Writing a simple Gaussian noise layer in Pytorch

2

Yes, you can move the mean by adding the mean to the output of the normal variable.
But, a maybe better way of doing it is to use the normal_ function as follows:

def gaussian(ins, is_training, mean, stddev):
    if is_training:
        noise = Variable(ins.data.new(ins.size()).normal_(mean, stddev))
        return ins + noise
    return ins
6 Likes