Code Machine Learning Pipelines - The Right Way.
)
Neuraxle is a Machine Learning (ML) library for building neat pipelines, providing the right abstractions to both ease research, development, and deployment of your ML applications.
Simply do:
pip install neuraxleOne of Neuraxle's most important goals is to make powerful machine learning pipelines easy to build and deploy. Using Neuraxle should be light, painless and obvious, yet without sacrificing powerfulness, performance, nor possibilities.
For example, you can build a pipeline composed of multiple steps as such:
p = Pipeline([
# A Pipeline is composed of multiple chained steps. Steps
# can alter the data before passing it to the next steps.
AddFeatures([
# Add (concatenate) features in parallel, that are
# themselves derived of the existing features:
PCA(n_components=2),
FastICA(n_components=2),
]),
RidgeModelStacking([
# Here is an ensemble of 4 models or feature extractors,
# That are themselves then fed to a ridge regression which
# will act as a judge to finalize the prediction.
LinearRegression(),
LogisticRegression(),
GradientBoostingRegressor(n_estimators=500),
GradientBoostingRegressor(max_depth=5),
KMeans(),
])
])
# Note: here all the steps were imported from scikit-learn,
# but the goal is that you can also define your own as needed.
# Also note that a pipeline is a step itself: you can nest them.
# The pipeline will learn on the data and acquire state.
p = p.fit(X_train, y_train)
# Once it learned, the pipeline can process new and
# unseen data for making predictions.
y_test_predicted = p.predict(X_test)Visit the examples to get more a feeling of how it works, and inspiration.
Here is how to use deep learning algorithms within a Neuraxle Pipeline.
Defining a Deep Learning pipeline is more complex. It needs a composition of many steps to:
- Loop on data for many epochs, but just during training.
- Shuffle the data, just during training.
- Use minibatches to process the data, which avoids to blow RAM. Your steps will fit incrementally.
- Process data that is 2D, 3D, 4D, or even 5D or ND, with transformers made for 2D data slices.
- Actually use your Deep Learning algorithm within your pipeline for it to learn and predict.
Below, we define a pipeline for time series classification using a LSTM RNN. It includes data preprocessing steps as well as the data flow management. Time Series data is 3D.
deep_learning_seq_classif_pipeline = EpochRepeater(Pipeline([
# X data shape: (batch, different_lengths, n_feature_columns)
# y data shape: (batch, different_lengths)
# Split X and Y into windows using
# an InputAndOutputTransformerMixin
# abstraction to transform y too:
SliceTimeSeries(window_size=128, last_label_as_seq_label=True),
# X data shape: (more_than_batch, 128, n_feature_columns)
# y data shape: (more_than_batch, 128)
TrainOnlyWrapper(DataShuffler(seed=42)),
MiniBatchSequentialPipeline([
# X data shape: (batch_size, 128, n_feature_columns)
# y data shape: (batch_size, 128)
# Loop on 2D slices of the batch's 3D time series
# data cube to apply 2D transformers:
ForEachDataInput(Pipeline([
# X data shape: (128, n_feature_columns)
# y data shape: (128)
# This step will load the lazy-loadable data
# into a brick:
ToNumpy(np_dtype=np.float32),
# Fill nan and inf values with 0:
DefaultValuesFiller(0.0),
# Transform the columns (that is the innermost
# axis/dim of data named `n_feature_columns`):
ColumnTransformer([
# Columns 0, 1, 2, 3 and 4 needs to be
# normalized by mean and variance (std):
(range(0, 5), MeanVarianceNormalizer()),
# Column 5 needs to have it's `log plus 1`
# value taken before normalization.
(5, Pipeline([
Log1P(),
MeanVarianceNormalizer()
]))
# Note that omited columns are discarded.
# Also, multiple transformers on a column will
# concatenate the results.
]),
# Transform the labels' indices to one-hot vectors.
OutputTransformerWrapper(
OneHotEncoder(nb_columns=6, name='labels'))
# X data shape: (128, n_feature_columns)
# y data shape: (128, 6)
])),
# X data shape: (batch_size, 128, n_feature_columns)
# y data shape: (batch_size, 128, 6)
# Classification with a deep neural network,
# using the Neuraxle-TensorFlow and/or
# Neuraxle-PyTorch extensions:
ClassificationLSTM(n_stacked=2, n_residual=3),
# X data shape: (batch_size, 128, 6)
# y data shape: (batch_size, 128, 6)
], batch_size=32),
# X data shape: (batch_size, 128, 6)
]), epochs=200, fit_only=True)Here we train and evaluate with a train-validation split. Note that automatic hyperparameter tuning would require only a few more lines of code: see our hyperparameter tuning example.
