Do you want to work on exciting technologies in NYC? Be part of something great.

Python is one of the best programming languages out there, with an extensive coverage in scientific computing: computer vision, artificial intelligence, mathematics, astronomy to name a few. Unsurprisingly, this holds true for machine learning as well.

Of course, it has some disadvantages too; one of which is that the tools and libraries for Python are scattered. If you are a unix-minded person, this works quite conveniently as every tool does one thing and does it well. However, this also requires you to know different libraries and tools, including their advantages and disadvantages, to be able to make a sound decision for the systems that you are building. Tools by themselves do not make a system or product better, but with the right tools we can work much more efficiently and be more productive. Therefore, knowing the right tools for your work domain is crucially important.

This post aims to list and describe the most useful machine learning tools and libraries that are available for Python. To make this list, we did not require the library to be written in Python; it was sufficient for it to have a Python interface. We also have a small section on Deep Learning at the end as it has received a fair amount of attention recently.

We do not aim to list all the machine learning libraries available in Python (the Python package index returns 139 results for “machine learning”) but rather the ones that we found useful and well-maintained to the best of our knowledge. Moreover, although some of modules could be used for various machine learning tasks, we included libraries whose main focus is machine learning. For example, although Scipy has some clustering algorithms, the main focus of this module is not machine learning but rather in being a comprehensive set of tools for scientific computing. Therefore, we excluded libraries like Scipy from our list (though we use it too!).

Another thing worth mentioning is that we also evaluated the library based on how it integrates with other scientific computing libraries because machine learning (either supervised or unsupervised) is part of a data processing system. If the library that you are using does not fit with your rest of data processing system, then you may find yourself spending a tremendous amount of time to creating intermediate layers between different libraries. It is important to have a great library in your toolset but it is also important for that library to integrate well with other libraries.

If you are great in another language but want to use Python packages, we also briefly go into how you could integrate with Python to use the libraries listed in the post.


Scikit Learn is our machine learning tool of choice at CB Insights. We use it for classification, feature selection, feature extraction and clustering. What we like most about it is that it has a consistent API which is easy to use while also providing a lot of evaluation, diagnostic and cross-validation methods out of the box (sound familiar? Python has batteries-included approach as well). The icing on the cake is that it uses Scipy data structures under the hood and fits quite well with the rest of scientific computing in Python with Scipy, Numpy, Pandas and Matplotlib packages. Therefore, if you want to visualize the performance of your classifiers (say, using a precision-recall graph or Receiver Operating Characteristics (ROC) curve) those could be quickly visualized with help of Matplotlib. Considering how much time is spent on cleaning and structuring the data, this makes it very convenient to use the library as it tightly integrates to other scientific computing packages.

Moreover, it has also limited Natural Language Processing feature extraction capabilities as well such as bag of words, tfidf, preprocessing (stop-words, custom preprocessing, analyzer). Moreover, if you want to quickly perform different benchmarks on toy datasets, it has a datasets module which provides common and useful datasets. You could also build toy datasets from these datasets for your own purposes to see if your model performs well before applying the model to the real-world dataset. For parameter optimization and tuning, it also provides grid search and random search. These features could not be accomplished if it did not have great community support or if it was not well-maintained. We look forward to its first stable release.


Statsmodels is another great library which focuses on statistical models and is used mainly for predictive and exploratory analysis. If you want to fit linear models, do statistical analysis, maybe a bit of predictive modeling, then Statsmodels is a great fit. The statistical tests it provides are quite comprehensive and cover validation tasks for most of the cases. If you are R or S user, it also accepts R syntax for some of its statistical models. It also accepts Numpy arrays as well as Pandas data-frames for its models making creating intermediate data structures a thing of the past!


PyMC is the tool of choice for Bayesians. It includes Bayesian models, statistical distributions and diagnostic tools for the convergence of models. It includes some hierarchical models as well. If you want to do Bayesian Analysis, you should check it out.


Shogun is a machine learning toolbox with a focus on Support Vector Machines (SVM) that is written in C++. It is actively developed and maintained, provides a Python interface and the Python interface is mostly documented well. However, we’ve found its API hard to use compared to Scikit-learn. Also, it does not provide many diagnostics or evaluation algorithms out of the box. However, its speed is a great advantage.


