\n                img_ids = np.arange(len(imgs)).tolist()\n                intrins = data[\"intrinsics\"]\n                traj = data[\"trajectories\"]\n                imgs_on_disk = sorted(os.listdir(osp.join(scene_dir, \"images\")))\n                imgs_on_disk = list(map(lambda x: x[:-4], imgs_on_disk))\n\n                dslr_ids = [\n                    i + offset\n                    for i in img_ids\n                    if imgs[i].startswith(\"DSC\") and imgs[i] in imgs_on_disk\n                ]\n                iphone_ids = [\n                    i + offset\n                    for i in img_ids\n                    if imgs[i].startswith(\"frame\") and imgs[i] in imgs_on_disk\n                ]\n\n                num_imgs = len(imgs)\n                assert max(dslr_ids) < min(iphone_ids)\n                assert \"image_collection\" in data\n\n                img_groups = []\n                img_id_ranges = []\n\n                for ref_id, group in data[\"image_collection\"].item().items():\n                    if len(group) + 1 < self.num_views:\n                        continue\n                    group.insert(0, (ref_id, 1.0))\n                    sorted_group = sorted(group, key=lambda x: x[1], reverse=True)\n                    group = [int(x[0] + offset) for x in sorted_group]\n                    img_groups.append(sorted(group))\n\n                    if imgs[ref_id].startswith(\"frame\"):\n                        img_id_ranges.append(dslr_ids)\n                    else:\n                        img_id_ranges.append(iphone_ids)\n\n                if len(img_groups) == 0:\n                    print(f\"Skipping {scene}\")\n                    continue\n                scenes.append(scene)\n                sceneids.extend([j] * num_imgs)\n                images.extend(imgs)\n                intrinsics.append(intrins)\n                trajectories.append(traj)\n\n                # offset groups\n                groups.extend(img_groups)\n                id_ranges.extend(img_id_ranges)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.intrinsics = np.concatenate(intrinsics, axis=0)\n        self.trajectories = np.concatenate(trajectories, axis=0)\n        self.id_ranges = id_ranges\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups) * 10\n\n    def get_image_num(self):\n        return self.image_num\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        image_idxs = self.groups[idx]\n        rand_val = rng.random()\n\n        image_idxs_video = self.id_ranges[idx]\n        cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        start_image_idxs = image_idxs_video[: len(image_idxs_video) - cut_off + 1]\n\n        if rand_val < 0.7 and len(start_image_idxs) > 0:\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs_video,\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            image_idxs = np.array(image_idxs_video)[pos]\n\n        else:\n            ordered_video = True\n            # ordered video with varying intervals\n            num_candidates = len(image_idxs)\n            max_id = min(num_candidates, int(num_views * (2 + 2 * rng.random())))\n            image_idxs = sorted(rng.permutation(image_idxs[:max_id])[:num_views])\n            if rand_val > 0.75:\n                ordered_video = False\n                image_idxs = rng.permutation(image_idxs)\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n\n            intrinsics = self.intrinsics[view_idx]\n            camera_pose = self.trajectories[view_idx]\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(scene_dir, \"images\", basename + \".jpg\"))\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(scene_dir, \"depth\", basename + \".png\"), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"ScanNet++\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"88b2400c9d9fe0e4d50364112194362c6bb1cda3decdaabc24ab12e74943c2c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannetpp.ScanNetpp_Multi","uri":"program://Human3R/class/src.dust3r.datasets.scannetpp.ScanNetpp_Multi#L14-L191","kind":"class","name":"ScanNetpp_Multi","path":"src/dust3r/datasets/scannetpp.py","language":"python","start_line":14,"end_line":191,"context_start_line":1,"context_end_line":191,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNetpp_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 3\n        super().__init__(*args, **kwargs)\n        assert self.split == \"train\"\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        with np.load(osp.join(self.ROOT, \"all_metadata.npz\")) as data:\n            self.scenes = data[\"scenes\"]\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        intrinsics = []\n        trajectories = []\n        groups = []\n        id_ranges = []\n        j = 0\n        self.image_num = 0\n        for scene in self.scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                imgs = data[\"images\"]\n                self.image_num += len(imgs)\n                img_ids = np.arange(len(imgs)).tolist()\n                intrins = data[\"intrinsics\"]\n                traj = data[\"trajectories\"]\n                imgs_on_disk = sorted(os.listdir(osp.join(scene_dir, \"images\")))\n                imgs_on_disk = list(map(lambda x: x[:-4], imgs_on_disk))\n\n                dslr_ids = [\n                    i + offset\n                    for i in img_ids\n                    if imgs[i].startswith(\"DSC\") and imgs[i] in imgs_on_disk\n                ]\n                iphone_ids = [\n                    i + offset\n                    for i in img_ids\n                    if imgs[i].startswith(\"frame\") and imgs[i] in imgs_on_disk\n                ]\n\n                num_imgs = len(imgs)\n                assert max(dslr_ids) < min(iphone_ids)\n                assert \"image_collection\" in data\n\n                img_groups = []\n                img_id_ranges = []\n\n                for ref_id, group in data[\"image_collection\"].item().items():\n                    if len(group) + 1 < self.num_views:\n                        continue\n                    group.insert(0, (ref_id, 1.0))\n                    sorted_group = sorted(group, key=lambda x: x[1], reverse=True)\n                    group = [int(x[0] + offset) for x in sorted_group]\n                    img_groups.append(sorted(group))\n\n                    if imgs[ref_id].startswith(\"frame\"):\n                        img_id_ranges.append(dslr_ids)\n                    else:\n                        img_id_ranges.append(iphone_ids)\n\n                if len(img_groups) == 0:\n                    print(f\"Skipping {scene}\")\n                    continue\n                scenes.append(scene)\n                sceneids.extend([j] * num_imgs)\n                images.extend(imgs)\n                intrinsics.append(intrins)\n                trajectories.append(traj)\n\n                # offset groups\n                groups.extend(img_groups)\n                id_ranges.extend(img_id_ranges)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.intrinsics = np.concatenate(intrinsics, axis=0)\n        self.trajectories = np.concatenate(trajectories, axis=0)\n        self.id_ranges = id_ranges\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups) * 10\n\n    def get_image_num(self):\n        return self.image_num\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        image_idxs = self.groups[idx]\n        rand_val = rng.random()\n\n        image_idxs_video = self.id_ranges[idx]\n        cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        start_image_idxs = image_idxs_video[: len(image_idxs_video) - cut_off + 1]\n\n        if rand_val < 0.7 and len(start_image_idxs) > 0:\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs_video,\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            image_idxs = np.array(image_idxs_video)[pos]\n\n        else:\n            ordered_video = True\n            # ordered video with varying intervals\n            num_candidates = len(image_idxs)\n            max_id = min(num_candidates, int(num_views * (2 + 2 * rng.random())))\n            image_idxs = sorted(rng.permutation(image_idxs[:max_id])[:num_views])\n            if rand_val > 0.75:\n                ordered_video = False\n                image_idxs = rng.permutation(image_idxs)\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n\n            intrinsics = self.intrinsics[view_idx]\n            camera_pose = self.trajectories[view_idx]\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(scene_dir, \"images\", basename + \".jpg\"))\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(scene_dir, \"depth\", basename + \".png\"), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"ScanNet++\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"88b2400c9d9fe0e4d50364112194362c6bb1cda3decdaabc24ab12e74943c2c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannetpp.__init__","uri":"program://Human3R/function/src.dust3r.datasets.scannetpp.__init__#L15-L22","kind":"function","name":"__init__","path":"src/dust3r/datasets/scannetpp.py","language":"python","start_line":15,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNetpp_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 3\n        super().__init__(*args, **kwargs)\n        assert self.split == \"train\"\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        with np.load(osp.join(self.ROOT, \"all_metadata.npz\")) as data:\n            self.scenes = data[\"scenes\"]\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        intrinsics = []\n        trajectories = []\n        groups = []\n        id_ranges = []\n        j = 0\n        self.image_num = 0\n        for scene in self.scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                imgs = data[\"images\"]","source_hash":"88b2400c9d9fe0e4d50364112194362c6bb1cda3decdaabc24ab12e74943c2c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannetpp._load_data","uri":"program://Human3R/function/src.dust3r.datasets.scannetpp._load_data#L24-L102","kind":"function","name":"_load_data","path":"src/dust3r/datasets/scannetpp.py","language":"python","start_line":24,"end_line":102,"context_start_line":4,"context_end_line":122,"code":"import itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNetpp_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 3\n        super().__init__(*args, **kwargs)\n        assert self.split == \"train\"\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        with np.load(osp.join(self.ROOT, \"all_metadata.npz\")) as data:\n            self.scenes = data[\"scenes\"]\n        offset = 0\n        scenes = []\n        sceneids = []\n        images = []\n        intrinsics = []\n        trajectories = []\n        groups = []\n        id_ranges = []\n        j = 0\n        self.image_num = 0\n        for scene in self.scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                imgs = data[\"images\"]\n                self.image_num += len(imgs)\n                img_ids = np.arange(len(imgs)).tolist()\n                intrins = data[\"intrinsics\"]\n                traj = data[\"trajectories\"]\n                imgs_on_disk = sorted(os.listdir(osp.join(scene_dir, \"images\")))\n                imgs_on_disk = list(map(lambda x: x[:-4], imgs_on_disk))\n\n                dslr_ids = [\n                    i + offset\n                    for i in img_ids\n                    if imgs[i].startswith(\"DSC\") and imgs[i] in imgs_on_disk\n                ]\n                iphone_ids = [\n                    i + offset\n                    for i in img_ids\n                    if imgs[i].startswith(\"frame\") and imgs[i] in imgs_on_disk\n                ]\n\n                num_imgs = len(imgs)\n                assert max(dslr_ids) < min(iphone_ids)\n                assert \"image_collection\" in data\n\n                img_groups = []\n                img_id_ranges = []\n\n                for ref_id, group in data[\"image_collection\"].item().items():\n                    if len(group) + 1 < self.num_views:\n                        continue\n                    group.insert(0, (ref_id, 1.0))\n                    sorted_group = sorted(group, key=lambda x: x[1], reverse=True)\n                    group = [int(x[0] + offset) for x in sorted_group]\n                    img_groups.append(sorted(group))\n\n                    if imgs[ref_id].startswith(\"frame\"):\n                        img_id_ranges.append(dslr_ids)\n                    else:\n                        img_id_ranges.append(iphone_ids)\n\n                if len(img_groups) == 0:\n                    print(f\"Skipping {scene}\")\n                    continue\n                scenes.append(scene)\n                sceneids.extend([j] * num_imgs)\n                images.extend(imgs)\n                intrinsics.append(intrins)\n                trajectories.append(traj)\n\n                # offset groups\n                groups.extend(img_groups)\n                id_ranges.extend(img_id_ranges)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.intrinsics = np.concatenate(intrinsics, axis=0)\n        self.trajectories = np.concatenate(trajectories, axis=0)\n        self.id_ranges = id_ranges\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups) * 10\n\n    def get_image_num(self):\n        return self.image_num\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        image_idxs = self.groups[idx]\n        rand_val = rng.random()\n\n        image_idxs_video = self.id_ranges[idx]\n        cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        start_image_idxs = image_idxs_video[: len(image_idxs_video) - cut_off + 1]\n\n        if rand_val < 0.7 and len(start_image_idxs) > 0:\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,","source_hash":"88b2400c9d9fe0e4d50364112194362c6bb1cda3decdaabc24ab12e74943c2c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannetpp.__len__","uri":"program://Human3R/function/src.dust3r.datasets.scannetpp.__len__#L104-L105","kind":"function","name":"__len__","path":"src/dust3r/datasets/scannetpp.py","language":"python","start_line":104,"end_line":105,"context_start_line":84,"context_end_line":125,"code":"                scenes.append(scene)\n                sceneids.extend([j] * num_imgs)\n                images.extend(imgs)\n                intrinsics.append(intrins)\n                trajectories.append(traj)\n\n                # offset groups\n                groups.extend(img_groups)\n                id_ranges.extend(img_id_ranges)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.intrinsics = np.concatenate(intrinsics, axis=0)\n        self.trajectories = np.concatenate(trajectories, axis=0)\n        self.id_ranges = id_ranges\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups) * 10\n\n    def get_image_num(self):\n        return self.image_num\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        image_idxs = self.groups[idx]\n        rand_val = rng.random()\n\n        image_idxs_video = self.id_ranges[idx]\n        cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        start_image_idxs = image_idxs_video[: len(image_idxs_video) - cut_off + 1]\n\n        if rand_val < 0.7 and len(start_image_idxs) > 0:\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs_video,\n                rng,","source_hash":"88b2400c9d9fe0e4d50364112194362c6bb1cda3decdaabc24ab12e74943c2c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannetpp.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.scannetpp.get_image_num#L107-L108","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/scannetpp.py","language":"python","start_line":107,"end_line":108,"context_start_line":87,"context_end_line":128,"code":"                intrinsics.append(intrins)\n                trajectories.append(traj)\n\n                # offset groups\n                groups.extend(img_groups)\n                id_ranges.extend(img_id_ranges)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.intrinsics = np.concatenate(intrinsics, axis=0)\n        self.trajectories = np.concatenate(trajectories, axis=0)\n        self.id_ranges = id_ranges\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups) * 10\n\n    def get_image_num(self):\n        return self.image_num\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        image_idxs = self.groups[idx]\n        rand_val = rng.random()\n\n        image_idxs_video = self.id_ranges[idx]\n        cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        start_image_idxs = image_idxs_video[: len(image_idxs_video) - cut_off + 1]\n\n        if rand_val < 0.7 and len(start_image_idxs) > 0:\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs_video,\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,","source_hash":"88b2400c9d9fe0e4d50364112194362c6bb1cda3decdaabc24ab12e74943c2c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannetpp._get_views","uri":"program://Human3R/function/src.dust3r.datasets.scannetpp._get_views#L110-L191","kind":"function","name":"_get_views","path":"src/dust3r/datasets/scannetpp.py","language":"python","start_line":110,"end_line":191,"context_start_line":90,"context_end_line":191,"code":"                # offset groups\n                groups.extend(img_groups)\n                id_ranges.extend(img_id_ranges)\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.intrinsics = np.concatenate(intrinsics, axis=0)\n        self.trajectories = np.concatenate(trajectories, axis=0)\n        self.id_ranges = id_ranges\n        self.groups = groups\n\n    def __len__(self):\n        return len(self.groups) * 10\n\n    def get_image_num(self):\n        return self.image_num\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        idx = idx // 10\n        image_idxs = self.groups[idx]\n        rand_val = rng.random()\n\n        image_idxs_video = self.id_ranges[idx]\n        cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3)\n        start_image_idxs = image_idxs_video[: len(image_idxs_video) - cut_off + 1]\n\n        if rand_val < 0.7 and len(start_image_idxs) > 0:\n            start_id = rng.choice(start_image_idxs)\n            pos, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                start_id,\n                image_idxs_video,\n                rng,\n                max_interval=self.max_interval,\n                video_prob=0.8,\n                fix_interval_prob=0.5,\n                block_shuffle=16,\n            )\n            image_idxs = np.array(image_idxs_video)[pos]\n\n        else:\n            ordered_video = True\n            # ordered video with varying intervals\n            num_candidates = len(image_idxs)\n            max_id = min(num_candidates, int(num_views * (2 + 2 * rng.random())))\n            image_idxs = sorted(rng.permutation(image_idxs[:max_id])[:num_views])\n            if rand_val > 0.75:\n                ordered_video = False\n                image_idxs = rng.permutation(image_idxs)\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n\n            intrinsics = self.intrinsics[view_idx]\n            camera_pose = self.trajectories[view_idx]\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(scene_dir, \"images\", basename + \".jpg\"))\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(scene_dir, \"depth\", basename + \".png\"), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"ScanNet++\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.99, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"88b2400c9d9fe0e4d50364112194362c6bb1cda3decdaabc24ab12e74943c2c3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.irs","uri":"program://Human3R/module/src.dust3r.datasets.irs#L1-L86","kind":"module","name":"src.dust3r.datasets.irs","path":"src/dust3r/datasets/irs.py","language":"python","start_line":1,"end_line":86,"context_start_line":1,"context_end_line":86,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass IRS(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap[depthmap > 200] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"irs\",\n                    label=img_name,\n                    instance=f\"{str(idx)}_{img_name}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"f6ffc7162a1dde1510b8500150f045f8244bdf033b105f8395f093c74818169d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.irs.IRS","uri":"program://Human3R/class/src.dust3r.datasets.irs.IRS#L14-L86","kind":"class","name":"IRS","path":"src/dust3r/datasets/irs.py","language":"python","start_line":14,"end_line":86,"context_start_line":1,"context_end_line":86,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass IRS(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap[depthmap > 200] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"irs\",\n                    label=img_name,\n                    instance=f\"{str(idx)}_{img_name}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"f6ffc7162a1dde1510b8500150f045f8244bdf033b105f8395f093c74818169d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.irs.__init__","uri":"program://Human3R/function/src.dust3r.datasets.irs.__init__#L15-L20","kind":"function","name":"__init__","path":"src/dust3r/datasets/irs.py","language":"python","start_line":15,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass IRS(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n","source_hash":"f6ffc7162a1dde1510b8500150f045f8244bdf033b105f8395f093c74818169d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.irs._load_data","uri":"program://Human3R/function/src.dust3r.datasets.irs._load_data#L22-L33","kind":"function","name":"_load_data","path":"src/dust3r/datasets/irs.py","language":"python","start_line":22,"end_line":33,"context_start_line":2,"context_end_line":53,"code":"import cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass IRS(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap[depthmap > 200] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)","source_hash":"f6ffc7162a1dde1510b8500150f045f8244bdf033b105f8395f093c74818169d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.irs.__len__","uri":"program://Human3R/function/src.dust3r.datasets.irs.__len__#L35-L36","kind":"function","name":"__len__","path":"src/dust3r/datasets/irs.py","language":"python","start_line":35,"end_line":36,"context_start_line":15,"context_end_line":56,"code":"    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap[depthmap > 200] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"","source_hash":"f6ffc7162a1dde1510b8500150f045f8244bdf033b105f8395f093c74818169d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.irs.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.irs.get_image_num#L38-L39","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/irs.py","language":"python","start_line":38,"end_line":39,"context_start_line":18,"context_end_line":59,"code":"        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap[depthmap > 200] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)","source_hash":"f6ffc7162a1dde1510b8500150f045f8244bdf033b105f8395f093c74818169d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.irs._get_views","uri":"program://Human3R/function/src.dust3r.datasets.irs._get_views#L41-L86","kind":"function","name":"_get_views","path":"src/dust3r/datasets/irs.py","language":"python","start_line":41,"end_line":86,"context_start_line":21,"context_end_line":86,"code":"\n    def _load_data(self):\n        scenes = os.listdir(self.ROOT)\n        img_names = []\n        for scene in scenes:\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")]\n            )\n            img_names.extend([(scene, basename) for basename in basenames])\n\n        self.img_names = img_names\n\n    def __len__(self):\n        return len(self.img_names)\n\n    def get_image_num(self):\n        return len(self.img_names)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        new_seed = rng.integers(0, 2**32) + idx\n        new_rng = np.random.default_rng(new_seed)\n        img_names = new_rng.choice(self.img_names, num_views, replace=False)\n\n        views = []\n        for v, img_name in enumerate(img_names):\n            # Load RGB image\n            scene, img_name = img_name\n            rgb_image = imread_cv2(osp.join(self.ROOT, scene, \"rgb\", f\"{img_name}.png\"))\n            depthmap = np.load(osp.join(self.ROOT, scene, \"depth\", f\"{img_name}.npy\"))\n            depthmap[depthmap > 200] = 0.0\n            depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0)\n\n            intrinsics = np.load(osp.join(self.ROOT, scene, \"cam\", f\"{img_name}.npz\"))[\n                \"intrinsics\"\n            ]\n            # camera pose is not provided, placeholder\n            camera_pose = np.eye(4)\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"irs\",\n                    label=img_name,\n                    instance=f\"{str(idx)}_{img_name}\",\n                    is_metric=self.is_metric,\n                    is_video=False,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=True,\n                    reset=True,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"f6ffc7162a1dde1510b8500150f045f8244bdf033b105f8395f093c74818169d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs","uri":"program://Human3R/module/src.dust3r.datasets.blendedmvs#L1-L305","kind":"module","name":"src.dust3r.datasets.blendedmvs","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":1,"end_line":305,"context_start_line":1,"context_end_line":305,"code":"import os.path as osp\nimport numpy as np\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport h5py\nfrom tqdm import tqdm\n\n\nclass BlendedMVS_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, split=None, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        # assert split is None\n        self._load_data()\n\n    def _load_data(self):\n        self.data_dict = self.read_h5_file(os.path.join(self.ROOT, \"new_overlap.h5\"))\n        self.num_imgs = sum(\n            [len(self.data_dict[s][\"basenames\"]) for s in self.data_dict.keys()]\n        )\n        self.num_scenes = len(self.data_dict.keys())\n        self.invalid_scenes = []\n        self.is_reachable_cache = {scene: {} for scene in self.data_dict.keys()}\n\n    def read_h5_file(self, h5_file_path):\n        data_dict = {}\n        self.all_ref_imgs = []\n        with h5py.File(h5_file_path, \"r\") as f:\n            for scene_dir in tqdm(f.keys()):\n                group = f[scene_dir]\n                basenames = group[\"basenames\"][:]\n                indices = group[\"indices\"][:]\n                values = group[\"values\"][:]\n                shape = group.attrs[\"shape\"]\n                # Reconstruct the sparse matrix\n                score_matrix = np.zeros(shape, dtype=np.float32)\n                score_matrix[indices[0], indices[1]] = values\n                data_dict[scene_dir] = {\n                    \"basenames\": basenames,\n                    \"score_matrix\": self.build_adjacency_list(score_matrix),\n                }\n                self.all_ref_imgs.extend(\n                    [(scene_dir, b) for b in range(len(basenames))]\n                )\n        return data_dict\n\n    @staticmethod\n    def build_adjacency_list(S, thresh=0.2):\n        adjacency_list = [[] for _ in range(len(S))]\n        S = S - thresh\n        S[S < 0] = 0\n        rows, cols = np.nonzero(S)\n        for i, j in zip(rows, cols):\n            adjacency_list[i].append((j, S[i][j]))\n        return adjacency_list\n\n    @staticmethod\n    def is_reachable(adjacency_list, start_index, k):\n        visited = set()\n        stack = [start_index]\n        while stack and len(visited) < k:\n            node = stack.pop()\n            if node not in visited:\n                visited.add(node)\n                for neighbor in adjacency_list[node]:\n                    if neighbor[0] not in visited:\n                        stack.append(neighbor[0])\n        return len(visited) >= k\n\n    @staticmethod\n    def random_sequence_no_revisit_with_backtracking(\n        adjacency_list, k, start_index, rng: np.random.Generator\n    ):\n        path = [start_index]\n        visited = set([start_index])\n\n        neighbor_iterators = []\n        # Initialize the iterator for the start index\n        neighbors = adjacency_list[start_index]\n        neighbor_idxs = [n[0] for n in neighbors]\n        neighbor_weights = [n[1] for n in neighbors]\n        neighbor_idxs = rng.choice(\n            neighbor_idxs,\n            size=len(neighbor_idxs),\n            replace=False,\n            p=np.array(neighbor_weights) / np.sum(neighbor_weights),\n        ).tolist()\n        neighbor_iterators.append(iter(neighbor_idxs))\n\n        while len(path) < k:\n            if not neighbor_iterators:\n                # No possible sequence\n                return None\n            current_iterator = neighbor_iterators[-1]\n            try:\n                next_index = next(current_iterator)\n                if next_index not in visited:\n                    path.append(next_index)\n                    visited.add(next_index)\n\n                    # Prepare iterator for the next node\n                    neighbors = adjacency_list[next_index]\n                    neighbor_idxs = [n[0] for n in neighbors]\n                    neighbor_weights = [n[1] for n in neighbors]\n                    neighbor_idxs = rng.choice(\n                        neighbor_idxs,\n                        size=len(neighbor_idxs),\n                        replace=False,\n                        p=np.array(neighbor_weights) / np.sum(neighbor_weights),\n                    ).tolist()\n                    neighbor_iterators.append(iter(neighbor_idxs))\n            except StopIteration:\n                # No more neighbors to try at this node, backtrack\n                neighbor_iterators.pop()\n                visited.remove(path.pop())\n        return path\n\n    @staticmethod\n    def random_sequence_with_optional_repeats(\n        adjacency_list,\n        k,\n        start_index,\n        rng: np.random.Generator,\n        max_k=None,\n        max_attempts=100,\n    ):\n        if max_k is None:\n            max_k = k\n        path = [start_index]\n        visited = set([start_index])\n        current_index = start_index\n        attempts = 0\n\n        while len(path) < max_k and attempts < max_attempts:\n            attempts += 1\n            neighbors = adjacency_list[current_index]\n            neighbor_idxs = [n[0] for n in neighbors]\n            neighbor_weights = [n[1] for n in neighbors]\n\n            if not neighbor_idxs:\n                # No neighbors, cannot proceed further\n                break\n\n            # Try to find unvisited neighbors\n            unvisited_neighbors = [\n                (idx, wgt)\n                for idx, wgt in zip(neighbor_idxs, neighbor_weights)\n                if idx not in visited\n            ]\n            if unvisited_neighbors:\n                # Select among unvisited neighbors\n                unvisited_idxs = [idx for idx, _ in unvisited_neighbors]\n                unvisited_weights = [wgt for _, wgt in unvisited_neighbors]\n                probabilities = np.array(unvisited_weights) / np.sum(unvisited_weights)\n                next_index = rng.choice(unvisited_idxs, p=probabilities)\n                visited.add(next_index)\n            else:\n                # All neighbors visited, but we need to reach length max_k\n                # So we can revisit nodes\n                probabilities = np.array(neighbor_weights) / np.sum(neighbor_weights)\n                next_index = rng.choice(neighbor_idxs, p=probabilities)\n\n            path.append(next_index)\n            current_index = next_index\n\n        if len(set(path)) >= k:\n            # If path is shorter than max_k, extend it by repeating existing elements\n            while len(path) < max_k:\n                # Randomly select nodes from the existing path to repeat\n                next_index = rng.choice(path)\n                path.append(next_index)\n            return path\n        else:\n            # Could not reach k unique nodes\n            return None\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def get_image_num(self):\n        return self.num_imgs\n\n    def get_stats(self):\n        return f\"{len(self)} imgs from {self.num_scenes} scenes\"\n\n    def generate_sequence(\n        self, scene, adj_list, num_views, start_index, rng, allow_repeat=False\n    ):\n        cutoff = num_views if not allow_repeat else max(num_views // 5, 3)\n        if start_index in self.is_reachable_cache[scene]:\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        else:\n            self.is_reachable_cache[scene][start_index] = self.is_reachable(\n                adj_list, start_index, cutoff\n            )\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        if not allow_repeat:\n            sequence = self.random_sequence_no_revisit_with_backtracking(\n                adj_list, cutoff, start_index, rng\n            )\n        else:\n            sequence = self.random_sequence_with_optional_repeats(\n                adj_list, cutoff, start_index, rng, max_k=num_views\n            )\n        if not sequence:\n            self.is_reachable_cache[scene][start_index] = False\n            print(\"Failed to generate a sequence without revisiting.\")\n        return sequence\n\n    def _get_views(self, idx, resolution, rng: np.random.Generator, num_views):\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n        invalid_seq = True\n        ordered_video = False\n\n        while invalid_seq:\n            basenames = self.data_dict[scene_info][\"basenames\"]\n            if (\n                sum(\n                    [\n                        (1 - int(x))\n                        for x in list(self.is_reachable_cache[scene_info].values())\n                    ]\n                )\n                > len(basenames) - self.num_views\n            ):\n                self.invalid_scenes.append(scene_info)\n            while scene_info in self.invalid_scenes:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n                basenames = self.data_dict[scene_info][\"basenames\"]\n\n            score_matrix = self.data_dict[scene_info][\"score_matrix\"]\n            imgs_idxs = self.generate_sequence(\n                scene_info, score_matrix, num_views, ref_img_idx, rng, self.allow_repeat\n            )\n\n            if imgs_idxs is None:\n                random_direction = 2 * rng.choice(2) - 1\n                for offset in range(1, len(basenames)):\n                    tentative_im_idx = (\n                        ref_img_idx + (random_direction * offset)\n                    ) % len(basenames)\n                    if (\n                        tentative_im_idx not in self.is_reachable_cache[scene_info]\n                        or self.is_reachable_cache[scene_info][tentative_im_idx]\n                    ):\n                        ref_img_idx = tentative_im_idx\n                        break\n            else:\n                invalid_seq = False\n        views = []\n        for view_idx in imgs_idxs:\n            scene_dir = osp.join(self.ROOT, scene_info)\n            impath = basenames[view_idx].decode(\"utf-8\")\n            image = imread_cv2(osp.join(scene_dir, impath + \".jpg\"))\n            depthmap = imread_cv2(osp.join(scene_dir, impath + \".exr\"))\n            camera_params = np.load(osp.join(scene_dir, impath + \".npz\"))\n\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.eye(4, dtype=np.float32)\n            camera_pose[:3, :3] = camera_params[\"R_cam2world\"]\n            camera_pose[:3, 3] = camera_params[\"t_cam2world\"]\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, impath)\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"BlendedMVS\",\n                    label=osp.relpath(scene_dir, self.ROOT),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    instance=osp.join(scene_dir, impath + \".jpg\"),\n                    quantile=np.array(0.97, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.BlendedMVS_Multi","uri":"program://Human3R/class/src.dust3r.datasets.blendedmvs.BlendedMVS_Multi#L13-L305","kind":"class","name":"BlendedMVS_Multi","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":13,"end_line":305,"context_start_line":1,"context_end_line":305,"code":"import os.path as osp\nimport numpy as np\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport h5py\nfrom tqdm import tqdm\n\n\nclass BlendedMVS_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, split=None, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        # assert split is None\n        self._load_data()\n\n    def _load_data(self):\n        self.data_dict = self.read_h5_file(os.path.join(self.ROOT, \"new_overlap.h5\"))\n        self.num_imgs = sum(\n            [len(self.data_dict[s][\"basenames\"]) for s in self.data_dict.keys()]\n        )\n        self.num_scenes = len(self.data_dict.keys())\n        self.invalid_scenes = []\n        self.is_reachable_cache = {scene: {} for scene in self.data_dict.keys()}\n\n    def read_h5_file(self, h5_file_path):\n        data_dict = {}\n        self.all_ref_imgs = []\n        with h5py.File(h5_file_path, \"r\") as f:\n            for scene_dir in tqdm(f.keys()):\n                group = f[scene_dir]\n                basenames = group[\"basenames\"][:]\n                indices = group[\"indices\"][:]\n                values = group[\"values\"][:]\n                shape = group.attrs[\"shape\"]\n                # Reconstruct the sparse matrix\n                score_matrix = np.zeros(shape, dtype=np.float32)\n                score_matrix[indices[0], indices[1]] = values\n                data_dict[scene_dir] = {\n                    \"basenames\": basenames,\n                    \"score_matrix\": self.build_adjacency_list(score_matrix),\n                }\n                self.all_ref_imgs.extend(\n                    [(scene_dir, b) for b in range(len(basenames))]\n                )\n        return data_dict\n\n    @staticmethod\n    def build_adjacency_list(S, thresh=0.2):\n        adjacency_list = [[] for _ in range(len(S))]\n        S = S - thresh\n        S[S < 0] = 0\n        rows, cols = np.nonzero(S)\n        for i, j in zip(rows, cols):\n            adjacency_list[i].append((j, S[i][j]))\n        return adjacency_list\n\n    @staticmethod\n    def is_reachable(adjacency_list, start_index, k):\n        visited = set()\n        stack = [start_index]\n        while stack and len(visited) < k:\n            node = stack.pop()\n            if node not in visited:\n                visited.add(node)\n                for neighbor in adjacency_list[node]:\n                    if neighbor[0] not in visited:\n                        stack.append(neighbor[0])\n        return len(visited) >= k\n\n    @staticmethod\n    def random_sequence_no_revisit_with_backtracking(\n        adjacency_list, k, start_index, rng: np.random.Generator\n    ):\n        path = [start_index]\n        visited = set([start_index])\n\n        neighbor_iterators = []\n        # Initialize the iterator for the start index\n        neighbors = adjacency_list[start_index]\n        neighbor_idxs = [n[0] for n in neighbors]\n        neighbor_weights = [n[1] for n in neighbors]\n        neighbor_idxs = rng.choice(\n            neighbor_idxs,\n            size=len(neighbor_idxs),\n            replace=False,\n            p=np.array(neighbor_weights) / np.sum(neighbor_weights),\n        ).tolist()\n        neighbor_iterators.append(iter(neighbor_idxs))\n\n        while len(path) < k:\n            if not neighbor_iterators:\n                # No possible sequence\n                return None\n            current_iterator = neighbor_iterators[-1]\n            try:\n                next_index = next(current_iterator)\n                if next_index not in visited:\n                    path.append(next_index)\n                    visited.add(next_index)\n\n                    # Prepare iterator for the next node\n                    neighbors = adjacency_list[next_index]\n                    neighbor_idxs = [n[0] for n in neighbors]\n                    neighbor_weights = [n[1] for n in neighbors]\n                    neighbor_idxs = rng.choice(\n                        neighbor_idxs,\n                        size=len(neighbor_idxs),\n                        replace=False,\n                        p=np.array(neighbor_weights) / np.sum(neighbor_weights),\n                    ).tolist()\n                    neighbor_iterators.append(iter(neighbor_idxs))\n            except StopIteration:\n                # No more neighbors to try at this node, backtrack\n                neighbor_iterators.pop()\n                visited.remove(path.pop())\n        return path\n\n    @staticmethod\n    def random_sequence_with_optional_repeats(\n        adjacency_list,\n        k,\n        start_index,\n        rng: np.random.Generator,\n        max_k=None,\n        max_attempts=100,\n    ):\n        if max_k is None:\n            max_k = k\n        path = [start_index]\n        visited = set([start_index])\n        current_index = start_index\n        attempts = 0\n\n        while len(path) < max_k and attempts < max_attempts:\n            attempts += 1\n            neighbors = adjacency_list[current_index]\n            neighbor_idxs = [n[0] for n in neighbors]\n            neighbor_weights = [n[1] for n in neighbors]\n\n            if not neighbor_idxs:\n                # No neighbors, cannot proceed further\n                break\n\n            # Try to find unvisited neighbors\n            unvisited_neighbors = [\n                (idx, wgt)\n                for idx, wgt in zip(neighbor_idxs, neighbor_weights)\n                if idx not in visited\n            ]\n            if unvisited_neighbors:\n                # Select among unvisited neighbors\n                unvisited_idxs = [idx for idx, _ in unvisited_neighbors]\n                unvisited_weights = [wgt for _, wgt in unvisited_neighbors]\n                probabilities = np.array(unvisited_weights) / np.sum(unvisited_weights)\n                next_index = rng.choice(unvisited_idxs, p=probabilities)\n                visited.add(next_index)\n            else:\n                # All neighbors visited, but we need to reach length max_k\n                # So we can revisit nodes\n                probabilities = np.array(neighbor_weights) / np.sum(neighbor_weights)\n                next_index = rng.choice(neighbor_idxs, p=probabilities)\n\n            path.append(next_index)\n            current_index = next_index\n\n        if len(set(path)) >= k:\n            # If path is shorter than max_k, extend it by repeating existing elements\n            while len(path) < max_k:\n                # Randomly select nodes from the existing path to repeat\n                next_index = rng.choice(path)\n                path.append(next_index)\n            return path\n        else:\n            # Could not reach k unique nodes\n            return None\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def get_image_num(self):\n        return self.num_imgs\n\n    def get_stats(self):\n        return f\"{len(self)} imgs from {self.num_scenes} scenes\"\n\n    def generate_sequence(\n        self, scene, adj_list, num_views, start_index, rng, allow_repeat=False\n    ):\n        cutoff = num_views if not allow_repeat else max(num_views // 5, 3)\n        if start_index in self.is_reachable_cache[scene]:\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        else:\n            self.is_reachable_cache[scene][start_index] = self.is_reachable(\n                adj_list, start_index, cutoff\n            )\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        if not allow_repeat:\n            sequence = self.random_sequence_no_revisit_with_backtracking(\n                adj_list, cutoff, start_index, rng\n            )\n        else:\n            sequence = self.random_sequence_with_optional_repeats(\n                adj_list, cutoff, start_index, rng, max_k=num_views\n            )\n        if not sequence:\n            self.is_reachable_cache[scene][start_index] = False\n            print(\"Failed to generate a sequence without revisiting.\")\n        return sequence\n\n    def _get_views(self, idx, resolution, rng: np.random.Generator, num_views):\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n        invalid_seq = True\n        ordered_video = False\n\n        while invalid_seq:\n            basenames = self.data_dict[scene_info][\"basenames\"]\n            if (\n                sum(\n                    [\n                        (1 - int(x))\n                        for x in list(self.is_reachable_cache[scene_info].values())\n                    ]\n                )\n                > len(basenames) - self.num_views\n            ):\n                self.invalid_scenes.append(scene_info)\n            while scene_info in self.invalid_scenes:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n                basenames = self.data_dict[scene_info][\"basenames\"]\n\n            score_matrix = self.data_dict[scene_info][\"score_matrix\"]\n            imgs_idxs = self.generate_sequence(\n                scene_info, score_matrix, num_views, ref_img_idx, rng, self.allow_repeat\n            )\n\n            if imgs_idxs is None:\n                random_direction = 2 * rng.choice(2) - 1\n                for offset in range(1, len(basenames)):\n                    tentative_im_idx = (\n                        ref_img_idx + (random_direction * offset)\n                    ) % len(basenames)\n                    if (\n                        tentative_im_idx not in self.is_reachable_cache[scene_info]\n                        or self.is_reachable_cache[scene_info][tentative_im_idx]\n                    ):\n                        ref_img_idx = tentative_im_idx\n                        break\n            else:\n                invalid_seq = False\n        views = []\n        for view_idx in imgs_idxs:\n            scene_dir = osp.join(self.ROOT, scene_info)\n            impath = basenames[view_idx].decode(\"utf-8\")\n            image = imread_cv2(osp.join(scene_dir, impath + \".jpg\"))\n            depthmap = imread_cv2(osp.join(scene_dir, impath + \".exr\"))\n            camera_params = np.load(osp.join(scene_dir, impath + \".npz\"))\n\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.eye(4, dtype=np.float32)\n            camera_pose[:3, :3] = camera_params[\"R_cam2world\"]\n            camera_pose[:3, 3] = camera_params[\"t_cam2world\"]\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, impath)\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"BlendedMVS\",\n                    label=osp.relpath(scene_dir, self.ROOT),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    instance=osp.join(scene_dir, impath + \".jpg\"),\n                    quantile=np.array(0.97, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.__init__","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.__init__#L16-L22","kind":"function","name":"__init__","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":16,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport numpy as np\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport h5py\nfrom tqdm import tqdm\n\n\nclass BlendedMVS_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, split=None, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        # assert split is None\n        self._load_data()\n\n    def _load_data(self):\n        self.data_dict = self.read_h5_file(os.path.join(self.ROOT, \"new_overlap.h5\"))\n        self.num_imgs = sum(\n            [len(self.data_dict[s][\"basenames\"]) for s in self.data_dict.keys()]\n        )\n        self.num_scenes = len(self.data_dict.keys())\n        self.invalid_scenes = []\n        self.is_reachable_cache = {scene: {} for scene in self.data_dict.keys()}\n\n    def read_h5_file(self, h5_file_path):\n        data_dict = {}\n        self.all_ref_imgs = []\n        with h5py.File(h5_file_path, \"r\") as f:\n            for scene_dir in tqdm(f.keys()):\n                group = f[scene_dir]\n                basenames = group[\"basenames\"][:]\n                indices = group[\"indices\"][:]\n                values = group[\"values\"][:]\n                shape = group.attrs[\"shape\"]","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs._load_data","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs._load_data#L24-L31","kind":"function","name":"_load_data","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":24,"end_line":31,"context_start_line":4,"context_end_line":51,"code":"import sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport h5py\nfrom tqdm import tqdm\n\n\nclass BlendedMVS_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, split=None, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        # assert split is None\n        self._load_data()\n\n    def _load_data(self):\n        self.data_dict = self.read_h5_file(os.path.join(self.ROOT, \"new_overlap.h5\"))\n        self.num_imgs = sum(\n            [len(self.data_dict[s][\"basenames\"]) for s in self.data_dict.keys()]\n        )\n        self.num_scenes = len(self.data_dict.keys())\n        self.invalid_scenes = []\n        self.is_reachable_cache = {scene: {} for scene in self.data_dict.keys()}\n\n    def read_h5_file(self, h5_file_path):\n        data_dict = {}\n        self.all_ref_imgs = []\n        with h5py.File(h5_file_path, \"r\") as f:\n            for scene_dir in tqdm(f.keys()):\n                group = f[scene_dir]\n                basenames = group[\"basenames\"][:]\n                indices = group[\"indices\"][:]\n                values = group[\"values\"][:]\n                shape = group.attrs[\"shape\"]\n                # Reconstruct the sparse matrix\n                score_matrix = np.zeros(shape, dtype=np.float32)\n                score_matrix[indices[0], indices[1]] = values\n                data_dict[scene_dir] = {\n                    \"basenames\": basenames,\n                    \"score_matrix\": self.build_adjacency_list(score_matrix),\n                }\n                self.all_ref_imgs.extend(\n                    [(scene_dir, b) for b in range(len(basenames))]","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.read_h5_file","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.read_h5_file#L33-L53","kind":"function","name":"read_h5_file","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":33,"end_line":53,"context_start_line":13,"context_end_line":73,"code":"class BlendedMVS_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, split=None, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        # assert split is None\n        self._load_data()\n\n    def _load_data(self):\n        self.data_dict = self.read_h5_file(os.path.join(self.ROOT, \"new_overlap.h5\"))\n        self.num_imgs = sum(\n            [len(self.data_dict[s][\"basenames\"]) for s in self.data_dict.keys()]\n        )\n        self.num_scenes = len(self.data_dict.keys())\n        self.invalid_scenes = []\n        self.is_reachable_cache = {scene: {} for scene in self.data_dict.keys()}\n\n    def read_h5_file(self, h5_file_path):\n        data_dict = {}\n        self.all_ref_imgs = []\n        with h5py.File(h5_file_path, \"r\") as f:\n            for scene_dir in tqdm(f.keys()):\n                group = f[scene_dir]\n                basenames = group[\"basenames\"][:]\n                indices = group[\"indices\"][:]\n                values = group[\"values\"][:]\n                shape = group.attrs[\"shape\"]\n                # Reconstruct the sparse matrix\n                score_matrix = np.zeros(shape, dtype=np.float32)\n                score_matrix[indices[0], indices[1]] = values\n                data_dict[scene_dir] = {\n                    \"basenames\": basenames,\n                    \"score_matrix\": self.build_adjacency_list(score_matrix),\n                }\n                self.all_ref_imgs.extend(\n                    [(scene_dir, b) for b in range(len(basenames))]\n                )\n        return data_dict\n\n    @staticmethod\n    def build_adjacency_list(S, thresh=0.2):\n        adjacency_list = [[] for _ in range(len(S))]\n        S = S - thresh\n        S[S < 0] = 0\n        rows, cols = np.nonzero(S)\n        for i, j in zip(rows, cols):\n            adjacency_list[i].append((j, S[i][j]))\n        return adjacency_list\n\n    @staticmethod\n    def is_reachable(adjacency_list, start_index, k):\n        visited = set()\n        stack = [start_index]\n        while stack and len(visited) < k:\n            node = stack.pop()\n            if node not in visited:\n                visited.add(node)\n                for neighbor in adjacency_list[node]:","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.build_adjacency_list","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.build_adjacency_list#L56-L63","kind":"function","name":"build_adjacency_list","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":56,"end_line":63,"context_start_line":36,"context_end_line":83,"code":"        with h5py.File(h5_file_path, \"r\") as f:\n            for scene_dir in tqdm(f.keys()):\n                group = f[scene_dir]\n                basenames = group[\"basenames\"][:]\n                indices = group[\"indices\"][:]\n                values = group[\"values\"][:]\n                shape = group.attrs[\"shape\"]\n                # Reconstruct the sparse matrix\n                score_matrix = np.zeros(shape, dtype=np.float32)\n                score_matrix[indices[0], indices[1]] = values\n                data_dict[scene_dir] = {\n                    \"basenames\": basenames,\n                    \"score_matrix\": self.build_adjacency_list(score_matrix),\n                }\n                self.all_ref_imgs.extend(\n                    [(scene_dir, b) for b in range(len(basenames))]\n                )\n        return data_dict\n\n    @staticmethod\n    def build_adjacency_list(S, thresh=0.2):\n        adjacency_list = [[] for _ in range(len(S))]\n        S = S - thresh\n        S[S < 0] = 0\n        rows, cols = np.nonzero(S)\n        for i, j in zip(rows, cols):\n            adjacency_list[i].append((j, S[i][j]))\n        return adjacency_list\n\n    @staticmethod\n    def is_reachable(adjacency_list, start_index, k):\n        visited = set()\n        stack = [start_index]\n        while stack and len(visited) < k:\n            node = stack.pop()\n            if node not in visited:\n                visited.add(node)\n                for neighbor in adjacency_list[node]:\n                    if neighbor[0] not in visited:\n                        stack.append(neighbor[0])\n        return len(visited) >= k\n\n    @staticmethod\n    def random_sequence_no_revisit_with_backtracking(\n        adjacency_list, k, start_index, rng: np.random.Generator\n    ):\n        path = [start_index]\n        visited = set([start_index])","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.is_reachable","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.is_reachable#L66-L76","kind":"function","name":"is_reachable","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":66,"end_line":76,"context_start_line":46,"context_end_line":96,"code":"                data_dict[scene_dir] = {\n                    \"basenames\": basenames,\n                    \"score_matrix\": self.build_adjacency_list(score_matrix),\n                }\n                self.all_ref_imgs.extend(\n                    [(scene_dir, b) for b in range(len(basenames))]\n                )\n        return data_dict\n\n    @staticmethod\n    def build_adjacency_list(S, thresh=0.2):\n        adjacency_list = [[] for _ in range(len(S))]\n        S = S - thresh\n        S[S < 0] = 0\n        rows, cols = np.nonzero(S)\n        for i, j in zip(rows, cols):\n            adjacency_list[i].append((j, S[i][j]))\n        return adjacency_list\n\n    @staticmethod\n    def is_reachable(adjacency_list, start_index, k):\n        visited = set()\n        stack = [start_index]\n        while stack and len(visited) < k:\n            node = stack.pop()\n            if node not in visited:\n                visited.add(node)\n                for neighbor in adjacency_list[node]:\n                    if neighbor[0] not in visited:\n                        stack.append(neighbor[0])\n        return len(visited) >= k\n\n    @staticmethod\n    def random_sequence_no_revisit_with_backtracking(\n        adjacency_list, k, start_index, rng: np.random.Generator\n    ):\n        path = [start_index]\n        visited = set([start_index])\n\n        neighbor_iterators = []\n        # Initialize the iterator for the start index\n        neighbors = adjacency_list[start_index]\n        neighbor_idxs = [n[0] for n in neighbors]\n        neighbor_weights = [n[1] for n in neighbors]\n        neighbor_idxs = rng.choice(\n            neighbor_idxs,\n            size=len(neighbor_idxs),\n            replace=False,\n            p=np.array(neighbor_weights) / np.sum(neighbor_weights),\n        ).tolist()\n        neighbor_iterators.append(iter(neighbor_idxs))","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.random_sequence_no_revisit_with_backtracking","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.random_sequence_no_revisit_with_backtracking#L79-L124","kind":"function","name":"random_sequence_no_revisit_with_backtracking","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":79,"end_line":124,"context_start_line":59,"context_end_line":144,"code":"        S[S < 0] = 0\n        rows, cols = np.nonzero(S)\n        for i, j in zip(rows, cols):\n            adjacency_list[i].append((j, S[i][j]))\n        return adjacency_list\n\n    @staticmethod\n    def is_reachable(adjacency_list, start_index, k):\n        visited = set()\n        stack = [start_index]\n        while stack and len(visited) < k:\n            node = stack.pop()\n            if node not in visited:\n                visited.add(node)\n                for neighbor in adjacency_list[node]:\n                    if neighbor[0] not in visited:\n                        stack.append(neighbor[0])\n        return len(visited) >= k\n\n    @staticmethod\n    def random_sequence_no_revisit_with_backtracking(\n        adjacency_list, k, start_index, rng: np.random.Generator\n    ):\n        path = [start_index]\n        visited = set([start_index])\n\n        neighbor_iterators = []\n        # Initialize the iterator for the start index\n        neighbors = adjacency_list[start_index]\n        neighbor_idxs = [n[0] for n in neighbors]\n        neighbor_weights = [n[1] for n in neighbors]\n        neighbor_idxs = rng.choice(\n            neighbor_idxs,\n            size=len(neighbor_idxs),\n            replace=False,\n            p=np.array(neighbor_weights) / np.sum(neighbor_weights),\n        ).tolist()\n        neighbor_iterators.append(iter(neighbor_idxs))\n\n        while len(path) < k:\n            if not neighbor_iterators:\n                # No possible sequence\n                return None\n            current_iterator = neighbor_iterators[-1]\n            try:\n                next_index = next(current_iterator)\n                if next_index not in visited:\n                    path.append(next_index)\n                    visited.add(next_index)\n\n                    # Prepare iterator for the next node\n                    neighbors = adjacency_list[next_index]\n                    neighbor_idxs = [n[0] for n in neighbors]\n                    neighbor_weights = [n[1] for n in neighbors]\n                    neighbor_idxs = rng.choice(\n                        neighbor_idxs,\n                        size=len(neighbor_idxs),\n                        replace=False,\n                        p=np.array(neighbor_weights) / np.sum(neighbor_weights),\n                    ).tolist()\n                    neighbor_iterators.append(iter(neighbor_idxs))\n            except StopIteration:\n                # No more neighbors to try at this node, backtrack\n                neighbor_iterators.pop()\n                visited.remove(path.pop())\n        return path\n\n    @staticmethod\n    def random_sequence_with_optional_repeats(\n        adjacency_list,\n        k,\n        start_index,\n        rng: np.random.Generator,\n        max_k=None,\n        max_attempts=100,\n    ):\n        if max_k is None:\n            max_k = k\n        path = [start_index]\n        visited = set([start_index])\n        current_index = start_index\n        attempts = 0\n\n        while len(path) < max_k and attempts < max_attempts:\n            attempts += 1\n            neighbors = adjacency_list[current_index]","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.random_sequence_with_optional_repeats","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.random_sequence_with_optional_repeats#L127-L183","kind":"function","name":"random_sequence_with_optional_repeats","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":127,"end_line":183,"context_start_line":107,"context_end_line":203,"code":"                    visited.add(next_index)\n\n                    # Prepare iterator for the next node\n                    neighbors = adjacency_list[next_index]\n                    neighbor_idxs = [n[0] for n in neighbors]\n                    neighbor_weights = [n[1] for n in neighbors]\n                    neighbor_idxs = rng.choice(\n                        neighbor_idxs,\n                        size=len(neighbor_idxs),\n                        replace=False,\n                        p=np.array(neighbor_weights) / np.sum(neighbor_weights),\n                    ).tolist()\n                    neighbor_iterators.append(iter(neighbor_idxs))\n            except StopIteration:\n                # No more neighbors to try at this node, backtrack\n                neighbor_iterators.pop()\n                visited.remove(path.pop())\n        return path\n\n    @staticmethod\n    def random_sequence_with_optional_repeats(\n        adjacency_list,\n        k,\n        start_index,\n        rng: np.random.Generator,\n        max_k=None,\n        max_attempts=100,\n    ):\n        if max_k is None:\n            max_k = k\n        path = [start_index]\n        visited = set([start_index])\n        current_index = start_index\n        attempts = 0\n\n        while len(path) < max_k and attempts < max_attempts:\n            attempts += 1\n            neighbors = adjacency_list[current_index]\n            neighbor_idxs = [n[0] for n in neighbors]\n            neighbor_weights = [n[1] for n in neighbors]\n\n            if not neighbor_idxs:\n                # No neighbors, cannot proceed further\n                break\n\n            # Try to find unvisited neighbors\n            unvisited_neighbors = [\n                (idx, wgt)\n                for idx, wgt in zip(neighbor_idxs, neighbor_weights)\n                if idx not in visited\n            ]\n            if unvisited_neighbors:\n                # Select among unvisited neighbors\n                unvisited_idxs = [idx for idx, _ in unvisited_neighbors]\n                unvisited_weights = [wgt for _, wgt in unvisited_neighbors]\n                probabilities = np.array(unvisited_weights) / np.sum(unvisited_weights)\n                next_index = rng.choice(unvisited_idxs, p=probabilities)\n                visited.add(next_index)\n            else:\n                # All neighbors visited, but we need to reach length max_k\n                # So we can revisit nodes\n                probabilities = np.array(neighbor_weights) / np.sum(neighbor_weights)\n                next_index = rng.choice(neighbor_idxs, p=probabilities)\n\n            path.append(next_index)\n            current_index = next_index\n\n        if len(set(path)) >= k:\n            # If path is shorter than max_k, extend it by repeating existing elements\n            while len(path) < max_k:\n                # Randomly select nodes from the existing path to repeat\n                next_index = rng.choice(path)\n                path.append(next_index)\n            return path\n        else:\n            # Could not reach k unique nodes\n            return None\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def get_image_num(self):\n        return self.num_imgs\n\n    def get_stats(self):\n        return f\"{len(self)} imgs from {self.num_scenes} scenes\"\n\n    def generate_sequence(\n        self, scene, adj_list, num_views, start_index, rng, allow_repeat=False\n    ):\n        cutoff = num_views if not allow_repeat else max(num_views // 5, 3)\n        if start_index in self.is_reachable_cache[scene]:\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.__len__","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.