Proof of Concept: Separating the Engine from AutoMLSearch #1975

freddyaboulton · 2021-03-12T15:46:27Z

Pull Request Description

Fixes #1914. Fixes #1972

I think there are good reasons to go this route:

Moves us away from using callbacks that have a reference to AutoMLSearch. The engine is solely responsible for submitting "work". AutoML is in charge of updating its progress and results.
Less code duplication. Our old interface required developers to implement logic for processing a whole batch of pipelines. There's a lot of core logic that goes into that implementation: when to use the ensemble split, when to terminate mid-batch, etc, that's hard to share across all engine implementations. With this approach, the engine only does "work". AutoML handles all the surrounding ml logic.
Lets users "bring their own engine" by allowing the engine to be passed as an AutoMLSearch argument. There is also less code to write to implement a new engine. Developers/users only need to worry about how to train/score/evaluate a single pipeline. AutoMLSearch handles the batch-wide logic.

Demo with dask

Cons (strategically placed at the bottom 😂):

Currently breaks how our console output/log works for the sequential case. I think we're going to have to do this for any approach we take though. See Dask Logging #1972. Edit: Pushed up a change to how the logging works that works well for general engines. The messages are written to the log once the evaluation job completes.
We need the engine to return the stacktrace if a computation fails. I think that's possible for the sequential case and dask but not sure how easy that is in the general case.
Still thinking about others. I imagine I'll have a better idea as I get the tests to pass.

Performance comparison: About 70% faster on AutoML

After creating the pull request: in order to pass the release_notes_updated check you will need to update the "Future Release" section of docs/source/release_notes.rst to include this pull request by adding :pr:123.

codecov · 2021-03-12T22:02:22Z

Codecov Report

Merging #1975 (9013565) into main (23fb534) will increase coverage by 0.1%.
The diff coverage is 100.0%.

@@            Coverage Diff            @@
##             main    #1975     +/-   ##
=========================================
+ Coverage   100.0%   100.0%   +0.1%     
=========================================
  Files         288      291      +3     
  Lines       23428    23698    +270     
=========================================
+ Hits        23418    23688    +270     
  Misses         10       10

Impacted Files	Coverage Δ
evalml/automl/callbacks.py	`100.0% <ø> (ø)`
evalml/automl/automl_search.py	`100.0% <100.0%> (ø)`
evalml/automl/engine/__init__.py	`100.0% <100.0%> (ø)`
evalml/automl/engine/dask_engine.py	`100.0% <100.0%> (ø)`
evalml/automl/engine/engine_base.py	`100.0% <100.0%> (ø)`
evalml/automl/engine/sequential_engine.py	`100.0% <100.0%> (ø)`
evalml/automl/utils.py	`100.0% <100.0%> (ø)`
evalml/tests/automl_tests/dask_test_utils.py	`100.0% <100.0%> (ø)`
evalml/tests/automl_tests/test_automl.py	`100.0% <100.0%> (ø)`
evalml/tests/automl_tests/test_automl_dask.py	`100.0% <100.0%> (ø)`
... and 10 more

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 23fb534...9013565. Read the comment docs.

freddyaboulton · 2021-03-16T15:31:50Z

evalml/automl/engine/ray_engine.py

@@ -0,0 +1,60 @@
+from evalml.automl.engine.engine_base import (


I'm including the ray engine to illustrate the flexibility of the proposed design. I realize ray integration is out-of-scope and could be done later.

Are we planning on keeping this then? I see a lot of lines not covered by codecov 👀

I'll get rid of it before merge!

freddyaboulton · 2021-03-19T19:51:15Z

evalml/automl/automl_search.py

@@ -562,14 +577,7 @@ def _should_continue(self):
        if self._interrupted:
            return False

-        # for add_to_rankings


These lines were added for add_to_rankings and baseline evaluation because the engine used to call the _should_continue_callback. The engine doesn't do that anymore so we can delete these lines!

freddyaboulton · 2021-03-23T22:57:31Z

Performance tests are here!

