update variational_api_quickstart for PyMC 6 - #882
Conversation
Part of pymc-devs#862. Five distinct breakages addressed with minimal source changes; full Restart & Run All produced fresh outputs and execution metadata. - Cells 9, 21, 50, 53: Approximation.sample(...) now requires an active model context in PyMC 6. Wrapped each call in 'with <model>:'. - Cell 21: az.plot_posterior was removed in ArviZ 1.0. Replaced with az.plot_dist(idata, var_names=[...]), matching the migration pattern used in pymc-devs#875. - Cells 28, 40, 84: pm.callbacks was removed; Tracker and CheckParametersConvergence now live in pm.variational.callbacks. - Cell 100: passing total_size=data.shape (tuple) to a multi-dim observed RV triggers an IndexError in get_scaling under symbolic_normalizing_constant. Changed to total_size=data.shape[0] (int), which is simpler and the intended API regardless. Verified under pymc 6.0.1, pytensor 3.0.3, arviz 1.1.0. All 105 cells execute cleanly under Restart & Run All; pre-commit passes.
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T07:12:05Z No need to use matplotlib inline Use az.style.use("arviz-variat") |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T07:12:06Z It seems you need ipywidgets |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T07:12:07Z Here you can use plot_dist. You can pass a dictionary where the keys are the names of the models. Check the arviz docs for details |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T07:12:07Z This should be plot_trace_dist. Alternatively we can use plot_rank_dist |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T07:12:08Z Same as previous comment |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T07:12:09Z Idem |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T07:12:09Z Are you always getting the same number? |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T07:12:10Z Kde from seaborn are bad. User arviz |
Per @aloctavodia's review: - Drop %matplotlib inline; apply az.style.use("arviz-variat") (cell 1). - Install ipywidgets locally; progress-bar warnings cleared from outputs (cell 5 thread). - Replace sns.kdeplot model comparisons with az.plot_dist dict form (cells 10, 50, 53). - az.plot_trace -> az.plot_trace_dist (cell 16). Kept plot_trace_dist over plot_rank_dist to stay consistent with the surrounding "Above are traces for x^2 and sin(x)" narrative. - Update markdown around repeated sample_node.eval(): in modern PyTensor the random graph is compiled, so repeated .eval() returns the cached value. Note this and point to sample_node(size=N) for fresh samples (cell 60). - Replace sns.kdeplot of raw numpy with az.convert_to_dataset + az.plot_dist; the first plot also switches to sample_node(size=N) so it reflects independent samples (cells 61, 64).
|
@aloctavodia thanks for the thorough review! Addressed all eight comments in the new commit:
Ready for re-review. |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T16:43:38Z Add a legend, pc.add_legend("model") and probably we just want to keep the kde (use visuals to remove elements) |
|
View / edit / reply to this conversation on ReviewNB aloctavodia commented on 2026-05-24T16:43:39Z Add semicolon |
Per @aloctavodia's followup: - Cell 10: add legend via pc.add_legend("model") and disable credible_interval/point_estimate visuals to keep only the KDE. - Cell 34: add trailing semicolon to advi.approx to suppress the bare-repr output. Applied the same legend + visuals pattern to the other two multi-model plot_dist calls (cells 50 and 53) for consistency.
|
Thanks @aloctavodia, applied both:
Applied the same legend + visuals pattern to the two other multi-model |
CI's black-jupyter wanted the trailing newline removed from the last line of three code cells; local hooks now match.
PyTensor's compiled function reads the shared RNG state without
advancing it, not because the value is "cached" — the previous
phrasing was inaccurate. Verified by inspecting RNG state across
.eval() calls (state identical) and by compiling an explicit
function with updates={rng_shared: next_rng}, which does advance
and produces fresh samples.
2 participants
Summary
Approximation.sample(...)calls inwith <model>:; PyMC 6 requires an active model context (cells 9, 21, 50, 53). Part of Updates examples for new major releases of pymc/pytensor/arviz #862.az.plot_posteriorwithaz.plot_dist(idata, var_names=[...])(cell 21).pm.callbacks.{Tracker,CheckParametersConvergence}references topm.variational.callbacks(cells 28, 40, 84).total_size=data.shape(tuple) todata.shape[0](int) — the tuple form hitsIndexErroringet_scaling(minibatch_rv.py:96) (cell 100).Verified with Restart & Run All under pymc 6.0.1 / pytensor 3.0.3 / arviz 1.1.0; pre-commit passes.