I use the key "fps" instead of "video_fps." ,but it still doesn't work !!!

[jackhexiangyu12]

#65

I use the key "fps" instead of "video_fps." ,but it still doesn't work !!!
messages = [
{
"role": "user",
"content": [
{
"type": "video",

"video": "IMG_3309.MOV",

"video": "test1.mp4",
"max_pixels": 360 * 420,
"fps": 2.0,
},
{"type": "text", "text": text1},
],
}
]

(Qwen2-VL-2B-Instruct1) hxy@thor:~$ python /media/ssd4/hxy/LLaVA-Video-72B-Qwen2/Qwen2-VL-2B-Instruct1.py
Qwen2VLRotaryEmbedding can now be fully parameterized by passing the model config through the config argument. All other arguments will be removed in v4.46
Loading checkpoint shards: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:01<00:00, 1.42it/s]
请输入问题：有什么
qwen-vl-utils using torchvision to read video.
Traceback (most recent call last):
File "/media/ssd4/hxy/LLaVA-Video-72B-Qwen2/Qwen2-VL-2B-Instruct1.py", line 91, in
image_inputs, video_inputs = process_vision_info(messages)
File "/home/hxy/anaconda3/envs/Qwen2-VL-2B-Instruct1/lib/python3.9/site-packages/qwen_vl_utils/vision_process.py", line 332, in process_vision_info
video_inputs.append(fetch_video(vision_info))
File "/home/hxy/anaconda3/envs/Qwen2-VL-2B-Instruct1/lib/python3.9/site-packages/qwen_vl_utils/vision_process.py", line 260, in fetch_video
video = VIDEO_READER_BACKENDSvideo_reader_backend
File "/home/hxy/anaconda3/envs/Qwen2-VL-2B-Instruct1/lib/python3.9/site-packages/qwen_vl_utils/vision_process.py", line 193, in _read_video_torchvision
total_frames, video_fps = video.size(0), info["video_fps"]
KeyError: 'video_fps'

[nlp4whp]

I fixed it by pip install "torchvision<0.19.0" or "torchvision>0.19.0"`

For me, this KeyError is from here: vedio, audio, info = io.read_video(...), info would be like {'video_fps': 25.0, 'audio_fps': 48000}
but when I used torchvision==0.19.0, info would be an empty dict {}