Upload 2711 files

.gitattributes

@@ -33,3 +33,17 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+Perceptrix/finetune/scripts/eval/local_data/reading_comprehension/coqa.jsonl filter=lfs diff=lfs merge=lfs -text
+Perceptrix/finetune/scripts/eval/local_data/reading_comprehension/narrative_qa.jsonl filter=lfs diff=lfs merge=lfs -text
+Perceptrix/finetune/scripts/eval/local_data/symbolic_problem_solving/bigbench_elementary_math_qa.jsonl filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/docs/source/nlp/dialogue_UML.png filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/docs/source/nlp/nemo_megatron/images/ddp.gif filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/docs/source/nlp/nemo_megatron/images/pnom.gif filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/docs/source/nlp/nemo_megatron/images/pp.gif filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/docs/source/nlp/nemo_megatron/images/tp.gif filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/docs/source/tts/images/fastpitch_model.png filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/docs/source/tts/images/radaligner_model.png filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/docs/source/tts/images/tacotron2_model.png filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/docs/source/tts/images/waveglow_model.png filter=lfs diff=lfs merge=lfs -text
+SoundScribe/SpeakerID/tools/speech_data_explorer/screenshot.png filter=lfs diff=lfs merge=lfs -text
+SoundScribe/voices/Vatsal.wav filter=lfs diff=lfs merge=lfs -text

CircumSpect/__init__.py

@@ -0,0 +1,2 @@
+from CircumSpect.vqa.vqa import answer_question, find_object_description
+from CircumSpect.object_detection.detect_objects import locate_object

CircumSpect/object_detection/detect_objects.py

@@ -0,0 +1,87 @@
+from CircumSpect.object_detection.groundingdino.util.inference import load_model, predict
+import CircumSpect.object_detection.groundingdino.datasets.transforms as T
+from torchvision.ops import box_convert
+from utils import setup_device
+from typing import Tuple, List
+import supervision as sv
+from io import BytesIO
+from PIL import Image
+import numpy as np
+import requests
+import torch
+import cv2
+
+model_v = load_model("CircumSpect/object_detection/groundingdino/config/GroundingDINO_SwinT_OGC.py",
+                     "CircumSpect/object_detection/weights/groundingdino_swint_ogc.pth")
+
+
+def load_image(image_path: str) -> Tuple[np.array, torch.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]),
+        ]
+    )
+    image_source = Image.open(BytesIO(requests.get(
+        image_path).content) if image_path.startswith("http") else image_path).convert("RGB")
+    image = np.asarray(image_source)
+    image_transformed, _ = transform(image_source, None)
+    return image, image_transformed
+
+
+def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor, phrases: List[str]) -> np.ndarray:
+    h, w, _ = image_source.shape
+    boxes = boxes * torch.Tensor([w, h, w, h])
+    xyxy = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xyxy").numpy()
+    detections = sv.Detections(xyxy=xyxy)
+
+    labels = [
+        f"{phrase} {logit:.2f}"
+        for phrase, logit
+        in zip(phrases, logits)
+    ]
+    coordinates = []
+    for i in list(detections):
+        left, top, right, bottom = list(list(i)[0])
+        coordinates.append((int((left+right)/2), int((top+bottom)/2)))
+
+    object_coordinates = [[" ".join(object_name.split()[:-1])+": "+object_name.split(
+    )[-1]+"%", coordinate] for object_name, coordinate in zip(labels, coordinates)]
+
+    box_annotator = sv.BoxAnnotator()
+    annotated_frame = cv2.cvtColor(image_source, cv2.COLOR_RGB2BGR)
+    annotated_frame = box_annotator.annotate(
+        scene=annotated_frame, detections=detections, labels=labels)
+    return annotated_frame, object_coordinates
+
+
+BOX_TRESHOLD = 0.35
+TEXT_TRESHOLD = 0.25
+
+device = setup_device()
+
+
+def locate_object(objects, image):
+    image_source, image = load_image(image)
+
+    boxes, logits, phrases = predict(
+        model=model_v,
+        image=image,
+        device=device,
+        caption=objects,
+        box_threshold=BOX_TRESHOLD,
+        text_threshold=TEXT_TRESHOLD
+    )
+
+    annotated_frame, object_coordinates = annotate(
+        image_source=image_source, boxes=boxes, logits=logits, phrases=phrases)
+    cv2.circle(annotated_frame, object_coordinates[0][1], 2, (255, 0, 0), 2)
+    cv2.imwrite("detected_objects.png", annotated_frame)
+    return annotated_frame, object_coordinates
+
+
+if __name__ == "__main__":
+    frame, coord = locate_object(
+        "https://images.nationalgeographic.org/image/upload/v1638890052/EducationHub/photos/robots-3d-landing-page.jpg", "drawer")
+    print(coord)

CircumSpect/object_detection/groundingdino/config/GroundingDINO_SwinB_cfg.py

@@ -0,0 +1,43 @@
+batch_size = 1
+modelname = "groundingdino"
+backbone = "swin_B_384_22k"
+position_embedding = "sine"
+pe_temperatureH = 20
+pe_temperatureW = 20
+return_interm_indices = [1, 2, 3]
+backbone_freeze_keywords = None
+enc_layers = 6
+dec_layers = 6
+pre_norm = False
+dim_feedforward = 2048
+hidden_dim = 256
+dropout = 0.0
+nheads = 8
+num_queries = 900
+query_dim = 4
+num_patterns = 0
+num_feature_levels = 4
+enc_n_points = 4
+dec_n_points = 4
+two_stage_type = "standard"
+two_stage_bbox_embed_share = False
+two_stage_class_embed_share = False
+transformer_activation = "relu"
+dec_pred_bbox_embed_share = True
+dn_box_noise_scale = 1.0
+dn_label_noise_ratio = 0.5
+dn_label_coef = 1.0
+dn_bbox_coef = 1.0
+embed_init_tgt = True
+dn_labelbook_size = 2000
+max_text_len = 256
+text_encoder_type = "bert-base-uncased"
+use_text_enhancer = True
+use_fusion_layer = True
+use_checkpoint = True
+use_transformer_ckpt = True
+use_text_cross_attention = True
+text_dropout = 0.0
+fusion_dropout = 0.0
+fusion_droppath = 0.1
+sub_sentence_present = True

CircumSpect/object_detection/groundingdino/config/GroundingDINO_SwinT_OGC.py

@@ -0,0 +1,43 @@
+batch_size = 1
+modelname = "groundingdino"
+backbone = "swin_T_224_1k"
+position_embedding = "sine"
+pe_temperatureH = 20
+pe_temperatureW = 20
+return_interm_indices = [1, 2, 3]
+backbone_freeze_keywords = None
+enc_layers = 6
+dec_layers = 6
+pre_norm = False
+dim_feedforward = 2048
+hidden_dim = 256
+dropout = 0.0
+nheads = 8
+num_queries = 900
+query_dim = 4
+num_patterns = 0
+num_feature_levels = 4
+enc_n_points = 4
+dec_n_points = 4
+two_stage_type = "standard"
+two_stage_bbox_embed_share = False
+two_stage_class_embed_share = False
+transformer_activation = "relu"
+dec_pred_bbox_embed_share = True
+dn_box_noise_scale = 1.0
+dn_label_noise_ratio = 0.5
+dn_label_coef = 1.0
+dn_bbox_coef = 1.0
+embed_init_tgt = True
+dn_labelbook_size = 2000
+max_text_len = 256
+text_encoder_type = "bert-base-uncased"
+use_text_enhancer = True
+use_fusion_layer = True
+use_checkpoint = True
+use_transformer_ckpt = True
+use_text_cross_attention = True
+text_dropout = 0.0
+fusion_dropout = 0.0
+fusion_droppath = 0.1
+sub_sentence_present = True

CircumSpect/object_detection/groundingdino/datasets/cocogrounding_eval.py

@@ -0,0 +1,269 @@
+# ------------------------------------------------------------------------
+# Grounding DINO. Midified by Shilong Liu.
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copyright (c) Aishwarya Kamath & Nicolas Carion. Licensed under the Apache License 2.0. All Rights Reserved
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+COCO evaluator that works in distributed mode.
+
+Mostly copy-paste from https://github.com/pytorch/vision/blob/edfd5a7/references/detection/coco_eval.py
+The difference is that there is less copy-pasting from pycocotools
+in the end of the file, as python3 can suppress prints with contextlib
+"""
+import contextlib
+import copy
+import os
+
+import numpy as np
+import pycocotools.mask as mask_util
+import torch
+from pycocotools.coco import COCO
+from pycocotools.cocoeval import COCOeval
+
+from groundingdino.util.misc import all_gather
+
+
+class CocoGroundingEvaluator(object):
+    def __init__(self, coco_gt, iou_types, useCats=True):
+        assert isinstance(iou_types, (list, tuple))
+        coco_gt = copy.deepcopy(coco_gt)
+        self.coco_gt = coco_gt
+
+        self.iou_types = iou_types
+        self.coco_eval = {}
+        for iou_type in iou_types:
+            self.coco_eval[iou_type] = COCOeval(coco_gt, iouType=iou_type)
+            self.coco_eval[iou_type].useCats = useCats
+
+        self.img_ids = []
+        self.eval_imgs = {k: [] for k in iou_types}
+        self.useCats = useCats
+
+    def update(self, predictions):
+        img_ids = list(np.unique(list(predictions.keys())))
+        self.img_ids.extend(img_ids)
+
+        for iou_type in self.iou_types:
+            results = self.prepare(predictions, iou_type)
+
+            # suppress pycocotools prints
+            with open(os.devnull, "w") as devnull:
+                with contextlib.redirect_stdout(devnull):
+                    coco_dt = COCO.loadRes(self.coco_gt, results) if results else COCO()
+
+            coco_eval = self.coco_eval[iou_type]
+
+            coco_eval.cocoDt = coco_dt
+            coco_eval.params.imgIds = list(img_ids)
+            coco_eval.params.useCats = self.useCats
+            img_ids, eval_imgs = evaluate(coco_eval)
+
+            self.eval_imgs[iou_type].append(eval_imgs)
+
+    def synchronize_between_processes(self):
+        for iou_type in self.iou_types:
+            self.eval_imgs[iou_type] = np.concatenate(self.eval_imgs[iou_type], 2)
+            create_common_coco_eval(self.coco_eval[iou_type], self.img_ids, self.eval_imgs[iou_type])
+
+    def accumulate(self):
+        for coco_eval in self.coco_eval.values():
+            coco_eval.accumulate()
+
+    def summarize(self):
+        for iou_type, coco_eval in self.coco_eval.items():
+            print("IoU metric: {}".format(iou_type))
+            coco_eval.summarize()
+
+    def prepare(self, predictions, iou_type):
+        if iou_type == "bbox":
+            return self.prepare_for_coco_detection(predictions)
+        elif iou_type == "segm":
+            return self.prepare_for_coco_segmentation(predictions)
+        elif iou_type == "keypoints":
+            return self.prepare_for_coco_keypoint(predictions)
+        else:
+            raise ValueError("Unknown iou type {}".format(iou_type))
+
+    def prepare_for_coco_detection(self, predictions):
+        coco_results = []
+        for original_id, prediction in predictions.items():
+            if len(prediction) == 0:
+                continue
+
+            boxes = prediction["boxes"]
+            boxes = convert_to_xywh(boxes).tolist()
+            scores = prediction["scores"].tolist()
+            labels = prediction["labels"].tolist()
+
+            coco_results.extend(
+                [
+                    {
+                        "image_id": original_id,
+                        "category_id": labels[k],
+                        "bbox": box,
+                        "score": scores[k],
+                    }
+                    for k, box in enumerate(boxes)
+                ]
+            )
+        return coco_results
+
+    def prepare_for_coco_segmentation(self, predictions):
+        coco_results = []
+        for original_id, prediction in predictions.items():
+            if len(prediction) == 0:
+                continue
+
+            scores = prediction["scores"]
+            labels = prediction["labels"]
+            masks = prediction["masks"]
+
+            masks = masks > 0.5
+
+            scores = prediction["scores"].tolist()
+            labels = prediction["labels"].tolist()
+
+            rles = [
+                mask_util.encode(np.array(mask[0, :, :, np.newaxis], dtype=np.uint8, order="F"))[0] 
+                for mask in masks
+            ]
+            for rle in rles:
+                rle["counts"] = rle["counts"].decode("utf-8")
+
+            coco_results.extend(
+                [
+                    {
+                        "image_id": original_id,
+                        "category_id": labels[k],
+                        "segmentation": rle,
+                        "score": scores[k],
+                    }
+                    for k, rle in enumerate(rles)
+                ]
+            )
+        return coco_results
+
+    def prepare_for_coco_keypoint(self, predictions):
+        coco_results = []
+        for original_id, prediction in predictions.items():
+            if len(prediction) == 0:
+                continue
+
+            boxes = prediction["boxes"]
+            boxes = convert_to_xywh(boxes).tolist()
+            scores = prediction["scores"].tolist()
+            labels = prediction["labels"].tolist()
+            keypoints = prediction["keypoints"]
+            keypoints = keypoints.flatten(start_dim=1).tolist()
+
+            coco_results.extend(
+                [
+                    {
+                        "image_id": original_id,
+                        "category_id": labels[k],
+                        "keypoints": keypoint,
+                        "score": scores[k],
+                    }
+                    for k, keypoint in enumerate(keypoints)
+                ]
+            )
+        return coco_results
+
+
+def convert_to_xywh(boxes):
+    xmin, ymin, xmax, ymax = boxes.unbind(1)
+    return torch.stack((xmin, ymin, xmax - xmin, ymax - ymin), dim=1)
+
+
+def merge(img_ids, eval_imgs):
+    all_img_ids = all_gather(img_ids)
+    all_eval_imgs = all_gather(eval_imgs)
+
+    merged_img_ids = []
+    for p in all_img_ids:
+        merged_img_ids.extend(p)
+
+    merged_eval_imgs = []
+    for p in all_eval_imgs:
+        merged_eval_imgs.append(p)
+
+    merged_img_ids = np.array(merged_img_ids)
+    merged_eval_imgs = np.concatenate(merged_eval_imgs, 2)
+
+    # keep only unique (and in sorted order) images
+    merged_img_ids, idx = np.unique(merged_img_ids, return_index=True)
+    merged_eval_imgs = merged_eval_imgs[..., idx]
+
+    return merged_img_ids, merged_eval_imgs
+
+
+def create_common_coco_eval(coco_eval, img_ids, eval_imgs):
+    img_ids, eval_imgs = merge(img_ids, eval_imgs)
+    img_ids = list(img_ids)
+    eval_imgs = list(eval_imgs.flatten())
+
+    coco_eval.evalImgs = eval_imgs
+    coco_eval.params.imgIds = img_ids
+    coco_eval._paramsEval = copy.deepcopy(coco_eval.params)
+
+
+#################################################################
+# From pycocotools, just removed the prints and fixed
+# a Python3 bug about unicode not defined
+#################################################################
+
+
+def evaluate(self):
+    """
+    Run per image evaluation on given images and store results (a list of dict) in self.evalImgs
+    :return: None
+    """
+    # tic = time.time()
+    # print('Running per image evaluation...')
+    p = self.params
+    # add backward compatibility if useSegm is specified in params
+    if p.useSegm is not None:
+        p.iouType = "segm" if p.useSegm == 1 else "bbox"
+        print("useSegm (deprecated) is not None. Running {} evaluation".format(p.iouType))
+    # print('Evaluate annotation type *{}*'.format(p.iouType))
+    p.imgIds = list(np.unique(p.imgIds))
+    if p.useCats:
+        p.catIds = list(np.unique(p.catIds))
+    p.maxDets = sorted(p.maxDets)
+    self.params = p
+
+    self._prepare()
+    # loop through images, area range, max detection number
+    catIds = p.catIds if p.useCats else [-1]
+
+    if p.iouType == "segm" or p.iouType == "bbox":
+        computeIoU = self.computeIoU
+    elif p.iouType == "keypoints":
+        computeIoU = self.computeOks
+    self.ious = {
+        (imgId, catId): computeIoU(imgId, catId) 
+        for imgId in p.imgIds 
+        for catId in catIds}
+
+    evaluateImg = self.evaluateImg
+    maxDet = p.maxDets[-1]
+    evalImgs = [
+        evaluateImg(imgId, catId, areaRng, maxDet) 
+        for catId in catIds 
+        for areaRng in p.areaRng 
+        for imgId in p.imgIds
+    ]
+    # this is NOT in the pycocotools code, but could be done outside
+    evalImgs = np.asarray(evalImgs).reshape(len(catIds), len(p.areaRng), len(p.imgIds))
+    self._paramsEval = copy.deepcopy(self.params)
+    # toc = time.time()
+    # print('DONE (t={:0.2f}s).'.format(toc-tic))
+    return p.imgIds, evalImgs
+
+
+#################################################################
+# end of straight copy from pycocotools, just removing the prints
+#################################################################

