Add Model objects to forecasting #1558

canerturkmen · 2022-02-23T15:01:37Z

Issue #, if available:

Earlier PR: #1538

Description of changes:

This PR adds model classes to autogluon.forecasting, porting over previous work and significantly refactoring:

adding an AbstractForecastingModel, partially inheriting from core's AbstractModel to modularize the code and remove code dupes. Most of the remaining code dupe is in the model_trial module, however these had to be re-implemented because of the "supervised model" signature (X, y) among other reasons.
Moving GluonTS specific functionality to AbstractGluonTSModel.
Fixing model serialization, HPO functionality, etc.
Adding tests for models
Improving style, adding docstrings, etc.

Coverage report:

Name                                                                      Stmts   Miss  Cover
---------------------------------------------------------------------------------------------
src/autogluon/forecasting/__init__.py                                         7      2    71%
src/autogluon/forecasting/models/__init__.py                                  2      0   100%
src/autogluon/forecasting/models/abstract/__init__.py                         2      0   100%
src/autogluon/forecasting/models/abstract/abstract_forecasting_model.py      95      8    92%
src/autogluon/forecasting/models/abstract/model_trial.py                     46     11    76%
src/autogluon/forecasting/models/gluonts/__init__.py                          2      0   100%
src/autogluon/forecasting/models/gluonts/abstract_gluonts.py                117      5    96%
src/autogluon/forecasting/models/gluonts/callback.py                         23      3    87%
src/autogluon/forecasting/models/gluonts/models.py                           26      2    92%
src/autogluon/forecasting/utils/__init__.py                                   0      0   100%
src/autogluon/forecasting/utils/metric_utils.py                              12      0   100%
src/autogluon/forecasting/utils/warning_filters.py                           19      0   100%
src/autogluon/forecasting/version.py                                          1      0   100%
---------------------------------------------------------------------------------------------
TOTAL                                                                       352     31    91%

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

canerturkmen · 2022-02-23T15:04:00Z

@Innixma @gradientsky @dcmaddix @jwmueller @huibinshen
cc: @yx1215 @rsnirwan

Innixma

Thanks for creating the PR! I will plan to do a full review hopefully next week after v0.4.0 releases.

forecasting/src/autogluon/forecasting/models/gluonts/models.py

szha · 2022-03-02T11:23:24Z

Job PR-1558-3 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-1558/3/index.html

dcmaddix

Thanks Caner! It looks good - we can always merge this in first and then add support for mean after. Great job on the thorough tests too.

forecasting/src/autogluon/forecasting/models/abstract/abstract_forecasting_model.py

forecasting/src/autogluon/forecasting/models/gluonts/models.py

Innixma

Looks very good overall!

forecasting/src/autogluon/forecasting/models/gluonts/abstract_gluonts.py

forecasting/setup.py

forecasting/src/autogluon/forecasting/models/gluonts/models.py

forecasting/tests/unittests/models/common.py

canerturkmen · 2022-03-23T10:20:31Z

Hello @Innixma @dcmaddix !

In the second revision of the PR, we've

taken care of the earlier comments (adding todos in some places)
partially taken care of the redundant gluonts model code by having a GenericGluonTSModel
enabled ProphetModel through GluonTS --- a big feature imho for small effort
added placeholders for how we will handle metadata in the library. Here, a metadata is a dictionary that maps feature types to columns in the data frame. Right now, metadata is supplied to models at construction time and model objects "own" their metadata. That said, there is still a fast follow item to this PR on how we consistently handle static features (in the data frame format without being redundant).

canerturkmen · 2022-03-23T10:21:01Z

cc @huibinshen

canerturkmen · 2022-03-23T15:55:39Z

@Innixma I can't tell why CI wasn't triggered.

Innixma · 2022-03-23T18:56:56Z

@canerturkmen Looked like a transient CI error, I retriggered CI, hopefully it passes.

szha · 2022-03-23T20:07:56Z

Job PR-1558-7 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-1558/7/index.html

Innixma

LGTM! Awesome contribution. A few minor things can be refined in future PRs once we align on how this integrates with Trainer, Learner, and Predictor.

Innixma · 2022-03-23T19:04:54Z

forecasting/src/autogluon/forecasting/models/gluonts/models.py

+        The class object of the GluonTS estimator to be used.
+    """
+
+    def __init__(self, gluonts_estimator_class: Type[GluonTSEstimator], **kwargs):


Should gluonts_estimator_class instead be a hyperparameter? How would the user specify a particular GluonTS model to train alongside others via Predictor?

