data generation

data_generation/add_all_json.py

+import json
+from argparse import ArgumentParser
+def open_json(path):
+    with open(path,"r") as f:
+        data=json.load(f)
+    return data
+def save_json(json_list,save_path):
+    with open(save_path, 'w') as file:
+        json.dump(json_list, file, indent=4)
+def caculate_IOU(box1,box2):
+    Ax1=box1[0]
+    Ay1=box1[1]
+    Ax2=box1[2]
+    Ay2=box1[3]
+
+    Bx1=box2[0]
+    By1=box2[1]
+    Bx2=box2[2]
+    By2=box2[3]
+
+    Ix1 = max(Ax1, Bx1)
+    Iy1 = max(Ay1, By1)
+    Ix2 = min(Ax2, Bx2)
+    Iy2 = min(Ay2, By2)
+    IntersectionArea = max(0, Ix2 - Ix1 + 1) * max(0, Iy2 - Iy1 + 1)
+    BoxAArea = (Ax2 - Ax1 + 1) * (Ay2 - Ay1 + 1)
+    BoxBArea = (Bx2 - Bx1 + 1) * (By2 - By1 + 1)
+    UnionArea = BoxAArea + BoxBArea - IntersectionArea
+    IOU = IntersectionArea / UnionArea
+    return IOU
+def _get_args():
+    parser = ArgumentParser()
+    parser.add_argument("--blip2_caption", type=str, default="./outputs/blip2_cap.json")
+    parser.add_argument("--ori_caption", type=str, default=None)
+    parser.add_argument("--grit", type=str, default="./outputs/grit_score.json")
+    parser.add_argument("--ppocr", type=str, default="./outputs/ppocr.json")
+    parser.add_argument("--sam_blip2", type=str, default="./outputs/sam_blip2_score.json")
+    parser.add_argument("--output", type=str, default="./outputs/ann_all.json")
+    args = parser.parse_args()
+    return args
+if __name__=="__main__":
+    json_save = []
+    args = _get_args()
+    if args.ori_caption is not None:
+        ori_cap = open_json(args.ori_caption)
+    blip2_caption = open_json(args.blip2_caption)
+    grit = open_json(args.grit)
+    ppocr = open_json(args.ppocr)
+    sam_blip2 = open_json(args.sam_blip2)
+    
+    blip2_caption_dict = {}
+    grit_dict = {}
+    ppocr_dict = {}
+    sam_blip2_dict = {}
+    for i in range(len(blip2_caption)):
+        img_id = blip2_caption[i]['img_id']
+        caption = blip2_caption[i]['blip2_caption']
+        blip2_caption_dict[img_id]=caption
+    for i in range(len(grit)):
+        img_id = grit[i]['img_id']
+        objects = grit[i]['objects']
+        caption = ""
+        for j in range(len(objects)):
+            if objects[j]['score']>0.4:
+                caption = caption + f"{objects[j]['caption']}: {objects[j]['box']}; "
+        grit_dict[img_id] = caption
+    
+    for i in range(len(ppocr)):
+        img_id = ppocr[i]['img_id']
+        objects = ppocr[i]['objects']
+        caption = ""
+        for j in range(len(objects)):
+            if objects[j]['score']>0.85:
+                caption = caption + f"{objects[j]['caption']}: {objects[j]['box']}; "
+        ppocr_dict[img_id] = caption
+    
+    for i in range(len(sam_blip2)):
+        img_id = sam_blip2[i]['img_id']
+        objects = sam_blip2[i]['objects']
+        caption = ""
+        iou_filter = {}
+        for j in range(len(objects)):
+            if objects[j]['score']>0.5:
+                if iou_filter.get(objects[j]['caption'], 0)==0:
+                    iou_filter[objects[j]['caption']]=objects[j]['box']
+                    caption = caption + f"{objects[j]['caption']}: {objects[j]['box']}; "
+                else:
+                    if caculate_IOU(iou_filter[objects[j]['caption']], objects[j]['box'])<0.6:
+                        caption = caption + f"{objects[j]['caption']}: {objects[j]['box']}; "
+        sam_blip2_dict[img_id] = caption
+    for key in blip2_caption_dict.keys():
+        ocr_result = ppocr_dict.get(key,"")
+        json_save.append({"img_id":key, "blip2cap":blip2_caption_dict[key], "grit":grit_dict[key], "ocr":ocr_result, "sam_blip":sam_blip2_dict[key]})
+    save_json(json_save, args.output)
\ No newline at end of file