CircumSpect/object_detection/groundingdino/datasets/transforms.py

@@ -0,0 +1,311 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+Transforms and data augmentation for both image + bbox.
+"""
+import os
+import random
+
+import PIL
+import torch
+import torchvision.transforms as T
+import torchvision.transforms.functional as F
+
+from CircumSpect.object_detection.groundingdino.util.box_ops import box_xyxy_to_cxcywh
+from CircumSpect.object_detection.groundingdino.util.misc import interpolate
+
+
+def crop(image, target, region):
+    cropped_image = F.crop(image, *region)
+
+    target = target.copy()
+    i, j, h, w = region
+
+    # should we do something wrt the original size?
+    target["size"] = torch.tensor([h, w])
+
+    fields = ["labels", "area", "iscrowd", "positive_map"]
+
+    if "boxes" in target:
+        boxes = target["boxes"]
+        max_size = torch.as_tensor([w, h], dtype=torch.float32)
+        cropped_boxes = boxes - torch.as_tensor([j, i, j, i])
+        cropped_boxes = torch.min(cropped_boxes.reshape(-1, 2, 2), max_size)
+        cropped_boxes = cropped_boxes.clamp(min=0)
+        area = (cropped_boxes[:, 1, :] - cropped_boxes[:, 0, :]).prod(dim=1)
+        target["boxes"] = cropped_boxes.reshape(-1, 4)
+        target["area"] = area
+        fields.append("boxes")
+
+    if "masks" in target:
+        # FIXME should we update the area here if there are no boxes?
+        target["masks"] = target["masks"][:, i : i + h, j : j + w]
+        fields.append("masks")
+
+    # remove elements for which the boxes or masks that have zero area
+    if "boxes" in target or "masks" in target:
+        # favor boxes selection when defining which elements to keep
+        # this is compatible with previous implementation
+        if "boxes" in target:
+            cropped_boxes = target["boxes"].reshape(-1, 2, 2)
+            keep = torch.all(cropped_boxes[:, 1, :] > cropped_boxes[:, 0, :], dim=1)
+        else:
+            keep = target["masks"].flatten(1).any(1)
+
+        for field in fields:
+            if field in target:
+                target[field] = target[field][keep]
+
+    if os.environ.get("IPDB_SHILONG_DEBUG", None) == "INFO":
+        # for debug and visualization only.
+        if "strings_positive" in target:
+            target["strings_positive"] = [
+                _i for _i, _j in zip(target["strings_positive"], keep) if _j
+            ]
+
+    return cropped_image, target
+
+
+def hflip(image, target):
+    flipped_image = F.hflip(image)
+
+    w, h = image.size
+
+    target = target.copy()
+    if "boxes" in target:
+        boxes = target["boxes"]
+        boxes = boxes[:, [2, 1, 0, 3]] * torch.as_tensor([-1, 1, -1, 1]) + torch.as_tensor(
+            [w, 0, w, 0]
+        )
+        target["boxes"] = boxes
+
+    if "masks" in target:
+        target["masks"] = target["masks"].flip(-1)
+
+    return flipped_image, target
+
+
+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 "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])
+
+    if "masks" in target:
+        target["masks"] = (
+            interpolate(target["masks"][:, None].float(), size, mode="nearest")[:, 0] > 0.5
+        )
+
+    return rescaled_image, target
+
+
+def pad(image, target, padding):
+    # assumes that we only pad on the bottom right corners
+    padded_image = F.pad(image, (0, 0, padding[0], padding[1]))
+    if target is None:
+        return padded_image, None
+    target = target.copy()
+    # should we do something wrt the original size?
+    target["size"] = torch.tensor(padded_image.size[::-1])
+    if "masks" in target:
+        target["masks"] = torch.nn.functional.pad(target["masks"], (0, padding[0], 0, padding[1]))
+    return padded_image, target
+
+
+class ResizeDebug(object):
+    def __init__(self, size):
+        self.size = size
+
+    def __call__(self, img, target):
+        return resize(img, target, self.size)
+
+
+class RandomCrop(object):
+    def __init__(self, size):
+        self.size = size
+
+    def __call__(self, img, target):
+        region = T.RandomCrop.get_params(img, self.size)
+        return crop(img, target, region)
+
+
+class RandomSizeCrop(object):
+    def __init__(self, min_size: int, max_size: int, respect_boxes: bool = False):
+        # respect_boxes:    True to keep all boxes
+        #                   False to tolerence box filter
+        self.min_size = min_size
+        self.max_size = max_size
+        self.respect_boxes = respect_boxes
+
+    def __call__(self, img: PIL.Image.Image, target: dict):
+        init_boxes = len(target["boxes"])
+        max_patience = 10
+        for i in range(max_patience):
+            w = random.randint(self.min_size, min(img.width, self.max_size))
+            h = random.randint(self.min_size, min(img.height, self.max_size))
+            region = T.RandomCrop.get_params(img, [h, w])
+            result_img, result_target = crop(img, target, region)
+            if (
+                not self.respect_boxes
+                or len(result_target["boxes"]) == init_boxes
+                or i == max_patience - 1
+            ):
+                return result_img, result_target
+        return result_img, result_target
+
+
+class CenterCrop(object):
+    def __init__(self, size):
+        self.size = size
+
+    def __call__(self, img, target):
+        image_width, image_height = img.size
+        crop_height, crop_width = self.size
+        crop_top = int(round((image_height - crop_height) / 2.0))
+        crop_left = int(round((image_width - crop_width) / 2.0))
+        return crop(img, target, (crop_top, crop_left, crop_height, crop_width))
+
+
+class RandomHorizontalFlip(object):
+    def __init__(self, p=0.5):
+        self.p = p
+
+    def __call__(self, img, target):
+        if random.random() < self.p:
+            return hflip(img, target)
+        return img, 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 RandomPad(object):
+    def __init__(self, max_pad):
+        self.max_pad = max_pad
+
+    def __call__(self, img, target):
+        pad_x = random.randint(0, self.max_pad)
+        pad_y = random.randint(0, self.max_pad)
+        return pad(img, target, (pad_x, pad_y))
+
+
+class RandomSelect(object):
+    """
+    Randomly selects between transforms1 and transforms2,
+    with probability p for transforms1 and (1 - p) for transforms2
+    """
+
+    def __init__(self, transforms1, transforms2, p=0.5):
+        self.transforms1 = transforms1
+        self.transforms2 = transforms2
+        self.p = p
+
+    def __call__(self, img, target):
+        if random.random() < self.p:
+            return self.transforms1(img, target)
+        return self.transforms2(img, target)
+
+
+class ToTensor(object):
+    def __call__(self, img, target):
+        return F.to_tensor(img), target
+
+
+class RandomErasing(object):
+    def __init__(self, *args, **kwargs):
+        self.eraser = T.RandomErasing(*args, **kwargs)
+
+    def __call__(self, img, target):
+        return self.eraser(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:]
+        if "boxes" in target:
+            boxes = target["boxes"]
+            boxes = box_xyxy_to_cxcywh(boxes)
+            boxes = boxes / torch.tensor([w, h, w, h], dtype=torch.float32)
+            target["boxes"] = boxes
+        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

CircumSpect/object_detection/groundingdino/models/GroundingDINO/__init__.py

@@ -0,0 +1,15 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copied from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+
+from .groundingdino import build_groundingdino

CircumSpect/object_detection/groundingdino/models/GroundingDINO/backbone/__init__.py

@@ -0,0 +1 @@
+from .backbone import build_backbone

CircumSpect/object_detection/groundingdino/models/GroundingDINO/backbone/backbone.py

@@ -0,0 +1,221 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copied from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+
+"""
+Backbone modules.
+"""
+
+from typing import Dict, List
+
+import torch
+import torch.nn.functional as F
+import torchvision
+from torch import nn
+from torchvision.models._utils import IntermediateLayerGetter
+
+from CircumSpect.object_detection.groundingdino.util.misc import NestedTensor, clean_state_dict, is_main_process
+
+from .position_encoding import build_position_encoding
+from .swin_transformer import build_swin_transformer
+
+
+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 BackboneBase(nn.Module):
+    def __init__(
+        self,
+        backbone: nn.Module,
+        train_backbone: bool,
+        num_channels: int,
+        return_interm_indices: list,
+    ):
+        super().__init__()
+        for name, parameter in backbone.named_parameters():
+            if (
+                not train_backbone
+                or "layer2" not in name
+                and "layer3" not in name
+                and "layer4" not in name
+            ):
+                parameter.requires_grad_(False)
+
+        return_layers = {}
+        for idx, layer_index in enumerate(return_interm_indices):
+            return_layers.update(
+                {"layer{}".format(5 - len(return_interm_indices) + idx): "{}".format(layer_index)}
+            )
+
+        # if len:
+        #     if use_stage1_feature:
+        #         return_layers = {"layer1": "0", "layer2": "1", "layer3": "2", "layer4": "3"}
+        #     else:
+        #         return_layers = {"layer2": "0", "layer3": "1", "layer4": "2"}
+        # else:
+        #     return_layers = {'layer4': "0"}
+        self.body = IntermediateLayerGetter(backbone, return_layers=return_layers)
+        self.num_channels = num_channels
+
+    def forward(self, tensor_list: NestedTensor):
+        xs = self.body(tensor_list.tensors)
+        out: Dict[str, NestedTensor] = {}
+        for name, x in xs.items():
+            m = tensor_list.mask
+            assert m is not None
+            mask = F.interpolate(m[None].float(), size=x.shape[-2:]).to(torch.bool)[0]
+            out[name] = NestedTensor(x, mask)
+        # import ipdb; ipdb.set_trace()
+        return out
+
+
+class Backbone(BackboneBase):
+    """ResNet backbone with frozen BatchNorm."""
+
+    def __init__(
+        self,
+        name: str,
+        train_backbone: bool,
+        dilation: bool,
+        return_interm_indices: list,
+        batch_norm=FrozenBatchNorm2d,
+    ):
+        if name in ["resnet18", "resnet34", "resnet50", "resnet101"]:
+            backbone = getattr(torchvision.models, name)(
+                replace_stride_with_dilation=[False, False, dilation],
+                pretrained=is_main_process(),
+                norm_layer=batch_norm,
+            )
+        else:
+            raise NotImplementedError("Why you can get here with name {}".format(name))
+        # num_channels = 512 if name in ('resnet18', 'resnet34') else 2048
+        assert name not in ("resnet18", "resnet34"), "Only resnet50 and resnet101 are available."
+        assert return_interm_indices in [[0, 1, 2, 3], [1, 2, 3], [3]]
+        num_channels_all = [256, 512, 1024, 2048]
+        num_channels = num_channels_all[4 - len(return_interm_indices) :]
+        super().__init__(backbone, train_backbone, num_channels, return_interm_indices)
+
+
+class Joiner(nn.Sequential):
+    def __init__(self, backbone, position_embedding):
+        super().__init__(backbone, position_embedding)
+
+    def forward(self, tensor_list: NestedTensor):
+        xs = self[0](tensor_list)
+        out: List[NestedTensor] = []
+        pos = []
+        for name, x in xs.items():
+            out.append(x)
+            # position encoding
+            pos.append(self[1](x).to(x.tensors.dtype))
+
+        return out, pos
+
+
+def build_backbone(args):
+    """
+    Useful args:
+        - backbone: backbone name
+        - lr_backbone:
+        - dilation
+        - return_interm_indices: available: [0,1,2,3], [1,2,3], [3]
+        - backbone_freeze_keywords:
+        - use_checkpoint: for swin only for now
+
+    """
+    position_embedding = build_position_encoding(args)
+    train_backbone = True
+    if not train_backbone:
+        raise ValueError("Please set lr_backbone > 0")
+    return_interm_indices = args.return_interm_indices
+    assert return_interm_indices in [[0, 1, 2, 3], [1, 2, 3], [3]]
+    args.backbone_freeze_keywords
+    use_checkpoint = getattr(args, "use_checkpoint", False)
+
+    if args.backbone in ["resnet50", "resnet101"]:
+        backbone = Backbone(
+            args.backbone,
+            train_backbone,
+            args.dilation,
+            return_interm_indices,
+            batch_norm=FrozenBatchNorm2d,
+        )
+        bb_num_channels = backbone.num_channels
+    elif args.backbone 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(args.backbone.split("_")[-2])
+        backbone = build_swin_transformer(
+            args.backbone,
+            pretrain_img_size=pretrain_img_size,
+            out_indices=tuple(return_interm_indices),
+            dilation=False,
+            use_checkpoint=use_checkpoint,
+        )
+
+        bb_num_channels = backbone.num_features[4 - len(return_interm_indices) :]
+    else:
+        raise NotImplementedError("Unknown backbone {}".format(args.backbone))
+
+    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, position_embedding)
+    model.num_channels = bb_num_channels
+    assert isinstance(
+        bb_num_channels, List
+    ), "bb_num_channels is expected to be a List but {}".format(type(bb_num_channels))
+    # import ipdb; ipdb.set_trace()
+    return model

