init utonia

.gitattributes

@@ -0,0 +1,28 @@
+ckpt/concerto_large.pth filter=lfs diff=lfs merge=lfs -text
+example filter=lfs diff=lfs merge=lfs -text
+example/pcd filter=lfs diff=lfs merge=lfs -text
+example/video filter=lfs diff=lfs merge=lfs -text
+example/pcd/hm3d_00012_kDgLKdMd5X8_1.ply filter=lfs diff=lfs merge=lfs -text
+example/pcd/hm3d_00012_kDgLKdMd5X8_2.ply filter=lfs diff=lfs merge=lfs -text
+example/pcd/s3dis_Area2_conferenceRoom1.png filter=lfs diff=lfs merge=lfs -text
+example/pcd/scannet_0603.png filter=lfs diff=lfs merge=lfs -text
+example/pcd/hm3d_00012_kDgLKdMd5X8_1.png filter=lfs diff=lfs merge=lfs -text
+example/pcd/scannet_0024.png filter=lfs diff=lfs merge=lfs -text
+example/pcd/s3dis_Area4_lobby1.png filter=lfs diff=lfs merge=lfs -text
+example/pcd/scannet_0024.ply filter=lfs diff=lfs merge=lfs -text
+example/pcd/hm3d_00012_kDgLKdMd5X8_2.png filter=lfs diff=lfs merge=lfs -text
+example/pcd/hm3d_00113_3goH1WRaCYC.ply filter=lfs diff=lfs merge=lfs -text
+example/pcd/hm3d_00113_3goH1WRaCYC.png filter=lfs diff=lfs merge=lfs -text
+example/pcd/s3dis_Area2_auditorium1.ply filter=lfs diff=lfs merge=lfs -text
+example/pcd/s3dis_Area2_auditorium1.png filter=lfs diff=lfs merge=lfs -text
+example/pcd/scannet_0603.ply filter=lfs diff=lfs merge=lfs -text
+example/pcd/s3dis_Area2_conferenceRoom1.ply filter=lfs diff=lfs merge=lfs -text
+example/pcd/s3dis_Area4_lobby1.ply filter=lfs diff=lfs merge=lfs -text
+example/pcd/scannetpp_2a1b555966.ply filter=lfs diff=lfs merge=lfs -text
+example/pcd/scannetpp_2a1b555966.png filter=lfs diff=lfs merge=lfs -text
+example/video/re10k_2.mp4 filter=lfs diff=lfs merge=lfs -text
+example/video/re10k_3.mp4 filter=lfs diff=lfs merge=lfs -text
+example/video/re10k_4.mp4 filter=lfs diff=lfs merge=lfs -text
+example/video/conference_room.mp4 filter=lfs diff=lfs merge=lfs -text
+example/video/office.mp4 filter=lfs diff=lfs merge=lfs -text
+example/video/re10k_1.mp4 filter=lfs diff=lfs merge=lfs -text

.github/CODE_OF_CONDUCT.md

@@ -0,0 +1,80 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+  advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+  address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+This Code of Conduct also applies outside the project spaces when there is a
+reasonable belief that an individual's behavior may have a negative impact on
+the project or its community.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <opensource-conduct@meta.com>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq

.github/CONTRIBUTING.md

@@ -0,0 +1,35 @@
+# Contributing to "sonata"
+
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+
+We welcome pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation in the code.
+4. Ensure the test suite passes.
+5. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Facebook's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+Facebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## License
+
+By contributing to "sonata", you agree that your contributions will be licensed under
+the [LICENSE](../LICENSE) file in the root directory of this source tree.

.gitignore

@@ -0,0 +1,18 @@
+image/
+__pycache__
+**/build/
+**/*.egg-info/
+**/dist/
+*.so
+exp
+weights
+data
+log
+**/ckpt/
+outputs/
+.vscode
+.idea
+*/.DS_Store
+**/*.out
+vggt/ckpt
+**/demo_output*

.gradio/certificate.pem

@@ -0,0 +1,31 @@
+-----BEGIN CERTIFICATE-----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=
+-----END CERTIFICATE-----

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -0,0 +1,11 @@
+---
+title: Utonia
+emoji: ✨
+colorFrom: gray
+colorTo: indigo
+sdk: gradio
+sdk_version: 6.1.0
+app_file: app.py
+pinned: false
+python_version: "3.10"
+---

app.py

@@ -0,0 +1,989 @@
+import gradio as gr
+import os
+import numpy as np
+import trimesh
+import torch
+import time
+import spaces
+import cv2
+import shutil
+from datetime import datetime
+import glob
+from einops import rearrange
+
+# Local imports
+from geometry_utils import (
+    Coord2zup,
+    extract_and_align_ground_plane,
+    pad_0001,
+    T_to_C,
+    im_distance_to_im_depth,
+    im_depth_to_point_cloud,
+)
+# VGGT specific imports
+from vggt.utils.load_fn import load_and_preprocess_images
+from vggt.utils.pose_enc import pose_encoding_to_extri_intri
+from vggt.utils.geometry import unproject_depth_map_to_point_map
+from vggt.models.vggt import VGGT
+
+import utonia
+utonia.utils.set_seed(53124)
+utonia_model = utonia.load("utonia", repo_id="Pointcept/Utonia")
+
+VGGT_model = VGGT()
+_URL = "https://huggingface.co/facebook/VGGT-1B/resolve/main/model.pt"
+VGGT_model.load_state_dict(torch.hub.load_state_dict_from_url(_URL))
+
+@spaces.GPU
+def _gpu_run_vggt_inference(images_tensor):
+    """
+    GPU-only function: Run VGGT model inference on preprocessed images.
+    Minimizes GPU time by only doing model inference and pose encoding conversion.
+    """
+    global VGGT_model
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    # Move images to GPU
+    images_tensor = images_tensor.to(device)
+    model = VGGT_model.to(device)
+    model.eval()
+
+    print("Running inference...")
+    with torch.no_grad():
+        if device == "cuda":
+            with torch.cuda.amp.autocast(dtype=torch.bfloat16):
+                predictions = model(images_tensor)
+        else:
+            predictions = model(images_tensor)
+
+    # Convert pose encoding to extrinsic and intrinsic matrices (GPU operation)
+    print("Converting pose encoding to extrinsic and intrinsic matrices...")
+    extrinsic, intrinsic = pose_encoding_to_extri_intri(predictions["pose_enc"], images_tensor.shape[-2:])
+    predictions["extrinsic"] = extrinsic
+    predictions["intrinsic"] = intrinsic
+
+    # Convert to numpy (still on GPU to minimize memory transfer)
+    for key in predictions.keys():
+        if isinstance(predictions[key], torch.Tensor):
+            predictions[key] = predictions[key].cpu().numpy().squeeze(0)
+
+    torch.cuda.empty_cache()
+    return predictions
+
+def run_model(target_dir) -> dict:
+    """
+    CPU-GPU hybrid: Handle CPU-intensive file I/O and call GPU function for inference.
+    """
+    print(f"Processing images from {target_dir}")
+
+    # Load and preprocess images (CPU)
+    image_names = glob.glob(os.path.join(target_dir, "images", "*"))
+    image_names = sorted(image_names)
+    print(f"Found {len(image_names)} images")
+    if len(image_names) == 0:
+        raise ValueError("No images found. Check your upload.")
+
+    images = load_and_preprocess_images(image_names)
+    print(f"Preprocessed images shape: {images.shape}")
+
+    # Call GPU function for inference
+    predictions = _gpu_run_vggt_inference(images)
+
+    # Post-processing (CPU)
+    print("Computing world points from depth map...")
+    depth_map = predictions["depth"]  # (S, H, W, 1)
+    world_points = unproject_depth_map_to_point_map(depth_map, predictions["extrinsic"], predictions["intrinsic"])
+    predictions["world_points_from_depth"] = world_points
+
+    return predictions
+
+def parse_frames(
+    target_dir,
+    conf_thres=3.0,
+    prediction_mode="Pointmap Regression",
+):
+    """
+    Perform reconstruction using the already-created target_dir/images.
+    """
+    if not os.path.isdir(target_dir) or target_dir == "None":
+        return None, "No valid target directory found. Please upload first.", None, None
+
+    start_time = time.time()
+
+    # Prepare frame_filter dropdown
+    target_dir_images = os.path.join(target_dir, "images")
+    all_files = sorted(os.listdir(target_dir_images)) if os.path.isdir(target_dir_images) else []
+    all_files = [f"{i}: {filename}" for i, filename in enumerate(all_files)]
+
+    print("Running run_model...")
+    with torch.no_grad():
+        predictions = run_model(target_dir)
+
+    # Save predictions
+    prediction_save_path = os.path.join(target_dir, "predictions.npz")
+    np.savez(prediction_save_path, **predictions)
+
+    # Convert pose encoding to extrinsic and intrinsic matrices
+    images = predictions["images"]
+    Ts, Ks = predictions["extrinsic"],predictions["intrinsic"]
+    Ts = pad_0001(Ts)
+    Ts_inv = np.linalg.inv(Ts)
+    Cs = np.array([T_to_C(T) for T in Ts])  # (n, 3)
+
+    # [1, 8, 294, 518, 3]
+    world_points = predictions["world_points"]
+
+    # Compute view direction for each pixel
+    # (b n h w c) - (n, 3)
+    view_dirs = world_points - rearrange(Cs, "n c -> n 1 1 c")
+    view_dirs = rearrange(view_dirs, "n h w c -> (n h w) c")
+    view_dirs = view_dirs / np.linalg.norm(view_dirs, axis=-1, keepdims=True)
+
+    # Extract points and colors
+    # [1, 8, 3, 294, 518]
+    img_num = world_points.shape[1]
+    images = predictions["images"]
+    points = rearrange(world_points, "n h w c -> (n h w) c")
+    colors = rearrange(images, "n c h w -> (n h w) c")
+    normals = np.zeros_like(points)
+
+    if prediction_mode=="Pointmap Branch":
+        world_points_conf = predictions["world_points_conf"]
+        conf = world_points_conf.reshape(-1)
+        points,Ts_inv,_ = Coord2zup(points, Ts_inv)
+        scale = 3 / (points[:, 2].max() - points[:, 2].min())
+        points *= scale
+        Ts_inv[:, :3, 3] *= scale
+
+        normals = -np.asarray(view_dirs)
+        normals = normals / np.clip(np.linalg.norm(normals, axis=-1, keepdims=True), 1e-8, None)
+        if conf_thres == 0.0:
+            conf_threshold = 0.0
+        else:
+            conf_threshold = np.percentile(conf, conf_thres)
+        conf_mask = (conf >= conf_threshold) & (conf > 1e-5)
+        points = points[conf_mask]
+        colors = colors[conf_mask]
+        normals = normals[conf_mask]
+
+        try:
+            points, colors, normals, _, _, _ = extract_and_align_ground_plane(
+                points=points,
+                colors=colors,
+                normals=normals,
+            )
+        except Exception as e:
+            print(f"cannot find ground, err:{e}")
+    elif prediction_mode=="Depthmap Branch":
+        # Integrate RGBD images into a TSDF volume and extract a mesh
+        # (n, h, w, 3)
+        im_colors = rearrange(images, "n c h w -> (n) h w c")
+        # (b, n, h, w, 3)
+        im_dists = world_points - rearrange(Cs, "n c -> n 1 1 c")
+        im_dists = np.linalg.norm(im_dists, axis=-1, keepdims=False)
+
+        # Convert distance to depth
+        im_depths = []  # (n, h, w, c)
+        for im_dist, K in zip(im_dists, Ks):
+            im_depth = im_distance_to_im_depth(im_dist, K)
+            im_depths.append(im_depth)
+        im_depths = np.stack(im_depths, axis=0)
+        points=[]
+        for K, T, im_depth in zip(Ks, Ts, im_depths):
+            point = im_depth_to_point_cloud(
+                im_depth=im_depth,
+                K=K,
+                T=T,
+                to_image=False,
+                ignore_invalid=False,
+            )
+            points.append(point)
+        points = np.vstack(points)
+        colors = im_colors.reshape(-1,3)
+        world_points_conf = predictions["depth_conf"]
+        conf = world_points_conf.reshape(-1)
+        if conf_thres == 0.0:
+            conf_threshold = 0.0
+        else:
+            conf_threshold = np.percentile(conf, conf_thres)
+        conf_mask = (conf >= conf_threshold) & (conf > 1e-5)
+        points = points[conf_mask]
+        colors = colors[conf_mask]
+        points,Ts_inv,_ = Coord2zup(points, Ts_inv)
+        scale_factor = 3./(np.max(points[:,2])-np.min(points[:,2]))
+        points *= scale_factor
+        Ts_inv[:, :3, 3] *= scale_factor
+        normals = np.zeros_like(points)
+        try:
+            points, colors, normals, _, _, _ = extract_and_align_ground_plane(
+                points=points,
+                colors=colors,
+                normals=normals,
+            )
+        except Exception as e:
+            print(f"cannot find ground, err:{e}")
+    original_points = np.asarray(points)
+    original_colors = np.asarray(colors)
+    original_normals = np.asarray(normals)
+    # Cleanup
+    del predictions
+    end_time = time.time()
+    print(f"Total time: {end_time - start_time:.2f} seconds")
+    return original_points, original_colors, original_normals
+
+def handle_uploads(input_file,input_video,conf_thres,frame_slider,prediction_mode):
+    """
+    Create a new 'target_dir' + 'images' subfolder, and place user-uploaded
+    images or extracted frames from video into it. Return (target_dir, image_paths).
+    """
+    start_time = time.time()
+
+    # Create a unique folder name
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
+    target_dir = f"demo_output/inputs_{timestamp}"
+    target_dir_images = os.path.join(target_dir, "images")
+    target_dir_pcds = os.path.join(target_dir, "pcds")
+
+    # Clean up if somehow that folder already exists
+    if os.path.exists(target_dir):
+        shutil.rmtree(target_dir)
+    os.makedirs(target_dir)
+    os.makedirs(target_dir_images)
+    os.makedirs(target_dir_pcds)
+    # Handle video
+    if input_video is not None:
+        print("processing video")
+        if isinstance(input_video, dict) and "name" in input_video:
+            video_path = input_video["name"]
+        else:
+            video_path = input_video
+
+        vs = cv2.VideoCapture(video_path)
+        fps = vs.get(cv2.CAP_PROP_FPS)
+        frame_interval = int(fps * frame_slider)  # 1 frame/sec
+
+        count = 0
+        video_frame_num = 0
+        image_paths = []
+        while True:
+            gotit, frame = vs.read()
+            if not gotit:
+                break
+            count += 1
+            if count % frame_interval == 0:
+                image_path = os.path.join(target_dir_images, f"{video_frame_num:06}.png")
+                cv2.imwrite(image_path, frame)
+                image_paths.append(image_path)
+                video_frame_num += 1
+        # Sort final images for gallery
+        image_paths = sorted(image_paths)
+        original_points, original_colors, original_normals = parse_frames(target_dir,conf_thres,prediction_mode)
+    if input_file is not None:
+        print("processing ply")
+        loaded = load_point_from_file(input_file)
+        if loaded is None:
+            raise ValueError("Failed to load input point cloud file")
+        original_points = loaded["coord"]
+        original_colors = loaded["color"]
+        original_normals = loaded["normal"]
+        image_paths = None
+    scene_3d = trimesh.Scene()
+    point_cloud_data = trimesh.PointCloud(vertices=original_points, colors=original_colors, vertex_normals=original_normals)
+    scene_3d.add_geometry(point_cloud_data)
+    original_temp = os.path.join(target_dir_pcds,"original.glb")
+    scene_3d.export(file_obj=original_temp)
+    np.save(os.path.join(target_dir_pcds, f"points.npy"), original_points)
+    np.save(os.path.join(target_dir_pcds, f"colors.npy"), original_colors)
+    np.save(os.path.join(target_dir_pcds, f"normals.npy"), original_normals)
+    end_time = time.time()
+    print(f"Files copied to {target_dir}; took {end_time - start_time:.3f} seconds")
+    return target_dir, image_paths,original_temp, end_time - start_time
+
+
+def load_point_from_file(input_file):
+    if input_file is None:
+        return None
+
+    file_path = input_file
+    if hasattr(input_file, "name"):
+        file_path = input_file.name
+    elif isinstance(input_file, dict) and "name" in input_file:
+        file_path = input_file["name"]
+
+    if not file_path:
+        return None
+
+    geometry = trimesh.load(file_path, process=False)
+    if isinstance(geometry, trimesh.Scene):
+        geometries = [g for g in geometry.geometry.values()]
+        if not geometries:
+            return None
+        geometry = geometries[0]
+
+    if isinstance(geometry, trimesh.PointCloud):
+        coord = np.asarray(geometry.vertices)
+        color = np.asarray(geometry.colors[:, :3]) if geometry.colors is not None and len(geometry.colors) else np.zeros_like(coord)
+        normal = np.zeros_like(coord)
+        if color.dtype != np.float32 and color.dtype != np.float64:
+            color = color.astype(np.float32) / 255.0
+        return {"coord": coord, "color": color, "normal": normal}
+
+    if isinstance(geometry, trimesh.Trimesh):
+        coord = np.asarray(geometry.vertices)
+        if geometry.visual is not None and hasattr(geometry.visual, "vertex_colors") and geometry.visual.vertex_colors is not None and len(geometry.visual.vertex_colors):
+            color = np.asarray(geometry.visual.vertex_colors[:, :3]).astype(np.float32) / 255.0
+        else:
+            color = np.zeros_like(coord)
+        normal = np.asarray(geometry.vertex_normals) if geometry.vertex_normals is not None and len(geometry.vertex_normals) else np.zeros_like(coord)
+        return {"coord": coord, "color": color, "normal": normal}
+
+    return None
+
+def update_gallery_on_upload(input_file,input_video,conf_thres,frame_slider,prediction_mode):
+    """
+    Whenever user uploads or changes files, immediately handle them
+    and show in the gallery. Return (target_dir, image_paths).
+    If nothing is uploaded, returns "None" and empty list.
+    """
+    if not input_video and not input_file:
+        return None, None, None, None
+    target_dir, image_paths,original_view, reconstruction_time = handle_uploads(input_file,input_video,conf_thres,frame_slider,prediction_mode)
+    if input_file is not None:
+        return original_view, target_dir, [], f"Upload and preprocess complete with {reconstruction_time:.3f} sec. Click \"PCA Generate\" to begin PCA processing."
+    if input_video is not None:
+        return original_view, target_dir, image_paths, f"Upload and preprocess complete with {reconstruction_time:.3f} sec. Click \"PCA Generate\" to begin PCA processing."
+
+
+def get_pca_color(feat, start = 0, brightness=1.25, center=True):
+    u, s, v = torch.pca_lowrank(feat, center=center, q=3*(start+1), niter=5)
+    projection = feat @ v
+    projection = projection[:, 3*start:3*(start+1)] * 0.6 + projection[:, 3*start:3*(start+1)] * 0.4
+    min_val = projection.min(dim=-2, keepdim=True)[0]
+    max_val = projection.max(dim=-2, keepdim=True)[0]
+    div = torch.clamp(max_val - min_val, min=1e-6)
+    color = (projection - min_val) / div * brightness
+    color = color.clamp(0.0, 1.0)
+    return color
+
+def clear_fields():
+    """
+    Clears the 3D viewer, the stored target_dir, and empties the gallery.
+    """
+    return None
+
+def PCAing_log(is_example, log_output):
+    """
+    Display a quick log message while waiting.
+    """
+    if is_example:
+        return log_output
+    return "Loading for Doing PCA..."
+
+def reset_log():
+    """
+    Reset a quick log message.
+    """
+    return "A new point cloud file or video is uploading and preprocessing..."
+
+
+@spaces.GPU
+def _gpu_utonia_forward_pca(point, utonia_model_, pca_slider, bright_slider):
+    """
+    GPU-only function: Run Utonia model forward pass and PCA in one place.
+    Uses inference_mode overall with a scoped disable for the forward call.
+    """
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+
+    # Move tensors and model to GPU
+    for key in point.keys():
+        if isinstance(point[key], torch.Tensor):
+            point[key] = point[key].to(device, non_blocking=True)
+
+    utonia_model_ = utonia_model_.to(device)
+    utonia_model_.eval()
+
+    with torch.inference_mode():
+        utonia_start_time = time.time()
+        # Disable inference_mode for model forward to avoid version counter issues
+        with torch.inference_mode(False):
+            point = utonia_model_(point)
+        utonia_end_time = time.time()
+
+        # Upcast point feature through hierarchical pooling
+        for _ in range(4):
+            assert "pooling_parent" in point.keys()
+            assert "pooling_inverse" in point.keys()
+            parent = point.pop("pooling_parent")
+            inverse = point.pop("pooling_inverse")
+            parent.feat = torch.cat([parent.feat, point.feat[inverse]], dim=-1)
+            point = parent
+        while "pooling_parent" in point.keys():
+            assert "pooling_inverse" in point.keys()
+            parent = point.pop("pooling_parent")
+            inverse = point.pop("pooling_inverse")
+            parent.feat = point.feat[inverse]
+            point = parent
+
+        pca_start_time = time.time()
+        pca_color = get_pca_color(point.feat, start=pca_slider, brightness=bright_slider, center=True)
+        pca_end_time = time.time()
+
+        # Inverse back to original scale
+        original_pca_color = pca_color[point.inverse]
+
+    processed_colors = original_pca_color.cpu().detach().numpy()
+    point_feat = point.feat.cpu().detach().numpy()
+    point_inverse = point.inverse.cpu().detach().numpy()
+    utonia_time = utonia_end_time - utonia_start_time
+    pca_time = pca_end_time - pca_start_time
+    return processed_colors, point_feat, point_inverse, utonia_time, pca_time
+
+def gradio_demo(target_dir, pca_slider, bright_slider, if_color=True, if_normal=True, scale_value=1.0, apply_z_positive=True, normalize_coord=False):
+    global utonia_model
+    target_dir_pcds = os.path.join(target_dir, "pcds")
+    if not os.path.isfile(os.path.join(target_dir_pcds, "points.npy")):
+        return None, "No point cloud available. Please upload data first."
+    
+    # CPU: Load point cloud data from disk
+    original_points = np.load(os.path.join(target_dir_pcds, "points.npy"))
+    if if_color:
+        original_colors = np.load(os.path.join(target_dir_pcds, "colors.npy"))
+    else:
+        original_colors = np.zeros_like(original_points)
+    if if_normal:
+        original_normals = np.load(os.path.join(target_dir_pcds, "normals.npy"))
+    else:
+        original_normals = np.zeros_like(original_points)
+    
+    processed_temp = os.path.join(target_dir_pcds, "processed.glb")
+
+    point = {"coord": original_points, "color": original_colors, "normal": original_normals}
+    original_coord = point["coord"].copy()
+
+    # CPU: Apply transform pipeline
+    transform = utonia.transform.default(scale=scale_value, apply_z_positive=apply_z_positive, normalize_coord=normalize_coord)
+    point = transform(point)
+
+    # GPU: Run Utonia forward + PCA together (inference_mode inside GPU function)
+    processed_colors, point_feat, point_inverse_cpu, utonia_time, pca_time = _gpu_utonia_forward_pca(
+        point, utonia_model, pca_slider, bright_slider
+    )
+
+    # CPU: Save features
+    np.save(os.path.join(target_dir_pcds, "feat.npy"), point_feat)
+    np.save(os.path.join(target_dir_pcds, "inverse.npy"), point_inverse_cpu)
+
+    # CPU: Build and save the 3D mesh
+    processed_points = original_coord
+    feat_3d = trimesh.Scene()
+    feat_data = trimesh.PointCloud(vertices=processed_points, colors=processed_colors, vertex_normals=original_normals)
+    feat_3d.add_geometry(feat_data)
+    feat_3d.export(processed_temp)
+
+    return processed_temp, f"Feature visualization process finished with {utonia_time:.3f} seconds using utonia inference and {pca_time:.3f} seconds using PCA. Updating visualization."
+
+@spaces.GPU
+def _gpu_pca_slider_compute(feat_array, inverse_array, pca_slider, bright_slider):
+    """
+    GPU-only function: Compute PCA colors for slider updates.
+    Minimal GPU allocation for only the essential computation.
+    """
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    # Move data to GPU inside GPU function
+    feat_tensor = torch.tensor(feat_array, device=device)
+    inverse_tensor = torch.tensor(inverse_array, device=device)
+    
+    pca_start_time = time.time()
+    pca_colors = get_pca_color(feat_tensor, start=pca_slider, brightness=bright_slider, center=True)
+    processed_colors = pca_colors[inverse_tensor].cpu().detach().numpy()
+    pca_end_time = time.time()
+    return processed_colors, (pca_end_time - pca_start_time)
+
+def utonia_slider_update(target_dir, pca_slider, bright_slider, is_example, log_output):
+    """
+    CPU-GPU hybrid: Handle file I/O on CPU, GPU for PCA computation only.
+    """
+    if is_example == "True":
+        return None, log_output
+    else:
+        target_dir_pcds = os.path.join(target_dir, "pcds")
+        if os.path.isfile(os.path.join(target_dir_pcds, "feat.npy")):
+            # CPU: Load data from disk
+            feat = np.load(os.path.join(target_dir_pcds, "feat.npy"))
+            inverse = np.load(os.path.join(target_dir_pcds, "inverse.npy"))
+            
+            # GPU: Compute PCA colors only (numpy arrays passed to GPU function)
+            processed_colors, pca_time = _gpu_pca_slider_compute(feat, inverse, pca_slider, bright_slider)
+            
+            # CPU: Load additional data and build mesh
+            processed_points = np.load(os.path.join(target_dir_pcds, "points.npy"))
+            processed_normals = np.load(os.path.join(target_dir_pcds, "normals.npy"))
+            processed_temp = os.path.join(target_dir_pcds, "processed.glb")
+            
+            feat_3d = trimesh.Scene()
+            feat_data = trimesh.PointCloud(vertices=processed_points, colors=processed_colors, vertex_normals=processed_normals)
+            feat_3d.add_geometry(feat_data)
+            feat_3d.export(processed_temp)
+            
+            log_output = f"Feature visualization process finished with {pca_time:.3f} seconds using PCA. Updating visualization."
+        else:
+            processed_temp = None
+            log_output = "No representations saved, please click PCA generate first."
+    return processed_temp, log_output
+
+BASE_URL = "https://huggingface.co/datasets/pointcept-bot/utonia_huggingface_demo/resolve/main/"
+def get_url(path):
+    return f"{BASE_URL}{path}"
+
+examples_object = [
+    [
+        get_url("object/0005df571e71437991594d0affec9c2b.png"),
+        get_url("object/0005df571e71437991594d0affec9c2b.ply"),
+        0, 1.2, "True", 1.0, True
+    ],
+    [
+        get_url("object/0023687e90394c3e97ab19b0160cafb3.png"),
+        get_url("object/0023687e90394c3e97ab19b0160cafb3.ply"),
+        0, 1.2, "True", 1.0, True
+    ],
+    [
+        get_url("object/0015eb3cf53b4339b2d0532cf912ab26.png"),
+        get_url("object/0015eb3cf53b4339b2d0532cf912ab26.ply"),
+        0, 1.2, "True", 1.0, True
+    ],
+    [
+        get_url("object/001a5201eddf4f3b98591598584673f5.png"),
+        get_url("object/001a5201eddf4f3b98591598584673f5.ply"),
+        0, 1.2, "True", 1.0, True
+    ],
+]
+
+examples_manipulation = [
+    [
+        get_url("manipulation/000021_AUTOLab_5d05c5aa_2023-11-17-23h-40m-52s-35_46.png"),
+        get_url("manipulation/000021_AUTOLab_5d05c5aa_2023-11-17-23h-40m-52s-35_46.ply"),
+        0, 1.2, "True", 4.0, False
+    ],
+    [
+        get_url("manipulation/000018_AUTOLab_44bb9c36_2023-11-23-20h-05m-45s-55_66.png"),
+        get_url("manipulation/000018_AUTOLab_44bb9c36_2023-11-23-20h-05m-45s-55_66.ply"),
+        1, 1.0, "True", 4.0, False
+    ],
+    [
+        get_url("manipulation/000037_IPRL_7790ec0a_2023-07-01-09h-37m-21s-15_26.png"),
+        get_url("manipulation/000037_IPRL_7790ec0a_2023-07-01-09h-37m-21s-15_26.ply"),
+        0, 1.2, "True", 4.0, False
+    ],
+    [
+        get_url("manipulation/000061_TRI_938130c4_2023-08-10-14h-40m-11s-70_81.png"),
+        get_url("manipulation/000061_TRI_938130c4_2023-08-10-14h-40m-11s-70_81.ply"),
+        2, 1.2, "True", 4.0, False
+    ],
+]
+
+examples_indoor = [
+    [
+        get_url("indoor/scene0024_00.png"),
+        get_url("indoor/scene0024_00.ply"),
+        0, 1.0, "True", 0.5, False
+    ],
+    [
+        get_url("indoor/scene0603_00.png"),
+        get_url("indoor/scene0603_00.ply"),
+        0, 1.0, "True", 0.5, False
+    ],
+    [
+        get_url("indoor/027cd6ea0f.png"),
+        get_url("indoor/027cd6ea0f.ply"),
+        0, 1.0, "True", 0.5, False
+    ],
+    [
+        get_url("indoor/2c7c10379b.png"),
+        get_url("indoor/2c7c10379b.ply"),
+        3, 1.0, "True", 0.5, False
+    ],
+]
+
+examples_outdoor = [
+    [
+        get_url("outdoor/segment-10455472356147194054_1560_000_1580_000_with_camera_labels.png"),
+        get_url("outdoor/segment-10455472356147194054_1560_000_1580_000_with_camera_labels.ply"),
+        1, 1.2, "True", 0.2, False
+    ],
+    [
+        get_url("outdoor/segment-10963653239323173269_1924_000_1944_000_with_camera_labels.png"),
+        get_url("outdoor/segment-10963653239323173269_1924_000_1944_000_with_camera_labels.ply"),
+        0, 1.2, "True", 0.2, False
+    ],
+    [
+        get_url("outdoor/segment-11718898130355901268_2300_000_2320_000_with_camera_labels.png"),
+        get_url("outdoor/segment-11718898130355901268_2300_000_2320_000_with_camera_labels.ply"),
+        0, 1.2, "True", 0.2, False
+    ],
+    [
+        get_url("outdoor/segment-11925224148023145510_1040_000_1060_000_with_camera_labels.png"),
+        get_url("outdoor/segment-11925224148023145510_1040_000_1060_000_with_camera_labels.ply"),
+        0, 1.2, "True", 0.2, False
+    ],
+]
+
+examples_video = [
+    [
+        get_url("video/re10k_1.mp4"),
+        10.0, 1, "Depthmap Branch", 2, 1.2, "True", 0.5, False
+    ],
+    [
+        get_url("video/re10k_2.mp4"),
+        20.0, 1, "Pointmap Branch", 2, 1.2, "True", 0.5, False
+    ],
+    [
+        get_url("video/re10k_3.mp4"),
+        10.0, 1, "Pointmap Branch", 1, 1.2, "True", 0.5, False
+    ],
+    [
+        get_url("video/re10k_4.mp4"),
+        10.0, 1, "Pointmap Branch", 1, 1., "True", 0.5, False
+    ],
+]
+
+def example_file_updated(
+    preview_imgs,
+    inputs,
+    pca_slider,
+    bright_slider,
+    is_example,
+    scale_slider,
+    normalize_coord,
+    url_input,
+):
+    pass
+
+def example_video_updated(
+    inputs,
+    conf_thres,
+    frame_slider,
+    prediction_mode,
+    pca_slider,
+    bright_slider,
+    is_example,
+    scale_slider,
+    normalize_coord,
+    url_input,
+    ):
+    pass
+
+with gr.Blocks(
+    css="""
+    .custom-log * {
+        font-style: italic;
+        font-size: 22px !important;
+        background-image: linear-gradient(120deg, #0ea5e9 0%, #6ee7b7 60%, #34d399 100%);
+        -webkit-background-clip: text;
+        background-clip: text;
+        font-weight: bold !important;
+        color: transparent !important;
+        text-align: center !important;
+        width: 800px;  
+        height: 100px; 
+    }
+    
+    .example-log * {
+        font-style: italic;
+        font-size: 16px !important;
+        background-image: linear-gradient(120deg, #0ea5e9 0%, #6ee7b7 60%, #34d399 100%);
+        -webkit-background-clip: text;
+        background-clip: text;
+        color: transparent !important;
+    }
+
+    .common-markdown * {
+        font-size: 22px !important;
+        -webkit-background-clip: text;
+        background-clip: text;
+        font-weight: bold !important;
+        color: #0ea5e9 !important;
+        text-align: center !important;
+    }
+
+    #big-box {
+        border: 3px solid #00bcd4; 
+        padding: 20px;            
+        background-color: transparent;  
+        border-radius: 15px;   
+    }
+    
+    #my_radio .wrap {
+        display: flex;
+        flex-wrap: nowrap;
+        justify-content: center;
+        align-items: center;
+    }
+
+    #my_radio .wrap label {
+        display: flex;
+        width: 50%;
+        justify-content: center;
+        align-items: center;
+        margin: 0;
+        padding: 10px 0;
+        box-sizing: border-box;
+    }
+    """,
+) as demo:
+    gr.HTML(
+    """
+    <h1>Utonia: Toward One Encoder for All Point Clouds</h1>
+    <div style="font-size: 16px; line-height: 1.5;">
+        <ol>
+            <details style="display:inline;">
+                <summary style="display:inline;"><h3>Getting Started:(<strong>Click to expand</strong>)</h3></summary>
+            <li><strong>Before Start: We recommend cloning this space to run locally on GPU</strong> for the limited GPU time.</li>
+            <li><strong>Upload Your Data:</strong> Use the "Upload Video" or "Upload Point Cloud" blocks on the left to provide your input. If you upload a video, it will be automatically split into individual frames with the specified frame gap by VGGT.</li>
+            <li>
+                <strong>[Optional] Adjust Video-Lifted Point Cloud:</strong>
+                Before reconstructing the video, you can fine-tune the VGGT lifting process using the options below
+                <details style="display:inline;">
+                    <summary style="display:inline;">(<strong>Click to expand</strong>)</summary>
+                    <ul>
+                        <li><em>Frame Gap / N Sec:</em> Adjust the frame interval.</li>
+                        <li><em>Confidence Threshold:</em> Adjust the point filtering based on confidence levels.</li>
+                        <li><em>Select Prediction Mode:</em> Choose between "Depthmap Branch" and "Pointmap Branch."</li>
+                    </ul>
+                </details>
+            </li>
+            <li><strong>PCA Generation:</strong> After reconstruction, click the "PCA Generate" button to start the representation extraction and PCA process.</li>
+            <li><strong>Clear:</strong> Click the "Clear" button to reset all content in the blocks.</li>
+            <li><strong>Point Cloud Preview:</strong> Your uploaded video or point cloud will be displayed in this block.</li>
+            <li><strong>PCA Result:</strong> The PCA point cloud will appear here. You can rotate, drag, and zoom to explore the model, and download the GLB file.</li>
+            <li>
+                <strong>[Optional] Adjust the Point Cloud Input:</strong>
+                <details style="display:inline;">
+                    <summary style="display:inline;">(<strong>Click to expand</strong>)</summary>
+                    <ul>
+                        <li><em>Input with Point Cloud Color/Normal:</em> If not checked, the corresponding information will be set to zeros.</li>
+                        <li><em>Z positive:</em> When enabled, the point cloud is transformed so that the Z-axis is strictly positive. This is the default recommendation for most inputs, except for outdoor scenes where maintaining the road at the XY plane is preferred.</li>
+                        <li><em>Normalize coord:</em> When enabled, the point cloud will be normalized to fit within the [-1,1] range before scaling. This is typically used for single objects.</li>
+                    </ul>
+                </details>
+            </li>
+            <li>
+                <strong>[Optional] Adjust PCA Visualization:</strong>
+                Fine-tune the PCA visualization using the options below
+                <details style="display:inline;">
+                    <summary style="display:inline;">(<strong>Click to expand</strong>)</summary>
+                    <ul>
+                        <li><em>PCA Start Dimension:</em> PCA reduces high-dimensional representations into 3D vectors. Adjust the PCA start dimension to change the range of the visualization. Increasing this value can help you see PCA visualization with less variance when the initial PCA dimension shows less diversity.</li>
+                        <li><em>PCA Brightness:</em> Adjust the brightness of the PCA visualization results.</li>
+                        <li><em>Notice:</em> As a linear dimension reduction method, PCA has its limitation. Sometimes, the visualization cannot fully exhibit the quality of representations.</li>
+                    </ul>
+                </details>
+            </li>
+            <li><strong>Adjust Scale Parameter:</strong> Adjust the scale of the point cloud. The default scale is 0.2 for outdoor, 0.5 for indoor, and 4.0 for manipulation with high resolution needs. Object data needs normalization. You can adjust it according to your desired granularity and memory limit.</li>
+            </details>
+        </ol>
+    </div>
+
+    """
+    )
+    _ = gr.Textbox(label="_", visible=False, value="False")
+    is_example = gr.Textbox(label="is_example", visible=False, value="False")
+    target_dir = gr.Textbox(label="Target Dir", visible=False, value="None")
+    preview_imgs = gr.Image(type="filepath",label="Preview Imgs", visible=False, value="None")
+    with gr.Row():
+        with gr.Column(scale=1,elem_id="big-box"):
+            input_file = gr.File(label="Upload Point Cloud", file_types=[".ply"])
+            input_video = gr.Video(label="Upload Video", interactive=True)
+            image_gallery = gr.Gallery(
+                label="Video Frame Preview",
+                columns=4,
+                height="300px",
+                # show_download_button=True,
+                object_fit="contain",
+                preview=True,
+            )
+
+            frame_slider = gr.Slider(minimum=0.1, maximum=10, value=1, step=0.1, 
+                             label="1 Frame/ N Sec", interactive=True)
+            conf_thres = gr.Slider(minimum=0, maximum=100, value=10, step=0.1, 
+                             label="Confidence", interactive=True)
+            prediction_mode = gr.Radio(
+                ["Depthmap Branch", "Pointmap Branch"],
+                label="Select a Prediction Mode",
+                value="Depthmap Branch",
+                scale=1,
+                elem_id="my_radio",
+            )
+            reconstruction_btn = gr.Button("Video Reconstruct")
+        with gr.Column(scale=2):
+            log_output = gr.Markdown(
+                "Please upload a video or point cloud ply file, then click \"PCA Generate\".", elem_classes=["custom-log"]
+            )
+            original_view = gr.Model3D(height=520, zoom_speed=0.5, pan_speed=0.5, label="Point Cloud Preview", camera_position = (90,None,None))
+            processed_view = gr.Model3D(height=520, zoom_speed=0.5, pan_speed=0.5, label="PCA Result", camera_position = (90,None,None))
+            with gr.Row():
+                if_color = gr.Checkbox(label="Input with point cloud color", value=True)
+                if_normal = gr.Checkbox(label="Input with point cloud normal", value=True)
+            with gr.Row():
+                normalize_coord = gr.Checkbox(label="Normalize coord", value=False)
+                apply_z_positive = gr.Checkbox(label="Z positive", value=True)
+            scale_slider = gr.Slider(minimum=0.001, maximum=5.0, value=1.0, step=0.0005, 
+                             label="Scale Parameter", interactive=True)
+            pca_slider = gr.Slider(minimum=0, maximum=5, value=0, step=1, 
+                             label="PCA Start Dimension", interactive=True)
+            bright_slider = gr.Slider(minimum=0.5, maximum=1.5, value=1.2, step=0.05, 
+                             label="PCA Brightness", interactive=True)
+            with gr.Row():
+                submit_btn = gr.Button("PCA Generate")
+                clear_btn = gr.ClearButton(
+                        [input_video, input_file, original_view, processed_view, log_output, target_dir, image_gallery],
+                        scale=1,
+                        elem_id="my_clear",
+                    )
+
+    gr.Markdown("Click any row to load an example.", elem_classes=["example-log"])
+    with gr.Row():
+        gr.Examples(
+            examples=examples_object,
+            inputs=[
+                preview_imgs,
+                input_file,
+                pca_slider,
+                bright_slider,
+                is_example,
+                scale_slider,
+                normalize_coord,
+            ],
+            outputs=[
+            ],
+            label = "Object Point Cloud Examples",
+            fn=example_file_updated,
+            cache_examples=False,
+            examples_per_page=50,
+        )
+    with gr.Row():
+        gr.Examples(
+            examples=examples_manipulation,
+            inputs=[
+                preview_imgs,
+                input_file,
+                pca_slider,
+                bright_slider,
+                is_example,
+                scale_slider,
+                normalize_coord,
+            ],
+            outputs=[
+            ],
+            label = "Manipulation Point Cloud Examples",
+            fn=example_file_updated,
+            cache_examples=False,
+            examples_per_page=50,
+        )
+    with gr.Row():
+        gr.Examples(
+            examples=examples_indoor,
+            inputs=[
+                preview_imgs,
+                input_file,
+                pca_slider,
+                bright_slider,
+                is_example,
+                scale_slider,
+                normalize_coord,
+            ],
+            outputs=[
+            ],
+            label = "Indoor Point Cloud Examples",
+            fn=example_file_updated,
+            cache_examples=False,
+            examples_per_page=50,
+        )
+    with gr.Row():
+        gr.Examples(
+            examples=examples_outdoor,
+            inputs=[
+                preview_imgs,
+                input_file,
+                pca_slider,
+                bright_slider,
+                is_example,
+                scale_slider,
+                normalize_coord,
+            ],
+            outputs=[
+            ],
+            label = "Outdoor Point Cloud Examples",
+            fn=example_file_updated,
+            cache_examples=False,
+            examples_per_page=50,
+        )
+
+    with gr.Row():
+        gr.Examples(
+            examples=examples_video,
+            inputs=[
+                input_video,
+                conf_thres,
+                frame_slider,
+                prediction_mode,
+                pca_slider,
+                bright_slider,
+                is_example,
+                scale_slider,
+                normalize_coord,
+            ],
+            outputs=[
+            ],
+            label = "Video Examples",
+            fn=example_video_updated,
+            cache_examples=False,
+            examples_per_page=50,
+        )
+
+    reconstruction_btn.click(
+        fn = update_gallery_on_upload,
+        inputs = [input_file,input_video,conf_thres,frame_slider,prediction_mode],
+        outputs = [original_view, target_dir, image_gallery, log_output]
+    )
+    submit_btn.click(fn=clear_fields, inputs=[], outputs=[processed_view]).then(
+        fn=PCAing_log, inputs=[is_example, log_output], outputs=[log_output]
+    ).then(
+        fn=gradio_demo,
+        inputs=[target_dir,pca_slider,bright_slider, if_color, if_normal, scale_slider, apply_z_positive, normalize_coord],
+        outputs=[processed_view,log_output],
+    ).then(
+        fn=lambda: "False", inputs=[], outputs=[is_example]  # set is_example to "False"
+    )
+
+    pca_slider.release(fn=clear_fields, inputs=[], outputs=[processed_view]).then(
+        fn=PCAing_log, inputs=[is_example, log_output], outputs=[log_output]
+    ).then(
+        fn=utonia_slider_update,
+        inputs=[target_dir,pca_slider,bright_slider,is_example,log_output],
+        outputs=[processed_view, log_output],
+    ).then(
+        fn=lambda: "False", inputs=[], outputs=[is_example]  # set is_example to "False"
+    )
+    bright_slider.release(fn=clear_fields, inputs=[], outputs=[processed_view]).then(
+        fn=PCAing_log, inputs=[is_example, log_output], outputs=[log_output]
+    ).then(
+        fn=utonia_slider_update,
+        inputs=[target_dir,pca_slider,bright_slider,is_example,log_output],
+        outputs=[processed_view, log_output],
+    ).then(
+        fn=lambda: "False", inputs=[], outputs=[is_example]  # set is_example to "False"
+    )
+    
+    input_file.change(fn=reset_log, inputs=[], outputs=[log_output]).then(
+        fn=update_gallery_on_upload,
+        inputs=[input_file,input_video, conf_thres,frame_slider,prediction_mode],
+        outputs=[original_view, target_dir, _, log_output],
+    )
+
+    input_video.change(fn=reset_log, inputs=[], outputs=[log_output]).then(
+        fn=update_gallery_on_upload,
+        inputs=[input_file,input_video, conf_thres,frame_slider,prediction_mode],
+        outputs=[original_view, target_dir, image_gallery, log_output],
+    )
+
+if __name__ == "__main__":
+    demo.queue(max_size=20).launch(show_error=True, share=True)

