NiftyNet

NiftyNet copied to clipboard

This issue was sparked by the implementation of SubPixel Convolution that is available in Fast.ai [1]

Things to do:

• [x] Restructure SubPixel Convolution since it does not fall in line with the original paper's description
• [x] Add the ICNR initialization [2] of the convolutional layer
• [x] Add AveragePool after the pixel shuffle [3]

[1] PixelShuffle ICNR [2] Checkerboard artifact free sub-pixel convolution [3] Checkerboard artifacts free convolutional neural networks

A new partial limitation has been created by this extension. If ICNR initialization is used then the upsample factor MUST be an integer.

Did the modification on the brach with the same number as the issue, waiting for the person that implemented ICNR to validate my generalization.