find_MAP fails when starting values passed

[fonnesbeck]

The example of using find_MAP in the discrete_find_MAP notebook is broken. It appears to be looking for a logodds transformed variable that it does not find:

with model:
    for i in range(n+1):
        s = {'p':0.5, 'surv_sim':i}
        map_est = pm.find_MAP(start=s, vars=model.vars)
        print('surv_sim: %i->%i, p: %f->%f, LogP:%f'%(s['surv_sim'],
                                                      map_est['surv_sim'],
                                                      s['p'],
                                                      map_est['p'],
                                                      model.logpc(map_est)))

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-11-1a981a0bba6e> in <module>()
      2     for i in range(n+1):
      3         s = {'p':0.5, 'surv_sim':i}
----> 4         map_est = pm.find_MAP(start=s, vars=model.vars)
      5         print('surv_sim: %i->%i, p: %f->%f, LogP:%f'%(s['surv_sim'],
      6                                                       map_est['surv_sim'],

/Users/fonnescj/Repositories/pymc3/pymc3/tuning/starting.py in find_MAP(start, vars, fmin, return_raw, model, *args, **kwargs)
     84             start), fprime=grad_logp_o, *args, **kwargs)
     85     else:
---> 86         r = fmin(logp_o, bij.map(start), *args, **kwargs)
     87 
     88     if isinstance(r, tuple):

/Users/fonnescj/Repositories/pymc3/pymc3/blocking.py in map(self, dpt)
     50         apt = np.empty(self.ordering.dimensions)
     51         for var, slc, _, _ in self.ordering.vmap:
---> 52             apt[slc] = dpt[var].ravel()
     53         return apt
     54 

KeyError: 'p_logodds_'

[springcoil]

I was able to replicate this. Been trying to understand what this bug is.

[ColCarroll]

Looks like it is just from supplying the start variable after it has been transformed -- adding 'p_logodds_': -0.6 to s fixes it.

It wouldn't be too hard to detect situations like this, but it might be too magical. I'll make a PR that gives an informative error message, and can maybe work on the "magic" fix, too.

[ColCarroll]

Closing -- will reopen if anyone thinks a "magical" solution would be better.