sample_posterior_predictive does not handle the case when the shared variable has only one sample.

[nokados]

Description of your problem

I want to make a prediction, so I set the value of a shared variable. It has shape (num_samples, num_features). It works in an unexpected way when num_samples is equal to 1. In this case, it generates results for the original value of the shared value, on which the model was trained.

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

x = np.random.randn(123, 2)
y = x[:,0] + x[:,1] > 0

with pm.Model() as model:
    x_shared = pm.Data('x_shared', x)
    coeff = pm.Normal('x', mu=0, sigma=1, shape=(2,))
    print(coeff.shape)
    logistic = pm.math.sigmoid(coeff[0] * x_shared[:,0] + coeff[1] * x_shared[:,1])
    pm.Bernoulli('obs', p=logistic, observed=y)

    # fit the model
    trace = pm.sample()

good_values = np.array([[0,0],
                       [0,1],
                       [1,1]])

bad_values = np.array([[1,2],])

print('good', good_values.shape)
print('bad', bad_values.shape, flush=True)

with model:
    # For good
    pm.set_data({'x_shared': good_values})
    post_pred = pm.sample_posterior_predictive(trace, samples=500)
    print('\ngood res', post_pred['obs'].shape)
    # For bad
    pm.set_data({'x_shared': bad_values})
    post_pred = pm.sample_posterior_predictive(trace, samples=500)
    print('\nbad res', post_pred['obs'].shape) # Expect (500, 1), get (500, 123) !!!

Output

2020-01-07 14:26:25,534: Auto-assigning NUTS sampler...
2020-01-07 14:26:25,535: Initializing NUTS using jitter+adapt_diag...
Shape.0
2020-01-07 14:26:26,115: Sequential sampling (2 chains in 1 job)
2020-01-07 14:26:26,116: NUTS: [x]
Sampling chain 0, 0 divergences: 100%|██████████| 1000/1000 [00:00<00:00, 1147.73it/s]
Sampling chain 1, 0 divergences: 100%|██████████| 1000/1000 [00:00<00:00, 1124.48it/s]
good (3, 2)
bad (1, 2)
/usr/local/lib/python3.6/dist-packages/pymc3/sampling.py:1247: UserWarning: samples parameter is smaller than nchains times ndraws, some draws and/or chains may not be represented in the returned posterior predictive sample
  "samples parameter is smaller than nchains times ndraws, some draws "
100%|██████████| 500/500 [00:00<00:00, 628.52it/s]
 25%|██▌       | 127/500 [00:00<00:00, 1260.77it/s]
good res (500, 3)
100%|██████████| 500/500 [00:00<00:00, 1273.05it/s]
bad res (500, 123)

Versions and main components

• PyMC3 Version: 3.8
• Theano Version: 1.0.4
• Python Version: 3.6.9
• Operating system: Linux Debian buster/sid
• How did you install PyMC3: pip

[rpgoldman]

My guess is that there is something in the code that is trying to turn a single-dimensional array into a scalar, and in this case that would be the wrong thing.

Ugh. Numpy. 🤮

[vtseng]

I ran into the same issue. Thanks for the question and the answer!

[vincent-bcg]

Hi,
thanks for pointing this out.
I had the same issue; I ended up passing 2 data points (duplicate) and discarding the duplicated results.
Looks like something like this:

dup_values = np.array([[1,2],[1,2]])
with model:
    pm.set_data({'x_shared': dup_values})
    post_pred = pm.sample_posterior_predictive(trace, samples=500)
    results = post_pred['obs'][:,0]

[rpgoldman]

Actually, isn't this issue the same as many others -- you can't reliably set a pm.Data to a new value with a different shape from its original value. That's because shapes can be compiled into the model and bad things happen when you change them.