update

app.py

@@ -0,0 +1,9 @@
+import gradio as gr
+import spaces
+
+@spaces.GPU
+def greet(name):
+    return "Hello " + name + "!!"
+
+demo = gr.Interface(fn=greet, inputs="text", outputs="text")
+demo.launch()

demo/gradio_demo2.py

@@ -0,0 +1,369 @@
+import gradio as gr
+import spaces
+from PIL import Image, ImageDraw, ImageFont
+import re
+import random
+import numpy as np
+from skimage.measure import label, regionprops
+from skimage.morphology import binary_dilation, disk
+from sam3.model_builder import build_sam3_image_model
+from sam3.model.sam3_image_processor import Sam3Processor
+from sam3.visualization_utils import plot_bbox, plot_mask, COLORS
+import matplotlib.pyplot as plt
+
+from detect_tools.upn import UPNWrapper
+from vlm_fo1.model.builder import load_pretrained_model
+from vlm_fo1.mm_utils import (
+    prepare_inputs,
+    extract_predictions_to_indexes,
+)
+from vlm_fo1.task_templates import *
+import torch
+import os
+from copy import deepcopy
+
+
+EXAMPLES = [
+    ["demo/sam3_examples/00000-72.jpg","airplane with letter AE on its body"],
+    ["demo/sam3_examples/00000-32.jpg","the lying cat which is not black"],
+    ["demo/sam3_examples/00000-22.jpg","person wearing a black top"],
+    ["demo/sam3_examples/000000378453.jpg", "zebra inside the mud puddle"],
+]
+
+
+def get_valid_examples():
+    valid_examples = []
+    demo_dir = os.path.dirname(os.path.abspath(__file__))
+    for example in EXAMPLES:
+        img_path = example[0]
+        full_path = os.path.join(demo_dir, img_path)
+        if os.path.exists(full_path):
+            valid_examples.append([
+                full_path,
+                example[1], 
+            ])
+        elif os.path.exists(img_path):
+            valid_examples.append([
+                img_path,
+                example[1], 
+            ])
+    return valid_examples
+
+
+def detect_model_upn(image, threshold=0.3):
+    proposals = upn_model.inference(image)
+    filtered_proposals = upn_model.filter(proposals, min_score=threshold)
+    picked_proposals = filtered_proposals['original_xyxy_boxes'][0][:100]
+    return picked_proposals
+
+
+def detect_model_sam3(image, text, threshold=0.3):
+    inference_state = sam3_processor.set_image(image)
+    output = sam3_processor.set_text_prompt(state=inference_state, prompt=text)
+    boxes, scores, masks = output["boxes"], output["scores"], output["masks"]   
+    sorted_indices = torch.argsort(scores, descending=True)
+    boxes = boxes[sorted_indices][:100, :]
+    scores = scores[sorted_indices][:100]
+    masks = masks[sorted_indices][:100]
+
+    output = {
+        "boxes": boxes,
+        "scores": scores,
+        "masks": masks,
+    }
+    return boxes.tolist(), scores.tolist(), masks.tolist(), output
+
+
+def multimodal_model(image, bboxes, text, scores=None):
+    if len(bboxes) == 0:
+        return None, {}, []
+    
+    if '<image>' in text:
+        print(text)
+        parts = [part.replace('\\n', '\n') for part in re.split(rf'(<image>)', text) if part.strip()]
+        print(parts)
+        content = []
+        for part in parts:
+            if part == '<image>':
+                content.append({"type": "image_url", "image_url": {"url": image}})
+            else:
+                content.append({"type": "text", "text": part})
+    else:
+        content = [{
+            "type": "image_url",
+            "image_url": {
+                "url": image
+            }
+            }, {
+                "type": "text",
+                "text": text
+            }]
+
+    messages = [
+        {
+            "role": "user",
+            "content": content,
+            "bbox_list": bboxes
+        }
+    ]
+    generation_kwargs = prepare_inputs(model_path, model, image_processors, tokenizer, messages,
+    max_tokens=4096, top_p=0.05, temperature=0.0, do_sample=False, image_size=1024)
+    with torch.inference_mode():
+        output_ids = model.generate(**generation_kwargs)
+        outputs = tokenizer.decode(output_ids[0, generation_kwargs['inputs'].shape[1]:]).strip()
+        print("========output========\n", outputs)
+
+    if '<ground>' in outputs:
+        prediction_dict = extract_predictions_to_indexes(outputs)
+    else:
+        match_pattern = r"<region(\d+)>"
+        matches = re.findall(match_pattern, outputs)
+        prediction_dict = {f"<region{m}>": {int(m)} for m in matches}
+
+    ans_bbox_json = []
+    ans_bbox_list = []
+    for k, v in prediction_dict.items():
+        for box_index in v:
+            box_index = int(box_index)
+            if box_index < len(bboxes):
+                current_bbox = bboxes[box_index]
+                item = {
+                    "region_index": f"<region{box_index}>",
+                    "xmin": current_bbox[0],
+                    "ymin": current_bbox[1],
+                    "xmax": current_bbox[2],
+                    "ymax": current_bbox[3],
+                    "label": k,
+                }
+                if scores is not None and box_index < len(scores):
+                    item["score"] = scores[box_index]
+                
+                ans_bbox_json.append(item)
+                ans_bbox_list.append(current_bbox)
+
+    return outputs, ans_bbox_json, ans_bbox_list
+
+
+def draw_sam3_results(img, results):
+    fig, ax = plt.subplots(figsize=(12, 8))
+    # fig.subplots_adjust(0, 0, 1, 1)
+    ax.imshow(img)
+    nb_objects = len(results["scores"])
+    print(f"found {nb_objects} object(s)")
+    for i in range(nb_objects):
+        color = COLORS[i % len(COLORS)]
+        plot_mask(results["masks"][i].squeeze(0).cpu(), color=color)
+        w, h = img.size
+        prob = results["scores"][i].item()
+        plot_bbox(
+            h,
+            w,
+            results["boxes"][i].cpu(),
+            text=f"(id={i}, {prob=:.2f})",
+            box_format="XYXY",
+            color=color,
+            relative_coords=False,
+        )
+    ax.axis("off")
+    fig.tight_layout(pad=0)
+    
+    # Convert matplotlib figure to PIL Image
+    fig.canvas.draw()
+    buf = fig.canvas.buffer_rgba()
+    pil_img = Image.frombytes('RGBA', fig.canvas.get_width_height(), buf)
+    plt.close(fig)
+    
+    return pil_img
+
+
+def draw_bboxes_simple(image, bboxes, labels=None):
+    image = image.copy()
+    draw = ImageDraw.Draw(image)
+
+    for bbox in bboxes:
+        draw.rectangle(bbox, outline="red", width=3)
+    return image
+
+@spaces.GPU
+def process(image, prompt, threshold=0.3):
+    if image is None:
+        error_msg = "Error: Please upload an image or select a valid example."
+        print(f"Error: image is None, original input type: {type(image)}")
+        return None, None, None, None, [], []
+    
+    try:
+        image = image.convert('RGB')
+    except Exception as e:
+        error_msg = f"Error: Cannot process image - {str(e)}"
+        return None, None, None, None, [], []
+
+    # --- SAM3 Pipeline ---
+    print("Running SAM3 Pipeline...")
+    sam3_bboxes, sam3_scores, masks, sam3_output = detect_model_sam3(image, prompt, threshold)
+
+    # Generate SAM3 outputs (Directly from SAM3, no VLM-FO1)
+    sam3_detection_image = draw_sam3_results(image, sam3_output)
+    
+    sam3_annotated_bboxes = []
+    sam3_ans_bbox_json = []
+    
+    img_width, img_height = image.size
+    for i, bbox in enumerate(sam3_bboxes):
+        xmin = max(0, min(img_width, int(bbox[0])))
+        ymin = max(0, min(img_height, int(bbox[1])))
+        xmax = max(0, min(img_width, int(bbox[2])))
+        ymax = max(0, min(img_height, int(bbox[3])))
+        score = sam3_scores[i]
+        
+        # Format label with score
+        label_text = f"{prompt} {score:.2f}"
+        
+        sam3_annotated_bboxes.append(
+            ((xmin, ymin, xmax, ymax), label_text)
+        )
+        
+        sam3_ans_bbox_json.append({
+            "region_index": i,
+            "xmin": bbox[0],
+            "ymin": bbox[1],
+            "xmax": bbox[2],
+            "ymax": bbox[3],
+            "label": prompt,
+            "score": score
+        })
+        
+    sam3_annotated_image = (image, sam3_annotated_bboxes)
+
+    # --- UPN Pipeline ---
+    print("Running UPN Pipeline...")
+    upn_bboxes = detect_model_upn(image, threshold=0.3) # Use default threshold for UPN
+    
+    fo1_prompt_upn = OD_template.format(prompt)
+    upn_bboxes = upn_bboxes[::-1]
+    upn_ans, upn_ans_bbox_json, upn_ans_bbox_list = multimodal_model(image, upn_bboxes, fo1_prompt_upn)
+    
+    upn_detection_image = draw_bboxes_simple(image, upn_bboxes)
+    
+    upn_annotated_bboxes = []
+    if len(upn_ans_bbox_json) > 0:
+        img_width, img_height = image.size
+        for item in upn_ans_bbox_json:
+            xmin = max(0, min(img_width, int(item['xmin'])))
+            ymin = max(0, min(img_height, int(item['ymin'])))
+            xmax = max(0, min(img_width, int(item['xmax'])))
+            ymax = max(0, min(img_height, int(item['ymax'])))
+            upn_annotated_bboxes.append(
+                ((xmin, ymin, xmax, ymax), item['label'])
+            )
+    upn_annotated_image = (image, upn_annotated_bboxes)
+
+
+    return sam3_annotated_image, sam3_detection_image, \
+           upn_annotated_image, upn_detection_image, upn_ans_bbox_json
+
+
+def update_btn(is_processing):
+    if is_processing:
+        return gr.update(value="Processing...", interactive=False)
+    else:
+        return gr.update(value="Submit", interactive=True)
+
+
+def launch_demo():
+    with gr.Blocks() as demo:
+        gr.Markdown("# 🚀 VLM-FO1 vs SAM3 Demo")
+        gr.Markdown("""
+        ### 📋 Instructions
+        Compare the detection performance of **SAM3** vs **VLM-FO1**.
+        
+        **How it works**
+        1. Upload or pick an example image.
+        2. Describe the target object in natural language.
+        3. Hit **Submit** to run both pipelines.
+        """)
+        
+        with gr.Row():
+            with gr.Column():
+                img_input_draw = gr.Image(
+                    label="Image Input",
+                    type="pil",
+                    sources=['upload'],
+                )
+
+                gr.Markdown("### Prompt")
+
+                prompt_input = gr.Textbox(
+                    label="Label Prompt", 
+                    lines=2,
+                )
+
+                submit_btn = gr.Button("Submit", variant="primary")
+                
+                examples = gr.Examples(
+                    examples=EXAMPLES,
+                    inputs=[img_input_draw, prompt_input],
+                    label="Click to load example",
+                    examples_per_page=5
+                )
+
+            with gr.Column():
+                gr.Markdown("### SAM3 Result")
+                with gr.Accordion("SAM3 Masks & Boxes", open=False):
+                    sam3_detection_output = gr.Image(label="SAM3 Visualization", height=300)
+                
+                sam3_final_output = gr.AnnotatedImage(label="SAM3 Detections", height=400)
+                # sam3_json_output = gr.JSON(label="SAM3 Output Data")
+
+            with gr.Column():
+                gr.Markdown("### VLM-FO1 Result")
+                with gr.Accordion("Bboxes Proposals", open=False):
+                    upn_detection_output = gr.Image(label="Bboxes", height=300)
+                
+                upn_final_output = gr.AnnotatedImage(label="VLM-FO1 Final", height=400)
+                upn_json_output = gr.JSON(label="VLM-FO1 Details")
+
+        submit_btn.click(
+            update_btn, 
+            inputs=[gr.State(True)], 
+            outputs=[submit_btn], 
+            queue=False
+        ).then(
+            process,
+            inputs=[img_input_draw, prompt_input],
+            outputs=[
+                sam3_final_output, sam3_detection_output,
+                upn_final_output, upn_detection_output, upn_json_output
+            ],
+            queue=True
+        ).then(
+            update_btn, 
+            inputs=[gr.State(False)], 
+            outputs=[submit_btn], 
+            queue=False
+        )
+    
+    return demo
+
+if __name__ == "__main__":
+    import os
+    exit_code = os.system(f"wget -c https://airesources.oss-cn-hangzhou.aliyuncs.com/lp/wheel/sam3.pt")
+
+    model_path = 'omlab/VLM-FO1_Qwen2.5-VL-3B-v01'
+    # sam3_model_path = './resources/sam3/sam3.pt'
+    upn_ckpt_path = "./resources/upn_large.pth"
+
+    # Load FO1
+    tokenizer, model, image_processors = load_pretrained_model(
+        model_path=model_path,
+        device="cuda:0",
+    )
+    
+    # Load SAM3
+    sam3_model = build_sam3_image_model(checkpoint_path='./sam3.pt', device="cuda",bpe_path='/home/user/app/resources/bpe_simple_vocab_16e6.txt.gz')
+    sam3_processor = Sam3Processor(sam3_model, confidence_threshold=0.0, device="cuda")
+    
+    # Load UPN
+    upn_model = UPNWrapper(upn_ckpt_path)
+
+    demo = launch_demo()
+    demo.launch()

detect_tools/upn/__init__.py

@@ -0,0 +1,45 @@
+from . import models
+from .builder import (
+    ARCHITECTURES,
+    BACKBONES,
+    DECODERS,
+    ENCODERS,
+    POS_EMBEDDINGS,
+    build_architecture,
+    build_backbone,
+    build_decoder,
+    build_encoder,
+    build_position_embedding,
+)
+from .inference_wrapper import UPNWrapper
+from .models.architecture import *
+from .models.backbone import *
+from .models.decoder import *
+from .models.encoder import *
+from .models.module import *
+from .models.utils import *
+
+__all__ = [
+    "BACKBONES",
+    "POS_EMBEDDINGS",
+    "ENCODERS",
+    "DECODERS",
+    "ARCHITECTURES",
+    "build_backbone",
+    "build_position_embedding",
+    "build_encoder",
+    "build_decoder",
+    "build_architecture",
+    "UPNWrapper",
+]
+
+__all__ += (
+    models.module.__all__
+    + models.utils.__all__
+    + models.architecture.__all__
+    + models.backbone.__all__
+    + models.encoder.__all__
+    + models.decoder.__all__
+    + models.module.__all__
+    + models.utils.__all__
+)

detect_tools/upn/builder.py

@@ -0,0 +1,39 @@
+from mmengine import Registry, build_from_cfg
+
+BACKBONES = Registry("backbone")
+POS_EMBEDDINGS = Registry("position_embedding")
+FUSERS = Registry("fuser")
+ENCODERS = Registry("encoder")
+DECODERS = Registry("decoder")
+ARCHITECTURES = Registry("architecture")
+
+
+def build_backbone(cfg):
+    """Build encoder."""
+    return build_from_cfg(cfg, BACKBONES)
+
+
+def build_position_embedding(cfg):
+    """Build position embedding."""
+    return build_from_cfg(cfg, POS_EMBEDDINGS)
+
+
+def build_fuser(cfg):
+    """Build fuser."""
+    return build_from_cfg(cfg, FUSERS)
+
+
+def build_encoder(cfg):
+    """Build encoder."""
+    return build_from_cfg(cfg, ENCODERS)
+
+
+def build_decoder(cfg):
+    """Build decoder."""
+    return build_from_cfg(cfg, DECODERS)
+
+
+def build_architecture(cfg):
+    """Build architecture."""
+
+    return build_from_cfg(cfg, ARCHITECTURES)

detect_tools/upn/configs/upn_large.py

@@ -0,0 +1,73 @@
+transformer_cfg = dict(
+    type="DeformableTransformer",
+    num_queries=900,
+    encoder_cfg=dict(
+        type="UPNEncoder",
+        encoder_layer_cfg=dict(
+            type="DeformableTransformerEncoderLayer",
+            activation="relu",
+            d_model=256,
+            dropout=0.0,
+            d_ffn=2048,
+            n_heads=8,
+            n_levels=5,
+        ),
+        d_model=256,
+        num_layers=6,
+        use_checkpoint=False,
+        use_transformer_ckpt=False,
+    ),
+    decoder_cfg=dict(
+        type="UPNDecoder",
+        decoder_layer_cfg=dict(
+            type="DeformableTransformerDecoderLayer",
+            activation="relu",
+            d_model=256,
+            n_heads=8,
+            dropout=0.0,
+            d_ffn=2048,
+            n_levels=5,
+        ),
+        d_model=256,
+        return_intermediate=True,
+        num_layers=6,
+        rm_dec_query_scale=True,
+        use_detached_boxes_dec_out=False,
+    ),
+    learnable_tgt_init=True,
+    random_refpoints_xy=False,
+    num_feature_levels=5,
+    two_stage_bbox_embed_share=False,
+    two_stage_class_embed_share=False,
+    two_stage_keep_all_tokens=False,
+    two_stage_learn_wh=False,
+    two_stage_type="standard",
+    binary_query_selection=False,
+)
+
+vision_backbone = dict(
+    type="SwinWrapper",
+    backbone_cfg="swin_L_384_22k",
+    lr_backbone=1e-05,
+    dilation=False,
+    return_interm_indices=[0, 1, 2, 3],
+    backbone_freeze_keywords=None,
+    backbone_ckpt_path=None,
+    use_checkpoint=False,
+    position_embedding_cfg=dict(
+        type="PositionEmbeddingSineHW",
+        normalize=True,
+        num_pos_feats=128,
+        temperatureH=20,
+        temperatureW=20,
+    ),
+)
+
+model = dict(
+    type="UPN",
+    vision_backbone_cfg=vision_backbone,
+    transformer_cfg=transformer_cfg,
+    num_queries=900,
+    dec_pred_bbox_embed_share=True,
+    dec_pred_class_embed_share=True,
+)

detect_tools/upn/inference_wrapper.py

@@ -0,0 +1,237 @@
+import copy
+import os
+from typing import Dict, List, Union
+
+import numpy as np
+import torch
+from mmengine import Config
+from PIL import Image
+from torchvision.ops import nms
+
+import detect_tools.upn.transforms.transform as T
+from detect_tools.upn import build_architecture
+from detect_tools.upn.models.module import nested_tensor_from_tensor_list
+
+
+def build_model(
+    ckpt_path: str,
+):
+    current_path = os.path.dirname(os.path.abspath(__file__))
+    config_file = f"configs/upn_large.py"
+    config_path = os.path.join(current_path, config_file)
+    model_cfg = Config.fromfile(config_path).model
+    model = build_architecture(model_cfg)
+    checkpoint = torch.load(ckpt_path, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    return model
+
+
+class UPNWrapper:
+    """A wrapper class for the UPN model.
+
+    Args:
+        ckpt_path (str): The path to the model checkpoint
+    """
+
+    def __init__(self, ckpt_path: str):
+
+        self.model = build_model(ckpt_path)
+        self.model.eval()
+        self.model.to("cuda")
+
+    def inference(
+        self,
+        image: List[Union[str, Image.Image]],
+        prompt_type: str = 'fine_grained_prompt',
+    ):
+        """Single image prediction.
+
+        Args:
+            image List[Union[str, Image.Image]]: A list of image path or
+                PIL.Image.Image object.
+            prompt_type (str): The type of prompt to use for the prediction. Choice in
+                ['fine_grained_prompt', 'coarse_grained_prompt'].
+
+        Returns:
+           Dict: Return dict in format:
+                {
+                    "original_xyxy_boxes": (np.ndarray): Original prediction boxes in shape (batch_size, 900, 4),
+                    "scores": (np.ndarray): Score in shape (batch_size, N)
+                }
+        """
+        if not isinstance(image, list):
+            image = [image]
+        input_images, image_sizes = self.construct_input(image)
+        outputs = self._inference(input_images, prompt_type)
+        post_processed_outputs = self.postprocess(outputs, image_sizes)
+        return post_processed_outputs
+
+    def _inference(self, input_images: List[torch.Tensor], prompt_type: str):
+        """Inference for T-Rex2
+
+        Args:
+            input_images (List[torch.Tensor]): Transformed Image
+
+        Retunrs:
+            (Dict): Return dict with keys:
+                - query_features: (torch.Tensor): Query features in shape (batch_size, N, 256)
+                - pred_boxes: (torch.Tensor): Normalized prediction boxes in shape (batch_size, N, 4),
+                    in cxcywh format
+        """
+        input_images = nested_tensor_from_tensor_list(input_images)
+        input_images = input_images.to("cuda")
+        with torch.no_grad():
+            outputs = self.model(input_images, prompt_type)
+        return outputs
+
+    def construct_input(self, image: List[Union[str, Image.Image]]):
+        """Construct input for the model
+
+        Args:
+            image (image: Union[List[Union[str, Image.Image]], torch.Tensor]): A list of image path or
+                PIL.Image.Image object. If the length of the list is more than 1, the model w`ill
+                perform batch inference.
+
+        Returns:
+            Tuple[torch.Tensor, List[List[int]]]: A tuple containing the
+                input images, and the sizes of the input images.
+        """
+        input_images = []
+        image_sizes = []
+        for _, img in enumerate(image):
+            if isinstance(img, str):
+                img = Image.open(img)
+            elif isinstance(img, Image.Image):
+                img = img
+            else:
+                raise ValueError(
+                    "image must be either a string or a PIL.Image.Image object"
+                )
+            W, H = img.size
+            image_sizes.append([H, W])
+            # add image in tensor format
+            input_images.append(self.transform_image(img))
+        return input_images, image_sizes
+
+    def transform_image(self, image_pil: Image) -> Image:
+        """apply a set of transformations to a cv2 load image.
+
+        Args:
+            image_path (str): The path to the image file.
+
+        Returns:
+            Tuple[PIL.Image, torch.Tensor]: A tuple containing the original PIL Image and the
+                transformed image as a PyTorch tensor.
+        """
+        transform = T.Compose(
+            [
+                T.RandomResize([800], max_size=1333),
+                T.ToTensor(),
+                T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+            ]
+        )
+        transformed_image, _ = transform(image_pil, None)  # 3, h, w
+        return transformed_image
+
+    def postprocess(
+        self,
+        outputs: Dict[str, torch.Tensor],
+        image_pil_sizes: List[List[int]] = None,
+    ):
+        boxes = outputs["pred_boxes"].cpu()
+        scores = (
+            outputs["pred_logits"].sigmoid().cpu() if "pred_logits" in outputs else None
+        )
+        normalized_xyxy_boxes = []
+        original_xyxy_boxes = []
+        for batch_idx, (H, W) in enumerate(image_pil_sizes):
+            batch_boxes = boxes[batch_idx]  # (num_queries, 4)
+            # from (cx, cy, w, h) to (x1, y1, x2, y2)
+            batch_boxes[:, 0] = batch_boxes[:, 0] - batch_boxes[:, 2] / 2
+            batch_boxes[:, 1] = batch_boxes[:, 1] - batch_boxes[:, 3] / 2
+            batch_boxes[:, 2] = batch_boxes[:, 0] + batch_boxes[:, 2]
+            batch_boxes[:, 3] = batch_boxes[:, 1] + batch_boxes[:, 3]
+            normalized_xyxy_boxes.append(copy.deepcopy(batch_boxes))
+            #  scale boxes
+            original_boxes = (
+                batch_boxes.clone()
+            )  # Copy the normalized boxes to scale to original sizes
+            original_boxes[:, 0] = original_boxes[:, 0] * W
+            original_boxes[:, 1] = original_boxes[:, 1] * H
+            original_boxes[:, 2] = original_boxes[:, 2] * W
+            original_boxes[:, 3] = original_boxes[:, 3] * H
+            original_xyxy_boxes.append(original_boxes)
+
+        original_xyxy_boxes = torch.stack(original_xyxy_boxes)
+        original_xyxy_boxes = original_xyxy_boxes.numpy()
+
+        # sort everything by score from highest to lowest
+        sorted_original_boxes = []
+        sorted_scores = []
+        for i in range(len(normalized_xyxy_boxes)):
+            scores_i = scores[i] if scores is not None else None
+            # sort in descending order
+            sorted_indices = scores_i.squeeze(-1).argsort(descending=True)
+            sorted_original_boxes.append(original_xyxy_boxes[i][sorted_indices])
+            sorted_scores.append(scores_i[sorted_indices])
+
+        original_xyxy_boxes = np.stack(sorted_original_boxes)
+        scores = torch.stack(sorted_scores)
+
+        return dict(
+            original_xyxy_boxes=original_xyxy_boxes,
+            scores=scores,
+        )
+
+    def filter(self, result: Dict, min_score: float, nms_value: float = 0.8):
+        """Filter the UPN detection result. Only keep boxes with score above min_score
+        and apply Non-Maximum Suppression (NMS) to filter overlapping boxes.
+
+        Args:
+            result (Dict): A dictionary containing detection results with 'original_xyxy_boxes' and 'scores'.
+            min_score (float): Minimum score threshold for keeping a box.
+            nms_value (float): NMS threshold for filtering boxes.
+
+        Returns:
+            Dict: Filtered result containing 'original_xyxy_boxes' and 'scores' with the filtered boxes.
+        """
+        filtered_result = {"original_xyxy_boxes": [], "scores": []}
+
+        for boxes, scores in zip(
+            np.array(result["original_xyxy_boxes"]), result["scores"].numpy()
+        ):
+            # Filter out boxes with score below min_score
+            keep = scores >= min_score
+            boxes = boxes[keep[:, 0]]
+            scores = scores[keep[:, 0]][:, 0]
+
+            if len(boxes) == 0:
+                return filtered_result
+
+            # Convert to torch tensors
+            boxes = torch.tensor(boxes, dtype=torch.float32)
+            scores = torch.tensor(scores, dtype=torch.float32)
+
+            # Apply Non-Maximum Suppression (NMS)
+            if nms_value > 0:
+                keep_indices = nms(boxes, scores, nms_value)
+            else:
+                keep_indices = torch.arange(len(boxes))
+
+            # Keep only the boxes that passed NMS
+            filtered_boxes = boxes[keep_indices].numpy().astype(np.int32)
+            filtered_scores = scores[keep_indices].numpy()
+
+            # Sort the boxes by score in descending order
+            sorted_indices = np.argsort(filtered_scores)[::-1]
+            filtered_boxes = filtered_boxes[sorted_indices]
+            filtered_scores = filtered_scores[sorted_indices]
+
+            # Round the scores to two decimal places
+            filtered_scores = [round(score, 2) for score in filtered_scores]
+
+            # Store the filtered boxes and scores in the result dictionary
+            filtered_result["original_xyxy_boxes"].append(filtered_boxes.tolist())
+            filtered_result["scores"].append(filtered_scores)
+
+        return filtered_result

detect_tools/upn/models/architecture/__init__.py

@@ -0,0 +1,4 @@
+from .deformable_transformer import DeformableTransformer
+from .upn_model import UPN
+
+__all__ = ["UPN", "DeformableTransformer"]

