How to enhance emcee.solver efficiency?

[tpos8]

Hi, when I try to use phoebe. Emcee to fit with my lightcurve, about 18000 points in this lightcurve. In this case, how can I do not change my computer and speed-up run_solver. I already set phoebe.multiprocessing_on() phoebe.multiprocessing_set_nprocs(16) to try to speed-up. but not fortunately, this step also need to spend a few days time to complete, So I want to ask for your help to search one method to speed-up
b.add_dataset('lc', times=time, fluxes=flux, sigmas=flux_err, passband='TESS:T', ld_mode='manual', pblum_mode='dataset-scaled', dataset='lc01') b.set_value('ld_mode_bol', component='primary', value='manual') b.set_value('ld_mode_bol', component='secondary', value='manual') b.run_compute(model='tess_observation')

[kecnry]

emcee with the phoebe backend is inherently computationally expensive. My best advice is to spend the time to ensure that your initializing distributions are as optimized as possible first and that your phase sampling is as coarse as is appropriate for your data/science. If you have access to external high-performance compute resources, that is also very helpful in allowing to multiprocess on more processors.

Besides that, there are a few considerations in your setup that you can take to ensure that each forward model is as efficient as possible. These two tutorials (hopefully we'll merge them sometime in the future) cover the most common areas where you can gain performance:

• Advanced Tutorial: Optimizing Performance with PHOEBE
• Workshop Tutorial: Optimizing Computations

Hope this helps!

[tpos8]

So, can I use the emcee Python package to fit with PHOEBE, instead of using emcee within PHOEBE? Will this step enhance the fitting efficiency?

[tpos8]

If this step can enhance efficiency, could you provide me with a case code for it?

[aprsa]

No, running emcee outside of phoebe will not enhance fitting efficiency. Kyle's suggestion was to make sure that the legwork before mcmc is done well, i.e. that you have a converged solution, that parameter priors are reasonable, and that any possible model simplifications are applied. We strive to make phoebe as general as possible, but the downside is that the computations are quite complex and, in consequence, costly to compute.

[tpos8]

OK, Thank for your help