geometry_utils.py

@@ -0,0 +1,197 @@
+import numpy as np
+from scipy.spatial.transform import Rotation as R
+
+
+def pad_0001(Ts):
+    Ts = np.asarray(Ts)
+    if Ts.shape[-2:] == (4, 4):
+        return Ts
+    if Ts.shape[-2:] != (3, 4):
+        raise ValueError("Ts must have shape (..., 3, 4) or (..., 4, 4)")
+    pad = np.zeros((*Ts.shape[:-2], 1, 4), dtype=Ts.dtype)
+    pad[..., 0, 3] = 1
+    return np.concatenate([Ts, pad], axis=-2)
+
+
+def T_to_C(T):
+    T = np.asarray(T)
+    if T.shape != (4, 4):
+        raise ValueError("T must be shape (4,4)")
+    Rm = T[:3, :3]
+    t = T[:3, 3]
+    return -Rm.T @ t
+
+
+def im_distance_to_im_depth(im_dist, K):
+    im_dist = np.asarray(im_dist)
+    H, W = im_dist.shape[:2]
+    ys, xs = np.indices((H, W), dtype=np.float32)
+    fx, fy = K[0, 0], K[1, 1]
+    cx, cy = K[0, 2], K[1, 2]
+    x = (xs - cx) / max(fx, 1e-8)
+    y = (ys - cy) / max(fy, 1e-8)
+    ray_norm = np.sqrt(x * x + y * y + 1.0)
+    return im_dist / np.clip(ray_norm, 1e-8, None)
+
+
+def im_depth_to_point_cloud(im_depth, K, T, to_image=False, ignore_invalid=False):
+    im_depth = np.asarray(im_depth)
+    H, W = im_depth.shape[:2]
+    ys, xs = np.indices((H, W), dtype=np.float32)
+
+    fx, fy = K[0, 0], K[1, 1]
+    cx, cy = K[0, 2], K[1, 2]
+
+    z = im_depth.reshape(-1)
+    x = (xs.reshape(-1) - cx) / max(fx, 1e-8) * z
+    y = (ys.reshape(-1) - cy) / max(fy, 1e-8) * z
+    pts_cam = np.stack([x, y, z], axis=-1)
+
+    T = np.asarray(T)
+    if T.shape != (4, 4):
+        raise ValueError("T must be shape (4,4)")
+    Rm = T[:3, :3]
+    t = T[:3, 3]
+
+    # Assume T is world->camera; invert to camera->world for point transform.
+    pts_world = (Rm.T @ (pts_cam - t).T).T
+
+    if ignore_invalid:
+        valid = np.isfinite(pts_world).all(axis=1) & (z > 0)
+        pts_world = pts_world[valid]
+
+    if to_image:
+        return pts_world.reshape(H, W, 3)
+    return pts_world
+
+def rotx(x, theta=90):
+    """
+    Rotate x by theta degrees around the x-axis
+    """
+    theta = np.deg2rad(theta)
+    rot_matrix = np.array(
+        [
+            [1, 0, 0, 0],
+            [0, np.cos(theta), -np.sin(theta), 0],
+            [0, np.sin(theta), np.cos(theta), 0],
+            [0, 0, 0, 1],
+        ]
+    )
+    return rot_matrix@ x
+
+
+def Coord2zup(points, extrinsics, normals = None):
+    """
+    Convert the dust3r coordinate system to the z-up coordinate system
+    """
+    points = np.concatenate([points, np.ones([points.shape[0], 1])], axis=1).T
+    points = rotx(points, -90)[:3].T
+    if normals is not None:
+        normals = np.concatenate([normals, np.ones([normals.shape[0], 1])], axis=1).T
+        normals = rotx(normals, -90)[:3].T
+        normals = normals / np.linalg.norm(normals, axis=1, keepdims=True)
+    t = np.min(points,axis=0)
+    points -= t
+    extrinsics = rotx(extrinsics, -90)
+    extrinsics[:, :3, 3] -= t.T
+    return points, extrinsics, normals
+
+def _ransac_plane(points, distance_threshold=0.01, ransac_n=3, num_iterations=1000):
+    if points.shape[0] < ransac_n:
+        raise ValueError("Not enough points for plane fitting.")
+
+    best_inliers = None
+    best_plane = None
+    rng = np.random.default_rng(42)
+
+    for _ in range(num_iterations):
+        sample_idx = rng.choice(points.shape[0], size=ransac_n, replace=False)
+        p0, p1, p2 = points[sample_idx]
+        normal = np.cross(p1 - p0, p2 - p0)
+        norm = np.linalg.norm(normal)
+        if norm < 1e-8:
+            continue
+        normal = normal / norm
+        d = -np.dot(normal, p0)
+
+        dist = np.abs(points @ normal + d)
+        inliers = np.where(dist < distance_threshold)[0]
+        if best_inliers is None or len(inliers) > len(best_inliers):
+            best_inliers = inliers
+            best_plane = np.array([normal[0], normal[1], normal[2], d], dtype=np.float64)
+
+    if best_inliers is None or best_plane is None:
+        raise ValueError("Failed to fit plane with RANSAC.")
+
+    return best_plane, best_inliers
+
+
+def extract_and_align_ground_plane(points,
+                                   colors=None,
+                                   normals=None,
+                                   height_percentile=20, 
+                                   ransac_distance_threshold=0.01, 
+                                   ransac_n=3, 
+                                   ransac_iterations=1000,
+                                   max_angle_degree=40,
+                                   max_trials=6):
+    points = np.asarray(points)
+    if points.ndim != 2 or points.shape[1] != 3:
+        raise ValueError("points must be shaped (N, 3)")
+
+    aligned_colors = np.asarray(colors) if colors is not None else None
+    aligned_normals = np.asarray(normals) if normals is not None else None
+
+    z_vals = points[:, 2]
+    z_thresh = np.percentile(z_vals, height_percentile)
+    low_indices = np.where(z_vals <= z_thresh)[0]
+
+    remaining_indices = low_indices.copy()
+
+    for trial in range(max_trials):
+        if len(remaining_indices) < ransac_n:
+            raise ValueError("Not enough points left to fit a plane.")
+
+        candidate_points = points[remaining_indices]
+        plane_model, inliers = _ransac_plane(
+            candidate_points,
+            distance_threshold=ransac_distance_threshold,
+            ransac_n=ransac_n,
+            num_iterations=ransac_iterations,
+        )
+        a, b, c, d = plane_model
+        normal = np.array([a, b, c])
+        normal /= np.linalg.norm(normal)
+
+        angle = np.arccos(np.clip(np.dot(normal, [0, 0, 1]), -1.0, 1.0)) * 180 / np.pi
+        if angle <= max_angle_degree:
+            inliers_global = remaining_indices[inliers]
+
+            target = np.array([0, 0, 1])
+            axis = np.cross(normal, target)
+            axis_norm = np.linalg.norm(axis)
+
+            if axis_norm < 1e-6:
+                rotation_matrix = np.eye(3)
+            else:
+                axis /= axis_norm
+                rot_angle = np.arccos(np.clip(np.dot(normal, target), -1.0, 1.0))
+                rotation = R.from_rotvec(axis * rot_angle)
+                rotation_matrix = rotation.as_matrix()
+
+            rotated_points = points @ rotation_matrix.T
+            ground_points_z = rotated_points[inliers_global, 2]
+            offset = np.mean(ground_points_z)
+            rotated_points[:, 2] -= offset
+
+            if aligned_normals is not None and len(aligned_normals) == len(points):
+                aligned_normals = aligned_normals @ rotation_matrix.T
+                aligned_normals = aligned_normals / np.clip(np.linalg.norm(aligned_normals, axis=-1, keepdims=True), 1e-8, None)
+
+            return rotated_points, aligned_colors, aligned_normals, inliers_global, rotation_matrix, offset
+
+        else:
+            rejected_indices = remaining_indices[inliers]
+            remaining_indices = np.setdiff1d(remaining_indices, rejected_indices)
+
+    raise ValueError("Failed to find a valid ground plane within max trials.")

requirements.txt

@@ -0,0 +1,27 @@
+# Base
+numpy<=1.26.4
+scipy
+addict
+timm
+psutil
+huggingface_hub
+opencv-python-headless
+einops
+ninja
+
+torch==2.4.0
+torchvision==0.19.0
+torchaudio==2.4.0
+
+# Extra dependencies
+--find-links https://data.pyg.org/whl/torch-2.4.0+cu121.html
+torch-scatter
+
+spconv-cu120
+
+https://github.com/Dao-AILab/flash-attention/releases/download/v2.6.3/flash_attn-2.6.3+cu123torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
+
+# Visualization
+trimesh
+camtools
+matplotlib

setup.py

@@ -0,0 +1,12 @@
+from setuptools import setup, find_packages
+
+
+setup(
+    name="utonia",
+    py_modules=["utonia"],
+    version="1.0",
+    description="",
+    author="Yujia Zhang",
+    packages=find_packages(exclude=["demo*"]),
+    include_package_data=True,
+)

utonia/__init__.py

@@ -0,0 +1,26 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from .model import load
+
+from . import model
+from . import module
+from . import structure
+from . import data
+from . import transform
+from . import utils
+from . import registry
+
+__all__ = ["load", "model", "module", "structure", "transform", "registry", "utils"]

utonia/data.py

@@ -0,0 +1,92 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import os
+import numpy as np
+import torch
+from collections.abc import Mapping, Sequence
+from huggingface_hub import hf_hub_download
+
+DATAS = [
+    "sample1",
+    "sample1_high_res",
+    "sample1_dino",
+    "sample2_outdoor",
+    "sample2_outdoor_multiframe",
+    "sample3_object",
+    "sample4_manipulation",
+    "sample5_hk",
+]
+
+
+def load(
+    name: str = "utonia_large",
+    download_root: str = None,
+):
+    if name in DATAS:
+        print(f"Loading data from HuggingFace: {name} ...")
+        data_path = hf_hub_download(
+            repo_id="pointcept/demo",
+            filename=f"{name}.npz",
+            repo_type="dataset",
+            revision="main",
+            local_dir=download_root or os.path.expanduser("~/.cache/utonia/data"),
+        )
+    elif os.path.isfile(name):
+        print(f"Loading data in local path: {name} ...")
+        data_path = name
+    else:
+        raise RuntimeError(f"Data {name} not found; available models = {DATAS}")
+    return dict(np.load(data_path))
+
+
+from torch.utils.data.dataloader import default_collate
+
+
+def collate_fn(batch):
+    """
+    collate function for point cloud which support dict and list,
+    'coord' is necessary to determine 'offset'
+    """
+    if not isinstance(batch, Sequence):
+        raise TypeError(f"{batch.dtype} is not supported.")
+
+    if isinstance(batch[0], torch.Tensor):
+        return torch.cat(list(batch))
+    elif isinstance(batch[0], str):
+        # str is also a kind of Sequence, judgement should before Sequence
+        return list(batch)
+    elif isinstance(batch[0], Sequence):
+        for data in batch:
+            data.append(torch.tensor([data[0].shape[0]]))
+        batch = [collate_fn(samples) for samples in zip(*batch)]
+        batch[-1] = torch.cumsum(batch[-1], dim=0).int()
+        return batch
+    elif isinstance(batch[0], Mapping):
+        batch = {
+            key: (
+                collate_fn([d[key] for d in batch])
+                if "offset" not in key
+                # offset -> bincount -> concat bincount-> concat offset
+                else torch.cumsum(
+                    collate_fn([d[key].diff(prepend=torch.tensor([0])) for d in batch]),
+                    dim=0,
+                )
+            )
+            for key in batch[0]
+        }
+        return batch
+    else:
+        return default_collate(batch)

utonia/model.py

@@ -0,0 +1,891 @@
+"""
+Point Transformer - V3 Mode3 - Utonia
+Pointcept detached version
+
+Author: Xiaoyang Wu (xiaoyang.wu.cs@gmail.com)
+Please cite our work if the code is helpful to you.
+"""
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import os
+import math
+from packaging import version
+from huggingface_hub import hf_hub_download, PyTorchModelHubMixin
+from addict import Dict
+import torch
+import torch.nn as nn
+from torch.nn.init import trunc_normal_
+import spconv.pytorch as spconv
+import torch_scatter
+from timm.layers import DropPath
+
+try:
+    import flash_attn
+except ImportError:
+    flash_attn = None
+
+from .structure import Point
+from .module import PointSequential, PointModule
+from .utils import offset2bincount
+
+MODELS = [
+    "utonia",
+    "utonia_linear_prob_head_sc",
+]
+
+
+class Point3DRoPE(nn.Module):
+    def __init__(self, head_dim, base=10000):
+        super().__init__()
+        assert (
+            head_dim % 3 == 0
+        ), f"Head dimension must be divisible by 3 for 3D RoPE, {head_dim}"
+
+        self.head_dim = head_dim
+        self.chunk_dim = head_dim // 3
+        self.base = base
+        inv_freq = 1.0 / (
+            self.base ** (torch.arange(0, self.chunk_dim, 2).float() / self.chunk_dim)
+        )
+        self.register_buffer("inv_freq", inv_freq)
+
+    def get_cos_sin(self, xyz):
+        x = xyz[:, 0:1]
+        y = xyz[:, 1:2]
+        z = xyz[:, 2:3]
+        freqs = self.inv_freq.unsqueeze(0)
+
+        emb_x = x * freqs
+        emb_y = y * freqs
+        emb_z = z * freqs
+
+        emb_x = torch.cat((emb_x, emb_x), dim=-1)
+        emb_y = torch.cat((emb_y, emb_y), dim=-1)
+        emb_z = torch.cat((emb_z, emb_z), dim=-1)
+        emb_3d = torch.cat((emb_x, emb_y, emb_z), dim=-1)
+        return emb_3d.cos().unsqueeze(1), emb_3d.sin().unsqueeze(1)
+
+    def rotate_half(self, x):
+        x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
+        return torch.cat((-x2, x1), dim=-1)
+
+    def forward(self, q, k, xyz):
+        cos, sin = self.get_cos_sin(xyz)
+        qs = torch.split(q, self.chunk_dim, dim=-1)
+        ks = torch.split(k, self.chunk_dim, dim=-1)
+
+        coss = torch.split(cos, self.chunk_dim, dim=-1)
+        sins = torch.split(sin, self.chunk_dim, dim=-1)
+
+        q_outs = []
+        k_outs = []
+
+        for i in range(3):
+            q_part = (qs[i] * coss[i]) + (self.rotate_half(qs[i]) * sins[i])
+            k_part = (ks[i] * coss[i]) + (self.rotate_half(ks[i]) * sins[i])
+
+            q_outs.append(q_part)
+            k_outs.append(k_part)
+
+        q_rot = torch.cat(q_outs, dim=-1)
+        k_rot = torch.cat(k_outs, dim=-1)
+
+        return q_rot, k_rot
+
+
+class LayerScale(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        init_values: float = 1e-5,
+        inplace: bool = False,
+    ) -> None:
+        super().__init__()
+        self.inplace = inplace
+        self.gamma = nn.Parameter(init_values * torch.ones(dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x.mul_(self.gamma) if self.inplace else x * self.gamma
+
+
+class RPE(torch.nn.Module):
+    def __init__(self, patch_size, num_heads):
+        super().__init__()
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.pos_bnd = int((4 * patch_size) ** (1 / 3) * 2)
+        self.rpe_num = 2 * self.pos_bnd + 1
+        self.rpe_table = torch.nn.Parameter(torch.zeros(3 * self.rpe_num, num_heads))
+        torch.nn.init.trunc_normal_(self.rpe_table, std=0.02)
+
+    def forward(self, coord):
+        idx = (
+            coord.clamp(-self.pos_bnd, self.pos_bnd)  # clamp into bnd
+            + self.pos_bnd  # relative position to positive index
+            + torch.arange(3, device=coord.device) * self.rpe_num  # x, y, z stride
+        )
+        out = self.rpe_table.index_select(0, idx.reshape(-1))
+        out = out.view(idx.shape + (-1,)).sum(3)
+        out = out.permute(0, 3, 1, 2)  # (N, K, K, H) -> (N, H, K, K)
+        return out
+
+
+class SerializedAttention(PointModule):
+    def __init__(
+        self,
+        channels,
+        num_heads,
+        patch_size,
+        qkv_bias=True,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+        order_index=0,
+        enable_rpe=False,
+        enable_flash=True,
+        upcast_attention=True,
+        upcast_softmax=True,
+        rope_base=10,
+        shift_coords=None,
+        jitter_coords=None,
+        rescale_coords=None,
+    ):
+        super().__init__()
+        assert channels % num_heads == 0
+        self.channels = channels
+        self.num_heads = num_heads
+        self.scale = qk_scale or (channels // num_heads) ** -0.5
+        self.order_index = order_index
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+        self.enable_rpe = enable_rpe
+        self.enable_flash = enable_flash
+        if enable_flash:
+            assert (
+                enable_rpe is False
+            ), "Set enable_rpe to False when enable Flash Attention"
+            assert (
+                upcast_attention is False
+            ), "Set upcast_attention to False when enable Flash Attention"
+            assert (
+                upcast_softmax is False
+            ), "Set upcast_softmax to False when enable Flash Attention"
+            assert flash_attn is not None, "Make sure flash_attn is installed."
+            self.patch_size = patch_size
+            self.attn_drop = attn_drop
+        else:
+            # when disable flash attention, we still don't want to use mask
+            # consequently, patch size will auto set to the
+            # min number of patch_size_max and number of points
+            self.patch_size_max = patch_size
+            self.patch_size = 0
+            self.attn_drop = torch.nn.Dropout(attn_drop)
+
+        self.qkv = torch.nn.Linear(channels, channels * 3, bias=qkv_bias)
+        self.proj = torch.nn.Linear(channels, channels)
+        self.proj_drop = torch.nn.Dropout(proj_drop)
+        self.softmax = torch.nn.Softmax(dim=-1)
+        self.rpe = RPE(patch_size, num_heads) if self.enable_rpe else None
+        self.rope = Point3DRoPE(head_dim=channels // num_heads, base=rope_base)
+        self.shift_coords = shift_coords
+        self.jitter_coords = jitter_coords
+        self.rescale_coords = rescale_coords
+
+    @torch.no_grad()
+    def get_rel_pos(self, point, order):
+        K = self.patch_size
+        rel_pos_key = f"rel_pos_{self.order_index}"
+        if rel_pos_key not in point.keys():
+            grid_coord = point.grid_coord[order]
+            grid_coord = grid_coord.reshape(-1, K, 3)
+            point[rel_pos_key] = grid_coord.unsqueeze(2) - grid_coord.unsqueeze(1)
+        return point[rel_pos_key]
+
+    @torch.no_grad()
+    def get_padding_and_inverse(self, point):
+        pad_key = "pad"
+        unpad_key = "unpad"
+        cu_seqlens_key = "cu_seqlens_key"
+        if (
+            pad_key not in point.keys()
+            or unpad_key not in point.keys()
+            or cu_seqlens_key not in point.keys()
+        ):
+            offset = point.offset
+            bincount = offset2bincount(offset)
+            bincount_pad = (
+                torch.div(
+                    bincount + self.patch_size - 1,
+                    self.patch_size,
+                    rounding_mode="trunc",
+                )
+                * self.patch_size
+            )
+            # only pad point when num of points larger than patch_size
+            mask_pad = bincount > self.patch_size
+            bincount_pad = ~mask_pad * bincount + mask_pad * bincount_pad
+            _offset = nn.functional.pad(offset, (1, 0))
+            _offset_pad = nn.functional.pad(torch.cumsum(bincount_pad, dim=0), (1, 0))
+            pad = torch.arange(_offset_pad[-1], device=offset.device)
+            unpad = torch.arange(_offset[-1], device=offset.device)
+            cu_seqlens = []
+            for i in range(len(offset)):
+                unpad[_offset[i] : _offset[i + 1]] += _offset_pad[i] - _offset[i]
+                if bincount[i] != bincount_pad[i]:
+                    pad[
+                        _offset_pad[i + 1]
+                        - self.patch_size
+                        + (bincount[i] % self.patch_size) : _offset_pad[i + 1]
+                    ] = pad[
+                        _offset_pad[i + 1]
+                        - 2 * self.patch_size
+                        + (bincount[i] % self.patch_size) : _offset_pad[i + 1]
+                        - self.patch_size
+                    ]
+                pad[_offset_pad[i] : _offset_pad[i + 1]] -= _offset_pad[i] - _offset[i]
+                cu_seqlens.append(
+                    torch.arange(
+                        _offset_pad[i],
+                        _offset_pad[i + 1],
+                        step=self.patch_size,
+                        dtype=torch.int32,
+                        device=offset.device,
+                    )
+                )
+            point[pad_key] = pad
+            point[unpad_key] = unpad
+            point[cu_seqlens_key] = nn.functional.pad(
+                torch.concat(cu_seqlens), (0, 1), value=_offset_pad[-1]
+            )
+        return point[pad_key], point[unpad_key], point[cu_seqlens_key]
+
+    def forward(self, point):
+        if not self.enable_flash:
+            self.patch_size = min(
+                offset2bincount(point.offset).min().tolist(), self.patch_size_max
+            )
+
+        H = self.num_heads
+        K = self.patch_size
+        C = self.channels
+
+        pad, unpad, cu_seqlens = self.get_padding_and_inverse(point)
+
+        order = point.serialized_order[self.order_index][pad]
+        inverse = unpad[point.serialized_inverse[self.order_index]]
+
+        # padding and reshape feat and batch for serialized point patch
+        qkv = self.qkv(point.feat)[order]
+
+        rope_coord = point.coord[order].clone()
+        if not self.enable_flash:
+            # encode and reshape qkv: (N', K, 3, H, C') => (3, N', H, K, C')
+            qkv = qkv.reshape(-1, 3, H, C // H)
+            q, k, v = qkv.unbind(dim=1)
+            q, k = self.rope(q, k, rope_coord)
+            qkv_roped = torch.stack([q, k, v], dim=1)
+            q, k, v = (
+                qkv.reshape(-1, K, 3, H, C // H).permute(2, 0, 3, 1, 4).unbind(dim=0)
+            )
+            # attn
+            if self.upcast_attention:
+                q = q.float()
+                k = k.float()
+            attn = (q * self.scale) @ k.transpose(-2, -1)  # (N', H, K, K)
+            if self.enable_rpe:
+                attn = attn + self.rpe(self.get_rel_pos(point, order))
+            if self.upcast_softmax:
+                attn = attn.float()
+            attn = self.softmax(attn)
+            attn = self.attn_drop(attn).to(qkv.dtype)
+            feat = (attn @ v).transpose(1, 2).reshape(-1, C)
+        else:
+            qkv = qkv.reshape(-1, 3, H, C // H)
+            q, k, v = qkv.unbind(dim=1)
+            q, k = self.rope(q, k, rope_coord)
+            qkv_roped = torch.stack([q, k, v], dim=1)
+            feat = flash_attn.flash_attn_varlen_qkvpacked_func(
+                qkv_roped.to(torch.bfloat16),
+                cu_seqlens,
+                max_seqlen=self.patch_size,
+                dropout_p=self.attn_drop if self.training else 0,
+                softmax_scale=self.scale,
+            ).reshape(-1, C)
+            feat = feat.to(qkv.dtype)
+        feat = feat[inverse]
+
+        # ffn
+        feat = self.proj(feat)
+        feat = self.proj_drop(feat)
+        point.feat = feat
+        return point
+
+
+class MLP(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels=None,
+        out_channels=None,
+        act_layer=nn.GELU,
+        drop=0.0,
+    ):
+        super().__init__()
+        out_channels = out_channels or in_channels
+        hidden_channels = hidden_channels or in_channels
+        self.fc1 = nn.Linear(in_channels, hidden_channels)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_channels, out_channels)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Block(PointModule):
+    def __init__(
+        self,
+        channels,
+        num_heads,
+        patch_size=48,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+        drop_path=0.0,
+        layer_scale=None,
+        norm_layer=nn.LayerNorm,
+        act_layer=nn.GELU,
+        pre_norm=True,
+        order_index=0,
+        cpe_indice_key=None,
+        enable_rpe=False,
+        enable_flash=True,
+        upcast_attention=True,
+        upcast_softmax=True,
+        rope_base=10,
+        shift_coords=None,
+        jitter_coords=None,
+        rescale_coords=None,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.pre_norm = pre_norm
+
+        self.cpe = PointSequential(
+            spconv.SubMConv3d(
+                channels,
+                channels,
+                kernel_size=3,
+                bias=True,
+                indice_key=cpe_indice_key,
+            ),
+            nn.Linear(channels, channels),
+            norm_layer(channels),
+        )
+
+        self.norm1 = PointSequential(norm_layer(channels))
+        self.ls1 = PointSequential(
+            LayerScale(channels, init_values=layer_scale)
+            if layer_scale is not None
+            else nn.Identity()
+        )
+        self.attn = SerializedAttention(
+            channels=channels,
+            patch_size=patch_size,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            attn_drop=attn_drop,
+            proj_drop=proj_drop,
+            order_index=order_index,
+            enable_rpe=enable_rpe,
+            enable_flash=enable_flash,
+            upcast_attention=upcast_attention,
+            upcast_softmax=upcast_softmax,
+            rope_base=rope_base,
+            shift_coords=shift_coords,
+            jitter_coords=jitter_coords,
+            rescale_coords=rescale_coords,
+        )
+        self.norm2 = PointSequential(norm_layer(channels))
+        self.ls2 = PointSequential(
+            LayerScale(channels, init_values=layer_scale)
+            if layer_scale is not None
+            else nn.Identity()
+        )
+        self.mlp = PointSequential(
+            MLP(
+                in_channels=channels,
+                hidden_channels=int(channels * mlp_ratio),
+                out_channels=channels,
+                act_layer=act_layer,
+                drop=proj_drop,
+            )
+        )
+        self.drop_path = PointSequential(
+            DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        )
+
+    def forward(self, point: Point):
+        shortcut = point.feat
+        point = self.cpe(point)
+        point.feat = shortcut + point.feat
+        shortcut = point.feat
+        if self.pre_norm:
+            point = self.norm1(point)
+        point = self.drop_path(self.ls1(self.attn(point)))
+        point.feat = shortcut + point.feat
+        if not self.pre_norm:
+            point = self.norm1(point)
+
+        shortcut = point.feat
+        if self.pre_norm:
+            point = self.norm2(point)
+        point = self.drop_path(self.ls2(self.mlp(point)))
+        point.feat = shortcut + point.feat
+        if not self.pre_norm:
+            point = self.norm2(point)
+        point.sparse_conv_feat = point.sparse_conv_feat.replace_feature(point.feat)
+        return point
+
+
+class GridPooling(PointModule):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        stride=2,
+        norm_layer=None,
+        act_layer=None,
+        reduce="max",
+        shuffle_orders=True,
+        traceable=True,  # record parent and cluster
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.stride = stride
+        assert reduce in ["sum", "mean", "min", "max"]
+        self.reduce = reduce
+        self.shuffle_orders = shuffle_orders
+        self.traceable = traceable
+
+        self.proj = nn.Linear(in_channels, out_channels)
+        if norm_layer is not None:
+            self.norm = PointSequential(norm_layer(out_channels))
+        if act_layer is not None:
+            self.act = PointSequential(act_layer())
+
+    def forward(self, point: Point):
+        if "grid_coord" in point.keys():
+            grid_coord = point.grid_coord
+        elif {"coord", "grid_size"}.issubset(point.keys()):
+            grid_coord = torch.div(
+                point.coord - point.coord.min(0)[0],
+                point.grid_size,
+                rounding_mode="trunc",
+            ).int()
+        else:
+            raise AssertionError(
+                "[gird_coord] or [coord, grid_size] should be include in the Point"
+            )
+        grid_coord = torch.div(grid_coord, self.stride, rounding_mode="trunc")
+        grid_coord = grid_coord | point.batch.view(-1, 1) << 48
+        grid_coord, cluster, counts = torch.unique(
+            grid_coord,
+            sorted=True,
+            return_inverse=True,
+            return_counts=True,
+            dim=0,
+        )
+        grid_coord = grid_coord & ((1 << 48) - 1)
+        # indices of point sorted by cluster, for torch_scatter.segment_csr
+        _, indices = torch.sort(cluster)
+        # index pointer for sorted point, for torch_scatter.segment_csr
+        idx_ptr = torch.cat([counts.new_zeros(1), torch.cumsum(counts, dim=0)])
+        # head_indices of each cluster, for reduce attr e.g. code, batch
+        head_indices = indices[idx_ptr[:-1]]
+        point_dict = Dict(
+            feat=torch_scatter.segment_csr(
+                self.proj(point.feat)[indices], idx_ptr, reduce=self.reduce
+            ),
+            coord=torch_scatter.segment_csr(
+                point.coord[indices], idx_ptr, reduce="mean"
+            ),
+            grid_coord=grid_coord,
+            batch=point.batch[head_indices],
+        )
+        if "origin_coord" in point.keys():
+            point_dict["origin_coord"] = torch_scatter.segment_csr(
+                point.origin_coord[indices], idx_ptr, reduce="mean"
+            )
+        if "condition" in point.keys():
+            point_dict["condition"] = point.condition
+        if "context" in point.keys():
+            point_dict["context"] = point.context
+        if "name" in point.keys():
+            point_dict["name"] = point.name
+        if "split" in point.keys():
+            point_dict["split"] = point.split
+        if "color" in point.keys():
+            point_dict["color"] = torch_scatter.segment_csr(
+                point.color[indices], idx_ptr, reduce="mean"
+            )
+        if "grid_size" in point.keys():
+            point_dict["grid_size"] = point.grid_size * self.stride
+
+        if self.traceable:
+            point_dict["pooling_inverse"] = cluster
+            point_dict["pooling_parent"] = point
+            point_dict["idx_ptr"] = idx_ptr
+        order = point.order
+        point = Point(point_dict)
+        if self.norm is not None:
+            point = self.norm(point)
+        if self.act is not None:
+            point = self.act(point)
+        point.serialization(order=order, shuffle_orders=self.shuffle_orders)
+        point.sparsify()
+        return point
+
+
+class GridUnpooling(PointModule):
+    def __init__(
+        self,
+        in_channels,
+        skip_channels,
+        out_channels,
+        norm_layer=None,
+        act_layer=None,
+        traceable=False,  # record parent and cluster
+    ):
+        super().__init__()
+        self.proj = PointSequential(nn.Linear(in_channels, out_channels))
+        self.proj_skip = PointSequential(nn.Linear(skip_channels, out_channels))
+
+        if norm_layer is not None:
+            self.proj.add(norm_layer(out_channels))
+            self.proj_skip.add(norm_layer(out_channels))
+
+        if act_layer is not None:
+            self.proj.add(act_layer())
+            self.proj_skip.add(act_layer())
+
+        self.traceable = traceable
+
+    def forward(self, point):
+        assert "pooling_parent" in point.keys()
+        assert "pooling_inverse" in point.keys()
+        parent = point.pop("pooling_parent")
+        inverse = point.pooling_inverse
+        feat = point.feat
+
+        parent = self.proj_skip(parent)
+        parent.feat = parent.feat + self.proj(point).feat[inverse]
+        parent.sparse_conv_feat = parent.sparse_conv_feat.replace_feature(parent.feat)
+
+        if self.traceable:
+            point.feat = feat
+            parent["unpooling_parent"] = point
+        return parent
+
+
+class Embedding(PointModule):
+    def __init__(
+        self,
+        in_channels,
+        embed_channels,
+        norm_layer=None,
+        act_layer=None,
+        mask_token=False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.embed_channels = embed_channels
+
+        self.stem = PointSequential(linear=nn.Linear(in_channels, embed_channels))
+        if norm_layer is not None:
+            self.stem.add(norm_layer(embed_channels), name="norm")
+        if act_layer is not None:
+            self.stem.add(act_layer(), name="act")
+
+        if mask_token:
+            self.mask_token = nn.Parameter(torch.zeros(1, embed_channels))
+        else:
+            self.mask_token = None
+
+    def forward(self, point: Point):
+        point = self.stem(point)
+        if "mask" in point.keys():
+            point.feat = torch.where(
+                point.mask.unsqueeze(-1),
+                self.mask_token.to(point.feat.dtype),
+                point.feat,
+            )
+        return point
+
+
+class PointTransformerV3(PointModule, PyTorchModelHubMixin):
+    def __init__(
+        self,
+        in_channels=6,
+        order=("z", "z-trans"),
+        stride=(2, 2, 2, 2),
+        enc_depths=(3, 3, 3, 12, 3),
+        enc_channels=(48, 96, 192, 384, 512),
+        enc_num_head=(3, 6, 12, 24, 32),
+        enc_patch_size=(1024, 1024, 1024, 1024, 1024),
+        dec_depths=(3, 3, 3, 3),
+        dec_channels=(96, 96, 192, 384),
+        dec_num_head=(6, 6, 12, 32),
+        dec_patch_size=(1024, 1024, 1024, 1024),
+        mlp_ratio=4,
+        qkv_bias=True,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+        drop_path=0.3,
+        layer_scale=None,
+        pre_norm=True,
+        shuffle_orders=True,
+        enable_rpe=False,
+        enable_flash=True,
+        upcast_attention=False,
+        upcast_softmax=False,
+        traceable=False,
+        mask_token=False,
+        enc_mode=False,
+        freeze_encoder=False,
+        rope_base=10,
+        shift_coords=None,
+        jitter_coords=None,
+        rescale_coords=None,
+    ):
+        super().__init__()
+        self.num_stages = len(enc_depths)
+        self.order = [order] if isinstance(order, str) else order
+        self.enc_mode = enc_mode
+        self.shuffle_orders = shuffle_orders
+        self.freeze_encoder = freeze_encoder
+
+        assert self.num_stages == len(stride) + 1
+        assert self.num_stages == len(enc_depths)
+        assert self.num_stages == len(enc_channels)
+        assert self.num_stages == len(enc_num_head)
+        assert self.num_stages == len(enc_patch_size)
+        assert self.enc_mode or self.num_stages == len(dec_depths) + 1
+        assert self.enc_mode or self.num_stages == len(dec_channels) + 1
+        assert self.enc_mode or self.num_stages == len(dec_num_head) + 1
+        assert self.enc_mode or self.num_stages == len(dec_patch_size) + 1
+
+        # normalization layer
+        ln_layer = nn.LayerNorm
+        # activation layers
+        act_layer = nn.GELU
+
+        self.embedding = Embedding(
+            in_channels=in_channels,
+            embed_channels=enc_channels[0],
+            norm_layer=ln_layer,
+            act_layer=act_layer,
+            mask_token=mask_token,
+        )
+
+        # encoder
+        enc_drop_path = [
+            x.item() for x in torch.linspace(0, drop_path, sum(enc_depths))
+        ]
+        self.enc = PointSequential()
+        for s in range(self.num_stages):
+            enc_drop_path_ = enc_drop_path[
+                sum(enc_depths[:s]) : sum(enc_depths[: s + 1])
+            ]
+            enc = PointSequential()
+            if s > 0:
+                enc.add(
+                    GridPooling(
+                        in_channels=enc_channels[s - 1],
+                        out_channels=enc_channels[s],
+                        stride=stride[s - 1],
+                        norm_layer=ln_layer,
+                        act_layer=act_layer,
+                    ),
+                    name="down",
+                )
+            for i in range(enc_depths[s]):
+                enc.add(
+                    Block(
+                        channels=enc_channels[s],
+                        num_heads=enc_num_head[s],
+                        patch_size=enc_patch_size[s],
+                        mlp_ratio=mlp_ratio,
+                        qkv_bias=qkv_bias,
+                        qk_scale=qk_scale,
+                        attn_drop=attn_drop,
+                        proj_drop=proj_drop,
+                        drop_path=enc_drop_path_[i],
+                        layer_scale=layer_scale,
+                        norm_layer=ln_layer,
+                        act_layer=act_layer,
+                        pre_norm=pre_norm,
+                        order_index=i % len(self.order),
+                        cpe_indice_key=f"stage{s}",
+                        enable_rpe=enable_rpe,
+                        enable_flash=enable_flash,
+                        upcast_attention=upcast_attention,
+                        upcast_softmax=upcast_softmax,
+                        rope_base=rope_base,
+                        shift_coords=shift_coords,
+                        jitter_coords=jitter_coords,
+                        rescale_coords=rescale_coords,
+                    ),
+                    name=f"block{i}",
+                )
+            if len(enc) != 0:
+                self.enc.add(module=enc, name=f"enc{s}")
+
+        # decoder
+        if not self.enc_mode:
+            dec_drop_path = [
+                x.item() for x in torch.linspace(0, drop_path, sum(dec_depths))
+            ]
+            self.dec = PointSequential()
+            dec_channels = list(dec_channels) + [enc_channels[-1]]
+            for s in reversed(range(self.num_stages - 1)):
+                dec_drop_path_ = dec_drop_path[
+                    sum(dec_depths[:s]) : sum(dec_depths[: s + 1])
+                ]
+                dec_drop_path_.reverse()
+                dec = PointSequential()
+                dec.add(
+                    GridUnpooling(
+                        in_channels=dec_channels[s + 1],
+                        skip_channels=enc_channels[s],
+                        out_channels=dec_channels[s],
+                        norm_layer=ln_layer,
+                        act_layer=act_layer,
+                        traceable=traceable,
+                    ),
+                    name="up",
+                )
+                for i in range(dec_depths[s]):
+                    dec.add(
+                        Block(
+                            channels=dec_channels[s],
+                            num_heads=dec_num_head[s],
+                            patch_size=dec_patch_size[s],
+                            mlp_ratio=mlp_ratio,
+                            qkv_bias=qkv_bias,
+                            qk_scale=qk_scale,
+                            attn_drop=attn_drop,
+                            proj_drop=proj_drop,
+                            drop_path=dec_drop_path_[i],
+                            layer_scale=layer_scale,
+                            norm_layer=ln_layer,
+                            act_layer=act_layer,
+                            pre_norm=pre_norm,
+                            order_index=i % len(self.order),
+                            cpe_indice_key=f"stage{s}",
+                            enable_rpe=enable_rpe,
+                            enable_flash=enable_flash,
+                            upcast_attention=upcast_attention,
+                            upcast_softmax=upcast_softmax,
+                            rope_base=rope_base,
+                            shift_coords=shift_coords,
+                            jitter_coords=jitter_coords,
+                            rescale_coords=rescale_coords,
+                        ),
+                        name=f"block{i}",
+                    )
+                self.dec.add(module=dec, name=f"dec{s}")
+        if self.freeze_encoder:
+            for p in self.embedding.parameters():
+                p.requires_grad = False
+            for p in self.enc.parameters():
+                p.requires_grad = False
+        self.apply(self._init_weights)
+
+    @staticmethod
+    def _init_weights(module):
+        if isinstance(module, nn.Linear):
+            trunc_normal_(module.weight, std=0.02)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, spconv.SubMConv3d):
+            trunc_normal_(module.weight, std=0.02)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+
+    def forward(self, data_dict):
+        point = Point(data_dict)
+        point = self.embedding(point)
+
+        point.serialization(order=self.order, shuffle_orders=self.shuffle_orders)
+        point.sparsify()
+
+        point = self.enc(point)
+        if not self.enc_mode:
+            point = self.dec(point)
+        return point
+
+
+def load(
+    name: str = "utonia",
+    repo_id="Pointcept/utonia",
+    download_root: str = None,
+    custom_config: dict = None,
+    ckpt_only: bool = False,
+):
+    if name in MODELS:
+        print(f"Loading checkpoint from HuggingFace: {name} ...")
+        ckpt_path = hf_hub_download(
+            repo_id=repo_id,
+            filename=f"{name}.pth",
+            repo_type="model",
+            revision="main",
+            local_dir=download_root or os.path.expanduser("~/.cache/utonia/ckpt"),
+        )
+    elif os.path.isfile(name):
+        print(f"Loading checkpoint in local path: {name} ...")
+        ckpt_path = name
+    else:
+        raise RuntimeError(f"Model {name} not found; available models = {MODELS}")
+
+    if version.parse(torch.__version__) >= version.parse("2.4"):
+        ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=True)
+    else:
+        ckpt = torch.load(ckpt_path, map_location="cpu")
+    if custom_config is not None:
+        for key, value in custom_config.items():
+            ckpt["config"][key] = value
+
+    if ckpt_only:
+        return ckpt
+    ckpt["config"]["drop_path"] = 0.0
+    model = PointTransformerV3(**ckpt["config"])
+    model.load_state_dict(ckpt["state_dict"])
+    n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    print(f"Model params: {n_parameters / 1e6:.2f}M")
+    return model