I see the need for this and will consider it in a future PR. Thanks!

Innixma · 2022-03-23T19:07:48Z

forecasting/src/autogluon/forecasting/models/gluonts/models.py

+        )
+
+
+# TODO: AutoGluon Tabular will be removed from GluonTS to avoid circular dependencies


Another option is that GluonTS can continue to depend on AutoGluon as an optional and lazily imported dependency that we explicitly don't use / import in AutoGluon Forecasting.

Meaning we would have AutoTabularModel implemented separately in both AG and GluonTS.

Another (temporary) hack may be if GluonTS just depends on autogluon.tabular sub-package instead of the full autogluon.
Long term though, the AutoTabularModel should probably be moved back into just AG as it's not a core part of GluonTS.

GluonTS does not have autogluon as a core dependency, but only as an optional requirement for users who want to use AG Tabular; so I think we are safe in the near future.

Innixma reviewed Feb 24, 2022

View reviewed changes

forecasting/src/autogluon/forecasting/models/gluonts/models.py Show resolved Hide resolved

canerturkmen force-pushed the add-forecasting-models branch from 6300a85 to a82ad69 Compare March 2, 2022 10:17

dcmaddix reviewed Mar 9, 2022

View reviewed changes

Innixma reviewed Mar 16, 2022

View reviewed changes

canerturkmen added 19 commits March 23, 2022 11:11

initial commit carrying over from main PR (autogluon#1538)

ff8309e

black everything and style edits

c2cb375

refactor the models package, add smoke tests for model init

e03adc1

added new unit tests for model and made fixes

731c663

added type hinting to abstract types

5772126

add docstrings

65380c2

changes in style, abstract model variable name and signature

be1fa8d

refactoring abstract models to allow HPO, style fixes

67ea930

improve constructor argument handling for gluonts models, enable HPO

b49f170

added tests, black everything

4224c23

fixes, modify and add tests

00802b0

added tests for load and save methods of models

0b1cf4d

added new tests, reogranized tests, refactored gluonts predict

050f59f

minor fixes to abstract model

ff032a1

address all PR comments except params_aux

421bf84

remove assignments to params_aux

921ede3

add and enable generic gluonts model

65eae63

added prophet to the gluonts model zoo

ace2f55

add placeholders for metadata dictionary and black everything

8fb8d46

canerturkmen force-pushed the add-forecasting-models branch from a82ad69 to 8fb8d46 Compare March 23, 2022 10:15

gracefully handle tests if prophet is not installed

5a32aaf

Innixma approved these changes Mar 24, 2022

View reviewed changes

Innixma merged commit 339e8f2 into autogluon:master Mar 24, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add Model objects to forecasting #1558

Add Model objects to forecasting #1558

canerturkmen commented Feb 23, 2022

canerturkmen commented Feb 23, 2022 •

edited

Loading

Innixma left a comment

szha commented Mar 2, 2022

dcmaddix left a comment

Innixma left a comment

canerturkmen commented Mar 23, 2022

canerturkmen commented Mar 23, 2022

canerturkmen commented Mar 23, 2022

Innixma commented Mar 23, 2022

szha commented Mar 23, 2022

Innixma left a comment

Innixma Mar 23, 2022

canerturkmen Mar 24, 2022

Innixma Mar 23, 2022

jwmueller Mar 24, 2022

canerturkmen Mar 24, 2022

		)


		# TODO: AutoGluon Tabular will be removed from GluonTS to avoid circular dependencies

Add Model objects to forecasting #1558

Add Model objects to forecasting #1558

Conversation

canerturkmen commented Feb 23, 2022

canerturkmen commented Feb 23, 2022 • edited Loading

Innixma left a comment

Choose a reason for hiding this comment

szha commented Mar 2, 2022

dcmaddix left a comment

Choose a reason for hiding this comment

Innixma left a comment

Choose a reason for hiding this comment

canerturkmen commented Mar 23, 2022

canerturkmen commented Mar 23, 2022

canerturkmen commented Mar 23, 2022

Innixma commented Mar 23, 2022

szha commented Mar 23, 2022

Innixma left a comment

Choose a reason for hiding this comment

Innixma Mar 23, 2022

Choose a reason for hiding this comment

canerturkmen Mar 24, 2022

Choose a reason for hiding this comment

Innixma Mar 23, 2022

Choose a reason for hiding this comment

jwmueller Mar 24, 2022

Choose a reason for hiding this comment

canerturkmen Mar 24, 2022

Choose a reason for hiding this comment

canerturkmen commented Feb 23, 2022 •

edited

Loading