MLBlocks is a simple framework for composing end-to-end tunable machine learning pipelines

Pipelines and primitives for machine learning and data science.

Free software: MIT license
Documentation: https://HDI-Project.github.io/mlblocks

Overview

At a high level:

Machine learning primitives are specified using standardized JSONs
User (or an external automated engine) specifies a list of primitives
The library transforms JSON specifications of machine learning primitives (blocks) into MLBlock instances, which expose tunable hyperparameters via MLHyperparams and composes a MLPipeline
The pipeline.fit and pipeline.predict functions then allow user to fit the pipeline to data and predict on a new set of data.

Project Structure

The MLBlocks consists of several modules and folders:

mlblocks.py: Defines the MLBlocks core class of the library.
mlpipeline.py: Defines the MLPipeline class that allows combining multiple MLBlocks.
mlhyperparam.py: Defines the MLHyperparam, an abstraction of an MLBlock tunable hyperparameter.
components: is a submodule that contains a collection of helper functions used to integrate primitives into MLBlocks, as well as some custom primitives.
parsers: defines the parsers: classes that initialize MLBlock instances from JSON primitives.
primitives: folder that contains the collection of JSON primitives.

Primitive JSONS

The primitive JSONs are the main component of our library. The format of said JSON files varies slightly depending on the model source library, but generally random_forest_classifier.json is a good starting example to look at. For neural keras primitives, refer to simple_cnn.json.

Components

The components subpackage provides the code for some auxiliary custom functions that are useful when creating pipelines. Each custom function should also have a corresponding primitive JSON. A useful example is the HOG featurization step for image pipelines, defined in components/functions/image/hog.py and primitives/HOG.json.

Parsers

Parsers provide the logic to create MLBlock instances from JSON primitive specifications. All parsers should extend the MLParser base class, particularly overriding the build_mlblock method. Other quality-of-life helper functions are provided in the MLParser class as well.

Installation

Install with pip

The simplest and recommended way to install MLBlocks is using pip:

pip install mlblocks

Install from sources

You can also clone the repository and install it from sources

git clone git@github.com:HDI-Project/MLBlocks.git
cd MLBlocks
make install

Usage

Initializing a pipeline

With MLBlocks, we can simply initialize a pipeline by passing it the list of MLBlocks that will compose it.

>>> from mlblocks.mlpipeline import MLPipeline
>>> image_pipeline = MLPipeline(['HOG', 'random_forest_classifier'])

Obtaining and updating hyperparameters

Upon initialization, a pipeline has random hyperparameter values. For a particular data science problem, we may want to set or view the values and attributes of particular hyperparameters. For example, we may need to pass in the current hyperparameter values of our pipeline into a third party tuner.

For tunable hyperparameters, we use get_tunable_hyperparams and update_tunable_hyperparams, in which we obtain a list of MLHyperparams and pass in a list of updated MLHyperparams respectively.

>>> tunable_hp = image_pipeline.get_tunable_hyperparams()
>>> print('\n'.join(map(str, tunable_hp)))
Hyperparameter: Name: num_orientations, Step Name: HOG, Type: int, Range: [9, 9], Value: 9
Hyperparameter: Name: num_cell_pixels, Step Name: HOG, Type: int, Range: [8, 8], Value: 8
Hyperparameter: Name: num_cells_block, Step Name: HOG, Type: int, Range: [3, 3], Value: 3
Hyperparameter: Name: criterion, Step Name: rf_classifier, Type: string, Range: ['entropy', 'gini'], Value: entropy
Hyperparameter: Name: max_features, Step Name: rf_classifier, Type: float, Range: [0.1, 1.0], Value: 0.9134616693335704
Hyperparameter: Name: max_depth, Step Name: rf_classifier, Type: int, Range: [2, 10], Value: 2
Hyperparameter: Name: min_samples_split, Step Name: rf_classifier, Type: int, Range: [2, 4], Value: 3
Hyperparameter: Name: min_samples_leaf, Step Name: rf_classifier, Type: int, Range: [1, 3], Value: 3
Hyperparameter: Name: n_estimators, Step Name: rf_classifier, Type: int_cat, Range: [100], Value: 100
Hyperparameter: Name: n_jobs, Step Name: rf_classifier, Type: int_cat, Range: [-1], Value: -1
>>> image_pipeline.update_tunable_hyperparams(tunable_hp)

If we only want to update the value of certain tunable hyperparameters, we can use the set_from_hyperparam_dict method, in which we provide a mapping of (step name, hyperparameter name) tuples to values to update to as input.

