""" 数据预处理脚本 支持提取文本、图像、音频等多模态数据 使用先进的Whisper ASR进行语音转文本 """ import os import cv2 import zipfile import subprocess import pickle import pandas as pd from pathlib import Path from typing import List, Optional, Dict from config import Config from asr_transcription import ASRTranscriber, get_transcription_text class DataPreprocessor: """数据预处理器""" def __init__(self, whisper_model_size: str = "base"): """ 初始化数据预处理器 Args: whisper_model_size: Whisper模型大小 (tiny/base/small/medium/large-v2) """ Config.create_dirs() self.asr_transcriber = None self.whisper_model_size = whisper_model_size # 如果启用文本模态，初始化ASR if Config.MODALITIES["text"]: try: self.asr_transcriber = ASRTranscriber(model_size=whisper_model_size) print(f"ASR转录器已初始化 (模型: {whisper_model_size})") except Exception as e: print(f"警告: ASR转录器初始化失败: {e}") print("文本模态将被禁用") Config.MODALITIES["text"] = False def extract_images_from_video( self, video_path: str, output_dir: str, max_frames: int = None ) -> List[str]: """ 从视频提取图像帧 Args: video_path: 视频文件路径 output_dir: 输出目录 max_frames: 最大提取帧数 Returns: 提取的图像路径列表 """ max_frames = max_frames or Config.MAX_FRAMES video_name = Path(video_path).stem # 创建输出目录 frame_dir = os.path.join(output_dir, video_name) os.makedirs(frame_dir, exist_ok=True) # 打开视频 cap = cv2.VideoCapture(video_path) if not cap.isOpened(): print(f"无法打开视频: {video_path}") return [] frame_paths = [] count = 0 while count < max_frames: ret, frame = cap.read() if not ret: break # 调整大小 frame = cv2.resize( frame, (Config.IMAGE_SIZE, Config.IMAGE_SIZE), interpolation=cv2.INTER_CUBIC ) # 保存帧 frame_path = os.path.join(frame_dir, f"frame_{count:03d}.jpg") cv2.imwrite(frame_path, frame) frame_paths.append(frame_path) count += 1 cap.release() return frame_paths def extract_audio_from_video( self, video_path: str, output_dir: str ) -> Optional[str]: """ 从视频提取音频 Args: video_path: 视频文件路径 output_dir: 输出目录 Returns: 音频文件路径 """ video_name = Path(video_path).stem audio_path = os.path.join(output_dir, f"{video_name}.wav") os.makedirs(output_dir, exist_ok=True) # 使用 ffmpeg 提取音频 command = [ "ffmpeg", "-i", video_path, "-ab", "320k", "-ac", "2", "-ar", str(Config.AUDIO_SAMPLE_RATE), "-vn", # 不包含视频 "-y", # 覆盖输出文件 audio_path ] try: subprocess.run( command, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL ) return audio_path except subprocess.CalledProcessError: print(f"提取音频失败: {video_path}") return None def extract_text_from_video( self, video_path: str, output_dir: str, use_audio: bool = True ) -> Optional[Dict]: """ 从视频提取文本转录（使用先进的Whisper ASR） Args: video_path: 视频文件路径 output_dir: 输出目录 use_audio: 如果为True，先提取音频再转录（更高效）；False则直接从视频转录 Returns: 包含转录信息的字典: { "text_path": "文本文件路径", "json_path": "JSON文件路径", "transcription": 转录结果字典 } """ if not Config.MODALITIES["text"]: return None video_name = Path(video_path).stem os.makedirs(output_dir, exist_ok=True) # 初始化ASR（如果尚未初始化） if self.asr_transcriber is None: try: self.asr_transcriber = ASRTranscriber(model_size=self.whisper_model_size) except Exception as e: print(f"无法初始化ASR转录器: {e}") return None try: # 方法1: 如果已有音频文件，直接转录音频（更快） if use_audio: # 先提取音频 audio_dir = os.path.join(os.path.dirname(output_dir), "audio_temp") os.makedirs(audio_dir, exist_ok=True) audio_path = os.path.join(audio_dir, f"{video_name}.wav") # 提取音频（如果不存在） if not os.path.exists(audio_path): command = [ "ffmpeg", "-i", video_path, "-ab", "320k", "-ac", "1", # 单声道（Whisper推荐） "-ar", "16000", # 16kHz（Whisper推荐） "-vn", "-y", audio_path ] try: subprocess.run( command, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL ) except subprocess.CalledProcessError: print(f"提取音频失败，尝试直接从视频转录") audio_path = None # 转录音频或视频 if audio_path and os.path.exists(audio_path): transcription = self.asr_transcriber.transcribe_audio(audio_path) else: transcription = self.asr_transcriber.transcribe_video(video_path) else: # 方法2: 直接从视频转录 transcription = self.asr_transcriber.transcribe_video(video_path) # 保存转录结果 json_path = os.path.join(output_dir, f"{video_name}_transcription.json") txt_path = os.path.join(output_dir, f"{video_name}_transcription.txt") # 保存JSON（包含完整信息） self.asr_transcriber.save_transcription(transcription, json_path, "json") # 保存纯文本 self.asr_transcriber.save_transcription(transcription, txt_path, "txt") print(f"转录完成: {video_name} (语言: {transcription.get('language', 'unknown')}, " f"文本长度: {len(transcription['text'])} 字符)") return { "text_path": txt_path, "json_path": json_path, "transcription": transcription } except Exception as e: print(f"转录失败 {video_name}: {e}") import traceback traceback.print_exc() return None def process_video( self, video_path: str, video_name: str, dataset_type: str = "training" ) -> Dict: """ 处理单个视频，提取所有模态数据 Args: video_path: 视频文件路径 video_name: 视频名称 dataset_type: 数据集类型 ("training" 或 "validation") Returns: 包含各模态数据路径的字典 """ result = { "video_name": video_name, "video_path": video_path, # 保存原始视频路径 "images": [], "audio_path": None, "text_path": None, "text_json_path": None, "transcription": None } # 提取图像 if Config.MODALITIES["vision"]: result["images"] = self.extract_images_from_video( video_path, os.path.join(Config.IMAGE_DATA_DIR, f"{dataset_type}Data"), max_frames=Config.MAX_FRAMES ) # 提取音频 if Config.MODALITIES["audio"]: result["audio_path"] = self.extract_audio_from_video( video_path, os.path.join(Config.AUDIO_DATA_DIR, f"{dataset_type}Data") ) # 提取文本转录（使用Whisper ASR） if Config.MODALITIES["text"]: text_result = self.extract_text_from_video( video_path, os.path.join(Config.TEXT_DATA_DIR, f"{dataset_type}Data"), use_audio=True # 使用已提取的音频（如果存在） ) if text_result: result["text_path"] = text_result["text_path"] result["text_json_path"] = text_result["json_path"] result["transcription"] = text_result["transcription"] return result def process_dataset( self, zip_files: List[str] = None, video_dir: str = None, dataset_type: str = "training" ): """ 处理整个数据集 Args: zip_files: zip文件路径列表（可选） video_dir: 视频目录路径（可选，如果提供则直接处理目录中的mp4文件） dataset_type: 数据集类型 ("training" 或 "validation") """ print(f"开始处理 {dataset_type} 数据集...") all_results = [] # 方式1: 如果提供了视频目录，直接处理目录中的mp4文件 if video_dir and os.path.exists(video_dir): print(f"从目录处理视频: {video_dir}") import glob video_files = glob.glob(os.path.join(video_dir, "*.mp4")) print(f"找到 {len(video_files)} 个视频文件") for video_path in video_files: video_name = Path(video_path).stem result = self.process_video(video_path, video_name, dataset_type) all_results.append(result) if len(all_results) % 10 == 0: print(f"已处理 {len(all_results)}/{len(video_files)} 个视频") # 显示转录统计 if Config.MODALITIES["text"]: transcribed_count = sum( 1 for r in all_results if r.get("transcription") is not None ) print(f" 其中 {transcribed_count} 个视频已完成转录") # 方式2: 从zip文件处理（原始方式） elif zip_files: for zip_file in zip_files: if not os.path.exists(zip_file): print(f"文件不存在: {zip_file}") continue