* Added DaskEngine unit tests * Added AutoMLSearch unit tests utilizing Dask Engine

… 1914-engine-redesign

kmax12

interesting stuff. left some comments based on a quick review

kmax12 · 2021-03-31T18:19:59Z

evalml/automl/automl_search.py

+        else:
+            self._engine = engine
+
+        self.automl_data = AutoMLData(self.ensembling_indices,


take this with a grain of salt since i only am doing a quick pass, but i wonder if AutoMLConfig or something like that might be a better name for this. When I see "data" I assume it is referring to X and y

kmax12 · 2021-03-31T18:28:09Z

evalml/automl/automl_search.py

@@ -538,8 +556,8 @@ def _find_best_pipeline(self):
                    train_indices = self.data_splitter.transform_sample(X_train, y_train)
                    X_train = X_train.iloc[train_indices]
                    y_train = y_train.iloc[train_indices]
-                best_pipeline = self._engine.train_pipeline(best_pipeline, X_train, y_train,
-                                                            self.optimize_thresholds, self.objective)
+                best_pipeline = train_pipeline(best_pipeline, X_train, y_train,


i wonder if we want this to be called through the engine?

a use case to imagine is that you are running evalml on your laptop and you've connected it to a remote dask cluster that has plenty of compute (perhaps the X and y are reference to dask dataframes where the data is remote too). you want all the training to happen on the remote dask cluster.

if i understand the code correctly, right now the search phase will happen on the remote cluster, but the final training of the pipeline would happen locally.

Good call! That was my intention but I think I forgot to actually make the change hehe.

kmax12 · 2021-03-31T18:37:16Z

evalml/automl/automl_search.py

+                    self._pre_evaluation_callback(pipeline)
+                    computation = self._engine.submit_evaluation_job(self.automl_data, pipeline, self.X_train, self.y_train)
+                    computations.append(computation)
+                while self._should_continue() and len(computations) > 0:


what is the behavior of this when using the dask engine and all pipelines in a batch are currently computing? does it just iterate as fast as it can over and over trying to find whatever computation has completed?

I wonder if this code can be simplified using a method like: https://docs.dask.org/en/latest/futures.html#distributed.as_completed

You're right about the current behavior! I thought about using/writing our own as_completed but the one tricky thing is that we want to stop the search as soon as one of our stopping criteria is met but as_completed will only pull off futures if they're done. I think this will make it hard to respect time based stopping rules.

For example, let's say the user has a 2 hour stop limit and fires off four jobs. The first three finish in 90 minutes but the fourth one won't finish until 6 hours.

If we do this:

for computation in as_completed(computations): if not self._should_continue(): break

We actually won't reach the break until hour 6 because that's when the final future finished.

I think we can definitely explore ways to consolidate/simplify the logic for pulling futures off of our queue but I'd rather leave that improvement for a separate PR.

i wonder if we need some sort of sleep call in the loop then. if you have a bunch of pipelines training for a while with none finishing, i dont think you want this loop iterating as fast as it can.

Great point Max. Yes, a short sleep here is probably a good idea for now.

Good call about distributed.as_completed here. I was going to suggest asyncio. Eventually we can make the whole automl API non-blocking ( #1047 ), but that is definitely out of scope for the current work though.

chukarsten

Awesome work Freddy, really great work. This is a huge step in the right direction toward separating out this code from AutoMLSearch. Take or leave the comment I left about the refactoring those similar while loops. I think they might be a little different and if the refactor looks less clear, then forget it.

chukarsten · 2021-03-17T18:48:26Z

evalml/automl/automl_search.py

@@ -541,8 +556,8 @@ def _find_best_pipeline(self):
                    train_indices = self.data_splitter.transform_sample(X_train, y_train)
                    X_train = X_train.iloc[train_indices]
                    y_train = y_train.iloc[train_indices]
-                best_pipeline = self._engine.train_pipeline(best_pipeline, X_train, y_train,
-                                                            self.optimize_thresholds, self.objective)
+                best_pipeline = train_pipeline(best_pipeline, X_train, y_train,


Is this taken away from the engine because it probably just makes sense to do it locally since it's one training computation?

