Pipelines and primitives for machine learning and data science.
Project description
An open source project from Data to AI Lab at MIT.
Pipelines and Primitives for Machine Learning and Data Science.
- Free software: MIT license
- Documentation: https://HDI-Project.github.io/MLBlocks
- Homepage: https://github.com/HDI-Project/MLBlocks
MLBlocks
MLBlocks is a simple framework for composing end-to-end tunable Machine Learning Pipelines by seamlessly combining tools from any python library with a simple, common and uniform interface.
Features include:
- Build Machine Learning Pipelines combining any Machine Learning Library in Python.
- Access a repository with hundreds of primitives and pipelines ready to be used with little to no python code to write, carefully curated by Machine Learning and Domain experts.
- Extract machine-readable information about which hyperparameters can be tuned and within which ranges, allowing automated integration with Hyperparameter Optimization tools like BTB.
- Complex multi-branch pipelines and DAG configurations, with unlimited number of inputs and outputs per primitive.
- Easy save and load Pipelines using JSON Annotations.
Install
Requirements
MLBlocks has been developed and tested on Python 3.5 and 3.6
Also, although it is not strictly required, the usage of a virtualenv is highly recommended in order to avoid interfering with other software installed in the system where MLBlocks is run.
These are the minimum commands needed to create a virtualenv using python3.6 for MLBlocks:
pip install virtualenv
virtualenv -p $(which python3.6) mlblocks-venv
Afterwards, you have to execute this command to have the virtualenv activated:
source mlblocks-venv/bin/activate
Remember about executing it every time you start a new console to work on MLBlocks!
Install with pip
After creating the virtualenv and activating it, we recommend using pip in order to install MLBlocks:
pip install mlblocks
This will pull and install the latest stable release from PyPi.
Install from source
Alternatively, with your virtualenv activated, you can clone the repository and install it from
source by running make install
on the stable
branch:
git clone git@github.com:HDI-Project/MLBlocks.git
cd MLBlocks
git checkout stable
make install
Install for Development
If you want to contribute to the project, a few more steps are required to make the project ready for development.
First, please head to the GitHub page of the project and make a fork of the project under you own username by clicking on the fork button on the upper right corner of the page.
Afterwards, clone your fork and create a branch from master with a descriptive name that includes the number of the issue that you are going to work on:
git clone git@github.com:{your username}/MLBlocks.git
cd MLBlocks
git branch issue-xx-cool-new-feature master
git checkout issue-xx-cool-new-feature
Finally, install the project with the following command, which will install some additional dependencies for code linting and testing.
make install-develop
Make sure to use them regularly while developing by running the commands make lint
and make test
.
MLPrimitives
In order to be usable, MLBlocks requires a compatible primitives library.
The official library, required in order to follow the following MLBlocks tutorial, is MLPrimitives, which you can install with this command:
pip install mlprimitives
Quickstart
Below there is a short example about how to use MLBlocks to create a simple pipeline, fit it using demo data and use it to make predictions.
Please make sure to also having installed MLPrimitives before following it.
For advance usage and more detailed explanation about each component, please have a look at the documentation
Creating a pipeline
With MLBlocks, creating a pipeline is as simple as specifying a list of primitives and passing
them to the MLPipeline
class.
>>> from mlblocks import MLPipeline
... primitives = [
... 'cv2.GaussianBlur',
... 'skimage.feature.hog',
... 'sklearn.ensemble.RandomForestClassifier'
... ]
>>> pipeline = MLPipeline(primitives)
Optionally, specific initialization arguments can be also set by specifying them in a dictionary:
>>> init_params = {
... 'skimage.feature.hog': {
... 'multichannel': True,
... 'visualize': False
... },
... 'sklearn.ensemble.RandomForestClassifier': {
... 'n_estimators': 100,
... }
... }
>>> pipeline = MLPipeline(primitives, init_params=init_params)
If you can see which hyperparameters a particular pipeline is using, you can do so by calling
its get_hyperparameters
method:
>>> import json
>>> hyperparameters = pipeline.get_hyperparameters()
>>> print(json.dumps(hyperparameters, indent=4))
{
"cv2.GaussianBlur#1": {
"ksize_width": 3,
"ksize_height": 3,
"sigma_x": 0,
"sigma_y": 0
},
"skimage.feature.hog#1": {
"multichannel": true,
"visualize": false,
"orientations": 9,
"pixels_per_cell_x": 8,
"pixels_per_cell_y": 8,
"cells_per_block_x": 3,
"cells_per_block_y": 3,
"block_norm": null
},
"sklearn.ensemble.RandomForestClassifier#1": {
"n_jobs": -1,
"n_estimators": 100,
"criterion": "entropy",
"max_features": null,
"max_depth": 10,
"min_samples_split": 0.1,
"min_samples_leaf": 0.1,
"class_weight": null
}
}
Making predictions
Once we have created the pipeline with the desired hyperparameters we can fit it and then use it to make predictions on new data.
To do this, we first call the fit
method passing the training data and the corresponding labels.
In this case in particular, we will be loading the handwritten digit classification dataset
from USPS using the mlblocks.datasets.load_usps
method, which returns a dataset object
ready to be played with.
