Fast Spectra Predictor Network (FSP-Net)

FSP-Net is an autoencoder with two halves: an encoder (blue box) and a decoder (red box).
The encoder is used to predict spectral parameters from spectra, while the decoder reconstructs spectra from spectral parameters.

See the Wiki for more information on how to use this repository.

The research paper can be found on MNRAS or the Arxiv.

Requirements

• Install dependencies:
  pip install -r requirements.txt
• PyTorch's dependencies¹:
  NVIDIA GPU with CUDA Toolkit ~= v11.6
• PyXspec dependency²:
  Xspec from HEASoft provided by NASA

System Specification

The system specs used to develop the network are:

• Operating system: Ubuntu
• CPU: Intel 6C/12T, 2.6 GHz - 5 GHz (i7-10750H)
• GPU: NVIDIA 4 GB VRAM, 1035 - 1200 MHz (1650 Ti Max-Q)

NVIDIA GPUs with more than 1 GB of VRAM³ are highly recommended.
All CPUs that can run PyTorch should work fine but will be significantly slower.
The code has not been developed for other brands of GPUs or NPUs, so it won't take advantage of them.
SSDs are also highly recommended as this can speed up data loading times.

Training Times

Training the decoder for 200 epochs with 100,000 spectra takes around 20 minutes on the GPU and 370 minutes on the CPU.
Training the encoder for 200 epochs with 10,800 spectra takes around 3 minutes on the GPU and 43 minutes on the CPU.

Data Compatibility

The network is designed for data from black hole X-ray binaries from the Neutron star Interior Composition Explorer (NICER) instrument.
It has not been tested on data collected by other instruments or for different sources, so retraining will likely be required.
However, the idea is that this network can be applied to several applications, so feel free to adapt it and train it to different use cases.

Footnotes

1. Only required for use with NVIDIA GPU ↩

2. Only required if using synthesize_spectra.py or spectrum_fit.py→pyxspec_test ↩

3. If your NVIDIA GPU has less than 1 GB of VRAM, set kwargs={} in spectrum_fit.py→initialization ↩