About GPU memory usage

[QAQEthan]

If non-local is applied to the low-level feature map, CUDA out of memory will happen.Is this due to the amount of memory required to compute the Attention matrix?
Looking forward to your reply

[buncybunny]

I'm also experiencing CUDA out of memory issue with non-local block. I'm trying to use non-local block at the top of my network, which is for bbox regression conv head in faster r-cnn. Do you guys have any ideas to address this?

[AlexHex7]

@Monkey-D-Luffy-star @vombategeht Hi~

The larger the size (height, width, depth) of feature maps is, the more memories the matrix multiplication will occupy.

When I encounter this problem，I will：

1. reduce the batch size
2. downsample the feature maps
3. move non-local block to high-level position
4. make some optimization. For example, follow the idea of papers:
   4.1. GCNet:Non-local Networks Meet Squeeze-Excitation Networks and Beyond
   4.2. Compact Generalized Non-local Network
5. follow the idea of transformer block: split tokens (height x width x depth) in several groups, then do self-attention in each group.
6. or directly try using transformer block

[QAQEthan]

@AlexHex7 Thx, benefit a lot.