To use nnU-Net as a framework and make changes to its components, please make sure to install it with the git clone
and pip install -e . commands so that a local copy of the code is created.
Changing components of nnU-Net needs to be done in different places, depending on whether these components belong to
the inferred, blueprint or empirical parameters. We cover some of the most common use cases below. They should give
you a good indication of where to start.
Generally it is recommended to look into the code where the thing you would like to change is currently implemented and then derive a strategy on how to change it. If you have any questions, feel free to open an issue on GitHub and we will help you as much as we can.
This section gives guidance on how to implement changes to loss function, training schedule, learning rates, optimizer, some architecture parameters, data augmentation etc. All these parameters are part of the nnU-Net trainer class, which we have already seen in the sections above. The default trainer class for 2D, 3D low resolution and 3D full resolution U-Net is nnUNetTrainerV2, the default for the 3D full resolution U-Net from the cascade is nnUNetTrainerV2CascadeFullRes. Trainer classes in nnU-Net inherit form each other, nnUNetTrainerV2CascadeFullRes for example has nnUNetTrainerV2 as parent class and only overrides cascade-specific code.
Due to the inheritance of trainer classes, changes can be integrated into nnU-Net quite easily and with minimal effort. Simply create a new trainer class (with some custom name), change the functionality you need to change and then specify this class (via its name) during training - done.
This process requires the new class to be located in a subfolder of nnunet.training.network_training! Do not save it somewhere else or nnU-Net will not be able to find it! Also don't use the same name twice! nnU-Net always picks the first trainer that matches the requested name.
Don't worry about overwriting results of another trainer class. nnU-Net always generates output folders that are named after the trainer class used to generate the results.
Due to the variety of possible changes to the blueprint parameters of nnU-Net, we here only present a summary of where to look for what kind of modification. During method development we have already created a large number of nnU-Net blueprint variations which should give a good indication of where to start:
| Type of modification | Examples |
|---|---|
| loss function | nnunet.training.network_training.loss_function.* |
| data augmentation | nnunet.training.network_training.data_augmentation.* |
| Optimizer, lr, momentum | nnunet.training.network_training.optimizer_and_lr.* |
| (Batch)Normalization | nnunet.training.network_training.architectural_variants.nnUNetTrainerV2_BN.py nnunet.training.network_training.architectural_variants.nnUNetTrainerV2_FRN.py nnunet.training.network_training.architectural_variants.nnUNetTrainerV2_GN.py nnunet.training.network_training.architectural_variants.nnUNetTrainerV2_NoNormalization_lr1en3.py |
| Nonlinearity | nnunet.training.network_training.architectural_variants.nnUNetTrainerV2_ReLU.py nnunet.training.network_training.architectural_variants.nnUNetTrainerV2_Mish.py |
| Architecture | nnunet.training.network_training.architectural_variants.nnUNetTrainerV2_3ConvPerStage.py nnunet.training.network_training.architectural_variants.nnUNetTrainerV2_ResencUNet |
| ... | (see nnunet.training.network_training and subfolders) |
The inferred parameters are determined based on the dataset fingerprint, a low dimensional representation of the properties
of the training cases. It captures, for example, the image shapes, voxel spacings and intensity information from
the training cases. The datset fingerprint is created by the DatasetAnalyzer (which is located in nnunet.preprocessing)
while running nnUNet_plan_and_preprocess.
nnUNet_plan_and_preprocess uses so called ExperimentPlanners for running the adaptation process. Default ExperimentPlanner
classes are ExperimentPlanner2D_v21 for the 2D U-Net and ExperimentPlanner3D_v21 for the 3D full resolution U-Net and the
U-Net cascade. Just like nnUNetTrainers, the ExperimentPlanners inherit from each other, resulting in minimal programming
effort to incorporate changes. Just like with the trainers, simply give your custom ExperimentPlanners a unique name and
save them in some subfolder of nnunet.experiment_planning. You can then specify your class names when running
nnUNet_plan_and_preprocess and nnU-Net will find them automatically. When inheriting form ExperimentPlanners, you MUST
overwrite the class variables self.data_identifier and self.plans_fname (just like for example
here).
