import argparse import torch import os import json from tqdm import tqdm import shortuuid from ola_vlm.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN from ola_vlm.conversation import conv_templates, SeparatorStyle from ola_vlm.model.builder import load_pretrained_model from ola_vlm.utils import disable_torch_init from ola_vlm.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path from torch.utils.data import Dataset, DataLoader from datasets import load_dataset 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] def load_jsonl(f): lines = open(f, encoding='utf-8').readlines() lines = [x.strip() for x in lines] if lines[-1] == '': lines = lines[:-1] data = [json.loads(x) for x in lines] return data def prepare_CVBench(path): dataset = load_jsonl(os.path.join(path, 'test.jsonl')) data = [] for i in range(len(dataset)): d = { "image": os.path.join(path, dataset[i]["filename"]), "question": dataset[i]["prompt"] + "\nOnly answer the option as the output. For example, if your answer is the option A, answer (A).", "answer": dataset[i]["answer"], "task": dataset[i]["task"], "source": dataset[i]["source"] } data.append(d) return data # Custom dataset class class CustomDataset(Dataset): def __init__(self, data, tokenizer, image_processor, model_config): self.questions = data self.tokenizer = tokenizer self.image_processor = image_processor self.model_config = model_config def __getitem__(self, index): d = self.questions[index] qs = d["question"] image_file = d["image"] ans = d["answer"] task = d["task"] source = d["source"] 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 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(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, image.size, ans, task, source def __len__(self): return len(self.questions) def collate_fn(batch): input_ids, image_tensors, image_sizes, answers, cats, cats_l2 = zip(*batch) input_ids = torch.stack(input_ids, dim=0) image_tensors = torch.stack(image_tensors, dim=0) return input_ids, image_tensors, image_sizes, answers, cats, cats_l2 # DataLoader def create_data_loader(questions, tokenizer, image_processor, model_config, batch_size=1, num_workers=4): assert batch_size == 1, "batch_size must be 1" dataset = CustomDataset(questions, tokenizer, image_processor, model_config) data_loader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False, collate_fn=collate_fn) 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) tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name) questions = prepare_CVBench(args.path) 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, tokenizer, image_processor, model.config) for (input_ids, image_tensor, image_sizes, answer, task, source), line in tqdm(zip(data_loader, questions), total=len(questions)): 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), image_sizes=image_sizes, 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) if not isinstance(output_ids, torch.Tensor): output_ids = output_ids.sequences outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip() ans_file.write(json.dumps({"prediction": outputs, "answer": answer, "question": line, "source": source, "task": task}) + "\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("--model-base", type=str, default=None) parser.add_argument("--path", type=str, default="CV-Bench") parser.add_argument("--answers-file", type=str, default="cv-bench_answer.jsonl") parser.add_argument("--conv-mode", type=str, default="llava_phi_3") 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=0.2) 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=128) args = parser.parse_args() eval_model(args)