sverdoot / maxent-gan

Experiments on sampling from GAN with probabilistic constraints

• Experiments on sampling from GAN with probabilistic constraints
  • Installation
    • StudioGAN:
  • Usage
  • Tutorials
  • Results

Installation

Create environment and set dependencies:

conda create -n constrained_gan python=3.8

curl -sSL https://install.python-poetry.org | python3 -
poetry config virtualenvs.create false

conda activate constrained_gan
conda install tensorflow-gpu==2.4.1
poetry install

Check whether TF can see GPU:

import tensorflow.compat.v1 as tf

tf.test.gpu_device_name()

If FID computation stucks that might be caused by conflict with scipy. Try running export MKL_NUM_THREADS=1.

StudioGAN:

./scripts/install_studiogan.sh

To use StudioGAN as is following is needed:

import sys
from pathlib import Path

from maxent_gan.utils.general_utils import ROOT_DIR

sys.path.append(Path(ROOT_DIR, 'studiogan'))

---

poetry install

Put CIFAR-10 into directory data/cifar10 using this script

import torchvision.datasets as dset

cifar = dset.CIFAR10(root='data/cifar10', download=True)

Make bash scripts runable

chmod +x -R scripts/*.sh

Usage

Download checkpoints:

dvc pull

To use wandb:

wandb login

python run.py configs/exp_configs/dcgan-inception.yml configs/targets/discriminator.yml \
    configs/gan_configs/cifar-10-dcgan.yml configs/feature_configs/inception.yml \
    configs/mcmc_configs/{ula/mala/isir/ex2mcmc/flex2mcmc}.yml configs/mcmc_exp.yml

Tutorials

• docs/maxent_sample_tutorial.md, notebooks/maxent_sample_tutorial.ipynb

• docs/maxent_train_tutorial.md, notebooks/maxent_train_tutorial.ipynb

Results