CircumSpect/object_detection/groundingdino/models/GroundingDINO/backbone/position_encoding.py

@@ -0,0 +1,186 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# DINO
+# Copyright (c) 2022 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copied from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+
+"""
+Various positional encodings for the transformer.
+"""
+import math
+
+import torch
+from torch import nn
+
+from CircumSpect.object_detection.groundingdino.util.misc import NestedTensor
+
+
+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.
+    """
+
+    def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=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):
+        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
+            # if os.environ.get("SHILONG_AMP", None) == '1':
+            #     eps = 1e-4
+            # else:
+            #     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
+
+
+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.
+    """
+
+    def __init__(
+        self, num_pos_feats=64, temperatureH=10000, temperatureW=10000, normalize=False, scale=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):
+        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)
+
+        # import ipdb; ipdb.set_trace()
+
+        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 * (torch.div(dim_tx, 2, rounding_mode='floor')) / 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 * (torch.div(dim_ty, 2, rounding_mode='floor')) / 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)
+
+        # import ipdb; ipdb.set_trace()
+
+        return pos
+
+
+class PositionEmbeddingLearned(nn.Module):
+    """
+    Absolute pos embedding, learned.
+    """
+
+    def __init__(self, num_pos_feats=256):
+        super().__init__()
+        self.row_embed = nn.Embedding(50, num_pos_feats)
+        self.col_embed = nn.Embedding(50, 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):
+        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

CircumSpect/object_detection/groundingdino/models/GroundingDINO/backbone/swin_transformer.py

@@ -0,0 +1,802 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# DINO
+# Copyright (c) 2022 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# --------------------------------------------------------
+# modified from https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/blob/master/mmdet/models/backbones/swin_transformer.py
+# --------------------------------------------------------
+
+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 CircumSpect.object_detection.groundingdino.util.misc 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
+
+
+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 init_weights(self, pretrained=None):
+    #     """Initialize the weights in backbone.
+    #     Args:
+    #         pretrained (str, optional): Path to pre-trained weights.
+    #             Defaults to None.
+    #     """
+
+    #     def _init_weights(m):
+    #         if isinstance(m, nn.Linear):
+    #             trunc_normal_(m.weight, std=.02)
+    #             if isinstance(m, nn.Linear) and m.bias is not None:
+    #                 nn.init.constant_(m.bias, 0)
+    #         elif isinstance(m, nn.LayerNorm):
+    #             nn.init.constant_(m.bias, 0)
+    #             nn.init.constant_(m.weight, 1.0)
+
+    #     if isinstance(pretrained, str):
+    #         self.apply(_init_weights)
+    #         logger = get_root_logger()
+    #         load_checkpoint(self, pretrained, strict=False, logger=logger)
+    #     elif pretrained is None:
+    #         self.apply(_init_weights)
+    #     else:
+    #         raise TypeError('pretrained must be a str or None')
+
+    def forward_raw(self, x):
+        """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)
+        # in:
+        #   torch.Size([2, 3, 1024, 1024])
+        # outs:
+        #   [torch.Size([2, 192, 256, 256]), torch.Size([2, 384, 128, 128]), \
+        #       torch.Size([2, 768, 64, 64]), torch.Size([2, 1536, 32, 32])]
+        return tuple(outs)
+
+    def forward(self, tensor_list: NestedTensor):
+        x = tensor_list.tensors
+
+        """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)
+
+            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)
+        # in:
+        #   torch.Size([2, 3, 1024, 1024])
+        # out:
+        #   [torch.Size([2, 192, 256, 256]), torch.Size([2, 384, 128, 128]), \
+        #       torch.Size([2, 768, 64, 64]), torch.Size([2, 1536, 32, 32])]
+
+        # 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
+
+
+if __name__ == "__main__":
+    model = build_swin_transformer("swin_L_384_22k", 384, dilation=True)
+    x = torch.rand(2, 3, 1024, 1024)
+    y = model.forward_raw(x)
+    import ipdb
+
+    ipdb.set_trace()
+    x = torch.rand(2, 3, 384, 384)
+    y = model.forward_raw(x)

CircumSpect/object_detection/groundingdino/models/GroundingDINO/bertwarper.py

@@ -0,0 +1,273 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from torch import Tensor, nn
+from torchvision.ops.boxes import nms
+from transformers import BertConfig, BertModel, BertPreTrainedModel
+from transformers.modeling_outputs import BaseModelOutputWithPoolingAndCrossAttentions
+
+
+class BertModelWarper(nn.Module):
+    def __init__(self, bert_model):
+        super().__init__()
+        # self.bert = bert_modelc
+
+        self.config = bert_model.config
+        self.embeddings = bert_model.embeddings
+        self.encoder = bert_model.encoder
+        self.pooler = bert_model.pooler
+
+        self.get_extended_attention_mask = bert_model.get_extended_attention_mask
+        self.invert_attention_mask = bert_model.invert_attention_mask
+        self.get_head_mask = bert_model.get_head_mask
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        """
+        output_attentions = (
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            batch_size, seq_length = input_shape
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size, seq_length = input_shape
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] if past_key_values is not None else 0
+        )
+
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                ((batch_size, seq_length + past_key_values_length)), device=device
+            )
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+            attention_mask, input_shape, device
+        )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+        # if os.environ.get('IPDB_SHILONG_DEBUG', None) == 'INFO':
+        #     import ipdb; ipdb.set_trace()
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class TextEncoderShell(nn.Module):
+    def __init__(self, text_encoder):
+        super().__init__()
+        self.text_encoder = text_encoder
+        self.config = self.text_encoder.config
+
+    def forward(self, **kw):
+        # feed into text encoder
+        return self.text_encoder(**kw)
+
+
+def generate_masks_with_special_tokens(tokenized, special_tokens_list, tokenizer):
+    """Generate attention mask between each pair of special tokens
+    Args:
+        input_ids (torch.Tensor): input ids. Shape: [bs, num_token]
+        special_tokens_mask (list): special tokens mask.
+    Returns:
+        torch.Tensor: attention mask between each special tokens.
+    """
+    input_ids = tokenized["input_ids"]
+    bs, num_token = input_ids.shape
+    # special_tokens_mask: bs, num_token. 1 for special tokens. 0 for normal tokens
+    special_tokens_mask = torch.zeros((bs, num_token), device=input_ids.device).bool()
+    for special_token in special_tokens_list:
+        special_tokens_mask |= input_ids == special_token
+
+    # idxs: each row is a list of indices of special tokens
+    idxs = torch.nonzero(special_tokens_mask)
+
+    # generate attention mask and positional ids
+    attention_mask = (
+        torch.eye(num_token, device=input_ids.device).bool().unsqueeze(0).repeat(bs, 1, 1)
+    )
+    position_ids = torch.zeros((bs, num_token), device=input_ids.device)
+    previous_col = 0
+    for i in range(idxs.shape[0]):
+        row, col = idxs[i]
+        if (col == 0) or (col == num_token - 1):
+            attention_mask[row, col, col] = True
+            position_ids[row, col] = 0
+        else:
+            attention_mask[row, previous_col + 1 : col + 1, previous_col + 1 : col + 1] = True
+            position_ids[row, previous_col + 1 : col + 1] = torch.arange(
+                0, col - previous_col, device=input_ids.device
+            )
+
+        previous_col = col
+
+    # # padding mask
+    # padding_mask = tokenized['attention_mask']
+    # attention_mask = attention_mask & padding_mask.unsqueeze(1).bool() & padding_mask.unsqueeze(2).bool()
+
+    return attention_mask, position_ids.to(torch.long)
+
+
+def generate_masks_with_special_tokens_and_transfer_map(tokenized, special_tokens_list, tokenizer):
+    """Generate attention mask between each pair of special tokens
+    Args:
+        input_ids (torch.Tensor): input ids. Shape: [bs, num_token]
+        special_tokens_mask (list): special tokens mask.
+    Returns:
+        torch.Tensor: attention mask between each special tokens.
+    """
+    input_ids = tokenized["input_ids"]
+    bs, num_token = input_ids.shape
+    # special_tokens_mask: bs, num_token. 1 for special tokens. 0 for normal tokens
+    special_tokens_mask = torch.zeros((bs, num_token), device=input_ids.device).bool()
+    for special_token in special_tokens_list:
+        special_tokens_mask |= input_ids == special_token
+
+    # idxs: each row is a list of indices of special tokens
+    idxs = torch.nonzero(special_tokens_mask)
+
+    # generate attention mask and positional ids
+    attention_mask = (
+        torch.eye(num_token, device=input_ids.device).bool().unsqueeze(0).repeat(bs, 1, 1)
+    )
+    position_ids = torch.zeros((bs, num_token), device=input_ids.device)
+    cate_to_token_mask_list = [[] for _ in range(bs)]
+    previous_col = 0
+    for i in range(idxs.shape[0]):
+        row, col = idxs[i]
+        if (col == 0) or (col == num_token - 1):
+            attention_mask[row, col, col] = True
+            position_ids[row, col] = 0
+        else:
+            attention_mask[row, previous_col + 1 : col + 1, previous_col + 1 : col + 1] = True
+            position_ids[row, previous_col + 1 : col + 1] = torch.arange(
+                0, col - previous_col, device=input_ids.device
+            )
+            c2t_maski = torch.zeros((num_token), device=input_ids.device).bool()
+            c2t_maski[previous_col + 1 : col] = True
+            cate_to_token_mask_list[row].append(c2t_maski)
+        previous_col = col
+
+    cate_to_token_mask_list = [
+        torch.stack(cate_to_token_mask_listi, dim=0)
+        for cate_to_token_mask_listi in cate_to_token_mask_list
+    ]
+
+    # # padding mask
+    # padding_mask = tokenized['attention_mask']
+    # attention_mask = attention_mask & padding_mask.unsqueeze(1).bool() & padding_mask.unsqueeze(2).bool()
+
+    return attention_mask, position_ids.to(torch.long), cate_to_token_mask_list

CircumSpect/object_detection/groundingdino/models/GroundingDINO/csrc/MsDeformAttn/ms_deform_attn.h

@@ -0,0 +1,64 @@
+/*!
+**************************************************************************************************
+* 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 "ms_deform_attn_cpu.h"
+
+#ifdef WITH_CUDA
+#include "ms_deform_attn_cuda.h"
+#endif
+
+namespace groundingdino {
+
+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.type().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.type().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");
+}
+
+} // namespace groundingdino

CircumSpect/object_detection/groundingdino/models/GroundingDINO/csrc/MsDeformAttn/ms_deform_attn_cpu.cpp

@@ -0,0 +1,43 @@
+/*!
+**************************************************************************************************
+* 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>
+
+namespace groundingdino {
+
+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");
+}
+
+} // namespace groundingdino

CircumSpect/object_detection/groundingdino/models/GroundingDINO/csrc/MsDeformAttn/ms_deform_attn_cpu.h

@@ -0,0 +1,35 @@
+/*!
+**************************************************************************************************
+* 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>
+
+namespace groundingdino {
+
+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);
+
+} // namespace groundingdino

CircumSpect/object_detection/groundingdino/models/GroundingDINO/csrc/MsDeformAttn/ms_deform_attn_cuda.cu

@@ -0,0 +1,156 @@
+/*!
+**************************************************************************************************
+* 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 "ms_deform_im2col_cuda.cuh"
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+namespace groundingdino {
+
+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.type().is_cuda(), "value must be a CUDA tensor");
+    AT_ASSERTM(spatial_shapes.type().is_cuda(), "spatial_shapes must be a CUDA tensor");
+    AT_ASSERTM(level_start_index.type().is_cuda(), "level_start_index must be a CUDA tensor");
+    AT_ASSERTM(sampling_loc.type().is_cuda(), "sampling_loc must be a CUDA tensor");
+    AT_ASSERTM(attn_weight.type().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.type(), "ms_deform_attn_forward_cuda", ([&] {
+            ms_deformable_im2col_cuda(at::cuda::getCurrentCUDAStream(),
+                value.data<scalar_t>() + n * im2col_step_ * per_value_size,
+                spatial_shapes.data<int64_t>(),
+                level_start_index.data<int64_t>(),
+                sampling_loc.data<scalar_t>() + n * im2col_step_ * per_sample_loc_size,
+                attn_weight.data<scalar_t>() + n * im2col_step_ * per_attn_weight_size,
+                batch_n, spatial_size, num_heads, channels, num_levels, num_query, num_point,
+                columns.data<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.type().is_cuda(), "value must be a CUDA tensor");
+    AT_ASSERTM(spatial_shapes.type().is_cuda(), "spatial_shapes must be a CUDA tensor");
+    AT_ASSERTM(level_start_index.type().is_cuda(), "level_start_index must be a CUDA tensor");
+    AT_ASSERTM(sampling_loc.type().is_cuda(), "sampling_loc must be a CUDA tensor");
+    AT_ASSERTM(attn_weight.type().is_cuda(), "attn_weight must be a CUDA tensor");
+    AT_ASSERTM(grad_output.type().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.type(), "ms_deform_attn_backward_cuda", ([&] {
+            ms_deformable_col2im_cuda(at::cuda::getCurrentCUDAStream(),
+                                    grad_output_g.data<scalar_t>(),
+                                    value.data<scalar_t>() + n * im2col_step_ * per_value_size,
+                                    spatial_shapes.data<int64_t>(),
+                                    level_start_index.data<int64_t>(),
+                                    sampling_loc.data<scalar_t>() + n * im2col_step_ * per_sample_loc_size,
+                                    attn_weight.data<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<scalar_t>() +  n * im2col_step_ * per_value_size,
+                                    grad_sampling_loc.data<scalar_t>() + n * im2col_step_ * per_sample_loc_size,
+                                    grad_attn_weight.data<scalar_t>() + n * im2col_step_ * per_attn_weight_size);
+
+        }));
+    }
+
+    return {
+        grad_value, grad_sampling_loc, grad_attn_weight
+    };
+}
+
+} // namespace groundingdino

CircumSpect/object_detection/groundingdino/models/GroundingDINO/csrc/MsDeformAttn/ms_deform_attn_cuda.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>
+
+namespace groundingdino {
+
+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);
+
+} // namespace groundingdino

CircumSpect/object_detection/groundingdino/models/GroundingDINO/csrc/MsDeformAttn/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));
+  }
+
+}

CircumSpect/object_detection/groundingdino/models/GroundingDINO/csrc/cuda_version.cu

@@ -0,0 +1,7 @@
+#include <cuda_runtime_api.h>
+
+namespace groundingdino {
+int get_cudart_version() {
+  return CUDART_VERSION;
+}
+} // namespace groundingdino

