https://github.com/breandan/hgan

Hyper volume maximization for GAN training
https://github.com/breandan/hgan

Last synced: 3 months ago
JSON representation

Hyper volume maximization for GAN training

• Host: GitHub
• URL: https://github.com/breandan/hgan
• Owner: breandan
• Created: 2018-03-31T15:21:03.000Z (over 8 years ago)
• Default Branch: master
• Last Pushed: 2018-03-31T15:34:29.000Z (over 8 years ago)
• Last Synced: 2025-06-04T19:18:41.438Z (about 1 year ago)
• Language: Python
• Size: 107 KB
• Stars: 0
• Watchers: 3
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Hyper Volume Generative Adversarial Network - hGAN

Replication of [Stabilizing GAN Training with Multiple Random Projections](https://arxiv.org/abs/1705.07831) and extension including training with multi-objective training via hyper volume maximization

## To run

Download the [cropped and aligned version of CelebA](http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html) 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