QUEST: A machine learning framework to generate quasar spectra

QUEST (Quasar Unsupervised Encoder and Synthesis Tool) is an implementation of a Variational Auto-Encoder (VAE) with the primary purpose of generating realistic quasar spectra and post-processing them to obtain synthetic quasar photometry. QUEST can also be used to reconstruct spectra with limited wavelength coverage, absorption systems, and even the continuum blueward of the Lyman-$\alpha$ emission line (with some caveats).

Check out the paper for a full breakdown of its capabilities and limitations.

Install instructions

We recommend installing QUEST in a dedicated virtual environment. We currently leverage Atelier to sample from a luminosity function. Unfortunately, due to the design of PyPI, we currently cannot include Atelier as a requirement. It is thus necessary to install it manually:

1. Create and activate a virtual environment (e.g., using venv):

   python -m venv venv_name
   source venv_name/bin/activate  # Linux/macOS
   

2. Install Atelier:

   git clone https://github.com/jtschindler/atelier
   cd Atelier
   pip install emcee
   pip install -e .
   

3. Install QUEST from source:

   git clone https://github.com/cosmic-dawn-group/QUEST.git
   cd QUEST
   pip install -e .
   

4. or Install from PyPI:

   pip install quest_qso
   

   Note: Updates on the PyPI version might lag slightly behind the main repository.

   A note of caution: QUEST has been tested as much as possible, but, especially at the beginning, there will be bugs and aspects to improve. Please report any issue you find using the GitHub Issues tab, or consider sending us an email (francesco.guarneri@uni-hamburg.de).

Environment variables

QUEST uses a few environment variables to set its output folders and ensure that it does not overuse resources on shared machines.

• QUEST_LOCALPATH — General cache directory. This is the primary folder used to download all cached files and save generated spectra/photometry in the examples. If downloaded using the utilities included in QUEST, this will also contain the datasets used to train the model.
• QUEST_LOG_TO_FILE - QUEST logs to the terminal by default. However, if this variable is set to True or 1, an additional log file will be created in QUEST_LOCALPATH.
• AM_I_ON_SHARED_SERVER - If set to True or 1, QUEST will limit its resource usage (see details in __init__.py -- make sure to customize this to your needs!).
• TORCH_SEED - Sets the overall seed for PyTorch. If this is not set, the seed defaults to 42. If negative, no seed is set. Otherwise, the seed will be set to the value of this variable.
• TORCH_DEBUG - Effectively sets torch.autograd.set_detect_anomaly(True). This should only be used to debug issues with the model, as it greatly slows down any PyTorch operation.

Environment variables can be set (for example, in bash) using the export command:

export QUEST_LOCALPATH="/path/to/your/cache/folder"

Usage

Head over to the examples folder, where we've included Jupyter notebooks showing how to load the model for inference, sample from it, or generate synthetic photometry. If you install the package via pip, the example folder will be located in the site-package of your environment. Reaching this folder is generally cumbersome, and making changes to any file of folder might cascade in a broken Python environment. In this case, we thus recommend downloading the Notebooks from GitHub, place them in a folder of choice, and run them from there.

Contributing

Contributions are more than welcome! Please open an issue to report problems, open PRs to contribute to the code, or just let us know if you have any feature requests! We are a small team but are happy to receive feedback!

License

See LICENSE in the repository root.