RajK853/GAN

Personal repository to learn about different types of GAN models.

GAN

Personal repository to learn about different types of GAN models using Keras.

Conda setup

1. Clone this repo.

git clone https://github.com/RajK853/GAN.git $SRC_DIR

2. Create and activate conda environment.

cd $SRC_DIR  
conda env create -f environment.yml    
conda activate gan-env

Implementations

GAN

Implementation of normal Generative Adversarial Network.

ACGAN

Implementation of Auxiliary Classifier Generative Adversarial Network.

BiGAN

Implementation of normal Bidirectional Generative Adversarial Network.

Example:

Train models

1. Create a YAML config file (let's say config_1.yaml) as:

default: &default_config
  epochs: 1000
  latent_size: 50
  batch_size: 128
  evaluate_interval: 5
  lr: 0.0003
  num_evaluates: 10

GAN_latent_50:
  <<: *default_config
  model: GAN

GAN_latent_100:
  <<: *default_config
  model: GAN
  latent_size: 100

ACGAN:
  <<: *default_config
  model: ACGAN

BiGAN:
  <<: *default_config
  model: BiGAN

2. Train the models by loading the parameters from the above YAML config file as:

python train.py config_1.yaml

The above config file will train GAN, ACGAN and BiGAN models with two different latent_size values for the GAN model only.

Any configuration with the key name with the prefix default will not be executed by default.

Feedforward layer configurations can be passed via layer_configs argument. Please look in example_configs directory for the sample YAML configuration file.