detect_tools/upn/models/architecture/deformable_transformer.py

@@ -0,0 +1,336 @@
+import math
+from typing import Dict, List
+
+import torch
+import torch.nn as nn
+
+from detect_tools.upn import ARCHITECTURES, build_decoder, build_encoder
+from detect_tools.upn.models.utils import (gen_encoder_output_proposals,
+                                      inverse_sigmoid)
+from detect_tools.upn.ops.modules import MSDeformAttn
+
+
+@ARCHITECTURES.register_module()
+class DeformableTransformer(nn.Module):
+    """Implementation of Deformable DETR.
+
+    Args:
+        encoder_cfg (Dict): Config for the TransformerEncoder.
+        decoder_cfg (Dict): Config for the TransformerDecoder.
+        num_queries (int): Number of queries. This is for matching part. Default: 900.
+        d_model (int): Dimension of the model. Default: 256.
+        num_feature_levels (int): Number of feature levels. Default: 1.
+        binary_query_selection (bool): Whether to use binary query selection. Default: False.
+            When using binary query selection, a linear with out channe =1 will be used to select
+            topk proposals. Otherwise, we will use ContrastiveAssign to select topk proposals.
+        learnable_tgt_init (bool): Whether to use learnable target init. Default: True. If False,
+            we will use the topk encoder features as the target init.
+        random_refpoints_xy (bool): Whether to use random refpoints xy. This is only used when
+            two_stage_type is not 'no'. Default: False. If True, we will use random refpoints xy.
+        two_stage_type (str): Type of two stage. Default: 'standard'. Options: 'no', 'standard'
+        two_stage_learn_wh (bool): Whether to learn the width and height of anchor boxes. Default: False.
+        two_stage_keep_all_tokens (bool): If False, the returned hs_enc, ref_enc, init_box_proposal
+            will only be the topk proposals. Otherwise, we will return all the proposals from the
+            encoder. Default: False.
+        two_stage_bbox_embed_share (bool): Whether to share the bbox embedding between the two stages.
+            Default: False.
+        two_stage_class_embed_share (bool): Whether to share the class embedding between the two stages.
+        rm_self_attn_layers (List[int]): The indices of the decoder layers to remove self-attention.
+            Default: None.
+        rm_detach (bool): Whether to detach the decoder output. Default: None.
+        embed_init_tgt (bool): If true, the target embedding is learnable. Otherwise, we will use
+            the topk encoder features as the target init. Default: True.
+    """
+
+    def __init__(
+        self,
+        encoder_cfg: Dict,
+        decoder_cfg: Dict,
+        mask_decoder_cfg: Dict = None,
+        num_queries: int = 900,
+        d_model: int = 256,
+        num_feature_levels: int = 4,
+        binary_query_selection: bool = False,
+        # init query (target)
+        learnable_tgt_init=True,
+        random_refpoints_xy=False,
+        # for two stage
+        two_stage_type: str = "standard",
+        two_stage_learn_wh: bool = False,
+        two_stage_keep_all_tokens: bool = False,
+        two_stage_bbox_embed_share: bool = False,
+        two_stage_class_embed_share: bool = False,
+        # evo of #anchors
+        rm_self_attn_layers: List[int] = None,
+        # for detach
+        rm_detach: bool = None,
+        with_encoder_out: bool = True,
+    ) -> None:
+        super().__init__()
+        self.binary_query_selection = binary_query_selection
+        self.num_queries = num_queries
+        self.num_feature_levels = num_feature_levels
+        self.rm_self_attn_layers = rm_self_attn_layers
+        self.d_model = d_model
+        self.two_stage_bbox_embed_share = two_stage_bbox_embed_share
+        self.two_stage_class_embed_share = two_stage_class_embed_share
+
+        if self.binary_query_selection:
+            self.binary_query_selection_layer = nn.Linear(d_model, 1)
+
+        # build encoder
+        self.encoder = build_encoder(encoder_cfg)
+
+        # build decoder
+        self.decoder = build_decoder(decoder_cfg)
+        self.num_decoder_layers = self.decoder.num_layers
+
+        # build sole mask decoder
+        if mask_decoder_cfg is not None:
+            self.mask_decoder = build_decoder(mask_decoder_cfg)
+        else:
+            self.mask_decoder = None
+        # level embedding
+        if num_feature_levels > 1:
+            self.level_embed = nn.Parameter(torch.Tensor(num_feature_levels, d_model))
+
+        # learnable target embedding
+        self.learnable_tgt_init = learnable_tgt_init
+        assert learnable_tgt_init, "why not learnable_tgt_init"
+
+        self.tgt_embed = nn.Embedding(num_queries, d_model)
+        nn.init.normal_(self.tgt_embed.weight.data)
+
+        # for two stage
+        # TODO: this part is really confusing
+        self.two_stage_type = two_stage_type
+        self.two_stage_learn_wh = two_stage_learn_wh
+        self.two_stage_keep_all_tokens = two_stage_keep_all_tokens
+        assert two_stage_type in [
+            "no",
+            "standard",
+        ], "unknown param {} of two_stage_type".format(two_stage_type)
+        self.with_encoder_out = with_encoder_out
+        if two_stage_type == "standard":
+            # anchor selection at the output of encoder
+            if with_encoder_out:
+                self.enc_output = nn.Linear(d_model, d_model)
+                self.enc_output_norm = nn.LayerNorm(d_model)
+
+            if two_stage_learn_wh:
+                # import ipdb; ipdb.set_trace()
+                self.two_stage_wh_embedding = nn.Embedding(1, 2)
+            else:
+                self.two_stage_wh_embedding = None
+
+        elif two_stage_type == "no":
+            self.init_ref_points(
+                num_queries, random_refpoints_xy
+            )  # init self.refpoint_embed
+
+        self.enc_out_class_embed = None  # this will be initialized outside of the model
+        self.enc_out_bbox_embed = None  # this will be initialized outside of the model
+
+        # remove some self_attn_layers or rm_detach
+        self._reset_parameters()
+
+        self.rm_self_attn_layers = rm_self_attn_layers
+        if rm_self_attn_layers is not None:
+            # assert len(rm_self_attn_layers) == num_decoder_layers
+            print(
+                "Removing the self-attn in {} decoder layers".format(
+                    rm_self_attn_layers
+                )
+            )
+            for lid, dec_layer in enumerate(self.decoder.layers):
+                if lid in rm_self_attn_layers:
+                    dec_layer.rm_self_attn_modules()
+
+        self.rm_detach = rm_detach
+        if self.rm_detach:
+            assert isinstance(rm_detach, list)
+            assert any([i in ["enc_ref", "enc_tgt", "dec"] for i in rm_detach])
+        self.decoder.rm_detach = rm_detach
+
+    def _reset_parameters(self):
+        for p in self.parameters():
+            if p.dim() > 1:
+                nn.init.xavier_uniform_(p)
+        for m in self.modules():
+            if isinstance(m, MSDeformAttn):
+                m._reset_parameters()
+        if self.num_feature_levels > 1 and self.level_embed is not None:
+            nn.init.normal_(self.level_embed)
+
+        if self.two_stage_learn_wh:
+            nn.init.constant_(
+                self.two_stage_wh_embedding.weight, math.log(0.05 / (1 - 0.05))
+            )
+
+    def init_ref_points(self, num_queries: int, random_refpoints_xy: bool = False):
+        """Initialize learnable reference points for each query.
+
+        Args:
+            num_queries (int): number of queries
+            random_refpoints_xy (bool, optional): whether to init the refpoints randomly.
+                Defaults to False.
+        """
+        self.refpoint_embed = nn.Embedding(num_queries, 4)
+        if random_refpoints_xy:
+            self.refpoint_embed.weight.data[:, :2].uniform_(0, 1)
+            self.refpoint_embed.weight.data[:, :2] = inverse_sigmoid(
+                self.refpoint_embed.weight.data[:, :2]
+            )
+            self.refpoint_embed.weight.data[:, :2].requires_grad = False
+
+    def get_valid_ratio(self, mask):
+        _, H, W = mask.shape
+        valid_H = torch.sum(~mask[:, :, 0], 1)
+        valid_W = torch.sum(~mask[:, 0, :], 1)
+        valid_ratio_h = valid_H.float() / H
+        valid_ratio_w = valid_W.float() / W
+        valid_ratio = torch.stack([valid_ratio_w, valid_ratio_h], -1)
+        return valid_ratio
+
+    def forward(
+        self,
+        src_flatten: torch.Tensor,
+        lvl_pos_embed_flatten: torch.Tensor,
+        level_start_index: List[int],
+        spatial_shapes: List[torch.Tensor],
+        valid_ratios: List[torch.Tensor],
+        mask_flatten: torch.Tensor,
+        prompt_type: str,
+    ) -> List[torch.Tensor]:
+        """Forward function."""
+        memory = self.encoder(
+            src_flatten,
+            pos=lvl_pos_embed_flatten,
+            level_start_index=level_start_index,
+            spatial_shapes=spatial_shapes,
+            valid_ratios=valid_ratios,
+            key_padding_mask=mask_flatten,
+        )
+        batch_size = src_flatten.shape[0]
+        crop_region_features = torch.zeros(batch_size, 1, self.d_model).to(
+            memory.device
+        )
+        if prompt_type == "fine_grained_prompt":
+            crop_region_features = (
+                self.fine_grained_prompt.weight[0]
+                .unsqueeze(0)
+                .unsqueeze(0)
+                .repeat(batch_size, 1, 1)
+            )
+        elif prompt_type == "coarse_grained_prompt":
+            crop_region_features = (
+                self.coarse_grained_prompt.weight[0]
+                .unsqueeze(0)
+                .unsqueeze(0)
+                .repeat(batch_size, 1, 1)
+            )
+        pad_mask = torch.ones(batch_size, 1).to(crop_region_features.device).bool()
+        self_attn_mask = torch.ones(batch_size, 1, 1).to(crop_region_features.device)
+        ref_dict = dict(
+            encoded_ref_feature=crop_region_features,
+            pad_mask=pad_mask,
+            self_attn_mask=self_attn_mask,
+            prompt_type="universal_prompt",
+        )
+
+        (
+            refpoint_embed,
+            tgt,
+            init_box_proposal,
+        ) = self.get_two_stage_proposal(memory, mask_flatten, spatial_shapes, ref_dict)
+
+        hs, references = self.decoder(
+            tgt=tgt.transpose(0, 1),
+            tgt_key_padding_mask=None,
+            memory=memory.transpose(0, 1),
+            memory_key_padding_mask=mask_flatten,
+            pos=lvl_pos_embed_flatten.transpose(0, 1),
+            refpoints_unsigmoid=refpoint_embed.transpose(0, 1),
+            level_start_index=level_start_index,
+            spatial_shapes=spatial_shapes,
+            valid_ratios=valid_ratios,
+            tgt_mask=None,
+            # we ~ the mask . False means use the token; True means pad the token
+        )
+        hs_enc = ref_enc = None
+        return (
+            hs,
+            references,
+            ref_dict,
+        )
+
+    def get_two_stage_proposal(
+        self,
+        memory: torch.Tensor,
+        mask_flatten: torch.Tensor,
+        spatial_shapes: List[torch.Tensor],
+        ref_dict: Dict,
+    ) -> List[torch.Tensor]:
+        """Two stage proposal generation for decoder
+
+        Args:
+            memory (torch.Tensor): Image encoded feature. [bs, n, 256]
+            mask_flatten (torch.Tensor): Flattened mask. [bs, n]
+            spatial_shapes (List[torch.Tensor]): Spatial shapes of each feature map. [bs, num_levels, 2]
+            refpoint_embed_dn (torch.Tensor): Denosing refpoint embedding. [bs, num_dn_queries, 256]
+            tgt_dn (torch.Tensor): Denosing target embedding. [bs, num_dn_queries, 256]
+            ref_dict (Dict): A dict containing all kinds of reference image related features.
+        """
+        bs = memory.shape[0]
+        input_hw = None
+        output_memory, output_proposals = gen_encoder_output_proposals(
+            memory, mask_flatten, spatial_shapes, input_hw
+        )
+        output_memory = self.enc_output_norm(self.enc_output(output_memory))
+
+        if self.binary_query_selection:  # Unused
+            topk_logits = self.binary_query_selection_layer(output_memory).squeeze(-1)
+        else:
+            if ref_dict is not None:
+                enc_outputs_class_unselected = self.enc_out_class_embed(
+                    output_memory, ref_dict
+                )  # this is not a linear layer for prediction. But contrastive similaryity, shape [B, len_image, len_text]
+            else:
+                enc_outputs_class_unselected = self.enc_out_class_embed(output_memory)
+            topk_logits = enc_outputs_class_unselected.max(-1)[
+                0
+            ]  # shape [B, len_image]
+        enc_outputs_coord_unselected = (
+            self.enc_out_bbox_embed(output_memory) + output_proposals
+        )  # (bs, \sum{hw}, 4) unsigmoid
+        topk = self.num_queries
+
+        try:
+            topk_proposals = torch.topk(topk_logits, topk, dim=1)[1]  # bs, nq
+        except:
+            raise ValueError(f"dadad {topk_logits.shape}")
+
+        # gather boxes
+        refpoint_embed_undetach = torch.gather(
+            enc_outputs_coord_unselected,
+            1,
+            topk_proposals.unsqueeze(-1).repeat(1, 1, 4),
+        )  # unsigmoid
+        refpoint_embed_ = refpoint_embed_undetach.detach()
+        init_box_proposal = torch.gather(
+            output_proposals, 1, topk_proposals.unsqueeze(-1).repeat(1, 1, 4)
+        ).sigmoid()  # sigmoid
+        # gather tgt
+        tgt_undetach = torch.gather(
+            output_memory, 1, topk_proposals.unsqueeze(-1).repeat(1, 1, self.d_model)
+        )
+        tgt_ = (
+            self.tgt_embed.weight[:, None, :].repeat(1, bs, 1).transpose(0, 1)
+        )  # nq, bs, d_model
+        refpoint_embed, tgt = refpoint_embed_, tgt_
+
+        return (
+            refpoint_embed,
+            tgt,
+            init_box_proposal,
+        )

detect_tools/upn/models/architecture/upn_model.py

@@ -0,0 +1,343 @@
+import copy
+from typing import Dict, List, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from detect_tools.upn import ARCHITECTURES, build_architecture, build_backbone
+from detect_tools.upn.models.module import (MLP, ContrastiveAssign, NestedTensor,
+                                       nested_tensor_from_tensor_list)
+from detect_tools.upn.models.utils import inverse_sigmoid
+
+
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
+
+
+@ARCHITECTURES.register_module()
+class UPN(nn.Module):
+    """Implementation of UPN"""
+
+    def __init__(
+        self,
+        vision_backbone_cfg: Dict,
+        transformer_cfg: Dict,
+        num_queries: int,
+        dec_pred_class_embed_share=True,
+        dec_pred_bbox_embed_share=True,
+        decoder_sa_type="sa",
+    ):
+        super().__init__()
+        # build vision backbone
+        self.backbone = build_backbone(vision_backbone_cfg)
+        # build transformer
+        self.transformer = build_architecture(transformer_cfg)
+
+        self.hidden_dim = self.transformer.d_model
+
+        # for dn training
+        self.num_queries = num_queries
+        self.num_feature_levels = self.transformer.num_feature_levels
+
+        # prepare projection layer for vision feature
+        self.input_proj = self.prepare_vision_feature_projection_layer(
+            self.backbone,
+            self.transformer.num_feature_levels,
+            self.hidden_dim,
+            self.transformer.two_stage_type,
+        )
+        # prepare prediction head
+        self.prepare_prediction_head(
+            dec_pred_class_embed_share,
+            dec_pred_bbox_embed_share,
+            self.hidden_dim,
+            self.transformer.num_decoder_layers,
+        )
+
+        self.decoder_sa_type = decoder_sa_type
+        assert decoder_sa_type in ["sa", "ca_label", "ca_content"]
+        # self.replace_sa_with_double_ca = replace_sa_with_double_ca
+
+        for layer in self.transformer.decoder.layers:
+            layer.label_embedding = None
+        self.label_embedding = None
+
+        # build a unversal token
+        self.transformer.fine_grained_prompt = nn.Embedding(1, self.hidden_dim)
+        self.transformer.coarse_grained_prompt = nn.Embedding(1, self.hidden_dim)
+
+        self._reset_parameters()
+
+    def forward(self, samples: NestedTensor, prompt_type: str = None) -> Dict:
+        """Foward function"""
+        self.device = samples.device
+
+        (
+            src_flatten,
+            lvl_pos_embed_flatten,
+            level_start_index,
+            spatial_shapes,
+            valid_ratios,
+            mask_flatten,
+        ) = self.forward_backbone_encoder(samples)
+
+        (
+            hs,
+            reference,
+            ref_dict,
+        ) = self.transformer(
+            src_flatten,
+            lvl_pos_embed_flatten,
+            level_start_index,
+            spatial_shapes,
+            valid_ratios,
+            mask_flatten,
+            prompt_type,
+        )
+
+        # deformable-detr-line anchor update
+        outputs_coord_list = []
+        outputs_class = []
+
+        for layer_idx, (layer_ref_sig, layer_bbox_embed, layer_hs) in enumerate(
+            zip(reference[:-1], self.bbox_embed, hs)
+        ):
+            layer_delta_unsig = layer_bbox_embed(layer_hs)
+            layer_outputs_unsig = layer_delta_unsig + inverse_sigmoid(layer_ref_sig)
+            layer_outputs_unsig = layer_outputs_unsig.sigmoid()
+            outputs_coord_list.append(layer_outputs_unsig)
+
+        outputs_coord_list = torch.stack(outputs_coord_list)
+
+        if ref_dict is None:
+            # build a mock outputs_class for mask_dn training
+            outputs_class = torch.zeros(
+                outputs_coord_list.shape[0],
+                outputs_coord_list.shape[1],
+                outputs_coord_list.shape[2],
+                self.hidden_dim,
+            )
+        else:
+            outputs_class = torch.stack(
+                [
+                    layer_cls_embed(layer_hs, ref_dict)
+                    for layer_cls_embed, layer_hs in zip(self.class_embed, hs)
+                ]
+            )
+
+        out = {
+            "pred_logits": outputs_class[-1],
+            "pred_boxes": outputs_coord_list[-1],
+        }
+        out["ref_dict"] = ref_dict
+        return out
+
+    def forward_backbone_encoder(self, samples: NestedTensor) -> Tuple:
+        # pass through backbone
+        if isinstance(samples, (list, torch.Tensor)):
+            samples = nested_tensor_from_tensor_list(samples)
+        features, poss = self.backbone(samples)
+        # project features
+        srcs = []
+        masks = []
+        for l, feat in enumerate(features):
+            src, mask = feat.decompose()
+            srcs.append(self.input_proj[l](src))  # downsample the feature map to 256
+            masks.append(mask)
+            assert mask is not None
+
+        if self.num_feature_levels > len(
+            srcs
+        ):  # add more feature levels by downsampling the last feature map
+            _len_srcs = len(srcs)
+            for l in range(_len_srcs, self.num_feature_levels):
+                if l == _len_srcs:
+                    src = self.input_proj[l](features[-1].tensors)
+                else:
+                    src = self.input_proj[l](srcs[-1])
+                m = samples.mask
+                mask = F.interpolate(m[None].float(), size=src.shape[-2:]).to(
+                    torch.bool
+                )[0]
+                pos_l = self.backbone.forward_pos_embed_only(
+                    NestedTensor(src, mask)
+                ).to(src.dtype)
+                srcs.append(src)
+                masks.append(mask)
+                poss.append(pos_l)
+
+        # prepare input for encoder with the following steps:
+        # 1. flatten the feature maps and masks
+        # 2. Add positional embedding and level embedding
+        # 3. Calculate the valid ratio of each feature map based on the mask
+        src_flatten = []
+        mask_flatten = []
+        lvl_pos_embed_flatten = []
+        spatial_shapes = []
+        for lvl, (src, mask, pos_embed) in enumerate(zip(srcs, masks, poss)):
+            bs, c, h, w = src.shape
+            spatial_shape = (h, w)
+            spatial_shapes.append(spatial_shape)
+
+            src = src.flatten(2).transpose(1, 2)  # bs, hw, c
+            mask = mask.flatten(1)  # bs, hw
+            pos_embed = pos_embed.flatten(2).transpose(1, 2)  # bs, hw, c
+            if self.num_feature_levels > 1 and self.transformer.level_embed is not None:
+                lvl_pos_embed = pos_embed + self.transformer.level_embed[lvl].view(
+                    1, 1, -1
+                )
+            else:
+                lvl_pos_embed = pos_embed
+            lvl_pos_embed_flatten.append(lvl_pos_embed)
+            src_flatten.append(src)
+            mask_flatten.append(mask)
+        src_flatten = torch.cat(src_flatten, 1)
+        mask_flatten = torch.cat(mask_flatten, 1)
+        lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)
+        spatial_shapes = torch.as_tensor(
+            spatial_shapes, dtype=torch.long, device=src_flatten.device
+        )
+        level_start_index = torch.cat(
+            (spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1])
+        )
+        valid_ratios = torch.stack(
+            [self.transformer.get_valid_ratio(m) for m in masks], 1
+        )
+
+        return (
+            src_flatten,
+            lvl_pos_embed_flatten,
+            level_start_index,
+            spatial_shapes,
+            valid_ratios,
+            mask_flatten,
+        )
+
+    def prepare_vision_feature_projection_layer(
+        self,
+        backbone: nn.Module,
+        num_feature_levels: int,
+        hidden_dim: int,
+        two_stage_type: str,
+    ) -> nn.ModuleList:
+        """Prepare projection layer to map backbone feature to hidden dim.
+
+        Args:
+            backbone (nn.Module): Backbone.
+            num_feature_levels (int): Number of feature levels.
+            hidden_dim (int): Hidden dim.
+            two_stage_type (str): Type of two stage.
+
+        Returns:
+            nn.ModuleList: Projection layer.
+        """
+        if num_feature_levels > 1:
+            num_backbone_outs = len(backbone.num_channels)
+            input_proj_list = []
+            for _ in range(num_backbone_outs):
+                in_channels = backbone.num_channels[_]
+                input_proj_list.append(
+                    nn.Sequential(
+                        nn.Conv2d(in_channels, hidden_dim, kernel_size=1),
+                        nn.GroupNorm(32, hidden_dim),
+                    )
+                )
+            for _ in range(num_feature_levels - num_backbone_outs):
+                input_proj_list.append(
+                    nn.Sequential(
+                        nn.Conv2d(
+                            in_channels, hidden_dim, kernel_size=3, stride=2, padding=1
+                        ),
+                        nn.GroupNorm(32, hidden_dim),
+                    )
+                )
+                in_channels = hidden_dim
+            input_proj = nn.ModuleList(input_proj_list)
+        else:
+            assert (
+                two_stage_type == "no"
+            ), "two_stage_type should be no if num_feature_levels=1 !!!"
+            input_proj = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        nn.Conv2d(backbone.num_channels[-1], hidden_dim, kernel_size=1),
+                        nn.GroupNorm(32, hidden_dim),
+                    )
+                ]
+            )
+        return input_proj
+
+    def prepare_prediction_head(
+        self,
+        dec_pred_class_embed_share: bool,
+        dec_pred_bbox_embed_share: bool,
+        hidden_dim: int,
+        num_decoder_layers: int,
+    ) -> Union[nn.ModuleList, nn.ModuleList]:
+        """Prepare prediction head. Including class embed and bbox embed.
+
+        Args:
+            dec_pred_class_embed_share (bool): Whether to share class embed for all decoder layers.
+            dec_pred_bbox_embed_share (bool): Whether to share bbox embed for all decoder layers.
+            im (int): Hidden dim.
+            num_decoder_layers (int): Number of decoder layers.
+
+        """
+        _class_embed = ContrastiveAssign()
+        _bbox_embed = MLP(hidden_dim, hidden_dim, 4, 3)
+        nn.init.constant_(_bbox_embed.layers[-1].weight.data, 0)
+        nn.init.constant_(_bbox_embed.layers[-1].bias.data, 0)
+        if dec_pred_bbox_embed_share:
+            box_embed_layerlist = [_bbox_embed for _ in range(num_decoder_layers)]
+        else:
+            box_embed_layerlist = [
+                copy.deepcopy(_bbox_embed) for i in range(num_decoder_layers)
+            ]
+        if dec_pred_class_embed_share:
+            class_embed_layerlist = [_class_embed for i in range(num_decoder_layers)]
+        else:
+            class_embed_layerlist = [
+                copy.deepcopy(_class_embed) for i in range(num_decoder_layers)
+            ]
+        bbox_embed = nn.ModuleList(box_embed_layerlist)
+        class_embed = nn.ModuleList(class_embed_layerlist)
+        self.bbox_embed = bbox_embed
+        self.class_embed = class_embed
+
+        # iniitalize bbox embed and class embed in transformer
+        self.transformer.decoder.bbox_embed = bbox_embed
+        self.transformer.decoder.class_embed = class_embed
+
+        if self.transformer.two_stage_type != "no":
+            if self.transformer.two_stage_bbox_embed_share:
+                assert dec_pred_class_embed_share and dec_pred_bbox_embed_share
+                self.transformer.enc_out_bbox_embed = _bbox_embed
+            else:
+                self.transformer.enc_out_bbox_embed = copy.deepcopy(_bbox_embed)
+
+            if self.transformer.two_stage_class_embed_share:
+                assert dec_pred_class_embed_share and dec_pred_bbox_embed_share
+                self.transformer.enc_out_class_embed = _class_embed
+            else:
+                self.transformer.enc_out_class_embed = copy.deepcopy(_class_embed)
+
+            self.refpoint_embed = None
+
+    def _reset_parameters(self):
+        # init input_proj
+        for proj in self.input_proj:
+
+            nn.init.xavier_uniform_(proj[0].weight, gain=1)
+            nn.init.constant_(proj[0].bias, 0)

detect_tools/upn/models/backbone/__init__.py

@@ -0,0 +1,4 @@
+from .swin import SwinTransformer
+from .wrapper import SwinWrapper
+
+__all__ = ["SwinWrapper", "SwinTransformer"]

detect_tools/upn/models/backbone/swin.py

@@ -0,0 +1,814 @@
+from typing import Dict, List
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+from detect_tools.upn import BACKBONES
+from detect_tools.upn.models.module import NestedTensor
+
+
+class Mlp(nn.Module):
+    """Multilayer perceptron."""
+
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        drop=0.0,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        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
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = (
+        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    )
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(
+        B, H // window_size, W // window_size, window_size, window_size, -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(
+        self,
+        dim,
+        window_size,
+        num_heads,
+        qkv_bias=True,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+    ):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
+        )  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = (
+            coords_flatten[:, :, None] - coords_flatten[:, None, :]
+        )  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(
+            1, 2, 0
+        ).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=0.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """Forward function.
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = (
+            self.qkv(x)
+            .reshape(B_, N, 3, self.num_heads, C // self.num_heads)
+            .permute(2, 0, 3, 1, 4)
+        )
+        q, k, v = (
+            qkv[0],
+            qkv[1],
+            qkv[2],
+        )  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = q @ k.transpose(-2, -1)
+
+        relative_position_bias = self.relative_position_bias_table[
+            self.relative_position_index.view(-1)
+        ].view(
+            self.window_size[0] * self.window_size[1],
+            self.window_size[0] * self.window_size[1],
+            -1,
+        )  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(
+            2, 0, 1
+        ).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
+                1
+            ).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class SwinTransformerBlock(nn.Module):
+    """Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        window_size=7,
+        shift_size=0,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        act_layer=nn.GELU,
+        norm_layer=nn.LayerNorm,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert (
+            0 <= self.shift_size < self.window_size
+        ), "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim,
+            window_size=to_2tuple(self.window_size),
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            attn_drop=attn_drop,
+            proj_drop=drop,
+        )
+
+        self.drop_path = 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 = Mlp(
+            in_features=dim,
+            hidden_features=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=drop,
+        )
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(
+                x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
+            )
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(
+            shifted_x, self.window_size
+        )  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(
+            -1, self.window_size * self.window_size, C
+        )  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(
+            x_windows, mask=attn_mask
+        )  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(
+                shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
+            )
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """Patch Merging Layer
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(
+        self,
+        dim,
+        depth,
+        num_heads,
+        window_size=7,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        norm_layer=nn.LayerNorm,
+        downsample=None,
+        use_checkpoint=False,
+    ):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList(
+            [
+                SwinTransformerBlock(
+                    dim=dim,
+                    num_heads=num_heads,
+                    window_size=window_size,
+                    shift_size=0 if (i % 2 == 0) else window_size // 2,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop,
+                    attn_drop=attn_drop,
+                    drop_path=(
+                        drop_path[i] if isinstance(drop_path, list) else drop_path
+                    ),
+                    norm_layer=norm_layer,
+                )
+                for i in range(depth)
+            ]
+        )
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """Forward function.
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        Hp = int(np.ceil(H / self.window_size)) * self.window_size
+        Wp = int(np.ceil(W / self.window_size)) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (
+            slice(0, -self.window_size),
+            slice(-self.window_size, -self.shift_size),
+            slice(-self.shift_size, None),
+        )
+        w_slices = (
+            slice(0, -self.window_size),
+            slice(-self.window_size, -self.shift_size),
+            slice(-self.shift_size, None),
+        )
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(
+            img_mask, self.window_size
+        )  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
+            attn_mask == 0, float(0.0)
+        )
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """Image to Patch Embedding
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size
+        )
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+@BACKBONES.register_module()
+class SwinTransformer(nn.Module):
+    """Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        dilation (bool): if True, the output size if 16x downsample, ow 32x downsample.
+    """
+
+    def __init__(
+        self,
+        pretrain_img_size=224,
+        patch_size=4,
+        in_chans=3,
+        embed_dim=96,
+        depths=[2, 2, 6, 2],
+        num_heads=[3, 6, 12, 24],
+        window_size=7,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop_rate=0.0,
+        attn_drop_rate=0.0,
+        drop_path_rate=0.2,
+        norm_layer=nn.LayerNorm,
+        ape=False,
+        patch_norm=True,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=-1,
+        dilation=False,
+        use_checkpoint=False,
+    ):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+        self.dilation = dilation
+
+        if use_checkpoint:
+            print("use_checkpoint!!!!!!!!!!!!!!!!!!!!!!!!")
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None,
+        )
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [
+                pretrain_img_size[0] // patch_size[0],
+                pretrain_img_size[1] // patch_size[1],
+            ]
+
+            self.absolute_pos_embed = nn.Parameter(
+                torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1])
+            )
+            trunc_normal_(self.absolute_pos_embed, std=0.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [
+            x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))
+        ]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        # prepare downsample list
+        downsamplelist = [PatchMerging for i in range(self.num_layers)]
+        downsamplelist[-1] = None
+        num_features = [int(embed_dim * 2**i) for i in range(self.num_layers)]
+        if self.dilation:
+            downsamplelist[-2] = None
+            num_features[-1] = int(embed_dim * 2 ** (self.num_layers - 1)) // 2
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(
+                # dim=int(embed_dim * 2 ** i_layer),
+                dim=num_features[i_layer],
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])],
+                norm_layer=norm_layer,
+                # downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                downsample=downsamplelist[i_layer],
+                use_checkpoint=use_checkpoint,
+            )
+            self.layers.append(layer)
+
+        # num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f"norm{i_layer}"
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+    def forward_raw(self, x: torch.Tensor) -> List[torch.Tensor]:
+        """Forward function."""
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(
+                self.absolute_pos_embed, size=(Wh, Ww), mode="bicubic"
+            )
+            x = (x + absolute_pos_embed).flatten(2).transpose(1, 2)  # B Wh*Ww C
+        else:
+            x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+
+        outs = []
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+            # import ipdb; ipdb.set_trace()
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f"norm{i}")
+                x_out = norm_layer(x_out)
+
+                out = (
+                    x_out.view(-1, H, W, self.num_features[i])
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+                outs.append(out)
+
+        return tuple(outs)
+
+    def forward(self, tensor_list: NestedTensor) -> Dict:
+        """Forward function.
+
+        Args:
+            tensor_list (NestedTensor): NestedTensor object containing tensors and masks.
+
+        Returns:
+            Dict: Dict containing output tensors. The structure is as follows.
+                - 0: NestedTensor from stage 0.
+                - 1: NestedTensor from stage 1.
+                - 2: NestedTensor from stage 2.
+                - 3: NestedTensor from stage 3.
+        """
+        x = tensor_list.tensors
+
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(
+                self.absolute_pos_embed, size=(Wh, Ww), mode="bicubic"
+            )
+            x = (x + absolute_pos_embed).flatten(2).transpose(1, 2)  # B Wh*Ww C
+        else:
+            x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+
+        outs = []
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f"norm{i}")
+                x_out = norm_layer(x_out)
+
+                out = (
+                    x_out.view(-1, H, W, self.num_features[i])
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+                outs.append(out)
+
+        # collect for nesttensors
+        outs_dict = {}
+        for idx, out_i in enumerate(outs):
+            m = tensor_list.mask
+            assert m is not None
+            mask = F.interpolate(m[None].float(), size=out_i.shape[-2:]).to(torch.bool)[
+                0
+            ]
+            outs_dict[idx] = NestedTensor(out_i, mask)
+
+        return outs_dict
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep layers freezed."""
+        super(SwinTransformer, self).train(mode)
+        self._freeze_stages()
+
+
+def build_swin_transformer(modelname, pretrain_img_size, **kw):
+    assert modelname in [
+        "swin_T_224_1k",
+        "swin_B_224_22k",
+        "swin_B_384_22k",
+        "swin_L_224_22k",
+        "swin_L_384_22k",
+    ]
+
+    model_para_dict = {
+        "swin_T_224_1k": dict(
+            embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7
+        ),
+        "swin_B_224_22k": dict(
+            embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=7
+        ),
+        "swin_B_384_22k": dict(
+            embed_dim=128,
+            depths=[2, 2, 18, 2],
+            num_heads=[4, 8, 16, 32],
+            window_size=12,
+        ),
+        "swin_L_224_22k": dict(
+            embed_dim=192,
+            depths=[2, 2, 18, 2],
+            num_heads=[6, 12, 24, 48],
+            window_size=7,
+        ),
+        "swin_L_384_22k": dict(
+            embed_dim=192,
+            depths=[2, 2, 18, 2],
+            num_heads=[6, 12, 24, 48],
+            window_size=12,
+        ),
+    }
+    kw_cgf = model_para_dict[modelname]
+    kw_cgf.update(kw)
+    model = SwinTransformer(pretrain_img_size=pretrain_img_size, **kw_cgf)
+    return model

