fix(pytorch): correct epoch-based learning rate decay behavior (#410)

fix(pytorch): correct epoch-based learning rate decay behavior (#410) #202