Merge branch 'main' of https://huggingface.co/spaces/Hasanmog/Peft-GroundingDINO

groundingdino/datasets/.ipynb_checkpoints/__init__-checkpoint.py

@@ -1,23 +0,0 @@
-# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
-import torch.utils.data
-import torchvision
-from .coco import build as build_coco
-
-
-def get_coco_api_from_dataset(dataset):
-    for _ in range(10):
-        # if isinstance(dataset, torchvision.datasets.CocoDetection):
-        #     break
-        if isinstance(dataset, torch.utils.data.Subset):
-            dataset = dataset.dataset
-    if isinstance(dataset, torchvision.datasets.CocoDetection):
-        return dataset.coco
-
-
-def build_dataset(image_set, args, datasetinfo):
-    if datasetinfo["dataset_mode"] == 'coco':
-        return build_coco(image_set, args, datasetinfo)
-    if datasetinfo["dataset_mode"] == 'odvg':
-        from .odvg import build_odvg
-        return build_odvg(image_set, args, datasetinfo)
-    raise ValueError(f'dataset {args.dataset_file} not supported')

groundingdino/datasets/.ipynb_checkpoints/coco-checkpoint.py

@@ -1,649 +0,0 @@
-# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
-"""
-COCO dataset which returns image_id for evaluation.
-
-Mostly copy-paste from https://github.com/pytorch/vision/blob/13b35ff/references/detection/coco_utils.py
-"""
-if __name__=="__main__":
-    # for debug only
-    import os, sys
-    sys.path.append(os.path.dirname(sys.path[0]))
-from torchvision.datasets.vision import VisionDataset
-
-import json
-from pathlib import Path
-import random
-import os
-from typing import Any, Callable, List, Optional, Tuple
-
-from PIL import Image
-
-import torch
-import torch.utils.data
-import torchvision
-from pycocotools import mask as coco_mask
-
-from datasets.data_util import preparing_dataset
-import datasets.transforms as T
-from util.box_ops import box_cxcywh_to_xyxy, box_iou
-
-__all__ = ['build']
-
-
-class label2compat():
-    def __init__(self) -> None:
-        self.category_map_str = {"1": 1, "2": 2, "3": 3, "4": 4, "5": 5, "6": 6, "7": 7, "8": 8, "9": 9, "10": 10, "11": 11, "13": 12, "14": 13, "15": 14, "16": 15, "17": 16, "18": 17, "19": 18, "20": 19, "21": 20, "22": 21, "23": 22, "24": 23, "25": 24, "27": 25, "28": 26, "31": 27, "32": 28, "33": 29, "34": 30, "35": 31, "36": 32, "37": 33, "38": 34, "39": 35, "40": 36, "41": 37, "42": 38, "43": 39, "44": 40, "46": 41, "47": 42, "48": 43, "49": 44, "50": 45, "51": 46, "52": 47, "53": 48, "54": 49, "55": 50, "56": 51, "57": 52, "58": 53, "59": 54, "60": 55, "61": 56, "62": 57, "63": 58, "64": 59, "65": 60, "67": 61, "70": 62, "72": 63, "73": 64, "74": 65, "75": 66, "76": 67, "77": 68, "78": 69, "79": 70, "80": 71, "81": 72, "82": 73, "84": 74, "85": 75, "86": 76, "87": 77, "88": 78, "89": 79, "90": 80}
-        self.category_map = {int(k):v for k,v in self.category_map_str.items()}
-
-    def __call__(self, target, img=None):
-        labels = target['labels']
-        res = torch.zeros(labels.shape, dtype=labels.dtype)
-        for idx, item in enumerate(labels):
-            res[idx] = self.category_map[item.item()] - 1
-        target['label_compat'] = res
-        if img is not None:
-            return target, img
-        else:
-            return target
-
-
-class label_compat2onehot():
-    def __init__(self, num_class=80, num_output_objs=1):
-        self.num_class = num_class
-        self.num_output_objs = num_output_objs
-        if num_output_objs != 1:
-            raise DeprecationWarning("num_output_objs!=1, which is only used for comparison")
-
-    def __call__(self, target, img=None):
-        labels = target['label_compat']
-        place_dict = {k:0 for k in range(self.num_class)}
-        if self.num_output_objs == 1:
-            res = torch.zeros(self.num_class)
-            for i in labels:
-                itm = i.item()
-                res[itm] = 1.0
-        else:
-            # compat with baseline
-            res = torch.zeros(self.num_class, self.num_output_objs)
-            for i in labels:
-                itm = i.item()
-                res[itm][place_dict[itm]] = 1.0
-                place_dict[itm] += 1
-        target['label_compat_onehot'] = res
-        if img is not None:
-            return target, img
-        else:
-            return target
-
-
-class box_label_catter():
-    def __init__(self):
-        pass
-
-    def __call__(self, target, img=None):
-        labels = target['label_compat']
-        boxes = target['boxes']
-        box_label = torch.cat((boxes, labels.unsqueeze(-1)), 1)
-        target['box_label'] = box_label
-        if img is not None:
-            return target, img
-        else:
-            return target
-
-
-class RandomSelectBoxlabels():
-    def __init__(self, num_classes, leave_one_out=False, blank_prob=0.8,
-                    prob_first_item = 0.0,
-                    prob_random_item = 0.0,
-                    prob_last_item = 0.8,
-                    prob_stop_sign = 0.2
-                ) -> None:
-        self.num_classes = num_classes
-        self.leave_one_out = leave_one_out
-        self.blank_prob = blank_prob
-
-        self.set_state(prob_first_item, prob_random_item, prob_last_item, prob_stop_sign)
-        
-
-    def get_state(self):
-        return [self.prob_first_item, self.prob_random_item, self.prob_last_item, self.prob_stop_sign]
-
-    def set_state(self, prob_first_item, prob_random_item, prob_last_item, prob_stop_sign):
-        sum_prob = prob_first_item + prob_random_item + prob_last_item + prob_stop_sign
-        assert sum_prob - 1 < 1e-6, \
-            f"Sum up all prob = {sum_prob}. prob_first_item:{prob_first_item}" \
-            + f"prob_random_item:{prob_random_item}, prob_last_item:{prob_last_item}" \
-            + f"prob_stop_sign:{prob_stop_sign}"
-
-        self.prob_first_item = prob_first_item
-        self.prob_random_item = prob_random_item
-        self.prob_last_item = prob_last_item
-        self.prob_stop_sign = prob_stop_sign
-        
-
-    def sample_for_pred_first_item(self, box_label: torch.FloatTensor):
-        box_label_known = torch.Tensor(0,5)
-        box_label_unknown = box_label
-        return box_label_known, box_label_unknown
-
-    def sample_for_pred_random_item(self, box_label: torch.FloatTensor):
-        n_select = int(random.random() * box_label.shape[0])
-        box_label = box_label[torch.randperm(box_label.shape[0])]
-        box_label_known = box_label[:n_select]
-        box_label_unknown = box_label[n_select:]
-        return box_label_known, box_label_unknown
-
-    def sample_for_pred_last_item(self, box_label: torch.FloatTensor):
-        box_label_perm = box_label[torch.randperm(box_label.shape[0])]
-        known_label_list = []
-        box_label_known = []
-        box_label_unknown = []
-        for item in box_label_perm:
-            label_i = item[4].item()
-            if label_i in known_label_list:
-                box_label_known.append(item)
-            else:
-                # first item
-                box_label_unknown.append(item)
-                known_label_list.append(label_i)
-        box_label_known = torch.stack(box_label_known) if len(box_label_known) > 0 else torch.Tensor(0,5)
-        box_label_unknown = torch.stack(box_label_unknown) if len(box_label_unknown) > 0 else torch.Tensor(0,5)
-        return box_label_known, box_label_unknown
-
-    def sample_for_pred_stop_sign(self, box_label: torch.FloatTensor):
-        box_label_unknown = torch.Tensor(0,5)
-        box_label_known = box_label
-        return box_label_known, box_label_unknown
-
-    def __call__(self, target, img=None):
-        box_label = target['box_label'] # K, 5
-
-        dice_number = random.random()
-
-        if dice_number < self.prob_first_item:
-            box_label_known, box_label_unknown = self.sample_for_pred_first_item(box_label)
-        elif dice_number < self.prob_first_item + self.prob_random_item:
-            box_label_known, box_label_unknown = self.sample_for_pred_random_item(box_label)
-        elif dice_number < self.prob_first_item + self.prob_random_item + self.prob_last_item:
-            box_label_known, box_label_unknown = self.sample_for_pred_last_item(box_label)
-        else:
-            box_label_known, box_label_unknown = self.sample_for_pred_stop_sign(box_label)
-
-        target['label_onehot_known'] = label2onehot(box_label_known[:,-1], self.num_classes)
-        target['label_onehot_unknown'] = label2onehot(box_label_unknown[:, -1], self.num_classes)
-        target['box_label_known'] = box_label_known
-        target['box_label_unknown'] = box_label_unknown
-
-        return target, img
-
-
-class RandomDrop():
-    def __init__(self, p=0.2) -> None:
-        self.p = p
-
-    def __call__(self, target, img=None):
-        known_box = target['box_label_known']
-        num_known_box = known_box.size(0)
-        idxs = torch.rand(num_known_box)
-        # indices = torch.randperm(num_known_box)[:int((1-self).p*num_known_box + 0.5 + random.random())]
-        target['box_label_known'] = known_box[idxs > self.p]
-        return target, img
-
-
-class BboxPertuber():
-    def __init__(self, max_ratio = 0.02, generate_samples = 1000) -> None:
-        self.max_ratio = max_ratio
-        self.generate_samples = generate_samples
-        self.samples = self.generate_pertube_samples()
-        self.idx = 0
-
-    def generate_pertube_samples(self):
-        import torch
-        samples = (torch.rand(self.generate_samples, 5) - 0.5) * 2 * self.max_ratio
-        return samples
-
-    def __call__(self, target, img):
-        known_box = target['box_label_known'] # Tensor(K,5), K known bbox
-        K = known_box.shape[0]
-        known_box_pertube = torch.zeros(K, 6) # 4:bbox, 1:prob, 1:label
-        if K == 0:
-            pass
-        else:
-            if self.idx + K > self.generate_samples:
-                self.idx = 0
-            delta = self.samples[self.idx: self.idx + K, :]
-            known_box_pertube[:, :4] = known_box[:, :4] + delta[:, :4]
-            iou = (torch.diag(box_iou(box_cxcywh_to_xyxy(known_box[:, :4]), box_cxcywh_to_xyxy(known_box_pertube[:, :4]))[0])) * (1 + delta[:, -1])
-            known_box_pertube[:, 4].copy_(iou)
-            known_box_pertube[:, -1].copy_(known_box[:, -1])
-
-        target['box_label_known_pertube'] = known_box_pertube
-        return target, img
-
-
-class RandomCutout():
-    def __init__(self, factor=0.5) -> None:
-        self.factor = factor
-
-    def __call__(self, target, img=None):
-        unknown_box = target['box_label_unknown']           # Ku, 5
-        known_box = target['box_label_known_pertube']       # Kk, 6
-        Ku = unknown_box.size(0)
-
-        known_box_add = torch.zeros(Ku, 6) # Ku, 6
-        known_box_add[:, :5] = unknown_box
-        known_box_add[:, 5].uniform_(0.5, 1) 
-        
-
-        known_box_add[:, :2] += known_box_add[:, 2:4] * (torch.rand(Ku, 2) - 0.5) / 2
-        known_box_add[:, 2:4] /= 2
-
-        target['box_label_known_pertube'] = torch.cat((known_box, known_box_add))
-        return target, img
-
-
-class RandomSelectBoxes():
-    def __init__(self, num_class=80) -> None:
-        Warning("This is such a slow function and will be deprecated soon!!!")
-        self.num_class = num_class
-
-    def __call__(self, target, img=None):
-        boxes = target['boxes']
-        labels = target['label_compat']
-
-        # transform to list of tensors
-        boxs_list = [[] for i in range(self.num_class)]
-        for idx, item in enumerate(boxes):
-            label = labels[idx].item()
-            boxs_list[label].append(item)
-        boxs_list_tensor = [torch.stack(i) if len(i) > 0 else torch.Tensor(0,4) for i in boxs_list]
-
-        # random selection
-        box_known = []
-        box_unknown = []
-        for idx, item in enumerate(boxs_list_tensor):
-            ncnt = item.shape[0]
-            nselect = int(random.random() * ncnt) # close in both sides, much faster than random.randint
-
-            item = item[torch.randperm(ncnt)]
-            # random.shuffle(item)
-            box_known.append(item[:nselect])
-            box_unknown.append(item[nselect:])
-
-        # box_known_tensor = [torch.stack(i) if len(i) > 0 else torch.Tensor(0,4) for i in box_known]
-        # box_unknown_tensor = [torch.stack(i) if len(i) > 0 else torch.Tensor(0,4) for i in box_unknown]
-        # print('box_unknown_tensor:', box_unknown_tensor)
-        target['known_box'] = box_known
-        target['unknown_box'] = box_unknown
-        return target, img
-
-
-def label2onehot(label, num_classes):
-    """
-    label: Tensor(K)
-    """
-    res = torch.zeros(num_classes)
-    for i in label:
-        itm = int(i.item())
-        res[itm] = 1.0
-    return res
-
-
-class MaskCrop():
-    def __init__(self) -> None:
-        pass
-
-    def __call__(self, target, img):
-        known_box = target['known_box']
-        h,w = img.shape[1:] # h,w
-        # imgsize = target['orig_size'] # h,w
-
-        scale = torch.Tensor([w, h, w, h])
-
-        # _cnt = 0
-        for boxes in known_box:
-            if boxes.shape[0] == 0:
-                continue
-            box_xyxy = box_cxcywh_to_xyxy(boxes) * scale
-            for box in box_xyxy:
-                x1, y1, x2, y2 = [int(i) for i in box.tolist()]
-                img[:, y1:y2, x1:x2] = 0
-                # _cnt += 1
-        # print("_cnt:", _cnt)
-        return target, img
-
-
-dataset_hook_register = {
-    'label2compat': label2compat,
-    'label_compat2onehot': label_compat2onehot,
-    'box_label_catter': box_label_catter,
-    'RandomSelectBoxlabels': RandomSelectBoxlabels,
-    'RandomSelectBoxes': RandomSelectBoxes,
-    'MaskCrop': MaskCrop,
-    'BboxPertuber': BboxPertuber,
-}
-
-
-class CocoDetection(torchvision.datasets.CocoDetection):
-    def __init__(self, img_folder, ann_file, transforms, return_masks, aux_target_hacks=None):
-        super(CocoDetection, self).__init__(img_folder, ann_file)
-        self._transforms = transforms
-        self.prepare = ConvertCocoPolysToMask(return_masks)
-        self.aux_target_hacks = aux_target_hacks
-
-    def change_hack_attr(self, hackclassname, attrkv_dict):
-        target_class = dataset_hook_register[hackclassname]
-        for item in self.aux_target_hacks:
-            if isinstance(item, target_class):
-                for k,v in attrkv_dict.items():
-                    setattr(item, k, v)
-
-    def get_hack(self, hackclassname):
-        target_class = dataset_hook_register[hackclassname]
-        for item in self.aux_target_hacks:
-            if isinstance(item, target_class):
-                return item
-            
-    def _load_image(self, id: int) -> Image.Image:
-        path = self.coco.loadImgs(id)[0]["file_name"]
-        abs_path = os.path.join(self.root, path)
-        return Image.open(abs_path).convert("RGB")
-
-    def __getitem__(self, idx):
-        """
-        Output:
-            - target: dict of multiple items
-                - boxes: Tensor[num_box, 4]. \
-                    Init type: x0,y0,x1,y1. unnormalized data.
-                    Final type: cx,cy,w,h. normalized data. 
-        """
-        try:
-            img, target = super(CocoDetection, self).__getitem__(idx)
-        except:
-            print("Error idx: {}".format(idx))
-            idx += 1
-            img, target = super(CocoDetection, self).__getitem__(idx)
-        image_id = self.ids[idx]
-        target = {'image_id': image_id, 'annotations': target}
-        img, target = self.prepare(img, target)
-        
-        if self._transforms is not None:
-            img, target = self._transforms(img, target)
-
-        # convert to needed format
-        if self.aux_target_hacks is not None:
-            for hack_runner in self.aux_target_hacks:
-                target, img = hack_runner(target, img=img)
-
-        return img, target
-
-
-def convert_coco_poly_to_mask(segmentations, height, width):
-    masks = []
-    for polygons in segmentations:
-        rles = coco_mask.frPyObjects(polygons, height, width)
-        mask = coco_mask.decode(rles)
-        if len(mask.shape) < 3:
-            mask = mask[..., None]
-        mask = torch.as_tensor(mask, dtype=torch.uint8)
-        mask = mask.any(dim=2)
-        masks.append(mask)
-    if masks:
-        masks = torch.stack(masks, dim=0)
-    else:
-        masks = torch.zeros((0, height, width), dtype=torch.uint8)
-    return masks
-
-
-class ConvertCocoPolysToMask(object):
-    def __init__(self, return_masks=False):
-        self.return_masks = return_masks
-
-    def __call__(self, image, target):
-        w, h = image.size
-
-        image_id = target["image_id"]
-        image_id = torch.tensor([image_id])
-
-        anno = target["annotations"]
-
-        anno = [obj for obj in anno if 'iscrowd' not in obj or obj['iscrowd'] == 0]
-
-        boxes = [obj["bbox"] for obj in anno]
-        # guard against no boxes via resizing
-        boxes = torch.as_tensor(boxes, dtype=torch.float32).reshape(-1, 4)
-        boxes[:, 2:] += boxes[:, :2]
-        boxes[:, 0::2].clamp_(min=0, max=w)
-        boxes[:, 1::2].clamp_(min=0, max=h)
-
-        classes = [obj["category_id"] for obj in anno]
-        classes = torch.tensor(classes, dtype=torch.int64)
-
-        if self.return_masks:
-            segmentations = [obj["segmentation"] for obj in anno]
-            masks = convert_coco_poly_to_mask(segmentations, h, w)
-
-        keypoints = None
-        if anno and "keypoints" in anno[0]:
-            keypoints = [obj["keypoints"] for obj in anno]
-            keypoints = torch.as_tensor(keypoints, dtype=torch.float32)
-            num_keypoints = keypoints.shape[0]
-            if num_keypoints:
-                keypoints = keypoints.view(num_keypoints, -1, 3)
-
-        keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0])
-        boxes = boxes[keep]
-        classes = classes[keep]
-        if self.return_masks:
-            masks = masks[keep]
-        if keypoints is not None:
-            keypoints = keypoints[keep]
-
-        target = {}
-        target["boxes"] = boxes
-        target["labels"] = classes
-        if self.return_masks:
-            target["masks"] = masks
-        target["image_id"] = image_id
-        if keypoints is not None:
-            target["keypoints"] = keypoints
-
-        # for conversion to coco api
-        area = torch.tensor([obj["area"] for obj in anno])
-        iscrowd = torch.tensor([obj["iscrowd"] if "iscrowd" in obj else 0 for obj in anno])
-        target["area"] = area[keep]
-        target["iscrowd"] = iscrowd[keep]
-
-        target["orig_size"] = torch.as_tensor([int(h), int(w)])
-        target["size"] = torch.as_tensor([int(h), int(w)])
-
-        return image, target
-
-
-def make_coco_transforms(image_set, fix_size=False, strong_aug=False, args=None):
-
-    normalize = T.Compose([
-        T.ToTensor(),
-        T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
-    ])
-
-    # config the params for data aug
-    scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800]
-    max_size = 1333
-    scales2_resize = [400, 500, 600]
-    scales2_crop = [384, 600]
-    
-    # update args from config files
-    scales = getattr(args, 'data_aug_scales', scales)
-    max_size = getattr(args, 'data_aug_max_size', max_size)
-    scales2_resize = getattr(args, 'data_aug_scales2_resize', scales2_resize)
-    scales2_crop = getattr(args, 'data_aug_scales2_crop', scales2_crop)
-
-    # resize them
-    data_aug_scale_overlap = getattr(args, 'data_aug_scale_overlap', None)
-    if data_aug_scale_overlap is not None and data_aug_scale_overlap > 0:
-        data_aug_scale_overlap = float(data_aug_scale_overlap)
-        scales = [int(i*data_aug_scale_overlap) for i in scales]
-        max_size = int(max_size*data_aug_scale_overlap)
-        scales2_resize = [int(i*data_aug_scale_overlap) for i in scales2_resize]
-        scales2_crop = [int(i*data_aug_scale_overlap) for i in scales2_crop]
-
-    datadict_for_print = {
-        'scales': scales,
-        'max_size': max_size,
-        'scales2_resize': scales2_resize,
-        'scales2_crop': scales2_crop
-    }
-    # print("data_aug_params:", json.dumps(datadict_for_print, indent=2))
-
-    if image_set == 'train':
-        if fix_size:
-            return T.Compose([
-                T.RandomHorizontalFlip(),
-                T.RandomResize([(max_size, max(scales))]),
-                # T.RandomResize([(512, 512)]),
-                normalize,
-            ])
-
-        if strong_aug:
-            import datasets.sltransform as SLT
-            
-            return T.Compose([
-                T.RandomHorizontalFlip(),
-                T.RandomSelect(
-                    T.RandomResize(scales, max_size=max_size),
-                    T.Compose([
-                        T.RandomResize(scales2_resize),
-                        T.RandomSizeCrop(*scales2_crop),
-                        T.RandomResize(scales, max_size=max_size),
-                    ])
-                ),
-                SLT.RandomSelectMulti([
-                    SLT.RandomCrop(),
-                    SLT.LightingNoise(),
-                    SLT.AdjustBrightness(2),
-                    SLT.AdjustContrast(2),
-                ]),
-                normalize,
-            ])
-        
-        return T.Compose([
-            T.RandomHorizontalFlip(),
-            T.RandomSelect(
-                T.RandomResize(scales, max_size=max_size),
-                T.Compose([
-                    T.RandomResize(scales2_resize),
-                    T.RandomSizeCrop(*scales2_crop),
-                    T.RandomResize(scales, max_size=max_size),
-                ])
-            ),
-            normalize,
-        ])
-
-    if image_set in ['val', 'eval_debug', 'train_reg', 'test']:
-
-        if os.environ.get("GFLOPS_DEBUG_SHILONG", False) == 'INFO':
-            print("Under debug mode for flops calculation only!!!!!!!!!!!!!!!!")
-            return T.Compose([
-                T.ResizeDebug((1280, 800)),
-                normalize,
-            ])   
-
-        return T.Compose([
-            T.RandomResize([max(scales)], max_size=max_size),
-            normalize,
-        ])
-
-
-
-    raise ValueError(f'unknown {image_set}')
-
-
-def get_aux_target_hacks_list(image_set, args):
-    if args.modelname in ['q2bs_mask', 'q2bs']:
-        aux_target_hacks_list = [
-            label2compat(), 
-            label_compat2onehot(), 
-            RandomSelectBoxes(num_class=args.num_classes)
-        ]
-        if args.masked_data and image_set == 'train':
-            # aux_target_hacks_list.append()
-            aux_target_hacks_list.append(MaskCrop())
-    elif args.modelname in ['q2bm_v2', 'q2bs_ce', 'q2op', 'q2ofocal', 'q2opclip', 'q2ocqonly']:
-        aux_target_hacks_list = [
-            label2compat(),
-            label_compat2onehot(),
-            box_label_catter(),
-            RandomSelectBoxlabels(num_classes=args.num_classes,
-                                    prob_first_item=args.prob_first_item,
-                                    prob_random_item=args.prob_random_item,
-                                    prob_last_item=args.prob_last_item,
-                                    prob_stop_sign=args.prob_stop_sign,
-                                    ),
-            BboxPertuber(max_ratio=0.02, generate_samples=1000),
-        ]
-    elif args.modelname in ['q2omask', 'q2osa']:
-        if args.coco_aug:
-            aux_target_hacks_list = [
-                label2compat(),
-                label_compat2onehot(),
-                box_label_catter(),
-                RandomSelectBoxlabels(num_classes=args.num_classes,
-                                        prob_first_item=args.prob_first_item,
-                                        prob_random_item=args.prob_random_item,
-                                        prob_last_item=args.prob_last_item,
-                                        prob_stop_sign=args.prob_stop_sign,
-                                        ),
-                RandomDrop(p=0.2),
-                BboxPertuber(max_ratio=0.02, generate_samples=1000),
-                RandomCutout(factor=0.5)
-            ]
-        else:
-            aux_target_hacks_list = [
-                label2compat(),
-                label_compat2onehot(),
-                box_label_catter(),
-                RandomSelectBoxlabels(num_classes=args.num_classes,
-                                        prob_first_item=args.prob_first_item,
-                                        prob_random_item=args.prob_random_item,
-                                        prob_last_item=args.prob_last_item,
-                                        prob_stop_sign=args.prob_stop_sign,
-                                        ),
-                BboxPertuber(max_ratio=0.02, generate_samples=1000),
-            ]
-    else:
-        aux_target_hacks_list = None
-
-    return aux_target_hacks_list
-
-
-def build(image_set, args, datasetinfo):
-    img_folder = datasetinfo["root"]
-    ann_file = datasetinfo["anno"]
-
-    # copy to local path
-    if os.environ.get('DATA_COPY_SHILONG') == 'INFO':
-        preparing_dataset(dict(img_folder=img_folder, ann_file=ann_file), image_set, args)
-
-    try:
-        strong_aug = args.strong_aug
-    except:
-        strong_aug = False
-    print(img_folder, ann_file)
-    dataset = CocoDetection(img_folder, ann_file, 
-            transforms=make_coco_transforms(image_set, fix_size=args.fix_size, strong_aug=strong_aug, args=args), 
-            return_masks=args.masks,
-            aux_target_hacks=None,
-        )
-    return dataset
-
-
-if __name__ == "__main__":
-    # Objects365 Val example
-    dataset_o365 = CocoDetection(
-            '/path/Objects365/train/',
-            "/path/Objects365/slannos/anno_preprocess_train_v2.json",
-            transforms=None,
-            return_masks=False,
-        )
-    print('len(dataset_o365):', len(dataset_o365))

