sample_posterior_predictive returns the wrong dimension

[FedericoV]

If you have questions about a specific use case, or you are not sure whether this is a bug or not, please post it to our discourse channel: https://discourse.pymc.io

Description of your problem

sample_posterior_predictive adds an extra dimension when sampling a multivariate observation

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

train_obs = model.observed_RVs[0]
print(train_obs.name, train_obs.eval().shape)

model = eu_forecaster.model_
advi = pm.ADVI(model=model)
approx = advi.fit(10)
trace = approx.sample(10)

posterior = pm.sample_posterior_predictive(trace=trace, model=model)

print(posterior['train_obs'].shape)

Please provide the full traceback.

train_obs (730, 56)
  0%|          | 0/10 [00:00<?, ?it/s]/home/ubuntu/anaconda3/lib/python3.7/site-packages/theano/tensor/subtensor.py:2320: FutureWarning: Using a non-tuple sequence for multidimensional indexing is deprecated; use `arr[tuple(seq)]` instead of `arr[seq]`. In the future this will be interpreted as an array index, `arr[np.array(seq)]`, which will result either in an error or a different result.
  out[0][inputs[2:]] = inputs[1]
/home/ubuntu/anaconda3/lib/python3.7/site-packages/theano/tensor/basic.py:6592: FutureWarning: Using a non-tuple sequence for multidimensional indexing is deprecated; use `arr[tuple(seq)]` instead of `arr[seq]`. In the future this will be interpreted as an array index, `arr[np.array(seq)]`, which will result either in an error or a different result.
  result[diagonal_slice] = x
/home/ubuntu/anaconda3/lib/python3.7/site-packages/theano/tensor/subtensor.py:2190: FutureWarning: Using a non-tuple sequence for multidimensional indexing is deprecated; use `arr[tuple(seq)]` instead of `arr[seq]`. In the future this will be interpreted as an array index, `arr[np.array(seq)]`, which will result either in an error or a different result.
  rval = inputs[0].__getitem__(inputs[1:])
Average Loss = 2.5178e+14: 100%|██████████| 10/10 [00:00<00:00, 120.98it/s]
Finished [100%]: Average Loss = 2.6755e+17
100%|██████████| 10/10 [00:06<00:00,  1.55it/s]
(10, 730, 730, 56)

Please provide any additional information below.
This is something that has only started showing up with the very latest version of PyMC3. I will try downgrading and see if that fixes it. The model still fits fine, and, looking at the parameter distributions, they look reasonable.

Versions and main components

• PyMC3 Version: 3.6
• Theano Version: '1.0.2'
• Python Version: 3.7
• Operating system:
• How did you install PyMC3: git clone

[FedericoV]

I can confirm that if I revert to commit (88acc7c) this does not appear, so it seems like it's a very recent issue.

[lucianopaz]

Could you post your model's code to better understand what distribution causes the error?
The most likely commit that could have introduced this is c73f97d, could you try if you see the same error there?
If yes, we will certainly need your model's code or an equivalent model that also has this error to be able to fix the problem.

[FedericoV]

The model unfortunately is fairly complicated and built by a class that does a lot of other stuff. I'll try to see if I can do a simpler version that reproduces it: the main thing is the likelihood is a full rank multivariate gaussian.

[lucianopaz]

Does it use a Mixture distribution at some point? Does it encode the covariance with LKJCholeskyCov distribution?

[FedericoV]

It does use the LKJCholeskyCov distribution, yup. No mixture.

…

On Wed, Jan 16, 2019 at 9:55 PM Luciano Paz ***@***.***> wrote: Does it use a Mixture distribution at some point? Does it encode the covariance with LKJCholeskyCov distribution? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#3343 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAmddrpWzUIEdE2linb-Q11TydvY3XPzks5vEBBKgaJpZM4aELip> .

[lucianopaz]

Then it may be an issue with the new LKJCholeskyCov.random method. When you post a simplified model that reproduces the problematic behavior, we will be able to start working on a fix

[lucianopaz]

@FedericoV, any luck with getting a minimal model that reproduces the error?

[lucianopaz]

@FedericoV, any update?

[junpenglao]

Should be fixed, close for now.