config_train_MSKCC.yaml

save_dir: "../results/MSKCC/" #directory to save all training loggings and checkpoints
run_id: null #specific id to refer to the training experiment (saving results). If null timestamp will be used.
n_gpu: 0 #number of GPUs to train the model (if 0 the model will be trained on CPU)

#define training data
training_data:

  # 1. define a train-validation split if needed
  validation_split: 0.2
  val_index:

  # 2. 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'

  # 3. 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:

  # 4. 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)

  data_loader:
    batch_size: 32
    shuffle: false
    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 training settings
training:

  seed: 3 #seed for reproducibility
  epochs: 125 #number of training epochs

  #define training loss (losses available in dmultipit.model.loss)
  loss:
    type: "BCELogitLoss"
    args:
  l2_penalty: 0.001 #regularization strength for l2 penalty (no penalty if null)
  attention_penalty: null #regularization strength for attention penlaty (see attention_norm attribute from the attention model)
  balanced_weights: True #whether to use balanced class weights for binary classification
  metrics: [ "balanced_accuracy_with_logit", "roc_auc" ] #metrics computed at each epoch (for training and validation if a validation set is specified)

  #define training optimizer
  optimizer:
    type: "Adam"
    args:
      lr: 0.01
  #    weight_decay: 0.005
  #     momentum: 0.9

  #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: True #if True only the best model according to the monitoring function is saved as checkpoint
  print_period: 2 #if not null the model performance metrics will be printed every *print_period* epochs
  verbosity: 1 #control the level of information
  tensorboard: true #allow tensorboad visualization
  monitor: "ema_min val_loss" #define the monitoring algorithm (first word) and the monitoring metric (second word)
                              #the monitoring metric should correspond to the metrics specified above or "loss" and
                              #the "val" prefix can be used to indicate that it should be computed on the validation set
                              #For the algorithm, "min" indicates that the best model corresponds to the minimum value
  ema_alpha: 0.6 #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: