dmri predict¶
Syntax¶
dmri predict FOLDER [options]
dmri predict --list-models [--repo-id REPOSITORY] [--revision REVISION]
FOLDER must contain data.nii.gz, nodif_brain_mask.nii.gz, bvals, and
bvecs.
Checkpoint¶
--model NAME: model subfolder in the repository. Default:msb3s_2_4_6_128.--repo-id OWNER/REPOSITORY: Hugging Face model repository. Default:manugloeck/dmri-pretrained.--revision REVISION: Hub branch, tag, or commit.--cache-dir PATH: Hugging Face cache location.--local-files-only: do not contact the Hub.--local-checkpoint PATH: local portable model bundle; bypasses the Hub.--list-models: list repository model subfolders and exit.
Sampling¶
--quality {very-fast,fast,balanced,high}: named sampling settings. Default:balanced. Each name selects aconf/predict/quality/group, so the numbers behind it are visible in the tree and recorded in the run's.hydraoutput.--num-steps INTEGER: ODE steps per theta sample.--mask-samples INTEGER: posterior mask samples.--theta-samples INTEGER: posterior parameter samples.--corrector {auto,none}: parameter correction policy.--precision {fp32,bf16,fp16}: network-forward precision. CPU fp16 falls back to fp32.--seed INTEGER: random seed. Default:1.
Model selection¶
--model-mode {per-sample,best,fixed}: sampled model per theta draw, one selected model per voxel, or one fixed model.--fixed-model {B1S,B2S,B3S}: fixed Ball3Stick-family model; implies fixed mode.
In per-sample mode, mask samples must be at least theta samples.
Hardware¶
--batch-size INTEGER: fixed voxels per batch; skips automatic sizing.--memory-fraction FLOAT: fraction of device memory JAX may preallocate.
CPU prediction prints a warning because even a few thousand voxels can take a long time. Use a supported GPU accelerator when possible.
Output and interaction¶
--output-subdir NAME: directory insideFOLDER. Default:dmri_output.--overwrite: replace an existing output directory.--no-viewer: do not writeview_results.html.--non-interactive: do not prompt for missing choices.--verbose: include detailed evaluation logs.
Interactive selection is used only when stdin and stdout are terminals. Use
arrow keys or j/k, Enter, or number keys.
Checkpoint and configuration¶
--checkpoint-which {latest,best,STEP}: which checkpoint to load. Default:best. Uselatestor a numeric step for bundles that provide those checkpoints.--set KEY=VALUE: set any configuration key directly. Repeatable, and applied last so it overrides both--qualityand the other flags.
--set is the escape hatch for anything without a dedicated flag:
dmri predict FOLDER --quality fast \
--set evaluation.sampling.theta.params.t_max=60 \
--set evaluation.export.theta.export_stds=true
Prediction composes conf/predict.yaml, which inherits the whole evaluation tree
and layers the prediction presets on top. See
Configuration.
See the prediction guide for workflow and output details.