detect_tools/upn/models/backbone/wrapper.py

@@ -0,0 +1,297 @@
+from typing import Dict, List, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from detect_tools.upn import BACKBONES, build_backbone, build_position_embedding
+from detect_tools.upn.models.module import NestedTensor
+from detect_tools.upn.models.utils import clean_state_dict
+
+
+class FrozenBatchNorm2d(torch.nn.Module):
+    """
+    BatchNorm2d where the batch statistics and the affine parameters are fixed.
+
+    Copy-paste from torchvision.misc.ops with added eps before rqsrt,
+    without which any other models than torchvision.models.resnet[18,34,50,101]
+    produce nans.
+    """
+
+    def __init__(self, n):
+        super(FrozenBatchNorm2d, self).__init__()
+        self.register_buffer("weight", torch.ones(n))
+        self.register_buffer("bias", torch.zeros(n))
+        self.register_buffer("running_mean", torch.zeros(n))
+        self.register_buffer("running_var", torch.ones(n))
+
+    def _load_from_state_dict(
+        self,
+        state_dict,
+        prefix,
+        local_metadata,
+        strict,
+        missing_keys,
+        unexpected_keys,
+        error_msgs,
+    ):
+        num_batches_tracked_key = prefix + "num_batches_tracked"
+        if num_batches_tracked_key in state_dict:
+            del state_dict[num_batches_tracked_key]
+
+        super(FrozenBatchNorm2d, self)._load_from_state_dict(
+            state_dict,
+            prefix,
+            local_metadata,
+            strict,
+            missing_keys,
+            unexpected_keys,
+            error_msgs,
+        )
+
+    def forward(self, x):
+        # move reshapes to the beginning
+        # to make it fuser-friendly
+        w = self.weight.reshape(1, -1, 1, 1)
+        b = self.bias.reshape(1, -1, 1, 1)
+        rv = self.running_var.reshape(1, -1, 1, 1)
+        rm = self.running_mean.reshape(1, -1, 1, 1)
+        eps = 1e-5
+        scale = w * (rv + eps).rsqrt()
+        bias = b - rm * scale
+        return x * scale + bias
+
+
+class Joiner(nn.Module):
+    """A wrapper for the backbone and the position embedding.
+
+    Args:
+        backbone_cfg (Dict): Config dict to build backbone.
+        position_embedding_cfg (Dict): Config dict to build position embedding.
+    """
+
+    def __init__(self, backbone: nn.Module, position_embedding: nn.Module) -> None:
+        super().__init__()
+        self.backbone = backbone
+        self.pos_embed = position_embedding
+
+    def forward(
+        self, tensor_list: NestedTensor
+    ) -> Union[List[NestedTensor], List[torch.Tensor]]:
+        """Forward function.
+
+        Args:
+            tensor_list (NestedTensor): NestedTensor wrapping the input tensor.
+
+        Returns:
+            [List[NestedTensor]: A list of feature map in NestedTensor format.
+            List[torch.Tensor]: A list of position encoding.
+        """
+
+        xs = self.backbone(tensor_list)
+        out: List[NestedTensor] = []
+        pos = []
+        for layer_idx, x in xs.items():
+            out.append(x)
+            # position encoding
+            pos.append(self.pos_embed(x).to(x.tensors.dtype))
+
+        return out, pos
+
+    def forward_pos_embed_only(self, x: NestedTensor) -> torch.Tensor:
+        """Forward function for position embedding only. This is used to generate additional layer
+
+        Args:
+            x (NestedTensor): NestedTensor wrapping the input tensor.
+
+        Returns:
+            [List[torch.Tensor]: A list of position encoding.
+        """
+        return self.pos_embed(x)
+
+
+@BACKBONES.register_module()
+class SwinWrapper(nn.Module):
+    """A wrapper for swin transformer.
+
+    Args:
+        backbone_cfg Union[Dict, str]: Config dict to build backbone. If given a str name, we
+            will call `get_swin_config` to get the config dict.
+        dilation (bool): Whether to use dilation in stage 4.
+        position_embedding_cfg (Dict): Config dict to build position embedding.
+        lr_backbone (float): Learning rate of the backbone.
+        return_interm_layers (List[int]): Which layers to return.
+        backbone_freeze_keywords (List[str]): List of keywords to freeze the backbone.
+        use_checkpoint (bool): Whether to use checkpoint. Default: False.
+        ckpt_path (str): Checkpoint path. Default: None.
+        use_pretrained_ckpt (bool): Whether to use pretrained checkpoint. Default: True.
+    """
+
+    def __init__(
+        self,
+        backbone_cfg: Union[Dict, str],
+        dilation: bool,
+        position_embedding_cfg: Dict,
+        lr_backbone: float,
+        return_interm_indices: List[int],
+        backbone_freeze_keywords: List[str],
+        use_checkpoint: bool = False,
+        backbone_ckpt_path: str = None,
+    ) -> None:
+        super(SwinWrapper, self).__init__()
+        pos_embedding = build_position_embedding(position_embedding_cfg)
+        train_backbone = lr_backbone > 0
+        if not train_backbone:
+            raise ValueError("Please set lr_backbone > 0")
+        assert return_interm_indices in [[0, 1, 2, 3], [1, 2, 3], [3]]
+
+        # build backbone
+        if isinstance(backbone_cfg, str):
+            assert (
+                backbone_cfg
+                in backbone_cfg
+                in [
+                    "swin_T_224_1k",
+                    "swin_B_224_22k",
+                    "swin_B_384_22k",
+                    "swin_L_224_22k",
+                    "swin_L_384_22k",
+                ]
+            )
+            pretrain_img_size = int(backbone_cfg.split("_")[-2])
+            backbone_cfg = get_swin_config(
+                backbone_cfg,
+                pretrain_img_size,
+                out_indices=tuple(return_interm_indices),
+                dilation=dilation,
+                use_checkpoint=use_checkpoint,
+            )
+        backbone = build_backbone(backbone_cfg)
+
+        # freeze some layers
+        if backbone_freeze_keywords is not None:
+            for name, parameter in backbone.named_parameters():
+                for keyword in backbone_freeze_keywords:
+                    if keyword in name:
+                        parameter.requires_grad_(False)
+                        break
+
+        # load checkpoint
+        if backbone_ckpt_path is not None:
+            print("Loading backbone checkpoint from {}".format(backbone_ckpt_path))
+            checkpoint = torch.load(backbone_ckpt_path, map_location="cpu")["model"]
+            from collections import OrderedDict
+
+            def key_select_function(keyname):
+                if "head" in keyname:
+                    return False
+                if dilation and "layers.3" in keyname:
+                    return False
+                return True
+
+            _tmp_st = OrderedDict(
+                {
+                    k: v
+                    for k, v in clean_state_dict(checkpoint).items()
+                    if key_select_function(k)
+                }
+            )
+            _tmp_st_output = backbone.load_state_dict(_tmp_st, strict=False)
+            print(str(_tmp_st_output))
+
+        bb_num_channels = backbone.num_features[4 - len(return_interm_indices) :]
+        assert len(bb_num_channels) == len(
+            return_interm_indices
+        ), f"len(bb_num_channels) {len(bb_num_channels)} != len(return_interm_indices) {len(return_interm_indices)}"
+
+        model = Joiner(backbone, pos_embedding)
+        model.num_channels = bb_num_channels
+        self.num_channels = bb_num_channels
+        self.model = model
+
+    def forward(
+        self, tensor_list: NestedTensor
+    ) -> Union[List[NestedTensor], List[torch.Tensor]]:
+        """Forward function.
+
+        Args:
+            tensor_list (NestedTensor): NestedTensor wrapping the input tensor.
+
+        Returns:
+            [List[NestedTensor]: A list of feature map in NestedTensor format.
+            List[torch.Tensor]: A list of position encoding.
+        """
+
+        return self.model(tensor_list)
+
+    def forward_pos_embed_only(self, tensor_list: NestedTensor) -> torch.Tensor:
+        """Forward function to get position embedding only.
+
+        Args:
+            tensor_list (NestedTensor): NestedTensor wrapping the input tensor.
+
+        Returns:
+            torch.Tensor: Position embedding.
+        """
+        return self.model.forward_pos_embed_only(tensor_list)
+
+
+def get_swin_config(modelname: str, pretrain_img_size: Tuple[int, int], **kw):
+    """Get swin config dict.
+
+    Args:
+        modelname (str): Name of the model.
+        pretrain_img_size (Tuple[int, int]): Image size of the pretrain model.
+        kw (Dict): Other key word arguments.
+
+    Returns:
+        Dict: Config dict.
+        str: Path to the pretrained checkpoint.
+    """
+    assert modelname in [
+        "swin_T_224_1k",
+        "swin_B_224_22k",
+        "swin_B_384_22k",
+        "swin_L_224_22k",
+        "swin_L_384_22k",
+    ]
+    model_para_dict = {
+        "swin_T_224_1k": dict(
+            type="SwinTransformer",
+            embed_dim=96,
+            depths=[2, 2, 6, 2],
+            num_heads=[3, 6, 12, 24],
+            window_size=7,
+        ),
+        "swin_B_224_22k": dict(
+            type="SwinTransformer",
+            embed_dim=128,
+            depths=[2, 2, 18, 2],
+            num_heads=[4, 8, 16, 32],
+            window_size=7,
+        ),
+        "swin_B_384_22k": dict(
+            type="SwinTransformer",
+            embed_dim=128,
+            depths=[2, 2, 18, 2],
+            num_heads=[4, 8, 16, 32],
+            window_size=12,
+        ),
+        "swin_L_224_22k": dict(
+            type="SwinTransformer",
+            embed_dim=192,
+            depths=[2, 2, 18, 2],
+            num_heads=[6, 12, 24, 48],
+            window_size=7,
+        ),
+        "swin_L_384_22k": dict(
+            type="SwinTransformer",
+            embed_dim=192,
+            depths=[2, 2, 18, 2],
+            num_heads=[6, 12, 24, 48],
+            window_size=12,
+        ),
+    }
+    kw_cgf = model_para_dict[modelname]
+    kw_cgf.update(kw)
+    kw_cgf.update(dict(pretrain_img_size=pretrain_img_size))
+    return kw_cgf

detect_tools/upn/models/decoder/__init__.py

@@ -0,0 +1,3 @@
+from .upn_decoder import UPNDecoder, DeformableTransformerDecoderLayer
+
+__all__ = ["UPNDecoder", "DeformableTransformerDecoderLayer"]

detect_tools/upn/models/decoder/upn_decoder.py

@@ -0,0 +1,378 @@
+from typing import Dict
+
+import torch
+import torch.nn as nn
+
+from detect_tools.upn import DECODERS, build_decoder
+from detect_tools.upn.models.module import MLP
+from detect_tools.upn.models.utils import (gen_sineembed_for_position,
+                                      get_activation_fn, get_clones,
+                                      inverse_sigmoid)
+from detect_tools.upn.ops.modules import MSDeformAttn
+
+
+@DECODERS.register_module()
+class DeformableTransformerDecoderLayer(nn.Module):
+    """Deformable Transformer Decoder Layer. This is a modified version in Grounding DINO.
+    After the query is attented to the image feature, it is further attented to the text feature.
+    The execute order is: self_attn -> cross_attn to text -> cross_attn to image -> ffn
+    Args:
+        d_model (int): The dimension of keys/values/queries in :class:`MultiheadAttention`.
+        d_ffn (int): The dimension of the feedforward network model.
+        dropout (float): Probability of an element to be zeroed.
+        activation (str): Activation function in the feedforward network.
+            'relu' and 'gelu' are supported.
+        n_levels (int): The number of levels in Multi-Scale Deformable Attention.
+        n_heads (int): Parallel attention heads.
+        n_points (int): Number of sampling points in Multi-Scale Deformable Attention.
+        ffn_extra_layernorm (bool): If True, add an extra layernorm after ffn.
+    """
+
+    def __init__(
+        self,
+        d_model: int = 256,
+        d_ffn: int = 1024,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        n_levels: int = 4,
+        n_heads: int = 8,
+        n_points: int = 4,
+        ffn_extra_layernorm: bool = False,
+    ) -> None:
+        super().__init__()
+
+        # cross attention for visual features
+        self.cross_attn = MSDeformAttn(d_model, n_levels, n_heads, n_points)
+        self.dropout1 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm1 = nn.LayerNorm(d_model)
+
+        # self attention for query
+        self.self_attn = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)
+        self.dropout2 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm2 = nn.LayerNorm(d_model)
+
+        # ffn
+        self.linear1 = nn.Linear(d_model, d_ffn)
+        self.activation = get_activation_fn(activation, d_model=d_ffn, batch_dim=1)
+        self.dropout3 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.linear2 = nn.Linear(d_ffn, d_model)
+        self.dropout4 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm3 = nn.LayerNorm(d_model)
+        if ffn_extra_layernorm:
+            raise NotImplementedError("ffn_extra_layernorm not implemented")
+            self.norm_ext = nn.LayerNorm(d_ffn)
+        else:
+            self.norm_ext = None
+
+        self.key_aware_proj = None
+
+    def rm_self_attn_modules(self):
+        self.self_attn = None
+        self.dropout2 = None
+        self.norm2 = None
+
+    @staticmethod
+    def with_pos_embed(tensor, pos):
+        return tensor if pos is None else tensor + pos
+
+    def forward_ffn(self, tgt):
+        tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
+
+        tgt = tgt + self.dropout4(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        tgt_query_pos: torch.Tensor = None,
+        tgt_reference_points: torch.Tensor = None,
+        memory: torch.Tensor = None,
+        memory_key_padding_mask: torch.Tensor = None,
+        memory_level_start_index: torch.Tensor = None,
+        memory_spatial_shapes: torch.Tensor = None,
+        self_attn_mask: torch.Tensor = None,
+        cross_attn_mask: torch.Tensor = None,
+    ) -> torch.Tensor:
+        """Forward function
+
+        Args:
+            tgt (torch.Tensor): Input target in shape (B, T, C)
+            tgt_query_pos (torch.Tensor): Positional encoding of the query.
+            tgt_query_sine_embed (torch.Tensor): Sine positional encoding of the query. Unused.
+            tgt_key_padding_mask (torch.Tensor): Mask for target feature in shape (B, T).
+            tgt_reference_points (torch.Tensor): Reference points for the query in shape (B, T, 4).
+            memory_text (torch.Tensor): Input text embeddings in shape (B, num_token, C).
+            text_attention_mask (torch.Tensor): Attention mask for text embeddings in shape
+                (B, num_token).
+            memory (torch.Tensor): Input image feature in shape (B, HW, C)
+            memory_key_padding_mask (torch.Tensor): Mask for image feature in shape (B, HW)
+            memory_level_start_index (torch.Tensor): Starting index of each level in memory.
+            memory_spatial_shapes (torch.Tensor): Spatial shape of each level in memory.
+            memory_pos (torch.Tensor): Positional encoding of memory. Unused.
+            self_attn_mask (torch.Tensor): Mask used for self-attention.
+            cross_attn_mask (torch.Tensor): Mask used for cross-attention.
+
+        Returns:
+            torch.Tensor: Output tensor in shape (B, T, C)
+        """
+        assert cross_attn_mask is None
+
+        # self attention
+        if self.self_attn is not None:
+            q = k = self.with_pos_embed(tgt, tgt_query_pos)
+            tgt2 = self.self_attn(q, k, tgt, attn_mask=self_attn_mask)[0]
+            tgt = tgt + self.dropout2(tgt2)
+            tgt = self.norm2(tgt)
+
+        # attend to image features
+        tgt2 = self.cross_attn(
+            self.with_pos_embed(tgt, tgt_query_pos).transpose(0, 1),
+            tgt_reference_points.transpose(0, 1).contiguous(),
+            memory.transpose(0, 1),
+            memory_spatial_shapes,
+            memory_level_start_index,
+            memory_key_padding_mask,
+        ).transpose(0, 1)
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+        # ffn
+        tgt = self.forward_ffn(tgt)
+
+        return tgt
+
+
+@DECODERS.register_module()
+class UPNDecoder(nn.Module):
+    """Decoder used in UPN. Each layer is a DeformableTransformerDecoderLayer. The query
+    will be abled to attend the image feature and text feature. The execute order is:
+    self_attn -> cross_attn to image -> ffn
+
+    Args:
+        decoder_layer_cfg (Dict): Config for the DeformableTransformerDecoderLayer.
+        num_layers (int): number of layers
+        norm (nn.Module, optional): normalization layer. Defaults to None.
+        return_intermediate (bool, optional): whether return intermediate results.
+            Defaults to False.
+        d_model (int, optional): dimension of the model. Defaults to 256.
+        query_dim (int, optional): dimension of the query. Defaults to 4.
+        modulate_hw_attn (bool, optional): whether modulate the attention weights
+            by the height and width of the image feature. Defaults to False.
+        num_feature_levels (int, optional): number of feature levels. Defaults to 1.
+        deformable_decoder (bool, optional): whether use deformable decoder. Defaults to False.
+        decoder_query_perturber ([type], optional): [description]. Defaults to None.
+        dec_layer_number ([type], optional): [description]. Defaults to None.
+        rm_dec_query_scale (bool, optional): [description]. Defaults to False.
+        dec_layer_share (bool, optional): [description]. Defaults to False.
+        dec_layer_dropout_prob ([type], optional): [description]. Defaults to None.
+    """
+
+    def __init__(
+        self,
+        decoder_layer_cfg: Dict,
+        num_layers: int,
+        norm: str = "layernorm",
+        return_intermediate: bool = True,
+        d_model: int = 256,
+        query_dim: int = 4,
+        modulate_hw_attn: bool = False,
+        num_feature_levels: int = 1,
+        deformable_decoder: bool = True,
+        decoder_query_perturber=None,
+        dec_layer_number=None,
+        rm_dec_query_scale: bool = True,
+        dec_layer_share: bool = False,
+        dec_layer_dropout_prob=None,
+        use_detached_boxes_dec_out: bool = False,
+    ):
+        super().__init__()
+
+        decoder_layer = build_decoder(decoder_layer_cfg)
+        if num_layers > 0:
+            self.layers = get_clones(
+                decoder_layer, num_layers, layer_share=dec_layer_share
+            )
+        else:
+            self.layers = []
+        self.num_layers = num_layers
+        if norm == "layernorm":
+            self.norm = nn.LayerNorm(d_model)
+        self.return_intermediate = return_intermediate
+        self.query_dim = query_dim
+        assert query_dim in [2, 4], "query_dim should be 2/4 but {}".format(query_dim)
+        self.num_feature_levels = num_feature_levels
+        self.use_detached_boxes_dec_out = use_detached_boxes_dec_out
+
+        self.ref_point_head = MLP(query_dim // 2 * d_model, d_model, d_model, 2)
+        self.ref_point_head_point = MLP(
+            d_model, d_model, d_model, 2
+        )  # for point reference only
+        if not deformable_decoder:
+            self.query_pos_sine_scale = MLP(d_model, d_model, d_model, 2)
+        else:
+            self.query_pos_sine_scale = None
+
+        if rm_dec_query_scale:
+            self.query_scale = None
+        else:
+            raise NotImplementedError
+            self.query_scale = MLP(d_model, d_model, d_model, 2)
+        self.bbox_embed = None
+        self.class_embed = None
+
+        self.d_model = d_model
+        self.modulate_hw_attn = modulate_hw_attn
+        self.deformable_decoder = deformable_decoder
+
+        if not deformable_decoder and modulate_hw_attn:
+            self.ref_anchor_head = MLP(d_model, d_model, 2, 2)
+        else:
+            self.ref_anchor_head = None
+
+        self.decoder_query_perturber = decoder_query_perturber
+        self.box_pred_damping = None
+
+        self.dec_layer_number = dec_layer_number
+        if dec_layer_number is not None:
+            assert isinstance(dec_layer_number, list)
+            assert len(dec_layer_number) == num_layers
+
+        self.dec_layer_dropout_prob = dec_layer_dropout_prob
+        if dec_layer_dropout_prob is not None:
+            assert isinstance(dec_layer_dropout_prob, list)
+            assert len(dec_layer_dropout_prob) == num_layers
+            for i in dec_layer_dropout_prob:
+                assert 0.0 <= i <= 1.0
+
+        self.rm_detach = None
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: torch.Tensor = None,
+        memory_mask: torch.Tensor = None,
+        tgt_key_padding_mask: torch.Tensor = None,
+        memory_key_padding_mask: torch.Tensor = None,
+        pos: torch.Tensor = None,
+        refpoints_unsigmoid: torch.Tensor = None,
+        level_start_index: torch.Tensor = None,
+        spatial_shapes: torch.Tensor = None,
+        valid_ratios: torch.Tensor = None,
+        memory_ref_image: torch.Tensor = None,
+        refImg_padding_mask: torch.Tensor = None,
+        memory_visual_prompt: torch.Tensor = None,
+    ):
+        """Forward function.
+
+        Args:
+            tgt (torch.Tensor): target feature, [bs, num_queries, d_model]
+            memory (torch.Tensor): Image feature, [bs, hw, d_model]
+            tgt_mask (torch.Tensor, optional): target mask for attention. Defaults to None.
+            memory_mask (torch.Tensor, optional): image mask for attention. Defaults to None.
+            tgt_key_padding_mask (torch.Tensor, optional): target mask for padding. Defaults to None.
+            memory_key_padding_mask (torch.Tensor, optional): image mask for padding. Defaults to None.
+            pos (torch.Tensor, optional): query position embedding
+            refpoints_unsigmoid (torch.Tensor, optional): reference points. Defaults to None.
+            level_start_index (torch.Tensor, optional): start index of each level. Defaults to None.
+            spatial_shapes (torch.Tensor, optional): spatial shape of each level. Defaults to None.
+            valid_ratios (torch.Tensor, optional): valid ratio of each level. Defaults to None.
+            memory_ref_image (torch.Tensor, optional): reference image feature, [bs, num_ref, d_model]. Defaults to None.
+            refImg_padding_mask (torch.Tensor, optional): padding mask for attention. Defaults to None.
+        """
+        output = tgt
+
+        intermediate = []
+        reference_points = refpoints_unsigmoid.sigmoid()
+        ref_points = [reference_points]
+
+        for layer_id, layer in enumerate(self.layers):
+
+            if reference_points.shape[-1] == 4:
+                reference_points_input = (
+                    reference_points[:, :, None]
+                    * torch.cat([valid_ratios, valid_ratios], -1)[None, :]
+                )  # nq, bs, nlevel, 4
+            else:
+                assert reference_points.shape[-1] == 2
+                reference_points_input = (
+                    reference_points[:, :, None] * valid_ratios[None, :]
+                )
+            query_sine_embed = gen_sineembed_for_position(
+                reference_points_input[:, :, 0, :]
+            )  # nq, bs, 256*2
+
+            # conditional query
+            if query_sine_embed.shape[-1] == 512:
+                raw_query_pos = (
+                    self.ref_point_head(query_sine_embed)
+                    + self.ref_point_head_point(
+                        torch.zeros_like(query_sine_embed)[:, :, :256]
+                    )
+                    * 0.0
+                )
+            else:
+                raw_query_pos = (
+                    self.ref_point_head_point(query_sine_embed)
+                    + self.ref_point_head(
+                        torch.zeros(
+                            query_sine_embed.shape[0],
+                            query_sine_embed.shape[1],
+                            512,
+                            device=query_sine_embed.device,
+                        )
+                    )
+                    * 0.0
+                )
+            pos_scale = self.query_scale(output) if self.query_scale is not None else 1
+            query_pos = pos_scale * raw_query_pos
+
+            # main process
+            output = layer(
+                tgt=output,
+                tgt_query_pos=query_pos,
+                tgt_reference_points=reference_points_input,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+                memory_level_start_index=level_start_index,
+                memory_spatial_shapes=spatial_shapes,
+                self_attn_mask=tgt_mask,
+                cross_attn_mask=memory_mask,
+            )
+            if output.isnan().any() | output.isinf().any():
+                print(f"output layer_id {layer_id} is nan")
+                try:
+                    num_nan = output.isnan().sum().item()
+                    num_inf = output.isinf().sum().item()
+                    print(f"num_nan {num_nan}, num_inf {num_inf}")
+                except Exception as e:
+                    print(e)
+
+            # iter update
+            if self.bbox_embed is not None:
+
+                reference_before_sigmoid = inverse_sigmoid(reference_points)
+                delta_unsig = self.bbox_embed[layer_id](output)
+                outputs_unsig = delta_unsig + reference_before_sigmoid
+                new_reference_points = outputs_unsig.sigmoid()
+
+                if self.rm_detach and "dec" in self.rm_detach:
+                    reference_points = new_reference_points
+                else:
+                    reference_points = new_reference_points.detach()
+
+                if self.use_detached_boxes_dec_out:
+                    ref_points.append(reference_points)
+                else:
+                    ref_points.append(new_reference_points)
+
+            if self.return_intermediate:
+                intermediate.append(self.norm(output))
+
+        if self.return_intermediate:
+            return [
+                [itm_out.transpose(0, 1) for itm_out in intermediate],
+                [itm_refpoint.transpose(0, 1) for itm_refpoint in ref_points],
+            ]
+        else:
+            return self.norm(output).transpose(0, 1)

detect_tools/upn/models/encoder/__init__.py

@@ -0,0 +1,3 @@
+from .upn_encoder import DeformableTransformerEncoderLayer, UPNEncoder
+
+__all__ = ["UPNEncoder", "DeformableTransformerEncoderLayer"]

detect_tools/upn/models/encoder/upn_encoder.py

@@ -0,0 +1,288 @@
+from typing import Dict
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint as checkpoint
+
+from detect_tools.upn import ENCODERS, build_encoder
+from detect_tools.upn.models.utils import get_activation_fn, get_clones
+from detect_tools.upn.ops.modules import MSDeformAttn
+
+
+@ENCODERS.register_module()
+class DeformableTransformerEncoderLayer(nn.Module):
+    """Deformable Transformer Encoder Layer.
+
+    Args:
+        d_model (int): The dimension of keys/values/queries in
+            :class:`MultiheadAttention`.
+        d_ffn (int): The dimension of the feedforward network model.
+        dropout (float): Probability of an element to be zeroed.
+        activation (str): Activation function in the feedforward network.
+            'relu' and 'gelu' are supported.
+        n_levels (int): The number of levels in Multi-Scale Deformable Attention.
+        n_heads (int): Parallel attention heads.
+        n_points (int): Number of sampling points in Multi-Scale Deformable Attention.
+        add_channel_attention (bool): If True, add channel attention.
+    """
+
+    def __init__(
+        self,
+        d_model: int = 256,
+        d_ffn: int = 1024,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        n_levels: int = 4,
+        n_heads: int = 8,
+        n_points: int = 4,
+        add_channel_attention: bool = False,
+    ) -> None:
+        super().__init__()
+
+        # self attention
+        self.self_attn = MSDeformAttn(d_model, n_levels, n_heads, n_points)
+        self.dropout1 = nn.Dropout(dropout)
+        self.norm1 = nn.LayerNorm(d_model)
+
+        # ffn
+        self.linear1 = nn.Linear(d_model, d_ffn)
+        self.activation = get_activation_fn(activation, d_model=d_ffn)
+        self.dropout2 = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(d_ffn, d_model)
+        self.dropout3 = nn.Dropout(dropout)
+        self.norm2 = nn.LayerNorm(d_model)
+
+        # channel attention
+        self.add_channel_attention = add_channel_attention
+        if add_channel_attention:
+            self.activ_channel = get_activation_fn("dyrelu", d_model=d_model)
+            self.norm_channel = nn.LayerNorm(d_model)
+
+    @staticmethod
+    def with_pos_embed(tensor, pos):
+        return tensor if pos is None else tensor + pos
+
+    def forward_ffn(self, src: torch.Tensor) -> torch.Tensor:
+        src2 = self.linear2(self.dropout2(self.activation(self.linear1(src))))
+        src = src + self.dropout3(src2)
+        src = self.norm2(src)
+        return src
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        pos: torch.Tensor,
+        reference_points: torch.Tensor,
+        spatial_shapes: torch.Tensor,
+        level_start_index: torch.Tensor,
+        key_padding_mask: torch.Tensor = None,
+    ) -> torch.Tensor:
+        """Forward function for `DeformableTransformerEncoderLayer`.
+
+        Args:
+            src (torch.Tensor): The input sequence of shape (S, N, E).
+            pos (torch.Tensor): The position embedding of shape (S, N, E).
+            reference_points (torch.Tensor): The reference points of shape (N, L, 2).
+            spatial_shapes (torch.Tensor): The spatial shapes of feature levels.
+            level_start_index (torch.Tensor): The start index of each level.
+            key_padding_mask (torch.Tensor): The mask for keys with shape (N, S).
+        """
+        # self attention
+        # import ipdb; ipdb.set_trace()
+        src2 = self.self_attn(
+            self.with_pos_embed(src, pos),
+            reference_points,
+            src,
+            spatial_shapes,
+            level_start_index,
+            key_padding_mask,
+        )
+        src = src + self.dropout1(src2)
+        src = self.norm1(src)
+
+        # ffn
+        src = self.forward_ffn(src)
+
+        # channel attn
+        if self.add_channel_attention:
+            src = self.norm_channel(src + self.activ_channel(src))
+
+        return src
+
+
+@ENCODERS.register_module()
+class UPNEncoder(nn.Module):
+    """Implementation of UPN Encoder.
+
+    Args:
+        num_layers (int): The number of layers in the TransformerEncoder.
+        d_model (int, optional): The dimension of the input feature. Defaults to 256.
+        encoder_layer_cfg (Dict): Config for the DeformableEncoderLayer.
+        use_checkpoint (bool, optional): Whether to use checkpoint in the fusion layer for
+            memory saving. Defaults to False.
+        use_transformer_ckpt (bool, optional): Whether to use checkpoint for the deformableencoder.
+        enc_layer_share (bool, optional): Whether to share the same memory for the encoder_layer.
+            Defaults to False. This is used for all the sub-layers in the basic block.
+    """
+
+    def __init__(
+        self,
+        num_layers: int,
+        d_model: int = 256,
+        encoder_layer_cfg: Dict = None,
+        use_checkpoint: bool = True,
+        use_transformer_ckpt: bool = True,
+        enc_layer_share: bool = False,
+        multi_level_encoder_fusion: str = None,
+    ):
+        super().__init__()
+        # prepare layers
+        self.layers = []
+        self.refImg_layers = []
+        self.fusion_layers = []
+        encoder_layer = build_encoder(encoder_layer_cfg)
+
+        self.multi_level_encoder_fusion = multi_level_encoder_fusion
+        self._initilize_memory_fusion_layers(
+            multi_level_encoder_fusion, num_layers, d_model
+        )
+
+        if num_layers > 0:
+            self.layers = get_clones(
+                encoder_layer, num_layers, layer_share=enc_layer_share
+            )
+        else:
+            self.layers = []
+            del encoder_layer
+
+        self.query_scale = None
+        self.num_layers = num_layers
+        self.d_model = d_model
+
+        self.use_checkpoint = use_checkpoint
+        self.use_transformer_ckpt = use_transformer_ckpt
+
+    def _initilize_memory_fusion_layers(self, fusion_type, num_layers, d_model):
+        if fusion_type is None:
+            self.memory_fusion_layer = None
+            return
+
+        assert fusion_type in ["dense_net_fusion", "stable_dense_fusion"]
+        if fusion_type == "stable_dense_fusion":
+            self.memory_fusion_layer = nn.Sequential(
+                nn.Linear(d_model * (num_layers + 1), d_model),
+                nn.LayerNorm(d_model),
+            )
+            nn.init.constant_(self.memory_fusion_layer[0].bias, 0)
+        elif fusion_type == "dense_net_fusion":
+            self.memory_fusion_layer = nn.ModuleList()
+            for i in range(num_layers):
+                self.memory_fusion_layer.append(
+                    nn.Sequential(
+                        nn.Linear(
+                            d_model * (i + 2), d_model
+                        ),  # from second encoder layer, 512 -> 256 / 3rd: 768 -> 256
+                        nn.LayerNorm(d_model),
+                    )
+                )
+            for layer in self.memory_fusion_layer:
+                nn.init.constant_(layer[0].bias, 0)
+        else:
+            raise NotImplementedError
+
+    @staticmethod
+    def get_reference_points(spatial_shapes, valid_ratios, device):
+        reference_points_list = []
+        for lvl, (H_, W_) in enumerate(spatial_shapes):
+
+            ref_y, ref_x = torch.meshgrid(
+                torch.linspace(0.5, H_ - 0.5, H_, dtype=torch.float32, device=device),
+                torch.linspace(0.5, W_ - 0.5, W_, dtype=torch.float32, device=device),
+            )
+            ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, lvl, 1] * H_)
+            ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, lvl, 0] * W_)
+            ref = torch.stack((ref_x, ref_y), -1)
+            reference_points_list.append(ref)
+        reference_points = torch.cat(reference_points_list, 1)
+        reference_points = reference_points[:, :, None] * valid_ratios[:, None]
+        return reference_points
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        pos: torch.Tensor,
+        spatial_shapes: torch.Tensor,
+        level_start_index: torch.Tensor,
+        valid_ratios: torch.Tensor,
+        key_padding_mask: torch.Tensor = None,
+    ):
+        """Forward function
+
+        Args:
+            src (torch.Tensor): Flattened Image features in shape [bs, sum(hi*wi), 256]
+            pos (torch.Tensor): Position embedding for image feature in shape [bs, sum(hi*wi), 256]
+            spatial_shapes (torch.Tensor): Spatial shape of each level in shape [num_level, 2]
+            level_start_index (torch.Tensor): Start index of each level in shape [num_level]
+            valid_ratios (torch.Tensor): Valid ratio of each level in shape [bs, num_level, 2]
+            key_padding_mask (torch.Tensor): Padding mask for image feature in shape [bs, sum(hi*wi)]
+            memory_refImg (torch.Tensor, optional): Text feature in shape [bs, n_ref, 256]. Defaults
+                to None.
+            refImg_padding_mask (torch.Tensor, optional): Padding mask for reference image feature
+                in shape [bs, n_text]. Defaults to None.
+            pos_refImg (torch.Tensor, optional): Position embedding for reference image in shape
+                [bs, n_ref, 256]. Defaults to None.
+            refImg_self_attention_masks (torch.Tensor, optional): Self attention mask for reference
+                image feature in shape [bs, n_ref, n_ref]. Defaults to None.
+        Outpus:
+            torch.Tensor: Encoded image feature in shape [bs, sum(hi*wi), 256]
+            torch.Tensor: Encoded reference image feature in shape [bs, n_ref, 256]
+        """
+
+        output = src
+        # preparation and reshape
+        if self.num_layers > 0:
+            reference_points = self.get_reference_points(
+                spatial_shapes, valid_ratios, device=src.device
+            )
+
+        # multi-level dense fusion
+        output_list = [output]
+        # main process
+        for layer_id, layer in enumerate(self.layers):
+            # main process
+            if self.use_transformer_ckpt:
+                output = checkpoint.checkpoint(
+                    layer,
+                    output,
+                    pos,
+                    reference_points,
+                    spatial_shapes,
+                    level_start_index,
+                    key_padding_mask,
+                )
+            else:
+                output = layer(
+                    src=output,
+                    pos=pos,
+                    reference_points=reference_points,
+                    spatial_shapes=spatial_shapes,
+                    level_start_index=level_start_index,
+                    key_padding_mask=key_padding_mask,
+                )
+
+                output_list.append(output)
+                if (
+                    self.multi_level_encoder_fusion is not None
+                    and self.multi_level_encoder_fusion == "dense_net_fusion"
+                ):
+                    output = self.memory_fusion_layer[layer_id](
+                        torch.cat(output_list, dim=-1)
+                    )
+
+        if (
+            self.multi_level_encoder_fusion is not None
+            and self.multi_level_encoder_fusion == "stable_dense_fusion"
+        ):
+            output = self.memory_fusion_layer(torch.cat(output_list, dim=-1))
+
+        return output

detect_tools/upn/models/module/__init__.py

@@ -0,0 +1,5 @@
+from .contrastive import ContrastiveAssign
+from .mlp import MLP
+from .nested_tensor import NestedTensor, nested_tensor_from_tensor_list
+
+__all__ = ["MLP", "NestedTensor", "nested_tensor_from_tensor_list", "ContrastiveAssign"]

detect_tools/upn/models/module/contrastive.py

@@ -0,0 +1,29 @@
+from typing import Dict
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class ContrastiveAssign(nn.Module):
+
+    def __init__(
+        self,
+        cal_bias: nn.Module = None,
+    ) -> None:
+        """Lanuage-Image Contrastive Assignment used to calculate the similarity between
+        the text and the image.
+
+        Args:
+            cal_bias (nn.Module, optional): The bias used to calculate the similarity.
+                Defaults to None.
+            max_text_len (int, optional): The max length of the text. Defaults to 256.
+        """
+        super().__init__()
+        self.cal_bias = cal_bias
+
+    def forward(self, x: torch.Tensor, ref_dict: Dict):
+
+        y = ref_dict["encoded_ref_feature"]
+        res = x @ y.transpose(-1, -2)
+        return res

detect_tools/upn/models/module/mlp.py

@@ -0,0 +1,18 @@
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class MLP(nn.Module):
+    """ Very simple multi-layer perceptron (also called FFN)"""
+
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(
+            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        return x

