Classes and functions to explore and reproduce the performance obtained by thousands of MLBlocks pipelines and templates across hundreds of datasets.
- Free software: MIT license
- Documentation: https://HDI-Project.github.io/piex
- Homepage: https://github.com/HDI-Project/piex
This repository contains a collection of classes and functions which allows a user to easily explore the results of a series of experiments run by team MIT using MLBlocks pipelines over a large collection of Datasets.
Along with this library we are releasing a number of fitted pipelines, their performance on cross validation, test data and metrics. The results of these experiments were stored in a Database and later on uploaded to Amazon S3, from where they can be downloaded and analyzed using the Pipeline Explorer.
We will continuously add more pipelines, templates and datasets to our experiments and make them publicly available to the community.
These can be used for the following purposes:
- Find what is the best score we found so far for a given dataset and task type (given the search space we defined and our tuners)
- Use information about pipeline performance to do meta learning
Current summary of our experiments is:
| # of | |
|---|---|
| datasets | 453 |
| pipelines | 2115907 |
| templates | 63 |
| tests | 2152 |
Before diving into the software usage, we briefly explain some concepts and terminology.
We call the smallest computational blocks used in a Machine Learning process primitives, which:
- Can be either classes or functions.
- Have some initialization arguments, which MLBlocks calls
init_params. - Have some tunable hyperparameters, which have types and a list or range of valid values.
Primitives can be combined to form what we call Templates, which:
- Have a list of primitives.
- Have some initialization arguments, which correspond to the initialization arguments of their primitives.
- Have some tunable hyperparameters, which correspond to the tunable hyperparameters of their primitives.
Templates can be used to build Pipelines by taking and fixing a set of valid hyperparameters for a Template. Hence, Pipelines:
- Have a list of primitives, which corresponds to the list of primitives of their template.
- Have some initialization arguments, which correspond to the initialization arguments of their template.
- Have some hyperparameter values, which fall within the ranges of valid tunable hyperparameters of their template.
A pipeline can be fitted and evaluated using the MLPipeline API in MLBlocks.
A collection of ~450 datasets was used covering 6 different data modalities and 17 task types.
Each dataset was split using a holdout method in two parts, training and testing, which were used respectively to find and fit the optimal pipeline for each dataset, and to later on evaluate the goodness-of-fit of each pipeline against a specific metric for each dataset.
This collection of datasets is stored in an Amazon S3 Bucket in the D3M format, including the training and testing partitioning, and can be downloaded both using piex or a web browser following this link: https://d3m-data-dai.s3.amazonaws.com/index.html
Throughout our description we will refer to a search process as an experiment or a test. An experiment/test is defined as follows:
- It is given a dataset and a task
- It is given a template
- It then searches using a Bayesian tuning algorithm (using a tuner from our BTB library). Tuning algorithm tests multiple pipelines derived from the template and tries to find the best set of hyperparameters possible for that template on each dataset.
- During the search process, a collection of information is stored in the database and is
available through piex. They are:
- Cross Validation score obtained over the training partition by each pipeline fitted during the search process.
- In parallel, at some points in time the best pipeline already found was validated against the testing data, and the obtained score was also stored in the database.
Each experiment was given one or more of the following configuration values:
- Timeout: Maximum time that the search process is allowed to run.
- Budget: Maximum number of tuning iterations that the search process is allowed to perform.
- Checkpoints: List of points in time, in seconds, where the best pipeline so far was scored against the testing data.
- Pipeline: The name of the template to use to build the pipelines.
- Tuner Type: The type of tuner to use,
gporuniform.
The simplest and recommended way to install the Pipeline Explorer is using pip:
pip install piexAlternatively, you can also clone the repository and install it from sources
git clone [email protected]:HDI-Project/piex.git
cd piex
pip install -e .The S3PipelineExplorer class provides methods to download the results from previous tests executions from S3, see which pipelines obtained the best scores and load them as a dictionary, ready to be used by an MLPipeline.
To start working with it, it needs to be given the name of the S3 Bucket from which the data will be downloaded.
