meliao / isp_baseline

Baseline Models for Solving the Inverse Scattering Problem

Overview

This repository currently contains four baseline deterministic models for solving the wideband inverse scattering problem. The models included are:

• SwitchNet
• Uncompressed Equivariant Model
• Compressed Equivariant Model
• Wideband Butterfly Network

Installation

Project Environment can be installed by running !pip install git+https://github.com/google-research/swirl-dynamics.git@main

Demos

Demos for these models can be found in the colabs folder.

Comments on the uncompressed and compressed rotationally equivariant models

-The two models are sensitive to the order of the source dimension (s) and the receiver dimension (r) in the far-field pattern data. If the models yield very poor results, try training with the perturbation data transposed.

-Using the warmup_cosine_decay_schedule scheduler to train the two models yields much better results (compared to the exponential_decay scheduler used in the TensorFlow codes).

Owen's Setup Notes

#Needs to be python 3.11 even though the docs say >=3.10
conda create -n jax_inv_scat python=3.11 
conda activate jax_inv_scat
pip install git+https://github.com/google-research/swirl-dynamics.git@main
# CUDA version of jax
pip install --upgrade "jax[cuda12_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
conda install jupyter matplotlib natsort 
conda install chardet
pip install pysteps

Owen's Data Generation Notes

Training Dataset

Here are the details for my reconstruction of the "Square (3, 5, 10)" dataset from the Wide-Band Butterfly Network paper. I was working in this private GitHub repository, which is a fork of Matt Li's repository.

https://github.com/meliao/Inverse_Scattering_ML_TF2

In MATLAB, I ran the following commands:

scatter_width_per_mesh_list = [3 5 10];
freq_list = [2.5 5 10];
dset = 'train';
gendata_multifreq_mixh(3,10, scatter_width_per_mesh_list, 'square', freq_list, dset)

Note that I used L=3, s=10 instead of the usual L=4, s=5. I don't believe this had a material effect, because these parameters are only used to determine the grid size = (2**L) * s.

After the MATLAB code completed (~ 3 hours) I ran the post-processing scripts:

# Go to the appropriate data dir
cd /Users/owen/projects/Inverse_Scattering_ML_TF2/data/xxxdata_L3s10_multifreq_square_3_5_10_h_freq_2.5_5_10
# post-process script to collect data
sh ../../post-processing-scripts/collect_data_2.5_5_10.sh
# post-process script to put data into HDF5
python ../../post-processing-scripts/merge_wavedatacsv_to_hdf5.py

These post-processing scripts were edited to reflect the different hostnames in Matt's compute environment and my own.

Test Dataset

I was using the 3,000 samples of the "Square (3, 5, 10)" dataset provided by Matt in the repository:

https://github.com/mtcli/Inverse_Scattering_ML_TF2/tree/master/public-example/testdata