detect_tools/upn/models/module/nested_tensor.py

@@ -0,0 +1,199 @@
+from typing import List, Union
+
+import torch
+import torchvision
+
+
+class NestedTensor(object):
+    """Define a NestedTensor class
+
+    Args:
+        tensors (torch.Tensor): Tensor with shape [batch, C, H, W] or [C, H, W]
+        mask (Union[torch.Tensor, str]): mask with shape [batch, H, W] or [H, W]. If mask
+            is 'auto', it will be generated automatically by summing the tensor along
+            the channel dimension. Mask is used to indicate the padding area.
+    """
+
+    def __init__(
+        self, tensors: torch.Tensor, mask: Union[torch.Tensor, str] = "auto"
+    ) -> None:
+        self.tensors = tensors
+        self.mask = mask
+        if mask == "auto":
+            self.mask = torch.zeros_like(tensors).to(tensors.device)
+            if self.mask.dim() == 3:
+                self.mask = self.mask.sum(0).to(bool)
+            elif self.mask.dim() == 4:
+                self.mask = self.mask.sum(1).to(bool)
+            else:
+                raise ValueError(
+                    "tensors dim must be 3 or 4 but {}({})".format(
+                        self.tensors.dim(), self.tensors.shape
+                    )
+                )
+
+    def imgsize(self) -> List[torch.Tensor]:
+        """get the img size of the tensor
+
+        Returns:
+            list[torch.Tensor]: list of tensor with shape [2] which is [H, W]
+        """
+        res = []
+        for i in range(self.tensors.shape[0]):
+            mask = self.mask[i]
+            maxH = (~mask).sum(0).max()
+            maxW = (~mask).sum(1).max()
+            res.append(torch.Tensor([maxH, maxW]))
+        return res
+
+    def to(self, device: torch.device):
+        """Move tensors and mask to the given device
+
+        Args:
+            device (torch.device): device to move
+
+        Returns:
+            NestedTensor: moved NestedTensor
+        """
+        cast_tensor = self.tensors.to(device)
+        mask = self.mask
+        if mask is not None:
+            assert mask is not None
+            cast_mask = mask.to(device)
+        else:
+            cast_mask = None
+        return NestedTensor(cast_tensor, cast_mask)
+
+    def to_img_list_single(
+        self, tensor: torch.Tensor, mask: torch.Tensor
+    ) -> torch.Tensor:
+        """remove the padding for one image
+
+        Args:
+            tensor (torch.Tensor): tensor with shape [C, H, W]
+            mask (torch.Tensor): mask with shape [H, W]
+
+        Returns:
+            torch.Tensor: tensor with shape [C, maxH, maxW]
+        """
+        assert tensor.dim() == 3, "dim of tensor should be 3 but {}".format(
+            tensor.dim()
+        )
+        maxH = (~mask).sum(0).max()
+        maxW = (~mask).sum(1).max()
+        img = tensor[:, :maxH, :maxW]
+        return img
+
+    def to_img_list(self) -> List[torch.Tensor]:
+        """remove the padding and convert to img list
+
+        Returns:
+            list[torch.Tensor]: list of tensor with shape [C, maxH, maxW]
+        """
+        if self.tensors.dim() == 3:
+            return self.to_img_list_single(self.tensors, self.mask)
+        else:
+            res = []
+            for i in range(self.tensors.shape[0]):
+                tensor_i = self.tensors[i]
+                mask_i = self.mask[i]
+                res.append(self.to_img_list_single(tensor_i, mask_i))
+            return res
+
+    @property
+    def device(self):
+        return self.tensors.device
+
+    def decompose(self):
+        return self.tensors, self.mask
+
+    def __repr__(self):
+        return str(self.tensors)
+
+    @property
+    def shape(self):
+        return {"tensors.shape": self.tensors.shape, "mask.shape": self.mask.shape}
+
+
+def _max_by_axis(the_list):
+    # type: (List[List[int]]) -> List[int]
+    maxes = the_list[0]
+    for sublist in the_list[1:]:
+        for index, item in enumerate(sublist):
+            maxes[index] = max(maxes[index], item)
+    return maxes
+
+
+def nested_tensor_from_tensor_list(
+    tensor_list: List[torch.Tensor], fixed_img_size=None
+):
+    if fixed_img_size is not None:
+        if isinstance(fixed_img_size, (list, tuple)):
+            assert (
+                len(fixed_img_size) == 2
+            ), "image size should be a tuple or list with two elements"
+        elif isinstance(fixed_img_size, int):
+            fixed_img_size = [fixed_img_size, fixed_img_size]
+
+    if tensor_list[0].ndim == 3:
+        if torchvision._is_tracing():
+            # nested_tensor_from_tensor_list() does not export well to ONNX
+            # call _onnx_nested_tensor_from_tensor_list() instead
+            return _onnx_nested_tensor_from_tensor_list(tensor_list)
+
+        # TODO make it support different-sized images
+        max_size = _max_by_axis([list(img.shape) for img in tensor_list])
+
+        if fixed_img_size is not None:
+            c, orig_h, orig_w = max_size
+            assert (
+                orig_h <= fixed_img_size[0] and orig_w <= fixed_img_size[1]
+            ), f"{orig_h} {orig_w} the fixed output image size should be larger than original image"
+            max_size = [c, fixed_img_size[0], fixed_img_size[1]]
+
+        # min_size = tuple(min(s) for s in zip(*[img.shape for img in tensor_list]))
+        batch_shape = [len(tensor_list)] + max_size
+        b, c, h, w = batch_shape
+        dtype = tensor_list[0].dtype
+        device = tensor_list[0].device
+        tensor = torch.zeros(batch_shape, dtype=dtype, device=device)
+        mask = torch.ones((b, h, w), dtype=torch.bool, device=device)
+        for img, pad_img, m in zip(tensor_list, tensor, mask):
+            pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img)
+            m[: img.shape[1], : img.shape[2]] = False
+    else:
+        raise ValueError("not supported")
+    return NestedTensor(tensor, mask)
+
+
+@torch.jit.unused
+def _onnx_nested_tensor_from_tensor_list(
+    tensor_list: List[torch.Tensor],
+) -> NestedTensor:
+    max_size = []
+    for i in range(tensor_list[0].dim()):
+        max_size_i = torch.max(
+            torch.stack([img.shape[i] for img in tensor_list]).to(torch.float32)
+        ).to(torch.int64)
+        max_size.append(max_size_i)
+    max_size = tuple(max_size)
+
+    padded_imgs = []
+    padded_masks = []
+    for img in tensor_list:
+        padding = [(s1 - s2) for s1, s2 in zip(max_size, tuple(img.shape))]
+        padded_img = torch.nn.functional.pad(
+            img, (0, padding[2], 0, padding[1], 0, padding[0])
+        )
+        padded_imgs.append(padded_img)
+
+        m = torch.zeros_like(img[0], dtype=torch.int, device=img.device)
+        padded_mask = torch.nn.functional.pad(
+            m, (0, padding[2], 0, padding[1]), "constant", 1
+        )
+        padded_masks.append(padded_mask.to(torch.bool))
+
+    tensor = torch.stack(padded_imgs)
+    mask = torch.stack(padded_masks)
+
+    return NestedTensor(tensor, mask=mask)

detect_tools/upn/models/utils/__init__.py

@@ -0,0 +1,23 @@
+from .detr_utils import (
+    PositionEmbeddingLearned,
+    PositionEmbeddingSine,
+    PositionEmbeddingSineHW,
+    clean_state_dict,
+    gen_encoder_output_proposals,
+    gen_sineembed_for_position,
+    get_activation_fn,
+    get_clones,
+    inverse_sigmoid,
+)
+
+__all__ = [
+    "inverse_sigmoid",
+    "gen_encoder_output_proposals",
+    "get_clones",
+    "gen_sineembed_for_position",
+    "get_activation_fn",
+    "clean_state_dict",
+    "PositionEmbeddingSine",
+    "PositionEmbeddingSineHW",
+    "PositionEmbeddingLearned",
+]

detect_tools/upn/models/utils/detr_utils.py

@@ -0,0 +1,415 @@
+import copy
+import math
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import nn
+
+from detect_tools.upn import POS_EMBEDDINGS
+from detect_tools.upn.models.module import NestedTensor
+
+
+@POS_EMBEDDINGS.register_module()
+class PositionEmbeddingSine(nn.Module):
+    """This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+
+    Args:
+        num_pos_feats (int): The channel of positional embeddings.
+        temperature (float): The temperature used in positional embeddings.
+        normalize (bool): Whether to normalize the positional embeddings.
+        scale (float): The scale factor of positional embeddings.
+    """
+
+    def __init__(
+        self,
+        num_pos_feats: int = 64,
+        temperature: int = 10000,
+        normalize: bool = False,
+        scale: float = None,
+    ) -> None:
+        super().__init__()
+        self.num_pos_feats = num_pos_feats
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+    def forward(self, tensor_list: NestedTensor) -> torch.Tensor:
+        """Forward function.
+
+        Args:
+            tensor_list (NestedTensor): NestedTensor wrapping the input tensor.
+
+        Returns:
+            torch.Tensor: Positional encoding in shape (B, num_pos_feats*2, H, W)
+        """
+        x = tensor_list.tensors
+        mask = tensor_list.mask
+        assert mask is not None
+        not_mask = ~mask
+        y_embed = not_mask.cumsum(1, dtype=torch.float32)
+        x_embed = not_mask.cumsum(2, dtype=torch.float32)
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos_y = torch.stack(
+            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        return pos
+
+
+@POS_EMBEDDINGS.register_module()
+class PositionEmbeddingSineHW(nn.Module):
+    """This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+
+    Args:
+        num_pos_feats (int): The channel of positional embeddings.
+        temperatureH (float): The temperature used in positional embeddings.
+        temperatureW (float): The temperature used in positional embeddings.
+        normalize (bool): Whether to normalize the positional embeddings.
+        scale (float): The scale factor of positional embeddings.
+    """
+
+    def __init__(
+        self,
+        num_pos_feats: int = 64,
+        temperatureH: int = 10000,
+        temperatureW: int = 10000,
+        normalize: bool = False,
+        scale: float = None,
+    ) -> None:
+        super().__init__()
+        self.num_pos_feats = num_pos_feats
+        self.temperatureH = temperatureH
+        self.temperatureW = temperatureW
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+    def forward(self, tensor_list: NestedTensor) -> torch.Tensor:
+        """Forward function.
+
+        Args:
+            tensor_list (NestedTensor): NestedTensor wrapping the input tensor.
+
+        Returns:
+            torch.Tensor: Positional encoding in shape (B, num_pos_feats*2, H, W)
+        """
+        x = tensor_list.tensors
+        mask = tensor_list.mask
+        assert mask is not None
+        not_mask = ~mask
+        y_embed = not_mask.cumsum(1, dtype=torch.float32)
+        x_embed = not_mask.cumsum(2, dtype=torch.float32)
+
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_tx = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_tx = self.temperatureW ** (2 * (dim_tx // 2) / self.num_pos_feats)
+        pos_x = x_embed[:, :, :, None] / dim_tx
+
+        dim_ty = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_ty = self.temperatureH ** (2 * (dim_ty // 2) / self.num_pos_feats)
+        pos_y = y_embed[:, :, :, None] / dim_ty
+
+        pos_x = torch.stack(
+            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos_y = torch.stack(
+            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+
+        return pos
+
+
+@POS_EMBEDDINGS.register_module()
+class PositionEmbeddingLearned(nn.Module):
+    """Absolute pos embedding, learned.
+
+    Args:
+        num_pos_feats (int): The channel dimension of positional embeddings.
+        num_row (int): The number of rows of the input feature map.
+        num_col (int): The number of columns of the input feature map.
+    """
+
+    def __init__(
+        self, num_row: int = 50, num_col: int = 50, num_pos_feats: int = 256
+    ) -> None:
+        super().__init__()
+        self.row_embed = nn.Embedding(num_row, num_pos_feats)
+        self.col_embed = nn.Embedding(num_col, num_pos_feats)
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        nn.init.uniform_(self.row_embed.weight)
+        nn.init.uniform_(self.col_embed.weight)
+
+    def forward(self, tensor_list: NestedTensor) -> torch.Tensor:
+        """Forward function.
+
+        Args:
+            tensor_list (NestedTensor): NestedTensor wrapping the input tensor.
+
+        Returns:
+            torch.Tensor: Positional encoding in shape (B, num_pos_feats*2, H, W)
+        """
+        x = tensor_list.tensors
+        h, w = x.shape[-2:]
+        i = torch.arange(w, device=x.device)
+        j = torch.arange(h, device=x.device)
+        x_emb = self.col_embed(i)
+        y_emb = self.row_embed(j)
+        pos = (
+            torch.cat(
+                [
+                    x_emb.unsqueeze(0).repeat(h, 1, 1),
+                    y_emb.unsqueeze(1).repeat(1, w, 1),
+                ],
+                dim=-1,
+            )
+            .permute(2, 0, 1)
+            .unsqueeze(0)
+            .repeat(x.shape[0], 1, 1, 1)
+        )
+        return pos
+
+
+def build_position_encoding(args):
+    N_steps = args.hidden_dim // 2
+    if args.position_embedding in ("v2", "sine"):
+        # TODO find a better way of exposing other arguments
+        position_embedding = PositionEmbeddingSineHW(
+            N_steps,
+            temperatureH=args.pe_temperatureH,
+            temperatureW=args.pe_temperatureW,
+            normalize=True,
+        )
+    elif args.position_embedding in ("v3", "learned"):
+        position_embedding = PositionEmbeddingLearned(N_steps)
+    else:
+        raise ValueError(f"not supported {args.position_embedding}")
+
+    return position_embedding
+
+
+def clean_state_dict(state_dict):
+    new_state_dict = OrderedDict()
+    for k, v in state_dict.items():
+        if k[:7] == "module.":
+            k = k[7:]  # remove `module.`
+        new_state_dict[k] = v
+    return new_state_dict
+
+
+def get_activation_fn(activation: str, d_model: int = 256, batch_dim: int = 0):
+    """Return an activation function given a string
+
+    Args:
+        activation (str): activation function name
+        d_model (int, optional): d_model. Defaults to 256.
+        batch_dim (int, optional): batch dimension. Defaults to 0.
+
+    Returns:
+        F: activation function
+    """
+    if activation == "relu":
+        return F.relu
+    if activation == "gelu":
+        return F.gelu
+    if activation == "glu":
+        return F.glu
+    if activation == "prelu":
+        return nn.PReLU()
+    if activation == "selu":
+        return F.selu
+
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def get_clones(module: nn.Module, N: int, layer_share: bool = False):
+    """Copy module N times
+
+    Args:
+        module (nn.Module): module to copy
+        N (int): number of copies
+        layer_share (bool, optional): share the same layer. If true, the modules will
+            share the same memory. Defaults to False.
+    """
+    if layer_share:
+        return nn.ModuleList([module for _ in range(N)])
+    else:
+        return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])
+
+
+def inverse_sigmoid(x, eps=1e-3):
+    x = x.clamp(min=0, max=1)
+    x1 = x.clamp(min=eps)
+    x2 = (1 - x).clamp(min=eps)
+    return torch.log(x1 / x2)
+
+
+def gen_sineembed_for_position(pos_tensor):
+    # n_query, bs, _ = pos_tensor.size()
+    # sineembed_tensor = torch.zeros(n_query, bs, 256)
+    scale = 2 * math.pi
+    dim_t = torch.arange(128, dtype=torch.float32, device=pos_tensor.device)
+    dim_t = 10000 ** (2 * (dim_t // 2) / 128)
+    x_embed = pos_tensor[:, :, 0] * scale
+    y_embed = pos_tensor[:, :, 1] * scale
+    pos_x = x_embed[:, :, None] / dim_t
+    pos_y = y_embed[:, :, None] / dim_t
+    pos_x = torch.stack(
+        (pos_x[:, :, 0::2].sin(), pos_x[:, :, 1::2].cos()), dim=3
+    ).flatten(2)
+    pos_y = torch.stack(
+        (pos_y[:, :, 0::2].sin(), pos_y[:, :, 1::2].cos()), dim=3
+    ).flatten(2)
+    if pos_tensor.size(-1) == 2:
+        pos = torch.cat((pos_y, pos_x), dim=2)
+    elif pos_tensor.size(-1) == 4:
+        w_embed = pos_tensor[:, :, 2] * scale
+        pos_w = w_embed[:, :, None] / dim_t
+        pos_w = torch.stack(
+            (pos_w[:, :, 0::2].sin(), pos_w[:, :, 1::2].cos()), dim=3
+        ).flatten(2)
+
+        h_embed = pos_tensor[:, :, 3] * scale
+        pos_h = h_embed[:, :, None] / dim_t
+        pos_h = torch.stack(
+            (pos_h[:, :, 0::2].sin(), pos_h[:, :, 1::2].cos()), dim=3
+        ).flatten(2)
+
+        pos = torch.cat((pos_y, pos_x, pos_w, pos_h), dim=2)
+    else:
+        raise ValueError("Unknown pos_tensor shape(-1):{}".format(pos_tensor.size(-1)))
+    return pos
+
+
+def get_sine_pos_embed(
+    pos_tensor: torch.Tensor,
+    num_pos_feats: int = 128,
+    temperature: int = 10000,
+    exchange_xy: bool = True,
+):
+    """generate sine position embedding from a position tensor
+    Args:
+        pos_tensor (torch.Tensor): shape: [..., n].
+        num_pos_feats (int): projected shape for each float in the tensor.
+        temperature (int): temperature in the sine/cosine function.
+        exchange_xy (bool, optional): exchange pos x and pos y. \
+            For example, input tensor is [x,y], the results will be [pos(y), pos(x)]. Defaults to True.
+    Returns:
+        pos_embed (torch.Tensor): shape: [..., n*num_pos_feats].
+    """
+    scale = 2 * math.pi
+    dim_t = torch.arange(num_pos_feats, dtype=torch.float32, device=pos_tensor.device)
+    dim_t = temperature ** (
+        2 * torch.div(dim_t, 2, rounding_mode="floor") / num_pos_feats
+    )
+
+    def sine_func(x: torch.Tensor):
+        sin_x = x * scale / dim_t
+        sin_x = torch.stack(
+            (sin_x[..., 0::2].sin(), sin_x[..., 1::2].cos()), dim=3
+        ).flatten(2)
+        return sin_x
+
+    pos_res = [
+        sine_func(x) for x in pos_tensor.split([1] * pos_tensor.shape[-1], dim=-1)
+    ]
+    if exchange_xy:
+        pos_res[0], pos_res[1] = pos_res[1], pos_res[0]
+    pos_res = torch.cat(pos_res, dim=-1)
+    return pos_res
+
+
+def gen_encoder_output_proposals(
+    memory: torch.Tensor,
+    memory_padding_mask: torch.Tensor,
+    spatial_shapes: torch.Tensor,
+    learnedwh=None,
+):
+    """
+    Input:
+        - memory: bs, \sum{hw}, d_model
+        - memory_padding_mask: bs, \sum{hw}
+        - spatial_shapes: nlevel, 2
+        - learnedwh: 2
+    Output:
+        - output_memory: bs, \sum{hw}, d_model
+        - output_proposals: bs, \sum{hw}, 4
+    """
+    N_, S_, C_ = memory.shape
+    base_scale = 4.0
+    proposals = []
+    _cur = 0
+    for lvl, (H_, W_) in enumerate(spatial_shapes):
+        mask_flatten_ = memory_padding_mask[:, _cur : (_cur + H_ * W_)].view(
+            N_, H_, W_, 1
+        )
+        valid_H = torch.sum(~mask_flatten_[:, :, 0, 0], 1)
+        valid_W = torch.sum(~mask_flatten_[:, 0, :, 0], 1)
+        grid_y, grid_x = torch.meshgrid(
+            torch.linspace(0, H_ - 1, H_, dtype=torch.float32, device=memory.device),
+            torch.linspace(0, W_ - 1, W_, dtype=torch.float32, device=memory.device),
+        )
+        grid = torch.cat([grid_x.unsqueeze(-1), grid_y.unsqueeze(-1)], -1)  # H_, W_, 2
+
+        scale = torch.cat([valid_W.unsqueeze(-1), valid_H.unsqueeze(-1)], 1).view(
+            N_, 1, 1, 2
+        )
+        grid = (grid.unsqueeze(0).expand(N_, -1, -1, -1) + 0.5) / scale
+
+        if learnedwh is not None:
+            wh = torch.ones_like(grid) * learnedwh.sigmoid() * (2.0**lvl)
+        else:
+            wh = torch.ones_like(grid) * 0.05 * (2.0**lvl)
+
+        proposal = torch.cat((grid, wh), -1).view(N_, -1, 4)
+        proposals.append(proposal)
+        _cur += H_ * W_
+
+    output_proposals = torch.cat(proposals, 1)
+    output_proposals_valid = (
+        (output_proposals > 0.01) & (output_proposals < 0.99)
+    ).all(-1, keepdim=True)
+    output_proposals = torch.log(output_proposals / (1 - output_proposals))  # unsigmoid
+    output_proposals = output_proposals.masked_fill(
+        memory_padding_mask.unsqueeze(-1), float("inf")
+    )
+    output_proposals = output_proposals.masked_fill(
+        ~output_proposals_valid, float("inf")
+    )
+
+    output_memory = memory
+    output_memory = output_memory.masked_fill(
+        memory_padding_mask.unsqueeze(-1), float(0)
+    )
+    output_memory = output_memory.masked_fill(~output_proposals_valid, float(0))
+
+    return output_memory, output_proposals

detect_tools/upn/ops/functions/__init__.py

@@ -0,0 +1,10 @@
+# ------------------------------------------------------------------------------------------------
+# Deformable DETR
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------------------------------
+# Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+# ------------------------------------------------------------------------------------------------
+
+from .ms_deform_attn_func import MSDeformAttnFunction
+

detect_tools/upn/ops/functions/ms_deform_attn_func.py

@@ -0,0 +1,61 @@
+# ------------------------------------------------------------------------------------------------
+# Deformable DETR
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------------------------------
+# Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+# ------------------------------------------------------------------------------------------------
+
+from __future__ import absolute_import
+from __future__ import print_function
+from __future__ import division
+
+import torch
+import torch.nn.functional as F
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+
+import MultiScaleDeformableAttention as MSDA
+
+
+class MSDeformAttnFunction(Function):
+    @staticmethod
+    def forward(ctx, value, value_spatial_shapes, value_level_start_index, sampling_locations, attention_weights, im2col_step):
+        ctx.im2col_step = im2col_step
+        output = MSDA.ms_deform_attn_forward(
+            value, value_spatial_shapes, value_level_start_index, sampling_locations, attention_weights, ctx.im2col_step)
+        ctx.save_for_backward(value, value_spatial_shapes, value_level_start_index, sampling_locations, attention_weights)
+        return output
+
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, grad_output):
+        value, value_spatial_shapes, value_level_start_index, sampling_locations, attention_weights = ctx.saved_tensors
+        grad_value, grad_sampling_loc, grad_attn_weight = \
+            MSDA.ms_deform_attn_backward(
+                value, value_spatial_shapes, value_level_start_index, sampling_locations, attention_weights, grad_output, ctx.im2col_step)
+
+        return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
+
+
+def ms_deform_attn_core_pytorch(value, value_spatial_shapes, sampling_locations, attention_weights):
+    # for debug and test only,
+    # need to use cuda version instead
+    N_, S_, M_, D_ = value.shape
+    _, Lq_, M_, L_, P_, _ = sampling_locations.shape
+    value_list = value.split([H_ * W_ for H_, W_ in value_spatial_shapes], dim=1)
+    sampling_grids = 2 * sampling_locations - 1
+    sampling_value_list = []
+    for lid_, (H_, W_) in enumerate(value_spatial_shapes):
+        # N_, H_*W_, M_, D_ -> N_, H_*W_, M_*D_ -> N_, M_*D_, H_*W_ -> N_*M_, D_, H_, W_
+        value_l_ = value_list[lid_].flatten(2).transpose(1, 2).reshape(N_*M_, D_, H_, W_)
+        # N_, Lq_, M_, P_, 2 -> N_, M_, Lq_, P_, 2 -> N_*M_, Lq_, P_, 2
+        sampling_grid_l_ = sampling_grids[:, :, :, lid_].transpose(1, 2).flatten(0, 1)
+        # N_*M_, D_, Lq_, P_
+        sampling_value_l_ = F.grid_sample(value_l_, sampling_grid_l_,
+                                          mode='bilinear', padding_mode='zeros', align_corners=False)
+        sampling_value_list.append(sampling_value_l_)
+    # (N_, Lq_, M_, L_, P_) -> (N_, M_, Lq_, L_, P_) -> (N_, M_, 1, Lq_, L_*P_)
+    attention_weights = attention_weights.transpose(1, 2).reshape(N_*M_, 1, Lq_, L_*P_)
+    output = (torch.stack(sampling_value_list, dim=-2).flatten(-2) * attention_weights).sum(-1).view(N_, M_*D_, Lq_)
+    return output.transpose(1, 2).contiguous()

detect_tools/upn/ops/modules/__init__.py

@@ -0,0 +1,9 @@
+# ------------------------------------------------------------------------------------------------
+# Deformable DETR
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------------------------------
+# Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+# ------------------------------------------------------------------------------------------------
+
+from .ms_deform_attn import MSDeformAttn

detect_tools/upn/ops/modules/ms_deform_attn.py

@@ -0,0 +1,204 @@
+# ------------------------------------------------------------------------------------------------
+# Deformable DETR
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------------------------------
+# Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+# ------------------------------------------------------------------------------------------------
+
+from __future__ import absolute_import
+from __future__ import print_function
+from __future__ import division
+
+import warnings
+import math, os
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn.init import xavier_uniform_, constant_
+
+try:
+    from ..functions import MSDeformAttnFunction
+except:
+    warnings.warn("Failed to import MSDeformAttnFunction.")
+
+
+def _is_power_of_2(n):
+    if (not isinstance(n, int)) or (n < 0):
+        raise ValueError(
+            "invalid input for _is_power_of_2: {} (type: {})".format(n, type(n))
+        )
+    return (n & (n - 1) == 0) and n != 0
+
+
+class MSDeformAttn(nn.Module):
+    def __init__(
+        self, d_model=256, n_levels=4, n_heads=8, n_points=4, use_4D_normalizer=False
+    ):
+        """
+        Multi-Scale Deformable Attention Module
+        :param d_model      hidden dimension
+        :param n_levels     number of feature levels
+        :param n_heads      number of attention heads
+        :param n_points     number of sampling points per attention head per feature level
+        """
+        super().__init__()
+        if d_model % n_heads != 0:
+            raise ValueError(
+                "d_model must be divisible by n_heads, but got {} and {}".format(
+                    d_model, n_heads
+                )
+            )
+        _d_per_head = d_model // n_heads
+        # you'd better set _d_per_head to a power of 2 which is more efficient in our CUDA implementation
+        if not _is_power_of_2(_d_per_head):
+            warnings.warn(
+                "You'd better set d_model in MSDeformAttn to make the dimension of each attention head a power of 2 "
+                "which is more efficient in our CUDA implementation."
+            )
+
+        self.im2col_step = 64
+
+        self.d_model = d_model
+        self.n_levels = n_levels
+        self.n_heads = n_heads
+        self.n_points = n_points
+
+        self.sampling_offsets = nn.Linear(d_model, n_heads * n_levels * n_points * 2)
+        self.attention_weights = nn.Linear(d_model, n_heads * n_levels * n_points)
+        self.value_proj = nn.Linear(d_model, d_model)
+        self.output_proj = nn.Linear(d_model, d_model)
+
+        self.use_4D_normalizer = use_4D_normalizer
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        constant_(self.sampling_offsets.weight.data, 0.0)
+        thetas = torch.arange(self.n_heads, dtype=torch.float32) * (
+            2.0 * math.pi / self.n_heads
+        )
+        grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
+        grid_init = (
+            (grid_init / grid_init.abs().max(-1, keepdim=True)[0])
+            .view(self.n_heads, 1, 1, 2)
+            .repeat(1, self.n_levels, self.n_points, 1)
+        )
+        for i in range(self.n_points):
+            grid_init[:, :, i, :] *= i + 1
+        with torch.no_grad():
+            self.sampling_offsets.bias = nn.Parameter(grid_init.view(-1))
+        constant_(self.attention_weights.weight.data, 0.0)
+        constant_(self.attention_weights.bias.data, 0.0)
+        xavier_uniform_(self.value_proj.weight.data)
+        constant_(self.value_proj.bias.data, 0.0)
+        xavier_uniform_(self.output_proj.weight.data)
+        constant_(self.output_proj.bias.data, 0.0)
+
+    @torch.cuda.amp.autocast(enabled=False)
+    def forward(
+        self,
+        query,
+        reference_points,
+        input_flatten,
+        input_spatial_shapes,
+        input_level_start_index,
+        input_padding_mask=None,
+    ):
+        """
+        :param query                       (N, Length_{query}, C)
+        :param reference_points            (N, Length_{query}, n_levels, 2), range in [0, 1], top-left (0,0), bottom-right (1, 1), including padding area
+                                        or (N, Length_{query}, n_levels, 4), add additional (w, h) to form reference boxes
+        :param input_flatten               (N, \sum_{l=0}^{L-1} H_l \cdot W_l, C)
+        :param input_spatial_shapes        (n_levels, 2), [(H_0, W_0), (H_1, W_1), ..., (H_{L-1}, W_{L-1})]
+        :param input_level_start_index     (n_levels, ), [0, H_0*W_0, H_0*W_0+H_1*W_1, H_0*W_0+H_1*W_1+H_2*W_2, ..., H_0*W_0+H_1*W_1+...+H_{L-1}*W_{L-1}]
+        :param input_padding_mask          (N, \sum_{l=0}^{L-1} H_l \cdot W_l), True for padding elements, False for non-padding elements
+
+        :return output                     (N, Length_{query}, C)
+        """
+        N, Len_q, _ = query.shape
+        N, Len_in, _ = input_flatten.shape
+        assert (input_spatial_shapes[:, 0] * input_spatial_shapes[:, 1]).sum() == Len_in
+
+        value = self.value_proj(input_flatten)
+        if input_padding_mask is not None:
+            value = value.masked_fill(input_padding_mask[..., None], float(0))
+        value = value.view(N, Len_in, self.n_heads, self.d_model // self.n_heads)
+        sampling_offsets = self.sampling_offsets(query).view(
+            N, Len_q, self.n_heads, self.n_levels, self.n_points, 2
+        )
+        attention_weights = self.attention_weights(query).view(
+            N, Len_q, self.n_heads, self.n_levels * self.n_points
+        )
+        attention_weights = F.softmax(attention_weights, -1).view(
+            N, Len_q, self.n_heads, self.n_levels, self.n_points
+        )
+        # N, Len_q, n_heads, n_levels, n_points, 2
+
+        # if os.environ.get('IPDB_DEBUG_SHILONG', False) == 'INFO':
+        #     import ipdb; ipdb.set_trace()
+
+        if reference_points.shape[-1] == 2:
+            offset_normalizer = torch.stack(
+                [input_spatial_shapes[..., 1], input_spatial_shapes[..., 0]], -1
+            )
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :]
+                + sampling_offsets / offset_normalizer[None, None, None, :, None, :]
+            )
+        elif reference_points.shape[-1] == 4:
+            if self.use_4D_normalizer:
+                offset_normalizer = torch.stack(
+                    [input_spatial_shapes[..., 1], input_spatial_shapes[..., 0]], -1
+                )
+                sampling_locations = (
+                    reference_points[:, :, None, :, None, :2]
+                    + sampling_offsets
+                    / offset_normalizer[None, None, None, :, None, :]
+                    * reference_points[:, :, None, :, None, 2:]
+                    * 0.5
+                )
+            else:
+                sampling_locations = (
+                    reference_points[:, :, None, :, None, :2]
+                    + sampling_offsets
+                    / self.n_points
+                    * reference_points[:, :, None, :, None, 2:]
+                    * 0.5
+                )
+        else:
+            raise ValueError(
+                "Last dim of reference_points must be 2 or 4, but get {} instead.".format(
+                    reference_points.shape[-1]
+                )
+            )
+
+        # if os.environ.get('IPDB_DEBUG_SHILONG', False) == 'INFO':
+        #     import ipdb; ipdb.set_trace()
+
+        # for amp
+        if value.dtype == torch.float16:
+            # for mixed precision
+            output = MSDeformAttnFunction.apply(
+                value.to(torch.float32),
+                input_spatial_shapes,
+                input_level_start_index,
+                sampling_locations.to(torch.float32),
+                attention_weights,
+                self.im2col_step,
+            )
+            output = output.to(torch.float16)
+            output = self.output_proj(output)
+            return output
+
+        output = MSDeformAttnFunction.apply(
+            value,
+            input_spatial_shapes,
+            input_level_start_index,
+            sampling_locations,
+            attention_weights,
+            self.im2col_step,
+        )
+        output = self.output_proj(output)
+        return output

