Hi, I am a newbie to PyTorch and I got some trouble when I am implementing one model related to 3D point cloud. The problem is that I use a point pair feature representation for each local patch, so the input size is [bs, 2048, 1024, 4], where 2048 is the number of interest points, and 1024 is the number of neighboring points per local patch, and 4 is the dimension of point pair feature. And the network architecture is relatively simple (just an Auto-Encoder), but even if I use a batch size 1, it still can not run because GPU out of memory. I use torchsummary to see the parameter size and intermediate variable size and find that the intermediate variable takes up some much memory. So I wonder is there any methods to reduce the size of intermediate variable?
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv1d-1 [2048, 64, 1024] 320
BatchNorm1d-2 [2048, 64, 1024] 128
ReLU-3 [2048, 64, 1024] 0
Conv1d-4 [2048, 128, 1024] 8,320
BatchNorm1d-5 [2048, 128, 1024] 256
ReLU-6 [2048, 128, 1024] 0
Conv1d-7 [2048, 256, 1024] 33,024
BatchNorm1d-8 [2048, 256, 1024] 512
ReLU-9 [2048, 256, 1024] 0
Linear-10 [2048, 1024, 512] 360,960
Linear-11 [2048, 1024, 512] 262,656
Encoder-12 [2048, 1, 512] 0
Conv1d-13 [2048, 256, 1024] 131,840
ReLU-14 [2048, 256, 1024] 0
Conv1d-15 [2048, 128, 1024] 32,896
ReLU-16 [2048, 128, 1024] 0
Conv1d-17 [2048, 64, 1024] 8,256
ReLU-18 [2048, 64, 1024] 0
Conv1d-19 [2048, 32, 1024] 2,080
ReLU-20 [2048, 32, 1024] 0
Conv1d-21 [2048, 4, 1024] 132
ReLU-22 [2048, 4, 1024] 0
Conv1d-23 [2048, 256, 1024] 132,352
ReLU-24 [2048, 256, 1024] 0
Conv1d-25 [2048, 128, 1024] 32,896
ReLU-26 [2048, 128, 1024] 0
Conv1d-27 [2048, 64, 1024] 8,256
ReLU-28 [2048, 64, 1024] 0
Conv1d-29 [2048, 32, 1024] 2,080
ReLU-30 [2048, 32, 1024] 0
Conv1d-31 [2048, 4, 1024] 132
ReLU-32 [2048, 4, 1024] 0
Decoder-33 [2048, 1024, 4] 0
================================================================
Total params: 1,017,096
Trainable params: 1,017,096
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 32.00
Forward/backward pass size (MB): 68936.00
Params size (MB): 3.88
Estimated Total Size (MB): 68971.88
----------------------------------------------------------------