groundingdino/datasets/.ipynb_checkpoints/dataset-checkpoint.py

@@ -1,44 +0,0 @@
-from __future__ import print_function
-
-import torch
-import torchvision.datasets as datasets
-from torch.utils.data import Dataset
-from PIL import Image
-from .tsv_io import TSVFile
-import numpy as np
-import base64
-import io
-
-
-class TSVDataset(Dataset):
-    """ TSV dataset for ImageNet 1K training
-    """    
-    def __init__(self, tsv_file, transform=None, target_transform=None):
-        self.tsv = TSVFile(tsv_file)
-        self.transform = transform
-        self.target_transform = target_transform
-
-    def __getitem__(self, index):
-        """
-        Args:
-            index (int): Index
-        Returns:
-            tuple: (image, target) where target is class_index of the target class.
-        """
-        row = self.tsv.seek(index)
-        image_data = base64.b64decode(row[-1])
-        image = Image.open(io.BytesIO(image_data))
-        image = image.convert('RGB')
-        target = int(row[1])
-
-        if self.transform is not None:
-            img = self.transform(image)
-        else:
-            img = image
-        if self.target_transform is not None:
-            target = self.target_transform(target)
-
-        return img, target
-
-    def __len__(self):
-        return self.tsv.num_rows()

groundingdino/datasets/.ipynb_checkpoints/odvg-checkpoint.py

@@ -1,258 +0,0 @@
-from torchvision.datasets.vision import VisionDataset
-import os.path
-from typing import Callable, Optional
-import json
-from PIL import Image
-import torch
-import random
-import os, sys
-sys.path.append(os.path.dirname(sys.path[0]))
-
-import datasets.transforms as T
-
-class ODVGDataset(VisionDataset):
-    """
-    Args:
-        root (string): Root directory where images are downloaded to.
-        anno (string): Path to json annotation file.
-        label_map_anno (string):  Path to json label mapping file. Only for Object Detection
-        transform (callable, optional): A function/transform that  takes in an PIL image
-            and returns a transformed version. E.g, ``transforms.PILToTensor``
-        target_transform (callable, optional): A function/transform that takes in the
-            target and transforms it.
-        transforms (callable, optional): A function/transform that takes input sample and its target as entry
-            and returns a transformed version.
-    """
-
-    def __init__(
-        self,
-        root: str,
-        anno: str,
-        label_map_anno: str = None,
-        max_labels: int = 80,
-        transform: Optional[Callable] = None,
-        target_transform: Optional[Callable] = None,
-        transforms: Optional[Callable] = None,
-    ) -> None:
-        super().__init__(root, transforms, transform, target_transform)
-        self.root = root
-        self.dataset_mode = "OD" if label_map_anno else "VG"
-        self.max_labels = max_labels
-        if self.dataset_mode == "OD":
-            self.load_label_map(label_map_anno)
-        self._load_metas(anno)
-        self.get_dataset_info()
-
-    def load_label_map(self, label_map_anno):
-        with open(label_map_anno, 'r') as file:
-            self.label_map = json.load(file)
-        self.label_index = set(self.label_map.keys())
-
-    def _load_metas(self, anno):
-      with open(anno, 'r') as f:
-          self.metas = json.load(f)
-
-
-    def get_dataset_info(self):
-        print(f"  == total images: {len(self)}")
-        if self.dataset_mode == "OD":
-            print(f"  == total labels: {len(self.label_map)}")
-
-    def __getitem__(self, index: int):
-        meta = self.metas[index]
-        rel_path = meta["filename"]
-        abs_path = os.path.join(self.root, rel_path)
-        if not os.path.exists(abs_path):
-            raise FileNotFoundError(f"{abs_path} not found.")
-        image = Image.open(abs_path).convert('RGB')
-        w, h = image.size
-        if self.dataset_mode == "OD":
-            anno = meta["detection"]
-            instances = [obj for obj in anno["instances"]]
-            boxes = [obj["bbox"] for obj in instances]
-            # generate vg_labels
-            # pos bbox labels
-            ori_classes = [str(obj["label"]) for obj in instances]
-            pos_labels = set(ori_classes)
-            # neg bbox labels 
-            neg_labels = self.label_index.difference(pos_labels)
-             
-            vg_labels = list(pos_labels)
-            num_to_add = min(len(neg_labels), self.max_labels-len(pos_labels))
-            if num_to_add > 0:
-                vg_labels.extend(random.sample(neg_labels, num_to_add))
-            
-            # shuffle
-            for i in range(len(vg_labels)-1, 0, -1):
-                j = random.randint(0, i)
-                vg_labels[i], vg_labels[j] = vg_labels[j], vg_labels[i]
-
-            caption_list = [self.label_map[lb] for lb in vg_labels]
-            caption_dict = {item:index for index, item in enumerate(caption_list)}
-
-            caption = ' . '.join(caption_list) + ' .'
-            classes = [caption_dict[self.label_map[str(obj["label"])]] for obj in instances]
-            boxes = torch.as_tensor(boxes, dtype=torch.float32).reshape(-1, 4)
-            classes = torch.tensor(classes, dtype=torch.int64)
-        elif self.dataset_mode == "VG":
-            anno = meta["Grounding"]
-            instances = [obj for obj in anno["regions"]] 
-            boxes = [obj["bbox"] for obj in instances]
-            caption_list = [obj["phrase"] for obj in instances]
-            c = list(zip(boxes, caption_list))
-            random.shuffle(c)
-            boxes[:], caption_list[:] = zip(*c)
-            uni_caption_list  = list(set(caption_list))
-            label_map = {}
-            for idx in range(len(uni_caption_list)):
-                label_map[uni_caption_list[idx]] = idx
-            classes = [label_map[cap] for cap in caption_list]
-            caption = ' . '.join(uni_caption_list) + ' .'
-            boxes = torch.as_tensor(boxes, dtype=torch.float32).reshape(-1, 4)
-            classes = torch.tensor(classes, dtype=torch.int64)
-            caption_list = uni_caption_list
-            # print("caption_list" , caption_list)
-            # print("caption" , caption)
-            # print("boxes" , boxes)
-        target = {}
-        target["image_id"] = rel_path.strip(".jpg")
-        target["size"] = torch.as_tensor([int(h), int(w)])
-        target["cap_list"] = caption_list
-        target["caption"] = caption
-        target["boxes"] = boxes
-        target["labels"] = classes
-        # print(" image_id " , target["image_id"])
-        # size, cap_list, caption, bboxes, labels
-
-        if self.transforms is not None:
-            image, target = self.transforms(image, target)
-
-        return image, target
-    
-
-    def __len__(self) -> int:
-        return len(self.metas)
-
-
-def make_coco_transforms(image_set, fix_size=False, strong_aug=False, args=None):
-
-    normalize = T.Compose([
-        T.ToTensor(),
-        T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
-    ])
-
-    # config the params for data aug
-    scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800]
-    max_size = 1333
-    scales2_resize = [400, 500, 600]
-    scales2_crop = [384, 600]
-    
-    # update args from config files
-    scales = getattr(args, 'data_aug_scales', scales)
-    max_size = getattr(args, 'data_aug_max_size', max_size)
-    scales2_resize = getattr(args, 'data_aug_scales2_resize', scales2_resize)
-    scales2_crop = getattr(args, 'data_aug_scales2_crop', scales2_crop)
-
-    # resize them
-    data_aug_scale_overlap = getattr(args, 'data_aug_scale_overlap', None)
-    if data_aug_scale_overlap is not None and data_aug_scale_overlap > 0:
-        data_aug_scale_overlap = float(data_aug_scale_overlap)
-        scales = [int(i*data_aug_scale_overlap) for i in scales]
-        max_size = int(max_size*data_aug_scale_overlap)
-        scales2_resize = [int(i*data_aug_scale_overlap) for i in scales2_resize]
-        scales2_crop = [int(i*data_aug_scale_overlap) for i in scales2_crop]
-
-    # datadict_for_print = {
-    #     'scales': scales,
-    #     'max_size': max_size,
-    #     'scales2_resize': scales2_resize,
-    #     'scales2_crop': scales2_crop
-    # }
-    # print("data_aug_params:", json.dumps(datadict_for_print, indent=2))
-
-    if image_set == 'train':
-        if fix_size:
-            return T.Compose([
-                T.RandomHorizontalFlip(),
-                T.RandomResize([(max_size, max(scales))]),
-                normalize,
-            ])
-
-        if strong_aug:
-            import datasets.sltransform as SLT
-            
-            return T.Compose([
-                T.RandomHorizontalFlip(),
-                T.RandomSelect(
-                    T.RandomResize(scales, max_size=max_size),
-                    T.Compose([
-                        T.RandomResize(scales2_resize),
-                        T.RandomSizeCrop(*scales2_crop),
-                        T.RandomResize(scales, max_size=max_size),
-                    ])
-                ),
-                SLT.RandomSelectMulti([
-                    SLT.RandomCrop(),
-                    SLT.LightingNoise(),
-                    SLT.AdjustBrightness(2),
-                    SLT.AdjustContrast(2),
-                ]),
-                normalize,
-            ])
-        
-        return T.Compose([
-            T.RandomHorizontalFlip(),
-            T.RandomSelect(
-                T.RandomResize(scales, max_size=max_size),
-                T.Compose([
-                    T.RandomResize(scales2_resize),
-                    T.RandomSizeCrop(*scales2_crop),
-                    T.RandomResize(scales, max_size=max_size),
-                ])
-            ),
-            normalize,
-        ])
-
-    if image_set in ['val', 'eval_debug', 'train_reg', 'test']:
-
-        if os.environ.get("GFLOPS_DEBUG_SHILONG", False) == 'INFO':
-            print("Under debug mode for flops calculation only!!!!!!!!!!!!!!!!")
-            return T.Compose([
-                T.ResizeDebug((1280, 800)),
-                normalize,
-            ])   
-
-        return T.Compose([
-            T.RandomResize([max(scales)], max_size=max_size),
-            normalize,
-        ])
-
-    raise ValueError(f'unknown {image_set}')
-
-def build_odvg(image_set, args, datasetinfo):
-    img_folder = datasetinfo["root"]
-    ann_file = datasetinfo["anno"]
-    label_map = datasetinfo["label_map"] if "label_map" in datasetinfo else None
-    try:
-        strong_aug = args.strong_aug
-    except:
-        strong_aug = False # False originally
-    print(img_folder, ann_file, label_map)
-    dataset = ODVGDataset(img_folder, ann_file, label_map, max_labels=args.max_labels,
-            transforms=make_coco_transforms(image_set, fix_size=args.fix_size, strong_aug=strong_aug, args=args), 
-    )
-    return dataset
-
-
-if __name__=="__main__":
-    dataset_vg = ODVGDataset("path/GRIT-20M/data/","path/GRIT-20M/anno/grit_odvg_10k.jsonl",)
-    print(len(dataset_vg))
-    data = dataset_vg[random.randint(0, 100)] 
-    print(data)
-    dataset_od = ODVGDataset("pathl/V3Det/",
-        "path/V3Det/annotations/v3det_2023_v1_all_odvg.jsonl",
-        "path/V3Det/annotations/v3det_label_map.json",
-    )
-    print(len(dataset_od))
-    data = dataset_od[random.randint(0, 100)] 
-    print(data)

groundingdino/datasets/.ipynb_checkpoints/transforms-checkpoint.py

@@ -1,285 +0,0 @@
-# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
-"""
-Transforms and data augmentation for both image + bbox.
-"""
-import random
-
-import PIL
-import torch
-import torchvision.transforms as T
-import torchvision.transforms.functional as F
-
-from util.box_ops import box_xyxy_to_cxcywh
-from 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"]
-
-    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:
-            target[field] = target[field][keep]
-
-    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):
-        self.min_size = min_size
-        self.max_size = max_size
-
-    def __call__(self, img: PIL.Image.Image, target: dict):
-        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])
-        return crop(img, target, region)
-
-
-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.))
-        crop_left = int(round((image_width - crop_width) / 2.))
-        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

groundingdino/datasets/__init__.py

@@ -1,23 +0,0 @@
-# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
-import torch.utils.data
-import torchvision
-from .coco import build as build_coco
-
-
-def get_coco_api_from_dataset(dataset):
-    for _ in range(10):
-        # if isinstance(dataset, torchvision.datasets.CocoDetection):
-        #     break
-        if isinstance(dataset, torch.utils.data.Subset):
-            dataset = dataset.dataset
-    if isinstance(dataset, torchvision.datasets.CocoDetection):
-        return dataset.coco
-
-
-def build_dataset(image_set, args, datasetinfo):
-    if datasetinfo["dataset_mode"] == 'coco':
-        return build_coco(image_set, args, datasetinfo)
-    if datasetinfo["dataset_mode"] == 'odvg':
-        from .odvg import build_odvg
-        return build_odvg(image_set, args, datasetinfo)
-    raise ValueError(f'dataset {args.dataset_file} not supported')

groundingdino/datasets/coco.py

@@ -1,649 +0,0 @@
-# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
-"""
-COCO dataset which returns image_id for evaluation.
-
-Mostly copy-paste from https://github.com/pytorch/vision/blob/13b35ff/references/detection/coco_utils.py
-"""
-if __name__=="__main__":
-    # for debug only
-    import os, sys
-    sys.path.append(os.path.dirname(sys.path[0]))
-from torchvision.datasets.vision import VisionDataset
-
-import json
-from pathlib import Path
-import random
-import os
-from typing import Any, Callable, List, Optional, Tuple
-
-from PIL import Image
-
-import torch
-import torch.utils.data
-import torchvision
-from pycocotools import mask as coco_mask
-
-from datasets.data_util import preparing_dataset
-import datasets.transforms as T
-from util.box_ops import box_cxcywh_to_xyxy, box_iou
-
-__all__ = ['build']
-
-
-class label2compat():
-    def __init__(self) -> None:
-        self.category_map_str = {"1": 1, "2": 2, "3": 3, "4": 4, "5": 5, "6": 6, "7": 7, "8": 8, "9": 9, "10": 10, "11": 11, "13": 12, "14": 13, "15": 14, "16": 15, "17": 16, "18": 17, "19": 18, "20": 19, "21": 20, "22": 21, "23": 22, "24": 23, "25": 24, "27": 25, "28": 26, "31": 27, "32": 28, "33": 29, "34": 30, "35": 31, "36": 32, "37": 33, "38": 34, "39": 35, "40": 36, "41": 37, "42": 38, "43": 39, "44": 40, "46": 41, "47": 42, "48": 43, "49": 44, "50": 45, "51": 46, "52": 47, "53": 48, "54": 49, "55": 50, "56": 51, "57": 52, "58": 53, "59": 54, "60": 55, "61": 56, "62": 57, "63": 58, "64": 59, "65": 60, "67": 61, "70": 62, "72": 63, "73": 64, "74": 65, "75": 66, "76": 67, "77": 68, "78": 69, "79": 70, "80": 71, "81": 72, "82": 73, "84": 74, "85": 75, "86": 76, "87": 77, "88": 78, "89": 79, "90": 80}
-        self.category_map = {int(k):v for k,v in self.category_map_str.items()}
-
-    def __call__(self, target, img=None):
-        labels = target['labels']
-        res = torch.zeros(labels.shape, dtype=labels.dtype)
-        for idx, item in enumerate(labels):
-            res[idx] = self.category_map[item.item()] - 1
-        target['label_compat'] = res
-        if img is not None:
-            return target, img
-        else:
-            return target
-
-
-class label_compat2onehot():
-    def __init__(self, num_class=80, num_output_objs=1):
-        self.num_class = num_class
-        self.num_output_objs = num_output_objs
-        if num_output_objs != 1:
-            raise DeprecationWarning("num_output_objs!=1, which is only used for comparison")
-
-    def __call__(self, target, img=None):
-        labels = target['label_compat']
-        place_dict = {k:0 for k in range(self.num_class)}
-        if self.num_output_objs == 1:
-            res = torch.zeros(self.num_class)
-            for i in labels:
-                itm = i.item()
-                res[itm] = 1.0
-        else:
-            # compat with baseline
-            res = torch.zeros(self.num_class, self.num_output_objs)
-            for i in labels:
-                itm = i.item()
-                res[itm][place_dict[itm]] = 1.0
-                place_dict[itm] += 1
-        target['label_compat_onehot'] = res
-        if img is not None:
-            return target, img
-        else:
-            return target
-
-
-class box_label_catter():
-    def __init__(self):
-        pass
-
-    def __call__(self, target, img=None):
-        labels = target['label_compat']
-        boxes = target['boxes']
-        box_label = torch.cat((boxes, labels.unsqueeze(-1)), 1)
-        target['box_label'] = box_label
-        if img is not None:
-            return target, img
-        else:
-            return target
-
-
-class RandomSelectBoxlabels():
-    def __init__(self, num_classes, leave_one_out=False, blank_prob=0.8,
-                    prob_first_item = 0.0,
-                    prob_random_item = 0.0,
-                    prob_last_item = 0.8,
-                    prob_stop_sign = 0.2
-                ) -> None:
-        self.num_classes = num_classes
-        self.leave_one_out = leave_one_out
-        self.blank_prob = blank_prob
-
-        self.set_state(prob_first_item, prob_random_item, prob_last_item, prob_stop_sign)
-        
-
-    def get_state(self):
-        return [self.prob_first_item, self.prob_random_item, self.prob_last_item, self.prob_stop_sign]
-
-    def set_state(self, prob_first_item, prob_random_item, prob_last_item, prob_stop_sign):
-        sum_prob = prob_first_item + prob_random_item + prob_last_item + prob_stop_sign
-        assert sum_prob - 1 < 1e-6, \
-            f"Sum up all prob = {sum_prob}. prob_first_item:{prob_first_item}" \
-            + f"prob_random_item:{prob_random_item}, prob_last_item:{prob_last_item}" \
-            + f"prob_stop_sign:{prob_stop_sign}"
-
-        self.prob_first_item = prob_first_item
-        self.prob_random_item = prob_random_item
-        self.prob_last_item = prob_last_item
-        self.prob_stop_sign = prob_stop_sign
-        
-
-    def sample_for_pred_first_item(self, box_label: torch.FloatTensor):
-        box_label_known = torch.Tensor(0,5)
-        box_label_unknown = box_label
-        return box_label_known, box_label_unknown
-
-    def sample_for_pred_random_item(self, box_label: torch.FloatTensor):
-        n_select = int(random.random() * box_label.shape[0])
-        box_label = box_label[torch.randperm(box_label.shape[0])]
-        box_label_known = box_label[:n_select]
-        box_label_unknown = box_label[n_select:]
-        return box_label_known, box_label_unknown
-
-    def sample_for_pred_last_item(self, box_label: torch.FloatTensor):
-        box_label_perm = box_label[torch.randperm(box_label.shape[0])]
-        known_label_list = []
-        box_label_known = []
-        box_label_unknown = []
-        for item in box_label_perm:
-            label_i = item[4].item()
-            if label_i in known_label_list:
-                box_label_known.append(item)
-            else:
-                # first item
-                box_label_unknown.append(item)
-                known_label_list.append(label_i)
-        box_label_known = torch.stack(box_label_known) if len(box_label_known) > 0 else torch.Tensor(0,5)
-        box_label_unknown = torch.stack(box_label_unknown) if len(box_label_unknown) > 0 else torch.Tensor(0,5)
-        return box_label_known, box_label_unknown
-
-    def sample_for_pred_stop_sign(self, box_label: torch.FloatTensor):
-        box_label_unknown = torch.Tensor(0,5)
-        box_label_known = box_label
-        return box_label_known, box_label_unknown
-
-    def __call__(self, target, img=None):
-        box_label = target['box_label'] # K, 5
-
-        dice_number = random.random()
-
-        if dice_number < self.prob_first_item:
-            box_label_known, box_label_unknown = self.sample_for_pred_first_item(box_label)
-        elif dice_number < self.prob_first_item + self.prob_random_item:
-            box_label_known, box_label_unknown = self.sample_for_pred_random_item(box_label)
-        elif dice_number < self.prob_first_item + self.prob_random_item + self.prob_last_item:
-            box_label_known, box_label_unknown = self.sample_for_pred_last_item(box_label)
-        else:
-            box_label_known, box_label_unknown = self.sample_for_pred_stop_sign(box_label)
-
-        target['label_onehot_known'] = label2onehot(box_label_known[:,-1], self.num_classes)
-        target['label_onehot_unknown'] = label2onehot(box_label_unknown[:, -1], self.num_classes)
-        target['box_label_known'] = box_label_known
-        target['box_label_unknown'] = box_label_unknown
-
-        return target, img
-
-
-class RandomDrop():
-    def __init__(self, p=0.2) -> None:
-        self.p = p
-
-    def __call__(self, target, img=None):
-        known_box = target['box_label_known']
-        num_known_box = known_box.size(0)
-        idxs = torch.rand(num_known_box)
-        # indices = torch.randperm(num_known_box)[:int((1-self).p*num_known_box + 0.5 + random.random())]
-        target['box_label_known'] = known_box[idxs > self.p]
-        return target, img
-
-
-class BboxPertuber():
-    def __init__(self, max_ratio = 0.02, generate_samples = 1000) -> None:
-        self.max_ratio = max_ratio
-        self.generate_samples = generate_samples
-        self.samples = self.generate_pertube_samples()
-        self.idx = 0
-
-    def generate_pertube_samples(self):
-        import torch
-        samples = (torch.rand(self.generate_samples, 5) - 0.5) * 2 * self.max_ratio
-        return samples
-
-    def __call__(self, target, img):
-        known_box = target['box_label_known'] # Tensor(K,5), K known bbox
-        K = known_box.shape[0]
-        known_box_pertube = torch.zeros(K, 6) # 4:bbox, 1:prob, 1:label
-        if K == 0:
-            pass
-        else:
-            if self.idx + K > self.generate_samples:
-                self.idx = 0
-            delta = self.samples[self.idx: self.idx + K, :]
-            known_box_pertube[:, :4] = known_box[:, :4] + delta[:, :4]
-            iou = (torch.diag(box_iou(box_cxcywh_to_xyxy(known_box[:, :4]), box_cxcywh_to_xyxy(known_box_pertube[:, :4]))[0])) * (1 + delta[:, -1])
-            known_box_pertube[:, 4].copy_(iou)
-            known_box_pertube[:, -1].copy_(known_box[:, -1])
-
-        target['box_label_known_pertube'] = known_box_pertube
-        return target, img
-
-
-class RandomCutout():
-    def __init__(self, factor=0.5) -> None:
-        self.factor = factor
-
-    def __call__(self, target, img=None):
-        unknown_box = target['box_label_unknown']           # Ku, 5
-        known_box = target['box_label_known_pertube']       # Kk, 6
-        Ku = unknown_box.size(0)
-
-        known_box_add = torch.zeros(Ku, 6) # Ku, 6
-        known_box_add[:, :5] = unknown_box
-        known_box_add[:, 5].uniform_(0.5, 1) 
-        
-
-        known_box_add[:, :2] += known_box_add[:, 2:4] * (torch.rand(Ku, 2) - 0.5) / 2
-        known_box_add[:, 2:4] /= 2
-
-        target['box_label_known_pertube'] = torch.cat((known_box, known_box_add))
-        return target, img
-
-
-class RandomSelectBoxes():
-    def __init__(self, num_class=80) -> None:
-        Warning("This is such a slow function and will be deprecated soon!!!")
-        self.num_class = num_class
-
-    def __call__(self, target, img=None):
-        boxes = target['boxes']
-        labels = target['label_compat']
-
-        # transform to list of tensors
-        boxs_list = [[] for i in range(self.num_class)]
-        for idx, item in enumerate(boxes):
-            label = labels[idx].item()
-            boxs_list[label].append(item)
-        boxs_list_tensor = [torch.stack(i) if len(i) > 0 else torch.Tensor(0,4) for i in boxs_list]
-
-        # random selection
-        box_known = []
-        box_unknown = []
-        for idx, item in enumerate(boxs_list_tensor):
-            ncnt = item.shape[0]
-            nselect = int(random.random() * ncnt) # close in both sides, much faster than random.randint
-
-            item = item[torch.randperm(ncnt)]
-            # random.shuffle(item)
-            box_known.append(item[:nselect])
-            box_unknown.append(item[nselect:])
-
-        # box_known_tensor = [torch.stack(i) if len(i) > 0 else torch.Tensor(0,4) for i in box_known]
-        # box_unknown_tensor = [torch.stack(i) if len(i) > 0 else torch.Tensor(0,4) for i in box_unknown]
-        # print('box_unknown_tensor:', box_unknown_tensor)
-        target['known_box'] = box_known
-        target['unknown_box'] = box_unknown
-        return target, img
-
-
-def label2onehot(label, num_classes):
-    """
-    label: Tensor(K)
-    """
-    res = torch.zeros(num_classes)
-    for i in label:
-        itm = int(i.item())
-        res[itm] = 1.0
-    return res
-
-
-class MaskCrop():
-    def __init__(self) -> None:
-        pass
-
-    def __call__(self, target, img):
-        known_box = target['known_box']
-        h,w = img.shape[1:] # h,w
-        # imgsize = target['orig_size'] # h,w
-
-        scale = torch.Tensor([w, h, w, h])
-
-        # _cnt = 0
-        for boxes in known_box:
-            if boxes.shape[0] == 0:
-                continue
-            box_xyxy = box_cxcywh_to_xyxy(boxes) * scale
-            for box in box_xyxy:
-                x1, y1, x2, y2 = [int(i) for i in box.tolist()]
-                img[:, y1:y2, x1:x2] = 0
-                # _cnt += 1
-        # print("_cnt:", _cnt)
-        return target, img
-
-
-dataset_hook_register = {
-    'label2compat': label2compat,
-    'label_compat2onehot': label_compat2onehot,
-    'box_label_catter': box_label_catter,
-    'RandomSelectBoxlabels': RandomSelectBoxlabels,
-    'RandomSelectBoxes': RandomSelectBoxes,
-    'MaskCrop': MaskCrop,
-    'BboxPertuber': BboxPertuber,
-}
-
-
-class CocoDetection(torchvision.datasets.CocoDetection):
-    def __init__(self, img_folder, ann_file, transforms, return_masks, aux_target_hacks=None):
-        super(CocoDetection, self).__init__(img_folder, ann_file)
-        self._transforms = transforms
-        self.prepare = ConvertCocoPolysToMask(return_masks)
-        self.aux_target_hacks = aux_target_hacks
-
-    def change_hack_attr(self, hackclassname, attrkv_dict):
-        target_class = dataset_hook_register[hackclassname]
-        for item in self.aux_target_hacks:
-            if isinstance(item, target_class):
-                for k,v in attrkv_dict.items():
-                    setattr(item, k, v)
-
-    def get_hack(self, hackclassname):
-        target_class = dataset_hook_register[hackclassname]
-        for item in self.aux_target_hacks:
-            if isinstance(item, target_class):
-                return item
-            
-    def _load_image(self, id: int) -> Image.Image:
-        path = self.coco.loadImgs(id)[0]["file_name"]
-        abs_path = os.path.join(self.root, path)
-        return Image.open(abs_path).convert("RGB")
-
-    def __getitem__(self, idx):
-        """
-        Output:
-            - target: dict of multiple items
-                - boxes: Tensor[num_box, 4]. \
-                    Init type: x0,y0,x1,y1. unnormalized data.
-                    Final type: cx,cy,w,h. normalized data. 
-        """
-        try:
-            img, target = super(CocoDetection, self).__getitem__(idx)
-        except:
-            print("Error idx: {}".format(idx))
-            idx += 1
-            img, target = super(CocoDetection, self).__getitem__(idx)
-        image_id = self.ids[idx]
-        target = {'image_id': image_id, 'annotations': target}
-        img, target = self.prepare(img, target)
-        
-        if self._transforms is not None:
-            img, target = self._transforms(img, target)
-
-        # convert to needed format
-        if self.aux_target_hacks is not None:
-            for hack_runner in self.aux_target_hacks:
-                target, img = hack_runner(target, img=img)
-
-        return img, target
-
-
-def convert_coco_poly_to_mask(segmentations, height, width):
-    masks = []
-    for polygons in segmentations:
-        rles = coco_mask.frPyObjects(polygons, height, width)
-        mask = coco_mask.decode(rles)
-        if len(mask.shape) < 3:
-            mask = mask[..., None]
-        mask = torch.as_tensor(mask, dtype=torch.uint8)
-        mask = mask.any(dim=2)
-        masks.append(mask)
-    if masks:
-        masks = torch.stack(masks, dim=0)
-    else:
-        masks = torch.zeros((0, height, width), dtype=torch.uint8)
-    return masks
-
-
-class ConvertCocoPolysToMask(object):
-    def __init__(self, return_masks=False):
-        self.return_masks = return_masks
-
-    def __call__(self, image, target):
-        w, h = image.size
-
-        image_id = target["image_id"]
-        image_id = torch.tensor([image_id])
-
-        anno = target["annotations"]
-
-        anno = [obj for obj in anno if 'iscrowd' not in obj or obj['iscrowd'] == 0]
-
-        boxes = [obj["bbox"] for obj in anno]
-        # guard against no boxes via resizing
-        boxes = torch.as_tensor(boxes, dtype=torch.float32).reshape(-1, 4)
-        boxes[:, 2:] += boxes[:, :2]
-        boxes[:, 0::2].clamp_(min=0, max=w)
-        boxes[:, 1::2].clamp_(min=0, max=h)
-
-        classes = [obj["category_id"] for obj in anno]
-        classes = torch.tensor(classes, dtype=torch.int64)
-
-        if self.return_masks:
-            segmentations = [obj["segmentation"] for obj in anno]
-            masks = convert_coco_poly_to_mask(segmentations, h, w)
-
-        keypoints = None
-        if anno and "keypoints" in anno[0]:
-            keypoints = [obj["keypoints"] for obj in anno]
-            keypoints = torch.as_tensor(keypoints, dtype=torch.float32)
-            num_keypoints = keypoints.shape[0]
-            if num_keypoints:
-                keypoints = keypoints.view(num_keypoints, -1, 3)
-
-        keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0])
-        boxes = boxes[keep]
-        classes = classes[keep]
-        if self.return_masks:
-            masks = masks[keep]
-        if keypoints is not None:
-            keypoints = keypoints[keep]
-
-        target = {}
-        target["boxes"] = boxes
-        target["labels"] = classes
-        if self.return_masks:
-            target["masks"] = masks
-        target["image_id"] = image_id
-        if keypoints is not None:
-            target["keypoints"] = keypoints
-
-        # for conversion to coco api
-        area = torch.tensor([obj["area"] for obj in anno])
-        iscrowd = torch.tensor([obj["iscrowd"] if "iscrowd" in obj else 0 for obj in anno])
-        target["area"] = area[keep]
-        target["iscrowd"] = iscrowd[keep]
-
-        target["orig_size"] = torch.as_tensor([int(h), int(w)])
-        target["size"] = torch.as_tensor([int(h), int(w)])
-
-        return image, target
-
-
-def make_coco_transforms(image_set, fix_size=False, strong_aug=False, args=None):
-
-    normalize = T.Compose([
-        T.ToTensor(),
-        T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
-    ])
-
-    # config the params for data aug
-    scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800]
-    max_size = 1333
-    scales2_resize = [400, 500, 600]
-    scales2_crop = [384, 600]
-    
-    # update args from config files
-    scales = getattr(args, 'data_aug_scales', scales)
-    max_size = getattr(args, 'data_aug_max_size', max_size)
-    scales2_resize = getattr(args, 'data_aug_scales2_resize', scales2_resize)
-    scales2_crop = getattr(args, 'data_aug_scales2_crop', scales2_crop)
-
-    # resize them
-    data_aug_scale_overlap = getattr(args, 'data_aug_scale_overlap', None)
-    if data_aug_scale_overlap is not None and data_aug_scale_overlap > 0:
-        data_aug_scale_overlap = float(data_aug_scale_overlap)
-        scales = [int(i*data_aug_scale_overlap) for i in scales]
-        max_size = int(max_size*data_aug_scale_overlap)
-        scales2_resize = [int(i*data_aug_scale_overlap) for i in scales2_resize]
-        scales2_crop = [int(i*data_aug_scale_overlap) for i in scales2_crop]
-
-    datadict_for_print = {
-        'scales': scales,
-        'max_size': max_size,
-        'scales2_resize': scales2_resize,
-        'scales2_crop': scales2_crop
-    }
-    # print("data_aug_params:", json.dumps(datadict_for_print, indent=2))
-
-    if image_set == 'train':
-        if fix_size:
-            return T.Compose([
-                T.RandomHorizontalFlip(),
-                T.RandomResize([(max_size, max(scales))]),
-                # T.RandomResize([(512, 512)]),
-                normalize,
-            ])
-
-        if strong_aug:
-            import datasets.sltransform as SLT
-            
-            return T.Compose([
-                T.RandomHorizontalFlip(),
-                T.RandomSelect(
-                    T.RandomResize(scales, max_size=max_size),
-                    T.Compose([
-                        T.RandomResize(scales2_resize),
-                        T.RandomSizeCrop(*scales2_crop),
-                        T.RandomResize(scales, max_size=max_size),
-                    ])
-                ),
-                SLT.RandomSelectMulti([
-                    SLT.RandomCrop(),
-                    SLT.LightingNoise(),
-                    SLT.AdjustBrightness(2),
-                    SLT.AdjustContrast(2),
-                ]),
-                normalize,
-            ])
-        
-        return T.Compose([
-            T.RandomHorizontalFlip(),
-            T.RandomSelect(
-                T.RandomResize(scales, max_size=max_size),
-                T.Compose([
-                    T.RandomResize(scales2_resize),
-                    T.RandomSizeCrop(*scales2_crop),
-                    T.RandomResize(scales, max_size=max_size),
-                ])
-            ),
-            normalize,
-        ])
-
-    if image_set in ['val', 'eval_debug', 'train_reg', 'test']:
-
-        if os.environ.get("GFLOPS_DEBUG_SHILONG", False) == 'INFO':
-            print("Under debug mode for flops calculation only!!!!!!!!!!!!!!!!")
-            return T.Compose([
-                T.ResizeDebug((1280, 800)),
-                normalize,
-            ])   
-
-        return T.Compose([
-            T.RandomResize([max(scales)], max_size=max_size),
-            normalize,
-        ])
-
-
-
-    raise ValueError(f'unknown {image_set}')
-
-
-def get_aux_target_hacks_list(image_set, args):
-    if args.modelname in ['q2bs_mask', 'q2bs']:
-        aux_target_hacks_list = [
-            label2compat(), 
-            label_compat2onehot(), 
-            RandomSelectBoxes(num_class=args.num_classes)
-        ]
-        if args.masked_data and image_set == 'train':
-            # aux_target_hacks_list.append()
-            aux_target_hacks_list.append(MaskCrop())
-    elif args.modelname in ['q2bm_v2', 'q2bs_ce', 'q2op', 'q2ofocal', 'q2opclip', 'q2ocqonly']:
-        aux_target_hacks_list = [
-            label2compat(),
-            label_compat2onehot(),
-            box_label_catter(),
-            RandomSelectBoxlabels(num_classes=args.num_classes,
-                                    prob_first_item=args.prob_first_item,
-                                    prob_random_item=args.prob_random_item,
-                                    prob_last_item=args.prob_last_item,
-                                    prob_stop_sign=args.prob_stop_sign,
-                                    ),
-            BboxPertuber(max_ratio=0.02, generate_samples=1000),
-        ]
-    elif args.modelname in ['q2omask', 'q2osa']:
-        if args.coco_aug:
-            aux_target_hacks_list = [
-                label2compat(),
-                label_compat2onehot(),
-                box_label_catter(),
-                RandomSelectBoxlabels(num_classes=args.num_classes,
-                                        prob_first_item=args.prob_first_item,
-                                        prob_random_item=args.prob_random_item,
-                                        prob_last_item=args.prob_last_item,
-                                        prob_stop_sign=args.prob_stop_sign,
-                                        ),
-                RandomDrop(p=0.2),
-                BboxPertuber(max_ratio=0.02, generate_samples=1000),
-                RandomCutout(factor=0.5)
-            ]
-        else:
-            aux_target_hacks_list = [
-                label2compat(),
-                label_compat2onehot(),
-                box_label_catter(),
-                RandomSelectBoxlabels(num_classes=args.num_classes,
-                                        prob_first_item=args.prob_first_item,
-                                        prob_random_item=args.prob_random_item,
-                                        prob_last_item=args.prob_last_item,
-                                        prob_stop_sign=args.prob_stop_sign,
-                                        ),
-                BboxPertuber(max_ratio=0.02, generate_samples=1000),
-            ]
-    else:
-        aux_target_hacks_list = None
-
-    return aux_target_hacks_list
-
-
-def build(image_set, args, datasetinfo):
-    img_folder = datasetinfo["root"]
-    ann_file = datasetinfo["anno"]
-
-    # copy to local path
-    if os.environ.get('DATA_COPY_SHILONG') == 'INFO':
-        preparing_dataset(dict(img_folder=img_folder, ann_file=ann_file), image_set, args)
-
-    try:
-        strong_aug = args.strong_aug
-    except:
-        strong_aug = False
-    print(img_folder, ann_file)
-    dataset = CocoDetection(img_folder, ann_file, 
-            transforms=make_coco_transforms(image_set, fix_size=args.fix_size, strong_aug=strong_aug, args=args), 
-            return_masks=args.masks,
-            aux_target_hacks=None,
-        )
-    return dataset
-
-
-if __name__ == "__main__":
-    # Objects365 Val example
-    dataset_o365 = CocoDetection(
-            '/path/Objects365/train/',
-            "/path/Objects365/slannos/anno_preprocess_train_v2.json",
-            transforms=None,
-            return_masks=False,
-        )
-    print('len(dataset_o365):', len(dataset_o365))

groundingdino/datasets/coco_eval.py

@@ -1,266 +0,0 @@
-# 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 os
-import contextlib
-import copy
-import numpy as np
-import torch
-
-from pycocotools.cocoeval import COCOeval
-from pycocotools.coco import COCO
-import pycocotools.mask as mask_util
-
-from util.misc import all_gather
-
-
-class CocoEvaluator(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()
-            if not isinstance(prediction["scores"], list):
-                scores = prediction["scores"].tolist()
-            else:
-                scores = prediction["scores"]
-            if not isinstance(prediction["labels"], list):
-                labels = prediction["labels"].tolist()
-            else:
-                labels = prediction["labels"]
-
-        
-            try:
-                coco_results.extend(
-                    [
-                        {
-                            "image_id": original_id,
-                            "category_id": labels[k],
-                            "bbox": box,
-                            "score": scores[k],
-                        }
-                        for k, box in enumerate(boxes)
-                    ]
-                )
-            except:
-                import ipdb; ipdb.set_trace()
-        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
-    '''
-    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))
-    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)
-
-    return p.imgIds, evalImgs
-
-#################################################################
-# end of straight copy from pycocotools, just removing the prints
-#################################################################

groundingdino/datasets/coco_panoptic.py

@@ -1,99 +0,0 @@
-# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
-import json
-from pathlib import Path
-
-import numpy as np
-import torch
-from PIL import Image
-
-from panopticapi.utils import rgb2id
-from util.box_ops import masks_to_boxes
-
-from .coco import make_coco_transforms
-
-
-class CocoPanoptic:
-    def __init__(self, img_folder, ann_folder, ann_file, transforms=None, return_masks=True):
-        with open(ann_file, 'r') as f:
-            self.coco = json.load(f)
-
-        # sort 'images' field so that they are aligned with 'annotations'
-        # i.e., in alphabetical order
-        self.coco['images'] = sorted(self.coco['images'], key=lambda x: x['id'])
-        # sanity check
-        if "annotations" in self.coco:
-            for img, ann in zip(self.coco['images'], self.coco['annotations']):
-                assert img['file_name'][:-4] == ann['file_name'][:-4]
-
-        self.img_folder = img_folder
-        self.ann_folder = ann_folder
-        self.ann_file = ann_file
-        self.transforms = transforms
-        self.return_masks = return_masks
-
-    def __getitem__(self, idx):
-        ann_info = self.coco['annotations'][idx] if "annotations" in self.coco else self.coco['images'][idx]
-        img_path = Path(self.img_folder) / ann_info['file_name'].replace('.png', '.jpg')
-        ann_path = Path(self.ann_folder) / ann_info['file_name']
-
-        img = Image.open(img_path).convert('RGB')
-        w, h = img.size
-        if "segments_info" in ann_info:
-            masks = np.asarray(Image.open(ann_path), dtype=np.uint32)
-            masks = rgb2id(masks)
-
-            ids = np.array([ann['id'] for ann in ann_info['segments_info']])
-            masks = masks == ids[:, None, None]
-
-            masks = torch.as_tensor(masks, dtype=torch.uint8)
-            labels = torch.tensor([ann['category_id'] for ann in ann_info['segments_info']], dtype=torch.int64)
-
-        target = {}
-        target['image_id'] = torch.tensor([ann_info['image_id'] if "image_id" in ann_info else ann_info["id"]])
-        if self.return_masks:
-            target['masks'] = masks
-        target['labels'] = labels
-
-        target["boxes"] = masks_to_boxes(masks)
-
-        target['size'] = torch.as_tensor([int(h), int(w)])
-        target['orig_size'] = torch.as_tensor([int(h), int(w)])
-        if "segments_info" in ann_info:
-            for name in ['iscrowd', 'area']:
-                target[name] = torch.tensor([ann[name] for ann in ann_info['segments_info']])
-
-        if self.transforms is not None:
-            img, target = self.transforms(img, target)
-
-        return img, target
-
-    def __len__(self):
-        return len(self.coco['images'])
-
-    def get_height_and_width(self, idx):
-        img_info = self.coco['images'][idx]
-        height = img_info['height']
-        width = img_info['width']
-        return height, width
-
-
-def build(image_set, args):
-    img_folder_root = Path(args.coco_path)
-    ann_folder_root = Path(args.coco_panoptic_path)
-    assert img_folder_root.exists(), f'provided COCO path {img_folder_root} does not exist'
-    assert ann_folder_root.exists(), f'provided COCO path {ann_folder_root} does not exist'
-    mode = 'panoptic'
-    PATHS = {
-        "train": ("train2017", Path("annotations") / f'{mode}_train2017.json'),
-        "val": ("val2017", Path("annotations") / f'{mode}_val2017.json'),
-    }
-
-    img_folder, ann_file = PATHS[image_set]
-    img_folder_path = img_folder_root / img_folder
-    ann_folder = ann_folder_root / f'{mode}_{img_folder}'
-    ann_file = ann_folder_root / ann_file
-
-    dataset = CocoPanoptic(img_folder_path, ann_folder, ann_file,
-                           transforms=make_coco_transforms(image_set), return_masks=args.masks)
-
-    return dataset

groundingdino/datasets/cocogrounding_eval.py

@@ -1,271 +0,0 @@
-# ------------------------------------------------------------------------
-# 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)
-        # import pdb;pdb.set_trace()
-        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...')
-
-    # import pdb;pdb.set_trace()
-    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
-#################################################################

groundingdino/datasets/data_util.py

@@ -1,170 +0,0 @@
-import os
-import os.path as osp
-import shutil
-import time
-import datetime
-
-import torch
-
-from util.slconfig import SLConfig
-
-class Error(OSError):
-    pass
-
-def slcopytree(src, dst, symlinks=False, ignore=None, copy_function=shutil.copyfile,
-             ignore_dangling_symlinks=False):
-    """
-    modified from shutil.copytree without copystat.
-    
-    Recursively copy a directory tree.
-
-    The destination directory must not already exist.
-    If exception(s) occur, an Error is raised with a list of reasons.
-
-    If the optional symlinks flag is true, symbolic links in the
-    source tree result in symbolic links in the destination tree; if
-    it is false, the contents of the files pointed to by symbolic
-    links are copied. If the file pointed by the symlink doesn't
-    exist, an exception will be added in the list of errors raised in
-    an Error exception at the end of the copy process.
-
-    You can set the optional ignore_dangling_symlinks flag to true if you
-    want to silence this exception. Notice that this has no effect on
-    platforms that don't support os.symlink.
-
-    The optional ignore argument is a callable. If given, it
-    is called with the `src` parameter, which is the directory
-    being visited by copytree(), and `names` which is the list of
-    `src` contents, as returned by os.listdir():
-
-        callable(src, names) -> ignored_names
-
-    Since copytree() is called recursively, the callable will be
-    called once for each directory that is copied. It returns a
-    list of names relative to the `src` directory that should
-    not be copied.
-
-    The optional copy_function argument is a callable that will be used
-    to copy each file. It will be called with the source path and the
-    destination path as arguments. By default, copy2() is used, but any
-    function that supports the same signature (like copy()) can be used.
-
-    """
-    errors = []
-    if os.path.isdir(src):
-        names = os.listdir(src)
-        if ignore is not None:
-            ignored_names = ignore(src, names)
-        else:
-            ignored_names = set()
-
-        os.makedirs(dst)
-        for name in names:
-            if name in ignored_names:
-                continue
-            srcname = os.path.join(src, name)
-            dstname = os.path.join(dst, name)
-            try:
-                if os.path.islink(srcname):
-                    linkto = os.readlink(srcname)
-                    if symlinks:
-                        # We can't just leave it to `copy_function` because legacy
-                        # code with a custom `copy_function` may rely on copytree
-                        # doing the right thing.
-                        os.symlink(linkto, dstname)
-                    else:
-                        # ignore dangling symlink if the flag is on
-                        if not os.path.exists(linkto) and ignore_dangling_symlinks:
-                            continue
-                        # otherwise let the copy occurs. copy2 will raise an error
-                        if os.path.isdir(srcname):
-                            slcopytree(srcname, dstname, symlinks, ignore,
-                                    copy_function)
-                        else:
-                            copy_function(srcname, dstname)
-                elif os.path.isdir(srcname):
-                    slcopytree(srcname, dstname, symlinks, ignore, copy_function)
-                else:
-                    # Will raise a SpecialFileError for unsupported file types
-                    copy_function(srcname, dstname)
-            # catch the Error from the recursive copytree so that we can
-            # continue with other files
-            except Error as err:
-                errors.extend(err.args[0])
-            except OSError as why:
-                errors.append((srcname, dstname, str(why)))
-    else:
-        copy_function(src, dst)
-
-    if errors:
-        raise Error(errors)
-    return dst
-
-def check_and_copy(src_path, tgt_path):
-    if os.path.exists(tgt_path):
-        return None
-
-    return slcopytree(src_path, tgt_path)
-
-
-def remove(srcpath):
-    if os.path.isdir(srcpath):
-        return shutil.rmtree(srcpath)
-    else:
-        return os.remove(srcpath)  
-
-
-def preparing_dataset(pathdict, image_set, args):
-    start_time = time.time()
-    dataset_file = args.dataset_file
-    data_static_info = SLConfig.fromfile('util/static_data_path.py')
-    static_dict = data_static_info[dataset_file][image_set]
-
-    copyfilelist = []
-    for k,tgt_v in pathdict.items():
-        if os.path.exists(tgt_v):
-            if args.local_rank == 0:
-                print("path <{}> exist. remove it!".format(tgt_v))
-                remove(tgt_v)
-            # continue
-        
-        if args.local_rank == 0:
-            src_v = static_dict[k]
-            assert isinstance(src_v, str)
-            if src_v.endswith('.zip'):
-                # copy
-                cp_tgt_dir = os.path.dirname(tgt_v)
-                filename = os.path.basename(src_v)
-                cp_tgt_path = os.path.join(cp_tgt_dir, filename)
-                print('Copy from <{}> to <{}>.'.format(src_v, cp_tgt_path))
-                os.makedirs(cp_tgt_dir, exist_ok=True)
-                check_and_copy(src_v, cp_tgt_path)          
-
-                # unzip
-                import zipfile
-                print("Starting unzip <{}>".format(cp_tgt_path))
-                with zipfile.ZipFile(cp_tgt_path, 'r') as zip_ref:
-                    zip_ref.extractall(os.path.dirname(cp_tgt_path))      
-
-                copyfilelist.append(cp_tgt_path)
-                copyfilelist.append(tgt_v)
-            else:
-                print('Copy from <{}> to <{}>.'.format(src_v, tgt_v))
-                os.makedirs(os.path.dirname(tgt_v), exist_ok=True)
-                check_and_copy(src_v, tgt_v)
-                copyfilelist.append(tgt_v)
-    
-    if len(copyfilelist) == 0:
-        copyfilelist = None
-    args.copyfilelist = copyfilelist
-        
-    if args.distributed:
-        torch.distributed.barrier()
-    total_time = time.time() - start_time
-    if copyfilelist:
-        total_time_str = str(datetime.timedelta(seconds=int(total_time)))
-        print('Data copy time {}'.format(total_time_str))
-    return copyfilelist
-
-
-    

groundingdino/datasets/dataset.py

@@ -1,44 +0,0 @@
-from __future__ import print_function
-
-import torch
-import torchvision.datasets as datasets
-from torch.utils.data import Dataset
-from PIL import Image
-from .tsv_io import TSVFile
-import numpy as np
-import base64
-import io
-
-
-class TSVDataset(Dataset):
-    """ TSV dataset for ImageNet 1K training
-    """    
-    def __init__(self, tsv_file, transform=None, target_transform=None):
-        self.tsv = TSVFile(tsv_file)
-        self.transform = transform
-        self.target_transform = target_transform
-
-    def __getitem__(self, index):
-        """
-        Args:
-            index (int): Index
-        Returns:
-            tuple: (image, target) where target is class_index of the target class.
-        """
-        row = self.tsv.seek(index)
-        image_data = base64.b64decode(row[-1])
-        image = Image.open(io.BytesIO(image_data))
-        image = image.convert('RGB')
-        target = int(row[1])
-
-        if self.transform is not None:
-            img = self.transform(image)
-        else:
-            img = image
-        if self.target_transform is not None:
-            target = self.target_transform(target)
-
-        return img, target
-
-    def __len__(self):
-        return self.tsv.num_rows()

groundingdino/datasets/odvg.py

@@ -1,258 +0,0 @@
-from torchvision.datasets.vision import VisionDataset
-import os.path
-from typing import Callable, Optional
-import json
-from PIL import Image
-import torch
-import random
-import os, sys
-sys.path.append(os.path.dirname(sys.path[0]))
-
-import datasets.transforms as T
-
-class ODVGDataset(VisionDataset):
-    """
-    Args:
-        root (string): Root directory where images are downloaded to.
-        anno (string): Path to json annotation file.
-        label_map_anno (string):  Path to json label mapping file. Only for Object Detection
-        transform (callable, optional): A function/transform that  takes in an PIL image
-            and returns a transformed version. E.g, ``transforms.PILToTensor``
-        target_transform (callable, optional): A function/transform that takes in the
-            target and transforms it.
-        transforms (callable, optional): A function/transform that takes input sample and its target as entry
-            and returns a transformed version.
-    """
-
-    def __init__(
-        self,
-        root: str,
-        anno: str,
-        label_map_anno: str = None,
-        max_labels: int = 80,
-        transform: Optional[Callable] = None,
-        target_transform: Optional[Callable] = None,
-        transforms: Optional[Callable] = None,
-    ) -> None:
-        super().__init__(root, transforms, transform, target_transform)
-        self.root = root
-        self.dataset_mode = "OD" if label_map_anno else "VG"
-        self.max_labels = max_labels
-        if self.dataset_mode == "OD":
-            self.load_label_map(label_map_anno)
-        self._load_metas(anno)
-        self.get_dataset_info()
-
-    def load_label_map(self, label_map_anno):
-        with open(label_map_anno, 'r') as file:
-            self.label_map = json.load(file)
-        self.label_index = set(self.label_map.keys())
-
-    def _load_metas(self, anno):
-      with open(anno, 'r') as f:
-          self.metas = json.load(f)
-
-
-    def get_dataset_info(self):
-        print(f"  == total images: {len(self)}")
-        if self.dataset_mode == "OD":
-            print(f"  == total labels: {len(self.label_map)}")
-
-    def __getitem__(self, index: int):
-        meta = self.metas[index]
-        rel_path = meta["filename"]
-        abs_path = os.path.join(self.root, rel_path)
-        if not os.path.exists(abs_path):
-            raise FileNotFoundError(f"{abs_path} not found.")
-        image = Image.open(abs_path).convert('RGB')
-        w, h = image.size
-        if self.dataset_mode == "OD":
-            anno = meta["detection"]
-            instances = [obj for obj in anno["instances"]]
-            boxes = [obj["bbox"] for obj in instances]
-            # generate vg_labels
-            # pos bbox labels
-            ori_classes = [str(obj["label"]) for obj in instances]
-            pos_labels = set(ori_classes)
-            # neg bbox labels 
-            neg_labels = self.label_index.difference(pos_labels)
-             
-            vg_labels = list(pos_labels)
-            num_to_add = min(len(neg_labels), self.max_labels-len(pos_labels))
-            if num_to_add > 0:
-                vg_labels.extend(random.sample(neg_labels, num_to_add))
-            
-            # shuffle
-            for i in range(len(vg_labels)-1, 0, -1):
-                j = random.randint(0, i)
-                vg_labels[i], vg_labels[j] = vg_labels[j], vg_labels[i]
-
-            caption_list = [self.label_map[lb] for lb in vg_labels]
-            caption_dict = {item:index for index, item in enumerate(caption_list)}
-
-            caption = ' . '.join(caption_list) + ' .'
-            classes = [caption_dict[self.label_map[str(obj["label"])]] for obj in instances]
-            boxes = torch.as_tensor(boxes, dtype=torch.float32).reshape(-1, 4)
-            classes = torch.tensor(classes, dtype=torch.int64)
-        elif self.dataset_mode == "VG":
-            anno = meta["Grounding"]
-            instances = [obj for obj in anno["regions"]] 
-            boxes = [obj["bbox"] for obj in instances]
-            caption_list = [obj["phrase"] for obj in instances]
-            c = list(zip(boxes, caption_list))
-            random.shuffle(c)
-            boxes[:], caption_list[:] = zip(*c)
-            uni_caption_list  = list(set(caption_list))
-            label_map = {}
-            for idx in range(len(uni_caption_list)):
-                label_map[uni_caption_list[idx]] = idx
-            classes = [label_map[cap] for cap in caption_list]
-            caption = ' . '.join(uni_caption_list) + ' .'
-            boxes = torch.as_tensor(boxes, dtype=torch.float32).reshape(-1, 4)
-            classes = torch.tensor(classes, dtype=torch.int64)
-            caption_list = uni_caption_list
-            # print("caption_list" , caption_list)
-            # print("caption" , caption)
-            # print("boxes" , boxes)
-        target = {}
-        target["image_id"] = rel_path.strip(".jpg")
-        target["size"] = torch.as_tensor([int(h), int(w)])
-        target["cap_list"] = caption_list
-        target["caption"] = caption
-        target["boxes"] = boxes
-        target["labels"] = classes
-        # print(" image_id " , target["image_id"])
-        # size, cap_list, caption, bboxes, labels
-
-        if self.transforms is not None:
-            image, target = self.transforms(image, target)
-
-        return image, target
-    
-
-    def __len__(self) -> int:
-        return len(self.metas)
-
-
-def make_coco_transforms(image_set, fix_size=False, strong_aug=False, args=None):
-
-    normalize = T.Compose([
-        T.ToTensor(),
-        T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
-    ])
-
-    # config the params for data aug
-    scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800]
-    max_size = 1333
-    scales2_resize = [400, 500, 600]
-    scales2_crop = [384, 600]
-    
-    # update args from config files
-    scales = getattr(args, 'data_aug_scales', scales)
-    max_size = getattr(args, 'data_aug_max_size', max_size)
-    scales2_resize = getattr(args, 'data_aug_scales2_resize', scales2_resize)
-    scales2_crop = getattr(args, 'data_aug_scales2_crop', scales2_crop)
-
-    # resize them
-    data_aug_scale_overlap = getattr(args, 'data_aug_scale_overlap', None)
-    if data_aug_scale_overlap is not None and data_aug_scale_overlap > 0:
-        data_aug_scale_overlap = float(data_aug_scale_overlap)
-        scales = [int(i*data_aug_scale_overlap) for i in scales]
-        max_size = int(max_size*data_aug_scale_overlap)
-        scales2_resize = [int(i*data_aug_scale_overlap) for i in scales2_resize]
-        scales2_crop = [int(i*data_aug_scale_overlap) for i in scales2_crop]
-
-    # datadict_for_print = {
-    #     'scales': scales,
-    #     'max_size': max_size,
-    #     'scales2_resize': scales2_resize,
-    #     'scales2_crop': scales2_crop
-    # }
-    # print("data_aug_params:", json.dumps(datadict_for_print, indent=2))
-
-    if image_set == 'train':
-        if fix_size:
-            return T.Compose([
-                T.RandomHorizontalFlip(),
-                T.RandomResize([(max_size, max(scales))]),
-                normalize,
-            ])
-
-        if strong_aug:
-            import datasets.sltransform as SLT
-            
-            return T.Compose([
-                T.RandomHorizontalFlip(),
-                T.RandomSelect(
-                    T.RandomResize(scales, max_size=max_size),
-                    T.Compose([
-                        T.RandomResize(scales2_resize),
-                        T.RandomSizeCrop(*scales2_crop),
-                        T.RandomResize(scales, max_size=max_size),
-                    ])
-                ),
-                SLT.RandomSelectMulti([
-                    SLT.RandomCrop(),
-                    SLT.LightingNoise(),
-                    SLT.AdjustBrightness(2),
-                    SLT.AdjustContrast(2),
-                ]),
-                normalize,
-            ])
-        
-        return T.Compose([
-            T.RandomHorizontalFlip(),
-            T.RandomSelect(
-                T.RandomResize(scales, max_size=max_size),
-                T.Compose([
-                    T.RandomResize(scales2_resize),
-                    T.RandomSizeCrop(*scales2_crop),
-                    T.RandomResize(scales, max_size=max_size),
-                ])
-            ),
-            normalize,
-        ])
-
-    if image_set in ['val', 'eval_debug', 'train_reg', 'test']:
-
-        if os.environ.get("GFLOPS_DEBUG_SHILONG", False) == 'INFO':
-            print("Under debug mode for flops calculation only!!!!!!!!!!!!!!!!")
-            return T.Compose([
-                T.ResizeDebug((1280, 800)),
-                normalize,
-            ])   
-
-        return T.Compose([
-            T.RandomResize([max(scales)], max_size=max_size),
-            normalize,
-        ])
-
-    raise ValueError(f'unknown {image_set}')
-
-def build_odvg(image_set, args, datasetinfo):
-    img_folder = datasetinfo["root"]
-    ann_file = datasetinfo["anno"]
-    label_map = datasetinfo["label_map"] if "label_map" in datasetinfo else None
-    try:
-        strong_aug = args.strong_aug
-    except:
-        strong_aug = False # False originally
-    print(img_folder, ann_file, label_map)
-    dataset = ODVGDataset(img_folder, ann_file, label_map, max_labels=args.max_labels,
-            transforms=make_coco_transforms(image_set, fix_size=args.fix_size, strong_aug=strong_aug, args=args), 
-    )
-    return dataset
-
-
-if __name__=="__main__":
-    dataset_vg = ODVGDataset("path/GRIT-20M/data/","path/GRIT-20M/anno/grit_odvg_10k.jsonl",)
-    print(len(dataset_vg))
-    data = dataset_vg[random.randint(0, 100)] 
-    print(data)
-    dataset_od = ODVGDataset("pathl/V3Det/",
-        "path/V3Det/annotations/v3det_2023_v1_all_odvg.jsonl",
-        "path/V3Det/annotations/v3det_label_map.json",
-    )
-    print(len(dataset_od))
-    data = dataset_od[random.randint(0, 100)] 
-    print(data)

groundingdino/datasets/panoptic_eval.py

@@ -1,44 +0,0 @@
-# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
-import json
-import os
-
-import util.misc as utils
-
-try:
-    from panopticapi.evaluation import pq_compute
-except ImportError:
-    pass
-
-
-class PanopticEvaluator(object):
-    def __init__(self, ann_file, ann_folder, output_dir="panoptic_eval"):
-        self.gt_json = ann_file
-        self.gt_folder = ann_folder
-        if utils.is_main_process():
-            if not os.path.exists(output_dir):
-                os.mkdir(output_dir)
-        self.output_dir = output_dir
-        self.predictions = []
-
-    def update(self, predictions):
-        for p in predictions:
-            with open(os.path.join(self.output_dir, p["file_name"]), "wb") as f:
-                f.write(p.pop("png_string"))
-
-        self.predictions += predictions
-
-    def synchronize_between_processes(self):
-        all_predictions = utils.all_gather(self.predictions)
-        merged_predictions = []
-        for p in all_predictions:
-            merged_predictions += p
-        self.predictions = merged_predictions
-
-    def summarize(self):
-        if utils.is_main_process():
-            json_data = {"annotations": self.predictions}
-            predictions_json = os.path.join(self.output_dir, "predictions.json")
-            with open(predictions_json, "w") as f:
-                f.write(json.dumps(json_data))
-            return pq_compute(self.gt_json, predictions_json, gt_folder=self.gt_folder, pred_folder=self.output_dir)
-        return None

