PyTorch Implementation of the Coupled Generative Adversarial Networks (CoGAN)
General
This is a PyTorch implementation of the Coupled Generative Adversarial Netowork algorithm. For more details please refer to our NIPS paper or our arXiv paper. Please cite the NIPS paper in your publications if you find the source code useful to your research.
Ming-Yu Liu, Oncel Tuzel "Coupled Generative Adversarial Networks" NIPS 2016
Installation
In your python package, install pytorch and torchvision. You also need yaml, Python Opencv and Google Logging to run the code.
Usage
Simple example
Train the CoGAN network to learn to generate digit images and the corresponding edges images of the digits images without the need of corresponding images in the two domains in the training dataset.
cd src;
python train_cogan_mnistedge.py --config ../exps/mnistedge_cogan.yaml;
After 5000 iterations, you will see the generation results in outputs/mnistedges_cogan/ and they should look like.
Domain adaptation using all training images
Train the CoGAN network to unsupervisedly adapt a digit classifier from the MNIST domain to the USPS domain by using all the images in the training sets. Use 60000 images from the MNIST training set when unsupervisedly adapting from MNIST to USPS. Use 7438 images from the USPS training set when unsupervisedly adapting from USPS to MNIST.
cd src;
python train_cogan_mnist2usps.py --config ../exps/mnist2usps_full_cogan.yaml;
python train_cogan_usps2mnist.py --config ../exps/usps2mnist_full_cogan.yaml;
You will see the accuracy of the adapted classifier in the test set in the target domain in the log file. The best accuracy in your log files should be something like
| Setting | MNIST to USPS | USPS to MNIST |
|---|---|---|
| CoGAN | 0.95XX | 0.93XX |
Domain adaptation using a subset of training images
Train the CoGAN network to unsupervisedly adapt a digit classifier from the MNIST domain to the USPS domain by using subsets of the training sets. Use 2000 images from the MNIST training set when unsupervisedly adapting from MNIST to USPS. Use 1800 images from the USPS training set when unsupervisedly adapting from USPS to MNIST.
cd src;
python train_cogan_mnist2usps.py --config ../exps/mnist2usps_small_cogan.yaml;
python train_cogan_usps2mnist.py --config ../exps/usps2mnist_small_cogan.yaml;
You will see the accuracy of the adapted classifier in the test set in the target domain in the log file. The best accuracy in your log files should be something like
| Setting | MNIST to USPS | USPS to MNIST |
|---|---|---|
| CoGAN | 0.94XX | 0.92XX |
CoGAN Network Architecture
CoGAN learn to generate corresponding smile and non-smile faces
CoGAN learn to generate corresponding faces with blond-hair and without non-blond-hair
CoGAN learn to generate corresponding faces with eye-glasses and without eye-glasses
CoGAN learn to generate corresponding RGB and depth images
Copyright 2017, Ming-Yu Liu All Rights Reserved
Permission to use, copy, modify, and distribute this software and its documentation for any non-commercial purpose is hereby granted without fee, provided that the above copyright notice appear in all copies and that both that copyright notice and this permission notice appear in supporting documentation, and that the name of the author not be used in advertising or publicity pertaining to distribution of the software without specific, written prior permission.
THE AUTHOR DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR ANY PARTICULAR PURPOSE. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.