__len__#L185-L186","kind":"function","name":"__len__","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":185,"end_line":186,"context_start_line":165,"context_end_line":206,"code":"            else:\n                # All neighbors visited, but we need to reach length max_k\n                # So we can revisit nodes\n                probabilities = np.array(neighbor_weights) / np.sum(neighbor_weights)\n                next_index = rng.choice(neighbor_idxs, p=probabilities)\n\n            path.append(next_index)\n            current_index = next_index\n\n        if len(set(path)) >= k:\n            # If path is shorter than max_k, extend it by repeating existing elements\n            while len(path) < max_k:\n                # Randomly select nodes from the existing path to repeat\n                next_index = rng.choice(path)\n                path.append(next_index)\n            return path\n        else:\n            # Could not reach k unique nodes\n            return None\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def get_image_num(self):\n        return self.num_imgs\n\n    def get_stats(self):\n        return f\"{len(self)} imgs from {self.num_scenes} scenes\"\n\n    def generate_sequence(\n        self, scene, adj_list, num_views, start_index, rng, allow_repeat=False\n    ):\n        cutoff = num_views if not allow_repeat else max(num_views // 5, 3)\n        if start_index in self.is_reachable_cache[scene]:\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        else:\n            self.is_reachable_cache[scene][start_index] = self.is_reachable(\n                adj_list, start_index, cutoff","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.get_image_num#L188-L189","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":188,"end_line":189,"context_start_line":168,"context_end_line":209,"code":"                probabilities = np.array(neighbor_weights) / np.sum(neighbor_weights)\n                next_index = rng.choice(neighbor_idxs, p=probabilities)\n\n            path.append(next_index)\n            current_index = next_index\n\n        if len(set(path)) >= k:\n            # If path is shorter than max_k, extend it by repeating existing elements\n            while len(path) < max_k:\n                # Randomly select nodes from the existing path to repeat\n                next_index = rng.choice(path)\n                path.append(next_index)\n            return path\n        else:\n            # Could not reach k unique nodes\n            return None\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def get_image_num(self):\n        return self.num_imgs\n\n    def get_stats(self):\n        return f\"{len(self)} imgs from {self.num_scenes} scenes\"\n\n    def generate_sequence(\n        self, scene, adj_list, num_views, start_index, rng, allow_repeat=False\n    ):\n        cutoff = num_views if not allow_repeat else max(num_views // 5, 3)\n        if start_index in self.is_reachable_cache[scene]:\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        else:\n            self.is_reachable_cache[scene][start_index] = self.is_reachable(\n                adj_list, start_index, cutoff\n            )\n            if not self.is_reachable_cache[scene][start_index]:\n                print(","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.get_stats","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.get_stats#L191-L192","kind":"function","name":"get_stats","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":191,"end_line":192,"context_start_line":171,"context_end_line":212,"code":"            path.append(next_index)\n            current_index = next_index\n\n        if len(set(path)) >= k:\n            # If path is shorter than max_k, extend it by repeating existing elements\n            while len(path) < max_k:\n                # Randomly select nodes from the existing path to repeat\n                next_index = rng.choice(path)\n                path.append(next_index)\n            return path\n        else:\n            # Could not reach k unique nodes\n            return None\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def get_image_num(self):\n        return self.num_imgs\n\n    def get_stats(self):\n        return f\"{len(self)} imgs from {self.num_scenes} scenes\"\n\n    def generate_sequence(\n        self, scene, adj_list, num_views, start_index, rng, allow_repeat=False\n    ):\n        cutoff = num_views if not allow_repeat else max(num_views // 5, 3)\n        if start_index in self.is_reachable_cache[scene]:\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        else:\n            self.is_reachable_cache[scene][start_index] = self.is_reachable(\n                adj_list, start_index, cutoff\n            )\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs.generate_sequence","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs.generate_sequence#L194-L224","kind":"function","name":"generate_sequence","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":194,"end_line":224,"context_start_line":174,"context_end_line":244,"code":"        if len(set(path)) >= k:\n            # If path is shorter than max_k, extend it by repeating existing elements\n            while len(path) < max_k:\n                # Randomly select nodes from the existing path to repeat\n                next_index = rng.choice(path)\n                path.append(next_index)\n            return path\n        else:\n            # Could not reach k unique nodes\n            return None\n\n    def __len__(self):\n        return len(self.all_ref_imgs)\n\n    def get_image_num(self):\n        return self.num_imgs\n\n    def get_stats(self):\n        return f\"{len(self)} imgs from {self.num_scenes} scenes\"\n\n    def generate_sequence(\n        self, scene, adj_list, num_views, start_index, rng, allow_repeat=False\n    ):\n        cutoff = num_views if not allow_repeat else max(num_views // 5, 3)\n        if start_index in self.is_reachable_cache[scene]:\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        else:\n            self.is_reachable_cache[scene][start_index] = self.is_reachable(\n                adj_list, start_index, cutoff\n            )\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        if not allow_repeat:\n            sequence = self.random_sequence_no_revisit_with_backtracking(\n                adj_list, cutoff, start_index, rng\n            )\n        else:\n            sequence = self.random_sequence_with_optional_repeats(\n                adj_list, cutoff, start_index, rng, max_k=num_views\n            )\n        if not sequence:\n            self.is_reachable_cache[scene][start_index] = False\n            print(\"Failed to generate a sequence without revisiting.\")\n        return sequence\n\n    def _get_views(self, idx, resolution, rng: np.random.Generator, num_views):\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n        invalid_seq = True\n        ordered_video = False\n\n        while invalid_seq:\n            basenames = self.data_dict[scene_info][\"basenames\"]\n            if (\n                sum(\n                    [\n                        (1 - int(x))\n                        for x in list(self.is_reachable_cache[scene_info].values())\n                    ]\n                )\n                > len(basenames) - self.num_views\n            ):\n                self.invalid_scenes.append(scene_info)\n            while scene_info in self.invalid_scenes:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.blendedmvs._get_views","uri":"program://Human3R/function/src.dust3r.datasets.blendedmvs._get_views#L226-L305","kind":"function","name":"_get_views","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":226,"end_line":305,"context_start_line":206,"context_end_line":305,"code":"                adj_list, start_index, cutoff\n            )\n            if not self.is_reachable_cache[scene][start_index]:\n                print(\n                    f\"Cannot reach {num_views} unique elements from index {start_index}.\"\n                )\n                return None\n        if not allow_repeat:\n            sequence = self.random_sequence_no_revisit_with_backtracking(\n                adj_list, cutoff, start_index, rng\n            )\n        else:\n            sequence = self.random_sequence_with_optional_repeats(\n                adj_list, cutoff, start_index, rng, max_k=num_views\n            )\n        if not sequence:\n            self.is_reachable_cache[scene][start_index] = False\n            print(\"Failed to generate a sequence without revisiting.\")\n        return sequence\n\n    def _get_views(self, idx, resolution, rng: np.random.Generator, num_views):\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n        invalid_seq = True\n        ordered_video = False\n\n        while invalid_seq:\n            basenames = self.data_dict[scene_info][\"basenames\"]\n            if (\n                sum(\n                    [\n                        (1 - int(x))\n                        for x in list(self.is_reachable_cache[scene_info].values())\n                    ]\n                )\n                > len(basenames) - self.num_views\n            ):\n                self.invalid_scenes.append(scene_info)\n            while scene_info in self.invalid_scenes:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n                basenames = self.data_dict[scene_info][\"basenames\"]\n\n            score_matrix = self.data_dict[scene_info][\"score_matrix\"]\n            imgs_idxs = self.generate_sequence(\n                scene_info, score_matrix, num_views, ref_img_idx, rng, self.allow_repeat\n            )\n\n            if imgs_idxs is None:\n                random_direction = 2 * rng.choice(2) - 1\n                for offset in range(1, len(basenames)):\n                    tentative_im_idx = (\n                        ref_img_idx + (random_direction * offset)\n                    ) % len(basenames)\n                    if (\n                        tentative_im_idx not in self.is_reachable_cache[scene_info]\n                        or self.is_reachable_cache[scene_info][tentative_im_idx]\n                    ):\n                        ref_img_idx = tentative_im_idx\n                        break\n            else:\n                invalid_seq = False\n        views = []\n        for view_idx in imgs_idxs:\n            scene_dir = osp.join(self.ROOT, scene_info)\n            impath = basenames[view_idx].decode(\"utf-8\")\n            image = imread_cv2(osp.join(scene_dir, impath + \".jpg\"))\n            depthmap = imread_cv2(osp.join(scene_dir, impath + \".exr\"))\n            camera_params = np.load(osp.join(scene_dir, impath + \".npz\"))\n\n            intrinsics = np.float32(camera_params[\"intrinsics\"])\n            camera_pose = np.eye(4, dtype=np.float32)\n            camera_pose[:3, :3] = camera_params[\"R_cam2world\"]\n            camera_pose[:3, 3] = camera_params[\"t_cam2world\"]\n\n            image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                image, depthmap, intrinsics, resolution, rng, info=(scene_dir, impath)\n            )\n\n            views.append(\n                dict(\n                    img=image,\n                    depthmap=depthmap,\n                    camera_pose=camera_pose,  # cam2world\n                    camera_intrinsics=intrinsics,\n                    dataset=\"BlendedMVS\",\n                    label=osp.relpath(scene_dir, self.ROOT),\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    instance=osp.join(scene_dir, impath + \".jpg\"),\n                    quantile=np.array(0.97, dtype=np.float32),\n                    img_mask=True,\n                    ray_mask=False,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.omniobject3d","uri":"program://Human3R/module/src.dust3r.datasets.omniobject3d#L1-L146","kind":"module","name":"src.dust3r.datasets.omniobject3d","path":"src/dust3r/datasets/omniobject3d.py","language":"python","start_line":1,"end_line":146,"context_start_line":1,"context_end_line":146,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nimport json\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass OmniObject3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False  # True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = [\n            d\n            for d in os.listdir(self.ROOT)\n            if os.path.isdir(os.path.join(self.ROOT, d))\n        ]\n        with open(os.path.join(self.ROOT, \"scale.json\"), \"r\") as f:\n            self.scales = json.load(f)\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=extract_number,\n            )\n\n            num_imgs = len(basenames)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene, start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=100, video_prob=0.0\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            scale = self.scales[self.scenes[scene_id]]\n            depthmap = depthmap / scale / 1000.0\n            camera_pose[:3, 3] = camera_pose[:3, 3] / scale / 1000.0\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.8, 0.15, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"OmniObject3D\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"0f11e6da03b9fe41e86f40fe1281e1ad5e89ec050643e731f0176236c6089282","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.omniobject3d.extract_number","uri":"program://Human3R/function/src.dust3r.datasets.omniobject3d.extract_number#L16-L20","kind":"function","name":"extract_number","path":"src/dust3r/datasets/omniobject3d.py","language":"python","start_line":16,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nimport json\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass OmniObject3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False  # True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = [\n            d\n            for d in os.listdir(self.ROOT)\n            if os.path.isdir(os.path.join(self.ROOT, d))\n        ]\n        with open(os.path.join(self.ROOT, \"scale.json\"), \"r\") as f:\n            self.scales = json.load(f)\n        offset = 0","source_hash":"0f11e6da03b9fe41e86f40fe1281e1ad5e89ec050643e731f0176236c6089282","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.omniobject3d.OmniObject3D_Multi","uri":"program://Human3R/class/src.dust3r.datasets.omniobject3d.OmniObject3D_Multi#L23-L146","kind":"class","name":"OmniObject3D_Multi","path":"src/dust3r/datasets/omniobject3d.py","language":"python","start_line":23,"end_line":146,"context_start_line":3,"context_end_line":146,"code":"import numpy as np\nimport itertools\nimport os\nimport sys\nimport json\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass OmniObject3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False  # True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = [\n            d\n            for d in os.listdir(self.ROOT)\n            if os.path.isdir(os.path.join(self.ROOT, d))\n        ]\n        with open(os.path.join(self.ROOT, \"scale.json\"), \"r\") as f:\n            self.scales = json.load(f)\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=extract_number,\n            )\n\n            num_imgs = len(basenames)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene, start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=100, video_prob=0.0\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            scale = self.scales[self.scenes[scene_id]]\n            depthmap = depthmap / scale / 1000.0\n            camera_pose[:3, 3] = camera_pose[:3, 3] / scale / 1000.0\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.8, 0.15, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"OmniObject3D\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"0f11e6da03b9fe41e86f40fe1281e1ad5e89ec050643e731f0176236c6089282","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.omniobject3d.__init__","uri":"program://Human3R/function/src.dust3r.datasets.omniobject3d.__init__#L24-L30","kind":"function","name":"__init__","path":"src/dust3r/datasets/omniobject3d.py","language":"python","start_line":24,"end_line":30,"context_start_line":4,"context_end_line":50,"code":"import itertools\nimport os\nimport sys\nimport json\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass OmniObject3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False  # True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = [\n            d\n            for d in os.listdir(self.ROOT)\n            if os.path.isdir(os.path.join(self.ROOT, d))\n        ]\n        with open(os.path.join(self.ROOT, \"scale.json\"), \"r\") as f:\n            self.scales = json.load(f)\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")","source_hash":"0f11e6da03b9fe41e86f40fe1281e1ad5e89ec050643e731f0176236c6089282","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.omniobject3d._load_data","uri":"program://Human3R/function/src.dust3r.datasets.omniobject3d._load_data#L32-L81","kind":"function","name":"_load_data","path":"src/dust3r/datasets/omniobject3d.py","language":"python","start_line":32,"end_line":81,"context_start_line":12,"context_end_line":101,"code":"from dust3r.utils.image import imread_cv2\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n\n\nclass OmniObject3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False  # True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):\n        self.scenes = [\n            d\n            for d in os.listdir(self.ROOT)\n            if os.path.isdir(os.path.join(self.ROOT, d))\n        ]\n        with open(os.path.join(self.ROOT, \"scale.json\"), \"r\") as f:\n            self.scales = json.load(f)\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, scene)\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=extract_number,\n            )\n\n            num_imgs = len(basenames)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene, start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=100, video_prob=0.0\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")","source_hash":"0f11e6da03b9fe41e86f40fe1281e1ad5e89ec050643e731f0176236c6089282","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.omniobject3d.__len__","uri":"program://Human3R/function/src.dust3r.datasets.omniobject3d.__len__#L83-L84","kind":"function","name":"__len__","path":"src/dust3r/datasets/omniobject3d.py","language":"python","start_line":83,"end_line":84,"context_start_line":63,"context_end_line":104,"code":"                continue\n            img_ids = list(np.arange(num_imgs) + offset)\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene, start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=100, video_prob=0.0\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n","source_hash":"0f11e6da03b9fe41e86f40fe1281e1ad5e89ec050643e731f0176236c6089282","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.omniobject3d.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.omniobject3d.get_image_num#L86-L87","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/omniobject3d.py","language":"python","start_line":86,"end_line":87,"context_start_line":66,"context_end_line":107,"code":"\n            start_img_ids.extend([(scene, id) for id in start_img_ids_])\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene, start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=100, video_prob=0.0\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image","source_hash":"0f11e6da03b9fe41e86f40fe1281e1ad5e89ec050643e731f0176236c6089282","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.omniobject3d._get_views","uri":"program://Human3R/function/src.dust3r.datasets.omniobject3d._get_views#L89-L146","kind":"function","name":"_get_views","path":"src/dust3r/datasets/omniobject3d.py","language":"python","start_line":89,"end_line":146,"context_start_line":69,"context_end_line":146,"code":"            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        scene, start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views, start_id, all_image_ids, rng, max_interval=100, video_prob=0.0\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            scale = self.scales[self.scenes[scene_id]]\n            depthmap = depthmap / scale / 1000.0\n            camera_pose[:3, 3] = camera_pose[:3, 3] / scale / 1000.0\n\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.8, 0.15, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"OmniObject3D\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"0f11e6da03b9fe41e86f40fe1281e1ad5e89ec050643e731f0176236c6089282","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannet","uri":"program://Human3R/module/src.dust3r.datasets.scannet#L1-L148","kind":"module","name":"src.dust3r.datasets.scannet","path":"src/dust3r/datasets/scannet.py","language":"python","start_line":1,"end_line":148,"context_start_line":1,"context_end_line":148,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scene_root = osp.join(\n            self.ROOT, \"scans_train\" if split == \"train\" else \"scans_test\"\n        )\n        self.scenes = [\n            scene for scene in os.listdir(self.scene_root) if scene.startswith(\"scene\")\n        ]\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.scene_root, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                basenames = data[\"images\"]\n                num_imgs = len(basenames)\n                img_ids = list(np.arange(num_imgs) + offset)\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                start_img_ids.extend(start_img_ids_)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                scenes.append(scene)\n                scene_img_list.append(img_ids)\n\n                # offset groups\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.6,\n            fix_interval_prob=0.6,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.scene_root, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"color\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(depth_dir, basename + \".png\"), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"ScanNet\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"89f549daf57cbe821d15f7050b7c9fba0ebc4e813aa4ca436245edda8a6f8a9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannet.ScanNet_Multi","uri":"program://Human3R/class/src.dust3r.datasets.scannet.ScanNet_Multi#L14-L148","kind":"class","name":"ScanNet_Multi","path":"src/dust3r/datasets/scannet.py","language":"python","start_line":14,"end_line":148,"context_start_line":1,"context_end_line":148,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scene_root = osp.join(\n            self.ROOT, \"scans_train\" if split == \"train\" else \"scans_test\"\n        )\n        self.scenes = [\n            scene for scene in os.listdir(self.scene_root) if scene.startswith(\"scene\")\n        ]\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.scene_root, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                basenames = data[\"images\"]\n                num_imgs = len(basenames)\n                img_ids = list(np.arange(num_imgs) + offset)\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                start_img_ids.extend(start_img_ids_)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                scenes.append(scene)\n                scene_img_list.append(img_ids)\n\n                # offset groups\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.6,\n            fix_interval_prob=0.6,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.scene_root, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"color\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(depth_dir, basename + \".png\"), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"ScanNet\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"89f549daf57cbe821d15f7050b7c9fba0ebc4e813aa4ca436245edda8a6f8a9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannet.__init__","uri":"program://Human3R/function/src.dust3r.datasets.scannet.__init__#L15-L22","kind":"function","name":"__init__","path":"src/dust3r/datasets/scannet.py","language":"python","start_line":15,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scene_root = osp.join(\n            self.ROOT, \"scans_train\" if split == \"train\" else \"scans_test\"\n        )\n        self.scenes = [\n            scene for scene in os.listdir(self.scene_root) if scene.startswith(\"scene\")\n        ]\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.scene_root, scene)\n            with np.load(","source_hash":"89f549daf57cbe821d15f7050b7c9fba0ebc4e813aa4ca436245edda8a6f8a9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannet._load_data","uri":"program://Human3R/function/src.dust3r.datasets.scannet._load_data#L24-L73","kind":"function","name":"_load_data","path":"src/dust3r/datasets/scannet.py","language":"python","start_line":24,"end_line":73,"context_start_line":4,"context_end_line":93,"code":"import itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scene_root = osp.join(\n            self.ROOT, \"scans_train\" if split == \"train\" else \"scans_test\"\n        )\n        self.scenes = [\n            scene for scene in os.listdir(self.scene_root) if scene.startswith(\"scene\")\n        ]\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.scene_root, scene)\n            with np.load(\n                osp.join(scene_dir, \"new_scene_metadata.npz\"), allow_pickle=True\n            ) as data:\n                basenames = data[\"images\"]\n                num_imgs = len(basenames)\n                img_ids = list(np.arange(num_imgs) + offset)\n                cut_off = (\n                    self.num_views\n                    if not self.allow_repeat\n                    else max(self.num_views // 3, 3)\n                )\n                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n\n                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                start_img_ids.extend(start_img_ids_)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                scenes.append(scene)\n                scene_img_list.append(img_ids)\n\n                # offset groups\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.6,\n            fix_interval_prob=0.6,\n            block_shuffle=16,\n        )","source_hash":"89f549daf57cbe821d15f7050b7c9fba0ebc4e813aa4ca436245edda8a6f8a9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannet.__len__","uri":"program://Human3R/function/src.dust3r.datasets.scannet.__len__#L75-L76","kind":"function","name":"__len__","path":"src/dust3r/datasets/scannet.py","language":"python","start_line":75,"end_line":76,"context_start_line":55,"context_end_line":96,"code":"                if num_imgs < cut_off:\n                    print(f\"Skipping {scene}\")\n                    continue\n\n                start_img_ids.extend(start_img_ids_)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                scenes.append(scene)\n                scene_img_list.append(img_ids)\n\n                # offset groups\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.6,\n            fix_interval_prob=0.6,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []","source_hash":"89f549daf57cbe821d15f7050b7c9fba0ebc4e813aa4ca436245edda8a6f8a9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannet.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.scannet.get_image_num#L78-L79","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/scannet.py","language":"python","start_line":78,"end_line":79,"context_start_line":58,"context_end_line":99,"code":"\n                start_img_ids.extend(start_img_ids_)\n                sceneids.extend([j] * num_imgs)\n                images.extend(basenames)\n                scenes.append(scene)\n                scene_img_list.append(img_ids)\n\n                # offset groups\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.6,\n            fix_interval_prob=0.6,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.scene_root, self.scenes[scene_id])","source_hash":"89f549daf57cbe821d15f7050b7c9fba0ebc4e813aa4ca436245edda8a6f8a9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.scannet._get_views","uri":"program://Human3R/function/src.dust3r.datasets.scannet._get_views#L81-L148","kind":"function","name":"_get_views","path":"src/dust3r/datasets/scannet.py","language":"python","start_line":81,"end_line":148,"context_start_line":61,"context_end_line":148,"code":"                images.extend(basenames)\n                scenes.append(scene)\n                scene_img_list.append(img_ids)\n\n                # offset groups\n                offset += num_imgs\n                j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=0.6,\n            fix_interval_prob=0.6,\n            block_shuffle=16,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.scene_root, self.scenes[scene_id])\n            rgb_dir = osp.join(scene_dir, \"color\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".jpg\"))\n            # Load depthmap\n            depthmap = imread_cv2(\n                osp.join(depth_dir, basename + \".png\"), cv2.IMREAD_UNCHANGED\n            )\n            depthmap = depthmap.astype(np.float32) / 1000\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.75, 0.2, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"ScanNet\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(0.98, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"89f549daf57cbe821d15f7050b7c9fba0ebc4e813aa4ca436245edda8a6f8a9b","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.cop3d","uri":"program://Human3R/module/src.dust3r.datasets.cop3d#L1-L110","kind":"module","name":"src.dust3r.datasets.cop3d","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":1,"end_line":110,"context_start_line":1,"context_end_line":110,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Cop3D_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"Cop3D\"\n        self.is_metric = False\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        # no depth, pseduo path just for getting the right resolution\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, impath, input_metadata):\n        # no depth, set to all ones\n        img = imread_cv2(impath, cv2.IMREAD_UNCHANGED)\n        depthmap = np.ones_like(img[..., 0], dtype=np.float32)\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]\n            if len(image_pool) < self.num_views:\n                print(\"Invalid scene!\")\n                self.invalid_scenes[scene_info] = True\n                continue\n\n            imgs_idxs, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                ref_img_idx,\n                image_pool,\n                rng,\n                max_interval=5,\n                video_prob=1.0,\n                fix_interval_prob=0.9,\n            )\n\n            views = []\n\n            for im_idx in imgs_idxs:\n                view_idx = image_pool[im_idx]\n                impath = self._get_impath(obj, instance, view_idx)\n                depthpath = self._get_depthpath(obj, instance, view_idx)\n\n                # load camera params\n                metadata_path = self._get_metadatapath(obj, instance, view_idx)\n                input_metadata = np.load(metadata_path)\n                camera_pose = input_metadata[\"camera_pose\"].astype(np.float32)\n                intrinsics = input_metadata[\"camera_intrinsics\"].astype(np.float32)\n\n                # load image and depth\n                rgb_image = imread_cv2(impath)\n                depthmap = self._read_depthmap(depthpath, input_metadata)\n\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=impath\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap,\n                        camera_pose=camera_pose,\n                        camera_intrinsics=intrinsics,\n                        dataset=self.dataset_label,\n                        label=osp.join(obj, instance),\n                        instance=osp.split(impath)[1],\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.96, dtype=np.float32),\n                        img_mask=True,\n                        ray_mask=False,\n                        camera_only=True,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n\n            if len(views) == num_views and not all(\n                [view[\"instance\"] == views[0][\"instance\"] for view in views]\n            ):\n                invalid_seq = False\n        return views","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.cop3d.Cop3D_Multi","uri":"program://Human3R/class/src.dust3r.datasets.cop3d.Cop3D_Multi#L12-L110","kind":"class","name":"Cop3D_Multi","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":12,"end_line":110,"context_start_line":1,"context_end_line":110,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Cop3D_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"Cop3D\"\n        self.is_metric = False\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        # no depth, pseduo path just for getting the right resolution\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, impath, input_metadata):\n        # no depth, set to all ones\n        img = imread_cv2(impath, cv2.IMREAD_UNCHANGED)\n        depthmap = np.ones_like(img[..., 0], dtype=np.float32)\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]\n            if len(image_pool) < self.num_views:\n                print(\"Invalid scene!\")\n                self.invalid_scenes[scene_info] = True\n                continue\n\n            imgs_idxs, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                ref_img_idx,\n                image_pool,\n                rng,\n                max_interval=5,\n                video_prob=1.0,\n                fix_interval_prob=0.9,\n            )\n\n            views = []\n\n            for im_idx in imgs_idxs:\n                view_idx = image_pool[im_idx]\n                impath = self._get_impath(obj, instance, view_idx)\n                depthpath = self._get_depthpath(obj, instance, view_idx)\n\n                # load camera params\n                metadata_path = self._get_metadatapath(obj, instance, view_idx)\n                input_metadata = np.load(metadata_path)\n                camera_pose = input_metadata[\"camera_pose\"].astype(np.float32)\n                intrinsics = input_metadata[\"camera_intrinsics\"].astype(np.float32)\n\n                # load image and depth\n                rgb_image = imread_cv2(impath)\n                depthmap = self._read_depthmap(depthpath, input_metadata)\n\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=impath\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap,\n                        camera_pose=camera_pose,\n                        camera_intrinsics=intrinsics,\n                        dataset=self.dataset_label,\n                        label=osp.join(obj, instance),\n                        instance=osp.split(impath)[1],\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.96, dtype=np.float32),\n                        img_mask=True,\n                        ray_mask=False,\n                        camera_only=True,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n\n            if len(views) == num_views and not all(\n                [view[\"instance\"] == views[0][\"instance\"] for view in views]\n            ):\n                invalid_seq = False\n        return views","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.cop3d.__init__","uri":"program://Human3R/function/src.dust3r.datasets.cop3d.__init__#L13-L16","kind":"function","name":"__init__","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":13,"end_line":16,"context_start_line":1,"context_end_line":36,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Cop3D_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"Cop3D\"\n        self.is_metric = False\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        # no depth, pseduo path just for getting the right resolution\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, impath, input_metadata):\n        # no depth, set to all ones\n        img = imread_cv2(impath, cv2.IMREAD_UNCHANGED)\n        depthmap = np.ones_like(img[..., 0], dtype=np.float32)\n        return depthmap\n","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.cop3d._get_metadatapath","uri":"program://Human3R/function/src.dust3r.datasets.cop3d._get_metadatapath#L18-L19","kind":"function","name":"_get_metadatapath","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":18,"end_line":19,"context_start_line":1,"context_end_line":39,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Cop3D_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"Cop3D\"\n        self.is_metric = False\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        # no depth, pseduo path just for getting the right resolution\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, impath, input_metadata):\n        # no depth, set to all ones\n        img = imread_cv2(impath, cv2.IMREAD_UNCHANGED)\n        depthmap = np.ones_like(img[..., 0], dtype=np.float32)\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.cop3d._get_impath","uri":"program://Human3R/function/src.dust3r.datasets.cop3d._get_impath#L21-L22","kind":"function","name":"_get_impath","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":21,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Cop3D_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"Cop3D\"\n        self.is_metric = False\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        # no depth, pseduo path just for getting the right resolution\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, impath, input_metadata):\n        # no depth, set to all ones\n        img = imread_cv2(impath, cv2.IMREAD_UNCHANGED)\n        depthmap = np.ones_like(img[..., 0], dtype=np.float32)\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.cop3d._get_depthpath","uri":"program://Human3R/function/src.dust3r.datasets.cop3d._get_depthpath#L24-L26","kind":"function","name":"_get_depthpath","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":24,"end_line":26,"context_start_line":4,"context_end_line":46,"code":"sys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Cop3D_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"Cop3D\"\n        self.is_metric = False\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        # no depth, pseduo path just for getting the right resolution\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, impath, input_metadata):\n        # no depth, set to all ones\n        img = imread_cv2(impath, cv2.IMREAD_UNCHANGED)\n        depthmap = np.ones_like(img[..., 0], dtype=np.float32)\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.cop3d._get_maskpath","uri":"program://Human3R/function/src.dust3r.datasets.cop3d._get_maskpath#L28-L29","kind":"function","name":"_get_maskpath","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":28,"end_line":29,"context_start_line":8,"context_end_line":49,"code":"from dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Cop3D_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"Cop3D\"\n        self.is_metric = False\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        # no depth, pseduo path just for getting the right resolution\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, impath, input_metadata):\n        # no depth, set to all ones\n        img = imread_cv2(impath, cv2.IMREAD_UNCHANGED)\n        depthmap = np.ones_like(img[..., 0], dtype=np.float32)\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]\n            if len(image_pool) < self.num_views:","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.cop3d._read_depthmap","uri":"program://Human3R/function/src.dust3r.datasets.cop3d._read_depthmap#L31-L35","kind":"function","name":"_read_depthmap","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":31,"end_line":35,"context_start_line":11,"context_end_line":55,"code":"\nclass Cop3D_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"Cop3D\"\n        self.is_metric = False\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        # no depth, pseduo path just for getting the right resolution\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, impath, input_metadata):\n        # no depth, set to all ones\n        img = imread_cv2(impath, cv2.IMREAD_UNCHANGED)\n        depthmap = np.ones_like(img[..., 0], dtype=np.float32)\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]\n            if len(image_pool) < self.num_views:\n                print(\"Invalid scene!\")\n                self.invalid_scenes[scene_info] = True\n                continue\n\n            imgs_idxs, ordered_video = self.get_seq_from_start_id(\n                num_views,","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.cop3d._get_views","uri":"program://Human3R/function/src.dust3r.datasets.cop3d._get_views#L37-L110","kind":"function","name":"_get_views","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":37,"end_line":110,"context_start_line":17,"context_end_line":110,"code":"\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_depthpath(self, obj, instance, view_idx):\n        # no depth, pseduo path just for getting the right resolution\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.jpg\")\n\n    def _get_maskpath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"masks\", f\"frame{view_idx:06n}.png\")\n\n    def _read_depthmap(self, impath, input_metadata):\n        # no depth, set to all ones\n        img = imread_cv2(impath, cv2.IMREAD_UNCHANGED)\n        depthmap = np.ones_like(img[..., 0], dtype=np.float32)\n        return depthmap\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        invalid_seq = True\n        scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n        while invalid_seq:\n            while self.invalid_scenes[scene_info]:\n                idx = rng.integers(low=0, high=len(self.all_ref_imgs))\n                scene_info, ref_img_idx = self.all_ref_imgs[idx]\n\n            obj, instance = scene_info\n\n            image_pool = self.scenes[obj, instance]\n            if len(image_pool) < self.num_views:\n                print(\"Invalid scene!\")\n                self.invalid_scenes[scene_info] = True\n                continue\n\n            imgs_idxs, ordered_video = self.get_seq_from_start_id(\n                num_views,\n                ref_img_idx,\n                image_pool,\n                rng,\n                max_interval=5,\n                video_prob=1.0,\n                fix_interval_prob=0.9,\n            )\n\n            views = []\n\n            for im_idx in imgs_idxs:\n                view_idx = image_pool[im_idx]\n                impath = self._get_impath(obj, instance, view_idx)\n                depthpath = self._get_depthpath(obj, instance, view_idx)\n\n                # load camera params\n                metadata_path = self._get_metadatapath(obj, instance, view_idx)\n                input_metadata = np.load(metadata_path)\n                camera_pose = input_metadata[\"camera_pose\"].astype(np.float32)\n                intrinsics = input_metadata[\"camera_intrinsics\"].astype(np.float32)\n\n                # load image and depth\n                rgb_image = imread_cv2(impath)\n                depthmap = self._read_depthmap(depthpath, input_metadata)\n\n                rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                    rgb_image, depthmap, intrinsics, resolution, rng=rng, info=impath\n                )\n\n                views.append(\n                    dict(\n                        img=rgb_image,\n                        depthmap=depthmap,\n                        camera_pose=camera_pose,\n                        camera_intrinsics=intrinsics,\n                        dataset=self.dataset_label,\n                        label=osp.join(obj, instance),\n                        instance=osp.split(impath)[1],\n                        is_metric=self.is_metric,\n                        is_video=ordered_video,\n                        quantile=np.array(0.96, dtype=np.float32),\n                        img_mask=True,\n                        ray_mask=False,\n                        camera_only=True,\n                        depth_only=False,\n                        single_view=False,\n                        reset=False,\n                    )\n                )\n\n            if len(views) == num_views and not all(\n                [view[\"instance\"] == views[0][\"instance\"] for view in views]\n            ):\n                invalid_seq = False\n        return views","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dynamic_replica","uri":"program://Human3R/module/src.dust3r.datasets.dynamic_replica#L1-L137","kind":"module","name":"src.dust3r.datasets.dynamic_replica","path":"src/dust3r/datasets/dynamic_replica.py","language":"python","start_line":1,"end_line":137,"context_start_line":1,"context_end_line":137,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DynamicReplica(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scenes = os.listdir(os.path.join(self.ROOT, split))\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, self.split, scene, \"left\")\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=lambda x: float(x),\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id], \"left\")\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.10, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"dynamic_replica\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"c2fa93cf842e3af15ccbc843f14fe72e4be88741e1200ec637cc18bddcc974a3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dynamic_replica.DynamicReplica","uri":"program://Human3R/class/src.dust3r.datasets.dynamic_replica.DynamicReplica#L14-L137","kind":"class","name":"DynamicReplica","path":"src/dust3r/datasets/dynamic_replica.py","language":"python","start_line":14,"end_line":137,"context_start_line":1,"context_end_line":137,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DynamicReplica(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scenes = os.listdir(os.path.join(self.ROOT, split))\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, self.split, scene, \"left\")\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=lambda x: float(x),\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id], \"left\")\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.10, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"dynamic_replica\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"c2fa93cf842e3af15ccbc843f14fe72e4be88741e1200ec637cc18bddcc974a3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dynamic_replica.__init__","uri":"program://Human3R/function/src.dust3r.datasets.dynamic_replica.__init__#L15-L22","kind":"function","name":"__init__","path":"src/dust3r/datasets/dynamic_replica.py","language":"python","start_line":15,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DynamicReplica(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scenes = os.listdir(os.path.join(self.ROOT, split))\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, self.split, scene, \"left\")\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=lambda x: float(x),\n            )\n            num_imgs = len(basenames)","source_hash":"c2fa93cf842e3af15ccbc843f14fe72e4be88741e1200ec637cc18bddcc974a3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dynamic_replica._load_data","uri":"program://Human3R/function/src.dust3r.datasets.dynamic_replica._load_data#L24-L66","kind":"function","name":"_load_data","path":"src/dust3r/datasets/dynamic_replica.py","language":"python","start_line":24,"end_line":66,"context_start_line":4,"context_end_line":86,"code":"import itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DynamicReplica(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data(self.split)\n\n    def _load_data(self, split):\n        self.scenes = os.listdir(os.path.join(self.ROOT, split))\n\n        offset = 0\n        scenes = []\n        sceneids = []\n        scene_img_list = []\n        images = []\n        start_img_ids = []\n\n        j = 0\n        for scene in tqdm(self.scenes):\n            scene_dir = osp.join(self.ROOT, self.split, scene, \"left\")\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            basenames = sorted(\n                [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(\".png\")],\n                key=lambda x: float(x),\n            )\n            num_imgs = len(basenames)\n            img_ids = list(np.arange(num_imgs) + offset)\n            cut_off = (\n                self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)\n            )\n            if num_imgs < cut_off:\n                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]","source_hash":"c2fa93cf842e3af15ccbc843f14fe72e4be88741e1200ec637cc18bddcc974a3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dynamic_replica.__len__","uri":"program://Human3R/function/src.dust3r.datasets.dynamic_replica.__len__#L68-L69","kind":"function","name":"__len__","path":"src/dust3r/datasets/dynamic_replica.py","language":"python","start_line":68,"end_line":69,"context_start_line":48,"context_end_line":89,"code":"                print(f\"Skipping {scene}\")\n                continue\n\n            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):","source_hash":"c2fa93cf842e3af15ccbc843f14fe72e4be88741e1200ec637cc18bddcc974a3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dynamic_replica.get_image_num","uri":"program://Human3R/function/src.dust3r.datasets.dynamic_replica.get_image_num#L71-L72","kind":"function","name":"get_image_num","path":"src/dust3r/datasets/dynamic_replica.py","language":"python","start_line":71,"end_line":72,"context_start_line":51,"context_end_line":92,"code":"            start_img_ids_ = img_ids[: num_imgs - cut_off + 1]\n            start_img_ids.extend(start_img_ids_)\n            sceneids.extend([j] * num_imgs)\n            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id], \"left\")\n            rgb_dir = osp.join(scene_dir, \"rgb\")","source_hash":"c2fa93cf842e3af15ccbc843f14fe72e4be88741e1200ec637cc18bddcc974a3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.dynamic_replica._get_views","uri":"program://Human3R/function/src.dust3r.datasets.dynamic_replica._get_views#L74-L137","kind":"function","name":"_get_views","path":"src/dust3r/datasets/dynamic_replica.py","language":"python","start_line":74,"end_line":137,"context_start_line":54,"context_end_line":137,"code":"            images.extend(basenames)\n            scenes.append(scene)\n            scene_img_list.append(img_ids)\n\n            # offset groups\n            offset += num_imgs\n            j += 1\n\n        self.scenes = scenes\n        self.sceneids = sceneids\n        self.images = images\n        self.start_img_ids = start_img_ids\n        self.scene_img_list = scene_img_list\n\n    def __len__(self):\n        return len(self.start_img_ids)\n\n    def get_image_num(self):\n        return len(self.images)\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        start_id = self.start_img_ids[idx]\n        all_image_ids = self.scene_img_list[self.sceneids[start_id]]\n        pos, ordered_video = self.get_seq_from_start_id(\n            num_views,\n            start_id,\n            all_image_ids,\n            rng,\n            max_interval=self.max_interval,\n            video_prob=1.0,\n            fix_interval_prob=1.0,\n        )\n        image_idxs = np.array(all_image_ids)[pos]\n\n        views = []\n        for v, view_idx in enumerate(image_idxs):\n            scene_id = self.sceneids[view_idx]\n            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id], \"left\")\n            rgb_dir = osp.join(scene_dir, \"rgb\")\n            depth_dir = osp.join(scene_dir, \"depth\")\n            cam_dir = osp.join(scene_dir, \"cam\")\n\n            basename = self.images[view_idx]\n\n            # Load RGB image\n            rgb_image = imread_cv2(osp.join(rgb_dir, basename + \".png\"))\n            # Load depthmap\n            depthmap = np.load(osp.join(depth_dir, basename + \".npy\"))\n            depthmap[~np.isfinite(depthmap)] = 0  # invalid\n\n            cam = np.load(osp.join(cam_dir, basename + \".npz\"))\n            camera_pose = cam[\"pose\"]\n            intrinsics = cam[\"intrinsics\"]\n            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(\n                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx\n            )\n\n            # generate img mask and raymap mask\n            img_mask, ray_mask = self.get_img_and_ray_masks(\n                self.is_metric, v, rng, p=[0.85, 0.10, 0.05]\n            )\n\n            views.append(\n                dict(\n                    img=rgb_image,\n                    depthmap=depthmap.astype(np.float32),\n                    camera_pose=camera_pose.astype(np.float32),\n                    camera_intrinsics=intrinsics.astype(np.float32),\n                    dataset=\"dynamic_replica\",\n                    label=self.scenes[scene_id] + \"_\" + basename,\n                    instance=f\"{str(idx)}_{str(view_idx)}\",\n                    is_metric=self.is_metric,\n                    is_video=ordered_video,\n                    quantile=np.array(1.0, dtype=np.float32),\n                    img_mask=img_mask,\n                    ray_mask=ray_mask,\n                    camera_only=False,\n                    depth_only=False,\n                    single_view=False,\n                    reset=False,\n                )\n            )\n        assert len(views) == num_views\n        return views","source_hash":"c2fa93cf842e3af15ccbc843f14fe72e4be88741e1200ec637cc18bddcc974a3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.corr","uri":"program://Human3R/module/src.dust3r.datasets.utils.corr#L1-L129","kind":"module","name":"src.dust3r.datasets.utils.corr","path":"src/dust3r/datasets/utils/corr.py","language":"python","start_line":1,"end_line":129,"context_start_line":1,"context_end_line":129,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nfrom dust3r.utils.device import to_numpy\nfrom dust3r.utils.geometry import inv, geotrf\n\n\ndef reproject_view(pts3d, view2):\n    shape = view2[\"pts3d\"].shape[:2]\n    return reproject(\n        pts3d, view2[\"camera_intrinsics\"], inv(view2[\"camera_pose\"]), shape\n    )\n\n\ndef reproject(pts3d, K, world2cam, shape):\n    H, W, THREE = pts3d.shape\n    assert THREE == 3\n\n    # reproject in camera2 space\n    with np.errstate(divide=\"ignore\", invalid=\"ignore\"):\n        pos = geotrf(K @ world2cam[:3], pts3d, norm=1, ncol=2)\n\n    # quantize to pixel positions\n    return (H, W), ravel_xy(pos, shape)\n\n\ndef ravel_xy(pos, shape):\n    H, W = shape\n    with np.errstate(invalid=\"ignore\"):\n        qx, qy = pos.reshape(-1, 2).round().astype(np.int32).T\n    quantized_pos = qx.clip(min=0, max=W - 1, out=qx) + W * qy.clip(\n        min=0, max=H - 1, out=qy\n    )\n    return quantized_pos\n\n\ndef unravel_xy(pos, shape):\n    # convert (x+W*y) back to 2d (x,y) coordinates\n    return np.unravel_index(pos, shape)[0].base[:, ::-1].copy()\n\n\ndef reciprocal_1d(corres_1_to_2, corres_2_to_1, ret_recip=False):\n    is_reciprocal1 = corres_2_to_1[corres_1_to_2] == np.arange(len(corres_1_to_2))\n    pos1 = is_reciprocal1.nonzero()[0]\n    pos2 = corres_1_to_2[pos1]\n    if ret_recip:\n        return is_reciprocal1, pos1, pos2\n    return pos1, pos2\n\n\ndef extract_correspondences_from_pts3d(\n    view1, view2, target_n_corres, rng=np.random, ret_xy=True, nneg=0\n):\n    view1, view2 = to_numpy((view1, view2))\n    # project pixels from image1 --> 3d points --> image2 pixels\n    shape1, corres1_to_2 = reproject_view(view1[\"pts3d\"], view2)\n    shape2, corres2_to_1 = reproject_view(view2[\"pts3d\"], view1)\n\n    # compute reciprocal correspondences:\n    # pos1 == valid pixels (correspondences) in image1\n    is_reciprocal1, pos1, pos2 = reciprocal_1d(\n        corres1_to_2, corres2_to_1, ret_recip=True\n    )\n    is_reciprocal2 = corres1_to_2[corres2_to_1] == np.arange(len(corres2_to_1))\n\n    if target_n_corres is None:\n        if ret_xy:\n            pos1 = unravel_xy(pos1, shape1)\n            pos2 = unravel_xy(pos2, shape2)\n        return pos1, pos2\n\n    available_negatives = min((~is_reciprocal1).sum(), (~is_reciprocal2).sum())\n    target_n_positives = int(target_n_corres * (1 - nneg))\n    n_positives = min(len(pos1), target_n_positives)\n    n_negatives = min(target_n_corres - n_positives, available_negatives)\n\n    if n_negatives + n_positives != target_n_corres:\n        # should be really rare => when there are not enough negatives\n        # in that case, break nneg and add a few more positives ?\n        n_positives = target_n_corres - n_negatives\n        assert n_positives <= len(pos1)\n\n    assert n_positives <= len(pos1)\n    assert n_positives <= len(pos2)\n    assert n_negatives <= (~is_reciprocal1).sum()\n    assert n_negatives <= (~is_reciprocal2).sum()\n    assert n_positives + n_negatives == target_n_corres\n\n    valid = np.ones(n_positives, dtype=bool)\n    if n_positives < len(pos1):\n        # random sub-sampling of valid correspondences\n        perm = rng.permutation(len(pos1))[:n_positives]\n        pos1 = pos1[perm]\n        pos2 = pos2[perm]\n\n    if n_negatives > 0:\n        # add false correspondences if not enough\n        def norm(p):\n            return p / p.sum()\n\n        pos1 = np.r_[\n            pos1,\n            rng.choice(\n                shape1[0] * shape1[1],\n                size=n_negatives,\n                replace=False,\n                p=norm(~is_reciprocal1),\n            ),\n        ]\n        pos2 = np.r_[\n            pos2,\n            rng.choice(\n                shape2[0] * shape2[1],\n                size=n_negatives,\n                replace=False,\n                p=norm(~is_reciprocal2),\n            ),\n        ]\n        valid = np.r_[valid, np.zeros(n_negatives, dtype=bool)]\n\n    # convert (x+W*y) back to 2d (x,y) coordinates\n    if ret_xy:\n        pos1 = unravel_xy(pos1, shape1)\n        pos2 = unravel_xy(pos2, shape2)\n    return pos1, pos2, valid","source_hash":"6195285c9dc0c477a03c0d7624a068d0b3dd52386da09aa9a74e1c4c466bdef8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.corr.reproject_view","uri":"program://Human3R/function/src.dust3r.datasets.utils.corr.reproject_view#L12-L16","kind":"function","name":"reproject_view","path":"src/dust3r/datasets/utils/corr.py","language":"python","start_line":12,"end_line":16,"context_start_line":1,"context_end_line":36,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nfrom dust3r.utils.device import to_numpy\nfrom dust3r.utils.geometry import inv, geotrf\n\n\ndef reproject_view(pts3d, view2):\n    shape = view2[\"pts3d\"].shape[:2]\n    return reproject(\n        pts3d, view2[\"camera_intrinsics\"], inv(view2[\"camera_pose\"]), shape\n    )\n\n\ndef reproject(pts3d, K, world2cam, shape):\n    H, W, THREE = pts3d.shape\n    assert THREE == 3\n\n    # reproject in camera2 space\n    with np.errstate(divide=\"ignore\", invalid=\"ignore\"):\n        pos = geotrf(K @ world2cam[:3], pts3d, norm=1, ncol=2)\n\n    # quantize to pixel positions\n    return (H, W), ravel_xy(pos, shape)\n\n\ndef ravel_xy(pos, shape):\n    H, W = shape\n    with np.errstate(invalid=\"ignore\"):\n        qx, qy = pos.reshape(-1, 2).round().astype(np.int32).T\n    quantized_pos = qx.clip(min=0, max=W - 1, out=qx) + W * qy.clip(\n        min=0, max=H - 1, out=qy","source_hash":"6195285c9dc0c477a03c0d7624a068d0b3dd52386da09aa9a74e1c4c466bdef8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.corr.reproject","uri":"program://Human3R/function/src.dust3r.datasets.utils.corr.reproject#L19-L28","kind":"function","name":"reproject","path":"src/dust3r/datasets/utils/corr.py","language":"python","start_line":19,"end_line":28,"context_start_line":1,"context_end_line":48,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nfrom dust3r.utils.device import to_numpy\nfrom dust3r.utils.geometry import inv, geotrf\n\n\ndef reproject_view(pts3d, view2):\n    shape = view2[\"pts3d\"].shape[:2]\n    return reproject(\n        pts3d, view2[\"camera_intrinsics\"], inv(view2[\"camera_pose\"]), shape\n    )\n\n\ndef reproject(pts3d, K, world2cam, shape):\n    H, W, THREE = pts3d.shape\n    assert THREE == 3\n\n    # reproject in camera2 space\n    with np.errstate(divide=\"ignore\", invalid=\"ignore\"):\n        pos = geotrf(K @ world2cam[:3], pts3d, norm=1, ncol=2)\n\n    # quantize to pixel positions\n    return (H, W), ravel_xy(pos, shape)\n\n\ndef ravel_xy(pos, shape):\n    H, W = shape\n    with np.errstate(invalid=\"ignore\"):\n        qx, qy = pos.reshape(-1, 2).round().astype(np.int32).T\n    quantized_pos = qx.clip(min=0, max=W - 1, out=qx) + W * qy.clip(\n        min=0, max=H - 1, out=qy\n    )\n    return quantized_pos\n\n\ndef unravel_xy(pos, shape):\n    # convert (x+W*y) back to 2d (x,y) coordinates\n    return np.unravel_index(pos, shape)[0].base[:, ::-1].copy()\n\n\ndef reciprocal_1d(corres_1_to_2, corres_2_to_1, ret_recip=False):\n    is_reciprocal1 = corres_2_to_1[corres_1_to_2] == np.arange(len(corres_1_to_2))\n    pos1 = is_reciprocal1.nonzero()[0]","source_hash":"6195285c9dc0c477a03c0d7624a068d0b3dd52386da09aa9a74e1c4c466bdef8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.corr.ravel_xy","uri":"program://Human3R/function/src.dust3r.datasets.utils.corr.ravel_xy#L31-L38","kind":"function","name":"ravel_xy","path":"src/dust3r/datasets/utils/corr.py","language":"python","start_line":31,"end_line":38,"context_start_line":11,"context_end_line":58,"code":"\ndef reproject_view(pts3d, view2):\n    shape = view2[\"pts3d\"].shape[:2]\n    return reproject(\n        pts3d, view2[\"camera_intrinsics\"], inv(view2[\"camera_pose\"]), shape\n    )\n\n\ndef reproject(pts3d, K, world2cam, shape):\n    H, W, THREE = pts3d.shape\n    assert THREE == 3\n\n    # reproject in camera2 space\n    with np.errstate(divide=\"ignore\", invalid=\"ignore\"):\n        pos = geotrf(K @ world2cam[:3], pts3d, norm=1, ncol=2)\n\n    # quantize to pixel positions\n    return (H, W), ravel_xy(pos, shape)\n\n\ndef ravel_xy(pos, shape):\n    H, W = shape\n    with np.errstate(invalid=\"ignore\"):\n        qx, qy = pos.reshape(-1, 2).round().astype(np.int32).T\n    quantized_pos = qx.clip(min=0, max=W - 1, out=qx) + W * qy.clip(\n        min=0, max=H - 1, out=qy\n    )\n    return quantized_pos\n\n\ndef unravel_xy(pos, shape):\n    # convert (x+W*y) back to 2d (x,y) coordinates\n    return np.unravel_index(pos, shape)[0].base[:, ::-1].copy()\n\n\ndef reciprocal_1d(corres_1_to_2, corres_2_to_1, ret_recip=False):\n    is_reciprocal1 = corres_2_to_1[corres_1_to_2] == np.arange(len(corres_1_to_2))\n    pos1 = is_reciprocal1.nonzero()[0]\n    pos2 = corres_1_to_2[pos1]\n    if ret_recip:\n        return is_reciprocal1, pos1, pos2\n    return pos1, pos2\n\n\ndef extract_correspondences_from_pts3d(\n    view1, view2, target_n_corres, rng=np.random, ret_xy=True, nneg=0\n):\n    view1, view2 = to_numpy((view1, view2))","source_hash":"6195285c9dc0c477a03c0d7624a068d0b3dd52386da09aa9a74e1c4c466bdef8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.corr.unravel_xy","uri":"program://Human3R/function/src.dust3r.datasets.utils.corr.unravel_xy#L41-L43","kind":"function","name":"unravel_xy","path":"src/dust3r/datasets/utils/corr.py","language":"python","start_line":41,"end_line":43,"context_start_line":21,"context_end_line":63,"code":"    assert THREE == 3\n\n    # reproject in camera2 space\n    with np.errstate(divide=\"ignore\", invalid=\"ignore\"):\n        pos = geotrf(K @ world2cam[:3], pts3d, norm=1, ncol=2)\n\n    # quantize to pixel positions\n    return (H, W), ravel_xy(pos, shape)\n\n\ndef ravel_xy(pos, shape):\n    H, W = shape\n    with np.errstate(invalid=\"ignore\"):\n        qx, qy = pos.reshape(-1, 2).round().astype(np.int32).T\n    quantized_pos = qx.clip(min=0, max=W - 1, out=qx) + W * qy.clip(\n        min=0, max=H - 1, out=qy\n    )\n    return quantized_pos\n\n\ndef unravel_xy(pos, shape):\n    # convert (x+W*y) back to 2d (x,y) coordinates\n    return np.unravel_index(pos, shape)[0].base[:, ::-1].copy()\n\n\ndef reciprocal_1d(corres_1_to_2, corres_2_to_1, ret_recip=False):\n    is_reciprocal1 = corres_2_to_1[corres_1_to_2] == np.arange(len(corres_1_to_2))\n    pos1 = is_reciprocal1.nonzero()[0]\n    pos2 = corres_1_to_2[pos1]\n    if ret_recip:\n        return is_reciprocal1, pos1, pos2\n    return pos1, pos2\n\n\ndef extract_correspondences_from_pts3d(\n    view1, view2, target_n_corres, rng=np.random, ret_xy=True, nneg=0\n):\n    view1, view2 = to_numpy((view1, view2))\n    # project pixels from image1 --> 3d points --> image2 pixels\n    shape1, corres1_to_2 = reproject_view(view1[\"pts3d\"], view2)\n    shape2, corres2_to_1 = reproject_view(view2[\"pts3d\"], view1)\n\n    # compute reciprocal correspondences:","source_hash":"6195285c9dc0c477a03c0d7624a068d0b3dd52386da09aa9a74e1c4c466bdef8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.corr.reciprocal_1d","uri":"program://Human3R/function/src.dust3r.datasets.utils.corr.reciprocal_1d#L46-L52","kind":"function","name":"reciprocal_1d","path":"src/dust3r/datasets/utils/corr.py","language":"python","start_line":46,"end_line":52,"context_start_line":26,"context_end_line":72,"code":"\n    # quantize to pixel positions\n    return (H, W), ravel_xy(pos, shape)\n\n\ndef ravel_xy(pos, shape):\n    H, W = shape\n    with np.errstate(invalid=\"ignore\"):\n        qx, qy = pos.reshape(-1, 2).round().astype(np.int32).T\n    quantized_pos = qx.clip(min=0, max=W - 1, out=qx) + W * qy.clip(\n        min=0, max=H - 1, out=qy\n    )\n    return quantized_pos\n\n\ndef unravel_xy(pos, shape):\n    # convert (x+W*y) back to 2d (x,y) coordinates\n    return np.unravel_index(pos, shape)[0].base[:, ::-1].copy()\n\n\ndef reciprocal_1d(corres_1_to_2, corres_2_to_1, ret_recip=False):\n    is_reciprocal1 = corres_2_to_1[corres_1_to_2] == np.arange(len(corres_1_to_2))\n    pos1 = is_reciprocal1.nonzero()[0]\n    pos2 = corres_1_to_2[pos1]\n    if ret_recip:\n        return is_reciprocal1, pos1, pos2\n    return pos1, pos2\n\n\ndef extract_correspondences_from_pts3d(\n    view1, view2, target_n_corres, rng=np.random, ret_xy=True, nneg=0\n):\n    view1, view2 = to_numpy((view1, view2))\n    # project pixels from image1 --> 3d points --> image2 pixels\n    shape1, corres1_to_2 = reproject_view(view1[\"pts3d\"], view2)\n    shape2, corres2_to_1 = reproject_view(view2[\"pts3d\"], view1)\n\n    # compute reciprocal correspondences:\n    # pos1 == valid pixels (correspondences) in image1\n    is_reciprocal1, pos1, pos2 = reciprocal_1d(\n        corres1_to_2, corres2_to_1, ret_recip=True\n    )\n    is_reciprocal2 = corres1_to_2[corres2_to_1] == np.arange(len(corres2_to_1))\n\n    if target_n_corres is None:\n        if ret_xy:\n            pos1 = unravel_xy(pos1, shape1)","source_hash":"6195285c9dc0c477a03c0d7624a068d0b3dd52386da09aa9a74e1c4c466bdef8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.corr.extract_correspondences_from_pts3d","uri":"program://Human3R/function/src.dust3r.datasets.utils.corr.extract_correspondences_from_pts3d#L55-L129","kind":"function","name":"extract_correspondences_from_pts3d","path":"src/dust3r/datasets/utils/corr.py","language":"python","start_line":55,"end_line":129,"context_start_line":35,"context_end_line":129,"code":"    quantized_pos = qx.clip(min=0, max=W - 1, out=qx) + W * qy.clip(\n        min=0, max=H - 1, out=qy\n    )\n    return quantized_pos\n\n\ndef unravel_xy(pos, shape):\n    # convert (x+W*y) back to 2d (x,y) coordinates\n    return np.unravel_index(pos, shape)[0].base[:, ::-1].copy()\n\n\ndef reciprocal_1d(corres_1_to_2, corres_2_to_1, ret_recip=False):\n    is_reciprocal1 = corres_2_to_1[corres_1_to_2] == np.arange(len(corres_1_to_2))\n    pos1 = is_reciprocal1.nonzero()[0]\n    pos2 = corres_1_to_2[pos1]\n    if ret_recip:\n        return is_reciprocal1, pos1, pos2\n    return pos1, pos2\n\n\ndef extract_correspondences_from_pts3d(\n    view1, view2, target_n_corres, rng=np.random, ret_xy=True, nneg=0\n):\n    view1, view2 = to_numpy((view1, view2))\n    # project pixels from image1 --> 3d points --> image2 pixels\n    shape1, corres1_to_2 = reproject_view(view1[\"pts3d\"], view2)\n    shape2, corres2_to_1 = reproject_view(view2[\"pts3d\"], view1)\n\n    # compute reciprocal correspondences:\n    # pos1 == valid pixels (correspondences) in image1\n    is_reciprocal1, pos1, pos2 = reciprocal_1d(\n        corres1_to_2, corres2_to_1, ret_recip=True\n    )\n    is_reciprocal2 = corres1_to_2[corres2_to_1] == np.arange(len(corres2_to_1))\n\n    if target_n_corres is None:\n        if ret_xy:\n            pos1 = unravel_xy(pos1, shape1)\n            pos2 = unravel_xy(pos2, shape2)\n        return pos1, pos2\n\n    available_negatives = min((~is_reciprocal1).sum(), (~is_reciprocal2).sum())\n    target_n_positives = int(target_n_corres * (1 - nneg))\n    n_positives = min(len(pos1), target_n_positives)\n    n_negatives = min(target_n_corres - n_positives, available_negatives)\n\n    if n_negatives + n_positives != target_n_corres:\n        # should be really rare => when there are not enough negatives\n        # in that case, break nneg and add a few more positives ?