This is my mistake. I think the right thing to do is have the engine train all pipelines but I missed this spot hehe.

chukarsten · 2021-03-31T15:38:41Z

evalml/automl/automl_search.py

                loop_interrupted = False
            except KeyboardInterrupt:
                loop_interrupted = True
                if self._handle_keyboard_interrupt():
-                    break
+                    self._interrupted = True


Have you tried testing the keyboard interrupt manually? Just curious. It seems like the probability of being able to keyboard interrupt the parallel evaluation will be low as you'd have to interrupt in the middle of the loop that's submitting the jobs. Once those futures are sent off to be evaluated, I don't think we can interrupt them anymore. I'm not stressing it that much. We might have to implement a different type of keyboard interrupt.

Yep I think you're right. There is a small window of time where if someone does a Ctrl-C, it won't be caught by this except, if that's what you're referring to. It doesn't bother me too much either.

@freddyaboulton it would be ideal if we could cancel the futures here when the user opts to exit.

chukarsten · 2021-03-31T16:02:59Z

evalml/automl/automl_search.py

+        while computations:
+            computation = computations.pop(0)
+            if computation.done():
+                try:
+                    fitted_pipeline = computation.get_result()
+                    fitted_pipelines[fitted_pipeline.name] = fitted_pipeline
+                except Exception as e:
+                    logger.error(f'Train error for {pipeline.name}: {str(e)}')
+                    tb = traceback.format_tb(sys.exc_info()[2])
+                    logger.error("Traceback:")
+                    logger.error("\n".join(tb))
+            else:
+                computations.append(computation)


IS this chunk worth pulling into a function? I think you use it at least 3 times.

It's not exactly the same because the thing you get back from get_result is not the same between training and scoring and the logic for handling exceptions is not the same as well.

This might be a good spot for @kmax12 's suggestion to use as_completed or something like it.

If we do that, I think this will look like:

for computation in as_completed(computations): try: fitted_pipeline = computation.get_result() fitted_pipelines[fitted_pipeline.name] = fitted_pipeline except Exception as e: logger.error(f'Train error for {pipeline.name}: {str(e)}') tb = traceback.format_tb(sys.exc_info()[2]) logger.error("Traceback:") logger.error("\n".join(tb))

I think at that point we're not reusing the while-loop logic so the differences between the two methods might be clearer.

If you're on board I'll file an issue after merge!

Yes please do! This proposal looks great. I am still confused how keyboard interrupt would fit with that example, but I like that time.sleep isn't needed.

Also FWIW, we do have the option to stop supporting keyboard interrupt if it is significantly hard to maintain. Its cool, but not worth too much of our time IMO.

Filed #2096 to track consolidating these two while loops!

chukarsten · 2021-03-31T16:27:11Z

evalml/automl/engine/sequential_engine.py

-        return new_pipeline_ids
+    def done(self) -> bool:
+        """Is the computation done?"""
+        return True


Is this definitely true, though?

chukarsten · 2021-03-31T16:28:40Z

evalml/tests/automl_tests/dask_testing.py

+
+
+# Top-level replacement for AutoML object to supply data for testing purposes.
+def err_call(*args, **kwargs):


Sorry - this is my bad...might want to drop an ignore in here for the time being.

I think providing just a docstring will make codecov happy so I'll push that up now.

chukarsten · 2021-03-31T16:35:38Z

evalml/tests/automl_tests/test_automl.py

+    X, y = X_y_binary
+
+    mock_input.side_effect = user_input
+    mock_future_get_result.side_effect = KeyboardInterruptOnKthPipeline(k=when_to_interrupt, starting_index=1)


OK, cool, I like this test a lot and I think that's really clever and I think that allays my suspicions about when the keyboard interrupt does its job 👍

chukarsten · 2021-03-31T18:51:09Z

evalml/tests/automl_tests/test_automl.py

-        else:
-            correct_output = automl._pipelines_per_batch
-        assert output.count(f"Batch {batch_number}: ") == correct_output
+    assert output.count("Batch Number") == max_batches


bchen1116

Looks good to me! Left two comments about renaming and adding doc string, but the tests are very solid! Love the comparisons between the performance between parallel and sequential.

