Support training-only components in pipelines and component graphs by angela97lin · Pull Request #2776 · alteryx/evalml

angela97lin · 2021-09-13T16:44:53Z

codecov · 2021-09-13T16:49:31Z

Codecov Report

Merging #2776 (54de18b) into main (d1e6afb) will increase coverage by 0.1%.
The diff coverage is 100.0%.

@@           Coverage Diff           @@
##            main   #2776     +/-   ##
=======================================
+ Coverage   99.8%   99.8%   +0.1%     
=======================================
  Files        298     298             
  Lines      27681   27731     +50     
=======================================
+ Hits       27613   27663     +50     
  Misses        68      68

Impacted Files	Coverage Δ
evalml/pipelines/component_graph.py	`99.8% <100.0%> (-<0.1%)`	⬇️
evalml/pipelines/components/component_base.py	`100.0% <100.0%> (ø)`
...valml/pipelines/components/estimators/estimator.py	`100.0% <100.0%> (ø)`
...ransformers/preprocessing/drop_rows_transformer.py	`100.0% <100.0%> (ø)`
...s/components/transformers/samplers/base_sampler.py	`89.4% <100.0%> (+0.3%)`	⬆️
...l/pipelines/components/transformers/transformer.py	`100.0% <100.0%> (ø)`
evalml/tests/component_tests/test_components.py	`100.0% <100.0%> (ø)`
...valml/tests/pipeline_tests/test_component_graph.py	`100.0% <100.0%> (ø)`
evalml/tests/pipeline_tests/test_pipelines.py	`99.9% <100.0%> (+0.1%)`	⬆️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d1e6afb...54de18b. Read the comment docs.

…681_training_only

eccabay · 2021-09-16T14:24:20Z


-        outputs = self._compute_features(self.compute_order, X, y, False)
+        outputs = self._compute_features(
+            self.compute_order, X, y, fit=False, evaluate_training=True


I may be misunderstanding something here - why should evaluate_training be true during transform, but not during fit?

@eccabay Good question! I think the language here is a bit confusing, but evaluate_training is a flag for determining whether or not we want to compute "training-only" components, such as the samplers or the DropRowsComponent when we are not fitting.

We decided that there should be a difference between transform, which takes the pipeline and calls transform on every component vs predict, where we want predictions on unseen data and should not evaluate training-only components. This enables us to keep our current behavior, but also allow users to use pipelines as just a sequence of transformations.

Does that make sense? (or did I confuse you further?) 😄

Ah yeah, that makes a lot of sense. Thanks! 😃

I think the confusing thing though is that it's not clear from the name that evaluate_training only applies during transform and predict and not fit.

Maybe this is what @eccabay was getting at, but when I read this code, I see evaluate_training=False in ComponentGraph.fit which makes me think that the training only components are not evaluated during fit but that's not true because evaluate_training only applies if fit=False:

if fit: output = component_instance.fit_transform(x_inputs, y_input) elif component_instance.training_only and evaluate_training is False: output = x_inputs, y_input else: output = component_instance.transform(x_inputs, y_input)

Maybe we should make this clearer somehow?

freddyaboulton

@angela97lin Thank you for doing this! I'm surprised this wasn't a bigger diff honestly!

I left some coverage suggestions as well as a comment on @eccabay 's original question - maybe we can make the parameter name/docstring clearer about what's happening?

I do think this change (the addition of training-only components) is the kind of thing that will eventually merit a redesign/refactor of the component graph. I think right now we encode whether we should run training-only components via the method that's called:

fit - yes
transform - yes
predict - no
compute_estimator_features - no
fit_features - yes

That kind of behavior is not really clear from the method names and we have a lot of methods for generating features that it's hard to keep it all straight (fit_features vs compute_estimator_features vs _fit_transform_features_helper vs _compute_features) especially in light of the training-only requirement.

I think this is good to merge but I wonder what your thoughts are on addressing the tech debt that I think has been accruing.

freddyaboulton · 2021-09-16T14:41:03Z


    name = "Drop Rows Transformer"
    modifies_target = True
+    training_only = True


Let's make this an abstract method?

The data coming out of the pipeline could be completely wrong if we default to false and that's wrong.

I've made this abstract, but Transformer still sets this value to False, so any subclass of Transformer will have a default value of False... which might defeat the purpose of having this as abstract 😅

I think this is okay though, since most of our components will have training_only as False for now... but open to thoughts and concerns :)

freddyaboulton · 2021-09-16T14:52:24Z


-        outputs = self._compute_features(self.compute_order, X, y, False)
+        outputs = self._compute_features(
+            self.compute_order, X, y, fit=False, evaluate_training=True


I think the confusing thing though is that it's not clear from the name that evaluate_training only applies during transform and predict and not fit.

