IndexError in GLM using NUTS

[fonnesbeck]

Running a GLM that worked in previous versions of PyMC3 now fails when NUTS sampling begins.

with pm.Model() as severity_model:
    
    formula = 'oxygen ~ ' + '+'.join(rsv_vars)
    GLM.from_formula(formula, data=variables, family=families.Binomial())
    severity_trace = pm.sample(2000)

ADVI tuning runs fine, but once sampling begins, an IndexError is raised.

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-63-9a1fc51f9f2d> in <module>()
      6     formula = 'oxygen ~ ' + '+'.join(rsv_vars)
      7     GLM.from_formula(formula, data=variables, family=families.Binomial())
----> 8     severity_trace = pm.sample(2000)

/Users/fonnescj/Repos/pymc3/pymc3/sampling.py in sample(draws, step, init, n_init, start, trace, chain, njobs, tune, nuts_kwargs, step_kwargs, progressbar, model, random_seed, live_plot, **kwargs)
    257         sample_func = _sample
    258 
--> 259     return sample_func(**sample_args)
    260 
    261 

/Users/fonnescj/Repos/pymc3/pymc3/sampling.py in _sample(draws, step, start, trace, chain, tune, progressbar, model, random_seed, live_plot, **kwargs)
    272     try:
    273         strace = None
--> 274         for it, strace in enumerate(sampling):
    275             if live_plot:
    276                 if it >= skip_first:

/Users/fonnescj/anaconda3/envs/dev/lib/python3.6/site-packages/tqdm/_tqdm.py in __iter__(self)
    831 """, fp_write=getattr(self.fp, 'write', sys.stderr.write))
    832 
--> 833             for obj in iterable:
    834                 yield obj
    835                 # Update and print the progressbar.

/Users/fonnescj/Repos/pymc3/pymc3/sampling.py in _iter_sample(draws, step, start, trace, chain, tune, model, random_seed)
    352         _update_start_vals(start, strace.point(-1), model)
    353     else:
--> 354         _update_start_vals(start, model.test_point, model)
    355 
    356     try:

/Users/fonnescj/Repos/pymc3/pymc3/sampling.py in _update_start_vals(a, b, model)
    467         for tname in b:
    468             if tname.startswith(name) and tname!=name:
--> 469                 transform_func = [d.transformation for d in model.deterministics if d.name==name][0]
    470                 b[tname] = transform_func.forward(a[name]).eval()
    471 

IndexError: list index out of range

The model also runs fine with alternative samplers (e.g. Metropolis).

Running current master on macOS and Anaconda Python 3.6.

[denadai2]

I found this error as well and I solved it wrapping a variable I added with Deterministic. Not sure about why.

(I don't know about GLM)

[aseyboldt]

Sounds like #2046 could be related. Does it also happen if you initialize Metropolis with mu from advi? Maybe it matters if the start value contains (un)transformed keys.

[twiecki]

@fonnesbeck Seems like that's code from your recent PR, no?

[fonnesbeck]

Fixed in the above PR.

[aseyboldt]

@fonnesbeck Nope, still here:

with model:
    b = pm.Normal('beta_raw', mu=0, sd=1, shape=n_counties)
    b = pm.Deterministic('beta', b * 2)
    trace = pm.sample(2000, tune=1000)

...
/home/adr/git/pymc3/pymc3/sampling.py in <listcomp>(.0)
    467         for tname in b:
    468             if tname.startswith(name) and tname!=name:
--> 469                 transform_func = [d.transformation for d in model.deterministics if d.name==name]
    470                 if transform_func:
    471                     b[tname] = transform_func[0].forward(a[name]).eval()

AttributeError: 'TensorVariable' object has no attribute 'transformation'

[ColCarroll]

I can add a fix later -- line 519 and 998 of model.py set the transform name. we should factor that out into a few tiny helper functions (transform_name, untransform_name, is_transformed_name).

I think it is ok if we leave some edge cases uncovered: right now we use

            transformed_name = "{}_{}_".format(name, transform.name)

maybe adding one more underscore to the end, then have a function is_transformed_name just check for two trailing underscores, and untransform_name be '_'.join(name.split('_')[:-3]). I'm ok not supporting people who name their variables beta__!

[aseyboldt]

Sounds reasonable to me.

[junpenglao]

This seems to also relate to #2109, as transform_func[0].forward(a[name]).eval() depends on some parent, and it's not pass to the transform_func during evaluation.

[junpenglao]

fixed in #2328