config.yml

model: # AR{ESNet model parameters
  input_shape: [256, 256]
  aenc: "arpesnet"
  kwargs:
    kernel_size: 11
    kernel_decay: 2
    n_layers: 1
    start_channels: 4
    max_channels: 32
    n_blocks: 5
    input_shape: [256, 256]
    relu: PReLU
    relu_kwargs: {num_parameters: 1}

preprocessing: # preprocessing steps, applied in order, before training
  Resize: [256, 256] # should be the same as model input_shape
  NormalizeMinMax: [0,100]

training_augmentations: # augmentations applied during training, in order
  RandomResizedCrop:
    size: [256, 256]
    scale: [0.8, 1.] #[0.6, 1.0]
    ratio:  [0.8, 1.2] # [0.5, 1.5]
  Flip:
    p: 0.5
  NormalizeMinMax: [0,100]

noise_augmentations: # augmentations applied during training, in order
  SetRandomPoissonExposure: [50_000,100_000_000]
  NormalizeMinMax: [0,100]

validation_augmentations:
  Resize: [256, 256]
  NormalizeMinMax: [0,100]

loss: # loss function and parameters
  criteria: ['mse'] # 'mse','psnr','ssim'... or list of them
  weights: [1.0] # weights for each loss function
  contractive: False # contractive autoencoder loss
  lam: 0.0001 # weight for contractive loss

optimizer: # optimizer and parameters
  name: 'Adam' 
  kwargs:  
    lr: 0.001
    weight_decay: 0.0

train: # training parameters
  split_ratio: [0.8,0.2] # train/test split
  batch_size: 32 # batch size
  n_epochs: 1000 # number of epochs
  shuffle: True
  drop_last: True
  denoiser: False

device: # device settings
  seed: 0 # random seed 
  use_gpu: True # use GPU if available
  copy_to_cuda: True # copy data to GPU, if False, data is copied to GPU on-the-fly and pin_memory is enabled