This folder contains jupyter notebooks and scripts necessary to reproduce the plots in the Nearly-NMF paper (link TBD). The following index describes which notebook is linked to which plot:

1. 01-toy.ipynb - This notebook contains the entirety of the toy problem, and the resultant plots, namely Figs 1-3. It also includes all the code to generate the dataset and the arbitrary shift Shift-NMF.
2. 02-spectra_plots.ipynb - Fig 4. Loads a portion of the dataset, and selects three spectra to plot.
3. 03-template_plots.ipynb - Fig 5. Loads the templates, and plots them, along with the sum of the weights per pixel.
4. 04-chi_2.ipynb - Fig 6. Loads $\chi^2$ data (from chi_2.py), and makes a histogram of their differences.
5. 05-timing_plots.ipynb - Fig 7. Loads the saved timing (from timing.py) to generate the plots. Includes both a horizontal and vertical version of the three panels joined as well as each individual panel.
6. 06-recon.ipynb - Fig 8. Loads the dataset and splits out the validation set, then plots reconstructions for both the noisy and noise-free templates.

The following is the process to generate the data and train templates used in the above plots:

1. Generate 200k simqso qsos: python run_simqso --seed 100921 --n_qsos 200000 --n_proc 128 -o qsos.fits
2. Generate the noisy qsos and their corresponding rebinned and rescaled noise-free truths: python add_noise.py --seed 100921 -o qsos -i qsos.fits --renorm
3. Train the templates: python train_templates.py --seed 100921 -i qsos_noisy.fits -o noisy --validate --algorithm both to train both Nearly and Shift-NMF on the noiseless dataset, saving with the suffix "noisy", and including fitting the templates to the validation dataset. Repeat the same commend with noiseless in place of noisy to fit to the noise-free truth dataset.
4. Use the NNLS solver to fit the trained templates to the validation dataset with python chi_2.py -i qsos_noisy.fits.