groundingdino/datasets/random_crop.py

@@ -1,135 +0,0 @@
-import PIL #version 1.2.0
-import torch
-import os
-import torchvision.transforms.functional as F
-import numpy as np
-import random
-
-
-def intersect(boxes1, boxes2):
-    '''
-        Find intersection of every box combination between two sets of box
-        boxes1: bounding boxes 1, a tensor of dimensions (n1, 4)
-        boxes2: bounding boxes 2, a tensor of dimensions (n2, 4)
-        
-        Out: Intersection each of boxes1 with respect to each of boxes2, 
-             a tensor of dimensions (n1, n2)
-    '''
-    n1 = boxes1.size(0)
-    n2 = boxes2.size(0)
-    max_xy =  torch.min(boxes1[:, 2:].unsqueeze(1).expand(n1, n2, 2),
-                        boxes2[:, 2:].unsqueeze(0).expand(n1, n2, 2))
-    
-    min_xy = torch.max(boxes1[:, :2].unsqueeze(1).expand(n1, n2, 2),
-                       boxes2[:, :2].unsqueeze(0).expand(n1, n2, 2))
-    inter = torch.clamp(max_xy - min_xy , min=0)  # (n1, n2, 2)
-    return inter[:, :, 0] * inter[:, :, 1]  #(n1, n2)
-def find_IoU(boxes1, boxes2):
-    '''
-        Find IoU between every boxes set of boxes 
-        boxes1: a tensor of dimensions (n1, 4) (left, top, right , bottom)
-        boxes2: a tensor of dimensions (n2, 4)
-        
-        Out: IoU each of boxes1 with respect to each of boxes2, a tensor of 
-             dimensions (n1, n2)
-        
-        Formula: 
-        (box1 ∩ box2) / (box1 u box2) = (box1 ∩ box2) / (area(box1) + area(box2) - (box1 ∩ box2 ))
-    '''
-    inter = intersect(boxes1, boxes2)
-    area_boxes1 = (boxes1[:, 2] - boxes1[:, 0]) * (boxes1[:, 3] - boxes1[:, 1])
-    area_boxes2 = (boxes2[:, 2] - boxes2[:, 0]) * (boxes2[:, 3] - boxes2[:, 1])
-    
-    area_boxes1 = area_boxes1.unsqueeze(1).expand_as(inter) #(n1, n2)
-    area_boxes2 = area_boxes2.unsqueeze(0).expand_as(inter)  #(n1, n2)
-    union = (area_boxes1 + area_boxes2 - inter)
-    return inter / union
-
-
-def random_crop(image, boxes, labels, difficulties=None):
-    '''
-        image: A PIL image
-        boxes: Bounding boxes, a tensor of dimensions (#objects, 4)
-        labels: labels of object, a tensor of dimensions (#objects)
-        difficulties: difficulties of detect object, a tensor of dimensions (#objects)
-        
-        Out: cropped image , new boxes, new labels, new difficulties
-    '''
-    if type(image) == PIL.Image.Image:
-        image = F.to_tensor(image)
-    original_h = image.size(1)
-    original_w = image.size(2)
-    
-    while True:
-        mode = random.choice([0.1, 0.3, 0.5, 0.9, None])
-        
-        if mode is None:
-            return F.to_pil_image(image), boxes, labels, difficulties
-        
-        new_image = image
-        new_boxes = boxes
-        new_difficulties = difficulties
-        new_labels = labels
-        for _ in range(50):
-            # Crop dimensions: [0.3, 1] of original dimensions
-            new_h = random.uniform(0.3*original_h, original_h)
-            new_w = random.uniform(0.3*original_w, original_w)
-            
-            # Aspect ratio constraint b/t .5 & 2
-            if new_h/new_w < 0.5 or new_h/new_w > 2:
-                continue
-            
-            #Crop coordinate
-            left = random.uniform(0, original_w - new_w)
-            right = left + new_w
-            top = random.uniform(0, original_h - new_h)
-            bottom = top + new_h
-            crop = torch.FloatTensor([int(left), int(top), int(right), int(bottom)])
-            
-            # Calculate IoU  between the crop and the bounding boxes
-            overlap = find_IoU(crop.unsqueeze(0), boxes) #(1, #objects)
-            overlap = overlap.squeeze(0)
-
-            # If not a single bounding box has a IoU of greater than the minimum, try again
-            if overlap.shape[0] == 0:
-                continue
-            if overlap.max().item() < mode:
-                continue
-            
-            #Crop
-            new_image = image[:, int(top):int(bottom), int(left):int(right)] #(3, new_h, new_w)
-            
-            #Center of bounding boxes
-            center_bb = (boxes[:, :2] + boxes[:, 2:])/2.0
-            
-            #Find bounding box has been had center in crop
-            center_in_crop = (center_bb[:, 0] >left) * (center_bb[:, 0] < right
-                             ) *(center_bb[:, 1] > top) * (center_bb[:, 1] < bottom)    #( #objects)
-            
-            if not center_in_crop.any():
-                continue
-            
-            #take matching bounding box
-            new_boxes = boxes[center_in_crop, :]
-            
-            #take matching labels
-            new_labels = labels[center_in_crop]
-            
-            #take matching difficulities
-            if difficulties is not None:
-                new_difficulties = difficulties[center_in_crop]
-            else:
-                new_difficulties = None
-            
-            #Use the box left and top corner or the crop's
-            new_boxes[:, :2] = torch.max(new_boxes[:, :2], crop[:2])
-            
-            #adjust to crop
-            new_boxes[:, :2] -= crop[:2]
-            
-            new_boxes[:, 2:] = torch.min(new_boxes[:, 2:],crop[2:])
-            
-            #adjust to crop
-            new_boxes[:, 2:] -= crop[:2]
-            
-            return F.to_pil_image(new_image), new_boxes, new_labels, new_difficulties

groundingdino/datasets/sltransform.py

@@ -1,247 +0,0 @@
-# modified from https://github.com/anhtuan85/Data-Augmentation-for-Object-Detection/blob/master/augmentation.ipynb
-
-import PIL #version 1.2.0
-from PIL import Image #version 6.1.0
-import torch
-import os
-import torchvision.transforms.functional as F
-import numpy as np
-import random
-
-from .random_crop import random_crop
-from util.box_ops import box_cxcywh_to_xyxy, box_xyxy_to_cxcywh
-
-class AdjustContrast:
-    def __init__(self, contrast_factor):
-        self.contrast_factor = contrast_factor
-
-    def __call__(self, img, target):
-        """
-        img (PIL Image or Tensor): Image to be adjusted.
-        """
-        _contrast_factor = ((random.random() + 1.0) / 2.0) * self.contrast_factor
-        img = F.adjust_contrast(img, _contrast_factor)
-        return img, target
-
-class AdjustBrightness:
-    def __init__(self, brightness_factor):
-        self.brightness_factor = brightness_factor
-
-    def __call__(self, img, target):
-        """
-        img (PIL Image or Tensor): Image to be adjusted.
-        """
-        _brightness_factor = ((random.random() + 1.0) / 2.0) * self.brightness_factor
-        img = F.adjust_brightness(img, _brightness_factor)
-        return img, target
-
-def lighting_noise(image):
-    '''
-        color channel swap in image
-        image: A PIL image
-    '''
-    new_image = image
-    perms = ((0, 1, 2), (0, 2, 1), (1, 0, 2), 
-             (1, 2, 0), (2, 0, 1), (2, 1, 0))
-    swap = perms[random.randint(0, len(perms)- 1)]
-    new_image = F.to_tensor(new_image)
-    new_image = new_image[swap, :, :]
-    new_image = F.to_pil_image(new_image)
-    return new_image
-
-class LightingNoise:
-    def __init__(self) -> None:
-        pass
-
-    def __call__(self, img, target):
-        return lighting_noise(img), target
-
-
-def rotate(image, boxes, angle):
-    '''
-        Rotate image and bounding box
-        image: A Pil image (w, h)
-        boxes: A tensors of dimensions (#objects, 4)
-        
-        Out: rotated image (w, h), rotated boxes
-    '''
-    new_image = image.copy()
-    new_boxes = boxes.clone()
-    
-    #Rotate image, expand = True
-    w = image.width
-    h = image.height
-    cx = w/2
-    cy = h/2
-    new_image = new_image.rotate(angle, expand=True)
-    angle = np.radians(angle)
-    alpha = np.cos(angle)
-    beta = np.sin(angle)
-    #Get affine matrix
-    AffineMatrix = torch.tensor([[alpha, beta, (1-alpha)*cx - beta*cy],
-                                 [-beta, alpha, beta*cx + (1-alpha)*cy]])
-    
-    #Rotation boxes
-    box_width = (boxes[:,2] - boxes[:,0]).reshape(-1,1)
-    box_height = (boxes[:,3] - boxes[:,1]).reshape(-1,1)
-    
-    #Get corners for boxes
-    x1 = boxes[:,0].reshape(-1,1)
-    y1 = boxes[:,1].reshape(-1,1)
-    
-    x2 = x1 + box_width
-    y2 = y1 
-    
-    x3 = x1
-    y3 = y1 + box_height
-    
-    x4 = boxes[:,2].reshape(-1,1)
-    y4 = boxes[:,3].reshape(-1,1)
-    
-    corners = torch.stack((x1,y1,x2,y2,x3,y3,x4,y4), dim= 1)
-    # corners.reshape(-1, 8)    #Tensors of dimensions (#objects, 8)
-    corners = corners.reshape(-1,2) #Tensors of dimension (4* #objects, 2)
-    corners = torch.cat((corners, torch.ones(corners.shape[0], 1)), dim= 1) #(Tensors of dimension (4* #objects, 3))
-    
-    cos = np.abs(AffineMatrix[0, 0])
-    sin = np.abs(AffineMatrix[0, 1])
-    
-    nW = int((h * sin) + (w * cos))
-    nH = int((h * cos) + (w * sin))
-    AffineMatrix[0, 2] += (nW / 2) - cx
-    AffineMatrix[1, 2] += (nH / 2) - cy
-    
-
-    #Apply affine transform
-    rotate_corners = torch.mm(AffineMatrix, corners.t().to(torch.float64)).t()
-    rotate_corners = rotate_corners.reshape(-1,8)
-    
-    x_corners = rotate_corners[:,[0,2,4,6]]
-    y_corners = rotate_corners[:,[1,3,5,7]]
-    
-    #Get (x_min, y_min, x_max, y_max)
-    x_min, _ = torch.min(x_corners, dim= 1)
-    x_min = x_min.reshape(-1, 1)
-    y_min, _ = torch.min(y_corners, dim= 1)
-    y_min = y_min.reshape(-1, 1)
-    x_max, _ = torch.max(x_corners, dim= 1)
-    x_max = x_max.reshape(-1, 1)
-    y_max, _ = torch.max(y_corners, dim= 1)
-    y_max = y_max.reshape(-1, 1)
-    
-    new_boxes = torch.cat((x_min, y_min, x_max, y_max), dim= 1)
-    
-    scale_x = new_image.width / w
-    scale_y = new_image.height / h
-    
-    #Resize new image to (w, h)
-
-    new_image = new_image.resize((w, h))
-    
-    #Resize boxes
-    new_boxes /= torch.Tensor([scale_x, scale_y, scale_x, scale_y])
-    new_boxes[:, 0] = torch.clamp(new_boxes[:, 0], 0, w)
-    new_boxes[:, 1] = torch.clamp(new_boxes[:, 1], 0, h)
-    new_boxes[:, 2] = torch.clamp(new_boxes[:, 2], 0, w)
-    new_boxes[:, 3] = torch.clamp(new_boxes[:, 3], 0, h)
-    return new_image, new_boxes
-
-# def convert_xywh_to_xyxy(boxes: torch.Tensor):
-#     _boxes = boxes.clone()
-#     box_xy = _boxes[:, :2]
-#     box_wh = _boxes[:, 2:]
-#     box_x1y1 = box_xy - box_wh/2 
-#     box_x2y2 = box_xy + box_wh/2
-#     box_xyxy = torch.cat((box_x1y1, box_x2y2), dim=-1)
-#     return box_xyxy
-
-class Rotate:
-    def __init__(self, angle=10) -> None:
-        self.angle = angle
-
-    def __call__(self, img, target):
-        w,h = img.size
-        whwh = torch.Tensor([w, h, w, h])
-        boxes_xyxy = box_cxcywh_to_xyxy(target['boxes']) * whwh
-        img, boxes_new = rotate(img, boxes_xyxy, self.angle)
-        target['boxes'] = box_xyxy_to_cxcywh(boxes_new).to(boxes_xyxy.dtype) / (whwh + 1e-3)
-        return img, target
-
-
-class RandomCrop:
-    def __init__(self) -> None:
-        pass
-
-    def __call__(self, img, target):
-        w,h = img.size
-        try:
-            boxes_xyxy = target['boxes']
-            labels = target['labels']
-            img, new_boxes, new_labels, _ = random_crop(img, boxes_xyxy, labels)
-            target['boxes'] = new_boxes
-            target['labels'] = new_labels
-        except Exception as e:
-            pass
-        return img, target
-
-
-class RandomCropDebug:
-    def __init__(self) -> None:
-        pass
-
-    def __call__(self, img, target):
-        boxes_xyxy = target['boxes'].clone()
-        labels = target['labels'].clone()
-        img, new_boxes, new_labels, _ = random_crop(img, boxes_xyxy, labels)
-        target['boxes'] = new_boxes
-        target['labels'] = new_labels
-
-
-        return img, target
-        
-class RandomSelectMulti(object):
-    """
-    Randomly selects between transforms1 and transforms2,
-    """
-    def __init__(self, transformslist, p=-1):
-        self.transformslist = transformslist
-        self.p = p
-        assert p == -1
-
-    def __call__(self, img, target):
-        if self.p == -1:
-            return random.choice(self.transformslist)(img, target)
-
-
-class Albumentations:
-    def __init__(self):
-        import albumentations as A
-        self.transform = A.Compose([
-            A.Blur(p=0.01),
-            A.MedianBlur(p=0.01),
-            A.ToGray(p=0.01),
-            A.CLAHE(p=0.01),
-            A.RandomBrightnessContrast(p=0.005),
-            A.RandomGamma(p=0.005),
-            A.ImageCompression(quality_lower=75, p=0.005)],
-            bbox_params=A.BboxParams(format='pascal_voc', label_fields=['class_labels']))
-
-    def __call__(self, img, target, p=1.0):
-        """
-        Input:
-            target['boxes']: xyxy, unnormalized data.
-        
-        """
-        boxes_raw = target['boxes']
-        labels_raw = target['labels']
-        img_np = np.array(img)
-        if self.transform and random.random() < p:
-            new_res = self.transform(image=img_np, bboxes=boxes_raw, class_labels=labels_raw)  # transformed
-            boxes_new = torch.Tensor(new_res['bboxes']).to(boxes_raw.dtype).reshape_as(boxes_raw)
-            img_np = new_res['image']
-            labels_new = torch.Tensor(new_res['class_labels']).to(labels_raw.dtype)
-        img_new = Image.fromarray(img_np)
-        target['boxes'] = boxes_new
-        target['labels'] = labels_new
-        
-        return img_new, target

groundingdino/datasets/transforms.py

@@ -1,285 +0,0 @@
-# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
-"""
-Transforms and data augmentation for both image + bbox.
-"""
-import random
-
-import PIL
-import torch
-import torchvision.transforms as T
-import torchvision.transforms.functional as F
-
-from util.box_ops import box_xyxy_to_cxcywh
-from 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"]
-
-    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:
-            target[field] = target[field][keep]
-
-    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):
-        self.min_size = min_size
-        self.max_size = max_size
-
-    def __call__(self, img: PIL.Image.Image, target: dict):
-        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])
-        return crop(img, target, region)
-
-
-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.))
-        crop_left = int(round((image_width - crop_width) / 2.))
-        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

groundingdino/models/GroundingDINO/.ipynb_checkpoints/bertwarper-checkpoint.py

@@ -1,273 +0,0 @@
-# ------------------------------------------------------------------------
-# 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

groundingdino/models/GroundingDINO/.ipynb_checkpoints/fuse_modules-checkpoint.py

@@ -1,298 +0,0 @@
-# ------------------------------------------------------------------------
-# 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
-import loralib as lora
-
-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
-        r = 12
-        # Object feature encoding
-        self.fc = lora.Linear(input_feat_size, output_feat_size,r=r, 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
-        r = 12
-        self.v_proj = lora.Linear(self.v_dim, self.embed_dim , r=r)
-        self.l_proj = lora.Linear(self.l_dim, self.embed_dim , r=r )
-        self.values_v_proj = lora.Linear(self.v_dim, self.embed_dim , r=r )
-        self.values_l_proj = lora.Linear(self.l_dim, self.embed_dim , r=r )
-
-        self.out_v_proj = lora.Linear(self.embed_dim, self.v_dim , r=r )
-        self.out_l_proj = lora.Linear(self.embed_dim, self.l_dim , r=r )
-
-        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)

groundingdino/models/GroundingDINO/.ipynb_checkpoints/groundingdino-checkpoint.py

