ADD Selection channel in another network

[ZirshaRiaz]

Hi, I really appreciate your research and efforts. I want to extend this work. I am trying to implementing selection channel on a CNN architecture. I am confused in coding section of m.alloc() and dealloc(). I have replaced conv2d with SelectiveConv as you mentioned in the paper.
I would be thankful for your concern and help. Thanks

[jh-jeong]

Hi, @ZirshaRiaz, thanks for your interest.

After replacing Conv2d to SelectiveConv2d, you can call m.dealloc() and m.realloc() during training whenever needed (e.g. for every epoch). This part can be found in L220 of main.py:

for m in model.modules():
    if type(m).__name__ in ['SelectiveConv2d']:
        if epoch < 0.5 * args['num_epochs']:
            m.dealloc()
            m.realloc()

Please let me know if I misunderstood your question.

[ZirshaRiaz]

Thanks for your fast response. I was doing this thing but I was facing issue in iterating through network. It seems working now. Let's see how it performs after training. Moreover, I would like to ask what if I add SelectiveConv2d in Unet. Will it work? and UNet also contains convTranspose layers. How to add it in Unet?

Again I really appreciate your concern.
Thanks,
Zirsha