import argparse import torch import os import json from tqdm import tqdm IMAGE_TOKEN_INDEX = -200 DEFAULT_IMAGE_TOKEN = "" DEFAULT_IMAGE_PATCH_TOKEN = "" DEFAULT_IM_START_TOKEN = "" DEFAULT_IM_END_TOKEN = "" IMAGE_PLACEHOLDER = "" # Added by Ferret DEFAULT_REGION_FEA_TOKEN = "" VOCAB_IMAGE_W = 1000 VOCAB_IMAGE_H = 1000 from conversation import conv_templates, SeparatorStyle from builder import load_pretrained_model from mm_utils import tokenizer_image_token, process_images from PIL import Image import math import pdb import numpy as np from copy import deepcopy from functools import partial def disable_torch_init(): """ Disable the redundant torch default initialization to accelerate model creation. """ import torch setattr(torch.nn.Linear, "reset_parameters", lambda self: None) setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None) def split_list(lst, n): """Split a list into n (roughly) equal-sized chunks""" chunk_size = math.ceil(len(lst) / n) # integer division return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)] def get_chunk(lst, n, k): chunks = split_list(lst, n) return chunks[k] def generate_mask_for_feature(coor, raw_w, raw_h, mask=None): if mask is not None: assert mask.shape[0] == raw_w and mask.shape[1] == raw_h coor_mask = np.zeros((raw_w, raw_h)) # Assume it samples a point. if len(coor) == 2: # Define window size span = 5 # Make sure the window does not exceed array bounds x_min = max(0, coor[0] - span) x_max = min(raw_w, coor[0] + span + 1) y_min = max(0, coor[1] - span) y_max = min(raw_h, coor[1] + span + 1) coor_mask[int(x_min):int(x_max), int(y_min):int(y_max)] = 1 assert (coor_mask==1).any(), f"coor: {coor}, raw_w: {raw_w}, raw_h: {raw_h}" elif len(coor) == 4: # Box input or Sketch input. coor_mask[coor[0]:coor[2]+1, coor[1]:coor[3]+1] = 1 if mask is not None: coor_mask = coor_mask * mask coor_mask = torch.from_numpy(coor_mask) try: assert len(coor_mask.nonzero()) != 0 except: pdb.set_trace() return coor_mask def get_task_from_file(file): box_in_tasks = ['widgetcaptions', 'taperception', 'ocr', 'icon_recognition', 'widget_classification', 'example_0'] # box_out_tasks = ['widget_listing', 'find_text', 'find_icons', 'find_widget', 'conversation_interaction'] # no_box = ['screen2words', 'detailed_description', 'conversation_perception', 'gpt4'] if any(task in file for task in box_in_tasks): return 'box_in' else: return 'no_box_in' # elif any(task in file for task in box_out_tasks): # return 'box_out' # elif any(task in file for task in no_box): # return 'no_box' def get_bbox_coor(box, ratio_w, ratio_h): return box[0] * ratio_w, box[1] * ratio_h, box[2] * ratio_w, box[3] * ratio_h def get_model_name_from_path(model_path): if 'gemma' in model_path: return 'ferret_gemma' elif 'llama' or 'vicuna' in model_path: return 'ferret_llama' else: raise ValueError(f"No model matched for {model_path}") class UIData: def __init__(self, data_path, image_path, args) -> None: self.obj_list = json.load(open(data_path, 'r')) self.image_path = image_path self.args = args self._ids = range(len(self.obj_list)) self.task = get_task_from_file(data_path) @property def ids(self): return deepcopy(self._ids) def __getitem__(self, idx): i = self.obj_list[idx] # image stuff image_path_i = os.path.join(self.image_path, i['image'].split('/')[-1]) image = Image.open(image_path_i).convert('RGB') q_turn = i['conversations'][0]['value'] if "" in q_turn: prompt = q_turn.split('\n')[1] else: prompt = q_turn i['question'] = prompt i['region_masks'] = None if self.task == 'box_in': ratio_w = VOCAB_IMAGE_W * 1.0 / i['image_w'] ratio_h = VOCAB_IMAGE_H * 1.0 / i['image_h'] box = i['box_x1y1x2y2'][0][0] box_x1, box_y1, box_x2, box_y2 = box box_x1_textvocab, box_y1_textvocab, box_x2_textvocab, box_y2_textvocab = get_bbox_coor(box=box, ratio_h=ratio_h, ratio_w=ratio_w) if self.args.region_format == 'box': region_coordinate_raw = [box_x1, box_y1, box_x2, box_y2] if args.add_region_feature: i['question'] = prompt.replace('', '[{}, {}, {}, {}] {}'.format(int(box_x1_textvocab), int(box_y1_textvocab), int(box_x2_textvocab), int(box_y2_textvocab), DEFAULT_REGION_FEA_TOKEN)) generated_mask = generate_mask_for_feature(region_coordinate_raw, raw_w=i['image_w'], raw_h=i['image_h'], mask=None) i['region_masks'] = [generated_mask] else: i['question'] = prompt.replace('', '[{}, {}, {}, {}]'.format(int(box_x1_textvocab), int(box_y1_textvocab), int(box_x2_textvocab), int(box_y2_textvocab))) else: raise NotImplementedError(f'{self.args.region_format} is