import argparse import itertools import json import os import random import time from functools import partial from typing import Optional import torch from tqdm import tqdm from PIL import Image from mplug_owl2.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN from mplug_owl2.conversation import conv_templates, SeparatorStyle from mplug_owl2.model.builder import load_pretrained_model from mplug_owl2.mm_utils import process_images, tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria from vqa import VQA from vqa_eval import VQAEval ds_collections = { 'vqav2_val': { 'train': 'data/vqav2/vqav2_train.jsonl', 'test': 'data/vqav2/vqav2_val.jsonl', 'question': 'data/vqav2/v2_OpenEnded_mscoco_val2014_questions.json', 'annotation': 'data/vqav2/v2_mscoco_val2014_annotations.json', 'metric': 'vqa_score', 'max_new_tokens': 10, }, 'vqav2_testdev': { 'train': 'data/vqav2/vqav2_train.jsonl', 'test': 'data/vqav2/vqav2_testdev.jsonl', 'metric': None, 'max_new_tokens': 10, }, 'okvqa_val': { 'train': 'data/okvqa/okvqa_train.jsonl', 'test': 'data/okvqa/okvqa_val.jsonl', 'question': 'data/okvqa/OpenEnded_mscoco_val2014_questions.json', 'annotation': 'data/okvqa/mscoco_val2014_annotations.json', 'metric': 'vqa_score', 'max_new_tokens': 10, }, 'textvqa_val': { 'train': 'data/textvqa/textvqa_train.jsonl', 'test': 'data/textvqa/textvqa_val.jsonl', 'question': 'data/textvqa/textvqa_val_questions.json', 'annotation': 'data/textvqa/textvqa_val_annotations.json', 'metric': 'vqa_score', 'max_new_tokens': 10, }, } def collate_fn(batches, tokenizer): questions = [_['question'] for _ in batches] question_ids = [_['question_id'] for _ in batches] annotations = [_['annotation'] for _ in batches] image_tensor = [_['image_tensor'] for _ in batches] input_ids = [] for input_text in questions: input_ids.append(tokenizer_image_token(input_text, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').tolist()) input_tokens_max_length = max([len(x) for x in input_ids]) pad_token_id = tokenizer.pad_token_id input_ids = [([pad_token_id] * (input_tokens_max_length - len(_)) + _) for _ in input_ids] # pad in the left input_ids = torch.LongTensor(input_ids) attention_mask = 1 - input_ids.eq(pad_token_id).long() image_tensor = torch.cat(image_tensor, dim=0) return question_ids, image_tensor, input_ids, attention_mask, annotations class VQADataset(torch.utils.data.Dataset): def __init__(self, train, test, prompt, image_processor, few_shot): self.test = json.load(open(test)) self.prompt = prompt self.image_processor = image_processor self.few_shot = few_shot if few_shot > 0: self.train = open(train).readlines() def __len__(self): return len(self.test) def __getitem__(self, idx): data = self.test[idx] image, question, question_id, annotation = data['image'], data[ 'question'], data['question_id'], data.get('answer', None) image = Image.open(image).convert('RGB') max_edge = max(image.size) image = image.resize((max_edge, max_edge)) # Resize here for best performance image_tensor = process_images([image], self.image_processor) return { 'image_tensor': image_tensor, 'question': self.prompt.format(question), 'question_id': question_id, 'annotation': annotation } class InferenceSampler(torch.utils.data.sampler.Sampler): def __init__(self, size): self._size = int(size) assert size > 0 self._rank = torch.distributed.get_rank() self._world_size = torch.distributed.get_world_size() self._local_indices = self._get_local_indices(size, self._world_size, self._rank) @staticmethod def _get_local_indices(total_size, world_size, rank): shard_size = total_size // world_size left = total_size % world_size shard_sizes = [shard_size + int(r < left) for r in range(world_size)] begin = sum(shard_sizes[:rank]) end = min(sum(shard_sizes[:rank + 1]), total_size) return range(begin, end) def __iter__(self): yield from self._local_indices def __len__(self): return len(self._local_indices) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--checkpoint', type=str, default='') parser.add_argument('--dataset', type=str, default='textvqa_val') parser.add_argument('--batch-size', type=int, default=1) parser.add_argument('--num-workers', type=int, default=1) parser.add_argument('--few-shot', type=int, default=0) parser.add_argument('--seed', type=int, default=0) args = parser.parse_args() torch.distributed.init_process_group( backend='nccl', world_size=int(os.getenv('WORLD_SIZE', '1')), rank=int(os.getenv('RANK', '0')), ) torch.cuda.set_device(int(os.getenv('LOCAL_RANK', 0))) model_path = args.checkpoint model_name = get_model_name_from_path(model_path) tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, None, model_name, load_8bit=False, load_4bit=False, device_map={"":f"cuda:{os.getenv('LOCAL_RANK', '0')}"}, device="cuda") tokenizer.padding_side = 'left' if not hasattr(tokenizer, 'pad_token_id'): tokenizer.pad_token_id = tokenizer.eos_token_id prompt = 'USER: <|image|>{} Answer the question using a single word or phrase. ASSISTANT: ' answer_processor = EvalAIAnswerProcessor() random.seed(args.seed) dataset = VQADataset( train=ds_collections[args.dataset]['train'], test=ds_collections[args.dataset]['test'], prompt=prompt, image_processor=image_processor, few_shot=args.few_shot, ) dataloader = torch.utils.data.DataLoader( dataset=dataset, sampler=InferenceSampler(len(dataset)), batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=True, drop_last=False, collate_fn=partial(collate_fn, tokenizer=tokenizer), ) outputs = [] for _, (question_ids, image_tensor, input_ids, attention_mask, annotations) in enumerate(tqdm(dataloader)): pred = model.generate( input_ids=input_ids.cuda(), attention_mask=attention_mask.cuda(), images=image_tensor.to(dtype=model.dtype).cuda(), do_sample=False, num_beams=1, max_new_tokens=ds_collections[args.dataset]['max_new_tokens'], min_new_tokens=1, length_penalty=1, num_return_sequences=1, output_hidden_states=True, use_cache=True, ) answers = [ tokenizer.decode(_[input_ids.size(1):].cpu(), skip_special_tokens=True).strip() for _ in pred ] for question_id, answer, annotation in zip(question_ids, answers, annotations): if args.dataset in ['vqav2_val', 'okvqa_val', 'textvqa_val']: outputs.append({ 'question_id': question_id, 'answer': answer, }) elif args.dataset == 'vqav2_testdev': outputs.append({ 'question_id': question_id, 'answer': answer_processor(answer), }) else: raise NotImplementedError torch.distributed.barrier() world_size = torch.distributed.get_world_size() merged_outputs = [None for _ in range(world_size)] torch.distributed.all_gather_object(merged_outputs, json.dumps(outputs)) merged_outputs = [json.loads(_) for _ in merged_outputs] merged_outputs = [_ for _ in itertools.chain.from_iterable(merged_outputs)] if torch.distributed.get_rank() == 0: print(f"Evaluating {args.dataset} ...") time_prefix = time.strftime('%y%m%d%H%M%S', time.localtime()) results_file = f'{args.dataset}_{time_prefix}_fs{args.few_shot}_s{args.seed}.json' json.dump(merged_outputs, open(results_file, 'w', encoding='utf-8'), ensure_ascii=False) if ds_collections[args.dataset]['metric'] == 'vqa_score': vqa = VQA(ds_collections[args.dataset]['annotation'], ds_collections[args.dataset]['question']) results = vqa.loadRes( resFile=results_file, quesFile=ds_collections[args.dataset]['question']) vqa_scorer = VQAEval(vqa, results, n=2) vqa_scorer.evaluate() print(vqa_scorer.accuracy) torch.distributed.barrier()