An Enhanced Version of the Progressive Growing of Patch Size Curriculum implemented into the nnUNet framework is now online
The updated version (submitted to Medical Image Analysis) can be found here: https://arxiv.org/pdf/2510.23241 The Github Repo can be found here: https://github.com/StefanFischer/nnUNet-ProgressiveGrowingOfPatchSize/tree/ProgressiveGrowingOfPatchSizeV2.6.2
MICCAI2024-Paper: https://papers.miccai.org/miccai-2024/paper/2008_paper.pdf
Arxiv-Paper: https://arxiv.org/abs/2407.07853
This is the implementation from our MICCAI2024 paper "Progressive Growing of Patch Size: Resource-Efficient Curriculum Learning for Dense Prediction Tasks". Abstract. In this work, we introduce Progressive Growing of Patch Size, a resource-efficient implicit curriculum learning approach for dense prediction tasks. Our curriculum approach is defined by growing the patch size during model training, which gradually increases the task’s difficulty. We integrated our curriculum into the nnU-Net framework and evaluated the methodology on all 10 tasks of the Medical Segmentation Decathlon. With our approach, we are able to substantially reduce runtime, computational costs, and CO2 emissions of network training compared to classical constant patch size training. In our experiments, the curriculum approach resulted in improved convergence.
The curriculum reduces the runtime on average by 50%, while achieving equivalent or slighty improved performance.
If you want to use the curriculum on top of nnUNetv2, follow this instruction! The curriculum is currently only applicable with nnunetv2=2.0.0. As the configuration manager of nnUNetv2 was changed in newer versions and is not yet compatible with this nnUnet trainer script!
Training code for PGPS/PGPSplus/RPSS can be found in directory:
- ./training/nnUNetTrainer/variants/sampling/nnUNetTrainer_PGPS.py
- ./training/nnUNetTrainer/variants/sampling/nnUNetTrainer_PGPSplus.py
- ./training/nnUNetTrainer/variants/sampling/nnUNetTrainer_RPSS.py
Other adjustments:
- ./utilities/plans_handling/plans_handler.py
To install the PGPS_nnUNet you can clone this GitHub Repo via:
git clone https://github.com/StefanFischer/PGPS_nnUNet.git
cd PGPS_nnUNet
pip install -e .
Besides you also need to define the environment variables of nnUNetv2: https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/setting_up_paths.md
Do training as standard nnUNet as specified here: https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/how_to_use_nnunet.md The only difference is to specifiy the PGPS/PGPS+/RPSS trainer:
# for PGPS
nnUNetv2_train DATASET_NAME_OR_ID 3d_fullres FOLD -tr nnUNetTrainer_PGPS
# for PGPS+
nnUNetv2_train DATASET_NAME_OR_ID 3d_fullres FOLD -tr nnUNetTrainer_PGPSplus
# for RPSS
nnUNetv2_train DATASET_NAME_OR_ID 3d_fullres FOLD -tr nnUNetTrainer_RPSS
For the best performance-runtime tradeoff use PGPS+ curriculum
It is also possible to track the C02 emissions of the training via codecarbon. For that uncomment the related code in the trainer scripts. For more information about codecarbon check their website: https://mlco2.github.io/codecarbon/installation.html