data_generation/amg.py

+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import cv2  # type: ignore
+
+from segment_anything import SamAutomaticMaskGenerator, sam_model_registry
+
+import argparse
+import json
+import os
+from typing import Any, Dict, List
+from tqdm import tqdm
+from PIL import Image
+parser = argparse.ArgumentParser(
+    description=(
+        "Runs automatic mask generation on an input image or directory of images, "
+        "and outputs masks as either PNGs or COCO-style RLEs. Requires open-cv, "
+        "as well as pycocotools if saving in RLE format."
+    )
+)
+
+parser.add_argument(
+    "--input",
+    type=str,
+    required=True,
+    help="Path to either a single input image or folder of images.",
+)
+
+parser.add_argument(
+    "--output",
+    type=str,
+    required=True,
+    help=(
+        "Path to the directory where masks will be output. Output will be either a folder "
+        "of PNGs per image or a single json with COCO-style masks."
+    ),
+)
+
+parser.add_argument(
+    "--model-type",
+    type=str,
+    required=True,
+    help="The type of model to load, in ['default', 'vit_h', 'vit_l', 'vit_b']",
+)
+
+parser.add_argument(
+    "--checkpoint",
+    type=str,
+    required=True,
+    help="The path to the SAM checkpoint to use for mask generation.",
+)
+
+parser.add_argument("--device", type=str, default="cuda", help="The device to run generation on.")
+
+parser.add_argument(
+    "--convert-to-rle",
+    action="store_true",
+    help=(
+        "Save masks as COCO RLEs in a single json instead of as a folder of PNGs. "
+        "Requires pycocotools."
+    ),
+)
+
+amg_settings = parser.add_argument_group("AMG Settings")
+
+amg_settings.add_argument(
+    "--points-per-side",
+    type=int,
+    default=None,
+    help="Generate masks by sampling a grid over the image with this many points to a side.",
+)
+
+amg_settings.add_argument(
+    "--points-per-batch",
+    type=int,
+    default=None,
+    help="How many input points to process simultaneously in one batch.",
+)
+
+amg_settings.add_argument(
+    "--pred-iou-thresh",
+    type=float,
+    default=None,
+    help="Exclude masks with a predicted score from the model that is lower than this threshold.",
+)
+
+amg_settings.add_argument(
+    "--stability-score-thresh",
+    type=float,
+    default=None,
+    help="Exclude masks with a stability score lower than this threshold.",
+)
+
+amg_settings.add_argument(
+    "--stability-score-offset",
+    type=float,
+    default=None,
+    help="Larger values perturb the mask more when measuring stability score.",
+)
+
+amg_settings.add_argument(
+    "--box-nms-thresh",
+    type=float,
+    default=None,
+    help="The overlap threshold for excluding a duplicate mask.",
+)
+
+amg_settings.add_argument(
+    "--crop-n-layers",
+    type=int,
+    default=None,
+    help=(
+        "If >0, mask generation is run on smaller crops of the image to generate more masks. "
+        "The value sets how many different scales to crop at."
+    ),
+)
+
+amg_settings.add_argument(
+    "--crop-nms-thresh",
+    type=float,
+    default=None,
+    help="The overlap threshold for excluding duplicate masks across different crops.",
+)
+
+amg_settings.add_argument(
+    "--crop-overlap-ratio",
+    type=int,
+    default=None,
+    help="Larger numbers mean image crops will overlap more.",
+)
+
+amg_settings.add_argument(
+    "--crop-n-points-downscale-factor",
+    type=int,
+    default=None,
+    help="The number of points-per-side in each layer of crop is reduced by this factor.",
+)
+
+amg_settings.add_argument(
+    "--min-mask-region-area",
+    type=int,
+    default=None,
+    help=(
+        "Disconnected mask regions or holes with area smaller than this value "
+        "in pixels are removed by postprocessing."
+    ),
+)
+amg_settings.add_argument(
+    "--rank",
+    type=int,
+    default=0,
+)
+
+
