Understanding U-Net shape constraints

[fepegar]

It was hard to me to understand shape constraints when I first tried NiftyNet. I've seen that others have issues as well (#144, #79, #93, #92, #83, #80, #125...), so I wrote a small tutorial:
https://nbviewer.jupyter.org/gist/fepegar/1fb865494cb44ac043c3189ec415d411

Probably DeepMedic and V-Net would be worth explaining as well.

[tvercaut]

Thanks, I think we should give this sort of documentation more visibility. This particular gist could be a nice first entry in the FAQ (see #137) that needs to be set up.

[yshvrdhn]

@fepegar Is the shape constraints the same for highresnet too ? If not could you point me to a source where I can understand the shape constraints for it ?

[fepegar]

@yshvrdhn, the constraints for HigRes3dNet are in the network README.

I'm not sure why the image size must be divisible by 8, maybe @wyli can help.

[yshvrdhn]

@fepegar thanks! Any idea how I can print the input image shape to the network in order to see what is the actual tensor shape of the input to the network in nifty net?

[fepegar]

The tensor shape of the input to the network is specified by you in the config file, it's the spatial_window_size.

[yshvrdhn]

Thanks! got it working.