RuntimeError: CUDA out of memory about stcn HOT 3 CLOSED

eval_generic does not store all the frames in the GPU simultaneously but still 4GB is not a lot and 1000 frames are too many. Increasing mem_every seems to be the easiest way.

from stcn.

Thank you for the response.
I tried using eval_generic, but I get "DefaultCPUAllocator: not enough memory" thrown by "images = torch.stack(images, 0)" in generic_test_dataset.py. I believe the images in this line contain all frames from a video and I was wondering if it would be possible to load the data in a different way so this problem could be circumvented.
Increasing mem_every doesn't seem to solve the problem as these errors occur before the inference core is initiated (the for data in progressbar line calls get_item on all videos and this is where the memory error occurs)

from stcn.

Ah, I see. That's on CPU memory, not GPU. You can rewrite the loader (generic_test_dataset.py) and eval_generic.py to load one frame from disk to CPU at a time.

from stcn.