Skip to content
/ lbs Public

A learned version of the Born Series for highly-scattering media

Notifications You must be signed in to change notification settings

ucl-bug/lbs

Repository files navigation

Learned Born Series

This repository contains the code for the paper

A Learned Born Series for Highly-Scattering Media

This work presents a method for solving the Helmholtz differential equation using a deep learning approach. We propose a modification to the existing convolutional Born series method to reduce the number of iterations required to solve the equation in highly-scattering media. This is achieved by transforming the linear operator into a non-linear one using a deep learning model. The method is tested on simulated examples, showing improved convergence compared to the original convolutional Born series method.

This repository can also be installed as a Python package using pip, to provide an implementation of the method in the Flax neural network library, as well as a Flax implementation of the Convergent Born Series by Osnabrugge et al., 2016.


Installation

To install the package, clone the repository and run

pip install -r requirements.txt
pip install -e .

This will install the package in editable mode, so that any changes to the code will be reflected in the installed package. From here, you have a Flax model of the bno. Anywhere you can write

from bno import BNO, WrappedBNO

and use it as a model/layer in your code. The WrappedBNO is made specifically for acoustic simulations, and takes care of transforming the output into a complex field.

Train

To train the network, run

python train.py --model bno

Training takes about 3/4 days to complete on a single GPU, but you get good results already after a few hours. There are several other arguments that can be passed to the script, which can be found by running

python train.py --help

Test

To test a network, modify the TRAIN_IDS variable with your run. The key is an arbitrary string, say my_model, while the value needs to be the run ID of the wandb run. Then run

python test.py --train_id my_model

To generate the figures from the paper, run

python make_figures --figure example --model my_model

where --figure can be one of example, iterations_error, show_iterations, show_pareto, and --model.

Citation

arXiv

If you use this repository in your research, please consider citing it as:

@article{stanziola2022learned,
  title={A Learned Born Series for Highly-Scattering Media},
  author={Stanziola, Antonio and Arridge, Simon and Cox, Ben T and Treeby, Bradley E},
  journal={arXiv preprint arXiv:2212.04948},
  year={2022}
}

About

A learned version of the Born Series for highly-scattering media

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published