\n        n_positives = target_n_corres - n_negatives\n        assert n_positives <= len(pos1)\n\n    assert n_positives <= len(pos1)\n    assert n_positives <= len(pos2)\n    assert n_negatives <= (~is_reciprocal1).sum()\n    assert n_negatives <= (~is_reciprocal2).sum()\n    assert n_positives + n_negatives == target_n_corres\n\n    valid = np.ones(n_positives, dtype=bool)\n    if n_positives < len(pos1):\n        # random sub-sampling of valid correspondences\n        perm = rng.permutation(len(pos1))[:n_positives]\n        pos1 = pos1[perm]\n        pos2 = pos2[perm]\n\n    if n_negatives > 0:\n        # add false correspondences if not enough\n        def norm(p):\n            return p / p.sum()\n\n        pos1 = np.r_[\n            pos1,\n            rng.choice(\n                shape1[0] * shape1[1],\n                size=n_negatives,\n                replace=False,\n                p=norm(~is_reciprocal1),\n            ),\n        ]\n        pos2 = np.r_[\n            pos2,\n            rng.choice(\n                shape2[0] * shape2[1],\n                size=n_negatives,\n                replace=False,\n                p=norm(~is_reciprocal2),\n            ),\n        ]\n        valid = np.r_[valid, np.zeros(n_negatives, dtype=bool)]\n\n    # convert (x+W*y) back to 2d (x,y) coordinates\n    if ret_xy:\n        pos1 = unravel_xy(pos1, shape1)\n        pos2 = unravel_xy(pos2, shape2)\n    return pos1, pos2, valid","source_hash":"6195285c9dc0c477a03c0d7624a068d0b3dd52386da09aa9a74e1c4c466bdef8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.corr.norm","uri":"program://Human3R/function/src.dust3r.datasets.utils.corr.norm#L102-L103","kind":"function","name":"norm","path":"src/dust3r/datasets/utils/corr.py","language":"python","start_line":102,"end_line":103,"context_start_line":82,"context_end_line":123,"code":"        # should be really rare => when there are not enough negatives\n        # in that case, break nneg and add a few more positives ?\n        n_positives = target_n_corres - n_negatives\n        assert n_positives <= len(pos1)\n\n    assert n_positives <= len(pos1)\n    assert n_positives <= len(pos2)\n    assert n_negatives <= (~is_reciprocal1).sum()\n    assert n_negatives <= (~is_reciprocal2).sum()\n    assert n_positives + n_negatives == target_n_corres\n\n    valid = np.ones(n_positives, dtype=bool)\n    if n_positives < len(pos1):\n        # random sub-sampling of valid correspondences\n        perm = rng.permutation(len(pos1))[:n_positives]\n        pos1 = pos1[perm]\n        pos2 = pos2[perm]\n\n    if n_negatives > 0:\n        # add false correspondences if not enough\n        def norm(p):\n            return p / p.sum()\n\n        pos1 = np.r_[\n            pos1,\n            rng.choice(\n                shape1[0] * shape1[1],\n                size=n_negatives,\n                replace=False,\n                p=norm(~is_reciprocal1),\n            ),\n        ]\n        pos2 = np.r_[\n            pos2,\n            rng.choice(\n                shape2[0] * shape2[1],\n                size=n_negatives,\n                replace=False,\n                p=norm(~is_reciprocal2),\n            ),\n        ]\n        valid = np.r_[valid, np.zeros(n_negatives, dtype=bool)]","source_hash":"6195285c9dc0c477a03c0d7624a068d0b3dd52386da09aa9a74e1c4c466bdef8","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping","uri":"program://Human3R/module/src.dust3r.datasets.utils.cropping#L1-L211","kind":"module","name":"src.dust3r.datasets.utils.cropping","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":1,"end_line":211,"context_start_line":1,"context_end_line":211,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# croppping utilities\n# --------------------------------------------------------\nimport PIL.Image\nimport os\n\nos.environ[\"OPENCV_IO_ENABLE_OPENEXR\"] = \"1\"\nimport cv2  # noqa\nimport numpy as np  # noqa\nfrom dust3r.utils.geometry import (\n    colmap_to_opencv_intrinsics,\n    opencv_to_colmap_intrinsics,\n)  # noqa\n\ntry:\n    lanczos = PIL.Image.Resampling.LANCZOS\n    bicubic = PIL.Image.Resampling.BICUBIC\nexcept AttributeError:\n    lanczos = PIL.Image.LANCZOS\n    bicubic = PIL.Image.BICUBIC\n\n\nclass ImageList:\n    \"\"\"Convenience class to aply the same operation to a whole set of images.\"\"\"\n\n    def __init__(self, images):\n        if not isinstance(images, (tuple, list, set)):\n            images = [images]\n        self.images = []\n        for image in images:\n            if not isinstance(image, PIL.Image.Image):\n                image = PIL.Image.fromarray(image)\n            self.images.append(image)\n\n    def __len__(self):\n        return len(self.images)\n\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):\n        return [getattr(im, func)(*args, **kwargs) for im in self.images]\n\n\ndef rescale_image_depthmap(\n    image, depthmap, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)\n    if depthmap is not None:\n        # can also use this with masks instead of depthmaps\n        assert tuple(depthmap.shape[:2]) == image.size[::-1]\n\n    # define output resolution\n    assert output_resolution.shape == (2,)\n    scale_final = max(output_resolution / image.size) + 1e-8\n    if scale_final >= 1 and not force:  # image is already smaller than what is asked\n        return (image.to_pil(), depthmap, camera_intrinsics)\n    output_resolution = np.floor(input_resolution * scale_final).astype(int)\n\n    # first rescale the image so that it contains the crop\n    image = image.resize(\n        output_resolution, resample=lanczos if scale_final < 1 else bicubic\n    )\n    if depthmap is not None:\n        depthmap = cv2.resize(\n            depthmap,\n            output_resolution,\n            fx=scale_final,\n            fy=scale_final,\n            interpolation=cv2.INTER_NEAREST,\n        )\n\n    # no offset here; simple rescaling\n    camera_intrinsics = camera_matrix_of_crop(\n        camera_intrinsics, input_resolution, output_resolution, scaling=scale_final\n    )\n\n    return image.to_pil(), depthmap, camera_intrinsics\n\n\ndef rescale_image_depthmap_mask(\n    image, depthmap, mask, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap, mask)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    mask = ImageList(mask)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)\n    if depthmap is not None:\n        # can also use this with masks instead of depthmaps\n        assert tuple(depthmap.shape[:2]) == image.size[::-1]\n\n    # define output resolution\n    assert output_resolution.shape == (2,)\n    scale_final = max(output_resolution / image.size) + 1e-8\n    if scale_final >= 1 and not force:  # image is already smaller than what is asked\n        return (image.to_pil(), depthmap, mask.to_pil(), camera_intrinsics)\n    output_resolution = np.floor(input_resolution * scale_final).astype(int)\n\n    # first rescale the image so that it contains the crop\n    image = image.resize(\n        output_resolution, resample=lanczos if scale_final < 1 else bicubic\n    )\n    if depthmap is not None:\n        depthmap = cv2.resize(\n            depthmap,\n            output_resolution,\n            fx=scale_final,\n            fy=scale_final,\n            interpolation=cv2.INTER_NEAREST,\n        )\n    if mask is not None:\n        mask = mask.resize(output_resolution, resample=PIL.Image.NEAREST)\n\n    # no offset here; simple rescaling\n    camera_intrinsics = camera_matrix_of_crop(\n        camera_intrinsics, input_resolution, output_resolution, scaling=scale_final\n    )\n\n    return image.to_pil(), depthmap, mask.to_pil(), camera_intrinsics\n\n\ndef camera_matrix_of_crop(\n    input_camera_matrix,\n    input_resolution,\n    output_resolution,\n    scaling=1,\n    offset_factor=0.5,\n    offset=None,\n):\n    # Margins to offset the origin\n    margins = np.asarray(input_resolution) * scaling - output_resolution\n    assert np.all(margins >= 0.0)\n    if offset is None:\n        offset = offset_factor * margins\n\n    # Generate new camera parameters\n    output_camera_matrix_colmap = opencv_to_colmap_intrinsics(input_camera_matrix)\n    output_camera_matrix_colmap[:2, :] *= scaling\n    output_camera_matrix_colmap[:2, 2] -= offset\n    output_camera_matrix = colmap_to_opencv_intrinsics(output_camera_matrix_colmap)\n\n    return output_camera_matrix\n\n\ndef crop_image_depthmap(image, depthmap, camera_intrinsics, crop_bbox):\n    \"\"\"\n    Return a crop of the input view.\n    \"\"\"\n    image = ImageList(image)\n    l, t, r, b = crop_bbox\n\n    image = image.crop((l, t, r, b))\n    if depthmap is not None:\n        depthmap = depthmap[t:b, l:r]\n\n    camera_intrinsics = camera_intrinsics.copy()\n    camera_intrinsics[0, 2] -= l\n    camera_intrinsics[1, 2] -= t\n\n    return image.to_pil(), depthmap, camera_intrinsics\n\n\ndef crop_image_depthmap_mask(image, depthmap, mask, camera_intrinsics, crop_bbox):\n    \"\"\"\n    Return a crop of the input view including mask.\n    \"\"\"\n    image = ImageList(image)\n    mask = ImageList(mask)\n    l, t, r, b = crop_bbox\n\n    image = image.crop((l, t, r, b))\n    mask = mask.crop((l, t, r, b))\n    depthmap = depthmap[t:b, l:r]\n\n    camera_intrinsics = camera_intrinsics.copy()\n    camera_intrinsics[0, 2] -= l\n    camera_intrinsics[1, 2] -= t\n\n    return image.to_pil(), depthmap, mask.to_pil(), camera_intrinsics\n\n\ndef bbox_from_intrinsics_in_out(\n    input_camera_matrix, output_camera_matrix, output_resolution\n):\n    out_width, out_height = output_resolution\n    l, t = np.int32(np.round(input_camera_matrix[:2, 2] - output_camera_matrix[:2, 2]))\n    crop_bbox = (l, t, l + out_width, t + out_height)\n    return crop_bbox","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.ImageList","uri":"program://Human3R/class/src.dust3r.datasets.utils.cropping.ImageList#L26-L57","kind":"class","name":"ImageList","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":26,"end_line":57,"context_start_line":6,"context_end_line":77,"code":"# --------------------------------------------------------\nimport PIL.Image\nimport os\n\nos.environ[\"OPENCV_IO_ENABLE_OPENEXR\"] = \"1\"\nimport cv2  # noqa\nimport numpy as np  # noqa\nfrom dust3r.utils.geometry import (\n    colmap_to_opencv_intrinsics,\n    opencv_to_colmap_intrinsics,\n)  # noqa\n\ntry:\n    lanczos = PIL.Image.Resampling.LANCZOS\n    bicubic = PIL.Image.Resampling.BICUBIC\nexcept AttributeError:\n    lanczos = PIL.Image.LANCZOS\n    bicubic = PIL.Image.BICUBIC\n\n\nclass ImageList:\n    \"\"\"Convenience class to aply the same operation to a whole set of images.\"\"\"\n\n    def __init__(self, images):\n        if not isinstance(images, (tuple, list, set)):\n            images = [images]\n        self.images = []\n        for image in images:\n            if not isinstance(image, PIL.Image.Image):\n                image = PIL.Image.fromarray(image)\n            self.images.append(image)\n\n    def __len__(self):\n        return len(self.images)\n\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):\n        return [getattr(im, func)(*args, **kwargs) for im in self.images]\n\n\ndef rescale_image_depthmap(\n    image, depthmap, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)\n    if depthmap is not None:\n        # can also use this with masks instead of depthmaps\n        assert tuple(depthmap.shape[:2]) == image.size[::-1]\n\n    # define output resolution\n    assert output_resolution.shape == (2,)\n    scale_final = max(output_resolution / image.size) + 1e-8\n    if scale_final >= 1 and not force:  # image is already smaller than what is asked\n        return (image.to_pil(), depthmap, camera_intrinsics)","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.rescale_image_depthmap","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.rescale_image_depthmap#L60-L98","kind":"function","name":"rescale_image_depthmap","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":60,"end_line":98,"context_start_line":40,"context_end_line":118,"code":"\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):\n        return [getattr(im, func)(*args, **kwargs) for im in self.images]\n\n\ndef rescale_image_depthmap(\n    image, depthmap, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)\n    if depthmap is not None:\n        # can also use this with masks instead of depthmaps\n        assert tuple(depthmap.shape[:2]) == image.size[::-1]\n\n    # define output resolution\n    assert output_resolution.shape == (2,)\n    scale_final = max(output_resolution / image.size) + 1e-8\n    if scale_final >= 1 and not force:  # image is already smaller than what is asked\n        return (image.to_pil(), depthmap, camera_intrinsics)\n    output_resolution = np.floor(input_resolution * scale_final).astype(int)\n\n    # first rescale the image so that it contains the crop\n    image = image.resize(\n        output_resolution, resample=lanczos if scale_final < 1 else bicubic\n    )\n    if depthmap is not None:\n        depthmap = cv2.resize(\n            depthmap,\n            output_resolution,\n            fx=scale_final,\n            fy=scale_final,\n            interpolation=cv2.INTER_NEAREST,\n        )\n\n    # no offset here; simple rescaling\n    camera_intrinsics = camera_matrix_of_crop(\n        camera_intrinsics, input_resolution, output_resolution, scaling=scale_final\n    )\n\n    return image.to_pil(), depthmap, camera_intrinsics\n\n\ndef rescale_image_depthmap_mask(\n    image, depthmap, mask, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap, mask)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    mask = ImageList(mask)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)\n    if depthmap is not None:\n        # can also use this with masks instead of depthmaps\n        assert tuple(depthmap.shape[:2]) == image.size[::-1]\n\n    # define output resolution\n    assert output_resolution.shape == (2,)\n    scale_final = max(output_resolution / image.size) + 1e-8\n    if scale_final >= 1 and not force:  # image is already smaller than what is asked","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.rescale_image_depthmap_mask","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.rescale_image_depthmap_mask#L101-L142","kind":"function","name":"rescale_image_depthmap_mask","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":101,"end_line":142,"context_start_line":81,"context_end_line":162,"code":"    image = image.resize(\n        output_resolution, resample=lanczos if scale_final < 1 else bicubic\n    )\n    if depthmap is not None:\n        depthmap = cv2.resize(\n            depthmap,\n            output_resolution,\n            fx=scale_final,\n            fy=scale_final,\n            interpolation=cv2.INTER_NEAREST,\n        )\n\n    # no offset here; simple rescaling\n    camera_intrinsics = camera_matrix_of_crop(\n        camera_intrinsics, input_resolution, output_resolution, scaling=scale_final\n    )\n\n    return image.to_pil(), depthmap, camera_intrinsics\n\n\ndef rescale_image_depthmap_mask(\n    image, depthmap, mask, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap, mask)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    mask = ImageList(mask)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)\n    if depthmap is not None:\n        # can also use this with masks instead of depthmaps\n        assert tuple(depthmap.shape[:2]) == image.size[::-1]\n\n    # define output resolution\n    assert output_resolution.shape == (2,)\n    scale_final = max(output_resolution / image.size) + 1e-8\n    if scale_final >= 1 and not force:  # image is already smaller than what is asked\n        return (image.to_pil(), depthmap, mask.to_pil(), camera_intrinsics)\n    output_resolution = np.floor(input_resolution * scale_final).astype(int)\n\n    # first rescale the image so that it contains the crop\n    image = image.resize(\n        output_resolution, resample=lanczos if scale_final < 1 else bicubic\n    )\n    if depthmap is not None:\n        depthmap = cv2.resize(\n            depthmap,\n            output_resolution,\n            fx=scale_final,\n            fy=scale_final,\n            interpolation=cv2.INTER_NEAREST,\n        )\n    if mask is not None:\n        mask = mask.resize(output_resolution, resample=PIL.Image.NEAREST)\n\n    # no offset here; simple rescaling\n    camera_intrinsics = camera_matrix_of_crop(\n        camera_intrinsics, input_resolution, output_resolution, scaling=scale_final\n    )\n\n    return image.to_pil(), depthmap, mask.to_pil(), camera_intrinsics\n\n\ndef camera_matrix_of_crop(\n    input_camera_matrix,\n    input_resolution,\n    output_resolution,\n    scaling=1,\n    offset_factor=0.5,\n    offset=None,\n):\n    # Margins to offset the origin\n    margins = np.asarray(input_resolution) * scaling - output_resolution\n    assert np.all(margins >= 0.0)\n    if offset is None:\n        offset = offset_factor * margins\n\n    # Generate new camera parameters\n    output_camera_matrix_colmap = opencv_to_colmap_intrinsics(input_camera_matrix)\n    output_camera_matrix_colmap[:2, :] *= scaling\n    output_camera_matrix_colmap[:2, 2] -= offset","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.camera_matrix_of_crop","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.camera_matrix_of_crop#L145-L165","kind":"function","name":"camera_matrix_of_crop","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":145,"end_line":165,"context_start_line":125,"context_end_line":185,"code":"    )\n    if depthmap is not None:\n        depthmap = cv2.resize(\n            depthmap,\n            output_resolution,\n            fx=scale_final,\n            fy=scale_final,\n            interpolation=cv2.INTER_NEAREST,\n        )\n    if mask is not None:\n        mask = mask.resize(output_resolution, resample=PIL.Image.NEAREST)\n\n    # no offset here; simple rescaling\n    camera_intrinsics = camera_matrix_of_crop(\n        camera_intrinsics, input_resolution, output_resolution, scaling=scale_final\n    )\n\n    return image.to_pil(), depthmap, mask.to_pil(), camera_intrinsics\n\n\ndef camera_matrix_of_crop(\n    input_camera_matrix,\n    input_resolution,\n    output_resolution,\n    scaling=1,\n    offset_factor=0.5,\n    offset=None,\n):\n    # Margins to offset the origin\n    margins = np.asarray(input_resolution) * scaling - output_resolution\n    assert np.all(margins >= 0.0)\n    if offset is None:\n        offset = offset_factor * margins\n\n    # Generate new camera parameters\n    output_camera_matrix_colmap = opencv_to_colmap_intrinsics(input_camera_matrix)\n    output_camera_matrix_colmap[:2, :] *= scaling\n    output_camera_matrix_colmap[:2, 2] -= offset\n    output_camera_matrix = colmap_to_opencv_intrinsics(output_camera_matrix_colmap)\n\n    return output_camera_matrix\n\n\ndef crop_image_depthmap(image, depthmap, camera_intrinsics, crop_bbox):\n    \"\"\"\n    Return a crop of the input view.\n    \"\"\"\n    image = ImageList(image)\n    l, t, r, b = crop_bbox\n\n    image = image.crop((l, t, r, b))\n    if depthmap is not None:\n        depthmap = depthmap[t:b, l:r]\n\n    camera_intrinsics = camera_intrinsics.copy()\n    camera_intrinsics[0, 2] -= l\n    camera_intrinsics[1, 2] -= t\n\n    return image.to_pil(), depthmap, camera_intrinsics\n\n","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.crop_image_depthmap","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.crop_image_depthmap#L168-L183","kind":"function","name":"crop_image_depthmap","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":168,"end_line":183,"context_start_line":148,"context_end_line":203,"code":"    output_resolution,\n    scaling=1,\n    offset_factor=0.5,\n    offset=None,\n):\n    # Margins to offset the origin\n    margins = np.asarray(input_resolution) * scaling - output_resolution\n    assert np.all(margins >= 0.0)\n    if offset is None:\n        offset = offset_factor * margins\n\n    # Generate new camera parameters\n    output_camera_matrix_colmap = opencv_to_colmap_intrinsics(input_camera_matrix)\n    output_camera_matrix_colmap[:2, :] *= scaling\n    output_camera_matrix_colmap[:2, 2] -= offset\n    output_camera_matrix = colmap_to_opencv_intrinsics(output_camera_matrix_colmap)\n\n    return output_camera_matrix\n\n\ndef crop_image_depthmap(image, depthmap, camera_intrinsics, crop_bbox):\n    \"\"\"\n    Return a crop of the input view.\n    \"\"\"\n    image = ImageList(image)\n    l, t, r, b = crop_bbox\n\n    image = image.crop((l, t, r, b))\n    if depthmap is not None:\n        depthmap = depthmap[t:b, l:r]\n\n    camera_intrinsics = camera_intrinsics.copy()\n    camera_intrinsics[0, 2] -= l\n    camera_intrinsics[1, 2] -= t\n\n    return image.to_pil(), depthmap, camera_intrinsics\n\n\ndef crop_image_depthmap_mask(image, depthmap, mask, camera_intrinsics, crop_bbox):\n    \"\"\"\n    Return a crop of the input view including mask.\n    \"\"\"\n    image = ImageList(image)\n    mask = ImageList(mask)\n    l, t, r, b = crop_bbox\n\n    image = image.crop((l, t, r, b))\n    mask = mask.crop((l, t, r, b))\n    depthmap = depthmap[t:b, l:r]\n\n    camera_intrinsics = camera_intrinsics.copy()\n    camera_intrinsics[0, 2] -= l\n    camera_intrinsics[1, 2] -= t\n\n    return image.to_pil(), depthmap, mask.to_pil(), camera_intrinsics\n","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.crop_image_depthmap_mask","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.crop_image_depthmap_mask#L186-L202","kind":"function","name":"crop_image_depthmap_mask","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":186,"end_line":202,"context_start_line":166,"context_end_line":211,"code":"\n\ndef crop_image_depthmap(image, depthmap, camera_intrinsics, crop_bbox):\n    \"\"\"\n    Return a crop of the input view.\n    \"\"\"\n    image = ImageList(image)\n    l, t, r, b = crop_bbox\n\n    image = image.crop((l, t, r, b))\n    if depthmap is not None:\n        depthmap = depthmap[t:b, l:r]\n\n    camera_intrinsics = camera_intrinsics.copy()\n    camera_intrinsics[0, 2] -= l\n    camera_intrinsics[1, 2] -= t\n\n    return image.to_pil(), depthmap, camera_intrinsics\n\n\ndef crop_image_depthmap_mask(image, depthmap, mask, camera_intrinsics, crop_bbox):\n    \"\"\"\n    Return a crop of the input view including mask.\n    \"\"\"\n    image = ImageList(image)\n    mask = ImageList(mask)\n    l, t, r, b = crop_bbox\n\n    image = image.crop((l, t, r, b))\n    mask = mask.crop((l, t, r, b))\n    depthmap = depthmap[t:b, l:r]\n\n    camera_intrinsics = camera_intrinsics.copy()\n    camera_intrinsics[0, 2] -= l\n    camera_intrinsics[1, 2] -= t\n\n    return image.to_pil(), depthmap, mask.to_pil(), camera_intrinsics\n\n\ndef bbox_from_intrinsics_in_out(\n    input_camera_matrix, output_camera_matrix, output_resolution\n):\n    out_width, out_height = output_resolution\n    l, t = np.int32(np.round(input_camera_matrix[:2, 2] - output_camera_matrix[:2, 2]))\n    crop_bbox = (l, t, l + out_width, t + out_height)\n    return crop_bbox","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.bbox_from_intrinsics_in_out","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.bbox_from_intrinsics_in_out#L205-L211","kind":"function","name":"bbox_from_intrinsics_in_out","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":205,"end_line":211,"context_start_line":185,"context_end_line":211,"code":"\ndef crop_image_depthmap_mask(image, depthmap, mask, camera_intrinsics, crop_bbox):\n    \"\"\"\n    Return a crop of the input view including mask.\n    \"\"\"\n    image = ImageList(image)\n    mask = ImageList(mask)\n    l, t, r, b = crop_bbox\n\n    image = image.crop((l, t, r, b))\n    mask = mask.crop((l, t, r, b))\n    depthmap = depthmap[t:b, l:r]\n\n    camera_intrinsics = camera_intrinsics.copy()\n    camera_intrinsics[0, 2] -= l\n    camera_intrinsics[1, 2] -= t\n\n    return image.to_pil(), depthmap, mask.to_pil(), camera_intrinsics\n\n\ndef bbox_from_intrinsics_in_out(\n    input_camera_matrix, output_camera_matrix, output_resolution\n):\n    out_width, out_height = output_resolution\n    l, t = np.int32(np.round(input_camera_matrix[:2, 2] - output_camera_matrix[:2, 2]))\n    crop_bbox = (l, t, l + out_width, t + out_height)\n    return crop_bbox","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.__init__","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.__init__#L29-L36","kind":"function","name":"__init__","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":29,"end_line":36,"context_start_line":9,"context_end_line":56,"code":"\nos.environ[\"OPENCV_IO_ENABLE_OPENEXR\"] = \"1\"\nimport cv2  # noqa\nimport numpy as np  # noqa\nfrom dust3r.utils.geometry import (\n    colmap_to_opencv_intrinsics,\n    opencv_to_colmap_intrinsics,\n)  # noqa\n\ntry:\n    lanczos = PIL.Image.Resampling.LANCZOS\n    bicubic = PIL.Image.Resampling.BICUBIC\nexcept AttributeError:\n    lanczos = PIL.Image.LANCZOS\n    bicubic = PIL.Image.BICUBIC\n\n\nclass ImageList:\n    \"\"\"Convenience class to aply the same operation to a whole set of images.\"\"\"\n\n    def __init__(self, images):\n        if not isinstance(images, (tuple, list, set)):\n            images = [images]\n        self.images = []\n        for image in images:\n            if not isinstance(image, PIL.Image.Image):\n                image = PIL.Image.fromarray(image)\n            self.images.append(image)\n\n    def __len__(self):\n        return len(self.images)\n\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.__len__","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.__len__#L38-L39","kind":"function","name":"__len__","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":38,"end_line":39,"context_start_line":18,"context_end_line":59,"code":"try:\n    lanczos = PIL.Image.Resampling.LANCZOS\n    bicubic = PIL.Image.Resampling.BICUBIC\nexcept AttributeError:\n    lanczos = PIL.Image.LANCZOS\n    bicubic = PIL.Image.BICUBIC\n\n\nclass ImageList:\n    \"\"\"Convenience class to aply the same operation to a whole set of images.\"\"\"\n\n    def __init__(self, images):\n        if not isinstance(images, (tuple, list, set)):\n            images = [images]\n        self.images = []\n        for image in images:\n            if not isinstance(image, PIL.Image.Image):\n                image = PIL.Image.fromarray(image)\n            self.images.append(image)\n\n    def __len__(self):\n        return len(self.images)\n\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):\n        return [getattr(im, func)(*args, **kwargs) for im in self.images]\n\n","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.to_pil","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.to_pil#L41-L42","kind":"function","name":"to_pil","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":41,"end_line":42,"context_start_line":21,"context_end_line":62,"code":"except AttributeError:\n    lanczos = PIL.Image.LANCZOS\n    bicubic = PIL.Image.BICUBIC\n\n\nclass ImageList:\n    \"\"\"Convenience class to aply the same operation to a whole set of images.\"\"\"\n\n    def __init__(self, images):\n        if not isinstance(images, (tuple, list, set)):\n            images = [images]\n        self.images = []\n        for image in images:\n            if not isinstance(image, PIL.Image.Image):\n                image = PIL.Image.fromarray(image)\n            self.images.append(image)\n\n    def __len__(self):\n        return len(self.images)\n\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):\n        return [getattr(im, func)(*args, **kwargs) for im in self.images]\n\n\ndef rescale_image_depthmap(\n    image, depthmap, camera_intrinsics, output_resolution, force=True\n):","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.size","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.size#L45-L48","kind":"function","name":"size","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":45,"end_line":48,"context_start_line":25,"context_end_line":68,"code":"\nclass ImageList:\n    \"\"\"Convenience class to aply the same operation to a whole set of images.\"\"\"\n\n    def __init__(self, images):\n        if not isinstance(images, (tuple, list, set)):\n            images = [images]\n        self.images = []\n        for image in images:\n            if not isinstance(image, PIL.Image.Image):\n                image = PIL.Image.fromarray(image)\n            self.images.append(image)\n\n    def __len__(self):\n        return len(self.images)\n\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):\n        return [getattr(im, func)(*args, **kwargs) for im in self.images]\n\n\ndef rescale_image_depthmap(\n    image, depthmap, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.resize","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.resize#L50-L51","kind":"function","name":"resize","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":50,"end_line":51,"context_start_line":30,"context_end_line":71,"code":"        if not isinstance(images, (tuple, list, set)):\n            images = [images]\n        self.images = []\n        for image in images:\n            if not isinstance(image, PIL.Image.Image):\n                image = PIL.Image.fromarray(image)\n            self.images.append(image)\n\n    def __len__(self):\n        return len(self.images)\n\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):\n        return [getattr(im, func)(*args, **kwargs) for im in self.images]\n\n\ndef rescale_image_depthmap(\n    image, depthmap, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)\n    if depthmap is not None:\n        # can also use this with masks instead of depthmaps\n        assert tuple(depthmap.shape[:2]) == image.size[::-1]","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping.crop","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping.crop#L53-L54","kind":"function","name":"crop","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":53,"end_line":54,"context_start_line":33,"context_end_line":74,"code":"        for image in images:\n            if not isinstance(image, PIL.Image.Image):\n                image = PIL.Image.fromarray(image)\n            self.images.append(image)\n\n    def __len__(self):\n        return len(self.images)\n\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):\n        return [getattr(im, func)(*args, **kwargs) for im in self.images]\n\n\ndef rescale_image_depthmap(\n    image, depthmap, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)\n    if depthmap is not None:\n        # can also use this with masks instead of depthmaps\n        assert tuple(depthmap.shape[:2]) == image.size[::-1]\n\n    # define output resolution\n    assert output_resolution.shape == (2,)","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.cropping._dispatch","uri":"program://Human3R/function/src.dust3r.datasets.utils.cropping._dispatch#L56-L57","kind":"function","name":"_dispatch","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":56,"end_line":57,"context_start_line":36,"context_end_line":77,"code":"            self.images.append(image)\n\n    def __len__(self):\n        return len(self.images)\n\n    def to_pil(self):\n        return tuple(self.images) if len(self.images) > 1 else self.images[0]\n\n    @property\n    def size(self):\n        sizes = [im.size for im in self.images]\n        assert all(sizes[0] == s for s in sizes)\n        return sizes[0]\n\n    def resize(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"resize\", *args, **kwargs))\n\n    def crop(self, *args, **kwargs):\n        return ImageList(self._dispatch(\"crop\", *args, **kwargs))\n\n    def _dispatch(self, func, *args, **kwargs):\n        return [getattr(im, func)(*args, **kwargs) for im in self.images]\n\n\ndef rescale_image_depthmap(\n    image, depthmap, camera_intrinsics, output_resolution, force=True\n):\n    \"\"\"Jointly rescale a (image, depthmap)\n    so that (out_width, out_height) >= output_res\n    \"\"\"\n    image = ImageList(image)\n    input_resolution = np.array(image.size)  # (W,H)\n    output_resolution = np.array(output_resolution)\n    if depthmap is not None:\n        # can also use this with masks instead of depthmaps\n        assert tuple(depthmap.shape[:2]) == image.size[::-1]\n\n    # define output resolution\n    assert output_resolution.shape == (2,)\n    scale_final = max(output_resolution / image.size) + 1e-8\n    if scale_final >= 1 and not force:  # image is already smaller than what is asked\n        return (image.to_pil(), depthmap, camera_intrinsics)","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.transforms","uri":"program://Human3R/module/src.dust3r.datasets.utils.transforms#L1-L80","kind":"module","name":"src.dust3r.datasets.utils.transforms","path":"src/dust3r/datasets/utils/transforms.py","language":"python","start_line":1,"end_line":80,"context_start_line":1,"context_end_line":80,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# DUST3R default transforms\n# --------------------------------------------------------\nimport torchvision.transforms as tvf\nfrom dust3r.utils.image import ImgNorm\n\n# define the standard image transforms\nColorJitter = tvf.Compose([tvf.ColorJitter(0.5, 0.5, 0.5, 0.1), ImgNorm])\n\n\ndef _check_input(value, center=1, bound=(0, float(\"inf\")), clip_first_on_zero=True):\n    if isinstance(value, (int, float)):\n        if value < 0:\n            raise ValueError(f\"If  is a single number, it must be non negative.\")\n        value = [center - float(value), center + float(value)]\n        if clip_first_on_zero:\n            value[0] = max(value[0], 0.0)\n    elif isinstance(value, (tuple, list)) and len(value) == 2:\n        value = [float(value[0]), float(value[1])]\n    else:\n        raise TypeError(f\"should be a single number or a list/tuple with length 2.\")\n\n    if not bound[0] <= value[0] <= value[1] <= bound[1]:\n        raise ValueError(f\"values should be between {bound}, but got {value}.\")\n\n    # if value is 0 or (1., 1.) for brightness/contrast/saturation\n    # or (0., 0.) for hue, do nothing\n    if value[0] == value[1] == center:\n        return None\n    else:\n        return tuple(value)\n\n\nimport torch\nimport torchvision.transforms.functional as F\n\n\ndef SeqColorJitter():\n    \"\"\"\n    Return a color jitter transform with same random parameters\n    \"\"\"\n    brightness = _check_input(0.5)\n    contrast = _check_input(0.5)\n    saturation = _check_input(0.5)\n    hue = _check_input(0.1, center=0, bound=(-0.5, 0.5), clip_first_on_zero=False)\n\n    fn_idx = torch.randperm(4)\n    brightness_factor = (\n        None\n        if brightness is None\n        else float(torch.empty(1).uniform_(brightness[0], brightness[1]))\n    )\n    contrast_factor = (\n        None\n        if contrast is None\n        else float(torch.empty(1).uniform_(contrast[0], contrast[1]))\n    )\n    saturation_factor = (\n        None\n        if saturation is None\n        else float(torch.empty(1).uniform_(saturation[0], saturation[1]))\n    )\n    hue_factor = None if hue is None else float(torch.empty(1).uniform_(hue[0], hue[1]))\n\n    def _color_jitter(img):\n        for fn_id in fn_idx:\n            if fn_id == 0 and brightness_factor is not None:\n                img = F.adjust_brightness(img, brightness_factor)\n            elif fn_id == 1 and contrast_factor is not None:\n                img = F.adjust_contrast(img, contrast_factor)\n            elif fn_id == 2 and saturation_factor is not None:\n                img = F.adjust_saturation(img, saturation_factor)\n            elif fn_id == 3 and hue_factor is not None:\n                img = F.adjust_hue(img, hue_factor)\n        return ImgNorm(img)\n\n    return _color_jitter","source_hash":"10308410a74d5b9ce263b77f0b94418bad3deaa7f1a3e0c330216538ba07d0d6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.transforms._check_input","uri":"program://Human3R/function/src.dust3r.datasets.utils.transforms._check_input#L14-L34","kind":"function","name":"_check_input","path":"src/dust3r/datasets/utils/transforms.py","language":"python","start_line":14,"end_line":34,"context_start_line":1,"context_end_line":54,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# DUST3R default transforms\n# --------------------------------------------------------\nimport torchvision.transforms as tvf\nfrom dust3r.utils.image import ImgNorm\n\n# define the standard image transforms\nColorJitter = tvf.Compose([tvf.ColorJitter(0.5, 0.5, 0.5, 0.1), ImgNorm])\n\n\ndef _check_input(value, center=1, bound=(0, float(\"inf\")), clip_first_on_zero=True):\n    if isinstance(value, (int, float)):\n        if value < 0:\n            raise ValueError(f\"If  is a single number, it must be non negative.\")\n        value = [center - float(value), center + float(value)]\n        if clip_first_on_zero:\n            value[0] = max(value[0], 0.0)\n    elif isinstance(value, (tuple, list)) and len(value) == 2:\n        value = [float(value[0]), float(value[1])]\n    else:\n        raise TypeError(f\"should be a single number or a list/tuple with length 2.\")\n\n    if not bound[0] <= value[0] <= value[1] <= bound[1]:\n        raise ValueError(f\"values should be between {bound}, but got {value}.\")\n\n    # if value is 0 or (1., 1.) for brightness/contrast/saturation\n    # or (0., 0.) for hue, do nothing\n    if value[0] == value[1] == center:\n        return None\n    else:\n        return tuple(value)\n\n\nimport torch\nimport torchvision.transforms.functional as F\n\n\ndef SeqColorJitter():\n    \"\"\"\n    Return a color jitter transform with same random parameters\n    \"\"\"\n    brightness = _check_input(0.5)\n    contrast = _check_input(0.5)\n    saturation = _check_input(0.5)\n    hue = _check_input(0.1, center=0, bound=(-0.5, 0.5), clip_first_on_zero=False)\n\n    fn_idx = torch.randperm(4)\n    brightness_factor = (\n        None\n        if brightness is None\n        else float(torch.empty(1).uniform_(brightness[0], brightness[1]))","source_hash":"10308410a74d5b9ce263b77f0b94418bad3deaa7f1a3e0c330216538ba07d0d6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.transforms.SeqColorJitter","uri":"program://Human3R/function/src.dust3r.datasets.utils.transforms.SeqColorJitter#L41-L80","kind":"function","name":"SeqColorJitter","path":"src/dust3r/datasets/utils/transforms.py","language":"python","start_line":41,"end_line":80,"context_start_line":21,"context_end_line":80,"code":"    elif isinstance(value, (tuple, list)) and len(value) == 2:\n        value = [float(value[0]), float(value[1])]\n    else:\n        raise TypeError(f\"should be a single number or a list/tuple with length 2.\")\n\n    if not bound[0] <= value[0] <= value[1] <= bound[1]:\n        raise ValueError(f\"values should be between {bound}, but got {value}.\")\n\n    # if value is 0 or (1., 1.) for brightness/contrast/saturation\n    # or (0., 0.) for hue, do nothing\n    if value[0] == value[1] == center:\n        return None\n    else:\n        return tuple(value)\n\n\nimport torch\nimport torchvision.transforms.functional as F\n\n\ndef SeqColorJitter():\n    \"\"\"\n    Return a color jitter transform with same random parameters\n    \"\"\"\n    brightness = _check_input(0.5)\n    contrast = _check_input(0.5)\n    saturation = _check_input(0.5)\n    hue = _check_input(0.1, center=0, bound=(-0.5, 0.5), clip_first_on_zero=False)\n\n    fn_idx = torch.randperm(4)\n    brightness_factor = (\n        None\n        if brightness is None\n        else float(torch.empty(1).uniform_(brightness[0], brightness[1]))\n    )\n    contrast_factor = (\n        None\n        if contrast is None\n        else float(torch.empty(1).uniform_(contrast[0], contrast[1]))\n    )\n    saturation_factor = (\n        None\n        if saturation is None\n        else float(torch.empty(1).uniform_(saturation[0], saturation[1]))\n    )\n    hue_factor = None if hue is None else float(torch.empty(1).uniform_(hue[0], hue[1]))\n\n    def _color_jitter(img):\n        for fn_id in fn_idx:\n            if fn_id == 0 and brightness_factor is not None:\n                img = F.adjust_brightness(img, brightness_factor)\n            elif fn_id == 1 and contrast_factor is not None:\n                img = F.adjust_contrast(img, contrast_factor)\n            elif fn_id == 2 and saturation_factor is not None:\n                img = F.adjust_saturation(img, saturation_factor)\n            elif fn_id == 3 and hue_factor is not None:\n                img = F.adjust_hue(img, hue_factor)\n        return ImgNorm(img)\n\n    return _color_jitter","source_hash":"10308410a74d5b9ce263b77f0b94418bad3deaa7f1a3e0c330216538ba07d0d6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.utils.transforms._color_jitter","uri":"program://Human3R/function/src.dust3r.datasets.utils.transforms._color_jitter#L68-L78","kind":"function","name":"_color_jitter","path":"src/dust3r/datasets/utils/transforms.py","language":"python","start_line":68,"end_line":78,"context_start_line":48,"context_end_line":80,"code":"    hue = _check_input(0.1, center=0, bound=(-0.5, 0.5), clip_first_on_zero=False)\n\n    fn_idx = torch.randperm(4)\n    brightness_factor = (\n        None\n        if brightness is None\n        else float(torch.empty(1).uniform_(brightness[0], brightness[1]))\n    )\n    contrast_factor = (\n        None\n        if contrast is None\n        else float(torch.empty(1).uniform_(contrast[0], contrast[1]))\n    )\n    saturation_factor = (\n        None\n        if saturation is None\n        else float(torch.empty(1).uniform_(saturation[0], saturation[1]))\n    )\n    hue_factor = None if hue is None else float(torch.empty(1).uniform_(hue[0], hue[1]))\n\n    def _color_jitter(img):\n        for fn_id in fn_idx:\n            if fn_id == 0 and brightness_factor is not None:\n                img = F.adjust_brightness(img, brightness_factor)\n            elif fn_id == 1 and contrast_factor is not None:\n                img = F.adjust_contrast(img, contrast_factor)\n            elif fn_id == 2 and saturation_factor is not None:\n                img = F.adjust_saturation(img, saturation_factor)\n            elif fn_id == 3 and hue_factor is not None:\n                img = F.adjust_hue(img, hue_factor)\n        return ImgNorm(img)\n\n    return _color_jitter","source_hash":"10308410a74d5b9ce263b77f0b94418bad3deaa7f1a3e0c330216538ba07d0d6","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.batched_sampler","uri":"program://Human3R/module/src.dust3r.datasets.base.batched_sampler#L1-L94","kind":"module","name":"src.dust3r.datasets.base.batched_sampler","path":"src/dust3r/datasets/base/batched_sampler.py","language":"python","start_line":1,"end_line":94,"context_start_line":1,"context_end_line":94,"code":"import numpy as np\nimport torch\nfrom accelerate import Accelerator\nimport torch.utils\nfrom torch.utils.data import BatchSampler, Sampler\nimport torch.utils.data\n\n\nclass CustomRandomSampler(Sampler):\n    \"\"\"Random sampling under a constraint: each sample in the batch has the same feature,\n    which is chosen randomly from a known pool of 'features' for each batch.\n\n    For instance, the 'feature' could be the image aspect-ratio.\n\n    The index returned is a tuple (sample_idx, feat_idx).\n    This sampler ensures that each series of `batch_size` indices has the same `feat_idx`.\n    \"\"\"\n\n    def __init__(\n        self,\n        dataset,\n        batch_size,\n        pool_size,\n        min_view_size,\n        max_view_size,\n        world_size,\n        warmup=1,\n        drop_last=True,\n    ):\n        self.batch_size = batch_size\n        self.pool_size = pool_size\n        self.min_view_size = min_view_size\n        self.max_view_size = max_view_size\n        self.drop_last = drop_last\n        self.len_dataset = N = len(dataset)\n        self.total_size = N\n\n        self.epoch = None\n        self.epochf = 0.0\n\n    def __len__(self):\n        return self.total_size\n\n    def set_epoch(self, epoch):\n        self.epoch = epoch\n\n    def __iter__(self):\n        if self.epoch is None:\n            raise ValueError(\n                \"Epoch number not set. Please call 'set_epoch(epoch)' before iterating.\"\n            )\n\n        seed = self.epoch + 788\n        rng = np.random.default_rng(seed=seed)\n        # random indices (will restart from 0 if not drop_last)\n        sample_idxs = np.arange(self.total_size)\n        rng.shuffle(sample_idxs)\n        # random feat_idxs (same across each batch)\n        n_batches = (self.total_size + self.batch_size - 1) // self.batch_size\n        if self.pool_size > 1:\n            p = np.ones(self.pool_size)\n            p[: self.pool_size // 2] *= 2\n            p = p / p.sum()\n            _feat_idxs = rng.choice(self.pool_size, size=n_batches, p=p)\n        else:\n            _feat_idxs = rng.integers(self.pool_size, size=n_batches)\n        _feat_idxs = np.broadcast_to(_feat_idxs[:, None], (n_batches, self.batch_size))\n        _feat_idxs = _feat_idxs.ravel()[: self.total_size]\n        _view_idxs = rng.integers(\n            self.min_view_size, self.max_view_size + 1, size=n_batches\n        )\n        _view_idxs = np.broadcast_to(_view_idxs[:, None], (n_batches, self.batch_size))\n        _view_idxs = _view_idxs.ravel()[: self.total_size]\n\n        idxs = np.c_[sample_idxs, _feat_idxs, _view_idxs] # id, selected type (eg., resolution), view number\n        yield from (tuple(idx) for idx in idxs)\n\n\nclass BatchedRandomSampler(BatchSampler):\n    \"\"\"Batch sampler that groups indices from RandomSampler into batches.\"\"\"\n\n    def __init__(self, sampler: CustomRandomSampler, batch_size, drop_last=True):\n        self.sampler = sampler  # An instance of RandomSampler\n        self.batch_size = batch_size\n        self.drop_last = drop_last\n\n    def set_epoch(self, epoch):\n        self.sampler.set_epoch(epoch)\n\n\ndef round_by(total, multiple, up=False):\n    if up:\n        total = total + multiple - 1\n    return (total // multiple) * multiple","source_hash":"bb63eb9b59cf6e180e044b6a187b54034aa8ae66718c8d7665d7e97d3a6dfae9","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.batched_sampler.CustomRandomSampler","uri":"program://Human3R/class/src.dust3r.datasets.base.batched_sampler.CustomRandomSampler#L9-L76","kind":"class","name":"CustomRandomSampler","path":"src/dust3r/datasets/base/batched_sampler.py","language":"python","start_line":9,"end_line":76,"context_start_line":1,"context_end_line":94,"code":"import numpy as np\nimport torch\nfrom accelerate import Accelerator\nimport torch.utils\nfrom torch.utils.data import BatchSampler, Sampler\nimport torch.utils.data\n\n\nclass CustomRandomSampler(Sampler):\n    \"\"\"Random sampling under a constraint: each sample in the batch has the same feature,\n    which is chosen randomly from a known pool of 'features' for each batch.\n\n    For instance, the 'feature' could be the image aspect-ratio.\n\n    The index returned is a tuple (sample_idx, feat_idx).\n    This sampler ensures that each series of `batch_size` indices has the same `feat_idx`.\n    \"\"\"\n\n    def __init__(\n        self,\n        dataset,\n        batch_size,\n        pool_size,\n        min_view_size,\n        max_view_size,\n        world_size,\n        warmup=1,\n        drop_last=True,\n    ):\n        self.batch_size = batch_size\n        self.pool_size = pool_size\n        self.min_view_size = min_view_size\n        self.max_view_size = max_view_size\n        self.drop_last = drop_last\n        self.len_dataset = N = len(dataset)\n        self.total_size = N\n\n        self.epoch = None\n        self.epochf = 0.0\n\n    def __len__(self):\n        return self.total_size\n\n    def set_epoch(self, epoch):\n        self.epoch = epoch\n\n    def __iter__(self):\n        if self.epoch is None:\n            raise ValueError(\n                \"Epoch number not set. Please call 'set_epoch(epoch)' before iterating.\"\n            )\n\n        seed = self.epoch + 788\n        rng = np.random.default_rng(seed=seed)\n        # random indices (will restart from 0 if not drop_last)\n        sample_idxs = np.arange(self.total_size)\n        rng.shuffle(sample_idxs)\n        # random feat_idxs (same across each batch)\n        n_batches = (self.total_size + self.batch_size - 1) // self.batch_size\n        if self.pool_size > 1:\n            p = np.ones(self.pool_size)\n            p[: self.pool_size // 2] *= 2\n            p = p / p.sum()\n            _feat_idxs = rng.choice(self.pool_size, size=n_batches, p=p)\n        else:\n            _feat_idxs = rng.integers(self.pool_size, size=n_batches)\n        _feat_idxs = np.broadcast_to(_feat_idxs[:, None], (n_batches, self.batch_size))\n        _feat_idxs = _feat_idxs.ravel()[: self.total_size]\n        _view_idxs = rng.integers(\n            self.min_view_size, self.max_view_size + 1, size=n_batches\n        )\n        _view_idxs = np.broadcast_to(_view_idxs[:, None], (n_batches, self.batch_size))\n        _view_idxs = _view_idxs.ravel()[: self.total_size]\n\n        idxs = np.c_[sample_idxs, _feat_idxs, _view_idxs] # id, selected type (eg., resolution), view number\n        yield from (tuple(idx) for idx in idxs)\n\n\nclass BatchedRandomSampler(BatchSampler):\n    \"\"\"Batch sampler that groups indices from RandomSampler into batches.\"\"\"\n\n    def __init__(self, sampler: CustomRandomSampler, batch_size, drop_last=True):\n        self.sampler = sampler  # An instance of RandomSampler\n        self.batch_size = batch_size\n        self.drop_last = drop_last\n\n    def set_epoch(self, epoch):\n        self.sampler.set_epoch(epoch)\n\n\ndef round_by(total, multiple, up=False):\n    if up:\n        total = total + multiple - 1\n    return (total // multiple) * multiple","source_hash":"bb63eb9b59cf6e180e044b6a187b54034aa8ae66718c8d7665d7e97d3a6dfae9","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.batched_sampler.BatchedRandomSampler","uri":"program://Human3R/class/src.dust3r.datasets.base.batched_sampler.BatchedRandomSampler#L79-L88","kind":"class","name":"BatchedRandomSampler","path":"src/dust3r/datasets/base/batched_sampler.py","language":"python","start_line":79,"end_line":88,"context_start_line":59,"context_end_line":94,"code":"        n_batches = (self.total_size + self.batch_size - 1) // self.batch_size\n        if self.pool_size > 1:\n            p = np.ones(self.pool_size)\n            p[: self.pool_size // 2] *= 2\n            p = p / p.sum()\n            _feat_idxs = rng.choice(self.pool_size, size=n_batches, p=p)\n        else:\n            _feat_idxs = rng.integers(self.pool_size, size=n_batches)\n        _feat_idxs = np.broadcast_to(_feat_idxs[:, None], (n_batches, self.batch_size))\n        _feat_idxs = _feat_idxs.ravel()[: self.total_size]\n        _view_idxs = rng.integers(\n            self.min_view_size, self.max_view_size + 1, size=n_batches\n        )\n        _view_idxs = np.broadcast_to(_view_idxs[:, None], (n_batches, self.batch_size))\n        _view_idxs = _view_idxs.ravel()[: self.total_size]\n\n        idxs = np.c_[sample_idxs, _feat_idxs, _view_idxs] # id, selected type (eg., resolution), view number\n        yield from (tuple(idx) for idx in idxs)\n\n\nclass BatchedRandomSampler(BatchSampler):\n    \"\"\"Batch sampler that groups indices from RandomSampler into batches.\"\"\"\n\n    def __init__(self, sampler: CustomRandomSampler, batch_size, drop_last=True):\n        self.sampler = sampler  # An instance of RandomSampler\n        self.batch_size = batch_size\n        self.drop_last = drop_last\n\n    def set_epoch(self, epoch):\n        self.sampler.set_epoch(epoch)\n\n\ndef round_by(total, multiple, up=False):\n    if up:\n        total = total + multiple - 1\n    return (total // multiple) * multiple","source_hash":"bb63eb9b59cf6e180e044b6a187b54034aa8ae66718c8d7665d7e97d3a6dfae9","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.batched_sampler.round_by","uri":"program://Human3R/function/src.dust3r.datasets.base.batched_sampler.round_by#L91-L94","kind":"function","name":"round_by","path":"src/dust3r/datasets/base/batched_sampler.py","language":"python","start_line":91,"end_line":94,"context_start_line":71,"context_end_line":94,"code":"        )\n        _view_idxs = np.broadcast_to(_view_idxs[:, None], (n_batches, self.batch_size))\n        _view_idxs = _view_idxs.ravel()[: self.total_size]\n\n        idxs = np.c_[sample_idxs, _feat_idxs, _view_idxs] # id, selected type (eg., resolution), view number\n        yield from (tuple(idx) for idx in idxs)\n\n\nclass BatchedRandomSampler(BatchSampler):\n    \"\"\"Batch sampler that groups indices from RandomSampler into batches.\"\"\"\n\n    def __init__(self, sampler: CustomRandomSampler, batch_size, drop_last=True):\n        self.sampler = sampler  # An instance of RandomSampler\n        self.batch_size = batch_size\n        self.drop_last = drop_last\n\n    def set_epoch(self, epoch):\n        self.sampler.set_epoch(epoch)\n\n\ndef round_by(total, multiple, up=False):\n    if up:\n        total = total + multiple - 1\n    return (total // multiple) * multiple","source_hash":"bb63eb9b59cf6e180e044b6a187b54034aa8ae66718c8d7665d7e97d3a6dfae9","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.batched_sampler.__init__","uri":"program://Human3R/function/src.dust3r.datasets.base.batched_sampler.__init__#L82-L85","kind":"function","name":"__init__","path":"src/dust3r/datasets/base/batched_sampler.py","language":"python","start_line":82,"end_line":85,"context_start_line":62,"context_end_line":94,"code":"            p[: self.pool_size // 2] *= 2\n            p = p / p.sum()\n            _feat_idxs = rng.choice(self.pool_size, size=n_batches, p=p)\n        else:\n            _feat_idxs = rng.integers(self.pool_size, size=n_batches)\n        _feat_idxs = np.broadcast_to(_feat_idxs[:, None], (n_batches, self.batch_size))\n        _feat_idxs = _feat_idxs.ravel()[: self.total_size]\n        _view_idxs = rng.integers(\n            self.min_view_size, self.max_view_size + 1, size=n_batches\n        )\n        _view_idxs = np.broadcast_to(_view_idxs[:, None], (n_batches, self.batch_size))\n        _view_idxs = _view_idxs.ravel()[: self.total_size]\n\n        idxs = np.c_[sample_idxs, _feat_idxs, _view_idxs] # id, selected type (eg., resolution), view number\n        yield from (tuple(idx) for idx in idxs)\n\n\nclass BatchedRandomSampler(BatchSampler):\n    \"\"\"Batch sampler that groups indices from RandomSampler into batches.\"\"\"\n\n    def __init__(self, sampler: CustomRandomSampler, batch_size, drop_last=True):\n        self.sampler = sampler  # An instance of RandomSampler\n        self.batch_size = batch_size\n        self.drop_last = drop_last\n\n    def set_epoch(self, epoch):\n        self.sampler.set_epoch(epoch)\n\n\ndef round_by(total, multiple, up=False):\n    if up:\n        total = total + multiple - 1\n    return (total // multiple) * multiple","source_hash":"bb63eb9b59cf6e180e044b6a187b54034aa8ae66718c8d7665d7e97d3a6dfae9","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.batched_sampler.__len__","uri":"program://Human3R/function/src.dust3r.datasets.base.batched_sampler.__len__#L41-L42","kind":"function","name":"__len__","path":"src/dust3r/datasets/base/batched_sampler.py","language":"python","start_line":41,"end_line":42,"context_start_line":21,"context_end_line":62,"code":"        dataset,\n        batch_size,\n        pool_size,\n        min_view_size,\n        max_view_size,\n        world_size,\n        warmup=1,\n        drop_last=True,\n    ):\n        self.batch_size = batch_size\n        self.pool_size = pool_size\n        self.min_view_size = min_view_size\n        self.max_view_size = max_view_size\n        self.drop_last = drop_last\n        self.len_dataset = N = len(dataset)\n        self.total_size = N\n\n        self.epoch = None\n        self.epochf = 0.0\n\n    def __len__(self):\n        return self.total_size\n\n    def set_epoch(self, epoch):\n        self.epoch = epoch\n\n    def __iter__(self):\n        if self.epoch is None:\n            raise ValueError(\n                \"Epoch number not set. Please call 'set_epoch(epoch)' before iterating.\"\n            )\n\n        seed = self.epoch + 788\n        rng = np.random.default_rng(seed=seed)\n        # random indices (will restart from 0 if not drop_last)\n        sample_idxs = np.arange(self.total_size)\n        rng.shuffle(sample_idxs)\n        # random feat_idxs (same across each batch)\n        n_batches = (self.total_size + self.batch_size - 1) // self.batch_size\n        if self.pool_size > 1:\n            p = np.ones(self.pool_size)\n            p[: self.pool_size // 2] *= 2","source_hash":"bb63eb9b59cf6e180e044b6a187b54034aa8ae66718c8d7665d7e97d3a6dfae9","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.batched_sampler.set_epoch","uri":"program://Human3R/function/src.dust3r.datasets.base.batched_sampler.set_epoch#L87-L88","kind":"function","name":"set_epoch","path":"src/dust3r/datasets/base/batched_sampler.py","language":"python","start_line":87,"end_line":88,"context_start_line":67,"context_end_line":94,"code":"        _feat_idxs = np.broadcast_to(_feat_idxs[:, None], (n_batches, self.batch_size))\n        _feat_idxs = _feat_idxs.ravel()[: self.total_size]\n        _view_idxs = rng.integers(\n            self.min_view_size, self.max_view_size + 1, size=n_batches\n        )\n        _view_idxs = np.broadcast_to(_view_idxs[:, None], (n_batches, self.batch_size))\n        _view_idxs = _view_idxs.ravel()[: self.total_size]\n\n        idxs = np.c_[sample_idxs, _feat_idxs, _view_idxs] # id, selected type (eg., resolution), view number\n        yield from (tuple(idx) for idx in idxs)\n\n\nclass BatchedRandomSampler(BatchSampler):\n    \"\"\"Batch sampler that groups indices from RandomSampler into batches.\"\"\"\n\n    def __init__(self, sampler: CustomRandomSampler, batch_size, drop_last=True):\n        self.sampler = sampler  # An instance of RandomSampler\n        self.batch_size = batch_size\n        self.drop_last = drop_last\n\n    def set_epoch(self, epoch):\n        self.sampler.set_epoch(epoch)\n\n\ndef round_by(total, multiple, up=False):\n    if up:\n        total = total + multiple - 1\n    return (total // multiple) * multiple","source_hash":"bb63eb9b59cf6e180e044b6a187b54034aa8ae66718c8d7665d7e97d3a6dfae9","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.batched_sampler.__iter__","uri":"program://Human3R/function/src.dust3r.datasets.base.batched_sampler.__iter__#L47-L76","kind":"function","name":"__iter__","path":"src/dust3r/datasets/base/batched_sampler.py","language":"python","start_line":47,"end_line":76,"context_start_line":27,"context_end_line":94,"code":"        warmup=1,\n        drop_last=True,\n    ):\n        self.batch_size = batch_size\n        self.pool_size = pool_size\n        self.min_view_size = min_view_size\n        self.max_view_size = max_view_size\n        self.drop_last = drop_last\n        self.len_dataset = N = len(dataset)\n        self.total_size = N\n\n        self.epoch = None\n        self.epochf = 0.0\n\n    def __len__(self):\n        return self.total_size\n\n    def set_epoch(self, epoch):\n        self.epoch = epoch\n\n    def __iter__(self):\n        if self.epoch is None:\n            raise ValueError(\n                \"Epoch number not set. Please call 'set_epoch(epoch)' before iterating.\"\n            )\n\n        seed = self.epoch + 788\n        rng = np.random.default_rng(seed=seed)\n        # random indices (will restart from 0 if not drop_last)\n        sample_idxs = np.arange(self.total_size)\n        rng.shuffle(sample_idxs)\n        # random feat_idxs (same across each batch)\n        n_batches = (self.total_size + self.batch_size - 1) // self.batch_size\n        if self.pool_size > 1:\n            p = np.ones(self.pool_size)\n            p[: self.pool_size // 2] *= 2\n            p = p / p.sum()\n            _feat_idxs = rng.choice(self.pool_size, size=n_batches, p=p)\n        else:\n            _feat_idxs = rng.integers(self.pool_size, size=n_batches)\n        _feat_idxs = np.broadcast_to(_feat_idxs[:, None], (n_batches, self.batch_size))\n        _feat_idxs = _feat_idxs.ravel()[: self.total_size]\n        _view_idxs = rng.integers(\n            self.min_view_size, self.max_view_size + 1, size=n_batches\n        )\n        _view_idxs = np.broadcast_to(_view_idxs[:, None], (n_batches, self.batch_size))\n        _view_idxs = _view_idxs.ravel()[: self.total_size]\n\n        idxs = np.c_[sample_idxs, _feat_idxs, _view_idxs] # id, selected type (eg., resolution), view number\n        yield from (tuple(idx) for idx in idxs)\n\n\nclass BatchedRandomSampler(BatchSampler):\n    \"\"\"Batch sampler that groups indices from RandomSampler into batches.\"\"\"\n\n    def __init__(self, sampler: CustomRandomSampler, batch_size, drop_last=True):\n        self.sampler = sampler  # An instance of RandomSampler\n        self.batch_size = batch_size\n        self.drop_last = drop_last\n\n    def set_epoch(self, epoch):\n        self.sampler.set_epoch(epoch)\n\n\ndef round_by(total, multiple, up=False):\n    if up:\n        total = total + multiple - 1\n    return (total // multiple) * multiple","source_hash":"bb63eb9b59cf6e180e044b6a187b54034aa8ae66718c8d7665d7e97d3a6dfae9","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset","uri":"program://Human3R/module/src.dust3r.datasets.base.base_multiview_dataset#L1-L607","kind":"module","name":"src.dust3r.datasets.base.base_multiview_dataset","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":1,"end_line":607,"context_start_line":1,"context_end_line":607,"code":"import PIL\nimport numpy as np\nimport torch\nimport random\nimport itertools\nfrom dust3r.datasets.base.easy_dataset import EasyDataset\nfrom dust3r.datasets.utils.transforms import ImgNorm, SeqColorJitter\nfrom dust3r.utils.geometry import depthmap_to_absolute_camera_coordinates\nimport dust3r.datasets.utils.cropping as cropping\nfrom dust3r.datasets.utils.corr import extract_correspondences_from_pts3d\nimport torchvision.transforms as tvf\n\ndef get_ray_map(c2w1, c2w2, intrinsics, h, w):\n    c2w = np.linalg.inv(c2w1) @ c2w2\n    i, j = np.meshgrid(np.arange(w), np.arange(h), indexing=\"xy\")\n    grid = np.stack([i, j, np.ones_like(i)], axis=-1)\n    ro = c2w[:3, 3]\n    rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n    rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n    rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n    ro = np.broadcast_to(ro, (h, w, 3))\n    ray_map = np.concatenate([ro, rd], axis=-1)\n    return ray_map\n\n\nclass BaseMultiViewDataset(EasyDataset):\n    \"\"\"Define all basic options.\n\n    Usage:\n        class MyDataset (BaseMultiViewDataset):\n            def _get_views(self, idx, rng):\n                # overload here\n                views = []\n                views.append(dict(img=, ...))\n                return views\n    \"\"\"\n\n    def __init__(\n        self,\n        *,  # only keyword arguments\n        num_views=None,\n        split=None,\n        resolution=None,  # square_size or (width, height) or list of [(width,height), ...]