CircumSpect/object_detection/groundingdino/models/GroundingDINO/csrc/vision.cpp

@@ -0,0 +1,58 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+#include "MsDeformAttn/ms_deform_attn.h"
+
+namespace groundingdino {
+
+#ifdef WITH_CUDA
+extern int get_cudart_version();
+#endif
+
+std::string get_cuda_version() {
+#ifdef WITH_CUDA
+  std::ostringstream oss;
+
+  // copied from
+  // https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/cuda/detail/CUDAHooks.cpp#L231
+  auto printCudaStyleVersion = [&](int v) {
+    oss << (v / 1000) << "." << (v / 10 % 100);
+    if (v % 10 != 0) {
+      oss << "." << (v % 10);
+    }
+  };
+  printCudaStyleVersion(get_cudart_version());
+  return oss.str();
+#else
+  return std::string("not available");
+#endif
+}
+
+// similar to
+// https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/Version.cpp
+std::string get_compiler_version() {
+  std::ostringstream ss;
+#if defined(__GNUC__)
+#ifndef __clang__
+  { ss << "GCC " << __GNUC__ << "." << __GNUC_MINOR__; }
+#endif
+#endif
+
+#if defined(__clang_major__)
+  {
+    ss << "clang " << __clang_major__ << "." << __clang_minor__ << "."
+       << __clang_patchlevel__;
+  }
+#endif
+
+#if defined(_MSC_VER)
+  { ss << "MSVC " << _MSC_FULL_VER; }
+#endif
+  return ss.str();
+}
+
+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");
+}
+
+} // namespace groundingdino

CircumSpect/object_detection/groundingdino/models/GroundingDINO/fuse_modules.py

@@ -0,0 +1,297 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from timm.models.layers import DropPath
+
+
+class FeatureResizer(nn.Module):
+    """
+    This class takes as input a set of embeddings of dimension C1 and outputs a set of
+    embedding of dimension C2, after a linear transformation, dropout and normalization (LN).
+    """
+
+    def __init__(self, input_feat_size, output_feat_size, dropout, do_ln=True):
+        super().__init__()
+        self.do_ln = do_ln
+        # Object feature encoding
+        self.fc = nn.Linear(input_feat_size, output_feat_size, bias=True)
+        self.layer_norm = nn.LayerNorm(output_feat_size, eps=1e-12)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, encoder_features):
+        x = self.fc(encoder_features)
+        if self.do_ln:
+            x = self.layer_norm(x)
+        output = self.dropout(x)
+        return output
+
+
+def l1norm(X, dim, eps=1e-8):
+    """L1-normalize columns of X"""
+    norm = torch.abs(X).sum(dim=dim, keepdim=True) + eps
+    X = torch.div(X, norm)
+    return X
+
+
+def l2norm(X, dim, eps=1e-8):
+    """L2-normalize columns of X"""
+    norm = torch.pow(X, 2).sum(dim=dim, keepdim=True).sqrt() + eps
+    X = torch.div(X, norm)
+    return X
+
+
+def func_attention(query, context, smooth=1, raw_feature_norm="softmax", eps=1e-8):
+    """
+    query: (n_context, queryL, d)
+    context: (n_context, sourceL, d)
+    """
+    batch_size_q, queryL = query.size(0), query.size(1)
+    batch_size, sourceL = context.size(0), context.size(1)
+
+    # Get attention
+    # --> (batch, d, queryL)
+    queryT = torch.transpose(query, 1, 2)
+
+    # (batch, sourceL, d)(batch, d, queryL)
+    # --> (batch, sourceL, queryL)
+    attn = torch.bmm(context, queryT)
+    if raw_feature_norm == "softmax":
+        # --> (batch*sourceL, queryL)
+        attn = attn.view(batch_size * sourceL, queryL)
+        attn = nn.Softmax()(attn)
+        # --> (batch, sourceL, queryL)
+        attn = attn.view(batch_size, sourceL, queryL)
+    elif raw_feature_norm == "l2norm":
+        attn = l2norm(attn, 2)
+    elif raw_feature_norm == "clipped_l2norm":
+        attn = nn.LeakyReLU(0.1)(attn)
+        attn = l2norm(attn, 2)
+    else:
+        raise ValueError("unknown first norm type:", raw_feature_norm)
+    # --> (batch, queryL, sourceL)
+    attn = torch.transpose(attn, 1, 2).contiguous()
+    # --> (batch*queryL, sourceL)
+    attn = attn.view(batch_size * queryL, sourceL)
+    attn = nn.Softmax()(attn * smooth)
+    # --> (batch, queryL, sourceL)
+    attn = attn.view(batch_size, queryL, sourceL)
+    # --> (batch, sourceL, queryL)
+    attnT = torch.transpose(attn, 1, 2).contiguous()
+
+    # --> (batch, d, sourceL)
+    contextT = torch.transpose(context, 1, 2)
+    # (batch x d x sourceL)(batch x sourceL x queryL)
+    # --> (batch, d, queryL)
+    weightedContext = torch.bmm(contextT, attnT)
+    # --> (batch, queryL, d)
+    weightedContext = torch.transpose(weightedContext, 1, 2)
+
+    return weightedContext, attnT
+
+
+class BiMultiHeadAttention(nn.Module):
+    def __init__(self, v_dim, l_dim, embed_dim, num_heads, dropout=0.1, cfg=None):
+        super(BiMultiHeadAttention, self).__init__()
+
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.head_dim = embed_dim // num_heads
+        self.v_dim = v_dim
+        self.l_dim = l_dim
+
+        assert (
+            self.head_dim * self.num_heads == self.embed_dim
+        ), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {self.num_heads})."
+        self.scale = self.head_dim ** (-0.5)
+        self.dropout = dropout
+
+        self.v_proj = nn.Linear(self.v_dim, self.embed_dim)
+        self.l_proj = nn.Linear(self.l_dim, self.embed_dim)
+        self.values_v_proj = nn.Linear(self.v_dim, self.embed_dim)
+        self.values_l_proj = nn.Linear(self.l_dim, self.embed_dim)
+
+        self.out_v_proj = nn.Linear(self.embed_dim, self.v_dim)
+        self.out_l_proj = nn.Linear(self.embed_dim, self.l_dim)
+
+        self.stable_softmax_2d = True
+        self.clamp_min_for_underflow = True
+        self.clamp_max_for_overflow = True
+
+        self._reset_parameters()
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def _reset_parameters(self):
+        nn.init.xavier_uniform_(self.v_proj.weight)
+        self.v_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.l_proj.weight)
+        self.l_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.values_v_proj.weight)
+        self.values_v_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.values_l_proj.weight)
+        self.values_l_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.out_v_proj.weight)
+        self.out_v_proj.bias.data.fill_(0)
+        nn.init.xavier_uniform_(self.out_l_proj.weight)
+        self.out_l_proj.bias.data.fill_(0)
+
+    def forward(self, v, l, attention_mask_v=None, attention_mask_l=None):
+        """_summary_
+
+        Args:
+            v (_type_): bs, n_img, dim
+            l (_type_): bs, n_text, dim
+            attention_mask_v (_type_, optional): _description_. bs, n_img
+            attention_mask_l (_type_, optional): _description_. bs, n_text
+
+        Returns:
+            _type_: _description_
+        """
+        # if os.environ.get('IPDB_SHILONG_DEBUG', None) == 'INFO':
+        #     import ipdb; ipdb.set_trace()
+        bsz, tgt_len, _ = v.size()
+
+        query_states = self.v_proj(v) * self.scale
+        key_states = self._shape(self.l_proj(l), -1, bsz)
+        value_v_states = self._shape(self.values_v_proj(v), -1, bsz)
+        value_l_states = self._shape(self.values_l_proj(l), -1, bsz)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_v_states = value_v_states.view(*proj_shape)
+        value_l_states = value_l_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))  # bs*nhead, nimg, ntxt
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is {attn_weights.size()}"
+            )
+
+        if self.stable_softmax_2d:
+            attn_weights = attn_weights - attn_weights.max()
+
+        if self.clamp_min_for_underflow:
+            attn_weights = torch.clamp(
+                attn_weights, min=-50000
+            )  # Do not increase -50000, data type half has quite limited range
+        if self.clamp_max_for_overflow:
+            attn_weights = torch.clamp(
+                attn_weights, max=50000
+            )  # Do not increase 50000, data type half has quite limited range
+
+        attn_weights_T = attn_weights.transpose(1, 2)
+        attn_weights_l = attn_weights_T - torch.max(attn_weights_T, dim=-1, keepdim=True)[0]
+        if self.clamp_min_for_underflow:
+            attn_weights_l = torch.clamp(
+                attn_weights_l, min=-50000
+            )  # Do not increase -50000, data type half has quite limited range
+        if self.clamp_max_for_overflow:
+            attn_weights_l = torch.clamp(
+                attn_weights_l, max=50000
+            )  # Do not increase 50000, data type half has quite limited range
+
+        # mask vison for language
+        if attention_mask_v is not None:
+            attention_mask_v = (
+                attention_mask_v[:, None, None, :].repeat(1, self.num_heads, 1, 1).flatten(0, 1)
+            )
+            attn_weights_l.masked_fill_(attention_mask_v, float("-inf"))
+
+        attn_weights_l = attn_weights_l.softmax(dim=-1)
+
+        # mask language for vision
+        if attention_mask_l is not None:
+            attention_mask_l = (
+                attention_mask_l[:, None, None, :].repeat(1, self.num_heads, 1, 1).flatten(0, 1)
+            )
+            attn_weights.masked_fill_(attention_mask_l, float("-inf"))
+        attn_weights_v = attn_weights.softmax(dim=-1)
+
+        attn_probs_v = F.dropout(attn_weights_v, p=self.dropout, training=self.training)
+        attn_probs_l = F.dropout(attn_weights_l, p=self.dropout, training=self.training)
+
+        attn_output_v = torch.bmm(attn_probs_v, value_l_states)
+        attn_output_l = torch.bmm(attn_probs_l, value_v_states)
+
+        if attn_output_v.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output_v` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {attn_output_v.size()}"
+            )
+
+        if attn_output_l.size() != (bsz * self.num_heads, src_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output_l` should be of size {(bsz, self.num_heads, src_len, self.head_dim)}, but is {attn_output_l.size()}"
+            )
+
+        attn_output_v = attn_output_v.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output_v = attn_output_v.transpose(1, 2)
+        attn_output_v = attn_output_v.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output_l = attn_output_l.view(bsz, self.num_heads, src_len, self.head_dim)
+        attn_output_l = attn_output_l.transpose(1, 2)
+        attn_output_l = attn_output_l.reshape(bsz, src_len, self.embed_dim)
+
+        attn_output_v = self.out_v_proj(attn_output_v)
+        attn_output_l = self.out_l_proj(attn_output_l)
+
+        return attn_output_v, attn_output_l
+
+
+# Bi-Direction MHA (text->image, image->text)
+class BiAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        v_dim,
+        l_dim,
+        embed_dim,
+        num_heads,
+        dropout=0.1,
+        drop_path=0.0,
+        init_values=1e-4,
+        cfg=None,
+    ):
+        """
+        Inputs:
+            embed_dim - Dimensionality of input and attention feature vectors
+            hidden_dim - Dimensionality of hidden layer in feed-forward network
+                         (usually 2-4x larger than embed_dim)
+            num_heads - Number of heads to use in the Multi-Head Attention block
+            dropout - Amount of dropout to apply in the feed-forward network
+        """
+        super(BiAttentionBlock, self).__init__()
+
+        # pre layer norm
+        self.layer_norm_v = nn.LayerNorm(v_dim)
+        self.layer_norm_l = nn.LayerNorm(l_dim)
+        self.attn = BiMultiHeadAttention(
+            v_dim=v_dim, l_dim=l_dim, embed_dim=embed_dim, num_heads=num_heads, dropout=dropout
+        )
+
+        # add layer scale for training stability
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        self.gamma_v = nn.Parameter(init_values * torch.ones((v_dim)), requires_grad=True)
+        self.gamma_l = nn.Parameter(init_values * torch.ones((l_dim)), requires_grad=True)
+
+    def forward(self, v, l, attention_mask_v=None, attention_mask_l=None):
+        v = self.layer_norm_v(v)
+        l = self.layer_norm_l(l)
+        delta_v, delta_l = self.attn(
+            v, l, attention_mask_v=attention_mask_v, attention_mask_l=attention_mask_l
+        )
+        # v, l = v + delta_v, l + delta_l
+        v = v + self.drop_path(self.gamma_v * delta_v)
+        l = l + self.drop_path(self.gamma_l * delta_l)
+        return v, l
+
+    # def forward(self, v:List[torch.Tensor], l, attention_mask_v=None, attention_mask_l=None)