utonia/module.py

@@ -0,0 +1,107 @@
+"""
+Point Modules
+Pointcept detached version
+
+Author: Xiaoyang Wu (xiaoyang.wu.cs@gmail.com)
+Please cite our work if the code is helpful to you.
+"""
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import sys
+import torch.nn as nn
+import spconv.pytorch as spconv
+from collections import OrderedDict
+
+from .structure import Point
+
+
+class PointModule(nn.Module):
+    r"""PointModule
+    placeholder, all module subclass from this will take Point in PointSequential.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+
+class PointSequential(PointModule):
+    r"""A sequential container.
+    Modules will be added to it in the order they are passed in the constructor.
+    Alternatively, an ordered dict of modules can also be passed in.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+        if len(args) == 1 and isinstance(args[0], OrderedDict):
+            for key, module in args[0].items():
+                self.add_module(key, module)
+        else:
+            for idx, module in enumerate(args):
+                self.add_module(str(idx), module)
+        for name, module in kwargs.items():
+            if sys.version_info < (3, 6):
+                raise ValueError("kwargs only supported in py36+")
+            if name in self._modules:
+                raise ValueError("name exists.")
+            self.add_module(name, module)
+
+    def __getitem__(self, idx):
+        if not (-len(self) <= idx < len(self)):
+            raise IndexError("index {} is out of range".format(idx))
+        if idx < 0:
+            idx += len(self)
+        it = iter(self._modules.values())
+        for i in range(idx):
+            next(it)
+        return next(it)
+
+    def __len__(self):
+        return len(self._modules)
+
+    def add(self, module, name=None):
+        if name is None:
+            name = str(len(self._modules))
+            if name in self._modules:
+                raise KeyError("name exists")
+        self.add_module(name, module)
+
+    def forward(self, input):
+        for k, module in self._modules.items():
+            # Point module
+            if isinstance(module, PointModule):
+                input = module(input)
+            # Spconv module
+            elif spconv.modules.is_spconv_module(module):
+                if isinstance(input, Point):
+                    input.sparse_conv_feat = module(input.sparse_conv_feat)
+                    input.feat = input.sparse_conv_feat.features
+                else:
+                    input = module(input)
+            # PyTorch module
+            else:
+                if isinstance(input, Point):
+                    input.feat = module(input.feat)
+                    if "sparse_conv_feat" in input.keys():
+                        input.sparse_conv_feat = input.sparse_conv_feat.replace_feature(
+                            input.feat
+                        )
+                elif isinstance(input, spconv.SparseConvTensor):
+                    if input.indices.shape[0] != 0:
+                        input = input.replace_feature(module(input.features))
+                else:
+                    input = module(input)
+        return input

utonia/registry.py

@@ -0,0 +1,340 @@
+# @lint-ignore-every LICENSELINT
+# Copyright (c) OpenMMLab. All rights reserved.
+import inspect
+import warnings
+from functools import partial
+from collections import abc
+
+
+def is_seq_of(seq, expected_type, seq_type=None):
+    """Check whether it is a sequence of some type.
+
+    Args:
+        seq (Sequence): The sequence to be checked.
+        expected_type (type): Expected type of sequence items.
+        seq_type (type, optional): Expected sequence type.
+
+    Returns:
+        bool: Whether the sequence is valid.
+    """
+    if seq_type is None:
+        exp_seq_type = abc.Sequence
+    else:
+        assert isinstance(seq_type, type)
+        exp_seq_type = seq_type
+    if not isinstance(seq, exp_seq_type):
+        return False
+    for item in seq:
+        if not isinstance(item, expected_type):
+            return False
+    return True
+
+
+def build_from_cfg(cfg, registry, default_args=None):
+    """Build a module from configs dict.
+
+    Args:
+        cfg (dict): Config dict. It should at least contain the key "type".
+        registry (:obj:`Registry`): The registry to search the type from.
+        default_args (dict, optional): Default initialization arguments.
+
+    Returns:
+        object: The constructed object.
+    """
+    if not isinstance(cfg, dict):
+        raise TypeError(f"cfg must be a dict, but got {type(cfg)}")
+    if "type" not in cfg:
+        if default_args is None or "type" not in default_args:
+            raise KeyError(
+                '`cfg` or `default_args` must contain the key "type", '
+                f"but got {cfg}\n{default_args}"
+            )
+    if not isinstance(registry, Registry):
+        raise TypeError(
+            "registry must be an mmcv.Registry object, " f"but got {type(registry)}"
+        )
+    if not (isinstance(default_args, dict) or default_args is None):
+        raise TypeError(
+            "default_args must be a dict or None, " f"but got {type(default_args)}"
+        )
+
+    args = cfg.copy()
+
+    if default_args is not None:
+        for name, value in default_args.items():
+            args.setdefault(name, value)
+
+    obj_type = args.pop("type")
+    if isinstance(obj_type, str):
+        obj_cls = registry.get(obj_type)
+        if obj_cls is None:
+            raise KeyError(f"{obj_type} is not in the {registry.name} registry")
+    elif inspect.isclass(obj_type):
+        obj_cls = obj_type
+    else:
+        raise TypeError(f"type must be a str or valid type, but got {type(obj_type)}")
+    try:
+        return obj_cls(**args)
+    except Exception as e:
+        # Normal TypeError does not print class name.
+        raise type(e)(f"{obj_cls.__name__}: {e}")
+
+
+class Registry:
+    """A registry to map strings to classes.
+
+    Registered object could be built from registry.
+    Example:
+        >>> MODELS = Registry('models')
+        >>> @MODELS.register_module()
+        >>> class ResNet:
+        >>>     pass
+        >>> resnet = MODELS.build(dict(type='ResNet'))
+
+    Please refer to
+    https://mmcv.readthedocs.io/en/latest/understand_mmcv/registry.html for
+    advanced usage.
+
+    Args:
+        name (str): Registry name.
+        build_func(func, optional): Build function to construct instance from
+            Registry, func:`build_from_cfg` is used if neither ``parent`` or
+            ``build_func`` is specified. If ``parent`` is specified and
+            ``build_func`` is not given,  ``build_func`` will be inherited
+            from ``parent``. Default: None.
+        parent (Registry, optional): Parent registry. The class registered in
+            children registry could be built from parent. Default: None.
+        scope (str, optional): The scope of registry. It is the key to search
+            for children registry. If not specified, scope will be the name of
+            the package where class is defined, e.g. mmdet, mmcls, mmseg.
+            Default: None.
+    """
+
+    def __init__(self, name, build_func=None, parent=None, scope=None):
+        self._name = name
+        self._module_dict = dict()
+        self._children = dict()
+        self._scope = self.infer_scope() if scope is None else scope
+
+        # self.build_func will be set with the following priority:
+        # 1. build_func
+        # 2. parent.build_func
+        # 3. build_from_cfg
+        if build_func is None:
+            if parent is not None:
+                self.build_func = parent.build_func
+            else:
+                self.build_func = build_from_cfg
+        else:
+            self.build_func = build_func
+        if parent is not None:
+            assert isinstance(parent, Registry)
+            parent._add_children(self)
+            self.parent = parent
+        else:
+            self.parent = None
+
+    def __len__(self):
+        return len(self._module_dict)
+
+    def __contains__(self, key):
+        return self.get(key) is not None
+
+    def __repr__(self):
+        format_str = (
+            self.__class__.__name__ + f"(name={self._name}, "
+            f"items={self._module_dict})"
+        )
+        return format_str
+
+    @staticmethod
+    def infer_scope():
+        """Infer the scope of registry.
+
+        The name of the package where registry is defined will be returned.
+
+        Example:
+            # in mmdet/models/backbone/resnet.py
+            >>> MODELS = Registry('models')
+            >>> @MODELS.register_module()
+            >>> class ResNet:
+            >>>     pass
+            The scope of ``ResNet`` will be ``mmdet``.
+
+
+        Returns:
+            scope (str): The inferred scope name.
+        """
+        # inspect.stack() trace where this function is called, the index-2
+        # indicates the frame where `infer_scope()` is called
+        filename = inspect.getmodule(inspect.stack()[2][0]).__name__
+        split_filename = filename.split(".")
+        return split_filename[0]
+
+    @staticmethod
+    def split_scope_key(key):
+        """Split scope and key.
+
+        The first scope will be split from key.
+
+        Examples:
+            >>> Registry.split_scope_key('mmdet.ResNet')
+            'mmdet', 'ResNet'
+            >>> Registry.split_scope_key('ResNet')
+            None, 'ResNet'
+
+        Return:
+            scope (str, None): The first scope.
+            key (str): The remaining key.
+        """
+        split_index = key.find(".")
+        if split_index != -1:
+            return key[:split_index], key[split_index + 1 :]
+        else:
+            return None, key
+
+    @property
+    def name(self):
+        return self._name
+
+    @property
+    def scope(self):
+        return self._scope
+
+    @property
+    def module_dict(self):
+        return self._module_dict
+
+    @property
+    def children(self):
+        return self._children
+
+    def get(self, key):
+        """Get the registry record.
+
+        Args:
+            key (str): The class name in string format.
+
+        Returns:
+            class: The corresponding class.
+        """
+        scope, real_key = self.split_scope_key(key)
+        if scope is None or scope == self._scope:
+            # get from self
+            if real_key in self._module_dict:
+                return self._module_dict[real_key]
+        else:
+            # get from self._children
+            if scope in self._children:
+                return self._children[scope].get(real_key)
+            else:
+                # goto root
+                parent = self.parent
+                while parent.parent is not None:
+                    parent = parent.parent
+                return parent.get(key)
+
+    def build(self, *args, **kwargs):
+        return self.build_func(*args, **kwargs, registry=self)
+
+    def _add_children(self, registry):
+        """Add children for a registry.
+
+        The ``registry`` will be added as children based on its scope.
+        The parent registry could build objects from children registry.
+
+        Example:
+            >>> models = Registry('models')
+            >>> mmdet_models = Registry('models', parent=models)
+            >>> @mmdet_models.register_module()
+            >>> class ResNet:
+            >>>     pass
+            >>> resnet = models.build(dict(type='mmdet.ResNet'))
+        """
+
+        assert isinstance(registry, Registry)
+        assert registry.scope is not None
+        assert (
+            registry.scope not in self.children
+        ), f"scope {registry.scope} exists in {self.name} registry"
+        self.children[registry.scope] = registry
+
+    def _register_module(self, module_class, module_name=None, force=False):
+        if not inspect.isclass(module_class):
+            raise TypeError("module must be a class, " f"but got {type(module_class)}")
+
+        if module_name is None:
+            module_name = module_class.__name__
+        if isinstance(module_name, str):
+            module_name = [module_name]
+        for name in module_name:
+            if not force and name in self._module_dict:
+                raise KeyError(f"{name} is already registered " f"in {self.name}")
+            self._module_dict[name] = module_class
+
+    def deprecated_register_module(self, cls=None, force=False):
+        warnings.warn(
+            "The old API of register_module(module, force=False) "
+            "is deprecated and will be removed, please use the new API "
+            "register_module(name=None, force=False, module=None) instead."
+        )
+        if cls is None:
+            return partial(self.deprecated_register_module, force=force)
+        self._register_module(cls, force=force)
+        return cls
+
+    def register_module(self, name=None, force=False, module=None):
+        """Register a module.
+
+        A record will be added to `self._module_dict`, whose key is the class
+        name or the specified name, and value is the class itself.
+        It can be used as a decorator or a normal function.
+
+        Example:
+            >>> backbones = Registry('backbone')
+            >>> @backbones.register_module()
+            >>> class ResNet:
+            >>>     pass
+
+            >>> backbones = Registry('backbone')
+            >>> @backbones.register_module(name='mnet')
+            >>> class MobileNet:
+            >>>     pass
+
+            >>> backbones = Registry('backbone')
+            >>> class ResNet:
+            >>>     pass
+            >>> backbones.register_module(ResNet)
+
+        Args:
+            name (str | None): The module name to be registered. If not
+                specified, the class name will be used.
+            force (bool, optional): Whether to override an existing class with
+                the same name. Default: False.
+            module (type): Module class to be registered.
+        """
+        if not isinstance(force, bool):
+            raise TypeError(f"force must be a boolean, but got {type(force)}")
+        # NOTE: This is a walkaround to be compatible with the old api,
+        # while it may introduce unexpected bugs.
+        if isinstance(name, type):
+            return self.deprecated_register_module(name, force=force)
+
+        # raise the error ahead of time
+        if not (name is None or isinstance(name, str) or is_seq_of(name, str)):
+            raise TypeError(
+                "name must be either of None, an instance of str or a sequence"
+                f"  of str, but got {type(name)}"
+            )
+
+        # use it as a normal method: x.register_module(module=SomeClass)
+        if module is not None:
+            self._register_module(module_class=module, module_name=name, force=force)
+            return module
+
+        # use it as a decorator: @x.register_module()
+        def _register(cls):
+            self._register_module(module_class=cls, module_name=name, force=force)
+            return cls
+
+        return _register

utonia/serialization/__init__.py

@@ -0,0 +1,8 @@
+from .default import (
+    encode,
+    decode,
+    z_order_encode,
+    z_order_decode,
+    hilbert_encode,
+    hilbert_decode,
+)

utonia/serialization/default.py

@@ -0,0 +1,82 @@
+"""
+Serialization Encoding
+Pointcept detached version
+
+Author: Xiaoyang Wu (xiaoyang.wu.cs@gmail.com)
+Please cite our work if the code is helpful to you.
+"""
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from .z_order import xyz2key as z_order_encode_
+from .z_order import key2xyz as z_order_decode_
+from .hilbert import encode as hilbert_encode_
+from .hilbert import decode as hilbert_decode_
+
+
+@torch.inference_mode()
+def encode(grid_coord, batch=None, depth=16, order="z"):
+    assert order in {"z", "z-trans", "hilbert", "hilbert-trans"}
+    if order == "z":
+        code = z_order_encode(grid_coord, depth=depth)
+    elif order == "z-trans":
+        code = z_order_encode(grid_coord[:, [1, 0, 2]], depth=depth)
+    elif order == "hilbert":
+        code = hilbert_encode(grid_coord, depth=depth)
+    elif order == "hilbert-trans":
+        code = hilbert_encode(grid_coord[:, [1, 0, 2]], depth=depth)
+    else:
+        raise NotImplementedError
+    if batch is not None:
+        batch = batch.long()
+        code = batch << depth * 3 | code
+    return code
+
+
+@torch.inference_mode()
+def decode(code, depth=16, order="z"):
+    assert order in {"z", "hilbert"}
+    batch = code >> depth * 3
+    code = code & ((1 << depth * 3) - 1)
+    if order == "z":
+        grid_coord = z_order_decode(code, depth=depth)
+    elif order == "hilbert":
+        grid_coord = hilbert_decode(code, depth=depth)
+    else:
+        raise NotImplementedError
+    return grid_coord, batch
+
+
+def z_order_encode(grid_coord: torch.Tensor, depth: int = 16):
+    x, y, z = grid_coord[:, 0].long(), grid_coord[:, 1].long(), grid_coord[:, 2].long()
+    # we block the support to batch, maintain batched code in Point class
+    code = z_order_encode_(x, y, z, b=None, depth=depth)
+    return code
+
+
+def z_order_decode(code: torch.Tensor, depth):
+    x, y, z = z_order_decode_(code, depth=depth)
+    grid_coord = torch.stack([x, y, z], dim=-1)  # (N,  3)
+    return grid_coord
+
+
+def hilbert_encode(grid_coord: torch.Tensor, depth: int = 16):
+    return hilbert_encode_(grid_coord, num_dims=3, num_bits=depth)
+
+
+def hilbert_decode(code: torch.Tensor, depth: int = 16):
+    return hilbert_decode_(code, num_dims=3, num_bits=depth)

utonia/serialization/hilbert.py

@@ -0,0 +1,309 @@
+"""
+Hilbert Order
+Modified from https://github.com/PrincetonLIPS/numpy-hilbert-curve
+
+Author: Xiaoyang Wu (xiaoyang.wu.cs@gmail.com), Kaixin Xu
+Please cite our work if the code is helpful to you.
+"""
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+
+
+def right_shift(binary, k=1, axis=-1):
+    """Right shift an array of binary values.
+
+    Parameters:
+    -----------
+     binary: An ndarray of binary values.
+
+     k: The number of bits to shift. Default 1.
+
+     axis: The axis along which to shift.  Default -1.
+
+    Returns:
+    --------
+     Returns an ndarray with zero prepended and the ends truncated, along
+     whatever axis was specified."""
+
+    # If we're shifting the whole thing, just return zeros.
+    if binary.shape[axis] <= k:
+        return torch.zeros_like(binary)
+
+    # Determine the padding pattern.
+    # padding = [(0,0)] * len(binary.shape)
+    # padding[axis] = (k,0)
+
+    # Determine the slicing pattern to eliminate just the last one.
+    slicing = [slice(None)] * len(binary.shape)
+    slicing[axis] = slice(None, -k)
+    shifted = torch.nn.functional.pad(
+        binary[tuple(slicing)], (k, 0), mode="constant", value=0
+    )
+
+    return shifted
+
+
+def binary2gray(binary, axis=-1):
+    """Convert an array of binary values into Gray codes.
+
+    This uses the classic X ^ (X >> 1) trick to compute the Gray code.
+
+    Parameters:
+    -----------
+     binary: An ndarray of binary values.
+
+     axis: The axis along which to compute the gray code. Default=-1.
+
+    Returns:
+    --------
+     Returns an ndarray of Gray codes.
+    """
+    shifted = right_shift(binary, axis=axis)
+
+    # Do the X ^ (X >> 1) trick.
+    gray = torch.logical_xor(binary, shifted)
+
+    return gray
+
+
+def gray2binary(gray, axis=-1):
+    """Convert an array of Gray codes back into binary values.
+
+    Parameters:
+    -----------
+     gray: An ndarray of gray codes.
+
+     axis: The axis along which to perform Gray decoding. Default=-1.
+
+    Returns:
+    --------
+     Returns an ndarray of binary values.
+    """
+
+    # Loop the log2(bits) number of times necessary, with shift and xor.
+    shift = 2 ** (torch.Tensor([gray.shape[axis]]).log2().ceil().int() - 1)
+    while shift > 0:
+        gray = torch.logical_xor(gray, right_shift(gray, shift))
+        shift = torch.div(shift, 2, rounding_mode="floor")
+    return gray
+
+
+def encode(locs, num_dims, num_bits):
+    """Decode an array of locations in a hypercube into a Hilbert integer.
+
+    This is a vectorized-ish version of the Hilbert curve implementation by John
+    Skilling as described in:
+
+    Skilling, J. (2004, April). Programming the Hilbert curve. In AIP Conference
+      Proceedings (Vol. 707, No. 1, pp. 381-387). American Institute of Physics.
+
+    Params:
+    -------
+     locs - An ndarray of locations in a hypercube of num_dims dimensions, in
+            which each dimension runs from 0 to 2**num_bits-1.  The shape can
+            be arbitrary, as long as the last dimension of the same has size
+            num_dims.
+
+     num_dims - The dimensionality of the hypercube. Integer.
+
+     num_bits - The number of bits for each dimension. Integer.
+
+    Returns:
+    --------
+     The output is an ndarray of uint64 integers with the same shape as the
+     input, excluding the last dimension, which needs to be num_dims.
+    """
+
+    # Keep around the original shape for later.
+    orig_shape = locs.shape
+    bitpack_mask = 1 << torch.arange(0, 8).to(locs.device)
+    bitpack_mask_rev = bitpack_mask.flip(-1)
+
+    if orig_shape[-1] != num_dims:
+        raise ValueError("""
+      The shape of locs was surprising in that the last dimension was of size
+      %d, but num_dims=%d.  These need to be equal.
+      """ % (orig_shape[-1], num_dims))
+
+    if num_dims * num_bits > 63:
+        raise ValueError("""
+      num_dims=%d and num_bits=%d for %d bits total, which can't be encoded
+      into a int64.  Are you sure you need that many points on your Hilbert
+      curve?
+      """ % (num_dims, num_bits, num_dims * num_bits))
+
+    # Treat the location integers as 64-bit unsigned and then split them up into
+    # a sequence of uint8s.  Preserve the association by dimension.
+    locs_uint8 = locs.long().view(torch.uint8).reshape((-1, num_dims, 8)).flip(-1)
+
+    # Now turn these into bits and truncate to num_bits.
+    gray = (
+        locs_uint8.unsqueeze(-1)
+        .bitwise_and(bitpack_mask_rev)
+        .ne(0)
+        .byte()
+        .flatten(-2, -1)[..., -num_bits:]
+    )
+
+    # Run the decoding process the other way.
+    # Iterate forwards through the bits.
+    for bit in range(0, num_bits):
+        # Iterate forwards through the dimensions.
+        for dim in range(0, num_dims):
+            # Identify which ones have this bit active.
+            mask = gray[:, dim, bit]
+
+            # Where this bit is on, invert the 0 dimension for lower bits.
+            gray[:, 0, bit + 1 :] = torch.logical_xor(
+                gray[:, 0, bit + 1 :], mask[:, None]
+            )
+
+            # Where the bit is off, exchange the lower bits with the 0 dimension.
+            to_flip = torch.logical_and(
+                torch.logical_not(mask[:, None]).repeat(1, gray.shape[2] - bit - 1),
+                torch.logical_xor(gray[:, 0, bit + 1 :], gray[:, dim, bit + 1 :]),
+            )
+            gray[:, dim, bit + 1 :] = torch.logical_xor(
+                gray[:, dim, bit + 1 :], to_flip
+            )
+            gray[:, 0, bit + 1 :] = torch.logical_xor(gray[:, 0, bit + 1 :], to_flip)
+
+    # Now flatten out.
+    gray = gray.swapaxes(1, 2).reshape((-1, num_bits * num_dims))
+
+    # Convert Gray back to binary.
+    hh_bin = gray2binary(gray)
+
+    # Pad back out to 64 bits.
+    extra_dims = 64 - num_bits * num_dims
+    padded = torch.nn.functional.pad(hh_bin, (extra_dims, 0), "constant", 0)
+
+    # Convert binary values into uint8s.
+    hh_uint8 = (
+        (padded.flip(-1).reshape((-1, 8, 8)) * bitpack_mask)
+        .sum(2)
+        .squeeze()
+        .type(torch.uint8)
+    )
+
+    # Convert uint8s into uint64s.
+    hh_uint64 = hh_uint8.view(torch.int64).squeeze()
+
+    return hh_uint64
+
+
+def decode(hilberts, num_dims, num_bits):
+    """Decode an array of Hilbert integers into locations in a hypercube.
+
+    This is a vectorized-ish version of the Hilbert curve implementation by John
+    Skilling as described in:
+
+    Skilling, J. (2004, April). Programming the Hilbert curve. In AIP Conference
+      Proceedings (Vol. 707, No. 1, pp. 381-387). American Institute of Physics.
+
+    Params:
+    -------
+     hilberts - An ndarray of Hilbert integers.  Must be an integer dtype and
+                cannot have fewer bits than num_dims * num_bits.
+
+     num_dims - The dimensionality of the hypercube. Integer.
+
+     num_bits - The number of bits for each dimension. Integer.
+
+    Returns:
+    --------
+     The output is an ndarray of unsigned integers with the same shape as hilberts
+     but with an additional dimension of size num_dims.
+    """
+
+    if num_dims * num_bits > 64:
+        raise ValueError("""
+      num_dims=%d and num_bits=%d for %d bits total, which can't be encoded
+      into a uint64.  Are you sure you need that many points on your Hilbert
+      curve?
+      """ % (num_dims, num_bits))
+
+    # Handle the case where we got handed a naked integer.
+    hilberts = torch.atleast_1d(hilberts)
+
+    # Keep around the shape for later.
+    orig_shape = hilberts.shape
+    bitpack_mask = 2 ** torch.arange(0, 8).to(hilberts.device)
+    bitpack_mask_rev = bitpack_mask.flip(-1)
+
+    # Treat each of the hilberts as a s equence of eight uint8.
+    # This treats all of the inputs as uint64 and makes things uniform.
+    hh_uint8 = (
+        hilberts.ravel().type(torch.int64).view(torch.uint8).reshape((-1, 8)).flip(-1)
+    )
+
+    # Turn these lists of uints into lists of bits and then truncate to the size
+    # we actually need for using Skilling's procedure.
+    hh_bits = (
+        hh_uint8.unsqueeze(-1)
+        .bitwise_and(bitpack_mask_rev)
+        .ne(0)
+        .byte()
+        .flatten(-2, -1)[:, -num_dims * num_bits :]
+    )
+
+    # Take the sequence of bits and Gray-code it.
+    gray = binary2gray(hh_bits)
+
+    # There has got to be a better way to do this.
+    # I could index them differently, but the eventual packbits likes it this way.
+    gray = gray.reshape((-1, num_bits, num_dims)).swapaxes(1, 2)
+
+    # Iterate backwards through the bits.
+    for bit in range(num_bits - 1, -1, -1):
+        # Iterate backwards through the dimensions.
+        for dim in range(num_dims - 1, -1, -1):
+            # Identify which ones have this bit active.
+            mask = gray[:, dim, bit]
+
+            # Where this bit is on, invert the 0 dimension for lower bits.
+            gray[:, 0, bit + 1 :] = torch.logical_xor(
+                gray[:, 0, bit + 1 :], mask[:, None]
+            )
+
+            # Where the bit is off, exchange the lower bits with the 0 dimension.
+            to_flip = torch.logical_and(
+                torch.logical_not(mask[:, None]),
+                torch.logical_xor(gray[:, 0, bit + 1 :], gray[:, dim, bit + 1 :]),
+            )
+            gray[:, dim, bit + 1 :] = torch.logical_xor(
+                gray[:, dim, bit + 1 :], to_flip
+            )
+            gray[:, 0, bit + 1 :] = torch.logical_xor(gray[:, 0, bit + 1 :], to_flip)
+
+    # Pad back out to 64 bits.
+    extra_dims = 64 - num_bits
+    padded = torch.nn.functional.pad(gray, (extra_dims, 0), "constant", 0)
+
+    # Now chop these up into blocks of 8.
+    locs_chopped = padded.flip(-1).reshape((-1, num_dims, 8, 8))
+
+    # Take those blocks and turn them unto uint8s.
+    # from IPython import embed; embed()
+    locs_uint8 = (locs_chopped * bitpack_mask).sum(3).squeeze().type(torch.uint8)
+
+    # Finally, treat these as uint64s.
+    flat_locs = locs_uint8.view(torch.int64)
+
+    # Return them in the expected shape.
+    return flat_locs.reshape((*orig_shape, num_dims))

utonia/serialization/z_order.py

@@ -0,0 +1,127 @@
+# @lint-ignore-every LICENSELINT
+# --------------------------------------------------------
+# Octree-based Sparse Convolutional Neural Networks
+# Copyright (c) 2022 Peng-Shuai Wang <wangps@hotmail.com>
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Peng-Shuai Wang
+# --------------------------------------------------------
+
+import torch
+from typing import Optional, Union
+
+
+class KeyLUT:
+    def __init__(self):
+        r256 = torch.arange(256, dtype=torch.int64)
+        r512 = torch.arange(512, dtype=torch.int64)
+        zero = torch.zeros(256, dtype=torch.int64)
+        device = torch.device("cpu")
+
+        self._encode = {
+            device: (
+                self.xyz2key(r256, zero, zero, 8),
+                self.xyz2key(zero, r256, zero, 8),
+                self.xyz2key(zero, zero, r256, 8),
+            )
+        }
+        self._decode = {device: self.key2xyz(r512, 9)}
+
+    def encode_lut(self, device=torch.device("cpu")):
+        if device not in self._encode:
+            cpu = torch.device("cpu")
+            self._encode[device] = tuple(e.to(device) for e in self._encode[cpu])
+        return self._encode[device]
+
+    def decode_lut(self, device=torch.device("cpu")):
+        if device not in self._decode:
+            cpu = torch.device("cpu")
+            self._decode[device] = tuple(e.to(device) for e in self._decode[cpu])
+        return self._decode[device]
+
+    def xyz2key(self, x, y, z, depth):
+        key = torch.zeros_like(x)
+        for i in range(depth):
+            mask = 1 << i
+            key = (
+                key
+                | ((x & mask) << (2 * i + 2))
+                | ((y & mask) << (2 * i + 1))
+                | ((z & mask) << (2 * i + 0))
+            )
+        return key
+
+    def key2xyz(self, key, depth):
+        x = torch.zeros_like(key)
+        y = torch.zeros_like(key)
+        z = torch.zeros_like(key)
+        for i in range(depth):
+            x = x | ((key & (1 << (3 * i + 2))) >> (2 * i + 2))
+            y = y | ((key & (1 << (3 * i + 1))) >> (2 * i + 1))
+            z = z | ((key & (1 << (3 * i + 0))) >> (2 * i + 0))
+        return x, y, z
+
+
+_key_lut = KeyLUT()
+
+
+def xyz2key(
+    x: torch.Tensor,
+    y: torch.Tensor,
+    z: torch.Tensor,
+    b: Optional[Union[torch.Tensor, int]] = None,
+    depth: int = 16,
+):
+    r"""Encodes :attr:`x`, :attr:`y`, :attr:`z` coordinates to the shuffled keys
+    based on pre-computed look up tables. The speed of this function is much
+    faster than the method based on for-loop.
+
+    Args:
+      x (torch.Tensor): The x coordinate.
+      y (torch.Tensor): The y coordinate.
+      z (torch.Tensor): The z coordinate.
+      b (torch.Tensor or int): The batch index of the coordinates, and should be
+          smaller than 32768. If :attr:`b` is :obj:`torch.Tensor`, the size of
+          :attr:`b` must be the same as :attr:`x`, :attr:`y`, and :attr:`z`.
+      depth (int): The depth of the shuffled key, and must be smaller than 17 (< 17).
+    """
+
+    EX, EY, EZ = _key_lut.encode_lut(x.device)
+    x, y, z = x.long(), y.long(), z.long()
+
+    mask = 255 if depth > 8 else (1 << depth) - 1
+    key = EX[x & mask] | EY[y & mask] | EZ[z & mask]
+    if depth > 8:
+        mask = (1 << (depth - 8)) - 1
+        key16 = EX[(x >> 8) & mask] | EY[(y >> 8) & mask] | EZ[(z >> 8) & mask]
+        key = key16 << 24 | key
+
+    if b is not None:
+        b = b.long()
+        key = b << 48 | key
+
+    return key
+
+
+def key2xyz(key: torch.Tensor, depth: int = 16):
+    r"""Decodes the shuffled key to :attr:`x`, :attr:`y`, :attr:`z` coordinates
+    and the batch index based on pre-computed look up tables.
+
+    Args:
+      key (torch.Tensor): The shuffled key.
+      depth (int): The depth of the shuffled key, and must be smaller than 17 (< 17).
+    """
+
+    DX, DY, DZ = _key_lut.decode_lut(key.device)
+    x, y, z = torch.zeros_like(key), torch.zeros_like(key), torch.zeros_like(key)
+
+    b = key >> 48
+    key = key & ((1 << 48) - 1)
+
+    n = (depth + 2) // 3
+    for i in range(n):
+        k = key >> (i * 9) & 511
+        x = x | (DX[k] << (i * 3))
+        y = y | (DY[k] << (i * 3))
+        z = z | (DZ[k] << (i * 3))
+
+    return x, y, z, b

utonia/structure.py

@@ -0,0 +1,159 @@
+"""
+Data structure for 3D point cloud
+
+Author: Xiaoyang Wu (xiaoyang.wu.cs@gmail.com)
+Please cite our work if the code is helpful to you.
+"""
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import spconv.pytorch as spconv
+from addict import Dict
+
+from .serialization import encode
+from .utils import offset2batch, batch2offset
+
+
+class Point(Dict):
+    """
+    Point Structure of Pointcept
+
+    A Point (point cloud) in Pointcept is a dictionary that contains various properties of
+    a batched point cloud. The property with the following names have a specific definition
+    as follows:
+
+    - "coord": original coordinate of point cloud;
+    - "grid_coord": grid coordinate for specific grid size (related to GridSampling);
+    Point also support the following optional attributes:
+    - "offset": if not exist, initialized as batch size is 1;
+    - "batch": if not exist, initialized as batch size is 1;
+    - "feat": feature of point cloud, default input of model;
+    - "grid_size": Grid size of point cloud (related to GridSampling);
+    (related to Serialization)
+    - "serialized_depth": depth of serialization, 2 ** depth * grid_size describe the maximum of point cloud range;
+    - "serialized_code": a list of serialization codes;
+    - "serialized_order": a list of serialization order determined by code;
+    - "serialized_inverse": a list of inverse mapping determined by code;
+    (related to Sparsify: SpConv)
+    - "sparse_shape": Sparse shape for Sparse Conv Tensor;
+    - "sparse_conv_feat": SparseConvTensor init with information provide by Point;
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # If one of "offset" or "batch" do not exist, generate by the existing one
+        if "batch" not in self.keys() and "offset" in self.keys():
+            self["batch"] = offset2batch(self.offset)
+        elif "offset" not in self.keys() and "batch" in self.keys():
+            self["offset"] = batch2offset(self.batch)
+
+    def serialization(self, order="z", depth=None, shuffle_orders=False):
+        """
+        Point Cloud Serialization
+
+        relay on ["grid_coord" or "coord" + "grid_size", "batch", "feat"]
+        """
+        self["order"] = order
+        assert "batch" in self.keys()
+        if "grid_coord" not in self.keys():
+            # if you don't want to operate GridSampling in data augmentation,
+            # please add the following augmentation into your pipeline:
+            # dict(type="Copy", keys_dict={"grid_size": 0.01}),
+            # (adjust `grid_size` to what your want)
+            assert {"grid_size", "coord"}.issubset(self.keys())
+
+            self["grid_coord"] = torch.div(
+                self.coord - self.coord.min(0)[0], self.grid_size, rounding_mode="trunc"
+            ).int()
+
+        if depth is None:
+            # Adaptive measure the depth of serialization cube (length = 2 ^ depth)
+            depth = int(self.grid_coord.max() + 1).bit_length()
+        self["serialized_depth"] = depth
+        # Maximum bit length for serialization code is 63 (int64)
+        assert depth * 3 + len(self.offset).bit_length() <= 63
+        # Here we follow OCNN and set the depth limitation to 16 (48bit) for the point position.
+        # Although depth is limited to less than 16, we can encode a 655.36^3 (2^16 * 0.01) meter^3
+        # cube with a grid size of 0.01 meter. We consider it is enough for the current stage.
+        # We can unlock the limitation by optimizing the z-order encoding function if necessary.
+        assert depth <= 16
+
+        # The serialization codes are arranged as following structures:
+        # [Order1 ([n]),
+        #  Order2 ([n]),
+        #   ...
+        #  OrderN ([n])] (k, n)
+        code = [
+            encode(self.grid_coord, self.batch, depth, order=order_) for order_ in order
+        ]
+        code = torch.stack(code)
+        order = torch.argsort(code)
+        inverse = torch.zeros_like(order).scatter_(
+            dim=1,
+            index=order,
+            src=torch.arange(0, code.shape[1], device=order.device).repeat(
+                code.shape[0], 1
+            ),
+        )
+
+        if shuffle_orders:
+            perm = torch.randperm(code.shape[0])
+            code = code[perm]
+            order = order[perm]
+            inverse = inverse[perm]
+
+        self["serialized_code"] = code
+        self["serialized_order"] = order
+        self["serialized_inverse"] = inverse
+
+    def sparsify(self, pad=96):
+        """
+        Point Cloud Serialization
+
+        Point cloud is sparse, here we use "sparsify" to specifically refer to
+        preparing "spconv.SparseConvTensor" for SpConv.
+
+        relay on ["grid_coord" or "coord" + "grid_size", "batch", "feat"]
+
+        pad: padding sparse for sparse shape.
+        """
+        assert {"feat", "batch"}.issubset(self.keys())
+        if "grid_coord" not in self.keys():
+            # if you don't want to operate GridSampling in data augmentation,
+            # please add the following augmentation into your pipeline:
+            # dict(type="Copy", keys_dict={"grid_size": 0.01}),
+            # (adjust `grid_size` to what your want)
+            assert {"grid_size", "coord"}.issubset(self.keys())
+            self["grid_coord"] = torch.div(
+                self.coord - self.coord.min(0)[0], self.grid_size, rounding_mode="trunc"
+            ).int()
+        if "sparse_shape" in self.keys():
+            sparse_shape = self.sparse_shape
+        else:
+            sparse_shape = torch.add(
+                torch.max(self.grid_coord, dim=0).values, pad
+            ).tolist()
+        sparse_conv_feat = spconv.SparseConvTensor(
+            features=self.feat,
+            indices=torch.cat(
+                [self.batch.unsqueeze(-1).int(), self.grid_coord.int()], dim=1
+            ).contiguous(),
+            spatial_shape=sparse_shape,
+            batch_size=self.batch[-1].tolist() + 1,
+        )
+        self["sparse_shape"] = sparse_shape
+        self["sparse_conv_feat"] = sparse_conv_feat

utonia/transform.py

@@ -0,0 +1,1226 @@
+"""
+3D point cloud augmentation
+
+Author: Xiaoyang Wu (xiaoyang.wu.cs@gmail.com)
+Please cite our work if the code is helpful to you.
+"""
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import random
+import numbers
+import scipy
+import scipy.ndimage
+import scipy.interpolate
+import scipy.stats
+import numpy as np
+import torch
+import copy
+from collections.abc import Sequence, Mapping
+
+from .registry import Registry
+
+TRANSFORMS = Registry("transforms")
+
+
+def index_operator(data_dict, index, duplicate=False):
+    # index selection operator for keys in "index_valid_keys"
+    # custom these keys by "Update" transform in config
+    if "index_valid_keys" not in data_dict:
+        data_dict["index_valid_keys"] = [
+            "coord",
+            "color",
+            "normal",
+            "strength",
+            "segment",
+            "instance",
+        ]
+    if not duplicate:
+        for key in data_dict["index_valid_keys"]:
+            if key in data_dict:
+                data_dict[key] = data_dict[key][index]
+        return data_dict
+    else:
+        data_dict_ = dict()
+        for key in data_dict.keys():
+            if key in data_dict["index_valid_keys"]:
+                data_dict_[key] = data_dict[key][index]
+            else:
+                data_dict_[key] = data_dict[key]
+        return data_dict_
+
+
+@TRANSFORMS.register_module()
+class Collect(object):
+    def __init__(self, keys, offset_keys_dict=None, **kwargs):
+        """
+        e.g. Collect(keys=[coord], feat_keys=[coord, color])
+        """
+        if offset_keys_dict is None:
+            offset_keys_dict = dict(offset="coord")
+        self.keys = keys
+        self.offset_keys = offset_keys_dict
+        self.kwargs = kwargs
+
+    def __call__(self, data_dict):
+        data = dict()
+        if isinstance(self.keys, str):
+            self.keys = [self.keys]
+        for key in self.keys:
+            data[key] = data_dict[key]
+        for key, value in self.offset_keys.items():
+            data[key] = torch.tensor([data_dict[value].shape[0]])
+        for name, keys in self.kwargs.items():
+            name = name.replace("_keys", "")
+            assert isinstance(keys, Sequence)
+            data[name] = torch.cat([data_dict[key].float() for key in keys], dim=1)
+        return data
+
+
+@TRANSFORMS.register_module()
+class Copy(object):
+    def __init__(self, keys_dict=None):
+        if keys_dict is None:
+            keys_dict = dict(coord="origin_coord", segment="origin_segment")
+        self.keys_dict = keys_dict
+
+    def __call__(self, data_dict):
+        for key, value in self.keys_dict.items():
+            if isinstance(data_dict[key], np.ndarray):
+                data_dict[value] = data_dict[key].copy()
+            elif isinstance(data_dict[key], torch.Tensor):
+                data_dict[value] = data_dict[key].clone().detach()
+            else:
+                data_dict[value] = copy.deepcopy(data_dict[key])
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class Update(object):
+    def __init__(self, keys_dict=None):
+        if keys_dict is None:
+            keys_dict = dict()
+        self.keys_dict = keys_dict
+
+    def __call__(self, data_dict):
+        for key, value in self.keys_dict.items():
+            data_dict[key] = value
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class ToTensor(object):
+    def __call__(self, data):
+        if isinstance(data, torch.Tensor):
+            return data
+        elif isinstance(data, str):
+            # note that str is also a kind of sequence, judgement should before sequence
+            return data
+        elif isinstance(data, int):
+            return torch.LongTensor([data])
+        elif isinstance(data, float):
+            return torch.FloatTensor([data])
+        elif isinstance(data, np.ndarray) and np.issubdtype(data.dtype, bool):
+            return torch.from_numpy(data)
+        elif isinstance(data, np.ndarray) and np.issubdtype(data.dtype, np.integer):
+            return torch.from_numpy(data).long()
+        elif isinstance(data, np.ndarray) and np.issubdtype(data.dtype, np.floating):
+            return torch.from_numpy(data).float()
+        elif isinstance(data, Mapping):
+            result = {sub_key: self(item) for sub_key, item in data.items()}
+            return result
+        elif isinstance(data, Sequence):
+            result = [self(item) for item in data]
+            return result
+        else:
+            raise TypeError(f"type {type(data)} cannot be converted to tensor.")
+
+
+@TRANSFORMS.register_module()
+class NormalizeColor(object):
+    def __call__(self, data_dict):
+        if "color" in data_dict.keys():
+            data_dict["color"] = data_dict["color"] / 255
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class NormalizeCoord(object):
+    def __call__(self, data_dict):
+        if "coord" in data_dict.keys():
+            # modified from pointnet2
+            centroid = np.mean(data_dict["coord"], axis=0)
+            data_dict["coord"] -= centroid
+            m = np.max(np.sqrt(np.sum(data_dict["coord"] ** 2, axis=1)))
+            data_dict["coord"] = data_dict["coord"] / m
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class PositiveShift(object):
+    def __call__(self, data_dict):
+        if "coord" in data_dict.keys():
+            coord_min = np.min(data_dict["coord"], 0)
+            data_dict["coord"] -= coord_min
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class CenterShift(object):
+    def __init__(self, apply_z=True):
+        self.apply_z = apply_z
+
+    def __call__(self, data_dict):
+        if "coord" in data_dict.keys():
+            x_min, y_min, z_min = data_dict["coord"].min(axis=0)
+            x_max, y_max, _ = data_dict["coord"].max(axis=0)
+            if self.apply_z:
+                shift = [(x_min + x_max) / 2, (y_min + y_max) / 2, z_min]
+            else:
+                shift = [(x_min + x_max) / 2, (y_min + y_max) / 2, 0]
+            data_dict["coord"] -= shift
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomShift(object):
+    def __init__(self, shift=((-0.2, 0.2), (-0.2, 0.2), (0, 0))):
+        self.shift = shift
+
+    def __call__(self, data_dict):
+        if "coord" in data_dict.keys():
+            shift_x = np.random.uniform(self.shift[0][0], self.shift[0][1])
+            shift_y = np.random.uniform(self.shift[1][0], self.shift[1][1])
+            shift_z = np.random.uniform(self.shift[2][0], self.shift[2][1])
+            data_dict["coord"] += [shift_x, shift_y, shift_z]
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class PointClip(object):
+    def __init__(self, point_cloud_range=(-80, -80, -3, 80, 80, 1)):
+        self.point_cloud_range = point_cloud_range
+
+    def __call__(self, data_dict):
+        if "coord" in data_dict.keys():
+            data_dict["coord"] = np.clip(
+                data_dict["coord"],
+                a_min=self.point_cloud_range[:3],
+                a_max=self.point_cloud_range[3:],
+            )
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomDropout(object):
+    def __init__(self, dropout_ratio=0.2, dropout_application_ratio=0.5):
+        """
+        upright_axis: axis index among x,y,z, i.e. 2 for z
+        """
+        self.dropout_ratio = dropout_ratio
+        self.dropout_application_ratio = dropout_application_ratio
+
+    def __call__(self, data_dict):
+        if random.random() < self.dropout_application_ratio:
+            n = len(data_dict["coord"])
+            idx = np.random.choice(n, int(n * (1 - self.dropout_ratio)), replace=False)
+            if "sampled_index" in data_dict:
+                # for ScanNet data efficient, we need to make sure labeled point is sampled.
+                idx = np.unique(np.append(idx, data_dict["sampled_index"]))
+                mask = np.zeros_like(data_dict["segment"]).astype(bool)
+                mask[data_dict["sampled_index"]] = True
+                data_dict["sampled_index"] = np.where(mask[idx])[0]
+            data_dict = index_operator(data_dict, idx)
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomRotate(object):
+    def __init__(self, angle=None, center=None, axis="z", always_apply=False, p=0.5):
+        self.angle = [-1, 1] if angle is None else angle
+        self.axis = axis
+        self.always_apply = always_apply
+        self.p = p if not self.always_apply else 1
+        self.center = center
+
+    def __call__(self, data_dict):
+        if random.random() > self.p:
+            return data_dict
+        angle = np.random.uniform(self.angle[0], self.angle[1]) * np.pi
+        rot_cos, rot_sin = np.cos(angle), np.sin(angle)
+        if self.axis == "x":
+            rot_t = np.array([[1, 0, 0], [0, rot_cos, -rot_sin], [0, rot_sin, rot_cos]])
+        elif self.axis == "y":
+            rot_t = np.array([[rot_cos, 0, rot_sin], [0, 1, 0], [-rot_sin, 0, rot_cos]])
+        elif self.axis == "z":
+            rot_t = np.array([[rot_cos, -rot_sin, 0], [rot_sin, rot_cos, 0], [0, 0, 1]])
+        else:
+            raise NotImplementedError
+        if "coord" in data_dict.keys():
+            if self.center is None:
+                x_min, y_min, z_min = data_dict["coord"].min(axis=0)
+                x_max, y_max, z_max = data_dict["coord"].max(axis=0)
+                center = [(x_min + x_max) / 2, (y_min + y_max) / 2, (z_min + z_max) / 2]
+            else:
+                center = self.center
+            data_dict["coord"] -= center
+            data_dict["coord"] = np.dot(data_dict["coord"], np.transpose(rot_t))
+            data_dict["coord"] += center
+        if "normal" in data_dict.keys():
+            data_dict["normal"] = np.dot(data_dict["normal"], np.transpose(rot_t))
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomRotateTargetAngle(object):
+    def __init__(
+        self, angle=(1 / 2, 1, 3 / 2), center=None, axis="z", always_apply=False, p=0.75
+    ):
+        self.angle = angle
+        self.axis = axis
+        self.always_apply = always_apply
+        self.p = p if not self.always_apply else 1
+        self.center = center
+
+    def __call__(self, data_dict):
+        if random.random() > self.p:
+            return data_dict
+        angle = np.random.choice(self.angle) * np.pi
+        rot_cos, rot_sin = np.cos(angle), np.sin(angle)
+        if self.axis == "x":
+            rot_t = np.array([[1, 0, 0], [0, rot_cos, -rot_sin], [0, rot_sin, rot_cos]])
+        elif self.axis == "y":
+            rot_t = np.array([[rot_cos, 0, rot_sin], [0, 1, 0], [-rot_sin, 0, rot_cos]])
+        elif self.axis == "z":
+            rot_t = np.array([[rot_cos, -rot_sin, 0], [rot_sin, rot_cos, 0], [0, 0, 1]])
+        else:
+            raise NotImplementedError
+        if "coord" in data_dict.keys():
+            if self.center is None:
+                x_min, y_min, z_min = data_dict["coord"].min(axis=0)
+                x_max, y_max, z_max = data_dict["coord"].max(axis=0)
+                center = [(x_min + x_max) / 2, (y_min + y_max) / 2, (z_min + z_max) / 2]
+            else:
+                center = self.center
+            data_dict["coord"] -= center
+            data_dict["coord"] = np.dot(data_dict["coord"], np.transpose(rot_t))
+            data_dict["coord"] += center
+        if "normal" in data_dict.keys():
+            data_dict["normal"] = np.dot(data_dict["normal"], np.transpose(rot_t))
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomScale(object):
+    def __init__(self, scale=None, anisotropic=False):
+        self.scale = scale if scale is not None else [0.95, 1.05]
+        self.anisotropic = anisotropic
+
+    def __call__(self, data_dict):
+        if "coord" in data_dict.keys():
+            scale = np.random.uniform(
+                self.scale[0], self.scale[1], 3 if self.anisotropic else 1
+            )
+            data_dict["coord"] *= scale
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomFlip(object):
+    def __init__(self, p=0.5):
+        self.p = p
+
+    def __call__(self, data_dict):
+        if np.random.rand() < self.p:
+            if "coord" in data_dict.keys():
+                data_dict["coord"][:, 0] = -data_dict["coord"][:, 0]
+            if "normal" in data_dict.keys():
+                data_dict["normal"][:, 0] = -data_dict["normal"][:, 0]
+        if np.random.rand() < self.p:
+            if "coord" in data_dict.keys():
+                data_dict["coord"][:, 1] = -data_dict["coord"][:, 1]
+            if "normal" in data_dict.keys():
+                data_dict["normal"][:, 1] = -data_dict["normal"][:, 1]
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomJitter(object):
+    def __init__(self, sigma=0.01, clip=0.05):
+        assert clip > 0
+        self.sigma = sigma
+        self.clip = clip
+
+    def __call__(self, data_dict):
+        if "coord" in data_dict.keys():
+            jitter = np.clip(
+                self.sigma * np.random.randn(data_dict["coord"].shape[0], 3),
+                -self.clip,
+                self.clip,
+            )
+            data_dict["coord"] += jitter
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class ClipGaussianJitter(object):
+    def __init__(self, scalar=0.02, store_jitter=False):
+        self.scalar = scalar
+        self.mean = np.mean(3)
+        self.cov = np.identity(3)
+        self.quantile = 1.96
+        self.store_jitter = store_jitter
+
+    def __call__(self, data_dict):
+        if "coord" in data_dict.keys():
+            jitter = np.random.multivariate_normal(
+                self.mean, self.cov, data_dict["coord"].shape[0]
+            )
+            jitter = self.scalar * np.clip(jitter / 1.96, -1, 1)
+            data_dict["coord"] += jitter
+            if self.store_jitter:
+                data_dict["jitter"] = jitter
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class ChromaticAutoContrast(object):
+    def __init__(self, p=0.2, blend_factor=None):
+        self.p = p
+        self.blend_factor = blend_factor
+
+    def __call__(self, data_dict):
+        if "color" in data_dict.keys() and np.random.rand() < self.p:
+            lo = np.min(data_dict["color"], 0, keepdims=True)
+            hi = np.max(data_dict["color"], 0, keepdims=True)
+            scale = 255 / (hi - lo)
+            contrast_feat = (data_dict["color"][:, :3] - lo) * scale
+            blend_factor = (
+                np.random.rand() if self.blend_factor is None else self.blend_factor
+            )
+            data_dict["color"][:, :3] = (1 - blend_factor) * data_dict["color"][
+                :, :3
+            ] + blend_factor * contrast_feat
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class ChromaticTranslation(object):
+    def __init__(self, p=0.95, ratio=0.05):
+        self.p = p
+        self.ratio = ratio
+
+    def __call__(self, data_dict):
+        if "color" in data_dict.keys() and np.random.rand() < self.p:
+            tr = (np.random.rand(1, 3) - 0.5) * 255 * 2 * self.ratio
+            data_dict["color"][:, :3] = np.clip(tr + data_dict["color"][:, :3], 0, 255)
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class ChromaticJitter(object):
+    def __init__(self, p=0.95, std=0.005):
+        self.p = p
+        self.std = std
+
+    def __call__(self, data_dict):
+        if "color" in data_dict.keys() and np.random.rand() < self.p:
+            noise = np.random.randn(data_dict["color"].shape[0], 3)
+            noise *= self.std * 255
+            data_dict["color"][:, :3] = np.clip(
+                noise + data_dict["color"][:, :3], 0, 255
+            )
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomColorGrayScale(object):
+    def __init__(self, p):
+        self.p = p
+
+    @staticmethod
+    def rgb_to_grayscale(color, num_output_channels=1):
+        if color.shape[-1] < 3:
+            raise TypeError(
+                "Input color should have at least 3 dimensions, but found {}".format(
+                    color.shape[-1]
+                )
+            )
+
+        if num_output_channels not in (1, 3):
+            raise ValueError("num_output_channels should be either 1 or 3")
+
+        r, g, b = color[..., 0], color[..., 1], color[..., 2]
+        gray = (0.2989 * r + 0.587 * g + 0.114 * b).astype(color.dtype)
+        gray = np.expand_dims(gray, axis=-1)
+
+        if num_output_channels == 3:
+            gray = np.broadcast_to(gray, color.shape)
+
+        return gray
+
+    def __call__(self, data_dict):
+        if np.random.rand() < self.p:
+            data_dict["color"] = self.rgb_to_grayscale(data_dict["color"], 3)
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomColorJitter(object):
+    """
+    Random Color Jitter for 3D point cloud (refer torchvision)
+    """
+
+    def __init__(self, brightness=0, contrast=0, saturation=0, hue=0, p=0.95):
+        self.brightness = self._check_input(brightness, "brightness")
+        self.contrast = self._check_input(contrast, "contrast")
+        self.saturation = self._check_input(saturation, "saturation")
+        self.hue = self._check_input(
+            hue, "hue", center=0, bound=(-0.5, 0.5), clip_first_on_zero=False
+        )
+        self.p = p
+
+    @staticmethod
+    def _check_input(
+        value, name, center=1, bound=(0, float("inf")), clip_first_on_zero=True
+    ):
+        if isinstance(value, numbers.Number):
+            if value < 0:
+                raise ValueError(
+                    "If {} is a single number, it must be non negative.".format(name)
+                )
+            value = [center - float(value), center + float(value)]
+            if clip_first_on_zero:
+                value[0] = max(value[0], 0.0)
+        elif isinstance(value, (tuple, list)) and len(value) == 2:
+            if not bound[0] <= value[0] <= value[1] <= bound[1]:
+                raise ValueError("{} values should be between {}".format(name, bound))
+        else:
+            raise TypeError(
+                "{} should be a single number or a list/tuple with length 2.".format(
+                    name
+                )
+            )
+
+        # if value is 0 or (1., 1.) for brightness/contrast/saturation
+        # or (0., 0.) for hue, do nothing
+        if value[0] == value[1] == center:
+            value = None
+        return value
+
+    @staticmethod
+    def blend(color1, color2, ratio):
+        ratio = float(ratio)
+        bound = 255.0
+        return (
+            (ratio * color1 + (1.0 - ratio) * color2)
+            .clip(0, bound)
+            .astype(color1.dtype)
+        )
+
+    @staticmethod
+    def rgb2hsv(rgb):
+        r, g, b = rgb[..., 0], rgb[..., 1], rgb[..., 2]
+        maxc = np.max(rgb, axis=-1)
+        minc = np.min(rgb, axis=-1)
+        eqc = maxc == minc
+        cr = maxc - minc
+        s = cr / (np.ones_like(maxc) * eqc + maxc * (1 - eqc))
+        cr_divisor = np.ones_like(maxc) * eqc + cr * (1 - eqc)
+        rc = (maxc - r) / cr_divisor
+        gc = (maxc - g) / cr_divisor
+        bc = (maxc - b) / cr_divisor
+
+        hr = (maxc == r) * (bc - gc)
+        hg = ((maxc == g) & (maxc != r)) * (2.0 + rc - bc)
+        hb = ((maxc != g) & (maxc != r)) * (4.0 + gc - rc)
+        h = hr + hg + hb
+        h = (h / 6.0 + 1.0) % 1.0
+        return np.stack((h, s, maxc), axis=-1)
+
+    @staticmethod
+    def hsv2rgb(hsv):
+        h, s, v = hsv[..., 0], hsv[..., 1], hsv[..., 2]
+        i = np.floor(h * 6.0)
+        f = (h * 6.0) - i
+        i = i.astype(np.int32)
+
+        p = np.clip((v * (1.0 - s)), 0.0, 1.0)
+        q = np.clip((v * (1.0 - s * f)), 0.0, 1.0)
+        t = np.clip((v * (1.0 - s * (1.0 - f))), 0.0, 1.0)
+        i = i % 6
+        mask = np.expand_dims(i, axis=-1) == np.arange(6)
+
+        a1 = np.stack((v, q, p, p, t, v), axis=-1)
+        a2 = np.stack((t, v, v, q, p, p), axis=-1)
+        a3 = np.stack((p, p, t, v, v, q), axis=-1)
+        a4 = np.stack((a1, a2, a3), axis=-1)
+
+        return np.einsum("...na, ...nab -> ...nb", mask.astype(hsv.dtype), a4)
+
+    def adjust_brightness(self, color, brightness_factor):
+        if brightness_factor < 0:
+            raise ValueError(
+                "brightness_factor ({}) is not non-negative.".format(brightness_factor)
+            )
+
+        return self.blend(color, np.zeros_like(color), brightness_factor)
+
+    def adjust_contrast(self, color, contrast_factor):
+        if contrast_factor < 0:
+            raise ValueError(
+                "contrast_factor ({}) is not non-negative.".format(contrast_factor)
+            )
+        mean = np.mean(RandomColorGrayScale.rgb_to_grayscale(color))
+        return self.blend(color, mean, contrast_factor)
+
+    def adjust_saturation(self, color, saturation_factor):
+        if saturation_factor < 0:
+            raise ValueError(
+                "saturation_factor ({}) is not non-negative.".format(saturation_factor)
+            )
+        gray = RandomColorGrayScale.rgb_to_grayscale(color)
+        return self.blend(color, gray, saturation_factor)
+
+    def adjust_hue(self, color, hue_factor):
+        if not (-0.5 <= hue_factor <= 0.5):
+            raise ValueError(
+                "hue_factor ({}) is not in [-0.5, 0.5].".format(hue_factor)
+            )
+        orig_dtype = color.dtype
+        hsv = self.rgb2hsv(color / 255.0)
+        h, s, v = hsv[..., 0], hsv[..., 1], hsv[..., 2]
+        h = (h + hue_factor) % 1.0
+        hsv = np.stack((h, s, v), axis=-1)
+        color_hue_adj = (self.hsv2rgb(hsv) * 255.0).astype(orig_dtype)
+        return color_hue_adj
+
+    @staticmethod
+    def get_params(brightness, contrast, saturation, hue):
+        fn_idx = torch.randperm(4)
+        b = (
+            None
+            if brightness is None
+            else np.random.uniform(brightness[0], brightness[1])
+        )
+        c = None if contrast is None else np.random.uniform(contrast[0], contrast[1])
+        s = (
+            None
+            if saturation is None
+            else np.random.uniform(saturation[0], saturation[1])
+        )
+        h = None if hue is None else np.random.uniform(hue[0], hue[1])
+        return fn_idx, b, c, s, h
+
+    def __call__(self, data_dict):
+        (
+            fn_idx,
+            brightness_factor,
+            contrast_factor,
+            saturation_factor,
+            hue_factor,
+        ) = self.get_params(self.brightness, self.contrast, self.saturation, self.hue)
+
+        for fn_id in fn_idx:
+            if (
+                fn_id == 0
+                and brightness_factor is not None
+                and np.random.rand() < self.p
+            ):
+                data_dict["color"] = self.adjust_brightness(
+                    data_dict["color"], brightness_factor
+                )
+            elif (
+                fn_id == 1 and contrast_factor is not None and np.random.rand() < self.p
+            ):
+                data_dict["color"] = self.adjust_contrast(
+                    data_dict["color"], contrast_factor
+                )
+            elif (
+                fn_id == 2
+                and saturation_factor is not None
+                and np.random.rand() < self.p
+            ):
+                data_dict["color"] = self.adjust_saturation(
+                    data_dict["color"], saturation_factor
+                )
+            elif fn_id == 3 and hue_factor is not None and np.random.rand() < self.p:
+                data_dict["color"] = self.adjust_hue(data_dict["color"], hue_factor)
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class HueSaturationTranslation(object):
+    @staticmethod
+    def rgb_to_hsv(rgb):
+        # Translated from source of colorsys.rgb_to_hsv
+        # r,g,b should be a numpy arrays with values between 0 and 255
+        # rgb_to_hsv returns an array of floats between 0.0 and 1.0.
+        rgb = rgb.astype("float")
+        hsv = np.zeros_like(rgb)
+        # in case an RGBA array was passed, just copy the A channel
+        hsv[..., 3:] = rgb[..., 3:]
+        r, g, b = rgb[..., 0], rgb[..., 1], rgb[..., 2]
+        maxc = np.max(rgb[..., :3], axis=-1)
+        minc = np.min(rgb[..., :3], axis=-1)
+        hsv[..., 2] = maxc
+        mask = maxc != minc
+        hsv[mask, 1] = (maxc - minc)[mask] / maxc[mask]
+        rc = np.zeros_like(r)
+        gc = np.zeros_like(g)
+        bc = np.zeros_like(b)
+        rc[mask] = (maxc - r)[mask] / (maxc - minc)[mask]
+        gc[mask] = (maxc - g)[mask] / (maxc - minc)[mask]
+        bc[mask] = (maxc - b)[mask] / (maxc - minc)[mask]
+        hsv[..., 0] = np.select(
+            [r == maxc, g == maxc], [bc - gc, 2.0 + rc - bc], default=4.0 + gc - rc
+        )
+        hsv[..., 0] = (hsv[..., 0] / 6.0) % 1.0
+        return hsv
+
+    @staticmethod
+    def hsv_to_rgb(hsv):
+        # Translated from source of colorsys.hsv_to_rgb
+        # h,s should be a numpy arrays with values between 0.0 and 1.0
+        # v should be a numpy array with values between 0.0 and 255.0
+        # hsv_to_rgb returns an array of uints between 0 and 255.
+        rgb = np.empty_like(hsv)
+        rgb[..., 3:] = hsv[..., 3:]
+        h, s, v = hsv[..., 0], hsv[..., 1], hsv[..., 2]
+        i = (h * 6.0).astype("uint8")
+        f = (h * 6.0) - i
+        p = v * (1.0 - s)
+        q = v * (1.0 - s * f)
+        t = v * (1.0 - s * (1.0 - f))
+        i = i % 6
+        conditions = [s == 0.0, i == 1, i == 2, i == 3, i == 4, i == 5]
+        rgb[..., 0] = np.select(conditions, [v, q, p, p, t, v], default=v)
+        rgb[..., 1] = np.select(conditions, [v, v, v, q, p, p], default=t)
+        rgb[..., 2] = np.select(conditions, [v, p, t, v, v, q], default=p)
+        return rgb.astype("uint8")
+
+    def __init__(self, hue_max=0.5, saturation_max=0.2):
+        self.hue_max = hue_max
+        self.saturation_max = saturation_max
+
+    def __call__(self, data_dict):
+        if "color" in data_dict.keys():
+            # Assume color[:, :3] is rgb
+            hsv = HueSaturationTranslation.rgb_to_hsv(data_dict["color"][:, :3])
+            hue_val = (np.random.rand() - 0.5) * 2 * self.hue_max
+            sat_ratio = 1 + (np.random.rand() - 0.5) * 2 * self.saturation_max
+            hsv[..., 0] = np.remainder(hue_val + hsv[..., 0] + 1, 1)
+            hsv[..., 1] = np.clip(sat_ratio * hsv[..., 1], 0, 1)
+            data_dict["color"][:, :3] = np.clip(
+                HueSaturationTranslation.hsv_to_rgb(hsv), 0, 255
+            )
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class RandomColorDrop(object):
+    def __init__(self, p=0.2, color_augment=0.0):
+        self.p = p
+        self.color_augment = color_augment
+
+    def __call__(self, data_dict):
+        if "color" in data_dict.keys() and np.random.rand() < self.p:
+            data_dict["color"] *= self.color_augment
+        return data_dict
+
+    def __repr__(self):
+        return "RandomColorDrop(color_augment: {}, p: {})".format(
+            self.color_augment, self.p
+        )
+
+
+@TRANSFORMS.register_module()
+class ElasticDistortion(object):
+    def __init__(self, distortion_params=None):
+        self.distortion_params = (
+            [[0.2, 0.4], [0.8, 1.6]] if distortion_params is None else distortion_params
+        )
+
+    @staticmethod
+    def elastic_distortion(coords, granularity, magnitude):
+        """
+        Apply elastic distortion on sparse coordinate space.
+        pointcloud: numpy array of (number of points, at least 3 spatial dims)
+        granularity: size of the noise grid (in same scale[m/cm] as the voxel grid)
+        magnitude: noise multiplier
+        """
+        blurx = np.ones((3, 1, 1, 1)).astype("float32") / 3
+        blury = np.ones((1, 3, 1, 1)).astype("float32") / 3
+        blurz = np.ones((1, 1, 3, 1)).astype("float32") / 3
+        coords_min = coords.min(0)
+
+        # Create Gaussian noise tensor of the size given by granularity.
+        noise_dim = ((coords - coords_min).max(0) // granularity).astype(int) + 3
+        noise = np.random.randn(*noise_dim, 3).astype(np.float32)
+
+        # Smoothing.
+        for _ in range(2):
+            noise = scipy.ndimage.filters.convolve(
+                noise, blurx, mode="constant", cval=0
+            )
+            noise = scipy.ndimage.filters.convolve(
+                noise, blury, mode="constant", cval=0
+            )
+            noise = scipy.ndimage.filters.convolve(
+                noise, blurz, mode="constant", cval=0
+            )
+
+        # Trilinear interpolate noise filters for each spatial dimensions.
+        ax = [
+            np.linspace(d_min, d_max, d)
+            for d_min, d_max, d in zip(
+                coords_min - granularity,
+                coords_min + granularity * (noise_dim - 2),
+                noise_dim,
+            )
+        ]
+        interp = scipy.interpolate.RegularGridInterpolator(
+            ax, noise, bounds_error=False, fill_value=0
+        )
+        coords += interp(coords) * magnitude
+        return coords
+
+    def __call__(self, data_dict):
+        if "coord" in data_dict.keys() and self.distortion_params is not None:
+            if random.random() < 0.95:
+                for granularity, magnitude in self.distortion_params:
+                    data_dict["coord"] = self.elastic_distortion(
+                        data_dict["coord"], granularity, magnitude
+                    )
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class GridSample(object):
+    def __init__(
+        self,
+        grid_size=0.05,
+        hash_type="fnv",
+        mode="train",
+        return_inverse=False,
+        return_grid_coord=False,
+        return_min_coord=False,
+        return_displacement=False,
+        project_displacement=False,
+    ):
+        self.grid_size = grid_size
+        self.hash = self.fnv_hash_vec if hash_type == "fnv" else self.ravel_hash_vec
+        assert mode in ["train", "test"]
+        self.mode = mode
+        self.return_inverse = return_inverse
+        self.return_grid_coord = return_grid_coord
+        self.return_min_coord = return_min_coord
+        self.return_displacement = return_displacement
+        self.project_displacement = project_displacement
+
+    def __call__(self, data_dict):
+        assert "coord" in data_dict.keys()
+        scaled_coord = data_dict["coord"] / np.array(self.grid_size)
+        grid_coord = np.floor(scaled_coord).astype(int)
+        min_coord = grid_coord.min(0)
+        grid_coord -= min_coord
+        scaled_coord -= min_coord
+        min_coord = min_coord * np.array(self.grid_size)
+        key = self.hash(grid_coord)
+        idx_sort = np.argsort(key)
+        key_sort = key[idx_sort]
+        _, inverse, count = np.unique(key_sort, return_inverse=True, return_counts=True)
+        if self.mode == "train":  # train mode
+            idx_select = (
+                np.cumsum(np.insert(count, 0, 0)[0:-1])
+                + np.random.randint(0, count.max(), count.size) % count
+            )
+            idx_unique = idx_sort[idx_select]
+            if "sampled_index" in data_dict:
+                # for ScanNet data efficient, we need to make sure labeled point is sampled.
+                idx_unique = np.unique(
+                    np.append(idx_unique, data_dict["sampled_index"])
+                )
+                mask = np.zeros_like(data_dict["segment"]).astype(bool)
+                mask[data_dict["sampled_index"]] = True
+                data_dict["sampled_index"] = np.where(mask[idx_unique])[0]
+            data_dict = index_operator(data_dict, idx_unique)
+            if self.return_inverse:
+                data_dict["inverse"] = np.zeros_like(inverse)
+                data_dict["inverse"][idx_sort] = inverse
+            if self.return_grid_coord:
+                data_dict["grid_coord"] = grid_coord[idx_unique]
+                data_dict["index_valid_keys"].append("grid_coord")
+            if self.return_min_coord:
+                data_dict["min_coord"] = min_coord.reshape([1, 3])
+            if self.return_displacement:
+                displacement = (
+                    scaled_coord - grid_coord - 0.5
+                )  # [0, 1] -> [-0.5, 0.5] displacement to center
+                if self.project_displacement:
+                    displacement = np.sum(
+                        displacement * data_dict["normal"], axis=-1, keepdims=True
+                    )
+                data_dict["displacement"] = displacement[idx_unique]
+                data_dict["index_valid_keys"].append("displacement")
+            return data_dict
+
+        elif self.mode == "test":  # test mode
+            data_part_list = []
+            for i in range(count.max()):
+                idx_select = np.cumsum(np.insert(count, 0, 0)[0:-1]) + i % count
+                idx_part = idx_sort[idx_select]
+                data_part = index_operator(data_dict, idx_part, duplicate=True)
+                data_part["index"] = idx_part
+                if self.return_inverse:
+                    data_part["inverse"] = np.zeros_like(inverse)
+                    data_part["inverse"][idx_sort] = inverse
+                if self.return_grid_coord:
+                    data_part["grid_coord"] = grid_coord[idx_part]
+                    data_dict["index_valid_keys"].append("grid_coord")
+                if self.return_min_coord:
+                    data_part["min_coord"] = min_coord.reshape([1, 3])
+                if self.return_displacement:
+                    displacement = (
+                        scaled_coord - grid_coord - 0.5
+                    )  # [0, 1] -> [-0.5, 0.5] displacement to center
+                    if self.project_displacement:
+                        displacement = np.sum(
+                            displacement * data_dict["normal"], axis=-1, keepdims=True
+                        )
+                    data_dict["displacement"] = displacement[idx_part]
+                    data_dict["index_valid_keys"].append("displacement")
+                data_part_list.append(data_part)
+            return data_part_list
+        else:
+            raise NotImplementedError
+
+    @staticmethod
+    def ravel_hash_vec(arr):
+        """
+        Ravel the coordinates after subtracting the min coordinates.
+        """
+        assert arr.ndim == 2
+        arr = arr.copy()
+        arr -= arr.min(0)
+        arr = arr.astype(np.uint64, copy=False)
+        arr_max = arr.max(0).astype(np.uint64) + 1
+
+        keys = np.zeros(arr.shape[0], dtype=np.uint64)
+        # Fortran style indexing
+        for j in range(arr.shape[1] - 1):
+            keys += arr[:, j]
+            keys *= arr_max[j + 1]
+        keys += arr[:, -1]
+        return keys
+
+    @staticmethod
+    def fnv_hash_vec(arr):
+        """
+        FNV64-1A
+        """
+        assert arr.ndim == 2
+        # Floor first for negative coordinates
+        arr = arr.copy()
+        arr = arr.astype(np.uint64, copy=False)
+        hashed_arr = np.uint64(14695981039346656037) * np.ones(
+            arr.shape[0], dtype=np.uint64
+        )
+        for j in range(arr.shape[1]):
+            hashed_arr *= np.uint64(1099511628211)
+            hashed_arr = np.bitwise_xor(hashed_arr, arr[:, j])
+        return hashed_arr
+
+
+@TRANSFORMS.register_module()
+class SphereCrop(object):
+    def __init__(self, point_max=80000, sample_rate=None, mode="random"):
+        self.point_max = point_max
+        self.sample_rate = sample_rate
+        assert mode in ["random", "center", "all"]
+        self.mode = mode
+
+    def __call__(self, data_dict):
+        point_max = (
+            int(self.sample_rate * data_dict["coord"].shape[0])
+            if self.sample_rate is not None
+            else self.point_max
+        )
+
+        assert "coord" in data_dict.keys()
+        if data_dict["coord"].shape[0] > point_max:
+            if self.mode == "random":
+                center = data_dict["coord"][
+                    np.random.randint(data_dict["coord"].shape[0])
+                ]
+            elif self.mode == "center":
+                center = data_dict["coord"][data_dict["coord"].shape[0] // 2]
+            else:
+                raise NotImplementedError
+            idx_crop = np.argsort(np.sum(np.square(data_dict["coord"] - center), 1))[
+                :point_max
+            ]
+            data_dict = index_operator(data_dict, idx_crop)
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class ShufflePoint(object):
+    def __call__(self, data_dict):
+        assert "coord" in data_dict.keys()
+        shuffle_index = np.arange(data_dict["coord"].shape[0])
+        np.random.shuffle(shuffle_index)
+        data_dict = index_operator(data_dict, shuffle_index)
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class CropBoundary(object):
+    def __call__(self, data_dict):
+        assert "segment" in data_dict
+        segment = data_dict["segment"].flatten()
+        mask = (segment != 0) * (segment != 1)
+        data_dict = index_operator(data_dict, mask)
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class ContrastiveViewsGenerator(object):
+    def __init__(
+        self,
+        view_keys=("coord", "color", "normal", "origin_coord"),
+        view_trans_cfg=None,
+    ):
+        self.view_keys = view_keys
+        self.view_trans = Compose(view_trans_cfg)
+
+    def __call__(self, data_dict):
+        view1_dict = dict()
+        view2_dict = dict()
+        for key in self.view_keys:
+            view1_dict[key] = data_dict[key].copy()
+            view2_dict[key] = data_dict[key].copy()
+        view1_dict = self.view_trans(view1_dict)
+        view2_dict = self.view_trans(view2_dict)
+        for key, value in view1_dict.items():
+            data_dict["view1_" + key] = value
+        for key, value in view2_dict.items():
+            data_dict["view2_" + key] = value
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class MultiViewGenerator(object):
+    def __init__(
+        self,
+        global_view_num=2,
+        global_view_scale=(0.4, 1.0),
+        local_view_num=4,
+        local_view_scale=(0.1, 0.4),
+        global_shared_transform=None,
+        global_transform=None,
+        local_transform=None,
+        max_size=65536,
+        center_height_scale=(0, 1),
+        shared_global_view=False,
+        view_keys=("coord", "origin_coord", "color", "normal"),
+    ):
+        self.global_view_num = global_view_num
+        self.global_view_scale = global_view_scale
+        self.local_view_num = local_view_num
+        self.local_view_scale = local_view_scale
+        self.global_shared_transform = Compose(global_shared_transform)
+        self.global_transform = Compose(global_transform)
+        self.local_transform = Compose(local_transform)
+        self.max_size = max_size
+        self.center_height_scale = center_height_scale
+        self.shared_global_view = shared_global_view
+        self.view_keys = view_keys
+        assert "coord" in view_keys
+
+    def get_view(self, point, center, scale):
+        coord = point["coord"]
+        max_size = min(self.max_size, coord.shape[0])
+        size = int(np.random.uniform(*scale) * max_size)
+        index = np.argsort(np.sum(np.square(coord - center), axis=-1))[:size]
+        view = dict(index=index)
+        for key in point.keys():
+            if key in self.view_keys:
+                view[key] = point[key][index]
+
+        if "index_valid_keys" in point.keys():
+            # inherit index_valid_keys from point
+            view["index_valid_keys"] = point["index_valid_keys"]
+        return view
+
+    def __call__(self, data_dict):
+        coord = data_dict["coord"]
+        point = self.global_shared_transform(copy.deepcopy(data_dict))
+        z_min = coord[:, 2].min()
+        z_max = coord[:, 2].max()
+        z_min_ = z_min + (z_max - z_min) * self.center_height_scale[0]
+        z_max_ = z_min + (z_max - z_min) * self.center_height_scale[1]
+        center_mask = np.logical_and(coord[:, 2] >= z_min_, coord[:, 2] <= z_max_)
+        # get major global view
+        major_center = coord[np.random.choice(np.where(center_mask)[0])]
+        major_view = self.get_view(point, major_center, self.global_view_scale)
+        major_coord = major_view["coord"]
+        # get global views: restrict the center of left global view within the major global view
+        if not self.shared_global_view:
+            global_views = [
+                self.get_view(
+                    point=point,
+                    center=major_coord[np.random.randint(major_coord.shape[0])],
+                    scale=self.global_view_scale,
+                )
+                for _ in range(self.global_view_num - 1)
+            ]
+        else:
+            global_views = [
+                {key: value.copy() for key, value in major_view.items()}
+                for _ in range(self.global_view_num - 1)
+            ]
+
+        global_views = [major_view] + global_views
+
+        # get local views: restrict the center of local view within the major global view
+        cover_mask = np.zeros_like(major_view["index"], dtype=bool)
+        local_views = []
+        for i in range(self.local_view_num):
+            if sum(~cover_mask) == 0:
+                # reset cover mask if all points are sampled
+                cover_mask[:] = False
+            local_view = self.get_view(
+                point=data_dict,
+                center=major_coord[np.random.choice(np.where(~cover_mask)[0])],
+                scale=self.local_view_scale,
+            )
+            local_views.append(local_view)
+            cover_mask[np.isin(major_view["index"], local_view["index"])] = True
+
+        # augmentation and concat
+        view_dict = {}
+        for global_view in global_views:
+            global_view.pop("index")
+            global_view = self.global_transform(global_view)
+            for key in self.view_keys:
+                if f"global_{key}" in view_dict.keys():
+                    view_dict[f"global_{key}"].append(global_view[key])
+                else:
+                    view_dict[f"global_{key}"] = [global_view[key]]
+        view_dict["global_offset"] = np.cumsum(
+            [data.shape[0] for data in view_dict["global_coord"]]
+        )
+        for local_view in local_views:
+            local_view.pop("index")
+            local_view = self.local_transform(local_view)
+            for key in self.view_keys:
+                if f"local_{key}" in view_dict.keys():
+                    view_dict[f"local_{key}"].append(local_view[key])
+                else:
+                    view_dict[f"local_{key}"] = [local_view[key]]
+        view_dict["local_offset"] = np.cumsum(
+            [data.shape[0] for data in view_dict["local_coord"]]
+        )
+        for key in view_dict.keys():
+            if "offset" not in key:
+                view_dict[key] = np.concatenate(view_dict[key], axis=0)
+        data_dict.update(view_dict)
+        return data_dict
+
+
+@TRANSFORMS.register_module()
+class InstanceParser(object):
+    def __init__(self, segment_ignore_index=(-1, 0, 1), instance_ignore_index=-1):
+        self.segment_ignore_index = segment_ignore_index
+        self.instance_ignore_index = instance_ignore_index
+
+    def __call__(self, data_dict):
+        coord = data_dict["coord"]
+        segment = data_dict["segment"]
+        instance = data_dict["instance"]
+        mask = ~np.in1d(segment, self.segment_ignore_index)
+        # mapping ignored instance to ignore index
+        instance[~mask] = self.instance_ignore_index
+        # reorder left instance
+        unique, inverse = np.unique(instance[mask], return_inverse=True)
+        instance_num = len(unique)
+        instance[mask] = inverse
+        # init instance information
+        centroid = np.ones((coord.shape[0], 3)) * self.instance_ignore_index
+        bbox = np.ones((instance_num, 8)) * self.instance_ignore_index
+        vacancy = [
+            index for index in self.segment_ignore_index if index >= 0
+        ]  # vacate class index
+
+        for instance_id in range(instance_num):
+            mask_ = instance == instance_id
+            coord_ = coord[mask_]
+            bbox_min = coord_.min(0)
+            bbox_max = coord_.max(0)
+            bbox_centroid = coord_.mean(0)
+            bbox_center = (bbox_max + bbox_min) / 2
+            bbox_size = bbox_max - bbox_min
+            bbox_theta = np.zeros(1, dtype=coord_.dtype)
+            bbox_class = np.array([segment[mask_][0]], dtype=coord_.dtype)
+            # shift class index to fill vacate class index caused by segment ignore index
+            bbox_class -= np.greater(bbox_class, vacancy).sum()
+
+            centroid[mask_] = bbox_centroid
+            bbox[instance_id] = np.concatenate(
+                [bbox_center, bbox_size, bbox_theta, bbox_class]
+            )  # 3 + 3 + 1 + 1 = 8
+        data_dict["instance"] = instance
+        data_dict["instance_centroid"] = centroid
+        data_dict["bbox"] = bbox
+        return data_dict
+
+
+class Compose(object):
+    def __init__(self, cfg=None):
+        self.cfg = cfg if cfg is not None else []
+        self.transforms = []
+        for t_cfg in self.cfg:
+            self.transforms.append(TRANSFORMS.build(t_cfg))
+
+    def __call__(self, data_dict):
+        for t in self.transforms:
+            data_dict = t(data_dict)
+        return data_dict
+
+
+def default(scale = 1.0, apply_z_positive = True, normalize_coord = False):
+    config = [
+        *([dict(type="NormalizeCoord")] if normalize_coord else []),
+        dict(type="RandomScale", scale=[scale, scale]),
+        *([dict(type="CenterShift", apply_z=True)] if apply_z_positive else []),
+        dict(
+            type="GridSample",
+            grid_size=0.01,
+            hash_type="fnv",
+            mode="train",
+            return_grid_coord=True,
+            return_inverse=True,
+        ),
+        dict(type="NormalizeColor"),
+        dict(type="ToTensor"),
+        dict(
+            type="Collect",
+            keys=("coord", "grid_coord", "color", "inverse"),
+            feat_keys=("coord", "color", "normal"),
+        ),
+    ]
+    return Compose(config)