\n        transform=ImgNorm,\n        aug_crop=False,\n        n_corres=0,\n        nneg=0,\n        seed=None,\n        allow_repeat=False,\n        seq_aug_crop=False,\n    ):\n        assert num_views is not None, \"undefined num_views\"\n        self.num_views = num_views\n        self.split = split\n        self._set_resolutions(resolution)\n\n        self.n_corres = n_corres\n        self.nneg = nneg\n        assert (\n            self.n_corres == \"all\"\n            or isinstance(self.n_corres, int)\n            or (\n                isinstance(self.n_corres, list) and len(self.n_corres) == self.num_views\n            )\n        ), f\"Error, n_corres should either be 'all', a single integer or a list of length {self.num_views}\"\n        assert (\n            self.nneg == 0 or self.n_corres != \"all\"\n        ), \"nneg should be 0 if n_corres is all\"\n\n        self.is_seq_color_jitter = False\n        if isinstance(transform, str):\n            transform = eval(transform)\n        if transform == SeqColorJitter:\n            transform = SeqColorJitter()\n            self.is_seq_color_jitter = True\n        self.transform = transform\n\n        self.aug_crop = aug_crop\n        self.seed = seed\n        self.allow_repeat = allow_repeat\n        self.seq_aug_crop = seq_aug_crop\n\n    def __len__(self):\n        return len(self.scenes)\n\n    @staticmethod\n    def efficient_random_intervals(\n        start,\n        num_elements,\n        interval_range,\n        fixed_interval_prob=0.8,\n        weights=None,\n        seed=42,\n    ):\n        if random.random() < fixed_interval_prob:\n            intervals = random.choices(interval_range, weights=weights) * (\n                num_elements - 1\n            )\n        else:\n            intervals = [\n                random.choices(interval_range, weights=weights)[0]\n                for _ in range(num_elements - 1)\n            ]\n        return list(itertools.accumulate([start] + intervals))\n\n    def sample_based_on_timestamps(self, i, timestamps, num_views, interval=1):\n        time_diffs = np.abs(timestamps - timestamps[i])\n        ids_candidate = np.where(time_diffs < interval)[0]\n        ids_candidate = np.sort(ids_candidate)\n        if (self.allow_repeat and len(ids_candidate) < num_views // 3) or (\n            len(ids_candidate) < num_views\n        ):\n            return []\n        ids_sel_list = []\n        ids_candidate_left = ids_candidate.copy()\n        while len(ids_candidate_left) >= num_views:\n            ids_sel = np.random.choice(ids_candidate_left, num_views, replace=False)\n            ids_sel_list.append(sorted(ids_sel))\n            ids_candidate_left = np.setdiff1d(ids_candidate_left, ids_sel)\n\n        if len(ids_candidate_left) > 0 and len(ids_candidate) >= num_views:\n            ids_sel = np.concatenate(\n                [\n                    ids_candidate_left,\n                    np.random.choice(\n                        np.setdiff1d(ids_candidate, ids_candidate_left),\n                        num_views - len(ids_candidate_left),\n                        replace=False,\n                    ),\n                ]\n            )\n            ids_sel_list.append(sorted(ids_sel))\n\n        if self.allow_repeat:\n            ids_sel_list.append(\n                sorted(np.random.choice(ids_candidate, num_views, replace=True))\n            )\n\n        # add sequences with fixed intervals (all possible intervals)\n        pos_i = np.where(ids_candidate == i)[0][0]\n        curr_interval = 1\n        stop = len(ids_candidate) < num_views\n        while not stop:\n            pos_sel = [pos_i]\n            count = 0\n            while len(pos_sel) < num_views:\n                if count % 2 == 0:\n                    curr_pos_i = pos_sel[-1] + curr_interval\n                    if curr_pos_i >= len(ids_candidate):\n                        stop = True\n                        break\n                    pos_sel.append(curr_pos_i)\n                else:\n                    curr_pos_i = pos_sel[0] - curr_interval\n                    if curr_pos_i < 0:\n                        stop = True\n                        break\n                    pos_sel.insert(0, curr_pos_i)\n                count += 1\n            if not stop and len(pos_sel) == num_views:\n                ids_sel = sorted([ids_candidate[pos] for pos in pos_sel])\n                if ids_sel not in ids_sel_list:\n                    ids_sel_list.append(ids_sel)\n            curr_interval += 1\n        return ids_sel_list\n\n    @staticmethod\n    def blockwise_shuffle(x, rng, block_shuffle):\n        if block_shuffle is None:\n            return rng.permutation(x).tolist()\n        else:\n            assert block_shuffle > 0\n            blocks = [x[i : i + block_shuffle] for i in range(0, len(x), block_shuffle)]\n            shuffled_blocks = [rng.permutation(block).tolist() for block in blocks]\n            shuffled_list = [item for block in shuffled_blocks for item in block]\n            return shuffled_list\n\n    def get_seq_from_start_id(\n        self,\n        num_views,\n        id_ref,\n        ids_all,\n        rng,\n        min_interval=1,\n        max_interval=25,\n        video_prob=0.5,\n        fix_interval_prob=0.5,\n        block_shuffle=None,\n    ):\n        \"\"\"\n        args:\n            num_views: number of views to return\n            id_ref: the reference id (first id)\n            ids_all: all the ids\n            rng: random number generator\n            max_interval: maximum interval between two views\n        returns:\n            pos: list of positions of the views in ids_all, i.e., index for ids_all\n            is_video: True if the views are consecutive\n        \"\"\"\n        assert min_interval > 0, f\"min_interval should be > 0, got {min_interval}\"\n        assert (\n            min_interval <= max_interval\n        ), f\"min_interval should be <= max_interval, got {min_interval} and {max_interval}\"\n        assert id_ref in ids_all\n        pos_ref = ids_all.index(id_ref)\n        all_possible_pos = np.arange(pos_ref, len(ids_all))\n\n        remaining_sum = len(ids_all) - 1 - pos_ref\n\n        if remaining_sum >= num_views - 1:\n            if remaining_sum == num_views - 1:\n                assert ids_all[-num_views] == id_ref\n                return [pos_ref + i for i in range(num_views)], True\n            max_interval = min(max_interval, 2 * remaining_sum // (num_views - 1))\n            intervals = [\n                rng.choice(range(min_interval, max_interval + 1))\n                for _ in range(num_views - 1)\n            ]\n\n            # if video or collection\n            if rng.random() < video_prob:\n                # if fixed interval or random\n                if rng.random() < fix_interval_prob:\n                    # regular interval\n                    fixed_interval = rng.choice(\n                        range(\n                            1,\n                            min(remaining_sum // (num_views - 1) + 1, max_interval + 1),\n                        )\n                    )\n                    intervals = [fixed_interval for _ in range(num_views - 1)]\n                is_video = True\n            else:\n                is_video = False\n\n            pos = list(itertools.accumulate([pos_ref] + intervals))\n            pos = [p for p in pos if p < len(ids_all)]\n            pos_candidates = [p for p in all_possible_pos if p not in pos]\n            pos = (\n                pos\n                + rng.choice(\n                    pos_candidates, num_views - len(pos), replace=False\n                ).tolist()\n            )\n\n            pos = (\n                sorted(pos)\n                if is_video\n                else self.blockwise_shuffle(pos, rng, block_shuffle)\n            )\n        else:\n            # assert self.allow_repeat\n            uniq_num = remaining_sum\n            new_pos_ref = rng.choice(np.arange(pos_ref + 1))\n            new_remaining_sum = len(ids_all) - 1 - new_pos_ref\n            new_max_interval = min(max_interval, new_remaining_sum // (uniq_num - 1))\n            new_intervals = [\n                rng.choice(range(1, new_max_interval + 1)) for _ in range(uniq_num - 1)\n            ]\n\n            revisit_random = rng.random()\n            video_random = rng.random()\n\n            if rng.random() < fix_interval_prob and video_random < video_prob:\n                # regular interval\n                fixed_interval = rng.choice(range(1, new_max_interval + 1))\n                new_intervals = [fixed_interval for _ in range(uniq_num - 1)]\n            pos = list(itertools.accumulate([new_pos_ref] + new_intervals))\n\n            is_video = False\n            if revisit_random < 0.5 or video_prob == 1.0:  # revisit, video / collection\n                is_video = video_random < video_prob\n                pos = (\n                    self.blockwise_shuffle(pos, rng, block_shuffle)\n                    if not is_video\n                    else pos\n                )\n                num_full_repeat = num_views // uniq_num\n                pos = (\n                    pos * num_full_repeat\n                    + pos[: num_views - len(pos) * num_full_repeat]\n                )\n            elif revisit_random < 0.9:  # random\n                pos = rng.choice(pos, num_views, replace=True)\n            else:  # ordered\n                pos = sorted(rng.choice(pos, num_views, replace=True))\n        assert len(pos) == num_views\n        return pos, is_video\n\n    def get_img_and_ray_masks(self, is_metric, v, rng, p=[0.8, 0.15, 0.05]):\n        # generate img mask and raymap mask\n        if v == 0 or (not is_metric):\n            img_mask = True\n            raymap_mask = False\n        else:\n            rand_val = rng.random()\n            if rand_val < p[0]:\n                img_mask = True\n                raymap_mask = False\n            elif rand_val < p[0] + p[1]:\n                img_mask = False\n                raymap_mask = True\n            else:\n                img_mask = True\n                raymap_mask = True\n        return img_mask, raymap_mask\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def __repr__(self):\n        resolutions_str = \"[\" + \";\".join(f\"{w}x{h}\" for w, h in self._resolutions) + \"]\"\n        return (\n            f\"\"\"{type(self).__name__}({self.get_stats()},\n            {self.num_views=},\n            {self.split=},\n            {self.seed=},\n            resolutions={resolutions_str},\n            {self.transform=})\"\"\".replace(\n                \"self.\", \"\"\n            )\n            .replace(\"\\n\", \"\")\n            .replace(\"   \", \"\")\n        )\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        raise NotImplementedError()\n\n    def __getitem__(self, idx):\n        # print(\"Receiving:\" , idx)\n        if isinstance(idx, (tuple, list, np.ndarray)):\n            # the idx is specifying the aspect-ratio\n            idx, ar_idx, nview = idx # start_id, selected type (eg., resolution), view number\n        else:\n            assert len(self._resolutions) == 1\n            ar_idx = 0\n            nview = self.num_views\n\n        assert nview >= 1 and nview <= self.num_views\n        # set-up the rng\n        if self.seed:  # reseed for each __getitem__\n            self._rng = np.random.default_rng(seed=self.seed + idx)\n        elif not hasattr(self, \"_rng\"):\n            seed = torch.randint(0, 2**32, (1,)).item()\n            self._rng = np.random.default_rng(seed=seed)\n\n        if self.aug_crop > 1 and self.seq_aug_crop:\n            self.delta_target_resolution = self._rng.integers(0, self.aug_crop)\n\n        # over-loaded code\n        resolution = self._resolutions[\n            ar_idx\n        ]  # DO NOT CHANGE THIS (compatible with BatchedRandomSampler)\n        views = self._get_views(idx, resolution, self._rng, nview)\n        assert len(views) == nview\n\n        if \"camera_pose\" not in views[0]:\n            views[0][\"camera_pose\"] = np.ones((4, 4), dtype=np.float32)\n        first_view_camera_pose = views[0][\"camera_pose\"]\n        transform = SeqColorJitter() if self.is_seq_color_jitter else self.transform\n\n        for v, view in enumerate(views):\n            assert (\n                \"pts3d\" not in view\n            ), f\"pts3d should not be there, they will be computed afterwards based on intrinsics+depthmap for view {view_name(view)}\"\n            view[\"idx\"] = (idx, ar_idx, v)\n\n            # encode the image\n            width, height = view[\"img\"].size\n\n            view[\"true_shape\"] = np.int32((height, width))\n            view[\"img\"] = transform(view[\"img\"])\n            view[\"sky_mask\"] = view[\"depthmap\"] < 0\n\n            assert \"camera_intrinsics\" in view\n            if \"camera_pose\" not in view:\n                view[\"camera_pose\"] = np.full((4, 4), np.nan, dtype=np.float32)\n            else:\n                assert np.isfinite(\n                    view[\"camera_pose\"]\n                ).all(), f\"NaN in camera pose for view {view_name(view)}\"\n\n            ray_map = get_ray_map(\n                first_view_camera_pose,\n                view[\"camera_pose\"],\n                view[\"camera_intrinsics\"],\n                height,\n                width,\n            )   # camera_pose: c2w\n            view[\"ray_map\"] = ray_map.astype(np.float32)\n\n            assert \"pts3d\" not in view\n            assert \"valid_mask\" not in view\n            assert np.isfinite(\n                view[\"depthmap\"]\n            ).all(), f\"NaN in depthmap for view {view_name(view)}\"\n            pts3d, valid_mask = depthmap_to_absolute_camera_coordinates(**view)\n\n            view[\"pts3d\"] = pts3d\n            view[\"valid_mask\"] = valid_mask & np.isfinite(pts3d).all(axis=-1)\n\n            # check all datatypes\n            for key, val in view.items():\n                res, err_msg = is_good_type(key, val)\n                assert res, f\"{err_msg} with {key}={val} for view {view_name(view)}\"\n            K = view[\"camera_intrinsics\"]\n\n        if self.n_corres > 0:\n            ref_view = views[0]\n            for view in views:\n                corres1, corres2, valid = extract_correspondences_from_pts3d(\n                    ref_view, view, self.n_corres, self._rng, nneg=self.nneg\n                )\n                view[\"corres\"] = (corres1, corres2)\n                view[\"valid_corres\"] = valid\n\n        # last thing done!\n        for view in views:\n            view[\"rng\"] = int.from_bytes(self._rng.bytes(4), \"big\")\n        return views\n\n    def _set_resolutions(self, resolutions):\n        assert resolutions is not None, \"undefined resolution\"\n\n        if not isinstance(resolutions, list):\n            resolutions = [resolutions]\n\n        self._resolutions = []\n        for resolution in resolutions:\n            if isinstance(resolution, int):\n                width = height = resolution\n            else:\n                width, height = resolution\n            assert isinstance(\n                width, int\n            ), f\"Bad type for {width=} {type(width)=}, should be int\"\n            assert isinstance(\n                height, int\n            ), f\"Bad type for {height=} {type(height)=}, should be int\"\n            self._resolutions.append((width, height))\n\n    def _crop_resize_if_necessary(\n        self, image, depthmap, intrinsics, resolution, rng=None, info=None\n    ):\n        \"\"\"This function:\n        - first downsizes the image with LANCZOS inteprolation,\n          which is better than bilinear interpolation\n        \"\"\"\n        if not isinstance(image, PIL.Image.Image):\n            image = PIL.Image.fromarray(image)\n\n        # downscale with lanczos interpolation so that image.size == resolution\n        # cropping centered on the principal point\n        W, H = image.size\n        cx, cy = intrinsics[:2, 2].round().astype(int)\n        min_margin_x = min(cx, W - cx)\n        min_margin_y = min(cy, H - cy)\n        assert min_margin_x > W / 5, f\"Bad principal point in view={info}\"\n        assert min_margin_y > H / 5, f\"Bad principal point in view={info}\"\n        # the new window will be a rectangle of size (2*min_margin_x, 2*min_margin_y) centered on (cx,cy)\n        l, t = cx - min_margin_x, cy - min_margin_y\n        r, b = cx + min_margin_x, cy + min_margin_y\n        crop_bbox = (l, t, r, b)\n        image, depthmap, intrinsics = cropping.crop_image_depthmap(\n            image, depthmap, intrinsics, crop_bbox\n        )\n\n        # transpose the resolution if necessary\n        W, H = image.size  # new size\n\n        # high-quality Lanczos down-scaling\n        target_resolution = np.array(resolution)\n        if self.aug_crop > 1:\n            target_resolution += (\n                rng.integers(0, self.aug_crop)\n                if not self.seq_aug_crop\n                else self.delta_target_resolution\n            )\n        image, depthmap, intrinsics = cropping.rescale_image_depthmap(\n            image, depthmap, intrinsics, target_resolution\n        )\n\n        # actual cropping (if necessary) with bilinear interpolation\n        intrinsics2 = cropping.camera_matrix_of_crop(\n            intrinsics, image.size, resolution, offset_factor=0.5\n        )\n        crop_bbox = cropping.bbox_from_intrinsics_in_out(\n            intrinsics, intrinsics2, resolution\n        )\n        image, depthmap, intrinsics2 = cropping.crop_image_depthmap(\n            image, depthmap, intrinsics, crop_bbox\n        )\n\n        return image, depthmap, intrinsics2\n\n    def _crop_resize_if_necessary_mask(\n        self, image, depthmap, mask, intrinsics, resolution, rng=None, info=None\n    ):\n        \"\"\"This function:\n        - first downsizes the image with LANCZOS inteprolation,\n          which is better than bilinear interpolation\n          also processes mask with the same transformations as image\n        \"\"\"\n        if not isinstance(image, PIL.Image.Image):\n            image = PIL.Image.fromarray(image)\n        if not isinstance(mask, PIL.Image.Image):\n            mask = PIL.Image.fromarray(mask)\n\n        # downscale with lanczos interpolation so that image.size == resolution\n        # cropping centered on the principal point\n        W, H = image.size\n        cx, cy = intrinsics[:2, 2].round().astype(int)\n        min_margin_x = min(cx, W - cx)\n        min_margin_y = min(cy, H - cy)\n        assert min_margin_x > W / 5, f\"Bad principal point in view={info}\"\n        assert min_margin_y > H / 5, f\"Bad principal point in view={info}\"\n        # the new window will be a rectangle of size (2*min_margin_x, 2*min_margin_y) centered on (cx,cy)\n        l, t = cx - min_margin_x, cy - min_margin_y\n        r, b = cx + min_margin_x, cy + min_margin_y\n        crop_bbox = (l, t, r, b)\n        image, depthmap, mask, intrinsics = cropping.cro\n# ... truncated ...","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.get_ray_map","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.get_ray_map#L13-L23","kind":"function","name":"get_ray_map","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":13,"end_line":23,"context_start_line":1,"context_end_line":43,"code":"import PIL\nimport numpy as np\nimport torch\nimport random\nimport itertools\nfrom dust3r.datasets.base.easy_dataset import EasyDataset\nfrom dust3r.datasets.utils.transforms import ImgNorm, SeqColorJitter\nfrom dust3r.utils.geometry import depthmap_to_absolute_camera_coordinates\nimport dust3r.datasets.utils.cropping as cropping\nfrom dust3r.datasets.utils.corr import extract_correspondences_from_pts3d\nimport torchvision.transforms as tvf\n\ndef get_ray_map(c2w1, c2w2, intrinsics, h, w):\n    c2w = np.linalg.inv(c2w1) @ c2w2\n    i, j = np.meshgrid(np.arange(w), np.arange(h), indexing=\"xy\")\n    grid = np.stack([i, j, np.ones_like(i)], axis=-1)\n    ro = c2w[:3, 3]\n    rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n    rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n    rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n    ro = np.broadcast_to(ro, (h, w, 3))\n    ray_map = np.concatenate([ro, rd], axis=-1)\n    return ray_map\n\n\nclass BaseMultiViewDataset(EasyDataset):\n    \"\"\"Define all basic options.\n\n    Usage:\n        class MyDataset (BaseMultiViewDataset):\n            def _get_views(self, idx, rng):\n                # overload here\n                views = []\n                views.append(dict(img=, ...))\n                return views\n    \"\"\"\n\n    def __init__(\n        self,\n        *,  # only keyword arguments\n        num_views=None,\n        split=None,\n        resolution=None,  # square_size or (width, height) or list of [(width,height), ...]","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.BaseMultiViewDataset","uri":"program://Human3R/class/src.dust3r.datasets.base.base_multiview_dataset.BaseMultiViewDataset#L26-L556","kind":"class","name":"BaseMultiViewDataset","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":26,"end_line":556,"context_start_line":6,"context_end_line":576,"code":"from dust3r.datasets.base.easy_dataset import EasyDataset\nfrom dust3r.datasets.utils.transforms import ImgNorm, SeqColorJitter\nfrom dust3r.utils.geometry import depthmap_to_absolute_camera_coordinates\nimport dust3r.datasets.utils.cropping as cropping\nfrom dust3r.datasets.utils.corr import extract_correspondences_from_pts3d\nimport torchvision.transforms as tvf\n\ndef get_ray_map(c2w1, c2w2, intrinsics, h, w):\n    c2w = np.linalg.inv(c2w1) @ c2w2\n    i, j = np.meshgrid(np.arange(w), np.arange(h), indexing=\"xy\")\n    grid = np.stack([i, j, np.ones_like(i)], axis=-1)\n    ro = c2w[:3, 3]\n    rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n    rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n    rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n    ro = np.broadcast_to(ro, (h, w, 3))\n    ray_map = np.concatenate([ro, rd], axis=-1)\n    return ray_map\n\n\nclass BaseMultiViewDataset(EasyDataset):\n    \"\"\"Define all basic options.\n\n    Usage:\n        class MyDataset (BaseMultiViewDataset):\n            def _get_views(self, idx, rng):\n                # overload here\n                views = []\n                views.append(dict(img=, ...))\n                return views\n    \"\"\"\n\n    def __init__(\n        self,\n        *,  # only keyword arguments\n        num_views=None,\n        split=None,\n        resolution=None,  # square_size or (width, height) or list of [(width,height), ...]\n        transform=ImgNorm,\n        aug_crop=False,\n        n_corres=0,\n        nneg=0,\n        seed=None,\n        allow_repeat=False,\n        seq_aug_crop=False,\n    ):\n        assert num_views is not None, \"undefined num_views\"\n        self.num_views = num_views\n        self.split = split\n        self._set_resolutions(resolution)\n\n        self.n_corres = n_corres\n        self.nneg = nneg\n        assert (\n            self.n_corres == \"all\"\n            or isinstance(self.n_corres, int)\n            or (\n                isinstance(self.n_corres, list) and len(self.n_corres) == self.num_views\n            )\n        ), f\"Error, n_corres should either be 'all', a single integer or a list of length {self.num_views}\"\n        assert (\n            self.nneg == 0 or self.n_corres != \"all\"\n        ), \"nneg should be 0 if n_corres is all\"\n\n        self.is_seq_color_jitter = False\n        if isinstance(transform, str):\n            transform = eval(transform)\n        if transform == SeqColorJitter:\n            transform = SeqColorJitter()\n            self.is_seq_color_jitter = True\n        self.transform = transform\n\n        self.aug_crop = aug_crop\n        self.seed = seed\n        self.allow_repeat = allow_repeat\n        self.seq_aug_crop = seq_aug_crop\n\n    def __len__(self):\n        return len(self.scenes)\n\n    @staticmethod\n    def efficient_random_intervals(\n        start,\n        num_elements,\n        interval_range,\n        fixed_interval_prob=0.8,\n        weights=None,\n        seed=42,\n    ):\n        if random.random() < fixed_interval_prob:\n            intervals = random.choices(interval_range, weights=weights) * (\n                num_elements - 1\n            )\n        else:\n            intervals = [\n                random.choices(interval_range, weights=weights)[0]\n                for _ in range(num_elements - 1)\n            ]\n        return list(itertools.accumulate([start] + intervals))\n\n    def sample_based_on_timestamps(self, i, timestamps, num_views, interval=1):\n        time_diffs = np.abs(timestamps - timestamps[i])\n        ids_candidate = np.where(time_diffs < interval)[0]\n        ids_candidate = np.sort(ids_candidate)\n        if (self.allow_repeat and len(ids_candidate) < num_views // 3) or (\n            len(ids_candidate) < num_views\n        ):\n            return []\n        ids_sel_list = []\n        ids_candidate_left = ids_candidate.copy()\n        while len(ids_candidate_left) >= num_views:\n            ids_sel = np.random.choice(ids_candidate_left, num_views, replace=False)\n            ids_sel_list.append(sorted(ids_sel))\n            ids_candidate_left = np.setdiff1d(ids_candidate_left, ids_sel)\n\n        if len(ids_candidate_left) > 0 and len(ids_candidate) >= num_views:\n            ids_sel = np.concatenate(\n                [\n                    ids_candidate_left,\n                    np.random.choice(\n                        np.setdiff1d(ids_candidate, ids_candidate_left),\n                        num_views - len(ids_candidate_left),\n                        replace=False,\n                    ),\n                ]\n            )\n            ids_sel_list.append(sorted(ids_sel))\n\n        if self.allow_repeat:\n            ids_sel_list.append(\n                sorted(np.random.choice(ids_candidate, num_views, replace=True))\n            )\n\n        # add sequences with fixed intervals (all possible intervals)\n        pos_i = np.where(ids_candidate == i)[0][0]\n        curr_interval = 1\n        stop = len(ids_candidate) < num_views\n        while not stop:\n            pos_sel = [pos_i]\n            count = 0\n            while len(pos_sel) < num_views:\n                if count % 2 == 0:\n                    curr_pos_i = pos_sel[-1] + curr_interval\n                    if curr_pos_i >= len(ids_candidate):\n                        stop = True\n                        break\n                    pos_sel.append(curr_pos_i)\n                else:\n                    curr_pos_i = pos_sel[0] - curr_interval\n                    if curr_pos_i < 0:\n                        stop = True\n                        break\n                    pos_sel.insert(0, curr_pos_i)\n                count += 1\n            if not stop and len(pos_sel) == num_views:\n                ids_sel = sorted([ids_candidate[pos] for pos in pos_sel])\n                if ids_sel not in ids_sel_list:\n                    ids_sel_list.append(ids_sel)\n            curr_interval += 1\n        return ids_sel_list\n\n    @staticmethod\n    def blockwise_shuffle(x, rng, block_shuffle):\n        if block_shuffle is None:\n            return rng.permutation(x).tolist()\n        else:\n            assert block_shuffle > 0\n            blocks = [x[i : i + block_shuffle] for i in range(0, len(x), block_shuffle)]\n            shuffled_blocks = [rng.permutation(block).tolist() for block in blocks]\n            shuffled_list = [item for block in shuffled_blocks for item in block]\n            return shuffled_list\n\n    def get_seq_from_start_id(\n        self,\n        num_views,\n        id_ref,\n        ids_all,\n        rng,\n        min_interval=1,\n        max_interval=25,\n        video_prob=0.5,\n        fix_interval_prob=0.5,\n        block_shuffle=None,\n    ):\n        \"\"\"\n        args:\n            num_views: number of views to return\n            id_ref: the reference id (first id)\n            ids_all: all the ids\n            rng: random number generator\n            max_interval: maximum interval between two views\n        returns:\n            pos: list of positions of the views in ids_all, i.e., index for ids_all\n            is_video: True if the views are consecutive\n        \"\"\"\n        assert min_interval > 0, f\"min_interval should be > 0, got {min_interval}\"\n        assert (\n            min_interval <= max_interval\n        ), f\"min_interval should be <= max_interval, got {min_interval} and {max_interval}\"\n        assert id_ref in ids_all\n        pos_ref = ids_all.index(id_ref)\n        all_possible_pos = np.arange(pos_ref, len(ids_all))\n\n        remaining_sum = len(ids_all) - 1 - pos_ref\n\n        if remaining_sum >= num_views - 1:\n            if remaining_sum == num_views - 1:\n                assert ids_all[-num_views] == id_ref\n                return [pos_ref + i for i in range(num_views)], True\n            max_interval = min(max_interval, 2 * remaining_sum // (num_views - 1))\n            intervals = [\n                rng.choice(range(min_interval, max_interval + 1))\n                for _ in range(num_views - 1)\n            ]\n\n            # if video or collection\n            if rng.random() < video_prob:\n                # if fixed interval or random\n                if rng.random() < fix_interval_prob:\n                    # regular interval\n                    fixed_interval = rng.choice(\n                        range(\n                            1,\n                            min(remaining_sum // (num_views - 1) + 1, max_interval + 1),\n                        )\n                    )\n                    intervals = [fixed_interval for _ in range(num_views - 1)]\n                is_video = True\n            else:\n                is_video = False\n\n            pos = list(itertools.accumulate([pos_ref] + intervals))\n            pos = [p for p in pos if p < len(ids_all)]\n            pos_candidates = [p for p in all_possible_pos if p not in pos]\n            pos = (\n                pos\n                + rng.choice(\n                    pos_candidates, num_views - len(pos), replace=False\n                ).tolist()\n            )\n\n            pos = (\n                sorted(pos)\n                if is_video\n                else self.blockwise_shuffle(pos, rng, block_shuffle)\n            )\n        else:\n            # assert self.allow_repeat\n            uniq_num = remaining_sum\n            new_pos_ref = rng.choice(np.arange(pos_ref + 1))\n            new_remaining_sum = len(ids_all) - 1 - new_pos_ref\n            new_max_interval = min(max_interval, new_remaining_sum // (uniq_num - 1))\n            new_intervals = [\n                rng.choice(range(1, new_max_interval + 1)) for _ in range(uniq_num - 1)\n            ]\n\n            revisit_random = rng.random()\n            video_random = rng.random()\n\n            if rng.random() < fix_interval_prob and video_random < video_prob:\n                # regular interval\n                fixed_interval = rng.choice(range(1, new_max_interval + 1))\n                new_intervals = [fixed_interval for _ in range(uniq_num - 1)]\n            pos = list(itertools.accumulate([new_pos_ref] + new_intervals))\n\n            is_video = False\n            if revisit_random < 0.5 or video_prob == 1.0:  # revisit, video / collection\n                is_video = video_random < video_prob\n                pos = (\n                    self.blockwise_shuffle(pos, rng, block_shuffle)\n                    if not is_video\n                    else pos\n                )\n                num_full_repeat = num_views // uniq_num\n                pos = (\n                    pos * num_full_repeat\n                    + pos[: num_views - len(pos) * num_full_repeat]\n                )\n            elif revisit_random < 0.9:  # random\n                pos = rng.choice(pos, num_views, replace=True)\n            else:  # ordered\n                pos = sorted(rng.choice(pos, num_views, replace=True))\n        assert len(pos) == num_views\n        return pos, is_video\n\n    def get_img_and_ray_masks(self, is_metric, v, rng, p=[0.8, 0.15, 0.05]):\n        # generate img mask and raymap mask\n        if v == 0 or (not is_metric):\n            img_mask = True\n            raymap_mask = False\n        else:\n            rand_val = rng.random()\n            if rand_val < p[0]:\n                img_mask = True\n                raymap_mask = False\n            elif rand_val < p[0] + p[1]:\n                img_mask = False\n                raymap_mask = True\n            else:\n                img_mask = True\n                raymap_mask = True\n        return img_mask, raymap_mask\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def __repr__(self):\n        resolutions_str = \"[\" + \";\".join(f\"{w}x{h}\" for w, h in self._resolutions) + \"]\"\n        return (\n            f\"\"\"{type(self).__name__}({self.get_stats()},\n            {self.num_views=},\n            {self.split=},\n            {self.seed=},\n            resolutions={resolutions_str},\n            {self.transform=})\"\"\".replace(\n                \"self.\", \"\"\n            )\n            .replace(\"\\n\", \"\")\n            .replace(\"   \", \"\")\n        )\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        raise NotImplementedError()\n\n    def __getitem__(self, idx):\n        # print(\"Receiving:\" , idx)\n        if isinstance(idx, (tuple, list, np.ndarray)):\n            # the idx is specifying the aspect-ratio\n            idx, ar_idx, nview = idx # start_id, selected type (eg., resolution), view number\n        else:\n            assert len(self._resolutions) == 1\n            ar_idx = 0\n            nview = self.num_views\n\n        assert nview >= 1 and nview <= self.num_views\n        # set-up the rng\n        if self.seed:  # reseed for each __getitem__\n            self._rng = np.random.default_rng(seed=self.seed + idx)\n        elif not hasattr(self, \"_rng\"):\n            seed = torch.randint(0, 2**32, (1,)).item()\n            self._rng = np.random.default_rng(seed=seed)\n\n        if self.aug_crop > 1 and self.seq_aug_crop:\n            self.delta_target_resolution = self._rng.integers(0, self.aug_crop)\n\n        # over-loaded code\n        resolution = self._resolutions[\n            ar_idx\n        ]  # DO NOT CHANGE THIS (compatible with BatchedRandomSampler)\n        views = self._get_views(idx, resolution, self._rng, nview)\n        assert len(views) == nview\n\n        if \"camera_pose\" not in views[0]:\n            views[0][\"camera_pose\"] = np.ones((4, 4), dtype=np.float32)\n        first_view_camera_pose = views[0][\"camera_pose\"]\n        transform = SeqColorJitter() if self.is_seq_color_jitter else self.transform\n\n        for v, view in enumerate(views):\n            assert (\n                \"pts3d\" not in view\n            ), f\"pts3d should not be there, they will be computed afterwards based on intrinsics+depthmap for view {view_name(view)}\"\n            view[\"idx\"] = (idx, ar_idx, v)\n\n            # encode the image\n            width, height = view[\"img\"].size\n\n            view[\"true_shape\"] = np.int32((height, width))\n            view[\"img\"] = transform(view[\"img\"])\n            view[\"sky_mask\"] = view[\"depthmap\"] < 0\n\n            assert \"camera_intrinsics\" in view\n            if \"camera_pose\" not in view:\n                view[\"camera_pose\"] = np.full((4, 4), np.nan, dtype=np.float32)\n            else:\n                assert np.isfinite(\n                    view[\"camera_pose\"]\n                ).all(), f\"NaN in camera pose for view {view_name(view)}\"\n\n            ray_map = get_ray_map(\n                first_view_camera_pose,\n                view[\"camera_pose\"],\n                view[\"camera_intrinsics\"],\n                height,\n                width,\n            )   # camera_pose: c2w\n            view[\"ray_map\"] = ray_map.astype(np.float32)\n\n            assert \"pts3d\" not in view\n            assert \"valid_mask\" not in view\n            assert np.isfinite(\n                view[\"depthmap\"]\n            ).all(), f\"NaN in depthmap for view {view_name(view)}\"\n            pts3d, valid_mask = depthmap_to_absolute_camera_coordinates(**view)\n\n            view[\"pts3d\"] = pts3d\n            view[\"valid_mask\"] = valid_mask & np.isfinite(pts3d).all(axis=-1)\n\n            # check all datatypes\n            for key, val in view.items():\n                res, err_msg = is_good_type(key, val)\n                assert res, f\"{err_msg} with {key}={val} for view {view_name(view)}\"\n            K = view[\"camera_intrinsics\"]\n\n        if self.n_corres > 0:\n            ref_view = views[0]\n            for view in views:\n                corres1, corres2, valid = extract_correspondences_from_pts3d(\n                    ref_view, view, self.n_corres, self._rng, nneg=self.nneg\n                )\n                view[\"corres\"] = (corres1, corres2)\n                view[\"valid_corres\"] = valid\n\n        # last thing done!\n        for view in views:\n            view[\"rng\"] = int.from_bytes(self._rng.bytes(4), \"big\")\n        return views\n\n    def _set_resolutions(self, resolutions):\n        assert resolutions is not None, \"undefined resolution\"\n\n        if not isinstance(resolutions, list):\n            resolutions = [resolutions]\n\n        self._resolutions = []\n        for resolution in resolutions:\n            if isinstance(resolution, int):\n                width = height = resolution\n            else:\n                width, height = resolution\n            assert isinstance(\n                width, int\n            ), f\"Bad type for {width=} {type(width)=}, should be int\"\n            assert isinstance(\n                height, int\n            ), f\"Bad type for {height=} {type(height)=}, should be int\"\n            self._resolutions.append((width, height))\n\n    def _crop_resize_if_necessary(\n        self, image, depthmap, intrinsics, resolution, rng=None, info=None\n    ):\n        \"\"\"This function:\n        - first downsizes the image with LANCZOS inteprolation,\n          which is better than bilinear interpolation\n        \"\"\"\n        if not isinstance(image, PIL.Image.Image):\n            image = PIL.Image.fromarray(image)\n\n        # downscale with lanczos interpolation so that image.size == resolution\n        # cropping centered on the principal point\n        W, H = image.size\n        cx, cy = intrinsics[:2, 2].round().astype(int)\n        min_margin_x = min(cx, W - cx)\n        min_margin_y = min(cy, H - cy)\n        assert min_margin_x > W / 5, f\"Bad principal point in view={info}\"\n        assert min_margin_y > H / 5, f\"Bad principal point in view={info}\"\n        # the new window will be a rectangle of size (2*min_margin_x, 2*min_margin_y) centered on (cx,cy)\n        l, t = cx - min_margin_x, cy - min_margin_y\n        r, b = cx + min_margin_x, cy + min_margin_y\n        crop_bbox = (l, t, r, b)\n        image, depthmap, intrinsics = cropping.crop_image_depthmap(\n            image, depthmap, intrinsics, crop_bbox\n        )\n\n        # transpose the resolution if necessary\n        W, H = image.size  # new size\n\n        # high-quality Lanczos down-scaling\n        target_resolution = np.array(resolution)\n        if self.aug_crop > 1:\n            target_resolution += (\n                rng.integers(0, self.aug_crop)\n                if not self.seq_aug_crop\n                else self.delta_target_resolution\n            )\n        image, depthmap, intrinsics = cropping.rescale_image_depthmap(\n            image, depthmap, intrinsics, target_resolution\n        )\n\n        # actual cropping (if necessary) with bilinear interpolation\n        intrinsics2 = cropping.camera_matrix_of_crop(\n            intrinsics, image.size, resolution, offset_factor=0.5\n        )\n        crop_bbox = cropping.bbox_from_intrinsics_in_out(\n            intrinsics, intrinsics2, resolution\n        )\n        image, depthmap, intrinsics2 = cropping.crop_image_depthmap(\n            image, depthmap, intrinsics, crop_bbox\n        )\n\n        return image, depthmap, intrinsics2\n\n    def _crop_resize_if_necessary_mask(\n        self, image, depthmap, mask, intrinsics, resolution, rng=None, info=None\n    ):\n        \"\"\"This function:\n        - first downsizes the image with LANCZOS inteprolation,\n          which is better than bilinear interpolation\n          also processes mask with the same transformations as image\n        \"\"\"\n        if not isinstance(image, PIL.Image.Image):\n            image = PIL.Image.fromarray(image)\n        if not isinstance(mask, PIL.Image.Image):\n            mask = PIL.Image.fromarray(mask)\n\n        # downscale with lanczos interpolation so that image.size == resolution\n        # cropping centered on the principal point\n        W, H = image.size\n        cx, cy = intrinsics[:2, 2].round().astype(int)\n        min_margin_x = min(cx, W - cx)\n        min_margin_y = min(cy, H - cy)\n        assert min_margin_x > W / 5, f\"Bad principal point in view={info}\"\n        assert min_margin_y > H / 5, f\"Bad principal point in view={info}\"\n        # the new window will be a rectangle of size (2*min_margin_x, 2*min_margin_y) centered on (cx,cy)\n        l, t = cx - min_margin_x, cy - min_margin_y\n        r, b = cx + min_margin_x, cy + min_margin_y\n        crop_bbox = (l, t, r, b)\n        image, depthmap, mask, intrinsics = cropping.crop_image_depthmap_mask(\n            image, depthmap, mask, intrinsics, crop\n# ... truncated ...","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.is_good_type","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.is_good_type#L559-L565","kind":"function","name":"is_good_type","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":559,"end_line":565,"context_start_line":539,"context_end_line":585,"code":"        )\n\n        # actual cropping (if necessary) with bilinear interpolation\n        intrinsics2 = cropping.camera_matrix_of_crop(\n            intrinsics, image.size, resolution, offset_factor=0.5\n        )\n        crop_bbox = cropping.bbox_from_intrinsics_in_out(\n            intrinsics, intrinsics2, resolution\n        )\n        image, depthmap, mask, intrinsics2 = cropping.crop_image_depthmap_mask(\n            image, depthmap, mask, intrinsics, crop_bbox\n        )\n\n        if isinstance(mask, PIL.Image.Image):\n            mask = tvf.ToTensor()(mask)\n            mask = (mask[0] > 0.5).float()\n\n        return image, depthmap, mask, intrinsics2\n\n\ndef is_good_type(key, v):\n    \"\"\"returns (is_good, err_msg)\"\"\"\n    if isinstance(v, (str, int, tuple)):\n        return True, None\n    if v.dtype not in (np.float32, torch.float32, bool, np.int32, np.int64, np.uint8):\n        return False, f\"bad {v.dtype=}\"\n    return True, None\n\n\ndef view_name(view, batch_index=None):\n    def sel(x):\n        return x[batch_index] if batch_index not in (None, slice(None)) else x\n\n    db = sel(view[\"dataset\"])\n    label = sel(view[\"label\"])\n    instance = sel(view[\"instance\"])\n    return f\"{db}/{label}/{instance}\"\n\n\ndef transpose_to_landscape(view):\n    height, width = view[\"true_shape\"]\n\n    if width < height:\n        # rectify portrait to landscape\n        assert view[\"img\"].shape == (3, height, width)\n        view[\"img\"] = view[\"img\"].swapaxes(1, 2)\n","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.view_name","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.view_name#L568-L575","kind":"function","name":"view_name","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":568,"end_line":575,"context_start_line":548,"context_end_line":595,"code":"        image, depthmap, mask, intrinsics2 = cropping.crop_image_depthmap_mask(\n            image, depthmap, mask, intrinsics, crop_bbox\n        )\n\n        if isinstance(mask, PIL.Image.Image):\n            mask = tvf.ToTensor()(mask)\n            mask = (mask[0] > 0.5).float()\n\n        return image, depthmap, mask, intrinsics2\n\n\ndef is_good_type(key, v):\n    \"\"\"returns (is_good, err_msg)\"\"\"\n    if isinstance(v, (str, int, tuple)):\n        return True, None\n    if v.dtype not in (np.float32, torch.float32, bool, np.int32, np.int64, np.uint8):\n        return False, f\"bad {v.dtype=}\"\n    return True, None\n\n\ndef view_name(view, batch_index=None):\n    def sel(x):\n        return x[batch_index] if batch_index not in (None, slice(None)) else x\n\n    db = sel(view[\"dataset\"])\n    label = sel(view[\"label\"])\n    instance = sel(view[\"instance\"])\n    return f\"{db}/{label}/{instance}\"\n\n\ndef transpose_to_landscape(view):\n    height, width = view[\"true_shape\"]\n\n    if width < height:\n        # rectify portrait to landscape\n        assert view[\"img\"].shape == (3, height, width)\n        view[\"img\"] = view[\"img\"].swapaxes(1, 2)\n\n        assert view[\"valid_mask\"].shape == (height, width)\n        view[\"valid_mask\"] = view[\"valid_mask\"].swapaxes(0, 1)\n\n        assert view[\"depthmap\"].shape == (height, width)\n        view[\"depthmap\"] = view[\"depthmap\"].swapaxes(0, 1)\n\n        assert view[\"pts3d\"].shape == (height, width, 3)\n        view[\"pts3d\"] = view[\"pts3d\"].swapaxes(0, 1)\n\n        # transpose x and y pixels","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.transpose_to_landscape","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.transpose_to_landscape#L578-L607","kind":"function","name":"transpose_to_landscape","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":578,"end_line":607,"context_start_line":558,"context_end_line":607,"code":"\ndef is_good_type(key, v):\n    \"\"\"returns (is_good, err_msg)\"\"\"\n    if isinstance(v, (str, int, tuple)):\n        return True, None\n    if v.dtype not in (np.float32, torch.float32, bool, np.int32, np.int64, np.uint8):\n        return False, f\"bad {v.dtype=}\"\n    return True, None\n\n\ndef view_name(view, batch_index=None):\n    def sel(x):\n        return x[batch_index] if batch_index not in (None, slice(None)) else x\n\n    db = sel(view[\"dataset\"])\n    label = sel(view[\"label\"])\n    instance = sel(view[\"instance\"])\n    return f\"{db}/{label}/{instance}\"\n\n\ndef transpose_to_landscape(view):\n    height, width = view[\"true_shape\"]\n\n    if width < height:\n        # rectify portrait to landscape\n        assert view[\"img\"].shape == (3, height, width)\n        view[\"img\"] = view[\"img\"].swapaxes(1, 2)\n\n        assert view[\"valid_mask\"].shape == (height, width)\n        view[\"valid_mask\"] = view[\"valid_mask\"].swapaxes(0, 1)\n\n        assert view[\"depthmap\"].shape == (height, width)\n        view[\"depthmap\"] = view[\"depthmap\"].swapaxes(0, 1)\n\n        assert view[\"pts3d\"].shape == (height, width, 3)\n        view[\"pts3d\"] = view[\"pts3d\"].swapaxes(0, 1)\n\n        # transpose x and y pixels\n        view[\"camera_intrinsics\"] = view[\"camera_intrinsics\"][[1, 0, 2]]\n\n        assert view[\"ray_map\"].shape == (height, width, 6)\n        view[\"ray_map\"] = view[\"ray_map\"].swapaxes(0, 1)\n\n        assert view[\"sky_mask\"].shape == (height, width)\n        view[\"sky_mask\"] = view[\"sky_mask\"].swapaxes(0, 1)\n\n        if \"corres\" in view:\n            # transpose correspondences x and y\n            view[\"corres\"][0] = view[\"corres\"][0][:, [1, 0]]\n            view[\"corres\"][1] = view[\"corres\"][1][:, [1, 0]]","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.__init__","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.__init__#L38-L81","kind":"function","name":"__init__","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":38,"end_line":81,"context_start_line":18,"context_end_line":101,"code":"    rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n    rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n    rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n    ro = np.broadcast_to(ro, (h, w, 3))\n    ray_map = np.concatenate([ro, rd], axis=-1)\n    return ray_map\n\n\nclass BaseMultiViewDataset(EasyDataset):\n    \"\"\"Define all basic options.\n\n    Usage:\n        class MyDataset (BaseMultiViewDataset):\n            def _get_views(self, idx, rng):\n                # overload here\n                views = []\n                views.append(dict(img=, ...))\n                return views\n    \"\"\"\n\n    def __init__(\n        self,\n        *,  # only keyword arguments\n        num_views=None,\n        split=None,\n        resolution=None,  # square_size or (width, height) or list of [(width,height), ...]\n        transform=ImgNorm,\n        aug_crop=False,\n        n_corres=0,\n        nneg=0,\n        seed=None,\n        allow_repeat=False,\n        seq_aug_crop=False,\n    ):\n        assert num_views is not None, \"undefined num_views\"\n        self.num_views = num_views\n        self.split = split\n        self._set_resolutions(resolution)\n\n        self.n_corres = n_corres\n        self.nneg = nneg\n        assert (\n            self.n_corres == \"all\"\n            or isinstance(self.n_corres, int)\n            or (\n                isinstance(self.n_corres, list) and len(self.n_corres) == self.num_views\n            )\n        ), f\"Error, n_corres should either be 'all', a single integer or a list of length {self.num_views}\"\n        assert (\n            self.nneg == 0 or self.n_corres != \"all\"\n        ), \"nneg should be 0 if n_corres is all\"\n\n        self.is_seq_color_jitter = False\n        if isinstance(transform, str):\n            transform = eval(transform)\n        if transform == SeqColorJitter:\n            transform = SeqColorJitter()\n            self.is_seq_color_jitter = True\n        self.transform = transform\n\n        self.aug_crop = aug_crop\n        self.seed = seed\n        self.allow_repeat = allow_repeat\n        self.seq_aug_crop = seq_aug_crop\n\n    def __len__(self):\n        return len(self.scenes)\n\n    @staticmethod\n    def efficient_random_intervals(\n        start,\n        num_elements,\n        interval_range,\n        fixed_interval_prob=0.8,\n        weights=None,\n        seed=42,\n    ):\n        if random.random() < fixed_interval_prob:\n            intervals = random.choices(interval_range, weights=weights) * (\n                num_elements - 1\n            )\n        else:\n            intervals = [\n                random.choices(interval_range, weights=weights)[0]","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.__len__","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.__len__#L83-L84","kind":"function","name":"__len__","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":83,"end_line":84,"context_start_line":63,"context_end_line":104,"code":"                isinstance(self.n_corres, list) and len(self.n_corres) == self.num_views\n            )\n        ), f\"Error, n_corres should either be 'all', a single integer or a list of length {self.num_views}\"\n        assert (\n            self.nneg == 0 or self.n_corres != \"all\"\n        ), \"nneg should be 0 if n_corres is all\"\n\n        self.is_seq_color_jitter = False\n        if isinstance(transform, str):\n            transform = eval(transform)\n        if transform == SeqColorJitter:\n            transform = SeqColorJitter()\n            self.is_seq_color_jitter = True\n        self.transform = transform\n\n        self.aug_crop = aug_crop\n        self.seed = seed\n        self.allow_repeat = allow_repeat\n        self.seq_aug_crop = seq_aug_crop\n\n    def __len__(self):\n        return len(self.scenes)\n\n    @staticmethod\n    def efficient_random_intervals(\n        start,\n        num_elements,\n        interval_range,\n        fixed_interval_prob=0.8,\n        weights=None,\n        seed=42,\n    ):\n        if random.random() < fixed_interval_prob:\n            intervals = random.choices(interval_range, weights=weights) * (\n                num_elements - 1\n            )\n        else:\n            intervals = [\n                random.choices(interval_range, weights=weights)[0]\n                for _ in range(num_elements - 1)\n            ]\n        return list(itertools.accumulate([start] + intervals))","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.efficient_random_intervals","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.efficient_random_intervals#L87-L104","kind":"function","name":"efficient_random_intervals","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":87,"end_line":104,"context_start_line":67,"context_end_line":124,"code":"            self.nneg == 0 or self.n_corres != \"all\"\n        ), \"nneg should be 0 if n_corres is all\"\n\n        self.is_seq_color_jitter = False\n        if isinstance(transform, str):\n            transform = eval(transform)\n        if transform == SeqColorJitter:\n            transform = SeqColorJitter()\n            self.is_seq_color_jitter = True\n        self.transform = transform\n\n        self.aug_crop = aug_crop\n        self.seed = seed\n        self.allow_repeat = allow_repeat\n        self.seq_aug_crop = seq_aug_crop\n\n    def __len__(self):\n        return len(self.scenes)\n\n    @staticmethod\n    def efficient_random_intervals(\n        start,\n        num_elements,\n        interval_range,\n        fixed_interval_prob=0.8,\n        weights=None,\n        seed=42,\n    ):\n        if random.random() < fixed_interval_prob:\n            intervals = random.choices(interval_range, weights=weights) * (\n                num_elements - 1\n            )\n        else:\n            intervals = [\n                random.choices(interval_range, weights=weights)[0]\n                for _ in range(num_elements - 1)\n            ]\n        return list(itertools.accumulate([start] + intervals))\n\n    def sample_based_on_timestamps(self, i, timestamps, num_views, interval=1):\n        time_diffs = np.abs(timestamps - timestamps[i])\n        ids_candidate = np.where(time_diffs < interval)[0]\n        ids_candidate = np.sort(ids_candidate)\n        if (self.allow_repeat and len(ids_candidate) < num_views // 3) or (\n            len(ids_candidate) < num_views\n        ):\n            return []\n        ids_sel_list = []\n        ids_candidate_left = ids_candidate.copy()\n        while len(ids_candidate_left) >= num_views:\n            ids_sel = np.random.choice(ids_candidate_left, num_views, replace=False)\n            ids_sel_list.append(sorted(ids_sel))\n            ids_candidate_left = np.setdiff1d(ids_candidate_left, ids_sel)\n\n        if len(ids_candidate_left) > 0 and len(ids_candidate) >= num_views:\n            ids_sel = np.concatenate(\n                [\n                    ids_candidate_left,","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.sample_based_on_timestamps","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.sample_based_on_timestamps#L106-L165","kind":"function","name":"sample_based_on_timestamps","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":106,"end_line":165,"context_start_line":86,"context_end_line":185,"code":"    @staticmethod\n    def efficient_random_intervals(\n        start,\n        num_elements,\n        interval_range,\n        fixed_interval_prob=0.8,\n        weights=None,\n        seed=42,\n    ):\n        if random.random() < fixed_interval_prob:\n            intervals = random.choices(interval_range, weights=weights) * (\n                num_elements - 1\n            )\n        else:\n            intervals = [\n                random.choices(interval_range, weights=weights)[0]\n                for _ in range(num_elements - 1)\n            ]\n        return list(itertools.accumulate([start] + intervals))\n\n    def sample_based_on_timestamps(self, i, timestamps, num_views, interval=1):\n        time_diffs = np.abs(timestamps - timestamps[i])\n        ids_candidate = np.where(time_diffs < interval)[0]\n        ids_candidate = np.sort(ids_candidate)\n        if (self.allow_repeat and len(ids_candidate) < num_views // 3) or (\n            len(ids_candidate) < num_views\n        ):\n            return []\n        ids_sel_list = []\n        ids_candidate_left = ids_candidate.copy()\n        while len(ids_candidate_left) >= num_views:\n            ids_sel = np.random.choice(ids_candidate_left, num_views, replace=False)\n            ids_sel_list.append(sorted(ids_sel))\n            ids_candidate_left = np.setdiff1d(ids_candidate_left, ids_sel)\n\n        if len(ids_candidate_left) > 0 and len(ids_candidate) >= num_views:\n            ids_sel = np.concatenate(\n                [\n                    ids_candidate_left,\n                    np.random.choice(\n                        np.setdiff1d(ids_candidate, ids_candidate_left),\n                        num_views - len(ids_candidate_left),\n                        replace=False,\n                    ),\n                ]\n            )\n            ids_sel_list.append(sorted(ids_sel))\n\n        if self.allow_repeat:\n            ids_sel_list.append(\n                sorted(np.random.choice(ids_candidate, num_views, replace=True))\n            )\n\n        # add sequences with fixed intervals (all possible intervals)\n        pos_i = np.where(ids_candidate == i)[0][0]\n        curr_interval = 1\n        stop = len(ids_candidate) < num_views\n        while not stop:\n            pos_sel = [pos_i]\n            count = 0\n            while len(pos_sel) < num_views:\n                if count % 2 == 0:\n                    curr_pos_i = pos_sel[-1] + curr_interval\n                    if curr_pos_i >= len(ids_candidate):\n                        stop = True\n                        break\n                    pos_sel.append(curr_pos_i)\n                else:\n                    curr_pos_i = pos_sel[0] - curr_interval\n                    if curr_pos_i < 0:\n                        stop = True\n                        break\n                    pos_sel.insert(0, curr_pos_i)\n                count += 1\n            if not stop and len(pos_sel) == num_views:\n                ids_sel = sorted([ids_candidate[pos] for pos in pos_sel])\n                if ids_sel not in ids_sel_list:\n                    ids_sel_list.append(ids_sel)\n            curr_interval += 1\n        return ids_sel_list\n\n    @staticmethod\n    def blockwise_shuffle(x, rng, block_shuffle):\n        if block_shuffle is None:\n            return rng.permutation(x).tolist()\n        else:\n            assert block_shuffle > 0\n            blocks = [x[i : i + block_shuffle] for i in range(0, len(x), block_shuffle)]\n            shuffled_blocks = [rng.permutation(block).tolist() for block in blocks]\n            shuffled_list = [item for block in shuffled_blocks for item in block]\n            return shuffled_list\n\n    def get_seq_from_start_id(\n        self,\n        num_views,\n        id_ref,\n        ids_all,\n        rng,\n        min_interval=1,\n        max_interval=25,","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.blockwise_shuffle","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.blockwise_shuffle#L168-L176","kind":"function","name":"blockwise_shuffle","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":168,"end_line":176,"context_start_line":148,"context_end_line":196,"code":"                    curr_pos_i = pos_sel[-1] + curr_interval\n                    if curr_pos_i >= len(ids_candidate):\n                        stop = True\n                        break\n                    pos_sel.append(curr_pos_i)\n                else:\n                    curr_pos_i = pos_sel[0] - curr_interval\n                    if curr_pos_i < 0:\n                        stop = True\n                        break\n                    pos_sel.insert(0, curr_pos_i)\n                count += 1\n            if not stop and len(pos_sel) == num_views:\n                ids_sel = sorted([ids_candidate[pos] for pos in pos_sel])\n                if ids_sel not in ids_sel_list:\n                    ids_sel_list.append(ids_sel)\n            curr_interval += 1\n        return ids_sel_list\n\n    @staticmethod\n    def blockwise_shuffle(x, rng, block_shuffle):\n        if block_shuffle is None:\n            return rng.permutation(x).tolist()\n        else:\n            assert block_shuffle > 0\n            blocks = [x[i : i + block_shuffle] for i in range(0, len(x), block_shuffle)]\n            shuffled_blocks = [rng.permutation(block).tolist() for block in blocks]\n            shuffled_list = [item for block in shuffled_blocks for item in block]\n            return shuffled_list\n\n    def get_seq_from_start_id(\n        self,\n        num_views,\n        id_ref,\n        ids_all,\n        rng,\n        min_interval=1,\n        max_interval=25,\n        video_prob=0.5,\n        fix_interval_prob=0.5,\n        block_shuffle=None,\n    ):\n        \"\"\"\n        args:\n            num_views: number of views to return\n            id_ref: the reference id (first id)\n            ids_all: all the ids\n            rng: random number generator\n            max_interval: maximum interval between two views","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.get_seq_from_start_id","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.get_seq_from_start_id#L178-L289","kind":"function","name":"get_seq_from_start_id","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":178,"end_line":289,"context_start_line":158,"context_end_line":309,"code":"                    pos_sel.insert(0, curr_pos_i)\n                count += 1\n            if not stop and len(pos_sel) == num_views:\n                ids_sel = sorted([ids_candidate[pos] for pos in pos_sel])\n                if ids_sel not in ids_sel_list:\n                    ids_sel_list.append(ids_sel)\n            curr_interval += 1\n        return ids_sel_list\n\n    @staticmethod\n    def blockwise_shuffle(x, rng, block_shuffle):\n        if block_shuffle is None:\n            return rng.permutation(x).tolist()\n        else:\n            assert block_shuffle > 0\n            blocks = [x[i : i + block_shuffle] for i in range(0, len(x), block_shuffle)]\n            shuffled_blocks = [rng.permutation(block).tolist() for block in blocks]\n            shuffled_list = [item for block in shuffled_blocks for item in block]\n            return shuffled_list\n\n    def get_seq_from_start_id(\n        self,\n        num_views,\n        id_ref,\n        ids_all,\n        rng,\n        min_interval=1,\n        max_interval=25,\n        video_prob=0.5,\n        fix_interval_prob=0.5,\n        block_shuffle=None,\n    ):\n        \"\"\"\n        args:\n            num_views: number of views to return\n            id_ref: the reference id (first id)\n            ids_all: all the ids\n            rng: random number generator\n            max_interval: maximum interval between two views\n        returns:\n            pos: list of positions of the views in ids_all, i.e., index for ids_all\n            is_video: True if the views are consecutive\n        \"\"\"\n        assert min_interval > 0, f\"min_interval should be > 0, got {min_interval}\"\n        assert (\n            min_interval <= max_interval\n        ), f\"min_interval should be <= max_interval, got {min_interval} and {max_interval}\"\n        assert id_ref in ids_all\n        pos_ref = ids_all.index(id_ref)\n        all_possible_pos = np.arange(pos_ref, len(ids_all))\n\n        remaining_sum = len(ids_all) - 1 - pos_ref\n\n        if remaining_sum >= num_views - 1:\n            if remaining_sum == num_views - 1:\n                assert ids_all[-num_views] == id_ref\n                return [pos_ref + i for i in range(num_views)], True\n            max_interval = min(max_interval, 2 * remaining_sum // (num_views - 1))\n            intervals = [\n                rng.choice(range(min_interval, max_interval + 1))\n                for _ in range(num_views - 1)\n            ]\n\n            # if video or collection\n            if rng.random() < video_prob:\n                # if fixed interval or random\n                if rng.random() < fix_interval_prob:\n                    # regular interval\n                    fixed_interval = rng.choice(\n                        range(\n                            1,\n                            min(remaining_sum // (num_views - 1) + 1, max_interval + 1),\n                        )\n                    )\n                    intervals = [fixed_interval for _ in range(num_views - 1)]\n                is_video = True\n            else:\n                is_video = False\n\n            pos = list(itertools.accumulate([pos_ref] + intervals))\n            pos = [p for p in pos if p < len(ids_all)]\n            pos_candidates = [p for p in all_possible_pos if p not in pos]\n            pos = (\n                pos\n                + rng.choice(\n                    pos_candidates, num_views - len(pos), replace=False\n                ).tolist()\n            )\n\n            pos = (\n                sorted(pos)\n                if is_video\n                else self.blockwise_shuffle(pos, rng, block_shuffle)\n            )\n        else:\n            # assert self.allow_repeat\n            uniq_num = remaining_sum\n            new_pos_ref = rng.choice(np.arange(pos_ref + 1))\n            new_remaining_sum = len(ids_all) - 1 - new_pos_ref\n            new_max_interval = min(max_interval, new_remaining_sum // (uniq_num - 1))\n            new_intervals = [\n                rng.choice(range(1, new_max_interval + 1)) for _ in range(uniq_num - 1)\n            ]\n\n            revisit_random = rng.random()\n            video_random = rng.random()\n\n            if rng.random() < fix_interval_prob and video_random < video_prob:\n                # regular interval\n                fixed_interval = rng.choice(range(1, new_max_interval + 1))\n                new_intervals = [fixed_interval for _ in range(uniq_num - 1)]\n            pos = list(itertools.accumulate([new_pos_ref] + new_intervals))\n\n            is_video = False\n            if revisit_random < 0.5 or video_prob == 1.0:  # revisit, video / collection\n                is_video = video_random < video_prob\n                pos = (\n                    self.blockwise_shuffle(pos, rng, block_shuffle)\n                    if not is_video\n                    else pos\n                )\n                num_full_repeat = num_views // uniq_num\n                pos = (\n                    pos * num_full_repeat\n                    + pos[: num_views - len(pos) * num_full_repeat]\n                )\n            elif revisit_random < 0.9:  # random\n                pos = rng.choice(pos, num_views, replace=True)\n            else:  # ordered\n                pos = sorted(rng.choice(pos, num_views, replace=True))\n        assert len(pos) == num_views\n        return pos, is_video\n\n    def get_img_and_ray_masks(self, is_metric, v, rng, p=[0.8, 0.15, 0.05]):\n        # generate img mask and raymap mask\n        if v == 0 or (not is_metric):\n            img_mask = True\n            raymap_mask = False\n        else:\n            rand_val = rng.random()\n            if rand_val < p[0]:\n                img_mask = True\n                raymap_mask = False\n            elif rand_val < p[0] + p[1]:\n                img_mask = False\n                raymap_mask = True\n            else:\n                img_mask = True\n                raymap_mask = True\n        return img_mask, raymap_mask\n\n    def get_stats(self):","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.get_img_and_ray_masks","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.get_img_and_ray_masks#L291-L307","kind":"function","name":"get_img_and_ray_masks","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":291,"end_line":307,"context_start_line":271,"context_end_line":327,"code":"            is_video = False\n            if revisit_random < 0.5 or video_prob == 1.0:  # revisit, video / collection\n                is_video = video_random < video_prob\n                pos = (\n                    self.blockwise_shuffle(pos, rng, block_shuffle)\n                    if not is_video\n                    else pos\n                )\n                num_full_repeat = num_views // uniq_num\n                pos = (\n                    pos * num_full_repeat\n                    + pos[: num_views - len(pos) * num_full_repeat]\n                )\n            elif revisit_random < 0.9:  # random\n                pos = rng.choice(pos, num_views, replace=True)\n            else:  # ordered\n                pos = sorted(rng.choice(pos, num_views, replace=True))\n        assert len(pos) == num_views\n        return pos, is_video\n\n    def get_img_and_ray_masks(self, is_metric, v, rng, p=[0.8, 0.15, 0.05]):\n        # generate img mask and raymap mask\n        if v == 0 or (not is_metric):\n            img_mask = True\n            raymap_mask = False\n        else:\n            rand_val = rng.random()\n            if rand_val < p[0]:\n                img_mask = True\n                raymap_mask = False\n            elif rand_val < p[0] + p[1]:\n                img_mask = False\n                raymap_mask = True\n            else:\n                img_mask = True\n                raymap_mask = True\n        return img_mask, raymap_mask\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def __repr__(self):\n        resolutions_str = \"[\" + \";\".join(f\"{w}x{h}\" for w, h in self._resolutions) + \"]\"\n        return (\n            f\"\"\"{type(self).__name__}({self.get_stats()},\n            {self.num_views=},\n            {self.split=},\n            {self.seed=},\n            resolutions={resolutions_str},\n            {self.transform=})\"\"\".replace(\n                \"self.\", \"\"\n            )\n            .replace(\"\\n\", \"\")\n            .replace(\"   \", \"\")\n        )\n\n    def _get_views(self, idx, resolution, rng, num_views):","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.get_stats","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.get_stats#L309-L310","kind":"function","name":"get_stats","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":309,"end_line":310,"context_start_line":289,"context_end_line":330,"code":"        return pos, is_video\n\n    def get_img_and_ray_masks(self, is_metric, v, rng, p=[0.8, 0.15, 0.05]):\n        # generate img mask and raymap mask\n        if v == 0 or (not is_metric):\n            img_mask = True\n            raymap_mask = False\n        else:\n            rand_val = rng.random()\n            if rand_val < p[0]:\n                img_mask = True\n                raymap_mask = False\n            elif rand_val < p[0] + p[1]:\n                img_mask = False\n                raymap_mask = True\n            else:\n                img_mask = True\n                raymap_mask = True\n        return img_mask, raymap_mask\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def __repr__(self):\n        resolutions_str = \"[\" + \";\".join(f\"{w}x{h}\" for w, h in self._resolutions) + \"]\"\n        return (\n            f\"\"\"{type(self).__name__}({self.get_stats()},\n            {self.num_views=},\n            {self.split=},\n            {self.seed=},\n            resolutions={resolutions_str},\n            {self.transform=})\"\"\".replace(\n                \"self.