+def write_masks_to_folder(masks: List[Dict[str, Any]], path: str) -> None:
+    header = "id,area,bbox_x0,bbox_y0,bbox_w,bbox_h,point_input_x,point_input_y,predicted_iou,stability_score,crop_box_x0,crop_box_y0,crop_box_w,crop_box_h"  # noqa
+    metadata = [header]
+    for i, mask_data in enumerate(masks):
+        mask = mask_data["segmentation"]
+        filename = f"{i}.png"
+        cv2.imwrite(os.path.join(path, filename), mask * 255)
+        mask_metadata = [
+            str(i),
+            str(mask_data["area"]),
+            *[str(x) for x in mask_data["bbox"]],
+            *[str(x) for x in mask_data["point_coords"][0]],
+            str(mask_data["predicted_iou"]),
+            str(mask_data["stability_score"]),
+            *[str(x) for x in mask_data["crop_box"]],
+        ]
+        row = ",".join(mask_metadata)
+        metadata.append(row)
+    metadata_path = os.path.join(path, "metadata.csv")
+    with open(metadata_path, "w") as f:
+        f.write("\n".join(metadata))
+    return
+
+
+def get_amg_kwargs(args):
+    amg_kwargs = {
+        "points_per_side": args.points_per_side,
+        "points_per_batch": args.points_per_batch,
+        "pred_iou_thresh": args.pred_iou_thresh,
+        "stability_score_thresh": args.stability_score_thresh,
+        "stability_score_offset": args.stability_score_offset,
+        "box_nms_thresh": args.box_nms_thresh,
+        "crop_n_layers": args.crop_n_layers,
+        "crop_nms_thresh": args.crop_nms_thresh,
+        "crop_overlap_ratio": args.crop_overlap_ratio,
+        "crop_n_points_downscale_factor": args.crop_n_points_downscale_factor,
+        "min_mask_region_area": args.min_mask_region_area,
+    }
+    amg_kwargs = {k: v for k, v in amg_kwargs.items() if v is not None}
+    return amg_kwargs
+
+def main(args: argparse.Namespace, image_list: List) -> None:
+    print("Loading model...")
+    sam = sam_model_registry[args.model_type](checkpoint=args.checkpoint)
+    _ = sam.to(device=f"cuda:{args.rank}")
+    output_mode = "coco_rle" if args.convert_to_rle else "binary_mask"
+    amg_kwargs = get_amg_kwargs(args)
+    generator = SamAutomaticMaskGenerator(sam, output_mode=output_mode, **amg_kwargs)
+    targets=image_list
+    num_empty=0
+    empty_list=[]
+    for i in tqdm(range(len(targets))):
+        t=targets[i]
+        if os.path.exists(t.replace('images', 'masks').replace('.jpg','.json')):
+            with open(t.replace('images', 'masks').replace('.jpg','.json'), 'r') as f:
+                data = json.load(f)
+            if data==[]:
+                num_empty+=1
+                empty_list.append(t)
+                print(f"{t} is empty process again")
+            else:
+                print(f"Exist '{t}'...")
+        print(f"Processing '{t}'...")
+        image = cv2.imread(t)
+        if image is None:
+            print(f"Could not load '{t}' as an image, skipping...")
+            continue
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+
+        masks = generator.generate(image)
+        name=t.split('/')[-1].split('.')[0]
+        save_base=os.path.join(args.output, name)
+        if output_mode == "binary_mask":
+            os.makedirs(save_base, exist_ok=False)
+            write_masks_to_folder(masks, save_base)
+        else:
+            save_file = save_base + ".json"
+            with open(save_file, "w") as f:
+                json.dump(masks, f)
+    print(f"empyt num {num_empty}")
+    print(empty_list)
+    print("Done!")
+def get_image_files(folder_path):  
+    image_files = []  
+    for root, dirs, files in os.walk(folder_path):  
+        for file in files:  
+            if file.endswith('.jpg') or file.endswith('.png'):  
+                image_files.append(os.path.join(root, file))  
+    return image_files
+if __name__ == "__main__":
+    args = parser.parse_args()
+    image_list=get_image_files(args.input)
+    print(f"len_image: {len(image_list)}")
+    main(args, image_list)
\ No newline at end of file

