This repository houses my implementations of the CoordConv layer as introduced in paper - An intriguing failing of convolutional neural networks and the CoordConv solution. The layer is implemented in TensorFlow and has taken some inspiration from the code for the layer as shown in the supplementary part of the paper.
CoordConv is an extension to Convoluton layers just to give the neural network extra understanding of the coordinates of the pixels. The layer keeps two properties of Convolution layer intact:
- few parameters
- efficient computation
But it gives a degree of freedom to network in case of translation invariance. It lets the network to decide whether it needs the said property or not.
 |
|---|
| Source: https://arxiv.org/pdf/1807.03247 |
This repository currently houses code for -
- Supervised Coordinate Classification over the NotSoClever dataset discussed in paper. Read more about it in the notebook here
To visualize the dataset, run
python visualize_coords.py. For some helpful command line arguments usepython visualize_coords.py -h
I have planned on working with this layer to make:
- Autoencoders
- GANs
- Time Series Related Tasks.