Multi-objective training of Generative Adversarial Networks with multiple discriminators

Code for reproducing the experiments reported on our paper that proposes a multi-objective optimization perspective on the training of GANs with multiple discriminators.

To run

Data

MNIST and Cifar-10 will be downloaded automatically

Stacked MMNIST has to be built and dumped into an .hdf file prior to training. Download the data from https://ufile.io/k854s, or build it by running:

python gen_data.py --data-size 50000 --data-path /path/to/download/mnist --out-file /out/file/path

CelebA: Download the cropped and aligned version of CelebA and unzip it

Cats dataset can be downloaded from: https://ufile.io/u6i98

Training generators

cd to the the folder corresponding to the desired dataset and run:

python train.py --help

to get the list of arguments.

Example for CelebA:

optional arguments:
  -h, --help            Show this help message and exit
  --batch-size N        Input batch size for training (default: 64)
  --epochs N            Number of epochs to train (default: 50)
  --lr LR               Learning rate (default: 0.0002)
  --beta1 lambda        Adam beta param (default: 0.5)
  --beta2 lambda        Adam beta param (default: 0.999)
  --ndiscriminators NDISCRIMINATORS
                        Number of discriminators. Default=8
  --checkpoint-epoch N  Epoch to load for checkpointing. If None, training
                        starts from scratch
  --checkpoint-path Path
                        Path for checkpointing
  --data-path Path      Path to data
  --workers WORKERS     Number of data loading workers
  --seed S              Random seed (default: 1)
  --save-every N        How many epochs to wait before logging training
                        status (default: 3)
  --train-mode {vanilla,hyper,gman,gman_grad,loss_delta,mgd}
                        Select train mode. Default is vanilla (simple average
                        of Ds losses)
  --disc-mode {RP,MD}   Multiple identical Ds with random projections (RP) or
                        Multiple distinct Ds without projection (MD)
  --nadir-slack nadir   Factor for nadir-point update. Only used in hyper mode
                        (default: 1.5)
  --alpha alpha         Used in GMAN and loss_del modes (default: 0.8)
  --no-cuda             Disables GPU use

Required args

For both cifar10_32 and 64, --fid-model-path has to be specified to allow for FID computation at train time. Download the model used in our experiments

Computing FID with a pre-trained model

Download the pretrained classifier for evaluation at https://ufile.io/8udto

Example for calculating FID using a ResNet for a generator trained on CIFAR-10:

cd to common and run

python compute_FID.py --model-path Path --data-stat-path Path --data-path Path --fid-model-path Path --model-cifar resnet --nsamples 100000 --ntests 10 --dataset cifar10

Computing coverage for Stacked MNIST

cd to stacked_mnist, download the pretrained classifier for evaluation at https://ufile.io/8udto (same as for FID). First, compute coverage for real data (for calculating the KL divergence), by running:

python coverage_real_data.py --classifier-path Path --data-path Path --out-file Path

Example:

python coverage.py --cp-folder Path --classifier-path Path --data-stat-path Path --out-file Path --n-tests 10 --n-samples 10000

Generating samples

cd to common and run

python gen_samples.py --cp-path Path --nsamples 10000 --out-path Path

TensorFlow implementation for computing Inception Score and FID using author's original implementation

Inception Score: https://github.com/openai/improved-gan/blob/master/inception_score/model.py

FID: https://github.com/bioinf-jku/TTUR

Tested with

• Python 3.6
• Pytorch > 0.4.1

Pending:

• Improve solver for inner optimization on MGD's update direction calculation

Citation

@article{albuquerque2019multi,
  title={Multi-objective training of Generative Adversarial Networks with multiple discriminators},
  author={Albuquerque, Isabela and Monteiro, Jo{\~a}o and Doan, Thang and Considine, Breandan and Falk, Tiago and Mitliagkas, Ioannis},
  journal={arXiv preprint arXiv:1901.08680},
  year={2019}
}