Code to accompany "Unlocking starlight subtraction in full data rate exoplanet imaging by efficiently updating Karhunen-Loève eigenimages" by Joseph D. Long (@joseph-long) and Jared R. Males (@jaredmales) (submitted).
This code is provided in hope that it will be useful and informative, but without documentation beyond that included in the source (and, of course, the paper). We would recommend implementing the algorithm from the text yourself, as there are many more features you would want in a research tool that would only complicate a proof of concept and benchmarking tool.
You will need several Python packages and a CUDA-capable NVIDIA GPU to run the tests and benchmarks. The conda environment used to perform the benchmarks is serialized as klipbenchmark_env.yml. If you use conda, conda create -f ./klipbenchmark_env.yml should leave you able to conda activate klipbenchmark and get an equivalent environment to the one we used. You can then run pytest from this directory and test the implementation. benchmark_snr.py is also relatively self-contained, and will download the example beta Pic NACO dataset itself when run.
Our implementation of the "SVD downdate" can be found in the minimal_downdate() function in starbgone.py. An example of an end-to-end calculation from data cube to SNR can be found in calc_snr() in benchmark_snr.py.
Benchmarks were conducted on the University of Arizona HPC cluster Ocelote, and PBS batch scripts to reproduce combined_out_benchmark_scaling.csv and out_benchmark_snr.csv are available upon request.
The example data of beta Pictoris b (naco_betapic_preproc_absil2013_gonzalez2017.npz) was initially published as Absil et al. 2013 and made available as part of https://github.com/carlgogo/VIP_extras (Gonzalez et al. 2017). It is stored in this repository to make the examples self-contained, but is not covered by the GPLv3 license. The original papers must be cited if it is reused.