@@ -1,857 +0,0 @@
-# ------------------------------------------------------------------------
-# 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 groundingdino.util import box_ops, get_tokenlizer
-from groundingdino.util.misc import (
-    NestedTensor,
-    accuracy,
-    get_world_size,
-    interpolate,
-    inverse_sigmoid,
-    is_dist_avail_and_initialized,
-    nested_tensor_from_tensor_list,
-)
-from groundingdino.util.utils import get_phrases_from_posmap
-from groundingdino.util.visualizer import COCOVisualizer
-from 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
-
-from .matcher import build_matcher
-
-
-
-
-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 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
-        )
-        one_hot_token = tokenized
-
-        (
-            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:
-            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
-        }
-
-
-        if isinstance(samples, (list, torch.Tensor)):
-            samples = nested_tensor_from_tensor_list(samples)
-        features, poss = self.backbone(samples)
-        srcs = []
-        masks = []
-        for l, feat in enumerate(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](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)
-                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, 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)
-
-
-        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]}
-
-        # Used to calculate losses
-        bs, len_td = text_dict['text_token_mask'].shape
-        out['text_mask']=torch.zeros(bs, self.max_text_len, dtype=torch.bool).to(
-            samples.device
-        )
-        for b in range(bs):
-            for j in range(len_td):
-                if text_dict['text_token_mask'][b][j] == True:
-                    out['text_mask'][b][j] = True
-
-        # for intermediate outputs
-        if self.aux_loss:
-            out['aux_outputs'] = self._set_aux_loss(outputs_class, outputs_coord_list)
-        out['token']=one_hot_token
-        # # 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}
-
-        # outputs['pred_logits'].shape
-        # torch.Size([4, 900, 256])
-
-        # outputs['pred_boxes'].shape
-        # torch.Size([4, 900, 4])
-
-        # outputs['text_mask'].shape
-        # torch.Size([256])
-
-        # outputs['text_mask']
-
-        # outputs['aux_outputs'][0].keys()
-        # dict_keys(['pred_logits', 'pred_boxes', 'one_hot', 'text_mask'])
-
-        # outputs['aux_outputs'][img_idx]
-
-        # outputs['token']
-        # <class 'transformers.tokenization_utils_base.BatchEncoding'>
-
-        # outputs['interm_outputs'].keys()
-        # dict_keys(['pred_logits', 'pred_boxes', 'one_hot', 'text_mask'])
-
-
-        # outputs['interm_outputs_for_matching_pre'].keys()
-        # dict_keys(['pred_logits', 'pred_boxes'])
-
-        # outputs['one_hot'].shape
-        # torch.Size([4, 900, 256])
-
-        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])
-        ]
-
-
-
-
-class SetCriterion(nn.Module):
-    def __init__(self, matcher, weight_dict, focal_alpha,focal_gamma, losses):
-        """ Create the criterion.
-        Parameters:
-            matcher: module able to compute a matching between targets and proposals
-            weight_dict: dict containing as key the names of the losses and as values their relative weight.
-            losses: list of all the losses to be applied. See get_loss for list of available losses.
-            focal_alpha: alpha in Focal Loss
-        """
-        super().__init__()
-        self.matcher = matcher
-        self.weight_dict = weight_dict
-        self.losses = losses
-        self.focal_alpha = focal_alpha
-        self.focal_gamma= focal_gamma
-
-    @torch.no_grad()
-    def loss_cardinality(self, outputs, targets, indices, num_boxes):
-        """ Compute the cardinality error, ie the absolute error in the number of predicted non-empty boxes
-        This is not really a loss, it is intended for logging purposes only. It doesn't propagate gradients
-        """
-
-        pred_logits = outputs['pred_logits']
-        device = pred_logits.device
-        tgt_lengths = torch.as_tensor([len(v["labels"]) for v in targets], device=device)
-        # Count the number of predictions that are NOT "no-object" (which is the last class)
-        card_pred = (pred_logits.argmax(-1) != pred_logits.shape[-1] - 1).sum(1)
-        card_err = F.l1_loss(card_pred.float(), tgt_lengths.float())
-        losses = {'cardinality_error': card_err}
-        return losses
-
-    def loss_boxes(self, outputs, targets, indices, num_boxes):
-        """Compute the losses related to the bounding boxes, the L1 regression loss and the GIoU loss
-           targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4]
-           The target boxes are expected in format (center_x, center_y, w, h), normalized by the image size.
-        """
-        assert 'pred_boxes' in outputs
-        idx = self._get_src_permutation_idx(indices)
-        src_boxes = outputs['pred_boxes'][idx]
-        target_boxes = torch.cat([t['boxes'][i] for t, (_, i) in zip(targets, indices)], dim=0)
-
-        loss_bbox = F.l1_loss(src_boxes, target_boxes, reduction='none')
-
-        losses = {}
-        losses['loss_bbox'] = loss_bbox.sum() / num_boxes
-
-        loss_giou = 1 - torch.diag(box_ops.generalized_box_iou(
-            box_ops.box_cxcywh_to_xyxy(src_boxes),
-            box_ops.box_cxcywh_to_xyxy(target_boxes)))
-        losses['loss_giou'] = loss_giou.sum() / num_boxes
-
-        # calculate the x,y and h,w loss
-        with torch.no_grad():
-            losses['loss_xy'] = loss_bbox[..., :2].sum() / num_boxes
-            losses['loss_hw'] = loss_bbox[..., 2:].sum() / num_boxes
-
-
-        return losses
-
-
-    def token_sigmoid_binary_focal_loss(self, outputs, targets, indices, num_boxes):
-        pred_logits=outputs['pred_logits']
-        new_targets=outputs['one_hot'].to(pred_logits.device)
-        text_mask=outputs['text_mask']
-
-        assert (new_targets.dim() == 3)
-        assert (pred_logits.dim() == 3)  # batch x from x to
-        
-        bs, n, _ = pred_logits.shape
-        alpha=self.focal_alpha
-        gamma=self.focal_gamma
-        if text_mask is not None:
-            # ODVG: each sample has different mask 
-            text_mask = text_mask.repeat(1, pred_logits.size(1)).view(outputs['text_mask'].shape[0],-1,outputs['text_mask'].shape[1])
-            pred_logits = torch.masked_select(pred_logits, text_mask)
-            new_targets = torch.masked_select(new_targets, text_mask)
-
-        new_targets=new_targets.float()
-        p = torch.sigmoid(pred_logits)
-        ce_loss = F.binary_cross_entropy_with_logits(pred_logits, new_targets, reduction="none")
-        p_t = p * new_targets + (1 - p) * (1 - new_targets)
-        loss = ce_loss * ((1 - p_t) ** gamma)
-
-        if alpha >= 0:
-            alpha_t = alpha * new_targets + (1 - alpha) * (1 - new_targets)
-            loss = alpha_t * loss
-
-        total_num_pos=0
-        for batch_indices in indices:
-            total_num_pos += len(batch_indices[0])
-        num_pos_avg_per_gpu = max(total_num_pos , 1.0)
-        loss=loss.sum()/num_pos_avg_per_gpu
-        
-        losses = {'loss_ce': loss}
-        return losses
-
-
-    def _get_src_permutation_idx(self, indices):
-        # permute predictions following indices
-        batch_idx = torch.cat([torch.full_like(src, i) for i, (src, _) in enumerate(indices)])
-        src_idx = torch.cat([src for (src, _) in indices])
-        return batch_idx, src_idx
-
-    def _get_tgt_permutation_idx(self, indices):
-        # permute targets following indices
-        batch_idx = torch.cat([torch.full_like(tgt, i) for i, (_, tgt) in enumerate(indices)])
-        tgt_idx = torch.cat([tgt for (_, tgt) in indices])
-        return batch_idx, tgt_idx
-
-    def get_loss(self, loss, outputs, targets, indices, num_boxes, **kwargs):
-        loss_map = {
-            'labels': self.token_sigmoid_binary_focal_loss,
-            'cardinality': self.loss_cardinality,
-            'boxes': self.loss_boxes,
-        }
-        assert loss in loss_map, f'do you really want to compute {loss} loss?'
-        return loss_map[loss](outputs, targets, indices, num_boxes, **kwargs)
-
-    def forward(self, outputs, targets, cat_list, caption, return_indices=False):
-        """ This performs the loss computation.
-        Parameters:
-             outputs: dict of tensors, see the output specification of the model for the format
-             targets: list of dicts, such that len(targets) == batch_size.
-                      The expected keys in each dict depends on the losses applied, see each loss' doc
-            
-             return_indices: used for vis. if True, the layer0-5 indices will be returned as well.
-        """
-        device=next(iter(outputs.values())).device
-        one_hot = torch.zeros(outputs['pred_logits'].size(),dtype=torch.int64) # torch.Size([bs, 900, 256])
-        token = outputs['token'] 
-        
-        label_map_list = []
-        indices = []
-        for j in range(len(cat_list)): # bs
-            label_map=[]
-            for i in range(len(cat_list[j])):
-                label_id=torch.tensor([i])
-                per_label=create_positive_map(token[j], label_id, cat_list[j], caption[j])
-                label_map.append(per_label)
-            label_map=torch.stack(label_map,dim=0).squeeze(1)
-            label_map_list.append(label_map)
-        for j in range(len(cat_list)): # bs
-            for_match = {
-                "pred_logits" : outputs['pred_logits'][j].unsqueeze(0),
-                "pred_boxes" : outputs['pred_boxes'][j].unsqueeze(0)
-            }
-            inds = self.matcher(for_match, [targets[j]], label_map_list[j])
-            indices.extend(inds)
-        # indices : A list of size batch_size, containing tuples of (index_i, index_j) where:
-        # - index_i is the indices of the selected predictions (in order)
-        # - index_j is the indices of the corresponding selected targets (in order)
-
-        # import pdb; pdb.set_trace()
-        tgt_ids = [v["labels"].cpu() for v in targets]
-        # len(tgt_ids) == bs
-        for i in range(len(indices)):
-            tgt_ids[i]=tgt_ids[i][indices[i][1]]
-            one_hot[i,indices[i][0]] = label_map_list[i][tgt_ids[i]].to(torch.long)
-        outputs['one_hot'] = one_hot
-        if return_indices:
-            indices0_copy = indices
-            indices_list = []
-
-        # Compute the average number of target boxes accross all nodes, for normalization purposes
-        num_boxes_list = [len(t["labels"]) for t in targets]
-        num_boxes = sum(num_boxes_list)
-        num_boxes = torch.as_tensor([num_boxes], dtype=torch.float, device=device)
-        if is_dist_avail_and_initialized():
-            torch.distributed.all_reduce(num_boxes)
-        num_boxes = torch.clamp(num_boxes / get_world_size(), min=1).item()
-
-        # Compute all the requested losses
-        losses = {}
-        for loss in self.losses:
-            losses.update(self.get_loss(loss, outputs, targets, indices, num_boxes))
-
-        # In case of auxiliary losses, we repeat this process with the output of each intermediate layer.
-        if 'aux_outputs' in outputs:
-            for idx, aux_outputs in enumerate(outputs['aux_outputs']):
-                indices = []
-                for j in range(len(cat_list)): # bs
-                    aux_output_single = {
-                        'pred_logits' : aux_outputs['pred_logits'][j].unsqueeze(0),
-                        'pred_boxes': aux_outputs['pred_boxes'][j].unsqueeze(0)
-                    }
-                    inds = self.matcher(aux_output_single, [targets[j]], label_map_list[j])
-                    indices.extend(inds)
-                one_hot_aux = torch.zeros(outputs['pred_logits'].size(),dtype=torch.int64)
-                tgt_ids = [v["labels"].cpu() for v in targets]
-                for i in range(len(indices)):
-                    tgt_ids[i]=tgt_ids[i][indices[i][1]]
-                    one_hot_aux[i,indices[i][0]] = label_map_list[i][tgt_ids[i]].to(torch.long)
-                aux_outputs['one_hot'] = one_hot_aux
-                aux_outputs['text_mask'] = outputs['text_mask']
-                if return_indices:
-                    indices_list.append(indices)
-                for loss in self.losses:
-                    kwargs = {}
-                    l_dict = self.get_loss(loss, aux_outputs, targets, indices, num_boxes, **kwargs)                
-                    l_dict = {k + f'_{idx}': v for k, v in l_dict.items()}
-                    losses.update(l_dict)
-
-        # interm_outputs loss
-        if 'interm_outputs' in outputs:
-            interm_outputs = outputs['interm_outputs']
-            indices = []
-            for j in range(len(cat_list)): # bs
-                interm_output_single = {
-                    'pred_logits' : interm_outputs['pred_logits'][j].unsqueeze(0),
-                    'pred_boxes': interm_outputs['pred_boxes'][j].unsqueeze(0)
-                }
-                inds = self.matcher(interm_output_single, [targets[j]], label_map_list[j])
-                indices.extend(inds)
-            one_hot_aux = torch.zeros(outputs['pred_logits'].size(),dtype=torch.int64)
-            tgt_ids = [v["labels"].cpu() for v in targets]
-            for i in range(len(indices)):
-                tgt_ids[i]=tgt_ids[i][indices[i][1]]
-                one_hot_aux[i,indices[i][0]] = label_map_list[i][tgt_ids[i]].to(torch.long)
-            interm_outputs['one_hot'] = one_hot_aux
-            interm_outputs['text_mask'] = outputs['text_mask']
-            if return_indices:
-                indices_list.append(indices)
-            for loss in self.losses:
-                kwargs = {}
-                l_dict = self.get_loss(loss, interm_outputs, targets, indices, num_boxes, **kwargs)
-                l_dict = {k + f'_interm': v for k, v in l_dict.items()}
-                losses.update(l_dict)
-
-        if return_indices:
-            indices_list.append(indices0_copy)
-            return losses, indices_list
-
-        return losses
-
-
-class PostProcess(nn.Module):
-    """ This module converts the model's output into the format expected by the coco api"""
-    def __init__(self, num_select=100,text_encoder_type='text_encoder_type', nms_iou_threshold=-1,use_coco_eval=False,args=None) -> None:
-        super().__init__()
-        self.num_select = num_select
-        self.tokenizer = get_tokenlizer.get_tokenlizer(text_encoder_type)
-        if args.use_coco_eval:
-            from pycocotools.coco import COCO
-            coco = COCO(args.coco_val_path)
-            category_dict = coco.loadCats(coco.getCatIds())
-            cat_list = [item['name'] for item in category_dict]
-        else:
-            cat_list=args.label_list
-        caption = " . ".join(cat_list) + ' .'
-        tokenized = self.tokenizer(caption, padding="longest", return_tensors="pt")
-        label_list = torch.arange(len(cat_list))
-        pos_map=create_positive_map(tokenized,label_list,cat_list,caption)
-        # build a mapping from label_id to pos_map
-        if args.use_coco_eval:
-            id_map = {0: 1, 1: 2, 2: 3, 3: 4, 4: 5, 5: 6, 6: 7, 7: 8, 8: 9, 9: 10, 10: 11, 11: 13, 12: 14, 13: 15, 14: 16, 15: 17, 16: 18, 17: 19, 18: 20, 19: 21, 20: 22, 21: 23, 22: 24, 23: 25, 24: 27, 25: 28, 26: 31, 27: 32, 28: 33, 29: 34, 30: 35, 31: 36, 32: 37, 33: 38, 34: 39, 35: 40, 36: 41, 37: 42, 38: 43, 39: 44, 40: 46,
-                    41: 47, 42: 48, 43: 49, 44: 50, 45: 51, 46: 52, 47: 53, 48: 54, 49: 55, 50: 56, 51: 57, 52: 58, 53: 59, 54: 60, 55: 61, 56: 62, 57: 63, 58: 64, 59: 65, 60: 67, 61: 70, 62: 72, 63: 73, 64: 74, 65: 75, 66: 76, 67: 77, 68: 78, 69: 79, 70: 80, 71: 81, 72: 82, 73: 84, 74: 85, 75: 86, 76: 87, 77: 88, 78: 89, 79: 90}
-            new_pos_map = torch.zeros((91, 256))
-            for k, v in id_map.items():
-                new_pos_map[v] = pos_map[k]
-            pos_map=new_pos_map
-
-
-        self.nms_iou_threshold=nms_iou_threshold
-        self.positive_map = pos_map
-
-    @torch.no_grad()
-    def forward(self, outputs, target_sizes, not_to_xyxy=False, test=False):
-        """ Perform the computation
-        Parameters:
-            outputs: raw outputs of the model
-            target_sizes: tensor of dimension [batch_size x 2] containing the size of each images of the batch
-                          For evaluation, this must be the original image size (before any data augmentation)
-                          For visualization, this should be the image size after data augment, but before padding
-        """
-        num_select = self.num_select
-        out_logits, out_bbox = outputs['pred_logits'], outputs['pred_boxes']
-
-
-        prob_to_token = out_logits.sigmoid()
-        pos_maps = self.positive_map.to(prob_to_token.device)
-        for label_ind in range(len(pos_maps)):
-            if pos_maps[label_ind].sum() != 0:
-                pos_maps[label_ind]=pos_maps[label_ind]/pos_maps[label_ind].sum()
-
-        prob_to_label = prob_to_token @ pos_maps.T
-
-        assert len(out_logits) == len(target_sizes)
-        assert target_sizes.shape[1] == 2
-
-        prob = prob_to_label
-        topk_values, topk_indexes = torch.topk(prob.view(prob.shape[0], -1), num_select, dim=1)
-        scores = topk_values
-        topk_boxes = torch.div(topk_indexes, prob.shape[2], rounding_mode='trunc')
-        labels = topk_indexes % prob.shape[2]
-        if not_to_xyxy:
-            boxes = out_bbox
-        else:
-            boxes = box_ops.box_cxcywh_to_xyxy(out_bbox)
-
-        # if test:
-        #     assert not not_to_xyxy
-        #     boxes[:,:,2:] = boxes[:,:,2:] - boxes[:,:,:2]
-        boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1,1,4))
-        
-        # and from relative [0, 1] to absolute [0, height] coordinates
-        img_h, img_w = target_sizes.unbind(1)
-        scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1)
-        boxes = boxes * scale_fct[:, None, :]
-
-        if self.nms_iou_threshold > 0:
-            item_indices = [nms(b, s, iou_threshold=self.nms_iou_threshold) for b,s in zip(boxes, scores)]
-
-            results = [{'scores': s[i], 'labels': l[i], 'boxes': b[i]} for s, l, b, i in zip(scores, labels, boxes, item_indices)]
-        else:
-            results = [{'scores': s, 'labels': l, 'boxes': b} for s, l, b in zip(scores, labels, boxes)]
-        results = [{'scores': s, 'labels': l, 'boxes': b} for s, l, b in zip(scores, labels, boxes)]
-        return results
-
-
-@MODULE_BUILD_FUNCS.registe_with_name(module_name="groundingdino")
-def build_groundingdino(args):
-    device = torch.device(args.device)
-    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=args.aux_loss,
-        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,
-    )
-
-
-
-    matcher = build_matcher(args)
-
-    # prepare weight dict
-    weight_dict = {'loss_ce': args.cls_loss_coef, 'loss_bbox': args.bbox_loss_coef}
-    weight_dict['loss_giou'] = args.giou_loss_coef
-    clean_weight_dict_wo_dn = copy.deepcopy(weight_dict)
-
-    
-
-    clean_weight_dict = copy.deepcopy(weight_dict)
-
-    # TODO this is a hack
-    if args.aux_loss:
-        aux_weight_dict = {}
-        for i in range(args.dec_layers - 1):
-            aux_weight_dict.update({k + f'_{i}': v for k, v in clean_weight_dict.items()})
-        weight_dict.update(aux_weight_dict)
-
-    if args.two_stage_type != 'no':
-        interm_weight_dict = {}
-        try:
-            no_interm_box_loss = args.no_interm_box_loss
-        except:
-            no_interm_box_loss = False
-        _coeff_weight_dict = {
-            'loss_ce': 1.0,
-            'loss_bbox': 1.0 if not no_interm_box_loss else 0.0,
-            'loss_giou': 1.0 if not no_interm_box_loss else 0.0,
-        }
-        try:
-            interm_loss_coef = args.interm_loss_coef
-        except:
-            interm_loss_coef = 1.0
-        interm_weight_dict.update({k + f'_interm': v * interm_loss_coef * _coeff_weight_dict[k] for k, v in clean_weight_dict_wo_dn.items()})
-        weight_dict.update(interm_weight_dict)
-
-    # losses = ['labels', 'boxes', 'cardinality']
-    losses = ['labels', 'boxes']
-
-    criterion = SetCriterion(matcher=matcher, weight_dict=weight_dict,
-                             focal_alpha=args.focal_alpha, focal_gamma=args.focal_gamma,losses=losses
-                             )
-    criterion.to(device)
-    postprocessors = {'bbox': PostProcess(num_select=args.num_select  , text_encoder_type=args.text_encoder_type,nms_iou_threshold=args.nms_iou_threshold,args=args)}
-
-    return model, criterion, postprocessors
-
-def create_positive_map(tokenized, tokens_positive,cat_list,caption):
-    """construct a map such that positive_map[i,j] = True iff box i is associated to token j"""
-    positive_map = torch.zeros((len(tokens_positive), 256), dtype=torch.float)
-
-    for j,label in enumerate(tokens_positive):
-
-        start_ind = caption.find(cat_list[label])
-        end_ind = start_ind + len(cat_list[label]) - 1
-        beg_pos = tokenized.char_to_token(start_ind)
-        try:
-            end_pos = tokenized.char_to_token(end_ind)
-        except:
-            end_pos = None
-        if end_pos is None:
-            try:
-                end_pos = tokenized.char_to_token(end_ind - 1)
-                if end_pos is None:
-                    end_pos = tokenized.char_to_token(end_ind - 2)
-            except:
-                end_pos = None
-        # except Exception as e:
-        #     print("beg:", beg, "end:", end)
-        #     print("token_positive:", tokens_positive)
-        #     # print("beg_pos:", beg_pos, "end_pos:", end_pos)
-        #     raise e
-        # if beg_pos is None:
-        #     try:
-        #         beg_pos = tokenized.char_to_token(beg + 1)
-        #         if beg_pos is None:
-        #             beg_pos = tokenized.char_to_token(beg + 2)
-        #     except:
-        #         beg_pos = None
-        # if end_pos is None:
-        #     try:
-        #         end_pos = tokenized.char_to_token(end - 2)
-        #         if end_pos is None:
-        #             end_pos = tokenized.char_to_token(end - 3)
-        #     except:
-        #         end_pos = None
-        if beg_pos is None or end_pos is None:
-            continue
-        if beg_pos < 0 or end_pos < 0:
-            continue
-        if beg_pos > end_pos:
-            continue
-        # assert beg_pos is not None and end_pos is not None
-        positive_map[j,beg_pos: end_pos + 1].fill_(1)
-    return positive_map 
-
-

