wgan.yaml

discriminator_params:
  feature_maps: 64
  image_channels: 3

generator_params:
  latent_dim: 256
  feature_maps: 64
  image_channels: 3

loss_params:
  top_k: False
  instance_noise: False
  latent_dim: 256 
  clip_value: 0.01

optimizer_params:
  learning_rate: 0.0002
  beta1: 0.5

exp_params:
  model_name: 'wgan'
  dataset: celeba
  data_path: "kaggle"
  refinement: null
  img_size: 64
  crop_size: 128
  batch_size: 256 # Better to have a square number

trainer_params:
  gpus: 1
  max_epochs: 10

# If top_k is true then we use these hyperparameters
top_k_params:
  k: 200
  gamma: 0.95

# If instance_noise is true then we use these hyperparameters
instance_noise_params:
  noise_level: 0.1
  gamma: 0.99

# Sampling algo if refinement is not null
sampling_params:
  eta: 0.01
  noise_factor: 0.001 
  num_steps: 50