GPU

[1995Emma]

I tried to run the code on two 2080Ti, however it raised "RuntimeError: CUDA out of memory. Tried to allocate 16.00 MiB (GPU 0; 11.00 GiB total capacity; 9.18 GiB already allocated; 11.94 MiB free; 9.23 GiB reserved in total by PyTorch)"
So I wonder if the code can only be trained on Titan RTX GPU? Or is there any ops that can solve this except the hyperparameters?

[HReynaud]

Hello,
With a batch size of 1 and keeping all the other hyperparameters, the network will take 12737MiB on any GPU during training and the 2080Ti has 10986MiB, so unfortunately the pretrained weights will not fit in these GPUs.
You could keep the network architecture untouched and reduce the number of frames (ds_max_length) to 64. This will have little effect on the performance at test time and reduce the memory footprint to 9307MiB, which would fit a 2080Ti.