titu1994/tf_GON

Tensorflow 2.x implementation of Gradient Origin Networks

Gradient Origin Network for Tensorflow 2.0

Tensorflow 2.0 port of Gradient Origin Networs paper from the original author's implementation found at - https://github.com/cwkx/GON.

Usage

MNIST

MNIST models (4.3k parameter models) can be trained in under 500 steps to good quality, so they dont have weights restoration. They take just a few minutes to train, even on CPU. Simply run mnist_example.py, and you it will create a directory called mnist/images which stores 3 kinds of images - reconstruction, sampled images and spherical interpolation between images.

CIFAR 10

CIFAR 10 models are much larger - 260k parameters, with a larger latent space (256 dim), and need to be trained for a lot longer (100k train steps). Therefore weights have been provided along with log of training.

Simply run cifar10_train.py which takes around 12 - 14 hours on a good GPU to train.

To evaluate results with pre-trained model, run cifar10_eval.py. It will create images under cifar10/images/*_eval.png

Requirements

pip install -r requirements.txt.

Citation

Please cite the original authors for their work -

@article{bondtaylor2020gradient,
  title={Gradient Origin Networks},
  author={Sam Bond-Taylor and Chris G. Willcocks},
  journal={arXiv preprint arXiv:2007.02798},
  year={2020}
}