config.yaml

general:
  label_type: &label_type # isup or bin or survival
  number_of_classes: &n_classes # 6 or 2 or 28
  image_directory: &directory # path to directory where images are
  seed: &seed # set for reproducibility
  image_channels: &img_channels # 3 for rgb
  use_mongo_db: # leave blank if not used, else specify the url (e.g. db_name:PORT_NUMBER)

# ------  ------  ------  ------  ------  ------  ------

data_generation:
  cache: # either None (leave blank) to cache to memory or /path/to/cache/folder where to cache dataset to memory
  # CSV FILES
  train_csv_file: #  full path to train.csv
  train_batch_size: # int
  valid_csv_file: #  full path to valid.csv
  valid_batch_size: # int
  test_csv_file: #  full path to test.csv (or also valid.csv during training)
  # WHERE TO FIND INFORMATION IN CSV FILE AND WHAT INFORMATION TO USE
  annotation_column: # column of the main class label, as it is called in csv file (isup, bin, relapse_time)
  additional_columns: # list of dict with
    # - internal_column_name: ['name_of_column_in_csv', 'input'] if used as additional input (like binary prediction)
    # - internal_column_name: ['name_of_column_in_csv', 'label'] if used as additional label (like censoring status)
    # - internal_column_name: ['name_of_column_in_csv', '-'] if just extra information (like age for inspection later)
    # - censored: ['censored', 'label']
  drop_cases: # list of lists
    # - ['value_to_drop', 'internal_column_name'] if cases should be dropped based on target label (e.g. label -1)
    # - ['value_to_drop', 'annotation_column'] if cases should be dropped based on additional column (e.g. age 0)
    # - ['-1', 'censored']
  # IMAGE PREPROCESSING
  patching:
    patch_size: # integer width and height of patch
    overlap: # integer (0)
    n_patches: &n_patches # integer how many patches to cut (if more than possible, additional patches are white)
                           # 0 if no patching is wanted
    order: # string: shuffle, original, ranked, shuffle_ranked
    keep_original: # False or integer, if in addition to patches, the original image should be kept in dataset too
  augmentation_config:
     rotation: # float, degree to rotate (if random, max degree)
     width_shift: # float
     height_shift: # float
     brightness: # float
     horizontal_flip: # True or False (if random, put true to randomly flip or not)
     vertical_flip: # True or False (if random, put true to randomly flip or not)
     fill_mode: # 'constant' to fill new pixels with random value (e.g. after rotation)
     cval: # color value to fill new pixels with , e.g. white: 255

  random_augmentation: # False or True: if the augmentation should be random
  resize_x: &resize_x
  resize_y: &resize_y
  resize: [*resize_x, *resize_y]   # leave blank, is filled automatically with values defined before
  # GENERAL INFORMATION - leave blank, is all filled with values defined above
  channels: *img_channels
  directory: *directory
  label_type: *label_type
  number_of_classes: *n_classes
  seed: *seed

# ------  ------  ------  ------  ------  ------  ------

model:
  name: #m_bin, m_isup, ecare_net
  base_model: # InceptionV3 or path/to/folder where model_json.json is stored
  additional_input: # list, e.g. ['bin0', 'bin1'] to add prediction for relapse in first two years yes+no
  dense_layer_nodes: # list of integers, e.g. [64]
  rnn_layer_nodes: # list of integers, e.g. [256]
  keras_model_params:
    weights: # None or 'imagenet'
    input_shape: [*resize_x, *resize_y, *img_channels] # input shape as list with [m,n,3] for rgb (no batch size) - blank!
  n_patches: *n_patches
  n_classes: *n_classes
  # only used for survival model
  mil_layer: # boolean
  self_attention: # boolean
  cut_off_layer: # int or name of layer after which model should be cut


# ------  ------  ------  ------  ------  ------  ------

training:
  epochs: # int, how many epochs to train
  initial_epoch: 0
  monitor_val: &monitor_val # e.g. 'val_loss'
  callbacks:
    - name: # e.g. 'EarlyStopping'
      params:
        patience: #100
        min_delta: #0.01
        monitor: #*monitor_val
        restore_best_weights: #true
  optimizer: # dict
    name: # String for tf.keras optimizer, e.g. 'Adam',
    params: # list of additional parameters as dict
      - learning_rate: # float
  loss_fn: # string, either custom or tf.keras function, e.g. 'binary_crossentropy'
  compile_metrics: # list of metrics, e.g. ['BinaryAccuracy']
  compile_attributes: # dictionary with additional attributes like class_weight
  model_save_path: # folder where model should be saved
  class_weight: # True or False, whether to weight classes based on how often they are present in dataset
  weighted_metrics:  # True or False, whether to weight metrics based on how often classes are present in dataset

# ------  ------  ------  ------  ------  ------  ------

evaluation:
  metrics:
    # - accuracy

inference:
  image_orig_size:
  inference_id: 79
  metrics: ['auc', 'brier', 'c_index']
  eval_csv: