Keras implementation of CoordConv for all Convolution layers
Keras implementation of CoordConv from the paper An intriguing failing of convolutional neural networks and the CoordConv solution.
Extends the
CoordinateChannelconcatenation from only 2D rank (images) to 1D (text / time series) and 3D tensors (video / voxels).
Import
coord.pyand call it before any convolution layer in order to attach the coordinate channels to the input.
There are 3 different versions of CoordinateChannel - 1D, 2D and 3D for each of
Conv1D,
Conv2Dand
Conv3D.
from coord import CoordinateChannel2Dprior to first conv
ip = Input(shape=(64, 64, 2)) x = CoordinateChannel2D()(ip) x = Conv2D(...)(x) # This defines the
CoordConv
from the paper. ... x = CoordinateChannel2D(use_radius=True)(x) x = Conv2D(...)(x) # This adds the 3rd channel for the radius.
The experiments folder contains the
Classificationof a 64x64 grid using the coordinate index as input as in the paper for both
Uniformand
Quadrantdatasets.
First, edit the
make_dataset.pyfile to change the
typeparameter - to either
uniformor
quadrant. This will generate 2 folders for the datasets and several numpy files.
The uniform dataset model can be trained and evaluated in less than 10 epochs using
train_uniform_classifier.py.
|Train | Test | Predictions |
|:---: | :---: | :-----------: |
| |
|
|
The quadrant dataset model can be trained and evaluated in less than 25 epochs using
train_quadrant_classifier.py
|Train | Test | Predictions |
|:---: | :---: | :-----------: |
| |
|
|
To see if the implementation of CoordConv index concatenation is correct, please refer to the numpy implementations in the
checksdirectory, for the implementation of all 3 versions.
This implementation of the coordinate channels creation differs slightly from the original paper.
The major difference is that for 2/3D Convolutions, it may not be the case that the height and width are the same for all layers. The original implementation would throw an error due to shape mismatch during the concatenation.
To over come this, the
np.ones()operation which occurs at the first of every channel is modified and a few transpose operations are added to account for this change.
This modification along with some transpose operations allows for height and width to be different and still work.
Theano is partially supported with the
coord_theano.pyscript and using passing a static batch size to the Input layer.