| import os |
| import json |
| import numpy as np |
|
|
| import torch |
| import clip |
| from tqdm import tqdm |
| from vbench.utils import load_video, load_dimension_info, clip_transform, read_frames_decord_by_fps, CACHE_DIR |
| from vbench.third_party.ViCLIP.viclip import ViCLIP |
| from vbench.third_party.ViCLIP.simple_tokenizer import SimpleTokenizer |
|
|
| from .distributed import ( |
| get_world_size, |
| get_rank, |
| all_gather, |
| barrier, |
| distribute_list_to_rank, |
| gather_list_of_dict, |
| ) |
|
|
|
|
| def get_text_features(model, input_text, tokenizer, text_feature_dict={}): |
| if input_text in text_feature_dict: |
| return text_feature_dict[input_text] |
| text_template= f"{input_text}" |
| with torch.no_grad(): |
| text_features = model.encode_text(text_template).float() |
| text_features /= text_features.norm(dim=-1, keepdim=True) |
| text_feature_dict[input_text] = text_features |
| return text_features |
|
|
| def get_vid_features(model, input_frames): |
| with torch.no_grad(): |
| clip_feat = model.encode_vision(input_frames,test=True).float() |
| clip_feat /= clip_feat.norm(dim=-1, keepdim=True) |
| return clip_feat |
|
|
| def get_predict_label(clip_feature, text_feats_tensor, top=5): |
| label_probs = (100.0 * clip_feature @ text_feats_tensor.T).softmax(dim=-1) |
| top_probs, top_labels = label_probs.cpu().topk(top, dim=-1) |
| return top_probs, top_labels |
|
|
| def temporal_style(clip_model, video_dict, tokenizer, device, sample="middle"): |
| sim = [] |
| video_results = [] |
| image_transform = clip_transform(224) |
| for info in tqdm(video_dict, disable=get_rank() > 0): |
| query = info['prompt'] |
| |
| video_list = info['video_list'] |
| for video_path in video_list: |
| cur_video = [] |
| with torch.no_grad(): |
| |
| images = read_frames_decord_by_fps(video_path, num_frames=8, sample=sample) |
| images = image_transform(images) |
| images = images.to(device) |
| clip_feat = get_vid_features(clip_model,images.unsqueeze(0)) |
| text_feat = get_text_features(clip_model, query, tokenizer) |
| logit_per_text = clip_feat @ text_feat.T |
| score_per_video = float(logit_per_text[0][0].cpu()) |
| sim.append(score_per_video) |
| video_results.append({'video_path': video_path, 'video_results': score_per_video}) |
| avg_score = np.mean(sim) |
| return avg_score, video_results |
|
|
| def compute_temporal_style(json_dir, device, submodules_list, **kwargs): |
| tokenizer = SimpleTokenizer(os.path.join(CACHE_DIR, "ViCLIP/bpe_simple_vocab_16e6.txt.gz")) |
| viclip = ViCLIP(tokenizer= tokenizer, **submodules_list).to(device) |
| _, video_dict = load_dimension_info(json_dir, dimension='temporal_style', lang='en') |
| video_dict = distribute_list_to_rank(video_dict) |
| all_results, video_results = temporal_style(viclip, video_dict, tokenizer, device) |
| if get_world_size() > 1: |
| video_results = gather_list_of_dict(video_results) |
| all_results = sum([d['video_results'] for d in video_results]) / len(video_results) |
| return all_results, video_results |
|
|