bchen1116 · 2021-04-01T14:33:19Z

evalml/automl/engine/engine_base.py

+    return {'cv_data': cv_data, 'training_time': training_time, 'cv_scores': cv_scores, 'cv_score_mean': cv_score_mean}, cv_pipeline, logger
+
+
+def evaluate_pipeline(pipeline, automl_data, X, y, logger):


can we add docstring for this and for score_pipeline

bchen1116 · 2021-04-01T14:59:23Z

evalml/automl/utils.py

@@ -110,3 +112,8 @@ def check_all_pipeline_names_unique(pipelines):
        plural, tense = ("s", "were") if len(duplicate_names) > 1 else ("", "was")
        duplicates = ", ".join([f"'{name}'" for name in sorted(duplicate_names)])
        raise ValueError(f"All pipeline names must be unique. The name{plural} {duplicates} {tense} repeated.")
+
+
+AutoMLData = namedtuple("AutoMLData", ["ensembling_indices", "data_splitter", "problem_type",


Agree with Max about naming here

Yep I also prefer AutoMLConfig

dsherry

@freddyaboulton this is amazing! So many good changes :D the dask engine code looks 💯 . The EngineComputation abstraction is super helpful. Well done to you and to @chukarsten 😁

I don't have many blocking comments. I agree with what Max said about adding a sleep in the main loop. I left a couple comments about the keyboard interrupt stuff in the main loop(s). I do think we should call that config object AutoMLConfig as suggested. And I left a couple comments on the test utils.

Also, I know dask is included in evalml as a core requirement through featuretools, but for clarity I think we should list dask in our core-requirements.txt after featuretools.

Since this is a significant change, please run the performance tests once more before merging as a protective measure.

dsherry · 2021-04-02T21:57:21Z

evalml/automl/automl_search.py

@@ -83,7 +89,8 @@ def __init__(self,
                 train_best_pipeline=True,
                 pipeline_parameters=None,
                 _ensembling_split_size=0.2,
-                 _pipelines_per_batch=5):
+                 _pipelines_per_batch=5,
+                 engine=None):


+1! If its time to make this method "public", let's doc it.

dsherry · 2021-04-02T21:58:16Z

evalml/automl/automl_search.py

+        else:
+            self._engine = engine
+
+        self.automl_data = AutoMLData(self.ensembling_indices,


dsherry · 2021-04-02T22:06:35Z

evalml/automl/automl_search.py

+                    self._pre_evaluation_callback(pipeline)
+                    computation = self._engine.submit_evaluation_job(self.automl_data, pipeline, self.X_train, self.y_train)
+                    computations.append(computation)
+                while self._should_continue() and len(computations) > 0:


Great point Max. Yes, a short sleep here is probably a good idea for now.

Good call about distributed.as_completed here. I was going to suggest asyncio. Eventually we can make the whole automl API non-blocking ( #1047 ), but that is definitely out of scope for the current work though.

dsherry · 2021-04-02T22:10:58Z

evalml/automl/automl_search.py

                loop_interrupted = False
            except KeyboardInterrupt:
                loop_interrupted = True
                if self._handle_keyboard_interrupt():
-                    break
+                    self._interrupted = True


Yep I think you're right. There is a small window of time where if someone does a Ctrl-C, it won't be caught by this except, if that's what you're referring to. It doesn't bother me too much either.

@freddyaboulton it would be ideal if we could cancel the futures here when the user opts to exit.

dsherry · 2021-04-02T22:16:50Z

evalml/automl/automl_search.py

+        self._pre_evaluation_callback(baseline)
+        logger.info(f"Evaluating Baseline Pipeline: {baseline.name}")
+        computation = self._engine.submit_evaluation_job(self.automl_data, baseline, self.X_train, self.y_train)
+        data, pipeline, job_log = computation.get_result()


Sweet

For future-proofing, could we do

result = computation.get_result() data = result.get('data') pipeline = result.get('pipeline') job_log = result.get('job_log')

Then we can add/reorder fields from the results in the future without making a change which breaks the API.

dsherry · 2021-04-02T23:04:36Z

evalml/tests/automl_tests/dask_testing.py

+                         additional_objectives=additional_objectives,
+                         optimize_thresholds=optimize_thresholds,
+                         error_callback=error_callback,
+                         random_seed=random_seed)


To avoid strange errors arising from accidentally mutating global state: could you please create this stuff in a test fixture, which runs once per test, instead of once at import-time as things stand currently?

