hypunet

This repo implements hypernetworks for parametrisation of the nnUNet library. Because of this, the majority of the syntax and documentation will follow exactly from nnUNet, however with 'nnunet' swapped out for 'hypunet' on commands. For this reason, we recommend referring to the nnUNet (v1) Documentation.

Installation

The library may be installed with the command:

pip install git+https://github.com/liamchalcroft/hypernets

Command set-up

export PATH=$PATH:~/.local/bin/
export hypunet_raw_data_base="~/hypunet/hypunet_raw_data_base"
export hypunet_preprocessed="~/hypunet/hypunet_preprocessed"
export RESULTS_FOLDER="~/hypunet/hypunet_trained_models"

Data pre-processing

Pre-training

VICReg

Baseline

hypunet_train 3d_fullres VICRegTrainer 002 all -n baseline-s25-v25-c0.25 -p hypunetPlans_pretrained_PLORASPRETRAIN --kwargs sim_loss_weight=25 var_loss_weight=25 cov_loss_weight=0.25

Hypernet

hypunet_train 3d_fullres VICRegTrainer 002 all -n hyper-0-s25-v25-c0.25 -p hypunetPlans_pretrained_PLORASPRETRAIN -meta 4 -hyper 0 --kwargs sim_loss_weight=25 var_loss_weight=25 cov_loss_weight=0.25

GradCache

Baseline

hypunet_train 3d_fullres GC_VICRegTrainer 002 all -n baseline-s25-v25-c0.25 -p hypunetPlans_pretrained_PLORASPRETRAIN --kwargs metabatch=16 sim_loss_weight=25 var_loss_weight=25 cov_loss_weight=0.25

Hypernet

hypunet_train 3d_fullres GC_VICRegTrainer 002 all -n hyper-0-s25-v25-c0.25 -p hypunetPlans_pretrained_PLORASPRETRAIN -meta 4 -hyper 0 --kwargs metabatch=16 sim_loss_weight=25 var_loss_weight=25 cov_loss_weight=0.25

Training

Standard

Baseline

hypunet_train 3d_fullres HyperTrainerV2 001 0 -n baseline

Hypernet

hypunet_train 3d_fullres HyperTrainerV2 001 0 -n hyper-0 -meta 4 -hyper 0

Pre-trained

Baseline

hypunet_train 3d_fullres HyperTrainerV2 001 0 -n baseline-s25-v25-c0.25 -p hypunetPlans_pretrained_PLORASPRETRAIN -pretrained_weights ~/hypunet/hypunet_trained_models/hypunet/3d_fullres/Task002_CLINICALPLORAS/GC_VICRegTrainer__hypunetPlans_pretrained_PLORASPRETRAIN/baseline-s25-v25-c0.25/model_final_checkpoint.model

Hypernet

hypunet_train 3d_fullres HyperTrainerV2 001 0 -n hyper-0-s25-v25-c0.25 -meta 4 -hyper 0 -p hypunetPlans_pretrained_PLORASPRETRAIN -pretrained_weights ~/hypunet/hypunet_trained_models/hypunet/3d_fullres/Task002_CLINICALPLORAS/GC_VICRegTrainer__hypunetPlans_pretrained_PLORASPRETRAIN/hyper-0-s25-v25-c0.25/model_final_checkpoint.model