nimbusml
is a Python module that provides Python bindings for ML.NET.
ML.NET was originally developed in Microsoft Research and is used across many product groups in Microsoft like Windows, Bing, PowerPoint, Excel, and others. nimbusml
was built to enable data science teams that are more familiar with Python to take advantage of ML.NET's functionality and performance.
nimbusml
enables training ML.NET pipelines or integrating ML.NET components directly into scikit-learn pipelines. It adheres to existing scikit-learn
conventions, allowing simple interoperability between nimbusml
and scikit-learn
components, while adding a suite of fast, highly optimized, and scalable algorithms, transforms, and components written in C++ and C#.
See examples below showing interoperability with scikit-learn
. A more detailed example in the documentation shows how to use a nimbusml
component in a scikit-learn
pipeline, and create a pipeline using only nimbusml
components.
nimbusml
supports numpy.ndarray
, scipy.sparse_cst
, and pandas.DataFrame
as inputs. In addition, nimbusml
also supports streaming from files without loading the dataset into memory with FileDataStream
, which allows training on data significantly exceeding memory.
Documentation can be found here and additional notebook samples can be found here.
nimbusml
runs on Windows, Linux, and macOS.
nimbusml
requires Python 2.7, 3.5, 3.6, 3.7 64 bit version only.
Install nimbusml
using pip
with:
pip install nimbusml
nimbusml
has been reported to work on Windows 10, MacOS 10.13, Ubuntu 14.04, Ubuntu 16.04, Ubuntu 18.04, CentOS 7, and RHEL 7.
Here is an example of how to train a model to predict sentiment from text samples (based on this ML.NET example). The full code for this example is here.
from nimbusml import Pipeline, FileDataStream
from nimbusml.datasets import get_dataset
from nimbusml.ensemble import FastTreesBinaryClassifier
from nimbusml.feature_extraction.text import NGramFeaturizer
train_file = get_dataset('gen_twittertrain').as_filepath()
test_file = get_dataset('gen_twittertest').as_filepath()
train_data = FileDataStream.read_csv(train_file, sep='\t')
test_data = FileDataStream.read_csv(test_file, sep='\t')
pipeline = Pipeline([ # nimbusml pipeline
NGramFeaturizer(columns={'Features': ['Text']}),
FastTreesBinaryClassifier(feature=['Features'], label='Label')
])
# fit and predict
pipeline.fit(train_data)
results = pipeline.predict(test_data)
Instead of creating an nimbusml
pipeline, you can also integrate components into scikit-learn pipelines:
from sklearn.pipeline import Pipeline
from nimbusml.datasets import get_dataset
from nimbusml.ensemble import FastTreesBinaryClassifier
from sklearn.feature_extraction.text import TfidfVectorizer
import pandas as pd
train_file = get_dataset('gen_twittertrain').as_filepath()
test_file = get_dataset('gen_twittertest').as_filepath()
train_data = pd.read_csv(train_file, sep='\t')
test_data = pd.read_csv(test_file, sep='\t')
pipeline = Pipeline([ # sklearn pipeline
('tfidf', TfidfVectorizer()), # sklearn transform
('clf', FastTreesBinaryClassifier()) # nimbusml learner
])
# fit and predict
pipeline.fit(train_data["Text"], train_data["Label"])
results = pipeline.predict(test_data["Text"])
Many additional examples and tutorials can be found in the documentation.
To build nimbusml
from source please visit our developer guide.
The contributions guide can be found here.
If you have an idea for a new feature or encounter a problem, please open an issue in this repository or ask your question on Stack Overflow.
NimbusML is licensed under the MIT license.