utonia/utils.py

@@ -0,0 +1,75 @@
+"""
+General utils
+
+Author: Xiaoyang Wu (xiaoyang.wu.cs@gmail.com)
+Please cite our work if the code is helpful to you.
+"""
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import os
+import random
+import numpy as np
+import torch
+import torch.backends.cudnn as cudnn
+from datetime import datetime
+
+
+@torch.no_grad()
+def offset2bincount(offset):
+    return torch.diff(
+        offset, prepend=torch.tensor([0], device=offset.device, dtype=torch.long)
+    )
+
+
+@torch.no_grad()
+def bincount2offset(bincount):
+    return torch.cumsum(bincount, dim=0)
+
+
+@torch.no_grad()
+def offset2batch(offset):
+    bincount = offset2bincount(offset)
+    return torch.arange(
+        len(bincount), device=offset.device, dtype=torch.long
+    ).repeat_interleave(bincount)
+
+
+@torch.no_grad()
+def batch2offset(batch):
+    return torch.cumsum(batch.bincount(), dim=0).long()
+
+
+def get_random_seed():
+    seed = (
+        os.getpid()
+        + int(datetime.now().strftime("%S%f"))
+        + int.from_bytes(os.urandom(2), "big")
+    )
+    return seed
+
+
+def set_seed(seed=None):
+    if seed is None:
+        seed = get_random_seed()
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    cudnn.benchmark = False
+    cudnn.deterministic = True
+    os.environ["PYTHONHASHSEED"] = str(seed)

vggt/heads/camera_head.py

@@ -0,0 +1,162 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from vggt.layers import Mlp
+from vggt.layers.block import Block
+from vggt.heads.head_act import activate_pose
+
+
+class CameraHead(nn.Module):
+    """
+    CameraHead predicts camera parameters from token representations using iterative refinement.
+
+    It applies a series of transformer blocks (the "trunk") to dedicated camera tokens.
+    """
+
+    def __init__(
+        self,
+        dim_in: int = 2048,
+        trunk_depth: int = 4,
+        pose_encoding_type: str = "absT_quaR_FoV",
+        num_heads: int = 16,
+        mlp_ratio: int = 4,
+        init_values: float = 0.01,
+        trans_act: str = "linear",
+        quat_act: str = "linear",
+        fl_act: str = "relu",  # Field of view activations: ensures FOV values are positive.
+    ):
+        super().__init__()
+
+        if pose_encoding_type == "absT_quaR_FoV":
+            self.target_dim = 9
+        else:
+            raise ValueError(f"Unsupported camera encoding type: {pose_encoding_type}")
+
+        self.trans_act = trans_act
+        self.quat_act = quat_act
+        self.fl_act = fl_act
+        self.trunk_depth = trunk_depth
+
+        # Build the trunk using a sequence of transformer blocks.
+        self.trunk = nn.Sequential(
+            *[
+                Block(
+                    dim=dim_in,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    init_values=init_values,
+                )
+                for _ in range(trunk_depth)
+            ]
+        )
+
+        # Normalizations for camera token and trunk output.
+        self.token_norm = nn.LayerNorm(dim_in)
+        self.trunk_norm = nn.LayerNorm(dim_in)
+
+        # Learnable empty camera pose token.
+        self.empty_pose_tokens = nn.Parameter(torch.zeros(1, 1, self.target_dim))
+        self.embed_pose = nn.Linear(self.target_dim, dim_in)
+
+        # Module for producing modulation parameters: shift, scale, and a gate.
+        self.poseLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim_in, 3 * dim_in, bias=True))
+
+        # Adaptive layer normalization without affine parameters.
+        self.adaln_norm = nn.LayerNorm(dim_in, elementwise_affine=False, eps=1e-6)
+        self.pose_branch = Mlp(
+            in_features=dim_in,
+            hidden_features=dim_in // 2,
+            out_features=self.target_dim,
+            drop=0,
+        )
+
+    def forward(self, aggregated_tokens_list: list, num_iterations: int = 4) -> list:
+        """
+        Forward pass to predict camera parameters.
+
+        Args:
+            aggregated_tokens_list (list): List of token tensors from the network;
+                the last tensor is used for prediction.
+            num_iterations (int, optional): Number of iterative refinement steps. Defaults to 4.
+
+        Returns:
+            list: A list of predicted camera encodings (post-activation) from each iteration.
+        """
+        # Use tokens from the last block for camera prediction.
+        tokens = aggregated_tokens_list[-1]
+
+        # Extract the camera tokens
+        pose_tokens = tokens[:, :, 0]
+        pose_tokens = self.token_norm(pose_tokens)
+
+        pred_pose_enc_list = self.trunk_fn(pose_tokens, num_iterations)
+        return pred_pose_enc_list
+
+    def trunk_fn(self, pose_tokens: torch.Tensor, num_iterations: int) -> list:
+        """
+        Iteratively refine camera pose predictions.
+
+        Args:
+            pose_tokens (torch.Tensor): Normalized camera tokens with shape [B, 1, C].
+            num_iterations (int): Number of refinement iterations.
+
+        Returns:
+            list: List of activated camera encodings from each iteration.
+        """
+        B, S, C = pose_tokens.shape  # S is expected to be 1.
+        pred_pose_enc = None
+        pred_pose_enc_list = []
+
+        for _ in range(num_iterations):
+            # Use a learned empty pose for the first iteration.
+            if pred_pose_enc is None:
+                module_input = self.embed_pose(self.empty_pose_tokens.expand(B, S, -1))
+            else:
+                # Detach the previous prediction to avoid backprop through time.
+                pred_pose_enc = pred_pose_enc.detach()
+                module_input = self.embed_pose(pred_pose_enc)
+
+            # Generate modulation parameters and split them into shift, scale, and gate components.
+            shift_msa, scale_msa, gate_msa = self.poseLN_modulation(module_input).chunk(3, dim=-1)
+
+            # Adaptive layer normalization and modulation.
+            pose_tokens_modulated = gate_msa * modulate(self.adaln_norm(pose_tokens), shift_msa, scale_msa)
+            pose_tokens_modulated = pose_tokens_modulated + pose_tokens
+
+            pose_tokens_modulated = self.trunk(pose_tokens_modulated)
+            # Compute the delta update for the pose encoding.
+            pred_pose_enc_delta = self.pose_branch(self.trunk_norm(pose_tokens_modulated))
+
+            if pred_pose_enc is None:
+                pred_pose_enc = pred_pose_enc_delta
+            else:
+                pred_pose_enc = pred_pose_enc + pred_pose_enc_delta
+
+            # Apply final activation functions for translation, quaternion, and field-of-view.
+            activated_pose = activate_pose(
+                pred_pose_enc,
+                trans_act=self.trans_act,
+                quat_act=self.quat_act,
+                fl_act=self.fl_act,
+            )
+            pred_pose_enc_list.append(activated_pose)
+
+        return pred_pose_enc_list
+
+
+def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
+    """
+    Modulate the input tensor using scaling and shifting parameters.
+    """
+    # modified from https://github.com/facebookresearch/DiT/blob/796c29e532f47bba17c5b9c5eb39b9354b8b7c64/models.py#L19
+    return x * (1 + scale) + shift

vggt/heads/dpt_head.py

@@ -0,0 +1,497 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+# Inspired by https://github.com/DepthAnything/Depth-Anything-V2
+
+
+import os
+from typing import List, Dict, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .head_act import activate_head
+from .utils import create_uv_grid, position_grid_to_embed
+
+
+class DPTHead(nn.Module):
+    """
+    DPT  Head for dense prediction tasks.
+
+    This implementation follows the architecture described in "Vision Transformers for Dense Prediction"
+    (https://arxiv.org/abs/2103.13413). The DPT head processes features from a vision transformer
+    backbone and produces dense predictions by fusing multi-scale features.
+
+    Args:
+        dim_in (int): Input dimension (channels).
+        patch_size (int, optional): Patch size. Default is 14.
+        output_dim (int, optional): Number of output channels. Default is 4.
+        activation (str, optional): Activation type. Default is "inv_log".
+        conf_activation (str, optional): Confidence activation type. Default is "expp1".
+        features (int, optional): Feature channels for intermediate representations. Default is 256.
+        out_channels (List[int], optional): Output channels for each intermediate layer.
+        intermediate_layer_idx (List[int], optional): Indices of layers from aggregated tokens used for DPT.
+        pos_embed (bool, optional): Whether to use positional embedding. Default is True.
+        feature_only (bool, optional): If True, return features only without the last several layers and activation head. Default is False.
+        down_ratio (int, optional): Downscaling factor for the output resolution. Default is 1.
+    """
+
+    def __init__(
+        self,
+        dim_in: int,
+        patch_size: int = 14,
+        output_dim: int = 4,
+        activation: str = "inv_log",
+        conf_activation: str = "expp1",
+        features: int = 256,
+        out_channels: List[int] = [256, 512, 1024, 1024],
+        intermediate_layer_idx: List[int] = [4, 11, 17, 23],
+        pos_embed: bool = True,
+        feature_only: bool = False,
+        down_ratio: int = 1,
+    ) -> None:
+        super(DPTHead, self).__init__()
+        self.patch_size = patch_size
+        self.activation = activation
+        self.conf_activation = conf_activation
+        self.pos_embed = pos_embed
+        self.feature_only = feature_only
+        self.down_ratio = down_ratio
+        self.intermediate_layer_idx = intermediate_layer_idx
+
+        self.norm = nn.LayerNorm(dim_in)
+
+        # Projection layers for each output channel from tokens.
+        self.projects = nn.ModuleList(
+            [
+                nn.Conv2d(
+                    in_channels=dim_in,
+                    out_channels=oc,
+                    kernel_size=1,
+                    stride=1,
+                    padding=0,
+                )
+                for oc in out_channels
+            ]
+        )
+
+        # Resize layers for upsampling feature maps.
+        self.resize_layers = nn.ModuleList(
+            [
+                nn.ConvTranspose2d(
+                    in_channels=out_channels[0], out_channels=out_channels[0], kernel_size=4, stride=4, padding=0
+                ),
+                nn.ConvTranspose2d(
+                    in_channels=out_channels[1], out_channels=out_channels[1], kernel_size=2, stride=2, padding=0
+                ),
+                nn.Identity(),
+                nn.Conv2d(
+                    in_channels=out_channels[3], out_channels=out_channels[3], kernel_size=3, stride=2, padding=1
+                ),
+            ]
+        )
+
+        self.scratch = _make_scratch(
+            out_channels,
+            features,
+            expand=False,
+        )
+
+        # Attach additional modules to scratch.
+        self.scratch.stem_transpose = None
+        self.scratch.refinenet1 = _make_fusion_block(features)
+        self.scratch.refinenet2 = _make_fusion_block(features)
+        self.scratch.refinenet3 = _make_fusion_block(features)
+        self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False)
+
+        head_features_1 = features
+        head_features_2 = 32
+
+        if feature_only:
+            self.scratch.output_conv1 = nn.Conv2d(head_features_1, head_features_1, kernel_size=3, stride=1, padding=1)
+        else:
+            self.scratch.output_conv1 = nn.Conv2d(
+                head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1
+            )
+            conv2_in_channels = head_features_1 // 2
+
+            self.scratch.output_conv2 = nn.Sequential(
+                nn.Conv2d(conv2_in_channels, head_features_2, kernel_size=3, stride=1, padding=1),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0),
+            )
+
+    def forward(
+        self,
+        aggregated_tokens_list: List[torch.Tensor],
+        images: torch.Tensor,
+        patch_start_idx: int,
+        frames_chunk_size: int = 8,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Forward pass through the DPT head, supports processing by chunking frames.
+        Args:
+            aggregated_tokens_list (List[Tensor]): List of token tensors from different transformer layers.
+            images (Tensor): Input images with shape [B, S, 3, H, W], in range [0, 1].
+            patch_start_idx (int): Starting index for patch tokens in the token sequence.
+                Used to separate patch tokens from other tokens (e.g., camera or register tokens).
+            frames_chunk_size (int, optional): Number of frames to process in each chunk.
+                If None or larger than S, all frames are processed at once. Default: 8.
+
+        Returns:
+            Tensor or Tuple[Tensor, Tensor]:
+                - If feature_only=True: Feature maps with shape [B, S, C, H, W]
+                - Otherwise: Tuple of (predictions, confidence) both with shape [B, S, 1, H, W]
+        """
+        B, S, _, H, W = images.shape
+
+        # If frames_chunk_size is not specified or greater than S, process all frames at once
+        if frames_chunk_size is None or frames_chunk_size >= S:
+            return self._forward_impl(aggregated_tokens_list, images, patch_start_idx)
+
+        # Otherwise, process frames in chunks to manage memory usage
+        assert frames_chunk_size > 0
+
+        # Process frames in batches
+        all_preds = []
+        all_conf = []
+
+        for frames_start_idx in range(0, S, frames_chunk_size):
+            frames_end_idx = min(frames_start_idx + frames_chunk_size, S)
+
+            # Process batch of frames
+            if self.feature_only:
+                chunk_output = self._forward_impl(
+                    aggregated_tokens_list, images, patch_start_idx, frames_start_idx, frames_end_idx
+                )
+                all_preds.append(chunk_output)
+            else:
+                chunk_preds, chunk_conf = self._forward_impl(
+                    aggregated_tokens_list, images, patch_start_idx, frames_start_idx, frames_end_idx
+                )
+                all_preds.append(chunk_preds)
+                all_conf.append(chunk_conf)
+
+        # Concatenate results along the sequence dimension
+        if self.feature_only:
+            return torch.cat(all_preds, dim=1)
+        else:
+            return torch.cat(all_preds, dim=1), torch.cat(all_conf, dim=1)
+
+    def _forward_impl(
+        self,
+        aggregated_tokens_list: List[torch.Tensor],
+        images: torch.Tensor,
+        patch_start_idx: int,
+        frames_start_idx: int = None,
+        frames_end_idx: int = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Implementation of the forward pass through the DPT head.
+
+        This method processes a specific chunk of frames from the sequence.
+
+        Args:
+            aggregated_tokens_list (List[Tensor]): List of token tensors from different transformer layers.
+            images (Tensor): Input images with shape [B, S, 3, H, W].
+            patch_start_idx (int): Starting index for patch tokens.
+            frames_start_idx (int, optional): Starting index for frames to process.
+            frames_end_idx (int, optional): Ending index for frames to process.
+
+        Returns:
+            Tensor or Tuple[Tensor, Tensor]: Feature maps or (predictions, confidence).
+        """
+        if frames_start_idx is not None and frames_end_idx is not None:
+            images = images[:, frames_start_idx:frames_end_idx]
+
+        B, S, _, H, W = images.shape
+
+        patch_h, patch_w = H // self.patch_size, W // self.patch_size
+
+        out = []
+        dpt_idx = 0
+
+        for layer_idx in self.intermediate_layer_idx:
+            x = aggregated_tokens_list[layer_idx][:, :, patch_start_idx:]
+
+            # Select frames if processing a chunk
+            if frames_start_idx is not None and frames_end_idx is not None:
+                x = x[:, frames_start_idx:frames_end_idx]
+
+            x = x.view(B * S, -1, x.shape[-1])
+
+            x = self.norm(x)
+
+            x = x.permute(0, 2, 1).reshape((x.shape[0], x.shape[-1], patch_h, patch_w))
+
+            x = self.projects[dpt_idx](x)
+            if self.pos_embed:
+                x = self._apply_pos_embed(x, W, H)
+            x = self.resize_layers[dpt_idx](x)
+
+            out.append(x)
+            dpt_idx += 1
+
+        # Fuse features from multiple layers.
+        out = self.scratch_forward(out)
+        # Interpolate fused output to match target image resolution.
+        out = custom_interpolate(
+            out,
+            (int(patch_h * self.patch_size / self.down_ratio), int(patch_w * self.patch_size / self.down_ratio)),
+            mode="bilinear",
+            align_corners=True,
+        )
+
+        if self.pos_embed:
+            out = self._apply_pos_embed(out, W, H)
+
+        if self.feature_only:
+            return out.view(B, S, *out.shape[1:])
+
+        out = self.scratch.output_conv2(out)
+        preds, conf = activate_head(out, activation=self.activation, conf_activation=self.conf_activation)
+
+        preds = preds.view(B, S, *preds.shape[1:])
+        conf = conf.view(B, S, *conf.shape[1:])
+        return preds, conf
+
+    def _apply_pos_embed(self, x: torch.Tensor, W: int, H: int, ratio: float = 0.1) -> torch.Tensor:
+        """
+        Apply positional embedding to tensor x.
+        """
+        patch_w = x.shape[-1]
+        patch_h = x.shape[-2]
+        pos_embed = create_uv_grid(patch_w, patch_h, aspect_ratio=W / H, dtype=x.dtype, device=x.device)
+        pos_embed = position_grid_to_embed(pos_embed, x.shape[1])
+        pos_embed = pos_embed * ratio
+        pos_embed = pos_embed.permute(2, 0, 1)[None].expand(x.shape[0], -1, -1, -1)
+        return x + pos_embed
+
+    def scratch_forward(self, features: List[torch.Tensor]) -> torch.Tensor:
+        """
+        Forward pass through the fusion blocks.
+
+        Args:
+            features (List[Tensor]): List of feature maps from different layers.
+
+        Returns:
+            Tensor: Fused feature map.
+        """
+        layer_1, layer_2, layer_3, layer_4 = features
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        out = self.scratch.refinenet4(layer_4_rn, size=layer_3_rn.shape[2:])
+        del layer_4_rn, layer_4
+
+        out = self.scratch.refinenet3(out, layer_3_rn, size=layer_2_rn.shape[2:])
+        del layer_3_rn, layer_3
+
+        out = self.scratch.refinenet2(out, layer_2_rn, size=layer_1_rn.shape[2:])
+        del layer_2_rn, layer_2
+
+        out = self.scratch.refinenet1(out, layer_1_rn)
+        del layer_1_rn, layer_1
+
+        out = self.scratch.output_conv1(out)
+        return out
+
+
+################################################################################
+# Modules
+################################################################################
+
+
+def _make_fusion_block(features: int, size: int = None, has_residual: bool = True, groups: int = 1) -> nn.Module:
+    return FeatureFusionBlock(
+        features,
+        nn.ReLU(inplace=True),
+        deconv=False,
+        bn=False,
+        expand=False,
+        align_corners=True,
+        size=size,
+        has_residual=has_residual,
+        groups=groups,
+    )
+
+
+def _make_scratch(in_shape: List[int], out_shape: int, groups: int = 1, expand: bool = False) -> nn.Module:
+    scratch = nn.Module()
+    out_shape1 = out_shape
+    out_shape2 = out_shape
+    out_shape3 = out_shape
+    if len(in_shape) >= 4:
+        out_shape4 = out_shape
+
+    if expand:
+        out_shape1 = out_shape
+        out_shape2 = out_shape * 2
+        out_shape3 = out_shape * 4
+        if len(in_shape) >= 4:
+            out_shape4 = out_shape * 8
+
+    scratch.layer1_rn = nn.Conv2d(
+        in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer2_rn = nn.Conv2d(
+        in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer3_rn = nn.Conv2d(
+        in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    if len(in_shape) >= 4:
+        scratch.layer4_rn = nn.Conv2d(
+            in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+        )
+    return scratch
+
+
+class ResidualConvUnit(nn.Module):
+    """Residual convolution module."""
+
+    def __init__(self, features, activation, bn, groups=1):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.bn = bn
+        self.groups = groups
+        self.conv1 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups)
+        self.conv2 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups)
+
+        self.norm1 = None
+        self.norm2 = None
+
+        self.activation = activation
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+
+        out = self.activation(x)
+        out = self.conv1(out)
+        if self.norm1 is not None:
+            out = self.norm1(out)
+
+        out = self.activation(out)
+        out = self.conv2(out)
+        if self.norm2 is not None:
+            out = self.norm2(out)
+
+        return self.skip_add.add(out, x)
+
+
+class FeatureFusionBlock(nn.Module):
+    """Feature fusion block."""
+
+    def __init__(
+        self,
+        features,
+        activation,
+        deconv=False,
+        bn=False,
+        expand=False,
+        align_corners=True,
+        size=None,
+        has_residual=True,
+        groups=1,
+    ):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock, self).__init__()
+
+        self.deconv = deconv
+        self.align_corners = align_corners
+        self.groups = groups
+        self.expand = expand
+        out_features = features
+        if self.expand == True:
+            out_features = features // 2
+
+        self.out_conv = nn.Conv2d(
+            features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=self.groups
+        )
+
+        if has_residual:
+            self.resConfUnit1 = ResidualConvUnit(features, activation, bn, groups=self.groups)
+
+        self.has_residual = has_residual
+        self.resConfUnit2 = ResidualConvUnit(features, activation, bn, groups=self.groups)
+
+        self.skip_add = nn.quantized.FloatFunctional()
+        self.size = size
+
+    def forward(self, *xs, size=None):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if self.has_residual:
+            res = self.resConfUnit1(xs[1])
+            output = self.skip_add.add(output, res)
+
+        output = self.resConfUnit2(output)
+
+        if (size is None) and (self.size is None):
+            modifier = {"scale_factor": 2}
+        elif size is None:
+            modifier = {"size": self.size}
+        else:
+            modifier = {"size": size}
+
+        output = custom_interpolate(output, **modifier, mode="bilinear", align_corners=self.align_corners)
+        output = self.out_conv(output)
+
+        return output
+
+
+def custom_interpolate(
+    x: torch.Tensor,
+    size: Tuple[int, int] = None,
+    scale_factor: float = None,
+    mode: str = "bilinear",
+    align_corners: bool = True,
+) -> torch.Tensor:
+    """
+    Custom interpolate to avoid INT_MAX issues in nn.functional.interpolate.
+    """
+    if size is None:
+        size = (int(x.shape[-2] * scale_factor), int(x.shape[-1] * scale_factor))
+
+    INT_MAX = 1610612736
+
+    input_elements = size[0] * size[1] * x.shape[0] * x.shape[1]
+
+    if input_elements > INT_MAX:
+        chunks = torch.chunk(x, chunks=(input_elements // INT_MAX) + 1, dim=0)
+        interpolated_chunks = [
+            nn.functional.interpolate(chunk, size=size, mode=mode, align_corners=align_corners) for chunk in chunks
+        ]
+        x = torch.cat(interpolated_chunks, dim=0)
+        return x.contiguous()
+    else:
+        return nn.functional.interpolate(x, size=size, mode=mode, align_corners=align_corners)

