Use set_data in forecast method
#451
Conversation
|
|
||
| with pm.Model(coords=temp_coords) as forecast_model: | ||
| (_, _, *matrices), grouped_outputs = self._kalman_filter_outputs_from_dummy_graph( | ||
| scenario=scenario, |
There was a problem hiding this comment.
Hey @jessegrabowski, sorry about the constant back and forth, but I think this addition here fixes the bug. The tests still fail but the error message is different than the expected dimension mismatch it is an assertion error:
FAILED tests/statespace/test_statespace.py::test_foreacast_valid_index - UserWarning: Skipping `CheckAndRaise` Op (assertion: The first dimension of a time varying matrix (the time dimension) must be equal to the first dimension of the data (the time dimension).) as JAX tracing would remove it.
Running the test test_foreacast_valid_index in a notebook works without any problems.
There was a problem hiding this comment.
It's just a warning, so we need to add a @pytest.filterwarnings decorator to the test.
I made this change then undid it because I thought it wasn't doing the right thing. I need to double-check, but the forecasting logic basically goes like this:
- The original user graph -- using the training data -- is reconstructed in its entirety.
- Using the original graph, we compute the value of the hidden states at the requested x0 for the forecasts. We do not want new scenario data here, otherwise the value of the requested x0 will be wrong.
- Starting from this x0, we construct a new graph which iterates the statespace equations forward for the requested number of time steps. If there are exogenous regressors, this is where we do want them.
I think I had the impression that this change put the scenario data in step (2), but I looked at it a few weeks ago now and I forget.
|
Closed in favor of #466 |
2 participants
Closes #424
There was some really complicated logic for clone-replacing data variables in the
forecastmethod for absolutely no reason. For certain models, this meant that the data was not being correctly updated, and coordinates didn't match after sampling.Just doing a basic
pm.set_datadoes the trick just fine.