detect_tools/upn/ops/modules/ms_deform_attn_key_aware.py

@@ -0,0 +1,130 @@
+# ------------------------------------------------------------------------------------------------
+# Deformable DETR
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------------------------------
+# Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+# ------------------------------------------------------------------------------------------------
+
+from __future__ import absolute_import
+from __future__ import print_function
+from __future__ import division
+
+import warnings
+import math, os
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn.init import xavier_uniform_, constant_
+
+try:
+    from ..functions import MSDeformAttnFunction
+except:
+    warnings.warn('Failed to import MSDeformAttnFunction.')
+
+
+def _is_power_of_2(n):
+    if (not isinstance(n, int)) or (n < 0):
+        raise ValueError("invalid input for _is_power_of_2: {} (type: {})".format(n, type(n)))
+    return (n & (n-1) == 0) and n != 0
+
+
+class MSDeformAttn(nn.Module):
+    def __init__(self, d_model=256, n_levels=4, n_heads=8, n_points=4, use_4D_normalizer=False):
+        """
+        Multi-Scale Deformable Attention Module
+        :param d_model      hidden dimension
+        :param n_levels     number of feature levels
+        :param n_heads      number of attention heads
+        :param n_points     number of sampling points per attention head per feature level
+        """
+        super().__init__()
+        if d_model % n_heads != 0:
+            raise ValueError('d_model must be divisible by n_heads, but got {} and {}'.format(d_model, n_heads))
+        _d_per_head = d_model // n_heads
+        # you'd better set _d_per_head to a power of 2 which is more efficient in our CUDA implementation
+        if not _is_power_of_2(_d_per_head):
+            warnings.warn("You'd better set d_model in MSDeformAttn to make the dimension of each attention head a power of 2 "
+                          "which is more efficient in our CUDA implementation.")
+
+        self.im2col_step = 64
+
+        self.d_model = d_model
+        self.n_levels = n_levels
+        self.n_heads = n_heads
+        self.n_points = n_points
+
+        self.sampling_offsets = nn.Linear(d_model, n_heads * n_levels * n_points * 2)
+        self.attention_weights = nn.Linear(d_model, n_heads * n_levels * n_points)
+        self.value_proj = nn.Linear(d_model, d_model)
+        self.output_proj = nn.Linear(d_model, d_model)
+
+        self.use_4D_normalizer = use_4D_normalizer
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        constant_(self.sampling_offsets.weight.data, 0.)
+        thetas = torch.arange(self.n_heads, dtype=torch.float32) * (2.0 * math.pi / self.n_heads)
+        grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
+        grid_init = (grid_init / grid_init.abs().max(-1, keepdim=True)[0]).view(self.n_heads, 1, 1, 2).repeat(1, self.n_levels, self.n_points, 1)
+        for i in range(self.n_points):
+            grid_init[:, :, i, :] *= i + 1
+        with torch.no_grad():
+            self.sampling_offsets.bias = nn.Parameter(grid_init.view(-1))
+        constant_(self.attention_weights.weight.data, 0.)
+        constant_(self.attention_weights.bias.data, 0.)
+        xavier_uniform_(self.value_proj.weight.data)
+        constant_(self.value_proj.bias.data, 0.)
+        xavier_uniform_(self.output_proj.weight.data)
+        constant_(self.output_proj.bias.data, 0.)
+
+    def forward(self, query, key, reference_points, input_flatten, input_spatial_shapes, input_level_start_index, input_padding_mask=None):
+        """
+        :param query                       (N, Length_{query}, C)
+        :param key                          (N, 1, C)
+        :param reference_points            (N, Length_{query}, n_levels, 2), range in [0, 1], top-left (0,0), bottom-right (1, 1), including padding area
+                                        or (N, Length_{query}, n_levels, 4), add additional (w, h) to form reference boxes
+        :param input_flatten               (N, \sum_{l=0}^{L-1} H_l \cdot W_l, C)
+        :param input_spatial_shapes        (n_levels, 2), [(H_0, W_0), (H_1, W_1), ..., (H_{L-1}, W_{L-1})]
+        :param input_level_start_index     (n_levels, ), [0, H_0*W_0, H_0*W_0+H_1*W_1, H_0*W_0+H_1*W_1+H_2*W_2, ..., H_0*W_0+H_1*W_1+...+H_{L-1}*W_{L-1}]
+        :param input_padding_mask          (N, \sum_{l=0}^{L-1} H_l \cdot W_l), True for padding elements, False for non-padding elements
+
+        :return output                     (N, Length_{query}, C)
+        """
+        N, Len_q, _ = query.shape
+        N, Len_in, _ = input_flatten.shape
+        assert (input_spatial_shapes[:, 0] * input_spatial_shapes[:, 1]).sum() == Len_in
+
+        value = self.value_proj(input_flatten)
+        if input_padding_mask is not None:
+            value = value.masked_fill(input_padding_mask[..., None], float(0))
+        value = value.view(N, Len_in, self.n_heads, self.d_model // self.n_heads)
+        sampling_offsets = self.sampling_offsets(query).view(N, Len_q, self.n_heads, self.n_levels, self.n_points, 2)
+        attention_weights = self.attention_weights(query).view(N, Len_q, self.n_heads, self.n_levels * self.n_points)
+        attention_weights = F.softmax(attention_weights, -1).view(N, Len_q, self.n_heads, self.n_levels, self.n_points)
+        # N, Len_q, n_heads, n_levels, n_points, 2
+
+        # if os.environ.get('IPDB_DEBUG_SHILONG', False) == 'INFO':
+        #     import ipdb; ipdb.set_trace()
+
+        if reference_points.shape[-1] == 2:
+            offset_normalizer = torch.stack([input_spatial_shapes[..., 1], input_spatial_shapes[..., 0]], -1)
+            sampling_locations = reference_points[:, :, None, :, None, :] \
+                                 + sampling_offsets / offset_normalizer[None, None, None, :, None, :]
+        elif reference_points.shape[-1] == 4:
+            if self.use_4D_normalizer:
+                offset_normalizer = torch.stack([input_spatial_shapes[..., 1], input_spatial_shapes[..., 0]], -1)
+                sampling_locations = reference_points[:, :, None, :, None, :2] \
+                                    + sampling_offsets / offset_normalizer[None, None, None, :, None, :] * reference_points[:, :, None, :, None, 2:] * 0.5
+            else:
+                sampling_locations = reference_points[:, :, None, :, None, :2] \
+                                    + sampling_offsets / self.n_points * reference_points[:, :, None, :, None, 2:] * 0.5
+        else:
+            raise ValueError(
+                'Last dim of reference_points must be 2 or 4, but get {} instead.'.format(reference_points.shape[-1]))
+        output = MSDeformAttnFunction.apply(
+            value, input_spatial_shapes, input_level_start_index, sampling_locations, attention_weights, self.im2col_step)
+        output = self.output_proj(output)
+        return output

detect_tools/upn/ops/setup.py

@@ -0,0 +1,73 @@
+# ------------------------------------------------------------------------------------------------
+# Deformable DETR
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------------------------------
+# Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+# ------------------------------------------------------------------------------------------------
+
+import os
+import glob
+
+import torch
+
+from torch.utils.cpp_extension import CUDA_HOME
+from torch.utils.cpp_extension import CppExtension
+from torch.utils.cpp_extension import CUDAExtension
+
+from setuptools import find_packages
+from setuptools import setup
+
+requirements = ["torch", "torchvision"]
+
+def get_extensions():
+    this_dir = os.path.dirname(os.path.abspath(__file__))
+    extensions_dir = os.path.join(this_dir, "src")
+
+    main_file = glob.glob(os.path.join(extensions_dir, "*.cpp"))
+    source_cpu = glob.glob(os.path.join(extensions_dir, "cpu", "*.cpp"))
+    source_cuda = glob.glob(os.path.join(extensions_dir, "cuda", "*.cu"))
+
+    sources = main_file + source_cpu
+    extension = CppExtension
+    extra_compile_args = {"cxx": []}
+    define_macros = []
+
+    # import ipdb; ipdb.set_trace()
+
+    if torch.cuda.is_available() and CUDA_HOME is not None:
+        extension = CUDAExtension
+        sources += source_cuda
+        define_macros += [("WITH_CUDA", None)]
+        extra_compile_args["nvcc"] = [
+            "-DCUDA_HAS_FP16=1",
+            "-D__CUDA_NO_HALF_OPERATORS__",
+            "-D__CUDA_NO_HALF_CONVERSIONS__",
+            "-D__CUDA_NO_HALF2_OPERATORS__",
+        ]
+    else:
+        raise NotImplementedError('Cuda is not availabel')
+
+    sources = [os.path.join(extensions_dir, s) for s in sources]
+    include_dirs = [extensions_dir]
+    ext_modules = [
+        extension(
+            "MultiScaleDeformableAttention",
+            sources,
+            include_dirs=include_dirs,
+            define_macros=define_macros,
+            extra_compile_args=extra_compile_args,
+        )
+    ]
+    return ext_modules
+
+setup(
+    name="MultiScaleDeformableAttention",
+    version="1.0",
+    author="Weijie Su",
+    url="https://github.com/fundamentalvision/Deformable-DETR",
+    description="PyTorch Wrapper for CUDA Functions of Multi-Scale Deformable Attention",
+    packages=find_packages(exclude=("configs", "tests",)),
+    ext_modules=get_extensions(),
+    cmdclass={"build_ext": torch.utils.cpp_extension.BuildExtension},
+)

detect_tools/upn/ops/src/cpu/ms_deform_attn_cpu.cpp

@@ -0,0 +1,41 @@
+/*!
+**************************************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************************************
+* Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+**************************************************************************************************
+*/
+
+#include <vector>
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+
+
+at::Tensor
+ms_deform_attn_cpu_forward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const int im2col_step)
+{
+    AT_ERROR("Not implement on cpu");
+}
+
+std::vector<at::Tensor>
+ms_deform_attn_cpu_backward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const at::Tensor &grad_output,
+    const int im2col_step)
+{
+    AT_ERROR("Not implement on cpu");
+}
+

detect_tools/upn/ops/src/cpu/ms_deform_attn_cpu.h

@@ -0,0 +1,33 @@
+/*!
+**************************************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************************************
+* Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+**************************************************************************************************
+*/
+
+#pragma once
+#include <torch/extension.h>
+
+at::Tensor
+ms_deform_attn_cpu_forward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const int im2col_step);
+
+std::vector<at::Tensor>
+ms_deform_attn_cpu_backward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const at::Tensor &grad_output,
+    const int im2col_step);
+
+

detect_tools/upn/ops/src/cuda/ms_deform_attn_cuda.cu

@@ -0,0 +1,153 @@
+/*!
+**************************************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************************************
+* Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+**************************************************************************************************
+*/
+
+#include <vector>
+#include "cuda/ms_deform_im2col_cuda.cuh"
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+
+at::Tensor ms_deform_attn_cuda_forward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const int im2col_step)
+{
+    AT_ASSERTM(value.is_contiguous(), "value tensor has to be contiguous");
+    AT_ASSERTM(spatial_shapes.is_contiguous(), "spatial_shapes tensor has to be contiguous");
+    AT_ASSERTM(level_start_index.is_contiguous(), "level_start_index tensor has to be contiguous");
+    AT_ASSERTM(sampling_loc.is_contiguous(), "sampling_loc tensor has to be contiguous");
+    AT_ASSERTM(attn_weight.is_contiguous(), "attn_weight tensor has to be contiguous");
+
+    AT_ASSERTM(value.is_cuda(), "value must be a CUDA tensor");
+    AT_ASSERTM(spatial_shapes.is_cuda(), "spatial_shapes must be a CUDA tensor");
+    AT_ASSERTM(level_start_index.is_cuda(), "level_start_index must be a CUDA tensor");
+    AT_ASSERTM(sampling_loc.is_cuda(), "sampling_loc must be a CUDA tensor");
+    AT_ASSERTM(attn_weight.is_cuda(), "attn_weight must be a CUDA tensor");
+
+    const int batch = value.size(0);
+    const int spatial_size = value.size(1);
+    const int num_heads = value.size(2);
+    const int channels = value.size(3);
+
+    const int num_levels = spatial_shapes.size(0);
+
+    const int num_query = sampling_loc.size(1);
+    const int num_point = sampling_loc.size(4);
+
+    const int im2col_step_ = std::min(batch, im2col_step);
+
+    AT_ASSERTM(batch % im2col_step_ == 0, "batch(%d) must divide im2col_step(%d)", batch, im2col_step_);
+    
+    auto output = at::zeros({batch, num_query, num_heads, channels}, value.options());
+
+    const int batch_n = im2col_step_;
+    auto output_n = output.view({batch/im2col_step_, batch_n, num_query, num_heads, channels});
+    auto per_value_size = spatial_size * num_heads * channels;
+    auto per_sample_loc_size = num_query * num_heads * num_levels * num_point * 2;
+    auto per_attn_weight_size = num_query * num_heads * num_levels * num_point;
+    for (int n = 0; n < batch/im2col_step_; ++n)
+    {
+        auto columns = output_n.select(0, n);
+        AT_DISPATCH_FLOATING_TYPES(value.scalar_type(), "ms_deform_attn_forward_cuda", ([&] {
+            ms_deformable_im2col_cuda(at::cuda::getCurrentCUDAStream(),
+                value.data_ptr<scalar_t>() + n * im2col_step_ * per_value_size,
+                spatial_shapes.data_ptr<int64_t>(),
+                level_start_index.data_ptr<int64_t>(),
+                sampling_loc.data_ptr<scalar_t>() + n * im2col_step_ * per_sample_loc_size,
+                attn_weight.data_ptr<scalar_t>() + n * im2col_step_ * per_attn_weight_size,
+                batch_n, spatial_size, num_heads, channels, num_levels, num_query, num_point,
+                columns.data_ptr<scalar_t>());
+
+        }));
+    }
+
+    output = output.view({batch, num_query, num_heads*channels});
+
+    return output;
+}
+
+
+std::vector<at::Tensor> ms_deform_attn_cuda_backward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const at::Tensor &grad_output,
+    const int im2col_step)
+{
+
+    AT_ASSERTM(value.is_contiguous(), "value tensor has to be contiguous");
+    AT_ASSERTM(spatial_shapes.is_contiguous(), "spatial_shapes tensor has to be contiguous");
+    AT_ASSERTM(level_start_index.is_contiguous(), "level_start_index tensor has to be contiguous");
+    AT_ASSERTM(sampling_loc.is_contiguous(), "sampling_loc tensor has to be contiguous");
+    AT_ASSERTM(attn_weight.is_contiguous(), "attn_weight tensor has to be contiguous");
+    AT_ASSERTM(grad_output.is_contiguous(), "grad_output tensor has to be contiguous");
+
+    AT_ASSERTM(value.is_cuda(), "value must be a CUDA tensor");
+    AT_ASSERTM(spatial_shapes.is_cuda(), "spatial_shapes must be a CUDA tensor");
+    AT_ASSERTM(level_start_index.is_cuda(), "level_start_index must be a CUDA tensor");
+    AT_ASSERTM(sampling_loc.is_cuda(), "sampling_loc must be a CUDA tensor");
+    AT_ASSERTM(attn_weight.is_cuda(), "attn_weight must be a CUDA tensor");
+    AT_ASSERTM(grad_output.is_cuda(), "grad_output must be a CUDA tensor");
+
+    const int batch = value.size(0);
+    const int spatial_size = value.size(1);
+    const int num_heads = value.size(2);
+    const int channels = value.size(3);
+
+    const int num_levels = spatial_shapes.size(0);
+
+    const int num_query = sampling_loc.size(1);
+    const int num_point = sampling_loc.size(4);
+
+    const int im2col_step_ = std::min(batch, im2col_step);
+
+    AT_ASSERTM(batch % im2col_step_ == 0, "batch(%d) must divide im2col_step(%d)", batch, im2col_step_);
+
+    auto grad_value = at::zeros_like(value);
+    auto grad_sampling_loc = at::zeros_like(sampling_loc);
+    auto grad_attn_weight = at::zeros_like(attn_weight);
+
+    const int batch_n = im2col_step_;
+    auto per_value_size = spatial_size * num_heads * channels;
+    auto per_sample_loc_size = num_query * num_heads * num_levels * num_point * 2;
+    auto per_attn_weight_size = num_query * num_heads * num_levels * num_point;
+    auto grad_output_n = grad_output.view({batch/im2col_step_, batch_n, num_query, num_heads, channels});
+    
+    for (int n = 0; n < batch/im2col_step_; ++n)
+    {
+        auto grad_output_g = grad_output_n.select(0, n);
+        AT_DISPATCH_FLOATING_TYPES(value.scalar_type(), "ms_deform_attn_backward_cuda", ([&] {
+            ms_deformable_col2im_cuda(at::cuda::getCurrentCUDAStream(),
+                                    grad_output_g.data_ptr<scalar_t>(),
+                                    value.data_ptr<scalar_t>() + n * im2col_step_ * per_value_size,
+                                    spatial_shapes.data_ptr<int64_t>(),
+                                    level_start_index.data_ptr<int64_t>(),
+                                    sampling_loc.data_ptr<scalar_t>() + n * im2col_step_ * per_sample_loc_size,
+                                    attn_weight.data_ptr<scalar_t>() + n * im2col_step_ * per_attn_weight_size,
+                                    batch_n, spatial_size, num_heads, channels, num_levels, num_query, num_point,
+                                    grad_value.data_ptr<scalar_t>() +  n * im2col_step_ * per_value_size,
+                                    grad_sampling_loc.data_ptr<scalar_t>() + n * im2col_step_ * per_sample_loc_size,
+                                    grad_attn_weight.data_ptr<scalar_t>() + n * im2col_step_ * per_attn_weight_size);
+
+        }));
+    }
+
+    return {
+        grad_value, grad_sampling_loc, grad_attn_weight
+    };
+}

detect_tools/upn/ops/src/cuda/ms_deform_attn_cuda.h

@@ -0,0 +1,30 @@
+/*!
+**************************************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************************************
+* Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+**************************************************************************************************
+*/
+
+#pragma once
+#include <torch/extension.h>
+
+at::Tensor ms_deform_attn_cuda_forward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const int im2col_step);
+
+std::vector<at::Tensor> ms_deform_attn_cuda_backward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const at::Tensor &grad_output,
+    const int im2col_step);
+

