Prediction on New Data Fails with Deterministic Variable

[BrianMiner]

I am not sure if this behavior is due to some theoretical limitation or not.....but the prediction of new data with sample_posterior_predictive() fails to recognize new new shared variables when the model was run with a Deterministic() parameter in the model. See example below from the quick start.

import theano

x = np.random.randn(100)
y = x > 0

x_shared = theano.shared(x)
y_shared = theano.shared(y)

with pm.Model() as model:
    coeff = pm.Normal('x', mu=0, sd=1)
    
    # Uncomment this and comment next line to see the issue
    #logistic = pm.Deterministic('p',pm.math.sigmoid(coeff * x_shared))
    logistic = pm.math.sigmoid(coeff * x_shared)
    
    pm.Bernoulli('obs', p=logistic, observed=y_shared)
    trace = pm.sample()
    
    
x_shared.set_value([-1, 0, 1.])
y_shared.set_value([0, 0, 0]) # dummy values

with model:
    post_pred = pm.sample_posterior_predictive(trace, samples=500)

post_pred['obs'].shape

[lucianopaz]

You're right. It's a bit disappointing not to be able to store p into the trace and then change the shared's shape. The thing is that if you use logistic = pm.Deterministic(...) the values computed for logistic while sampling the posterior, get stored into the trace. These are stored as numpy.ndarrays and their shape is given by the x_shared.shape's at the time you call pm.sample. If you checkout the stored values shape with trace['p'].shape you'll see that you get (2000, 100), with the last axis being equal to len(x). The shape of the parameter's value stored in the trace then has precedence over the shape of the shared array.

Currently, the only ways around this are:

1. To not use the pm.Deterministic, just as you did.
2. To remove p from the trace after sampling by calling trace.remove_values('p'). Note that this may be dangerous so it is better to work on a deepcopy of the original trace.

[dsvolk]

This issue seems to be introduced in the recent pymc3==3.6 update. Try downgrading to pymc3==3.5.

With pymc3==3.5, I've been experimenting with predictive sampling of a model with a Deterministic for a few months, and everything worked fine.

[dsvolk]

Tested this example with pymc3==3.5 and it works fine both with Deterministic and without:

import numpy as np
import theano
import pymc3 as pm

print('Starting...')
x = np.random.randn(100)
y = x > 0

x_shared = theano.shared(x)
y_shared = theano.shared(y)

with pm.Model() as model:
    coeff = pm.Normal('x', mu=0, sd=1)
    
    # Uncomment this and comment next line to see the issue
    #logistic = pm.Deterministic('p',pm.math.sigmoid(coeff * x_shared))
    logistic = pm.math.sigmoid(coeff * x_shared)
    
    pm.Bernoulli('obs', p=logistic, observed=y_shared)
    trace = pm.sample()
    
    
x_shared.set_value([-1, 0, 1.])
y_shared.set_value([0, 0, 0]) # dummy values

with model:
    post_pred = pm.sample_ppc(trace, samples=500)

print(post_pred['obs'].shape)

[lucianopaz]

Thanks @dsvolk! I looked at the differences between the released draw_values function and found that I had included a statement that was too permissive with what could be taken from the supplied point. I will shortly submit a PR with the change.