\", \"\"\n            )\n            .replace(\"\\n\", \"\")\n            .replace(\"   \", \"\")\n        )\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        raise NotImplementedError()\n\n    def __getitem__(self, idx):","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.__repr__","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.__repr__#L312-L325","kind":"function","name":"__repr__","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":312,"end_line":325,"context_start_line":292,"context_end_line":345,"code":"        # generate img mask and raymap mask\n        if v == 0 or (not is_metric):\n            img_mask = True\n            raymap_mask = False\n        else:\n            rand_val = rng.random()\n            if rand_val < p[0]:\n                img_mask = True\n                raymap_mask = False\n            elif rand_val < p[0] + p[1]:\n                img_mask = False\n                raymap_mask = True\n            else:\n                img_mask = True\n                raymap_mask = True\n        return img_mask, raymap_mask\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def __repr__(self):\n        resolutions_str = \"[\" + \";\".join(f\"{w}x{h}\" for w, h in self._resolutions) + \"]\"\n        return (\n            f\"\"\"{type(self).__name__}({self.get_stats()},\n            {self.num_views=},\n            {self.split=},\n            {self.seed=},\n            resolutions={resolutions_str},\n            {self.transform=})\"\"\".replace(\n                \"self.\", \"\"\n            )\n            .replace(\"\\n\", \"\")\n            .replace(\"   \", \"\")\n        )\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        raise NotImplementedError()\n\n    def __getitem__(self, idx):\n        # print(\"Receiving:\" , idx)\n        if isinstance(idx, (tuple, list, np.ndarray)):\n            # the idx is specifying the aspect-ratio\n            idx, ar_idx, nview = idx # start_id, selected type (eg., resolution), view number\n        else:\n            assert len(self._resolutions) == 1\n            ar_idx = 0\n            nview = self.num_views\n\n        assert nview >= 1 and nview <= self.num_views\n        # set-up the rng\n        if self.seed:  # reseed for each __getitem__\n            self._rng = np.random.default_rng(seed=self.seed + idx)\n        elif not hasattr(self, \"_rng\"):\n            seed = torch.randint(0, 2**32, (1,)).item()","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset._get_views","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset._get_views#L327-L328","kind":"function","name":"_get_views","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":327,"end_line":328,"context_start_line":307,"context_end_line":348,"code":"        return img_mask, raymap_mask\n\n    def get_stats(self):\n        return f\"{len(self)} groups of views\"\n\n    def __repr__(self):\n        resolutions_str = \"[\" + \";\".join(f\"{w}x{h}\" for w, h in self._resolutions) + \"]\"\n        return (\n            f\"\"\"{type(self).__name__}({self.get_stats()},\n            {self.num_views=},\n            {self.split=},\n            {self.seed=},\n            resolutions={resolutions_str},\n            {self.transform=})\"\"\".replace(\n                \"self.\", \"\"\n            )\n            .replace(\"\\n\", \"\")\n            .replace(\"   \", \"\")\n        )\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        raise NotImplementedError()\n\n    def __getitem__(self, idx):\n        # print(\"Receiving:\" , idx)\n        if isinstance(idx, (tuple, list, np.ndarray)):\n            # the idx is specifying the aspect-ratio\n            idx, ar_idx, nview = idx # start_id, selected type (eg., resolution), view number\n        else:\n            assert len(self._resolutions) == 1\n            ar_idx = 0\n            nview = self.num_views\n\n        assert nview >= 1 and nview <= self.num_views\n        # set-up the rng\n        if self.seed:  # reseed for each __getitem__\n            self._rng = np.random.default_rng(seed=self.seed + idx)\n        elif not hasattr(self, \"_rng\"):\n            seed = torch.randint(0, 2**32, (1,)).item()\n            self._rng = np.random.default_rng(seed=seed)\n\n        if self.aug_crop > 1 and self.seq_aug_crop:","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.__getitem__","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.__getitem__#L330-L421","kind":"function","name":"__getitem__","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":330,"end_line":421,"context_start_line":310,"context_end_line":441,"code":"        return f\"{len(self)} groups of views\"\n\n    def __repr__(self):\n        resolutions_str = \"[\" + \";\".join(f\"{w}x{h}\" for w, h in self._resolutions) + \"]\"\n        return (\n            f\"\"\"{type(self).__name__}({self.get_stats()},\n            {self.num_views=},\n            {self.split=},\n            {self.seed=},\n            resolutions={resolutions_str},\n            {self.transform=})\"\"\".replace(\n                \"self.\", \"\"\n            )\n            .replace(\"\\n\", \"\")\n            .replace(\"   \", \"\")\n        )\n\n    def _get_views(self, idx, resolution, rng, num_views):\n        raise NotImplementedError()\n\n    def __getitem__(self, idx):\n        # print(\"Receiving:\" , idx)\n        if isinstance(idx, (tuple, list, np.ndarray)):\n            # the idx is specifying the aspect-ratio\n            idx, ar_idx, nview = idx # start_id, selected type (eg., resolution), view number\n        else:\n            assert len(self._resolutions) == 1\n            ar_idx = 0\n            nview = self.num_views\n\n        assert nview >= 1 and nview <= self.num_views\n        # set-up the rng\n        if self.seed:  # reseed for each __getitem__\n            self._rng = np.random.default_rng(seed=self.seed + idx)\n        elif not hasattr(self, \"_rng\"):\n            seed = torch.randint(0, 2**32, (1,)).item()\n            self._rng = np.random.default_rng(seed=seed)\n\n        if self.aug_crop > 1 and self.seq_aug_crop:\n            self.delta_target_resolution = self._rng.integers(0, self.aug_crop)\n\n        # over-loaded code\n        resolution = self._resolutions[\n            ar_idx\n        ]  # DO NOT CHANGE THIS (compatible with BatchedRandomSampler)\n        views = self._get_views(idx, resolution, self._rng, nview)\n        assert len(views) == nview\n\n        if \"camera_pose\" not in views[0]:\n            views[0][\"camera_pose\"] = np.ones((4, 4), dtype=np.float32)\n        first_view_camera_pose = views[0][\"camera_pose\"]\n        transform = SeqColorJitter() if self.is_seq_color_jitter else self.transform\n\n        for v, view in enumerate(views):\n            assert (\n                \"pts3d\" not in view\n            ), f\"pts3d should not be there, they will be computed afterwards based on intrinsics+depthmap for view {view_name(view)}\"\n            view[\"idx\"] = (idx, ar_idx, v)\n\n            # encode the image\n            width, height = view[\"img\"].size\n\n            view[\"true_shape\"] = np.int32((height, width))\n            view[\"img\"] = transform(view[\"img\"])\n            view[\"sky_mask\"] = view[\"depthmap\"] < 0\n\n            assert \"camera_intrinsics\" in view\n            if \"camera_pose\" not in view:\n                view[\"camera_pose\"] = np.full((4, 4), np.nan, dtype=np.float32)\n            else:\n                assert np.isfinite(\n                    view[\"camera_pose\"]\n                ).all(), f\"NaN in camera pose for view {view_name(view)}\"\n\n            ray_map = get_ray_map(\n                first_view_camera_pose,\n                view[\"camera_pose\"],\n                view[\"camera_intrinsics\"],\n                height,\n                width,\n            )   # camera_pose: c2w\n            view[\"ray_map\"] = ray_map.astype(np.float32)\n\n            assert \"pts3d\" not in view\n            assert \"valid_mask\" not in view\n            assert np.isfinite(\n                view[\"depthmap\"]\n            ).all(), f\"NaN in depthmap for view {view_name(view)}\"\n            pts3d, valid_mask = depthmap_to_absolute_camera_coordinates(**view)\n\n            view[\"pts3d\"] = pts3d\n            view[\"valid_mask\"] = valid_mask & np.isfinite(pts3d).all(axis=-1)\n\n            # check all datatypes\n            for key, val in view.items():\n                res, err_msg = is_good_type(key, val)\n                assert res, f\"{err_msg} with {key}={val} for view {view_name(view)}\"\n            K = view[\"camera_intrinsics\"]\n\n        if self.n_corres > 0:\n            ref_view = views[0]\n            for view in views:\n                corres1, corres2, valid = extract_correspondences_from_pts3d(\n                    ref_view, view, self.n_corres, self._rng, nneg=self.nneg\n                )\n                view[\"corres\"] = (corres1, corres2)\n                view[\"valid_corres\"] = valid\n\n        # last thing done!\n        for view in views:\n            view[\"rng\"] = int.from_bytes(self._rng.bytes(4), \"big\")\n        return views\n\n    def _set_resolutions(self, resolutions):\n        assert resolutions is not None, \"undefined resolution\"\n\n        if not isinstance(resolutions, list):\n            resolutions = [resolutions]\n\n        self._resolutions = []\n        for resolution in resolutions:\n            if isinstance(resolution, int):\n                width = height = resolution\n            else:\n                width, height = resolution\n            assert isinstance(\n                width, int\n            ), f\"Bad type for {width=} {type(width)=}, should be int\"\n            assert isinstance(\n                height, int\n            ), f\"Bad type for {height=} {type(height)=}, should be int\"\n            self._resolutions.append((width, height))","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset._set_resolutions","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset._set_resolutions#L423-L441","kind":"function","name":"_set_resolutions","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":423,"end_line":441,"context_start_line":403,"context_end_line":461,"code":"            # check all datatypes\n            for key, val in view.items():\n                res, err_msg = is_good_type(key, val)\n                assert res, f\"{err_msg} with {key}={val} for view {view_name(view)}\"\n            K = view[\"camera_intrinsics\"]\n\n        if self.n_corres > 0:\n            ref_view = views[0]\n            for view in views:\n                corres1, corres2, valid = extract_correspondences_from_pts3d(\n                    ref_view, view, self.n_corres, self._rng, nneg=self.nneg\n                )\n                view[\"corres\"] = (corres1, corres2)\n                view[\"valid_corres\"] = valid\n\n        # last thing done!\n        for view in views:\n            view[\"rng\"] = int.from_bytes(self._rng.bytes(4), \"big\")\n        return views\n\n    def _set_resolutions(self, resolutions):\n        assert resolutions is not None, \"undefined resolution\"\n\n        if not isinstance(resolutions, list):\n            resolutions = [resolutions]\n\n        self._resolutions = []\n        for resolution in resolutions:\n            if isinstance(resolution, int):\n                width = height = resolution\n            else:\n                width, height = resolution\n            assert isinstance(\n                width, int\n            ), f\"Bad type for {width=} {type(width)=}, should be int\"\n            assert isinstance(\n                height, int\n            ), f\"Bad type for {height=} {type(height)=}, should be int\"\n            self._resolutions.append((width, height))\n\n    def _crop_resize_if_necessary(\n        self, image, depthmap, intrinsics, resolution, rng=None, info=None\n    ):\n        \"\"\"This function:\n        - first downsizes the image with LANCZOS inteprolation,\n          which is better than bilinear interpolation\n        \"\"\"\n        if not isinstance(image, PIL.Image.Image):\n            image = PIL.Image.fromarray(image)\n\n        # downscale with lanczos interpolation so that image.size == resolution\n        # cropping centered on the principal point\n        W, H = image.size\n        cx, cy = intrinsics[:2, 2].round().astype(int)\n        min_margin_x = min(cx, W - cx)\n        min_margin_y = min(cy, H - cy)\n        assert min_margin_x > W / 5, f\"Bad principal point in view={info}\"\n        assert min_margin_y > H / 5, f\"Bad principal point in view={info}\"\n        # the new window will be a rectangle of size (2*min_margin_x, 2*min_margin_y) centered on (cx,cy)","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset._crop_resize_if_necessary","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset._crop_resize_if_necessary#L443-L495","kind":"function","name":"_crop_resize_if_necessary","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":443,"end_line":495,"context_start_line":423,"context_end_line":515,"code":"    def _set_resolutions(self, resolutions):\n        assert resolutions is not None, \"undefined resolution\"\n\n        if not isinstance(resolutions, list):\n            resolutions = [resolutions]\n\n        self._resolutions = []\n        for resolution in resolutions:\n            if isinstance(resolution, int):\n                width = height = resolution\n            else:\n                width, height = resolution\n            assert isinstance(\n                width, int\n            ), f\"Bad type for {width=} {type(width)=}, should be int\"\n            assert isinstance(\n                height, int\n            ), f\"Bad type for {height=} {type(height)=}, should be int\"\n            self._resolutions.append((width, height))\n\n    def _crop_resize_if_necessary(\n        self, image, depthmap, intrinsics, resolution, rng=None, info=None\n    ):\n        \"\"\"This function:\n        - first downsizes the image with LANCZOS inteprolation,\n          which is better than bilinear interpolation\n        \"\"\"\n        if not isinstance(image, PIL.Image.Image):\n            image = PIL.Image.fromarray(image)\n\n        # downscale with lanczos interpolation so that image.size == resolution\n        # cropping centered on the principal point\n        W, H = image.size\n        cx, cy = intrinsics[:2, 2].round().astype(int)\n        min_margin_x = min(cx, W - cx)\n        min_margin_y = min(cy, H - cy)\n        assert min_margin_x > W / 5, f\"Bad principal point in view={info}\"\n        assert min_margin_y > H / 5, f\"Bad principal point in view={info}\"\n        # the new window will be a rectangle of size (2*min_margin_x, 2*min_margin_y) centered on (cx,cy)\n        l, t = cx - min_margin_x, cy - min_margin_y\n        r, b = cx + min_margin_x, cy + min_margin_y\n        crop_bbox = (l, t, r, b)\n        image, depthmap, intrinsics = cropping.crop_image_depthmap(\n            image, depthmap, intrinsics, crop_bbox\n        )\n\n        # transpose the resolution if necessary\n        W, H = image.size  # new size\n\n        # high-quality Lanczos down-scaling\n        target_resolution = np.array(resolution)\n        if self.aug_crop > 1:\n            target_resolution += (\n                rng.integers(0, self.aug_crop)\n                if not self.seq_aug_crop\n                else self.delta_target_resolution\n            )\n        image, depthmap, intrinsics = cropping.rescale_image_depthmap(\n            image, depthmap, intrinsics, target_resolution\n        )\n\n        # actual cropping (if necessary) with bilinear interpolation\n        intrinsics2 = cropping.camera_matrix_of_crop(\n            intrinsics, image.size, resolution, offset_factor=0.5\n        )\n        crop_bbox = cropping.bbox_from_intrinsics_in_out(\n            intrinsics, intrinsics2, resolution\n        )\n        image, depthmap, intrinsics2 = cropping.crop_image_depthmap(\n            image, depthmap, intrinsics, crop_bbox\n        )\n\n        return image, depthmap, intrinsics2\n\n    def _crop_resize_if_necessary_mask(\n        self, image, depthmap, mask, intrinsics, resolution, rng=None, info=None\n    ):\n        \"\"\"This function:\n        - first downsizes the image with LANCZOS inteprolation,\n          which is better than bilinear interpolation\n          also processes mask with the same transformations as image\n        \"\"\"\n        if not isinstance(image, PIL.Image.Image):\n            image = PIL.Image.fromarray(image)\n        if not isinstance(mask, PIL.Image.Image):\n            mask = PIL.Image.fromarray(mask)\n\n        # downscale with lanczos interpolation so that image.size == resolution\n        # cropping centered on the principal point\n        W, H = image.size\n        cx, cy = intrinsics[:2, 2].round().astype(int)\n        min_margin_x = min(cx, W - cx)\n        min_margin_y = min(cy, H - cy)","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset._crop_resize_if_necessary_mask","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset._crop_resize_if_necessary_mask#L497-L556","kind":"function","name":"_crop_resize_if_necessary_mask","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":497,"end_line":556,"context_start_line":477,"context_end_line":576,"code":"                if not self.seq_aug_crop\n                else self.delta_target_resolution\n            )\n        image, depthmap, intrinsics = cropping.rescale_image_depthmap(\n            image, depthmap, intrinsics, target_resolution\n        )\n\n        # actual cropping (if necessary) with bilinear interpolation\n        intrinsics2 = cropping.camera_matrix_of_crop(\n            intrinsics, image.size, resolution, offset_factor=0.5\n        )\n        crop_bbox = cropping.bbox_from_intrinsics_in_out(\n            intrinsics, intrinsics2, resolution\n        )\n        image, depthmap, intrinsics2 = cropping.crop_image_depthmap(\n            image, depthmap, intrinsics, crop_bbox\n        )\n\n        return image, depthmap, intrinsics2\n\n    def _crop_resize_if_necessary_mask(\n        self, image, depthmap, mask, intrinsics, resolution, rng=None, info=None\n    ):\n        \"\"\"This function:\n        - first downsizes the image with LANCZOS inteprolation,\n          which is better than bilinear interpolation\n          also processes mask with the same transformations as image\n        \"\"\"\n        if not isinstance(image, PIL.Image.Image):\n            image = PIL.Image.fromarray(image)\n        if not isinstance(mask, PIL.Image.Image):\n            mask = PIL.Image.fromarray(mask)\n\n        # downscale with lanczos interpolation so that image.size == resolution\n        # cropping centered on the principal point\n        W, H = image.size\n        cx, cy = intrinsics[:2, 2].round().astype(int)\n        min_margin_x = min(cx, W - cx)\n        min_margin_y = min(cy, H - cy)\n        assert min_margin_x > W / 5, f\"Bad principal point in view={info}\"\n        assert min_margin_y > H / 5, f\"Bad principal point in view={info}\"\n        # the new window will be a rectangle of size (2*min_margin_x, 2*min_margin_y) centered on (cx,cy)\n        l, t = cx - min_margin_x, cy - min_margin_y\n        r, b = cx + min_margin_x, cy + min_margin_y\n        crop_bbox = (l, t, r, b)\n        image, depthmap, mask, intrinsics = cropping.crop_image_depthmap_mask(\n            image, depthmap, mask, intrinsics, crop_bbox\n        )\n\n        # transpose the resolution if necessary\n        W, H = image.size  # new size\n\n        # high-quality Lanczos down-scaling\n        target_resolution = np.array(resolution)\n        if self.aug_crop > 1:\n            target_resolution += (\n                rng.integers(0, self.aug_crop)\n                if not self.seq_aug_crop\n                else self.delta_target_resolution\n            )\n        image, depthmap, mask, intrinsics = cropping.rescale_image_depthmap_mask(\n            image, depthmap, mask, intrinsics, target_resolution\n        )\n\n        # actual cropping (if necessary) with bilinear interpolation\n        intrinsics2 = cropping.camera_matrix_of_crop(\n            intrinsics, image.size, resolution, offset_factor=0.5\n        )\n        crop_bbox = cropping.bbox_from_intrinsics_in_out(\n            intrinsics, intrinsics2, resolution\n        )\n        image, depthmap, mask, intrinsics2 = cropping.crop_image_depthmap_mask(\n            image, depthmap, mask, intrinsics, crop_bbox\n        )\n\n        if isinstance(mask, PIL.Image.Image):\n            mask = tvf.ToTensor()(mask)\n            mask = (mask[0] > 0.5).float()\n\n        return image, depthmap, mask, intrinsics2\n\n\ndef is_good_type(key, v):\n    \"\"\"returns (is_good, err_msg)\"\"\"\n    if isinstance(v, (str, int, tuple)):\n        return True, None\n    if v.dtype not in (np.float32, torch.float32, bool, np.int32, np.int64, np.uint8):\n        return False, f\"bad {v.dtype=}\"\n    return True, None\n\n\ndef view_name(view, batch_index=None):\n    def sel(x):\n        return x[batch_index] if batch_index not in (None, slice(None)) else x\n\n    db = sel(view[\"dataset\"])\n    label = sel(view[\"label\"])\n    instance = sel(view[\"instance\"])\n    return f\"{db}/{label}/{instance}\"\n","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.base_multiview_dataset.sel","uri":"program://Human3R/function/src.dust3r.datasets.base.base_multiview_dataset.sel#L569-L570","kind":"function","name":"sel","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":569,"end_line":570,"context_start_line":549,"context_end_line":590,"code":"            image, depthmap, mask, intrinsics, crop_bbox\n        )\n\n        if isinstance(mask, PIL.Image.Image):\n            mask = tvf.ToTensor()(mask)\n            mask = (mask[0] > 0.5).float()\n\n        return image, depthmap, mask, intrinsics2\n\n\ndef is_good_type(key, v):\n    \"\"\"returns (is_good, err_msg)\"\"\"\n    if isinstance(v, (str, int, tuple)):\n        return True, None\n    if v.dtype not in (np.float32, torch.float32, bool, np.int32, np.int64, np.uint8):\n        return False, f\"bad {v.dtype=}\"\n    return True, None\n\n\ndef view_name(view, batch_index=None):\n    def sel(x):\n        return x[batch_index] if batch_index not in (None, slice(None)) else x\n\n    db = sel(view[\"dataset\"])\n    label = sel(view[\"label\"])\n    instance = sel(view[\"instance\"])\n    return f\"{db}/{label}/{instance}\"\n\n\ndef transpose_to_landscape(view):\n    height, width = view[\"true_shape\"]\n\n    if width < height:\n        # rectify portrait to landscape\n        assert view[\"img\"].shape == (3, height, width)\n        view[\"img\"] = view[\"img\"].swapaxes(1, 2)\n\n        assert view[\"valid_mask\"].shape == (height, width)\n        view[\"valid_mask\"] = view[\"valid_mask\"].swapaxes(0, 1)\n\n        assert view[\"depthmap\"].shape == (height, width)\n        view[\"depthmap\"] = view[\"depthmap\"].swapaxes(0, 1)","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset","uri":"program://Human3R/module/src.dust3r.datasets.base.easy_dataset#L1-L198","kind":"module","name":"src.dust3r.datasets.base.easy_dataset","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":1,"end_line":198,"context_start_line":1,"context_end_line":198,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nfrom dust3r.datasets.base.batched_sampler import (\n    BatchedRandomSampler,\n    CustomRandomSampler,\n)\nimport torch\n\n\nclass EasyDataset:\n    \"\"\"a dataset that you can easily resize and combine.\n    Examples:\n    ---------\n        2 * dataset ==> duplicate each element 2x\n\n        10 @ dataset ==> set the size to 10 (random sampling, duplicates if necessary)\n\n        dataset1 + dataset2 ==> concatenate datasets\n    \"\"\"\n\n    def __add__(self, other):\n        return CatDataset([self, other])\n\n    def __rmul__(self, factor):\n        return MulDataset(factor, self)\n\n    def __rmatmul__(self, factor):\n        return ResizedDataset(factor, self)\n\n    def set_epoch(self, epoch):\n        pass  # nothing to do by default\n\n    def make_sampler(\n        self, batch_size, shuffle=True, drop_last=True, world_size=1, rank=0, fixed_length=False\n    ):\n        if not (shuffle):\n            raise NotImplementedError()  # cannot deal yet\n        num_of_aspect_ratios = len(self._resolutions)\n        num_of_views = self.num_views\n        sampler = CustomRandomSampler(\n            self,\n            batch_size,\n            num_of_aspect_ratios,\n            4 if not fixed_length else num_of_views,\n            num_of_views,\n            world_size,\n            warmup=1,\n            drop_last=drop_last,\n        )\n        return BatchedRandomSampler(sampler, batch_size, drop_last)\n\n\nclass MulDataset(EasyDataset):\n    \"\"\"Artifically augmenting the size of a dataset.\"\"\"\n\n    multiplicator: int\n\n    def __init__(self, multiplicator, dataset):\n        assert isinstance(multiplicator, int) and multiplicator > 0\n        self.multiplicator = multiplicator\n        self.dataset = dataset\n\n    def __len__(self):\n        return self.multiplicator * len(self.dataset)\n\n    def __repr__(self):\n        return f\"{self.multiplicator}*{repr(self.dataset)}\"\n\n    def __getitem__(self, idx):\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n            return self.dataset[idx // self.multiplicator, other, another]\n        else:\n            return self.dataset[idx // self.multiplicator]\n\n    @property\n    def _resolutions(self):\n        return self.dataset._resolutions\n\n    @property\n    def num_views(self):\n        return self.dataset.num_views\n\n\nclass ResizedDataset(EasyDataset):\n    \"\"\"Artifically changing the size of a dataset.\"\"\"\n\n    new_size: int\n\n    def __init__(self, new_size, dataset):\n        assert isinstance(new_size, int) and new_size > 0\n        self.new_size = new_size\n        self.dataset = dataset\n\n    def __len__(self):\n        return self.new_size\n\n    def __repr__(self):\n        size_str = str(self.new_size)\n        for i in range((len(size_str) - 1) // 3):\n            sep = -4 * i - 3\n            size_str = size_str[:sep] + \"_\" + size_str[sep:]\n        return f\"{size_str} @ {repr(self.dataset)}\"\n\n    def set_epoch(self, epoch):\n        # this random shuffle only depends on the epoch\n        rng = np.random.default_rng(seed=epoch + 777)\n\n        # shuffle all indices\n        perm = rng.permutation(len(self.dataset))\n\n        # rotary extension until target size is met\n        shuffled_idxs = np.concatenate(\n            [perm] * (1 + (len(self) - 1) // len(self.dataset))\n        )\n        self._idxs_mapping = shuffled_idxs[: self.new_size]\n\n        assert len(self._idxs_mapping) == self.new_size\n\n    def __getitem__(self, idx):\n        assert hasattr(\n            self, \"_idxs_mapping\"\n        ), \"You need to call dataset.set_epoch() to use ResizedDataset.__getitem__()\"\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n            return self.dataset[self._idxs_mapping[idx], other, another] # start_id, selected type (eg., resolution), view number\n        else:\n            return self.dataset[self._idxs_mapping[idx]]\n\n    @property\n    def _resolutions(self):\n        return self.dataset._resolutions\n\n    @property\n    def num_views(self):\n        return self.dataset.num_views\n\n\nclass CatDataset(EasyDataset):\n    \"\"\"Concatenation of several datasets\"\"\"\n\n    def __init__(self, datasets):\n        for dataset in datasets:\n            assert isinstance(dataset, EasyDataset)\n        self.datasets = datasets\n        self._cum_sizes = np.cumsum([len(dataset) for dataset in datasets])\n\n    def __len__(self):\n        return self._cum_sizes[-1]\n\n    def __repr__(self):\n        # remove uselessly long transform\n        return \" + \".join(\n            repr(dataset).replace(\n                \",transform=Compose( ToTensor() Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)))\",\n                \"\",\n            )\n            for dataset in self.datasets\n        )\n\n    def set_epoch(self, epoch):\n        for dataset in self.datasets:\n            dataset.set_epoch(epoch)\n\n    def __getitem__(self, idx):\n        other = None\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n\n        if not (0 <= idx < len(self)):\n            raise IndexError()\n\n        db_idx = np.searchsorted(self._cum_sizes, idx, \"right\")\n        dataset = self.datasets[db_idx]\n        new_idx = idx - (self._cum_sizes[db_idx - 1] if db_idx > 0 else 0)\n\n        if other is not None and another is not None:\n            new_idx = (new_idx, other, another)\n        return dataset[new_idx]\n\n    @property\n    def _resolutions(self):\n        resolutions = self.datasets[0]._resolutions\n        for dataset in self.datasets[1:]:\n            assert tuple(dataset._resolutions) == tuple(resolutions)\n        return resolutions\n\n    @property\n    def num_views(self):\n        num_views = self.datasets[0].num_views\n        for dataset in self.datasets[1:]:\n            assert dataset.num_views == num_views\n        return num_views","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.EasyDataset","uri":"program://Human3R/class/src.dust3r.datasets.base.easy_dataset.EasyDataset#L15-L55","kind":"class","name":"EasyDataset","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":15,"end_line":55,"context_start_line":1,"context_end_line":75,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nfrom dust3r.datasets.base.batched_sampler import (\n    BatchedRandomSampler,\n    CustomRandomSampler,\n)\nimport torch\n\n\nclass EasyDataset:\n    \"\"\"a dataset that you can easily resize and combine.\n    Examples:\n    ---------\n        2 * dataset ==> duplicate each element 2x\n\n        10 @ dataset ==> set the size to 10 (random sampling, duplicates if necessary)\n\n        dataset1 + dataset2 ==> concatenate datasets\n    \"\"\"\n\n    def __add__(self, other):\n        return CatDataset([self, other])\n\n    def __rmul__(self, factor):\n        return MulDataset(factor, self)\n\n    def __rmatmul__(self, factor):\n        return ResizedDataset(factor, self)\n\n    def set_epoch(self, epoch):\n        pass  # nothing to do by default\n\n    def make_sampler(\n        self, batch_size, shuffle=True, drop_last=True, world_size=1, rank=0, fixed_length=False\n    ):\n        if not (shuffle):\n            raise NotImplementedError()  # cannot deal yet\n        num_of_aspect_ratios = len(self._resolutions)\n        num_of_views = self.num_views\n        sampler = CustomRandomSampler(\n            self,\n            batch_size,\n            num_of_aspect_ratios,\n            4 if not fixed_length else num_of_views,\n            num_of_views,\n            world_size,\n            warmup=1,\n            drop_last=drop_last,\n        )\n        return BatchedRandomSampler(sampler, batch_size, drop_last)\n\n\nclass MulDataset(EasyDataset):\n    \"\"\"Artifically augmenting the size of a dataset.\"\"\"\n\n    multiplicator: int\n\n    def __init__(self, multiplicator, dataset):\n        assert isinstance(multiplicator, int) and multiplicator > 0\n        self.multiplicator = multiplicator\n        self.dataset = dataset\n\n    def __len__(self):\n        return self.multiplicator * len(self.dataset)\n\n    def __repr__(self):\n        return f\"{self.multiplicator}*{repr(self.dataset)}\"\n\n    def __getitem__(self, idx):\n        if isinstance(idx, tuple):","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.MulDataset","uri":"program://Human3R/class/src.dust3r.datasets.base.easy_dataset.MulDataset#L58-L87","kind":"class","name":"MulDataset","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":58,"end_line":87,"context_start_line":38,"context_end_line":107,"code":"    def make_sampler(\n        self, batch_size, shuffle=True, drop_last=True, world_size=1, rank=0, fixed_length=False\n    ):\n        if not (shuffle):\n            raise NotImplementedError()  # cannot deal yet\n        num_of_aspect_ratios = len(self._resolutions)\n        num_of_views = self.num_views\n        sampler = CustomRandomSampler(\n            self,\n            batch_size,\n            num_of_aspect_ratios,\n            4 if not fixed_length else num_of_views,\n            num_of_views,\n            world_size,\n            warmup=1,\n            drop_last=drop_last,\n        )\n        return BatchedRandomSampler(sampler, batch_size, drop_last)\n\n\nclass MulDataset(EasyDataset):\n    \"\"\"Artifically augmenting the size of a dataset.\"\"\"\n\n    multiplicator: int\n\n    def __init__(self, multiplicator, dataset):\n        assert isinstance(multiplicator, int) and multiplicator > 0\n        self.multiplicator = multiplicator\n        self.dataset = dataset\n\n    def __len__(self):\n        return self.multiplicator * len(self.dataset)\n\n    def __repr__(self):\n        return f\"{self.multiplicator}*{repr(self.dataset)}\"\n\n    def __getitem__(self, idx):\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n            return self.dataset[idx // self.multiplicator, other, another]\n        else:\n            return self.dataset[idx // self.multiplicator]\n\n    @property\n    def _resolutions(self):\n        return self.dataset._resolutions\n\n    @property\n    def num_views(self):\n        return self.dataset.num_views\n\n\nclass ResizedDataset(EasyDataset):\n    \"\"\"Artifically changing the size of a dataset.\"\"\"\n\n    new_size: int\n\n    def __init__(self, new_size, dataset):\n        assert isinstance(new_size, int) and new_size > 0\n        self.new_size = new_size\n        self.dataset = dataset\n\n    def __len__(self):\n        return self.new_size\n\n    def __repr__(self):\n        size_str = str(self.new_size)\n        for i in range((len(size_str) - 1) // 3):\n            sep = -4 * i - 3\n            size_str = size_str[:sep] + \"_\" + size_str[sep:]","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.ResizedDataset","uri":"program://Human3R/class/src.dust3r.datasets.base.easy_dataset.ResizedDataset#L90-L141","kind":"class","name":"ResizedDataset","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":90,"end_line":141,"context_start_line":70,"context_end_line":161,"code":"\n    def __repr__(self):\n        return f\"{self.multiplicator}*{repr(self.dataset)}\"\n\n    def __getitem__(self, idx):\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n            return self.dataset[idx // self.multiplicator, other, another]\n        else:\n            return self.dataset[idx // self.multiplicator]\n\n    @property\n    def _resolutions(self):\n        return self.dataset._resolutions\n\n    @property\n    def num_views(self):\n        return self.dataset.num_views\n\n\nclass ResizedDataset(EasyDataset):\n    \"\"\"Artifically changing the size of a dataset.\"\"\"\n\n    new_size: int\n\n    def __init__(self, new_size, dataset):\n        assert isinstance(new_size, int) and new_size > 0\n        self.new_size = new_size\n        self.dataset = dataset\n\n    def __len__(self):\n        return self.new_size\n\n    def __repr__(self):\n        size_str = str(self.new_size)\n        for i in range((len(size_str) - 1) // 3):\n            sep = -4 * i - 3\n            size_str = size_str[:sep] + \"_\" + size_str[sep:]\n        return f\"{size_str} @ {repr(self.dataset)}\"\n\n    def set_epoch(self, epoch):\n        # this random shuffle only depends on the epoch\n        rng = np.random.default_rng(seed=epoch + 777)\n\n        # shuffle all indices\n        perm = rng.permutation(len(self.dataset))\n\n        # rotary extension until target size is met\n        shuffled_idxs = np.concatenate(\n            [perm] * (1 + (len(self) - 1) // len(self.dataset))\n        )\n        self._idxs_mapping = shuffled_idxs[: self.new_size]\n\n        assert len(self._idxs_mapping) == self.new_size\n\n    def __getitem__(self, idx):\n        assert hasattr(\n            self, \"_idxs_mapping\"\n        ), \"You need to call dataset.set_epoch() to use ResizedDataset.__getitem__()\"\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n            return self.dataset[self._idxs_mapping[idx], other, another] # start_id, selected type (eg., resolution), view number\n        else:\n            return self.dataset[self._idxs_mapping[idx]]\n\n    @property\n    def _resolutions(self):\n        return self.dataset._resolutions\n\n    @property\n    def num_views(self):\n        return self.dataset.num_views\n\n\nclass CatDataset(EasyDataset):\n    \"\"\"Concatenation of several datasets\"\"\"\n\n    def __init__(self, datasets):\n        for dataset in datasets:\n            assert isinstance(dataset, EasyDataset)\n        self.datasets = datasets\n        self._cum_sizes = np.cumsum([len(dataset) for dataset in datasets])\n\n    def __len__(self):\n        return self._cum_sizes[-1]\n\n    def __repr__(self):\n        # remove uselessly long transform\n        return \" + \".join(\n            repr(dataset).replace(\n                \",transform=Compose( ToTensor() Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)))\",\n                \"\",","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.CatDataset","uri":"program://Human3R/class/src.dust3r.datasets.base.easy_dataset.CatDataset#L144-L198","kind":"class","name":"CatDataset","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":144,"end_line":198,"context_start_line":124,"context_end_line":198,"code":"\n    def __getitem__(self, idx):\n        assert hasattr(\n            self, \"_idxs_mapping\"\n        ), \"You need to call dataset.set_epoch() to use ResizedDataset.__getitem__()\"\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n            return self.dataset[self._idxs_mapping[idx], other, another] # start_id, selected type (eg., resolution), view number\n        else:\n            return self.dataset[self._idxs_mapping[idx]]\n\n    @property\n    def _resolutions(self):\n        return self.dataset._resolutions\n\n    @property\n    def num_views(self):\n        return self.dataset.num_views\n\n\nclass CatDataset(EasyDataset):\n    \"\"\"Concatenation of several datasets\"\"\"\n\n    def __init__(self, datasets):\n        for dataset in datasets:\n            assert isinstance(dataset, EasyDataset)\n        self.datasets = datasets\n        self._cum_sizes = np.cumsum([len(dataset) for dataset in datasets])\n\n    def __len__(self):\n        return self._cum_sizes[-1]\n\n    def __repr__(self):\n        # remove uselessly long transform\n        return \" + \".join(\n            repr(dataset).replace(\n                \",transform=Compose( ToTensor() Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)))\",\n                \"\",\n            )\n            for dataset in self.datasets\n        )\n\n    def set_epoch(self, epoch):\n        for dataset in self.datasets:\n            dataset.set_epoch(epoch)\n\n    def __getitem__(self, idx):\n        other = None\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n\n        if not (0 <= idx < len(self)):\n            raise IndexError()\n\n        db_idx = np.searchsorted(self._cum_sizes, idx, \"right\")\n        dataset = self.datasets[db_idx]\n        new_idx = idx - (self._cum_sizes[db_idx - 1] if db_idx > 0 else 0)\n\n        if other is not None and another is not None:\n            new_idx = (new_idx, other, another)\n        return dataset[new_idx]\n\n    @property\n    def _resolutions(self):\n        resolutions = self.datasets[0]._resolutions\n        for dataset in self.datasets[1:]:\n            assert tuple(dataset._resolutions) == tuple(resolutions)\n        return resolutions\n\n    @property\n    def num_views(self):\n        num_views = self.datasets[0].num_views\n        for dataset in self.datasets[1:]:\n            assert dataset.num_views == num_views\n        return num_views","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.__add__","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.__add__#L26-L27","kind":"function","name":"__add__","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":26,"end_line":27,"context_start_line":6,"context_end_line":47,"code":"\nimport numpy as np\nfrom dust3r.datasets.base.batched_sampler import (\n    BatchedRandomSampler,\n    CustomRandomSampler,\n)\nimport torch\n\n\nclass EasyDataset:\n    \"\"\"a dataset that you can easily resize and combine.\n    Examples:\n    ---------\n        2 * dataset ==> duplicate each element 2x\n\n        10 @ dataset ==> set the size to 10 (random sampling, duplicates if necessary)\n\n        dataset1 + dataset2 ==> concatenate datasets\n    \"\"\"\n\n    def __add__(self, other):\n        return CatDataset([self, other])\n\n    def __rmul__(self, factor):\n        return MulDataset(factor, self)\n\n    def __rmatmul__(self, factor):\n        return ResizedDataset(factor, self)\n\n    def set_epoch(self, epoch):\n        pass  # nothing to do by default\n\n    def make_sampler(\n        self, batch_size, shuffle=True, drop_last=True, world_size=1, rank=0, fixed_length=False\n    ):\n        if not (shuffle):\n            raise NotImplementedError()  # cannot deal yet\n        num_of_aspect_ratios = len(self._resolutions)\n        num_of_views = self.num_views\n        sampler = CustomRandomSampler(\n            self,\n            batch_size,","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.__rmul__","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.__rmul__#L29-L30","kind":"function","name":"__rmul__","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":29,"end_line":30,"context_start_line":9,"context_end_line":50,"code":"    BatchedRandomSampler,\n    CustomRandomSampler,\n)\nimport torch\n\n\nclass EasyDataset:\n    \"\"\"a dataset that you can easily resize and combine.\n    Examples:\n    ---------\n        2 * dataset ==> duplicate each element 2x\n\n        10 @ dataset ==> set the size to 10 (random sampling, duplicates if necessary)\n\n        dataset1 + dataset2 ==> concatenate datasets\n    \"\"\"\n\n    def __add__(self, other):\n        return CatDataset([self, other])\n\n    def __rmul__(self, factor):\n        return MulDataset(factor, self)\n\n    def __rmatmul__(self, factor):\n        return ResizedDataset(factor, self)\n\n    def set_epoch(self, epoch):\n        pass  # nothing to do by default\n\n    def make_sampler(\n        self, batch_size, shuffle=True, drop_last=True, world_size=1, rank=0, fixed_length=False\n    ):\n        if not (shuffle):\n            raise NotImplementedError()  # cannot deal yet\n        num_of_aspect_ratios = len(self._resolutions)\n        num_of_views = self.num_views\n        sampler = CustomRandomSampler(\n            self,\n            batch_size,\n            num_of_aspect_ratios,\n            4 if not fixed_length else num_of_views,\n            num_of_views,","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.__rmatmul__","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.__rmatmul__#L32-L33","kind":"function","name":"__rmatmul__","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":32,"end_line":33,"context_start_line":12,"context_end_line":53,"code":"import torch\n\n\nclass EasyDataset:\n    \"\"\"a dataset that you can easily resize and combine.\n    Examples:\n    ---------\n        2 * dataset ==> duplicate each element 2x\n\n        10 @ dataset ==> set the size to 10 (random sampling, duplicates if necessary)\n\n        dataset1 + dataset2 ==> concatenate datasets\n    \"\"\"\n\n    def __add__(self, other):\n        return CatDataset([self, other])\n\n    def __rmul__(self, factor):\n        return MulDataset(factor, self)\n\n    def __rmatmul__(self, factor):\n        return ResizedDataset(factor, self)\n\n    def set_epoch(self, epoch):\n        pass  # nothing to do by default\n\n    def make_sampler(\n        self, batch_size, shuffle=True, drop_last=True, world_size=1, rank=0, fixed_length=False\n    ):\n        if not (shuffle):\n            raise NotImplementedError()  # cannot deal yet\n        num_of_aspect_ratios = len(self._resolutions)\n        num_of_views = self.num_views\n        sampler = CustomRandomSampler(\n            self,\n            batch_size,\n            num_of_aspect_ratios,\n            4 if not fixed_length else num_of_views,\n            num_of_views,\n            world_size,\n            warmup=1,\n            drop_last=drop_last,","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.set_epoch","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.set_epoch#L166-L168","kind":"function","name":"set_epoch","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":166,"end_line":168,"context_start_line":146,"context_end_line":188,"code":"\n    def __init__(self, datasets):\n        for dataset in datasets:\n            assert isinstance(dataset, EasyDataset)\n        self.datasets = datasets\n        self._cum_sizes = np.cumsum([len(dataset) for dataset in datasets])\n\n    def __len__(self):\n        return self._cum_sizes[-1]\n\n    def __repr__(self):\n        # remove uselessly long transform\n        return \" + \".join(\n            repr(dataset).replace(\n                \",transform=Compose( ToTensor() Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)))\",\n                \"\",\n            )\n            for dataset in self.datasets\n        )\n\n    def set_epoch(self, epoch):\n        for dataset in self.datasets:\n            dataset.set_epoch(epoch)\n\n    def __getitem__(self, idx):\n        other = None\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n\n        if not (0 <= idx < len(self)):\n            raise IndexError()\n\n        db_idx = np.searchsorted(self._cum_sizes, idx, \"right\")\n        dataset = self.datasets[db_idx]\n        new_idx = idx - (self._cum_sizes[db_idx - 1] if db_idx > 0 else 0)\n\n        if other is not None and another is not None:\n            new_idx = (new_idx, other, another)\n        return dataset[new_idx]\n\n    @property\n    def _resolutions(self):\n        resolutions = self.datasets[0]._resolutions","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.make_sampler","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.make_sampler#L38-L55","kind":"function","name":"make_sampler","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":38,"end_line":55,"context_start_line":18,"context_end_line":75,"code":"    ---------\n        2 * dataset ==> duplicate each element 2x\n\n        10 @ dataset ==> set the size to 10 (random sampling, duplicates if necessary)\n\n        dataset1 + dataset2 ==> concatenate datasets\n    \"\"\"\n\n    def __add__(self, other):\n        return CatDataset([self, other])\n\n    def __rmul__(self, factor):\n        return MulDataset(factor, self)\n\n    def __rmatmul__(self, factor):\n        return ResizedDataset(factor, self)\n\n    def set_epoch(self, epoch):\n        pass  # nothing to do by default\n\n    def make_sampler(\n        self, batch_size, shuffle=True, drop_last=True, world_size=1, rank=0, fixed_length=False\n    ):\n        if not (shuffle):\n            raise NotImplementedError()  # cannot deal yet\n        num_of_aspect_ratios = len(self._resolutions)\n        num_of_views = self.num_views\n        sampler = CustomRandomSampler(\n            self,\n            batch_size,\n            num_of_aspect_ratios,\n            4 if not fixed_length else num_of_views,\n            num_of_views,\n            world_size,\n            warmup=1,\n            drop_last=drop_last,\n        )\n        return BatchedRandomSampler(sampler, batch_size, drop_last)\n\n\nclass MulDataset(EasyDataset):\n    \"\"\"Artifically augmenting the size of a dataset.\"\"\"\n\n    multiplicator: int\n\n    def __init__(self, multiplicator, dataset):\n        assert isinstance(multiplicator, int) and multiplicator > 0\n        self.multiplicator = multiplicator\n        self.dataset = dataset\n\n    def __len__(self):\n        return self.multiplicator * len(self.dataset)\n\n    def __repr__(self):\n        return f\"{self.multiplicator}*{repr(self.dataset)}\"\n\n    def __getitem__(self, idx):\n        if isinstance(idx, tuple):","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.__init__","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.__init__#L147-L151","kind":"function","name":"__init__","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":147,"end_line":151,"context_start_line":127,"context_end_line":171,"code":"            self, \"_idxs_mapping\"\n        ), \"You need to call dataset.set_epoch() to use ResizedDataset.__getitem__()\"\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n            return self.dataset[self._idxs_mapping[idx], other, another] # start_id, selected type (eg., resolution), view number\n        else:\n            return self.dataset[self._idxs_mapping[idx]]\n\n    @property\n    def _resolutions(self):\n        return self.dataset._resolutions\n\n    @property\n    def num_views(self):\n        return self.dataset.num_views\n\n\nclass CatDataset(EasyDataset):\n    \"\"\"Concatenation of several datasets\"\"\"\n\n    def __init__(self, datasets):\n        for dataset in datasets:\n            assert isinstance(dataset, EasyDataset)\n        self.datasets = datasets\n        self._cum_sizes = np.cumsum([len(dataset) for dataset in datasets])\n\n    def __len__(self):\n        return self._cum_sizes[-1]\n\n    def __repr__(self):\n        # remove uselessly long transform\n        return \" + \".join(\n            repr(dataset).replace(\n                \",transform=Compose( ToTensor() Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)))\",\n                \"\",\n            )\n            for dataset in self.datasets\n        )\n\n    def set_epoch(self, epoch):\n        for dataset in self.datasets:\n            dataset.set_epoch(epoch)\n\n    def __getitem__(self, idx):\n        other = None","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.__len__","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.__len__#L153-L154","kind":"function","name":"__len__","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":153,"end_line":154,"context_start_line":133,"context_end_line":174,"code":"            return self.dataset[self._idxs_mapping[idx]]\n\n    @property\n    def _resolutions(self):\n        return self.dataset._resolutions\n\n    @property\n    def num_views(self):\n        return self.dataset.num_views\n\n\nclass CatDataset(EasyDataset):\n    \"\"\"Concatenation of several datasets\"\"\"\n\n    def __init__(self, datasets):\n        for dataset in datasets:\n            assert isinstance(dataset, EasyDataset)\n        self.datasets = datasets\n        self._cum_sizes = np.cumsum([len(dataset) for dataset in datasets])\n\n    def __len__(self):\n        return self._cum_sizes[-1]\n\n    def __repr__(self):\n        # remove uselessly long transform\n        return \" + \".join(\n            repr(dataset).replace(\n                \",transform=Compose( ToTensor() Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)))\",\n                \"\",\n            )\n            for dataset in self.datasets\n        )\n\n    def set_epoch(self, epoch):\n        for dataset in self.datasets:\n            dataset.set_epoch(epoch)\n\n    def __getitem__(self, idx):\n        other = None\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.__repr__","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.__repr__#L156-L164","kind":"function","name":"__repr__","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":156,"end_line":164,"context_start_line":136,"context_end_line":184,"code":"    def _resolutions(self):\n        return self.dataset._resolutions\n\n    @property\n    def num_views(self):\n        return self.dataset.num_views\n\n\nclass CatDataset(EasyDataset):\n    \"\"\"Concatenation of several datasets\"\"\"\n\n    def __init__(self, datasets):\n        for dataset in datasets:\n            assert isinstance(dataset, EasyDataset)\n        self.datasets = datasets\n        self._cum_sizes = np.cumsum([len(dataset) for dataset in datasets])\n\n    def __len__(self):\n        return self._cum_sizes[-1]\n\n    def __repr__(self):\n        # remove uselessly long transform\n        return \" + \".join(\n            repr(dataset).replace(\n                \",transform=Compose( ToTensor() Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)))\",\n                \"\",\n            )\n            for dataset in self.datasets\n        )\n\n    def set_epoch(self, epoch):\n        for dataset in self.datasets:\n            dataset.set_epoch(epoch)\n\n    def __getitem__(self, idx):\n        other = None\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n\n        if not (0 <= idx < len(self)):\n            raise IndexError()\n\n        db_idx = np.searchsorted(self._cum_sizes, idx, \"right\")\n        dataset = self.datasets[db_idx]\n        new_idx = idx - (self._cum_sizes[db_idx - 1] if db_idx > 0 else 0)\n\n        if other is not None and another is not None:\n            new_idx = (new_idx, other, another)\n        return dataset[new_idx]","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.__getitem__","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.__getitem__#L170-L184","kind":"function","name":"__getitem__","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":170,"end_line":184,"context_start_line":150,"context_end_line":198,"code":"        self.datasets = datasets\n        self._cum_sizes = np.cumsum([len(dataset) for dataset in datasets])\n\n    def __len__(self):\n        return self._cum_sizes[-1]\n\n    def __repr__(self):\n        # remove uselessly long transform\n        return \" + \".join(\n            repr(dataset).replace(\n                \",transform=Compose( ToTensor() Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)))\",\n                \"\",\n            )\n            for dataset in self.datasets\n        )\n\n    def set_epoch(self, epoch):\n        for dataset in self.datasets:\n            dataset.set_epoch(epoch)\n\n    def __getitem__(self, idx):\n        other = None\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n\n        if not (0 <= idx < len(self)):\n            raise IndexError()\n\n        db_idx = np.searchsorted(self._cum_sizes, idx, \"right\")\n        dataset = self.datasets[db_idx]\n        new_idx = idx - (self._cum_sizes[db_idx - 1] if db_idx > 0 else 0)\n\n        if other is not None and another is not None:\n            new_idx = (new_idx, other, another)\n        return dataset[new_idx]\n\n    @property\n    def _resolutions(self):\n        resolutions = self.datasets[0]._resolutions\n        for dataset in self.datasets[1:]:\n            assert tuple(dataset._resolutions) == tuple(resolutions)\n        return resolutions\n\n    @property\n    def num_views(self):\n        num_views = self.datasets[0].num_views\n        for dataset in self.datasets[1:]:\n            assert dataset.num_views == num_views\n        return num_views","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset._resolutions","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset._resolutions#L187-L191","kind":"function","name":"_resolutions","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":187,"end_line":191,"context_start_line":167,"context_end_line":198,"code":"        for dataset in self.datasets:\n            dataset.set_epoch(epoch)\n\n    def __getitem__(self, idx):\n        other = None\n        if isinstance(idx, tuple):\n            idx, other, another = idx\n\n        if not (0 <= idx < len(self)):\n            raise IndexError()\n\n        db_idx = np.searchsorted(self._cum_sizes, idx, \"right\")\n        dataset = self.datasets[db_idx]\n        new_idx = idx - (self._cum_sizes[db_idx - 1] if db_idx > 0 else 0)\n\n        if other is not None and another is not None:\n            new_idx = (new_idx, other, another)\n        return dataset[new_idx]\n\n    @property\n    def _resolutions(self):\n        resolutions = self.datasets[0]._resolutions\n        for dataset in self.datasets[1:]:\n            assert tuple(dataset._resolutions) == tuple(resolutions)\n        return resolutions\n\n    @property\n    def num_views(self):\n        num_views = self.datasets[0].num_views\n        for dataset in self.datasets[1:]:\n            assert dataset.num_views == num_views\n        return num_views","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.datasets.base.easy_dataset.num_views","uri":"program://Human3R/function/src.dust3r.datasets.base.easy_dataset.num_views#L194-L198","kind":"function","name":"num_views","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":194,"end_line":198,"context_start_line":174,"context_end_line":198,"code":"\n        if not (0 <= idx < len(self)):\n            raise IndexError()\n\n        db_idx = np.searchsorted(self._cum_sizes, idx, \"right\")\n        dataset = self.datasets[db_idx]\n        new_idx = idx - (self._cum_sizes[db_idx - 1] if db_idx > 0 else 0)\n\n        if other is not None and another is not None:\n            new_idx = (new_idx, other, another)\n        return dataset[new_idx]\n\n    @property\n    def _resolutions(self):\n        resolutions = self.datasets[0]._resolutions\n        for dataset in self.datasets[1:]:\n            assert tuple(dataset._resolutions) == tuple(resolutions)\n        return resolutions\n\n    @property\n    def num_views(self):\n        num_views = self.datasets[0].num_views\n        for dataset in self.datasets[1:]:\n            assert dataset.num_views == num_views\n        return num_views","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head","uri":"program://Human3R/module/src.dust3r.heads.dpt_head#L1-L671","kind":"module","name":"src.dust3r.heads.dpt_head","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":1,"end_line":671,"context_start_line":1,"context_end_line":671,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nfrom einops import rearrange\nfrom typing import List\nimport torch\nimport torch.nn as nn\nfrom dust3r.heads.postprocess import (\n    postprocess,\n    postprocess_desc,\n    postprocess_rgb,\n    postprocess_pose_conf,\n    postprocess_pose,\n    reg_dense_conf,\n    postprocess_smpl,\n    postprocess_score,\n)\nimport dust3r.utils.path_to_croco  # noqa: F401\nfrom models.dpt_block import DPTOutputAdapter, Interpolate  # noqa\nfrom dust3r.utils.camera import pose_encoding_to_camera, PoseDecoder\nfrom dust3r.blocks import ConditionModulationBlock\nfrom torch.utils.checkpoint import checkpoint\nfrom dust3r.smpl_model import SMPLDecoder, MEAN_PARAMS, regression_mlp\nimport numpy as np\n\n\nclass DPTOutputAdapter_fix(DPTOutputAdapter):\n    \"\"\"\n    Adapt croco's DPTOutputAdapter implementation for dust3r:\n    remove duplicated weigths, and fix forward for dust3r\n    \"\"\"\n\n    def init(self, dim_tokens_enc=768):\n        super().init(dim_tokens_enc)\n\n        del self.act_1_postprocess\n        del self.act_2_postprocess\n        del self.act_3_postprocess\n        del self.act_4_postprocess\n\n    def forward(self, encoder_tokens: List[torch.Tensor], image_size=None, ret_feat=False):\n        assert (\n            self.dim_tokens_enc is not None\n        ), \"Need to call init(dim_tokens_enc) function first\"\n\n        image_size = self.image_size if image_size is None else image_size\n        H, W = image_size\n\n        N_H = H // (self.stride_level * self.P_H)\n        N_W = W // (self.stride_level * self.P_W)\n\n        layers = [encoder_tokens[hook] for hook in self.hooks]\n\n        layers = [self.adapt_tokens(l) for l in layers]\n\n        layers = [\n            rearrange(l, \"b (nh nw) c -> b c nh nw\", nh=N_H, nw=N_W) for l in layers\n        ]\n\n        layers = [self.act_postprocess[idx](l) for idx, l in enumerate(layers)]\n\n        layers = [self.scratch.layer_rn[idx](l) for idx, l in enumerate(layers)]\n\n        path_4 = self.scratch.refinenet4(layers[3])[\n            :, :, : layers[2].shape[2], : layers[2].shape[3]\n        ]\n        path_3 = self.scratch.refinenet3(path_4, layers[2])\n        path_2 = self.scratch.refinenet2(path_3, layers[1])\n        path_1 = self.scratch.refinenet1(path_2, layers[0])\n\n        out = self.head(path_1)\n        \n        if ret_feat:\n            return out, path_1\n\n        return out\n\n\nclass PixelwiseTaskWithDPT(nn.Module):\n    \"\"\"DPT module for dust3r, can return 3D points + confidence for all pixels\"\"\"\n\n    def __init__(\n        self,\n        *,\n        n_cls_token=0,\n        hooks_idx=None,\n        dim_tokens=None,\n        output_width_ratio=1,\n        num_channels=1,\n        postprocess=None,\n        depth_mode=None,\n        conf_mode=None,\n        **kwargs\n    ):\n        super(PixelwiseTaskWithDPT, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.postprocess = postprocess\n        self.depth_mode = depth_mode\n        self.conf_mode = conf_mode\n\n        assert n_cls_token == 0, \"Not implemented\"\n        dpt_args = dict(\n            output_width_ratio=output_width_ratio, num_channels=num_channels, **kwargs\n        )\n        if hooks_idx is not None:\n            dpt_args.update(hooks=hooks_idx)\n        self.dpt = DPTOutputAdapter_fix(**dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt.init(**dpt_init_args)\n\n    def forward(self, x, img_info):\n        out = self.dpt(x, image_size=(img_info[0], img_info[1]))\n        if self.postprocess:\n            out = self.postprocess(out, self.depth_mode, self.conf_mode)\n        return out\n\n\ndef create_dpt_head(net, has_conf=False):\n    \"\"\"\n    return PixelwiseTaskWithDPT for given net params\n    \"\"\"\n    assert net.dec_depth > 9\n    l2 = net.dec_depth\n    feature_dim = 256\n    last_dim = feature_dim // 2\n    out_nchan = 3\n    ed = net.enc_embed_dim\n    dd = net.dec_embed_dim\n    return PixelwiseTaskWithDPT(\n        num_channels=out_nchan + has_conf,\n        feature_dim=feature_dim,\n        last_dim=last_dim,\n        hooks_idx=[0, l2 * 2 // 4, l2 * 3 // 4, l2],\n        dim_tokens=[ed, dd, dd, dd],\n        postprocess=postprocess,\n        depth_mode=net.depth_mode,\n        conf_mode=net.conf_mode,\n        head_type=\"regression\",\n    )\n\n\nclass DPTPts3dPose(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super(DPTPts3dPose, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim\n        dd = net.dec_embed_dim\n        hooks_idx = [0, 1, 2, 3]\n        dim_tokens = [ed, dd, dd, dd]\n        head_type = \"regression\"\n        output_width_ratio = 1\n\n        pts_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=pts_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        rgb_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=rgb_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        if hooks_idx is not None:\n            pts_dpt_args.update(hooks=hooks_idx)\n            rgb_dpt_args.update(hooks=hooks_idx)\n\n        self.dpt_self = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_self.init(**dpt_init_args)\n\n        self.final_transform = nn.ModuleList(\n            [\n                ConditionModulationBlock(\n                    net.dec_embed_dim,\n                    net.dec_num_heads,\n                    mlp_ratio=4.0,\n                    qkv_bias=True,\n                    rope=net.rope,\n                )\n                for _ in range(2)\n            ]\n        )\n\n        self.dpt_cross = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_cross.init(**dpt_init_args)\n\n        if has_rgb:\n            self.dpt_rgb = DPTOutputAdapter_fix(**rgb_dpt_args)\n            dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n            self.dpt_rgb.init(**dpt_init_args)\n\n        if has_pose:\n            in_dim = net.dec_embed_dim\n            self.pose_head = PoseDecoder(hidden_size=in_dim)\n\n    def forward(self, x, img_info, **kwargs):\n        if self.has_pose:\n            pose_token = x[-1][:, 0].clone()\n            token = x[-1][:, 1:]\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n\n            token_cross = token.clone()\n            for blk in self.final_transform:\n                token_cross = blk(token_cross, pose_token, kwargs.get(\"pos\"))\n            x = x[:-1] + [token]\n            x_cross = x[:-1] + [token_cross]\n\n        with torch.cuda.amp.autocast(enabled=False):\n            self_out = checkpoint(\n                self.dpt_self,\n                x,\n                image_size=(img_info[0], img_info[1]),\n                use_reentrant=False,\n            )\n\n            final_output = postprocess(self_out, self.depth_mode, self.conf_mode)\n            final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n            final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n            if self.has_rgb:\n                rgb_out = checkpoint(\n                    self.dpt_rgb,\n                    x,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                rgb_output = postprocess_rgb(rgb_out)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n                cross_out = checkpoint(\n                    self.dpt_cross,\n                    x_cross,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                tmp = postprocess(cross_out, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n        return final_output\n\n\nclass DPTPts3dPoseSMPL(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False, has_msk=False):\n        super(DPTPts3dPoseSMPL, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n        self.has_msk = has_msk\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim\n        dd = net.dec_embed_dim\n        hooks_idx = [0, 1, 2, 3]\n        dim_tokens = [ed, dd, dd, dd]\n        head_type = \"regression\"\n        output_width_ratio = 1\n\n        pts_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=pts_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        rgb_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=rgb_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        if hooks_idx is not None:\n            pts_dpt_args.update(hooks=hooks_idx)\n            rgb_dpt_args.update(hooks=hooks_idx)\n\n        self.dpt_self = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_self.init(**dpt_init_args)\n\n        self.final_transform = nn.ModuleList(\n            [\n                ConditionModulationBlock(\n                    net.dec_embed_dim,\n                    net.dec_num_heads,\n                    mlp_ratio=4.0,\n                    qkv_bias=True,\n                    rope=net.rope,\n                )\n                for _ in range(2)\n            ]\n        )\n\n        self.dpt_cross = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_cross.init(**dpt_init_args)\n\n        if has_rgb:\n            self.dpt_rgb = DPTOutputAdapter_fix(**rgb_dpt_args)\n            dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n            self.dpt_rgb.init(**dpt_init_args)\n\n        if has_pose:\n            in_dim = net.dec_embed_dim\n            self.pose_head = PoseDecoder(hidden_size=in_dim)\n            self.in_dim = in_dim\n\n        # MHMR Heads - Detection\n        backbone_dim = net.backbone_dim\n        self.bb_patch_size = net.bb_patch_size\n        self.mlp_classif = regression_mlp([backbone_dim, backbone_dim, 1]) # bg or human\n        self.mlp_offset = regression_mlp([backbone_dim, backbone_dim, 2]) # offset\n        if has_msk:\n            self.mlp_msk = SMPLDecoder(\n                hidden_size=backbone_dim, target_dim=self.bb_patch_size**2, num_layers=2, mlp_ratio=1)\n\n        # feature fuse\n        self.mlp_fuse = SMPLDecoder(hidden_size=ed+backbone_dim, target_dim=dd, num_layers=2, mlp_ratio=4)\n\n        # SMPL\n        self.joint_rep_type, self.joint_rep_dim = '6d', 6\n        self.nrot = 53\n        self.num_body_joints = self.nrot - 1\n\n        npose = self.joint_rep_dim * (self.num_body_joints + 1)\n        self.npose = npose\n\n        self.input_is_mean_shape = True\n        self.num_betas = 10\n        assert self.num_betas in [10, 11]\n\n        # SMPL param heads\n        self.deccam = SMPLDecoder(\n                        hidden_size=in_dim, \n                        target_dim=3, \n                        num_layers=2, \n                        mlp_ratio=4)\n        self.decpose, self.decshape, self.decexpression = [\n            SMPLDecoder(\n                hidden_size=in_dim+backbone_dim, \n                target_dim=od, \n                num_layers=2, \n                mlp_ratio=4) for od in [self.npose, self.num_betas, 10]]\n\n        self.set_smpl_init()\n\n\n    def set_smpl_init(self):\n        \"\"\" Fetch saved SMPL parameters and register buffers.