A reference implementation of lazy resampling for pytorch, meant to accompany "Lazy Resampling: fast and information preserving preprocessing for deep learning".
This repository is a placeholder for a standalone reference implementation. The current reference implementation used in the paper is available at https://github.com/atbenmurray/MONAI/tree/lr_development.
This repository additionally contains the python notebooks and instructions for replicating the results of the experiments described in the paper.
It is recommended that you use conda to create a virtual environment in which to run both the jupyter notebook and the model training. The instructions will use the environment name lazyresampling but feel free to use an alternative name. Please follow these instructions for installing miniconda if you don't already have it.
Note that other python versions may be used but these instructions have been validated for python 3.10.
conda create --name lazyresampling python=3.10
conda activate lazyresampling
As mentioned, the reference implementation for Lazy Resampling is temporarily a branch on a fork of MONAI, but will be moved to this repository in due course. It should be installed as follows:
git clone git@github.com:atbenmurray/MONAI lazyresampling
cd lazyresampling
git checkout lr_development
pip install -e .
The reference network is a unet implemention installed as follows:
git clone git@github.com:atbenmurray/relight
cd relight
pip install -e .
The reference network can be trained by calling the following script:
python train_unet.py -S <seed_number> -L
-S sets the seed number for training; -L specifies that the preprocessing should run in lazy mode, whereas omitting it runs preprocessing in traditional mode.