Skip to content

Latest commit

 

History

History
82 lines (53 loc) · 3.26 KB

pretraining_and_finetuning.md

File metadata and controls

82 lines (53 loc) · 3.26 KB
Raw

Pretraining with nnU-Net

Intro

So far nnU-Net only supports supervised pre-training, meaning that you train a regular nnU-Net on some source dataset and then use the final network weights as initialization for your target dataset.

As a reminder, many training hyperparameters such as patch size and network topology differ between datasets as a result of the automated dataset analysis and experiment planning nnU-Net is known for. So, out of the box, it is not possible to simply take the network weights from some dataset and then reuse them for another.

Consequently, the plans need to be aligned between the two tasks. In this README we show how this can be achieved and how the resulting weights can then be used for initialization.

Terminology

Throughout this README we use the following terminology:

  • source dataset is the dataset you intend to run the pretraining on
  • target dataset is the dataset you are interested in; the one you wish to fine tune on

Pretraining on the source dataset

In order to obtain matching network topologies we need to transfer the plans from one dataset to another. Since we are only interested in the target dataset, we first need to run experiment planning (and preprocessing) for it:

nnUNetv2_plan_and_preprocess -d TARGET_DATASET

Then we need to extract the dataset fingerprint of the source dataset, if not yet available:

nnUNetv2_extract_fingerprint -d SOURCE_DATASET

Now we can take the plans from the target dataset and transfer it to the source:

nnUNetv2_move_plans_between_datasets -s SOURCE_DATSET -t TARGET_DATASET -sp SOURCE_PLANS_IDENTIFIER -tp TARGET_PLANS_IDENTIFIER

SOURCE_PLANS_IDENTIFIER is hereby probably nnUNetPlans unless you changed the experiment planner in nnUNetv2_plan_and_preprocess. For TARGET_PLANS_IDENTIFIER we recommend you set something custom in order to not overwrite default plans.

Note that EVERYTHING is transferred between the datasets. Not just the network topology, batch size and patch size but also the normalization scheme! Therefore, a transfer between datasets that use different normalization schemes may not work well (but it could, depending on the schemes!).

Note on CT normalization: Yes, also the clip values, mean and std are transferred!

Now you can run the preprocessing on the source task:

nnUNetv2_preprocess -d SOURCE_DATSET -plans_name TARGET_PLANS_IDENTIFIER

And run the training as usual:

nnUNetv2_train SOURCE_DATSET CONFIG all -p TARGET_PLANS_IDENTIFIER

Note how we use the 'all' fold to train on all available data. For pretraining it does not make sense to split the data.

Using pretrained weights

Once pretraining is completed (or you obtain compatible weights by other means) you can use them to initialize your model:

nnUNetv2_train TARGET_DATASET CONFIG FOLD -pretrained_weights PATH_TO_CHECKPOINT

Specify the checkpoint in PATH_TO_CHECKPOINT.

When loading pretrained weights, all layers except the segmentation layers will be used!

So far there are no specific nnUNet trainers for fine tuning, so the current recommendation is to just use nnUNetTrainer. You can however easily write your own trainers with learning rate ramp up, fine-tuning of segmentation heads or shorter training time.