import os import sys import pathlib import json import argparse import hashlib from tqdm import tqdm from video import VideoData from utils import load_json, save_json # 是否默认使用全部帧进行评测，若为 False，则默认采样16帧，因为我的电脑跑全部帧会爆内存 DEFAULT_ALL_FRAMES = True # --model_input_dir和--model_output_dir参数，即被评测模型的输入和输出文件夹路径 # 评测脚本会遍历输出文件夹中的视频文件，并根据对应的输入文件夹中的同名图片、音频、文本文件，进行评测 parser = argparse.ArgumentParser(description="Evaluate video generation models.") parser.add_argument('--model_input_dir', type=str, required=True, help='Path to the input directory of the evaluated model. This directory should contain the reference images, audio, and text files corresponding to the generated videos.') parser.add_argument('--model_output_dir', type=str, required=True, help='Path to the output directory of the evaluated model. This directory should contain the generated video files.') # results_dir参数，指定评测结果的保存路径 parser.add_argument('--results_dir', type=str, default='./evaluation_results', help='Path to the directory where evaluation results will be saved.') # --evaluate_subjects参数，指定需要运行的评测主题，多个主题用逗号分隔 # 例如：--evaluate_subjects dino_consistency,audio_visual_synchrony parser.add_argument('--evaluate_subjects', type=str, required=True, help='Comma-separated list of evaluation subjects to run, e.g., "dino_consistency,audio_visual_synchrony".') # --device参数，指定运行评测时使用的设备，默认为'cuda' parser.add_argument('--device', type=str, default='cuda', help='Device to use for evaluation, e.g., "cuda" or "cpu". Default is "cuda".') # --batch_size参数，指定评测时的批处理大小，默认为16 parser.add_argument('--batch_size', type=int, default=16, help='Batch size to use during evaluation. Default is 16.') if DEFAULT_ALL_FRAMES: # --sampling 参数，指定评测时的视频采样数量，默认为0，表示使用全部帧 parser.add_argument('--sampling', type=int, default=0, help='Number of frames to sample from each video during evaluation. Default is 0 (use all frames).') else: # --sampling 参数，指定评测时的视频采样数量，默认为16，0表示使用全部帧 parser.add_argument('--sampling', type=int, default=16, help='Number of frames to sample from each video during evaluation. Default is 16.') # --model_args参数，传递给评测主题模型的额外参数，格式为JSON字符串，如果有多个评测主题，可以为每个主题传递不同的参数，用分号分隔 parser.add_argument('--model_args', type=str, default='{}', help='Additional arguments for the evaluation subject models in JSON format.') # --filter 参数，指定在评测前只过滤出文件名包含该字符串的文件进行评测，默认为空表示不过滤 parser.add_argument('--filter', type=str, default='', help='Filter to only evaluate videos whose filenames contain this string. Default is empty (no filter).') # --group_id 参数，指定当前评测的组号（从0开始），需要与--group_total配合使用 parser.add_argument('--group_id', type=int, default=None, help='Group ID for hash-based file partitioning (0-indexed). Must be used with --group_total.') # --group_total 参数，指定总共分成多少组，需要与--group_id配合使用 parser.add_argument('--group_total', type=int, default=None, help='Total number of groups for hash-based file partitioning. Must be used with --group_id.') args = parser.parse_args() def main(): model_input_dir = pathlib.Path(args.model_input_dir) model_output_dir = pathlib.Path(args.model_output_dir) results_dir = pathlib.Path(args.results_dir) results_dir.mkdir(parents=True, exist_ok=True) evaluate_subjects = [subj.strip() for subj in args.evaluate_subjects.split(',')] device = args.device batch_size = args.batch_size sampling = args.sampling model_args_list = [json.loads(arg) for arg in args.model_args.split(';')] # 验证分组参数 if (args.group_id is not None) != (args.group_total is not None): raise ValueError("--group_id and --group_total must be used together.") if args.group_id is not None: if args.group_id < 0 or args.group_id >= args.group_total: raise ValueError(f"--group_id must be between 0 and {args.group_total - 1}.") if args.group_total <= 0: raise ValueError("--group_total must be greater than 0.") print(f"Using hash-based file partitioning: group {args.group_id} of {args.group_total}") # 构建VideoData列表 data_list = [] for video_file in model_output_dir.glob('*.mp4'): video_filename = video_file.stem if args.filter and args.filter not in video_filename: continue # 如果启用了分组功能，根据文件名哈希取模判断是否属于当前组 if args.group_id is not None: # 使用 hashlib.md5 确保在不同进程和机器上哈希值一致 file_hash = int(hashlib.md5(video_filename.encode('utf-8')).hexdigest(), 16) group = file_hash % args.group_total if group != args.group_id: continue # 假设对应的参考图像、音频和文本文件与视频文件同名但扩展名不同 image_file = model_input_dir / f"{video_filename}.png" audio_file = model_input_dir / f"{video_filename}.wav" text_file = model_input_dir / f"{video_filename}.json" video_data = VideoData( video_path=str(video_file), audio_path=str(audio_file) if image_file.exists() else None, text_path=str(text_file) if text_file.exists() else None, image_path=str(image_file) if image_file.exists() else None ) data_list.append(video_data) # data_list = data_list[:3] # 逐个评测主题运行评测 for subject, model_args in zip(evaluate_subjects, model_args_list): print(f"Running evaluation for subject: {subject}") subject_module = __import__(f"subjects.{subject}", fromlist=['evaluate']) data_list = subject_module.evaluate( data_list, device=device, batch_size=batch_size, model_args=model_args, sampling=sampling ) # 保存评测结果 results = [data.to_dict() for data in data_list] # 如果使用了分组功能，在文件名中包含组信息 if args.group_id is not None: results_path = results_dir / f"evaluation_results_{'-'.join(evaluate_subjects)}_group{args.group_id}of{args.group_total}.json" else: results_path = results_dir / f"evaluation_results_{'-'.join(evaluate_subjects)}.json" save_json(results, str(results_path)) print(f"Evaluation results saved to {results_path}") print(f"Total videos evaluated: {len(data_list)}") if __name__ == "__main__": main() # Example usage: # python evaluate.py --model_input_dir /mnt/f/temp/video-eval-mock/inputs --model_output_dir /mnt/f/temp/video-eval-mock/results --evaluate_subjects dino_consistency --model_args '{"model_name": "facebook/dinov3-convnext-tiny-pretrain-lvd1689m"}'