20x Performance degradation when using theano shared variables

[jvans1]

Description of your problem

Hi,

I noticed a pretty significant performance hit when using theano shared variables. Please correct me if I'm doing something wrong. If this is a bug, I am happy to dig into this a bit more if someone can perhaps point me in the right direction

Please provide a minimal, self-contained, and reproducible example.

import pymc3 as pm
import numpy as np
Y = 95
N = 100
with pm.Model() as binomial_model1:
    pct = pm.Beta("pct", alpha=2, beta=2)
    pm.Binomial("obs", n=N, p=pct, observed=Y)
    binomial_traces1 = pm.sample(2000, tune=500, cores=2)

%%timeit
pm.sample_posterior_predictive(binomial_traces1, samples=5000, model=binomial_model1, progressbar=False)

This returns:
1.66 s ± 15 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

When I do the same thing but used theano shared variables I see the perf hit:

Y = 95
N = 100
with pm.Model() as binomial_model2:
    Ys = pm.Data('Ys', Y)
    ns = pm.Data('Ns', N)
    pct = pm.Beta("pct", alpha=2, beta=2)
    pm.Binomial("obs", n=ns, p=pct, observed=Ys)
    binomial_traces2 = pm.sample(2000, tune=500, cores=2)

%%timeit
pm.sample_posterior_predictive(binomial_traces2, samples=5000, model=binomial_model2, progressbar=False)

This results in:

31.7 s ± 498 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

My notebook for this is here

Versions and main components

• PyMC3 Version: 3.8
• Theano Version: 1.0.4
• Python Version: 3.7.4
• Operating system: Ubuntu 18.04
• How did you install PyMC3: conda

[rpgoldman]

Interestingly, I ran this notebook on my laptop and compared the two versions on my fast_sample_posterior_predictive (see #3597), which is vectorized. I also see a slowdown, but the overall time cost is not as bad:

Without pm.Data:

1.66 s ± 15 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) -- sample_posterior_predictive
1.54 ms ± 91.3 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) -- fast spp

With pm.Data:

31.5 s ± 1.19 s per loop (mean ± std. dev. of 7 runs, 1 loop each)
13.1 ms ± 412 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

I don't know theano very well. Am I right in believing that this means it does not do constant folding?

[twiecki]

I would suspect that is expected, there are many optimizations that could be made if you know the length of an array. In Theano, they can't anticipate when a shared array will change its length. However, in PyMC3 we actually can as we know it will stay constant during inference. Not sure if there is a way to exploit that, however. We could ask the Theano guys.