kobiso / cbam-tensorflow-slim Goto Github PK

Hello！I have a question: should I train my model until the loss becomes the end learning rate(0.0001)? If I stop training my model when the loss is 0.027, then my model will perform badly?

Hello,
thank you for your code.
I am wondering if you can provide the pretrained model of the resnet_v2 with cbam attention model?
Thanks!!!

Dear kobiso:

Thanks for your re-implementations of cbam-tensorflow-slim and SENet-tensorflow-slim.

However, I find that the position you add cbam block is different from the paper "CBAM: Convolutional Block Attention Module" for the inception-resnet-v2. The same situation also happens for the inception-resnet-v2 in SENet-tensorflow-slim.

From my point of view, in both papers, i.e., "CBAM: Convolutional Block Attention Module" and "Squeeze-and-Excitation Networks", the attention modules should be placed between the original blocks of the inception-resnet-v2. So for the inception-resnet-v2, code of adding cbam and SENet modules would be as:

net += scaled_up
if activation_fn:
net = activation_fn(net)
if cbam_block:
net=cbam_block(net, 'cbam_block')
if se_block:
net = se_block(net, 'se_block')
which means that the attention modules would be added after "net += scaled_up".

Also in the inception-resnet-v2 of cbam-tensorflow-slim, you add the attention module in the scaled_up branch as " scaled_up= cbam_block(scaled_up, 'cbam_block')" before "net += scaled_up", while in the the inception-resnet-v2 of SENet-tensorflow-slim, you add the attention module in the identity branch as " net = se_block(net, 'se_block')" before "net += scaled_up".

I don't know if you make these modifications because these modifications show better performance in your own re-implementations. Or, do I make mistakes in understanding the way of adding these two attention modules to the inception-resnet-v2?

Thank you again for your hard work.

All the best.

Can you share some code related to visualization。thank you