For this examples, we will be using the ml-pipelines-2018 bucket, where the results
of the experiments run for the Machine Learning Bazaar paper can be found.
from piex.explorer import S3PipelineExplorer
piex = S3PipelineExplorer('ml-pipelines-2018')The get_datasets method returns a pandas.DataFrame with information about the
available datasets, their data modalities, task types and task subtypes.
datasets = piex.get_datasets()
datasets.shape(453, 4)
datasets.head()| dataset | data_modality | task_type | task_subtype | |
|---|---|---|---|---|
| 314 | 124_120_mnist | image | classification | multi_class |
| 315 | 124_138_cifar100 | image | classification | multi_class |
| 316 | 124_153_svhn_cropped | image | classification | multi_class |
| 317 | 124_174_cifar10 | image | classification | multi_class |
| 318 | 124_178_coil100 | image | classification | multi_class |
datasets = piex.get_datasets(data_modality='multi_table', task_type='regression')
datasets.head()| dataset | data_modality | task_type | task_subtype | |
|---|---|---|---|---|
| 311 | uu2_gp_hyperparameter_estimation | multi_table | regression | multivariate |
| 312 | uu3_world_development_indicators | multi_table | regression | univariate |
The list of tests that have been executed can be obtained with the method get_tests.
This method returns a pandas.DataFrame that contains a row for each experiment that has been run on each dataset.
This dataset includes information about the dataset, the configuration used for the experiment, such as the
template, the checkpoints or the budget, and information about the execution, such as the timestamp, the exact
software version, the host that executed the test and whether there was an error or not.
Just like the get_datasets, any keyword arguments will be used to filter the results.
import pandas as pd
tests = piex.get_tests()
tests.head().T| 0 | 1 | 2 | 3 | 4 | |
|---|---|---|---|---|---|
| budget | NaN | NaN | NaN | NaN | NaN |
| checkpoints | [900, 1800, 3600, 7200] | [900, 1800, 3600, 7200] | [900, 1800, 3600, 7200] | [900, 1800, 3600, 7200] | [900, 1800, 3600, 7200] |
| commit | 4c7c29f | 4c7c29f | 4c7c29f | 4c7c29f | 4c7c29f |
| dataset | 196_autoMpg | 26_radon_seed | LL0_1027_esl | LL0_1028_swd | LL0_1030_era |
| docker | False | False | False | False | False |
| error | NaN | NaN | NaN | NaN | NaN |
| hostname | ec2-52-14-97-167.us-east-2.compute.amazonaws.com | ec2-18-223-109-53.us-east-2.compute.amazonaws.com | ec2-18-217-79-23.us-east-2.compute.amazonaws.com | ec2-18-217-239-54.us-east-2.compute.amazonaws.com | ec2-18-225-32-252.us-east-2.compute.amazonaws.com |
| image | NaN | NaN | NaN | NaN | NaN |
| insert_ts | 2018-10-24 20:05:01.872 | 2018-10-24 20:05:02.778 | 2018-10-24 20:05:02.879 | 2018-10-24 20:05:02.980 | 2018-10-24 20:05:03.081 |
| pipeline | categorical_encoder/imputer/standard_scaler/xg... | categorical_encoder/imputer/standard_scaler/xg... | categorical_encoder/imputer/standard_scaler/xg... | categorical_encoder/imputer/standard_scaler/xg... | categorical_encoder/imputer/standard_scaler/xg... |
| status | done | done | done | done | done |
| test_id | 20181024200501872083 | 20181024200501872083 | 20181024200501872083 | 20181024200501872083 | 20181024200501872083 |
| timeout | NaN | NaN | NaN | NaN | NaN |
| traceback | NaN | NaN | NaN | NaN | NaN |
| tuner_type | NaN | NaN | NaN | NaN | NaN |
| update_ts | 2018-10-24 22:05:55.386 | 2018-10-24 22:05:57.508 | 2018-10-24 22:05:56.337 | 2018-10-24 22:05:56.112 | 2018-10-24 22:05:56.164 |
| data_modality | single_table | single_table | single_table | single_table | single_table |
| task_type | regression | regression | regression | regression | regression |
| task_subtype | univariate | univariate | univariate | univariate | univariate |
| metric | meanSquaredError | rootMeanSquaredError | meanSquaredError | meanSquaredError | meanSquaredError |
| dataset_id | 196_autoMpg_dataset_TRAIN | 26_radon_seed_dataset_TRAIN | LL0_1027_esl_dataset_TRAIN | LL0_1028_swd_dataset_TRAIN | LL0_1030_era_dataset_TRAIN |
