This repository includes source code for training and using the Born Fourier Neural Operator (BornFNO) model proposed in our ICASSP 2023 paper, Deep Born Operator Learning for Reflection Tomographic Imaging by Qingqing Zhao, Yanting Ma, Petros Boufounos, Saleh Nabi, Hassan Mansour.

Please click here to read the paper.

If you use any part of this code for your work, we ask that you include the following citation:

@InProceedings{Zhao_2023ICASSP,
  author =	 {Qingqing Zhao and Yanting Ma and Petros Boufounos and Saleh Nabi and Hassan Mansour},
  title =	 {Deep Born Operator Learning for Reflection Tomographic Imaging},
  booktitle =	 {Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year =	 2023,
  month =	 June
}

1. Environment Setup
2. Download the Dataset and Pretrained Models
3. Training and Evaluating the Forward Models
4. Training the Autoencoder Prior Network
5. Solving the Inverse Problem

To setup a conda environment use these commands

conda env create -f environment.yml
conda activate deepGPR

Dataset and the pretrained model can be download here. Unzip the data.zip and place v1 it in the data folder. In this dataset, we consider two layer structures with two cylinders of various radius embedded completely in the second layer. More setup detail could be found in the paper, section 4.1.
The structure of h5 file:

data_frequency.h5
    ├── 0
    │   └── eps
        └── f_field
    ├── 1
        └── eps
        └── f_field
    ├── ...

Unzip the model_zoo.zip and place it in the ./ folder.

To start multiple training in parallel, run following commends. The default setting includes FNO/BornFNO with 1/5 layers.

cd forward_model
python run_fno.py

To train one BornFNO model only, use following commands. Modify layer_num, type if necessary,

cd forward_model
python train_fno.py --type="born" --layer_num=5 --modes=12 --width=128 --lr=1e-3 --batch_size=64 --dataset=v1 --add_noise=0 --step_size=700

The log and the trained model will be saved in model_zoo/forward_model/. To evaluate the trained model, run forward_model/eval.ipynb.

cd prior
python experiment_prior.py --sig=0.001 --latent_size=64 --gpu=6

The log and the trained model will be saved in model_zoo/priors/.

cd inverse_problem
python run_results_no_prior.py # solve inverse problem without learned autoencoder prior
python run_results_with_prior.py # solve inverse problem with learned autoencoder prior

Run inverse_problem/demo.ipynb for a quick demo and visualization.

See CONTRIBUTING.md for our policy on contributions.

Released under AGPL-3.0-or-later license, as found in the LICENSE.md file.

All files, except as noted below:

Copyright (C) 2022-2023 Mitsubishi Electric Research Laboratories (MERL).

SPDX-License-Identifier: AGPL-3.0-or-later

forward_model/Adam.py was adapted from an earlier release of neuraloperator https://github.com/neuraloperator/neuraloperator and forward_model/model.py contains functions that were adapted and modified from an earlier release of neuraloperator https://github.com/neuraloperator/neuraloperator (MIT license as found in LICENSES/MIT.txt):

Copyright (C) 2022-2023 Mitsubishi Electric Research Laboratories (MERL)
Copyright (C) 2023 NeuralOperator developers

SPDX-License-Identifier: AGPL-3.0-or-later
SPDX-License-Identifier: MIT

prior/encoder_model.py was adapted and modified from https://github.com/AntixK/PyTorch-VAE/ (Apache-2.0 license as found in LICENSES/Apache-2.0.md)

Copyright (C) 2022-2023 Mitsubishi Electric Research Laboratories (MERL)
Copyright (C) 2020 Anand Krishnamoorthy Subramanian

SPDX-License-Identifier: AGPL-3.0-or-later
SPDX-License-Identifier: Apache-2.0