XaviXva/Video-LLaVA

1import argparse2import torch3import os4import json5from tqdm import tqdm6import shortuuid7 8from llava.constants import X_TOKEN_INDEX, DEFAULT_X_TOKEN, DEFAULT_X_START_TOKEN, DEFAULT_X_END_TOKEN9from llava.conversation import conv_templates, SeparatorStyle10from llava.model.builder import load_pretrained_model11from llava.utils import disable_torch_init12from llava.mm_utils import tokenizer_X_token, process_images, get_model_name_from_path13from torch.utils.data import Dataset, DataLoader14 15from PIL import Image16import math17 18 19def split_list(lst, n):20    """Split a list into n (roughly) equal-sized chunks"""21    chunk_size = math.ceil(len(lst) / n)  # integer division22    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]23 24 25def get_chunk(lst, n, k):26    chunks = split_list(lst, n)27    return chunks[k]28 29 30# Custom dataset class31class CustomDataset(Dataset):32    def __init__(self, questions, image_folder, tokenizer, image_processor, model_config):33        self.questions = questions34        self.image_folder = image_folder35        self.tokenizer = tokenizer36        self.image_processor = image_processor37        self.model_config = model_config38 39    def __getitem__(self, index):40        line = self.questions[index]41        image_file = line["image"]42        qs = line["text"]43        # if self.model_config.mm_use_im_start_end:44        #     qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs45        # else:46        #     qs = DEFAULT_IMAGE_TOKEN + '\n' + qs47 48        if self.model_config.mm_use_x_start_end:49            qs = DEFAULT_X_START_TOKEN['IMAGE'] + DEFAULT_X_TOKEN['IMAGE'] + DEFAULT_X_END_TOKEN['IMAGE'] + '\n' + qs50        else:51            qs = DEFAULT_X_TOKEN['IMAGE'] + '\n' + qs52        conv = conv_templates[args.conv_mode].copy()53        conv.append_message(conv.roles[0], qs)54        conv.append_message(conv.roles[1], None)55        prompt = conv.get_prompt()56 57        image = Image.open(os.path.join(self.image_folder, image_file)).convert('RGB')58        image_tensor = process_images([image], self.image_processor, self.model_config)[0]59 60        input_ids = tokenizer_X_token(prompt, self.tokenizer, X_TOKEN_INDEX['IMAGE'], return_tensors='pt')61 62        return input_ids, image_tensor63 64    def __len__(self):65        return len(self.questions)66 67 68# DataLoader69def create_data_loader(questions, image_folder, tokenizer, image_processor, model_config, batch_size=1, num_workers=4):70    assert batch_size == 1, "batch_size must be 1"71    dataset = CustomDataset(questions, image_folder, tokenizer, image_processor, model_config)72    data_loader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False)73    return data_loader74 75 76def eval_model(args):77    # Model78    disable_torch_init()79    model_path = os.path.expanduser(args.model_path)80    model_name = get_model_name_from_path(model_path)81    tokenizer, model, processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)82 83    questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")]84    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)85    answers_file = os.path.expanduser(args.answers_file)86    os.makedirs(os.path.dirname(answers_file), exist_ok=True)87    ans_file = open(answers_file, "w")88 89    if 'plain' in model_name and 'finetune' not in model_name.lower() and 'mmtag' not in args.conv_mode:90        args.conv_mode = args.conv_mode + '_mmtag'91        print(f'It seems that this is a plain model, but it is not using a mmtag prompt, auto switching to {args.conv_mode}.')92 93    data_loader = create_data_loader(questions, args.image_folder, tokenizer, processor['image'], model.config)94 95    for (input_ids, image_tensor), line in tqdm(zip(data_loader, questions), total=len(questions)):96        idx = line["question_id"]97        cur_prompt = line["text"]98 99        stop_str = conv_templates[args.conv_mode].sep if conv_templates[args.conv_mode].sep_style != SeparatorStyle.TWO else conv_templates[args.conv_mode].sep2100        input_ids = input_ids.to(device='cuda', non_blocking=True)101 102        with torch.inference_mode():103            output_ids = model.generate(104                input_ids,105                images=[[image_tensor[0].to(dtype=torch.float16, device='cuda', non_blocking=True)], ['image']],106                do_sample=True if args.temperature > 0 else False,107                temperature=args.temperature,108                top_p=args.top_p,109                num_beams=args.num_beams,110                max_new_tokens=128,111                use_cache=True)112 113        input_token_len = input_ids.shape[1]114        n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()115        if n_diff_input_output > 0:116            print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')117        outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]118        outputs = outputs.strip()119        if outputs.endswith(stop_str):120            outputs = outputs[:-len(stop_str)]121        outputs = outputs.strip()122 123        ans_id = shortuuid.uuid()124        ans_file.write(json.dumps({"question_id": idx,125                                   "prompt": cur_prompt,126                                   "text": outputs,127                                   "answer_id": ans_id,128                                   "model_id": model_name,129                                   "metadata": {}}) + "\n")130        # ans_file.flush()131    ans_file.close()132 133if __name__ == "__main__":134    parser = argparse.ArgumentParser()135    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")136    parser.add_argument("--model-base", type=str, default=None)137    parser.add_argument("--image-folder", type=str, default="")138    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")139    parser.add_argument("--answers-file", type=str, default="answer.jsonl")140    parser.add_argument("--conv-mode", type=str, default="llava_v1")141    parser.add_argument("--num-chunks", type=int, default=1)142    parser.add_argument("--chunk-idx", type=int, default=0)143    parser.add_argument("--temperature", type=float, default=0.2)144    parser.add_argument("--top_p", type=float, default=None)145    parser.add_argument("--num_beams", type=int, default=1)146    args = parser.parse_args()147 148    eval_model(args)149