RuntimeError: Given groups=1, weight of size [400, 256, 1, 1], expected input[64, 3, 1, 1] to have 256 channels, but got 3 channels instead

[trungmanhhuynh]

I got this error when testing using pre-trained model. Please help ''' mmskl configs/recognition/st_gcn_aaai18/ntu-rgbd-xsub/test.yaml ''' Error: ''' RuntimeError: Given groups=1, weight of size [400, 256, 1, 1], expected input[64, 3, 1, 1] to have 256 channels, but got 3 channels instead '''

[trungmanhhuynh]

Really dont know, but you added --gpus 1 to the command, it will work ''' mmskl configs/recognition/st_gcn_aaai18/ntu-rgbd-xsub/test.yaml --gpus 1 '''

The code does not work properly with multiple gpus ?

[yosagaf]

Try using single GPU to see : CUDA_VISIBLE_DEVICES=0 mmskl configs/recognition/st_gcn_aaai18/ntu-rgbd-xsub/test.yaml

[yysijie]

@trungmanhhuynh, @yosagaf, This bug is caused by Pytorch 1.5.0. nn.Parameter and nn.DataParallel can not be used together. (Please refer to https://github.com/pytorch/pytorch/issues/36035 for more information) I recommend you to roll back to Pytorch 1.2.0. Please install it by:

# CUDA 9.2
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=9.2 -c pytorch

# CUDA 10.0
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch

We will solve this problem in the next version of mmskeleton.

[trungmanhhuynh]

Thank you, @yysijie @yosagaf

[JiabenChen]

I wonder have you fixed this problem...? @yysijie

[TOMMYWHY]

Hi yan, do you have any schedule to fix it? @yysijie

[TOMMYWHY]

Hi yan, the #36035 has been fixed by torch. However, the latest torch(1.11) has the same error.... @yysijie