config.yaml

# General settings
experiment_name: ""   # name of the experiment
device: 0   # -1 for CPU, otherwise GPU id
seed: 27  # random seed
data_base_path: ""    # path to the base directory containing the datasets


# Training
training:
  epochs: 10    # number of epochs
  lr: !!float 1e-3    # learning rate
  batch_size: 16     # batch size
  num_workers: 20   # number of workers for data loading
  log_images_frequency: 1000  # log input images every n batches
  resume_training: false    # resume training and logging of the experiment with the same name

  data:
    patch_size: 224   # patch size for training
    max_distortions: 4  # maximum number of distortions to apply. Must be in the range [0, 7]
    num_levels: 5   # number of distortion levels to consider. Must be in the range [1, 5]
    pristine_prob: 0.05   # probability of not distorting images during training

  optimizer:
    name: SGD   # optimizer name
    momentum: 0.9   # momentum
    weight_decay: !!float 1e-4    # weight decay

  lr_scheduler:
    name: CosineAnnealingWarmRestarts   # learning rate scheduler name
    T_0: 1    # T_0 for CosineAnnealingWarmRestarts
    T_mult: 2   # T_mult for CosineAnnealingWarmRestarts
    eta_min: !!float 1e-6   # eta_min for CosineAnnealingWarmRestarts


# Validation
validation:
  frequency: 1    # validate every frequency epochs
  num_splits: 10   # number of splits to consider for each dataset
  alpha: 0.1  # alpha value for the regression
  visualize: true   # visualize embeddings with t-SNE for KADID10K dataset
  visualization:
    tsne:
      n_components: 3   # number of components for t-SNE
      perplexity: 30    # perplexity for t-SNE
      n_iter: 1000    # number of iterations for t-SNE
    umap:
      n_components: 3   # number of components for UMAP
      n_neighbors: 25   # number of neighbors for UMAP
      min_dist: 0.2   # min_dist for UMAP
      metric: euclidean   # metric for UMAP
  datasets:  # datasets to use for validation
    - live
    - csiq
    - tid2013
    - kadid10k
    - flive
    - spaq


# Test
test:
  batch_size: 16    # batch size
  num_workers: 20   # number of workers for data loading
  num_splits: 10   # number of splits to consider for each dataset
  grid_search: true   # if True, grid search on the validation splits is used to find the best alpha value for the regression
  alpha: 0.1  # alpha value for the regression when grid search is not used
  crop_size: 224  # crop size for inference
  datasets:  # datasets to use for test
    - live
    - csiq
    - tid2013
    - kadid10k
    - flive
    - spaq


# Model
model:
  temperature: 0.1    # temperature for the NT-Xent loss
  encoder:    # encoder parameters
    embedding_dim: 128    # embedding dimension
    pretrained: true    # if True, use ImageNet pretrained weights
    use_norm: true    # if True, normalize the embeddings


# Logging
logging:
  use_wandb: true   # if True, use wandb for logging
  wandb:
    online: true    # if True, log online to wandb
    project: ""   # wandb project name
    entity: ""    # wandb entity name