data_generation/blip2.py

+import torch
+from PIL import Image
+from lavis.models import load_model_and_preprocess
+from argparse import ArgumentParser
+import os
+from torch.utils.data import Dataset, DataLoader
+from tqdm import tqdm
+import torchvision.utils as vutils
+import json
+def save_json(json_list,save_path):
+    with open(save_path, 'w') as file:
+        json.dump(json_list, file, indent=4)
+def get_image_files(folder_path):  
+    image_files = []  
+    for root, dirs, files in os.walk(folder_path):  
+        for file in files:  
+            if file.endswith('.jpg') or file.endswith('.png'):  
+                image_files.append(os.path.join(root, file))  
+    return image_files
+def _get_args():
+    parser = ArgumentParser()
+    parser.add_argument("--image_folder", type=str, default="./images")
+    parser.add_argument("--output_path", type=str, default="./outputs/blip2_cap.json")
+    parser.add_argument("--batch_size", type=int, default=2)
+    parser.add_argument("--device", type=str, default="cuda:0")
+    args = parser.parse_args()
+    return args
+class lazydataset(Dataset):
+    def __init__(self, data_path, processor) -> None:
+        super(lazydataset).__init__()
+        self.image_paths = get_image_files(data_path)
+        self.processor = processor
+    def __len__(self):
+        return len(self.image_paths)
+    def __getitem__(self, i):
+        image_path = self.image_paths[i]
+        raw_image = Image.open(image_path).convert('RGB')
+        image = self.processor["eval"](raw_image)
+        return {'image':image, 'img_id': image_path.split('/')[-1]}
+def collate_fn(batch):
+    image = [item['image'].squeeze(0) for item in batch]
+    image = torch.stack(image)
+    img_id = [item['img_id'] for item in batch]
+    return {'image':image, 'img_id':img_id}
+
+if __name__=="__main__":
+    json_save = []
+    args = _get_args()
+    device = args.device
+    model, vis_processors, _ = load_model_and_preprocess(name="blip2_opt", model_type="pretrain_opt2.7b", is_eval=True, device=device)
+    dataset = lazydataset(data_path=args.image_folder, processor = vis_processors)
+    dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=False, num_workers=0, collate_fn=collate_fn)
+    for batch in tqdm(dataloader):
+        image = batch['image'].to(device)
+        captions = model.generate({"image": image})
+        img_id = batch['img_id']
+        for i in range(len(img_id)):
+            json_save.append({'img_id':img_id[i],'blip2_caption':captions[i]})
+    save_json(json_save, args.output_path)

