import argparse import torch import os import json from tqdm import tqdm import shortuuid from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN from llava.conversation import conv_templates, SeparatorStyle from llava.model.builder import load_pretrained_model from llava.utils import disable_torch_init from llava.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path from llava.eval.m4c_evaluator import EvalAIAnswerProcessor from torch.utils.data import Dataset, DataLoader from PIL import Image import math 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] # Custom dataset class class CustomDataset(Dataset): def __init__(self, questions, image_folder, tokenizer, image_processor, model_config): self.questions = questions self.image_folder = image_folder self.tokenizer = tokenizer self.image_processor = image_processor self.model_config = model_config def __getitem__(self, index): line = self.questions[index] image_file = line["image_id"] # qs = line["text"] qs = line["question"] if self.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 qs += PROMPT_SUFFIX 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() image = Image.open(os.path.join(self.image_folder, image_file)).convert('RGB') image_tensor = process_images([image], self.image_processor, self.model_config)[0] input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt') return input_ids, image_tensor def __len__(self): return len(self.questions) # DataLoader def create_data_loader(questions, image_folder, tokenizer, image_processor, model_config, batch_size=1, num_workers=4): assert batch_size == 1, "batch_size must be 1" dataset = CustomDataset(questions, image_folder, tokenizer, image_processor, model_config) data_loader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False) return data_loader def eval_model(args): # Model disable_torch_init() model_path = os.path.expanduser(args.model_path) model_name = get_model_name_from_path(model_path) model, image_processor, tokenizer, context_len = load_pretrained_model(model_path, args.model_base, model_name, pretrained_rob_path=args.pretrained_rob_path) # print(model.transformer()) # model.to('cuda') # rand_input = torch.rand((1, 3, 224, 224), dtype=torch.half).to('cuda') # # rand_input = torch.rand((1, 3, 336, 336), dtype=torch.half).to('cuda') # op = model.get_model().get_vision_tower()(rand_input) # print(op.size()) # exit() # questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")] questions = json.load(open(os.path.expanduser(args.question_file), "r")) questions = questions["questions"] questions = get_chunk(questions, args.num_chunks, args.chunk_idx) answers_file = os.path.expanduser(args.answers_file) os.makedirs(os.path.dirname(answers_file), exist_ok=True) ans_file = open(answers_file, "w") if 'plain' in model_name and 'finetune' not in model_name.lower() and 'mmtag' not in args.conv_mode: args.conv_mode = args.conv_mode + '_mmtag' print(f'It seems that this is a plain model, but it is not using a mmtag prompt, auto switching to {args.conv_mode}.') data_loader = create_data_loader(questions, args.image_folder, tokenizer, image_processor, model.config) answer_processor = EvalAIAnswerProcessor() i = 0 all_outputs = [] for (input_ids, image_tensor), line in tqdm(zip(data_loader, questions), total=len(questions)): # idx = line["question_id"] # cur_prompt = line["text"] if i >= 1000: break q_id = line["question_id"] cur_prompt = line["question"] + PROMPT_SUFFIX stop_str = conv_templates[args.conv_mode].sep if conv_templates[args.conv_mode].sep_style != SeparatorStyle.TWO else conv_templates[args.conv_mode].sep2 input_ids = input_ids.to(device='cuda', non_blocking=True) with torch.inference_mode(): output_ids = model.generate( input_ids, images=image_tensor.to(dtype=torch.float16, device='cuda', non_blocking=True), 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=128, use_cache=True) input_token_len = input_ids.shape[1] n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item() if n_diff_input_output > 0: print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids') outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0] outputs = outputs.strip() if outputs.endswith(stop_str): outputs = outputs[:-len(stop_str)] outputs = outputs.strip() # ans_id = shortuuid.uuid() # ans_file.write(json.dumps({"question_id": q_id, # "prompt": cur_prompt, # "answer": outputs, # "answer_id": ans_id, # "model_id": model_name, # "metadata": {}}) + "\n") all_outputs.append( {"question_id": q_id, "question": cur_prompt, "answer": answer_processor(outputs), "answer_orig": outputs} ) print(f"\n[question] {cur_prompt}") print(f"[answer] {outputs}") i += 1 # ans_file.flush() json.dump(all_outputs, ans_file, indent=4) ans_file.close() print(f'Wrote {len(all_outputs)} answers to {answers_file}') if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--model-path", type=str, default="liuhaotian/llava-v1.5-7b") parser.add_argument("--pretrained_rob_path", type=str, default=None, help='Pass None, openai or path-to-rob-ckpt') # "/data/naman_deep_singh/project_multimodal/clip-finetune/sbatch/ViT-L-14_openai_imagenet_txtSup_False_vit-l-unsup-clean-0p1-eps4-3adv-lr1e-4-wd-1e-3_f8o0v/checkpoints/final.pt") # /mnt/nsingh/project_multimodal/models/ViT-L-14_openai_imagenet_txtSup_False_vit-l-unsup-clean-0p1-eps4-3adv-lr1e-4-wd-1e-3_f8o0v/checkpoints/final.pt parser.add_argument("--model-base", type=str, default=None) parser.add_argument("--image-folder", type=str, default="/mnt/datasets/vizwiz/val") parser.add_argument("--question-file", type=str, default="/mnt/datasets/vizwiz/val_questions_vqa_format.json") parser.add_argument("--answers-file", type=str, help="for output", default="/mnt/cschlarmann37/scratch.json") parser.add_argument("--conv-mode", type=str, default="vicuna_v1") parser.add_argument("--num-chunks", type=int, default=1) parser.add_argument("--chunk-idx", type=int, default=0) parser.add_argument("--temperature", type=float, default=0.) parser.add_argument("--top_p", type=float, default=None) parser.add_argument("--num_beams", type=int, default=1) args = parser.parse_args() DATASET_NAME = "vizwiz" if DATASET_NAME == "vizwiz": PROMPT_SUFFIX = "\nWhen the provided information is insufficient, respond with 'Unanswerable'.\nAnswer the question using a single word or phrase." else: PROMPT_SUFFIX = "" print(f"Unknown dataset: {DATASET_NAME}, using no prompt suffix.") eval_model(args)