detect_tools/upn/ops/src/cuda/ms_deform_im2col_cuda.cuh

@@ -0,0 +1,1327 @@
+/*!
+**************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************
+* Modified from DCN (https://github.com/msracver/Deformable-ConvNets)
+* Copyright (c) 2018 Microsoft
+**************************************************************************
+*/
+
+#include <cstdio>
+#include <algorithm>
+#include <cstring>
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+
+#include <THC/THCAtomics.cuh>
+
+#define CUDA_KERNEL_LOOP(i, n)                          \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x;   \
+      i < (n);                                          \
+      i += blockDim.x * gridDim.x)
+
+const int CUDA_NUM_THREADS = 1024;
+inline int GET_BLOCKS(const int N, const int num_threads)
+{
+  return (N + num_threads - 1) / num_threads;
+}
+
+
+template <typename scalar_t>
+__device__ scalar_t ms_deform_attn_im2col_bilinear(const scalar_t* &bottom_data, 
+                                                   const int &height, const int &width, const int &nheads, const int &channels,
+                                                   const scalar_t &h, const scalar_t &w, const int &m, const int &c)
+{
+  const int h_low = floor(h);
+  const int w_low = floor(w);
+  const int h_high = h_low + 1;
+  const int w_high = w_low + 1;
+
+  const scalar_t lh = h - h_low;
+  const scalar_t lw = w - w_low;
+  const scalar_t hh = 1 - lh, hw = 1 - lw;
+
+  const int w_stride = nheads * channels;
+  const int h_stride = width * w_stride;
+  const int h_low_ptr_offset = h_low * h_stride;
+  const int h_high_ptr_offset = h_low_ptr_offset + h_stride;
+  const int w_low_ptr_offset = w_low * w_stride;
+  const int w_high_ptr_offset = w_low_ptr_offset + w_stride;
+  const int base_ptr = m * channels + c;
+
+  scalar_t v1 = 0;
+  if (h_low >= 0 && w_low >= 0)
+  {
+    const int ptr1 = h_low_ptr_offset + w_low_ptr_offset + base_ptr;
+    v1 = bottom_data[ptr1];
+  }
+  scalar_t v2 = 0;
+  if (h_low >= 0 && w_high <= width - 1)
+  {
+    const int ptr2 = h_low_ptr_offset + w_high_ptr_offset + base_ptr;
+    v2 = bottom_data[ptr2];
+  }
+  scalar_t v3 = 0;
+  if (h_high <= height - 1 && w_low >= 0)
+  {
+    const int ptr3 = h_high_ptr_offset + w_low_ptr_offset + base_ptr;
+    v3 = bottom_data[ptr3];
+  }
+  scalar_t v4 = 0;
+  if (h_high <= height - 1 && w_high <= width - 1)
+  {
+    const int ptr4 = h_high_ptr_offset + w_high_ptr_offset + base_ptr;
+    v4 = bottom_data[ptr4];
+  }
+
+  const scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw;
+
+  const scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+  return val;
+}
+
+
+template <typename scalar_t>
+__device__ void ms_deform_attn_col2im_bilinear(const scalar_t* &bottom_data, 
+                                                   const int &height, const int &width, const int &nheads, const int &channels,
+                                                   const scalar_t &h, const scalar_t &w, const int &m, const int &c,
+                                                   const scalar_t &top_grad,
+                                                   const scalar_t &attn_weight,
+                                                   scalar_t* &grad_value, 
+                                                   scalar_t* grad_sampling_loc,
+                                                   scalar_t* grad_attn_weight)
+{
+  const int h_low = floor(h);
+  const int w_low = floor(w);
+  const int h_high = h_low + 1;
+  const int w_high = w_low + 1;
+
+  const scalar_t lh = h - h_low;
+  const scalar_t lw = w - w_low;
+  const scalar_t hh = 1 - lh, hw = 1 - lw;
+
+  const int w_stride = nheads * channels;
+  const int h_stride = width * w_stride;
+  const int h_low_ptr_offset = h_low * h_stride;
+  const int h_high_ptr_offset = h_low_ptr_offset + h_stride;
+  const int w_low_ptr_offset = w_low * w_stride;
+  const int w_high_ptr_offset = w_low_ptr_offset + w_stride;
+  const int base_ptr = m * channels + c;
+
+  const scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw;
+  const scalar_t top_grad_value = top_grad * attn_weight;
+  scalar_t grad_h_weight = 0, grad_w_weight = 0;
+
+  scalar_t v1 = 0;
+  if (h_low >= 0 && w_low >= 0)
+  {
+    const int ptr1 = h_low_ptr_offset + w_low_ptr_offset + base_ptr;
+    v1 = bottom_data[ptr1];
+    grad_h_weight -= hw * v1;
+    grad_w_weight -= hh * v1;
+    atomicAdd(grad_value+ptr1, w1*top_grad_value);
+  }
+  scalar_t v2 = 0;
+  if (h_low >= 0 && w_high <= width - 1)
+  {
+    const int ptr2 = h_low_ptr_offset + w_high_ptr_offset + base_ptr;
+    v2 = bottom_data[ptr2];
+    grad_h_weight -= lw * v2;
+    grad_w_weight += hh * v2;
+    atomicAdd(grad_value+ptr2, w2*top_grad_value);
+  }
+  scalar_t v3 = 0;
+  if (h_high <= height - 1 && w_low >= 0)
+  {
+    const int ptr3 = h_high_ptr_offset + w_low_ptr_offset + base_ptr;
+    v3 = bottom_data[ptr3];
+    grad_h_weight += hw * v3;
+    grad_w_weight -= lh * v3;
+    atomicAdd(grad_value+ptr3, w3*top_grad_value); 
+  }
+  scalar_t v4 = 0;
+  if (h_high <= height - 1 && w_high <= width - 1)
+  {
+    const int ptr4 = h_high_ptr_offset + w_high_ptr_offset + base_ptr;
+    v4 = bottom_data[ptr4];
+    grad_h_weight += lw * v4;
+    grad_w_weight += lh * v4;
+    atomicAdd(grad_value+ptr4, w4*top_grad_value);
+  }
+
+  const scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+  *grad_attn_weight = top_grad * val;
+  *grad_sampling_loc = width * grad_w_weight * top_grad_value;
+  *(grad_sampling_loc + 1) = height * grad_h_weight * top_grad_value;
+}
+
+
+template <typename scalar_t>
+__device__ void ms_deform_attn_col2im_bilinear_gm(const scalar_t* &bottom_data, 
+                                                   const int &height, const int &width, const int &nheads, const int &channels,
+                                                   const scalar_t &h, const scalar_t &w, const int &m, const int &c,
+                                                   const scalar_t &top_grad,
+                                                   const scalar_t &attn_weight,
+                                                   scalar_t* &grad_value, 
+                                                   scalar_t* grad_sampling_loc,
+                                                   scalar_t* grad_attn_weight)
+{
+  const int h_low = floor(h);
+  const int w_low = floor(w);
+  const int h_high = h_low + 1;
+  const int w_high = w_low + 1;
+
+  const scalar_t lh = h - h_low;
+  const scalar_t lw = w - w_low;
+  const scalar_t hh = 1 - lh, hw = 1 - lw;
+
+  const int w_stride = nheads * channels;
+  const int h_stride = width * w_stride;
+  const int h_low_ptr_offset = h_low * h_stride;
+  const int h_high_ptr_offset = h_low_ptr_offset + h_stride;
+  const int w_low_ptr_offset = w_low * w_stride;
+  const int w_high_ptr_offset = w_low_ptr_offset + w_stride;
+  const int base_ptr = m * channels + c;
+
+  const scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw;
+  const scalar_t top_grad_value = top_grad * attn_weight;
+  scalar_t grad_h_weight = 0, grad_w_weight = 0;
+
+  scalar_t v1 = 0;
+  if (h_low >= 0 && w_low >= 0)
+  {
+    const int ptr1 = h_low_ptr_offset + w_low_ptr_offset + base_ptr;
+    v1 = bottom_data[ptr1];
+    grad_h_weight -= hw * v1;
+    grad_w_weight -= hh * v1;
+    atomicAdd(grad_value+ptr1, w1*top_grad_value);
+  }
+  scalar_t v2 = 0;
+  if (h_low >= 0 && w_high <= width - 1)
+  {
+    const int ptr2 = h_low_ptr_offset + w_high_ptr_offset + base_ptr;
+    v2 = bottom_data[ptr2];
+    grad_h_weight -= lw * v2;
+    grad_w_weight += hh * v2;
+    atomicAdd(grad_value+ptr2, w2*top_grad_value);
+  }
+  scalar_t v3 = 0;
+  if (h_high <= height - 1 && w_low >= 0)
+  {
+    const int ptr3 = h_high_ptr_offset + w_low_ptr_offset + base_ptr;
+    v3 = bottom_data[ptr3];
+    grad_h_weight += hw * v3;
+    grad_w_weight -= lh * v3;
+    atomicAdd(grad_value+ptr3, w3*top_grad_value); 
+  }
+  scalar_t v4 = 0;
+  if (h_high <= height - 1 && w_high <= width - 1)
+  {
+    const int ptr4 = h_high_ptr_offset + w_high_ptr_offset + base_ptr;
+    v4 = bottom_data[ptr4];
+    grad_h_weight += lw * v4;
+    grad_w_weight += lh * v4;
+    atomicAdd(grad_value+ptr4, w4*top_grad_value);
+  }
+
+  const scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+  atomicAdd(grad_attn_weight, top_grad * val); 
+  atomicAdd(grad_sampling_loc, width * grad_w_weight * top_grad_value);
+  atomicAdd(grad_sampling_loc + 1, height * grad_h_weight * top_grad_value);
+}
+
+
+template <typename scalar_t>
+__global__ void ms_deformable_im2col_gpu_kernel(const int n,
+                                                const scalar_t *data_value, 
+                                                const int64_t *data_spatial_shapes,
+                                                const int64_t *data_level_start_index, 
+                                                const scalar_t *data_sampling_loc,
+                                                const scalar_t *data_attn_weight,
+                                                const int batch_size, 
+                                                const int spatial_size, 
+                                                const int num_heads,
+                                                const int channels, 
+                                                const int num_levels,
+                                                const int num_query,
+                                                const int num_point,
+                                                scalar_t *data_col)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    int _temp = index;
+    const int c_col = _temp % channels;
+    _temp /= channels;
+    const int sampling_index = _temp; 
+    const int m_col = _temp % num_heads;
+    _temp /= num_heads;
+    const int q_col = _temp % num_query;
+    _temp /= num_query;
+    const int b_col = _temp;
+
+    scalar_t *data_col_ptr = data_col + index;
+    int data_weight_ptr = sampling_index * num_levels * num_point;
+    int data_loc_w_ptr = data_weight_ptr << 1;
+    const int qid_stride = num_heads * channels;
+    const int data_value_ptr_init_offset = b_col * spatial_size * qid_stride;
+    scalar_t col = 0;
+    
+    for (int l_col=0; l_col < num_levels; ++l_col)
+    {
+      const int level_start_id = data_level_start_index[l_col];
+      const int spatial_h_ptr = l_col << 1;
+      const int spatial_h = data_spatial_shapes[spatial_h_ptr];
+      const int spatial_w = data_spatial_shapes[spatial_h_ptr + 1];
+      const scalar_t *data_value_ptr = data_value + (data_value_ptr_init_offset + level_start_id * qid_stride);
+      for (int p_col=0; p_col < num_point; ++p_col)
+      {
+        const scalar_t loc_w = data_sampling_loc[data_loc_w_ptr];
+        const scalar_t loc_h = data_sampling_loc[data_loc_w_ptr + 1];
+        const scalar_t weight = data_attn_weight[data_weight_ptr];
+
+        const scalar_t h_im = loc_h * spatial_h - 0.5;
+        const scalar_t w_im = loc_w * spatial_w - 0.5;
+
+        if (h_im > -1 && w_im > -1 && h_im < spatial_h && w_im < spatial_w)
+        {
+          col += ms_deform_attn_im2col_bilinear(data_value_ptr, spatial_h, spatial_w, num_heads, channels, h_im, w_im, m_col, c_col) * weight;
+        }
+
+        data_weight_ptr += 1;
+        data_loc_w_ptr += 2;
+      }
+    }
+    *data_col_ptr = col;
+  }
+}
+
+template <typename scalar_t, unsigned int blockSize>
+__global__ void ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v1(const int n,
+                                                const scalar_t *grad_col,
+                                                const scalar_t *data_value,
+                                                const int64_t *data_spatial_shapes,
+                                                const int64_t *data_level_start_index, 
+                                                const scalar_t *data_sampling_loc,
+                                                const scalar_t *data_attn_weight,
+                                                const int batch_size, 
+                                                const int spatial_size, 
+                                                const int num_heads,
+                                                const int channels, 
+                                                const int num_levels,
+                                                const int num_query,
+                                                const int num_point,
+                                                scalar_t *grad_value,
+                                                scalar_t *grad_sampling_loc,
+                                                scalar_t *grad_attn_weight)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    __shared__ scalar_t cache_grad_sampling_loc[blockSize * 2];
+    __shared__ scalar_t cache_grad_attn_weight[blockSize];
+    unsigned int tid = threadIdx.x;
+    int _temp = index;
+    const int c_col = _temp % channels;
+    _temp /= channels;
+    const int sampling_index = _temp; 
+    const int m_col = _temp % num_heads;
+    _temp /= num_heads;
+    const int q_col = _temp % num_query;
+    _temp /= num_query;
+    const int b_col = _temp;
+
+    const scalar_t top_grad = grad_col[index];
+
+    int data_weight_ptr = sampling_index * num_levels * num_point;
+    int data_loc_w_ptr = data_weight_ptr << 1;
+    const int grad_sampling_ptr = data_weight_ptr;
+    grad_sampling_loc += grad_sampling_ptr << 1;
+    grad_attn_weight += grad_sampling_ptr;
+    const int grad_weight_stride = 1;
+    const int grad_loc_stride = 2;
+    const int qid_stride = num_heads * channels;
+    const int data_value_ptr_init_offset = b_col * spatial_size * qid_stride;
+
+    for (int l_col=0; l_col < num_levels; ++l_col)
+    {
+      const int level_start_id = data_level_start_index[l_col];
+      const int spatial_h_ptr = l_col << 1;
+      const int spatial_h = data_spatial_shapes[spatial_h_ptr];
+      const int spatial_w = data_spatial_shapes[spatial_h_ptr + 1];
+      const int value_ptr_offset = data_value_ptr_init_offset + level_start_id * qid_stride;
+      const scalar_t *data_value_ptr = data_value + value_ptr_offset;
+      scalar_t *grad_value_ptr = grad_value + value_ptr_offset;
+
+      for (int p_col=0; p_col < num_point; ++p_col)
+      {
+        const scalar_t loc_w = data_sampling_loc[data_loc_w_ptr];
+        const scalar_t loc_h = data_sampling_loc[data_loc_w_ptr + 1];
+        const scalar_t weight = data_attn_weight[data_weight_ptr];
+
+        const scalar_t h_im = loc_h * spatial_h - 0.5;
+        const scalar_t w_im = loc_w * spatial_w - 0.5;
+        *(cache_grad_sampling_loc+(threadIdx.x << 1)) = 0;
+        *(cache_grad_sampling_loc+((threadIdx.x << 1) + 1)) = 0;
+        *(cache_grad_attn_weight+threadIdx.x)=0;
+        if (h_im > -1 && w_im > -1 && h_im < spatial_h && w_im < spatial_w)
+        {
+          ms_deform_attn_col2im_bilinear(
+            data_value_ptr, spatial_h, spatial_w, num_heads, channels, h_im, w_im, m_col, c_col,
+            top_grad, weight, grad_value_ptr, 
+            cache_grad_sampling_loc+(threadIdx.x << 1), cache_grad_attn_weight+threadIdx.x);
+        }
+        
+        __syncthreads();
+        if (tid == 0)
+        {
+          scalar_t _grad_w=cache_grad_sampling_loc[0], _grad_h=cache_grad_sampling_loc[1], _grad_a=cache_grad_attn_weight[0];
+          int sid=2;
+          for (unsigned int tid = 1; tid < blockSize; ++tid)
+          {
+            _grad_w += cache_grad_sampling_loc[sid];
+            _grad_h += cache_grad_sampling_loc[sid + 1];
+            _grad_a += cache_grad_attn_weight[tid];
+            sid += 2;
+          }
+          
+          
+          *grad_sampling_loc = _grad_w;
+          *(grad_sampling_loc + 1) = _grad_h;
+          *grad_attn_weight = _grad_a;
+        }
+        __syncthreads();
+
+        data_weight_ptr += 1;
+        data_loc_w_ptr += 2;
+        grad_attn_weight += grad_weight_stride;
+        grad_sampling_loc += grad_loc_stride;
+      }
+    }
+  }
+}
+
+
+template <typename scalar_t, unsigned int blockSize>
+__global__ void ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v2(const int n,
+                                                const scalar_t *grad_col,
+                                                const scalar_t *data_value,
+                                                const int64_t *data_spatial_shapes,
+                                                const int64_t *data_level_start_index, 
+                                                const scalar_t *data_sampling_loc,
+                                                const scalar_t *data_attn_weight,
+                                                const int batch_size, 
+                                                const int spatial_size, 
+                                                const int num_heads,
+                                                const int channels, 
+                                                const int num_levels,
+                                                const int num_query,
+                                                const int num_point,
+                                                scalar_t *grad_value,
+                                                scalar_t *grad_sampling_loc,
+                                                scalar_t *grad_attn_weight)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    __shared__ scalar_t cache_grad_sampling_loc[blockSize * 2];
+    __shared__ scalar_t cache_grad_attn_weight[blockSize];
+    unsigned int tid = threadIdx.x;
+    int _temp = index;
+    const int c_col = _temp % channels;
+    _temp /= channels;
+    const int sampling_index = _temp; 
+    const int m_col = _temp % num_heads;
+    _temp /= num_heads;
+    const int q_col = _temp % num_query;
+    _temp /= num_query;
+    const int b_col = _temp;
+
+    const scalar_t top_grad = grad_col[index];
+
+    int data_weight_ptr = sampling_index * num_levels * num_point;
+    int data_loc_w_ptr = data_weight_ptr << 1;
+    const int grad_sampling_ptr = data_weight_ptr;
+    grad_sampling_loc += grad_sampling_ptr << 1;
+    grad_attn_weight += grad_sampling_ptr;
+    const int grad_weight_stride = 1;
+    const int grad_loc_stride = 2;
+    const int qid_stride = num_heads * channels;
+    const int data_value_ptr_init_offset = b_col * spatial_size * qid_stride;
+
+    for (int l_col=0; l_col < num_levels; ++l_col)
+    {
+      const int level_start_id = data_level_start_index[l_col];
+      const int spatial_h_ptr = l_col << 1;
+      const int spatial_h = data_spatial_shapes[spatial_h_ptr];
+      const int spatial_w = data_spatial_shapes[spatial_h_ptr + 1];
+      const int value_ptr_offset = data_value_ptr_init_offset + level_start_id * qid_stride;
+      const scalar_t *data_value_ptr = data_value + value_ptr_offset;
+      scalar_t *grad_value_ptr = grad_value + value_ptr_offset;
+
+      for (int p_col=0; p_col < num_point; ++p_col)
+      {
+        const scalar_t loc_w = data_sampling_loc[data_loc_w_ptr];
+        const scalar_t loc_h = data_sampling_loc[data_loc_w_ptr + 1];
+        const scalar_t weight = data_attn_weight[data_weight_ptr];
+
+        const scalar_t h_im = loc_h * spatial_h - 0.5;
+        const scalar_t w_im = loc_w * spatial_w - 0.5;
+        *(cache_grad_sampling_loc+(threadIdx.x << 1)) = 0;
+        *(cache_grad_sampling_loc+((threadIdx.x << 1) + 1)) = 0;
+        *(cache_grad_attn_weight+threadIdx.x)=0;
+        if (h_im > -1 && w_im > -1 && h_im < spatial_h && w_im < spatial_w)
+        {
+          ms_deform_attn_col2im_bilinear(
+            data_value_ptr, spatial_h, spatial_w, num_heads, channels, h_im, w_im, m_col, c_col,
+            top_grad, weight, grad_value_ptr, 
+            cache_grad_sampling_loc+(threadIdx.x << 1), cache_grad_attn_weight+threadIdx.x);
+        }
+        
+        __syncthreads();
+
+        for (unsigned int s=blockSize/2; s>0; s>>=1)
+        {
+          if (tid < s) {
+            const unsigned int xid1 = tid << 1;
+            const unsigned int xid2 = (tid + s) << 1;
+            cache_grad_attn_weight[tid] += cache_grad_attn_weight[tid + s];
+            cache_grad_sampling_loc[xid1] += cache_grad_sampling_loc[xid2];
+            cache_grad_sampling_loc[xid1 + 1] += cache_grad_sampling_loc[xid2 + 1];
+          }
+          __syncthreads();
+        }
+
+        if (tid == 0)
+        { 
+          *grad_sampling_loc = cache_grad_sampling_loc[0];
+          *(grad_sampling_loc + 1) = cache_grad_sampling_loc[1];
+          *grad_attn_weight = cache_grad_attn_weight[0];
+        }
+        __syncthreads();
+
+        data_weight_ptr += 1;
+        data_loc_w_ptr += 2;
+        grad_attn_weight += grad_weight_stride;
+        grad_sampling_loc += grad_loc_stride;
+      }
+    }
+  }
+}
+
+
+template <typename scalar_t>
+__global__ void ms_deformable_col2im_gpu_kernel_shm_reduce_v1(const int n,
+                                                const scalar_t *grad_col,
+                                                const scalar_t *data_value,
+                                                const int64_t *data_spatial_shapes,
+                                                const int64_t *data_level_start_index, 
+                                                const scalar_t *data_sampling_loc,
+                                                const scalar_t *data_attn_weight,
+                                                const int batch_size, 
+                                                const int spatial_size, 
+                                                const int num_heads,
+                                                const int channels, 
+                                                const int num_levels,
+                                                const int num_query,
+                                                const int num_point,
+                                                scalar_t *grad_value,
+                                                scalar_t *grad_sampling_loc,
+                                                scalar_t *grad_attn_weight)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    extern __shared__ int _s[];
+    scalar_t* cache_grad_sampling_loc = (scalar_t*)_s;
+    scalar_t* cache_grad_attn_weight = cache_grad_sampling_loc + 2 * blockDim.x;
+    unsigned int tid = threadIdx.x;
+    int _temp = index;
+    const int c_col = _temp % channels;
+    _temp /= channels;
+    const int sampling_index = _temp; 
+    const int m_col = _temp % num_heads;
+    _temp /= num_heads;
+    const int q_col = _temp % num_query;
+    _temp /= num_query;
+    const int b_col = _temp;
+
+    const scalar_t top_grad = grad_col[index];
+
+    int data_weight_ptr = sampling_index * num_levels * num_point;
+    int data_loc_w_ptr = data_weight_ptr << 1;
+    const int grad_sampling_ptr = data_weight_ptr;
+    grad_sampling_loc += grad_sampling_ptr << 1;
+    grad_attn_weight += grad_sampling_ptr;
+    const int grad_weight_stride = 1;
+    const int grad_loc_stride = 2;
+    const int qid_stride = num_heads * channels;
+    const int data_value_ptr_init_offset = b_col * spatial_size * qid_stride;
+
+    for (int l_col=0; l_col < num_levels; ++l_col)
+    {
+      const int level_start_id = data_level_start_index[l_col];
+      const int spatial_h_ptr = l_col << 1;
+      const int spatial_h = data_spatial_shapes[spatial_h_ptr];
+      const int spatial_w = data_spatial_shapes[spatial_h_ptr + 1];
+      const int value_ptr_offset = data_value_ptr_init_offset + level_start_id * qid_stride;
+      const scalar_t *data_value_ptr = data_value + value_ptr_offset;
+      scalar_t *grad_value_ptr = grad_value + value_ptr_offset;
+
+      for (int p_col=0; p_col < num_point; ++p_col)
+      {
+        const scalar_t loc_w = data_sampling_loc[data_loc_w_ptr];
+        const scalar_t loc_h = data_sampling_loc[data_loc_w_ptr + 1];
+        const scalar_t weight = data_attn_weight[data_weight_ptr];
+
+        const scalar_t h_im = loc_h * spatial_h - 0.5;
+        const scalar_t w_im = loc_w * spatial_w - 0.5;
+        *(cache_grad_sampling_loc+(threadIdx.x << 1)) = 0;
+        *(cache_grad_sampling_loc+((threadIdx.x << 1) + 1)) = 0;
+        *(cache_grad_attn_weight+threadIdx.x)=0;
+        if (h_im > -1 && w_im > -1 && h_im < spatial_h && w_im < spatial_w)
+        {
+          ms_deform_attn_col2im_bilinear(
+            data_value_ptr, spatial_h, spatial_w, num_heads, channels, h_im, w_im, m_col, c_col,
+            top_grad, weight, grad_value_ptr, 
+            cache_grad_sampling_loc+(threadIdx.x << 1), cache_grad_attn_weight+threadIdx.x);
+        }
+        
+        __syncthreads();
+        if (tid == 0)
+        {
+          scalar_t _grad_w=cache_grad_sampling_loc[0], _grad_h=cache_grad_sampling_loc[1], _grad_a=cache_grad_attn_weight[0];
+          int sid=2;
+          for (unsigned int tid = 1; tid < blockDim.x; ++tid)
+          {
+            _grad_w += cache_grad_sampling_loc[sid];
+            _grad_h += cache_grad_sampling_loc[sid + 1];
+            _grad_a += cache_grad_attn_weight[tid];
+            sid += 2;
+          }
+          
+          
+          *grad_sampling_loc = _grad_w;
+          *(grad_sampling_loc + 1) = _grad_h;
+          *grad_attn_weight = _grad_a;
+        }
+        __syncthreads();
+
+        data_weight_ptr += 1;
+        data_loc_w_ptr += 2;
+        grad_attn_weight += grad_weight_stride;
+        grad_sampling_loc += grad_loc_stride;
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+__global__ void ms_deformable_col2im_gpu_kernel_shm_reduce_v2(const int n,
+                                                const scalar_t *grad_col,
+                                                const scalar_t *data_value,
+                                                const int64_t *data_spatial_shapes,
+                                                const int64_t *data_level_start_index, 
+                                                const scalar_t *data_sampling_loc,
+                                                const scalar_t *data_attn_weight,
+                                                const int batch_size, 
+                                                const int spatial_size, 
+                                                const int num_heads,
+                                                const int channels, 
+                                                const int num_levels,
+                                                const int num_query,
+                                                const int num_point,
+                                                scalar_t *grad_value,
+                                                scalar_t *grad_sampling_loc,
+                                                scalar_t *grad_attn_weight)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    extern __shared__ int _s[];
+    scalar_t* cache_grad_sampling_loc = (scalar_t*)_s;
+    scalar_t* cache_grad_attn_weight = cache_grad_sampling_loc + 2 * blockDim.x;
+    unsigned int tid = threadIdx.x;
+    int _temp = index;
+    const int c_col = _temp % channels;
+    _temp /= channels;
+    const int sampling_index = _temp; 
+    const int m_col = _temp % num_heads;
+    _temp /= num_heads;
+    const int q_col = _temp % num_query;
+    _temp /= num_query;
+    const int b_col = _temp;
+
+    const scalar_t top_grad = grad_col[index];
+
+    int data_weight_ptr = sampling_index * num_levels * num_point;
+    int data_loc_w_ptr = data_weight_ptr << 1;
+    const int grad_sampling_ptr = data_weight_ptr;
+    grad_sampling_loc += grad_sampling_ptr << 1;
+    grad_attn_weight += grad_sampling_ptr;
+    const int grad_weight_stride = 1;
+    const int grad_loc_stride = 2;
+    const int qid_stride = num_heads * channels;
+    const int data_value_ptr_init_offset = b_col * spatial_size * qid_stride;
+
+    for (int l_col=0; l_col < num_levels; ++l_col)
+    {
+      const int level_start_id = data_level_start_index[l_col];
+      const int spatial_h_ptr = l_col << 1;
+      const int spatial_h = data_spatial_shapes[spatial_h_ptr];
+      const int spatial_w = data_spatial_shapes[spatial_h_ptr + 1];
+      const int value_ptr_offset = data_value_ptr_init_offset + level_start_id * qid_stride;
+      const scalar_t *data_value_ptr = data_value + value_ptr_offset;
+      scalar_t *grad_value_ptr = grad_value + value_ptr_offset;
+
+      for (int p_col=0; p_col < num_point; ++p_col)
+      {
+        const scalar_t loc_w = data_sampling_loc[data_loc_w_ptr];
+        const scalar_t loc_h = data_sampling_loc[data_loc_w_ptr + 1];
+        const scalar_t weight = data_attn_weight[data_weight_ptr];
+
+        const scalar_t h_im = loc_h * spatial_h - 0.5;
+        const scalar_t w_im = loc_w * spatial_w - 0.5;
+        *(cache_grad_sampling_loc+(threadIdx.x << 1)) = 0;
+        *(cache_grad_sampling_loc+((threadIdx.x << 1) + 1)) = 0;
+        *(cache_grad_attn_weight+threadIdx.x)=0;
+        if (h_im > -1 && w_im > -1 && h_im < spatial_h && w_im < spatial_w)
+        {
+          ms_deform_attn_col2im_bilinear(
+            data_value_ptr, spatial_h, spatial_w, num_heads, channels, h_im, w_im, m_col, c_col,
+            top_grad, weight, grad_value_ptr, 
+            cache_grad_sampling_loc+(threadIdx.x << 1), cache_grad_attn_weight+threadIdx.x);
+        }
+        
+        __syncthreads();
+
+        for (unsigned int s=blockDim.x/2, spre=blockDim.x; s>0; s>>=1, spre>>=1)
+        {
+          if (tid < s) {
+            const unsigned int xid1 = tid << 1;
+            const unsigned int xid2 = (tid + s) << 1;
+            cache_grad_attn_weight[tid] += cache_grad_attn_weight[tid + s];
+            cache_grad_sampling_loc[xid1] += cache_grad_sampling_loc[xid2];
+            cache_grad_sampling_loc[xid1 + 1] += cache_grad_sampling_loc[xid2 + 1];
+            if (tid + (s << 1) < spre)
+            {
+              cache_grad_attn_weight[tid] += cache_grad_attn_weight[tid + (s << 1)];
+              cache_grad_sampling_loc[xid1] += cache_grad_sampling_loc[xid2 + (s << 1)];
+              cache_grad_sampling_loc[xid1 + 1] += cache_grad_sampling_loc[xid2 + 1 + (s << 1)];
+            } 
+          }
+          __syncthreads();
+        }
+
+        if (tid == 0)
+        {
+          *grad_sampling_loc = cache_grad_sampling_loc[0];
+          *(grad_sampling_loc + 1) = cache_grad_sampling_loc[1];
+          *grad_attn_weight = cache_grad_attn_weight[0];
+        }
+        __syncthreads();
+
+        data_weight_ptr += 1;
+        data_loc_w_ptr += 2;
+        grad_attn_weight += grad_weight_stride;
+        grad_sampling_loc += grad_loc_stride;
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+__global__ void ms_deformable_col2im_gpu_kernel_shm_reduce_v2_multi_blocks(const int n,
+                                                const scalar_t *grad_col,
+                                                const scalar_t *data_value,
+                                                const int64_t *data_spatial_shapes,
+                                                const int64_t *data_level_start_index, 
+                                                const scalar_t *data_sampling_loc,
+                                                const scalar_t *data_attn_weight,
+                                                const int batch_size, 
+                                                const int spatial_size, 
+                                                const int num_heads,
+                                                const int channels, 
+                                                const int num_levels,
+                                                const int num_query,
+                                                const int num_point,
+                                                scalar_t *grad_value,
+                                                scalar_t *grad_sampling_loc,
+                                                scalar_t *grad_attn_weight)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    extern __shared__ int _s[];
+    scalar_t* cache_grad_sampling_loc = (scalar_t*)_s;
+    scalar_t* cache_grad_attn_weight = cache_grad_sampling_loc + 2 * blockDim.x;
+    unsigned int tid = threadIdx.x;
+    int _temp = index;
+    const int c_col = _temp % channels;
+    _temp /= channels;
+    const int sampling_index = _temp; 
+    const int m_col = _temp % num_heads;
+    _temp /= num_heads;
+    const int q_col = _temp % num_query;
+    _temp /= num_query;
+    const int b_col = _temp;
+
+    const scalar_t top_grad = grad_col[index];
+
+    int data_weight_ptr = sampling_index * num_levels * num_point;
+    int data_loc_w_ptr = data_weight_ptr << 1;
+    const int grad_sampling_ptr = data_weight_ptr;
+    grad_sampling_loc += grad_sampling_ptr << 1;
+    grad_attn_weight += grad_sampling_ptr;
+    const int grad_weight_stride = 1;
+    const int grad_loc_stride = 2;
+    const int qid_stride = num_heads * channels;
+    const int data_value_ptr_init_offset = b_col * spatial_size * qid_stride;
+
+    for (int l_col=0; l_col < num_levels; ++l_col)
+    {
+      const int level_start_id = data_level_start_index[l_col];
+      const int spatial_h_ptr = l_col << 1;
+      const int spatial_h = data_spatial_shapes[spatial_h_ptr];
+      const int spatial_w = data_spatial_shapes[spatial_h_ptr + 1];
+      const int value_ptr_offset = data_value_ptr_init_offset + level_start_id * qid_stride;
+      const scalar_t *data_value_ptr = data_value + value_ptr_offset;
+      scalar_t *grad_value_ptr = grad_value + value_ptr_offset;
+
+      for (int p_col=0; p_col < num_point; ++p_col)
+      {
+        const scalar_t loc_w = data_sampling_loc[data_loc_w_ptr];
+        const scalar_t loc_h = data_sampling_loc[data_loc_w_ptr + 1];
+        const scalar_t weight = data_attn_weight[data_weight_ptr];
+
+        const scalar_t h_im = loc_h * spatial_h - 0.5;
+        const scalar_t w_im = loc_w * spatial_w - 0.5;
+        *(cache_grad_sampling_loc+(threadIdx.x << 1)) = 0;
+        *(cache_grad_sampling_loc+((threadIdx.x << 1) + 1)) = 0;
+        *(cache_grad_attn_weight+threadIdx.x)=0;
+        if (h_im > -1 && w_im > -1 && h_im < spatial_h && w_im < spatial_w)
+        {
+          ms_deform_attn_col2im_bilinear(
+            data_value_ptr, spatial_h, spatial_w, num_heads, channels, h_im, w_im, m_col, c_col,
+            top_grad, weight, grad_value_ptr, 
+            cache_grad_sampling_loc+(threadIdx.x << 1), cache_grad_attn_weight+threadIdx.x);
+        }
+        
+        __syncthreads();
+
+        for (unsigned int s=blockDim.x/2, spre=blockDim.x; s>0; s>>=1, spre>>=1)
+        {
+          if (tid < s) {
+            const unsigned int xid1 = tid << 1;
+            const unsigned int xid2 = (tid + s) << 1;
+            cache_grad_attn_weight[tid] += cache_grad_attn_weight[tid + s];
+            cache_grad_sampling_loc[xid1] += cache_grad_sampling_loc[xid2];
+            cache_grad_sampling_loc[xid1 + 1] += cache_grad_sampling_loc[xid2 + 1];
+            if (tid + (s << 1) < spre)
+            {
+              cache_grad_attn_weight[tid] += cache_grad_attn_weight[tid + (s << 1)];
+              cache_grad_sampling_loc[xid1] += cache_grad_sampling_loc[xid2 + (s << 1)];
+              cache_grad_sampling_loc[xid1 + 1] += cache_grad_sampling_loc[xid2 + 1 + (s << 1)];
+            }
+          }
+          __syncthreads();
+        }
+
+        if (tid == 0)
+        {
+          atomicAdd(grad_sampling_loc, cache_grad_sampling_loc[0]);
+          atomicAdd(grad_sampling_loc + 1, cache_grad_sampling_loc[1]);
+          atomicAdd(grad_attn_weight, cache_grad_attn_weight[0]);
+        }
+        __syncthreads();
+
+        data_weight_ptr += 1;
+        data_loc_w_ptr += 2;
+        grad_attn_weight += grad_weight_stride;
+        grad_sampling_loc += grad_loc_stride;
+      }
+    }
+  }
+}
+
+
+template <typename scalar_t>
+__global__ void ms_deformable_col2im_gpu_kernel_gm(const int n,
+                                                const scalar_t *grad_col,
+                                                const scalar_t *data_value,
+                                                const int64_t *data_spatial_shapes,
+                                                const int64_t *data_level_start_index, 
+                                                const scalar_t *data_sampling_loc,
+                                                const scalar_t *data_attn_weight,
+                                                const int batch_size, 
+                                                const int spatial_size, 
+                                                const int num_heads,
+                                                const int channels, 
+                                                const int num_levels,
+                                                const int num_query,
+                                                const int num_point,
+                                                scalar_t *grad_value,
+                                                scalar_t *grad_sampling_loc,
+                                                scalar_t *grad_attn_weight)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    int _temp = index;
+    const int c_col = _temp % channels;
+    _temp /= channels;
+    const int sampling_index = _temp; 
+    const int m_col = _temp % num_heads;
+    _temp /= num_heads;
+    const int q_col = _temp % num_query;
+    _temp /= num_query;
+    const int b_col = _temp;
+
+    const scalar_t top_grad = grad_col[index];
+
+    int data_weight_ptr = sampling_index * num_levels * num_point;
+    int data_loc_w_ptr = data_weight_ptr << 1;
+    const int grad_sampling_ptr = data_weight_ptr;
+    grad_sampling_loc += grad_sampling_ptr << 1;
+    grad_attn_weight += grad_sampling_ptr;
+    const int grad_weight_stride = 1;
+    const int grad_loc_stride = 2;
+    const int qid_stride = num_heads * channels;
+    const int data_value_ptr_init_offset = b_col * spatial_size * qid_stride;
+
+    for (int l_col=0; l_col < num_levels; ++l_col)
+    {
+      const int level_start_id = data_level_start_index[l_col];
+      const int spatial_h_ptr = l_col << 1;
+      const int spatial_h = data_spatial_shapes[spatial_h_ptr];
+      const int spatial_w = data_spatial_shapes[spatial_h_ptr + 1];
+      const int value_ptr_offset = data_value_ptr_init_offset + level_start_id * qid_stride;
+      const scalar_t *data_value_ptr = data_value + value_ptr_offset;
+      scalar_t *grad_value_ptr = grad_value + value_ptr_offset;
+
+      for (int p_col=0; p_col < num_point; ++p_col)
+      {
+        const scalar_t loc_w = data_sampling_loc[data_loc_w_ptr];
+        const scalar_t loc_h = data_sampling_loc[data_loc_w_ptr + 1];
+        const scalar_t weight = data_attn_weight[data_weight_ptr];
+
+        const scalar_t h_im = loc_h * spatial_h - 0.5;
+        const scalar_t w_im = loc_w * spatial_w - 0.5;
+        if (h_im > -1 && w_im > -1 && h_im < spatial_h && w_im < spatial_w)
+        {
+          ms_deform_attn_col2im_bilinear_gm(
+            data_value_ptr, spatial_h, spatial_w, num_heads, channels, h_im, w_im, m_col, c_col,
+            top_grad, weight, grad_value_ptr, 
+            grad_sampling_loc, grad_attn_weight);
+        }
+        data_weight_ptr += 1;
+        data_loc_w_ptr += 2;
+        grad_attn_weight += grad_weight_stride;
+        grad_sampling_loc += grad_loc_stride;
+      }
+    }
+  }
+}
+
+
+template <typename scalar_t>
+void ms_deformable_im2col_cuda(cudaStream_t stream,
+                              const scalar_t* data_value,
+                              const int64_t* data_spatial_shapes, 
+                              const int64_t* data_level_start_index, 
+                              const scalar_t* data_sampling_loc,
+                              const scalar_t* data_attn_weight,
+                              const int batch_size,
+                              const int spatial_size, 
+                              const int num_heads, 
+                              const int channels, 
+                              const int num_levels, 
+                              const int num_query,
+                              const int num_point,
+                              scalar_t* data_col)
+{
+  const int num_kernels = batch_size * num_query * num_heads * channels;
+  const int num_actual_kernels = batch_size * num_query * num_heads * channels;
+  const int num_threads = CUDA_NUM_THREADS;
+  ms_deformable_im2col_gpu_kernel<scalar_t>
+      <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+          0, stream>>>(
+      num_kernels, data_value, data_spatial_shapes, data_level_start_index, data_sampling_loc, data_attn_weight, 
+      batch_size, spatial_size, num_heads, channels, num_levels, num_query, num_point, data_col);
+  
+  cudaError_t err = cudaGetLastError();
+  if (err != cudaSuccess)
+  {
+    printf("error in ms_deformable_im2col_cuda: %s\n", cudaGetErrorString(err));
+  }
+
+}
+
+template <typename scalar_t>
+void ms_deformable_col2im_cuda(cudaStream_t stream,
+                              const scalar_t* grad_col,
+                              const scalar_t* data_value,
+                              const int64_t * data_spatial_shapes,
+                              const int64_t * data_level_start_index,
+                              const scalar_t * data_sampling_loc,
+                              const scalar_t * data_attn_weight,
+                              const int batch_size, 
+                              const int spatial_size, 
+                              const int num_heads,
+                              const int channels, 
+                              const int num_levels,
+                              const int num_query,
+                              const int num_point, 
+                              scalar_t* grad_value,
+                              scalar_t* grad_sampling_loc,
+                              scalar_t* grad_attn_weight)
+{
+  const int num_threads = (channels > CUDA_NUM_THREADS)?CUDA_NUM_THREADS:channels;
+  const int num_kernels = batch_size * num_query * num_heads * channels;
+  const int num_actual_kernels = batch_size * num_query * num_heads * channels;
+  if (channels > 1024)
+  {
+    if ((channels & 1023) == 0)
+    {
+      ms_deformable_col2im_gpu_kernel_shm_reduce_v2_multi_blocks<scalar_t>
+          <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+              num_threads*3*sizeof(scalar_t), stream>>>(
+                        num_kernels, 
+                        grad_col,
+                        data_value,
+                        data_spatial_shapes,
+                        data_level_start_index, 
+                        data_sampling_loc,
+                        data_attn_weight,
+                        batch_size, 
+                        spatial_size, 
+                        num_heads,
+                        channels, 
+                        num_levels,
+                        num_query,
+                        num_point,
+                        grad_value,
+                        grad_sampling_loc,
+                        grad_attn_weight);
+    }
+    else
+    {
+      ms_deformable_col2im_gpu_kernel_gm<scalar_t>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+    }
+  }
+  else{
+    switch(channels)
+    {
+      case 1:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v1<scalar_t, 1>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 2:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v1<scalar_t, 2>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 4:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v1<scalar_t, 4>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 8:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v1<scalar_t, 8>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 16:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v1<scalar_t, 16>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 32:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v1<scalar_t, 32>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 64:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v2<scalar_t, 64>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 128:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v2<scalar_t, 128>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 256:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v2<scalar_t, 256>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 512:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v2<scalar_t, 512>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      case 1024:
+        ms_deformable_col2im_gpu_kernel_shm_blocksize_aware_reduce_v2<scalar_t, 1024>
+        <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+            0, stream>>>(
+                      num_kernels, 
+                      grad_col,
+                      data_value,
+                      data_spatial_shapes,
+                      data_level_start_index, 
+                      data_sampling_loc,
+                      data_attn_weight,
+                      batch_size, 
+                      spatial_size, 
+                      num_heads,
+                      channels, 
+                      num_levels,
+                      num_query,
+                      num_point,
+                      grad_value,
+                      grad_sampling_loc,
+                      grad_attn_weight);
+        break;
+      default:
+        if (channels < 64)
+        {
+          ms_deformable_col2im_gpu_kernel_shm_reduce_v1<scalar_t>
+          <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+              num_threads*3*sizeof(scalar_t), stream>>>(
+                        num_kernels, 
+                        grad_col,
+                        data_value,
+                        data_spatial_shapes,
+                        data_level_start_index, 
+                        data_sampling_loc,
+                        data_attn_weight,
+                        batch_size, 
+                        spatial_size, 
+                        num_heads,
+                        channels, 
+                        num_levels,
+                        num_query,
+                        num_point,
+                        grad_value,
+                        grad_sampling_loc,
+                        grad_attn_weight);
+        }
+        else
+        {
+          ms_deformable_col2im_gpu_kernel_shm_reduce_v2<scalar_t>
+          <<<GET_BLOCKS(num_actual_kernels, num_threads), num_threads,
+              num_threads*3*sizeof(scalar_t), stream>>>(
+                        num_kernels, 
+                        grad_col,
+                        data_value,
+                        data_spatial_shapes,
+                        data_level_start_index, 
+                        data_sampling_loc,
+                        data_attn_weight,
+                        batch_size, 
+                        spatial_size, 
+                        num_heads,
+                        channels, 
+                        num_levels,
+                        num_query,
+                        num_point,
+                        grad_value,
+                        grad_sampling_loc,
+                        grad_attn_weight);
+        }
+    }
+  }
+  cudaError_t err = cudaGetLastError();
+  if (err != cudaSuccess)
+  {
+    printf("error in ms_deformable_col2im_cuda: %s\n", cudaGetErrorString(err));
+  }
+
+}

detect_tools/upn/ops/src/ms_deform_attn.h

@@ -0,0 +1,62 @@
+/*!
+**************************************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************************************
+* Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+**************************************************************************************************
+*/
+
+#pragma once
+
+#include "cpu/ms_deform_attn_cpu.h"
+
+#ifdef WITH_CUDA
+#include "cuda/ms_deform_attn_cuda.h"
+#endif
+
+
+at::Tensor
+ms_deform_attn_forward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const int im2col_step)
+{
+    if (value.is_cuda())
+    {
+#ifdef WITH_CUDA
+        return ms_deform_attn_cuda_forward(
+            value, spatial_shapes, level_start_index, sampling_loc, attn_weight, im2col_step);
+#else
+        AT_ERROR("Not compiled with GPU support");
+#endif
+    }
+    AT_ERROR("Not implemented on the CPU");
+}
+
+std::vector<at::Tensor>
+ms_deform_attn_backward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const at::Tensor &grad_output,
+    const int im2col_step)
+{
+    if (value.is_cuda())
+    {
+#ifdef WITH_CUDA
+        return ms_deform_attn_cuda_backward(
+            value, spatial_shapes, level_start_index, sampling_loc, attn_weight, grad_output, im2col_step);
+#else
+        AT_ERROR("Not compiled with GPU support");
+#endif
+    }
+    AT_ERROR("Not implemented on the CPU");
+}
+

detect_tools/upn/ops/src/vision.cpp

@@ -0,0 +1,16 @@
+/*!
+**************************************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************************************
+* Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+**************************************************************************************************
+*/
+
+#include "ms_deform_attn.h"
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("ms_deform_attn_forward", &ms_deform_attn_forward, "ms_deform_attn_forward");
+  m.def("ms_deform_attn_backward", &ms_deform_attn_backward, "ms_deform_attn_backward");
+}

