Hi,
I want to process parameters of a network only in forward pass and do not have backprop for the op. For example, add noise to the every parameter of every layer in net. But have no effect in backprop. How can I achieve this?? Also if I try to modify self.conv.weight.data then I run into CUDA related errors.
I’m very new to pytorch so sorry if its a silly question.
class net(nn.Module):
def __init__(self):
self.conv = nn.Conv2d(...)
def forward(self,x):
self.conv.weight = self.conv.weight + normal(mean=0,stddev=0.2)
return self.conv(x)