CLI reference
icu-benchmarks [train | evaluate] -d DATA_DIR -t TASK [options]
train is the default command. evaluate is the same as --eval: it evaluates the model in --source-dir without
training. Run uv run icu-benchmarks train -h for the help of your installed version.
Experiment
Option
Description
-d, --data-dir DATA_DIR
Required. Folder with the cohort's parquet files.
-t, --task TASK
Required. Task config in configs/tasks/, e.g. BinaryClassification, Regression, DatasetImputation.
-n, --name NAME
Name of the (target) dataset, used for the log folder.
-tn, --task-name TASK_NAME
Name of the task, used for the log folder, e.g. Mortality24.
-m, --model MODEL
Model config in configs/prediction_models/, e.g. LGBMClassifier, GRU.
-e, --experiment EXPERIMENT
Experiment config in configs/experiments/<EXPERIMENT>.gin, used instead of the task and model configs.
-l, --log-dir LOG_DIR
Log directory for results and model weights.
-s, --seed SEED
Random seed for preprocessing, tuning and training.
-hp, --hyperparams HP [HP ...]
Gin bindings that override the configs, e.g. -hp train_common.epochs=10 'LGBMClassifier.max_depth=5'. Pass all bindings after a single -hp.
--label LABEL
Label to use when the task config lists several.
-mo, --modalities MO [MO ...]
Modalities to use, as keys of the task's modality_mapping (e.g. DYNAMIC, STATIC).
Data and preprocessing
Option
Description
-gc, --generate_cache
Save the preprocessed data to <data-dir>/cache/.
-lc, --load_cache
Load the preprocessed data from the cache if it exists.
-p, --preprocessor PATH
Python file with a CustomPreprocessor class to use instead of the task's preprocessor; see Preprocessing .
-imp, --pretrained-imputation PATH
Checkpoint of a trained imputation model to impute the data with; see Imputation .
-db, --debug
Load only part of the data, for quick tests.
--complete-train
Train on all the data and skip testing.
Training and tuning
Option
Description
--tune
Tune the hyperparameters before training; see Hyperparameter tuning .
--hp-checkpoint PATH
An earlier run's hyperparameter_tuning_logs.db, relative to <log-dir>/<name>/<task-name>/<model>/ (or absolute). With --tune, tuning continues from it; without, training uses its best trial.
--cpu
Use the CPU even if a GPU is available.
--reproducible
Make PyTorch deterministic.
-wd, --wandb-sweep
Log to Weights & Biases and read the sweep's parameters; see Weights & Biases .
Evaluation and transfer
Option
Description
--eval
Only evaluate the model from --source-dir, skip training.
-ft, --fine-tune N
Load the model from --source-dir and fine-tune it on N training stays of the target data.
-sa, --samples N
Train with N training stays; with --source-dir, reuse that run's configuration.
-sn, --source-name NAME
Name of the source dataset, used for the log folder.
--source-dir PATH
Fold folder of the source run with train_config.gin and the model weights.
See Evaluation and transfer .
Output
Option
Description
-v, --verbose
Log debug messages.
-pl, --plot
Save common plots.
--explain-features
Save feature attributions of the test set: SHAP values for tree-based models, attributions and their quality metrics for deep learning models (needs --extra explain); see Feature explanations .
Boolean flags also have a --no- form, e.g. --no-tune.