global_cov:log_amp reasonable values for the sigma-contours plot?

Hi,

We are trialling starfish to see if we can implement some of your ideas here into our groups existing radiative transfer work.

We have your example files (setup.ipynb/single.ipynb) and it all runs smoothly apart from the sigma-contours are enormous (1e8) on the residuals plot. Example below.

We have managed to reduce this error through changing the log_amp parameter in the SpectrumModel(...) function to -11 (instead of 38) and changing the priors "global_cov:log_amp" value to (-11, 3) instead of (38,1). For us, this produces a similar looking plot to the readthedocs example. Numbers and plot below

Is a log_amp value of -11 considered to be okay/reasonable and normal behaviour?

As this seems significantly different to the original.
I'm not sure if you have changed something in the code since the readthedocs was written that might account for this difference?

The only other difference we had noted was the numbers outputted from training the emulator was different, such as the log likelihood. Does this all sound okay? Numbers from initial and secondary training below.

1st

2nd

This was verified by my supervisor on a separate machine. We both produced identical log-likelihood numbers of the emulator training session and in turn, identical log_amp numbers. Starfish version 0.4.2

Many thanks,
Austen

Start by running setup.ipynb and single.ipynb with Starfish 0.4.2, then inspect SpectrumModel(...) and the global_cov:log_amp prior used in those examples. Compare the sigma-contours and emulator-training outputs with the ReadTheDocs example; done means determining whether the differing log_amp and likelihood values are expected or identifying the cause of the discrepancy.