Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Merge statespace module from http://github.com/jessegrabowski/pymc_st…
…atespace (pymc-devs#174) * move http://github.com/jessegrabowski/pymc_statespace to pymc-experimental * Make tests compatible with float32 * Fix remaining tests that failed when dtype is float32 * Fix remaining tests that failed when dtype is float32 * Replace print with `logger.info` * Use `getattr` to access model variables in `gather_required_random_variables` * Reduce code duplication, eliminate use of `pathlib` in favor of explict long paths * Refactor `PyMCStateSpace` and `PytensorRepresentation` to no longer require data in `__init__` * Refactor kalman filters to remove singleton dimensions everywhere possible * Add tests for sarimax, add first pass at state space distribution * Add more helpers to statespace class for building PyMC model * Changes to SARIMAX to make states interpretable, plus bad draft pymc distributions for kf outputs * Distribution attempt 2 * Distribution attempt 3 * Fix broken tests * Fix remaining tests * Preserve name/shape information after updates to matrices * Add test for `SequenceMvNormal` * Add logp as parameter to `SequenceMvNormal` * Add logp argument to `SequenceMvNormal` test * Refactor shape checking to use `type.shape` Refactor `statespace.build_statespace_graph()` to use new distribution, preparing for simulation refactor * Refactor shape checking to use `type.shape` Refactor `statespace.build_statespace_graph()` to use new distribution, preparing for simulation refactor * Add float32 test skip back to `test_filters_match_statsmodel_output` * Refactor `LinearGaussianStateSpace` for use in prior/posterior predictive sampling * Infer coordinates for statespace matrices from parameters * Refactor posterior predictive sampling, coords, and data handling * Remove typecheck with `Union` * Fix float32 error in tests * Fix float32 error in tests * Fix bugs in SARMIAX, add measurement error option to SARIMAX, update notebooks. * always add `TIME_DIM` to model coords, even when data aren't given dims * always add `TIME_DIM` to model coords, even when data aren't given dims * Add forecast method to `Statespace`, add coord/dim info to `VARMAX`, update ARIMA notebook to show forecasting * Add `impulse_response_function` method to `Statespace`, update `ARIMA` example notebook * Update VARMAX notebook, remove `pm.DiracDelta` from `Statespace.forecast` (JAX doesn't support it) * Update VARMAX example notebook * Add docstrings to statespace methods and utilities * More docstrings * documentation * More docs, fix failing tests * Implement conditional and unconditonal sampling from prior idata * Trying to get the API docs to render * Trying to get the API docs to render * Revert "Trying to get the API docs to render" This reverts commit 3f766d3. * Trying to get the API docs to render, revert local change to `pyproject.toml` * Trying to get the API docs to render * Remove `specify_broadcast` in `SingleTimeseriesFilter.update` * Fixing docstrings * Fixing docstrings * create toctree hierarchy * create toctree hierarchy * create toctree hierarchy * docstrings * docstrings * docstrings * Making a table in a docstring * Making a table in a docstring * More docs * More docstrings * More docstrings * More docstrings * Remove `BayesianLocalLevel`, add a new `StructuralTimeSeries` module * Remove `BayesianLocalLevel`, add a new `StructuralTimeSeries` module * Bug fixes in `structural`, new example notebook * Bug fixes in `structural`, update structural example * Kalman filter no longer returns a forecast on the predicted states Predicted state dimensions are no longer incorrectly shifted backwards 1 step Remove `extended_time` dim from everywhere Seed all tests * Fix bugs in IRF API Update VARMAX example notebook * Fix bugs in IRF API Update VARMAX example notebook * Add `airpass.csv` test data Update example notebooks More docstrings * More docstrings * More docstrings * More docstrings * More docstrings * Fix VARMAX tests * More docs * More docs, allow user to name states in seasonal components * more docs, rename `test_structural_model` to `test_structural` * Allow non-integer seasonal length in `FrequencySeasonality` (breaks tests) * Seed `sample_posterior` and `sample_prior` tests, unbreak `test_structural.py` * Set test seed as a global constant * Remove numba dependency Co-authored-by: Ricardo Vieira <[email protected]> * Add helper function to `StructuralTimeSeries` to reconstruct components from hidden states * Add `fast_eval` as test utility * Changes from review feedback * Try to fix failing test * Catch expected test warnings with `pytest.mark.filterwarnings` * Scope fixtures in sampling tests to speed them up a bit * More fixes from review * Trying to fix the failing test * Remove numba from `windows-environment-test.yml` * Test for NaNs in sampled statespace objects * Redo the `Custom Statespace` notebooks * Add `ImputationWarning` to `mask_missing_values_in_data` * Add `ImputationWarning` to `mask_missing_values_in_data` Raise if data contains the missing mask value Add `missing_variable_fill` argument to `build_statespace_graph` * Add off-diagonal averaging to covariance `stabilize` function. * Covariance matrices output by kalman filter not always PSD * Add some stability tricks to Kalman filters Add test that all filter outputs are PSD * Add `measurement_error` kwarg to `PyMCStateSpace` * Allow `eig == 0.0` to pass PSD test * Add stability tests to VARMAX and SARIMAX * LGSS distribution doesn't need to know anything about measurement error Add jitter to `H` in unconditional_sampling when there is no measurement error Stabilize `P_filtered` and `P_smoothed` as every kalman step Loosen `atol` and `rtol` for kalman filter tests when comparing with statsmodels to account for all the diagonal jittering * Re-run Structural Timeseries Modeling.ipynb with some new features * Delete out-of-date example notebook * Updates to Making a Custom Statespace Model.ipynb * Link to the bVAR example notebook in the VARMAX Example.ipynb * Make the IRF plots more readable in ARIMA Example.ipynb * Increase `JITTER_DEFAULT` when pytensor is in `float32` mode * Increase `JITTER_DEFAULT` when pytensor is in `float32` mode * Add stabilization to univariate filter covariance * Trying to get the last tests to pass * Different stability strategy for UnivariateFilter * Skip overly sensitive float32 tests * Remove `update` function `make_symbolic_graph` is now the core method for `PyMCStateSpace` Add shape and name info to `KalmanFilter` outputs * All tests pass Update Structural Timeseries Modeling.ipynb to reflect refactor * Update Making a Custom Statespace Model.ipynb to reflect refactor * Relax float32 test tolerance in `test_structural.py` * Set dtype on numpy arrays used in test_structural.py * Remove dictionary merge with pipe Tweak tolerances for test in `test_structural.py` * Disable all PSD tests for univariate filter when floatX=float32 * Remove `variable_by_shape` helper, use `pt.tensor` directly. * Remove unused helper functions Add tests for impulse_response_function and forecast Add tests for SARIMAX in "interpretable" mode * Adjust new tests for float32 * Add test for equivalence between SARIMA representations Add reference to Harvey (1989) in SARIMA docstring * Add test for equivalence between SARIMA representations Add reference to Harvey (1989) in SARIMA docstring * Use `self.mode` in all kalman filter scans * Begin adding seasonal components to SARIMAX.py * Implement seasonal lags and differences in SARIMAX * Add an Exogenous Regression component to `structural.py` * Expand support for exogenous variables in statespace models Re-run notebooks * Tweak tests * Tweak tests --------- Co-authored-by: Ricardo Vieira <[email protected]>
- Loading branch information
1 parent
15c88e8
commit 67a9695