\"\"\"\n        mean_params = np.load(MEAN_PARAMS)\n        if self.nrot == 53:\n            init_body_pose = torch.eye(3).reshape(1,3,3).repeat(self.nrot,1,1)[:,:,:2].flatten(1).reshape(1, -1)\n            init_body_pose[:,:24*6] = torch.from_numpy(mean_params['pose'][:]).float() # global_orient+body_pose from SMPL\n        else:\n            init_body_pose = torch.from_numpy(mean_params['pose'].astype(np.float32)).unsqueeze(0)\n\n        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)\n        init_betas_kid = torch.cat([init_betas, torch.zeros_like(init_betas[:,[0]])],1)\n        init_expression = 0. * torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n\n        if self.num_betas == 11:\n            init_betas = torch.cat([init_betas, torch.zeros_like(init_betas[:,:1])], 1)\n\n        self.register_buffer('init_body_pose', init_body_pose)\n        self.register_buffer('init_betas', init_betas)\n        self.register_buffer('init_betas_kid', init_betas_kid)\n        self.register_buffer('init_cam', init_cam)\n        self.register_buffer('init_expression', init_expression)\n        \n    def detect_mhmr(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            scores = postprocess_score(self.mlp_classif(x)) # per token detection score.\n        return scores\n\n    def segment(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            msks = postprocess_score(self.mlp_msk(x))\n        return msks\n    \n    def forward(self, x, img_info, **kwargs):\n        if self.has_pose:\n            pose_token = x[-1][:, 0].clone()\n            n_humans_i = kwargs.get(\"n_humans\")\n            token = x[-1][:, 1:]\n\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n                if n_humans_i > 0:\n                    smpl_token = kwargs.get(\"smpl_token\")   # CUT3R smpl token (bs, 10, 768)\n                    pred_body_pose = self.decpose(smpl_token) + self.init_body_pose\n                    pred_betas = self.decshape(smpl_token) + self.init_betas\n                    pred_cam = self.deccam(smpl_token[..., :self.in_dim])\n                    pred_expression = self.decexpression(smpl_token) + self.init_expression\n                    pred_smpl = [pred_body_pose, pred_betas, pred_cam, pred_expression]\n                    \n            token_cross = token.clone()\n            for blk in self.final_transform:\n                token_cross = blk(token_cross, pose_token, kwargs.get(\"pos\"))\n            x = x[:-1] + [token]\n            x_cross = x[:-1] + [token_cross]\n\n        with torch.cuda.amp.autocast(enabled=False):\n            self_out = checkpoint(\n                self.dpt_self,\n                x,\n                image_size=(img_info[0], img_info[1]),\n                # ret_feat=True,\n                use_reentrant=False,\n            )\n\n            final_output = postprocess(self_out, self.depth_mode, self.conf_mode)\n            final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n            final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n            if self.has_rgb:\n                rgb_out = checkpoint(\n                    self.dpt_rgb,\n                    x,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                rgb_output = postprocess_rgb(rgb_out)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n                cross_out = checkpoint(\n                    self.dpt_cross,\n                    x_cross,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                tmp = postprocess(cross_out, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n            \n            if n_humans_i > 0:\n                smpl_out = postprocess_smpl(pred_smpl, self.depth_mode)\n                final_output.update(smpl_out)\n\n        return final_output\n\n\nclass NaiveDPTPts3dPoseSMPL(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super(NaiveDPTPts3dPoseSMPL, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim\n        dd = net.dec_embed_dim\n        hooks_idx = [0, 1, 2, 3]\n        dim_tokens = [ed, dd, dd, dd]\n        head_type = \"regression\"\n        output_width_ratio = 1\n\n        pts_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=pts_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        rgb_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=rgb_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        if hooks_idx is not None:\n            pts_dpt_args.update(hooks=hooks_idx)\n            rgb_dpt_args.update(hooks=hooks_idx)\n\n        self.dpt_self = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_self.init(**dpt_init_args)\n\n        self.final_transform = nn.ModuleList(\n            [\n                ConditionModulationBlock(\n                    net.dec_embed_dim,\n                    net.dec_num_heads,\n                    mlp_ratio=4.0,\n                    qkv_bias=True,\n                    rope=net.rope,\n                )\n                for _ in range(2)\n            ]\n        )\n\n        self.dpt_cross = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_cross.init(**dpt_init_args)\n\n        if has_rgb:\n            self.dpt_rgb = DPTOutputAdapter_fix(**rgb_dpt_args)\n            dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n            self.dpt_rgb.init(**dpt_\n# ... truncated ...","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":true}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.DPTOutputAdapter_fix","uri":"program://Human3R/class/src.dust3r.heads.dpt_head.DPTOutputAdapter_fix#L30-L79","kind":"class","name":"DPTOutputAdapter_fix","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":30,"end_line":79,"context_start_line":10,"context_end_line":99,"code":"import torch.nn as nn\nfrom dust3r.heads.postprocess import (\n    postprocess,\n    postprocess_desc,\n    postprocess_rgb,\n    postprocess_pose_conf,\n    postprocess_pose,\n    reg_dense_conf,\n    postprocess_smpl,\n    postprocess_score,\n)\nimport dust3r.utils.path_to_croco  # noqa: F401\nfrom models.dpt_block import DPTOutputAdapter, Interpolate  # noqa\nfrom dust3r.utils.camera import pose_encoding_to_camera, PoseDecoder\nfrom dust3r.blocks import ConditionModulationBlock\nfrom torch.utils.checkpoint import checkpoint\nfrom dust3r.smpl_model import SMPLDecoder, MEAN_PARAMS, regression_mlp\nimport numpy as np\n\n\nclass DPTOutputAdapter_fix(DPTOutputAdapter):\n    \"\"\"\n    Adapt croco's DPTOutputAdapter implementation for dust3r:\n    remove duplicated weigths, and fix forward for dust3r\n    \"\"\"\n\n    def init(self, dim_tokens_enc=768):\n        super().init(dim_tokens_enc)\n\n        del self.act_1_postprocess\n        del self.act_2_postprocess\n        del self.act_3_postprocess\n        del self.act_4_postprocess\n\n    def forward(self, encoder_tokens: List[torch.Tensor], image_size=None, ret_feat=False):\n        assert (\n            self.dim_tokens_enc is not None\n        ), \"Need to call init(dim_tokens_enc) function first\"\n\n        image_size = self.image_size if image_size is None else image_size\n        H, W = image_size\n\n        N_H = H // (self.stride_level * self.P_H)\n        N_W = W // (self.stride_level * self.P_W)\n\n        layers = [encoder_tokens[hook] for hook in self.hooks]\n\n        layers = [self.adapt_tokens(l) for l in layers]\n\n        layers = [\n            rearrange(l, \"b (nh nw) c -> b c nh nw\", nh=N_H, nw=N_W) for l in layers\n        ]\n\n        layers = [self.act_postprocess[idx](l) for idx, l in enumerate(layers)]\n\n        layers = [self.scratch.layer_rn[idx](l) for idx, l in enumerate(layers)]\n\n        path_4 = self.scratch.refinenet4(layers[3])[\n            :, :, : layers[2].shape[2], : layers[2].shape[3]\n        ]\n        path_3 = self.scratch.refinenet3(path_4, layers[2])\n        path_2 = self.scratch.refinenet2(path_3, layers[1])\n        path_1 = self.scratch.refinenet1(path_2, layers[0])\n\n        out = self.head(path_1)\n        \n        if ret_feat:\n            return out, path_1\n\n        return out\n\n\nclass PixelwiseTaskWithDPT(nn.Module):\n    \"\"\"DPT module for dust3r, can return 3D points + confidence for all pixels\"\"\"\n\n    def __init__(\n        self,\n        *,\n        n_cls_token=0,\n        hooks_idx=None,\n        dim_tokens=None,\n        output_width_ratio=1,\n        num_channels=1,\n        postprocess=None,\n        depth_mode=None,\n        conf_mode=None,\n        **kwargs\n    ):\n        super(PixelwiseTaskWithDPT, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.PixelwiseTaskWithDPT","uri":"program://Human3R/class/src.dust3r.heads.dpt_head.PixelwiseTaskWithDPT#L82-L118","kind":"class","name":"PixelwiseTaskWithDPT","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":82,"end_line":118,"context_start_line":62,"context_end_line":138,"code":"\n        layers = [self.act_postprocess[idx](l) for idx, l in enumerate(layers)]\n\n        layers = [self.scratch.layer_rn[idx](l) for idx, l in enumerate(layers)]\n\n        path_4 = self.scratch.refinenet4(layers[3])[\n            :, :, : layers[2].shape[2], : layers[2].shape[3]\n        ]\n        path_3 = self.scratch.refinenet3(path_4, layers[2])\n        path_2 = self.scratch.refinenet2(path_3, layers[1])\n        path_1 = self.scratch.refinenet1(path_2, layers[0])\n\n        out = self.head(path_1)\n        \n        if ret_feat:\n            return out, path_1\n\n        return out\n\n\nclass PixelwiseTaskWithDPT(nn.Module):\n    \"\"\"DPT module for dust3r, can return 3D points + confidence for all pixels\"\"\"\n\n    def __init__(\n        self,\n        *,\n        n_cls_token=0,\n        hooks_idx=None,\n        dim_tokens=None,\n        output_width_ratio=1,\n        num_channels=1,\n        postprocess=None,\n        depth_mode=None,\n        conf_mode=None,\n        **kwargs\n    ):\n        super(PixelwiseTaskWithDPT, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.postprocess = postprocess\n        self.depth_mode = depth_mode\n        self.conf_mode = conf_mode\n\n        assert n_cls_token == 0, \"Not implemented\"\n        dpt_args = dict(\n            output_width_ratio=output_width_ratio, num_channels=num_channels, **kwargs\n        )\n        if hooks_idx is not None:\n            dpt_args.update(hooks=hooks_idx)\n        self.dpt = DPTOutputAdapter_fix(**dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt.init(**dpt_init_args)\n\n    def forward(self, x, img_info):\n        out = self.dpt(x, image_size=(img_info[0], img_info[1]))\n        if self.postprocess:\n            out = self.postprocess(out, self.depth_mode, self.conf_mode)\n        return out\n\n\ndef create_dpt_head(net, has_conf=False):\n    \"\"\"\n    return PixelwiseTaskWithDPT for given net params\n    \"\"\"\n    assert net.dec_depth > 9\n    l2 = net.dec_depth\n    feature_dim = 256\n    last_dim = feature_dim // 2\n    out_nchan = 3\n    ed = net.enc_embed_dim\n    dd = net.dec_embed_dim\n    return PixelwiseTaskWithDPT(\n        num_channels=out_nchan + has_conf,\n        feature_dim=feature_dim,\n        last_dim=last_dim,\n        hooks_idx=[0, l2 * 2 // 4, l2 * 3 // 4, l2],\n        dim_tokens=[ed, dd, dd, dd],\n        postprocess=postprocess,","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.create_dpt_head","uri":"program://Human3R/function/src.dust3r.heads.dpt_head.create_dpt_head#L121-L142","kind":"function","name":"create_dpt_head","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":121,"end_line":142,"context_start_line":101,"context_end_line":162,"code":"        self.depth_mode = depth_mode\n        self.conf_mode = conf_mode\n\n        assert n_cls_token == 0, \"Not implemented\"\n        dpt_args = dict(\n            output_width_ratio=output_width_ratio, num_channels=num_channels, **kwargs\n        )\n        if hooks_idx is not None:\n            dpt_args.update(hooks=hooks_idx)\n        self.dpt = DPTOutputAdapter_fix(**dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt.init(**dpt_init_args)\n\n    def forward(self, x, img_info):\n        out = self.dpt(x, image_size=(img_info[0], img_info[1]))\n        if self.postprocess:\n            out = self.postprocess(out, self.depth_mode, self.conf_mode)\n        return out\n\n\ndef create_dpt_head(net, has_conf=False):\n    \"\"\"\n    return PixelwiseTaskWithDPT for given net params\n    \"\"\"\n    assert net.dec_depth > 9\n    l2 = net.dec_depth\n    feature_dim = 256\n    last_dim = feature_dim // 2\n    out_nchan = 3\n    ed = net.enc_embed_dim\n    dd = net.dec_embed_dim\n    return PixelwiseTaskWithDPT(\n        num_channels=out_nchan + has_conf,\n        feature_dim=feature_dim,\n        last_dim=last_dim,\n        hooks_idx=[0, l2 * 2 // 4, l2 * 3 // 4, l2],\n        dim_tokens=[ed, dd, dd, dd],\n        postprocess=postprocess,\n        depth_mode=net.depth_mode,\n        conf_mode=net.conf_mode,\n        head_type=\"regression\",\n    )\n\n\nclass DPTPts3dPose(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super(DPTPts3dPose, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim\n        dd = net.dec_embed_dim","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.DPTPts3dPose","uri":"program://Human3R/class/src.dust3r.heads.dpt_head.DPTPts3dPose#L145-L267","kind":"class","name":"DPTPts3dPose","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":145,"end_line":267,"context_start_line":125,"context_end_line":287,"code":"    assert net.dec_depth > 9\n    l2 = net.dec_depth\n    feature_dim = 256\n    last_dim = feature_dim // 2\n    out_nchan = 3\n    ed = net.enc_embed_dim\n    dd = net.dec_embed_dim\n    return PixelwiseTaskWithDPT(\n        num_channels=out_nchan + has_conf,\n        feature_dim=feature_dim,\n        last_dim=last_dim,\n        hooks_idx=[0, l2 * 2 // 4, l2 * 3 // 4, l2],\n        dim_tokens=[ed, dd, dd, dd],\n        postprocess=postprocess,\n        depth_mode=net.depth_mode,\n        conf_mode=net.conf_mode,\n        head_type=\"regression\",\n    )\n\n\nclass DPTPts3dPose(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super(DPTPts3dPose, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim\n        dd = net.dec_embed_dim\n        hooks_idx = [0, 1, 2, 3]\n        dim_tokens = [ed, dd, dd, dd]\n        head_type = \"regression\"\n        output_width_ratio = 1\n\n        pts_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=pts_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        rgb_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=rgb_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        if hooks_idx is not None:\n            pts_dpt_args.update(hooks=hooks_idx)\n            rgb_dpt_args.update(hooks=hooks_idx)\n\n        self.dpt_self = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_self.init(**dpt_init_args)\n\n        self.final_transform = nn.ModuleList(\n            [\n                ConditionModulationBlock(\n                    net.dec_embed_dim,\n                    net.dec_num_heads,\n                    mlp_ratio=4.0,\n                    qkv_bias=True,\n                    rope=net.rope,\n                )\n                for _ in range(2)\n            ]\n        )\n\n        self.dpt_cross = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_cross.init(**dpt_init_args)\n\n        if has_rgb:\n            self.dpt_rgb = DPTOutputAdapter_fix(**rgb_dpt_args)\n            dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n            self.dpt_rgb.init(**dpt_init_args)\n\n        if has_pose:\n            in_dim = net.dec_embed_dim\n            self.pose_head = PoseDecoder(hidden_size=in_dim)\n\n    def forward(self, x, img_info, **kwargs):\n        if self.has_pose:\n            pose_token = x[-1][:, 0].clone()\n            token = x[-1][:, 1:]\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n\n            token_cross = token.clone()\n            for blk in self.final_transform:\n                token_cross = blk(token_cross, pose_token, kwargs.get(\"pos\"))\n            x = x[:-1] + [token]\n            x_cross = x[:-1] + [token_cross]\n\n        with torch.cuda.amp.autocast(enabled=False):\n            self_out = checkpoint(\n                self.dpt_self,\n                x,\n                image_size=(img_info[0], img_info[1]),\n                use_reentrant=False,\n            )\n\n            final_output = postprocess(self_out, self.depth_mode, self.conf_mode)\n            final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n            final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n            if self.has_rgb:\n                rgb_out = checkpoint(\n                    self.dpt_rgb,\n                    x,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                rgb_output = postprocess_rgb(rgb_out)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n                cross_out = checkpoint(\n                    self.dpt_cross,\n                    x_cross,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                tmp = postprocess(cross_out, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n        return final_output\n\n\nclass DPTPts3dPoseSMPL(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False, has_msk=False):\n        super(DPTPts3dPoseSMPL, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n        self.has_msk = has_msk\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.DPTPts3dPoseSMPL","uri":"program://Human3R/class/src.dust3r.heads.dpt_head.DPTPts3dPoseSMPL#L270-L482","kind":"class","name":"DPTPts3dPoseSMPL","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":270,"end_line":482,"context_start_line":250,"context_end_line":502,"code":"                    use_reentrant=False,\n                )\n                rgb_output = postprocess_rgb(rgb_out)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n                cross_out = checkpoint(\n                    self.dpt_cross,\n                    x_cross,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                tmp = postprocess(cross_out, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n        return final_output\n\n\nclass DPTPts3dPoseSMPL(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False, has_msk=False):\n        super(DPTPts3dPoseSMPL, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n        self.has_msk = has_msk\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim\n        dd = net.dec_embed_dim\n        hooks_idx = [0, 1, 2, 3]\n        dim_tokens = [ed, dd, dd, dd]\n        head_type = \"regression\"\n        output_width_ratio = 1\n\n        pts_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=pts_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        rgb_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=rgb_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        if hooks_idx is not None:\n            pts_dpt_args.update(hooks=hooks_idx)\n            rgb_dpt_args.update(hooks=hooks_idx)\n\n        self.dpt_self = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_self.init(**dpt_init_args)\n\n        self.final_transform = nn.ModuleList(\n            [\n                ConditionModulationBlock(\n                    net.dec_embed_dim,\n                    net.dec_num_heads,\n                    mlp_ratio=4.0,\n                    qkv_bias=True,\n                    rope=net.rope,\n                )\n                for _ in range(2)\n            ]\n        )\n\n        self.dpt_cross = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_cross.init(**dpt_init_args)\n\n        if has_rgb:\n            self.dpt_rgb = DPTOutputAdapter_fix(**rgb_dpt_args)\n            dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n            self.dpt_rgb.init(**dpt_init_args)\n\n        if has_pose:\n            in_dim = net.dec_embed_dim\n            self.pose_head = PoseDecoder(hidden_size=in_dim)\n            self.in_dim = in_dim\n\n        # MHMR Heads - Detection\n        backbone_dim = net.backbone_dim\n        self.bb_patch_size = net.bb_patch_size\n        self.mlp_classif = regression_mlp([backbone_dim, backbone_dim, 1]) # bg or human\n        self.mlp_offset = regression_mlp([backbone_dim, backbone_dim, 2]) # offset\n        if has_msk:\n            self.mlp_msk = SMPLDecoder(\n                hidden_size=backbone_dim, target_dim=self.bb_patch_size**2, num_layers=2, mlp_ratio=1)\n\n        # feature fuse\n        self.mlp_fuse = SMPLDecoder(hidden_size=ed+backbone_dim, target_dim=dd, num_layers=2, mlp_ratio=4)\n\n        # SMPL\n        self.joint_rep_type, self.joint_rep_dim = '6d', 6\n        self.nrot = 53\n        self.num_body_joints = self.nrot - 1\n\n        npose = self.joint_rep_dim * (self.num_body_joints + 1)\n        self.npose = npose\n\n        self.input_is_mean_shape = True\n        self.num_betas = 10\n        assert self.num_betas in [10, 11]\n\n        # SMPL param heads\n        self.deccam = SMPLDecoder(\n                        hidden_size=in_dim, \n                        target_dim=3, \n                        num_layers=2, \n                        mlp_ratio=4)\n        self.decpose, self.decshape, self.decexpression = [\n            SMPLDecoder(\n                hidden_size=in_dim+backbone_dim, \n                target_dim=od, \n                num_layers=2, \n                mlp_ratio=4) for od in [self.npose, self.num_betas, 10]]\n\n        self.set_smpl_init()\n\n\n    def set_smpl_init(self):\n        \"\"\" Fetch saved SMPL parameters and register buffers.\"\"\"\n        mean_params = np.load(MEAN_PARAMS)\n        if self.nrot == 53:\n            init_body_pose = torch.eye(3).reshape(1,3,3).repeat(self.nrot,1,1)[:,:,:2].flatten(1).reshape(1, -1)\n            init_body_pose[:,:24*6] = torch.from_numpy(mean_params['pose'][:]).float() # global_orient+body_pose from SMPL\n        else:\n            init_body_pose = torch.from_numpy(mean_params['pose'].astype(np.float32)).unsqueeze(0)\n\n        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)\n        init_betas_kid = torch.cat([init_betas, torch.zeros_like(init_betas[:,[0]])],1)\n        init_expression = 0. * torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n\n        if self.num_betas == 11:\n            init_betas = torch.cat([init_betas, torch.zeros_like(init_betas[:,:1])], 1)\n\n        self.register_buffer('init_body_pose', init_body_pose)\n        self.register_buffer('init_betas', init_betas)\n        self.register_buffer('init_betas_kid', init_betas_kid)\n        self.register_buffer('init_cam', init_cam)\n        self.register_buffer('init_expression', init_expression)\n        \n    def detect_mhmr(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            scores = postprocess_score(self.mlp_classif(x)) # per token detection score.\n        return scores\n\n    def segment(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            msks = postprocess_score(self.mlp_msk(x))\n        return msks\n    \n    def forward(self, x, img_info, **kwargs):\n        if self.has_pose:\n            pose_token = x[-1][:, 0].clone()\n            n_humans_i = kwargs.get(\"n_humans\")\n            token = x[-1][:, 1:]\n\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n                if n_humans_i > 0:\n                    smpl_token = kwargs.get(\"smpl_token\")   # CUT3R smpl token (bs, 10, 768)\n                    pred_body_pose = self.decpose(smpl_token) + self.init_body_pose\n                    pred_betas = self.decshape(smpl_token) + self.init_betas\n                    pred_cam = self.deccam(smpl_token[..., :self.in_dim])\n                    pred_expression = self.decexpression(smpl_token) + self.init_expression\n                    pred_smpl = [pred_body_pose, pred_betas, pred_cam, pred_expression]\n                    \n            token_cross = token.clone()\n            for blk in self.final_transform:\n                token_cross = blk(token_cross, pose_token, kwargs.get(\"pos\"))\n            x = x[:-1] + [token]\n            x_cross = x[:-1] + [token_cross]\n\n        with torch.cuda.amp.autocast(enabled=False):\n            self_out = checkpoint(\n                self.dpt_self,\n                x,\n                image_size=(img_info[0], img_info[1]),\n                # ret_feat=True,\n                use_reentrant=False,\n            )\n\n            final_output = postprocess(self_out, self.depth_mode, self.conf_mode)\n            final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n            final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n            if self.has_rgb:\n                rgb_out = checkpoint(\n                    self.dpt_rgb,\n                    x,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                rgb_output = postprocess_rgb(rgb_out)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n                cross_out = checkpoint(\n                    self.dpt_cross,\n                    x_cross,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                tmp = postprocess(cross_out, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n            \n            if n_humans_i > 0:\n                smpl_out = postprocess_smpl(pred_smpl, self.depth_mode)\n                final_output.update(smpl_out)\n\n        return final_output\n\n\nclass NaiveDPTPts3dPoseSMPL(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super(NaiveDPTPts3dPoseSMPL, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim\n        dd = net.dec_embed_dim","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.NaiveDPTPts3dPoseSMPL","uri":"program://Human3R/class/src.dust3r.heads.dpt_head.NaiveDPTPts3dPoseSMPL#L485-L671","kind":"class","name":"NaiveDPTPts3dPoseSMPL","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":485,"end_line":671,"context_start_line":465,"context_end_line":671,"code":"            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n                cross_out = checkpoint(\n                    self.dpt_cross,\n                    x_cross,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                tmp = postprocess(cross_out, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n            \n            if n_humans_i > 0:\n                smpl_out = postprocess_smpl(pred_smpl, self.depth_mode)\n                final_output.update(smpl_out)\n\n        return final_output\n\n\nclass NaiveDPTPts3dPoseSMPL(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super(NaiveDPTPts3dPoseSMPL, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim\n        dd = net.dec_embed_dim\n        hooks_idx = [0, 1, 2, 3]\n        dim_tokens = [ed, dd, dd, dd]\n        head_type = \"regression\"\n        output_width_ratio = 1\n\n        pts_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=pts_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        rgb_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=rgb_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        if hooks_idx is not None:\n            pts_dpt_args.update(hooks=hooks_idx)\n            rgb_dpt_args.update(hooks=hooks_idx)\n\n        self.dpt_self = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_self.init(**dpt_init_args)\n\n        self.final_transform = nn.ModuleList(\n            [\n                ConditionModulationBlock(\n                    net.dec_embed_dim,\n                    net.dec_num_heads,\n                    mlp_ratio=4.0,\n                    qkv_bias=True,\n                    rope=net.rope,\n                )\n                for _ in range(2)\n            ]\n        )\n\n        self.dpt_cross = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_cross.init(**dpt_init_args)\n\n        if has_rgb:\n            self.dpt_rgb = DPTOutputAdapter_fix(**rgb_dpt_args)\n            dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n            self.dpt_rgb.init(**dpt_init_args)\n\n        if has_pose:\n            in_dim = net.dec_embed_dim\n            self.pose_head = PoseDecoder(hidden_size=in_dim)\n\n        # MHMR Heads - Detection\n        backbone_dim = net.backbone_dim\n        self.bb_patch_size = net.bb_patch_size\n        self.mlp_classif = regression_mlp([backbone_dim, backbone_dim, 1]) # bg or human\n        self.mlp_offset = regression_mlp([backbone_dim, backbone_dim, 2]) # offset\n\n        # SMPL\n        self.joint_rep_type, self.joint_rep_dim = '6d', 6\n        self.nrot = 53\n        self.num_body_joints = self.nrot - 1\n\n        npose = self.joint_rep_dim * (self.num_body_joints + 1)\n        self.npose = npose\n\n        self.input_is_mean_shape = True\n        self.num_betas = 10\n        assert self.num_betas in [10, 11]\n\n        # MHMR Heads - SMPL\n        self.decpose, self.decshape, self.deccam, self.decexpression = [\n            nn.Linear(1024, od) for od in [self.npose, self.num_betas, 3, 10]] # MLP(1024, x)\n\n        self.set_smpl_init()\n\n    def set_smpl_init(self):\n        \"\"\" Fetch saved SMPL parameters and register buffers.\"\"\"\n        mean_params = np.load(MEAN_PARAMS)\n        if self.nrot == 53:\n            init_body_pose = torch.eye(3).reshape(1,3,3).repeat(self.nrot,1,1)[:,:,:2].flatten(1).reshape(1, -1)\n            init_body_pose[:,:24*6] = torch.from_numpy(mean_params['pose'][:]).float() # global_orient+body_pose from SMPL\n        else:\n            init_body_pose = torch.from_numpy(mean_params['pose'].astype(np.float32)).unsqueeze(0)\n\n        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)\n        init_betas_kid = torch.cat([init_betas, torch.zeros_like(init_betas[:,[0]])],1)\n        init_expression = 0. * torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n\n        if self.num_betas == 11:\n            init_betas = torch.cat([init_betas, torch.zeros_like(init_betas[:,:1])], 1)\n\n        self.register_buffer('init_body_pose', init_body_pose)\n        self.register_buffer('init_betas', init_betas)\n        self.register_buffer('init_betas_kid', init_betas_kid)\n        self.register_buffer('init_cam', init_cam)\n        self.register_buffer('init_expression', init_expression)\n        \n    def detect_mhmr(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            scores = postprocess_score(self.mlp_classif(x)) # per token detection score.\n        return scores\n      \n    def forward(self, x, img_info, **kwargs):\n        if self.has_pose:\n            pose_token = x[-1][:, 0].clone()\n            n_humans_i = kwargs.get(\"n_humans\")\n            token = x[-1][:, 1:]\n\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n                if n_humans_i > 0:\n                    smpl_token = kwargs.get(\"smpl_token\") \n                    decoders = [self.decpose, self.decshape, self.deccam, self.decexpression]\n                    inits = [self.init_body_pose, self.init_betas, self.init_cam, self.init_expression]\n                    pred_smpl = [d(smpl_token) + i for d, i in zip(decoders, inits)]\n\n            token_cross = token.clone()\n            for blk in self.final_transform:\n                token_cross = blk(token_cross, pose_token, kwargs.get(\"pos\"))\n            x = x[:-1] + [token]\n            x_cross = x[:-1] + [token_cross]\n\n        with torch.cuda.amp.autocast(enabled=False):\n            self_out = checkpoint(\n                self.dpt_self,\n                x,\n                image_size=(img_info[0], img_info[1]),\n                use_reentrant=False,\n            )\n\n            final_output = postprocess(self_out, self.depth_mode, self.conf_mode)\n            final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n            final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n            if self.has_rgb:\n                rgb_out = checkpoint(\n                    self.dpt_rgb,\n                    x,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                rgb_output = postprocess_rgb(rgb_out)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n                cross_out = checkpoint(\n                    self.dpt_cross,\n                    x_cross,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                tmp = postprocess(cross_out, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n            \n            if n_humans_i > 0:\n                smpl_out = postprocess_smpl(pred_smpl, self.depth_mode, naive_mode=True)\n                final_output.update(smpl_out)\n\n        return final_output","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.init","uri":"program://Human3R/function/src.dust3r.heads.dpt_head.init#L36-L42","kind":"function","name":"init","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":36,"end_line":42,"context_start_line":16,"context_end_line":62,"code":"    postprocess_pose,\n    reg_dense_conf,\n    postprocess_smpl,\n    postprocess_score,\n)\nimport dust3r.utils.path_to_croco  # noqa: F401\nfrom models.dpt_block import DPTOutputAdapter, Interpolate  # noqa\nfrom dust3r.utils.camera import pose_encoding_to_camera, PoseDecoder\nfrom dust3r.blocks import ConditionModulationBlock\nfrom torch.utils.checkpoint import checkpoint\nfrom dust3r.smpl_model import SMPLDecoder, MEAN_PARAMS, regression_mlp\nimport numpy as np\n\n\nclass DPTOutputAdapter_fix(DPTOutputAdapter):\n    \"\"\"\n    Adapt croco's DPTOutputAdapter implementation for dust3r:\n    remove duplicated weigths, and fix forward for dust3r\n    \"\"\"\n\n    def init(self, dim_tokens_enc=768):\n        super().init(dim_tokens_enc)\n\n        del self.act_1_postprocess\n        del self.act_2_postprocess\n        del self.act_3_postprocess\n        del self.act_4_postprocess\n\n    def forward(self, encoder_tokens: List[torch.Tensor], image_size=None, ret_feat=False):\n        assert (\n            self.dim_tokens_enc is not None\n        ), \"Need to call init(dim_tokens_enc) function first\"\n\n        image_size = self.image_size if image_size is None else image_size\n        H, W = image_size\n\n        N_H = H // (self.stride_level * self.P_H)\n        N_W = W // (self.stride_level * self.P_W)\n\n        layers = [encoder_tokens[hook] for hook in self.hooks]\n\n        layers = [self.adapt_tokens(l) for l in layers]\n\n        layers = [\n            rearrange(l, \"b (nh nw) c -> b c nh nw\", nh=N_H, nw=N_W) for l in layers\n        ]\n","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.forward","uri":"program://Human3R/function/src.dust3r.heads.dpt_head.forward#L612-L671","kind":"function","name":"forward","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":612,"end_line":671,"context_start_line":592,"context_end_line":671,"code":"\n        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)\n        init_betas_kid = torch.cat([init_betas, torch.zeros_like(init_betas[:,[0]])],1)\n        init_expression = 0. * torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n\n        if self.num_betas == 11:\n            init_betas = torch.cat([init_betas, torch.zeros_like(init_betas[:,:1])], 1)\n\n        self.register_buffer('init_body_pose', init_body_pose)\n        self.register_buffer('init_betas', init_betas)\n        self.register_buffer('init_betas_kid', init_betas_kid)\n        self.register_buffer('init_cam', init_cam)\n        self.register_buffer('init_expression', init_expression)\n        \n    def detect_mhmr(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            scores = postprocess_score(self.mlp_classif(x)) # per token detection score.\n        return scores\n      \n    def forward(self, x, img_info, **kwargs):\n        if self.has_pose:\n            pose_token = x[-1][:, 0].clone()\n            n_humans_i = kwargs.get(\"n_humans\")\n            token = x[-1][:, 1:]\n\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n                if n_humans_i > 0:\n                    smpl_token = kwargs.get(\"smpl_token\") \n                    decoders = [self.decpose, self.decshape, self.deccam, self.decexpression]\n                    inits = [self.init_body_pose, self.init_betas, self.init_cam, self.init_expression]\n                    pred_smpl = [d(smpl_token) + i for d, i in zip(decoders, inits)]\n\n            token_cross = token.clone()\n            for blk in self.final_transform:\n                token_cross = blk(token_cross, pose_token, kwargs.get(\"pos\"))\n            x = x[:-1] + [token]\n            x_cross = x[:-1] + [token_cross]\n\n        with torch.cuda.amp.autocast(enabled=False):\n            self_out = checkpoint(\n                self.dpt_self,\n                x,\n                image_size=(img_info[0], img_info[1]),\n                use_reentrant=False,\n            )\n\n            final_output = postprocess(self_out, self.depth_mode, self.conf_mode)\n            final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n            final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n            if self.has_rgb:\n                rgb_out = checkpoint(\n                    self.dpt_rgb,\n                    x,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                rgb_output = postprocess_rgb(rgb_out)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n                cross_out = checkpoint(\n                    self.dpt_cross,\n                    x_cross,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                tmp = postprocess(cross_out, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n            \n            if n_humans_i > 0:\n                smpl_out = postprocess_smpl(pred_smpl, self.depth_mode, naive_mode=True)\n                final_output.update(smpl_out)\n\n        return final_output","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.__init__","uri":"program://Human3R/function/src.dust3r.heads.dpt_head.__init__#L486-L582","kind":"function","name":"__init__","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":486,"end_line":582,"context_start_line":466,"context_end_line":602,"code":"                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n                cross_out = checkpoint(\n                    self.dpt_cross,\n                    x_cross,\n                    image_size=(img_info[0], img_info[1]),\n                    use_reentrant=False,\n                )\n                tmp = postprocess(cross_out, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n            \n            if n_humans_i > 0:\n                smpl_out = postprocess_smpl(pred_smpl, self.depth_mode)\n                final_output.update(smpl_out)\n\n        return final_output\n\n\nclass NaiveDPTPts3dPoseSMPL(nn.Module):\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super(NaiveDPTPts3dPoseSMPL, self).__init__()\n        self.return_all_layers = True  # backbone needs to return all layers\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        pts_channels = 3 + has_conf\n        rgb_channels = has_rgb * 3\n        feature_dim = 256\n        last_dim = feature_dim // 2\n        ed = net.enc_embed_dim\n        dd = net.dec_embed_dim\n        hooks_idx = [0, 1, 2, 3]\n        dim_tokens = [ed, dd, dd, dd]\n        head_type = \"regression\"\n        output_width_ratio = 1\n\n        pts_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=pts_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        rgb_dpt_args = dict(\n            output_width_ratio=output_width_ratio,\n            num_channels=rgb_channels,\n            feature_dim=feature_dim,\n            last_dim=last_dim,\n            dim_tokens=dim_tokens,\n            hooks_idx=hooks_idx,\n            head_type=head_type,\n        )\n        if hooks_idx is not None:\n            pts_dpt_args.update(hooks=hooks_idx)\n            rgb_dpt_args.update(hooks=hooks_idx)\n\n        self.dpt_self = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_self.init(**dpt_init_args)\n\n        self.final_transform = nn.ModuleList(\n            [\n                ConditionModulationBlock(\n                    net.dec_embed_dim,\n                    net.dec_num_heads,\n                    mlp_ratio=4.0,\n                    qkv_bias=True,\n                    rope=net.rope,\n                )\n                for _ in range(2)\n            ]\n        )\n\n        self.dpt_cross = DPTOutputAdapter_fix(**pts_dpt_args)\n        dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n        self.dpt_cross.init(**dpt_init_args)\n\n        if has_rgb:\n            self.dpt_rgb = DPTOutputAdapter_fix(**rgb_dpt_args)\n            dpt_init_args = {} if dim_tokens is None else {\"dim_tokens_enc\": dim_tokens}\n            self.dpt_rgb.init(**dpt_init_args)\n\n        if has_pose:\n            in_dim = net.dec_embed_dim\n            self.pose_head = PoseDecoder(hidden_size=in_dim)\n\n        # MHMR Heads - Detection\n        backbone_dim = net.backbone_dim\n        self.bb_patch_size = net.bb_patch_size\n        self.mlp_classif = regression_mlp([backbone_dim, backbone_dim, 1]) # bg or human\n        self.mlp_offset = regression_mlp([backbone_dim, backbone_dim, 2]) # offset\n\n        # SMPL\n        self.joint_rep_type, self.joint_rep_dim = '6d', 6\n        self.nrot = 53\n        self.num_body_joints = self.nrot - 1\n\n        npose = self.joint_rep_dim * (self.num_body_joints + 1)\n        self.npose = npose\n\n        self.input_is_mean_shape = True\n        self.num_betas = 10\n        assert self.num_betas in [10, 11]\n\n        # MHMR Heads - SMPL\n        self.decpose, self.decshape, self.deccam, self.decexpression = [\n            nn.Linear(1024, od) for od in [self.npose, self.num_betas, 3, 10]] # MLP(1024, x)\n\n        self.set_smpl_init()\n\n    def set_smpl_init(self):\n        \"\"\" Fetch saved SMPL parameters and register buffers.\"\"\"\n        mean_params = np.load(MEAN_PARAMS)\n        if self.nrot == 53:\n            init_body_pose = torch.eye(3).reshape(1,3,3).repeat(self.nrot,1,1)[:,:,:2].flatten(1).reshape(1, -1)\n            init_body_pose[:,:24*6] = torch.from_numpy(mean_params['pose'][:]).float() # global_orient+body_pose from SMPL\n        else:\n            init_body_pose = torch.from_numpy(mean_params['pose'].astype(np.float32)).unsqueeze(0)\n\n        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)\n        init_betas_kid = torch.cat([init_betas, torch.zeros_like(init_betas[:,[0]])],1)\n        init_expression = 0. * torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n\n        if self.num_betas == 11:\n            init_betas = torch.cat([init_betas, torch.zeros_like(init_betas[:,:1])], 1)\n\n        self.register_buffer('init_body_pose', init_body_pose)\n        self.register_buffer('init_betas', init_betas)","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.set_smpl_init","uri":"program://Human3R/function/src.dust3r.heads.dpt_head.set_smpl_init#L584-L605","kind":"function","name":"set_smpl_init","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":584,"end_line":605,"context_start_line":564,"context_end_line":625,"code":"        self.mlp_offset = regression_mlp([backbone_dim, backbone_dim, 2]) # offset\n\n        # SMPL\n        self.joint_rep_type, self.joint_rep_dim = '6d', 6\n        self.nrot = 53\n        self.num_body_joints = self.nrot - 1\n\n        npose = self.joint_rep_dim * (self.num_body_joints + 1)\n        self.npose = npose\n\n        self.input_is_mean_shape = True\n        self.num_betas = 10\n        assert self.num_betas in [10, 11]\n\n        # MHMR Heads - SMPL\n        self.decpose, self.decshape, self.deccam, self.decexpression = [\n            nn.Linear(1024, od) for od in [self.npose, self.num_betas, 3, 10]] # MLP(1024, x)\n\n        self.set_smpl_init()\n\n    def set_smpl_init(self):\n        \"\"\" Fetch saved SMPL parameters and register buffers.\"\"\"\n        mean_params = np.load(MEAN_PARAMS)\n        if self.nrot == 53:\n            init_body_pose = torch.eye(3).reshape(1,3,3).repeat(self.nrot,1,1)[:,:,:2].flatten(1).reshape(1, -1)\n            init_body_pose[:,:24*6] = torch.from_numpy(mean_params['pose'][:]).float() # global_orient+body_pose from SMPL\n        else:\n            init_body_pose = torch.from_numpy(mean_params['pose'].astype(np.float32)).unsqueeze(0)\n\n        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)\n        init_betas_kid = torch.cat([init_betas, torch.zeros_like(init_betas[:,[0]])],1)\n        init_expression = 0. * torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n\n        if self.num_betas == 11:\n            init_betas = torch.cat([init_betas, torch.zeros_like(init_betas[:,:1])], 1)\n\n        self.register_buffer('init_body_pose', init_body_pose)\n        self.register_buffer('init_betas', init_betas)\n        self.register_buffer('init_betas_kid', init_betas_kid)\n        self.register_buffer('init_cam', init_cam)\n        self.register_buffer('init_expression', init_expression)\n        \n    def detect_mhmr(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            scores = postprocess_score(self.mlp_classif(x)) # per token detection score.\n        return scores\n      \n    def forward(self, x, img_info, **kwargs):\n        if self.has_pose:\n            pose_token = x[-1][:, 0].clone()\n            n_humans_i = kwargs.get(\"n_humans\")\n            token = x[-1][:, 1:]\n\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n                if n_humans_i > 0:\n                    smpl_token = kwargs.get(\"smpl_token\") \n                    decoders = [self.decpose, self.decshape, self.deccam, self.decexpression]\n                    inits = [self.init_body_pose, self.init_betas, self.init_cam, self.init_expression]\n                    pred_smpl = [d(smpl_token) + i for d, i in zip(decoders, inits)]\n","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.detect_mhmr","uri":"program://Human3R/function/src.dust3r.heads.dpt_head.detect_mhmr#L607-L610","kind":"function","name":"detect_mhmr","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":607,"end_line":610,"context_start_line":587,"context_end_line":630,"code":"        if self.nrot == 53:\n            init_body_pose = torch.eye(3).reshape(1,3,3).repeat(self.nrot,1,1)[:,:,:2].flatten(1).reshape(1, -1)\n            init_body_pose[:,:24*6] = torch.from_numpy(mean_params['pose'][:]).float() # global_orient+body_pose from SMPL\n        else:\n            init_body_pose = torch.from_numpy(mean_params['pose'].astype(np.float32)).unsqueeze(0)\n\n        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)\n        init_betas_kid = torch.cat([init_betas, torch.zeros_like(init_betas[:,[0]])],1)\n        init_expression = 0. * torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n\n        if self.num_betas == 11:\n            init_betas = torch.cat([init_betas, torch.zeros_like(init_betas[:,:1])], 1)\n\n        self.register_buffer('init_body_pose', init_body_pose)\n        self.register_buffer('init_betas', init_betas)\n        self.register_buffer('init_betas_kid', init_betas_kid)\n        self.register_buffer('init_cam', init_cam)\n        self.register_buffer('init_expression', init_expression)\n        \n    def detect_mhmr(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            scores = postprocess_score(self.mlp_classif(x)) # per token detection score.\n        return scores\n      \n    def forward(self, x, img_info, **kwargs):\n        if self.has_pose:\n            pose_token = x[-1][:, 0].clone()\n            n_humans_i = kwargs.get(\"n_humans\")\n            token = x[-1][:, 1:]\n\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n                if n_humans_i > 0:\n                    smpl_token = kwargs.get(\"smpl_token\") \n                    decoders = [self.decpose, self.decshape, self.deccam, self.decexpression]\n                    inits = [self.init_body_pose, self.init_betas, self.init_cam, self.init_expression]\n                    pred_smpl = [d(smpl_token) + i for d, i in zip(decoders, inits)]\n\n            token_cross = token.clone()\n            for blk in self.final_transform:\n                token_cross = blk(token_cross, pose_token, kwargs.get(\"pos\"))\n            x = x[:-1] + [token]\n            x_cross = x[:-1] + [token_cross]","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.dpt_head.segment","uri":"program://Human3R/function/src.dust3r.heads.dpt_head.segment#L415-L418","kind":"function","name":"segment","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":415,"end_line":418,"context_start_line":395,"context_end_line":438,"code":"\n        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)\n        init_betas_kid = torch.cat([init_betas, torch.zeros_like(init_betas[:,[0]])],1)\n        init_expression = 0. * torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)\n\n        if self.num_betas == 11:\n            init_betas = torch.cat([init_betas, torch.zeros_like(init_betas[:,:1])], 1)\n\n        self.register_buffer('init_body_pose', init_body_pose)\n        self.register_buffer('init_betas', init_betas)\n        self.register_buffer('init_betas_kid', init_betas_kid)\n        self.register_buffer('init_cam', init_cam)\n        self.register_buffer('init_expression', init_expression)\n        \n    def detect_mhmr(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            scores = postprocess_score(self.mlp_classif(x)) # per token detection score.\n        return scores\n\n    def segment(self, x):\n        with torch.cuda.amp.autocast(enabled=False):\n            msks = postprocess_score(self.mlp_msk(x))\n        return msks\n    \n    def forward(self, x, img_info, **kwargs):\n        if self.has_pose:\n            pose_token = x[-1][:, 0].clone()\n            n_humans_i = kwargs.get(\"n_humans\")\n            token = x[-1][:, 1:]\n\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n                if n_humans_i > 0:\n                    smpl_token = kwargs.get(\"smpl_token\")   # CUT3R smpl token (bs, 10, 768)\n                    pred_body_pose = self.decpose(smpl_token) + self.init_body_pose\n                    pred_betas = self.decshape(smpl_token) + self.init_betas\n                    pred_cam = self.deccam(smpl_token[..., :self.in_dim])\n                    pred_expression = self.decexpression(smpl_token) + self.init_expression\n                    pred_smpl = [pred_body_pose, pred_betas, pred_cam, pred_expression]\n                    \n            token_cross = token.clone()\n            for blk in self.final_transform:\n                token_cross = blk(token_cross, pose_token, kwargs.get(\"pos\"))","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess","uri":"program://Human3R/module/src.dust3r.heads.postprocess#L1-L210","kind":"module","name":"src.dust3r.heads.postprocess","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":1,"end_line":210,"context_start_line":1,"context_end_line":210,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn.functional as F\nimport roma\n\ndef postprocess(out, depth_mode, conf_mode, pos_z=False):\n    \"\"\"\n    extract 3D points/confidence from prediction head output\n    \"\"\"\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode, pos_z=pos_z))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n    return res\n\n\ndef postprocess_rgb(out, eps=1e-6):\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = torch.sigmoid(fmap) * (1 - 2 * eps) + eps\n    res = (res - 0.5) * 2\n    return dict(rgb=res)\n\n\ndef postprocess_score(scores, eps=1e-6):\n    out = torch.sigmoid(scores) * (1 - 2 * eps) + eps\n    return out\n\n\ndef postprocess_pose(out, mode, inverse=False):\n    \"\"\"\n    extract pose from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n\n    no_bounds = (vmin == -float(\"inf\")) and (vmax == float(\"inf\"))\n    assert no_bounds\n    trans = out[..., 0:3]\n    quats = out[..., 3:7]\n\n    if mode == \"linear\":\n        if no_bounds:\n            return trans  # [-inf, +inf]\n        return trans.clip(min=vmin, max=vmax)\n\n    d = trans.norm(dim=-1, keepdim=True)\n\n    if mode == \"square\":\n        if inverse:\n            scale = d / d.square().clip(min=1e-8)\n        else:\n            scale = d.square() / d.clip(min=1e-8)\n\n    if mode == \"exp\":\n        if inverse:\n            scale = d / torch.expm1(d).clip(min=1e-8)\n        else:\n            scale = torch.expm1(d) / d.clip(min=1e-8)\n\n    trans = trans * scale\n    quats = standardize_quaternion(quats)\n\n    return torch.cat([trans, quats], dim=-1)\n\n\ndef postprocess_smpl(out, mode, pos_z=False, naive_mode=False):\n    \"\"\"\n    extract SMPL parameters from prediction head output\n    \"\"\"\n    pred_body_pose, pred_betas, pred_cam, pred_expression = out\n    bs, max_humans = pred_body_pose.shape[:2]\n\n    res = dict(smpl_shape=pred_betas)\n    if naive_mode:\n        res['smpl_transl'] = pred_cam # MHMR\n    else:\n        res['smpl_transl'] = reg_dense_depth(pred_cam, mode=mode, pos_z=pos_z)\n\n    res['smpl_rotmat'] = rot6d_to_rotmat(\n        pred_body_pose.view(-1, pred_body_pose.shape[-1]), \n        naive_mode=naive_mode,\n        ).view(bs, max_humans, -1, 3, 3)\n    res['smpl_expression'] = pred_expression\n\n    return res\n\n\ndef postprocess_pose_conf(out):\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,1\n    return dict(pose_conf=torch.sigmoid(fmap))\n\n\ndef postprocess_desc(out, depth_mode, conf_mode, desc_dim, double_channel=False):\n    \"\"\"\n    extract 3D points/confidence from prediction head output\n    \"\"\"\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n\n    if double_channel:\n        res[\"pts3d_self\"] = reg_dense_depth(\n            fmap[\n                :, :, :, 3 + int(conf_mode is not None) : 6 + int(conf_mode is not None)\n            ],\n            mode=depth_mode,\n        )\n        if conf_mode is not None:\n            res[\"conf_self\"] = reg_dense_conf(\n                fmap[:, :, :, 6 + int(conf_mode is not None)], mode=conf_mode\n            )\n\n    start = (\n        3\n        + int(conf_mode is not None)\n        + int(double_channel) * (3 + int(conf_mode is not None))\n    )\n    res[\"desc\"] = reg_desc(fmap[:, :, :, start : start + desc_dim], mode=\"norm\")\n    res[\"desc_conf\"] = reg_dense_conf(fmap[:, :, :, start + desc_dim], mode=conf_mode)\n    assert start + desc_dim + 1 == fmap.shape[-1]\n\n    return res\n\n\ndef reg_desc(desc, mode=\"norm\"):\n    if \"norm\" in mode:\n        desc = desc / desc.norm(dim=-1, keepdim=True)\n    else:\n        raise ValueError(f\"Unknown desc mode {mode}\")\n    return desc\n\n\ndef reg_dense_depth(xyz, mode, pos_z=False):\n    \"\"\"\n    extract 3D points from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n\n    no_bounds = (vmin == -float(\"inf\")) and (vmax == float(\"inf\"))\n    assert no_bounds\n\n    if mode == \"linear\":\n        if no_bounds:\n            return xyz  # [-inf, +inf]\n        return xyz.clip(min=vmin, max=vmax)\n\n    if pos_z:\n        sign = torch.sign(xyz[..., -1:])\n        xyz *= sign\n    d = xyz.norm(dim=-1, keepdim=True)\n    xyz = xyz / d.clip(min=1e-8)\n\n    if mode == \"square\":\n        return xyz * d.square()\n\n    if mode == \"exp\":\n        return xyz * torch.expm1(d)\n\n    raise ValueError(f\"bad {mode=}\")\n\n\ndef reg_dense_conf(x, mode):\n    \"\"\"\n    extract confidence from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n    if mode == \"exp\":\n        return vmin + x.exp().clip(max=vmax - vmin)\n    if mode == \"sigmoid\":\n        return (vmax - vmin) * torch.sigmoid(x) + vmin\n    raise ValueError(f\"bad {mode=}\")\n\n\ndef standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert a unit quaternion to a standard form: one in which the real\n    part is non negative.\n\n    Args:\n        quaternions: Quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Standardized quaternions as tensor of shape (..., 4).\n    \"\"\"\n    quaternions = F.normalize(quaternions, p=2, dim=-1)\n    return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions)\n\n\ndef rot6d_to_rotmat(x, naive_mode=False):\n    \"\"\"\n    6D rotation representation to 3x3 rotation matrix.\n    Args:\n        x: (nvh,num_joints*6) Batch of 6-D rotation representations.\n    Returns:\n        torch.Tensor: Batch of corresponding rotation matrices with shape (nvh*num_joints,3,3).\n    \"\"\"\n    if naive_mode:\n        x = x.reshape(-1,2,3).permute(0, 2, 1).contiguous() # (nvh, num_joints*6) -> (nvh*num_joints, 3, 2) # inherited from MHMR\n    else:\n        x = x.reshape(-1,3,2).contiguous() # (nvh, num_joints*6) -> (nvh*num_joints, 3, 2)\n    y = roma.special_gramschmidt(x, epsilon=1e-6) # (nvh*num_joints, 3, 3)\n    return y","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.postprocess","uri":"program://Human3R/function/src.dust3r.heads.postprocess.postprocess#L11-L20","kind":"function","name":"postprocess","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":11,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn.functional as F\nimport roma\n\ndef postprocess(out, depth_mode, conf_mode, pos_z=False):\n    \"\"\"\n    extract 3D points/confidence from prediction head output\n    \"\"\"\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode, pos_z=pos_z))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n    return res\n\n\ndef postprocess_rgb(out, eps=1e-6):\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = torch.sigmoid(fmap) * (1 - 2 * eps) + eps\n    res = (res - 0.5) * 2\n    return dict(rgb=res)\n\n\ndef postprocess_score(scores, eps=1e-6):\n    out = torch.sigmoid(scores) * (1 - 2 * eps) + eps\n    return out\n\n\ndef postprocess_pose(out, mode, inverse=False):\n    \"\"\"\n    extract pose from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.postprocess_rgb","uri":"program://Human3R/function/src.dust3r.heads.postprocess.postprocess_rgb#L23-L27","kind":"function","name":"postprocess_rgb","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":23,"end_line":27,"context_start_line":3,"context_end_line":47,"code":"#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn.functional as F\nimport roma\n\ndef postprocess(out, depth_mode, conf_mode, pos_z=False):\n    \"\"\"\n    extract 3D points/confidence from prediction head output\n    \"\"\"\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode, pos_z=pos_z))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n    return res\n\n\ndef postprocess_rgb(out, eps=1e-6):\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = torch.sigmoid(fmap) * (1 - 2 * eps) + eps\n    res = (res - 0.5) * 2\n    return dict(rgb=res)\n\n\ndef postprocess_score(scores, eps=1e-6):\n    out = torch.sigmoid(scores) * (1 - 2 * eps) + eps\n    return out\n\n\ndef postprocess_pose(out, mode, inverse=False):\n    \"\"\"\n    extract pose from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n\n    no_bounds = (vmin == -float(\"inf\")) and (vmax == float(\"inf\"))\n    assert no_bounds\n    trans = out[..., 0:3]\n    quats = out[..., 3:7]\n\n    if mode == \"linear\":\n        if no_bounds:","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.postprocess_score","uri":"program://Human3R/function/src.dust3r.heads.postprocess.postprocess_score#L30-L32","kind":"function","name":"postprocess_score","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":30,"end_line":32,"context_start_line":10,"context_end_line":52,"code":"\ndef postprocess(out, depth_mode, conf_mode, pos_z=False):\n    \"\"\"\n    extract 3D points/confidence from prediction head output\n    \"\"\"\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode, pos_z=pos_z))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n    return res\n\n\ndef postprocess_rgb(out, eps=1e-6):\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = torch.sigmoid(fmap) * (1 - 2 * eps) + eps\n    res = (res - 0.5) * 2\n    return dict(rgb=res)\n\n\ndef postprocess_score(scores, eps=1e-6):\n    out = torch.sigmoid(scores) * (1 - 2 * eps) + eps\n    return out\n\n\ndef postprocess_pose(out, mode, inverse=False):\n    \"\"\"\n    extract pose from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n\n    no_bounds = (vmin == -float(\"inf\")) and (vmax == float(\"inf\"))\n    assert no_bounds\n    trans = out[..., 0:3]\n    quats = out[..., 3:7]\n\n    if mode == \"linear\":\n        if no_bounds:\n            return trans  # [-inf, +inf]\n        return trans.clip(min=vmin, max=vmax)\n\n    d = trans.norm(dim=-1, keepdim=True)\n","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.postprocess_pose","uri":"program://Human3R/function/src.dust3r.heads.postprocess.postprocess_pose#L35-L68","kind":"function","name":"postprocess_pose","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":35,"end_line":68,"context_start_line":15,"context_end_line":88,"code":"    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode, pos_z=pos_z))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n    return res\n\n\ndef postprocess_rgb(out, eps=1e-6):\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = torch.sigmoid(fmap) * (1 - 2 * eps) + eps\n    res = (res - 0.5) * 2\n    return dict(rgb=res)\n\n\ndef postprocess_score(scores, eps=1e-6):\n    out = torch.sigmoid(scores) * (1 - 2 * eps) + eps\n    return out\n\n\ndef postprocess_pose(out, mode, inverse=False):\n    \"\"\"\n    extract pose from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n\n    no_bounds = (vmin == -float(\"inf\")) and (vmax == float(\"inf\"))\n    assert no_bounds\n    trans = out[..., 0:3]\n    quats = out[..., 3:7]\n\n    if mode == \"linear\":\n        if no_bounds:\n            return trans  # [-inf, +inf]\n        return trans.clip(min=vmin, max=vmax)\n\n    d = trans.norm(dim=-1, keepdim=True)\n\n    if mode == \"square\":\n        if inverse:\n            scale = d / d.square().clip(min=1e-8)\n        else:\n            scale = d.square() / d.clip(min=1e-8)\n\n    if mode == \"exp\":\n        if inverse:\n            scale = d / torch.expm1(d).clip(min=1e-8)\n        else:\n            scale = torch.expm1(d) / d.clip(min=1e-8)\n\n    trans = trans * scale\n    quats = standardize_quaternion(quats)\n\n    return torch.cat([trans, quats], dim=-1)\n\n\ndef postprocess_smpl(out, mode, pos_z=False, naive_mode=False):\n    \"\"\"\n    extract SMPL parameters from prediction head output\n    \"\"\"\n    pred_body_pose, pred_betas, pred_cam, pred_expression = out\n    bs, max_humans = pred_body_pose.shape[:2]\n\n    res = dict(smpl_shape=pred_betas)\n    if naive_mode:\n        res['smpl_transl'] = pred_cam # MHMR\n    else:\n        res['smpl_transl'] = reg_dense_depth(pred_cam, mode=mode, pos_z=pos_z)\n\n    res['smpl_rotmat'] = rot6d_to_rotmat(\n        pred_body_pose.view(-1, pred_body_pose.shape[-1]), \n        naive_mode=naive_mode,\n        ).view(bs, max_humans, -1, 3, 3)\n    res['smpl_expression'] = pred_expression","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.postprocess_smpl","uri":"program://Human3R/function/src.dust3r.heads.postprocess.postprocess_smpl#L71-L90","kind":"function","name":"postprocess_smpl","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":71,"end_line":90,"context_start_line":51,"context_end_line":110,"code":"    d = trans.norm(dim=-1, keepdim=True)\n\n    if mode == \"square\":\n        if inverse:\n            scale = d / d.square().clip(min=1e-8)\n        else:\n            scale = d.square() / d.clip(min=1e-8)\n\n    if mode == \"exp\":\n        if inverse:\n            scale = d / torch.expm1(d).clip(min=1e-8)\n        else:\n            scale = torch.expm1(d) / d.clip(min=1e-8)\n\n    trans = trans * scale\n    quats = standardize_quaternion(quats)\n\n    return torch.cat([trans, quats], dim=-1)\n\n\ndef postprocess_smpl(out, mode, pos_z=False, naive_mode=False):\n    \"\"\"\n    extract SMPL parameters from prediction head output\n    \"\"\"\n    pred_body_pose, pred_betas, pred_cam, pred_expression = out\n    bs, max_humans = pred_body_pose.shape[:2]\n\n    res = dict(smpl_shape=pred_betas)\n    if naive_mode:\n        res['smpl_transl'] = pred_cam # MHMR\n    else:\n        res['smpl_transl'] = reg_dense_depth(pred_cam, mode=mode, pos_z=pos_z)\n\n    res['smpl_rotmat'] = rot6d_to_rotmat(\n        pred_body_pose.view(-1, pred_body_pose.shape[-1]), \n        naive_mode=naive_mode,\n        ).view(bs, max_humans, -1, 3, 3)\n    res['smpl_expression'] = pred_expression\n\n    return res\n\n\ndef postprocess_pose_conf(out):\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,1\n    return dict(pose_conf=torch.sigmoid(fmap))\n\n\ndef postprocess_desc(out, depth_mode, conf_mode, desc_dim, double_channel=False):\n    \"\"\"\n    extract 3D points/confidence from prediction head output\n    \"\"\"\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n\n    if double_channel:\n        res[\"pts3d_self\"] = reg_dense_depth(\n            fmap[","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.postprocess_pose_conf","uri":"program://Human3R/function/src.dust3r.heads.postprocess.postprocess_pose_conf#L93-L95","kind":"function","name":"postprocess_pose_conf","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":93,"end_line":95,"context_start_line":73,"context_end_line":115,"code":"    extract SMPL parameters from prediction head output\n    \"\"\"\n    pred_body_pose, pred_betas, pred_cam, pred_expression = out\n    bs, max_humans = pred_body_pose.shape[:2]\n\n    res = dict(smpl_shape=pred_betas)\n    if naive_mode:\n        res['smpl_transl'] = pred_cam # MHMR\n    else:\n        res['smpl_transl'] = reg_dense_depth(pred_cam, mode=mode, pos_z=pos_z)\n\n    res['smpl_rotmat'] = rot6d_to_rotmat(\n        pred_body_pose.view(-1, pred_body_pose.shape[-1]), \n        naive_mode=naive_mode,\n        ).view(bs, max_humans, -1, 3, 3)\n    res['smpl_expression'] = pred_expression\n\n    return res\n\n\ndef postprocess_pose_conf(out):\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,1\n    return dict(pose_conf=torch.sigmoid(fmap))\n\n\ndef postprocess_desc(out, depth_mode, conf_mode, desc_dim, double_channel=False):\n    \"\"\"\n    extract 3D points/confidence from prediction head output\n    \"\"\"\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n\n    if double_channel:\n        res[\"pts3d_self\"] = reg_dense_depth(\n            fmap[\n                :, :, :, 3 + int(conf_mode is not None) : 6 + int(conf_mode is not None)\n            ],\n            mode=depth_mode,\n        )\n        if conf_mode is not None:","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.postprocess_desc","uri":"program://Human3R/function/src.dust3r.heads.postprocess.postprocess_desc#L98-L129","kind":"function","name":"postprocess_desc","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":98,"end_line":129,"context_start_line":78,"context_end_line":149,"code":"    res = dict(smpl_shape=pred_betas)\n    if naive_mode:\n        res['smpl_transl'] = pred_cam # MHMR\n    else:\n        res['smpl_transl'] = reg_dense_depth(pred_cam, mode=mode, pos_z=pos_z)\n\n    res['smpl_rotmat'] = rot6d_to_rotmat(\n        pred_body_pose.view(-1, pred_body_pose.shape[-1]), \n        naive_mode=naive_mode,\n        ).view(bs, max_humans, -1, 3, 3)\n    res['smpl_expression'] = pred_expression\n\n    return res\n\n\ndef postprocess_pose_conf(out):\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,1\n    return dict(pose_conf=torch.sigmoid(fmap))\n\n\ndef postprocess_desc(out, depth_mode, conf_mode, desc_dim, double_channel=False):\n    \"\"\"\n    extract 3D points/confidence from prediction head output\n    \"\"\"\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n\n    if double_channel:\n        res[\"pts3d_self\"] = reg_dense_depth(\n            fmap[\n                :, :, :, 3 + int(conf_mode is not None) : 6 + int(conf_mode is not None)\n            ],\n            mode=depth_mode,\n        )\n        if conf_mode is not None:\n            res[\"conf_self\"] = reg_dense_conf(\n                fmap[:, :, :, 6 + int(conf_mode is not None)], mode=conf_mode\n            )\n\n    start = (\n        3\n        + int(conf_mode is not None)\n        + int(double_channel) * (3 + int(conf_mode is not None))\n    )\n    res[\"desc\"] = reg_desc(fmap[:, :, :, start : start + desc_dim], mode=\"norm\")\n    res[\"desc_conf\"] = reg_dense_conf(fmap[:, :, :, start + desc_dim], mode=conf_mode)\n    assert start + desc_dim + 1 == fmap.shape[-1]\n\n    return res\n\n\ndef reg_desc(desc, mode=\"norm\"):\n    if \"norm\" in mode:\n        desc = desc / desc.norm(dim=-1, keepdim=True)\n    else:\n        raise ValueError(f\"Unknown desc mode {mode}\")\n    return desc\n\n\ndef reg_dense_depth(xyz, mode, pos_z=False):\n    \"\"\"\n    extract 3D points from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n\n    no_bounds = (vmin == -float(\"inf\")) and (vmax == float(\"inf\"))\n    assert no_bounds\n\n    if mode == \"linear\":","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.reg_desc","uri":"program://Human3R/function/src.dust3r.heads.postprocess.reg_desc#L132-L137","kind":"function","name":"reg_desc","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":132,"end_line":137,"context_start_line":112,"context_end_line":157,"code":"            ],\n            mode=depth_mode,\n        )\n        if conf_mode is not None:\n            res[\"conf_self\"] = reg_dense_conf(\n                fmap[:, :, :, 6 + int(conf_mode is not None)], mode=conf_mode\n            )\n\n    start = (\n        3\n        + int(conf_mode is not None)\n        + int(double_channel) * (3 + int(conf_mode is not None))\n    )\n    res[\"desc\"] = reg_desc(fmap[:, :, :, start : start + desc_dim], mode=\"norm\")\n    res[\"desc_conf\"] = reg_dense_conf(fmap[:, :, :, start + desc_dim], mode=conf_mode)\n    assert start + desc_dim + 1 == fmap.shape[-1]\n\n    return res\n\n\ndef reg_desc(desc, mode=\"norm\"):\n    if \"norm\" in mode:\n        desc = desc / desc.norm(dim=-1, keepdim=True)\n    else:\n        raise ValueError(f\"Unknown desc mode {mode}\")\n    return desc\n\n\ndef reg_dense_depth(xyz, mode, pos_z=False):\n    \"\"\"\n    extract 3D points from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n\n    no_bounds = (vmin == -float(\"inf\")) and (vmax == float(\"inf\"))\n    assert no_bounds\n\n    if mode == \"linear\":\n        if no_bounds:\n            return xyz  # [-inf, +inf]\n        return xyz.clip(min=vmin, max=vmax)\n\n    if pos_z:\n        sign = torch.sign(xyz[..., -1:])\n        xyz *= sign\n    d = xyz.norm(dim=-1, keepdim=True)","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.reg_dense_depth","uri":"program://Human3R/function/src.dust3r.heads.postprocess.reg_dense_depth#L140-L166","kind":"function","name":"reg_dense_depth","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":140,"end_line":166,"context_start_line":120,"context_end_line":186,"code":"    start = (\n        3\n        + int(conf_mode is not None)\n        + int(double_channel) * (3 + int(conf_mode is not None))\n    )\n    res[\"desc\"] = reg_desc(fmap[:, :, :, start : start + desc_dim], mode=\"norm\")\n    res[\"desc_conf\"] = reg_dense_conf(fmap[:, :, :, start + desc_dim], mode=conf_mode)\n    assert start + desc_dim + 1 == fmap.shape[-1]\n\n    return res\n\n\ndef reg_desc(desc, mode=\"norm\"):\n    if \"norm\" in mode:\n        desc = desc / desc.norm(dim=-1, keepdim=True)\n    else:\n        raise ValueError(f\"Unknown desc mode {mode}\")\n    return desc\n\n\ndef reg_dense_depth(xyz, mode, pos_z=False):\n    \"\"\"\n    extract 3D points from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n\n    no_bounds = (vmin == -float(\"inf\")) and (vmax == float(\"inf\"))\n    assert no_bounds\n\n    if mode == \"linear\":\n        if no_bounds:\n            return xyz  # [-inf, +inf]\n        return xyz.clip(min=vmin, max=vmax)\n\n    if pos_z:\n        sign = torch.sign(xyz[..., -1:])\n        xyz *= sign\n    d = xyz.norm(dim=-1, keepdim=True)\n    xyz = xyz / d.clip(min=1e-8)\n\n    if mode == \"square\":\n        return xyz * d.square()\n\n    if mode == \"exp\":\n        return xyz * torch.expm1(d)\n\n    raise ValueError(f\"bad {mode=}\")\n\n\ndef reg_dense_conf(x, mode):\n    \"\"\"\n    extract confidence from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n    if mode == \"exp\":\n        return vmin + x.exp().clip(max=vmax - vmin)\n    if mode == \"sigmoid\":\n        return (vmax - vmin) * torch.sigmoid(x) + vmin\n    raise ValueError(f\"bad {mode=}\")\n\n\ndef standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert a unit quaternion to a standard form: one in which the real\n    part is non negative.