Questions about GPU memory usage in non local blocks.

[seominseok0429]

I am an undergraduate student studying Deep Learning in Korea. I am impressed by your paper and doing experiments.

I thought gpu memory usage was only related to the parameters of the model, but using non-local blocks would use too much gpu memory.

Adding two non-local blocks to the model of the i3d inception structure, the total number of parameters increased by about 7%.

However, gpu memory usage has increased significantly. For example, in the i3d inception structure, six batches were available for each gpu, but adding two non-local blocks maximized two batches for each gpu.

Can you tell me if this is common or if I miscoded it?

(input size=224x224, time=64)