Gensim is defined as “topic modeling for humans”. As its homepage describes, its main focus is Latent Dirichlet Allocation (LDA) and its variants. Different from other packages, it has support for Natural Language Processing which makes it easier to combine NLP pipeline with other machine learning algorithms. If your domain is in NLP and you want to do clustering and basic classification, you may want to check it out. Recently, they introduced Recurrent Neural Network based text representation called word2vec from Google to their API as well. This library is written purely in Python.


Orange is the only library that has a Graphical User Interface (GUI) among the libraries listed in this post. It is also quite comprehensive in terms of classification, clustering and feature selection methods and has some cross-validation methods. It is better than Scikit-learn in some aspects (classification methods, some preprocessing capabilities) as well, but it does not fit well with the rest of the scientific computing ecosystem (Numpy, Scipy, Matplotlib, Pandas) as nicely as Scikit-learn.

Having a GUI is an important advantage over other libraries however. You could visualize cross-validation results, models and feature selection methods (you need to install Graphviz for some of the capabilities separately). Orange has its own data structures for most of the algorithms so you need to wrap the data into Orange-compatible data structures which makes the learning curve steeper.


PyMVPA is another statistical learning library which is similar to Scikit-learn in terms of its API. It has cross-validation and diagnostic tools as well, but it is not as comprehensive as Scikit-learn.

Deep Learning

Even though deep learning is a subsection Machine Learning, we created a separate section for this field as it has received tremendous attention recently with various acqui-hires by Google and Facebook.


Theano is the most mature of deep learning library. It provides nice data structures (tensors) to represent layers of neural networks and they are efficient in terms of linear algebra similar to Numpy arrays. One caution is that, its API may not be very intuitive, which increases learning curve for users. There are a lot of libraries which build on top of Theano exploiting its data structures. It has support for GPU programming out of the box as well.


There is another library built on top of Theano, called PyLearn2 which brings modularity and configurability to Theano where you could create your neural network through different configuration files so that it would be easier to experiment different parameters. Arguably, it provides more modularity by separating the parameters and properties of neural network to the configuration file.


Decaf is a recently released deep learning library from UC Berkeley which has state of art neural network implementations which are tested on the Imagenet classification competition.


If you want to use excellent Scikit-learn library api in deep learning as well, Nolearn wraps Decaf to make the life easier for you. It is a wrapper on top of Decaf and it is compatible(mostly) with Scikit-learn, which makes Decaf even more awesome.


OverFeat is a recent winner of Dogs vs Cats (kaggle competition) which is written in C++ but it comes with a Python wrapper as well(along with Matlab and Lua). It uses GPU through Torch library so it is quite fast. It also won the detection and localization competition in ImageNet classification. If your main domain is in computer vision, you may want to check it out.


Hebel is another neural network library comes along with GPU support out of the box. You could determine the properties of your neural networks through YAML files(similar to Pylearn2) which provides a nice way to separate your neural network from the code and quickly run your models. Since it has been recently developed, documentation is lacking in terms of depth and breadth. It is also limited in terms of neural network models as it only has one type of neural network model(feed-forward). However, it is written in pure Python and it will be nice library as it has a lot of utility functions such as schedulers and monitors which we did not see any library provides such functionalities.


NeuroLab is another neural network library which has nice api(similar to Matlab’s api if you are familiar) It has different variants of Recurrent Neural Network(RNN) implementation unlike other libraries. If you want to use RNN, this library might be one of the best choice with its simple API.

Integration with other languages

You do not know any Python but great in another language? Do not despair! One of the strengths of Python (among many other) is that it is a perfect glue language that you could use your tool of choice programming language with these libraries through access from Python. Following packages for respective programming languages could be used to combine Python with other programming languages:

Inactive Libraries

These are the libraries that did not release any updates for more than one year, we are listing them because some may find it useful, but it is unlikely that these libraries will be maintained for bug fixes and especially enhancements in the future:

If we are missing one of your favorite packages in Python for machine learning, feel free to let us know in the comments. We will gladly add that library to our blog post as well.

  • Arindam Paul

    great compilation. I have already bookmarked the site. I have been using scikit extensively and love this because of it’s wide range of ML primitives and it’s consistency across APIs.
    You have mentioned integration with other languages. Is there any hook available for JavaScript/NodeJs? just wondering…

  • bugra

    I am aware that there is a package called node-python (, but its numerical support is not very good and I highly doubt that this package supports numerical libraries namely Numpy, Scipy and Scikit-Learn.