\n\n    Args:","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.reg_dense_conf","uri":"program://Human3R/function/src.dust3r.heads.postprocess.reg_dense_conf#L169-L178","kind":"function","name":"reg_dense_conf","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":169,"end_line":178,"context_start_line":149,"context_end_line":198,"code":"    if mode == \"linear\":\n        if no_bounds:\n            return xyz  # [-inf, +inf]\n        return xyz.clip(min=vmin, max=vmax)\n\n    if pos_z:\n        sign = torch.sign(xyz[..., -1:])\n        xyz *= sign\n    d = xyz.norm(dim=-1, keepdim=True)\n    xyz = xyz / d.clip(min=1e-8)\n\n    if mode == \"square\":\n        return xyz * d.square()\n\n    if mode == \"exp\":\n        return xyz * torch.expm1(d)\n\n    raise ValueError(f\"bad {mode=}\")\n\n\ndef reg_dense_conf(x, mode):\n    \"\"\"\n    extract confidence from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n    if mode == \"exp\":\n        return vmin + x.exp().clip(max=vmax - vmin)\n    if mode == \"sigmoid\":\n        return (vmax - vmin) * torch.sigmoid(x) + vmin\n    raise ValueError(f\"bad {mode=}\")\n\n\ndef standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert a unit quaternion to a standard form: one in which the real\n    part is non negative.\n\n    Args:\n        quaternions: Quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Standardized quaternions as tensor of shape (..., 4).\n    \"\"\"\n    quaternions = F.normalize(quaternions, p=2, dim=-1)\n    return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions)\n\n\ndef rot6d_to_rotmat(x, naive_mode=False):\n    \"\"\"","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.standardize_quaternion","uri":"program://Human3R/function/src.dust3r.heads.postprocess.standardize_quaternion#L181-L194","kind":"function","name":"standardize_quaternion","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":181,"end_line":194,"context_start_line":161,"context_end_line":210,"code":"        return xyz * d.square()\n\n    if mode == \"exp\":\n        return xyz * torch.expm1(d)\n\n    raise ValueError(f\"bad {mode=}\")\n\n\ndef reg_dense_conf(x, mode):\n    \"\"\"\n    extract confidence from prediction head output\n    \"\"\"\n    mode, vmin, vmax = mode\n    if mode == \"exp\":\n        return vmin + x.exp().clip(max=vmax - vmin)\n    if mode == \"sigmoid\":\n        return (vmax - vmin) * torch.sigmoid(x) + vmin\n    raise ValueError(f\"bad {mode=}\")\n\n\ndef standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert a unit quaternion to a standard form: one in which the real\n    part is non negative.\n\n    Args:\n        quaternions: Quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Standardized quaternions as tensor of shape (..., 4).\n    \"\"\"\n    quaternions = F.normalize(quaternions, p=2, dim=-1)\n    return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions)\n\n\ndef rot6d_to_rotmat(x, naive_mode=False):\n    \"\"\"\n    6D rotation representation to 3x3 rotation matrix.\n    Args:\n        x: (nvh,num_joints*6) Batch of 6-D rotation representations.\n    Returns:\n        torch.Tensor: Batch of corresponding rotation matrices with shape (nvh*num_joints,3,3).\n    \"\"\"\n    if naive_mode:\n        x = x.reshape(-1,2,3).permute(0, 2, 1).contiguous() # (nvh, num_joints*6) -> (nvh*num_joints, 3, 2) # inherited from MHMR\n    else:\n        x = x.reshape(-1,3,2).contiguous() # (nvh, num_joints*6) -> (nvh*num_joints, 3, 2)\n    y = roma.special_gramschmidt(x, epsilon=1e-6) # (nvh*num_joints, 3, 3)\n    return y","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.postprocess.rot6d_to_rotmat","uri":"program://Human3R/function/src.dust3r.heads.postprocess.rot6d_to_rotmat#L197-L210","kind":"function","name":"rot6d_to_rotmat","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":197,"end_line":210,"context_start_line":177,"context_end_line":210,"code":"        return (vmax - vmin) * torch.sigmoid(x) + vmin\n    raise ValueError(f\"bad {mode=}\")\n\n\ndef standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:\n    \"\"\"\n    Convert a unit quaternion to a standard form: one in which the real\n    part is non negative.\n\n    Args:\n        quaternions: Quaternions with real part first,\n            as tensor of shape (..., 4).\n\n    Returns:\n        Standardized quaternions as tensor of shape (..., 4).\n    \"\"\"\n    quaternions = F.normalize(quaternions, p=2, dim=-1)\n    return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions)\n\n\ndef rot6d_to_rotmat(x, naive_mode=False):\n    \"\"\"\n    6D rotation representation to 3x3 rotation matrix.\n    Args:\n        x: (nvh,num_joints*6) Batch of 6-D rotation representations.\n    Returns:\n        torch.Tensor: Batch of corresponding rotation matrices with shape (nvh*num_joints,3,3).\n    \"\"\"\n    if naive_mode:\n        x = x.reshape(-1,2,3).permute(0, 2, 1).contiguous() # (nvh, num_joints*6) -> (nvh*num_joints, 3, 2) # inherited from MHMR\n    else:\n        x = x.reshape(-1,3,2).contiguous() # (nvh, num_joints*6) -> (nvh*num_joints, 3, 2)\n    y = roma.special_gramschmidt(x, epsilon=1e-6) # (nvh*num_joints, 3, 3)\n    return y","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.linear_head","uri":"program://Human3R/module/src.dust3r.heads.linear_head#L1-L346","kind":"module","name":"src.dust3r.heads.linear_head","path":"src/dust3r/heads/linear_head.py","language":"python","start_line":1,"end_line":346,"context_start_line":1,"context_end_line":346,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom dust3r.heads.postprocess import (\n    postprocess,\n    postprocess_desc,\n    postprocess_rgb,\n    postprocess_pose_conf,\n    postprocess_pose,\n    reg_dense_conf,\n)\nimport dust3r.utils.path_to_croco  # noqa\nfrom models.blocks import Mlp  # noqa\nfrom dust3r.utils.geometry import geotrf\nfrom dust3r.utils.camera import pose_encoding_to_camera, PoseDecoder\nfrom dust3r.blocks import ConditionModulationBlock\n\n\nclass LinearPts3d(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(\n        self, net, has_conf=False, has_depth=False, has_rgb=False, has_pose_conf=False\n    ):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose_conf = has_pose_conf\n        self.has_depth = has_depth\n        self.proj = Mlp(\n            net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2\n        )\n        if has_depth:\n            self.self_proj = Mlp(\n                net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2\n            )\n        if has_rgb:\n            self.rgb_proj = Mlp(net.dec_embed_dim, out_features=3 * self.patch_size**2)\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape):\n        H, W = img_shape\n        tokens = decout[-1]\n        B, S, D = tokens.shape\n\n        feat = self.proj(tokens)  # B,S,D\n        feat = feat.transpose(-1, -2).view(\n            B, -1, H // self.patch_size, W // self.patch_size\n        )\n        feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n\n        final_output = postprocess(feat, self.depth_mode, self.conf_mode)\n        final_output[\"pts3d_in_other_view\"] = final_output.pop(\"pts3d\")\n\n        if self.has_depth:\n            self_feat = self.self_proj(tokens)  # B,S,D\n            self_feat = self_feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            self_feat = F.pixel_shuffle(self_feat, self.patch_size)  # B,3,H,W\n            self_3d_output = postprocess(self_feat, self.depth_mode, self.conf_mode)\n            self_3d_output[\"pts3d_in_self_view\"] = self_3d_output.pop(\"pts3d\")\n            self_3d_output[\"conf_self\"] = self_3d_output.pop(\"conf\")\n            final_output.update(self_3d_output)\n\n        if self.has_rgb:\n            rgb_feat = self.rgb_proj(tokens)\n            rgb_feat = rgb_feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size)  # B,3,H,W\n            rgb_output = postprocess_rgb(rgb_feat)\n            final_output.update(rgb_output)\n\n        if self.has_pose_conf:\n            pose_conf = self.pose_conf_proj(tokens)\n            pose_conf = pose_conf.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            pose_conf = F.pixel_shuffle(pose_conf, self.patch_size)\n            pose_conf_output = postprocess_pose_conf(pose_conf)\n            final_output.update(pose_conf_output)\n\n        return final_output\n\n\nclass LinearPts3d_Desc(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(\n        self,\n        net,\n        has_conf=False,\n        has_depth=False,\n        local_feat_dim=24,\n        hidden_dim_factor=4.0,\n    ):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.has_conf = has_conf\n        self.double_channel = has_depth\n        self.local_feat_dim = local_feat_dim\n\n        if not has_depth:\n            self.proj = nn.Linear(\n                net.dec_embed_dim, (3 + has_conf) * self.patch_size**2\n            )\n        else:\n            self.proj = nn.Linear(\n                net.dec_embed_dim, (3 + has_conf) * 2 * self.patch_size**2\n            )\n        idim = net.enc_embed_dim + net.dec_embed_dim\n        self.head_local_features = Mlp(\n            in_features=idim,\n            hidden_features=int(hidden_dim_factor * idim),\n            out_features=(self.local_feat_dim + 1) * self.patch_size**2,\n        )\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape):\n        H, W = img_shape\n        tokens = decout[-1]\n        B, S, D = tokens.shape\n\n        feat = self.proj(tokens)  # B,S,D\n        feat = feat.transpose(-1, -2).view(\n            B, -1, H // self.patch_size, W // self.patch_size\n        )\n        feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n\n        enc_output, dec_output = decout[0], decout[-1]\n        cat_output = torch.cat([enc_output, dec_output], dim=-1)\n        local_features = self.head_local_features(cat_output)  # B,S,D\n        local_features = local_features.transpose(-1, -2).view(\n            B, -1, H // self.patch_size, W // self.patch_size\n        )\n        local_features = F.pixel_shuffle(local_features, self.patch_size)  # B,d,H,W\n        feat = torch.cat([feat, local_features], dim=1)\n\n        return postprocess_desc(\n            feat,\n            self.depth_mode,\n            self.conf_mode,\n            self.local_feat_dim,\n            self.double_channel,\n        )\n\n\nclass LinearPts3dPoseDirect(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        self.proj = Mlp(\n            net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2\n        )\n        if has_rgb:\n            self.rgb_proj = Mlp(net.dec_embed_dim, out_features=3 * self.patch_size**2)\n        if has_pose:\n            self.pose_head = PoseDecoder(hidden_size=net.dec_embed_dim)\n        if has_conf:\n            self.cross_conf_proj = Mlp(\n                net.dec_embed_dim, out_features=self.patch_size**2\n            )\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape):\n        H, W = img_shape\n        tokens = decout[-1]\n        if self.has_pose:\n            pose_token = tokens[:, 0]\n            tokens = tokens[:, 1:]\n        B, S, D = tokens.shape\n\n        feat = self.proj(tokens)  # B,S,D\n        feat = feat.transpose(-1, -2).view(\n            B, -1, H // self.patch_size, W // self.patch_size\n        )\n        feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n        final_output = postprocess(feat, self.depth_mode, self.conf_mode)\n        final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n        final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n        if self.has_rgb:\n            rgb_feat = self.rgb_proj(tokens)\n            rgb_feat = rgb_feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size)  # B,3,H,W\n            rgb_output = postprocess_rgb(rgb_feat)\n            final_output.update(rgb_output)\n\n        if self.has_pose:\n            pose = self.pose_head(pose_token)\n            pose = postprocess_pose(pose, self.pose_mode)\n            final_output[\"camera_pose\"] = pose  # B,7\n            final_output[\"pts3d_in_other_view\"] = geotrf(\n                pose_encoding_to_camera(final_output[\"camera_pose\"]),\n                final_output[\"pts3d_in_self_view\"],\n            )\n\n        if self.has_conf:\n            cross_conf = self.cross_conf_proj(tokens)\n            cross_conf = cross_conf.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            cross_conf = F.pixel_shuffle(cross_conf, self.patch_size)[:, 0]\n            final_output[\"conf\"] = reg_dense_conf(cross_conf, mode=self.conf_mode)\n        return final_output\n\n\nclass LinearPts3dPose(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(\n        self, net, has_conf=False, has_rgb=False, has_pose=False, mlp_ratio=4.0\n    ):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        self.proj = Mlp(\n            net.dec_embed_dim,\n            hidden_features=int(mlp_ratio * net.dec_embed_dim),\n            out_features=(3 + has_conf) * self.patch_size**2,\n        )\n        if has_rgb:\n            self.rgb_proj = Mlp(\n                net.dec_embed_dim,\n                hidden_features=int(mlp_ratio * net.dec_embed_dim),\n                out_features=3 * self.patch_size**2,\n            )\n        if has_pose:\n            self.pose_head = PoseDecoder(hidden_size=net.dec_embed_dim)\n            self.final_transform = nn.ModuleList(\n                [\n                    ConditionModulationBlock(\n                        net.dec_embed_dim,\n                        net.dec_num_heads,\n                        mlp_ratio=4.0,\n                        qkv_bias=True,\n                        rope=net.rope,\n                    )\n                    for _ in range(2)\n                ]\n            )\n            self.cross_proj = Mlp(\n                net.dec_embed_dim,\n                hidden_features=int(mlp_ratio * net.dec_embed_dim),\n                out_features=(3 + has_conf) * self.patch_size**2,\n            )\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape, **kwargs):\n        H, W = img_shape\n        tokens = decout[-1]\n        if self.has_pose:\n            pose_token = tokens[:, 0]\n            tokens = tokens[:, 1:]\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n            cross_tokens = tokens\n            for blk in self.final_transform:\n                cross_tokens = blk(cross_tokens, pose_token, kwargs.get(\"pos\"))\n\n        with torch.cuda.amp.autocast(enabled=False):\n            B, S, D = tokens.shape\n\n            feat = self.proj(tokens)  # B,S,D\n            feat = feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n            final_output = postprocess(\n                feat, self.depth_mode, self.conf_mode, pos_z=True\n            )\n            final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n            final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n            if self.has_rgb:\n                rgb_feat = self.rgb_proj(tokens)\n                rgb_feat = rgb_feat.transpose(-1, -2).view(\n                    B, -1, H // self.patch_size, W // self.patch_size\n                )\n                rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size)  # B,3,H,W\n                rgb_output = postprocess_rgb(rgb_feat)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n\n                cross_feat = self.cross_proj(cross_tokens)  # B,S,D\n                cross_feat = cross_feat.transpose(-1, -2).view(\n                    B, -1, H // self.patch_size, W // self.patch_size\n                )\n                cross_feat = F.pixel_shuffle(cross_feat, self.patch_size)  # B,3,H,W\n                tmp = postprocess(cross_feat, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n\n            return final_output","source_hash":"c52a465d11c1b8ee9145a379e8801682fe1db5ac9f076e5a55de364ec9efa424","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.linear_head.LinearPts3d","uri":"program://Human3R/class/src.dust3r.heads.linear_head.LinearPts3d#L25-L98","kind":"class","name":"LinearPts3d","path":"src/dust3r/heads/linear_head.py","language":"python","start_line":25,"end_line":98,"context_start_line":5,"context_end_line":118,"code":"# modified from DUSt3R\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom dust3r.heads.postprocess import (\n    postprocess,\n    postprocess_desc,\n    postprocess_rgb,\n    postprocess_pose_conf,\n    postprocess_pose,\n    reg_dense_conf,\n)\nimport dust3r.utils.path_to_croco  # noqa\nfrom models.blocks import Mlp  # noqa\nfrom dust3r.utils.geometry import geotrf\nfrom dust3r.utils.camera import pose_encoding_to_camera, PoseDecoder\nfrom dust3r.blocks import ConditionModulationBlock\n\n\nclass LinearPts3d(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(\n        self, net, has_conf=False, has_depth=False, has_rgb=False, has_pose_conf=False\n    ):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose_conf = has_pose_conf\n        self.has_depth = has_depth\n        self.proj = Mlp(\n            net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2\n        )\n        if has_depth:\n            self.self_proj = Mlp(\n                net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2\n            )\n        if has_rgb:\n            self.rgb_proj = Mlp(net.dec_embed_dim, out_features=3 * self.patch_size**2)\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape):\n        H, W = img_shape\n        tokens = decout[-1]\n        B, S, D = tokens.shape\n\n        feat = self.proj(tokens)  # B,S,D\n        feat = feat.transpose(-1, -2).view(\n            B, -1, H // self.patch_size, W // self.patch_size\n        )\n        feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n\n        final_output = postprocess(feat, self.depth_mode, self.conf_mode)\n        final_output[\"pts3d_in_other_view\"] = final_output.pop(\"pts3d\")\n\n        if self.has_depth:\n            self_feat = self.self_proj(tokens)  # B,S,D\n            self_feat = self_feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            self_feat = F.pixel_shuffle(self_feat, self.patch_size)  # B,3,H,W\n            self_3d_output = postprocess(self_feat, self.depth_mode, self.conf_mode)\n            self_3d_output[\"pts3d_in_self_view\"] = self_3d_output.pop(\"pts3d\")\n            self_3d_output[\"conf_self\"] = self_3d_output.pop(\"conf\")\n            final_output.update(self_3d_output)\n\n        if self.has_rgb:\n            rgb_feat = self.rgb_proj(tokens)\n            rgb_feat = rgb_feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size)  # B,3,H,W\n            rgb_output = postprocess_rgb(rgb_feat)\n            final_output.update(rgb_output)\n\n        if self.has_pose_conf:\n            pose_conf = self.pose_conf_proj(tokens)\n            pose_conf = pose_conf.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            pose_conf = F.pixel_shuffle(pose_conf, self.patch_size)\n            pose_conf_output = postprocess_pose_conf(pose_conf)\n            final_output.update(pose_conf_output)\n\n        return final_output\n\n\nclass LinearPts3d_Desc(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(\n        self,\n        net,\n        has_conf=False,\n        has_depth=False,\n        local_feat_dim=24,\n        hidden_dim_factor=4.0,\n    ):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode","source_hash":"c52a465d11c1b8ee9145a379e8801682fe1db5ac9f076e5a55de364ec9efa424","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.linear_head.LinearPts3d_Desc","uri":"program://Human3R/class/src.dust3r.heads.linear_head.LinearPts3d_Desc#L101-L167","kind":"class","name":"LinearPts3d_Desc","path":"src/dust3r/heads/linear_head.py","language":"python","start_line":101,"end_line":167,"context_start_line":81,"context_end_line":187,"code":"            rgb_feat = self.rgb_proj(tokens)\n            rgb_feat = rgb_feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size)  # B,3,H,W\n            rgb_output = postprocess_rgb(rgb_feat)\n            final_output.update(rgb_output)\n\n        if self.has_pose_conf:\n            pose_conf = self.pose_conf_proj(tokens)\n            pose_conf = pose_conf.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            pose_conf = F.pixel_shuffle(pose_conf, self.patch_size)\n            pose_conf_output = postprocess_pose_conf(pose_conf)\n            final_output.update(pose_conf_output)\n\n        return final_output\n\n\nclass LinearPts3d_Desc(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(\n        self,\n        net,\n        has_conf=False,\n        has_depth=False,\n        local_feat_dim=24,\n        hidden_dim_factor=4.0,\n    ):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.has_conf = has_conf\n        self.double_channel = has_depth\n        self.local_feat_dim = local_feat_dim\n\n        if not has_depth:\n            self.proj = nn.Linear(\n                net.dec_embed_dim, (3 + has_conf) * self.patch_size**2\n            )\n        else:\n            self.proj = nn.Linear(\n                net.dec_embed_dim, (3 + has_conf) * 2 * self.patch_size**2\n            )\n        idim = net.enc_embed_dim + net.dec_embed_dim\n        self.head_local_features = Mlp(\n            in_features=idim,\n            hidden_features=int(hidden_dim_factor * idim),\n            out_features=(self.local_feat_dim + 1) * self.patch_size**2,\n        )\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape):\n        H, W = img_shape\n        tokens = decout[-1]\n        B, S, D = tokens.shape\n\n        feat = self.proj(tokens)  # B,S,D\n        feat = feat.transpose(-1, -2).view(\n            B, -1, H // self.patch_size, W // self.patch_size\n        )\n        feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n\n        enc_output, dec_output = decout[0], decout[-1]\n        cat_output = torch.cat([enc_output, dec_output], dim=-1)\n        local_features = self.head_local_features(cat_output)  # B,S,D\n        local_features = local_features.transpose(-1, -2).view(\n            B, -1, H // self.patch_size, W // self.patch_size\n        )\n        local_features = F.pixel_shuffle(local_features, self.patch_size)  # B,d,H,W\n        feat = torch.cat([feat, local_features], dim=1)\n\n        return postprocess_desc(\n            feat,\n            self.depth_mode,\n            self.conf_mode,\n            self.local_feat_dim,\n            self.double_channel,\n        )\n\n\nclass LinearPts3dPoseDirect(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        self.proj = Mlp(\n            net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2","source_hash":"c52a465d11c1b8ee9145a379e8801682fe1db5ac9f076e5a55de364ec9efa424","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.linear_head.LinearPts3dPoseDirect","uri":"program://Human3R/class/src.dust3r.heads.linear_head.LinearPts3dPoseDirect#L170-L243","kind":"class","name":"LinearPts3dPoseDirect","path":"src/dust3r/heads/linear_head.py","language":"python","start_line":170,"end_line":243,"context_start_line":150,"context_end_line":263,"code":"        feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n\n        enc_output, dec_output = decout[0], decout[-1]\n        cat_output = torch.cat([enc_output, dec_output], dim=-1)\n        local_features = self.head_local_features(cat_output)  # B,S,D\n        local_features = local_features.transpose(-1, -2).view(\n            B, -1, H // self.patch_size, W // self.patch_size\n        )\n        local_features = F.pixel_shuffle(local_features, self.patch_size)  # B,d,H,W\n        feat = torch.cat([feat, local_features], dim=1)\n\n        return postprocess_desc(\n            feat,\n            self.depth_mode,\n            self.conf_mode,\n            self.local_feat_dim,\n            self.double_channel,\n        )\n\n\nclass LinearPts3dPoseDirect(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        self.proj = Mlp(\n            net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2\n        )\n        if has_rgb:\n            self.rgb_proj = Mlp(net.dec_embed_dim, out_features=3 * self.patch_size**2)\n        if has_pose:\n            self.pose_head = PoseDecoder(hidden_size=net.dec_embed_dim)\n        if has_conf:\n            self.cross_conf_proj = Mlp(\n                net.dec_embed_dim, out_features=self.patch_size**2\n            )\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape):\n        H, W = img_shape\n        tokens = decout[-1]\n        if self.has_pose:\n            pose_token = tokens[:, 0]\n            tokens = tokens[:, 1:]\n        B, S, D = tokens.shape\n\n        feat = self.proj(tokens)  # B,S,D\n        feat = feat.transpose(-1, -2).view(\n            B, -1, H // self.patch_size, W // self.patch_size\n        )\n        feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n        final_output = postprocess(feat, self.depth_mode, self.conf_mode)\n        final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n        final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n        if self.has_rgb:\n            rgb_feat = self.rgb_proj(tokens)\n            rgb_feat = rgb_feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size)  # B,3,H,W\n            rgb_output = postprocess_rgb(rgb_feat)\n            final_output.update(rgb_output)\n\n        if self.has_pose:\n            pose = self.pose_head(pose_token)\n            pose = postprocess_pose(pose, self.pose_mode)\n            final_output[\"camera_pose\"] = pose  # B,7\n            final_output[\"pts3d_in_other_view\"] = geotrf(\n                pose_encoding_to_camera(final_output[\"camera_pose\"]),\n                final_output[\"pts3d_in_self_view\"],\n            )\n\n        if self.has_conf:\n            cross_conf = self.cross_conf_proj(tokens)\n            cross_conf = cross_conf.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            cross_conf = F.pixel_shuffle(cross_conf, self.patch_size)[:, 0]\n            final_output[\"conf\"] = reg_dense_conf(cross_conf, mode=self.conf_mode)\n        return final_output\n\n\nclass LinearPts3dPose(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(\n        self, net, has_conf=False, has_rgb=False, has_pose=False, mlp_ratio=4.0\n    ):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n","source_hash":"c52a465d11c1b8ee9145a379e8801682fe1db5ac9f076e5a55de364ec9efa424","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.linear_head.LinearPts3dPose","uri":"program://Human3R/class/src.dust3r.heads.linear_head.LinearPts3dPose#L246-L346","kind":"class","name":"LinearPts3dPose","path":"src/dust3r/heads/linear_head.py","language":"python","start_line":246,"end_line":346,"context_start_line":226,"context_end_line":346,"code":"\n        if self.has_pose:\n            pose = self.pose_head(pose_token)\n            pose = postprocess_pose(pose, self.pose_mode)\n            final_output[\"camera_pose\"] = pose  # B,7\n            final_output[\"pts3d_in_other_view\"] = geotrf(\n                pose_encoding_to_camera(final_output[\"camera_pose\"]),\n                final_output[\"pts3d_in_self_view\"],\n            )\n\n        if self.has_conf:\n            cross_conf = self.cross_conf_proj(tokens)\n            cross_conf = cross_conf.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            cross_conf = F.pixel_shuffle(cross_conf, self.patch_size)[:, 0]\n            final_output[\"conf\"] = reg_dense_conf(cross_conf, mode=self.conf_mode)\n        return final_output\n\n\nclass LinearPts3dPose(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(\n        self, net, has_conf=False, has_rgb=False, has_pose=False, mlp_ratio=4.0\n    ):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        self.proj = Mlp(\n            net.dec_embed_dim,\n            hidden_features=int(mlp_ratio * net.dec_embed_dim),\n            out_features=(3 + has_conf) * self.patch_size**2,\n        )\n        if has_rgb:\n            self.rgb_proj = Mlp(\n                net.dec_embed_dim,\n                hidden_features=int(mlp_ratio * net.dec_embed_dim),\n                out_features=3 * self.patch_size**2,\n            )\n        if has_pose:\n            self.pose_head = PoseDecoder(hidden_size=net.dec_embed_dim)\n            self.final_transform = nn.ModuleList(\n                [\n                    ConditionModulationBlock(\n                        net.dec_embed_dim,\n                        net.dec_num_heads,\n                        mlp_ratio=4.0,\n                        qkv_bias=True,\n                        rope=net.rope,\n                    )\n                    for _ in range(2)\n                ]\n            )\n            self.cross_proj = Mlp(\n                net.dec_embed_dim,\n                hidden_features=int(mlp_ratio * net.dec_embed_dim),\n                out_features=(3 + has_conf) * self.patch_size**2,\n            )\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape, **kwargs):\n        H, W = img_shape\n        tokens = decout[-1]\n        if self.has_pose:\n            pose_token = tokens[:, 0]\n            tokens = tokens[:, 1:]\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n            cross_tokens = tokens\n            for blk in self.final_transform:\n                cross_tokens = blk(cross_tokens, pose_token, kwargs.get(\"pos\"))\n\n        with torch.cuda.amp.autocast(enabled=False):\n            B, S, D = tokens.shape\n\n            feat = self.proj(tokens)  # B,S,D\n            feat = feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n            final_output = postprocess(\n                feat, self.depth_mode, self.conf_mode, pos_z=True\n            )\n            final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n            final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n            if self.has_rgb:\n                rgb_feat = self.rgb_proj(tokens)\n                rgb_feat = rgb_feat.transpose(-1, -2).view(\n                    B, -1, H // self.patch_size, W // self.patch_size\n                )\n                rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size)  # B,3,H,W\n                rgb_output = postprocess_rgb(rgb_feat)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n\n                cross_feat = self.cross_proj(cross_tokens)  # B,S,D\n                cross_feat = cross_feat.transpose(-1, -2).view(\n                    B, -1, H // self.patch_size, W // self.patch_size\n                )\n                cross_feat = F.pixel_shuffle(cross_feat, self.patch_size)  # B,3,H,W\n                tmp = postprocess(cross_feat, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n\n            return final_output","source_hash":"c52a465d11c1b8ee9145a379e8801682fe1db5ac9f076e5a55de364ec9efa424","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.linear_head.__init__","uri":"program://Human3R/function/src.dust3r.heads.linear_head.__init__#L252-L293","kind":"function","name":"__init__","path":"src/dust3r/heads/linear_head.py","language":"python","start_line":252,"end_line":293,"context_start_line":232,"context_end_line":313,"code":"                pose_encoding_to_camera(final_output[\"camera_pose\"]),\n                final_output[\"pts3d_in_self_view\"],\n            )\n\n        if self.has_conf:\n            cross_conf = self.cross_conf_proj(tokens)\n            cross_conf = cross_conf.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            cross_conf = F.pixel_shuffle(cross_conf, self.patch_size)[:, 0]\n            final_output[\"conf\"] = reg_dense_conf(cross_conf, mode=self.conf_mode)\n        return final_output\n\n\nclass LinearPts3dPose(nn.Module):\n    \"\"\"\n    Linear head for dust3r\n    Each token outputs: - 16x16 3D points (+ confidence)\n    \"\"\"\n\n    def __init__(\n        self, net, has_conf=False, has_rgb=False, has_pose=False, mlp_ratio=4.0\n    ):\n        super().__init__()\n        self.patch_size = net.patch_embed.patch_size[0]\n        self.depth_mode = net.depth_mode\n        self.conf_mode = net.conf_mode\n        self.pose_mode = net.pose_mode\n        self.has_conf = has_conf\n        self.has_rgb = has_rgb\n        self.has_pose = has_pose\n\n        self.proj = Mlp(\n            net.dec_embed_dim,\n            hidden_features=int(mlp_ratio * net.dec_embed_dim),\n            out_features=(3 + has_conf) * self.patch_size**2,\n        )\n        if has_rgb:\n            self.rgb_proj = Mlp(\n                net.dec_embed_dim,\n                hidden_features=int(mlp_ratio * net.dec_embed_dim),\n                out_features=3 * self.patch_size**2,\n            )\n        if has_pose:\n            self.pose_head = PoseDecoder(hidden_size=net.dec_embed_dim)\n            self.final_transform = nn.ModuleList(\n                [\n                    ConditionModulationBlock(\n                        net.dec_embed_dim,\n                        net.dec_num_heads,\n                        mlp_ratio=4.0,\n                        qkv_bias=True,\n                        rope=net.rope,\n                    )\n                    for _ in range(2)\n                ]\n            )\n            self.cross_proj = Mlp(\n                net.dec_embed_dim,\n                hidden_features=int(mlp_ratio * net.dec_embed_dim),\n                out_features=(3 + has_conf) * self.patch_size**2,\n            )\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape, **kwargs):\n        H, W = img_shape\n        tokens = decout[-1]\n        if self.has_pose:\n            pose_token = tokens[:, 0]\n            tokens = tokens[:, 1:]\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n            cross_tokens = tokens\n            for blk in self.final_transform:\n                cross_tokens = blk(cross_tokens, pose_token, kwargs.get(\"pos\"))\n\n        with torch.cuda.amp.autocast(enabled=False):\n            B, S, D = tokens.shape\n\n            feat = self.proj(tokens)  # B,S,D","source_hash":"c52a465d11c1b8ee9145a379e8801682fe1db5ac9f076e5a55de364ec9efa424","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.linear_head.setup","uri":"program://Human3R/function/src.dust3r.heads.linear_head.setup#L295-L296","kind":"function","name":"setup","path":"src/dust3r/heads/linear_head.py","language":"python","start_line":295,"end_line":296,"context_start_line":275,"context_end_line":316,"code":"        if has_pose:\n            self.pose_head = PoseDecoder(hidden_size=net.dec_embed_dim)\n            self.final_transform = nn.ModuleList(\n                [\n                    ConditionModulationBlock(\n                        net.dec_embed_dim,\n                        net.dec_num_heads,\n                        mlp_ratio=4.0,\n                        qkv_bias=True,\n                        rope=net.rope,\n                    )\n                    for _ in range(2)\n                ]\n            )\n            self.cross_proj = Mlp(\n                net.dec_embed_dim,\n                hidden_features=int(mlp_ratio * net.dec_embed_dim),\n                out_features=(3 + has_conf) * self.patch_size**2,\n            )\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape, **kwargs):\n        H, W = img_shape\n        tokens = decout[-1]\n        if self.has_pose:\n            pose_token = tokens[:, 0]\n            tokens = tokens[:, 1:]\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n            cross_tokens = tokens\n            for blk in self.final_transform:\n                cross_tokens = blk(cross_tokens, pose_token, kwargs.get(\"pos\"))\n\n        with torch.cuda.amp.autocast(enabled=False):\n            B, S, D = tokens.shape\n\n            feat = self.proj(tokens)  # B,S,D\n            feat = feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )","source_hash":"c52a465d11c1b8ee9145a379e8801682fe1db5ac9f076e5a55de364ec9efa424","truncated":false}
{"repo_id":"Human3R","entity_id":"py:src.dust3r.heads.linear_head.forward","uri":"program://Human3R/function/src.dust3r.heads.linear_head.forward#L298-L346","kind":"function","name":"forward","path":"src/dust3r/heads/linear_head.py","language":"python","start_line":298,"end_line":346,"context_start_line":278,"context_end_line":346,"code":"                [\n                    ConditionModulationBlock(\n                        net.dec_embed_dim,\n                        net.dec_num_heads,\n                        mlp_ratio=4.0,\n                        qkv_bias=True,\n                        rope=net.rope,\n                    )\n                    for _ in range(2)\n                ]\n            )\n            self.cross_proj = Mlp(\n                net.dec_embed_dim,\n                hidden_features=int(mlp_ratio * net.dec_embed_dim),\n                out_features=(3 + has_conf) * self.patch_size**2,\n            )\n\n    def setup(self, croconet):\n        pass\n\n    def forward(self, decout, img_shape, **kwargs):\n        H, W = img_shape\n        tokens = decout[-1]\n        if self.has_pose:\n            pose_token = tokens[:, 0]\n            tokens = tokens[:, 1:]\n            with torch.cuda.amp.autocast(enabled=False):\n                pose = self.pose_head(pose_token)\n            cross_tokens = tokens\n            for blk in self.final_transform:\n                cross_tokens = blk(cross_tokens, pose_token, kwargs.get(\"pos\"))\n\n        with torch.cuda.amp.autocast(enabled=False):\n            B, S, D = tokens.shape\n\n            feat = self.proj(tokens)  # B,S,D\n            feat = feat.transpose(-1, -2).view(\n                B, -1, H // self.patch_size, W // self.patch_size\n            )\n            feat = F.pixel_shuffle(feat, self.patch_size)  # B,3,H,W\n            final_output = postprocess(\n                feat, self.depth_mode, self.conf_mode, pos_z=True\n            )\n            final_output[\"pts3d_in_self_view\"] = final_output.pop(\"pts3d\")\n            final_output[\"conf_self\"] = final_output.pop(\"conf\")\n\n            if self.has_rgb:\n                rgb_feat = self.rgb_proj(tokens)\n                rgb_feat = rgb_feat.transpose(-1, -2).view(\n                    B, -1, H // self.patch_size, W // self.patch_size\n                )\n                rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size)  # B,3,H,W\n                rgb_output = postprocess_rgb(rgb_feat)\n                final_output.update(rgb_output)\n\n            if self.has_pose:\n                pose = postprocess_pose(pose, self.pose_mode)\n                final_output[\"camera_pose\"] = pose  # B,7\n\n                cross_feat = self.cross_proj(cross_tokens)  # B,S,D\n                cross_feat = cross_feat.transpose(-1, -2).view(\n                    B, -1, H // self.patch_size, W // self.patch_size\n                )\n                cross_feat = F.pixel_shuffle(cross_feat, self.patch_size)  # B,3,H,W\n                tmp = postprocess(cross_feat, self.depth_mode, self.conf_mode)\n                final_output[\"pts3d_in_other_view\"] = tmp.pop(\"pts3d\")\n                final_output[\"conf\"] = tmp.pop(\"conf\")\n\n            return final_output","source_hash":"c52a465d11c1b8ee9145a379e8801682fe1db5ac9f076e5a55de364ec9efa424","truncated":false}
{"repo_id":"Human3R","entity_id":"file:demo.py","uri":"program://Human3R/file/demo.py","kind":"file","name":"demo.py","path":"demo.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"#!/usr/bin/env python3\n\"\"\"\nModified from CUT3R: https://github.com/CUT3R/CUT3R\n\nOnline Human-Scene Reconstruction Inference and Visualization Script\n\nThis script performs inference using the ARCroco3DStereo model and visualizes the\nresulting 3D scene point clouds and SMPLX sequences with the SceneHumanViewer. \nUse the command-line arguments to adjust parameters \nsuch as the model checkpoint path, image sequence directory, image size, device, etc.\n\nExample:\n    python demo.py --model_path src/human3r.pth --size 512 \\\n        --seq_path examples/GoodMornin1.mp4 --subsample 1 --vis_threshold 2 \\\n        --downsample_factor 1 --use_ttt3r --reset_interval 100\n\"\"\"\n\nimport os\nimport numpy as np\nimport torch\nimport time","source_hash":"d7f1b65a79599b5081acf2f9a0b80bd6656913e66c74093a514651e32499b044","truncated":false}
{"repo_id":"Human3R","entity_id":"file:add_ckpt_path.py","uri":"program://Human3R/file/add_ckpt_path.py","kind":"file","name":"add_ckpt_path.py","path":"add_ckpt_path.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":9,"code":"import sys\nimport os\nimport os.path as path\n\n\ndef add_path_to_dust3r(ckpt):\n    HERE_PATH = os.path.dirname(os.path.abspath(ckpt))\n    # workaround for sibling import\n    sys.path.insert(0, HERE_PATH)","source_hash":"40c45aec68241a855e4a09e6a7c95aeadb4ff6fc5fa40568ee98b586b5dc6d5a","truncated":false}
{"repo_id":"Human3R","entity_id":"file:viser_utils.py","uri":"program://Human3R/file/viser_utils.py","kind":"file","name":"viser_utils.py","path":"viser_utils.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import torch\nimport os\nfrom matplotlib.figure import Figure\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nimport matplotlib as mpl\nimport cv2\nimport numpy as np\nimport matplotlib.cm as cm\nimport viser\nimport viser.transforms as tf\nimport time\nimport trimesh\nimport dataclasses\nfrom scipy.spatial.transform import Rotation\nfrom skimage.morphology import binary_dilation, binary_erosion, disk\nfrom src.dust3r.viz import (\n    add_scene_cam,\n    CAM_COLORS,\n    OPENGL,\n    pts3d_to_trimesh,","source_hash":"c46b574d08f322872fafef6e14f9e12bca9ee8145d6792846847a5cfdb05ead6","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/dataset/prepare_3dpw.py","uri":"program://Human3R/file/eval/dataset/prepare_3dpw.py","kind":"file","name":"eval/dataset/prepare_3dpw.py","path":"eval/dataset/prepare_3dpw.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Modified from Multi-HMR\n# Preprocess 3DPW dataset for human evaluation, and saves the annotations\n# files (i.e., 3dpw_test.pkl) in [ANNOT_DIR].\n# Usage: python -m eval.dataset.prepare_3dpw \"create_annots()\"\n\nimport os\nos.environ[\"PYOPENGL_PLATFORM\"] = \"egl\"\nos.environ['EGL_DEVICE_ID'] = '0'\n\nimport pickle\nimport torch\nimport smplx\nfrom tqdm import tqdm\nimport sys\nimport numpy as np\nfrom PIL import Image, ImageFile\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nsys.path.append(os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(__file__))), 'src'))\nfrom dust3r.smpl_model import SMPLX_DIR\n","source_hash":"eefff66a8480dc590d2e891a31922dff5fed33d1a2119d898a226143535d8fef","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/relpose/metadata.py","uri":"program://Human3R/file/eval/relpose/metadata.py","kind":"file","name":"eval/relpose/metadata.py","path":"eval/relpose/metadata.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os\nimport glob\nfrom tqdm import tqdm\n\n# Define the merged dataset metadata dictionary\ndataset_metadata = {\n    \"davis\": {\n        \"img_path\": \"data/davis/DAVIS/JPEGImages/480p\",\n        \"mask_path\": \"data/davis/DAVIS/masked_images/480p\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: None,\n        \"traj_format\": None,\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: os.path.join(mask_path, seq),\n        \"skip_condition\": None,\n        \"process_func\": None,  # Not used in mono depth estimation\n    },\n    \"kitti\": {\n        \"img_path\": \"data/kitti/depth_selection/val_selection_cropped/image_gathered\",  # Default path\n        \"mask_path\": None,","source_hash":"fed1eca7775b0155dd84b26e08fca02e97e4ac8bdd84a396f5767f62c64cd478","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/relpose/utils.py","uri":"program://Human3R/file/eval/relpose/utils.py","kind":"file","name":"eval/relpose/utils.py","path":"eval/relpose/utils.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"from copy import deepcopy\nimport cv2\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport roma\nfrom copy import deepcopy\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom eval.relpose.evo_utils import *\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\n\n# from checkpoints.dust3r.viz import colorize_np, colorize\n","source_hash":"f58d9558865899bb3fbaeb167358e5f119ff3a029bf29cddb7f9094b6a8071c3","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/relpose/launch.py","uri":"program://Human3R/file/eval/relpose/launch.py","kind":"file","name":"eval/relpose/launch.py","path":"eval/relpose/launch.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os\nimport sys\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nimport math\nimport cv2\nimport numpy as np\nimport torch\nimport argparse\n\nfrom copy import deepcopy\nfrom eval.relpose.metadata import dataset_metadata\nfrom eval.relpose.utils import *\n\nfrom accelerate import PartialState\nfrom add_ckpt_path import add_path_to_dust3r\n\nfrom tqdm import tqdm\n\n\ndef get_args_parser():","source_hash":"9427c203a9f9a49ee297b62876dd88027e5fc38e89eafb54f2042f1aee1bb8b7","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/relpose/evo_utils.py","uri":"program://Human3R/file/eval/relpose/evo_utils.py","kind":"file","name":"eval/relpose/evo_utils.py","path":"eval/relpose/evo_utils.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"\nimport os\nimport re\nfrom copy import deepcopy\nfrom pathlib import Path\n\nimport evo.main_ape as main_ape\nimport evo.main_rpe as main_rpe\n\nimport matplotlib.pyplot as plt\nimport numpy as np\nfrom evo.core import sync\nfrom evo.core.metrics import PoseRelation, Unit\nfrom evo.core.trajectory import PosePath3D, PoseTrajectory3D\nfrom evo.tools import file_interface, plot\nfrom scipy.spatial.transform import Rotation\nfrom evo.core import metrics\n\n\ndef sintel_cam_read(filename):\n    \"\"\"Read camera data, return (M,N) tuple.","source_hash":"b157c2d2c619716d80bae786606f4d81c5d14820cd0a10f3ad8b093d6368c915","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/video_depth/metadata.py","uri":"program://Human3R/file/eval/video_depth/metadata.py","kind":"file","name":"eval/video_depth/metadata.py","path":"eval/video_depth/metadata.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os\nimport glob\nfrom tqdm import tqdm\n\n# Define the merged dataset metadata dictionary\ndataset_metadata = {\n    \"davis\": {\n        \"img_path\": \"data/davis/DAVIS/JPEGImages/480p\",\n        \"mask_path\": \"data/davis/DAVIS/masked_images/480p\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(img_path, seq),\n        \"gt_traj_func\": lambda img_path, anno_path, seq: None,\n        \"traj_format\": None,\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_seq_func\": lambda mask_path, seq: os.path.join(mask_path, seq),\n        \"skip_condition\": None,\n        \"process_func\": None,  # Not used in mono depth estimation\n    },\n    \"kitti\": {\n        \"img_path\": \"data/kitti/depth_selection/val_selection_cropped/image_gathered\",  # Default path\n        \"mask_path\": None,","source_hash":"ffd74aa6378d590ead9dc876de366133264ba355cc9575aa84610d3c63c51c0c","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/video_depth/utils.py","uri":"program://Human3R/file/eval/video_depth/utils.py","kind":"file","name":"eval/video_depth/utils.py","path":"eval/video_depth/utils.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"from copy import deepcopy\nimport cv2\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport roma\nfrom copy import deepcopy\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\n\n\ndef save_focals(cam_dict, path):\n    # convert focal to txt","source_hash":"465aa4f9585546500272cc92d5aed6eb4cd319db9955c91991a13f60a6c1c287","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/video_depth/launch.py","uri":"program://Human3R/file/eval/video_depth/launch.py","kind":"file","name":"eval/video_depth/launch.py","path":"eval/video_depth/launch.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os\nimport sys\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nimport math\nimport cv2\nimport numpy as np\nimport torch\nimport argparse\n\nfrom copy import deepcopy\nfrom eval.video_depth.metadata import dataset_metadata\nfrom eval.video_depth.utils import save_depth_maps\nfrom accelerate import PartialState\nfrom add_ckpt_path import add_path_to_dust3r\nimport time\nfrom tqdm import tqdm\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()","source_hash":"3f359473cd914d2450b9679c39667f98d970a03cdaf1d817c168b6c112cbfc1f","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/video_depth/tools.py","uri":"program://Human3R/file/eval/video_depth/tools.py","kind":"file","name":"eval/video_depth/tools.py","path":"eval/video_depth/tools.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import torch\nimport numpy as np\nimport cv2\nimport glob\nimport argparse\nfrom pathlib import Path\nfrom tqdm import tqdm\nfrom copy import deepcopy\nfrom scipy.optimize import minimize\nimport os\nfrom collections import defaultdict\n\n\ndef group_by_directory(pathes, idx=-1):\n    \"\"\"\n    Groups the file paths based on the second-to-last directory in their paths.\n\n    Parameters:\n    - pathes (list): List of file paths.\n\n    Returns:","source_hash":"079171f863616133446644309f22940572375c7a1ce61131084e01dbd38b0863","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/video_depth/eval_depth.py","uri":"program://Human3R/file/eval/video_depth/eval_depth.py","kind":"file","name":"eval/video_depth/eval_depth.py","path":"eval/video_depth/eval_depth.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os\nimport sys\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nfrom eval.video_depth.tools import depth_evaluation, group_by_directory\nimport numpy as np\nimport cv2\nfrom tqdm import tqdm\nimport glob\nfrom PIL import Image\nimport argparse\nimport json\nfrom eval.video_depth.metadata import dataset_metadata\n\n\ndef get_args_parser():\n    parser = argparse.ArgumentParser()\n\n    parser.add_argument(\n        \"--output_dir\",\n        type=str,","source_hash":"924683ea2cb1871eefa5d1266bb3aed3d887aaabf173d23b4a20b537325d0b34","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/global_human/metadata.py","uri":"program://Human3R/file/eval/global_human/metadata.py","kind":"file","name":"eval/global_human/metadata.py","path":"eval/global_human/metadata.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os\nimport torch\nimport numpy as np\nimport pickle\nfrom eval.global_human.data_utils import *\n\n# Define the merged dataset metadata dictionary\n\ndef create_emdb(split):\n    return {\n        \"img_path\": \"/path/to/EMDB\",\n        \"dir_path_func\": lambda img_path, seq: os.path.join(f\"{img_path}/{seq}/images\"),\n        \"seq_list\": None,\n        \"full_seq\": True,\n        \"mask_path_func\": lambda filelist: [],\n        \"split\": split,\n        \"subsample\": 1,\n        \"get_view_func\": lambda inputs: load_view_emdb(*inputs),\n        \"get_seq_func\": lambda img_path, split, annots: get_seq_emdb(img_path, split, annots),\n        \"get_annot_func\": lambda img_path, split: get_annot_emdb(img_path, split),\n        \"is_global\": lambda split: {1: False, 2: True}[split],","source_hash":"433a0caf5c126d16f2360372e5053572f4509421411f7fa4c229bbcad9a6ef0f","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/global_human/data_utils.py","uri":"program://Human3R/file/eval/global_human/data_utils.py","kind":"file","name":"eval/global_human/data_utils.py","path":"eval/global_human/data_utils.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Modified from GVHMR [https://github.com/zju3dv/GVHMR].\n# Load EMDB data\n\nimport pickle\nimport torch\n\nEMDB1_LIST = [\n    'P8/69_outdoor_cartwheel/P8_69_outdoor_cartwheel_data.pkl', # 656\n    'P5/42_indoor_dancing/P5_42_indoor_dancing_data.pkl', # 1291\n    'P6/51_outdoor_dancing/P6_51_outdoor_dancing_data.pkl',  # 1427\n    'P2/23_outdoor_hug_tree/P2_23_outdoor_hug_tree_data.pkl',  # 1086\n    'P6/49_outdoor_big_stairs_down/P6_49_outdoor_big_stairs_down_data.pkl',  # DUPLICATE 1559\n\n    'P7/59_outdoor_rom/P7_59_outdoor_rom_data.pkl', # 1839\n    'P3/31_outdoor_workout/P3_31_outdoor_workout_data.pkl',  # 1216\n    'P3/33_outdoor_soccer_warmup_b/P3_33_outdoor_soccer_warmup_b_data.pkl',  # 1433\n    'P7/57_outdoor_rock_chair/P7_57_outdoor_rock_chair_data.pkl', # DUPLICATE 1558\n\n    'P3/32_outdoor_soccer_warmup_a/P3_32_outdoor_soccer_warmup_a_data.pkl',  # 1084\n    'P8/64_outdoor_skateboard/P8_64_outdoor_skateboard_data.pkl',   # DUPLICATE 1704\n    'P7/60_outdoor_workout/P7_60_outdoor_workout_data.pkl',  # 1693","source_hash":"2a695c1ada5890d3f4d5a97b6195f7c76acea26e923ca9a1fceeb91cc7860c14","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/global_human/utils.py","uri":"program://Human3R/file/eval/global_human/utils.py","kind":"file","name":"eval/global_human/utils.py","path":"eval/global_human/utils.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"from copy import deepcopy\nimport cv2\n\nimport os\nimport re\nimport numpy as np\nimport torch\nimport roma\nimport math\nimport tqdm\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom matplotlib.backends.backend_agg import FigureCanvasAgg\nfrom scipy.spatial.transform import Rotation\nfrom eval.relpose.evo_utils import *\nfrom PIL import Image\nimport imageio.v2 as iio\nfrom matplotlib.figure import Figure\nfrom itertools import product\n","source_hash":"e8f234fd3597419474135b8591f007ef166f375f1186d5447152895c3fd5a168","truncated":false}
{"repo_id":"Human3R","entity_id":"file:eval/global_human/launch.py","uri":"program://Human3R/file/eval/global_human/launch.py","kind":"file","name":"eval/global_human/launch.py","path":"eval/global_human/launch.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os\nimport sys\nimport gc\n\nsys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__))))\nimport numpy as np\nimport torch\nimport argparse\n\nfrom copy import deepcopy\nfrom eval.global_human.metadata import dataset_metadata\nfrom eval.global_human.utils import *\n\nfrom accelerate import PartialState\nfrom add_ckpt_path import add_path_to_dust3r\n\nfrom collections import defaultdict, Counter\nfrom tqdm import tqdm\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n","source_hash":"bceabf4fb251e4769e4a2becbd1c6be29eea554469dd3703feb1538f3ecfb9da","truncated":false}
{"repo_id":"Human3R","entity_id":"file:datasets_preprocess/preprocess_bedlam.py","uri":"program://Human3R/file/datasets_preprocess/preprocess_bedlam.py","kind":"file","name":"datasets_preprocess/preprocess_bedlam.py","path":"datasets_preprocess/preprocess_bedlam.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"#!/usr/bin/env python3\n\"\"\"\nModified from CUT3R [https://github.com/CUT3R/CUT3R].\n\nProcess Bedlam scenes by computing camera intrinsics and extrinsics\nfrom extracted data. The script reads per-scene CSV and image/depth files,\ncomputes the necessary camera parameters, and saves the resulting camera\nfiles (as .npz files) in an output directory.\nNote: CUT3R filtered out HDRI scenes and closeup scenes.\nWe also filter out the sequences without SMPLX annotations following Multi-HMR.\n\nUsage:\n    python preprocess_bedlam.py --root /path/to/extracted_data \\\n                             --outdir /path/to/processed_bedlam \\\n                             --annot_dir /path/to/bedlam/processed_labels\n\"\"\"\n\nimport os\nimport sys\nimport cv2\nimport numpy as np","source_hash":"9f9ebd197547b6d54e576a00b4cbde475564207dd2f52228126212b30b1e3a54","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/train.py","uri":"program://Human3R/file/src/train.py","kind":"file","name":"src/train.py","path":"src/train.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# --------------------------------------------------------\n# training code for Human3R\n# --------------------------------------------------------\n# References:\n# CUT3R: https://github.com/CUT3R/CUT3R\n# DUSt3R: https://github.com/naver/dust3r\n# --------------------------------------------------------\nimport argparse\nimport datetime\nimport json\nimport numpy as np\nimport os\nimport sys\nimport time\nimport math\nfrom collections import defaultdict\nfrom pathlib import Path\nfrom typing import Sized\n\nimport torch\nimport torch.backends.cudnn as cudnn","source_hash":"614cbc639f064b4c3f72dd4b99ee1abb1988318dbd55c0cd789c01194bdfcfc7","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/mhmr/blocks/camera_embed.py","uri":"program://Human3R/file/src/mhmr/blocks/camera_embed.py","kind":"file","name":"src/mhmr/blocks/camera_embed.py","path":"src/mhmr/blocks/camera_embed.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\nimport numpy as np\n\nclass FourierPositionEncoding(nn.Module):\n    def __init__(self, n, num_bands, max_resolution):\n        \"\"\"\n        Module that generate Fourier encoding - no learning involved\n        \"\"\"\n        super().__init__()\n\n        self.num_bands = num_bands\n        self.max_resolution = [max_resolution] * n\n    \n    @property\n    def channels(self):\n        \"\"\"","source_hash":"756359c5c2186151b8e64c7ff20d4581419015dee1f45e8fc689f7c5ec0a7b7b","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/mhmr/blocks/__init__.py","uri":"program://Human3R/file/src/mhmr/blocks/__init__.py","kind":"file","name":"src/mhmr/blocks/__init__.py","path":"src/mhmr/blocks/__init__.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":6,"code":"\nfrom .camera_embed import FourierPositionEncoding\n\nfrom .dinov2 import Dinov2Backbone\n\nfrom .cross_attn_transformer import TransformerDecoder","source_hash":"15566db4aab8509abd35de3753f6e9e2207fad27f3b34fe7445f0f2cee8326d9","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/mhmr/blocks/cross_attn_transformer.py","uri":"program://Human3R/file/src/mhmr/blocks/cross_attn_transformer.py","kind":"file","name":"src/mhmr/blocks/cross_attn_transformer.py","path":"src/mhmr/blocks/cross_attn_transformer.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nfrom typing import Callable, Optional\nimport torch\nfrom torch import nn\nfrom inspect import isfunction\nfrom einops import rearrange\n\nclass AdaptiveLayerNorm1D(torch.nn.Module):\n    \"\"\"\n    Code modified from https://github.com/shubham-goel/4D-Humans/blob/a0def798c7eac811a63c8220fcc22d983b39785e/hmr2/models/components/t_cond_mlp.py#L7\n    \"\"\"\n    def __init__(self, data_dim: int, norm_cond_dim: int):\n        super().__init__()\n        if data_dim <= 0:\n            raise ValueError(f\"data_dim must be positive, but got {data_dim}\")\n        if norm_cond_dim <= 0:\n            raise ValueError(f\"norm_cond_dim must be positive, but got {norm_cond_dim}\")\n        self.norm = torch.nn.LayerNorm(","source_hash":"22fde2e72e2d4f36cc37ee5a415d958d7f20e9aafc0db4033db90309d6a6d8bc","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/mhmr/blocks/dinov2.py","uri":"program://Human3R/file/src/mhmr/blocks/dinov2.py","kind":"file","name":"src/mhmr/blocks/dinov2.py","path":"src/mhmr/blocks/dinov2.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\n\nclass Dinov2Backbone(nn.Module):\n    def __init__(self, name='dinov2_vitb14', pretrained=False, *args, **kwargs):\n        super().__init__()\n        self.name = name\n        self.encoder = torch.hub.load('facebookresearch/dinov2', self.name, pretrained=pretrained)\n        self.patch_size = self.encoder.patch_size\n        self.embed_dim = self.encoder.embed_dim\n\n    def forward(self, x):\n        \"\"\"\n        Encode a RGB image using a ViT-backbone\n        Args:\n            - x: torch.Tensor of shape [bs,3,w,h]\n        Return:","source_hash":"4b239e93d13eb9ee975f749374c8b36531fb220fb767a8488946258cdd3e630b","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/pretrain.py","uri":"program://Human3R/file/src/croco/pretrain.py","kind":"file","name":"src/croco/pretrain.py","path":"src/croco/pretrain.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# Pre-training CroCo\n# --------------------------------------------------------\n# References:\n# MAE: https://github.com/facebookresearch/mae\n# DeiT: https://github.com/facebookresearch/deit\n# BEiT: https://github.com/microsoft/unilm/tree/master/beit\n# --------------------------------------------------------\nimport argparse\nimport datetime\nimport json\nimport numpy as np\nimport os\nimport sys\nimport time\nimport math\nfrom pathlib import Path\nfrom typing import Iterable","source_hash":"d9ab7f1f3c1d4175e5eb4be181de5b2941ddb31fe6a9904212936beaa0d4ca1a","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/blocks.py","uri":"program://Human3R/file/src/croco/models/blocks.py","kind":"file","name":"src/croco/models/blocks.py","path":"src/croco/models/blocks.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# Main encoder/decoder blocks\n# --------------------------------------------------------\n# References:\n# timm\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/helpers.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/mlp.py\n# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/patch_embed.py\n\n\nimport torch\nimport torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc","source_hash":"aac24e819cfe9ff2072c93b91cf5059978a406aad986abe510ffe0fc4a66ab71","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/criterion.py","uri":"program://Human3R/file/src/croco/models/criterion.py","kind":"file","name":"src/croco/models/criterion.py","path":"src/croco/models/criterion.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# Criterion to train CroCo\n# --------------------------------------------------------\n# References:\n# MAE: https://github.com/facebookresearch/mae\n# --------------------------------------------------------\n\nimport torch\n\n\nclass MaskedMSE(torch.nn.Module):\n\n    def __init__(self, norm_pix_loss=False, masked=True):\n        \"\"\"\n        norm_pix_loss: normalize each patch by their pixel mean and variance\n        masked: compute loss over the masked patches only\n        \"\"\"\n        super().__init__()","source_hash":"19d837dec5326843d0e4c03ffae6cab0f9274f0c3354894e7fcd9762c97ef47c","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/croco_downstream.py","uri":"program://Human3R/file/src/croco/models/croco_downstream.py","kind":"file","name":"src/croco/models/croco_downstream.py","path":"src/croco/models/croco_downstream.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# CroCo model for downstream tasks\n# --------------------------------------------------------\n\nimport torch\n\nfrom .croco import CroCoNet\n\n\ndef croco_args_from_ckpt(ckpt):\n    if \"croco_kwargs\" in ckpt:  # CroCo v2 released models\n        return ckpt[\"croco_kwargs\"]\n    elif \"args\" in ckpt and hasattr(\n        ckpt[\"args\"], \"model\"\n    ):  # pretrained using the official code release\n        s = ckpt[\n            \"args\"\n        ].model  # eg \"CroCoNet(enc_embed_dim=1024, enc_num_heads=16, enc_depth=24)\"","source_hash":"5c85337ca98a97940c6c4600c624ae2eec9af8435ce6b7ad5275e837bf27f160","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/dpt_block.py","uri":"program://Human3R/file/src/croco/models/dpt_block.py","kind":"file","name":"src/croco/models/dpt_block.py","path":"src/croco/models/dpt_block.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# DPT head for ViTs\n# --------------------------------------------------------\n# References:\n# https://github.com/isl-org/DPT\n# https://github.com/EPFL-VILAB/MultiMAE/blob/main/multimae/output_adapters.py\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom einops import rearrange, repeat\nfrom typing import Union, Tuple, Iterable, List, Optional, Dict\n\n\ndef pair(t):\n    return t if isinstance(t, tuple) else (t, t)\n\n","source_hash":"42d7fe6afaff9c6ca6acf1b808eddb6e6d480722b95d04c08da8ce27cd17d476","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/pos_embed.py","uri":"program://Human3R/file/src/croco/models/pos_embed.py","kind":"file","name":"src/croco/models/pos_embed.py","path":"src/croco/models/pos_embed.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# Position embedding utils\n# --------------------------------------------------------\n\n\nimport numpy as np\n\nimport torch\n\n\n# --------------------------------------------------------\n# 2D sine-cosine position embedding\n# References:\n# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py\n# Transformer: https://github.com/tensorflow/models/blob/master/official/nlp/transformer/model_utils.py\n# MoCo v3: https://github.com/facebookresearch/moco-v3\n# --------------------------------------------------------","source_hash":"11141e25c64044391924b5fe731453d580d810a47afb2b692ae5d46a37e71c2e","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/croco.py","uri":"program://Human3R/file/src/croco/models/croco.py","kind":"file","name":"src/croco/models/croco.py","path":"src/croco/models/croco.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# CroCo model during pretraining\n# --------------------------------------------------------\n\n\nimport torch\nimport torch.nn as nn\n\ntorch.backends.cuda.matmul.allow_tf32 = True  # for gpu >= Ampere and pytorch >= 1.12\nfrom functools import partial\n\nfrom models.blocks import Block, DecoderBlock, PatchEmbed\nfrom models.pos_embed import get_2d_sincos_pos_embed, RoPE2D\nfrom models.masking import RandomMask\n\nfrom transformers import PretrainedConfig\nfrom transformers import PreTrainedModel","source_hash":"ecdfa14cceba7c99007a2548a804f2c3ecac85c29f12c9228a92838ac38aabfa","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/head_downstream.py","uri":"program://Human3R/file/src/croco/models/head_downstream.py","kind":"file","name":"src/croco/models/head_downstream.py","path":"src/croco/models/head_downstream.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Heads for downstream tasks\n# --------------------------------------------------------\n\n\"\"\"\nA head is a module where the __init__ defines only the head hyperparameters.\nA method setup(croconet) takes a CroCoNet and set all layers according to the head and croconet attributes.\nThe forward takes the features as well as a dictionary img_info containing the keys 'width' and 'height'\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom .dpt_block import DPTOutputAdapter\n\n\nclass PixelwiseTaskWithDPT(nn.Module):\n    \"\"\"DPT module for CroCo.\n    by default, hooks_idx will be equal to:","source_hash":"b02fdbfeb55dd3ffa53c2442a566df2f33f43bda7ffb84237665ef78f615dad8","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/masking.py","uri":"program://Human3R/file/src/croco/models/masking.py","kind":"file","name":"src/croco/models/masking.py","path":"src/croco/models/masking.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\n# --------------------------------------------------------\n# Masking utils\n# --------------------------------------------------------\n\nimport torch\nimport torch.nn as nn\n\n\nclass RandomMask(nn.Module):\n    \"\"\"\n    random masking\n    \"\"\"\n\n    def __init__(self, num_patches, mask_ratio):\n        super().__init__()\n        self.num_patches = num_patches\n        self.num_mask = int(mask_ratio * self.num_patches)","source_hash":"541cf1fac15d432d40fd89502e820103309d59aa1bb8e54d1088aea1fe72a460","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/curope/setup.py","uri":"program://Human3R/file/src/croco/models/curope/setup.py","kind":"file","name":"src/croco/models/curope/setup.py","path":"src/croco/models/curope/setup.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nfrom setuptools import setup\nfrom torch import cuda\nfrom torch.utils.cpp_extension import BuildExtension, CUDAExtension\n\n# compile for all possible CUDA architectures\nall_cuda_archs = cuda.get_gencode_flags().replace(\"compute=\", \"arch=\").split()\n# alternatively, you can list cuda archs that you want, eg:\n# all_cuda_archs = [\n# '-gencode', 'arch=compute_70,code=sm_70',\n# '-gencode', 'arch=compute_75,code=sm_75',\n# '-gencode', 'arch=compute_80,code=sm_80',\n# '-gencode', 'arch=compute_86,code=sm_86'\n# ]\n\nsetup(\n    name=\"curope\",\n    ext_modules=[\n        CUDAExtension(","source_hash":"7b85998c132be2ece2eef307d088355e73c39b326c874518862b34f1fd209057","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/curope/curope2d.py","uri":"program://Human3R/file/src/croco/models/curope/curope2d.py","kind":"file","name":"src/croco/models/curope/curope2d.py","path":"src/croco/models/curope/curope2d.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport torch\n\ntry:\n    import curope as _kernels  # run `python setup.py install`\nexcept ModuleNotFoundError:\n    from . import curope as _kernels  # run `python setup.py build_ext --inplace`\n\n\nclass cuRoPE2D_func(torch.autograd.Function):\n\n    @staticmethod\n    def forward(ctx, tokens, positions, base, F0=1):\n        ctx.save_for_backward(positions)\n        ctx.saved_base = base\n        ctx.saved_F0 = F0\n        # tokens = tokens.clone() # uncomment this if inplace doesn't work\n        _kernels.rope_2d(tokens, positions, base, F0)\n        ctx.mark_dirty(tokens)","source_hash":"dc49348fc210445a9b12252e1faf2ec678346e3ae3ef9735867eb39556739377","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/models/curope/__init__.py","uri":"program://Human3R/file/src/croco/models/curope/__init__.py","kind":"file","name":"src/croco/models/curope/__init__.py","path":"src/croco/models/curope/__init__.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":4,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nfrom .curope2d import cuRoPE2D","source_hash":"d30b867f4f8b45f019c2b9a24ef38146867e09fc0d86c55232cc56e5b1d4938e","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/utils/misc.py","uri":"program://Human3R/file/src/croco/utils/misc.py","kind":"file","name":"src/croco/utils/misc.py","path":"src/croco/utils/misc.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# utilitary functions for CroCo\n# --------------------------------------------------------\n# References:\n# MAE: https://github.com/facebookresearch/mae\n# DeiT: https://github.com/facebookresearch/deit\n# BEiT: https://github.com/microsoft/unilm/tree/master/beit\n# --------------------------------------------------------\n\nimport builtins\nimport datetime\nimport os\nimport time\nimport math\nimport json\nfrom collections import defaultdict, deque\nfrom pathlib import Path\nimport numpy as np","source_hash":"b715e338c54b6454e0197d257d2135b6c7a850dfee8eb28eeb160b3f93437677","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/stereoflow/criterion.py","uri":"program://Human3R/file/src/croco/stereoflow/criterion.py","kind":"file","name":"src/croco/stereoflow/criterion.py","path":"src/croco/stereoflow/criterion.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Losses, metrics per batch, metrics per dataset\n# --------------------------------------------------------\n\nimport torch\nfrom torch import nn\nimport torch.nn.functional as F\n\n\ndef _get_gtnorm(gt):\n    if gt.size(1) == 1:  # stereo\n        return gt\n    # flow\n    return torch.sqrt(torch.sum(gt**2, dim=1, keepdims=True))  # Bx1xHxW\n\n\n############ losses without confidence\n","source_hash":"29d8fc10728d56a1c64269372fbba477fb01f405e9ff1fe1fdb0df00c7c13337","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/stereoflow/augmentor.py","uri":"program://Human3R/file/src/croco/stereoflow/augmentor.py","kind":"file","name":"src/croco/stereoflow/augmentor.py","path":"src/croco/stereoflow/augmentor.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Data augmentation for training stereo and flow\n# --------------------------------------------------------\n\n# References\n# https://github.com/autonomousvision/unimatch/blob/master/dataloader/stereo/transforms.py\n# https://github.com/autonomousvision/unimatch/blob/master/dataloader/flow/transforms.py\n\n\nimport numpy as np\nimport random\nfrom PIL import Image\n\nimport cv2\n\ncv2.setNumThreads(0)\ncv2.ocl.setUseOpenCL(False)\n","source_hash":"728502a9a6732b3970367a266e8f9f8bad2421273f4a528df0e4a2012777ba59","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/stereoflow/train.py","uri":"program://Human3R/file/src/croco/stereoflow/train.py","kind":"file","name":"src/croco/stereoflow/train.py","path":"src/croco/stereoflow/train.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Main training function\n# --------------------------------------------------------\n\nimport argparse\nimport datetime\nimport json\nimport numpy as np\nimport os\nimport sys\nimport time\n\nimport torch\nimport torch.distributed as dist\nimport torch.backends.cudnn as cudnn\nfrom torch.utils.tensorboard import SummaryWriter\nimport torchvision.transforms as transforms\nimport torchvision.datasets as datasets","source_hash":"504e0fbf6bf43cfb177a5b689add022db755bc954edd509fef0acb4b2b9693ad","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/stereoflow/datasets_stereo.py","uri":"program://Human3R/file/src/croco/stereoflow/datasets_stereo.py","kind":"file","name":"src/croco/stereoflow/datasets_stereo.py","path":"src/croco/stereoflow/datasets_stereo.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Dataset structure for stereo\n# --------------------------------------------------------\n\nimport sys, os\nimport os.path as osp\nimport pickle\nimport numpy as np\nfrom PIL import Image\nimport json\nimport h5py\nfrom glob import glob\nimport cv2\n\nimport torch\nfrom torch.utils import data\n\nfrom .augmentor import StereoAugmentor","source_hash":"5229e5eb7a4605a55cd0194c229be69974f1b1cc820c00638c155ad484da4912","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/stereoflow/test.py","uri":"program://Human3R/file/src/croco/stereoflow/test.py","kind":"file","name":"src/croco/stereoflow/test.py","path":"src/croco/stereoflow/test.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Main test function\n# --------------------------------------------------------\n\nimport os\nimport argparse\nimport pickle\nfrom PIL import Image\nimport numpy as np\nfrom tqdm import tqdm\n\nimport torch\nfrom torch.utils.data import DataLoader\n\nimport utils.misc as misc\nfrom models.croco_downstream import CroCoDownstreamBinocular\nfrom models.head_downstream import PixelwiseTaskWithDPT\n","source_hash":"9b2037c221595588a9c1684ecf9bedc5786069a7378f136984998d469c1ea252","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/stereoflow/engine.py","uri":"program://Human3R/file/src/croco/stereoflow/engine.py","kind":"file","name":"src/croco/stereoflow/engine.py","path":"src/croco/stereoflow/engine.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Main function for training one epoch or testing\n# --------------------------------------------------------\n\nimport math\nimport sys\nfrom typing import Iterable\nimport numpy as np\nimport torch\nimport torchvision\n\nfrom utils import misc as misc\n\n\ndef split_prediction_conf(predictions, with_conf=False):\n    if not with_conf:\n        return predictions, None\n    conf = predictions[:, -1:, :, :]","source_hash":"0c98d5eca3d81cd24092f37d1a04a1f0d94716098c6112af566ec203bfdf90b8","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/stereoflow/datasets_flow.py","uri":"program://Human3R/file/src/croco/stereoflow/datasets_flow.py","kind":"file","name":"src/croco/stereoflow/datasets_flow.py","path":"src/croco/stereoflow/datasets_flow.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n# --------------------------------------------------------\n# Dataset structure for flow\n# --------------------------------------------------------\n\nimport os\nimport os.path as osp\nimport pickle\nimport numpy as np\nimport struct\nfrom PIL import Image\nimport json\nimport h5py\nimport torch\nfrom torch.utils import data\n\nfrom .augmentor import FlowAugmentor\nfrom .datasets_stereo import _read_img, img_to_tensor, dataset_to_root, _read_pfm\nfrom copy import deepcopy","source_hash":"94b8a487e1d90cc3495840ffe1385c20ebcec9a3d53918dcba9cf200bbcc1917","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/datasets/pairs_dataset.py","uri":"program://Human3R/file/src/croco/datasets/pairs_dataset.py","kind":"file","name":"src/croco/datasets/pairs_dataset.py","path":"src/croco/datasets/pairs_dataset.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nfrom torch.utils.data import Dataset\nfrom PIL import Image\n\nfrom datasets.transforms import get_pair_transforms\n\n\ndef load_image(impath):\n    return Image.open(impath)\n\n\ndef load_pairs_from_cache_file(fname, root=\"\"):\n    assert os.path.isfile(\n        fname\n    ), \"cannot parse pairs from {:s}, file does not exist\".format(fname)\n    with open(fname, \"r\") as fid:\n        lines = fid.read().strip().splitlines()\n    pairs = [","source_hash":"e7deac836d7e51024a28ea5ed1f3559fd3ae243f447fa1f364ab22bc7c8854d0","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/datasets/transforms.py","uri":"program://Human3R/file/src/croco/datasets/transforms.py","kind":"file","name":"src/croco/datasets/transforms.py","path":"src/croco/datasets/transforms.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport torch\nimport torchvision.transforms\nimport torchvision.transforms.functional as F\n\n# \"Pair\": apply a transform on a pair\n# \"Both\": apply the exact same transform to both images\n\n\nclass ComposePair(torchvision.transforms.Compose):\n    def __call__(self, img1, img2):\n        for t in self.transforms:\n            img1, img2 = t(img1, img2)\n        return img1, img2\n\n\nclass NormalizeBoth(torchvision.transforms.Normalize):\n    def forward(self, img1, img2):\n        img1 = super().forward(img1)","source_hash":"e2ead70dbd3cb762930ea459b1ab3bab397a9f0ca3cb395f42400745532fc1d2","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/datasets/crops/extract_crops_from_images.py","uri":"program://Human3R/file/src/croco/datasets/crops/extract_crops_from_images.py","kind":"file","name":"src/croco/datasets/crops/extract_crops_from_images.py","path":"src/croco/datasets/crops/extract_crops_from_images.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# Extracting crops for pre-training\n# --------------------------------------------------------\n\nimport os\nimport argparse\nfrom tqdm import tqdm\nfrom PIL import Image\nimport functools\nfrom multiprocessing import Pool\nimport math\n\n\ndef arg_parser():\n    parser = argparse.ArgumentParser(\n        \"Generate cropped image pairs from image crop list\"\n    )\n","source_hash":"40d0deaa505957e33e29b690825ad05305136b4769615911ad01756c49e028cf","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/datasets/habitat_sim/paths.py","uri":"program://Human3R/file/src/croco/datasets/habitat_sim/paths.py","kind":"file","name":"src/croco/datasets/habitat_sim/paths.py","path":"src/croco/datasets/habitat_sim/paths.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\"\"\"\nPaths to Habitat-Sim scenes\n\"\"\"\n\nimport os\nimport json\nimport collections\nfrom tqdm import tqdm\n\n\n# Hardcoded path to the different scene datasets\nSCENES_DATASET = {\n    \"hm3d\": \"./data/habitat-sim-data/scene_datasets/hm3d/\",\n    \"gibson\": \"./data/habitat-sim-data/scene_datasets/gibson/\",\n    \"habitat-test-scenes\": \"./data/habitat-sim/scene_datasets/habitat-test-scenes/\",\n    \"replica_cad_baked_lighting\": \"./data/habitat-sim/scene_datasets/replica_cad_baked_lighting/\",\n    \"replica_cad\": \"./data/habitat-sim/scene_datasets/replica_cad/\",\n    \"replica\": \"./data/habitat-sim/scene_datasets/ReplicaDataset/\",","source_hash":"2c46a5b329717acf7032810cc9ed6a556007a2ba566ff7374849e5614a3ff37a","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/datasets/habitat_sim/generate_multiview_images.py","uri":"program://Human3R/file/src/croco/datasets/habitat_sim/generate_multiview_images.py","kind":"file","name":"src/croco/datasets/habitat_sim/generate_multiview_images.py","path":"src/croco/datasets/habitat_sim/generate_multiview_images.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nfrom tqdm import tqdm\nimport argparse\nimport PIL.Image\nimport numpy as np\nimport json\nfrom datasets.habitat_sim.multiview_habitat_sim_generator import (\n    MultiviewHabitatSimGenerator,\n    NoNaviguableSpaceError,\n)\nfrom datasets.habitat_sim.paths import list_scenes_available\nimport cv2\nimport quaternion\nimport shutil\n\n\ndef generate_multiview_images_for_scene(\n    scene_dataset_config_file,","source_hash":"f1f29c975905b469f700a147e5acbd9462535dcf6798a041df74d8f91d6498ea","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","uri":"program://Human3R/file/src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","kind":"file","name":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","path":"src/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\nimport os\nimport numpy as np\nimport quaternion\nimport habitat_sim\nimport json\nfrom sklearn.neighbors import NearestNeighbors\nimport cv2\n\n# OpenCV to habitat camera convention transformation\nR_OPENCV2HABITAT = np.stack(\n    (habitat_sim.geo.RIGHT, -habitat_sim.geo.UP, habitat_sim.geo.FRONT), axis=0\n)\nR_HABITAT2OPENCV = R_OPENCV2HABITAT.T\nDEG2RAD = np.pi / 180\n\n\ndef compute_camera_intrinsics(height, width, hfov):\n    f = width / 2 / np.tan(hfov / 2 * np.pi / 180)","source_hash":"e3ac040aa682c5b0181b2820551b018e7e8dbe1b992d511a2a597461ccf782aa","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/datasets/habitat_sim/pack_metadata_files.py","uri":"program://Human3R/file/src/croco/datasets/habitat_sim/pack_metadata_files.py","kind":"file","name":"src/croco/datasets/habitat_sim/pack_metadata_files.py","path":"src/croco/datasets/habitat_sim/pack_metadata_files.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\"\"\"\nUtility script to pack metadata files of the dataset in order to be able to re-generate it elsewhere.\n\"\"\"\nimport os\nimport glob\nfrom tqdm import tqdm\nimport shutil\nimport json\nfrom datasets.habitat_sim.paths import *\nimport argparse\nimport collections\n\nif __name__ == \"__main__\":\n    parser = argparse.ArgumentParser()\n    parser.add_argument(\"input_dir\")\n    parser.add_argument(\"output_dir\")\n    args = parser.parse_args()\n\n    input_dirname = args.input_dir","source_hash":"229db302fb8332cbe6635e5f2f05976fe5610332b4d27e24898290f4b5f85c33","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/datasets/habitat_sim/generate_from_metadata_files.py","uri":"program://Human3R/file/src/croco/datasets/habitat_sim/generate_from_metadata_files.py","kind":"file","name":"src/croco/datasets/habitat_sim/generate_from_metadata_files.py","path":"src/croco/datasets/habitat_sim/generate_from_metadata_files.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\"\"\"\nScript generating commandlines to generate image pairs from metadata files.\n\"\"\"\nimport os\nimport glob\nfrom tqdm import tqdm\nimport argparse\n\nif __name__ == \"__main__\":\n    parser = argparse.ArgumentParser()\n    parser.add_argument(\"--input_dir\", required=True)\n    parser.add_argument(\"--output_dir\", required=True)\n    parser.add_argument(\n        \"--prefix\",\n        default=\"\",\n        help=\"Commanline prefix, useful e.g. to setup environment.\",\n    )\n    args = parser.parse_args()","source_hash":"256b884779bccbbf23b8f17af2003a93da03cbf4288ab8361399ec3e0b0cb944","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/croco/datasets/habitat_sim/generate_from_metadata.py","uri":"program://Human3R/file/src/croco/datasets/habitat_sim/generate_from_metadata.py","kind":"file","name":"src/croco/datasets/habitat_sim/generate_from_metadata.py","path":"src/croco/datasets/habitat_sim/generate_from_metadata.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n\n\"\"\"\nScript to generate image pairs for a given scene reproducing poses provided in a metadata file.\n\"\"\"\nimport os\nfrom datasets.habitat_sim.multiview_habitat_sim_generator import (\n    MultiviewHabitatSimGenerator,\n)\nfrom datasets.habitat_sim.paths import SCENES_DATASET\nimport argparse\nimport quaternion\nimport PIL.Image\nimport cv2\nimport json\nfrom tqdm import tqdm\n\n\ndef generate_multiview_images_from_metadata(\n    metadata_filename,","source_hash":"e1f03764fc02c17b46a0194c5751d5d582ebdb1f4e00c8e89c6e84d55ffb27c3","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/blocks.py","uri":"program://Human3R/file/src/dust3r/blocks.py","kind":"file","name":"src/dust3r/blocks.py","path":"src/dust3r/blocks.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn as nn\n\nfrom itertools import repeat\nimport collections.abc\nfrom torch.nn.functional import scaled_dot_product_attention\nfrom functools import partial\n\n\ndef _ntuple(n):\n    def parse(x):\n        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):\n            return x\n        return tuple(repeat(x, n))\n","source_hash":"0d17a8b47db9398e3fc027c99a73c291ef4d61a02cba4c7b608102973ee7f253","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/losses.py","uri":"program://Human3R/file/src/dust3r/losses.py","kind":"file","name":"src/dust3r/losses.py","path":"src/dust3r/losses.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"from copy import copy, deepcopy\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom dust3r.utils.geometry import (\n    inv,\n    geotrf,\n    normalize_pointcloud_group,\n    get_group_pointcloud_center_scale,\n    to_euclidean_dist,\n)\nimport numpy as np\nfrom dust3r.utils.camera import (\n    pose_encoding_to_camera,\n    camera_to_pose_encoding,\n    relative_pose_absT_quatR,\n)\nfrom dust3r.utils.image import unpad_image\nfrom dust3r.utils import SMPL_Layer\n","source_hash":"212216eb52d000b7be1b1b95001e437c928a54fe15a86dab121ea09a83f81829","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/smpl_model.py","uri":"program://Human3R/file/src/dust3r/smpl_model.py","kind":"file","name":"src/dust3r/smpl_model.py","path":"src/dust3r/smpl_model.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# modified from Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nimport numpy as np\nimport smplx\nfrom smplx.joint_names import JOINT_NAMES\nfrom dust3r.utils.geometry import (\n    perspective_projection, \n    resize_camera_intrinsics,\n    get_camera_parameters\n)\nfrom dust3r.utils.image import pad_image\nimport roma\nimport pickle\nimport os\n\ncurrent_dir = os.path.dirname(os.path.abspath(__file__))\nsrc_dir = os.path.dirname(current_dir)\nSMPLX_DIR = os.path.join(src_dir, 'models')","source_hash":"8d80005cba608f086e4b862436ebeec8697aeee5faa48afcd69a3fdfbeb730ce","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/inference.py","uri":"program://Human3R/file/src/dust3r/inference.py","kind":"file","name":"src/dust3r/inference.py","path":"src/dust3r/inference.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import tqdm\nimport torch\nfrom dust3r.utils.device import to_cpu, collate_with_cat\nfrom dust3r.utils.misc import invalid_to_nans\nfrom dust3r.utils.geometry import depthmap_to_pts3d, geotrf\nfrom dust3r.model import ARCroco3DStereo\nfrom dust3r.smpl_model import SMPLModel\nfrom accelerate import Accelerator\nimport re\n\n\ndef custom_sort_key(key):\n    text = key.split(\"/\")\n    if len(text) > 1:\n        text, num = text[0], text[-1]\n        return (text, int(num))\n    else:\n        return (key, -1)\n\n\ndef merge_chunk_dict(old_dict, curr_dict, add_number):","source_hash":"1e097b766f9665ca0e9bb9ba069bc06bb8f8f4a8a857ddecb762357bd9ce9b86","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/patch_embed.py","uri":"program://Human3R/file/src/dust3r/patch_embed.py","kind":"file","name":"src/dust3r/patch_embed.py","path":"src/dust3r/patch_embed.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport dust3r.utils.path_to_croco  # noqa: F401\nfrom models.blocks import PatchEmbed  # noqa\n\n\ndef get_patch_embed(patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3):\n    assert patch_embed_cls in [\"PatchEmbedDust3R\", \"ManyAR_PatchEmbed\"]\n    patch_embed = eval(patch_embed_cls)(img_size, patch_size, in_chans, enc_embed_dim)\n    return patch_embed\n\n\nclass PatchEmbedDust3R(PatchEmbed):\n    def forward(self, x, **kw):\n        B, C, H, W = x.shape\n        assert (","source_hash":"c0dae36e876d2124f0e31403a3a306db7e6d5bd69d8e9c6f16bbfb81c82eef3d","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/model.py","uri":"program://Human3R/file/src/dust3r/model.py","kind":"file","name":"src/dust3r/model.py","path":"src/dust3r/model.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import sys\nimport os\n\nsys.path.append(os.path.dirname(os.path.dirname(__file__)))\nfrom collections import OrderedDict\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.checkpoint import checkpoint\nfrom copy import deepcopy\nfrom functools import partial\nfrom typing import Optional, Tuple, List, Any\nfrom dataclasses import dataclass\nfrom transformers import PretrainedConfig\nfrom transformers import PreTrainedModel\nfrom transformers.modeling_outputs import BaseModelOutput\nfrom transformers.file_utils import ModelOutput\nimport time\nfrom dust3r.utils.misc import (\n    fill_default_args,\n    freeze_all_params,","source_hash":"8721e924da7d99cfa78e046e82630d63b6b2b18a8ee4df29d6b755bee3d43e9b","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/viz.py","uri":"program://Human3R/file/src/dust3r/viz.py","kind":"file","name":"src/dust3r/viz.py","path":"src/dust3r/viz.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport PIL.Image\nimport numpy as np\nfrom scipy.spatial.transform import Rotation\nimport torch\nimport cv2\nimport matplotlib as mpl\nimport matplotlib.cm as cm\nimport matplotlib.pyplot as plt\nfrom dust3r.utils.geometry import (\n    geotrf,\n    get_med_dist_between_poses,\n    depthmap_to_absolute_camera_coordinates,\n)\nfrom dust3r.utils.device import to_numpy\nfrom dust3r.utils.image import rgb, img_to_arr","source_hash":"ef736c54ff88a04474b10becf0da754a6deab94a803d5d338aeba91722ea8915","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/post_process.py","uri":"program://Human3R/file/src/dust3r/post_process.py","kind":"file","name":"src/dust3r/post_process.py","path":"src/dust3r/post_process.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nimport torch\nfrom dust3r.utils.geometry import xy_grid\n\n\ndef estimate_focal_knowing_depth(\n    pts3d, pp, focal_mode=\"median\", min_focal=0.0, max_focal=np.inf\n):\n    \"\"\"Reprojection method, for when the absolute depth is known:\n    1) estimate the camera focal using a robust estimator\n    2) reproject points onto true rays, minimizing a certain error\n    \"\"\"\n    B, H, W, THREE = pts3d.shape\n    assert THREE == 3\n","source_hash":"e2ef308ee09ae547f64c3e8e5314cc6dd60a5ff37b2e548812af53f8f1ad1f37","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/device.py","uri":"program://Human3R/file/src/dust3r/utils/device.py","kind":"file","name":"src/dust3r/utils/device.py","path":"src/dust3r/utils/device.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nimport torch\n\n\ndef todevice(batch, device, callback=None, non_blocking=False):\n    \"\"\"Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy).\n\n    batch: list, tuple, dict of tensors or other things\n    device: pytorch device or 'numpy'\n    callback: function that would be called on every sub-elements.\n    \"\"\"\n    if callback:\n        batch = callback(batch)\n\n    if isinstance(batch, dict):","source_hash":"d6eed957d20ec85bcb142727736b58f5c034322d6fc610964924f0b95d3e977c","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/misc.py","uri":"program://Human3R/file/src/dust3r/utils/misc.py","kind":"file","name":"src/dust3r/utils/misc.py","path":"src/dust3r/utils/misc.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\n\n\ndef fill_default_args(kwargs, func):\n    import inspect  # a bit hacky but it works reliably\n\n    signature = inspect.signature(func)\n\n    for k, v in signature.parameters.items():\n        if v.default is inspect.Parameter.empty:\n            continue\n        kwargs.setdefault(k, v.default)\n\n    return kwargs\n","source_hash":"19af638d6626e7ae0be97cbe997af80b0e833fbbf9554fc0e268cd06df32ddf8","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/camera.py","uri":"program://Human3R/file/src/dust3r/utils/camera.py","kind":"file","name":"src/dust3r/utils/camera.py","path":"src/dust3r/utils/camera.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"from typing import Optional\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom croco.models.blocks import Mlp\nfrom dust3r.heads.postprocess import postprocess_pose\n\ninf = float(\"inf\")\n\n\nclass PoseDecoder(nn.Module):\n    def __init__(\n        self,\n        hidden_size=768,\n        mlp_ratio=4,\n        pose_encoding_type=\"absT_quaR\",\n    ):\n        super().__init__()\n","source_hash":"f18a3ecc54492ae2878f285ad756fb1d26dbc0c4146bcf76d3f7ed2665c35d73","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/render.py","uri":"program://Human3R/file/src/dust3r/utils/render.py","kind":"file","name":"src/dust3r/utils/render.py","path":"src/dust3r/utils/render.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os\nos.environ['PYOPENGL_PLATFORM'] = 'egl'\n\nimport torch\nfrom gsplat import rasterization\nfrom dust3r.utils.geometry import inv, geotrf\nfrom dust3r.utils.image import unpad_image\nimport numpy as np\ntry:\n    import pyrender\nexcept:\n    import pyrender\n\nimport trimesh\nfrom PIL import Image\n\ndef render(\n    intrinsics: torch.Tensor,\n    pts3d: torch.Tensor,\n    rgbs: torch.Tensor | None = None,\n    scale: float = 0.002,","source_hash":"90a8f9d4510f97514fa7c7a7883b7d20b06567f8d036b66bba6dd74d06532bbe","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/smpl_layer.py","uri":"program://Human3R/file/src/dust3r/utils/smpl_layer.py","kind":"file","name":"src/dust3r/utils/smpl_layer.py","path":"src/dust3r/utils/smpl_layer.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Multi-HMR\n# Copyright (c) 2024-present NAVER Corp.\n# CC BY-NC-SA 4.0 license\n\nimport torch\nfrom torch import nn\nimport smplx\nimport torch\nfrom dust3r.utils.geometry import inverse_perspective_projection, perspective_projection\nimport roma\nfrom dust3r.smpl_model import SMPLX_DIR\nfrom smplx.joint_names import JOINT_NAMES\n\nclass SMPL_Layer(nn.Module):\n    \"\"\"\n    Extension of the SMPL Layer with information about the camera for (inverse) projection the camera plane.\n    \"\"\"\n    def __init__(self, \n                 type='smplx', \n                 gender='neutral', \n                 num_betas=10,","source_hash":"60e448f46c122d3463d5e52753ff8f19c232dff613351f87cb7d5fbdf603ffeb","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/__init__.py","uri":"program://Human3R/file/src/dust3r/utils/__init__.py","kind":"file","name":"src/dust3r/utils/__init__.py","path":"src/dust3r/utils/__init__.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":5,"code":"from .image import normalize_rgb, denormalize_rgb\n\nfrom .render import render_meshes, vis_heatmap, OPENCV_TO_OPENGL_CAMERA_CONVENTION\n\nfrom .smpl_layer import SMPL_Layer","source_hash":"1488593a748b82adfd66c71ee813beccc91c07f323d1f9668dae3700bd4652bc","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/path_to_croco.py","uri":"program://Human3R/file/src/dust3r/utils/path_to_croco.py","kind":"file","name":"src/dust3r/utils/path_to_croco.py","path":"src/dust3r/utils/path_to_croco.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport sys\nimport os.path as path\n\nHERE_PATH = path.normpath(path.dirname(__file__))\nCROCO_REPO_PATH = path.normpath(path.join(HERE_PATH, \"../../croco\"))\nCROCO_MODELS_PATH = path.join(CROCO_REPO_PATH, \"models\")\n\nif path.isdir(CROCO_MODELS_PATH):\n\n    sys.path.insert(0, CROCO_REPO_PATH)\nelse:\n    raise ImportError(\n        f\"croco is not initialized, could not find: {CROCO_MODELS_PATH}.\\n \"\n        \"Did you forget to run 'git submodule update --init --recursive' ?\"\n    )","source_hash":"3d8c68320432fd7e3527f6bb94390f2a44195957ee5460539c1d807faf35367e","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/geometry.py","uri":"program://Human3R/file/src/dust3r/utils/geometry.py","kind":"file","name":"src/dust3r/utils/geometry.py","path":"src/dust3r/utils/geometry.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport numpy as np\nfrom scipy.spatial import cKDTree as KDTree\n\nfrom dust3r.utils.misc import invalid_to_zeros, invalid_to_nans\nfrom dust3r.utils.device import to_numpy\n\n\ndef xy_grid(\n    W,\n    H,\n    device=None,\n    origin=(0, 0),\n    unsqueeze=None,\n    cat_dim=-1,","source_hash":"f0c00b21eca5fade7426f98871613e336952e49f42b25445508379a07bafd8b3","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/parallel.py","uri":"program://Human3R/file/src/dust3r/utils/parallel.py","kind":"file","name":"src/dust3r/utils/parallel.py","path":"src/dust3r/utils/parallel.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nfrom tqdm import tqdm\nfrom multiprocessing.dummy import Pool as ThreadPool\nfrom multiprocessing import cpu_count\n\n\ndef parallel_threads(\n    function,\n    args,\n    workers=0,\n    star_args=False,\n    kw_args=False,\n    front_num=1,\n    Pool=ThreadPool,\n    **tqdm_kw\n):","source_hash":"8af1ba7e2e52a7daf1af51a955b98f7b5de3e24e70e66ab5dd3555c8df48aefb","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/utils/image.py","uri":"program://Human3R/file/src/dust3r/utils/image.py","kind":"file","name":"src/dust3r/utils/image.py","path":"src/dust3r/utils/image.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport os\nimport torch\nimport numpy as np\nimport PIL.Image\nfrom PIL.ImageOps import exif_transpose\nimport torchvision.transforms as tvf\n\nos.environ[\"OPENCV_IO_ENABLE_OPENEXR\"] = \"1\"\nimport cv2  # noqa\n\ntry:\n    from pillow_heif import register_heif_opener  # noqa\n\n    register_heif_opener()\n    heif_support_enabled = True","source_hash":"eb8ee2ea11cee81674abdb5ae5f4b9bec138009e635773a98fba533984e2cbd3","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/mvimgnet.py","uri":"program://Human3R/file/src/dust3r/datasets/mvimgnet.py","kind":"file","name":"src/dust3r/datasets/mvimgnet.py","path":"src/dust3r/datasets/mvimgnet.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVImgNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 32\n        super().__init__(*args, **kwargs)\n","source_hash":"a75252d8106f05fe95cf409028aba31d93c3a7b1dce68b0b3d4bbdd6c0833c95","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/urbansyn.py","uri":"program://Human3R/file/src/dust3r/datasets/urbansyn.py","kind":"file","name":"src/dust3r/datasets/urbansyn.py","path":"src/dust3r/datasets/urbansyn.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass UrbanSyn(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n","source_hash":"1f5f40b60354c47212c541c812dd9e041a5c442d92bd1c61fd85262a60702b22","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/spring.py","uri":"program://Human3R/file/src/dust3r/datasets/spring.py","kind":"file","name":"src/dust3r/datasets/spring.py","path":"src/dust3r/datasets/spring.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Spring(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n","source_hash":"98f3c2bb4565d2eff0cf26b79f50759cfe2e138117b4dae20ff675297ef4a088","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/uasol.py","uri":"program://Human3R/file/src/dust3r/datasets/uasol.py","kind":"file","name":"src/dust3r/datasets/uasol.py","path":"src/dust3r/datasets/uasol.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n","source_hash":"f729162f30ae3c3895e0f11af5cbd767d7a2236a5554455d31fd914706a1b17d","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/bedlam.py","uri":"program://Human3R/file/src/dust3r/datasets/bedlam.py","kind":"file","name":"src/dust3r/datasets/bedlam.py","path":"src/dust3r/datasets/bedlam.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport numpy as np\nimport os\nimport sys\nimport pickle\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\ninvalid_seqs = [\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000042\",\n    \"20221024_10_100_batch01handhair_zoom_suburb_d_seq_000059\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000079\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000978\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000081\",\n    \"20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000268\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000089\",\n    \"20221013_3_250_batch01hand_orbit_bigOffice_seq_000189\",\n    \"20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000034\",\n    \"20221019_3-8_1000_highbmihand_static_suburb_d_seq_000889\",","source_hash":"813bf1fcea87d92249abdcd53500ae1fd82ea9b86bf2a9bc98d3a9c1ca4b29f0","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/realestate10k.py","uri":"program://Human3R/file/src/dust3r/datasets/realestate10k.py","kind":"file","name":"src/dust3r/datasets/realestate10k.py","path":"src/dust3r/datasets/realestate10k.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass RE10K_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 128\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()","source_hash":"d56877dac5303ffef371832f370bdc4e24c13d8fe45cec26c645e7707d470747","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/threedkb.py","uri":"program://Human3R/file/src/dust3r/datasets/threedkb.py","kind":"file","name":"src/dust3r/datasets/threedkb.py","path":"src/dust3r/datasets/threedkb.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ThreeDKenBurns(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n","source_hash":"5e16195ffff7cf350365c24fb986de0c2f6c73050243f3a5c131ae9c27b4b0eb","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/co3d.py","uri":"program://Human3R/file/src/dust3r/datasets/co3d.py","kind":"file","name":"src/dust3r/datasets/co3d.py","path":"src/dust3r/datasets/co3d.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport json\nimport itertools\nfrom collections import deque\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\nimport time\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Co3d_Multi(BaseMultiViewDataset):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        super().__init__(*args, **kwargs)\n        assert mask_bg in (True, False, \"rand\")\n        self.mask_bg = mask_bg","source_hash":"9eaf54792831c38cdae571e0cd7f73c7b46140c1747de84f76e7b86acb5db33e","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/hypersim.py","uri":"program://Human3R/file/src/dust3r/datasets/hypersim.py","kind":"file","name":"src/dust3r/datasets/hypersim.py","path":"src/dust3r/datasets/hypersim.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HyperSim_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n","source_hash":"43d4851eab94fb053c0d919adca3d534f7ac04532b35396ddb3adbf974cfda74","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/arkitscenes_highres.py","uri":"program://Human3R/file/src/dust3r/datasets/arkitscenes_highres.py","kind":"file","name":"src/dust3r/datasets/arkitscenes_highres.py","path":"src/dust3r/datasets/arkitscenes_highres.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\nimport h5py\nimport math\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ARKitScenesHighRes_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 8\n        self.is_metric = True\n        super().__init__(*args, **kwargs)","source_hash":"f79e8f322b84b677ad3ac4f2c2a30de694db17110ad601a2da641bfd1592b14c","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/mapfree.py","uri":"program://Human3R/file/src/dust3r/datasets/mapfree.py","kind":"file","name":"src/dust3r/datasets/mapfree.py","path":"src/dust3r/datasets/mapfree.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nimport pickle\nimport h5py\nfrom tqdm import tqdm\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MapFree_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT","source_hash":"fbc4d12bb9f2e06915a7065115ac3a623c5368febabaa61ba1a6ff704625e137","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/synscapes.py","uri":"program://Human3R/file/src/dust3r/datasets/synscapes.py","kind":"file","name":"src/dust3r/datasets/synscapes.py","path":"src/dust3r/datasets/synscapes.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SynScapes(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n","source_hash":"2c2e047055922150a529b6b5b8b41ad47bd66810fbd6d7806051aa3e05b0f03b","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/pointodyssey.py","uri":"program://Human3R/file/src/dust3r/datasets/pointodyssey.py","kind":"file","name":"src/dust3r/datasets/pointodyssey.py","path":"src/dust3r/datasets/pointodyssey.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass PointOdyssey_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n        assert self.split in [\"train\", \"test\", \"val\"]","source_hash":"5ead728f8a1a89568d38f8a7ca8c1423c9cb41194ebbf01a5271ddfa9854dd9d","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/tartanair.py","uri":"program://Human3R/file/src/dust3r/datasets/tartanair.py","kind":"file","name":"src/dust3r/datasets/tartanair.py","path":"src/dust3r/datasets/tartanair.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass TartanAir_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 20","source_hash":"351cf02bfb408b6fe9c9aa1481dcdfb0f31c260bfe11dd796c8d028cb42ad56e","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/__init__.py","uri":"program://Human3R/file/src/dust3r/datasets/__init__.py","kind":"file","name":"src/dust3r/datasets/__init__.py","path":"src/dust3r/datasets/__init__.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"from .utils.transforms import *\nfrom .base.batched_sampler import BatchedRandomSampler  # noqa\nfrom .arkitscenes import ARKitScenes_Multi  # noqa\nfrom .arkitscenes_highres import ARKitScenesHighRes_Multi\nfrom .bedlam import BEDLAM_Multi\nfrom .blendedmvs import BlendedMVS_Multi  # noqa\nfrom .co3d import Co3d_Multi  # noqa\nfrom .cop3d import Cop3D_Multi\nfrom .dl3dv import DL3DV_Multi\nfrom .dynamic_replica import DynamicReplica\nfrom .eden import EDEN_Multi\nfrom .hypersim import HyperSim_Multi\nfrom .hoi4d import HOI4D_Multi\nfrom .irs import IRS\nfrom .mapfree import MapFree_Multi\nfrom .megadepth import MegaDepth_Multi  # noqa\nfrom .mp3d import MP3D_Multi\nfrom .mvimgnet import MVImgNet_Multi\nfrom .mvs_synth import MVS_Synth_Multi\nfrom .omniobject3d import OmniObject3D_Multi\nfrom .pointodyssey import PointOdyssey_Multi","source_hash":"44d163045a71f2c769262f07efada2266648f7232f40ca7d20129297c127d83a","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/unreal4k.py","uri":"program://Human3R/file/src/dust3r/datasets/unreal4k.py","kind":"file","name":"src/dust3r/datasets/unreal4k.py","path":"src/dust3r/datasets/unreal4k.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\nR_conv = np.array([[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]).astype(\n    np.float32\n)\n\n\nclass UnReal4K_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):","source_hash":"1b5528a7239f01051125b7378898ff95fc0b7b7813894d237c9124e38e69f0ce","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/smartportraits.py","uri":"program://Human3R/file/src/dust3r/datasets/smartportraits.py","kind":"file","name":"src/dust3r/datasets/smartportraits.py","path":"src/dust3r/datasets/smartportraits.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass SmartPortraits_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n","source_hash":"6acc2076c1e4b6866dfada9c3fb9848333e2f83012c98fd955253d8dac19092b","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/hoi4d.py","uri":"program://Human3R/file/src/dust3r/datasets/hoi4d.py","kind":"file","name":"src/dust3r/datasets/hoi4d.py","path":"src/dust3r/datasets/hoi4d.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nsys.path.append(osp.join(osp.dirname(__file__), '..','..'))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass HOI4D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n\n    def _load_data(self):","source_hash":"471296779fbd64093c0aee3d7f279d1e02e1366abb07afef990621489b72e3f5","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/wildrgbd.py","uri":"program://Human3R/file/src/dust3r/datasets/wildrgbd.py","kind":"file","name":"src/dust3r/datasets/wildrgbd.py","path":"src/dust3r/datasets/wildrgbd.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass WildRGBD_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"WildRGBD\"\n        self.is_metric = True\n        # load all scenes\n        self.scenes.pop((\"box\", \"scenes/scene_257\"), None)\n        self.scene_list = list(self.scenes.keys())\n        cut_off = (\n            self.num_views if not self.allow_repeat else max(self.num_views // 3, 3)","source_hash":"0ab18544678679ec0be20bbe08a43abcd20e2f7cd6cdd712fd50f45ed1fed206","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/vkitti2.py","uri":"program://Human3R/file/src/dust3r/datasets/vkitti2.py","kind":"file","name":"src/dust3r/datasets/vkitti2.py","path":"src/dust3r/datasets/vkitti2.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport numpy as np\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass VirtualKITTI2_Multi(BaseMultiViewDataset):\n\n    def __init__(self, ROOT, *args, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 5","source_hash":"1a1d9ca9ee4628332411dadbb333eb4b6f131d681ade6f4aa15100a8e2b58706","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/waymo.py","uri":"program://Human3R/file/src/dust3r/datasets/waymo.py","kind":"file","name":"src/dust3r/datasets/waymo.py","path":"src/dust3r/datasets/waymo.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport os\nimport numpy as np\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport h5py\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Waymo_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.max_interval = 8\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        assert self.split is None","source_hash":"c4b92fdb09e48cd8a65e7d20cd6e3f68686f262d74b443dc37823999c18f67c3","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/mp3d.py","uri":"program://Human3R/file/src/dust3r/datasets/mp3d.py","kind":"file","name":"src/dust3r/datasets/mp3d.py","path":"src/dust3r/datasets/mp3d.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MP3D_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n\n        self.loaded_data = self._load_data()","source_hash":"3e20428e776f42c93423aa2aabfdbd174febbe6a92c9470d83afd3688050258b","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/megadepth.py","uri":"program://Human3R/file/src/dust3r/datasets/megadepth.py","kind":"file","name":"src/dust3r/datasets/megadepth.py","path":"src/dust3r/datasets/megadepth.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MegaDepth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        super().__init__(*args, **kwargs)\n        self._load_data(self.split)\n        self.is_metric = False\n        if self.split is None:\n            pass\n        elif self.split == \"train\":\n            self.select_scene((\"0015\", \"0022\"), opposite=True)","source_hash":"ef2736d369ccf3fb3b75739934b58297c64fe33eaf4a725060217482e63fc336","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/dl3dv.py","uri":"program://Human3R/file/src/dust3r/datasets/dl3dv.py","kind":"file","name":"src/dust3r/datasets/dl3dv.py","path":"src/dust3r/datasets/dl3dv.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DL3DV_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, split, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.max_interval = 20\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n","source_hash":"a6df42342e9cf8ec382750f73d79bb007161e1bf5a7dbd4845f22291feafb2e1","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/mvs_synth.py","uri":"program://Human3R/file/src/dust3r/datasets/mvs_synth.py","kind":"file","name":"src/dust3r/datasets/mvs_synth.py","path":"src/dust3r/datasets/mvs_synth.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass MVS_Synth_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = False\n        self.max_interval = 4\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()","source_hash":"c87b10ff5331934399cb4c883a051b6432c67c316edab551daee3998ab8d36f3","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/eden.py","uri":"program://Human3R/file/src/dust3r/datasets/eden.py","kind":"file","name":"src/dust3r/datasets/eden.py","path":"src/dust3r/datasets/eden.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass EDEN_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n","source_hash":"67146a2ea020d46842392a85258e5848bc14f0821b47a7b6b5ba083db06cb4ff","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/arkitscenes.py","uri":"program://Human3R/file/src/dust3r/datasets/arkitscenes.py","kind":"file","name":"src/dust3r/datasets/arkitscenes.py","path":"src/dust3r/datasets/arkitscenes.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport os\nimport sys\nimport itertools\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\ndef stratified_sampling(indices, num_samples, rng=None):\n    if num_samples > len(indices):\n        raise ValueError(\"num_samples cannot exceed the number of available indices.\")\n    elif num_samples == len(indices):\n        return indices\n\n    sorted_indices = sorted(indices)\n    stride = len(sorted_indices) / num_samples","source_hash":"190b16bd1d3f8f27e36812fdbecfba6b979028b589cc3a5dab5afe430f181efd","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/scannetpp.py","uri":"program://Human3R/file/src/dust3r/datasets/scannetpp.py","kind":"file","name":"src/dust3r/datasets/scannetpp.py","path":"src/dust3r/datasets/scannetpp.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\n\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNetpp_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 3\n        super().__init__(*args, **kwargs)\n        assert self.split == \"train\"","source_hash":"88b2400c9d9fe0e4d50364112194362c6bb1cda3decdaabc24ab12e74943c2c3","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/irs.py","uri":"program://Human3R/file/src/dust3r/datasets/irs.py","kind":"file","name":"src/dust3r/datasets/irs.py","path":"src/dust3r/datasets/irs.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass IRS(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = True\n        super().__init__(*args, **kwargs)\n        self.loaded_data = self._load_data()\n","source_hash":"f6ffc7162a1dde1510b8500150f045f8244bdf033b105f8395f093c74818169d","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/blendedmvs.py","uri":"program://Human3R/file/src/dust3r/datasets/blendedmvs.py","kind":"file","name":"src/dust3r/datasets/blendedmvs.py","path":"src/dust3r/datasets/blendedmvs.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport numpy as np\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport h5py\nfrom tqdm import tqdm\n\n\nclass BlendedMVS_Multi(BaseMultiViewDataset):\n    \"\"\"Dataset of outdoor street scenes, 5 images each time\"\"\"\n\n    def __init__(self, *args, ROOT, split=None, **kwargs):\n        self.ROOT = ROOT\n        self.video = False\n        self.is_metric = False\n        super().__init__(*args, **kwargs)\n        # assert split is None","source_hash":"c3e62991f5aae1bd1b7134ff000e3e6bce266305f7d894e2a396f1e84f289cd1","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/omniobject3d.py","uri":"program://Human3R/file/src/dust3r/datasets/omniobject3d.py","kind":"file","name":"src/dust3r/datasets/omniobject3d.py","path":"src/dust3r/datasets/omniobject3d.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\nimport json\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\nimport re\n\n\ndef extract_number(filename):\n    match = re.search(r\"\\d+\", filename)\n    if match:\n        return int(match.group())\n    return 0\n","source_hash":"0f11e6da03b9fe41e86f40fe1281e1ad5e89ec050643e731f0176236c6089282","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/scannet.py","uri":"program://Human3R/file/src/dust3r/datasets/scannet.py","kind":"file","name":"src/dust3r/datasets/scannet.py","path":"src/dust3r/datasets/scannet.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass ScanNet_Multi(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 30\n        super().__init__(*args, **kwargs)\n","source_hash":"89f549daf57cbe821d15f7050b7c9fba0ebc4e813aa4ca436245edda8a6f8a9b","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/cop3d.py","uri":"program://Human3R/file/src/dust3r/datasets/cop3d.py","kind":"file","name":"src/dust3r/datasets/cop3d.py","path":"src/dust3r/datasets/cop3d.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nimport cv2\nimport numpy as np\n\nfrom dust3r.datasets.co3d import Co3d_Multi\nfrom dust3r.utils.image import imread_cv2\n\n\nclass Cop3D_Multi(Co3d_Multi):\n    def __init__(self, mask_bg=\"rand\", *args, ROOT, **kwargs):\n        super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs)\n        self.dataset_label = \"Cop3D\"\n        self.is_metric = False\n\n    def _get_metadatapath(self, obj, instance, view_idx):\n        return osp.join(self.ROOT, obj, instance, \"images\", f\"frame{view_idx:06n}.npz\")\n\n    def _get_impath(self, obj, instance, view_idx):","source_hash":"011da64ae58ff38ab993c38860a465a5eef35ad83c08cdd3f079b6540acd485d","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/dynamic_replica.py","uri":"program://Human3R/file/src/dust3r/datasets/dynamic_replica.py","kind":"file","name":"src/dust3r/datasets/dynamic_replica.py","path":"src/dust3r/datasets/dynamic_replica.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import os.path as osp\nimport cv2\nimport numpy as np\nimport itertools\nimport os\nimport sys\n\nsys.path.append(osp.join(osp.dirname(__file__), \"..\", \"..\"))\nfrom tqdm import tqdm\nfrom dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset\nfrom dust3r.utils.image import imread_cv2\n\n\nclass DynamicReplica(BaseMultiViewDataset):\n    def __init__(self, *args, ROOT, **kwargs):\n        self.ROOT = ROOT\n        self.video = True\n        self.is_metric = True\n        self.max_interval = 16\n        super().__init__(*args, **kwargs)\n","source_hash":"c2fa93cf842e3af15ccbc843f14fe72e4be88741e1200ec637cc18bddcc974a3","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/utils/corr.py","uri":"program://Human3R/file/src/dust3r/datasets/utils/corr.py","kind":"file","name":"src/dust3r/datasets/utils/corr.py","path":"src/dust3r/datasets/utils/corr.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nfrom dust3r.utils.device import to_numpy\nfrom dust3r.utils.geometry import inv, geotrf\n\n\ndef reproject_view(pts3d, view2):\n    shape = view2[\"pts3d\"].shape[:2]\n    return reproject(\n        pts3d, view2[\"camera_intrinsics\"], inv(view2[\"camera_pose\"]), shape\n    )\n\n\ndef reproject(pts3d, K, world2cam, shape):\n    H, W, THREE = pts3d.shape\n    assert THREE == 3","source_hash":"6195285c9dc0c477a03c0d7624a068d0b3dd52386da09aa9a74e1c4c466bdef8","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/utils/cropping.py","uri":"program://Human3R/file/src/dust3r/datasets/utils/cropping.py","kind":"file","name":"src/dust3r/datasets/utils/cropping.py","path":"src/dust3r/datasets/utils/cropping.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# croppping utilities\n# --------------------------------------------------------\nimport PIL.Image\nimport os\n\nos.environ[\"OPENCV_IO_ENABLE_OPENEXR\"] = \"1\"\nimport cv2  # noqa\nimport numpy as np  # noqa\nfrom dust3r.utils.geometry import (\n    colmap_to_opencv_intrinsics,\n    opencv_to_colmap_intrinsics,\n)  # noqa\n\ntry:\n    lanczos = PIL.Image.Resampling.LANCZOS\n    bicubic = PIL.Image.Resampling.BICUBIC\nexcept AttributeError:","source_hash":"ab8892a252cbd895a9c5b56b6d08d134abd6a585bf7c7a79e4f094b8eeedde3e","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/utils/__init__.py","uri":"program://Human3R/file/src/dust3r/datasets/utils/__init__.py","kind":"file","name":"src/dust3r/datasets/utils/__init__.py","path":"src/dust3r/datasets/utils/__init__.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":2,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).","source_hash":"b75338cd2bc70ca0d9dac551d4727a0e11a8098d15a85487691a538d0284c473","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/utils/transforms.py","uri":"program://Human3R/file/src/dust3r/datasets/utils/transforms.py","kind":"file","name":"src/dust3r/datasets/utils/transforms.py","path":"src/dust3r/datasets/utils/transforms.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# DUST3R default transforms\n# --------------------------------------------------------\nimport torchvision.transforms as tvf\nfrom dust3r.utils.image import ImgNorm\n\n# define the standard image transforms\nColorJitter = tvf.Compose([tvf.ColorJitter(0.5, 0.5, 0.5, 0.1), ImgNorm])\n\n\ndef _check_input(value, center=1, bound=(0, float(\"inf\")), clip_first_on_zero=True):\n    if isinstance(value, (int, float)):\n        if value < 0:\n            raise ValueError(f\"If  is a single number, it must be non negative.\")\n        value = [center - float(value), center + float(value)]\n        if clip_first_on_zero:\n            value[0] = max(value[0], 0.0)\n    elif isinstance(value, (tuple, list)) and len(value) == 2:","source_hash":"10308410a74d5b9ce263b77f0b94418bad3deaa7f1a3e0c330216538ba07d0d6","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/base/batched_sampler.py","uri":"program://Human3R/file/src/dust3r/datasets/base/batched_sampler.py","kind":"file","name":"src/dust3r/datasets/base/batched_sampler.py","path":"src/dust3r/datasets/base/batched_sampler.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import numpy as np\nimport torch\nfrom accelerate import Accelerator\nimport torch.utils\nfrom torch.utils.data import BatchSampler, Sampler\nimport torch.utils.data\n\n\nclass CustomRandomSampler(Sampler):\n    \"\"\"Random sampling under a constraint: each sample in the batch has the same feature,\n    which is chosen randomly from a known pool of 'features' for each batch.\n\n    For instance, the 'feature' could be the image aspect-ratio.\n\n    The index returned is a tuple (sample_idx, feat_idx).\n    This sampler ensures that each series of `batch_size` indices has the same `feat_idx`.\n    \"\"\"\n\n    def __init__(\n        self,\n        dataset,","source_hash":"bb63eb9b59cf6e180e044b6a187b54034aa8ae66718c8d7665d7e97d3a6dfae9","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/base/base_multiview_dataset.py","uri":"program://Human3R/file/src/dust3r/datasets/base/base_multiview_dataset.py","kind":"file","name":"src/dust3r/datasets/base/base_multiview_dataset.py","path":"src/dust3r/datasets/base/base_multiview_dataset.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"import PIL\nimport numpy as np\nimport torch\nimport random\nimport itertools\nfrom dust3r.datasets.base.easy_dataset import EasyDataset\nfrom dust3r.datasets.utils.transforms import ImgNorm, SeqColorJitter\nfrom dust3r.utils.geometry import depthmap_to_absolute_camera_coordinates\nimport dust3r.datasets.utils.cropping as cropping\nfrom dust3r.datasets.utils.corr import extract_correspondences_from_pts3d\nimport torchvision.transforms as tvf\n\ndef get_ray_map(c2w1, c2w2, intrinsics, h, w):\n    c2w = np.linalg.inv(c2w1) @ c2w2\n    i, j = np.meshgrid(np.arange(w), np.arange(h), indexing=\"xy\")\n    grid = np.stack([i, j, np.ones_like(i)], axis=-1)\n    ro = c2w[:3, 3]\n    rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T\n    rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3)\n    rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True)\n    ro = np.broadcast_to(ro, (h, w, 3))","source_hash":"d72fe7d7e7c68bbe92ebb540d6c00508fb922d57635200a6d6e5e1ae1b47c6f2","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/datasets/base/easy_dataset.py","uri":"program://Human3R/file/src/dust3r/datasets/base/easy_dataset.py","kind":"file","name":"src/dust3r/datasets/base/easy_dataset.py","path":"src/dust3r/datasets/base/easy_dataset.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport numpy as np\nfrom dust3r.datasets.base.batched_sampler import (\n    BatchedRandomSampler,\n    CustomRandomSampler,\n)\nimport torch\n\n\nclass EasyDataset:\n    \"\"\"a dataset that you can easily resize and combine.\n    Examples:\n    ---------\n        2 * dataset ==> duplicate each element 2x\n\n        10 @ dataset ==> set the size to 10 (random sampling, duplicates if necessary)","source_hash":"408ff24c50c42b40eccacf9839b65b5631ebb1a79c2d19c4b83acc0832a727da","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/heads/dpt_head.py","uri":"program://Human3R/file/src/dust3r/heads/dpt_head.py","kind":"file","name":"src/dust3r/heads/dpt_head.py","path":"src/dust3r/heads/dpt_head.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nfrom einops import rearrange\nfrom typing import List\nimport torch\nimport torch.nn as nn\nfrom dust3r.heads.postprocess import (\n    postprocess,\n    postprocess_desc,\n    postprocess_rgb,\n    postprocess_pose_conf,\n    postprocess_pose,\n    reg_dense_conf,\n    postprocess_smpl,\n    postprocess_score,\n)\nimport dust3r.utils.path_to_croco  # noqa: F401","source_hash":"502143f1a42c7ee9816e10a757c9eefcbd3d1c4e9a9e3561287d144824730665","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/heads/postprocess.py","uri":"program://Human3R/file/src/dust3r/heads/postprocess.py","kind":"file","name":"src/dust3r/heads/postprocess.py","path":"src/dust3r/heads/postprocess.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn.functional as F\nimport roma\n\ndef postprocess(out, depth_mode, conf_mode, pos_z=False):\n    \"\"\"\n    extract 3D points/confidence from prediction head output\n    \"\"\"\n    fmap = out.permute(0, 2, 3, 1)  # B,H,W,3\n    res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode, pos_z=pos_z))\n\n    if conf_mode is not None:\n        res[\"conf\"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode)\n    return res\n","source_hash":"2d58bfe5e8657102951e0e8296b46351e810baa67f79af3e3b3489522e6739e3","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/heads/linear_head.py","uri":"program://Human3R/file/src/dust3r/heads/linear_head.py","kind":"file","name":"src/dust3r/heads/linear_head.py","path":"src/dust3r/heads/linear_head.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom dust3r.heads.postprocess import (\n    postprocess,\n    postprocess_desc,\n    postprocess_rgb,\n    postprocess_pose_conf,\n    postprocess_pose,\n    reg_dense_conf,\n)\nimport dust3r.utils.path_to_croco  # noqa\nfrom models.blocks import Mlp  # noqa\nfrom dust3r.utils.geometry import geotrf\nfrom dust3r.utils.camera import pose_encoding_to_camera, PoseDecoder","source_hash":"c52a465d11c1b8ee9145a379e8801682fe1db5ac9f076e5a55de364ec9efa424","truncated":false}
{"repo_id":"Human3R","entity_id":"file:src/dust3r/heads/__init__.py","uri":"program://Human3R/file/src/dust3r/heads/__init__.py","kind":"file","name":"src/dust3r/heads/__init__.py","path":"src/dust3r/heads/__init__.py","language":"python","start_line":1,"end_line":1,"context_start_line":1,"context_end_line":21,"code":"# Copyright (C) 2024-present Naver Corporation. All rights reserved.\n# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).\n#\n# --------------------------------------------------------\n# modified from DUSt3R\n\nfrom .linear_head import LinearPts3d, LinearPts3d_Desc, LinearPts3dPose\nfrom .dpt_head import DPTPts3dPose, DPTPts3dPoseSMPL, NaiveDPTPts3dPoseSMPL\n\n\ndef head_factory(\n    head_type,\n    output_mode,\n    net,\n    has_conf=False,\n    has_depth=False,\n    has_rgb=False,\n    has_pose_conf=False,\n    has_pose=False,\n    has_msk=False,\n):","source_hash":"1cf6d96243e44d334f72585fdab515361d7ee1450d1e448da99ddf4d468fefb8","truncated":false}