Skip to content

copick training nnunet

torch

Plan, preprocess, and train nnUNet on a CoPick dataset.

Plugin command — copick-torch

This command is provided by the copick-torch plugin, not copick core. Install it to make this command available:

pip install copick-torch

See the plugin system guide for details.

Usage

copick training nnunet [OPTIONS]

Description

Runs nnUNet planning and preprocessing followed by training on a dataset that already exists in nnunet_raw (produced by copick convert nnunet). The command first invokes nnUNetv2_plan_and_preprocess to fingerprint the dataset and plan patch sizes, then runs nnUNetv2_train once per requested fold. Pass --skip-preprocess to reuse an existing plan.

The --model flag selects the architecture and the matching trainer class: nnunet (standard nnUNet), resnecl (Residual Encoder Large), and the MedNeXt variants (mednext_s, mednext_b, mednext_m, mednext_l). MedNeXt trainers require the MedNeXt package to be installed. If a fold checkpoint already exists in the results directory, training resumes from it automatically.

Options

Option Type Default Description
-id, --dataset-id integer 1 nnUNet dataset ID (must match the one used in prepare)
-n, --dataset-name text required nnUNet dataset name (must match the one used in prepare)
-r, --raw path required Path to nnunet_raw directory
-pre, --preprocessed path required Path to nnunet_preprocessed directory
-o, --output path required Path to nnunet_results directory
-cfg, --configuration choice (3d_fullres | 3d_lowres | 3d_cascade_fullres) 3d_fullres nnUNet configuration to train
-f, --folds text 0 Folds to train, e.g. 0 or 0,1,2,3,4
-m, --model choice (nnunet | resnecl | mednext_s | mednext_b | mednext_m | mednext_l) nnunet Model architecture to train.
-skip, --skip-preprocess boolean flag False Skip nnUNetv2_plan_and_preprocess (useful if already done).

Examples

# Plan, preprocess, and train the default nnUNet on fold 0
copick training nnunet -n my_dataset -id 1 \
    -r ./nnUNet_raw -pre ./nnUNet_preprocessed -o ./nnUNet_results

# Train a MedNeXt Medium model across all five folds
copick training nnunet -n my_dataset -id 1 \
    -r ./nnUNet_raw -pre ./nnUNet_preprocessed -o ./nnUNet_results \
    --model mednext_m --folds 0,1,2,3,4

# Reuse an existing plan and resume training, skipping preprocessing
copick training nnunet -n my_dataset -id 1 \
    -r ./nnUNet_raw -pre ./nnUNet_preprocessed -o ./nnUNet_results \
    --skip-preprocess

See also