Release 1.0.2

[WardBrian]

Summary:

I think we should plan on another release in the near future.

Since 1.0.1 the following feature PRs have been merged:

• Model formatting #539
• Support . syntax for Stan variable access #543
• Update default precision for summary() #551
• Add community page to docs #553
• Detect failures in generated quantities #557
• Make a subdirectory of the temporary directory per-runset #569
• Feature/555 compute diagnostics #577
• Add an interactive option to install_cmdstan #579
• Docs/570 install instructions #578

I think the following issues/PRs would also be nice before a new patch release:

• macOS force arch #556

[mitzimorris]

hi Brian, re issue #555 -
here's what you need to get from the sample -

vars = fit.method_variables()
max(vars['treedepth__'].max(axis=0))
sum(vars['divergent__'].sum(axis=0))

[mitzimorris]

for #554, the logic would be similar to what you just did for standalone generated quantities - check for any error messages in the console and then issue a message with the link to the HMC warnings.

[WardBrian]

here's what you need to get from the sample -

My question is where do we want to do this? New methods on the fit objects?

[mitzimorris]

yes
maybe max_treedepth and divergences ?
open to better names if you can think of them.

[WardBrian]

It seems like cmdstanr also reports some diagnostics after each call to sample

[mitzimorris]

we could also (maybe) report eBFMI - estimated Bayesian Fraction of Missing Information
this compares std dev of energy__ column to sqrt( .5 * number-of-unconstrained-params), which is the dimension of the mass matrix.

[WardBrian]

If you could sketch out more in the original issue what diagnostics (e.g a list and how to calculate them, like above) and where we want them that would be helpful. We can expose them as functions the user would have to call themselves, but it seems like cmdstanr is running them automatically and I'm not sure which behavior is most useful

[mitzimorris]

I added a comment on #555 and would be happy to tackle it now

[WardBrian]

I can add a note to the docs about the interactive install and then I think we are good to run this release

[WardBrian]

@mitzimorris - would you like to do a rc for this, or just the release?

[mitzimorris]

what did we do for the 1.0.1 release?

[WardBrian]

It looks like we did not do release candidates: https://pypi.org/project/cmdstanpy/#history

[mitzimorris]

OK, let's follow that precedent.

[WardBrian]

Starting release now, I've preemptively pinned a few versions in the documentation dependencies to work around the issue you found with newer versions of the Sphinx themes. We don't really want that changing without us changing it intentionally, anyway

[WardBrian]

Release notes:

CmdStanPy 1.0.2

• CmdStanPy can now format (and canonicalize) your Stan files with CmdStanModel.format()
• Stan variables can now be accessed from fit objects using the . syntax when no naming conflicts occur. For example, previous code fit.stan_variable("my_cool_variable") can now be written fit.my_cool_variable
• CmdStanPy is more robust to running in threaded environments and tries harder to not overwrite its own output files
• The install_cmdstan script can now be run in interactive mode using --interactive/-i
• CmdStanPy now computes some diagnostics after running HMC and will warn you about post-warmup divergences and treedepth exceptions
• Runtime exceptions in the generated quantities block should be recognized better now.
• The default level of precision to fit.summary() is now 6, as it is when stansummary is used from the command line.\
• Various documentation improvements

[WardBrian]

Release is out! Conda usually takes a day or so to propagate through

[mitzimorris]

this is a nice set of new features - worth an announcement on Discourse?

[WardBrian]

We can post a new thread or bump the old one we used for 1.0.0 (I posted 1.0.1 there, too)