data_generation/chatgpt.py

+import json
+import openai
+from argparse import ArgumentParser
+#You can replace this step with Open Source LMM
+openai.api_base = ""
+openai.api_key = ''
+def save_json(json_list,save_path):
+    with open(save_path, 'w') as file:
+        json.dump(json_list, file, indent=4)
+def _get_args():
+    parser = ArgumentParser()
+    parser.add_argument("--ann_path", type=str, default="./outputs/ann_all.json")
+    parser.add_argument("--output_path", type=str, default="./outputs/detailed_caption.json")
+    args = parser.parse_args()
+    return args
+
+def get_question(ann):
+    sentence1 = "I want you to act as an intelligent image captioner. You should generate a descriptive, coherent and logical description of the image based on the given descriptions from different people for the same image. The position is represented by normalized top-left and bottom-right coordinates."
+    sentence2 = "\n Overall Image Caption: "
+    sentece3 = "\n Dense Caption1 (Region Description: Area Location): "
+    sentece4 = "\n Dense Caption2 (Region Description: Area Location): "
+    sentece5 = "\n Texts in the image: "
+    sentence6 = "\n There are some rules for your response: Provide context of the image. \n Merging the descriptions of the same object at the same position and the texts belonging to the same sentence. \n Show main objects with their attributes (e.g. position, color, shape). \n Show relative position between main objects. \n Less than 6 sentences. \n Do not show any numbers or coordinates. \n Do not describe any individual letter."
+    
+    if ann['ocr']!="":
+        question = f"{sentence1}{sentence2}{ann['blip2cap']}{sentece3}{ann['grit']}{sentece4}{ann['sam_blip']}{sentece5}{ann['ocr']}{sentence6}"
+    else:
+        question = f"{sentence1}{sentence2}{ann['blip2cap']}{sentece3}{ann['grit']}{sentece4}{ann['sam_blip']}{sentence6}"
+    return question
+
+if __name__=="__main__":
+    args = _get_args()
+    json_save=[]
+    with open(args.ann_path,"r") as f:
+        data = json.load(f)
+    
+    for i in range(len(data)):
+        question = get_question(data[i])
+        response = openai.ChatCompletion.create(
+            model="gpt-3.5-turbo", #gpt-4-1106-preview
+            messages=[
+            {"role": "user", "content": question},
+            ]
+        )
+        answer = response['choices'][0]['message']['content']
+        json_save.append({"img_id":data[i]['img_id'],"detailed_caption":answer})
+    save_json(json_save, args.output_path)
\ No newline at end of file

