scikit-learn for OCaml

ocaml-sklearn allows using Python's scikit-learn machine learning library from OCaml.

Read the online scikit-learn OCaml API documentation here.

If you are not familiar with scikit-learn, consult its Python getting started documentation and user guide.

As of version 0.22-0.3.0, most classes and functions from scikit-learn and Numpy should be usable. Many examples have been ported from Python to OCaml successfully (see below). However, the APIs have not yet proved stable and will probably evolve in the next releases.

Example : support vector regression with RBF kernel

module Np = Np.Numpy
let n_samples, n_features = 10, 5 in
Np.Random.seed 0;
let y = Np.Random.uniform ~size:[n_samples] () in
let x = Np.Random.uniform ~size:[n_samples; n_features] () in
let open Sklearn.Svm in
let clf = SVR.create ~c:1.0 ~epsilon:0.2 () in
Format.printf "%a\n" SVR.pp @@ SVR.fit clf ~x ~y;
Format.printf "%a\n" Np.pp @@ SVR.support_vectors_ clf;;

This outputs:

SVR(C=1.0, cache_size=200, coef0=0.0, degree=3, epsilon=0.2, gamma='scale',
    kernel='rbf', max_iter=-1, shrinking=True, tol=0.001, verbose=False)
[[0.14509922 0.16277752 0.99033894 0.84013554 0.96508279]
 [0.8865312  0.80655193 0.07459775 0.36058768 0.22130337]
 [0.21844203 0.09612442 0.49908686 0.1154579  0.98202969]
 [0.07306658 0.97225754 0.20558949 0.16423512 0.57400651]
 [0.08153976 0.41462111 0.66190418 0.70208221 0.3600998 ]
 [0.20502873 0.04244781 0.21800856 0.28184598 0.4282653 ]
 [0.89211037 0.51466381 0.23432621 0.29850877 0.13323457]]

There are more examples in examples/auto, for instance examples/auto/svm.ml.

Installation

opam install sklearn

Finding Python's scikit-learn at runtime

You do not need a Python installation when compiling your OCaml program using ocaml-sklearn. However, when running, your program will need to load the sklearn, numpy and scipy Python libraries, so these must be installed where the OCaml program is deployed.

A version of ocaml-sklearn is tied to a version of Python's scikit-learn, numpy and scipy. For instance, a version of ocaml-sklearn for Python's scikit-learn 0.22.2 will refuse to initialize (by throwing an exception) if scikit-learn's version is not 0.22 (it can however be 0.22.1, 0.22.2 or 0.22.2.post1).

One way to make sure you run with the right versions is to create a virtualenv, install scikit-learn the Python packages inside, and run your OCaml program in the activated virtualenv.

Do this once to create the virtualenv in directory .venv and install scikit-learn, numpy and scipy inside:

python3 -mvenv .venv
source .venv/bin/activate
pip install scikit-learn==%%SKLEARN_FULL_VERSION%% numpy==%%NUMPY_FULL_VERSION%% scipy==%%SCIPY_FULL_VERSION%% pytest

Then run your compiled OCaml program inside the virtualenv:

source .venv/bin/activate
./my_ocaml_program.exe

API

We attempt to bind all of scikit-learn's APIs. However, not all of the APIs are currently tested, and some are probably hard to use or unusable at the moment.

Each Python module or class gets its own OCaml module. For instance Python class sklearn.svm.SVC can be found in OCaml module Sklearn.Svm.SVC. This module has a create function to construct an SVC and functions corresponding to the Python methods and attributes.

Most data is passed in and out of sklearn through module Ndarray (in module Np.Np.Numpy).

You should generally build a dense array using the constructors in Np.Numpy:

module Np = Np.Numpy
let x = Np.matrixi [|[| 1; 2 |]; [| 3; 4 |]|]

To get data out of an Ndarray, use to_int_array, to_float_array or to_string_array (all of these return a flattened copy of the data, and will raise an exception if the data type is wrong).

Attributes are exposed read-only, each with two getters: one that raises Not_found if the attribute is None, and the other that returns an option.

Bunches (as returned from the sklearn.datasets APIs) are exposed as objects.

Arguments taking string values are converted (in most cases) to polymorphic variants.

Each module has a conversion function to Py.Object.t (called to_pyobject), so that you can always escape and use pyml directly if the API provided here is incomplete.

No attempt is made to expose features marked as deprecated.

Development notes

ocaml-sklearn's sources are generated using a Python program (see lib/skdoc.py) that loads up sklearn and uses introspection to generate bindings based on pyml. To determine types, it parses scikit-learn's documentation.

Dev tl;dr

sudo apt install python-3.8 python3.8-venv libpython3.8
python3.8 -mvenv .venv
source .venv/bin/activate
pip install -r requirements-dev.txt
opam switch create . 4.11.1 --deps-only
eval `opam env`
dune runtest

Python requirements

Make sure you have Python >= 3.8, skdoc.py needs it.

The requirements for developing (not using) the bindings are in file requirements-dev.txt. Install it using:

sudo apt install python-3.8 python3.8-venv libpython3.8
python3.8 -mvenv .venv
source .venv/bin/activate
pip install -r requirements-dev.txt

Running tests

dune runtest

The tests are in examples/auto. They are based on examples extracted from the Python documentation. A good way to develop is to pick one of the files and start porting examples.

The following examples have been ported completely:

• calibration
• cluster
• compose
• ensemble
• impute
• manifold
• metrics
• model_selection
• neighbors
• neural_network
• pipeline
• preprocessing
• svm
• tree

The following examples still need to be ported:

• covariance
• cross_decomposition
• datasets
• decomposition
• discriminant_analysis
• feature_extraction
• feature_selection
• gaussian_process
• inspection
• isotonic
• kernel_approximation
• linear_model
• multiclass
• multioutput
• naive_bayes
• random_projection
• semi_supervised
• utils

Generating documentation

lib/build-doc

Documentation can then be found in html_doc/. Serve it locally with something like:

python3 -mhttp.server --directory html_doc
xdg-open http://localhost:8000

License

BSD-3. See file LICENSE.