>>> hp_dict = {('rf_classifier', 'max_depth'): 9}
>>> image_pipeline.set_from_hyperparam_dict(hp_dict)
>>> updated_hp = image_pipeline.get_tunable_hyperparams()
>>> print('\n'.join(map(str, updated_hp)))
Hyperparameter: Name: num_orientations, Step Name: HOG, Type: int, Range: [9, 9], Value: 9
Hyperparameter: Name: num_cell_pixels, Step Name: HOG, Type: int, Range: [8, 8], Value: 8
Hyperparameter: Name: num_cells_block, Step Name: HOG, Type: int, Range: [3, 3], Value: 3
Hyperparameter: Name: criterion, Step Name: rf_classifier, Type: string, Range: ['entropy', 'gini'], Value: entropy
Hyperparameter: Name: max_features, Step Name: rf_classifier, Type: float, Range: [0.1, 1.0], Value: 0.9134616693335704
Hyperparameter: Name: max_depth, Step Name: rf_classifier, Type: int, Range: [2, 10], Value: 9
Hyperparameter: Name: min_samples_split, Step Name: rf_classifier, Type: int, Range: [2, 4], Value: 3
Hyperparameter: Name: min_samples_leaf, Step Name: rf_classifier, Type: int, Range: [1, 3], Value: 3
Hyperparameter: Name: n_estimators, Step Name: rf_classifier, Type: int_cat, Range: [100], Value: 100
Hyperparameter: Name: n_jobs, Step Name: rf_classifier, Type: int_cat, Range: [-1], Value: -1

Sometimes, we might want to obtain and update fixed hyperparameters. We can use analogous get_fixed_hyperparams and update_fixed_hyperparams. Similarly to the set_from_hyperparam_dict, the outputs of and input to these functions respectively are mappings of (step name, hyperparameter name) tuples to hyperparameter values.

>>> hp_to_update = {('rf_classifier', 'bootstrap'): True}
>>> image_pipeline.update_fixed_hyperparams(hp_to_update)
>>> image_pipeline.get_fixed_hyperparams()
{('rf_classifier', 'bootstrap'): True}

Making predictions

Once we have set the appropriate hyperparameters for our pipeline, we can make predictions on a dataset.

To do this, we first call the fit method if necessary. This takes in training data and labels as well as any other parameters each individual step may use during fitting. These are specified as mappings from (step name, fit parameter name) tuples to fit parameter values.

>>> from sklearn.datasets import fetch_mldata
>>> from sklearn.model_selection import train_test_split
>>> mnist = fetch_mldata('MNIST original')
>>> X, X_test, y, y_test = train_test_split(mnist.data, mnist.target, train_size=1000, test_size=300)
>>> optional_fit_params = {('rf_classifier', 'sample_weight'): None}
>>> image_pipeline.fit(X, y, optional_fit_params)

Once we have fit our model to our data, we can simply make predictions. From these predictions, we can do useful things, such as obtain an accuracy score.

>>> from sklearn.metrics import accuracy_score
>>> predicted_y_val = image_pipeline.predict(X_test)
>>> score = accuracy_score(y_test, predicted_y_val)
>>> print(score)
0.85

History

In its first iteration in 2015, MLBlocks was designed for only multi table, multi entity temporal data. A good reference to see our design rationale at that time is Bryan Collazo’s thesis:

Machine learning blocks. Bryan Collazo. Masters thesis, MIT EECS, 2015.

With recent availability of a multitude of libraries and tools, we decided it was time to integrate them and expand the library to address other data types: images, text, graph, time series and integrate with deep learning libraries.

essobi/MLBlocks