a bug?

[kanezaki]

a bug?

[kanezaki]

This is not a bug. “self.bn0” is a batch normalization layer for each convolutional layer component in a standard use case. It still works even if you disable them. “self.bn1” is the batch normalization layer that I particularly mentioned in my paper, which is mandatory for the proposed method. (It is actually uncommon to insert a batch normalization layer after the final classification layer, though.) Thanks!

[kktsubota]

Thank you for your reply.

I can solve the question thanks to you.
What I meant is that self.bn0 is used twice when args.nConv == 2 and it seems to be weird. In a standard case, each layer is constructed for each use case (an example of a standard use case is here). So, in this case, I think constructing self.bn2 like self.bn0 and using self.bn2 instead of self.bn0 in L61 is usual. I thought this works unexpectedly but I confirmed this also works well.

I have another question.
When args.nConv > 2 is used, this code uses self.conv2 more than one time. This means the weight of self.conv2 is shared for all the Component in Feature Extractor. (Component and Feature Extractor comes from fig. 1 in your paper.) In usual case, I think each conv layer should have its own weight.
So I want to confirm whether this is your intent or not.

[kanezaki]

Oh, I got what you meant. You're right, I didn't intend to share the weights among different conv layers. I fixed the code and now it seems it's working perfect. Please check it out if you have time. Thanks!!