CircumSpect/object_detection/groundingdino/models/GroundingDINO/groundingdino.py

@@ -0,0 +1,412 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR model and criterion classes.
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Modified from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+# Modified from Deformable DETR (https://github.com/fundamentalvision/Deformable-DETR)
+# Copyright (c) 2020 SenseTime. All Rights Reserved.
+# ------------------------------------------------------------------------
+import copy
+from typing import List
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torchvision.ops.boxes import nms
+from transformers import AutoTokenizer, BertModel, BertTokenizer, RobertaModel, RobertaTokenizerFast
+
+from CircumSpect.object_detection.groundingdino.util import box_ops, get_tokenlizer
+from CircumSpect.object_detection.groundingdino.util.misc import (
+    NestedTensor,
+    accuracy,
+    get_world_size,
+    interpolate,
+    inverse_sigmoid,
+    is_dist_avail_and_initialized,
+    nested_tensor_from_tensor_list,
+)
+from CircumSpect.object_detection.groundingdino.util.utils import get_phrases_from_posmap
+from CircumSpect.object_detection.groundingdino.util.visualizer import COCOVisualizer
+from CircumSpect.object_detection.groundingdino.util.vl_utils import create_positive_map_from_span
+
+from ..registry import MODULE_BUILD_FUNCS
+from .backbone import build_backbone
+from .bertwarper import (
+    BertModelWarper,
+    generate_masks_with_special_tokens,
+    generate_masks_with_special_tokens_and_transfer_map,
+)
+from .transformer import build_transformer
+from .utils import MLP, ContrastiveEmbed, sigmoid_focal_loss
+
+
+class GroundingDINO(nn.Module):
+    """This is the Cross-Attention Detector module that performs object detection"""
+
+    def __init__(
+        self,
+        backbone,
+        transformer,
+        num_queries,
+        aux_loss=False,
+        iter_update=False,
+        query_dim=2,
+        num_feature_levels=1,
+        nheads=8,
+        # two stage
+        two_stage_type="no",  # ['no', 'standard']
+        dec_pred_bbox_embed_share=True,
+        two_stage_class_embed_share=True,
+        two_stage_bbox_embed_share=True,
+        num_patterns=0,
+        dn_number=100,
+        dn_box_noise_scale=0.4,
+        dn_label_noise_ratio=0.5,
+        dn_labelbook_size=100,
+        text_encoder_type="bert-base-uncased",
+        sub_sentence_present=True,
+        max_text_len=256,
+    ):
+        """Initializes the model.
+        Parameters:
+            backbone: torch module of the backbone to be used. See backbone.py
+            transformer: torch module of the transformer architecture. See transformer.py
+            num_queries: number of object queries, ie detection slot. This is the maximal number of objects
+                         Conditional DETR can detect in a single image. For COCO, we recommend 100 queries.
+            aux_loss: True if auxiliary decoding losses (loss at each decoder layer) are to be used.
+        """
+        super().__init__()
+        self.num_queries = num_queries
+        self.transformer = transformer
+        self.hidden_dim = hidden_dim = transformer.d_model
+        self.num_feature_levels = num_feature_levels
+        self.nheads = nheads
+        self.max_text_len = 256
+        self.sub_sentence_present = sub_sentence_present
+
+        # setting query dim
+        self.query_dim = query_dim
+        assert query_dim == 4
+
+        # for dn training
+        self.num_patterns = num_patterns
+        self.dn_number = dn_number
+        self.dn_box_noise_scale = dn_box_noise_scale
+        self.dn_label_noise_ratio = dn_label_noise_ratio
+        self.dn_labelbook_size = dn_labelbook_size
+
+        # bert
+        self.tokenizer = get_tokenlizer.get_tokenlizer(text_encoder_type)
+        self.bert = get_tokenlizer.get_pretrained_language_model(text_encoder_type)
+        self.bert.pooler.dense.weight.requires_grad_(False)
+        self.bert.pooler.dense.bias.requires_grad_(False)
+        self.bert = BertModelWarper(bert_model=self.bert)
+
+        self.feat_map = nn.Linear(self.bert.config.hidden_size, self.hidden_dim, bias=True)
+        nn.init.constant_(self.feat_map.bias.data, 0)
+        nn.init.xavier_uniform_(self.feat_map.weight.data)
+        # freeze
+
+        # special tokens
+        self.specical_tokens = self.tokenizer.convert_tokens_to_ids(["[CLS]", "[SEP]", ".", "?"])
+
+        # prepare input projection layers
+        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
+            self.input_proj = nn.ModuleList(input_proj_list)
+        else:
+            assert two_stage_type == "no", "two_stage_type should be no if num_feature_levels=1 !!!"
+            self.input_proj = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        nn.Conv2d(backbone.num_channels[-1], hidden_dim, kernel_size=1),
+                        nn.GroupNorm(32, hidden_dim),
+                    )
+                ]
+            )
+
+        self.backbone = backbone
+        self.aux_loss = aux_loss
+        self.box_pred_damping = box_pred_damping = None
+
+        self.iter_update = iter_update
+        assert iter_update, "Why not iter_update?"
+
+        # prepare pred layers
+        self.dec_pred_bbox_embed_share = dec_pred_bbox_embed_share
+        # prepare class & box embed
+        _class_embed = ContrastiveEmbed()
+
+        _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 i in range(transformer.num_decoder_layers)]
+        else:
+            box_embed_layerlist = [
+                copy.deepcopy(_bbox_embed) for i in range(transformer.num_decoder_layers)
+            ]
+        class_embed_layerlist = [_class_embed for i in range(transformer.num_decoder_layers)]
+        self.bbox_embed = nn.ModuleList(box_embed_layerlist)
+        self.class_embed = nn.ModuleList(class_embed_layerlist)
+        self.transformer.decoder.bbox_embed = self.bbox_embed
+        self.transformer.decoder.class_embed = self.class_embed
+
+        # two stage
+        self.two_stage_type = two_stage_type
+        assert two_stage_type in ["no", "standard"], "unknown param {} of two_stage_type".format(
+            two_stage_type
+        )
+        if two_stage_type != "no":
+            if two_stage_bbox_embed_share:
+                assert 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 two_stage_class_embed_share:
+                assert 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
+
+        self._reset_parameters()
+
+    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)
+
+    def set_image_tensor(self, samples: NestedTensor):
+        if isinstance(samples, (list, torch.Tensor)):
+            samples = nested_tensor_from_tensor_list(samples)
+        self.features, self.poss = self.backbone(samples)
+
+    def unset_image_tensor(self):
+        if hasattr(self, 'features'):
+            del self.features
+        if hasattr(self,'poss'):
+            del self.poss 
+
+    def set_image_features(self, features , poss):
+        self.features = features
+        self.poss = poss
+
+    def init_ref_points(self, use_num_queries):
+        self.refpoint_embed = nn.Embedding(use_num_queries, self.query_dim)
+
+    def forward(self, samples: NestedTensor, targets: List = None, **kw):
+        """The forward expects a NestedTensor, which consists of:
+           - samples.tensor: batched images, of shape [batch_size x 3 x H x W]
+           - samples.mask: a binary mask of shape [batch_size x H x W], containing 1 on padded pixels
+
+        It returns a dict with the following elements:
+           - "pred_logits": the classification logits (including no-object) for all queries.
+                            Shape= [batch_size x num_queries x num_classes]
+           - "pred_boxes": The normalized boxes coordinates for all queries, represented as
+                           (center_x, center_y, width, height). These values are normalized in [0, 1],
+                           relative to the size of each individual image (disregarding possible padding).
+                           See PostProcess for information on how to retrieve the unnormalized bounding box.
+           - "aux_outputs": Optional, only returned when auxilary losses are activated. It is a list of
+                            dictionnaries containing the two above keys for each decoder layer.
+        """
+        if targets is None:
+            captions = kw["captions"]
+        else:
+            captions = [t["caption"] for t in targets]
+
+        # encoder texts
+        tokenized = self.tokenizer(captions, padding="longest", return_tensors="pt").to(
+            samples.device
+        )
+        (
+            text_self_attention_masks,
+            position_ids,
+            cate_to_token_mask_list,
+        ) = generate_masks_with_special_tokens_and_transfer_map(
+            tokenized, self.specical_tokens, self.tokenizer
+        )
+
+        if text_self_attention_masks.shape[1] > self.max_text_len:
+            text_self_attention_masks = text_self_attention_masks[
+                :, : self.max_text_len, : self.max_text_len
+            ]
+            position_ids = position_ids[:, : self.max_text_len]
+            tokenized["input_ids"] = tokenized["input_ids"][:, : self.max_text_len]
+            tokenized["attention_mask"] = tokenized["attention_mask"][:, : self.max_text_len]
+            tokenized["token_type_ids"] = tokenized["token_type_ids"][:, : self.max_text_len]
+
+        # extract text embeddings
+        if self.sub_sentence_present:
+            tokenized_for_encoder = {k: v for k, v in tokenized.items() if k != "attention_mask"}
+            tokenized_for_encoder["attention_mask"] = text_self_attention_masks
+            tokenized_for_encoder["position_ids"] = position_ids
+        else:
+            # import ipdb; ipdb.set_trace()
+            tokenized_for_encoder = tokenized
+
+        bert_output = self.bert(**tokenized_for_encoder)  # bs, 195, 768
+
+        encoded_text = self.feat_map(bert_output["last_hidden_state"])  # bs, 195, d_model
+        text_token_mask = tokenized.attention_mask.bool()  # bs, 195
+        # text_token_mask: True for nomask, False for mask
+        # text_self_attention_masks: True for nomask, False for mask
+
+        if encoded_text.shape[1] > self.max_text_len:
+            encoded_text = encoded_text[:, : self.max_text_len, :]
+            text_token_mask = text_token_mask[:, : self.max_text_len]
+            position_ids = position_ids[:, : self.max_text_len]
+            text_self_attention_masks = text_self_attention_masks[
+                :, : self.max_text_len, : self.max_text_len
+            ]
+
+        text_dict = {
+            "encoded_text": encoded_text,  # bs, 195, d_model
+            "text_token_mask": text_token_mask,  # bs, 195
+            "position_ids": position_ids,  # bs, 195
+            "text_self_attention_masks": text_self_attention_masks,  # bs, 195,195
+        }
+
+        # import ipdb; ipdb.set_trace()
+        if isinstance(samples, (list, torch.Tensor)):
+            samples = nested_tensor_from_tensor_list(samples)
+        if not hasattr(self, 'features') or not hasattr(self, 'poss'):
+            self.set_image_tensor(samples)
+
+        srcs = []
+        masks = []
+        for l, feat in enumerate(self.features):
+            src, mask = feat.decompose()
+            srcs.append(self.input_proj[l](src))
+            masks.append(mask)
+            assert mask is not None
+        if self.num_feature_levels > len(srcs):
+            _len_srcs = len(srcs)
+            for l in range(_len_srcs, self.num_feature_levels):
+                if l == _len_srcs:
+                    src = self.input_proj[l](self.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[1](NestedTensor(src, mask)).to(src.dtype)
+                srcs.append(src)
+                masks.append(mask)
+                self.poss.append(pos_l)
+
+        input_query_bbox = input_query_label = attn_mask = dn_meta = None
+        hs, reference, hs_enc, ref_enc, init_box_proposal = self.transformer(
+            srcs, masks, input_query_bbox, self.poss, input_query_label, attn_mask, text_dict
+        )
+
+        # deformable-detr-like anchor update
+        outputs_coord_list = []
+        for dec_lid, (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)
+
+        # output
+        outputs_class = torch.stack(
+            [
+                layer_cls_embed(layer_hs, text_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]}
+
+        # # for intermediate outputs
+        # if self.aux_loss:
+        #     out['aux_outputs'] = self._set_aux_loss(outputs_class, outputs_coord_list)
+
+        # # for encoder output
+        # if hs_enc is not None:
+        #     # prepare intermediate outputs
+        #     interm_coord = ref_enc[-1]
+        #     interm_class = self.transformer.enc_out_class_embed(hs_enc[-1], text_dict)
+        #     out['interm_outputs'] = {'pred_logits': interm_class, 'pred_boxes': interm_coord}
+        #     out['interm_outputs_for_matching_pre'] = {'pred_logits': interm_class, 'pred_boxes': init_box_proposal}
+        unset_image_tensor = kw.get('unset_image_tensor', True)
+        if unset_image_tensor:
+            self.unset_image_tensor() ## If necessary
+        return out
+
+    @torch.jit.unused
+    def _set_aux_loss(self, outputs_class, outputs_coord):
+        # this is a workaround to make torchscript happy, as torchscript
+        # doesn't support dictionary with non-homogeneous values, such
+        # as a dict having both a Tensor and a list.
+        return [
+            {"pred_logits": a, "pred_boxes": b}
+            for a, b in zip(outputs_class[:-1], outputs_coord[:-1])
+        ]
+
+
+@MODULE_BUILD_FUNCS.registe_with_name(module_name="groundingdino")
+def build_groundingdino(args):
+
+    backbone = build_backbone(args)
+    transformer = build_transformer(args)
+
+    dn_labelbook_size = args.dn_labelbook_size
+    dec_pred_bbox_embed_share = args.dec_pred_bbox_embed_share
+    sub_sentence_present = args.sub_sentence_present
+
+    model = GroundingDINO(
+        backbone,
+        transformer,
+        num_queries=args.num_queries,
+        aux_loss=True,
+        iter_update=True,
+        query_dim=4,
+        num_feature_levels=args.num_feature_levels,
+        nheads=args.nheads,
+        dec_pred_bbox_embed_share=dec_pred_bbox_embed_share,
+        two_stage_type=args.two_stage_type,
+        two_stage_bbox_embed_share=args.two_stage_bbox_embed_share,
+        two_stage_class_embed_share=args.two_stage_class_embed_share,
+        num_patterns=args.num_patterns,
+        dn_number=0,
+        dn_box_noise_scale=args.dn_box_noise_scale,
+        dn_label_noise_ratio=args.dn_label_noise_ratio,
+        dn_labelbook_size=dn_labelbook_size,
+        text_encoder_type=args.text_encoder_type,
+        sub_sentence_present=sub_sentence_present,
+        max_text_len=args.max_text_len,
+    )
+
+    return model
+

CircumSpect/object_detection/groundingdino/models/GroundingDINO/ms_deform_attn.py

@@ -0,0 +1,413 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# 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/fundamentalvision/Deformable-DETR/blob/main/models/ops/functions/ms_deform_attn_func.py
+# https://github.com/fundamentalvision/Deformable-DETR/blob/main/models/ops/modules/ms_deform_attn.py
+# https://github.com/open-mmlab/mmcv/blob/master/mmcv/ops/multi_scale_deform_attn.py
+# ------------------------------------------------------------------------------------------------
+
+import math
+import warnings
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+from torch.nn.init import constant_, xavier_uniform_
+
+try:
+    from groundingdino import _C
+except:
+    warnings.warn("Failed to load custom C++ ops. Running on CPU mode Only!")
+
+
+# helpers
+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 MultiScaleDeformableAttnFunction(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 = _C.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 = _C.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 multi_scale_deformable_attn_pytorch(
+    value: torch.Tensor,
+    value_spatial_shapes: torch.Tensor,
+    sampling_locations: torch.Tensor,
+    attention_weights: torch.Tensor,
+) -> torch.Tensor:
+
+    bs, _, num_heads, embed_dims = value.shape
+    _, num_queries, num_heads, num_levels, num_points, _ = 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 level, (H_, W_) in enumerate(value_spatial_shapes):
+        # bs, H_*W_, num_heads, embed_dims ->
+        # bs, H_*W_, num_heads*embed_dims ->
+        # bs, num_heads*embed_dims, H_*W_ ->
+        # bs*num_heads, embed_dims, H_, W_
+        value_l_ = (
+            value_list[level].flatten(2).transpose(1, 2).reshape(bs * num_heads, embed_dims, H_, W_)
+        )
+        # bs, num_queries, num_heads, num_points, 2 ->
+        # bs, num_heads, num_queries, num_points, 2 ->
+        # bs*num_heads, num_queries, num_points, 2
+        sampling_grid_l_ = sampling_grids[:, :, :, level].transpose(1, 2).flatten(0, 1)
+        # bs*num_heads, embed_dims, num_queries, num_points
+        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_)
+    # (bs, num_queries, num_heads, num_levels, num_points) ->
+    # (bs, num_heads, num_queries, num_levels, num_points) ->
+    # (bs, num_heads, 1, num_queries, num_levels*num_points)
+    attention_weights = attention_weights.transpose(1, 2).reshape(
+        bs * num_heads, 1, num_queries, num_levels * num_points
+    )
+    output = (
+        (torch.stack(sampling_value_list, dim=-2).flatten(-2) * attention_weights)
+        .sum(-1)
+        .view(bs, num_heads * embed_dims, num_queries)
+    )
+    return output.transpose(1, 2).contiguous()
+
+
+class MultiScaleDeformableAttention(nn.Module):
+    """Multi-Scale Deformable Attention Module used in Deformable-DETR
+
+    `Deformable DETR: Deformable Transformers for End-to-End Object Detection.
+    <https://arxiv.org/pdf/2010.04159.pdf>`_.
+
+    Args:
+        embed_dim (int): The embedding dimension of Attention. Default: 256.
+        num_heads (int): The number of attention heads. Default: 8.
+        num_levels (int): The number of feature map used in Attention. Default: 4.
+        num_points (int): The number of sampling points for each query
+            in each head. Default: 4.
+        img2col_steps (int): The step used in image_to_column. Defualt: 64.
+            dropout (float): Dropout layer used in output. Default: 0.1.
+        batch_first (bool): if ``True``, then the input and output tensor will be
+            provided as `(bs, n, embed_dim)`. Default: False. `(n, bs, embed_dim)`
+    """
+
+    def __init__(
+        self,
+        embed_dim: int = 256,
+        num_heads: int = 8,
+        num_levels: int = 4,
+        num_points: int = 4,
+        img2col_step: int = 64,
+        batch_first: bool = False,
+    ):
+        super().__init__()
+        if embed_dim % num_heads != 0:
+            raise ValueError(
+                "embed_dim must be divisible by num_heads, but got {} and {}".format(
+                    embed_dim, num_heads
+                )
+            )
+        head_dim = embed_dim // num_heads
+
+        self.batch_first = batch_first
+
+        if not _is_power_of_2(head_dim):
+            warnings.warn(
+                """
+                You'd better set d_model in MSDeformAttn to make sure that
+                each dim of the attention head a power of 2, which is more efficient.
+                """
+            )
+
+        self.im2col_step = img2col_step
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.num_levels = num_levels
+        self.num_points = num_points
+        self.sampling_offsets = nn.Linear(embed_dim, num_heads * num_levels * num_points * 2)
+        self.attention_weights = nn.Linear(embed_dim, num_heads * num_levels * num_points)
+        self.value_proj = nn.Linear(embed_dim, embed_dim)
+        self.output_proj = nn.Linear(embed_dim, embed_dim)
+
+        self.init_weights()
+
+    def _reset_parameters(self):
+        return self.init_weights()
+
+    def init_weights(self):
+        """
+        Default initialization for Parameters of Module.
+        """
+        constant_(self.sampling_offsets.weight.data, 0.0)
+        thetas = torch.arange(self.num_heads, dtype=torch.float32) * (
+            2.0 * math.pi / self.num_heads
+        )
+        grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
+        grid_init = (
+            (grid_init / grid_init.abs().max(-1, keepdim=True)[0])
+            .view(self.num_heads, 1, 1, 2)
+            .repeat(1, self.num_levels, self.num_points, 1)
+        )
+        for i in range(self.num_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)
+
+    def freeze_sampling_offsets(self):
+        print("Freeze sampling offsets")
+        self.sampling_offsets.weight.requires_grad = False
+        self.sampling_offsets.bias.requires_grad = False
+
+    def freeze_attention_weights(self):
+        print("Freeze attention weights")
+        self.attention_weights.weight.requires_grad = False
+        self.attention_weights.bias.requires_grad = False
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: Optional[torch.Tensor] = None,
+        value: Optional[torch.Tensor] = None,
+        query_pos: Optional[torch.Tensor] = None,
+        key_padding_mask: Optional[torch.Tensor] = None,
+        reference_points: Optional[torch.Tensor] = None,
+        spatial_shapes: Optional[torch.Tensor] = None,
+        level_start_index: Optional[torch.Tensor] = None,
+        **kwargs
+    ) -> torch.Tensor:
+
+        """Forward Function of MultiScaleDeformableAttention
+
+        Args:
+            query (torch.Tensor): Query embeddings with shape
+                `(num_query, bs, embed_dim)`
+            key (torch.Tensor): Key embeddings with shape
+                `(num_key, bs, embed_dim)`
+            value (torch.Tensor): Value embeddings with shape
+                `(num_key, bs, embed_dim)`
+            query_pos (torch.Tensor): The position embedding for `query`. Default: None.
+            key_padding_mask (torch.Tensor): ByteTensor for `query`, with shape `(bs, num_key)`,
+                indicating which elements within `key` to be ignored in attention.
+            reference_points (torch.Tensor): The normalized reference points
+                with shape `(bs, num_query, num_levels, 2)`,
+                all elements is range in [0, 1], top-left (0, 0),
+                bottom-right (1, 1), including padding are.
+                or `(N, Length_{query}, num_levels, 4)`, add additional
+                two dimensions `(h, w)` to form reference boxes.
+            spatial_shapes (torch.Tensor): Spatial shape of features in different levels.
+                With shape `(num_levels, 2)`, last dimension represents `(h, w)`.
+            level_start_index (torch.Tensor): The start index of each level. A tensor with
+                shape `(num_levels, )` which can be represented as
+                `[0, h_0 * w_0, h_0 * w_0 + h_1 * w_1, ...]`.
+
+        Returns:
+            torch.Tensor: forward results with shape `(num_query, bs, embed_dim)`
+        """
+
+        if value is None:
+            value = query
+
+        if query_pos is not None:
+            query = query + query_pos
+
+        if not self.batch_first:
+            # change to (bs, num_query ,embed_dims)
+            query = query.permute(1, 0, 2)
+            value = value.permute(1, 0, 2)
+
+        bs, num_query, _ = query.shape
+        bs, num_value, _ = value.shape
+
+        assert (spatial_shapes[:, 0] * spatial_shapes[:, 1]).sum() == num_value
+
+        value = self.value_proj(value)
+        if key_padding_mask is not None:
+            value = value.masked_fill(key_padding_mask[..., None], float(0))
+        value = value.view(bs, num_value, self.num_heads, -1)
+        sampling_offsets = self.sampling_offsets(query).view(
+            bs, num_query, self.num_heads, self.num_levels, self.num_points, 2
+        )
+        attention_weights = self.attention_weights(query).view(
+            bs, num_query, self.num_heads, self.num_levels * self.num_points
+        )
+        attention_weights = attention_weights.softmax(-1)
+        attention_weights = attention_weights.view(
+            bs,
+            num_query,
+            self.num_heads,
+            self.num_levels,
+            self.num_points,
+        )
+
+        # bs, num_query, num_heads, num_levels, num_points, 2
+        if reference_points.shape[-1] == 2:
+            offset_normalizer = torch.stack([spatial_shapes[..., 1], spatial_shapes[..., 0]], -1)
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :]
+                + sampling_offsets / offset_normalizer[None, None, None, :, None, :]
+            )
+        elif reference_points.shape[-1] == 4:
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :2]
+                + sampling_offsets
+                / self.num_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 torch.cuda.is_available() and value.is_cuda:
+            halffloat = False
+            if value.dtype == torch.float16:
+                halffloat = True
+                value = value.float()
+                sampling_locations = sampling_locations.float()
+                attention_weights = attention_weights.float()
+
+            output = MultiScaleDeformableAttnFunction.apply(
+                value,
+                spatial_shapes,
+                level_start_index,
+                sampling_locations,
+                attention_weights,
+                self.im2col_step,
+            )
+
+            if halffloat:
+                output = output.half()
+        else:
+            output = multi_scale_deformable_attn_pytorch(
+                value, spatial_shapes, sampling_locations, attention_weights
+            )
+
+        output = self.output_proj(output)
+
+        if not self.batch_first:
+            output = output.permute(1, 0, 2)
+
+        return output
+
+
+def create_dummy_class(klass, dependency, message=""):
+    """
+    When a dependency of a class is not available, create a dummy class which throws ImportError
+    when used.
+
+    Args:
+        klass (str): name of the class.
+        dependency (str): name of the dependency.
+        message: extra message to print
+    Returns:
+        class: a class object
+    """
+    err = "Cannot import '{}', therefore '{}' is not available.".format(dependency, klass)
+    if message:
+        err = err + " " + message
+
+    class _DummyMetaClass(type):
+        # throw error on class attribute access
+        def __getattr__(_, __):  # noqa: B902
+            raise ImportError(err)
+
+    class _Dummy(object, metaclass=_DummyMetaClass):
+        # throw error on constructor
+        def __init__(self, *args, **kwargs):
+            raise ImportError(err)
+
+    return _Dummy
+
+
+def create_dummy_func(func, dependency, message=""):
+    """
+    When a dependency of a function is not available, create a dummy function which throws
+    ImportError when used.
+
+    Args:
+        func (str): name of the function.
+        dependency (str or list[str]): name(s) of the dependency.
+        message: extra message to print
+    Returns:
+        function: a function object
+    """
+    err = "Cannot import '{}', therefore '{}' is not available.".format(dependency, func)
+    if message:
+        err = err + " " + message
+
+    if isinstance(dependency, (list, tuple)):
+        dependency = ",".join(dependency)
+
+    def _dummy(*args, **kwargs):
+        raise ImportError(err)
+
+    return _dummy

