data:
  dataset: LSUN
  category: church_outdoor
  image_size: 64
  channels: 3
  logit_transform: false
  uniform_dequantization: false
  gaussian_dequantization: false
  random_flip: true
  rescaled: false
  num_workers: 32

model:
  sigma_begin: 140
  num_classes: 788
  ema: true
  ema_rate: 0.999
  spec_norm: false
  sigma_dist: geometric
  sigma_end: 0.01
  normalization: InstanceNorm++
  nonlinearity: elu
  ngf: 128
  unet: false
  dropout: 0.0

sampling:
  noise_first: false
  clamp: false
  experiment_name: 'logits_evolution_tracking_best_point'

training:
  log_all_sigmas: true  # whether to send training information to tensorboard

optim:
  weight_decay: 0.000
  optimizer: Adam
  lr: 0.0001
  beta1: 0.9
  beta2: 0.999
  adv_beta1: -0.5
  adv_beta2: 0.9
  amsgrad: false
  eps: 0.00000001
  momentum: 0.9

fast_fid:
  batch_size: 50
  num_samples: 1000

adversarial:
  lambda_dae: 1           # multiplier term for Lp loss function (only used in GAN setting)
  lambda_D: 1             # multiplier term for GAN loss of the discriminator'
  lambda_G_gan: 1         # multiplier term for GAN loss of the generator
  D_steps: 1              # Discriminator steps per Generator step
  adv_loss: LSGAN         # 'GAN, LSGAN, HingeGAN, RpGAN, RaGAN, RaLSGAN, RaHingeGAN'
  arch: 2                 # 0 is DCGAN_D0, 1 is DCGAN_D1, 2 is BigGAN
  spectral: false         # If True, spectral normalization in D
  no_batch_norm_D: false  # Not active in BigGAN
  adv_clamp: true     # If True, do not clamp output of score network before giving to discriminator

biggan:
  ch: 64          # Number of channels
  thin: false     # If True, use thin Discriminator (D_ch = 0 with D_thin = True leads to )
  kernel_size: 3
  attn: '64'       # Number of attention filters (If 0, do not use self-attention)
  n_classes: 1    # Number of classes of the dataset (If = 1 leads to Unconditional GAN)  # inactive
  init: xavier    # Type of init,: ortho, xavier, N02