not supported.') return image, i, image.size def eval_model(args): # Data dataset = UIData(data_path=args.data_path, image_path=args.image_path, args=args) data_ids = dataset.ids # Model disable_torch_init() model_path = os.path.expanduser(args.model_path) model_name = get_model_name_from_path(model_path) tokenizer, model, image_processor, context_len = \ load_pretrained_model(model_path, args.model_base, model_name) chunk_data_ids = get_chunk(data_ids, args.num_chunks, args.chunk_idx) answers_folder = os.path.expanduser(args.answers_file) os.makedirs(answers_folder, exist_ok=True) answers_file = os.path.join(answers_folder, f'{args.chunk_idx}_of_{args.num_chunks}.jsonl') ans_file = open(answers_file, "w") for i, id in enumerate(tqdm(chunk_data_ids)): img, ann, image_size = dataset[id] image_path = ann['image'] qs = ann["question"] cur_prompt = qs if "" in qs: qs = qs.split('\n')[1] if model.config.mm_use_im_start_end: qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs else: qs = DEFAULT_IMAGE_TOKEN + '\n' + qs conv = conv_templates[args.conv_mode].copy() conv.append_message(conv.roles[0], qs) conv.append_message(conv.roles[1], None) prompt = conv.get_prompt() input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda() if model.config.image_aspect_ratio == "square_nocrop": image_tensor = image_processor.preprocess(img, return_tensors='pt', do_resize=True, do_center_crop=False, size=[args.image_h, args.image_w])['pixel_values'][0] elif model.config.image_aspect_ratio == "anyres": image_process_func = partial(image_processor.preprocess, return_tensors='pt', do_resize=True, do_center_crop=False, size=[args.image_h, args.image_w]) image_tensor = process_images([img], image_processor, model.config, image_process_func=image_process_func)[0] else: image_tensor = process_images([img], image_processor, model.config)[0] images = image_tensor.unsqueeze(0).to(args.data_type).cuda() region_masks = ann['region_masks'] if region_masks is not None: region_masks = [[region_mask_i.cuda().half() for region_mask_i in region_masks]] else: region_masks = None with torch.inference_mode(): model.orig_forward = model.forward model.forward = partial( model.orig_forward, region_masks=region_masks ) output_ids = model.generate( input_ids, images=images, region_masks=region_masks, image_sizes=[image_size], do_sample=True if args.temperature > 0 else False, temperature=args.temperature, top_p=args.top_p, num_beams=args.num_beams, max_new_tokens=args.max_new_tokens, use_cache=True) model.forward = model.orig_forward outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0] outputs = outputs.strip() if 'label' in ann: label = ann['label'] elif len(ann['conversations']) > 1: label = ann['conversations'][1]['value'] else: label = None ans_file.write(json.dumps({"id":ann['id'], # +1 offset "image_path":image_path, "prompt": cur_prompt, "text": outputs, "label": label, }) + "\n") ans_file.flush() ans_file.close() if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--model_path", type=str, default="facebook/opt-350m") parser.add_argument("--vision_model_path", type=str, default=None) parser.add_argument("--model_base", type=str, default=None) parser.add_argument("--image_path", type=str, default="") parser.add_argument("--data_path", type=str, default="") parser.add_argument("--answers_file", type=str, default="") parser.add_argument("--conv_mode", type=str, default="ferret_gemma_instruct", help="[ferret_gemma_instruct,ferret_llama_3,ferret_vicuna_v1]") parser.add_argument("--num_chunks", type=int, default=1) parser.add_argument("--chunk_idx", type=int, default=0) parser.add_argument("--image_w", type=int, default=336) # 224 parser.add_argument("--image_h", type=int, default=336) # 224 parser.add_argument("--add_region_feature", action="store_true") parser.add_argument("--region_format", type=str, default="point", choices=["point", "box", "segment", "free_shape"]) parser.add_argument("--no_coor", action="store_true") parser.add_argument("--temperature", type=float, default=0.001) parser.add_argument("--top_p", type=float, default=None) parser.add_argument("--num_beams", type=int, default=1) parser.add_argument("--max_new_tokens", type=int, default=1024) parser.add_argument("--data_type", type=str, default='fp16', choices=['fp16', 'bf16', 'fp32']) args = parser.parse_args() if args.data_type == 'fp16': args.data_type = torch.float16 elif args.data_type == 'bf16': args.data_type = torch.bfloat16 else: args.data_type = torch.float32 eval_model(args)