from time_r1.utils.qwen_vl_utils_prepare import floor_by_factor, FRAME_FACTOR, smart_resize, ceil_by_factor import decord import torch import os import tqdm import glob import multiprocessing from torchvision import io, transforms from torchvision.transforms import InterpolationMode from functools import partial from time_r1.utils.io import load_jsonl import debugpy # try: # # 5678 is the default attach port in the VS Code debug configurations. Unless a host and port are specified, host defaults to 127.0.0.1 # debugpy.listen(("localhost", 9511)) # print("Waiting for debugger attach") # debugpy.wait_for_client() # except Exception as e: # pass import os os.environ['DECORD_EOF_RETRY_MAX'] = '20480' SAVE_DIR = '/data/shuimu.chen/videomarathon/downloaded_videos/activitynet_192_50_frame_cache_unit8' VIDEO_PATH = "/data/shuimu.chen/videomarathon/downloaded_videos/activitynet" FPS_MIN_FRAMES = 4 FPS_MAX_FRAMES = 50 # 建议根据你的显存情况调整，60-80 是常用范围 FRAME_FACTOR = 2 # def get_video_tensor(video_path, target_fps=2, image_factor = 28, min_pixels = 28 * 28 * 4, max_pixels = 28 * 28 * 128): # """ # 将视频以固定帧率提前抽帧、解码保存为tensor，用于后续训练 # """ # vr = decord.VideoReader(video_path) # total_frames, video_fps = len(vr), vr.get_avg_fps() # min_nframes = ceil_by_factor(FPS_MIN_FRAMES, FRAME_FACTOR) # ：ceil_by_factor 是为了保证帧数是 FRAME_FACTOR 的倍数 # max_frames = floor_by_factor(min(FPS_MAX_FRAMES, total_frames), FRAME_FACTOR) # nframes = total_frames / video_fps * target_fps # 理论帧数 # if nframes > total_frames: # print(f"smart_nframes: nframes[{nframes}] > total_frames[{total_frames}]") # nframes = min(min(max(nframes, min_nframes), max_frames), total_frames) # nframes = floor_by_factor(nframes, FRAME_FACTOR) # frame_idx = torch.linspace(0, total_frames - 1, nframes).round().long().tolist() # 均匀采样 # frame_tensor = vr.get_batch(frame_idx).asnumpy() # frame_tensor = torch.tensor(frame_tensor).permute(0, 3, 1, 2) # Convert to TCHW format # sample_fps = nframes / max(total_frames, 1e-6) * video_fps # height, width = frame_tensor.shape[2], frame_tensor.shape[3] # resized_height, resized_width = smart_resize( # height, # width, # factor=image_factor, # min_pixels=min_pixels, # max_pixels=max_pixels, # ) # frame_tensor = transforms.functional.resize( # frame_tensor, # [resized_height, resized_width], # interpolation=InterpolationMode.BICUBIC, # antialias=True, # ) # if frame_tensor.dtype != torch.uint8: # frame_tensor = frame_tensor.clamp(0, 255).to(torch.uint8) # frame_cache = { # "frame_tensor": frame_tensor, # "fps": sample_fps, # } # return frame_cache import cv2 # 必须引入 cv2 import numpy as np # ... 其他引用保持不变 ... def get_video_tensor(video_path, target_fps=2, image_factor=28, min_pixels=28 * 28 * 4, max_pixels=28 * 28 * 128): """ 内存优化版： 1. 优先 Decord (CPU 模式) 2. 失败回退 OpenCV (流式读取，不占内存) """ # ------------------------------------------------------------------ # 1. 尝试读取视频元数据 & 选择读取器 # ------------------------------------------------------------------ vr = None cap = None method = "unknown" total_frames = 0 video_fps = 0.0 # === 方案 A: Decord (CPU) === try: vr = decord.VideoReader(video_path, ctx=decord.cpu(0), num_threads=1) total_frames = len(vr) video_fps = vr.get_avg_fps() method = "decord" except Exception: # === 方案 B: OpenCV (流式读取，内存安全) === try: cap = cv2.VideoCapture(video_path) if cap.isOpened(): total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) video_fps = cap.get(cv2.CAP_PROP_FPS) method = "cv2" except Exception: pass # 如果两者都挂了，或者读不到帧数，直接报错 if total_frames <= 0 or video_fps <= 0: if cap: cap.release() raise RuntimeError(f"Failed to read video (or empty): {video_path}") # ------------------------------------------------------------------ # 2. 计算采样逻辑 (保持不变) # ------------------------------------------------------------------ min_nframes = ceil_by_factor(FPS_MIN_FRAMES, FRAME_FACTOR) max_frames_limit = floor_by_factor(min(FPS_MAX_FRAMES, total_frames), FRAME_FACTOR) nframes = total_frames / video_fps * target_fps nframes = min(min(max(nframes, min_nframes), max_frames_limit), total_frames) nframes = floor_by_factor(nframes, FRAME_FACTOR) nframes = int(nframes) # 均匀采样索引 frame_idx = torch.linspace(0, total_frames - 1, nframes).round().long() # ------------------------------------------------------------------ # 3. 提取帧数据 (内存安全的方式) # ------------------------------------------------------------------ frames_list = [] if method == "decord": # Decord: 批量读取 (高效) try: frame_tensor = vr.get_batch(frame_idx.tolist()).asnumpy() frame_tensor = torch.tensor(frame_tensor).permute(0, 3, 1, 2) # (T, H, W, C) -> (T, C, H, W) except Exception as e: # 如果 decord 虽然打开了但在 get_batch 挂了 (比如坏帧)，尝试切回 cv2 method = "cv2" cap = cv2.VideoCapture(video_path) # 重新打开 if method == "cv2": # OpenCV: 逐帧 Seek 读取 (内存占用极低) for idx in frame_idx.tolist(): cap.set(cv2.CAP_PROP_POS_FRAMES, idx) ret, frame = cap.read() if ret: # BGR -> RGB frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) frames_list.append(frame) else: # 如果读取失败（坏帧），补上一帧或者全黑帧 if len(frames_list) > 0: frames_list.append(frames_list[-1]) else: frames_list.append(np.zeros((224, 224, 3), dtype=np.uint8)) cap.release() if len(frames_list) == 0: raise RuntimeError(f"OpenCV failed to read any frames: {video_path}") # Stack 起来: List[H,W,C] -> (T, H, W, C) frame_tensor = torch.tensor(np.array(frames_list)).permute(0, 3, 1, 2) # (T, C, H, W) # ------------------------------------------------------------------ # 4. 后处理 (Resize + Uint8转换) # ------------------------------------------------------------------ sample_fps = nframes / max(total_frames, 1e-6) * video_fps frame_tensor = frame_tensor.float() height, width = frame_tensor.shape[2], frame_tensor.shape[3] resized_height, resized_width = smart_resize( height, width, factor=image_factor, min_pixels=min_pixels, max_pixels=max_pixels, ) frame_tensor = transforms.functional.resize( frame_tensor, [resized_height, resized_width], interpolation=InterpolationMode.BICUBIC, antialias=True, ) if frame_tensor.dtype != torch.uint8: frame_tensor = frame_tensor.clamp(0, 255).to(torch.uint8) frame_cache = { "frame_tensor": frame_tensor, "fps": sample_fps, } return frame_cache def process_single_video(video_path, target_fps=1, image_factor=28, min_pixels=28*28*128, max_pixels=28*28*256): os.makedirs(SAVE_DIR, exist_ok=True) # 【修复1】文件名处理：建议保留部分目录结构或使用Hash避免重名 # 这里简单处理：如果有重名风险，建议用 hashlib 对路径做 hash 作为前缀 video_filename = os.path.basename(video_path) save_path = os.path.join(SAVE_DIR, video_filename + '.frame_cache') # 如果已经存在且文件正常，跳过（双重保险） if os.path.exists(save_path): return None # 表示已跳过 try: frame_cache = get_video_tensor(video_path, target_fps, image_factor, min_pixels, max_pixels) torch.save(frame_cache, save_path) return None # 成功 except Exception as e: # 【修复2】返回错误信息而不是直接打印，方便主进程统计 return f"Error: {video_path} -> {str(e)}" def prepare_frame_cache(video_root = None, dataset_path=None, num_workers=8, target_fps=2, overwrite=False, image_factor=28, min_pixels=28*28*128, max_pixels=28*28*256): video_root = VIDEO_PATH dataset_path = None if dataset_path is not None: # 修改这里：直接使用json.load而不是load_jsonl import json with open(dataset_path, 'r', encoding='utf-8') as f: video_data = json.load(f) # 这是JSON数组 # 提取video_path字段，并去重（同一个视频可能有多条记录） video_paths = set() # 使用set去重 for v in video_data: if "video_path" in v: video_paths.add(v["video_path"]) elif "video" in v: # 如果有video字段，可能需要拼接路径 video_path = os.path.join(video_root, v["video"]) video_paths.add(video_path) video_list = list(video_paths) else: video_list = glob.glob(os.path.join(video_root, "*.mp4")) # 【修复3】修正跳过逻辑，检查目标目录 if not overwrite: print("Checking existing files in target directory...") # 预先计算目标文件名集合 existing_files = set(os.listdir(SAVE_DIR)) if os.path.exists(SAVE_DIR) else set() new_video_list = [] skipped_count = 0 duplicate_name_warning = [] seen_basenames = set() for v in video_list: basename = os.path.basename(v) target_name = basename + '.frame_cache' # 检查重名冲突 if basename in seen_basenames: duplicate_name_warning.append(v) seen_basenames.add(basename) if target_name in existing_files: skipped_count += 1 else: new_video_list.append(v) video_list = new_video_list print(f"Skipped {skipped_count} existing files.") if duplicate_name_warning: print(f"⚠️ 警告: 发现 {len(duplicate_name_warning)} 个同名不同路径的视频！这会导致覆盖问题。") # 打印前5个例子 for d in duplicate_name_warning[:5]: print(f" - Duplicate basename: {d}") print(f"Starting processing {len(video_list)} videos...") failed_videos = [] with multiprocessing.Pool(processes=num_workers) as pool: func = partial(process_single_video, target_fps=target_fps, image_factor=image_factor, min_pixels=min_pixels, max_pixels=max_pixels) # 使用 tqdm 接收返回值 for result in tqdm.tqdm(pool.imap_unordered(func, video_list), total=len(video_list)): if result is not None: # 如果 result 是字符串，说明报错了 failed_videos.append(result) print("\n" + "="*30) print("Processing Complete.") if failed_videos: print(f"❌ {len(failed_videos)} videos failed to process:") for err in failed_videos: print(err) else: print("✅ All videos processed successfully.") if __name__ == "__main__": import fire fire.Fire(prepare_frame_cache)