You could even move it to evalml/tests/conftest.py.

dsherry · 2021-04-02T23:09:05Z

evalml/tests/automl_tests/test_automl_search_regression.py

@@ -82,7 +82,7 @@ def add_result_callback(results, trained_pipeline, automl_obj, counts=counts):
                          add_result_callback=add_result_callback, n_jobs=1)
    automl.search()

-    assert counts["start_iteration_callback"] == max_iterations
+    assert counts["start_iteration_callback"] == len(get_estimators("regression")) + 1


Interesting, why did this have to change? Were we not calling this for the baseline pipeline before? That seems like a positive change to me.

dsherry · 2021-04-02T23:10:18Z

evalml/tests/automl_tests/test_dask_engine.py

+class TestDaskEngine(unittest.TestCase):
+
+    @classmethod
+    def setUpClass(cls) -> None:


Remove the -> None ?

dsherry · 2021-04-02T23:11:05Z

evalml/tests/automl_tests/test_dask_engine.py

+
+    @classmethod
+    def setUpClass(cls) -> None:
+        cls.client = Client()


Nice. This starts a local dask cluster, once for all of the tests in this class, correct? Any idea how many cores it'll take, how many workers it'll have?

@dsherry instantiating with no arguments created a Client() based off a LocalCluster() with number of workers related to the number of available cores on the machine running it.

dsherry · 2021-04-02T23:11:37Z

evalml/tests/automl_tests/test_dask_engine.py

+
+    @classmethod
+    def tearDownClass(cls) -> None:
+        cls.client.close()


Very nice

I suggest moving this to be at the top near setup so its obvious to new readers what's going on.

dsherry · 2021-04-02T23:33:40Z

evalml/automl/automl_search.py

-                                        pre_evaluation_callback=self._pre_evaluation_callback,
-                                        post_evaluation_callback=self._post_evaluation_callback)
+        if not engine:
+            self._engine = SequentialEngine()


@freddyaboulton @chukarsten :

I am thinking that sooner or later we should switch this to default to DaskEngine with a local dask cluster. What would we need to do in order to make that happen? By my accounting: make sure the team feels good about this move, list dask in our requirements, update our documentation to explain what's going on and how to configure a remote cluster, and confirm that the perf testing we've done is adequate to make this change.

I was kind of thinking the same thing working through all this. It seems that, in the grand scheme of things, a DaskEngine with 1 worker eliminates the need for a SequentialEngine. And yeah, that sounds about right, but the proper documentation for how to limit the DaskEngine to a certain number of workers and perhaps an expectation of how many default workers there should be are required.

So currently the AutoMLSearch is designed to take in an engine argument. The EngineBase object expects a client argument, in the case of a DaskEngine. And the client object, when it's not run as a default instance, takes a cluster argument. The cluster is where you lay out all the details behind the resources and amount of parallelism you want to use. As such, I think we need to answer the questions 1.) what are the default settings for workers we want for the DaskEngine? and 2.) what are the default settings for workers we want for the AutoMLSearch? We could either go the way of trying to make it as fast as possible for the user by default or protect their resource by default. I think whichever of the two philosophies (faster vs. resource conservative) we choose, we should probably be consistent between DaskEngine and AutoMLSearch.

…and sleep in automl search loop.

freddyaboulton · 2021-04-07T15:14:38Z

Updated perf tests results are here. Same performance improvements after making the requested changes so I think this is good to merge!

