tinychart/eval/eval_model.py

import argparse
import time

import torch
import os
import json
from tqdm import tqdm
import shortuuid

from tinychart.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
from tinychart.conversation import conv_templates, SeparatorStyle
from tinychart.model.builder import load_pretrained_model
from tinychart.utils import disable_torch_init
from tinychart.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path, KeywordsStoppingCriteria
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]

class EvalDataset(Dataset):
    def __init__(self, data_items, image_folder, tokenizer, image_processor, model_config):
        self.data_items = data_items
        self.image_folder = image_folder
        self.tokenizer = tokenizer
        self.image_processor = image_processor
        self.model_config = model_config

    def __getitem__(self, index):
        line = self.data_items[index]
        image_file = line["image"]
        qs = line["conversations"][0]["value"]
        # 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(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, image.size

    def __len__(self):
        return len(self.data_items)


def collate_fn(batch):
    input_ids, image_tensors, image_sizes = 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


# 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 = EvalDataset(questions, image_folder, 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):
    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)

    all_data = json.load(open(args.data_path, "r"))
    all_data = get_chunk(all_data, args.num_chunks, args.chunk_idx)
    answers_file = os.path.expanduser(args.output_path)
    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(all_data, args.image_folder, tokenizer, image_processor, model.config)
    for (input_ids, image_tensor, image_sizes), line in tqdm(zip(data_loader, all_data), total=len(all_data)):
        idx = line["id"]
        cur_prompt = line["conversations"][0]["value"]
        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),
                pad_token_id=tokenizer.pad_token_id,
                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,
                min_new_tokens=args.min_new_tokens,
                use_cache=True)

        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
        ans_id = shortuuid.uuid()
        ans_file.write(json.dumps({"id": idx,
                                   "question": cur_prompt,
                                   "gt_answer": line["conversations"][1]["value"],
                                   "model_answer": outputs}) + "\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("--image_folder", type=str, default="")
    parser.add_argument("--data_path", type=str, default="./data/test_chartqa+cot_shuffle.json")
    parser.add_argument("--output_path", type=str, default="./output/")
    parser.add_argument("--conv_mode", type=str, default="phi")
    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=0.0)
    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("--min_new_tokens", type=int, default=0)
    args = parser.parse_args()

    eval_model(args)