Use this method need the super memory? about vsfa HOT 6 CLOSED

Input:
!python /content/VSFA/test_demo.py --video_path='/content/VSFA/ACMMM_2019_supplementary_44/video pair/better-MAH_02693_8s.mp4' --model_path=/content/VSFA/models/VSFA.pt
output:
RuntimeError: CUDA out of memory. Tried to allocate 3.96 GiB (GPU 0; 15.90 GiB total capacity; 9.73 GiB already allocated; 2.47 GiB free; 12.72 GiB reserved in total by PyTorch)

from vsfa.

This question is my computer configuration or using way?

from vsfa.

@JavanTang You should read the test_demo.py carefully.
Or at least you can run python test_demo.py --help to get more information.
Actually, you can set frame_batch_size to 16, then it only needs about 9G GPU memory.

The supplementary video pairs are just used for illustrating the motivation of this paper.
We provide the test_demo.py with the test.mp4 example (which has a smaller resolution than my supplementary video pairs, so it can be run with default frame_batch_size, 32).

To save GPU memory, a smaller frame_batch_size is preferred.

from vsfa.

@lidq92 Well. Video length increases, can increase memory? Try to test 32 seconds video on my side, but my memory more than 12 gb, is this normal?

from vsfa.

@JavanTang Read the test_demo.py, or at least I strongly suggest you run python test_demo.py --help to get more information.

It is obvious that the video spatial resolution has an effect on the GPU memory.
"It only needs about 9G GPU memory for a 1080P video with setting frame_batch_size=16.

from vsfa.

@lidq92 Ok, thank you

from vsfa.