… 1914-engine-redesign

…skEngine`` #1975. - Added optional ``engine`` argument to ``AutoMLSearch`` #1975 - Added a warning about how time series support is still in beta when a user passes in a time series problem to ``AutoMLSearch`` #2118 - Added ``NaturalLanguageNaNDataCheck`` data check #2122 - Added ValueError to ``partial_dependence`` to prevent users from computing partial dependence on columns with all NaNs #2120 - Added standard deviation of cv scores to rankings table #2154 - Fixed ``BalancedClassificationDataCVSplit``, ``BalancedClassificationDataTVSplit``, and ``BalancedClassificationSampler`` to use ``minority:majority`` ratio instead of ``majority:minority`` #2077 - Fixed bug where two-way partial dependence plots with categorical variables were not working correctly #2117 - Fixed bug where ``hyperparameters`` were not displaying properly for pipelines with a list ``component_graph`` and duplicate components #2133 - Fixed bug where ``pipeline_parameters`` argument in ``AutoMLSearch`` was not applied to pipelines passed in as ``allowed_pipelines`` #2133 - Fixed bug where ``AutoMLSearch`` was not applying custom hyperparameters to pipelines with a list ``component_graph`` and duplicate components #2133 - Removed ``hyperparameter_ranges`` from Undersampler and renamed ``balanced_ratio`` to ``sampling_ratio`` for samplers #2113 - Renamed ``TARGET_BINARY_NOT_TWO_EXAMPLES_PER_CLASS`` data check message code to ``TARGET_MULTICLASS_NOT_TWO_EXAMPLES_PER_CLASS`` #2126 - Modified one-way partial dependence plots of categorical features to display data with a bar plot #2117 - Renamed ``score`` column for ``automl.rankings`` as ``mean_cv_score`` #2135 - Fixed ``conf.py`` file #2112 - Added a sentence to the automl user guide stating that our support for time series problems is still in beta. #2118 - Fixed documentation demos #2139 - Update test badge in README to use GitHub Actions #2150 - Fixed ``test_describe_pipeline`` for ``pandas`` ``v1.2.4`` #2129 - Added a GitHub Action for building the conda package #1870 #2148 .. warning:: - Renamed ``balanced_ratio`` to ``sampling_ratio`` for the ``BalancedClassificationDataCVSplit``, ``BalancedClassificationDataTVSplit``, ``BalancedClassficationSampler``, and Undersampler #2113 - Deleted the "errors" key from automl results #1975 - Deleted the ``raise_and_save_error_callback`` and the ``log_and_save_error_callback`` #1975 - Fixed ``BalancedClassificationDataCVSplit``, ``BalancedClassificationDataTVSplit``, and ``BalancedClassificationSampler`` to use minority:majority ratio instead of majority:minority #2077

freddyaboulton added 4 commits March 11, 2021 18:06

Sequential Engine implementation

520b657

Adding DaskEngine.

a40a612

Fixing some unit tests.

f33610c

Fixing unit tests.

df4ab9e

freddyaboulton added 6 commits March 15, 2021 10:56

Fixing test_callback for minimal dependencies.

e6bea3d

Making logging work better for non-sequential engines.

5e47439

Scattering data in DaskEngine.

f0da542

Merging main.

5745998

Adding ray engine.

4ad4071

Linting ray engine.

873a681

freddyaboulton commented Mar 16, 2021

View reviewed changes

freddyaboulton added 3 commits March 16, 2021 17:01

Deleting code that should not be covered.

e6e7066

Merge branch 'main' into 1914-engine-redesign

b9e1a31

Deleting save error callback tests in automl.

5f133d3

freddyaboulton mentioned this pull request Mar 18, 2021

Dask Logging #1972

Closed

Deleting unneeded lines from AutoMLSearch._should_continue.

2023fc3

freddyaboulton commented Mar 19, 2021

View reviewed changes

freddyaboulton added 3 commits March 22, 2021 18:11

Updating woodwork.

8860bda

Fixing leak.

505e9c0

Use name instead of custom name.

f41b2d5

freddyaboulton and others added 6 commits March 23, 2021 19:07

Merge branch 'main' into 1914-engine-redesign

ef2ed31

sorting imports

24de1e0

1914 engine redesign dask (#1990)

5c5046e

* Added DaskEngine unit tests * Added AutoMLSearch unit tests utilizing Dask Engine

Removing callbacks from api ref.

