pure-predict speeds up and slims down machine learning prediction applications. It is
a foundational tool for serverless inference or small batch prediction with popular machine
learning frameworks like scikit-learn and fasttext.
It implements the predict methods of these frameworks in pure Python.
Primary Use Cases
The primary use case for pure-predict is the following scenario:
- A model is trained in an environment without strong container footprint constraints. Perhaps a long running “offline” job on one or many machines where installing a number of python packages from PyPI is not at all problematic.
- At prediction time the model needs to be served behind an API. Typical access patterns are to request a prediction for one “record” (one “row” in a numpy array or one string of text to classify) per request or a mini-batch of records per request.
- Preferred infrastructure for the prediction service is either serverless (AWS Lambda) or a container service where the memory footprint of the container is constrained.
- The fitted model object’s artifacts needed for prediction (coefficients, weights, vocabulary, decision tree artifacts, etc.) are relatively small (10s to 100s of MBs).
In this scenario, a container service with a large dependency footprint can be overkill for a microservice, particularly if the access patterns favor the pricing model of a serverless application. Additionally, for smaller models and single record predictions per request, the numpy and scipy functionality in the prediction methods of popular machine learning frameworks work against the application in terms of latency, underperforming pure python in some cases.
It is a Python package for machine learning prediction distributed under
the Apache 2.0 software license.
It contains multiple subpackages which mirror their open source
counterpart (scikit-learn, fasttext, etc.). Each subpackage has utilities to
convert a fitted machine learning model into a custom object containing prediction methods
that mirror their native counterparts, but converted to pure python. Additionally, all
relevant model artifacts needed for prediction are converted to pure python.
A pure-predict model object can then be pickled and later
unpickled without any 3rd party dependencies other than pure-predict.
This eliminates the need to have large dependency packages installed in order to
make predictions with fitted machine learning models using popular open source packages for
training models. These dependencies (numpy, scipy, scikit-learn, fasttext, etc.)
are large in size and not always necessary to make fast and accurate
Additionally, they rely on C extensions that may not be ideal for serverless applications with a python runtime.
Quick Start Example
In a python enviornment with scikit-learn and its dependencies installed:
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_iris
from pure_sklearn.map import convert_estimator
# fit sklearn estimator
X, y = load_iris(return_X_y=True)
clf = RandomForestClassifier()
# convert to pure python estimator
clf_pure_predict = convert_estimator(clf)
with open("model.pkl", "wb") as f:
# make prediction with sklearn estimator
y_pred = clf.predict([[0.25, 2.0, 8.3, 1.0]])
In a python enviornment with only pure-predict installed:
# load pickled model
with open("model.pkl", "rb") as f:
clf = pickle.load(f)
# make prediction with pure-predict object
y_pred = clf.predict([[0.25, 2.0, 8.3, 1.0]])
Prediction in pure python for a subset of scikit-learn estimators and transformers.
- linear models - supports the majority of linear models for classification
- trees - decision trees, random forests, gradient boosting and xgboost
- naive bayes - a number of popular naive bayes classifiers
- svm - linear SVC
- preprocessing - normalization and onehot/ordinal encoders
- impute - simple imputation
- feature extraction - text (tfidf, count vectorizer, hashing vectorizer) and dictionary vectorization
- pipeline - pipelines and feature unions
Sparse data - supports a custom pure python sparse data object - sparse data is handled as would be expected by the relevent transformers and estimators
Prediction in pure python for fasttext.
- supervised - predicts labels for supervised models; no support for quantized models (blocked by this issue)
- unsupervised - lookup of word or sentence embeddings given input text
- pure_sklearn has been tested with scikit-learn versions >= 0.20 – certain functionality may work with lower versions but are not guaranteed. Some functionality is explicitly not supported for certain scikit-learn versions and exceptions will be raised as appropriate.
- xgboost requires version >= 0.82 for support with pure_sklearn.
- pure-predict is not supported with Python 2.
The easiest way to install pure-predict is with pip:
pip install --upgrade pure-predict
You can also download the source code:
git clone https://github.com/Ibotta/pure-predict.git
With pytest installed, you can run tests locally:
The package contains examples
on how to use pure-predict in practice.
Calls for Contributors
Contributing to pure-predict is welcomed by any contributors. Specific calls for contribution are as follows:
- Examples, tests and documentation – particularly more detailed examples with performance testing of various estimators under various constraints.
- Adding more pure_sklearn estimators. The scikit-learn package is extensive and only partially covered by pure_sklearn. Regression tasks in particular missing from pure_sklearn. Clustering, dimensionality reduction, nearest neighbors, feature selection, non-linear SVM, and more are also omitted and would be good candidates for extending pure_sklearn.
- General efficiency. There is likely low hanging fruit for improving the efficiency of the numpy and scipy functionality that has been ported to pure-predict.
- Threading could be considered to improve performance – particularly for making predictions with multiple records.
- A public AWS lambda layer containing pure-predict.
The project was started at Ibotta
Inc. on the machine learning
team and open sourced in 2020. It is currently maintained by the machine
learning team at Ibotta.