plans-and-configuration.md

Plans and Configuration Reference

This page is the concise reference for nnUNetPlans.json and configuration-level customization.

What a plans file does

The plans file defines how nnU-Net preprocesses the dataset and how each configuration is trained.

It contains:

• global dataset-level settings
• one or more named configurations under configurations

Important global fields

• image_reader_writer: reader/writer class used for this dataset
• label_manager: label-handling class
• transpose_forward and transpose_backward: axis reordering used internally
• dataset_name: dataset identity metadata

Important configuration fields

• spacing: target spacing
• patch_size: training patch size
• batch_size: batch size
• data_identifier: name of the corresponding preprocessed data folder
• preprocessor_name: preprocessor class
• normalization_schemes: normalization mapping per channel
• network_arch_class_name: architecture class
• conv_kernel_sizes: encoder kernel layout
• n_conv_per_stage_encoder
• n_conv_per_stage_decoder
• num_pool_per_axis
• pool_op_kernel_sizes

Special configuration relationships

• inherits_from: reuse another configuration and override selected fields
• previous_stage: previous stage of a cascade
• next_stage: next stage of a cascade

When you must rerun preprocessing

You generally need to rerun preprocessing if you change anything that affects the prepared data, for example:

• spacing
• preprocessor_name
• normalization_schemes
• resampling functions
• any change that requires a new data_identifier

You generally do not need to rerun preprocessing for training-only changes such as:

• batch_size
• some architecture-only settings that still reuse the same prepared data

Common customization patterns

Increase batch size

"3d_fullres_bs40": {
  "inherits_from": "3d_fullres",
  "batch_size": 40
}

Add a custom preprocessor

"3d_fullres_custom_preproc": {
  "inherits_from": "3d_fullres",
  "preprocessor_name": "MyPreprocessor",
  "data_identifier": "3d_fullres_custom_preproc"
}

Change target spacing

"3d_fullres_custom_spacing": {
  "inherits_from": "3d_fullres",
  "spacing": [1.0, 1.0, 2.5],
  "data_identifier": "3d_fullres_custom_spacing"
}

Related pages

• Train models
• Intensity normalization in nnU-Net
• Extending nnU-Net

Detailed legacy page

• Modifying the nnU-Net configurations