Maybe this is what @eccabay was getting at, but when I read this code, I see evaluate_training=False in ComponentGraph.fit which makes me think that the training only components are not evaluated during fit but that's not true because evaluate_training only applies if fit=False:

if fit: output = component_instance.fit_transform(x_inputs, y_input) elif component_instance.training_only and evaluate_training is False: output = x_inputs, y_input else: output = component_instance.transform(x_inputs, y_input)

Maybe we should make this clearer somehow?

freddyaboulton · 2021-09-16T15:02:40Z

        pipeline.transform(X, y)
+
+
+@patch("evalml.pipelines.components.LogisticRegressionClassifier.fit")


@angela97lin Can you please add coverage for fit_features and compute_estimator_features?

The expectation is that the training only components are run during fit_features and not compute_estimator_features

Of course, great idea 😁

chukarsten

LGTM. Just a nit about using the same X,y three times, but very clear overall. Thanks for doing this!

chukarsten · 2021-09-16T15:26:46Z

+    X = pd.DataFrame(
+        {
+            "a": [i for i in range(9)] + [np.nan],
+            "b": [i % 3 for i in range(10)],
+            "c": [i % 7 for i in range(10)],
+        }
+    )
+    y = pd.Series([0] * 5 + [1] * 5)


Admittedly this is on my backburner's backburner - but are able to either use an existing X/y pair for these tests or create a module level test fixture for these additional tests?

Thank you for holding me accountable, done!

chukarsten · 2021-09-16T15:28:33Z

+        parameters={"Drop Rows Transformer": {"indices_to_drop": [0, 9]}},
+    )
+    pipeline.fit(X, y)
+    assert len(mock_fit.call_args[0][0]) == 8


There's a time when your brains are too big for me. This is one of them, haha. Why should this be true if the issue is completed successfully?

Hahaha, I think this issue is pretty confusing! But it's 8 here because during fit / training time, we evaluate all components, regardless of whether or not they're training_only. So the DropRowsTransformer will drop the two rows during fit --> a total of 8 rows left!

chukarsten · 2021-09-16T15:30:54Z

+    )
+    pipeline.fit(X, y)
+    preds = pipeline.predict(X)
+    assert len(preds) == 10


Is the jist here that no rows have been dropped during prediction since DropRowsTransformer is evaluate_training=True?

angela97lin · 2021-09-17T02:53:16Z

@freddyaboulton I think your instinct is on point! It seems like there's a lot of confusion amongst everyone, which probably indicates that something is not super clear, either in the implementation or in the design 😅

I renamed evaluate_training to evaluate_training_only_components to be as explicit as possible about what the flag is, and filed #2795 with some of my thoughts about the different methods. TLDR is, I wonder if there's a need for all of these methods, when in reality, we've been using some for just time series work and internal implementations. I'll add test coverage in this PR for those methods regardless though since we still have them around 😁

init

ccee608

angela97lin self-assigned this Sep 13, 2021

angela97lin added 9 commits September 13, 2021 14:33

move logic to component base and update component graph

c6b922b

Merge branch 'main' into 2681_training_only

3f1f231

linting

fb93e57

added test but hmm...

186aa69

Merge branch 'main' into 2681_training_only

fe33b7c

linting

c1612b0

Merge branch 'main' into 2681_training_only

767f2f5

add evaluate_training flag and update tests

2d6ce05

Merge branch '2681_training_only' of github.com:alteryx/evalml into 2…

e0e92f5

…681_training_only

angela97lin marked this pull request as ready for review September 15, 2021 20:16

angela97lin requested review from ParthivNaresh, bchen1116, chukarsten, dsherry, eccabay, freddyaboulton and jeremyliweishih September 15, 2021 20:16

eccabay reviewed Sep 16, 2021

View reviewed changes

freddyaboulton approved these changes Sep 16, 2021

View reviewed changes

chukarsten approved these changes Sep 16, 2021

View reviewed changes

angela97lin added 2 commits September 16, 2021 22:37

address pr comments

698eddf

merge main

93d8cd1

angela97lin mentioned this pull request Sep 17, 2021

Refactor component graph impl #2795

Closed

angela97lin added 2 commits September 16, 2021 23:27

address comments and add tests

13dac30

fix mock components in tests

54de18b

angela97lin merged commit 0440bec into main Sep 17, 2021

angela97lin deleted the 2681_training_only branch September 17, 2021 04:20

chukarsten mentioned this pull request Oct 1, 2021

Release v0.34.0 #2864

Merged

		pipeline.transform(X, y)


		@patch("evalml.pipelines.components.LogisticRegressionClassifier.fit")

Uh oh!

Conversation

angela97lin commented Sep 13, 2021

Uh oh!

codecov Bot commented Sep 13, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

freddyaboulton left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chukarsten left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

angela97lin commented Sep 17, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

codecov Bot commented Sep 13, 2021 •

edited

Loading

angela97lin commented Sep 17, 2021 •

edited

Loading