SDPoint not working on CIFAR #1
Description
Hello,
I am trying to run the given code for CIFAR10/100. However, It gives me the following error during training the model:
RuntimeError: adaptive_avg_pool2d_backward(): Expected grad_output to have non-zero size for non-batch dimensions, but grad_output has sizes [128, 512, 0, 0] with dimension 2 being empty
As per my understanding, by downsampling at a certain point, the feature map shape reaches 0x0, which results in this error. I am using resnet18 but the error stays even if I change the model to resnext or preresnet. Can you share your insights if you face the same issue while running the model?
Or maybe, if you have the implementation of the CIFAR code, you can share it here.