Configuration

The configuration file is in the format of a .json file. Most settings must be set explicitly in the configuration file.

The configuration file is split in several sections. There are no requirements for ordering within these sections, or for ordering of the sections themselves.

General

• wandb_project_name

  The name of the weights and biases project. Weights and biases is setup in the project for experiment tracking and logging.

• experiment_name

  The name of the weights and biases experiment.

• experiment_tags

  List of tags to place on the run in weights and biases.

• checkpoint_dir

  The local directory where the model checkpoints are stored.

• batch_size

  Number of training examples passed in one training step.

• epochs

  The number of full dataset forward iterations the model should be trained with.

• num_workers

  Number of workers processes for the data loading. If 0, multi-process data loading is disabled.

• gpus

  The number of gpus to be used. If 0, the training runs on CPU.

• prediction_count

  The number of sample predictions to be generated for the validation set. This option is intended for use cases where you want to assess a model’s quality using sample predictions.

• prediction_dir

  The local directory to store the sample predictions.

• random_state

  Constant to ensure reproducibility of random operations.

[model_config] section

The model_config section specifies parameters to setup the segmentation model architecture and losses.

• architecture

  The name of the architecture to use. Allowable values are: "u_net".

Note

If the model architecture of your choice is not yet included in the framework, it can be added by subclassing models.pytorch_model.PytorchModel().

• optimizer

  The name of the algorithm used to calculate the loss and update the weights. Allowable values are: "adam" and "sgd" (gradient descent).

• loss_config

  Dictionary with loss parameters. Mandatory is the key "type" with one of the allowable values: "cross_entropy", "dice", "cross_entropy_dice", "general_dice", "fp", "fp_dice", "focal". More detailed documentation and configuration options of the losses can be looked up in functional.losses.

• learning_rate

  The step size at each iteration while moving towards a minimum of the loss. Defaults to 0.0001.

• num_levels

  Number of levels (encoder and decoder blocks) in the U-Net. Defaults to 4.

• dim

  The dimensionality of the U-Net. Allowable values are: 2 and 3. Defaults to 2.

• model_selection_criterion

  The criterion for selecting the best model for checkpointing. Defaults to "loss".

• train_metrics

  A list with the names of the metrics that should be computed and logged in each training and validation epoch of the training loop. Available options: "dice_score", "sensitivity", "specificity", "hausdorff95". Defaults to ["dice_score"].

• train_metric_confidence_levels

  A list of confidence levels for which the metrics specified in the train_metrics parameter should be computed in the training loop (trainer.fit()). This parameter is used only for multi-label classification tasks. Defaults to [0.5].

• test_metrics

  A list with the names of the metrics that should be computed and logged in the model validation or testing loop (trainer.validate(), trainer.test()). Available options: "dice_score", "sensitivity", "specificity", "hausdorff95". Defaults to ["dice_score", "sensitivity", "specificity", "hausdorff95"].

• test_metric_confidence_levels

  A list of confidence levels for which the metrics specified in the test_metrics parameter should be computed in the validation or testing loop. This parameter is used only for multi-label classification tasks. Defaults to [0.5].

[dataset_config] section

The dataset_config section specifies parameters to setup the dataset and data loading.

• dataset

  The name of the dataset to use. Allowable values are: "brats", "decathlon" and "bcss".

• data_dir

  The directory where the data of the selected dataset resides.

• cache_size

  Number of images to keep in memory between epochs to speed-up data loading. Defaults to 0.

Note

Further mandatory or optional fields can be found in the documentation of the respective data module. Available data modules as of now are datasets.decathlon_data_module.DecathlonDataModule(), datasets.brats_data_module.BraTSDataModule() and datasets.bcss_data_module.BCSSDataModule().

[active_learning_config] section

The active_learning_config section specifies parameters to run the active learning loop.

• active_learning_mode

  Enable/Disabled Active Learning Pipeline. Defaults to False. If False, the model is trained on the full training dataset.

• reset_weights

  Enable/Disable resetting of weights after every active learning iteration. Defaults to False.

• initial_training_set_size

  Initial size of the training set if the active learning mode is activated. Defaults to 1.

• iterations

  Iteration times how often the active learning pipeline should be executed. If None, the active learning pipeline is run until the whole dataset is labeled. Defaults to None.

• items_to_label

  Number of items that should be selected for labeling in each active learning iteration. Defaults to 1.

• batch_size_unlabeled_set

  Batch size for the unlabeled set. Defaults to batch_size.

• heatmaps_per_iteration

  Number of heatmaps to be generated per active learning iteration. This option is intended for uses cases where you want to assess the quality of a sampling strategy using heatmaps of the model’s predictions. Defaults to 0.

[strategy_config] section

The strategy_config section specifies parameters to setup the strategy to query for new examples to be labeled.

• type

  Name of the sampling strategy to use. Allowable values are: "random", "interpolation", "uncertainty", "representativeness_distance", "representativeness_clustering" and "representativeness_uncertainty".

• description

  Detailed description about the configuration of the strategy. The information is logged to make experiments clearer.

Note

Further mandatory or optional fields can be found in the documentation of the respective strategy. Available strategies and their documentations can be found in the query_strategies package.