Config (Hydra)

maldet uses Hydra for configuration. Each stage reads a single YAML file passed via --config.

Required top-level keys

defaults: [_self_]
stage: train          # train | evaluate | predict

paths:
  config_dir: .
  output_dir: ./output
  samples_root: ./samples
  source_model: ./output/model

data:
  train_csv: ./train.csv
  test_csv: ./test.csv
  predict_csv: ./predict.csv

model:
  _target_: mydet.models.make_rf
  n_estimators: 100

model._target_ is resolved by hydra.utils.instantiate. Any extra keys under model: become keyword arguments to the factory function.

Path injection

paths.output_dir, paths.samples_root, and paths.source_model are injected by the StageRunner before each stage. Override on the command line:

maldet run train --config config.yaml \
  paths.output_dir=/mnt/runs/001 \
  paths.samples_root=/data/elf

CLI overrides

Any config key can be overridden with dot-notation:

maldet run train --config config.yaml model.n_estimators=200

Multirun sweeps

Hydra multirun is triggered with --multirun (or -m):

maldet run train --config config.yaml --multirun \
  model.n_estimators=50,100,200

Each combination runs in its own subdirectory under output_dir. Combine with hydra/launcher=joblib for local parallelism.