detect_tools/upn/ops/test.py

@@ -0,0 +1,89 @@
+# ------------------------------------------------------------------------------------------------
+# Deformable DETR
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------------------------------
+# Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+# ------------------------------------------------------------------------------------------------
+
+from __future__ import absolute_import
+from __future__ import print_function
+from __future__ import division
+
+import time
+import torch
+import torch.nn as nn
+from torch.autograd import gradcheck
+
+from functions.ms_deform_attn_func import MSDeformAttnFunction, ms_deform_attn_core_pytorch
+
+
+N, M, D = 1, 2, 2
+Lq, L, P = 2, 2, 2
+shapes = torch.as_tensor([(6, 4), (3, 2)], dtype=torch.long).cuda()
+level_start_index = torch.cat((shapes.new_zeros((1, )), shapes.prod(1).cumsum(0)[:-1]))
+S = sum([(H*W).item() for H, W in shapes])
+
+
+torch.manual_seed(3)
+
+
+@torch.no_grad()
+def check_forward_equal_with_pytorch_double():
+    value = torch.rand(N, S, M, D).cuda() * 0.01
+    sampling_locations = torch.rand(N, Lq, M, L, P, 2).cuda()
+    attention_weights = torch.rand(N, Lq, M, L, P).cuda() + 1e-5
+    attention_weights /= attention_weights.sum(-1, keepdim=True).sum(-2, keepdim=True)
+    im2col_step = 2
+    output_pytorch = ms_deform_attn_core_pytorch(value.double(), shapes, sampling_locations.double(), attention_weights.double()).detach().cpu()
+    output_cuda = MSDeformAttnFunction.apply(value.double(), shapes, level_start_index, sampling_locations.double(), attention_weights.double(), im2col_step).detach().cpu()
+    fwdok = torch.allclose(output_cuda, output_pytorch)
+    max_abs_err = (output_cuda - output_pytorch).abs().max()
+    max_rel_err = ((output_cuda - output_pytorch).abs() / output_pytorch.abs()).max()
+
+    print(f'* {fwdok} check_forward_equal_with_pytorch_double: max_abs_err {max_abs_err:.2e} max_rel_err {max_rel_err:.2e}')
+
+
+@torch.no_grad()
+def check_forward_equal_with_pytorch_float():
+    value = torch.rand(N, S, M, D).cuda() * 0.01
+    sampling_locations = torch.rand(N, Lq, M, L, P, 2).cuda()
+    attention_weights = torch.rand(N, Lq, M, L, P).cuda() + 1e-5
+    attention_weights /= attention_weights.sum(-1, keepdim=True).sum(-2, keepdim=True)
+    im2col_step = 2
+    output_pytorch = ms_deform_attn_core_pytorch(value, shapes, sampling_locations, attention_weights).detach().cpu()
+    output_cuda = MSDeformAttnFunction.apply(value, shapes, level_start_index, sampling_locations, attention_weights, im2col_step).detach().cpu()
+    fwdok = torch.allclose(output_cuda, output_pytorch, rtol=1e-2, atol=1e-3)
+    max_abs_err = (output_cuda - output_pytorch).abs().max()
+    max_rel_err = ((output_cuda - output_pytorch).abs() / output_pytorch.abs()).max()
+
+    print(f'* {fwdok} check_forward_equal_with_pytorch_float: max_abs_err {max_abs_err:.2e} max_rel_err {max_rel_err:.2e}')
+
+
+def check_gradient_numerical(channels=4, grad_value=True, grad_sampling_loc=True, grad_attn_weight=True):
+
+    value = torch.rand(N, S, M, channels).cuda() * 0.01
+    sampling_locations = torch.rand(N, Lq, M, L, P, 2).cuda()
+    attention_weights = torch.rand(N, Lq, M, L, P).cuda() + 1e-5
+    attention_weights /= attention_weights.sum(-1, keepdim=True).sum(-2, keepdim=True)
+    im2col_step = 2
+    func = MSDeformAttnFunction.apply
+
+    value.requires_grad = grad_value
+    sampling_locations.requires_grad = grad_sampling_loc
+    attention_weights.requires_grad = grad_attn_weight
+
+    gradok = gradcheck(func, (value.double(), shapes, level_start_index, sampling_locations.double(), attention_weights.double(), im2col_step))
+
+    print(f'* {gradok} check_gradient_numerical(D={channels})')
+
+
+if __name__ == '__main__':
+    check_forward_equal_with_pytorch_double()
+    check_forward_equal_with_pytorch_float()
+
+    for channels in [30, 32, 64, 71, 1025, 2048, 3096]:
+        check_gradient_numerical(channels, True, True, True)
+
+
+

detect_tools/upn/requirments.txt

@@ -0,0 +1 @@
+mmengine==0.8.2

detect_tools/upn/transforms/transform.py

@@ -0,0 +1,142 @@
+import random
+import torch
+import torchvision.transforms.functional as F
+
+
+def resize(image, target, size, max_size=None):
+    # size can be min_size (scalar) or (w, h) tuple
+
+    def get_size_with_aspect_ratio(image_size, size, max_size=None):
+        w, h = image_size
+        if max_size is not None:
+            min_original_size = float(min((w, h)))
+            max_original_size = float(max((w, h)))
+            if max_original_size / min_original_size * size > max_size:
+                size = int(round(max_size * min_original_size / max_original_size))
+
+        if (w <= h and w == size) or (h <= w and h == size):
+            return (h, w)
+
+        if w < h:
+            ow = size
+            oh = int(size * h / w)
+        else:
+            oh = size
+            ow = int(size * w / h)
+
+        return (oh, ow)
+
+    def get_size(image_size, size, max_size=None):
+        if isinstance(size, (list, tuple)):
+            return size[::-1]
+        else:
+            return get_size_with_aspect_ratio(image_size, size, max_size)
+
+    size = get_size(image.size, size, max_size)
+    rescaled_image = F.resize(image, size)
+
+    if target is None:
+        return rescaled_image, None
+
+    ratios = tuple(
+        float(s) / float(s_orig) for s, s_orig in zip(rescaled_image.size, image.size)
+    )
+    ratio_width, ratio_height = ratios
+
+    target = target.copy()
+    if "exampler_box" in target:
+        boxes = target["exampler_box"]
+        if isinstance(boxes, torch.Tensor):
+            scaled_boxes = boxes * torch.as_tensor(
+                [ratio_width, ratio_height, ratio_width, ratio_height]
+            )
+            target["exampler_box"] = scaled_boxes
+        elif isinstance(boxes, dict):
+            for k, v in boxes.items():
+                scaled_boxes = v * torch.as_tensor(
+                    [ratio_width, ratio_height, ratio_width, ratio_height]
+                )
+                target["exampler_box"][k] = scaled_boxes
+
+    if "demo_pos_exampler_box" in target:
+        boxes = target["demo_pos_exampler_box"]
+        scaled_boxes = boxes * torch.as_tensor(
+            [ratio_width, ratio_height, ratio_width, ratio_height]
+        )
+        target["demo_pos_exampler_box"] = scaled_boxes
+
+    if "demo_neg_exampler_box" in target:
+        boxes = target["demo_neg_exampler_box"]
+        scaled_boxes = boxes * torch.as_tensor(
+            [ratio_width, ratio_height, ratio_width, ratio_height]
+        )
+        target["demo_neg_exampler_box"] = scaled_boxes
+
+    if "boxes" in target:
+        boxes = target["boxes"]
+        scaled_boxes = boxes * torch.as_tensor(
+            [ratio_width, ratio_height, ratio_width, ratio_height]
+        )
+        target["boxes"] = scaled_boxes
+
+    if "area" in target:
+        area = target["area"]
+        scaled_area = area * (ratio_width * ratio_height)
+        target["area"] = scaled_area
+
+    h, w = size
+    target["size"] = torch.tensor([h, w])
+
+    return rescaled_image, target
+
+
+class RandomResize(object):
+
+    def __init__(self, sizes, max_size=None):
+        assert isinstance(sizes, (list, tuple))
+        self.sizes = sizes
+        self.max_size = max_size
+
+    def __call__(self, img, target=None):
+        size = random.choice(self.sizes)
+        return resize(img, target, size, self.max_size)
+
+
+class ToTensor(object):
+
+    def __call__(self, img, target):
+        return F.to_tensor(img), target
+
+
+class Normalize(object):
+
+    def __init__(self, mean, std):
+        self.mean = mean
+        self.std = std
+
+    def __call__(self, image, target=None):
+        image = F.normalize(image, mean=self.mean, std=self.std)
+        if target is None:
+            return image, None
+        target = target.copy()
+        h, w = image.shape[-2:]
+        return image, target
+
+
+class Compose(object):
+
+    def __init__(self, transforms):
+        self.transforms = transforms
+
+    def __call__(self, image, target):
+        for t in self.transforms:
+            image, target = t(image, target)
+        return image, target
+
+    def __repr__(self):
+        format_string = self.__class__.__name__ + "("
+        for t in self.transforms:
+            format_string += "\n"
+            format_string += "    {0}".format(t)
+        format_string += "\n)"
+        return format_string

requirements.txt

@@ -0,0 +1,19 @@
+torch==2.6.0
+torchvision==0.21.0
+transformers==4.50.1
+https://airesources.oss-cn-hangzhou.aliyuncs.com/lp/wheel/sam3-0.1.0-py3-none-any.whl
+timm==1.0.9
+accelerate==1.4.0
+gradio
+mmengine==0.8.2
+einops
+ninja
+scikit-image
+decord
+scikit-learn
+matplotlib
+modelscope
+https://airesources.oss-cn-hangzhou.aliyuncs.com/lp/wheel/multiscaledeformableattention-1.0-cp310-cp310-linux_x86_64.whl
+https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
+pycocotools
+opencv-python

vlm_fo1/__init__.py

@@ -0,0 +1 @@
+

vlm_fo1/constants.py

@@ -0,0 +1,29 @@
+LOGDIR = "."
+
+global DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+# Model Constants
+IGNORE_INDEX = -100
+IMAGE_TOKEN_INDEX = -200 #151656 #151655 #-200
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"
+
+# For Qwen2_5_VL
+QWEN2_5_VL_IMAGE_TOKEN = "<|image_pad|>"
+QWEN2_5_VL_IMAGE_TOKEN_INDEX = 151655
+
+# For regions
+DEFAULT_REGION_TOKEN = "<region<i>>"
+DEFAULT_REGION_FEATURE_TOKEN = "<regionfeat>"
+DEFAULT_REGION_INDEX = -300 #151654 #151654 #-300
+
+# For Grounding
+DEFAULT_GROUNDING_START = "<ground>"
+DEFAULT_GROUNDING_END = "</ground>"
+DEFAULT_GROUNDING_OBJECTS_START = "<objects>"
+DEFAULT_GROUNDING_OBJECTS_END = "</objects>"
+
+# For Think
+DEFAULT_THINK_START = "<think>"
+DEFAULT_THINK_END = "</think>"

vlm_fo1/mm_utils.py

@@ -0,0 +1,660 @@
+from PIL import Image
+from PIL import ImageDraw, ImageOps
+from io import BytesIO
+import base64
+import re 
+import torch
+from transformers import StoppingCriteria
+from vlm_fo1.constants import IMAGE_TOKEN_INDEX, DEFAULT_REGION_INDEX
+import requests
+from vlm_fo1.constants import (
+    IMAGE_TOKEN_INDEX,
+    DEFAULT_IMAGE_TOKEN,
+    DEFAULT_IM_START_TOKEN,
+    DEFAULT_IM_END_TOKEN,
+    IGNORE_INDEX,
+    DEFAULT_REGION_TOKEN, 
+    DEFAULT_REGION_FEATURE_TOKEN
+)
+import torch
+from transformers import TextStreamer
+import random
+import re
+from typing import List, Tuple
+import io
+import base64
+
+
+def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
+    """
+    Tokenizes prompts containing <image> or <image_0>... special tokens.
+
+    If the prompt uses <image_0>, <image_1>, ..., each is replaced with a placeholder index (-200).
+    If the prompt uses <image>, it is replaced with image_token_index.
+
+    Args:
+        prompt (str): The prompt potentially containing image tokens.
+        tokenizer: The tokenizer object.
+        image_token_index (int): Token id to use when encountering <image> token.
+        return_tensors (Optional[str]): If 'pt', return a torch tensor.
+
+    Returns:
+        List[int] or torch.Tensor: The tokenized input with image token indices inserted appropriately.
+    """
+    if "<image_0>" in prompt:
+        # Case: prompt contains indexed image tokens like <image_0>, <image_1>, etc.
+        image_token_pattern = re.compile(r"<image_(\d+)>")
+        prompt_chunks = re.split(r'<image_[0-9]+>', prompt)
+        image_tags = image_token_pattern.findall(prompt)
+
+        input_ids = []
+        for i, chunk in enumerate(prompt_chunks):
+            input_ids.extend(tokenizer(chunk).input_ids)
+            if i < len(image_tags):
+                # Insert placeholder where <image_n> token was.
+                input_ids.append(-200)
+    else:
+        # Case: prompt contains plain <image> tokens.
+        prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]
+
+        def insert_separator(X, sep):
+            # Helper function to insert a separator token between chunks.
+            return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]
+
+        input_ids = []
+        offset = 0
+        # If first chunk starts with <bos> token, make sure to keep it only once.
+        if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
+            offset = 1
+            input_ids.append(prompt_chunks[0][0])
+
+        # Insert image_token_index between chunks.
+        for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
+            input_ids.extend(x[offset:])
+    # Optionally convert output to PyTorch tensor.
+    if return_tensors is not None:
+        if return_tensors == 'pt':
+            return torch.tensor(input_ids, dtype=torch.long)
+        else:
+            raise ValueError(f'Unsupported tensor type: {return_tensors}')
+
+    return input_ids
+
+def tokenizer_image_region_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, region_token_index=DEFAULT_REGION_INDEX, return_tensors=None):
+    """
+    Tokenizes prompts containing both <image> and <regionfeat> delimiters, inserting specified token indices.
+
+    Each <image> chunk is split, and within that chunk, <regionfeat> locations receive region_token_index.
+
+    Args:
+        prompt (str): The prompt with <image> and <regionfeat> delimiters.
+        tokenizer: The tokenizer object.
+        image_token_index (int): Insert this at <image> splits.
+        region_token_index (int): Insert this at <regionfeat> splits.
+        return_tensors (Optional[str]): If 'pt', return torch tensor.
+
+    Returns:
+        List[int] or torch.Tensor: The tokenized input with region/image tokens placed.
+    """
+    # Split by <image> tags first.
+    image_chunks = prompt.split('<image>')
+    
+    prompt_chunks = []
+    for chunk in image_chunks:
+        # Split each image chunk by <regionfeat>.
+        obj_chunks = chunk.split('<regionfeat>')
+        # Tokenize each subchunk.
+        token_chunks = [tokenizer(c).input_ids for c in obj_chunks]
+        prompt_chunks.append(token_chunks)
+    
+    input_ids = []
+    offset = 0
+
+    # If first chunk starts with <bos> token, include only once.
+    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and len(prompt_chunks[0][0]) > 0 and prompt_chunks[0][0][0] == tokenizer.bos_token_id:
+        offset = 1
+        input_ids.append(prompt_chunks[0][0][0])
+    
+    # Stitch together all chunks with region/image tokens at appropriate locations.
+    for i, chunk_group in enumerate(prompt_chunks):
+        if len(chunk_group) > 0:
+            input_ids.extend(chunk_group[0][offset:])
+        for chunk in chunk_group[1:]:
+            input_ids.append(region_token_index)
+            input_ids.extend(chunk)
+        # Insert <image> token except after the last image chunk.
+        if i < len(prompt_chunks) - 1:
+            input_ids.append(image_token_index)
+    # Optionally convert to PyTorch tensor.
+    if return_tensors is not None:
+        if return_tensors == 'pt':
+            return torch.tensor(input_ids, dtype=torch.long)
+        else:
+            raise ValueError(f'Unsupported tensor type: {return_tensors}')
+
+    return input_ids
+
+class KeywordsStoppingCriteria(StoppingCriteria):
+    """
+    Implements custom stopping criteria for generation based on keywords:
+    If the generated output contains any of the keywords, generation stops.
+    """
+    def __init__(self, keywords, tokenizer, input_ids):
+        self.keywords = keywords
+        self.keyword_ids = []
+        self.max_keyword_len = 0
+        for keyword in keywords:
+            cur_keyword_ids = tokenizer(keyword).input_ids
+            # Remove BOS if present except for single token
+            if len(cur_keyword_ids) > 1 and cur_keyword_ids[0] == tokenizer.bos_token_id:
+                cur_keyword_ids = cur_keyword_ids[1:]
+            if len(cur_keyword_ids) > self.max_keyword_len:
+                self.max_keyword_len = len(cur_keyword_ids)
+            self.keyword_ids.append(torch.tensor(cur_keyword_ids))
+        self.tokenizer = tokenizer
+        # Track the generation start length
+        self.start_len = input_ids.shape[1]
+
+    def call_for_batch(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        """
+        Checks if a keyword exists in the latest generated output ids for a single batch element.
+        """
+        offset = min(output_ids.shape[1] - self.start_len, self.max_keyword_len)
+        self.keyword_ids = [keyword_id.to(output_ids.device) for keyword_id in self.keyword_ids]
+        for keyword_id in self.keyword_ids:
+            truncated_output_ids = output_ids[0, -keyword_id.shape[0]:]
+            if torch.equal(truncated_output_ids, keyword_id):
+                return True
+        outputs = self.tokenizer.batch_decode(output_ids[:, -offset:], skip_special_tokens=True)[0]
+        for keyword in self.keywords:
+            if keyword in outputs:
+                return True
+        return False
+
+    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        """
+        Checks for keywords in each batch item; stops when all have satisfied the keyword condition.
+        """
+        outputs = []
+        for i in range(output_ids.shape[0]):
+            outputs.append(self.call_for_batch(output_ids[i].unsqueeze(0), scores))
+        return all(outputs)
+
+def load_image(image_file):
+    """
+    Loads an image from a local path, base64 string, URL, or PIL.Image.
+
+    If the input image is smaller than 28x28, it will be resized to at least that size.
+
+    Args:
+        image_file (str or PIL.Image.Image): Image source.
+
+    Returns:
+        PIL.Image.Image: Loaded image in RGB mode, at least 28x28 in size.
+    """
+    if isinstance(image_file, Image.Image):
+        image = image_file.convert("RGB")
+    # Case: load from URL
+    elif image_file.startswith("http") or image_file.startswith("https"):
+        response = requests.get(image_file)
+        image = Image.open(BytesIO(response.content)).convert("RGB")
+    # Case: load from base64-encoded string
+    elif image_file.startswith("data:image/"):
+        image = image_file.replace("data:image/jpeg;base64,", "")
+        image_data = base64.b64decode(image)
+        image = Image.open(BytesIO(image_data)).convert("RGB")
+    elif isinstance(image_file, str):
+        # Case: load from local file path
+        image = Image.open(image_file).convert("RGB")
+    else:
+        raise ValueError(f"Unsupported image type: {type(image_file)}")
+    
+    # Ensure minimum size 28x28
+    if image.width < 28 or image.height < 28:
+        image = image.resize((max(28, image.width), max(28, image.height)))
+    return image
+
+def image_to_base64(img_pil):
+    """
+    Encodes a PIL Image as JPEG in base64 format.
+
+    Args:
+        img_pil (PIL.Image.Image): Source image.
+
+    Returns:
+        str: base64-encoded JPEG image string.
+    """
+    with io.BytesIO() as buffer:
+        img_pil.save(buffer, format="JPEG")
+        base64_image = base64.b64encode(buffer.getvalue()).decode('utf-8')
+    return base64_image
+
+def draw_bboxes_and_save(
+    image: Image.Image,
+    fo1_bboxes: dict = {},
+    detection_bboxes: List[Tuple[int, int, int, int]] = [],
+    output_path: str = 'output.jpg',
+    color: str = 'red',
+    total_color: str = 'green',
+    width: int = 2
+) -> None:
+    """
+    Draws bounding boxes (both ground-truth/proposed and detection) on a PIL image and saves result.
+
+    Args:
+        image (PIL.Image.Image): Input PIL image object.
+        fo1_bboxes (dict): Label -> List[bbox] mapping for annotation bboxes.
+        detection_bboxes (List[Tuple]): List of detection bounding boxes; each bbox is (x_min, y_min, x_max, y_max).
+        output_path (str): Path to save the output image.
+        color (str): Color for fo1_bboxes.
+        total_color (str): Color for detection_bboxes.
+        width (int): Rectangle outline width.
+
+    Returns:
+        None
+    """
+    draw = ImageDraw.Draw(image)
+    
+    # Draw detection boxes with `total_color`
+    for bbox in detection_bboxes:
+        if len(bbox) != 4:
+            print(f"Warning: skip the invalid bbox {bbox}")
+            continue
+        shape = [(bbox[0], bbox[1]), (bbox[2], bbox[3])]
+        draw.rectangle(shape, outline=total_color, width=width)
+
+    # Draw annotated bboxes with labels and `color`
+    for bbox_label, bbox_list in fo1_bboxes.items():
+        for bbox in bbox_list:
+            if len(bbox) != 4:
+                print(f"Warning: skip the invalid bbox {bbox}")
+                continue
+            shape = [(bbox[0], bbox[1]), (bbox[2], bbox[3])]
+            draw.rectangle(shape, outline=color, width=width)
+            draw.text((bbox[0], bbox[1]), bbox_label, fill=color)
+
+    # Save output image (catching common IO exceptions).
+    try:
+        image.save(output_path)
+        print(f"The image has been successfully saved to: {output_path}")
+    except IOError as e:
+        print(f"Error: failed to save the image to {output_path}. Reason: {e}")
+
+def adjust_bbox(bbox_list, original_h, original_w, resize_h, resize_w):
+    """
+    Adjusts bounding boxes from original image size to resized image size, compensating for scaling.
+
+    Args:
+        bbox_list (List[List[float]]): List of original boxes [x1, y1, x2, y2].
+        original_h (int): Original image height.
+        original_w (int): Original image width.
+        resize_h (int): Resized image height.
+        resize_w (int): Resized image width.
+
+    Returns:
+        List[List[float]]: Bounding boxes transformed to resized image coordinates.
+    """
+    output_list = []
+    def adjust_bbox_range(bbox, width, height):
+        # Ensure all coordinates are within the original image border.
+        x1, y1, x2, y2 = bbox
+        x1 = max(0, min(width, x1))
+        y1 = max(0, min(height, y1))
+        x2 = max(0, min(width, x2))
+        y2 = max(0, min(height, y2))
+        return [x1, y1, x2, y2]
+
+    for bbox in bbox_list:
+        bbox = adjust_bbox_range(bbox, original_w, original_h)
+        bbox[0] = bbox[0] * resize_w / original_w
+        bbox[1] = bbox[1] * resize_h / original_h
+        bbox[2] = bbox[2] * resize_w / original_w
+        bbox[3] = bbox[3] * resize_h / original_h
+        output_list.append(bbox)
+    return output_list
+
+def extract_predictions_to_bboxes(prediction: str, bbox_list):
+    """
+    Parse prediction string in the expected format and map each ground label
+    to its corresponding bounding boxes using bbox_list.
+
+    Args:
+        prediction (str): Model output string with <ground>...<objects>... markup.
+        bbox_list (List[List[float]]): Full list of predicted or reference bounding boxes.
+
+    Returns:
+        dict: label -> list of bboxes
+    """
+    label_to_indexes = {}
+    label_to_bboxes = {}
+
+    match_pattern = r"<ground>(.*?)<\/ground><objects>(.*?)<\/objects>"
+    matches = re.findall(match_pattern, prediction)
+
+    for label_text, indexes in matches:
+        label_text = label_text.strip()
+        indexes_tags = re.findall(r"<region\d+>", indexes)
+        region_indexes = set([int(index.split("<region")[-1].split(">")[0]) for index in indexes_tags])
+        if label_text not in label_to_indexes:
+            label_to_indexes[label_text] = region_indexes
+        else:
+            label_to_indexes[label_text] = label_to_indexes[label_text] | region_indexes
+
+    for label, indexes in label_to_indexes.items():
+        label_to_bboxes[label] = [bbox_list[index] for index in indexes]
+
+    return label_to_bboxes
+
+def extract_predictions_to_indexes(prediction: str):
+    """
+    Parse prediction string, returning label -> set-of-indexes mapping.
+
+    Args:
+        prediction (str): Model prediction output.
+
+    Returns:
+        dict: label -> set(int)
+    """
+    label_to_indexes = {}
+    match_pattern = r"<ground>(.*?)<\/ground><objects>(.*?)<\/objects>"
+    matches = re.findall(match_pattern, prediction)
+
+    for label_text, indexes in matches:
+        label_text = label_text.strip()
+        indexes_tags = re.findall(r"<region\d+>", indexes)
+        region_indexes = set([int(index.split("<region")[-1].split(">")[0]) for index in indexes_tags])
+        if label_text not in label_to_indexes:
+            label_to_indexes[label_text] = region_indexes
+        else:
+            label_to_indexes[label_text] = label_to_indexes[label_text] | region_indexes
+
+    return label_to_indexes
+
+def resize_shortest_edge_images_and_bboxes(
+    image_list: List[Image.Image],
+    bbox_lists: List,
+    candidate_sizes: List[int] = [], 
+    max_size: int = 2048
+    ):
+    """
+    Randomly selects a size for the shortest edge, and proportionally resizes both images and bounding boxes.
+
+    The function maintains the image aspect ratio and ensures that the resized dimensions do not exceed the specified max_size.
+    Bounding boxes are transformed accordingly.
+
+    Args:
+        image_list (List[Image.Image]): A list of PIL Image objects.
+        bbox_lists (List[List[List[float]]]): A list of lists of bounding boxes per image.
+        candidate_sizes (List[int]): Optional list of sizes to choose the target short edge from.
+        max_size (int): Maximum allowed long edge after resizing.
+
+    Returns:
+        Tuple[List[Image.Image], List[List[List[float]]]]:
+            ([resized_image1, ...], [bbox_list1, ...]) - Possibly shape will match original (see below)
+
+    Raises:
+        ValueError: on input list length mismatch or emptiness.
+    """
+    bbox_tensor = torch.tensor(bbox_lists)
+    # Normalize input: wrap bbox_lists into list-of-list, if needed.
+    if len(bbox_tensor.shape) == 2 and bbox_tensor.shape[1] == 4:
+        bbox_lists = [bbox_lists]
+
+    if not image_list or not bbox_lists:
+        raise ValueError("Input lists cannot be empty.")
+    if len(image_list) != len(bbox_lists):
+        raise ValueError("The lengths of the image list and the bounding box list must be the same.")
+
+    # Randomly select short edge size (if given candidate sizes)
+    if len(candidate_sizes) > 0:
+        target_size = random.choice(candidate_sizes)
+    else:
+        target_size = None
+
+    resized_images = []
+    transformed_bbox_lists = []
+
+    # Process each image and its corresponding bbox list
+    for img, bboxes in zip(image_list, bbox_lists):
+        original_width, original_height = img.size
+
+        # Determine scaling factor to bring short edge to target_size
+        shortest_side = min(original_width, original_height)
+        if target_size:
+            scale = target_size / shortest_side
+        else:
+            scale = 1.0
+
+        # Propose new height and width with this scale
+        new_height, new_width = int(original_height * scale), int(original_width * scale)
+
+        # If resulting long edge exceeds max_size, rescale down so that it fits.
+        longest_side = max(new_height, new_width)
+        if longest_side > max_size:
+            scale = max_size / longest_side
+            new_height, new_width = int(new_height * scale), int(new_width * scale)
+        # Ensure images are at least 28x28 (model may expect it)
+        new_width = max(28, new_width)
+        new_height = max(28, new_height)
+
+        # Resize image, using BICUBIC for quality if shape changes
+        if new_width == original_width and new_height == original_height:
+            resized_img = img
+        else:
+            resized_img = img.resize((new_width, new_height), Image.Resampling.BICUBIC)
+        resized_images.append(resized_img)
+
+        # Transform bounding boxes
+        current_transformed_bboxes = []
+        scale_ratio_x = new_width / original_width
+        scale_ratio_y = new_height / original_height
+        for bbox in bboxes:
+            x1, y1, x2, y2 = bbox
+            new_x1 = x1 * scale_ratio_x
+            new_y1 = y1 * scale_ratio_y
+            new_x2 = x2 * scale_ratio_x
+            new_y2 = y2 * scale_ratio_y
+            current_transformed_bboxes.append([new_x1, new_y1, new_x2, new_y2])
+        transformed_bbox_lists.append(current_transformed_bboxes)
+
+    # If original input was a single image (not list), unpack.
+    if len(bbox_tensor.shape) == 2 and bbox_tensor.shape[1] == 4:
+        return resized_images, transformed_bbox_lists[0]
+    else:
+        return resized_images, transformed_bbox_lists
+
+def make_message_context(tokenizer, message, chat_format="chatml"):
+    """
+    Given a message dict, construct the prompt, tokenized context tokens, image URLs, and bbox_list.
+
+    Handles both standard string 'content' and multi-part (list) content, appropriately placing image/region tokens.
+
+    Args:
+        tokenizer: tokenizer object
+        message (dict): Contains role, content, and optionally bbox_list.
+        chat_format (str): Optionally select chat format (default 'chatml').
+
+    Returns:
+        tuple: (inp, context_tokens, image_urls, bbox_list)
+    """
+    image_urls = []
+    if chat_format == "chatml":
+        im_start, im_end = "<|im_start|>", "<|im_end|>"
+        im_start_tokens = [151644]
+        im_end_tokens = [151645]
+        nl_tokens = tokenizer.encode("\n")
+        role = message["role"]
+        content = message["content"]
+        bbox_list = message.get("bbox_list", None)
+
+        if role == "system":
+            inp = f"{im_start}{role}\n{content}{im_end}\n"
+            context_tokens = tokenizer.encode(
+                role, allowed_special=set()) + nl_tokens + tokenizer.encode(content, allowed_special=set())
+            context_tokens = im_start_tokens + context_tokens + im_end_tokens
+
+        if role == "user":
+            if isinstance(content, str):
+                # Plain string message
+                inp = f"{im_start}{role}\n{content}{im_end}\n"
+                context_tokens = tokenizer.encode(
+                    role, allowed_special=set()) + nl_tokens + tokenizer.encode(content,
+                                                                                allowed_special=set())
+                context_tokens = im_start_tokens + context_tokens + im_end_tokens
+            if isinstance(content, list):
+                # Multi-part message (text and image_url parts, maybe region tokens)
+                inp = f"{im_start}{role}\n"
+                image_count = 1
+                for message_part in content:
+                    if message_part["type"] == "text":
+                        inp += f"{message_part['text']}"
+
+                    if message_part["type"] == "image_url":
+                        # Insert special vision/image tokens, possibly region tokens
+                        inp += DEFAULT_IM_START_TOKEN + '<image>' + DEFAULT_IM_END_TOKEN + '\n'
+                        # If regions exist, add per-region special token.
+                        if bbox_list and len(bbox_list) > 0:
+                            for idx, bbox in enumerate(bbox_list):
+                                inp += DEFAULT_REGION_TOKEN.replace('<i>', str(idx)) + DEFAULT_REGION_FEATURE_TOKEN
+                            inp += '\n'
+
+                        image_urls.append(message_part['image_url']['url'])
+                        image_count += 1
+                inp += f"{im_end}\n"
+
+                # Choose tokenizer logic based on whether bbox (region) list exists
+                if bbox_list and len(bbox_list) > 0:
+                    context_tokens = tokenizer_image_region_token(inp, tokenizer)
+                else:
+                    context_tokens = tokenizer_image_token(inp, tokenizer, image_token_index=IMAGE_TOKEN_INDEX)
+        return inp, context_tokens, image_urls, bbox_list
+
+def prepare_inputs(model_name, model, image_processors, tokenizer, messages, device="cuda", max_tokens=512, top_p=1.0, temperature=0.0, do_sample=False, image_size=None):
+    """
+    Fully prepares keyword arguments for model.generate (and compatible API) from messages and model specs.
+
+    Handles prompt assembly, tokenization, image loading/preprocessing, region support, streaming, etc.
+    Supports specific tweak for Qwen2.5-VL style vision tokens.
+
+    Args:
+        model_name (str): Model identifier string.
+        model: Model/config object.
+        image_processors (tuple): (primary, auxiliary) image processors.
+        tokenizer: Tokenizer object.
+        messages (list): Multi-message input list (chat history).
+        device (str): Target (usually 'cuda' or 'cpu').
+        max_tokens, top_p, temperature, do_sample: Standard generation kwargs.
+
+    Returns:
+        dict: ready-to-use argument dict for model.generate().
+    """
+    # For Qwen2.5-VL, patch vision special tokens globally.
+    if 'qwen2.5-vl' in model_name.lower() or 'qwen2_5_vl' in model_name.lower():
+        global DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+        DEFAULT_IM_START_TOKEN = "<|vision_start|>"
+        DEFAULT_IM_END_TOKEN = "<|vision_end|>"
+
+    primary_image_processor, auxiliary_image_processor = image_processors
+
+    prompt = ""
+    input_tokens = []
+    image_urls = []
+    # Compose prompt and accumulate all components from provided messages
+    for message in messages:
+        inp, context_tokens, image_urls, bbox_list = make_message_context(tokenizer, message)
+        prompt += inp
+        input_tokens.extend(context_tokens)
+
+    # Ensure a system prompt at start, if not already present.
+    if "system" not in prompt:
+        system_content = "system\nYou are a helpful assistant."
+        system_prompt = "<|im_start|>" + system_content + "<|im_end|>" + "\n"
+        prompt = system_prompt + prompt
+        system_tokens = [151644] + tokenizer(system_content).input_ids + [151645] + tokenizer("\n").input_ids
+        input_tokens = system_tokens + input_tokens
+
+    # Ensure prompt ends with assistant's turn.
+    if not prompt.endswith("<|im_start|>assistant"):
+        last_assistant_prompt = "<|im_start|>" + "assistant" + "\n"
+        prompt += last_assistant_prompt
+        # last_assistant_tokens = [6] + self.tokenizer("assistant\n").input_ids
+        last_assistant_tokens = [151644] + tokenizer("assistant\n").input_ids
+        input_tokens.extend(last_assistant_tokens)
+
+    primary_images_tensor = None
+    auxiliary_images_tensor = None
+    primary_image_grid_thws = None
+    if image_urls:
+        # Load images, resize them, and update bbox_list downstream
+        images = [load_image(i) for i in image_urls]
+        if image_size is not None:
+            images, bbox_list = resize_shortest_edge_images_and_bboxes(images, bbox_list, candidate_sizes=[image_size], max_size=2048)
+        else:
+            images, bbox_list = resize_shortest_edge_images_and_bboxes(images, bbox_list, max_size=2048)
+
+    
+        # When region-indexed tokens are enabled
+        if getattr(model.config, 'mm_use_region_index_token', False):
+            origin_image_size = [image.size for image in images]
+            aux_images = images.copy()
+            auxiliary_images_tensor = [auxiliary_image_processor.preprocess(i, return_tensors='pt')['pixel_values'][0].to(device) for i in aux_images]
+
+            if bbox_list and len(bbox_list) > 0:
+                # Limit number of bbox (for computational constraints, etc.)
+                bbox_list = bbox_list[:100]
+                resize_h, resize_w = auxiliary_images_tensor[0].shape[-2:]
+                original_w, original_h = origin_image_size[0]
+                # Adjust bbox to match resized images (post pre-processing)
+                bbox_list = adjust_bbox(bbox_list, original_h, original_w, resize_h, resize_w)
+                bbox_list = [torch.tensor(bbox_list)]
+            else:
+                bbox_list = None
+        else:
+            auxiliary_images_tensor = None
+
+    # Preprocess primary images for main vision model branch
+    primary_images = []
+    primary_image_grid_thws = []
+    for im in images:
+        processed_data = primary_image_processor.preprocess(im, return_tensors="pt")
+        image_i = processed_data['pixel_values']
+        image_grid_thw_i = processed_data['image_grid_thw']
+        primary_images.append(image_i)
+        primary_image_grid_thws.append(image_grid_thw_i)
+    primary_images_tensor = [image_i.to(device) for image_i in primary_images]
+
+    # For Qwen-style, force specific end-token as stopping criterion
+    if "qwen" in model_name.lower():
+        input_ids = torch.tensor([input_tokens]).to(device)
+        keywords = ["<|im_end|>"]
+
+    stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
+    streamer = TextStreamer(
+        tokenizer, skip_prompt=True, skip_special_tokens=True
+    )
+
+    # Default: greedy decoding if temperature=0. Else: enable sampling.
+    if temperature == 0.0:
+        do_sample = False
+    else:
+        do_sample = True
+
+    print("question:================\n", prompt, "\n=================")
+    # print("input ids:========", input_ids, "========")
+    generation_kwargs = dict(
+        inputs=input_ids,
+        images=primary_images_tensor,
+        images_aux=auxiliary_images_tensor,
+        image_grid_thws=primary_image_grid_thws,
+        bbox_list=bbox_list,
+        do_sample=do_sample,
+        temperature=temperature,
+        max_new_tokens=max_tokens,
+        streamer=streamer,
+        top_p=top_p,
+        use_cache=True,
+        stopping_criteria=[stopping_criteria],
+        pad_token_id=tokenizer.pad_token_id
+    )
+    return generation_kwargs
+

