Evaluation CLI¶

dmri eval runs the Hydra-configured evaluation pipeline. It can read a local training run, a local portable checkpoint bundle, or a checkpoint bundle from Hugging Face; process real or generated data; sample model masks and parameters; compute configured metrics; and export NIfTI and JSON files.

Use dmri predict for a simpler command that validates a standard input folder, downloads a pretrained model by default, creates an output folder, and disables evaluation metrics. Use dmri eval when you need synthetic data, metrics, non-default exporters, or direct Hydra overrides.

Synopsis¶

dmri eval [HYDRA_OVERRIDES...]

Useful inspection commands:

dmri eval --help
dmri eval +experiment/eval=eval_b3s_best_model_selection --cfg job --resolve

dmri_eval remains as a deprecated command alias.

Minimal Real-Data Run¶

dmri eval +experiment/eval=eval_b3s_best_model_selection \
  checkpoint.model_name=my_run/2026-08-02_12-00-00 \
  evaluation.input.path=/data/sub-001/dwi

The input directory must contain data.nii.gz, nodif_brain_mask.nii.gz, bvals, and bvecs, unless their names are overridden under evaluation.input.

Common Overrides¶

Purpose	Override
Local run below `results/`	`checkpoint.model_name=<run>/<timestamp>`
Local portable bundle	`checkpoint.path=<parent> checkpoint.model_name=<bundle-name>`
Hugging Face bundle	`checkpoint.pretrained.repo_id=<owner/repo> checkpoint.model_name=<name>`
Best or latest checkpoint	`checkpoint.which=best` or `checkpoint.which=latest`
Direct input directory	`evaluation.input.path=/absolute/path`
Input below a data root	`evaluation.input.data_root=/data evaluation.input.data_folder=sub-001/dwi`
Synthetic input	`evaluation/input=synthetic`
Parameter samples	`evaluation.sampling.theta.num_samples=50`
Mask samples	`evaluation.sampling.mask.n_samples=50`
Corrector	`evaluation/theta/corrector@evaluation.sampling.theta.corrector=auto`
Precision	`evaluation.precision=fp32`
Fixed batch size	`evaluation.batch_size=1024`
Explicit export root	`run.output_dir=/absolute/output`
Disable theta metrics	`~evaluation.export.theta.metrics`

evaluation.input.path takes precedence over data_folder. Without path, the input is <data_root>/<data_folder>; data_root defaults to data/ below the directory from which the command was started.

Outputs¶

By default, exports are written below the selected checkpoint location. For a normal run or a Hub checkpoint, that is:

<checkpoint.results_root>/<checkpoint.results_folder>/<checkpoint.model_name>/

The optional results_folder segment is omitted when it is null. With checkpoint.path=<parent>, the default is instead <parent>/<checkpoint.model_name>/.

For example, the best-model preset writes parameter maps to ball3stick_inference_results_best_model_selection/ and model-selection maps to ball3stick_model_selection_results/. Set run.output_dir to put all export subdirectories elsewhere. hydra.run.dir controls Hydra's own run metadata; it does not control these exports.

See the evaluation guide for checkpoint layouts, pipeline modes, preprocessing, metrics, and hardware settings.