How much GPU memory is needed when evaluating DAVIS?

I tried ‘eval_video_segmentation.py’ with one 16GB V100, but CUDA out of memory always occurs when processing the 28th video, i.e. ‘shooting’.

RuntimeError: CUDA out of memory. Tried to allocate 8.90 GiB (GPU 0; 15.78 GiB total capacity; 3.46 GiB already allocated; 1.59 GiB free; 12.99 GiB reserved in total by PyTorch)

However, the length of ‘shooting’ is only 39, which is shorter than the previous videos. I am confused about why the out of memory happens here.

Did you use 32GB V100 for inference?

[mathildecaron31]

Closing the issue, feel free to re-open if you have further question.

[pansanity666]

Ha yes you’re right I removed this feature when refactoring the code for simplicity. I should remove the argument to avoid confusion… Well in that case you can either

• run code on cpu for the shooting video
• train on GPU with more memory (I used 32Gb V100 GPU for 8x8 models)
• resize the shorter side to a smaller resolution (by changing scale_size there https://github.com/facebookresearch/dino/blob/ba9edd18db78a99193005ef991e04d63984b25a8/eval_video_segmentation.py#L224 and there https://github.com/facebookresearch/dino/blob/ba9edd18db78a99193005ef991e04d63984b25a8/eval_video_segmentation.py#L197 )

What model are you using ? ViT-small 16x16 fits into memory with a 16Gb GPU in my experiments.

I am using ViT-small 8x8, and 16x16 works well in my experiments too. May be the torch.bmm operation leads to the OOM. I will try the cpu inference for shooting. Thank you for your suggestion and reply.

Best,