# Wrap the pipeline by a validation strategy,
# this could have been Cross Validation as well:
training_pipeline = ValidationSplitWrapper(
deep_learning_seq_classif_pipeline,
val_size=0.1,
scoring_function=sklearn.metrics.accuracy_score
)
# Fitting and evaluating the pipeline.
# X_train data shape: (batch, different_lengths, n_feature_columns)
# y_train data shape: (batch, different_lengths)
training_pipeline.fit(X_train, y_train)
# Note that X_train and y_train can be lazy loaders.
print('Train accuracy: {}'.format(
training_pipeline.scores_train_mean))
print('Validation accuracy: {}'.format(
training_pipeline.scores_validation_mean))
# Recover the pipeline in test mode:
production_pipeline = training_pipeline.get_step()
production_pipeline.set_train(False)Deploying your deep learning app to a JSON REST API. Refer to Flask's deployment documentation for more info on deployment servers and security.
# Will now serve the pipeline to a REST API as an example:
# Note that having saved the pipeline to disk
# (for reloading this in another file) would be easy, too, using savers.
app = FlaskRestApiWrapper(
json_decoder=YourCustomJSONDecoderFor2DArray(),
wrapped=production_pipeline,
json_encoder=YourCustomJSONEncoderOfOutputs()
).get_app()
app.run(debug=False, port=5000)This could be ran from another distant computer to call your app:
p = APICaller(
json_encoder=YourCustomJSONEncoderOfInputs(),
url="http://127.0.0.1:5000/",
json_decoder=YourCustomJSONDecoderOfOutputs()
)
y_pred = p.predict(X_test)
print(y_pred)Note that we'll soon have better remote proxy design patterns for distant pipelines, and distant parallel processing and distant parallel training.
Most research projects don't ever get to production. However, you want your project to be production-ready and already adaptable (clean) by the time you finish it. You also want things to be simple so that you can get started quickly.
Most existing machine learning pipeline frameworks are either too simple or too complicated for medium-scale projects. Neuraxle is balanced for medium-scale projects, providing simple, yet powerful abstractions that are ready to be used.
Neuraxle is built as a framework that enables you to define your own pipeline steps.
This means that you can use scikit-learn, Keras, TensorFlow, PyTorch, Hyperopt, Ray and/or any other machine learning library you like within and throughout your Neuraxle pipelines.
Neuraxle offer multiple parallel processing features. One magical thing that we did are Savers. Savers allow you to define how a step can be serialized. This way, it's possible to avoid Python's parallel processing limitations and pitfalls.
Let's suppose that your pipeline has a step that imports code from another library and that this code isn't serializable (e.g.: some code written in C++ and interacting with the GPUs or anything funky). To make this step serializable, just define a saver which will tell the step how to dump itself to disk and reload itself. This will allow the step to be sent to a remote computer or to be threadable by reloading the save. The save can be dumped to a RAM disk for more performance and avoid truly writing to disks.
Neuraxle is compatible with most other ML and DL libraries. We're currently already writing savers for PyTorch and TensorFlow in the Neuraxle-PyTorch and Neuraxle-TensorFlow extensions of this project.
Although Neuraxle is not limited to just time series processing projects, it's especially good for those projects, as one of the goals of Neuraxle is to provides a few abstractions that are useful for time series projects, as Time Series data is often 3D or even ND.
With the various abstractions that Neuraxle provides, it's easy to get started building a time-series processing project. Neuraxle is also the backbone of the Neuraxio Time Series project, which is a premium software package built on top of Neuraxle for business boost their time series machine learning projects by providing out-of-the-box specialized pipeline steps. Some of those specialized steps are featured in the Deep Learning Pipelines section above.
Note: the Neuraxio Time Series project is different from the Neuraxle project, those are separate projects. Neuraxio is commited to build open-source software, and defines itself as an open-source company. Learn more on Neuraxle's license. The Neuraxle library is free and will always stay free, while Neuraxio Time Series is a premium add-on to Neuraxle.
