GitLab

I believe that this reshaping -- which flattened the samples -- was a hold-over from the pointwise variant.  It was causing test failures.

Local class declared after reference.

The contextvars library is used to make the vectorized version of sample_posterior_predictive compatible with the legacy version.
contextvars was added in python 3.7, but there is a compatibility library for 3.6
This is an attempt to ensure it will be loaded in python 3.6

When sampling fast_posterior_predictive gets some values with a degenerate extra dimension (e.g., 1 row of 200 values, instead of just 200 values: (1, 200) instead of (200,) ).  Trim these off.
It would be better if I could figure out where these are coming from and avoid that.

More carefully check the results of posterior predictive sampling with pm.Data.

Per @lucianopaz suggestion, I have renamed my new version as fast_sample_posterior_predictive(), and restored the old version.
This is not just nice for testing, but also is critical because we cannot currently vectorize the sampling for multivariate distributions.

Created a new _TraceDict class that represents a trace as a dictionary of variable names to numpy arrays, and use that as the internal representation, per @lucianopaz suggestion.
Made this a class so that it can have varnames and a length, like the MultiTrace.
Added a couple of translation tests.

Incorporate fix from @lucianopaz (98635eb5) and made corresponding changes to posterior_predictive.

Needed to change how theano functions were computed in vectorized variant of spp.

In general, the recursive calls to `draw_values` were not handled properly because I had bugs in the way I handled `ContextManager` objects. I believe those bugs are all fixed now.

The new version of `sample_posterior_predictive()` does not support the `size` argument, so tests using it are commented out.
Also, we don't support asking for a number of samples that is not a round multiple of the number of chains, so in some cases I bumped the number of samples to fit the input trace object.

This is work in progress only -- there are issues with recursive calls to `draw_values()` that occur inside calls to `sample_posterior_predictive`.

Redirect top level calls, and redirect recursive calls to draw_values that occur inside sample_posterior_predictive.

This is intended to make work on sample posterior predictive easier by making the arguments it takes more predictable.

Updating number of samples

* add regression test for #3819

* check for tune stop before making a draw
closes #3819

* add some type annotations

* fix reference to deprecated arguments, fix typos

* fix typos

* fix typos

For python 3.7+ we need patsy >= 0.5.1
Apparently something about tokenizing interior to python has changed.

The centering term was wrong. I was getting infs for the marginal log likelihood from my model and I dug into the code and found the centering term was wrong. The original term was too large, causing overflows unnecessarily.

* [WIP] Improve target criterion for step size adaptation

close #3655
See https://github.com/stan-dev/stan/issues/2789 for detail analysis.

* debug test fail attempt 1

* bug fix

* fix tests

* formatting edit

* fix test

Closes #3802