SNN with BatchNorm layers doesn't give the correct validation accuracy when net.eval() is set (there is some conflicting issue between net.eval() and BatchNorm2d in SNNs,)

[gwgknudayanga]

Hi Experts,

I trained a SNN which has BatchNorm2d layers also.
But at the validation it always gives very low accuracy though the training accuracy is good.
When i remove the net.eval() (still keeping the torch.no_grad()) at the validation time, then validation accuracy is good.
When I removed the batchnorm layers this issue involve with net.eval() doesn't exist. It seems there is some conflicting issue between net.eval() and BatchNorm2d in SNNs,

Thank you.
Rgds.
Udayanga