Indexing problem when using SQLite backend

[DeliciousHair]

Hello!

Not quite sure how to describe this, so I'll just go with an example to demonstrate, specifically the intro model from the tutorial page.

import numpy as np
from pymc3 import *
from scipy import optimize

np.random.seed(123)
alpha, sigma = 1, 1
beta = [1, 2.5]
size = 100

X1 = np.linspace(0, 1, size)
X2 = np.linspace(0,.2, size)
Y = alpha + beta[0]*X1 + beta[1]*X2 + np.random.randn(size)*sigma

basic_model = Model()

with basic_model:
    alpha = Normal('alpha', mu=0, sd=10)
    beta = Normal('beta', mu=0, sd=10, shape=2)
    sigma = HalfNormal('sigma', sd=1)
    mu = alpha + beta[0]*X1 + beta[1]*X2

    Y_obs = Normal('Y_obs', mu=mu, sd=sigma, observed=Y)

    start = find_MAP(fmin=optimize.fmin_powell)
    step = NUTS(scaling=start)
    backend = backends.SQLite('demo_sampling.sqlite')
    trace = sample(500, step, start=start, trace=backend)

This works fine. The problem arises when I try to use the traceplot method with some indexing:

In [17]: traceplot(trace[500::5])
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
 in ()
----> 1 traceplot(trace[500::5])

/usr/local/lib/python2.7/dist-packages/pymc3/backends/base.pyc in __getitem__(self, idx)
    159     def __getitem__(self, idx):
    160         if isinstance(idx, slice):
--> 161             return self._slice(idx)
    162 
    163         try:

/usr/local/lib/python2.7/dist-packages/pymc3/backends/base.pyc in _slice(self, idx)
    215         """Return a new MultiTrace object sliced according to `idx`."""
    216         new_traces = [trace._slice(idx) for trace in self._traces.values()]
--> 217         return MultiTrace(new_traces)
    218 
    219     def point(self, idx, chain=None):

/usr/local/lib/python2.7/dist-packages/pymc3/backends/base.pyc in __init__(self, traces)
    140         self._traces = {}
    141         for trace in traces:
--> 142             if trace.chain in self._traces:
    143                 raise ValueError("Chains are not unique.")
    144             self._traces[trace.chain] = trace

AttributeError: 'NoneType' object has no attribute 'chain'

Clearly, there is something wrong with how the traces are being read --traces not being passed to MultiTraces or something perhaps? But I cannot seem to figure this out as calling like this when using the default RAM backend works fine. Any ideas where I might want to look?

[twiecki]

Thanks for reporting. @kyleam any ideas?

[kyleam]

Thanks. The SQLite backend doesn't support slicing (there should be a warning before the error you reported), so the output should have been checked for Nones. I've pushed a commit that does this, but that doesn't make your snippet work (instead you are now going to be passing a None to traceplot). For that, I've opened a PR that allows SQLite and Text to be sliced.

[DeliciousHair]

Thanks @kyleam, that makes perfect sense! Digging through backends/sqlite.py I see this is indeed mentioned, but I would also mention that I did not receive the warning message you alluded to. Will check out your PR to see how this fix works :-)