config_cv_MSKCC.yaml

save_dir: "../results/MSKCC/" #directory to save all cross-validation results
exp_id: null #specific id to refer to the cv experiment (saving results). If null timestamp will be used.
n_gpu: 0 #number of GPUs (if 0 the model will be trained and tested on CPU)

#define number of jobs to run in parallel (useful when there are several repeats of the cv scheme!)
parallelization:
  n_jobs_repeats: 1

#define cross-validation data
cross_val_data:

  # Define a train-validation split if needed
  validation_split: null # 0.15

  # 1. Load raw data from your files
  loader:
    type: "load_MSKCC_multimoda"
    args:
      clinical_file: "../data/MSKCC/clinicals.csv"
      radiomics_file: "../data/MSKCC/radiomics.csv"
      pathomics_file: "../data/MSKCC/pathomics.csv"
      omics_file: "../data/MSKCC/omics.csv"
      pdl1_file: "../data/MSKCC/pdl1.csv"

      clinical_features: ['age', 'pack_years', 'ecog', 'albumin', 'dnlr', 'brain_mets', 'liver_mets', 'tumor_burden',
                          'therapy_line', 'recieves_combo_therapy', 'site_lung', 'recieves_pdl1_therapy', 'hist_adeno']
      radiomic_features:
      omics_features:
      pathomics_features:
      outcome: 'RECIST'

  # 2. Build a dataset from the loaded raw data
  dataset: "MSKCCDataset"

  # processing steps to apply on the different data modalities
  processing:

    radiomics_PL:
      scaler:
        strategy: 'robust'

    radiomics_LN:
      scaler:
        strategy: 'robust'

    radiomics_PC:
      scaler:
        strategy: 'robust'

    pathomics:
      scaler:
        strategy: 'robust'

    omics:
      scaler:
        strategy: 'robust'

    pdl1:
      scaler:
        strategy: 'robust'

    clinicals:
      imputer:
        numericals: [1]
        categoricals: [10]
      scaler:
        strategy: 'robust'

    #multimodal processing step (if any), such as dmultipit.dataset.transformers.CustomSelection
    multimodal:

   # 3. Define data loader with or without oversampling and data augmentation
  sampler: False #whether to use oversampling for the minority class during training (only for classification !)
  drop_modalities: False #whether to use data augmentation during training (dropping randomly modalties)

  training_data_loader:
    batch_size: 256
    shuffle: true
    num_workers: 0

#define specific radiomic transformer for MSKCC data (see dmultipit.dataset.transformers.MSKCCRadiomicsTransform)
#Note: lesion_type argument and order argument (in the config_architecture file) shoud specify the same radiomic lesion types
radiomics_transform:
  lesion_type: ['PC', 'LN', 'PL']
  robustness_cutoff: 0.15
  outlier_cutoff: 6
  l1_C: 0.1
  aggregation: 'mean'

#define cross-validation settings
cross_val:
  n_repeats: 1 #number of repetitions for cross validation scheme
  seed: 3 #fix seed for reproducibility
  strategy: "StratifiedKFold" #cross-validation strategy
  n_splits: 10 #number of folds

#define training settings
training:

  epochs: 125 #number of training epochs

  #define training loss (losses available in dmultipit.model.loss)
  loss:
    type: "BCELogitLoss"
    args:
  l2_penalty: 0.001
  attention_penalty: null
  balanced_weights: True
  metrics: [ "balanced_accuracy", "roc_auc" ]

  #define training optimizer
  optimizer:
    type: "Adam"
    args:
      lr: 0.01
#      weight_decay: 0.005

  #define learning rate scheduler (optional)
  lr_scheduler:
    type:

  #define monitoring/checkpoint parameters
  save_period: null #if not null a model checkpoint will be saved every *save_period* epochs
  save_best_only: null  #if True only the best model according to the monitoring function is saved as checkpoint
  print_period: null #if not null the model performance metrics will be printed every *print_period* epochs
  verbosity: 1 #control the level of information
  tensorboard: false #allow tensorboad visualization
  monitor: "off" #define monitoring strategy
  ema_alpha: 0.9 #parameter for the exponential moving average (ema) smoothing algortihm
  early_stop: 50 #number of epochs to wait for model improvement before stopping (no early stopping if 0)

  #define pseudo-labelling strategies (not used for MSKCC data !)
  pseudo_labelling: False

  unlabelled_loss:
    type:
    args:

  unlabelled_scheduler:
    type:

#define test settings
testing:
  verbosity: 0
  loss:
    type: "BCELogitLoss"
    args:
  metrics: ["balanced_accuracy_with_logit", "roc_auc"]