Single-View Refractive Index Tomography using Neural Fields.
This code was built and tested on Ubuntu 22.04 and NVIDIA RTX A6000 GPU.
There are quite a few packages used in this project. Notably jax requires a bit of care for installation. Otherwise, packages (output from pip freeze) are included in requirements.txt. After installing jax, you should be able to install these by running
pip install -r requirements.txt
However, package management with python can be finicky so you may have to fiddle around with the installation if you run into errors.
To run neural field recovery, you will have to download a series of simulated measurement files for each experiment. These include data such as camera settings, the true refractive field, simulated refracted image measurements, etc. These can be downloaded at this link (~600MB). After downloading, move the archive into the exp directory:
mkdir exp
mv /path/to/exp.tar.gz exp/exp.tar.gz
tar -xvzf exp/exp.tar.gz
Included in this repository are a couple of tutorial notebooks for generating true refractive and intensity fields, as well as rendering refracted images, which can be found in the tutorials directory.
There are also three scripts included for neural field recovery, as well as the corresponding simulated measurements from our paper. They can be run as follows:
# Gas flow recovery (Fig. 4):
python train_nf_gas.py -b gasflow -e test
# Recovery of Smooth Refractive Fields (Fig. 5):
python train_nf.py -b matern_s2 -e test
python train_nf.py -b matern_s8 -e test
python train_nf.py -b matern_s9 -e test
# Sensitivity to Light Source Density (Fig. 6):
python train_nf.py -b src_density_50 -e test
python train_nf.py -b src_density_100 -e test
python train_nf.py -b src_density_250 -e test
# Simulated Dark Matter Halo Recovery (Fig. 1):
python train_nf_dm.py -b dark_matter -e test
These scripts include optional checkpointing functions and training visualization plots using tensorboard.