breandan / hGAN

Hyper volume maximization for GAN training

Hyper Volume Generative Adversarial Network - hGAN

Replication of Stabilizing GAN Training with Multiple Random Projections and extension including training with multi-objective training via hyper volume maximization

To run

Download the cropped and aligned version of CelebA and unzip it

python train.py --ndiscriminators 12

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 is 5
  --hyper-mode          enables training with hypervolume maximization
  --nadir-factor nadir  Factor of the max disc loss to initialize nadir point
                        (default: 50.0)
  --no-cuda             Disables GPU use

Tested with

• Python 3.6
• Pytorch 0.3.0

To do

• Scheduler for the nadir point

Collaborators: Isabela Albuquerque, Breandan Considine