| problem_id | 196_autoMpg_problem_TRAIN | 26_radon_seed_problem_TRAIN | LL0_1027_esl_problem_TRAIN | LL0_1028_swd_problem_TRAIN | LL0_1030_era_problem_TRAIN |
| target | class | log_radon | out1 | Out1 | out1 |
| size | 24 | 160 | 16 | 52 | 32 |
| size_human | 24K | 160K | 16K | 52K | 32K |
| test_features | 7 | 28 | 4 | 10 | 4 |
| test_samples | 100 | 183 | 100 | 199 | 199 |
| train_features | 7 | 28 | 4 | 10 | 4 |
| train_samples | 298 | 736 | 388 | 801 | 801 |
pd.DataFrame(tests.groupby(['data_modality', 'task_type']).size(), columns=['count'])| count | ||
|---|---|---|
| data_modality | task_type | |
| graph | community_detection | 5 |
| graph_matching | 18 | |
| link_prediction | 2 | |
| vertex_nomination | 2 | |
| image | classification | 57 |
| regression | 1 | |
| multi_table | classification | 1 |
| regression | 1 | |
| single_table | classification | 1405 |
| collaborative_filtering | 1 | |
| regression | 430 | |
| time_series_forecasting | 175 | |
| text | classification | 17 |
| timeseries | classification | 37 |
tests = piex.get_tests(data_modality='graph', task_type='link_prediction')
tests[['dataset', 'pipeline', 'checkpoints', 'test_id']]| dataset | pipeline | checkpoints | test_id | |
|---|---|---|---|---|
| 1716 | 59_umls | NaN | [900, 1800, 3600, 7200] | 20181031040541366347 |
| 2141 | 59_umls | graph/link_prediction/random_forest_classifier | [900, 1800, 3600, 7200] | 20181031182305995728 |
The results of the experiments can be seen using the get_experiment_results method.
These results include both the cross validation score obtained by the pipeline during the tuning, as well as the score obtained by this pipeline once it has been fitted using the training data and then used to make predictions over the test data.
Just like the get_datasets, any keyword arguments will be used to filter the results,
including the test_id.
results = piex.get_test_results(test_id='20181031182305995728')
results[['test_id', 'pipeline', 'score', 'cv_score', 'elapsed', 'iterations']]| test_id | pipeline | score | cv_score | elapsed | iterations | |
|---|---|---|---|---|---|---|
| 7464 | 20181031182305995728 | graph/link_prediction/random_forest_classifier | 0.499853 | 0.843175 | 900.255511 | 435.0 |
| 7465 | 20181031182305995728 | graph/link_prediction/random_forest_classifier | 0.499853 | 0.854603 | 1800.885417 | 805.0 |
| 7466 | 20181031182305995728 | graph/link_prediction/random_forest_classifier | 0.499853 | 0.854603 | 3600.005072 | 1432.0 |
| 7467 | 20181031182305995728 | graph/link_prediction/random_forest_classifier | 0.785568 | 0.860000 | 7200.225256 | 2366.0 |
Information about the best pipeline for a dataset can be obtained using the get_best_pipeline method.
This method returns a pandas.Series object with information about the pipeline that obtained the
best cross validation score during the tuning, as well as the template that was used to build it.
Note: This call will download some data in the background the first time that it is run, so it might take a while to return.
piex.get_best_pipeline('185_baseball')id 17385666-31da-4b6e-ab7f-8ac7080a4d55
dataset 185_baseball_dataset_TRAIN
metric f1Macro
name categorical_encoder/imputer/standard_scaler/xg...
rank 0.307887
score 0.692113
template 5bd0ce5249e71569e8bf8003
test_id 20181024234726559170
pipeline categorical_encoder/imputer/standard_scaler/xg...
data_modality single_table
task_type classification
Name: 1149699, dtype: object
Apart from obtaining this information, we can use the load_best_pipeline method
to load its JSON specification, ready to be using in an mlblocks.MLPipeline object.
pipeline = piex.load_best_pipeline('185_baseball')
pipeline['primitives']['mlprimitives.feature_extraction.CategoricalEncoder',
'sklearn.preprocessing.Imputer',
'sklearn.preprocessing.StandardScaler',
'mlprimitives.preprocessing.ClassEncoder',
'xgboost.XGBClassifier',
'mlprimitives.preprocessing.ClassDecoder']
Just like the best pipeline, the best template for a given dataset can be obtained using
the get_best_template method.
