AverageUnpooling layer for PyTorch (Proposal)

[revanurambareesh]

🚀 Feature

PyTorch supports MaxPooling, AveragePooling layers and also supports MaxUnpooling layer. However AverageUnpooling layer is currently not supported in PyTorch.

Motivation

As a researcher I find that having AverageUnpooling layer as part of PyTorch will be very useful for following reasons:

1. Common Use-case: CNN autoencoder with average pooling. Other use-cases include image reconstruction.
2. The average unpooling layer is also used (although not as extensively as MaxUnpooling) in research and I strongly feel that this layer should exist in PyTorch.
3. This will also help address average unpooling queries, such as one posted here in discuss.pytorch.org

Pitch

Implementing AverageUnpooling layer similar to https://github.com/HyeonwooNoh/caffe/blob/master/src/caffe/layers/unpooling_layer.cpp

EDIT: PyTorch's interpolation suits my requirements. Hence closed the issue.

[fmassa]

Isn't AverageUnpooling very similar to F.interpolate?

[zou3519]

What exactly are the semantics of AverageUnpooling?

[revanurambareesh]

Yeah, It is very similar to F.interpolate, but with padding. I was thinking that having it as a layer with the name AvgUnpool would be beneficial. However, since it is semantically similar to interpolation, I am closing the issue. Pytorch's interpolation suffices my requirements. Thanks!