groundingdino/models/GroundingDINO/.ipynb_checkpoints/matcher-checkpoint.py

@@ -1,218 +0,0 @@
-# ------------------------------------------------------------------------
-# DINO
-# Copyright (c) 2022 IDEA. All Rights Reserved.
-# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
-# ------------------------------------------------------------------------
-# Modules to compute the matching cost and solve the corresponding LSAP.
-# 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 torch, os
-from torch import nn
-from scipy.optimize import linear_sum_assignment
-
-from util.box_ops import box_cxcywh_to_xyxy, generalized_box_iou
-
-
-class HungarianMatcher(nn.Module):
-    """This class computes an assignment between the targets and the predictions of the network
-    For efficiency reasons, the targets don't include the no_object. Because of this, in general,
-    there are more predictions than targets. In this case, we do a 1-to-1 matching of the best predictions,
-    while the others are un-matched (and thus treated as non-objects).
-    """
-
-    def __init__(self, cost_class: float = 1, cost_bbox: float = 1, cost_giou: float = 1, focal_alpha = 0.25):
-        """Creates the matcher
-        Params:
-            cost_class: This is the relative weight of the classification error in the matching cost
-            cost_bbox: This is the relative weight of the L1 error of the bounding box coordinates in the matching cost
-            cost_giou: This is the relative weight of the giou loss of the bounding box in the matching cost
-        """
-        super().__init__()
-        self.cost_class = cost_class
-        self.cost_bbox = cost_bbox
-        self.cost_giou = cost_giou
-        assert cost_class != 0 or cost_bbox != 0 or cost_giou != 0, "all costs cant be 0"
-
-        self.focal_alpha = focal_alpha
-
-    @torch.no_grad()
-    def forward(self, outputs, targets, label_map):
-        """ Performs the matching
-        Params:
-            outputs: This is a dict that contains at least these entries:
-                 "pred_logits": Tensor of dim [batch_size, num_queries, num_classes] with the classification logits
-                 "pred_boxes": Tensor of dim [batch_size, num_queries, 4] with the predicted box coordinates
-            targets: This is a list of targets (len(targets) = batch_size), where each target is a dict containing:
-                 "labels": Tensor of dim [num_target_boxes] (where num_target_boxes is the number of ground-truth
-                           objects in the target) containing the class labels
-                 "boxes": Tensor of dim [num_target_boxes, 4] containing the target box coordinates
-        Returns:
-            A list of size batch_size, containing tuples of (index_i, index_j) where:
-                - index_i is the indices of the selected predictions (in order)
-                - index_j is the indices of the corresponding selected targets (in order)
-            For each batch element, it holds:
-                len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
-        """
-
-        bs, num_queries = outputs["pred_logits"].shape[:2]
-
-        # We flatten to compute the cost matrices in a batch
-        out_prob = outputs["pred_logits"].flatten(0, 1).sigmoid()  # [batch_size * num_queries, num_classes]
-        out_bbox = outputs["pred_boxes"].flatten(0, 1)  # [batch_size * num_queries, 4]
-
-        # Also concat the target labels and boxes
-        tgt_ids = torch.cat([v["labels"] for v in targets])
-        tgt_bbox = torch.cat([v["boxes"] for v in targets])
-
-        # Compute the classification cost.
-        alpha = self.focal_alpha
-        gamma = 2.0
-
-        new_label_map=label_map[tgt_ids.cpu()]
-
-        neg_cost_class = (1 - alpha) * (out_prob ** gamma) * (-(1 - out_prob + 1e-8).log())
-        pos_cost_class = alpha * ((1 - out_prob) ** gamma) * (-(out_prob + 1e-8).log())
-        new_label_map=new_label_map.to(pos_cost_class.device)
-        
-        cost_bbox = torch.cdist(out_bbox, tgt_bbox, p=1)
-
-        # cost_class=(pos_cost_class @ new_label_map.T - neg_cost_class@ new_label_map.T)
-        cost_class=[]
-        for idx_map in new_label_map:       
-            idx_map = idx_map / idx_map.sum()
-            cost_class.append(pos_cost_class @ idx_map - neg_cost_class@ idx_map)
-        if cost_class:
-            cost_class=torch.stack(cost_class,dim=0).T
-        else:
-            cost_class=torch.zeros_like(cost_bbox)
-        # Compute the L1 cost between boxes
-        
-
-        # Compute the giou cost betwen boxes
-        cost_giou = -generalized_box_iou(box_cxcywh_to_xyxy(out_bbox), box_cxcywh_to_xyxy(tgt_bbox))
-        # import pdb;pdb.set_trace()
-        # Final cost matrix
-        C = self.cost_bbox * cost_bbox + self.cost_class * cost_class + self.cost_giou * cost_giou
-        C = C.view(bs, num_queries, -1).cpu()
-        C[torch.isnan(C)] = 0.0
-        C[torch.isinf(C)] = 0.0
-
-        sizes = [len(v["boxes"]) for v in targets]
-        try:
-            indices = [linear_sum_assignment(c[i]) for i, c in enumerate(C.split(sizes, -1))]
-        except:
-            print("warning: use SimpleMinsumMatcher")
-            indices = []
-            device = C.device
-            for i, (c, _size) in enumerate(zip(C.split(sizes, -1), sizes)):
-                weight_mat = c[i]
-                idx_i = weight_mat.min(0)[1]
-                idx_j = torch.arange(_size).to(device)
-                indices.append((idx_i, idx_j))
-        return [(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64)) for i, j in indices]
-
-
-class SimpleMinsumMatcher(nn.Module):
-    """This class computes an assignment between the targets and the predictions of the network
-    For efficiency reasons, the targets don't include the no_object. Because of this, in general,
-    there are more predictions than targets. In this case, we do a 1-to-1 matching of the best predictions,
-    while the others are un-matched (and thus treated as non-objects).
-    """
-
-    def __init__(self, cost_class: float = 1, cost_bbox: float = 1, cost_giou: float = 1, focal_alpha = 0.25):
-        """Creates the matcher
-        Params:
-            cost_class: This is the relative weight of the classification error in the matching cost
-            cost_bbox: This is the relative weight of the L1 error of the bounding box coordinates in the matching cost
-            cost_giou: This is the relative weight of the giou loss of the bounding box in the matching cost
-        """
-        super().__init__()
-        self.cost_class = cost_class
-        self.cost_bbox = cost_bbox
-        self.cost_giou = cost_giou
-        assert cost_class != 0 or cost_bbox != 0 or cost_giou != 0, "all costs cant be 0"
-
-        self.focal_alpha = focal_alpha
-
-    @torch.no_grad()
-    def forward(self, outputs, targets):
-        """ Performs the matching
-        Params:
-            outputs: This is a dict that contains at least these entries:
-                 "pred_logits": Tensor of dim [batch_size, num_queries, num_classes] with the classification logits
-                 "pred_boxes": Tensor of dim [batch_size, num_queries, 4] with the predicted box coordinates
-            targets: This is a list of targets (len(targets) = batch_size), where each target is a dict containing:
-                 "labels": Tensor of dim [num_target_boxes] (where num_target_boxes is the number of ground-truth
-                           objects in the target) containing the class labels
-                 "boxes": Tensor of dim [num_target_boxes, 4] containing the target box coordinates
-        Returns:
-            A list of size batch_size, containing tuples of (index_i, index_j) where:
-                - index_i is the indices of the selected predictions (in order)
-                - index_j is the indices of the corresponding selected targets (in order)
-            For each batch element, it holds:
-                len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
-        """
-
-        bs, num_queries = outputs["pred_logits"].shape[:2]
-
-        # We flatten to compute the cost matrices in a batch
-        out_prob = outputs["pred_logits"].flatten(0, 1).sigmoid()  # [batch_size * num_queries, num_classes]
-        out_bbox = outputs["pred_boxes"].flatten(0, 1)  # [batch_size * num_queries, 4]
-
-        # Also concat the target labels and boxes
-        tgt_ids = torch.cat([v["labels"] for v in targets])
-        tgt_bbox = torch.cat([v["boxes"] for v in targets])
-
-        # Compute the classification cost.
-        alpha = self.focal_alpha
-        gamma = 2.0
-        neg_cost_class = (1 - alpha) * (out_prob ** gamma) * (-(1 - out_prob + 1e-8).log())
-        pos_cost_class = alpha * ((1 - out_prob) ** gamma) * (-(out_prob + 1e-8).log())
-        cost_class = pos_cost_class[:, tgt_ids] - neg_cost_class[:, tgt_ids]
-
-        # Compute the L1 cost between boxes
-        cost_bbox = torch.cdist(out_bbox, tgt_bbox, p=1)
-            
-        # Compute the giou cost betwen boxes            
-        cost_giou = -generalized_box_iou(box_cxcywh_to_xyxy(out_bbox), box_cxcywh_to_xyxy(tgt_bbox))
-
-        # Final cost matrix
-        
-        C = self.cost_bbox * cost_bbox + self.cost_class * cost_class + self.cost_giou * cost_giou
-        C = C.view(bs, num_queries, -1)
-
-        sizes = [len(v["boxes"]) for v in targets]
-        indices = []
-        device = C.device
-        for i, (c, _size) in enumerate(zip(C.split(sizes, -1), sizes)):
-            weight_mat = c[i]
-            idx_i = weight_mat.min(0)[1]
-            idx_j = torch.arange(_size).to(device)
-            indices.append((idx_i, idx_j))
-
-        return [(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64)) for i, j in indices]
-
-
-def build_matcher(args):
-    assert args.matcher_type in ['HungarianMatcher', 'SimpleMinsumMatcher'], "Unknown args.matcher_type: {}".format(args.matcher_type)
-    if args.matcher_type == 'HungarianMatcher':
-        return HungarianMatcher(
-            cost_class=args.set_cost_class, cost_bbox=args.set_cost_bbox, cost_giou=args.set_cost_giou,
-            focal_alpha=args.focal_alpha
-        )
-    elif args.matcher_type == 'SimpleMinsumMatcher':
-        return SimpleMinsumMatcher(
-            cost_class=args.set_cost_class, cost_bbox=args.set_cost_bbox, cost_giou=args.set_cost_giou,
-            focal_alpha=args.focal_alpha
-        )    
-    else:
-        raise NotImplementedError("Unknown args.matcher_type: {}".format(args.matcher_type))

groundingdino/models/GroundingDINO/.ipynb_checkpoints/ms_deform_attn-checkpoint.py

@@ -1,416 +0,0 @@
-# ------------------------------------------------------------------------
-# 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_
-import loralib as lora
-
-try:
-    # from groundingdino import _C
-    import MultiScaleDeformableAttention as _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
-        r = 12
-        self.sampling_offsets = lora.Linear(embed_dim, num_heads * num_levels * num_points * 2 , r=r )
-        self.attention_weights = lora.Linear(embed_dim, num_heads * num_levels * num_points , r=r)
-        self.value_proj = lora.Linear(embed_dim, embed_dim , r=r)
-        self.output_proj = lora.Linear(embed_dim, embed_dim , r=r)
-
-        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

groundingdino/models/GroundingDINO/.ipynb_checkpoints/transformer-checkpoint.py

@@ -1,969 +0,0 @@
-# ------------------------------------------------------------------------
-# 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 groundingdino.util.misc import inverse_sigmoid
-import loralib as lora
-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
-            r = 24
-            self.enc_output = lora.Linear(d_model, d_model , r=r)
-            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":  #把encoder的输出作为proposal
-            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
-        #########################################################
-
-        #memory  torch.Size([2, 16320, 256])
-
-        # import pdb;pdb.set_trace()
-        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))
-
-        # import pdb;pdb.set_trace()
-
-        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)
-        r =12
-        # ffn
-        self.linear1 = lora.Linear(d_model, d_ffn , r=r )
-        self.activation = _get_activation_fn(activation, d_model=d_ffn)
-        self.dropout2 = nn.Dropout(dropout)
-        self.linear2 = lora.Linear(d_ffn, d_model , r=r)
-        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
-        r = 12
-        self.linear1 = lora.Linear(d_model, d_ffn , r=r)
-        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 = lora.Linear(d_ffn, d_model , r=r )
-        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,
-    )

groundingdino/models/GroundingDINO/.ipynb_checkpoints/transformer_vanilla-checkpoint.py

@@ -1,125 +0,0 @@
-# ------------------------------------------------------------------------
-# 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
-import loralib as lora
-
-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)
-        r = 12
-        # Implementation of Feedforward model
-        self.linear1 = lora.Linear(d_model, dim_feedforward , r=r)
-        self.dropout = nn.Dropout(dropout)
-        self.linear2 = lora.Linear(dim_feedforward, d_model , r=r)
-
-        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

groundingdino/models/GroundingDINO/.ipynb_checkpoints/utils-checkpoint.py

@@ -1,274 +0,0 @@
-# ------------------------------------------------------------------------
-# 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
-import loralib as lora
-
-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__()
-        r = 12
-        self.num_layers = num_layers
-        h = [hidden_dim] * (num_layers - 1)
-        self.layers = nn.ModuleList(
-            lora.Linear(n, k , r=r) 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)
-        # print(x)  #torch.Size([2, 16320, 256])
-        # print(text_dict)
-
-        # import pdb;pdb.set_trace()
-        y = text_dict["encoded_text"]  #torch.Size([2, 195, 256])
-        text_token_mask = text_dict["text_token_mask"]
-
-        res = x @ y.transpose(-1, -2)
-        res.masked_fill_(~text_token_mask[:, None, :], float("-inf"))
-        # 接着，对res进行掩码操作，将未使用的文本token（即padding的token）对应的得分置为负无穷float("-inf")。这是为了在计算相似度时，排除padding部分的影响。
-
-
-        # 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  #torch.Size([2, 16320, 195])
-
-        return new_res

groundingdino/models/GroundingDINO/backbone/.ipynb_checkpoints/__init__-checkpoint.py

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

groundingdino/models/GroundingDINO/backbone/.ipynb_checkpoints/backbone-checkpoint.py

@@ -1,221 +0,0 @@
-# ------------------------------------------------------------------------
-# 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 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

groundingdino/models/GroundingDINO/backbone/.ipynb_checkpoints/position_encoding-checkpoint.py

@@ -1,186 +0,0 @@
-# ------------------------------------------------------------------------
-# 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 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

groundingdino/models/GroundingDINO/backbone/.ipynb_checkpoints/swin_transformer-checkpoint.py

@@ -1,804 +0,0 @@
-# ------------------------------------------------------------------------
-# 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_
-import loralib as lora
-from 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__()
-        r = 12
-        out_features = out_features or in_features
-        hidden_features = hidden_features or in_features
-        self.fc1 = lora.Linear(in_features, hidden_features , r=r)
-        self.act = act_layer()
-        self.fc2 = lora.Linear(hidden_features, out_features , r=r)
-        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)
-        r = 12
-        self.qkv = lora.Linear(dim, dim * 3, r=r ,  bias=qkv_bias)
-        self.attn_drop = nn.Dropout(attn_drop)
-        self.proj = lora.Linear(dim, dim , r=r)
-        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__()
-        r = 24
-        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)