3f20962

Merge branch '1914-engine-redesign' of github.com:alteryx/evalml into…

02e249e

… 1914-engine-redesign

Deleting un-used lines in test_automl.py

cee718e

freddyaboulton marked this pull request as ready for review March 24, 2021 16:28

auto-assign bot assigned freddyaboulton Mar 24, 2021

kmax12 reviewed Mar 31, 2021

View reviewed changes

chukarsten approved these changes Mar 31, 2021

View reviewed changes

freddyaboulton added 2 commits March 31, 2021 15:11

Use engine to train best pipeline

370aca4

Adding docstring explaining how SequentialComputation works

8e14924

bchen1116 approved these changes Apr 1, 2021

View reviewed changes

dsherry approved these changes Apr 2, 2021

View reviewed changes

dsherry reviewed Apr 2, 2021

View reviewed changes

freddyaboulton mentioned this pull request Apr 6, 2021

Use as_completed in train_pipelines/score_pipelines #2096

Open

freddyaboulton and others added 9 commits April 6, 2021 13:36

PR changes - Don't pop in AutoMLSearch while loop. Adding docstrings …

9cb5fe8

…and sleep in automl search loop.

Merge branch 'main' into 1914-engine-redesign

41d0a0f

Linting

bc1abc4

Updating latest_dependencies_version.txt to include dask.

1a0e93e

Updated docstrings in dask_engine.

4ffb5a9

Cache rename.

1e6b93e

Updated the name of the dask testing utils file.

51c5a61

Added cancel test.

48dd8bc

Merge branch 'main' into 1914-engine-redesign

6ddd828

freddyaboulton added 5 commits April 7, 2021 13:57

Resend data if cancelled

31bc8d1

Merge branch 'main' into 1914-engine-redesign

36516e1

Merge branch '1914-engine-redesign' of github.com:alteryx/evalml into…

d6a08d5

… 1914-engine-redesign

Trigger Docs

d711caf

Merge branch 'main' into 1914-engine-redesign

9013565

freddyaboulton merged commit 0b06740 into main Apr 8, 2021

freddyaboulton mentioned this pull request Apr 16, 2021

Don't use filepath in callback's logged message #2149

Closed

chukarsten mentioned this pull request Apr 20, 2021

Release v0.23.0 #2165

Merged

freddyaboulton deleted the 1914-engine-redesign branch June 3, 2021 14:02

freddyaboulton mentioned this pull request Jun 16, 2021

Speed up unit tests #2365

Merged

angela97lin mentioned this pull request Jun 28, 2021

Ensure logging works with Parallel Engines #2451

Open

		@@ -0,0 +1,60 @@
		from evalml.automl.engine.engine_base import (



		# Top-level replacement for AutoML object to supply data for testing purposes.
		def err_call(args, *kwargs):

		return {'cv_data': cv_data, 'training_time': training_time, 'cv_scores': cv_scores, 'cv_score_mean': cv_score_mean}, cv_pipeline, logger


		def evaluate_pipeline(pipeline, automl_data, X, y, logger):

Proof of Concept: Separating the Engine from AutoMLSearch #1975

Proof of Concept: Separating the Engine from AutoMLSearch #1975

Conversation

freddyaboulton commented Mar 12, 2021 • edited

Pull Request Description

Demo with dask

Performance comparison: About 70% faster on AutoML

codecov bot commented Mar 12, 2021 • edited

Codecov Report

freddyaboulton Mar 16, 2021 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

freddyaboulton commented Mar 23, 2021 • edited

kmax12 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kmax12 Mar 31, 2021 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kmax12 Apr 1, 2021 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

chukarsten left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

freddyaboulton Mar 31, 2021 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bchen1116 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dsherry left a comment • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

freddyaboulton commented Apr 7, 2021

freddyaboulton commented Mar 12, 2021 •

edited

codecov bot commented Mar 12, 2021 •

edited

freddyaboulton Mar 16, 2021 •

edited

freddyaboulton commented Mar 23, 2021 •

edited

kmax12 Mar 31, 2021 •

edited

kmax12 Apr 1, 2021 •

edited

freddyaboulton Mar 31, 2021 •

edited

dsherry left a comment •

edited