Neural Posterior Estimation for Pulsar Timing Arrays
Code to reproduce the results in https://arxiv.org/abs/2310.12209
D. Shih, M. Freytsis, S. R. Taylor, J. A. Dror and N. Smyth, "Fast Parameter Inference on Pulsar Timing Arrays with Normalizing Flows," [arXiv:2310.12209 [astro-ph.IM]].
Requires pytorch, nflows.
Download the training data from here:
https://zenodo.org/doi/10.5281/zenodo.10906129
Train model for 100 epochs:
python train_model.py
Prints out val loss and saves each epoch.
Currently set up to train on 3 GPUs with DataParallel. Takes approximately 15 minutes / epoch on 3 Tesla P100 or 3 GeForce GTX 1080 Ti cards.