Do we need to train the emulator when reading chains from file?

Chris-Pedersen / LaCE_manager

Cosmological emulator for the 1D flux power spectrum of the Lyman-alpha forest

3 stars 1 forks source link

Do we need to train the emulator when reading chains from file? #68

Closed andreufont closed 2 years ago

andreufont commented 2 years ago

I noticed that when using the compare_corners function we spend some time training the emulator. Why do we need to do any training? The hyperparameters are stored in the info.txt file, so we could just read these... Am I missing something?

andreufont commented 2 years ago

It is not a priorty, but leaving it here in case we ever have the time / energy to implement the reading of hyperparameters

andreufont commented 2 years ago

ok, this is actually getting quite annoying when plotting pre-computed chains, specially with the k_bin emulator that is slower to train. I might try to tackle it next week.

andreufont commented 2 years ago

This seems to only be a problem at NERSC, where training the emulator takes ~ one minute instead of the ~ one second that it takes on my laptop (!). Probably related to the cython warnings I get when I import GPy at NERSC... I will investigate.

andreufont commented 2 years ago

Addressed in #72