config_training_explained.md

In this file, the input parameters of config_training.json are explained:

• epochs (int): number of training epochs

• lambda_loss (float): value that weights the two terms (dice and BCE) of the hybrid loss; must be between 0 and 1

• batch_size (int): batch size to use during training

• lr (float): learning rate to use during training

• conv_filters (list): number of filters to use in the convolution layers of the UNET

• fold_to_do (int): training fold that will be done; has to be one of [1,2,3,4,5]; we do one fold at a time so more folds can be done in parallel if more than one GPU is available

• use_validation_data (str): if "True", validation data is created to monitor the training curves

• percentage_validation_subs (float): percentage of subjects to keep for validation; must be between 0. and 0.3

• n_parallel_jobs (int): number of CPUs to create the training dataset in parallel (the higher, the faster!); if set to -1, all available CPUs are used

• data_path (str): path to directory containing dataset of patches (created during step 1 of the pipeline)

• input_ds_identifier (str): unique name given to rename output folders

• path_previous_weights_for_pretraining (str): path where previous weights are stored (used for pretraining); if empty, no pretraining is done

• train_test_split_to_replicate (str): path to directory containing the test subjects of each CV split; this can be found in the github directory /Forced_CV_for_reproducibility. There is no need to specify the fold since this will be added automatically using fold_to_do. This argument is needed in order to re-create the exact same cross-validation split used for the original paper. In this way, results are truly comparable.