| import decord |
| import numpy as np |
| from qwen_vl_utils import process_vision_info |
|
|
|
|
| def build_video_item(path, fps, max_frames, max_pixels): |
| return {"type": "video", "video": str(path), "fps": fps, |
| "max_frames": max_frames, "max_pixels": max_pixels} |
|
|
|
|
| def prepare(messages, video_path): |
| out = process_vision_info(messages, return_video_kwargs=True) |
| videos = out[1] |
| vk = out[2] if len(out) > 2 else {} |
| vk = {k: (v[0] if isinstance(v, list) and len(v) == 1 else v) for k, v in vk.items()} |
| vr = decord.VideoReader(str(video_path), num_threads=1) |
| n_native = len(vr) |
| native_fps = float(vr.get_avg_fps()) |
| n = videos[0].shape[0] |
| idx = np.linspace(0, n_native - 1, n).round().astype(int) |
| meta = [{"fps": native_fps, "duration": float(n_native / native_fps), |
| "total_num_frames": int(n_native), "frames_indices": idx}] |
| vk["video_metadata"] = meta |
| return videos, vk |
|
|