data_generation/grit.py

+import argparse
+import multiprocessing as mp
+import os
+import time
+import cv2
+from tqdm import tqdm
+import sys
+from detectron2.config import get_cfg
+from detectron2.data.detection_utils import read_image
+from detectron2.utils.logger import setup_logger
+sys.path.insert(0, 'grit/third_party/CenterNet2/projects/CenterNet2/')
+sys.path.append('./grit')
+from centernet.config import add_centernet_config
+from grit.config import add_grit_config
+from grit.predictor import VisualizationDemo, BatchVisualizationDemo
+from torch.utils.data import Dataset, DataLoader
+from PIL import Image
+import json
+
+def save_json(json_list,save_path):
+    with open(save_path, 'w') as file:
+        json.dump(json_list, file, indent=4)
+def get_image_files(folder_path):  
+    image_files = []  
+    for root, dirs, files in os.walk(folder_path):  
+        for file in files:  
+            if file.endswith('.jpg') or file.endswith('.png'):  
+                image_files.append(os.path.join(root, file))  
+    return image_files
+
+# constants
+WINDOW_NAME = "GRiT"
+
+def norm_xy(xy, width, height):
+    xy[0]=round(xy[0]/width,3)
+    xy[2]=round(xy[2]/width,3)
+    xy[1]=round(xy[1]/height,3)
+    xy[3]=round(xy[3]/height,3)
+    return xy
+def dense_pred_to_normcaption(predictions, width, height):
+    boxes = predictions["instances"].pred_boxes if predictions["instances"].has("pred_boxes") else None
+    object_description = predictions["instances"].pred_object_descriptions.data
+    objects = []
+    for i in range(len(object_description)):
+        xy = [a for a in boxes[i].tensor.cpu().detach().numpy()[0]]
+        box = norm_xy(xy, width, height)
+        objects.append({"caption":object_description[i],"box":box})
+    return objects
+
+def setup_cfg(args):
+    cfg = get_cfg()
+    if args.cpu:
+        cfg.MODEL.DEVICE="cpu"
+    cfg.MODEL.DEVICE=args.device
+    add_centernet_config(cfg)
+    add_grit_config(cfg)
+    cfg.merge_from_file(args.config_file)
+    cfg.merge_from_list(args.opts)
+    # Set score_threshold for builtin models
+    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = args.confidence_threshold
+    cfg.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH = args.confidence_threshold
+    if args.test_task:
+        cfg.MODEL.TEST_TASK = args.test_task
+    cfg.MODEL.BEAM_SIZE = 1
+    cfg.MODEL.ROI_HEADS.SOFT_NMS_ENABLED = False
+    cfg.USE_ACT_CHECKPOINT = False
+    cfg.freeze()
+    return cfg
+
+
+def _get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--config-file",
+        default="./grit/configs/GRiT_B_DenseCap_ObjectDet.yaml",
+        metavar="FILE",
+        help="path to config file",
+    )
+    parser.add_argument("--cpu", type=bool, default=False)
+    parser.add_argument("--device", type=str, default="cuda:0")
+    parser.add_argument(
+        "--confidence-threshold",
+        type=float,
+        default=0.5,
+        help="Minimum score for instance predictions to be shown",
+    )
+    parser.add_argument(
+        "--test-task",
+        type=str,
+        default='DenseCap',
+        help="Choose a task to have GRiT perform",
+    )
+    parser.add_argument(
+        "--opts",
+        help="Modify config options using the command-line 'KEY VALUE' pairs",
+        default=["MODEL.WEIGHTS", "./grit/model_weight/grit_b_densecap.pth"],
+        nargs=argparse.REMAINDER,
+    )
+    parser.add_argument("--image_folder", type=str, default="./images")
+    parser.add_argument("--output_path", type=str, default="./outputs/grit.json")
+    parser.add_argument("--batch_size", type=int, default=2)
+    args = parser.parse_args()
+    return args
+
+class lazydataset(Dataset):
+    def __init__(self, data_path) -> None:
+        super(lazydataset).__init__()
+        self.image_paths = get_image_files(data_path)
+    def __len__(self):
+        return len(self.image_paths)
+    def __getitem__(self, i):
+        image_path = self.image_paths[i]
+        image = read_image(image_path, format="BGR")
+        return {'image':image, 'img_id': image_path.split('/')[-1]}
+def collate_fn(batch):
+    image = [item['image'] for item in batch]
+    img_id = [item['img_id'] for item in batch]
+    return {'image':image, 'img_id':img_id}
+if __name__ == "__main__":
+    json_save=[]
+    args = _get_args()
+    cfg = setup_cfg(args)
+    demo = BatchVisualizationDemo(cfg)
+    dataset=lazydataset(args.image_folder)
+    dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=False, num_workers=0,collate_fn=collate_fn)
+    for batch in tqdm(dataloader):
+        predictions = demo.run_on_images(batch['image'])
+        for i in range(len(predictions)):
+            height, width = batch['image'][i].shape[0], batch['image'][i].shape[1]
+            objects = dense_pred_to_normcaption(predictions[i], width, height)
+            json_save.append({"img_id":batch['img_id'][i], "objects":objects})
+    save_json(json_save, args.output_path)
\ No newline at end of file

data_generation/grit/code_source.md

+These codes come from https://github.com/JialianW/GRiT.git.
\ No newline at end of file

