Some questions and a suggestion

[vinsis]

Hi Yunpeng, just read your paper and have a couple of quick questions:

• Why do you implement batchnorm + relu before the convolution in the BN_AC_CONV3D class here?

• In your paper you say: we set the number of the first-layer output channels to be 4 times smaller than its input channels, ... Hence, shouldn't this line be written as shown below?

# current version: num_ix = int(num_mid/4)
num_ix = int(num_in/4)

I believe self.conv_i1 and self.conv_i2 are the layers for the multiplexer. Or am I getting it wrong?

• This is a theoretical question: is the role of multiplexer to rearrange information across channels so as to minimize information loss for each fiber?

Lastly, a small suggestion: since you are using PyTorch v0.4, you need not use Variable anymore. Hence you can write this line as:

data = torch.randn(1,3,16,224,224)

Thank you.

[cypw]

Hi @vinsis,

Thank you for the questions.

• We put "batchnorm + relu" before each convolution layer since it is more clear and easier to design. It is actually following this paper from Kaiming. We haven't tried putting them after the convolution, but it should give the same or even better performamce, see here.

• You are right. Please refer to the code for the correct version. We will update the paper soon. Thank you for pointing it out.

• Multiplexer can pass information across fibers, so that features learned by one fiber can be seen by other fibers. It has the effect of increasing the overall learning capacity and minimize information loss.

• Thank you for the suggestion.

Thank you.

[vinsis]

Thanks @cypw for replying. Closing the issue now.