One of the core goal of this framework is to enable easy automatic machine learning, and also meta-learning. It should be easy to train a meta-optimizer on many different tasks: the optimizer is a model itself that maps features of datasets and features of the hyperparameter space to a guessed performance score to predict the best hyperparameters. Hyperparameter spaces are easily defined with a range, and are only coupled to their respective pipeline steps, rather than being coupled to the whole pipeline, which enable class reuse and more modularity.
Everything that works in sklearn is also useable in Neuraxle. Neuraxle
is built in a way that does not replace what already exists. Therefore,
Neuraxle adds more power to scikit-lean by providing neat abstractions,
and neuraxle is even retrocompatible with sklean if it ever needed to be
included in an already-existing sklearn pipeline (you can do that by
using .tosklearn() on your Neuraxle pipeline). We believe that
Neuraxle helps scikit-learn, and also scikit-learn will help Neuraxle.
Neuraxle is best used with scikit-learn.
Also, the top core developers of scikit-learn, Andreas C. Müller, gave a talk in which he lists the elements that are yet to be done in scikit-learn. He refers to building bigger pipelines with automatic machine learning, meta learning, improving the abstractions of the search spaces, and he also points out that it would be possible do achieve that in another library which could reuse scikit-learn. Neuraxle is here to solve those problems that are actually shared by the open-source community in general. Let's move forward with Neuraxle: join Neuraxle's community.
Apache Beam is a big, multi-language project and hence is complicated. Neuraxle is pythonic and user-friendly: it's easy to get started.
Also, it seems that Apache Beam has GPL and MPL dependencies, which means Apache Beam might itself be copyleft (?). Neuraxle doesn't have such copyleft dependencies.
spaCy has copyleft dependencies or may download copyleft content, and it is built only for Natural Language Processing (NLP) projects. Neuraxle is open to any kind of machine learning projects and isn't an NLP-first project.
Kubeflow is cloud-first, using Kubernetes and is more oriented towards devops. Neuraxle isn't built as a cloud-first solution and isn't tied to Kubernetes. Neuraxle instead offers many parallel processing features, such as the ability to be scaled on many cores of a computer, and even on a computer cluster (e.g.: in the cloud using any cloud provider) with joblib, using dask's distributed library as a joblib backend. A Neuraxle project is best deployed as a microservice within your software environment, and you can fully control and customize how you deploy your project (e.g.: coding yourself a pipeline step that does json conversion to accept http requests).
Join our Slack workspace and our Gitter! We <3 collaborators. You can also subscribe to our mailing list where we will post updates and news.
For technical questions, we recommend posting them on
StackOverflow
first with neuraxle in the tags (amongst probably python and
machine-learning), and then opening an
issue to link to your
Stack Overflow question.
For suggestions, comments, and issues, don't hesitate to open an issue.
For contributors, we recommend using the PyCharm code editor and to let it manage the virtual environment, with the default code auto-formatter, and using pytest as a test runner. To contribute, first fork the project, then do your changes, and then open a pull request in the main repository. Please make your pull request(s) editable, such as for us to add you to the list of contributors if you didn't add the entry, for example. Ensure that all tests run before opening a pull request. You'll also agree that your contributions will be licensed under the Apache 2.0 License, which is required for everyone to be able to use your open-source contributions.
Neuraxle is licensed under the Apache License, Version 2.0.
You may cite our extended abstract that was presented at the Montreal Artificial Intelligence Symposium (MAIS) 2019. Here is the bibtex code to cite:
@misc{neuraxle,
author = {Chevalier, Guillaume and Brillant, Alexandre and Hamel, Eric},
year = {2019},
month = {09},
pages = {},
title = {Neuraxle - A Python Framework for Neat Machine Learning Pipelines},
doi = {10.13140/RG.2.2.33135.59043}
}Thanks to everyone who contributed to the project:
- Guillaume Chevalier: https://github.com/guillaume-chevalier
- Alexandre Brillant: https://github.com/alexbrillant
- Éric Hamel: https://github.com/Eric2Hamel
- Jérôme Blanchet: https://github.com/JeromeBlanchet
- Michaël Lévesque-Dion: https://github.com/mlevesquedion
- Philippe Racicot: https://github.com/Vaunorage
We thank these organisations for generously supporting the project:
- Neuraxio Inc.: https://github.com/Neuraxio
- Umanéo Technologies Inc.: https://www.umaneo.com/
- Solution Nexam Inc.: https://nexam.io/
- La Cité, LP: https://www.lacitelp.com/
- Kimoby: https://www.kimoby.com/
