pymc-devs/pymc v6.0.0

on GitHub

What's Changed

See the announcement post for an overview of what's new in PyMC 6.0.

Major Changes 🛠

• Depend on PyTensor 3.0 major release by @ricardoV94 in #8199
  • Default backend is now numba
    • set pytensor.config.linker = "cvm" to return to the old default.
  • Several imports have changed location and already deprecated functions were removed
  • Low level changes that shouldn't impact most users:
    • Autodiff L_op and R_op methods are now deprecated in favor of pull_back and push_forward
    • test_value machinery is deprecated
    • low-level random API changed
• PyMC is now safely pip install-able
• Make nutpie the default nuts sampler (if installed) by @ricardoV94 in #8248
  • pip install pymc[nutpie] is now available
• Use arviz 1.0 by @aloctavodia in #8019
  • arviz.InferenceData replaced by xarray.DataTree
  • plot_trace replaced by plot_trace_dist (although plot_rank_dist is recommended instead)
  • default credible interval changed from 0.94 highest-density interval to 0.89 equal-tailed interval
  • InferenceData.extend replaced by DataTree.update
  • See the full migration guide for more details
• Allow samplers to choose default number of tuning steps by @ricardoV94 in #8289
  • Default pymc nuts sampler remains at 1000 tuning steps. Nutpie defaults to 400. Set tune explicitly to control it.
• Change sample_posterior_predictive API wrt to volatility by @ricardoV94 in #8209
  • Introduce sample_vars and freeze_vars argument to explicitly control which variables are resampled and which are reused from trace
  • Variables that are deemed "volatile" because they depend on data that has changed or variables being sampled will issue a warning instead of being resampled by default. Users should assign these variables to either sample_vars or freeze_vars to suppress the warning.
  • var_names has no effect on sampling semantics and only controls what variables are saved in the trace.
  • See docstrings for detailed examples.
• Remove deprecated samples argument in sample_prior_predictive by @williambdean in #8204
• Remove experimental warning from dims module by @ricardoV94 in #8137
• Remove several functions and objects from PyMC root namespace by @ricardoV94 in #6973

New Features 🎉

• Richer textual representation: show Data variables, fold constants, and use semantic separators by @drbenvincent in #8205
• Add Model.table() method by @fonnesbeck in #8254
• Implement correct unconstraining transform for LKJCorr by @johncant in #7380
• Implement unconstraining transform for Wishart by @ricardoV94 in #8246
• Add logccdf to LogNormal distribution by @vybhav72954 in #8166
• Apply PyMC brand theme to Model.display mermaid diagram by @williambdean in #8294

Bugfixes 🪲

• Validate coordinate and variables don't have the same names by @Chirag3841 in #8093

Documentation 📖

• compile_pymc -> compile by @williambdean in #8168
• docs: fix RST in Model.debug docstring numbered list by @anevolbap in #8281
• Updated pymc overview notebook to v6 by @fonnesbeck in #8284
• Improve HalfFlat class docstring by @KRYSTALM7 in #8211

Maintenance 🔧

• Support blackjax 1.5 progress bars by @zaxtax in #8241
• VI Approximation's sample method uses model contexts by @zaxtax in #7940
• Reject variable names with / (#8256) by @ricardoV94 in #8261
• Log external sampler and variables by @ricardoV94 in #8275
• Speedup PyMC import by @ricardoV94 in #8276
• compute_log_prior: make subsequent arguments keyword only by @vkverma9534 in #8123

New Contributors

• @vkverma9534 made their first contribution in #8123
• @drbenvincent made their first contribution in #8205
• @vybhav72954 made their first contribution in #8166
• @johncant made their first contribution in #7380
• @KRYSTALM7 made their first contribution in #8211
• @MukundaKatta made their first contribution in #8251
• @anevolbap made their first contribution in #8281

Full Changelog: v5.28.0...v6.0.0