soniasargol/monotone-GAN

Conditional Sampling With Monotone GANs

Reproducing results in Conditional Sampling With Monotone GANs.

To start running the examples, clone the repository:

$ cd monotone-gans/

Installation

Run the following commands in the command line to install the necessary libraries and setup the conda envirement:

conda env create -f environment.yml
source activate mgan

Synthetic examples

Run the script below for creating the synthetic examples. Figures will be saved in the figs/ directory. Checkpoints of the network weights will be saved in the checkpoints/ directory.

$ python src/monotone-gan.py  --example simple --experiment synthetic_4 --equation 4

$ python src/monotone-gan.py  --example simple --experiment synthetic_5 --equation 5

$ python src/monotone-gan.py  --example simple --experiment synthetic_6 --equation 6

In all the above cases, the command line input argument --phase inference can be used to run the inference phase once there exists a saved checkpoint.

MNIST Inpainting

Run the command below to train the MNIST inpainting network. The training will take about 12 hours on a small GPU.

python src/monotone-gan.py  --example mnist --experiment mnist_inpainting --wd 0.0 --batch_size 128 --max_epoch 300 --lr 0.0002

Figures will be saved in the figs/ directory. Checkpoints of the network weights will be saved in the checkpoints/ directory. The command line input argument --phase inference can be used to run the inference phase once there exists a saved checkpoint.

Author

Sonia Sargolzaei