Add prophet regressor #1704
Add prophet regressor #1704
Conversation
Codecov Report
@@ Coverage Diff @@
## main #1704 +/- ##
=========================================
+ Coverage 100.0% 100.0% +0.1%
=========================================
Files 247 249 +2
Lines 19598 19740 +142
=========================================
+ Hits 19589 19731 +142
Misses 9 9
Continue to review full report at Codecov.
|
| @@ -514,3 +514,38 @@ def save_plot(fig, filepath=None, format='png', interactive=False, return_filepa | |||
|
|
|||
| if return_filepath: | |||
| return filepath | |||
|
|
|||
|
|
|||
| class suppress_stdout_stderr(object): | |||
There was a problem hiding this comment.
Prophet doesn't have a flag for verbosity so we need this context manager to remove their training output:
There was a problem hiding this comment.
I think we should add a link to the sktime original source
There was a problem hiding this comment.
@freddyaboulton the sktime source grabbed it from the issue so I think linking the issue is fine here
| @@ -9,3 +9,5 @@ matplotlib>=3.3.3 | |||
| graphviz>=0.13 | |||
| seaborn>=0.11.1 | |||
| category_encoders>=2.0.0 | |||
| pystan | |||
There was a problem hiding this comment.
I left this open for now until we know what versions work with evalml. I can also just use latest versions for now.
| estimators_to_check = [estimator for estimator in _all_estimators() if estimator not in [StackedEnsembleClassifier, StackedEnsembleRegressor, TimeSeriesBaselineEstimator]] + [test_estimator_needs_fitting_false] | ||
| for component_class in estimators_to_check: | ||
| if not component_class.needs_fitting: | ||
| continue | ||
|
|
||
| component = helper_functions.safe_init_component_with_njobs_1(component_class) | ||
| if component.supported_problem_types == [ProblemTypes.TIME_SERIES_REGRESSION]: |
There was a problem hiding this comment.
Prophet does not follow sklearn's estimator API (and this is our first estimator that only does time series problem types) and tests in test_components.py break so I've added some work arounds here for now. Happy to discuss any of these changes.
There was a problem hiding this comment.
Nice! I think it's fine to have this code outside a util function since it's only used twice and only in our tests.
| component_obj=prophet_regressor, | ||
| random_state=random_state) | ||
|
|
||
| def build_prophet_df(self, X, y=None): |
There was a problem hiding this comment.
Prophet requires input to be a single dataframe with a ds column for dates and yfor targets.
| @@ -163,6 +163,15 @@ jobs: | |||
| conda config --add channels conda-forge | |||
| conda activate curr_py | |||
| conda install numba -q -y | |||
| - run: | |||
There was a problem hiding this comment.
Need this for windows installation. In the PR to add prophet to automl we should also include installation instructions in our docs.
|
PR should be good to go. Windows py3.6 tests hang on automl tests for some reason though. I'll take a look into it. |
jeremyliweishih
requested review from
freddyaboulton,
dsherry,
chukarsten and
angela97lin
and removed request for
freddyaboulton
evalml/pipelines/components/estimators/regressors/prophet_regressor.py
Outdated
Show resolved
Hide resolved
|
Still stuck on windows py3.6 tests taking 1hr+. Related to this issue(facebook/prophet#122) but it doesn't seem like their suggestions are working. Will keep working on it. |
This reverts commit 82f21f1.
Fixes #1499.