data_generation/grit/configs/Base.yaml

+MODEL:
+  META_ARCHITECTURE: "GRiT"
+  MASK_ON: True
+  PROPOSAL_GENERATOR:
+    NAME: "CenterNet"
+  FPN:
+    IN_FEATURES: ["layer3", "layer4", "layer5"]
+  PIXEL_MEAN: [123.675, 116.280, 103.530]
+  PIXEL_STD: [58.395, 57.12, 57.375]
+  ROI_HEADS:
+    NAME: GRiTROIHeadsAndTextDecoder
+    IN_FEATURES: ["p3", "p4", "p5"]
+    IOU_THRESHOLDS: [0.6]
+    NUM_CLASSES: 1
+    SCORE_THRESH_TEST: 0.02
+    NMS_THRESH_TEST: 0.5
+    OBJECT_FEAT_POOLER_RES: 14
+  ROI_BOX_CASCADE_HEAD:
+    IOUS: [0.6, 0.7, 0.8]
+  ROI_BOX_HEAD:
+    NAME: "FastRCNNConvFCHead"
+    NUM_FC: 2
+    POOLER_RESOLUTION: 7
+    CLS_AGNOSTIC_BBOX_REG: True
+    MULT_PROPOSAL_SCORE: True
+  ROI_MASK_HEAD:
+    NAME: "MaskRCNNConvUpsampleHead"
+    NUM_CONV: 4
+    POOLER_RESOLUTION: 14
+    CLS_AGNOSTIC_MASK: True
+  CENTERNET:
+    NUM_CLASSES: 1
+    REG_WEIGHT: 1.
+    NOT_NORM_REG: True
+    ONLY_PROPOSAL: True
+    WITH_AGN_HM: True
+    INFERENCE_TH: 0.0001
+    PRE_NMS_TOPK_TRAIN: 4000
+    POST_NMS_TOPK_TRAIN: 2000
+    PRE_NMS_TOPK_TEST: 1000
+    POST_NMS_TOPK_TEST: 256
+    NMS_TH_TRAIN: 0.9
+    NMS_TH_TEST: 0.9
+    POS_WEIGHT: 0.5
+    NEG_WEIGHT: 0.5
+    IGNORE_HIGH_FP: 0.85
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+DATALOADER:
+  SAMPLER_TRAIN: "MultiDatasetSampler"
+  DATASET_RATIO: [1]
+  DATASET_INPUT_SIZE: [1024]
+  DATASET_INPUT_SCALE: [[0.1, 2.0]]
+  FILTER_EMPTY_ANNOTATIONS: False
+  NUM_WORKERS: 8
+TEST:
+  DETECTIONS_PER_IMAGE: 256
+SOLVER:
+  LR_SCHEDULER_NAME: "WarmupCosineLR"
+  CHECKPOINT_PERIOD: 10000
+  WARMUP_ITERS: 1000
+  WARMUP_FACTOR: 0.001
+  USE_CUSTOM_SOLVER: True
+  OPTIMIZER: "ADAMW"
+  MAX_ITER: 180000
+  IMS_PER_BATCH: 64
+  BASE_LR: 0.00008
+  VIT_LAYER_DECAY: True
+  CLIP_GRADIENTS:
+    ENABLED: True
+INPUT:
+  FORMAT: RGB
+  CUSTOM_AUG: EfficientDetResizeCrop
+  TRAIN_SIZE: 640
+USE_ACT_CHECKPOINT: True
+VERSION: 2
\ No newline at end of file

data_generation/grit/configs/GRiT_B_DenseCap.yaml

+_BASE_: "Base.yaml"
+MODEL:
+  TRAIN_TASK: ["DenseCap"]
+  TEST_TASK: "DenseCap"
+  MASK_ON: False
+  ROI_HEADS:
+    SOFT_NMS_ENABLED: False
+  BEAM_SIZE: 1
+  WEIGHTS: "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_base.pth"
+  BACKBONE:
+    NAME: build_vit_fpn_backbone
+  VIT_LAYERS: 12
+SOLVER:
+  VIT_LAYER_DECAY_RATE: 0.7
+DATASETS:
+  TRAIN: ("vg_train",)
+  TEST: ("vg_test",)
+DATALOADER:
+  DATASET_BS: 2
+OUTPUT_DIR: "./output/GRiT_B_DenseCap"
\ No newline at end of file

