This repository contains official PyTorch code for implementing Fourier convolutions and FourierNets/FourierUNets from the paper: FourierNets enable the design of highly non-local optical encoders for computational imaging.
We include PyTorch and JAX implementations of
- Fourier convolutions and multiscale Fourier convolutions
- FourierNets and FourierUNets
What is not included:
- Scripts to recreate experiments from the paper. If you want to reproduce those experiments, you can obtain training/testing code from TuragaLab/snapshotscope.
- This repository does not contain the simulation code required to run the experiments. The simulation package can be obtained from TuragaLab/snapshotscope.
- This repository does not include the data required to run the experiments. The data can be obtained from Figshare (coming soon).
There are two steps to installation, depending on whether you are interested in only the Fourier convolution implementations or also the simulation package required to run the experiment scripts. Either way, first make sure that you've installed PyTorch or Jax and its necessary dependencies for your device.
We have tested fouriernet on Python 3.7 with PyTorch 1.7. Newer versions of PyTorch will remove the old FFT interface, and cause this software to fail.
To install only the Fourier convolution architectures contained in this package, you can simply:
pip install git+https://github.com/TuragaLab/fouriernet
We have tested snapshotscope on Python 3.7 with PyTorch 1.7. Newer versions of PyTorch will remove the old FFT interface, and cause this software to fail.
To install the simulation library (required for running the experiment scripts), you can run:
pip install git+https://github.com/TuragaLab/snapshotscope