Skip to content

run-inference.md

Latest commit

 

History

History
89 lines (56 loc) · 2.59 KB

run-inference.md

File metadata and controls

89 lines (56 loc) · 2.59 KB
Raw

Run Inference

This guide covers prediction, optional ensembling, and postprocessing.

Before you start

Input images must match the trained dataset's naming convention and file endings. See:

If you previously ran nnUNetv2_find_best_configuration, use the commands it generated in inference_instructions.txt whenever possible.

Predict with a trained configuration

nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d DATASET_NAME_OR_ID -c CONFIGURATION

If you want to ensemble probability outputs from multiple configurations, add --save_probabilities:

nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d DATASET_NAME_OR_ID -c CONFIGURATION --save_probabilities

By default, inference uses the 5 trained folds as an ensemble. If you trained the all fold and want to use only that model:

nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d DATASET_NAME_OR_ID -c CONFIGURATION -f all

Ensemble multiple configuration outputs

nnUNetv2_ensemble -i FOLDER1 FOLDER2 -o OUTPUT_FOLDER -np NUM_PROCESSES

The input folders must contain probability files produced with --save_probabilities.

Apply postprocessing

nnUNetv2_apply_postprocessing \
  -i FOLDER_WITH_PREDICTIONS \
  -o OUTPUT_FOLDER \
  --pp_pkl_file POSTPROCESSING_FILE \
  -plans_json PLANS_FILE \
  -dataset_json DATASET_JSON_FILE

For single-configuration predictions, plans.json and dataset.json are usually copied automatically. For ensemble outputs, provide them explicitly.

Predict from a model folder

If you want to run inference directly from an exported or copied model folder:

nnUNetv2_predict_from_modelfolder -i INPUT_FOLDER -o OUTPUT_FOLDER -m MODEL_FOLDER

Export and import your own trained model

To move a trained model to another machine:

  1. Export it:
nnUNetv2_export_model_to_zip -d DATASET_NAME_OR_ID -o MODEL.zip
  1. Install it on the target machine:
nnUNetv2_install_pretrained_model_from_zip MODEL.zip

The target machine still needs a compatible nnU-Net installation and all dependencies.

If the model was trained with a custom nnUNetTrainer subclass, the target machine also needs that trainer class to be importable. See:

Public pretrained models

The old page on pretrained-model inference remains here:

Check that page for the current status before relying on it.