vggt/heads/head_act.py

@@ -0,0 +1,125 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import torch
+import torch.nn.functional as F
+
+
+def activate_pose(pred_pose_enc, trans_act="linear", quat_act="linear", fl_act="linear"):
+    """
+    Activate pose parameters with specified activation functions.
+
+    Args:
+        pred_pose_enc: Tensor containing encoded pose parameters [translation, quaternion, focal length]
+        trans_act: Activation type for translation component
+        quat_act: Activation type for quaternion component
+        fl_act: Activation type for focal length component
+
+    Returns:
+        Activated pose parameters tensor
+    """
+    T = pred_pose_enc[..., :3]
+    quat = pred_pose_enc[..., 3:7]
+    fl = pred_pose_enc[..., 7:]  # or fov
+
+    T = base_pose_act(T, trans_act)
+    quat = base_pose_act(quat, quat_act)
+    fl = base_pose_act(fl, fl_act)  # or fov
+
+    pred_pose_enc = torch.cat([T, quat, fl], dim=-1)
+
+    return pred_pose_enc
+
+
+def base_pose_act(pose_enc, act_type="linear"):
+    """
+    Apply basic activation function to pose parameters.
+
+    Args:
+        pose_enc: Tensor containing encoded pose parameters
+        act_type: Activation type ("linear", "inv_log", "exp", "relu")
+
+    Returns:
+        Activated pose parameters
+    """
+    if act_type == "linear":
+        return pose_enc
+    elif act_type == "inv_log":
+        return inverse_log_transform(pose_enc)
+    elif act_type == "exp":
+        return torch.exp(pose_enc)
+    elif act_type == "relu":
+        return F.relu(pose_enc)
+    else:
+        raise ValueError(f"Unknown act_type: {act_type}")
+
+
+def activate_head(out, activation="norm_exp", conf_activation="expp1"):
+    """
+    Process network output to extract 3D points and confidence values.
+
+    Args:
+        out: Network output tensor (B, C, H, W)
+        activation: Activation type for 3D points
+        conf_activation: Activation type for confidence values
+
+    Returns:
+        Tuple of (3D points tensor, confidence tensor)
+    """
+    # Move channels from last dim to the 4th dimension => (B, H, W, C)
+    fmap = out.permute(0, 2, 3, 1)  # B,H,W,C expected
+
+    # Split into xyz (first C-1 channels) and confidence (last channel)
+    xyz = fmap[:, :, :, :-1]
+    conf = fmap[:, :, :, -1]
+
+    if activation == "norm_exp":
+        d = xyz.norm(dim=-1, keepdim=True).clamp(min=1e-8)
+        xyz_normed = xyz / d
+        pts3d = xyz_normed * torch.expm1(d)
+    elif activation == "norm":
+        pts3d = xyz / xyz.norm(dim=-1, keepdim=True)
+    elif activation == "exp":
+        pts3d = torch.exp(xyz)
+    elif activation == "relu":
+        pts3d = F.relu(xyz)
+    elif activation == "inv_log":
+        pts3d = inverse_log_transform(xyz)
+    elif activation == "xy_inv_log":
+        xy, z = xyz.split([2, 1], dim=-1)
+        z = inverse_log_transform(z)
+        pts3d = torch.cat([xy * z, z], dim=-1)
+    elif activation == "sigmoid":
+        pts3d = torch.sigmoid(xyz)
+    elif activation == "linear":
+        pts3d = xyz
+    else:
+        raise ValueError(f"Unknown activation: {activation}")
+
+    if conf_activation == "expp1":
+        conf_out = 1 + conf.exp()
+    elif conf_activation == "expp0":
+        conf_out = conf.exp()
+    elif conf_activation == "sigmoid":
+        conf_out = torch.sigmoid(conf)
+    else:
+        raise ValueError(f"Unknown conf_activation: {conf_activation}")
+
+    return pts3d, conf_out
+
+
+def inverse_log_transform(y):
+    """
+    Apply inverse log transform: sign(y) * (exp(|y|) - 1)
+
+    Args:
+        y: Input tensor
+
+    Returns:
+        Transformed tensor
+    """
+    return torch.sign(y) * (torch.expm1(torch.abs(y)))

vggt/heads/track_head.py

@@ -0,0 +1,108 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch.nn as nn
+from .dpt_head import DPTHead
+from .track_modules.base_track_predictor import BaseTrackerPredictor
+
+
+class TrackHead(nn.Module):
+    """
+    Track head that uses DPT head to process tokens and BaseTrackerPredictor for tracking.
+    The tracking is performed iteratively, refining predictions over multiple iterations.
+    """
+
+    def __init__(
+        self,
+        dim_in,
+        patch_size=14,
+        features=128,
+        iters=4,
+        predict_conf=True,
+        stride=2,
+        corr_levels=7,
+        corr_radius=4,
+        hidden_size=384,
+    ):
+        """
+        Initialize the TrackHead module.
+
+        Args:
+            dim_in (int): Input dimension of tokens from the backbone.
+            patch_size (int): Size of image patches used in the vision transformer.
+            features (int): Number of feature channels in the feature extractor output.
+            iters (int): Number of refinement iterations for tracking predictions.
+            predict_conf (bool): Whether to predict confidence scores for tracked points.
+            stride (int): Stride value for the tracker predictor.
+            corr_levels (int): Number of correlation pyramid levels
+            corr_radius (int): Radius for correlation computation, controlling the search area.
+            hidden_size (int): Size of hidden layers in the tracker network.
+        """
+        super().__init__()
+
+        self.patch_size = patch_size
+
+        # Feature extractor based on DPT architecture
+        # Processes tokens into feature maps for tracking
+        self.feature_extractor = DPTHead(
+            dim_in=dim_in,
+            patch_size=patch_size,
+            features=features,
+            feature_only=True,  # Only output features, no activation
+            down_ratio=2,  # Reduces spatial dimensions by factor of 2
+            pos_embed=False,
+        )
+
+        # Tracker module that predicts point trajectories
+        # Takes feature maps and predicts coordinates and visibility
+        self.tracker = BaseTrackerPredictor(
+            latent_dim=features,  # Match the output_dim of feature extractor
+            predict_conf=predict_conf,
+            stride=stride,
+            corr_levels=corr_levels,
+            corr_radius=corr_radius,
+            hidden_size=hidden_size,
+        )
+
+        self.iters = iters
+
+    def forward(self, aggregated_tokens_list, images, patch_start_idx, query_points=None, iters=None):
+        """
+        Forward pass of the TrackHead.
+
+        Args:
+            aggregated_tokens_list (list): List of aggregated tokens from the backbone.
+            images (torch.Tensor): Input images of shape (B, S, C, H, W) where:
+                                   B = batch size, S = sequence length.
+            patch_start_idx (int): Starting index for patch tokens.
+            query_points (torch.Tensor, optional): Initial query points to track.
+                                                  If None, points are initialized by the tracker.
+            iters (int, optional): Number of refinement iterations. If None, uses self.iters.
+
+        Returns:
+            tuple:
+                - coord_preds (torch.Tensor): Predicted coordinates for tracked points.
+                - vis_scores (torch.Tensor): Visibility scores for tracked points.
+                - conf_scores (torch.Tensor): Confidence scores for tracked points (if predict_conf=True).
+        """
+        B, S, _, H, W = images.shape
+
+        # Extract features from tokens
+        # feature_maps has shape (B, S, C, H//2, W//2) due to down_ratio=2
+        feature_maps = self.feature_extractor(aggregated_tokens_list, images, patch_start_idx)
+
+        # Use default iterations if not specified
+        if iters is None:
+            iters = self.iters
+
+        # Perform tracking using the extracted features
+        coord_preds, vis_scores, conf_scores = self.tracker(
+            query_points=query_points,
+            fmaps=feature_maps,
+            iters=iters,
+        )
+
+        return coord_preds, vis_scores, conf_scores

vggt/heads/track_modules/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

vggt/heads/track_modules/base_track_predictor.py

@@ -0,0 +1,209 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+from einops import rearrange, repeat
+
+
+from .blocks import EfficientUpdateFormer, CorrBlock
+from .utils import sample_features4d, get_2d_embedding, get_2d_sincos_pos_embed
+from .modules import Mlp
+
+
+class BaseTrackerPredictor(nn.Module):
+    def __init__(
+        self,
+        stride=1,
+        corr_levels=5,
+        corr_radius=4,
+        latent_dim=128,
+        hidden_size=384,
+        use_spaceatt=True,
+        depth=6,
+        max_scale=518,
+        predict_conf=True,
+    ):
+        super(BaseTrackerPredictor, self).__init__()
+        """
+        The base template to create a track predictor
+        
+        Modified from https://github.com/facebookresearch/co-tracker/
+        and https://github.com/facebookresearch/vggsfm
+        """
+
+        self.stride = stride
+        self.latent_dim = latent_dim
+        self.corr_levels = corr_levels
+        self.corr_radius = corr_radius
+        self.hidden_size = hidden_size
+        self.max_scale = max_scale
+        self.predict_conf = predict_conf
+
+        self.flows_emb_dim = latent_dim // 2
+
+        self.corr_mlp = Mlp(
+            in_features=self.corr_levels * (self.corr_radius * 2 + 1) ** 2,
+            hidden_features=self.hidden_size,
+            out_features=self.latent_dim,
+        )
+
+        self.transformer_dim = self.latent_dim + self.latent_dim + self.latent_dim + 4
+
+        self.query_ref_token = nn.Parameter(torch.randn(1, 2, self.transformer_dim))
+
+        space_depth = depth if use_spaceatt else 0
+        time_depth = depth
+
+        self.updateformer = EfficientUpdateFormer(
+            space_depth=space_depth,
+            time_depth=time_depth,
+            input_dim=self.transformer_dim,
+            hidden_size=self.hidden_size,
+            output_dim=self.latent_dim + 2,
+            mlp_ratio=4.0,
+            add_space_attn=use_spaceatt,
+        )
+
+        self.fmap_norm = nn.LayerNorm(self.latent_dim)
+        self.ffeat_norm = nn.GroupNorm(1, self.latent_dim)
+
+        # A linear layer to update track feats at each iteration
+        self.ffeat_updater = nn.Sequential(nn.Linear(self.latent_dim, self.latent_dim), nn.GELU())
+
+        self.vis_predictor = nn.Sequential(nn.Linear(self.latent_dim, 1))
+
+        if predict_conf:
+            self.conf_predictor = nn.Sequential(nn.Linear(self.latent_dim, 1))
+
+    def forward(self, query_points, fmaps=None, iters=6, return_feat=False, down_ratio=1, apply_sigmoid=True):
+        """
+        query_points: B x N x 2, the number of batches, tracks, and xy
+        fmaps: B x S x C x HH x WW, the number of batches, frames, and feature dimension.
+                note HH and WW is the size of feature maps instead of original images
+        """
+        B, N, D = query_points.shape
+        B, S, C, HH, WW = fmaps.shape
+
+        assert D == 2, "Input points must be 2D coordinates"
+
+        # apply a layernorm to fmaps here
+        fmaps = self.fmap_norm(fmaps.permute(0, 1, 3, 4, 2))
+        fmaps = fmaps.permute(0, 1, 4, 2, 3)
+
+        # Scale the input query_points because we may downsample the images
+        # by down_ratio or self.stride
+        # e.g., if a 3x1024x1024 image is processed to a 128x256x256 feature map
+        # its query_points should be query_points/4
+        if down_ratio > 1:
+            query_points = query_points / float(down_ratio)
+
+        query_points = query_points / float(self.stride)
+
+        # Init with coords as the query points
+        # It means the search will start from the position of query points at the reference frames
+        coords = query_points.clone().reshape(B, 1, N, 2).repeat(1, S, 1, 1)
+
+        # Sample/extract the features of the query points in the query frame
+        query_track_feat = sample_features4d(fmaps[:, 0], coords[:, 0])
+
+        # init track feats by query feats
+        track_feats = query_track_feat.unsqueeze(1).repeat(1, S, 1, 1)  # B, S, N, C
+        # back up the init coords
+        coords_backup = coords.clone()
+
+        fcorr_fn = CorrBlock(fmaps, num_levels=self.corr_levels, radius=self.corr_radius)
+
+        coord_preds = []
+
+        # Iterative Refinement
+        for _ in range(iters):
+            # Detach the gradients from the last iteration
+            # (in my experience, not very important for performance)
+            coords = coords.detach()
+
+            fcorrs = fcorr_fn.corr_sample(track_feats, coords)
+
+            corr_dim = fcorrs.shape[3]
+            fcorrs_ = fcorrs.permute(0, 2, 1, 3).reshape(B * N, S, corr_dim)
+            fcorrs_ = self.corr_mlp(fcorrs_)
+
+            # Movement of current coords relative to query points
+            flows = (coords - coords[:, 0:1]).permute(0, 2, 1, 3).reshape(B * N, S, 2)
+
+            flows_emb = get_2d_embedding(flows, self.flows_emb_dim, cat_coords=False)
+
+            # (In my trials, it is also okay to just add the flows_emb instead of concat)
+            flows_emb = torch.cat([flows_emb, flows / self.max_scale, flows / self.max_scale], dim=-1)
+
+            track_feats_ = track_feats.permute(0, 2, 1, 3).reshape(B * N, S, self.latent_dim)
+
+            # Concatenate them as the input for the transformers
+            transformer_input = torch.cat([flows_emb, fcorrs_, track_feats_], dim=2)
+
+            # 2D positional embed
+            # TODO: this can be much simplified
+            pos_embed = get_2d_sincos_pos_embed(self.transformer_dim, grid_size=(HH, WW)).to(query_points.device)
+            sampled_pos_emb = sample_features4d(pos_embed.expand(B, -1, -1, -1), coords[:, 0])
+
+            sampled_pos_emb = rearrange(sampled_pos_emb, "b n c -> (b n) c").unsqueeze(1)
+
+            x = transformer_input + sampled_pos_emb
+
+            # Add the query ref token to the track feats
+            query_ref_token = torch.cat(
+                [self.query_ref_token[:, 0:1], self.query_ref_token[:, 1:2].expand(-1, S - 1, -1)], dim=1
+            )
+            x = x + query_ref_token.to(x.device).to(x.dtype)
+
+            # B, N, S, C
+            x = rearrange(x, "(b n) s d -> b n s d", b=B)
+
+            # Compute the delta coordinates and delta track features
+            delta, _ = self.updateformer(x)
+
+            # BN, S, C
+            delta = rearrange(delta, " b n s d -> (b n) s d", b=B)
+            delta_coords_ = delta[:, :, :2]
+            delta_feats_ = delta[:, :, 2:]
+
+            track_feats_ = track_feats_.reshape(B * N * S, self.latent_dim)
+            delta_feats_ = delta_feats_.reshape(B * N * S, self.latent_dim)
+
+            # Update the track features
+            track_feats_ = self.ffeat_updater(self.ffeat_norm(delta_feats_)) + track_feats_
+
+            track_feats = track_feats_.reshape(B, N, S, self.latent_dim).permute(0, 2, 1, 3)  # BxSxNxC
+
+            # B x S x N x 2
+            coords = coords + delta_coords_.reshape(B, N, S, 2).permute(0, 2, 1, 3)
+
+            # Force coord0 as query
+            # because we assume the query points should not be changed
+            coords[:, 0] = coords_backup[:, 0]
+
+            # The predicted tracks are in the original image scale
+            if down_ratio > 1:
+                coord_preds.append(coords * self.stride * down_ratio)
+            else:
+                coord_preds.append(coords * self.stride)
+
+        # B, S, N
+        vis_e = self.vis_predictor(track_feats.reshape(B * S * N, self.latent_dim)).reshape(B, S, N)
+        if apply_sigmoid:
+            vis_e = torch.sigmoid(vis_e)
+
+        if self.predict_conf:
+            conf_e = self.conf_predictor(track_feats.reshape(B * S * N, self.latent_dim)).reshape(B, S, N)
+            if apply_sigmoid:
+                conf_e = torch.sigmoid(conf_e)
+        else:
+            conf_e = None
+
+        if return_feat:
+            return coord_preds, vis_e, track_feats, query_track_feat, conf_e
+        else:
+            return coord_preds, vis_e, conf_e

vggt/heads/track_modules/blocks.py

@@ -0,0 +1,246 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+# Modified from https://github.com/facebookresearch/co-tracker/
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .utils import bilinear_sampler
+from .modules import Mlp, AttnBlock, CrossAttnBlock, ResidualBlock
+
+
+class EfficientUpdateFormer(nn.Module):
+    """
+    Transformer model that updates track estimates.
+    """
+
+    def __init__(
+        self,
+        space_depth=6,
+        time_depth=6,
+        input_dim=320,
+        hidden_size=384,
+        num_heads=8,
+        output_dim=130,
+        mlp_ratio=4.0,
+        add_space_attn=True,
+        num_virtual_tracks=64,
+    ):
+        super().__init__()
+
+        self.out_channels = 2
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.add_space_attn = add_space_attn
+
+        # Add input LayerNorm before linear projection
+        self.input_norm = nn.LayerNorm(input_dim)
+        self.input_transform = torch.nn.Linear(input_dim, hidden_size, bias=True)
+
+        # Add output LayerNorm before final projection
+        self.output_norm = nn.LayerNorm(hidden_size)
+        self.flow_head = torch.nn.Linear(hidden_size, output_dim, bias=True)
+        self.num_virtual_tracks = num_virtual_tracks
+
+        if self.add_space_attn:
+            self.virual_tracks = nn.Parameter(torch.randn(1, num_virtual_tracks, 1, hidden_size))
+        else:
+            self.virual_tracks = None
+
+        self.time_blocks = nn.ModuleList(
+            [
+                AttnBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    attn_class=nn.MultiheadAttention,
+                )
+                for _ in range(time_depth)
+            ]
+        )
+
+        if add_space_attn:
+            self.space_virtual_blocks = nn.ModuleList(
+                [
+                    AttnBlock(
+                        hidden_size,
+                        num_heads,
+                        mlp_ratio=mlp_ratio,
+                        attn_class=nn.MultiheadAttention,
+                    )
+                    for _ in range(space_depth)
+                ]
+            )
+            self.space_point2virtual_blocks = nn.ModuleList(
+                [CrossAttnBlock(hidden_size, hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(space_depth)]
+            )
+            self.space_virtual2point_blocks = nn.ModuleList(
+                [CrossAttnBlock(hidden_size, hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(space_depth)]
+            )
+            assert len(self.time_blocks) >= len(self.space_virtual2point_blocks)
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+            torch.nn.init.trunc_normal_(self.flow_head.weight, std=0.001)
+
+        self.apply(_basic_init)
+
+    def forward(self, input_tensor, mask=None):
+        # Apply input LayerNorm
+        input_tensor = self.input_norm(input_tensor)
+        tokens = self.input_transform(input_tensor)
+
+        init_tokens = tokens
+
+        B, _, T, _ = tokens.shape
+
+        if self.add_space_attn:
+            virtual_tokens = self.virual_tracks.repeat(B, 1, T, 1)
+            tokens = torch.cat([tokens, virtual_tokens], dim=1)
+
+        _, N, _, _ = tokens.shape
+
+        j = 0
+        for i in range(len(self.time_blocks)):
+            time_tokens = tokens.contiguous().view(B * N, T, -1)  # B N T C -> (B N) T C
+
+            time_tokens = self.time_blocks[i](time_tokens)
+
+            tokens = time_tokens.view(B, N, T, -1)  # (B N) T C -> B N T C
+            if self.add_space_attn and (i % (len(self.time_blocks) // len(self.space_virtual_blocks)) == 0):
+                space_tokens = tokens.permute(0, 2, 1, 3).contiguous().view(B * T, N, -1)  # B N T C -> (B T) N C
+                point_tokens = space_tokens[:, : N - self.num_virtual_tracks]
+                virtual_tokens = space_tokens[:, N - self.num_virtual_tracks :]
+
+                virtual_tokens = self.space_virtual2point_blocks[j](virtual_tokens, point_tokens, mask=mask)
+                virtual_tokens = self.space_virtual_blocks[j](virtual_tokens)
+                point_tokens = self.space_point2virtual_blocks[j](point_tokens, virtual_tokens, mask=mask)
+
+                space_tokens = torch.cat([point_tokens, virtual_tokens], dim=1)
+                tokens = space_tokens.view(B, T, N, -1).permute(0, 2, 1, 3)  # (B T) N C -> B N T C
+                j += 1
+
+        if self.add_space_attn:
+            tokens = tokens[:, : N - self.num_virtual_tracks]
+
+        tokens = tokens + init_tokens
+
+        # Apply output LayerNorm before final projection
+        tokens = self.output_norm(tokens)
+        flow = self.flow_head(tokens)
+
+        return flow, None
+
+
+class CorrBlock:
+    def __init__(self, fmaps, num_levels=4, radius=4, multiple_track_feats=False, padding_mode="zeros"):
+        """
+        Build a pyramid of feature maps from the input.
+
+        fmaps: Tensor (B, S, C, H, W)
+        num_levels: number of pyramid levels (each downsampled by factor 2)
+        radius: search radius for sampling correlation
+        multiple_track_feats: if True, split the target features per pyramid level
+        padding_mode: passed to grid_sample / bilinear_sampler
+        """
+        B, S, C, H, W = fmaps.shape
+        self.S, self.C, self.H, self.W = S, C, H, W
+        self.num_levels = num_levels
+        self.radius = radius
+        self.padding_mode = padding_mode
+        self.multiple_track_feats = multiple_track_feats
+
+        # Build pyramid: each level is half the spatial resolution of the previous
+        self.fmaps_pyramid = [fmaps]  # level 0 is full resolution
+        current_fmaps = fmaps
+        for i in range(num_levels - 1):
+            B, S, C, H, W = current_fmaps.shape
+            # Merge batch & sequence dimensions
+            current_fmaps = current_fmaps.reshape(B * S, C, H, W)
+            # Avg pool down by factor 2
+            current_fmaps = F.avg_pool2d(current_fmaps, kernel_size=2, stride=2)
+            _, _, H_new, W_new = current_fmaps.shape
+            current_fmaps = current_fmaps.reshape(B, S, C, H_new, W_new)
+            self.fmaps_pyramid.append(current_fmaps)
+
+        # Precompute a delta grid (of shape (2r+1, 2r+1, 2)) for sampling.
+        # This grid is added to the (scaled) coordinate centroids.
+        r = self.radius
+        dx = torch.linspace(-r, r, 2 * r + 1, device=fmaps.device, dtype=fmaps.dtype)
+        dy = torch.linspace(-r, r, 2 * r + 1, device=fmaps.device, dtype=fmaps.dtype)
+        # delta: for every (dy,dx) displacement (i.e. Δx, Δy)
+        self.delta = torch.stack(torch.meshgrid(dy, dx, indexing="ij"), dim=-1)  # shape: (2r+1, 2r+1, 2)
+
+    def corr_sample(self, targets, coords):
+        """
+        Instead of storing the entire correlation pyramid, we compute each level's correlation
+        volume, sample it immediately, then discard it. This saves GPU memory.
+
+        Args:
+          targets: Tensor (B, S, N, C) — features for the current targets.
+          coords: Tensor (B, S, N, 2) — coordinates at full resolution.
+
+        Returns:
+          Tensor (B, S, N, L) where L = num_levels * (2*radius+1)**2 (concatenated sampled correlations)
+        """
+        B, S, N, C = targets.shape
+
+        # If you have multiple track features, split them per level.
+        if self.multiple_track_feats:
+            targets_split = torch.split(targets, C // self.num_levels, dim=-1)
+
+        out_pyramid = []
+        for i, fmaps in enumerate(self.fmaps_pyramid):
+            # Get current spatial resolution H, W for this pyramid level.
+            B, S, C, H, W = fmaps.shape
+            # Reshape feature maps for correlation computation:
+            # fmap2s: (B, S, C, H*W)
+            fmap2s = fmaps.view(B, S, C, H * W)
+            # Choose appropriate target features.
+            fmap1 = targets_split[i] if self.multiple_track_feats else targets  # shape: (B, S, N, C)
+
+            # Compute correlation directly
+            corrs = compute_corr_level(fmap1, fmap2s, C)
+            corrs = corrs.view(B, S, N, H, W)
+
+            # Prepare sampling grid:
+            # Scale down the coordinates for the current level.
+            centroid_lvl = coords.reshape(B * S * N, 1, 1, 2) / (2**i)
+            # Make sure our precomputed delta grid is on the same device/dtype.
+            delta_lvl = self.delta.to(coords.device).to(coords.dtype)
+            # Now the grid for grid_sample is:
+            # coords_lvl = centroid_lvl + delta_lvl   (broadcasted over grid)
+            coords_lvl = centroid_lvl + delta_lvl.view(1, 2 * self.radius + 1, 2 * self.radius + 1, 2)
+
+            # Sample from the correlation volume using bilinear interpolation.
+            # We reshape corrs to (B * S * N, 1, H, W) so grid_sample acts over each target.
+            corrs_sampled = bilinear_sampler(
+                corrs.reshape(B * S * N, 1, H, W), coords_lvl, padding_mode=self.padding_mode
+            )
+            # The sampled output is (B * S * N, 1, 2r+1, 2r+1). Flatten the last two dims.
+            corrs_sampled = corrs_sampled.view(B, S, N, -1)  # Now shape: (B, S, N, (2r+1)^2)
+            out_pyramid.append(corrs_sampled)
+
+        # Concatenate all levels along the last dimension.
+        out = torch.cat(out_pyramid, dim=-1).contiguous()
+        return out
+
+
+def compute_corr_level(fmap1, fmap2s, C):
+    # fmap1: (B, S, N, C)
+    # fmap2s: (B, S, C, H*W)
+    corrs = torch.matmul(fmap1, fmap2s)  # (B, S, N, H*W)
+    corrs = corrs.view(fmap1.shape[0], fmap1.shape[1], fmap1.shape[2], -1)  # (B, S, N, H*W)
+    return corrs / math.sqrt(C)

vggt/heads/track_modules/modules.py

@@ -0,0 +1,218 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+from typing import Callable
+import collections
+from torch import Tensor
+from itertools import repeat
+
+
+# From PyTorch internals
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
+            return tuple(x)
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    return val if exists(val) else d
+
+
+to_2tuple = _ntuple(2)
+
+
+class ResidualBlock(nn.Module):
+    """
+    ResidualBlock: construct a block of two conv layers with residual connections
+    """
+
+    def __init__(self, in_planes, planes, norm_fn="group", stride=1, kernel_size=3):
+        super(ResidualBlock, self).__init__()
+
+        self.conv1 = nn.Conv2d(
+            in_planes,
+            planes,
+            kernel_size=kernel_size,
+            padding=1,
+            stride=stride,
+            padding_mode="zeros",
+        )
+        self.conv2 = nn.Conv2d(
+            planes,
+            planes,
+            kernel_size=kernel_size,
+            padding=1,
+            padding_mode="zeros",
+        )
+        self.relu = nn.ReLU(inplace=True)
+
+        num_groups = planes // 8
+
+        if norm_fn == "group":
+            self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+            self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+            if not stride == 1:
+                self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+
+        elif norm_fn == "batch":
+            self.norm1 = nn.BatchNorm2d(planes)
+            self.norm2 = nn.BatchNorm2d(planes)
+            if not stride == 1:
+                self.norm3 = nn.BatchNorm2d(planes)
+
+        elif norm_fn == "instance":
+            self.norm1 = nn.InstanceNorm2d(planes)
+            self.norm2 = nn.InstanceNorm2d(planes)
+            if not stride == 1:
+                self.norm3 = nn.InstanceNorm2d(planes)
+
+        elif norm_fn == "none":
+            self.norm1 = nn.Sequential()
+            self.norm2 = nn.Sequential()
+            if not stride == 1:
+                self.norm3 = nn.Sequential()
+        else:
+            raise NotImplementedError
+
+        if stride == 1:
+            self.downsample = None
+        else:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride),
+                self.norm3,
+            )
+
+    def forward(self, x):
+        y = x
+        y = self.relu(self.norm1(self.conv1(y)))
+        y = self.relu(self.norm2(self.conv2(y)))
+
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        return self.relu(x + y)
+
+
+class Mlp(nn.Module):
+    """MLP as used in Vision Transformer, MLP-Mixer and related networks"""
+
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        norm_layer=None,
+        bias=True,
+        drop=0.0,
+        use_conv=False,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        bias = to_2tuple(bias)
+        drop_probs = to_2tuple(drop)
+        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
+
+        self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.fc2 = linear_layer(hidden_features, out_features, bias=bias[1])
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+class AttnBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        num_heads,
+        attn_class: Callable[..., nn.Module] = nn.MultiheadAttention,
+        mlp_ratio=4.0,
+        **block_kwargs
+    ):
+        """
+        Self attention block
+        """
+        super().__init__()
+
+        self.norm1 = nn.LayerNorm(hidden_size)
+        self.norm2 = nn.LayerNorm(hidden_size)
+
+        self.attn = attn_class(embed_dim=hidden_size, num_heads=num_heads, batch_first=True, **block_kwargs)
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, drop=0)
+
+    def forward(self, x, mask=None):
+        # Prepare the mask for PyTorch's attention (it expects a different format)
+        # attn_mask = mask if mask is not None else None
+        # Normalize before attention
+        x = self.norm1(x)
+
+        # PyTorch's MultiheadAttention returns attn_output, attn_output_weights
+        # attn_output, _ = self.attn(x, x, x, attn_mask=attn_mask)
+
+        attn_output, _ = self.attn(x, x, x)
+
+        # Add & Norm
+        x = x + attn_output
+        x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class CrossAttnBlock(nn.Module):
+    def __init__(self, hidden_size, context_dim, num_heads=1, mlp_ratio=4.0, **block_kwargs):
+        """
+        Cross attention block
+        """
+        super().__init__()
+
+        self.norm1 = nn.LayerNorm(hidden_size)
+        self.norm_context = nn.LayerNorm(hidden_size)
+        self.norm2 = nn.LayerNorm(hidden_size)
+
+        self.cross_attn = nn.MultiheadAttention(
+            embed_dim=hidden_size, num_heads=num_heads, batch_first=True, **block_kwargs
+        )
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, drop=0)
+
+    def forward(self, x, context, mask=None):
+        # Normalize inputs
+        x = self.norm1(x)
+        context = self.norm_context(context)
+
+        # Apply cross attention
+        # Note: nn.MultiheadAttention returns attn_output, attn_output_weights
+        attn_output, _ = self.cross_attn(x, context, context, attn_mask=mask)
+
+        # Add & Norm
+        x = x + attn_output
+        x = x + self.mlp(self.norm2(x))
+        return x

vggt/heads/track_modules/utils.py

@@ -0,0 +1,226 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Modified from https://github.com/facebookresearch/vggsfm
+# and https://github.com/facebookresearch/co-tracker/tree/main
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from typing import Optional, Tuple, Union
+
+
+def get_2d_sincos_pos_embed(embed_dim: int, grid_size: Union[int, Tuple[int, int]], return_grid=False) -> torch.Tensor:
+    """
+    This function initializes a grid and generates a 2D positional embedding using sine and cosine functions.
+    It is a wrapper of get_2d_sincos_pos_embed_from_grid.
+    Args:
+    - embed_dim: The embedding dimension.
+    - grid_size: The grid size.
+    Returns:
+    - pos_embed: The generated 2D positional embedding.
+    """
+    if isinstance(grid_size, tuple):
+        grid_size_h, grid_size_w = grid_size
+    else:
+        grid_size_h = grid_size_w = grid_size
+    grid_h = torch.arange(grid_size_h, dtype=torch.float)
+    grid_w = torch.arange(grid_size_w, dtype=torch.float)
+    grid = torch.meshgrid(grid_w, grid_h, indexing="xy")
+    grid = torch.stack(grid, dim=0)
+    grid = grid.reshape([2, 1, grid_size_h, grid_size_w])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if return_grid:
+        return (
+            pos_embed.reshape(1, grid_size_h, grid_size_w, -1).permute(0, 3, 1, 2),
+            grid,
+        )
+    return pos_embed.reshape(1, grid_size_h, grid_size_w, -1).permute(0, 3, 1, 2)
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim: int, grid: torch.Tensor) -> torch.Tensor:
+    """
+    This function generates a 2D positional embedding from a given grid using sine and cosine functions.
+
+    Args:
+    - embed_dim: The embedding dimension.
+    - grid: The grid to generate the embedding from.
+
+    Returns:
+    - emb: The generated 2D positional embedding.
+    """
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = torch.cat([emb_h, emb_w], dim=2)  # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim: int, pos: torch.Tensor) -> torch.Tensor:
+    """
+    This function generates a 1D positional embedding from a given grid using sine and cosine functions.
+
+    Args:
+    - embed_dim: The embedding dimension.
+    - pos: The position to generate the embedding from.
+
+    Returns:
+    - emb: The generated 1D positional embedding.
+    """
+    assert embed_dim % 2 == 0
+    omega = torch.arange(embed_dim // 2, dtype=torch.double)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    emb = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb[None].float()
+
+
+def get_2d_embedding(xy: torch.Tensor, C: int, cat_coords: bool = True) -> torch.Tensor:
+    """
+    This function generates a 2D positional embedding from given coordinates using sine and cosine functions.
+
+    Args:
+    - xy: The coordinates to generate the embedding from.
+    - C: The size of the embedding.
+    - cat_coords: A flag to indicate whether to concatenate the original coordinates to the embedding.
+
+    Returns:
+    - pe: The generated 2D positional embedding.
+    """
+    B, N, D = xy.shape
+    assert D == 2
+
+    x = xy[:, :, 0:1]
+    y = xy[:, :, 1:2]
+    div_term = (torch.arange(0, C, 2, device=xy.device, dtype=torch.float32) * (1000.0 / C)).reshape(1, 1, int(C / 2))
+
+    pe_x = torch.zeros(B, N, C, device=xy.device, dtype=torch.float32)
+    pe_y = torch.zeros(B, N, C, device=xy.device, dtype=torch.float32)
+
+    pe_x[:, :, 0::2] = torch.sin(x * div_term)
+    pe_x[:, :, 1::2] = torch.cos(x * div_term)
+
+    pe_y[:, :, 0::2] = torch.sin(y * div_term)
+    pe_y[:, :, 1::2] = torch.cos(y * div_term)
+
+    pe = torch.cat([pe_x, pe_y], dim=2)  # (B, N, C*3)
+    if cat_coords:
+        pe = torch.cat([xy, pe], dim=2)  # (B, N, C*3+3)
+    return pe
+
+
+def bilinear_sampler(input, coords, align_corners=True, padding_mode="border"):
+    r"""Sample a tensor using bilinear interpolation
+
+    `bilinear_sampler(input, coords)` samples a tensor :attr:`input` at
+    coordinates :attr:`coords` using bilinear interpolation. It is the same
+    as `torch.nn.functional.grid_sample()` but with a different coordinate
+    convention.
+
+    The input tensor is assumed to be of shape :math:`(B, C, H, W)`, where
+    :math:`B` is the batch size, :math:`C` is the number of channels,
+    :math:`H` is the height of the image, and :math:`W` is the width of the
+    image. The tensor :attr:`coords` of shape :math:`(B, H_o, W_o, 2)` is
+    interpreted as an array of 2D point coordinates :math:`(x_i,y_i)`.
+
+    Alternatively, the input tensor can be of size :math:`(B, C, T, H, W)`,
+    in which case sample points are triplets :math:`(t_i,x_i,y_i)`. Note
+    that in this case the order of the components is slightly different
+    from `grid_sample()`, which would expect :math:`(x_i,y_i,t_i)`.
+
+    If `align_corners` is `True`, the coordinate :math:`x` is assumed to be
+    in the range :math:`[0,W-1]`, with 0 corresponding to the center of the
+    left-most image pixel :math:`W-1` to the center of the right-most
+    pixel.
+
+    If `align_corners` is `False`, the coordinate :math:`x` is assumed to
+    be in the range :math:`[0,W]`, with 0 corresponding to the left edge of
+    the left-most pixel :math:`W` to the right edge of the right-most
+    pixel.
+
+    Similar conventions apply to the :math:`y` for the range
+    :math:`[0,H-1]` and :math:`[0,H]` and to :math:`t` for the range
+    :math:`[0,T-1]` and :math:`[0,T]`.
+
+    Args:
+        input (Tensor): batch of input images.
+        coords (Tensor): batch of coordinates.
+        align_corners (bool, optional): Coordinate convention. Defaults to `True`.
+        padding_mode (str, optional): Padding mode. Defaults to `"border"`.
+
+    Returns:
+        Tensor: sampled points.
+    """
+    coords = coords.detach().clone()
+    ############################################################
+    # IMPORTANT:
+    coords = coords.to(input.device).to(input.dtype)
+    ############################################################
+
+    sizes = input.shape[2:]
+
+    assert len(sizes) in [2, 3]
+
+    if len(sizes) == 3:
+        # t x y -> x y t to match dimensions T H W in grid_sample
+        coords = coords[..., [1, 2, 0]]
+
+    if align_corners:
+        scale = torch.tensor(
+            [2 / max(size - 1, 1) for size in reversed(sizes)], device=coords.device, dtype=coords.dtype
+        )
+    else:
+        scale = torch.tensor([2 / size for size in reversed(sizes)], device=coords.device, dtype=coords.dtype)
+
+    coords.mul_(scale)  # coords = coords * scale
+    coords.sub_(1)  # coords = coords - 1
+
+    return F.grid_sample(input, coords, align_corners=align_corners, padding_mode=padding_mode)
+
+
+def sample_features4d(input, coords):
+    r"""Sample spatial features
+
+    `sample_features4d(input, coords)` samples the spatial features
+    :attr:`input` represented by a 4D tensor :math:`(B, C, H, W)`.
+
+    The field is sampled at coordinates :attr:`coords` using bilinear
+    interpolation. :attr:`coords` is assumed to be of shape :math:`(B, R,
+    2)`, where each sample has the format :math:`(x_i, y_i)`. This uses the
+    same convention as :func:`bilinear_sampler` with `align_corners=True`.
+
+    The output tensor has one feature per point, and has shape :math:`(B,
+    R, C)`.
+
+    Args:
+        input (Tensor): spatial features.
+        coords (Tensor): points.
+
+    Returns:
+        Tensor: sampled features.
+    """
+
+    B, _, _, _ = input.shape
+
+    # B R 2 -> B R 1 2
+    coords = coords.unsqueeze(2)
+
+    # B C R 1
+    feats = bilinear_sampler(input, coords)
+
+    return feats.permute(0, 2, 1, 3).view(B, -1, feats.shape[1] * feats.shape[3])  # B C R 1 -> B R C