>>> from mlblocks.datasets import load_usps
>>> dataset = load_usps()
>>> X_train, X_test, y_train, y_test = dataset.get_splits(1)
>>> pipeline.fit(X_train, y_train)
Once we have fitted our model to our data, we can call the predict
method passing new data
to obtain predictions from the pipeline.
>>> predictions = pipeline.predict(X_test)
>>> predictions
array([3, 2, 1, ..., 1, 1, 2])
What's Next?
If you want to learn more about how to tune the pipeline hyperparameters, save and load the pipelines using JSON annotations or build complex multi-branched pipelines, please check our documentation.
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.
Changelog
0.3.3 - 2019-09-09
- Improved intermediate outputs management - Issue #105 by @csala
0.3.2 - 2019-08-12
- Allow passing fit and produce arguments as
init_params
- Issue #96 by @csala - Support optional fit and produce args and arg defaults - Issue #95 by @csala
- Isolate primitives from their hyperparameters dictionary - Issue #94 by @csala
- Add functions to explore the available primitives and pipelines - Issue #90 by @csala
- Add primitive caching - Issue #22 by @csala
0.3.1 - Pipelines Discovery
- Support flat hyperparameter dictionaries - Issue #92 by @csala
- Load pipelines by name and register them as
entry_points
- Issue #88 by @csala - Implement partial re-fit -Issue #61 by @csala
- Move argument parsing to MLBlock - Issue #86 by @csala
- Allow getting intermediate outputs - Issue #58 by @csala
0.3.0 - New Primitives Discovery
- New primitives discovery system based on
entry_points
. - Conditional Hyperparameters filtering in MLBlock initialization.
- Improved logging and exception reporting.
0.2.4 - New Datasets and Unit Tests
- Add a new multi-table dataset.
- Add Unit Tests up to 50% coverage.
- Improve documentation.
- Fix minor bug in newsgroups dataset.
0.2.3 - Demo Datasets
- Add new methods to Dataset class.
- Add documentation for the datasets module.
0.2.2 - MLPipeline Load/Save
- Implement save and load methods for MLPipelines
- Add more datasets
0.2.1 - New Documentation
- Add mlblocks.datasets module with demo data download functions.
- Extensive documentation, including multiple pipeline examples.
0.2.0 - New MLBlocks API
A new MLBlocks API and Primitive format.
This is a summary of the changes:
- Primitives JSONs and Python code has been moved to a different repository, called MLPrimitives
- Optional usage of multiple JSON primitive folders.
- JSON format has been changed to allow more flexibility and features:
- input and output arguments, as well as argument types, can be specified for each method
- both classes and function as primitives are supported
- multitype and conditional hyperparameters fully supported
- data modalities and primitive classifiers introduced
- metadata such as documentation, description and author fields added
- Parsers are removed, and now the MLBlock class is responsible for loading and reading the JSON primitive.
- Multiple blocks of the same primitive are supported within the same pipeline.
- Arbitrary inputs and outputs for both pipelines and blocks are allowed.
- Shared variables during pipeline execution, usable by multiple blocks.
0.1.9 - Bugfix Release
- Disable some NetworkX functions for incompatibilities with some types of graphs.
0.1.8 - New primitives and some improvements
- Improve the NetworkX primitives.
- Add String Vectorization and Datetime Featurization primitives.
- Refactor some Keras primitives to work with single dimension
y
arrays and be compatible withpickle
. - Add XGBClassifier and XGBRegressor primitives.
- Add some
keras.applications
pretrained networks as preprocessing primitives. - Add helper class to allow function primitives.
0.1.7 - Nested hyperparams dicts
- Support passing hyperparams as nested dicts.
0.1.6 - Text and Graph Pipelines
- Add LSTM classifier and regressor primitives.
- Add OneHotEncoder and MultiLabelEncoder primitives.
- Add several NetworkX graph featurization primitives.
- Add
community.best_partition
primitive.
0.1.5 - Collaborative Filtering Pipelines
- Add LightFM primitive.
0.1.4 - Image pipelines improved
- Allow passing
init_params
onMLPipeline
creation. - Fix bug with MLHyperparam types and Keras.
- Rename
produce_params
aspredict_params
. - Add SingleCNN Classifier and Regressor primitives.
- Simplify and improve Trivial Predictor
0.1.3 - Multi Table pipelines improved
- Improve RandomForest primitive ranges
- Improve DFS primitive
- Add Tree Based Feature Selection primitives
- Fix bugs in TrivialPredictor
- Improved documentation
0.1.2 - Bugfix release
- Fix bug in TrivialMedianPredictor
- Fix bug in OneHotLabelEncoder
0.1.1 - Single Table pipelines improved
- New project structure and primitives for integration into MIT-TA2.
- MIT-TA2 default pipelines and single table pipelines fully working.
0.1.0
- First release on PyPI.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for mlblocks-0.3.3-py2.py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | e35b7312f0d548e69ca2a40c8c0312eaf4e7d900f0a6bfcd3264ccf3f582014a |
|
MD5 | 802f9d7dd61a3e4078944f9b742bef5e |
|
BLAKE2b-256 | 3750d4beca217ea3c52b22a3ac9ad51a8a2366599d331ac0159bbcf1f735d7b3 |