data_generation/grit/configs/GRiT_B_DenseCap_ObjectDet.yaml

+_BASE_: "Base.yaml"
+MODEL:
+  TRAIN_TASK: ["ObjectDet", "DenseCap"]
+  TEST_TASK: "DenseCap" # DenseCap or ObjectDet: Choose one for testing
+  MASK_ON: True
+  ROI_HEADS:
+    SOFT_NMS_ENABLED: False
+  BEAM_SIZE: 1
+  WEIGHTS: "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_base.pth"
+  BACKBONE:
+    NAME: build_vit_fpn_backbone
+  VIT_LAYERS: 12
+SOLVER:
+  VIT_LAYER_DECAY_RATE: 0.7
+DATASETS:
+  TRAIN: ("GRiT_coco2017_train", "vg_train")
+  TEST: ("coco_2017_test-dev",)
+DATALOADER:
+  DATASET_RATIO: [1, 1]
+  DATASET_BS: 2
+  DATASET_INPUT_SIZE: [1024, 1024]
+  DATASET_INPUT_SCALE: [[0.1, 2.0], [0.1, 2.0]]
+OUTPUT_DIR: "./output/GRiT_B_DenseCap_ObjectDet"
\ No newline at end of file

data_generation/grit/configs/GRiT_B_ObjectDet.yaml

+_BASE_: "Base.yaml"
+MODEL:
+  TRAIN_TASK: ["ObjectDet"]
+  TEST_TASK: "ObjectDet"
+  MASK_ON: True
+  ROI_HEADS:
+    SOFT_NMS_ENABLED: True
+  BEAM_SIZE: 3
+  WEIGHTS: "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_base.pth"
+  BACKBONE:
+    NAME: build_vit_fpn_backbone
+  VIT_LAYERS: 12
+SOLVER:
+  VIT_LAYER_DECAY_RATE: 0.7
+DATASETS:
+  TRAIN: ("GRiT_coco2017_train",)
+  TEST: ("coco_2017_val",)
+DATALOADER:
+  DATASET_BS: 2
+OUTPUT_DIR: "./output/GRiT_B_ObjectDet"
\ No newline at end of file

data_generation/grit/configs/GRiT_H_ObjectDet.yaml

+_BASE_: "Base.yaml"
+MODEL:
+  TRAIN_TASK: ["ObjectDet"]
+  TEST_TASK: "ObjectDet"
+  MASK_ON: True
+  ROI_HEADS:
+    SOFT_NMS_ENABLED: True
+  BEAM_SIZE: 3
+  WEIGHTS: "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_huge_p14to16.pth"
+  BACKBONE:
+    NAME: build_vit_fpn_backbone_huge
+  VIT_LAYERS: 32
+SOLVER:
+  MAX_ITER: 135000
+  VIT_LAYER_DECAY_RATE: 0.9
+DATASETS:
+  TRAIN: ("GRiT_coco2017_train",)
+  TEST: ("coco_2017_val",)
+DATALOADER:
+  DATASET_BS: 1
+OUTPUT_DIR: "./output/GRiT_H_ObjectDet"
\ No newline at end of file

data_generation/grit/configs/GRiT_L_ObjectDet.yaml

+_BASE_: "Base.yaml"
+MODEL:
+  TRAIN_TASK: ["ObjectDet"]
+  TEST_TASK: "ObjectDet"
+  MASK_ON: True
+  ROI_HEADS:
+    SOFT_NMS_ENABLED: True
+  BEAM_SIZE: 3
+  WEIGHTS: "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_large.pth"
+  BACKBONE:
+    NAME: build_vit_fpn_backbone_large
+  VIT_LAYERS: 24
+SOLVER:
+  VIT_LAYER_DECAY_RATE: 0.8
+DATASETS:
+  TRAIN: ("GRiT_coco2017_train",)
+  TEST: ("coco_2017_val",)
+DATALOADER:
+  DATASET_BS: 1
+OUTPUT_DIR: "./output/GRiT_L_ObjectDet"
\ No newline at end of file

data_generation/grit/grit/__init__.py

+from .modeling.meta_arch import grit
+from .modeling.roi_heads import grit_roi_heads
+from .modeling.backbone import vit
+
+from .data.datasets import object365
+from .data.datasets import vg
+from .data.datasets import grit_coco
\ No newline at end of file