    You could use yhat’s api for combining node.js and Scikit-learn. Here is a post

    how they explain integration of Scikit-learn and their api written in node.js, you may find it useful. Note that you need to register at yhat (username and api_key is required).

  • Danny Bickson

    I would add Pyspark and Yelp’s mrjob which are focused on big data analytics. iPython is an educational/sharing tool. And most important: GraphLab’s Python beta: :-)

  • bugra

    I intentionally excluded analytics libraries to be honest. I have two reasons;

    1) Not all of them have or even support ML in their ecosystem.

    2) They are more of a part of data processing pipeline rather than ML per se.

    Otherwise, Luigi ( and Apache Pig ( could be added as well as the libraries you suggested(MrJob and PySpark).

    GraphLab seems quite interesting, I will try to learn more about it and update the post accordingly. Thanks for letting us know!

  • Radim

    Interesting list Bugra; a few corrections re. gensim (I’m the author :-)

    1. no classification capabilities in gensim at all (nor are they planned); focus is purely on unsupervised learning

    2. main advantage is processing very large datasets (data streaming + online learning algorithms, larger than RAM)

    3. gensim builds on NumPy/BLAS and Cython for performance.

    Happy coding, Python rocks!

  • Joshua

    You may want to consider adding NLTK to the list. It has classification and clustering methods, in addition to a robust Natural Language Processing suite.

    The relevant packages for machine learning are:

    NLTK Cluster

    NLTK Classify

    I’ve had the pleasure of training a NBC a few times with NLTK Classify. It’s pretty easy to figure out, and having access to all the other NLP Tools makes it pretty efficient when you’re using lexical features for the classifier.

  • Pietro Berkes (MDP developer)

    MDP is still actively maintained, and we release patches and contributed improvements once or twice per year. It’s Python 3 compatible etc.

    While it’s true that the number of new features has stagnated from its heydays, MDP still provides implementation of algorithms that cannot be found elsewhere (e.g., Slow Feature Analysis, some flavors of ICA, and Neural Nets), and still has a decent number of users because of that.

  • Dan Blanchard

    If you like scikit-learn, you guys may also like our wrapper around it, SKLL (Sci-Kit Learn Laboratory). It’s FOSS and it simplifies a lot of common workflows for scikit-learn (namely ones involving running repeated experiments with pregenerated features).

    I gave a talk about it at PyData NYC 2013:

  • bugra

    I actually saw your comment just now, sorry about the late response.

    I was not aware that the library is actively maintained and Python 3 compatible, I will update the blog post accordingly.

    I know that MDP was one of the first Python libraries that implemented Deep Belief Network, some Neural Nets and RBM and also widely adopted by neuralscience community.

  • bugra

    I actually attended your talk in November at Pydata, which was quite good. Hopefully SKLL will be helpful for the future readers who want to make their data-processing and machine learning workflow easy.

  • Anh Nguyen

    The popular Caffe has also a Python wrapper!

  • bugra

    At the time when this article was written, there was not a stable Python wrapper, thanks for letting us know!

    Even though they discontinued the Decaffe, they have a Python wrapper for Caffe which could be grabbed from GitHub:

    Hopefully, it will be useful for future readers.

  • Radim

    The description of “gensim” is wrong in almost every respect.

    It doesn’t have “basic classification capabilities” — it has no classification capabilities at all. It focuses purely on unsupervised analysis and topic modeling.

    It’s also highly optimized C/Fortran code (via NumPy and SciPy), with some modules written in extremely optimized Cython for performance.

    It’s main selling point, however, is memory independence. Unlike all the other packages, it uses data streaming (generators, iterators) to process extremely large inputs (larger than RAM).

    I realize these blog posts are popular SEO fodder, but it’s a shame not to do even basic research :)

  • bugra

    Thanks for the comment and feedback.
    I erred in saying gensim has basic classification capabilities in the post and removed the sentence from the post.

  • Alice Zheng

    GraphLab Create. It’s a new ML platform with a Python frontend, designed to scale out of core and has fast and robust data/feature engineering ops and ML models. It also allows for deployment of arbitrary Python objects. More info here: