diff --git a/app/services/material.py b/app/services/material.py index 2a84f85..e48ffa4 100644 --- a/app/services/material.py +++ b/app/services/material.py @@ -4,6 +4,7 @@ import random import traceback from urllib.parse import urlencode from datetime import datetime +import json import requests from typing import List @@ -332,14 +333,25 @@ def save_clip_video(timestamp: str, origin_video: str, save_dir: str = "") -> di video_id = f"vid-{timestamp.replace(':', '-').replace(',', '_')}" video_path = os.path.join(save_dir, f"{video_id}.mp4") + # 如果视频已存在，直接返回 if os.path.exists(video_path) and os.path.getsize(video_path) > 0: - logger.info(f"video already exists: {video_path}") + logger.info(f"视频已存在: {video_path}") return {timestamp: video_path} try: - # 加载视频获取总时长 - video = VideoFileClip(origin_video) - total_duration = video.duration + # 检查视频是否存在 + if not os.path.exists(origin_video): + logger.error(f"源视频文件不存在: {origin_video}") + return {} + + # 获取视频总时长 + try: + probe_cmd = ["ffprobe", "-v", "error", "-show_entries", "format=duration", + "-of", "default=noprint_wrappers=1:nokey=1", origin_video] + total_duration = float(subprocess.check_output(probe_cmd).decode('utf-8').strip()) + except subprocess.CalledProcessError as e: + logger.error(f"获取视频时长失败: {str(e)}") + return {} # 解析时间戳 start_str, end_str = timestamp.split('-') @@ -349,7 +361,6 @@ def save_clip_video(timestamp: str, origin_video: str, save_dir: str = "") -> di # 验证时间段 if start >= total_duration: logger.warning(f"起始时间 {format_timestamp(start)} ({start:.3f}秒) 超出视频总时长 {format_timestamp(total_duration)} ({total_duration:.3f}秒)") - video.close() return {} if end > total_duration: @@ -358,57 +369,161 @@ def save_clip_video(timestamp: str, origin_video: str, save_dir: str = "") -> di if end <= start: logger.warning(f"结束时间 {format_timestamp(end)} 必须大于起始时间 {format_timestamp(start)}") - video.close() return {} - # 剪辑视频 + # 计算剪辑时长 duration = end - start logger.info(f"开始剪辑视频: {format_timestamp(start)} - {format_timestamp(end)}，时长 {format_timestamp(duration)}") - # 剪辑视频 - subclip = video.subclip(start, end) + # 检测可用的硬件加速选项 + hwaccel = _detect_hardware_acceleration() - try: - # 检查视频是否有音频轨道并写入文件 - subclip.write_videofile( - video_path, - codec='libx264', - audio_codec='aac', - temp_audiofile='temp-audio.m4a', - remove_temp=True, - audio=(subclip.audio is not None), - logger=None - ) + # 构建ffmpeg命令 + ffmpeg_cmd = ["ffmpeg", "-y"] + + # 添加硬件加速参数（如果可用） + if hwaccel: + if hwaccel == "cuda": + ffmpeg_cmd.extend(["-hwaccel", "cuda"]) + elif hwaccel == "videotoolbox": # macOS + ffmpeg_cmd.extend(["-hwaccel", "videotoolbox"]) + elif hwaccel == "qsv": # Intel Quick Sync + ffmpeg_cmd.extend(["-hwaccel", "qsv"]) + elif hwaccel == "vaapi": # Linux VA-API + ffmpeg_cmd.extend(["-hwaccel", "vaapi", "-vaapi_device", "/dev/dri/renderD128"]) + elif hwaccel == "dxva2": # Windows DXVA2 + ffmpeg_cmd.extend(["-hwaccel", "dxva2"]) + logger.info(f"使用硬件加速: {hwaccel}") + + # 设置输入选项和精确剪辑时间范围 + ffmpeg_cmd.extend([ + "-ss", str(start), # 从这个时间点开始 + "-t", str(duration), # 剪辑的持续时间 + "-i", origin_video, # 输入文件 + "-map_metadata", "-1" # 移除元数据 + ]) + + # 设置视频编码参数 + if hwaccel == "cuda": + ffmpeg_cmd.extend(["-c:v", "h264_nvenc", "-preset", "p4", "-profile:v", "high"]) + elif hwaccel == "videotoolbox": + ffmpeg_cmd.extend(["-c:v", "h264_videotoolbox", "-profile:v", "high"]) + elif hwaccel == "qsv": + ffmpeg_cmd.extend(["-c:v", "h264_qsv", "-preset", "medium"]) + elif hwaccel == "vaapi": + ffmpeg_cmd.extend(["-c:v", "h264_vaapi", "-profile", "high"]) + else: + ffmpeg_cmd.extend(["-c:v", "libx264", "-preset", "medium", "-profile:v", "high"]) - # 验证生成的视频文件 - if os.path.exists(video_path) and os.path.getsize(video_path) > 0: - with VideoFileClip(video_path) as clip: - if clip.duration > 0 and clip.fps > 0: - return {timestamp: video_path} - - raise ValueError("视频文件验证失败") + # 音频编码参数（检查是否有音频流） + audio_check_cmd = ["ffprobe", "-i", origin_video, "-show_streams", "-select_streams", "a", + "-loglevel", "error", "-print_format", "json"] + audio_result = subprocess.run(audio_check_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True) + audio_info = json.loads(audio_result.stdout) if audio_result.stdout else {"streams": []} + has_audio = len(audio_info.get("streams", [])) > 0 + + if has_audio: + ffmpeg_cmd.extend(["-c:a", "aac", "-b:a", "128k"]) + else: + ffmpeg_cmd.extend(["-an"]) # 没有音频 - except Exception as e: - logger.warning(f"视频文件处理失败: {video_path} => {str(e)}") + # 设置输出视频参数 + ffmpeg_cmd.extend([ + "-pix_fmt", "yuv420p", # 兼容性更好的颜色格式 + "-movflags", "+faststart", # 优化MP4文件结构以便快速开始播放 + video_path # 输出文件 + ]) + + # 执行ffmpeg命令 + logger.debug(f"执行命令: {' '.join(ffmpeg_cmd)}") + + process = subprocess.run( + ffmpeg_cmd, + stdout=subprocess.PIPE, + stderr=subprocess.PIPE, + text=True, + check=False # 不抛出异常，我们会检查返回码 + ) + + # 检查是否成功 + if process.returncode != 0: + logger.error(f"视频剪辑失败: {process.stderr}") if os.path.exists(video_path): os.remove(video_path) + return {} + + # 验证生成的视频文件 + if os.path.exists(video_path) and os.path.getsize(video_path) > 0: + # 检查视频是否可播放 + probe_cmd = ["ffprobe", "-v", "error", video_path] + validate_result = subprocess.run(probe_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE) + + if validate_result.returncode == 0: + logger.info(f"视频剪辑成功: {video_path}") + return {timestamp: video_path} - except Exception as e: - logger.warning(f"视频剪辑失败: \n{str(traceback.format_exc())}") + logger.error("视频文件验证失败") if os.path.exists(video_path): os.remove(video_path) - finally: - # 确保视频对象被正确关闭 - try: - video.close() - if 'subclip' in locals(): - subclip.close() - except: - pass + return {} + + except Exception as e: + logger.error(f"视频剪辑过程中发生错误: \n{str(traceback.format_exc())}") + if os.path.exists(video_path): + os.remove(video_path) + return {} return {} +def _detect_hardware_acceleration() -> str: + """ + 检测系统可用的硬件加速器 + + Returns: + str: 可用的硬件加速类型，如果没有找到返回空字符串 + """ + import platform + system = platform.system().lower() + + # 测试常见的硬件加速类型 + acceleration_types = [] + + if system == 'darwin': # macOS + acceleration_types = ["videotoolbox"] + elif system == 'linux': + acceleration_types = ["vaapi", "cuda", "nvenc"] + elif system == 'windows': + acceleration_types = ["cuda", "nvenc", "dxva2", "qsv"] + + for accel in acceleration_types: + test_cmd = [ + "ffmpeg", + "-hide_banner", + "-loglevel", "error", + "-hwaccel", accel, + "-i", "/dev/null", # 这不是实际文件，但是足以测试硬件加速器是否可用 + "-f", "null", + "-" + ] + + try: + result = subprocess.run(test_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, timeout=1) + # 某些硬件加速器会报错，但仍然可以使用，我们主要检查的是CUDA和类似的错误 + stderr = result.stderr.decode('utf-8', errors='ignore') + if result.returncode == 0 or ( + "No such file or directory" in stderr and + not any(x in stderr for x in ["Invalid", "Error", "not supported"]) + ): + logger.info(f"检测到可用的硬件加速器: {accel}") + return accel + except (subprocess.SubprocessError, OSError): + continue + + logger.info("未检测到可用的硬件加速器，将使用软件编码") + return "" + + def clip_videos(task_id: str, timestamp_terms: List[str], origin_video: str, progress_callback=None) -> dict: """ 剪辑视频 diff --git a/config.example.toml b/config.example.toml index 7576350..762651b 100644 --- a/config.example.toml +++ b/config.example.toml @@ -172,12 +172,7 @@ speech_region="" [frames] - skip_seconds = 0 - # threshold（差异阈值）用于判断两个连续帧之间是否发生了场景切换 - # 较小的阈值（如 20）：更敏感，能捕捉到细微的场景变化，但可能会误判，关键帧图片更多 - # 较大的阈值（如 40）：更保守，只捕捉明显的场景切换，但可能会漏掉渐变场景，关键帧图片更少 - # 默认值 30：在实践中是一个比较平衡的选择 - threshold = 30 - version = "v2" + # 提取关键帧的间隔时间 + frame_interval_input = 3 # 大模型单次处理的关键帧数量 vision_batch_size = 10