print(f"处理: {zip_file}") # 解压zip文件 with zipfile.ZipFile(zip_file, 'r') as archive: archive.extractall(f"./unzippedData/{dataset_type}") # 处理每个视频 for file_name in archive.namelist(): if file_name.endswith('.mp4'): video_path = os.path.join( f"./unzippedData/{dataset_type}", file_name ) if os.path.exists(video_path): video_name = Path(file_name).stem result = self.process_video(video_path, video_name, dataset_type) all_results.append(result) if len(all_results) % 10 == 0: print(f"已处理 {len(all_results)} 个视频") # 显示转录统计 if Config.MODALITIES["text"]: transcribed_count = sum( 1 for r in all_results if r.get("transcription") is not None ) print(f" 其中 {transcribed_count} 个视频已完成转录") # 保存处理结果 output_file = os.path.join( Config.OUTPUT_DIR, f"{dataset_type}_data_info.pkl" ) with open(output_file, "wb") as f: pickle.dump(all_results, f) print(f"处理完成！共处理 {len(all_results)} 个视频") print(f"结果已保存到: {output_file}") return all_results def load_annotations(self, annotation_file: str) -> pd.DataFrame: """ 加载标注文件 Args: annotation_file: 标注文件路径 Returns: 标注DataFrame """ with open(annotation_file, "rb") as f: pickle_data = pickle.load(f, encoding="latin1") df = pd.DataFrame(pickle_data) df.reset_index(inplace=True) if "interview" in df.columns: del df["interview"] df.columns = [ "VideoName", "ValueExtraversion", "ValueNeuroticism", "ValueAgreeableness", "ValueConscientiousness", "ValueOpenness", ] return df def main(): """主函数：处理训练和验证数据""" import argparse parser = argparse.ArgumentParser(description="数据预处理：提取多模态数据") parser.add_argument( "--whisper-model", type=str, default="base", choices=["tiny", "base", "small", "medium", "large-v2"], help="Whisper模型大小 (默认: base)" ) parser.add_argument( "--skip-training", action="store_true", help="跳过训练数据处理" ) parser.add_argument( "--skip-validation", action="store_true", help="跳过验证数据处理" ) parser.add_argument( "--video-dir", type=str, default=None, help="直接指定视频目录（而不是zip文件）" ) args = parser.parse_args() print("=" * 60) print("数据预处理：多模态数据提取") print("=" * 60) print(f"Whisper模型: {args.whisper_model}") print(f"激活的模态: {Config.get_active_modalities()}") print("=" * 60) preprocessor = DataPreprocessor(whisper_model_size=args.whisper_model) # 如果指定了视频目录，直接使用 if args.video_dir: print(f"\n使用视频目录: {args.video_dir}") preprocessor.process_dataset(video_dir=args.video_dir, dataset_type="training") else: # 尝试从first-impressions-v2目录处理 train_dir = "../First-Impression/data/first-impressions-v2/train" val_dir = "../First-Impression/data/first-impressions-v2/validation" if os.path.exists(train_dir): # 处理训练数据 if not args.skip_training: print("\n开始处理训练数据...") preprocessor.process_dataset(video_dir=train_dir, dataset_type="training") # 处理验证数据 if not args.skip_validation: print("\n开始处理验证数据...") preprocessor.process_dataset(video_dir=val_dir, dataset_type="validation") else: # 回退到zip文件方式 if not args.skip_training: print("\n开始处理训练数据（从zip文件）...") training_zips = [ f"../First-Impression/data/training80_{i:02d}.zip" for i in range(1, 76) ] preprocessor.process_dataset(training_zips, dataset_type="training") if not args.skip_validation: print("\n开始处理验证数据（从zip文件）...") validation_zips = [ f"../First-Impression/data/validation80_{i:02d}.zip" for i in range(1, 26) ] preprocessor.process_dataset(validation_zips, dataset_type="validation") print("\n" + "=" * 60) print("数据预处理完成！") print("=" * 60) if __name__ == "__main__": main()