If you omit this step the planner will overwrite the plans file and the preprocessed data of the planner it inherits from.
To train with your custom configuration, simply specify the correct plans identifier with -p when you call the
nnUNet_train command. The plans file also contains the data_identifier specified in your ExperimentPlanner, so the
trainer class will automatically know what data should be used.
Possible adaptations to the inferred parameters could include a different way of prioritizing batch size vs patch size (currently, nnU-Net prioritizies patch size), a different handling of the spacing information for architecture template instantiation, changing the definition of target spacing, or using different strategies for finding the 3d low resolution U-Net configuration.
The folders located in nnunet.experiment_planning contain several example ExperimentPlanner that modify various aspects of the inferred parameters. You can use them as inspiration for your own.
If you wish to run a different preprocessing, you most likely will have to implement your own Preprocessor class.
The preprocessor class that is used by some ExperimentPlanner is specified in its preprocessor_name class variable. The
default is self.preprocessor_name = "GenericPreprocessor" for 3D and PreprocessorFor2D for 2D (the 2D preprocessor
ignores the target spacing for the first axis to ensure that images are only resampled in the axes that will make up the training samples).
GenericPreprocessor (and all custom Preprocessors you implement) must be located in nnunet.preprocessing. The
preprocessor_name is saved in the plans file (by ExperimentPlanner), so that the
nnUNetTrainer knows which preprocessor must be used during inference to match the preprocessing of the training data.
Modifications to the preprocessing pipeline could be the addition of bias field correction to MRI images, a different CT preprocessing scheme or a different way of resampling segmentations and image data for anisotropic cases. An example is provided here.
When implementing a custom preprocessor, you should also create a custom ExperimentPlanner that uses it (via self.preprocessor_name). This experiment planner must also use a matching data_identifier and plans_fname to ensure no other data is overwritten.
Changing the network architecture in nnU-Net is easy, but not self-explanatory. Any new segmentation network you implement needs to understand what nnU-Net requests from it (wrt how many downsampling operations are done, whether deep supervision is used, what the convolutional kernel sizes are supposed to be). It needs to be able to dynamiccaly change its topology, just like our implementation of the Generic_UNet. Furthermore, it must be able to generate a value that can be used to estimate memory consumption. What we have implemented for Generic_UNet effectively counts the number of voxels found in all feature maps that are present in a given configuration. Although this estimation disregards the number of parameters we have found it to work quite well. Unless you implement an architecture with unreasonably high number of parameters, the large majority of the VRAM used during training will be occupied by feature maps, so parameters can be (mostly) disregarded. For implementing your own network, it is key to understand that the number we are computing here cannot be interpreted directly as memory consumption (other factors than the feature maps of the convolutions also play a role, such as instance normalization. This is furthermore very hard to predict because there are also several different algorithms for running the convolutions, each with its own memory requirement. We train models with cudnn.benchmark=True, so it is impossible to predict which algorithm is used). So instead, to approch this problem in the most straightforward way, we manually identify the largest configuration we can fit in the GPU of choice (manually define the dowmsampling, patch size etc) and use this value (-10% or so to be save) as reference in the ExperimentPlanner that uses this architecture.
To illustrate this process, we have implemented a U-Net with a residual encoder (see FabiansUNet in generic_modular_residual_UNet.py). This UNet has a class variable called use_this_for_3D_configuration. This value was found with the code located in find_3d_configuration (same python file). The corresponding ExperimentPlanner ExperimentPlanner3DFabiansResUNet_v21 compares this value to values generated for the currently configured network topology (which are also computed by FabiansUNet.compute_approx_vram_consumption) to ensure that the GPU memory target is met.