This returns just the name of the template that was used to build the best pipeline.
template_name = piex.get_best_template('185_baseball')
template_name'categorical_encoder/imputer/standard_scaler/xgbclassifier'
This can be later on used to explore the template, obtaining its default hyperparameters:
defaults = piex.get_default_hyperparameters(template_name)
defaults{'mlprimitives.feature_extraction.CategoricalEncoder#1': {'copy': True,
'features': 'auto',
'max_labels': 0},
'sklearn.preprocessing.Imputer#1': {'missing_values': 'NaN',
'axis': 0,
'copy': True,
'strategy': 'mean'},
'sklearn.preprocessing.StandardScaler#1': {'with_mean': True,
'with_std': True},
'mlprimitives.preprocessing.ClassEncoder#1': {},
'xgboost.XGBClassifier#1': {'n_jobs': -1,
'n_estimators': 100,
'max_depth': 3,
'learning_rate': 0.1,
'gamma': 0,
'min_child_weight': 1},
'mlprimitives.preprocessing.ClassDecoder#1': {}}
Or obtaining the corresponding tunable ranges, ready to be used with a tuner:
tunable = piex.get_tunable_hyperparameters(template_name)
tunable{'mlprimitives.feature_extraction.CategoricalEncoder#1': {'max_labels': {'type': 'int',
'default': 0,
'range': [0, 100]}},
'sklearn.preprocessing.Imputer#1': {'strategy': {'type': 'str',
'default': 'mean',
'values': ['mean', 'median', 'most_frequent']}},
'sklearn.preprocessing.StandardScaler#1': {'with_mean': {'type': 'bool',
'default': True},
'with_std': {'type': 'bool', 'default': True}},
'mlprimitives.preprocessing.ClassEncoder#1': {},
'xgboost.XGBClassifier#1': {'n_estimators': {'type': 'int',
'default': 100,
'range': [10, 1000]},
'max_depth': {'type': 'int', 'default': 3, 'range': [3, 10]},
'learning_rate': {'type': 'float', 'default': 0.1, 'range': [0, 1]},
'gamma': {'type': 'float', 'default': 0, 'range': [0, 1]},
'min_child_weight': {'type': 'int', 'default': 1, 'range': [1, 10]}},
'mlprimitives.preprocessing.ClassDecoder#1': {}}
The S3PipelineExplorer class also allows cross validating templates and pipelines over any of the datasets.
The simplest use case is cross validating a pipeline over a dataset.
For this, we must pass the ID of the pipeline and the name of the dataset to the method score_pipeline.
The dataset can be the one that was used during the experiments or a different one.
piex.score_pipeline(pipeline['id'], '185_baseball')(0.6921128080904511, 0.09950216269594728)
piex.score_pipeline(pipeline['id'], 'uu4_SPECT')(0.8897656842904123, 0.037662864373452655)
Optionally, the cross validation configuration can be changed
piex.score_pipeline(pipeline['id'], 'uu4_SPECT', n_splits=3, random_state=43)(0.8869488536155202, 0.019475563687443638)
A Template can also be tested over any dataset by passing its name, the dataset and, optionally, the cross validation specification. You have to make sure to choose template that is relevant for the task/data modality for which you want to use it.
If no hyperparameters are passed, the default ones will be used:
piex.score_template(template_name, 'uu4_SPECT', n_splits=3, random_state=43)(0.8555346666968675, 0.028343173498423108)
You can get the default hyperparameters, and update the hyperparameters by setting values in the dictionary:
With this anyone can tune the templates that we have for different task/data modality types using their own AutoML routine. If you choose to do so, let us know the score you are getting and the pipeline and we will add to our database.
hyperparameters = piex.get_default_hyperparameters(template_name)
hyperparameters['xgboost.XGBClassifier#1']['learning_rate'] = 1
piex.score_template(template_name, 'uu4_SPECT', hyperparameters, n_splits=3, random_state=43)(0.8754554700753094, 0.019151608028236813)