adding debug mode to inferer

[arik-shurygin]

as mentioned in #244 it is helpful to have sanity checks on our likelihood functions. This function added to the mechanistic_inferer shows off a basic example of how to implement such a sanity check. Any changes to self.likelihood must also be added to the _debug_likelihood function.

Due to the layout of how self.likelihood function we can not just pass it alone into bx_model.mcmc.numpyro_nuts.debug so overrides are needed separately.

[kokbent]

wait i don't understand what it means any changes to likelihood must also be made to _debug_likelihood... they seem to just be calling self.likelihood anyway?

[SamuelBrand1]

Nice way to weave bayeux into the MechanisticInferer class.

My request is to add a test which checks if we would have picked up the issue we had when abstol/reltol wasn't aggressive enough, e.g. would this debug mode have captured potential problems without having to run chains etc.

[arik-shurygin]

@SamuelBrand1 added a test and passed kwargs so that _debug_inferer does not need to be updated every time likelihood is changed

[SamuelBrand1]

@SamuelBrand1 added a test and passed kwargs so that _debug_inferer does not need to be updated every time likelihood is changed

Looks good.

Running it locally give:

Checking test_point shape ✓
Computing test point log density ✓
Loading keyword arguments... ✓
Checking it is possible to compute an initial state ✓
Checking initial state is has no NaN ✓
Computing initial state log density ×
Transforming model to R^n ×
Computing transformed state log density shape ×
Computing gradients of transformed log density ×

Should I be interested in the fails here?

[arik-shurygin]

@SamuelBrand1 added a test and passed kwargs so that _debug_inferer does not need to be updated every time likelihood is changed

Looks good.

Running it locally give:

Checking test_point shape ✓
Computing test point log density ✓
Loading keyword arguments... ✓
Checking it is possible to compute an initial state ✓
Checking initial state is has no NaN ✓
Computing initial state log density ×
Transforming model to R^n ×
Computing transformed state log density shape ×
Computing gradients of transformed log density ×

Should I be interested in the fails here?

to be honest I am not sure. @kokbent @tjhladish I am curious your thoughts

[SamuelBrand1]

Upon reflection whether the specific test model has debug fails is not in scope for this PR.

[SamuelBrand1]

Handy extra functionality. If the debug fails on the test version of the model are relevant is another issue.