Unpooling and deconvolution layers

[Daemon2017]

Hello!

Is it possible to create a segmentation network using your library? As I know, for that I need an unpooling and deconvolutional layers.

[cbovar]

Hi!

Currently there is no unpooling or deconvolution. It's something I have planned to do but it's low on my priority list.

PR welcome :)

[tzaeru]

Def interested about this too. I'm not sure if my own skill/time (the two tend to be related!) is enough to add this feature, but many of my use cases would need proper learning-enabled upsampling in the form of transposed convolution.

[cbovar]

I think I will get cracking on it. It would be very useful for generative networks.
IIRC, the maths are very similar to convolutions (some indexes swapped) and cudnn already handles that.

[cbovar]

Some useful links:

• Convolution and Transposed Convolution algos: https://arxiv.org/pdf/1603.07285.pdf
• Convolution, GPU double: https://github.com/cbovar/ConvNetSharp/blob/master/src/ConvNetSharp.Volume.GPU/Double/Volume.cs#L245
• Convolution, GPU single: https://github.com/cbovar/ConvNetSharp/blob/master/src/ConvNetSharp.Volume.GPU/Single/Volume.cs#L246
• Convolution, CPU double: https://github.com/cbovar/ConvNetSharp/blob/master/src/ConvNetSharp.Volume/Double/Volume.cs#L145
• Convolution, CPU single: https://github.com/cbovar/ConvNetSharp/blob/master/src/ConvNetSharp.Volume/Single/Volume.cs#L142
• Tests: https://github.com/cbovar/ConvNetSharp/blob/master/src/ConvNetSharp.Volume.Tests/VolumeTests.cs#L252