# coord-conv-pytorch

**Repository Path**: kaluo_zZ/coord-conv-pytorch

## Basic Information

- **Project Name**: coord-conv-pytorch
- **Description**: All convolution operators replaced with coordinate convolutions 
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2021-06-29
- **Last Updated**: 2021-06-29

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# 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](image8.jpg "From https://eng.uber.com/coordconv/")

Taken and subsequent adaptions added from mkocabas.


Added support for:
1. cuda.FloatTensors
2. Transpose convolution