vggt/heads/utils.py

@@ -0,0 +1,108 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+
+
+def position_grid_to_embed(pos_grid: torch.Tensor, embed_dim: int, omega_0: float = 100) -> torch.Tensor:
+    """
+    Convert 2D position grid (HxWx2) to sinusoidal embeddings (HxWxC)
+
+    Args:
+        pos_grid: Tensor of shape (H, W, 2) containing 2D coordinates
+        embed_dim: Output channel dimension for embeddings
+
+    Returns:
+        Tensor of shape (H, W, embed_dim) with positional embeddings
+    """
+    H, W, grid_dim = pos_grid.shape
+    assert grid_dim == 2
+    pos_flat = pos_grid.reshape(-1, grid_dim)  # Flatten to (H*W, 2)
+
+    # Process x and y coordinates separately
+    emb_x = make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 0], omega_0=omega_0)  # [1, H*W, D/2]
+    emb_y = make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 1], omega_0=omega_0)  # [1, H*W, D/2]
+
+    # Combine and reshape
+    emb = torch.cat([emb_x, emb_y], dim=-1)  # [1, H*W, D]
+
+    return emb.view(H, W, embed_dim)  # [H, W, D]
+
+
+def make_sincos_pos_embed(embed_dim: int, pos: torch.Tensor, omega_0: float = 100) -> torch.Tensor:
+    """
+    This function generates a 1D positional embedding from a given grid using sine and cosine functions.
+
+    Args:
+    - embed_dim: The embedding dimension.
+    - pos: The position to generate the embedding from.
+
+    Returns:
+    - emb: The generated 1D positional embedding.
+    """
+    assert embed_dim % 2 == 0
+    omega = torch.arange(embed_dim // 2, dtype=torch.double, device=pos.device)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / omega_0**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    emb = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb.float()
+
+
+# Inspired by https://github.com/microsoft/moge
+
+
+def create_uv_grid(
+    width: int, height: int, aspect_ratio: float = None, dtype: torch.dtype = None, device: torch.device = None
+) -> torch.Tensor:
+    """
+    Create a normalized UV grid of shape (width, height, 2).
+
+    The grid spans horizontally and vertically according to an aspect ratio,
+    ensuring the top-left corner is at (-x_span, -y_span) and the bottom-right
+    corner is at (x_span, y_span), normalized by the diagonal of the plane.
+
+    Args:
+        width (int): Number of points horizontally.
+        height (int): Number of points vertically.
+        aspect_ratio (float, optional): Width-to-height ratio. Defaults to width/height.
+        dtype (torch.dtype, optional): Data type of the resulting tensor.
+        device (torch.device, optional): Device on which the tensor is created.
+
+    Returns:
+        torch.Tensor: A (width, height, 2) tensor of UV coordinates.
+    """
+    # Derive aspect ratio if not explicitly provided
+    if aspect_ratio is None:
+        aspect_ratio = float(width) / float(height)
+
+    # Compute normalized spans for X and Y
+    diag_factor = (aspect_ratio**2 + 1.0) ** 0.5
+    span_x = aspect_ratio / diag_factor
+    span_y = 1.0 / diag_factor
+
+    # Establish the linspace boundaries
+    left_x = -span_x * (width - 1) / width
+    right_x = span_x * (width - 1) / width
+    top_y = -span_y * (height - 1) / height
+    bottom_y = span_y * (height - 1) / height
+
+    # Generate 1D coordinates
+    x_coords = torch.linspace(left_x, right_x, steps=width, dtype=dtype, device=device)
+    y_coords = torch.linspace(top_y, bottom_y, steps=height, dtype=dtype, device=device)
+
+    # Create 2D meshgrid (width x height) and stack into UV
+    uu, vv = torch.meshgrid(x_coords, y_coords, indexing="xy")
+    uv_grid = torch.stack((uu, vv), dim=-1)
+
+    return uv_grid

vggt/layers/__init__.py

@@ -0,0 +1,11 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .mlp import Mlp
+from .patch_embed import PatchEmbed
+from .swiglu_ffn import SwiGLUFFN, SwiGLUFFNFused
+from .block import NestedTensorBlock
+from .attention import MemEffAttention

vggt/layers/attention.py

@@ -0,0 +1,98 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+# References:
+#   https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
+#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/models/vision_transformer.py
+
+import logging
+import os
+import warnings
+
+from torch import Tensor
+from torch import nn
+import torch.nn.functional as F
+
+XFORMERS_AVAILABLE = False
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        proj_bias: bool = True,
+        attn_drop: float = 0.0,
+        proj_drop: float = 0.0,
+        norm_layer: nn.Module = nn.LayerNorm,
+        qk_norm: bool = False,
+        fused_attn: bool = True,  # use F.scaled_dot_product_attention or not
+        rope=None,
+    ) -> None:
+        super().__init__()
+        assert dim % num_heads == 0, "dim should be divisible by num_heads"
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim**-0.5
+        self.fused_attn = fused_attn
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim, bias=proj_bias)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.rope = rope
+
+    def forward(self, x: Tensor, pos=None) -> Tensor:
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv.unbind(0)
+        q, k = self.q_norm(q), self.k_norm(k)
+
+        if self.rope is not None:
+            q = self.rope(q, pos)
+            k = self.rope(k, pos)
+
+        if self.fused_attn:
+            x = F.scaled_dot_product_attention(
+                q,
+                k,
+                v,
+                dropout_p=self.attn_drop.p if self.training else 0.0,
+            )
+        else:
+            q = q * self.scale
+            attn = q @ k.transpose(-2, -1)
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+            x = attn @ v
+
+        x = x.transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class MemEffAttention(Attention):
+    def forward(self, x: Tensor, attn_bias=None, pos=None) -> Tensor:
+        assert pos is None
+        if not XFORMERS_AVAILABLE:
+            if attn_bias is not None:
+                raise AssertionError("xFormers is required for using nested tensors")
+            return super().forward(x)
+
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
+
+        q, k, v = unbind(qkv, 2)
+
+        x = memory_efficient_attention(q, k, v, attn_bias=attn_bias)
+        x = x.reshape([B, N, C])
+
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x

vggt/layers/block.py

@@ -0,0 +1,259 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+# References:
+#   https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
+#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/layers/patch_embed.py
+
+import logging
+import os
+from typing import Callable, List, Any, Tuple, Dict
+import warnings
+
+import torch
+from torch import nn, Tensor
+
+from .attention import Attention
+from .drop_path import DropPath
+from .layer_scale import LayerScale
+from .mlp import Mlp
+
+
+XFORMERS_AVAILABLE = False
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        proj_bias: bool = True,
+        ffn_bias: bool = True,
+        drop: float = 0.0,
+        attn_drop: float = 0.0,
+        init_values=None,
+        drop_path: float = 0.0,
+        act_layer: Callable[..., nn.Module] = nn.GELU,
+        norm_layer: Callable[..., nn.Module] = nn.LayerNorm,
+        attn_class: Callable[..., nn.Module] = Attention,
+        ffn_layer: Callable[..., nn.Module] = Mlp,
+        qk_norm: bool = False,
+        fused_attn: bool = True,  # use F.scaled_dot_product_attention or not
+        rope=None,
+    ) -> None:
+        super().__init__()
+
+        self.norm1 = norm_layer(dim)
+
+        self.attn = attn_class(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            proj_bias=proj_bias,
+            attn_drop=attn_drop,
+            proj_drop=drop,
+            qk_norm=qk_norm,
+            fused_attn=fused_attn,
+            rope=rope,
+        )
+
+        self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        self.drop_path1 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = ffn_layer(
+            in_features=dim,
+            hidden_features=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=drop,
+            bias=ffn_bias,
+        )
+        self.ls2 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        self.drop_path2 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.sample_drop_ratio = drop_path
+
+    def forward(self, x: Tensor, pos=None) -> Tensor:
+        def attn_residual_func(x: Tensor, pos=None) -> Tensor:
+            return self.ls1(self.attn(self.norm1(x), pos=pos))
+
+        def ffn_residual_func(x: Tensor) -> Tensor:
+            return self.ls2(self.mlp(self.norm2(x)))
+
+        if self.training and self.sample_drop_ratio > 0.1:
+            # the overhead is compensated only for a drop path rate larger than 0.1
+            x = drop_add_residual_stochastic_depth(
+                x,
+                pos=pos,
+                residual_func=attn_residual_func,
+                sample_drop_ratio=self.sample_drop_ratio,
+            )
+            x = drop_add_residual_stochastic_depth(
+                x,
+                residual_func=ffn_residual_func,
+                sample_drop_ratio=self.sample_drop_ratio,
+            )
+        elif self.training and self.sample_drop_ratio > 0.0:
+            x = x + self.drop_path1(attn_residual_func(x, pos=pos))
+            x = x + self.drop_path1(ffn_residual_func(x))  # FIXME: drop_path2
+        else:
+            x = x + attn_residual_func(x, pos=pos)
+            x = x + ffn_residual_func(x)
+        return x
+
+
+def drop_add_residual_stochastic_depth(
+    x: Tensor,
+    residual_func: Callable[[Tensor], Tensor],
+    sample_drop_ratio: float = 0.0,
+    pos=None,
+) -> Tensor:
+    # 1) extract subset using permutation
+    b, n, d = x.shape
+    sample_subset_size = max(int(b * (1 - sample_drop_ratio)), 1)
+    brange = (torch.randperm(b, device=x.device))[:sample_subset_size]
+    x_subset = x[brange]
+
+    # 2) apply residual_func to get residual
+    if pos is not None:
+        # if necessary, apply rope to the subset
+        pos = pos[brange]
+        residual = residual_func(x_subset, pos=pos)
+    else:
+        residual = residual_func(x_subset)
+
+    x_flat = x.flatten(1)
+    residual = residual.flatten(1)
+
+    residual_scale_factor = b / sample_subset_size
+
+    # 3) add the residual
+    x_plus_residual = torch.index_add(x_flat, 0, brange, residual.to(dtype=x.dtype), alpha=residual_scale_factor)
+    return x_plus_residual.view_as(x)
+
+
+def get_branges_scales(x, sample_drop_ratio=0.0):
+    b, n, d = x.shape
+    sample_subset_size = max(int(b * (1 - sample_drop_ratio)), 1)
+    brange = (torch.randperm(b, device=x.device))[:sample_subset_size]
+    residual_scale_factor = b / sample_subset_size
+    return brange, residual_scale_factor
+
+
+def add_residual(x, brange, residual, residual_scale_factor, scaling_vector=None):
+    if scaling_vector is None:
+        x_flat = x.flatten(1)
+        residual = residual.flatten(1)
+        x_plus_residual = torch.index_add(x_flat, 0, brange, residual.to(dtype=x.dtype), alpha=residual_scale_factor)
+    else:
+        x_plus_residual = scaled_index_add(
+            x, brange, residual.to(dtype=x.dtype), scaling=scaling_vector, alpha=residual_scale_factor
+        )
+    return x_plus_residual
+
+
+attn_bias_cache: Dict[Tuple, Any] = {}
+
+
+def get_attn_bias_and_cat(x_list, branges=None):
+    """
+    this will perform the index select, cat the tensors, and provide the attn_bias from cache
+    """
+    batch_sizes = [b.shape[0] for b in branges] if branges is not None else [x.shape[0] for x in x_list]
+    all_shapes = tuple((b, x.shape[1]) for b, x in zip(batch_sizes, x_list))
+    if all_shapes not in attn_bias_cache.keys():
+        seqlens = []
+        for b, x in zip(batch_sizes, x_list):
+            for _ in range(b):
+                seqlens.append(x.shape[1])
+        attn_bias = fmha.BlockDiagonalMask.from_seqlens(seqlens)
+        attn_bias._batch_sizes = batch_sizes
+        attn_bias_cache[all_shapes] = attn_bias
+
+    if branges is not None:
+        cat_tensors = index_select_cat([x.flatten(1) for x in x_list], branges).view(1, -1, x_list[0].shape[-1])
+    else:
+        tensors_bs1 = tuple(x.reshape([1, -1, *x.shape[2:]]) for x in x_list)
+        cat_tensors = torch.cat(tensors_bs1, dim=1)
+
+    return attn_bias_cache[all_shapes], cat_tensors
+
+
+def drop_add_residual_stochastic_depth_list(
+    x_list: List[Tensor],
+    residual_func: Callable[[Tensor, Any], Tensor],
+    sample_drop_ratio: float = 0.0,
+    scaling_vector=None,
+) -> Tensor:
+    # 1) generate random set of indices for dropping samples in the batch
+    branges_scales = [get_branges_scales(x, sample_drop_ratio=sample_drop_ratio) for x in x_list]
+    branges = [s[0] for s in branges_scales]
+    residual_scale_factors = [s[1] for s in branges_scales]
+
+    # 2) get attention bias and index+concat the tensors
+    attn_bias, x_cat = get_attn_bias_and_cat(x_list, branges)
+
+    # 3) apply residual_func to get residual, and split the result
+    residual_list = attn_bias.split(residual_func(x_cat, attn_bias=attn_bias))  # type: ignore
+
+    outputs = []
+    for x, brange, residual, residual_scale_factor in zip(x_list, branges, residual_list, residual_scale_factors):
+        outputs.append(add_residual(x, brange, residual, residual_scale_factor, scaling_vector).view_as(x))
+    return outputs
+
+
+class NestedTensorBlock(Block):
+    def forward_nested(self, x_list: List[Tensor]) -> List[Tensor]:
+        """
+        x_list contains a list of tensors to nest together and run
+        """
+        assert isinstance(self.attn, MemEffAttention)
+
+        if self.training and self.sample_drop_ratio > 0.0:
+
+            def attn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
+                return self.attn(self.norm1(x), attn_bias=attn_bias)
+
+            def ffn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
+                return self.mlp(self.norm2(x))
+
+            x_list = drop_add_residual_stochastic_depth_list(
+                x_list,
+                residual_func=attn_residual_func,
+                sample_drop_ratio=self.sample_drop_ratio,
+                scaling_vector=self.ls1.gamma if isinstance(self.ls1, LayerScale) else None,
+            )
+            x_list = drop_add_residual_stochastic_depth_list(
+                x_list,
+                residual_func=ffn_residual_func,
+                sample_drop_ratio=self.sample_drop_ratio,
+                scaling_vector=self.ls2.gamma if isinstance(self.ls1, LayerScale) else None,
+            )
+            return x_list
+        else:
+
+            def attn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
+                return self.ls1(self.attn(self.norm1(x), attn_bias=attn_bias))
+
+            def ffn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
+                return self.ls2(self.mlp(self.norm2(x)))
+
+            attn_bias, x = get_attn_bias_and_cat(x_list)
+            x = x + attn_residual_func(x, attn_bias=attn_bias)
+            x = x + ffn_residual_func(x)
+            return attn_bias.split(x)
+
+    def forward(self, x_or_x_list):
+        if isinstance(x_or_x_list, Tensor):
+            return super().forward(x_or_x_list)
+        elif isinstance(x_or_x_list, list):
+            if not XFORMERS_AVAILABLE:
+                raise AssertionError("xFormers is required for using nested tensors")
+            return self.forward_nested(x_or_x_list)
+        else:
+            raise AssertionError

vggt/layers/drop_path.py

@@ -0,0 +1,34 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+# References:
+#   https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
+#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/layers/drop.py
+
+
+from torch import nn
+
+
+def drop_path(x, drop_prob: float = 0.0, training: bool = False):
+    if drop_prob == 0.0 or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+    if keep_prob > 0.0:
+        random_tensor.div_(keep_prob)
+    output = x * random_tensor
+    return output
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
+
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)

vggt/layers/layer_scale.py

@@ -0,0 +1,27 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+# Modified from: https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vision_transformer.py#L103-L110
+
+from typing import Union
+
+import torch
+from torch import Tensor
+from torch import nn
+
+
+class LayerScale(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        init_values: Union[float, Tensor] = 1e-5,
+        inplace: bool = False,
+    ) -> None:
+        super().__init__()
+        self.inplace = inplace
+        self.gamma = nn.Parameter(init_values * torch.ones(dim))
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x.mul_(self.gamma) if self.inplace else x * self.gamma

vggt/layers/mlp.py

@@ -0,0 +1,40 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+# References:
+#   https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
+#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/layers/mlp.py
+
+
+from typing import Callable, Optional
+
+from torch import Tensor, nn
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: Optional[int] = None,
+        out_features: Optional[int] = None,
+        act_layer: Callable[..., nn.Module] = nn.GELU,
+        drop: float = 0.0,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x

vggt/layers/patch_embed.py

@@ -0,0 +1,88 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+# References:
+#   https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
+#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/layers/patch_embed.py
+
+from typing import Callable, Optional, Tuple, Union
+
+from torch import Tensor
+import torch.nn as nn
+
+
+def make_2tuple(x):
+    if isinstance(x, tuple):
+        assert len(x) == 2
+        return x
+
+    assert isinstance(x, int)
+    return (x, x)
+
+
+class PatchEmbed(nn.Module):
+    """
+    2D image to patch embedding: (B,C,H,W) -> (B,N,D)
+
+    Args:
+        img_size: Image size.
+        patch_size: Patch token size.
+        in_chans: Number of input image channels.
+        embed_dim: Number of linear projection output channels.
+        norm_layer: Normalization layer.
+    """
+
+    def __init__(
+        self,
+        img_size: Union[int, Tuple[int, int]] = 224,
+        patch_size: Union[int, Tuple[int, int]] = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        norm_layer: Optional[Callable] = None,
+        flatten_embedding: bool = True,
+    ) -> None:
+        super().__init__()
+
+        image_HW = make_2tuple(img_size)
+        patch_HW = make_2tuple(patch_size)
+        patch_grid_size = (
+            image_HW[0] // patch_HW[0],
+            image_HW[1] // patch_HW[1],
+        )
+
+        self.img_size = image_HW
+        self.patch_size = patch_HW
+        self.patches_resolution = patch_grid_size
+        self.num_patches = patch_grid_size[0] * patch_grid_size[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.flatten_embedding = flatten_embedding
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_HW, stride=patch_HW)
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+
+    def forward(self, x: Tensor) -> Tensor:
+        _, _, H, W = x.shape
+        patch_H, patch_W = self.patch_size
+
+        assert H % patch_H == 0, f"Input image height {H} is not a multiple of patch height {patch_H}"
+        assert W % patch_W == 0, f"Input image width {W} is not a multiple of patch width: {patch_W}"
+
+        x = self.proj(x)  # B C H W
+        H, W = x.size(2), x.size(3)
+        x = x.flatten(2).transpose(1, 2)  # B HW C
+        x = self.norm(x)
+        if not self.flatten_embedding:
+            x = x.reshape(-1, H, W, self.embed_dim)  # B H W C
+        return x
+
+    def flops(self) -> float:
+        Ho, Wo = self.patches_resolution
+        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
+        if self.norm is not None:
+            flops += Ho * Wo * self.embed_dim
+        return flops

vggt/layers/rope.py

@@ -0,0 +1,188 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+
+# Implementation of 2D Rotary Position Embeddings (RoPE).
+
+# This module provides a clean implementation of 2D Rotary Position Embeddings,
+# which extends the original RoPE concept to handle 2D spatial positions.
+
+# Inspired by:
+#         https://github.com/meta-llama/codellama/blob/main/llama/model.py
+#         https://github.com/naver-ai/rope-vit
+
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Dict, Tuple
+
+
+class PositionGetter:
+    """Generates and caches 2D spatial positions for patches in a grid.
+
+    This class efficiently manages the generation of spatial coordinates for patches
+    in a 2D grid, caching results to avoid redundant computations.
+
+    Attributes:
+        position_cache: Dictionary storing precomputed position tensors for different
+            grid dimensions.
+    """
+
+    def __init__(self):
+        """Initializes the position generator with an empty cache."""
+        self.position_cache: Dict[Tuple[int, int], torch.Tensor] = {}
+
+    def __call__(self, batch_size: int, height: int, width: int, device: torch.device) -> torch.Tensor:
+        """Generates spatial positions for a batch of patches.
+
+        Args:
+            batch_size: Number of samples in the batch.
+            height: Height of the grid in patches.
+            width: Width of the grid in patches.
+            device: Target device for the position tensor.
+
+        Returns:
+            Tensor of shape (batch_size, height*width, 2) containing y,x coordinates
+            for each position in the grid, repeated for each batch item.
+        """
+        if (height, width) not in self.position_cache:
+            y_coords = torch.arange(height, device=device)
+            x_coords = torch.arange(width, device=device)
+            positions = torch.cartesian_prod(y_coords, x_coords)
+            self.position_cache[height, width] = positions
+
+        cached_positions = self.position_cache[height, width]
+        return cached_positions.view(1, height * width, 2).expand(batch_size, -1, -1).clone()
+
+
+class RotaryPositionEmbedding2D(nn.Module):
+    """2D Rotary Position Embedding implementation.
+
+    This module applies rotary position embeddings to input tokens based on their
+    2D spatial positions. It handles the position-dependent rotation of features
+    separately for vertical and horizontal dimensions.
+
+    Args:
+        frequency: Base frequency for the position embeddings. Default: 100.0
+        scaling_factor: Scaling factor for frequency computation. Default: 1.0
+
+    Attributes:
+        base_frequency: Base frequency for computing position embeddings.
+        scaling_factor: Factor to scale the computed frequencies.
+        frequency_cache: Cache for storing precomputed frequency components.
+    """
+
+    def __init__(self, frequency: float = 100.0, scaling_factor: float = 1.0):
+        """Initializes the 2D RoPE module."""
+        super().__init__()
+        self.base_frequency = frequency
+        self.scaling_factor = scaling_factor
+        self.frequency_cache: Dict[Tuple, Tuple[torch.Tensor, torch.Tensor]] = {}
+
+    def _compute_frequency_components(
+        self, dim: int, seq_len: int, device: torch.device, dtype: torch.dtype
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Computes frequency components for rotary embeddings.
+
+        Args:
+            dim: Feature dimension (must be even).
+            seq_len: Maximum sequence length.
+            device: Target device for computations.
+            dtype: Data type for the computed tensors.
+
+        Returns:
+            Tuple of (cosine, sine) tensors for frequency components.
+        """
+        cache_key = (dim, seq_len, device, dtype)
+        if cache_key not in self.frequency_cache:
+            # Compute frequency bands
+            exponents = torch.arange(0, dim, 2, device=device).float() / dim
+            inv_freq = 1.0 / (self.base_frequency**exponents)
+
+            # Generate position-dependent frequencies
+            positions = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
+            angles = torch.einsum("i,j->ij", positions, inv_freq)
+
+            # Compute and cache frequency components
+            angles = angles.to(dtype)
+            angles = torch.cat((angles, angles), dim=-1)
+            cos_components = angles.cos().to(dtype)
+            sin_components = angles.sin().to(dtype)
+            self.frequency_cache[cache_key] = (cos_components, sin_components)
+
+        return self.frequency_cache[cache_key]
+
+    @staticmethod
+    def _rotate_features(x: torch.Tensor) -> torch.Tensor:
+        """Performs feature rotation by splitting and recombining feature dimensions.
+
+        Args:
+            x: Input tensor to rotate.
+
+        Returns:
+            Rotated feature tensor.
+        """
+        feature_dim = x.shape[-1]
+        x1, x2 = x[..., : feature_dim // 2], x[..., feature_dim // 2 :]
+        return torch.cat((-x2, x1), dim=-1)
+
+    def _apply_1d_rope(
+        self, tokens: torch.Tensor, positions: torch.Tensor, cos_comp: torch.Tensor, sin_comp: torch.Tensor
+    ) -> torch.Tensor:
+        """Applies 1D rotary position embeddings along one dimension.
+
+        Args:
+            tokens: Input token features.
+            positions: Position indices.
+            cos_comp: Cosine components for rotation.
+            sin_comp: Sine components for rotation.
+
+        Returns:
+            Tokens with applied rotary position embeddings.
+        """
+        # Embed positions with frequency components
+        cos = F.embedding(positions, cos_comp)[:, None, :, :]
+        sin = F.embedding(positions, sin_comp)[:, None, :, :]
+
+        # Apply rotation
+        return (tokens * cos) + (self._rotate_features(tokens) * sin)
+
+    def forward(self, tokens: torch.Tensor, positions: torch.Tensor) -> torch.Tensor:
+        """Applies 2D rotary position embeddings to input tokens.
+
+        Args:
+            tokens: Input tensor of shape (batch_size, n_heads, n_tokens, dim).
+                   The feature dimension (dim) must be divisible by 4.
+            positions: Position tensor of shape (batch_size, n_tokens, 2) containing
+                      the y and x coordinates for each token.
+
+        Returns:
+            Tensor of same shape as input with applied 2D rotary position embeddings.
+
+        Raises:
+            AssertionError: If input dimensions are invalid or positions are malformed.
+        """
+        # Validate inputs
+        assert tokens.size(-1) % 2 == 0, "Feature dimension must be even"
+        assert positions.ndim == 3 and positions.shape[-1] == 2, "Positions must have shape (batch_size, n_tokens, 2)"
+
+        # Compute feature dimension for each spatial direction
+        feature_dim = tokens.size(-1) // 2
+
+        # Get frequency components
+        max_position = int(positions.max()) + 1
+        cos_comp, sin_comp = self._compute_frequency_components(feature_dim, max_position, tokens.device, tokens.dtype)
+
+        # Split features for vertical and horizontal processing
+        vertical_features, horizontal_features = tokens.chunk(2, dim=-1)
+
+        # Apply RoPE separately for each dimension
+        vertical_features = self._apply_1d_rope(vertical_features, positions[..., 0], cos_comp, sin_comp)
+        horizontal_features = self._apply_1d_rope(horizontal_features, positions[..., 1], cos_comp, sin_comp)
+
+        # Combine processed features
+        return torch.cat((vertical_features, horizontal_features), dim=-1)

vggt/layers/swiglu_ffn.py

@@ -0,0 +1,72 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+import os
+from typing import Callable, Optional
+import warnings
+
+from torch import Tensor, nn
+import torch.nn.functional as F
+
+
+class SwiGLUFFN(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: Optional[int] = None,
+        out_features: Optional[int] = None,
+        act_layer: Callable[..., nn.Module] = None,
+        drop: float = 0.0,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.w12 = nn.Linear(in_features, 2 * hidden_features, bias=bias)
+        self.w3 = nn.Linear(hidden_features, out_features, bias=bias)
+
+    def forward(self, x: Tensor) -> Tensor:
+        x12 = self.w12(x)
+        x1, x2 = x12.chunk(2, dim=-1)
+        hidden = F.silu(x1) * x2
+        return self.w3(hidden)
+
+
+XFORMERS_ENABLED = os.environ.get("XFORMERS_DISABLED") is None
+# try:
+#     if XFORMERS_ENABLED:
+#         from xformers.ops import SwiGLU
+
+#         XFORMERS_AVAILABLE = True
+#         warnings.warn("xFormers is available (SwiGLU)")
+#     else:
+#         warnings.warn("xFormers is disabled (SwiGLU)")
+#         raise ImportError
+# except ImportError:
+SwiGLU = SwiGLUFFN
+XFORMERS_AVAILABLE = False
+
+# warnings.warn("xFormers is not available (SwiGLU)")
+
+
+class SwiGLUFFNFused(SwiGLU):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: Optional[int] = None,
+        out_features: Optional[int] = None,
+        act_layer: Callable[..., nn.Module] = None,
+        drop: float = 0.0,
+        bias: bool = True,
+    ) -> None:
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        hidden_features = (int(hidden_features * 2 / 3) + 7) // 8 * 8
+        super().__init__(
+            in_features=in_features,
+            hidden_features=hidden_features,
+            out_features=out_features,
+            bias=bias,
+        )

vggt/layers/vision_transformer.py

@@ -0,0 +1,407 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+# References:
+#   https://github.com/facebookresearch/dino/blob/main/vision_transformer.py
+#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/models/vision_transformer.py
+
+from functools import partial
+import math
+import logging
+from typing import Sequence, Tuple, Union, Callable
+
+import torch
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint
+from torch.nn.init import trunc_normal_
+from . import Mlp, PatchEmbed, SwiGLUFFNFused, MemEffAttention, NestedTensorBlock as Block
+
+logger = logging.getLogger("dinov2")
+
+
+def named_apply(fn: Callable, module: nn.Module, name="", depth_first=True, include_root=False) -> nn.Module:
+    if not depth_first and include_root:
+        fn(module=module, name=name)
+    for child_name, child_module in module.named_children():
+        child_name = ".".join((name, child_name)) if name else child_name
+        named_apply(fn=fn, module=child_module, name=child_name, depth_first=depth_first, include_root=True)
+    if depth_first and include_root:
+        fn(module=module, name=name)
+    return module
+
+
+class BlockChunk(nn.ModuleList):
+    def forward(self, x):
+        for b in self:
+            x = b(x)
+        return x
+
+
+class DinoVisionTransformer(nn.Module):
+    def __init__(
+        self,
+        img_size=224,
+        patch_size=16,
+        in_chans=3,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        ffn_bias=True,
+        proj_bias=True,
+        drop_path_rate=0.0,
+        drop_path_uniform=False,
+        init_values=None,  # for layerscale: None or 0 => no layerscale
+        embed_layer=PatchEmbed,
+        act_layer=nn.GELU,
+        block_fn=Block,
+        ffn_layer="mlp",
+        block_chunks=1,
+        num_register_tokens=0,
+        interpolate_antialias=False,
+        interpolate_offset=0.1,
+        qk_norm=False,
+    ):
+        """
+        Args:
+            img_size (int, tuple): input image size
+            patch_size (int, tuple): patch size
+            in_chans (int): number of input channels
+            embed_dim (int): embedding dimension
+            depth (int): depth of transformer
+            num_heads (int): number of attention heads
+            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
+            qkv_bias (bool): enable bias for qkv if True
+            proj_bias (bool): enable bias for proj in attn if True
+            ffn_bias (bool): enable bias for ffn if True
+            drop_path_rate (float): stochastic depth rate
+            drop_path_uniform (bool): apply uniform drop rate across blocks
+            weight_init (str): weight init scheme
+            init_values (float): layer-scale init values
+            embed_layer (nn.Module): patch embedding layer
+            act_layer (nn.Module): MLP activation layer
+            block_fn (nn.Module): transformer block class
+            ffn_layer (str): "mlp", "swiglu", "swiglufused" or "identity"
+            block_chunks: (int) split block sequence into block_chunks units for FSDP wrap
+            num_register_tokens: (int) number of extra cls tokens (so-called "registers")
+            interpolate_antialias: (str) flag to apply anti-aliasing when interpolating positional embeddings
+            interpolate_offset: (float) work-around offset to apply when interpolating positional embeddings
+        """
+        super().__init__()
+        norm_layer = partial(nn.LayerNorm, eps=1e-6)
+
+        # tricky but makes it work
+        self.use_checkpoint = False
+        #
+
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        self.num_tokens = 1
+        self.n_blocks = depth
+        self.num_heads = num_heads
+        self.patch_size = patch_size
+        self.num_register_tokens = num_register_tokens
+        self.interpolate_antialias = interpolate_antialias
+        self.interpolate_offset = interpolate_offset
+
+        self.patch_embed = embed_layer(img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim))
+        assert num_register_tokens >= 0
+        self.register_tokens = (
+            nn.Parameter(torch.zeros(1, num_register_tokens, embed_dim)) if num_register_tokens else None
+        )
+
+        if drop_path_uniform is True:
+            dpr = [drop_path_rate] * depth
+        else:
+            dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+
+        if ffn_layer == "mlp":
+            logger.info("using MLP layer as FFN")
+            ffn_layer = Mlp
+        elif ffn_layer == "swiglufused" or ffn_layer == "swiglu":
+            logger.info("using SwiGLU layer as FFN")
+            ffn_layer = SwiGLUFFNFused
+        elif ffn_layer == "identity":
+            logger.info("using Identity layer as FFN")
+
+            def f(*args, **kwargs):
+                return nn.Identity()
+
+            ffn_layer = f
+        else:
+            raise NotImplementedError
+
+        blocks_list = [
+            block_fn(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                proj_bias=proj_bias,
+                ffn_bias=ffn_bias,
+                drop_path=dpr[i],
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                ffn_layer=ffn_layer,
+                init_values=init_values,
+                qk_norm=qk_norm,
+            )
+            for i in range(depth)
+        ]
+        if block_chunks > 0:
+            self.chunked_blocks = True
+            chunked_blocks = []
+            chunksize = depth // block_chunks
+            for i in range(0, depth, chunksize):
+                # this is to keep the block index consistent if we chunk the block list
+                chunked_blocks.append([nn.Identity()] * i + blocks_list[i : i + chunksize])
+            self.blocks = nn.ModuleList([BlockChunk(p) for p in chunked_blocks])
+        else:
+            self.chunked_blocks = False
+            self.blocks = nn.ModuleList(blocks_list)
+
+        self.norm = norm_layer(embed_dim)
+        self.head = nn.Identity()
+
+        self.mask_token = nn.Parameter(torch.zeros(1, embed_dim))
+
+        self.init_weights()
+
+    def init_weights(self):
+        trunc_normal_(self.pos_embed, std=0.02)
+        nn.init.normal_(self.cls_token, std=1e-6)
+        if self.register_tokens is not None:
+            nn.init.normal_(self.register_tokens, std=1e-6)
+        named_apply(init_weights_vit_timm, self)
+
+    def interpolate_pos_encoding(self, x, w, h):
+        previous_dtype = x.dtype
+        npatch = x.shape[1] - 1
+        N = self.pos_embed.shape[1] - 1
+        if npatch == N and w == h:
+            return self.pos_embed
+        pos_embed = self.pos_embed.float()
+        class_pos_embed = pos_embed[:, 0]
+        patch_pos_embed = pos_embed[:, 1:]
+        dim = x.shape[-1]
+        w0 = w // self.patch_size
+        h0 = h // self.patch_size
+        M = int(math.sqrt(N))  # Recover the number of patches in each dimension
+        assert N == M * M
+        kwargs = {}
+        if self.interpolate_offset:
+            # Historical kludge: add a small number to avoid floating point error in the interpolation, see https://github.com/facebookresearch/dino/issues/8
+            # Note: still needed for backward-compatibility, the underlying operators are using both output size and scale factors
+            sx = float(w0 + self.interpolate_offset) / M
+            sy = float(h0 + self.interpolate_offset) / M
+            kwargs["scale_factor"] = (sx, sy)
+        else:
+            # Simply specify an output size instead of a scale factor
+            kwargs["size"] = (w0, h0)
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed.reshape(1, M, M, dim).permute(0, 3, 1, 2),
+            mode="bicubic",
+            antialias=self.interpolate_antialias,
+            **kwargs,
+        )
+        assert (w0, h0) == patch_pos_embed.shape[-2:]
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1).to(previous_dtype)
+
+    def prepare_tokens_with_masks(self, x, masks=None):
+        B, nc, w, h = x.shape
+        x = self.patch_embed(x)
+        if masks is not None:
+            x = torch.where(masks.unsqueeze(-1), self.mask_token.to(x.dtype).unsqueeze(0), x)
+
+        x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
+        x = x + self.interpolate_pos_encoding(x, w, h)
+
+        if self.register_tokens is not None:
+            x = torch.cat(
+                (
+                    x[:, :1],
+                    self.register_tokens.expand(x.shape[0], -1, -1),
+                    x[:, 1:],
+                ),
+                dim=1,
+            )
+
+        return x
+
+    def forward_features_list(self, x_list, masks_list):
+        x = [self.prepare_tokens_with_masks(x, masks) for x, masks in zip(x_list, masks_list)]
+
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint(blk, x, use_reentrant=self.use_reentrant)
+            else:
+                x = blk(x)
+
+        all_x = x
+        output = []
+        for x, masks in zip(all_x, masks_list):
+            x_norm = self.norm(x)
+            output.append(
+                {
+                    "x_norm_clstoken": x_norm[:, 0],
+                    "x_norm_regtokens": x_norm[:, 1 : self.num_register_tokens + 1],
+                    "x_norm_patchtokens": x_norm[:, self.num_register_tokens + 1 :],
+                    "x_prenorm": x,
+                    "masks": masks,
+                }
+            )
+        return output
+
+    def forward_features(self, x, masks=None):
+        if isinstance(x, list):
+            return self.forward_features_list(x, masks)
+
+        x = self.prepare_tokens_with_masks(x, masks)
+
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint(blk, x, use_reentrant=self.use_reentrant)
+            else:
+                x = blk(x)
+
+        x_norm = self.norm(x)
+        return {
+            "x_norm_clstoken": x_norm[:, 0],
+            "x_norm_regtokens": x_norm[:, 1 : self.num_register_tokens + 1],
+            "x_norm_patchtokens": x_norm[:, self.num_register_tokens + 1 :],
+            "x_prenorm": x,
+            "masks": masks,
+        }
+
+    def _get_intermediate_layers_not_chunked(self, x, n=1):
+        x = self.prepare_tokens_with_masks(x)
+        # If n is an int, take the n last blocks. If it's a list, take them
+        output, total_block_len = [], len(self.blocks)
+        blocks_to_take = range(total_block_len - n, total_block_len) if isinstance(n, int) else n
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if i in blocks_to_take:
+                output.append(x)
+        assert len(output) == len(blocks_to_take), f"only {len(output)} / {len(blocks_to_take)} blocks found"
+        return output
+
+    def _get_intermediate_layers_chunked(self, x, n=1):
+        x = self.prepare_tokens_with_masks(x)
+        output, i, total_block_len = [], 0, len(self.blocks[-1])
+        # If n is an int, take the n last blocks. If it's a list, take them
+        blocks_to_take = range(total_block_len - n, total_block_len) if isinstance(n, int) else n
+        for block_chunk in self.blocks:
+            for blk in block_chunk[i:]:  # Passing the nn.Identity()
+                x = blk(x)
+                if i in blocks_to_take:
+                    output.append(x)
+                i += 1
+        assert len(output) == len(blocks_to_take), f"only {len(output)} / {len(blocks_to_take)} blocks found"
+        return output
+
+    def get_intermediate_layers(
+        self,
+        x: torch.Tensor,
+        n: Union[int, Sequence] = 1,  # Layers or n last layers to take
+        reshape: bool = False,
+        return_class_token: bool = False,
+        norm=True,
+    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]]]:
+        if self.chunked_blocks:
+            outputs = self._get_intermediate_layers_chunked(x, n)
+        else:
+            outputs = self._get_intermediate_layers_not_chunked(x, n)
+        if norm:
+            outputs = [self.norm(out) for out in outputs]
+        class_tokens = [out[:, 0] for out in outputs]
+        outputs = [out[:, 1 + self.num_register_tokens :] for out in outputs]
+        if reshape:
+            B, _, w, h = x.shape
+            outputs = [
+                out.reshape(B, w // self.patch_size, h // self.patch_size, -1).permute(0, 3, 1, 2).contiguous()
+                for out in outputs
+            ]
+        if return_class_token:
+            return tuple(zip(outputs, class_tokens))
+        return tuple(outputs)
+
+    def forward(self, *args, is_training=True, **kwargs):
+        ret = self.forward_features(*args, **kwargs)
+        if is_training:
+            return ret
+        else:
+            return self.head(ret["x_norm_clstoken"])
+
+
+def init_weights_vit_timm(module: nn.Module, name: str = ""):
+    """ViT weight initialization, original timm impl (for reproducibility)"""
+    if isinstance(module, nn.Linear):
+        trunc_normal_(module.weight, std=0.02)
+        if module.bias is not None:
+            nn.init.zeros_(module.bias)
+
+
+def vit_small(patch_size=16, num_register_tokens=0, **kwargs):
+    model = DinoVisionTransformer(
+        patch_size=patch_size,
+        embed_dim=384,
+        depth=12,
+        num_heads=6,
+        mlp_ratio=4,
+        block_fn=partial(Block, attn_class=MemEffAttention),
+        num_register_tokens=num_register_tokens,
+        **kwargs,
+    )
+    return model
+
+
+def vit_base(patch_size=16, num_register_tokens=0, **kwargs):
+    model = DinoVisionTransformer(
+        patch_size=patch_size,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        mlp_ratio=4,
+        block_fn=partial(Block, attn_class=MemEffAttention),
+        num_register_tokens=num_register_tokens,
+        **kwargs,
+    )
+    return model
+
+
+def vit_large(patch_size=16, num_register_tokens=0, **kwargs):
+    model = DinoVisionTransformer(
+        patch_size=patch_size,
+        embed_dim=1024,
+        depth=24,
+        num_heads=16,
+        mlp_ratio=4,
+        block_fn=partial(Block, attn_class=MemEffAttention),
+        num_register_tokens=num_register_tokens,
+        **kwargs,
+    )
+    return model
+
+
+def vit_giant2(patch_size=16, num_register_tokens=0, **kwargs):
+    """
+    Close to ViT-giant, with embed-dim 1536 and 24 heads => embed-dim per head 64
+    """
+    model = DinoVisionTransformer(
+        patch_size=patch_size,
+        embed_dim=1536,
+        depth=40,
+        num_heads=24,
+        mlp_ratio=4,
+        block_fn=partial(Block, attn_class=MemEffAttention),
+        num_register_tokens=num_register_tokens,
+        **kwargs,
+    )
+    return model

