Caveats on the use of tinyGP#

The tinyGP package is a lightweight alternative to the george package or the internally implemented Gaussian Process regression.

Models implemented using tinyGP are:

  • tinygp_quasiperiodic, as an alternative to gp_quasiperiodic

  • tinygp_multidimensional_quasiperiodic, as an alternative to gp_multidimensional_quasiperiodic

The tinyGP alternatives deliver the same results as their traditional counterparts (as far as I know). The main advantages over the traditional are faster loglikelihood computation and the support of GPU acceleration.

Likely due to the use of jax and a slightly different approach compared to the other models, at the moment PyORBIT is not able to start the emcee analysis after the pyDE step, or reload an intermediate emcee analysis, unless the code is stopped and launch again. The code will not throw any error, it will simply be stuck without performing any further analysis.

To circumvent this problem, I introduced a flag called safe_reload in the parameter section of the configuration file.

1parameters:
2  Tref: 59200.00
3  safe_reload: True

If the safe_reload flag is set to True, the code will exit after completing the global search with pyDE. It is then necessary to launch PyORBIT again, but this task can be accomplished by writing a simple bash script where PyORBIT is just executed twice on the same configuration file.

Following the example in the PyORBIT_examples/gaussian_processes repository, we prepare a simple text file code exec_all_fits.source including these lines:

pyorbit_run emcee RV_tinyGPframework_2p.yaml > RV_tinyGPframework_2p_emcee_run.log
pyorbit_run emcee RV_tinyGPframework_2p.yaml >> RV_tinyGPframework_2p_emcee_run.log

source exec_all_fits.source

The first line will execute the PyDE step, the second line will launch the emcee analysis. Note that in the second line, we have replaced > with >>, so that the terminal output of the second run will not overwrite the previous terminal output file, but the output will be appended at the end to the existing file.

If you are using the nsave option to save the temporary emcee results after a given number of steps, you will need to launch PyORBIT a number of time equal to the total number of steps nsteps divided by the number of temporary steps nsave, plus an additional run for PyDE. Consider the example below:

In this case, you would need to launch PyORBIT for nsteps/nsave = 3 times, plus one for PyDE. The resulting script would look like this:

pyorbit_run emcee RV_tinyGPframework_2p.yaml > RV_tinyGPframework_2p_emcee_run.log
pyorbit_run emcee RV_tinyGPframework_2p.yaml >> RV_tinyGPframework_2p_emcee_run.log
pyorbit_run emcee RV_tinyGPframework_2p.yaml >> RV_tinyGPframework_2p_emcee_run.log
pyorbit_run emcee RV_tinyGPframework_2p.yaml >> RV_tinyGPframework_2p_emcee_run.log

You can add as a last line a pyorbit_results call to automatically generate the results files.