Machine Enhanced Reconstruction Learning and Interpretation Networks (MERLIN) - merlinth
Project description
MERLIN - Machine Enhanced Reconstruction Learning and Interpretation Networks
This repository contains machine learning (ML) tools for PyTorch, TensorFlow and Python in three modules:
merlinth
: ML extensions to PyTorchmerlintf
: ML extensions to TensorFlowmerlinpy
: ML extensions to Python
If you use this code, please cite
@inproceedings{HammernikKuestner2022,
title={Machine Enhanced Reconstruction Learning and Interpretation Networks (MERLIN)},
author={Hammernik, K. and K{\"u}stner, T.},
booktitle={Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM)},
year={2022}
}
Requirements
git clone https://github.com/midas-tum/optox.git
cd optox
python3 install.py
follow build instructions on the github.
Installation
PyPi
pip3 install merlinpy-mri merlinth-mri merlintf-mri
In case you want to use the sampling codes (C++), please use the direct way installation below for direct compilation according to your system setup.
Direct way
git clone https://github.com/midas-tum/merlin.git
python3 install.py
Verification
Run unittests to ensure proper working of sub-modules
python3 -m unittest discover -s merlinpy.test
python3 -m unittest discover -s merlinth.test
python3 -m unittest discover -s merlintf.test
Contents
!!! Attention !!! This package is work in progress and still under construction. Major changes in structure will appear. If you experience any issues, if you have any feature requests or if you found any bugs, please let us know and raise an issue and/or pull request in github :)
Please watch the Issues
space and look for the latest updates regularly! :)
merlinth
merlinth
|--- layers # Data-driven regularizer following (https://github.com/VLOGroup/tdv), extended to complex-valued layers and similar setup as layers in `merlintf.keras`
|-- Complex-valued convolutions
|-- Complex-valued activations
|-- Complex-valued pooling
|-- Complex-valued normalization
|-- FFT operations
|-- Data consistency
|-- ...
|-- losses # Common and custom loss functions
|-- models # Model zoo
|-- Fields-of-Experts (FOE) regularizer
|-- Total deep variation (TDV) regularizer
|-- UNet
|-- optim # Custom optimizers such as BlockAdam
merlintf
merlintf
|-- keras
|-- layers # basic building blocks, focusing on complex valued operations
|-- Complex-valued convolutions
|-- Complex-valued activations
|-- Complex-valued pooling
|-- Complex-valued normalization
|-- FFT operations
|-- Data consistency
|-- ...
|-- models # several layers are put together into networks for complex-valued processing (2-channel-real networks, complex networks)
|-- Convolutional Neural Network
|-- Fields-of-Experts (FOE) regularizer
|-- Total deep variation (TDV) regularizer
|-- UNet
|-- optimizers # custom optimizers
|-- optim # custom optimizers
merlinpy
merlinpy
|-- datapipeline # collection of datapipelines and transform functions
|-- sampling # subsampling codes and sampling trajectories
|-- fastmri # dataloader and processing related to fastMRI database
|-- losses # losses/metrics
|-- recon # conventional reconstructions
|-- wandb # logging via wandb.ai
Common mistakes and best practices
writing own keras modules and layers
tf.keras.Model
cannot hold any trainable parameters. All trainable weights have to be defined intf.keras.layers.Layers
. Wrong implementation will cause weird behaviour when saving and re-loading the model weights!- Do not define weights in the
__init__
function. Weights should be only created and initialized in thedef build(self, input_shape)
function of theLayer
. Wrong implementation will cause weird behaviour when saving and re-loading the model weights! - The online documentation is a good orientation point to write own modules.
Make use of keras
Constraints
andInitializers
.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for merlinth_mri-0.4.0-py2.py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 19f27ebe3f5bb041ccef05f280cc0bdcee96891dc1987361809f1b2b9c5d80ad |
|
MD5 | 0c4cb052b79eaf69ca2c3769b96b7648 |
|
BLAKE2b-256 | a2fea7263c4c6dcaa3172e42b0ddc45c9469957c3b098343ae6ae2541357b468 |