vggt/models/aggregator.py

@@ -0,0 +1,331 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional, Tuple, Union, List, Dict, Any
+
+from vggt.layers import PatchEmbed
+from vggt.layers.block import Block
+from vggt.layers.rope import RotaryPositionEmbedding2D, PositionGetter
+from vggt.layers.vision_transformer import vit_small, vit_base, vit_large, vit_giant2
+
+logger = logging.getLogger(__name__)
+
+_RESNET_MEAN = [0.485, 0.456, 0.406]
+_RESNET_STD = [0.229, 0.224, 0.225]
+
+
+class Aggregator(nn.Module):
+    """
+    The Aggregator applies alternating-attention over input frames,
+    as described in VGGT: Visual Geometry Grounded Transformer.
+
+
+    Args:
+        img_size (int): Image size in pixels.
+        patch_size (int): Size of each patch for PatchEmbed.
+        embed_dim (int): Dimension of the token embeddings.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        mlp_ratio (float): Ratio of MLP hidden dim to embedding dim.
+        num_register_tokens (int): Number of register tokens.
+        block_fn (nn.Module): The block type used for attention (Block by default).
+        qkv_bias (bool): Whether to include bias in QKV projections.
+        proj_bias (bool): Whether to include bias in the output projection.
+        ffn_bias (bool): Whether to include bias in MLP layers.
+        patch_embed (str): Type of patch embed. e.g., "conv" or "dinov2_vitl14_reg".
+        aa_order (list[str]): The order of alternating attention, e.g. ["frame", "global"].
+        aa_block_size (int): How many blocks to group under each attention type before switching. If not necessary, set to 1.
+        qk_norm (bool): Whether to apply QK normalization.
+        rope_freq (int): Base frequency for rotary embedding. -1 to disable.
+        init_values (float): Init scale for layer scale.
+    """
+
+    def __init__(
+        self,
+        img_size=518,
+        patch_size=14,
+        embed_dim=1024,
+        depth=24,
+        num_heads=16,
+        mlp_ratio=4.0,
+        num_register_tokens=4,
+        block_fn=Block,
+        qkv_bias=True,
+        proj_bias=True,
+        ffn_bias=True,
+        patch_embed="dinov2_vitl14_reg",
+        aa_order=["frame", "global"],
+        aa_block_size=1,
+        qk_norm=True,
+        rope_freq=100,
+        init_values=0.01,
+    ):
+        super().__init__()
+
+        self.__build_patch_embed__(patch_embed, img_size, patch_size, num_register_tokens, embed_dim=embed_dim)
+
+        # Initialize rotary position embedding if frequency > 0
+        self.rope = RotaryPositionEmbedding2D(frequency=rope_freq) if rope_freq > 0 else None
+        self.position_getter = PositionGetter() if self.rope is not None else None
+
+        self.frame_blocks = nn.ModuleList(
+            [
+                block_fn(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    proj_bias=proj_bias,
+                    ffn_bias=ffn_bias,
+                    init_values=init_values,
+                    qk_norm=qk_norm,
+                    rope=self.rope,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+        self.global_blocks = nn.ModuleList(
+            [
+                block_fn(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    proj_bias=proj_bias,
+                    ffn_bias=ffn_bias,
+                    init_values=init_values,
+                    qk_norm=qk_norm,
+                    rope=self.rope,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+        self.depth = depth
+        self.aa_order = aa_order
+        self.patch_size = patch_size
+        self.aa_block_size = aa_block_size
+
+        # Validate that depth is divisible by aa_block_size
+        if self.depth % self.aa_block_size != 0:
+            raise ValueError(f"depth ({depth}) must be divisible by aa_block_size ({aa_block_size})")
+
+        self.aa_block_num = self.depth // self.aa_block_size
+
+        # Note: We have two camera tokens, one for the first frame and one for the rest
+        # The same applies for register tokens
+        self.camera_token = nn.Parameter(torch.randn(1, 2, 1, embed_dim))
+        self.register_token = nn.Parameter(torch.randn(1, 2, num_register_tokens, embed_dim))
+
+        # The patch tokens start after the camera and register tokens
+        self.patch_start_idx = 1 + num_register_tokens
+
+        # Initialize parameters with small values
+        nn.init.normal_(self.camera_token, std=1e-6)
+        nn.init.normal_(self.register_token, std=1e-6)
+
+        # Register normalization constants as buffers
+        for name, value in (
+            ("_resnet_mean", _RESNET_MEAN),
+            ("_resnet_std", _RESNET_STD),
+        ):
+            self.register_buffer(
+                name,
+                torch.FloatTensor(value).view(1, 1, 3, 1, 1),
+                persistent=False,
+            )
+
+    def __build_patch_embed__(
+        self,
+        patch_embed,
+        img_size,
+        patch_size,
+        num_register_tokens,
+        interpolate_antialias=True,
+        interpolate_offset=0.0,
+        block_chunks=0,
+        init_values=1.0,
+        embed_dim=1024,
+    ):
+        """
+        Build the patch embed layer. If 'conv', we use a
+        simple PatchEmbed conv layer. Otherwise, we use a vision transformer.
+        """
+
+        if "conv" in patch_embed:
+            self.patch_embed = PatchEmbed(img_size=img_size, patch_size=patch_size, in_chans=3, embed_dim=embed_dim)
+        else:
+            vit_models = {
+                "dinov2_vitl14_reg": vit_large,
+                "dinov2_vitb14_reg": vit_base,
+                "dinov2_vits14_reg": vit_small,
+                "dinov2_vitg2_reg": vit_giant2,
+            }
+
+            self.patch_embed = vit_models[patch_embed](
+                img_size=img_size,
+                patch_size=patch_size,
+                num_register_tokens=num_register_tokens,
+                interpolate_antialias=interpolate_antialias,
+                interpolate_offset=interpolate_offset,
+                block_chunks=block_chunks,
+                init_values=init_values,
+            )
+
+            # Disable gradient updates for mask token
+            if hasattr(self.patch_embed, "mask_token"):
+                self.patch_embed.mask_token.requires_grad_(False)
+
+    def forward(
+        self,
+        images: torch.Tensor,
+    ) -> Tuple[List[torch.Tensor], int]:
+        """
+        Args:
+            images (torch.Tensor): Input images with shape [B, S, 3, H, W], in range [0, 1].
+                B: batch size, S: sequence length, 3: RGB channels, H: height, W: width
+
+        Returns:
+            (list[torch.Tensor], int):
+                The list of outputs from the attention blocks,
+                and the patch_start_idx indicating where patch tokens begin.
+        """
+        B, S, C_in, H, W = images.shape
+
+        if C_in != 3:
+            raise ValueError(f"Expected 3 input channels, got {C_in}")
+
+        # Normalize images and reshape for patch embed
+        images = (images - self._resnet_mean) / self._resnet_std
+
+        # Reshape to [B*S, C, H, W] for patch embedding
+        images = images.view(B * S, C_in, H, W)
+        patch_tokens = self.patch_embed(images)
+
+        if isinstance(patch_tokens, dict):
+            patch_tokens = patch_tokens["x_norm_patchtokens"]
+
+        _, P, C = patch_tokens.shape
+
+        # Expand camera and register tokens to match batch size and sequence length
+        camera_token = slice_expand_and_flatten(self.camera_token, B, S)
+        register_token = slice_expand_and_flatten(self.register_token, B, S)
+
+        # Concatenate special tokens with patch tokens
+        tokens = torch.cat([camera_token, register_token, patch_tokens], dim=1)
+
+        pos = None
+        if self.rope is not None:
+            pos = self.position_getter(B * S, H // self.patch_size, W // self.patch_size, device=images.device)
+
+        if self.patch_start_idx > 0:
+            # do not use position embedding for special tokens (camera and register tokens)
+            # so set pos to 0 for the special tokens
+            pos = pos + 1
+            pos_special = torch.zeros(B * S, self.patch_start_idx, 2).to(images.device).to(pos.dtype)
+            pos = torch.cat([pos_special, pos], dim=1)
+
+        # update P because we added special tokens
+        _, P, C = tokens.shape
+
+        frame_idx = 0
+        global_idx = 0
+        output_list = []
+
+        for _ in range(self.aa_block_num):
+            for attn_type in self.aa_order:
+                if attn_type == "frame":
+                    tokens, frame_idx, frame_intermediates = self._process_frame_attention(
+                        tokens, B, S, P, C, frame_idx, pos=pos
+                    )
+                elif attn_type == "global":
+                    tokens, global_idx, global_intermediates = self._process_global_attention(
+                        tokens, B, S, P, C, global_idx, pos=pos
+                    )
+                else:
+                    raise ValueError(f"Unknown attention type: {attn_type}")
+
+            for i in range(len(frame_intermediates)):
+                # concat frame and global intermediates, [B x S x P x 2C]
+                concat_inter = torch.cat([frame_intermediates[i], global_intermediates[i]], dim=-1)
+                output_list.append(concat_inter)
+
+        del concat_inter
+        del frame_intermediates
+        del global_intermediates
+        return output_list, self.patch_start_idx
+
+    def _process_frame_attention(self, tokens, B, S, P, C, frame_idx, pos=None):
+        """
+        Process frame attention blocks. We keep tokens in shape (B*S, P, C).
+        """
+        # If needed, reshape tokens or positions:
+        if tokens.shape != (B * S, P, C):
+            tokens = tokens.view(B, S, P, C).view(B * S, P, C)
+
+        if pos is not None and pos.shape != (B * S, P, 2):
+            pos = pos.view(B, S, P, 2).view(B * S, P, 2)
+
+        intermediates = []
+
+        # by default, self.aa_block_size=1, which processes one block at a time
+        for _ in range(self.aa_block_size):
+            tokens = self.frame_blocks[frame_idx](tokens, pos=pos)
+            frame_idx += 1
+            intermediates.append(tokens.view(B, S, P, C))
+
+        return tokens, frame_idx, intermediates
+
+    def _process_global_attention(self, tokens, B, S, P, C, global_idx, pos=None):
+        """
+        Process global attention blocks. We keep tokens in shape (B, S*P, C).
+        """
+        if tokens.shape != (B, S * P, C):
+            tokens = tokens.view(B, S, P, C).view(B, S * P, C)
+
+        if pos is not None and pos.shape != (B, S * P, 2):
+            pos = pos.view(B, S, P, 2).view(B, S * P, 2)
+
+        intermediates = []
+
+        # by default, self.aa_block_size=1, which processes one block at a time
+        for _ in range(self.aa_block_size):
+            tokens = self.global_blocks[global_idx](tokens, pos=pos)
+            global_idx += 1
+            intermediates.append(tokens.view(B, S, P, C))
+
+        return tokens, global_idx, intermediates
+
+
+def slice_expand_and_flatten(token_tensor, B, S):
+    """
+    Processes specialized tokens with shape (1, 2, X, C) for multi-frame processing:
+    1) Uses the first position (index=0) for the first frame only
+    2) Uses the second position (index=1) for all remaining frames (S-1 frames)
+    3) Expands both to match batch size B
+    4) Concatenates to form (B, S, X, C) where each sequence has 1 first-position token
+       followed by (S-1) second-position tokens
+    5) Flattens to (B*S, X, C) for processing
+
+    Returns:
+        torch.Tensor: Processed tokens with shape (B*S, X, C)
+    """
+
+    # Slice out the "query" tokens => shape (1, 1, ...)
+    query = token_tensor[:, 0:1, ...].expand(B, 1, *token_tensor.shape[2:])
+    # Slice out the "other" tokens => shape (1, S-1, ...)
+    others = token_tensor[:, 1:, ...].expand(B, S - 1, *token_tensor.shape[2:])
+    # Concatenate => shape (B, S, ...)
+    combined = torch.cat([query, others], dim=1)
+
+    # Finally flatten => shape (B*S, ...)
+    combined = combined.view(B * S, *combined.shape[2:])
+    return combined

vggt/models/vggt.py

@@ -0,0 +1,81 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+from huggingface_hub import PyTorchModelHubMixin  # used for model hub
+
+from vggt.models.aggregator import Aggregator
+from vggt.heads.camera_head import CameraHead
+from vggt.heads.dpt_head import DPTHead
+from vggt.heads.track_head import TrackHead
+
+
+class VGGT(nn.Module, PyTorchModelHubMixin):
+    def __init__(self, img_size=518, patch_size=14, embed_dim=1024):
+        super().__init__()
+
+        self.aggregator = Aggregator(img_size=img_size, patch_size=patch_size, embed_dim=embed_dim)
+        self.camera_head = CameraHead(dim_in=2 * embed_dim)
+        self.point_head = DPTHead(dim_in=2 * embed_dim, output_dim=4, activation="inv_log", conf_activation="expp1")
+        self.depth_head = DPTHead(dim_in=2 * embed_dim, output_dim=2, activation="exp", conf_activation="expp1")
+        self.track_head = TrackHead(dim_in=2 * embed_dim, patch_size=patch_size)
+
+    def forward(
+        self,
+        images: torch.Tensor,
+        query_points: torch.Tensor = None,
+    ):
+        """
+        Forward pass of the VGGT model.
+
+        Args:
+            images (torch.Tensor): Input images with shape [S, 3, H, W] or [B, S, 3, H, W], in range [0, 1].
+                B: batch size, S: sequence length, 3: RGB channels, H: height, W: width
+            query_points (torch.Tensor, optional): Query points for tracking, in pixel coordinates.
+                Shape: [N, 2] or [B, N, 2], where N is the number of query points.
+                Default: None
+        """
+        # If without batch dimension, add it
+        if len(images.shape) == 4:
+            images = images.unsqueeze(0)
+        if query_points is not None and len(query_points.shape) == 2:
+            query_points = query_points.unsqueeze(0)
+
+        aggregated_tokens_list, patch_start_idx = self.aggregator(images)
+
+        predictions = {}
+
+        with torch.cuda.amp.autocast(enabled=False):
+            if self.camera_head is not None:
+                pose_enc_list = self.camera_head(aggregated_tokens_list)
+                predictions["pose_enc"] = pose_enc_list[-1]  # pose encoding of the last iteration
+
+            if self.depth_head is not None:
+                depth, depth_conf = self.depth_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
+                )
+                predictions["depth"] = depth
+                predictions["depth_conf"] = depth_conf
+
+            if self.point_head is not None:
+                pts3d, pts3d_conf = self.point_head(
+                    aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
+                )
+                predictions["world_points"] = pts3d
+                predictions["world_points_conf"] = pts3d_conf
+
+        if self.track_head is not None and query_points is not None:
+            track_list, vis, conf = self.track_head(
+                aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx, query_points=query_points
+            )
+            predictions["track"] = track_list[-1]  # track of the last iteration
+            predictions["vis"] = vis
+            predictions["conf"] = conf
+
+        predictions["images"] = images
+
+        return predictions

vggt/utils/geometry.py

@@ -0,0 +1,167 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import torch
+import numpy as np
+from typing import Tuple
+
+
+def unproject_depth_map_to_point_map(
+    depth_map: np.ndarray, extrinsics_cam: np.ndarray, intrinsics_cam: np.ndarray
+) -> np.ndarray:
+    """
+    Unproject a batch of depth maps to 3D world coordinates.
+
+    Args:
+        depth_map (np.ndarray): Batch of depth maps of shape (S, H, W, 1) or (S, H, W)
+        extrinsics_cam (np.ndarray): Batch of camera extrinsic matrices of shape (S, 3, 4)
+        intrinsics_cam (np.ndarray): Batch of camera intrinsic matrices of shape (S, 3, 3)
+
+    Returns:
+        np.ndarray: Batch of 3D world coordinates of shape (S, H, W, 3)
+    """
+    if isinstance(depth_map, torch.Tensor):
+        depth_map = depth_map.cpu().numpy()
+    if isinstance(extrinsics_cam, torch.Tensor):
+        extrinsics_cam = extrinsics_cam.cpu().numpy()
+    if isinstance(intrinsics_cam, torch.Tensor):
+        intrinsics_cam = intrinsics_cam.cpu().numpy()
+
+    world_points_list = []
+    for frame_idx in range(depth_map.shape[0]):
+        cur_world_points, _, _ = depth_to_world_coords_points(
+            depth_map[frame_idx].squeeze(-1), extrinsics_cam[frame_idx], intrinsics_cam[frame_idx]
+        )
+        world_points_list.append(cur_world_points)
+    world_points_array = np.stack(world_points_list, axis=0)
+
+    return world_points_array
+
+
+def depth_to_world_coords_points(
+    depth_map: np.ndarray,
+    extrinsic: np.ndarray,
+    intrinsic: np.ndarray,
+    eps=1e-8,
+) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """
+    Convert a depth map to world coordinates.
+
+    Args:
+        depth_map (np.ndarray): Depth map of shape (H, W).
+        intrinsic (np.ndarray): Camera intrinsic matrix of shape (3, 3).
+        extrinsic (np.ndarray): Camera extrinsic matrix of shape (3, 4). OpenCV camera coordinate convention, cam from world.
+
+    Returns:
+        tuple[np.ndarray, np.ndarray]: World coordinates (H, W, 3) and valid depth mask (H, W).
+    """
+    if depth_map is None:
+        return None, None, None
+
+    # Valid depth mask
+    point_mask = depth_map > eps
+
+    # Convert depth map to camera coordinates
+    cam_coords_points = depth_to_cam_coords_points(depth_map, intrinsic)
+
+    # Multiply with the inverse of extrinsic matrix to transform to world coordinates
+    # extrinsic_inv is 4x4 (note closed_form_inverse_OpenCV is batched, the output is (N, 4, 4))
+    cam_to_world_extrinsic = closed_form_inverse_se3(extrinsic[None])[0]
+
+    R_cam_to_world = cam_to_world_extrinsic[:3, :3]
+    t_cam_to_world = cam_to_world_extrinsic[:3, 3]
+
+    # Apply the rotation and translation to the camera coordinates
+    world_coords_points = np.dot(cam_coords_points, R_cam_to_world.T) + t_cam_to_world  # HxWx3, 3x3 -> HxWx3
+    # world_coords_points = np.einsum("ij,hwj->hwi", R_cam_to_world, cam_coords_points) + t_cam_to_world
+
+    return world_coords_points, cam_coords_points, point_mask
+
+
+def depth_to_cam_coords_points(depth_map: np.ndarray, intrinsic: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+    """
+    Convert a depth map to camera coordinates.
+
+    Args:
+        depth_map (np.ndarray): Depth map of shape (H, W).
+        intrinsic (np.ndarray): Camera intrinsic matrix of shape (3, 3).
+
+    Returns:
+        tuple[np.ndarray, np.ndarray]: Camera coordinates (H, W, 3)
+    """
+    H, W = depth_map.shape
+    assert intrinsic.shape == (3, 3), "Intrinsic matrix must be 3x3"
+    assert intrinsic[0, 1] == 0 and intrinsic[1, 0] == 0, "Intrinsic matrix must have zero skew"
+
+    # Intrinsic parameters
+    fu, fv = intrinsic[0, 0], intrinsic[1, 1]
+    cu, cv = intrinsic[0, 2], intrinsic[1, 2]
+
+    # Generate grid of pixel coordinates
+    u, v = np.meshgrid(np.arange(W), np.arange(H))
+
+    # Unproject to camera coordinates
+    x_cam = (u - cu) * depth_map / fu
+    y_cam = (v - cv) * depth_map / fv
+    z_cam = depth_map
+
+    # Stack to form camera coordinates
+    cam_coords = np.stack((x_cam, y_cam, z_cam), axis=-1).astype(np.float32)
+
+    return cam_coords
+
+
+def closed_form_inverse_se3(se3, R=None, T=None):
+    """
+    Compute the inverse of each 4x4 (or 3x4) SE3 matrix in a batch.
+
+    If `R` and `T` are provided, they must correspond to the rotation and translation
+    components of `se3`. Otherwise, they will be extracted from `se3`.
+
+    Args:
+        se3: Nx4x4 or Nx3x4 array or tensor of SE3 matrices.
+        R (optional): Nx3x3 array or tensor of rotation matrices.
+        T (optional): Nx3x1 array or tensor of translation vectors.
+
+    Returns:
+        Inverted SE3 matrices with the same type and device as `se3`.
+
+    Shapes:
+        se3: (N, 4, 4)
+        R: (N, 3, 3)
+        T: (N, 3, 1)
+    """
+    # Check if se3 is a numpy array or a torch tensor
+    is_numpy = isinstance(se3, np.ndarray)
+
+    # Validate shapes
+    if se3.shape[-2:] != (4, 4) and se3.shape[-2:] != (3, 4):
+        raise ValueError(f"se3 must be of shape (N,4,4), got {se3.shape}.")
+
+    # Extract R and T if not provided
+    if R is None:
+        R = se3[:, :3, :3]  # (N,3,3)
+    if T is None:
+        T = se3[:, :3, 3:]  # (N,3,1)
+
+    # Transpose R
+    if is_numpy:
+        # Compute the transpose of the rotation for NumPy
+        R_transposed = np.transpose(R, (0, 2, 1))
+        # -R^T t for NumPy
+        top_right = -np.matmul(R_transposed, T)
+        inverted_matrix = np.tile(np.eye(4), (len(R), 1, 1))
+    else:
+        R_transposed = R.transpose(1, 2)  # (N,3,3)
+        top_right = -torch.bmm(R_transposed, T)  # (N,3,1)
+        inverted_matrix = torch.eye(4, 4)[None].repeat(len(R), 1, 1)
+        inverted_matrix = inverted_matrix.to(R.dtype).to(R.device)
+
+    inverted_matrix[:, :3, :3] = R_transposed
+    inverted_matrix[:, :3, 3:] = top_right
+
+    return inverted_matrix

vggt/utils/load_fn.py

@@ -0,0 +1,111 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from PIL import Image
+from torchvision import transforms as TF
+
+
+def load_and_preprocess_images(image_path_list):
+    """
+    A quick start function to load and preprocess images for model input.
+    This assumes the images should have the same shape for easier batching, but our model can also work well with different shapes.
+
+    Args:
+        image_path_list (list): List of paths to image files
+
+    Returns:
+        torch.Tensor: Batched tensor of preprocessed images with shape (N, 3, H, W)
+
+    Raises:
+        ValueError: If the input list is empty
+
+    Notes:
+        - Images with different dimensions will be padded with white (value=1.0)
+        - A warning is printed when images have different shapes
+        - The function ensures width=518px while maintaining aspect ratio
+        - Height is adjusted to be divisible by 14 for compatibility with model requirements
+    """
+    # Check for empty list
+    if len(image_path_list) == 0:
+        raise ValueError("At least 1 image is required")
+
+    images = []
+    shapes = set()
+    to_tensor = TF.ToTensor()
+
+    # First process all images and collect their shapes
+    for image_path in image_path_list:
+
+        # Open image
+        img = Image.open(image_path)
+
+        # If there's an alpha channel, blend onto white background:
+        if img.mode == 'RGBA':
+            # Create white background
+            background = Image.new('RGBA', img.size, (255, 255, 255, 255))
+            # Alpha composite onto the white background
+            img = Image.alpha_composite(background, img)
+
+        # Now convert to "RGB" (this step assigns white for transparent areas)
+        img = img.convert("RGB")
+
+        width, height = img.size
+        new_width = 518
+
+        # Calculate height maintaining aspect ratio, divisible by 14
+        new_height = round(height * (new_width / width) / 14) * 14
+
+        # Resize with new dimensions (width, height)
+
+        img = img.resize((new_width, new_height), Image.Resampling.BICUBIC)
+        img = to_tensor(img)  # Convert to tensor (0, 1)
+
+        # Center crop height if it's larger than 518
+
+        if new_height > 518:
+            start_y = (new_height - 518) // 2
+            img = img[:, start_y : start_y + 518, :]
+
+        shapes.add((img.shape[1], img.shape[2]))
+        images.append(img)
+
+    # Check if we have different shapes
+    # In theory our model can also work well with different shapes
+
+    if len(shapes) > 1:
+        print(f"Warning: Found images with different shapes: {shapes}")
+        # Find maximum dimensions
+        max_height = max(shape[0] for shape in shapes)
+        max_width = max(shape[1] for shape in shapes)
+
+        # Pad images if necessary
+        padded_images = []
+        for img in images:
+            h_padding = max_height - img.shape[1]
+            w_padding = max_width - img.shape[2]
+
+            if h_padding > 0 or w_padding > 0:
+                pad_top = h_padding // 2
+                pad_bottom = h_padding - pad_top
+                pad_left = w_padding // 2
+                pad_right = w_padding - pad_left
+
+                img = torch.nn.functional.pad(
+                    img, (pad_left, pad_right, pad_top, pad_bottom), mode="constant", value=1.0
+                )
+            padded_images.append(img)
+        images = padded_images
+
+    images = torch.stack(images)  # concatenate images
+
+    # Ensure correct shape when single image
+    if len(image_path_list) == 1:
+        # Verify shape is (1, C, H, W)
+        if images.dim() == 3:
+            images = images.unsqueeze(0)
+
+    return images

vggt/utils/pose_enc.py

@@ -0,0 +1,73 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from .rotation import quat_to_mat, mat_to_quat
+
+
+def extri_intri_to_pose_encoding(
+    extrinsics,
+    intrinsics,
+    image_size_hw=None,  # e.g., (256, 512)
+    pose_encoding_type="absT_quaR_FoV",
+    min_focal_length=0.1,
+    max_focal_length=10,
+):
+
+    # extrinsics: BxSx3x4
+    # intrinsics: BxSx3x3
+
+    if pose_encoding_type == "absT_quaR_FoV":
+        R = extrinsics[:, :, :3, :3]  # BxSx3x3
+        T = extrinsics[:, :, :3, 3]  # BxSx3
+
+        quat = mat_to_quat(R)
+        # R_reverse = quat_to_mat(quat)
+        # Note the order of h and w here
+        H, W = image_size_hw
+        fov_h = 2 * torch.atan((H / 2) / intrinsics[..., 1, 1])
+        fov_w = 2 * torch.atan((W / 2) / intrinsics[..., 0, 0])
+        pose_encoding = torch.cat([T, quat, fov_h[..., None], fov_w[..., None]], dim=-1).float()
+    else:
+        raise NotImplementedError
+
+    return pose_encoding
+
+
+def pose_encoding_to_extri_intri(
+    pose_encoding,
+    image_size_hw=None,  # e.g., (256, 512)
+    min_focal_length=0.1,
+    max_focal_length=10,
+    pose_encoding_type="absT_quaR_FoV",
+    build_intrinsics=True,
+):
+
+    intrinsics = None
+
+    if pose_encoding_type == "absT_quaR_FoV":
+        T = pose_encoding[..., :3]
+        quat = pose_encoding[..., 3:7]
+        fov_h = pose_encoding[..., 7]
+        fov_w = pose_encoding[..., 8]
+
+        R = quat_to_mat(quat)
+        extrinsics = torch.cat([R, T[..., None]], dim=-1)
+
+        if build_intrinsics:
+            H, W = image_size_hw
+            fy = (H / 2.0) / torch.tan(fov_h / 2.0)
+            fx = (W / 2.0) / torch.tan(fov_w / 2.0)
+            intrinsics = torch.zeros(pose_encoding.shape[:2] + (3, 3), device=pose_encoding.device)
+            intrinsics[..., 0, 0] = fx
+            intrinsics[..., 1, 1] = fy
+            intrinsics[..., 0, 2] = W / 2
+            intrinsics[..., 1, 2] = H / 2
+            intrinsics[..., 2, 2] = 1.0  # Set the homogeneous coordinate to 1
+    else:
+        raise NotImplementedError
+
+    return extrinsics, intrinsics

vggt/utils/rotation.py

@@ -0,0 +1,138 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Modified from PyTorch3D, https://github.com/facebookresearch/pytorch3d
+
+import torch
+import numpy as np
+import torch.nn.functional as F
+
+
+def quat_to_mat(quaternions: torch.Tensor) -> torch.Tensor:
+    """
+    Quaternion Order: XYZW or say ijkr, scalar-last
+
+    Convert rotations given as quaternions to rotation matrices.
+    Args:
+        quaternions: quaternions with real part last,
+            as tensor of shape (..., 4).
+
+    Returns:
+        Rotation matrices as tensor of shape (..., 3, 3).
+    """
+    i, j, k, r = torch.unbind(quaternions, -1)
+    # pyre-fixme[58]: `/` is not supported for operand types `float` and `Tensor`.
+    two_s = 2.0 / (quaternions * quaternions).sum(-1)
+
+    o = torch.stack(
+        (
+            1 - two_s * (j * j + k * k),
+            two_s * (i * j - k * r),
+            two_s * (i * k + j * r),
+            two_s * (i * j + k * r),
+            1 - two_s * (i * i + k * k),
+            two_s * (j * k - i * r),
+            two_s * (i * k - j * r),
+            two_s * (j * k + i * r),
+            1 - two_s * (i * i + j * j),
+        ),
+        -1,
+    )
+    return o.reshape(quaternions.shape[:-1] + (3, 3))
+
+
+def mat_to_quat(matrix: torch.Tensor) -> torch.Tensor:
+    """
+    Convert rotations given as rotation matrices to quaternions.
+
+    Args:
+        matrix: Rotation matrices as tensor of shape (..., 3, 3).
+
+    Returns:
+        quaternions with real part last, as tensor of shape (..., 4).
+        Quaternion Order: XYZW or say ijkr, scalar-last
+    """
+    if matrix.size(-1) != 3 or matrix.size(-2) != 3:
+        raise ValueError(f"Invalid rotation matrix shape {matrix.shape}.")
+
+    batch_dim = matrix.shape[:-2]
+    m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind(matrix.reshape(batch_dim + (9,)), dim=-1)
+
+    q_abs = _sqrt_positive_part(
+        torch.stack(
+            [
+                1.0 + m00 + m11 + m22,
+                1.0 + m00 - m11 - m22,
+                1.0 - m00 + m11 - m22,
+                1.0 - m00 - m11 + m22,
+            ],
+            dim=-1,
+        )
+    )
+
+    # we produce the desired quaternion multiplied by each of r, i, j, k
+    quat_by_rijk = torch.stack(
+        [
+            # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
+            #  `int`.
+            torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1),
+            # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
+            #  `int`.
+            torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1),
+            # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
+            #  `int`.
+            torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1),
+            # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
+            #  `int`.
+            torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1),
+        ],
+        dim=-2,
+    )
+
+    # We floor here at 0.1 but the exact level is not important; if q_abs is small,
+    # the candidate won't be picked.
+    flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device)
+    quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr))
+
+    # if not for numerical problems, quat_candidates[i] should be same (up to a sign),
+    # forall i; we pick the best-conditioned one (with the largest denominator)
+    out = quat_candidates[F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, :].reshape(batch_dim + (4,))
+
+    # Convert from rijk to ijkr
+    out = out[..., [1, 2, 3, 0]]
+
+    out = standardize_quaternion(out)
+
+    return out
+
+
+def _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:
+    """
+    Returns torch.sqrt(torch.max(0, x))
+    but with a zero subgradient where x is 0.
+    """
+    ret = torch.zeros_like(x)
+    positive_mask = x > 0
+    if torch.is_grad_enabled():
+        ret[positive_mask] = torch.sqrt(x[positive_mask])
+    else:
+        ret = torch.where(positive_mask, torch.sqrt(x), ret)
+    return ret
+
+
+def standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:
+    """
+    Convert a unit quaternion to a standard form: one in which the real
+    part is non negative.
+
+    Args:
+        quaternions: Quaternions with real part last,
+            as tensor of shape (..., 4).
+
+    Returns:
+        Standardized quaternions as tensor of shape (..., 4).
+    """
+    return torch.where(quaternions[..., 3:4] < 0, -quaternions, quaternions)