This repository contains code to train a neural network to remove 1/f striped noise from JWST images. The demo.ipynb file shows how to load the optimized weights and apply the Unet model to predict a clean template of the striped noise, which can then be used to subtract from the original image.
To run demo.ipynb, the following packages along with their dependencies are required:
- numpy:
pip install numpy. - matplotlib:
pip install matplotlib. - jupyter notebook:
pip install notebook. - astropy:
pip install astropy. - tensorflow:
pip install tensorflow.-
If you encounter issues on macOS with the installation of tensorflow, you can instead follow the instructions here. Specifically, if you have a Mac with the M1 chip, I would try creating a conda environment and installing all the packages in the following order:
conda create --name mlstripes python=3.9 conda activate mlstripes pip install numpy pip install matplotlib pip install jupyter notebook pip install astropy conda install -c apple tensorflow-deps python3 -m pip install tensorflow-macos
-
If you would like to train your own model using the provided scripts, then you also need to install:
- tensorflow addons:
pip install tensorflow-addons. - colorednoise:
pip install colorednoise.
These additional packages are used to simulate training data and to perform data augmentation.
Since the network's weights exceeded the file size limit on Github, I uploaded the weights here. To run demo.ipynb with these weights, simply drag and drop the weights folder in 'scripts'.
If you have any questions, feel free to send me an email at [email protected] .