Doubling channels before pooling

[alexklibisz]

Hello,

Looking at your code in model.py, it seems that you overlooked this statement on page 4 of the Cicek paper:

Like suggested in [13] we avoid bottlenecks by doubling the number of channels already before max pooling. We also adopt this scheme in the synthesis path.

I annotated the figure below:

I believe this wasn't the case in 2D UNet, so it's likely a copy-paste artifact. I'm not sure what difference it will make in terms of performance, but I figured it's worth mentioning.

[ellisdg]

Nice catch!
Does #12 fix the issue?

[alexklibisz]

Thank you. It does for the downward pass (i.e. left side of figure). However the bottom "block" should be Pool -> Conv 256 -> Conv 512 -> Deconv 512 -> Concat. It's not apparent in the paper or figure, but if you look at the prototxt you'll see it. It's also helpful to visualize the prototxt, which you can do here.

I might submit a PR later today that fixes it. That being said - I'm not sure if it will improve performance. So it goes with deep nets. :)