CircumSpect/object_detection/groundingdino/models/GroundingDINO/transformer.py

@@ -0,0 +1,959 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# DINO
+# Copyright (c) 2022 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Conditional DETR Transformer class.
+# Copyright (c) 2021 Microsoft. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Modified from DETR (https://github.com/facebookresearch/detr)
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# ------------------------------------------------------------------------
+
+from typing import Optional
+
+import torch
+import torch.utils.checkpoint as checkpoint
+from torch import Tensor, nn
+
+from CircumSpect.object_detection.groundingdino.util.misc import inverse_sigmoid
+
+from .fuse_modules import BiAttentionBlock
+from .ms_deform_attn import MultiScaleDeformableAttention as MSDeformAttn
+from .transformer_vanilla import TransformerEncoderLayer
+from .utils import (
+    MLP,
+    _get_activation_fn,
+    _get_clones,
+    gen_encoder_output_proposals,
+    gen_sineembed_for_position,
+    get_sine_pos_embed,
+)
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        d_model=256,
+        nhead=8,
+        num_queries=300,
+        num_encoder_layers=6,
+        num_unicoder_layers=0,
+        num_decoder_layers=6,
+        dim_feedforward=2048,
+        dropout=0.0,
+        activation="relu",
+        normalize_before=False,
+        return_intermediate_dec=False,
+        query_dim=4,
+        num_patterns=0,
+        # for deformable encoder
+        num_feature_levels=1,
+        enc_n_points=4,
+        dec_n_points=4,
+        # init query
+        learnable_tgt_init=False,
+        # two stage
+        two_stage_type="no",  # ['no', 'standard', 'early', 'combine', 'enceachlayer', 'enclayer1']
+        embed_init_tgt=False,
+        # for text
+        use_text_enhancer=False,
+        use_fusion_layer=False,
+        use_checkpoint=False,
+        use_transformer_ckpt=False,
+        use_text_cross_attention=False,
+        text_dropout=0.1,
+        fusion_dropout=0.1,
+        fusion_droppath=0.0,
+    ):
+        super().__init__()
+        self.num_feature_levels = num_feature_levels
+        self.num_encoder_layers = num_encoder_layers
+        self.num_unicoder_layers = num_unicoder_layers
+        self.num_decoder_layers = num_decoder_layers
+        self.num_queries = num_queries
+        assert query_dim == 4
+
+        # choose encoder layer type
+        encoder_layer = DeformableTransformerEncoderLayer(
+            d_model, dim_feedforward, dropout, activation, num_feature_levels, nhead, enc_n_points
+        )
+
+        if use_text_enhancer:
+            text_enhance_layer = TransformerEncoderLayer(
+                d_model=d_model,
+                nhead=nhead // 2,
+                dim_feedforward=dim_feedforward // 2,
+                dropout=text_dropout,
+            )
+        else:
+            text_enhance_layer = None
+
+        if use_fusion_layer:
+            feature_fusion_layer = BiAttentionBlock(
+                v_dim=d_model,
+                l_dim=d_model,
+                embed_dim=dim_feedforward // 2,
+                num_heads=nhead // 2,
+                dropout=fusion_dropout,
+                drop_path=fusion_droppath,
+            )
+        else:
+            feature_fusion_layer = None
+
+        encoder_norm = nn.LayerNorm(d_model) if normalize_before else None
+        assert encoder_norm is None
+        self.encoder = TransformerEncoder(
+            encoder_layer,
+            num_encoder_layers,
+            d_model=d_model,
+            num_queries=num_queries,
+            text_enhance_layer=text_enhance_layer,
+            feature_fusion_layer=feature_fusion_layer,
+            use_checkpoint=use_checkpoint,
+            use_transformer_ckpt=use_transformer_ckpt,
+        )
+
+        # choose decoder layer type
+        decoder_layer = DeformableTransformerDecoderLayer(
+            d_model,
+            dim_feedforward,
+            dropout,
+            activation,
+            num_feature_levels,
+            nhead,
+            dec_n_points,
+            use_text_cross_attention=use_text_cross_attention,
+        )
+
+        decoder_norm = nn.LayerNorm(d_model)
+        self.decoder = TransformerDecoder(
+            decoder_layer,
+            num_decoder_layers,
+            decoder_norm,
+            return_intermediate=return_intermediate_dec,
+            d_model=d_model,
+            query_dim=query_dim,
+            num_feature_levels=num_feature_levels,
+        )
+
+        self.d_model = d_model
+        self.nhead = nhead
+        self.dec_layers = num_decoder_layers
+        self.num_queries = num_queries  # useful for single stage model only
+        self.num_patterns = num_patterns
+        if not isinstance(num_patterns, int):
+            Warning("num_patterns should be int but {}".format(type(num_patterns)))
+            self.num_patterns = 0
+
+        if num_feature_levels > 1:
+            if self.num_encoder_layers > 0:
+                self.level_embed = nn.Parameter(torch.Tensor(num_feature_levels, d_model))
+            else:
+                self.level_embed = None
+
+        self.learnable_tgt_init = learnable_tgt_init
+        assert learnable_tgt_init, "why not learnable_tgt_init"
+        self.embed_init_tgt = embed_init_tgt
+        if (two_stage_type != "no" and embed_init_tgt) or (two_stage_type == "no"):
+            self.tgt_embed = nn.Embedding(self.num_queries, d_model)
+            nn.init.normal_(self.tgt_embed.weight.data)
+        else:
+            self.tgt_embed = None
+
+        # for two stage
+        self.two_stage_type = two_stage_type
+        assert two_stage_type in ["no", "standard"], "unknown param {} of two_stage_type".format(
+            two_stage_type
+        )
+        if two_stage_type == "standard":
+            # anchor selection at the output of encoder
+            self.enc_output = nn.Linear(d_model, d_model)
+            self.enc_output_norm = nn.LayerNorm(d_model)
+            self.two_stage_wh_embedding = None
+
+        if two_stage_type == "no":
+            self.init_ref_points(num_queries)  # init self.refpoint_embed
+
+        self.enc_out_class_embed = None
+        self.enc_out_bbox_embed = None
+
+        self._reset_parameters()
+
+    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)
+
+    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 init_ref_points(self, use_num_queries):
+        self.refpoint_embed = nn.Embedding(use_num_queries, 4)
+
+    def forward(self, srcs, masks, refpoint_embed, pos_embeds, tgt, attn_mask=None, text_dict=None):
+        """
+        Input:
+            - srcs: List of multi features [bs, ci, hi, wi]
+            - masks: List of multi masks [bs, hi, wi]
+            - refpoint_embed: [bs, num_dn, 4]. None in infer
+            - pos_embeds: List of multi pos embeds [bs, ci, hi, wi]
+            - tgt: [bs, num_dn, d_model]. None in infer
+
+        """
+        # prepare input for encoder
+        src_flatten = []
+        mask_flatten = []
+        lvl_pos_embed_flatten = []
+        spatial_shapes = []
+        for lvl, (src, mask, pos_embed) in enumerate(zip(srcs, masks, pos_embeds)):
+            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.level_embed is not None:
+                lvl_pos_embed = pos_embed + self.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)  # bs, \sum{hxw}, c
+        mask_flatten = torch.cat(mask_flatten, 1)  # bs, \sum{hxw}
+        lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)  # bs, \sum{hxw}, c
+        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.get_valid_ratio(m) for m in masks], 1)
+
+        # two stage
+        enc_topk_proposals = enc_refpoint_embed = None
+
+        #########################################################
+        # Begin Encoder
+        #########################################################
+        memory, memory_text = 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,
+            memory_text=text_dict["encoded_text"],
+            text_attention_mask=~text_dict["text_token_mask"],
+            # we ~ the mask . False means use the token; True means pad the token
+            position_ids=text_dict["position_ids"],
+            text_self_attention_masks=text_dict["text_self_attention_masks"],
+        )
+        #########################################################
+        # End Encoder
+        # - memory: bs, \sum{hw}, c
+        # - mask_flatten: bs, \sum{hw}
+        # - lvl_pos_embed_flatten: bs, \sum{hw}, c
+        # - enc_intermediate_output: None or (nenc+1, bs, nq, c) or (nenc, bs, nq, c)
+        # - enc_intermediate_refpoints: None or (nenc+1, bs, nq, c) or (nenc, bs, nq, c)
+        #########################################################
+        text_dict["encoded_text"] = memory_text
+        # if os.environ.get("SHILONG_AMP_INFNAN_DEBUG") == '1':
+        #     if memory.isnan().any() | memory.isinf().any():
+        #         import ipdb; ipdb.set_trace()
+
+        if self.two_stage_type == "standard":
+            output_memory, output_proposals = gen_encoder_output_proposals(
+                memory, mask_flatten, spatial_shapes
+            )
+            output_memory = self.enc_output_norm(self.enc_output(output_memory))
+
+            if text_dict is not None:
+                enc_outputs_class_unselected = self.enc_out_class_embed(output_memory, text_dict)
+            else:
+                enc_outputs_class_unselected = self.enc_out_class_embed(output_memory)
+
+            topk_logits = enc_outputs_class_unselected.max(-1)[0]
+            enc_outputs_coord_unselected = (
+                self.enc_out_bbox_embed(output_memory) + output_proposals
+            )  # (bs, \sum{hw}, 4) unsigmoid
+            topk = self.num_queries
+
+            topk_proposals = torch.topk(topk_logits, topk, dim=1)[1]  # bs, nq
+
+            # 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)
+            )
+            if self.embed_init_tgt:
+                tgt_ = (
+                    self.tgt_embed.weight[:, None, :].repeat(1, bs, 1).transpose(0, 1)
+                )  # nq, bs, d_model
+            else:
+                tgt_ = tgt_undetach.detach()
+
+            if refpoint_embed is not None:
+                refpoint_embed = torch.cat([refpoint_embed, refpoint_embed_], dim=1)
+                tgt = torch.cat([tgt, tgt_], dim=1)
+            else:
+                refpoint_embed, tgt = refpoint_embed_, tgt_
+
+        elif self.two_stage_type == "no":
+            tgt_ = (
+                self.tgt_embed.weight[:, None, :].repeat(1, bs, 1).transpose(0, 1)
+            )  # nq, bs, d_model
+            refpoint_embed_ = (
+                self.refpoint_embed.weight[:, None, :].repeat(1, bs, 1).transpose(0, 1)
+            )  # nq, bs, 4
+
+            if refpoint_embed is not None:
+                refpoint_embed = torch.cat([refpoint_embed, refpoint_embed_], dim=1)
+                tgt = torch.cat([tgt, tgt_], dim=1)
+            else:
+                refpoint_embed, tgt = refpoint_embed_, tgt_
+
+            if self.num_patterns > 0:
+                tgt_embed = tgt.repeat(1, self.num_patterns, 1)
+                refpoint_embed = refpoint_embed.repeat(1, self.num_patterns, 1)
+                tgt_pat = self.patterns.weight[None, :, :].repeat_interleave(
+                    self.num_queries, 1
+                )  # 1, n_q*n_pat, d_model
+                tgt = tgt_embed + tgt_pat
+
+            init_box_proposal = refpoint_embed_.sigmoid()
+
+        else:
+            raise NotImplementedError("unknown two_stage_type {}".format(self.two_stage_type))
+        #########################################################
+        # End preparing tgt
+        # - tgt: bs, NQ, d_model
+        # - refpoint_embed(unsigmoid): bs, NQ, d_model
+        #########################################################
+
+        #########################################################
+        # Begin Decoder
+        #########################################################
+        hs, references = self.decoder(
+            tgt=tgt.transpose(0, 1),
+            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=attn_mask,
+            memory_text=text_dict["encoded_text"],
+            text_attention_mask=~text_dict["text_token_mask"],
+            # we ~ the mask . False means use the token; True means pad the token
+        )
+        #########################################################
+        # End Decoder
+        # hs: n_dec, bs, nq, d_model
+        # references: n_dec+1, bs, nq, query_dim
+        #########################################################
+
+        #########################################################
+        # Begin postprocess
+        #########################################################
+        if self.two_stage_type == "standard":
+            hs_enc = tgt_undetach.unsqueeze(0)
+            ref_enc = refpoint_embed_undetach.sigmoid().unsqueeze(0)
+        else:
+            hs_enc = ref_enc = None
+        #########################################################
+        # End postprocess
+        # hs_enc: (n_enc+1, bs, nq, d_model) or (1, bs, nq, d_model) or (n_enc, bs, nq, d_model) or None
+        # ref_enc: (n_enc+1, bs, nq, query_dim) or (1, bs, nq, query_dim) or (n_enc, bs, nq, d_model) or None
+        #########################################################
+
+        return hs, references, hs_enc, ref_enc, init_box_proposal
+        # hs: (n_dec, bs, nq, d_model)
+        # references: sigmoid coordinates. (n_dec+1, bs, bq, 4)
+        # hs_enc: (n_enc+1, bs, nq, d_model) or (1, bs, nq, d_model) or None
+        # ref_enc: sigmoid coordinates. \
+        #           (n_enc+1, bs, nq, query_dim) or (1, bs, nq, query_dim) or None
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self,
+        encoder_layer,
+        num_layers,
+        d_model=256,
+        num_queries=300,
+        enc_layer_share=False,
+        text_enhance_layer=None,
+        feature_fusion_layer=None,
+        use_checkpoint=False,
+        use_transformer_ckpt=False,
+    ):
+        """_summary_
+
+        Args:
+            encoder_layer (_type_): _description_
+            num_layers (_type_): _description_
+            norm (_type_, optional): _description_. Defaults to None.
+            d_model (int, optional): _description_. Defaults to 256.
+            num_queries (int, optional): _description_. Defaults to 300.
+            enc_layer_share (bool, optional): _description_. Defaults to False.
+
+        """
+        super().__init__()
+        # prepare layers
+        self.layers = []
+        self.text_layers = []
+        self.fusion_layers = []
+        if num_layers > 0:
+            self.layers = _get_clones(encoder_layer, num_layers, layer_share=enc_layer_share)
+
+            if text_enhance_layer is not None:
+                self.text_layers = _get_clones(
+                    text_enhance_layer, num_layers, layer_share=enc_layer_share
+                )
+            if feature_fusion_layer is not None:
+                self.fusion_layers = _get_clones(
+                    feature_fusion_layer, num_layers, layer_share=enc_layer_share
+                )
+        else:
+            self.layers = []
+            del encoder_layer
+
+            if text_enhance_layer is not None:
+                self.text_layers = []
+                del text_enhance_layer
+            if feature_fusion_layer is not None:
+                self.fusion_layers = []
+                del feature_fusion_layer
+
+        self.query_scale = None
+        self.num_queries = num_queries
+        self.num_layers = num_layers
+        self.d_model = d_model
+
+        self.use_checkpoint = use_checkpoint
+        self.use_transformer_ckpt = use_transformer_ckpt
+
+    @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,
+        # for images
+        src: Tensor,
+        pos: Tensor,
+        spatial_shapes: Tensor,
+        level_start_index: Tensor,
+        valid_ratios: Tensor,
+        key_padding_mask: Tensor,
+        # for texts
+        memory_text: Tensor = None,
+        text_attention_mask: Tensor = None,
+        pos_text: Tensor = None,
+        text_self_attention_masks: Tensor = None,
+        position_ids: Tensor = None,
+    ):
+        """
+        Input:
+            - src: [bs, sum(hi*wi), 256]
+            - pos: pos embed for src. [bs, sum(hi*wi), 256]
+            - spatial_shapes: h,w of each level [num_level, 2]
+            - level_start_index: [num_level] start point of level in sum(hi*wi).
+            - valid_ratios: [bs, num_level, 2]
+            - key_padding_mask: [bs, sum(hi*wi)]
+
+            - memory_text: bs, n_text, 256
+            - text_attention_mask: bs, n_text
+                False for no padding; True for padding
+            - pos_text: bs, n_text, 256
+
+            - position_ids: bs, n_text
+        Intermedia:
+            - reference_points: [bs, sum(hi*wi), num_level, 2]
+        Outpus:
+            - output: [bs, sum(hi*wi), 256]
+        """
+
+        output = src
+
+        # preparation and reshape
+        if self.num_layers > 0:
+            reference_points = self.get_reference_points(
+                spatial_shapes, valid_ratios, device=src.device
+            )
+
+        if self.text_layers:
+            # generate pos_text
+            bs, n_text, text_dim = memory_text.shape
+            if pos_text is None and position_ids is None:
+                pos_text = (
+                    torch.arange(n_text, device=memory_text.device)
+                    .float()
+                    .unsqueeze(0)
+                    .unsqueeze(-1)
+                    .repeat(bs, 1, 1)
+                )
+                pos_text = get_sine_pos_embed(pos_text, num_pos_feats=256, exchange_xy=False)
+            if position_ids is not None:
+                pos_text = get_sine_pos_embed(
+                    position_ids[..., None], num_pos_feats=256, exchange_xy=False
+                )
+
+        # main process
+        for layer_id, layer in enumerate(self.layers):
+            # if output.isnan().any() or memory_text.isnan().any():
+            #     if os.environ.get('IPDB_SHILONG_DEBUG', None) == 'INFO':
+            #         import ipdb; ipdb.set_trace()
+            if self.fusion_layers:
+                if self.use_checkpoint:
+                    output, memory_text = checkpoint.checkpoint(
+                        self.fusion_layers[layer_id],
+                        output,
+                        memory_text,
+                        key_padding_mask,
+                        text_attention_mask,
+                    )
+                else:
+                    output, memory_text = self.fusion_layers[layer_id](
+                        v=output,
+                        l=memory_text,
+                        attention_mask_v=key_padding_mask,
+                        attention_mask_l=text_attention_mask,
+                    )
+
+            if self.text_layers:
+                memory_text = self.text_layers[layer_id](
+                    src=memory_text.transpose(0, 1),
+                    src_mask=~text_self_attention_masks,  # note we use ~ for mask here
+                    src_key_padding_mask=text_attention_mask,
+                    pos=(pos_text.transpose(0, 1) if pos_text is not None else None),
+                ).transpose(0, 1)
+
+            # 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,
+                )
+
+        return output, memory_text
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(
+        self,
+        decoder_layer,
+        num_layers,
+        norm=None,
+        return_intermediate=False,
+        d_model=256,
+        query_dim=4,
+        num_feature_levels=1,
+    ):
+        super().__init__()
+        if num_layers > 0:
+            self.layers = _get_clones(decoder_layer, num_layers)
+        else:
+            self.layers = []
+        self.num_layers = num_layers
+        self.norm = norm
+        self.return_intermediate = return_intermediate
+        assert return_intermediate, "support return_intermediate only"
+        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.ref_point_head = MLP(query_dim // 2 * d_model, d_model, d_model, 2)
+        self.query_pos_sine_scale = None
+
+        self.query_scale = None
+        self.bbox_embed = None
+        self.class_embed = None
+
+        self.d_model = d_model
+
+        self.ref_anchor_head = None
+
+    def forward(
+        self,
+        tgt,
+        memory,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        refpoints_unsigmoid: Optional[Tensor] = None,  # num_queries, bs, 2
+        # for memory
+        level_start_index: Optional[Tensor] = None,  # num_levels
+        spatial_shapes: Optional[Tensor] = None,  # bs, num_levels, 2
+        valid_ratios: Optional[Tensor] = None,
+        # for text
+        memory_text: Optional[Tensor] = None,
+        text_attention_mask: Optional[Tensor] = None,
+    ):
+        """
+        Input:
+            - tgt: nq, bs, d_model
+            - memory: hw, bs, d_model
+            - pos: hw, bs, d_model
+            - refpoints_unsigmoid: nq, bs, 2/4
+            - valid_ratios/spatial_shapes: bs, nlevel, 2
+        """
+        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
+            raw_query_pos = self.ref_point_head(query_sine_embed)  # nq, bs, 256
+            pos_scale = self.query_scale(output) if self.query_scale is not None else 1
+            query_pos = pos_scale * raw_query_pos
+            # if os.environ.get("SHILONG_AMP_INFNAN_DEBUG") == '1':
+            #     if query_pos.isnan().any() | query_pos.isinf().any():
+            #         import ipdb; ipdb.set_trace()
+
+            # main process
+            output = layer(
+                tgt=output,
+                tgt_query_pos=query_pos,
+                tgt_query_sine_embed=query_sine_embed,
+                tgt_key_padding_mask=tgt_key_padding_mask,
+                tgt_reference_points=reference_points_input,
+                memory_text=memory_text,
+                text_attention_mask=text_attention_mask,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+                memory_level_start_index=level_start_index,
+                memory_spatial_shapes=spatial_shapes,
+                memory_pos=pos,
+                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)
+                    # if os.environ.get("SHILONG_AMP_INFNAN_DEBUG") == '1':
+                    #     import ipdb; ipdb.set_trace()
+
+            # iter update
+            if self.bbox_embed is not None:
+                # box_holder = self.bbox_embed(output)
+                # box_holder[..., :self.query_dim] += inverse_sigmoid(reference_points)
+                # new_reference_points = box_holder[..., :self.query_dim].sigmoid()
+
+                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()
+
+                reference_points = new_reference_points.detach()
+                # if layer_id != self.num_layers - 1:
+                ref_points.append(new_reference_points)
+
+            intermediate.append(self.norm(output))
+
+        return [
+            [itm_out.transpose(0, 1) for itm_out in intermediate],
+            [itm_refpoint.transpose(0, 1) for itm_refpoint in ref_points],
+        ]
+
+
+class DeformableTransformerEncoderLayer(nn.Module):
+    def __init__(
+        self,
+        d_model=256,
+        d_ffn=1024,
+        dropout=0.1,
+        activation="relu",
+        n_levels=4,
+        n_heads=8,
+        n_points=4,
+    ):
+        super().__init__()
+
+        # self attention
+        self.self_attn = MSDeformAttn(
+            embed_dim=d_model,
+            num_levels=n_levels,
+            num_heads=n_heads,
+            num_points=n_points,
+            batch_first=True,
+        )
+        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)
+
+    @staticmethod
+    def with_pos_embed(tensor, pos):
+        return tensor if pos is None else tensor + pos
+
+    def forward_ffn(self, src):
+        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, pos, reference_points, spatial_shapes, level_start_index, key_padding_mask=None
+    ):
+        # self attention
+        # import ipdb; ipdb.set_trace()
+        src2 = self.self_attn(
+            query=self.with_pos_embed(src, pos),
+            reference_points=reference_points,
+            value=src,
+            spatial_shapes=spatial_shapes,
+            level_start_index=level_start_index,
+            key_padding_mask=key_padding_mask,
+        )
+        src = src + self.dropout1(src2)
+        src = self.norm1(src)
+
+        # ffn
+        src = self.forward_ffn(src)
+
+        return src
+
+
+class DeformableTransformerDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        d_model=256,
+        d_ffn=1024,
+        dropout=0.1,
+        activation="relu",
+        n_levels=4,
+        n_heads=8,
+        n_points=4,
+        use_text_feat_guide=False,
+        use_text_cross_attention=False,
+    ):
+        super().__init__()
+
+        # cross attention
+        self.cross_attn = MSDeformAttn(
+            embed_dim=d_model,
+            num_levels=n_levels,
+            num_heads=n_heads,
+            num_points=n_points,
+            batch_first=True,
+        )
+        self.dropout1 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm1 = nn.LayerNorm(d_model)
+
+        # cross attention text
+        if use_text_cross_attention:
+            self.ca_text = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)
+            self.catext_dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+            self.catext_norm = nn.LayerNorm(d_model)
+
+        # self attention
+        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)
+
+        self.key_aware_proj = None
+        self.use_text_feat_guide = use_text_feat_guide
+        assert not use_text_feat_guide
+        self.use_text_cross_attention = use_text_cross_attention
+
+    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):
+        with torch.cuda.amp.autocast(enabled=False):
+            tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout4(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward(
+        self,
+        # for tgt
+        tgt: Optional[Tensor],  # nq, bs, d_model
+        tgt_query_pos: Optional[Tensor] = None,  # pos for query. MLP(Sine(pos))
+        tgt_query_sine_embed: Optional[Tensor] = None,  # pos for query. Sine(pos)
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        tgt_reference_points: Optional[Tensor] = None,  # nq, bs, 4
+        memory_text: Optional[Tensor] = None,  # bs, num_token, d_model
+        text_attention_mask: Optional[Tensor] = None,  # bs, num_token
+        # for memory
+        memory: Optional[Tensor] = None,  # hw, bs, d_model
+        memory_key_padding_mask: Optional[Tensor] = None,
+        memory_level_start_index: Optional[Tensor] = None,  # num_levels
+        memory_spatial_shapes: Optional[Tensor] = None,  # bs, num_levels, 2
+        memory_pos: Optional[Tensor] = None,  # pos for memory
+        # sa
+        self_attn_mask: Optional[Tensor] = None,  # mask used for self-attention
+        cross_attn_mask: Optional[Tensor] = None,  # mask used for cross-attention
+    ):
+        """
+        Input:
+            - tgt/tgt_query_pos: nq, bs, d_model
+            -
+        """
+        assert cross_attn_mask is None
+
+        # self attention
+        if self.self_attn is not None:
+            # import ipdb; ipdb.set_trace()
+            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)
+
+        if self.use_text_cross_attention:
+            tgt2 = self.ca_text(
+                self.with_pos_embed(tgt, tgt_query_pos),
+                memory_text.transpose(0, 1),
+                memory_text.transpose(0, 1),
+                key_padding_mask=text_attention_mask,
+            )[0]
+            tgt = tgt + self.catext_dropout(tgt2)
+            tgt = self.catext_norm(tgt)
+
+        tgt2 = self.cross_attn(
+            query=self.with_pos_embed(tgt, tgt_query_pos).transpose(0, 1),
+            reference_points=tgt_reference_points.transpose(0, 1).contiguous(),
+            value=memory.transpose(0, 1),
+            spatial_shapes=memory_spatial_shapes,
+            level_start_index=memory_level_start_index,
+            key_padding_mask=memory_key_padding_mask,
+        ).transpose(0, 1)
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+
+        # ffn
+        tgt = self.forward_ffn(tgt)
+
+        return tgt
+
+
+def build_transformer(args):
+    return Transformer(
+        d_model=args.hidden_dim,
+        dropout=args.dropout,
+        nhead=args.nheads,
+        num_queries=args.num_queries,
+        dim_feedforward=args.dim_feedforward,
+        num_encoder_layers=args.enc_layers,
+        num_decoder_layers=args.dec_layers,
+        normalize_before=args.pre_norm,
+        return_intermediate_dec=True,
+        query_dim=args.query_dim,
+        activation=args.transformer_activation,
+        num_patterns=args.num_patterns,
+        num_feature_levels=args.num_feature_levels,
+        enc_n_points=args.enc_n_points,
+        dec_n_points=args.dec_n_points,
+        learnable_tgt_init=True,
+        # two stage
+        two_stage_type=args.two_stage_type,  # ['no', 'standard', 'early']
+        embed_init_tgt=args.embed_init_tgt,
+        use_text_enhancer=args.use_text_enhancer,
+        use_fusion_layer=args.use_fusion_layer,
+        use_checkpoint=args.use_checkpoint,
+        use_transformer_ckpt=args.use_transformer_ckpt,
+        use_text_cross_attention=args.use_text_cross_attention,
+        text_dropout=args.text_dropout,
+        fusion_dropout=args.fusion_dropout,
+        fusion_droppath=args.fusion_droppath,
+    )