vlm_fo1/model/__init__.py

@@ -0,0 +1 @@
+from .language_model.omchat_qwen2_5_vl import OmChatQwen25VLForCausalLM, OmChatQwen25VLConfig

vlm_fo1/model/builder.py

@@ -0,0 +1,89 @@
+from transformers import AutoTokenizer
+import torch
+from vlm_fo1.model import *
+from safetensors.torch import load_file
+import os
+
+
+def load_pretrained_model(model_path, load_8bit=False, load_4bit=False, device="cuda"):
+    """
+    Loads a pretrained model along with its vision towers (and associated image processors).
+    This function supports loading in 8bit/4bit precision and explicit device placement.
+
+    Args:
+        model_path (str): Path to the pretrained model directory.
+        load_8bit (bool): Whether to load the model in 8bit mode.
+        load_4bit (bool): Whether to load the model in 4bit mode.
+        device (str): Device to load model onto, e.g., "cuda" or "cpu".
+
+    Returns:
+        tuple: (tokenizer, model, image_processor)
+    """
+    kwargs = {"device_map": device}
+
+    # Set model loading parameters for quantization or floating point
+    if load_8bit:
+        kwargs['load_in_8bit'] = True
+    elif load_4bit:
+        kwargs['load_in_4bit'] = True
+    else:
+        kwargs['torch_dtype'] = torch.bfloat16
+
+    # print(model_path)
+
+    # Only proceed for vlm-fo1 models
+    if 'vlm-fo1' in model_path.lower():
+        # Load tokenizer (slow tokenizer enforced)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
+        # If this is the Qwen2.5-VL variant, load with additional kwargs
+        if 'qwen2.5-vl' in model_path.lower() or 'qwen2_5_vl' in model_path.lower():
+            model, loading_info = OmChatQwen25VLForCausalLM.from_pretrained(
+                model_path,
+                low_cpu_mem_usage=True,
+                output_loading_info=True,
+                attn_implementation="flash_attention_2",
+                **kwargs
+            )
+            # print(f'OmChatQwen25VLForCausalLM loading_info: {loading_info}')
+        # (For other variants of vlm-fo1, model loading detail may need additional condition.)
+
+    if 'vlm-fo1' in model_path.lower():
+        # --- Vision Tower Loading ---
+        # Load the main vision tower weights from model_path if it is not yet loaded
+        primary_vision_tower = model.get_vision_tower()
+        if primary_vision_tower and not primary_vision_tower.is_loaded:
+            primary_vision_tower.load_model(model_path=model_path, is_train=False)
+            primary_vision_tower.to(device=device, dtype=torch.bfloat16)  # Move to correct device/dtype
+
+        # Grab primary image processor from vision tower, if present
+        if primary_vision_tower:
+            primary_image_processor = primary_vision_tower.image_processor
+
+        # --- Auxiliary Vision Tower Handling (Qwen2.5-VL case only) ---
+        if 'qwen2.5-vl' in model_path.lower() or 'qwen2_5_vl' in model_path.lower():
+            try:
+                aux_image_size = model.config.aux_image_size
+            except Exception:
+                # If aux_image_size is missing from config fallback to 768
+                aux_image_size = 768
+
+            aux_image_aspect_ratio = model.config.aux_image_aspect_ratio
+            aux_vision_tower = model.get_vision_tower_aux()
+            # Only load if not already loaded
+            if aux_vision_tower and not aux_vision_tower.is_loaded:
+                aux_vision_tower.load_model(image_size=aux_image_size, is_train=False, aspect_ratio=aux_image_aspect_ratio)
+                aux_vision_tower.to(device=device, dtype=torch.bfloat16)
+
+        # Get auxiliary image processor if there is an aux vision tower
+        if aux_vision_tower:
+            aux_image_processor = aux_vision_tower.image_processor
+        else:
+            image_processor = None  # Set to None if there is no auxiliary vision tower
+
+        # image_processor returned as a tuple of (primary, aux)
+        image_processor = (primary_image_processor, aux_image_processor)
+
+    # Set model to eval mode and move to correct device before returning
+    model.eval()
+    model.to(device=device, dtype=torch.bfloat16)
+    return tokenizer, model, image_processor

vlm_fo1/model/language_model/omchat_qwen2_5_vl.py

@@ -0,0 +1,576 @@
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from transformers import Qwen2_5_VLConfig, AutoConfig, AutoModelForCausalLM
+from vlm_fo1.model.multimodal_encoder.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLModel, Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLCausalLMOutputWithPast
+from vlm_fo1.model.multimodal_encoder.qwen2_5_vl_encoder import Qwen2_5_VlVisionTower
+from vlm_fo1.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_REGION_INDEX, QWEN2_5_VL_IMAGE_TOKEN, QWEN2_5_VL_IMAGE_TOKEN_INDEX
+
+from ..omchat_arch import OmChatMetaModel, OmChatMetaForCausalLM
+
+# Custom config which extends Qwen2_5_VLConfig for OmChat multimodal model
+class OmChatQwen25VLConfig(Qwen2_5_VLConfig):
+    model_type = "omchat_qwen2_5_vl"
+    rotary_type = "normal_rotary"
+    multi_scale_im = None
+    vision_tower_aux = None
+
+# Core model definition: inherits from OmChat and Qwen multimodal base
+class OmChatQwen25VLModel(OmChatMetaModel, Qwen2_5_VLModel):
+    config_class = OmChatQwen25VLConfig
+
+    def __init__(self, config: Qwen2_5_VLConfig):
+        super(OmChatQwen25VLModel, self).__init__(config)
+
+# Main class for multimodal CausalLM
+class OmChatQwen25VLForCausalLM(Qwen2_5_VLForConditionalGeneration, OmChatMetaForCausalLM):
+    config_class = OmChatQwen25VLConfig
+
+    def __init__(self, config, delay_load=True):
+        # Ensure config has delay_load property
+        if not hasattr(config, 'delay_load'):
+            config.delay_load = delay_load
+        super(Qwen2_5_VLForConditionalGeneration, self).__init__(config)
+        self.model = OmChatQwen25VLModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.rope_deltas = None  # cache rope_deltas here
+
+        self.post_init()
+
+    # Encode input images into feature representations
+    def encode_images(self, images, images_grid_thw=None):
+        # If vision_tower is Qwen2.5-specific, use its custom forward signature
+        if isinstance(self.get_model().get_vision_tower(), Qwen2_5_VlVisionTower):
+            image_features = self.get_model().get_vision_tower()(images, images_grid_thw)
+            image_features, image_grid_thws, multi_level_features = image_features
+            # If multiple images, handle concatenation
+            if type(image_features) is list:
+                # List has items of shape (1, seq_len, dim)
+                token_length_list = [i.shape[1] for i in image_features]
+                image_features = torch.cat(image_features, dim=1)      # Concatenate to (1, total_seq_len, dim)
+        else:
+            image_features = self.get_model().get_vision_tower()(images)
+            image_grid_thws = None
+            multi_level_features = None
+
+        image_features = self.get_model().mm_projector(image_features)
+
+        # Split concatenated image features back by original lengths (for multi-image case)
+        if isinstance(self.get_model().get_vision_tower(), Qwen2_5_VlVisionTower):
+            start = 0
+            new_image_features = []
+            # Split according to token_length_list
+            for length in token_length_list:
+                end = start + length
+                new_image_features.append(image_features[:, start:end, :].squeeze(0))
+                start = end
+            image_features = new_image_features
+        
+        return image_features, image_grid_thws, multi_level_features
+    
+    # Encode region regions (bounding boxes) into features, optionally using auxiliary vision tower
+    def encode_regions(self, images, bbox_list, vt_multi_level_features=None, vt_images_size=None):
+        aux_image_features_list = self.get_model().get_vision_tower_aux()(images)
+        region_features = []
+        if getattr(self.config, "mm_use_vision_tower_region_feature", False):
+            image_features_list = vt_multi_level_features
+            for batch_idx, (image_features, aux_image_features) in enumerate(zip(image_features_list, aux_image_features_list)):
+
+                if getattr(self.config, "mm_use_simpleFPN_for_vt", False):
+                    multilevel_visual_feats = image_features[-1]
+                else:
+                    multilevel_visual_feats = image_features
+                multilevel_aux_visual_feats = aux_image_features["image_features"]
+                boxes = bbox_list[batch_idx]
+
+                # If no boxes provided, use dummy box (covers tiny region)
+                if boxes is None or len(boxes) == 0:
+                    boxes = torch.tensor([[0, 10, 0, 10]], device=multilevel_aux_visual_feats[0].device, dtype=torch.float32)
+
+                boxes = boxes.to(torch.float32).to(multilevel_aux_visual_feats[0].device)
+                current_image_height, current_image_width = images[batch_idx].shape[-2:]
+                original_height, original_width = vt_images_size[batch_idx]
+                # Scale bounding boxes from original image size to processed size
+                scale_height = original_height / current_image_height
+                scale_width = original_width / current_image_width
+                vt_boxes = boxes * torch.tensor([scale_width, scale_height, scale_width, scale_height], device=boxes.device)
+
+                extracted_region_feat = self.get_model().object_vp_extractor(
+                    aux_multi_level_features=multilevel_aux_visual_feats,
+                    vt_multi_level_features=multilevel_visual_feats,
+                    aux_boxes=[boxes],
+                    vt_boxes=[vt_boxes]
+                ).squeeze(0).to(multilevel_aux_visual_feats[0].dtype)
+                region_feat = self.get_model().mm_projector_aux(extracted_region_feat) # [num_bbox, 2048]
+                region_features.append(region_feat)
+        else:
+            # Extract region features only from auxiliary vision tower
+            for batch_idx, image_features in enumerate(aux_image_features_list):
+                multilevel_visual_feats = image_features["image_features"]
+                last_feat = image_features["last_feat"]
+                boxes = bbox_list[batch_idx]
+
+                if boxes is None or len(boxes) == 0:
+                    boxes = torch.tensor([[0, 10, 0, 10]], device=multilevel_visual_feats[0].device, dtype=torch.float32)
+               
+                multi_level_aux_features = multilevel_visual_feats
+                boxes = boxes.to(torch.float32).to(multi_level_aux_features[0].device)
+                extracted_region_feat = self.get_model().object_vp_extractor(
+                    multi_level_aux_features,
+                    [boxes],
+                ).squeeze(0).to(multi_level_aux_features[0].dtype)
+                region_feat = self.get_model().mm_projector_aux(extracted_region_feat) # [num_bbox, 2880]
+                region_features.append(region_feat)
+
+        return region_features
+    
+    def get_model(self):
+        # Getter for model. Used to access backbone/model internals.
+        return self.model
+
+    # Convert sequence of input_ids/labels/images/boxes to multimodal embedding and associated masks/ids for transformer input.
+    def prepare_inputs_labels_for_qwen2_5_vl_multimodal(
+        self, input_ids, position_ids, attention_mask, past_key_values, labels, images, images_aux=None, bbox_list=None, image_grid_thws=None
+    ):
+        # ========================== Above this line, input parsing and batching =============================
+        vision_tower = self.get_vision_tower()
+        video_tower = self.get_video_tower()
+        vision_tower_aux = self.get_vision_tower_aux()
+        # Fast-path for non-multimodal case or first step in generation (i.e. only one token in input)
+        if (vision_tower is None and video_tower is None) or images is None or input_ids.shape[1] == 1:
+            if past_key_values is not None and (vision_tower is not None or video_tower is not None) and images is not None and input_ids.shape[1] == 1:
+
+                target_shape = past_key_values[-1][-1].shape[-2] + 1
+                attention_mask = torch.cat((attention_mask, torch.ones(
+                    (attention_mask.shape[0], target_shape - attention_mask.shape[1]),
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device
+                )), dim=1)
+
+                position_ids=None
+                cache_position = torch.tensor([target_shape - 1],device=attention_mask.device)
+            return input_ids, position_ids, attention_mask, past_key_values, None, labels, None, cache_position
+
+        # Indices for images (3D or 2D tensors) and videos (4D tensors)
+        image_idx = [idx for idx, img in enumerate(images) if img.ndim == 3 or img.ndim == 2]
+        is_all_image = len(image_idx) == len(images)
+        video_idx = [idx for idx, vid in enumerate(images) if vid.ndim == 4]
+        
+        # Stack image and video tensors accordingly for mini-batch processing
+        if isinstance(vision_tower, Qwen2_5_VlVisionTower):
+            images_minibatch = [images[idx] for idx in image_idx] if len(image_idx) > 0 else []  # list of [c,h,w], can have variable shapes
+        else:
+            images_minibatch = torch.stack([images[idx] for idx in image_idx]) if len(image_idx) > 0 else []  # tensor [mini_b, c, h, w]
+        videos_minibatch = torch.stack([images[idx] for idx in video_idx]) if len(video_idx) > 0 else []  # tensor [mini_b, c, t, h, w]
+
+        # Auxiliary batch for region encoding, if relevant
+        if vision_tower_aux is not None and images_aux is not None:
+            images_minibatch_aux = [images_aux[idx].unsqueeze(0) for idx in image_idx] if len(image_idx) > 0 else []  # list of [1, c, h, w]
+
+        # tmp_image_features will be indexed to scatter extracted image/video features into original batch positions
+        tmp_image_features = [None] * (len(image_idx) + len(video_idx))
+        if getattr(images_minibatch, 'ndim', 0) == 4 or (type(images_minibatch) is list and len(images_minibatch) > 0):  # batch consists of images, [mini_b, c, h, w]
+            if vision_tower is not None:
+                image_features_minibatch, image_grid_thws_minibatch, vt_multi_level_features_minibatch = self.encode_images(images_minibatch, image_grid_thws)  # [mini_b, l, c]
+            else:
+                image_features_minibatch = torch.randn(1).to(self.device)  # dummy feature for video-only training under tuning
+
+            # Map extracted image features back to their places in the original batch
+            for i, pos in enumerate(image_idx):
+                tmp_image_features[pos] = image_features_minibatch[i]
+            
+            # Handle auxiliary region features if enabled and boxes provided
+            if vision_tower_aux is not None and bbox_list is not None and len(bbox_list) > 0:
+                if isinstance(self.get_model().get_vision_tower(), Qwen2_5_VlVisionTower):
+                    patch_size = self.get_model().get_vision_tower().config.patch_size
+                    vt_images_size_minibatch = [im_grid_thw[0][-2:]*patch_size for im_grid_thw in image_grid_thws]
+                    region_features = self.encode_regions(images_minibatch_aux, bbox_list, vt_multi_level_features_minibatch, vt_images_size_minibatch)  # [mini_b, l, c]
+            else:
+                region_features = None
+
+        # Same as above, but for video features if any
+        if getattr(videos_minibatch, 'ndim', 0) == 5:  # batch consists of videos, [mini_b, c, t, h, w]
+            video_features_minibatch = self.encode_videos(videos_minibatch)  # fake list [mini_b, t, l, c]
+            for i, pos in enumerate(video_idx):
+                tmp_image_features[pos] = video_features_minibatch[i]
+
+        # Flatten image feature slot list to proper order for current batch
+        new_tmp = []
+        for image in tmp_image_features:
+            # If multi-image per item, flatten out
+            if isinstance(image, list):
+                t = len(image)
+                for i in range(t):
+                    new_tmp.append(image[i])
+            else:
+                new_tmp.append(image)
+        image_features = new_tmp
+
+        # =========================== Now, build multimodal input & target sequences =========================
+
+        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
+            raise NotImplementedError
+
+        _labels = labels
+        _position_ids = position_ids
+        _attention_mask = attention_mask
+
+        # Default construction of masks etc.
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
+        else:
+            attention_mask = attention_mask.bool()
+        if position_ids is None:
+            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
+        if labels is None:
+           labels = torch.full_like(input_ids, IGNORE_INDEX)
+
+        # For each batch item, strip padded tokens based on attention_mask
+        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
+        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
+
+        # If neither region auxiliary nor bboxes present: process classic image-text input
+        if vision_tower_aux is None and (bbox_list is None or all(x is None for x in bbox_list)):
+            new_input_embeds = []
+            new_labels = []
+            new_input_ids = []
+            cur_image_idx = 0
+            image_nums_in_batch = []
+            
+            for batch_idx, cur_input_ids in enumerate(input_ids):
+                num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
+                image_nums_in_batch.append(num_images)
+                # If there are no image markers, just get text features
+                if num_images == 0:
+                    cur_image_features = image_features[cur_image_idx]
+                    cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
+                    cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
+                    new_input_embeds.append(cur_input_embeds)
+                    new_labels.append(labels[batch_idx])
+                    new_input_ids.append(cur_input_ids)
+                    cur_image_idx += 1
+                    continue
+
+                # Split on image token indices: replace them with image features after conversion
+                image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
+                cur_input_ids_noim = []
+                cur_labels = labels[batch_idx]
+                cur_labels_noim = []
+                for i in range(len(image_token_indices) - 1):
+                    cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
+                    cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
+                split_sizes = [x.shape[0] for x in cur_labels_noim]
+                cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
+                cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
+
+                cur_new_input_embeds = []
+                cur_new_labels = []
+                cur_new_input_ids = []
+                for i in range(num_images + 1):
+                    # Interleave text and image features
+                    cur_new_input_embeds.append(cur_input_embeds_no_im[i])
+                    cur_new_labels.append(cur_labels_noim[i])
+                    cur_new_input_ids.append(cur_input_ids_noim[i])
+                    if i < num_images:
+                        cur_image_features = image_features[cur_image_idx].to(self.device)
+                        cur_image_idx += 1
+                        cur_new_input_embeds.append(cur_image_features)
+                        cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
+                        cur_new_input_ids.append(torch.full((cur_image_features.shape[0],), self.config.image_token_id, device=cur_labels.device, dtype=cur_labels.dtype))
+                cur_new_input_embeds = torch.cat(cur_new_input_embeds)
+                cur_new_labels = torch.cat(cur_new_labels)
+                cur_new_input_ids = torch.cat(cur_new_input_ids)
+
+                new_input_embeds.append(cur_new_input_embeds)
+                new_labels.append(cur_new_labels)
+                new_input_ids.append(cur_new_input_ids)
+        # If region markers or region features enabled in config
+        else:
+            new_input_embeds = []
+            new_labels = []
+            new_input_ids = []
+            cur_image_idx = 0
+            image_nums_in_batch = []
+            
+            for batch_idx, cur_input_ids in enumerate(input_ids):
+                cur_region_idx = 0 
+                # Detect image and region special token counts
+                num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
+                num_regions = (cur_input_ids == DEFAULT_REGION_INDEX).sum() if DEFAULT_REGION_INDEX in cur_input_ids else 0
+                image_nums_in_batch.append(num_images)
+
+                # If no markers, just do text embedding for this item
+                if num_images == 0 and num_regions == 0:
+                    cur_image_features = image_features[cur_image_idx]
+                    cur_region_features = region_features[cur_region_idx]
+                    cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
+                    cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0], cur_region_features[0:0]], dim=0)
+                    new_input_embeds.append(cur_input_embeds)
+                    new_labels.append(labels[batch_idx])
+                    new_input_ids.append(cur_input_ids)
+                    cur_image_idx += 1
+                    continue
+
+                # Get all special marker indices (image/region)
+                image_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist()
+                region_indices = torch.where(cur_input_ids == DEFAULT_REGION_INDEX)[0].tolist() if num_regions > 0 else []
+                all_special_indices = sorted([-1] + image_indices + region_indices + [cur_input_ids.shape[0]])
+
+                # Split out plain text chunks between special markers
+                cur_input_ids_segments = []
+                cur_labels = labels[batch_idx]
+                cur_labels_segments = []
+                
+                for i in range(len(all_special_indices) - 1):
+                    cur_input_ids_segments.append(cur_input_ids[all_special_indices[i]+1:all_special_indices[i+1]])
+                    cur_labels_segments.append(cur_labels[all_special_indices[i]+1:all_special_indices[i+1]])
+
+                # Project text ids to word embeddings
+                split_sizes = [x.shape[0] for x in cur_labels_segments]
+                cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_segments))
+                if num_regions == 0 and vision_tower_aux is not None and region_features is not None:
+                    cur_region_features = region_features[cur_region_idx]
+                    temp_input_embeds = torch.cat([cur_input_embeds, cur_region_features[0:0]], dim=0)
+                    cur_input_embeds = temp_input_embeds
+                
+                cur_input_embeds_segments = torch.split(cur_input_embeds, split_sizes, dim=0)
+
+                # Reassemble text and image/region segments in order
+                cur_new_input_embeds = []
+                cur_new_labels = []
+                cur_new_input_ids = []
+                
+                for i in range(len(all_special_indices) - 1):
+                    # Insert current text segment
+                    cur_new_input_embeds.append(cur_input_embeds_segments[i])
+                    cur_new_labels.append(cur_labels_segments[i])
+                    cur_new_input_ids.append(cur_input_ids_segments[i])
+                    # If next is image, insert feature representation
+                    if all_special_indices[i+1] in image_indices:
+                        cur_image_features = image_features[cur_image_idx].to(self.device)
+                        cur_image_idx += 1
+                        cur_new_input_embeds.append(cur_image_features)
+                        cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
+                        cur_new_input_ids.append(torch.full((cur_image_features.shape[0],), self.config.image_token_id, device=cur_labels.device, dtype=cur_labels.dtype))
+                        
+                    # If next is region token, insert extracted region features
+                    elif all_special_indices[i+1] in region_indices:
+                        cur_region_features = region_features[batch_idx][cur_region_idx].to(self.device).unsqueeze(0)
+                        cur_region_idx += 1
+                        cur_new_input_embeds.append(cur_region_features)
+                        
+                        cur_new_labels.append(torch.full((cur_region_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
+                        cur_new_input_ids.append(torch.full((cur_region_features.shape[0],), DEFAULT_REGION_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
+                # Combine for this batch item
+                cur_new_input_embeds = torch.cat(cur_new_input_embeds)
+                cur_new_labels = torch.cat(cur_new_labels)
+                cur_new_input_ids = torch.cat(cur_new_input_ids)
+                new_input_embeds.append(cur_new_input_embeds)
+                new_labels.append(cur_new_labels)
+                new_input_ids.append(cur_new_input_ids)
+        # Truncate sequences to maximum model length, if image+region tokens caused overflow
+        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
+        if tokenizer_model_max_length is not None:
+            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
+            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
+
+        # Pad sequences in the batch to same length; compute batch masks
+        max_len = max(x.shape[0] for x in new_input_embeds)
+        batch_size = len(new_input_embeds)
+
+        new_input_embeds_padded = []
+        new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
+        new_input_ids_padded = torch.full((batch_size, max_len), self.config.bos_token_id, dtype=new_input_ids[0].dtype, device=new_input_ids[0].device)
+        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
+        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
+
+        # Left or right padding as per config; fill padded tensors
+        for i, (cur_new_embed, cur_new_labels, cur_new_input_ids) in enumerate(zip(new_input_embeds, new_labels, new_input_ids)):
+            cur_len = cur_new_embed.shape[0]
+            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
+                # Left pad: add zeros before text tokens/features
+                new_input_embeds_padded.append(torch.cat((
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
+                    cur_new_embed
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, -cur_len:] = cur_new_labels
+                    attention_mask[i, -cur_len:] = True
+                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+            else:
+                # Right pad: add zeros after text tokens/features
+                new_input_embeds_padded.append(torch.cat((
+                    cur_new_embed,
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, :cur_len] = cur_new_labels
+                    new_input_ids_padded[i, :cur_len] = cur_new_input_ids
+                    attention_mask[i, :cur_len] = True
+                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+
+        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
+        new_input_ids = new_input_ids_padded
+
+        # Only set new_labels if original labels were not None
+        if _labels is None:
+            new_labels = None
+        else:
+            new_labels = new_labels_padded
+
+        # Similarly handle provided attention_mask/position_ids overrides
+        if _attention_mask is None:
+            attention_mask = None
+        else:
+            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
+
+        if _position_ids is None:
+            position_ids = None
+
+        # For Qwen2.5 vision towers, use and concatenate image_grid_thws for positional computations
+        if isinstance(self.get_model().get_vision_tower(), Qwen2_5_VlVisionTower):
+            image_grid_thws = []
+            cur_image_idx = 0
+            for num_images in image_nums_in_batch:
+                if num_images == 0:
+                    cur_image_idx += 1
+                    continue
+                image_grid_thws += image_grid_thws_minibatch[cur_image_idx:cur_image_idx+num_images]
+                cur_image_idx += num_images
+            
+            if len(image_grid_thws) > 0:
+                image_grid_thws = torch.cat(image_grid_thws, dim=0)
+            else:
+                image_grid_thws = None
+            
+            rope_index_kwargs = {
+                "input_ids": new_input_ids,
+                "image_grid_thw": image_grid_thws,
+                "video_grid_thw": None,
+                "attention_mask": attention_mask,
+            }
+
+            # Compute new position_ids and rope_deltas for transformer (for rotary embeddings)
+            position_ids, rope_deltas = self.get_rope_index(**rope_index_kwargs)
+            cache_position = torch.arange(new_input_embeds.shape[1], device=new_input_embeds.device)
+        else:
+            rope_deltas = None
+            cache_position = None
+        # Final output is a tuple mimicking HuggingFace prepare_inputs_for_generation return
+        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels, rope_deltas, cache_position
+    
+    # Patch forward() of HF CausalLM to allow multimodal embedding with images/regions
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        pixel_values_videos: Optional[torch.FloatTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        rope_deltas: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        second_per_grid_ts: Optional[torch.Tensor] = None,
+        images: Optional[torch.FloatTensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        bbox_list: Optional[torch.FloatTensor] = None,
+        image_grid_thws: Optional[torch.FloatTensor] = None,
+    ) -> Union[Tuple, Qwen2_5_VLCausalLMOutputWithPast]:
+        
+        if inputs_embeds is None:
+            (
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                inputs_embeds,
+                labels,
+                rope_deltas,
+                cache_position
+            ) = self.prepare_inputs_labels_for_qwen2_5_vl_multimodal(
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                labels,
+                images,
+                images_aux,
+                bbox_list,
+                image_grid_thws
+            )
+        
+        if rope_deltas is not None:
+            self.rope_deltas = rope_deltas
+
+        # Call base CausalLM forward, with possibly replaced multimodal embeddings
+        out = super().forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            rope_deltas=rope_deltas,
+            cache_position=cache_position,
+            second_per_grid_ts=second_per_grid_ts,
+            return_dict=return_dict
+        )
+        return out
+
+    # Prepare model input dict for autoregressive generation (for use with generation methods like generate())
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        pixel_values=None,
+        pixel_values_videos=None,
+        image_grid_thw=None,
+        video_grid_thw=None,
+        second_per_grid_ts=None,
+        images: Optional[torch.FloatTensor] = None,
+        images_aux: Optional[torch.FloatTensor] = None,
+        bbox_list: Optional[torch.FloatTensor] = None,
+        image_grid_thws: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ):
+        # Wrap parent logic so extra multimodal kwargs are preserved
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            cache_position=cache_position,
+            pixel_values=pixel_values,
+            pixel_values_videos=pixel_values_videos,
+            image_grid_thw=image_grid_thw,
+            video_grid_thw=video_grid_thw,
+            second_per_grid_ts=second_per_grid_ts,
+            images=images,
+            images_aux=images_aux,
+            bbox_list=bbox_list,
+            image_grid_thws=image_grid_thws,
+        )   
+        return model_inputs
+    
+# Register our config and model with HuggingFace transformers registry
+AutoConfig.register("omchat_qwen2_5_vl", OmChatQwen25VLConfig)
+AutoModelForCausalLM.register(OmChatQwen25VLConfig, OmChatQwen25VLForCausalLM)

vlm_fo1/model/multimodal_encoder/base_encoder.py

@@ -0,0 +1,33 @@
+import torch
+import torch.nn as nn
+
+
+class AbsVisionTower(nn.Module):
+    @torch.no_grad()
+    def forward(self, images):
+        raise NotImplementedError
+
+    @property
+    def dummy_feature(self):
+        raise NotImplementedError
+
+    @property
+    def dtype(self):
+        raise NotImplementedError
+
+    @property
+    def device(self):
+        raise NotImplementedError
+
+    @property
+    def config(self):
+        raise NotImplementedError
+
+
+    @property
+    def hidden_size(self):
+        raise NotImplementedError
+
+    @property
+    def num_patches(self):
+        raise NotImplementedError

vlm_fo1/model/multimodal_encoder/builder.py

@@ -0,0 +1,38 @@
+# Builders for different vision tower backbones (MM encoder visual modules)
+from .qwen2_5_vl_encoder import Qwen2_5_VlVisionTower    # Main Qwen2.5 vision tower
+from .davit_aux_encoder import DavitVisionTower as DavitVisionTowerAux  # Auxiliary DaViT vision tower
+
+def build_vision_tower(vision_tower_cfg, **kwargs):
+    """
+    Use model config to construct the main vision tower.
+
+    vision_tower_cfg: should have attribute mm_vision_tower
+    Returns: instance of configured vision backbone
+    """
+    vision_tower_name = getattr(vision_tower_cfg, 'mm_vision_tower', None)
+    # print(vision_tower_cfg)  # Debug print of the config being used
+    
+    # Check for the Qwen2.5-VL vision model in tower name
+    if "qwen2.5-vl" in vision_tower_name.lower():
+        return Qwen2_5_VlVisionTower(vision_tower_name, args=vision_tower_cfg, **kwargs) 
+
+    # Raise a clear error for unknown towers
+    raise ValueError(f'Unknown vision tower: {vision_tower_name}')
+
+def build_vision_tower_aux(vision_tower_cfg, **kwargs):
+    """
+    Use model config to construct the auxiliary (helper) vision tower.
+
+    vision_tower_cfg: should have attribute mm_vision_tower_aux
+    Returns: instance of configured auxiliary vision backbone
+    """
+    vision_tower_aux = getattr(vision_tower_cfg, 'mm_vision_tower_aux', None)
+    # Optionally print config for debugging
+    # print(vision_tower_cfg)
+
+    # Check for the DaViT auxiliary vision model in tower name
+    if 'davit' in vision_tower_aux.lower():
+        return DavitVisionTowerAux(vision_tower_aux, args=vision_tower_cfg, **kwargs)
+
+    # Raise a clear error if tower type is unknown
+    raise ValueError(f'Unknown aux vision tower: {vision_tower_aux}')