Errors when obtaining gradients for `nn.batch_norm`

[Lyken17]

Thanks for the information! This is exactly what I need.