CircumSpect/object_detection/groundingdino/models/GroundingDINO/transformer_vanilla.py

@@ -0,0 +1,123 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copyright (c) Aishwarya Kamath & Nicolas Carion. Licensed under the Apache License 2.0. All Rights Reserved
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+DETR Transformer class.
+
+Copy-paste from torch.nn.Transformer with modifications:
+    * positional encodings are passed in MHattention
+    * extra LN at the end of encoder is removed
+    * decoder returns a stack of activations from all decoding layers
+"""
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+
+from .utils import (
+    MLP,
+    _get_activation_fn,
+    _get_clones,
+    gen_encoder_output_proposals,
+    gen_sineembed_for_position,
+    sigmoid_focal_loss,
+)
+
+
+class TextTransformer(nn.Module):
+    def __init__(self, num_layers, d_model=256, nheads=8, dim_feedforward=2048, dropout=0.1):
+        super().__init__()
+        self.num_layers = num_layers
+        self.d_model = d_model
+        self.nheads = nheads
+        self.dim_feedforward = dim_feedforward
+        self.norm = None
+
+        single_encoder_layer = TransformerEncoderLayer(
+            d_model=d_model, nhead=nheads, dim_feedforward=dim_feedforward, dropout=dropout
+        )
+        self.layers = _get_clones(single_encoder_layer, num_layers)
+
+    def forward(self, memory_text: torch.Tensor, text_attention_mask: torch.Tensor):
+        """
+
+        Args:
+            text_attention_mask: bs, num_token
+            memory_text: bs, num_token, d_model
+
+        Raises:
+            RuntimeError: _description_
+
+        Returns:
+            output: bs, num_token, d_model
+        """
+
+        output = memory_text.transpose(0, 1)
+
+        for layer in self.layers:
+            output = layer(output, src_key_padding_mask=text_attention_mask)
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output.transpose(0, 1)
+
+
+class TransformerEncoderLayer(nn.Module):
+    def __init__(
+        self,
+        d_model,
+        nhead,
+        dim_feedforward=2048,
+        dropout=0.1,
+        activation="relu",
+        normalize_before=False,
+    ):
+        super().__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+        self.normalize_before = normalize_before
+        self.nhead = nhead
+
+    def with_pos_embed(self, tensor, pos: Optional[Tensor]):
+        return tensor if pos is None else tensor + pos
+
+    def forward(
+        self,
+        src,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+    ):
+        # repeat attn mask
+        if src_mask.dim() == 3 and src_mask.shape[0] == src.shape[1]:
+            # bs, num_q, num_k
+            src_mask = src_mask.repeat(self.nhead, 1, 1)
+
+        q = k = self.with_pos_embed(src, pos)
+
+        src2 = self.self_attn(q, k, value=src, attn_mask=src_mask)[0]
+
+        # src2 = self.self_attn(q, k, value=src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask)[0]
+        src = src + self.dropout1(src2)
+        src = self.norm1(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = src + self.dropout2(src2)
+        src = self.norm2(src)
+        return src

CircumSpect/object_detection/groundingdino/models/GroundingDINO/utils.py

@@ -0,0 +1,268 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+
+import copy
+import math
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+
+
+def _get_clones(module, N, layer_share=False):
+    # import ipdb; ipdb.set_trace()
+    if layer_share:
+        return nn.ModuleList([module for i in range(N)])
+    else:
+        return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+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: Tensor, memory_padding_mask: Tensor, spatial_shapes: 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
+    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)
+
+        # import ipdb; ipdb.set_trace()
+
+        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:
+            # import ipdb; ipdb.set_trace()
+            wh = torch.ones_like(grid) * learnedwh.sigmoid() * (2.0**lvl)
+        else:
+            wh = torch.ones_like(grid) * 0.05 * (2.0**lvl)
+
+        # scale = torch.cat([W_[None].unsqueeze(-1), H_[None].unsqueeze(-1)], 1).view(1, 1, 1, 2).repeat(N_, 1, 1, 1)
+        # grid = (grid.unsqueeze(0).expand(N_, -1, -1, -1) + 0.5) / scale
+        # wh = torch.ones_like(grid) / scale
+        proposal = torch.cat((grid, wh), -1).view(N_, -1, 4)
+        proposals.append(proposal)
+        _cur += H_ * W_
+    # import ipdb; ipdb.set_trace()
+    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))
+
+    # output_memory = output_memory.masked_fill(memory_padding_mask.unsqueeze(-1), float('inf'))
+    # output_memory = output_memory.masked_fill(~output_proposals_valid, float('inf'))
+
+    return output_memory, output_proposals
+
+
+class RandomBoxPerturber:
+    def __init__(
+        self, x_noise_scale=0.2, y_noise_scale=0.2, w_noise_scale=0.2, h_noise_scale=0.2
+    ) -> None:
+        self.noise_scale = torch.Tensor(
+            [x_noise_scale, y_noise_scale, w_noise_scale, h_noise_scale]
+        )
+
+    def __call__(self, refanchors: Tensor) -> Tensor:
+        nq, bs, query_dim = refanchors.shape
+        device = refanchors.device
+
+        noise_raw = torch.rand_like(refanchors)
+        noise_scale = self.noise_scale.to(device)[:query_dim]
+
+        new_refanchors = refanchors * (1 + (noise_raw - 0.5) * noise_scale)
+        return new_refanchors.clamp_(0, 1)
+
+
+def sigmoid_focal_loss(
+    inputs, targets, num_boxes, alpha: float = 0.25, gamma: float = 2, no_reduction=False
+):
+    """
+    Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
+    Args:
+        inputs: A float tensor of arbitrary shape.
+                The predictions for each example.
+        targets: A float tensor with the same shape as inputs. Stores the binary
+                 classification label for each element in inputs
+                (0 for the negative class and 1 for the positive class).
+        alpha: (optional) Weighting factor in range (0,1) to balance
+                positive vs negative examples. Default = -1 (no weighting).
+        gamma: Exponent of the modulating factor (1 - p_t) to
+               balance easy vs hard examples.
+    Returns:
+        Loss tensor
+    """
+    prob = inputs.sigmoid()
+    ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
+    p_t = prob * targets + (1 - prob) * (1 - targets)
+    loss = ce_loss * ((1 - p_t) ** gamma)
+
+    if alpha >= 0:
+        alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
+        loss = alpha_t * loss
+
+    if no_reduction:
+        return loss
+
+    return loss.mean(1).sum() / num_boxes
+
+
+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
+
+
+def _get_activation_fn(activation, d_model=256, batch_dim=0):
+    """Return an activation function given a string"""
+    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 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 * (torch.div(dim_t, 2, rounding_mode='floor')) / 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
+
+
+class ContrastiveEmbed(nn.Module):
+    def __init__(self, max_text_len=256):
+        """
+        Args:
+            max_text_len: max length of text.
+        """
+        super().__init__()
+        self.max_text_len = max_text_len
+
+    def forward(self, x, text_dict):
+        """_summary_
+
+        Args:
+            x (_type_): _description_
+            text_dict (_type_): _description_
+            {
+                'encoded_text': encoded_text, # bs, 195, d_model
+                'text_token_mask': text_token_mask, # bs, 195
+                        # True for used tokens. False for padding tokens
+            }
+        Returns:
+            _type_: _description_
+        """
+        assert isinstance(text_dict, dict)
+
+        y = text_dict["encoded_text"]
+        text_token_mask = text_dict["text_token_mask"]
+
+        res = x @ y.transpose(-1, -2)
+        res.masked_fill_(~text_token_mask[:, None, :], float("-inf"))
+
+        # padding to max_text_len
+        new_res = torch.full((*res.shape[:-1], self.max_text_len), float("-inf"), device=res.device)
+        new_res[..., : res.shape[-1]] = res
+
+        return new_res

CircumSpect/object_detection/groundingdino/models/__init__.py

@@ -0,0 +1,18 @@
+# ------------------------------------------------------------------------
+# Grounding DINO
+# url: https://github.com/IDEA-Research/GroundingDINO
+# Copyright (c) 2023 IDEA. All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+# ------------------------------------------------------------------------
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .GroundingDINO import build_groundingdino
+
+
+def build_model(args):
+    # we use register to maintain models from catdet6 on.
+    from .registry import MODULE_BUILD_FUNCS
+
+    assert args.modelname in MODULE_BUILD_FUNCS._module_dict
+    build_func = MODULE_BUILD_FUNCS.get(args.modelname)
+    model = build_func(args)
+    return model