AttributeError: 'CustomSignal' object has no attribute '_loglikelihood'

[thjsal]

Calling the likelihood check function seem to result in AttributeError: 'CustomSignal' object has no attribute '_loglikelihood' error at least in the following cases:

• If calling originally the star object with the same parameter vector p (by star(p)) as that which is provided to the likelihood check function (likelihood.check(None, [-3.27536126e+04], 1.0e-6,physical_points=[p])), and not clearing the cache (likelihood.clear_cache()) before the likelihood check or possibly not calculating the likelihood by calling the likelihood object directly first (likelihood(p)). This happens because the likelihood check then thinks that likelihood was already calculated (and the attribute _loglikelihood would then exist) even though it was actually not.

• As in the previous case, but if the parameter vector provided is out of the prior bounds, calling just first likelihood(p) does not help since the likelihood will not be fully calculated and the attribute _loglikelihood will not be introduced. Clearing the cache before the likelihood check will anyway solve the problem.

I would suggest either handling this error better to give a more helpful error message about the likely cause of the issue or clearing the cache always by default before doing the rest of the likelihood check (which I think would always prevent the error from happening, at least for these cases). Clearing the cache could be automatically done at the beginning of the check() function in Likelihood.py by adding a line: self.clear_cache(). Unless people think it is good to keep the possibility for cached likelihood checks if parameters are unchanged?

[sguillot]

This error rings a bell. I think I've encountered it too!

[thjsal]

@sguillot yep! I noticed your slack thread with Tom. Apparently this same error can also happen when starting to sample with MultiNest if not using the option likelihood.externally_updated = True. Clearing the cache before sampling is not enough in that case unless that done every time inside the call function of the Likelihood class.

For simple likelihood checks that I mentioned above the externally updating would also fix the problem, but for some reason externally_updated is set to always False in the check() function at the moment.