II_Project

Code used in the Researing Intership in Intensity Interferometry, where I adapted the pix2pix GAN (Generative Adversarial Network) to work on the phase retrieval problem, specifially on the power spectrum from a fast-rotating star.

• generate_ellipsoids.py : Generates images of ellipsoids with different sizes, shapes and angular rotations. Output location: 'Data/original/'

• generate_sampling_mask.py : Generates the sparse sampling mask based on the telescope layout at MAGIC (+LST) at the Northern CTA site. Output location: 'Data/masks/128px/'

• ff2d_calc_ellipsoids.py: Creates the input data for the GAN, based on the output of 'generate_ellipsoids.py'. It is important that the input sizes of the images match. This includes the optional sparse sampling (mask based on 'generate_sampling_mask.py') and introduced Salt-and-Pepper noise. Output location: 'Data/Ellipsoids_128px_N_tele/'

• functions.py: All the functions used in the three scripts above. Must be in the same directory to work.

• Ellipsoids_v0.X.ipynb: Different versions of the GAN. The important two versions are Version_0.7 (64x64px images) and Version_0.8 (128x128px images). When running an existing model, it is very important that the architecture machtes with the one used in the model, otherwise there might be an error (and the generated images will not work properly).

• Analysis.ipynb: Analysis of the model performance for different hyperparameters. Downloads the logs from Tensorboard.dev and creates a pandas dataframe. This can be used to plot the different losses for different parameters.

To replicate this setup, please:

1. Install the packages, e.g. with pip install requirements.txt

2. Run all the scripts in this order: generate_ellipsoids, generate_sampling_mask, ff2d_calc_ellipsoids:

3. Adapt the parameters in the start of the Ellipsoids.v0.X notebook. And start the training!

If you only want to see the results and saved checkpoints:

1. Install requirements.

2. Open the Ellipsoids_v0.X.ipynb notebook, adapt the parameters based on the checkpoint name, and select the correct path. Then run the notebook with 'run_training=False'.

The Spiral-Galaxy directory contains some early tests where the same model has been applied to Spiral Galaxies. There were some issues though, as the training dataset was too heterogenous. Might re-do some of these models. It also contains a script which automatically grabs any image from APOD which contains certain keywords in the description.

If you get an error when loading the checkpoint:

• Check whether the model architecture (i.e. the filtersize (default: 4) and the number of layers is correct).