/coord-conv-pytorch

All convolution operators replaced with coordinate convolutions

Primary LanguagePython

Coordinate Convolution (Pytorch implementation)

Given a very simple task of predicting the $(i,j)^{th}$ coordinate of a single pixel on a black background, convolutional nets fail. Uber AI propose a simple solution of adding the $i^{th}$ and $j^{th}$ coordinate as additional channels to each input tensor.

The image below (taken from https://eng.uber.com/coordconv/) shows how this is implemented.

Coordinate convolution

Taken and subsequent adaptions added from mkocabas.

Added support for:

  1. cuda.FloatTensors
  2. Transpose convolution