Skip to main content

Python wrapper for the Lolo machine learning library

Project description

Python Wrapper for Lolo

``lolopy`` implements a Python interface to the `Lolo machine learning
library <>`__.

Lolo is a Scala library that contains a variety of machine learning
algorithms, with a particular focus on algorithms that provide robust
uncertainty estimates. ``lolopy`` gives access to these algorithms as
scikit-learn compatible interfaces and automatically manages the
interface between Python and the JVM (i.e., you can use ``lolopy``
without knowing that it is running on the JVM)


``lolopy`` is available on PyPi. Install it by calling:

``pip install lolopy``

To use ``lolopy``, you will also need to install Java JRE >= 1.8 on your
system. The ``lolopy`` PyPi package contains the compiled ``lolo``
library, so it is ready to use after installation.


Lolopy requires Python >= 3.6, Java JDK >= 1.8, and Maven to be
installed on your system when developing lolopy.

Before developing ``lolopy``, compile ``lolo`` on your system using
Maven. We have provided a ``Makefile`` that contains the needed
operations. To build and install ``lolopy`` call ``make`` in this


The ``RandomForestRegressor`` class most clearly demonstrates the use of
``lolopy``. This class is based on the `Random Forest with
Jackknife-based uncertainty estimates of Wagner et
al <>`__, which -
in effect - uses the variance between different trees in the forest to
produce estimates of the uncertainty of each prediction. Using this
algorithm is as simple as using the `RandomForestRegressor from
scikit-learn <>`__:

.. code:: python

from lolopy.learners import RandomForestRegressor

rf = RandomForestRegressor(), y)
y_pred, y_std = rf.predict(X, return_std=True)

The results of this code is to produce the predicted values (``y_pred``)
and their uncertainties (``y_std``).

See the ```examples`` <./examples>`__ folder for more examples and

You may need to increase the amount of memory available to ``lolopy``
when using it on larger dataset sizes. Setting the maximum memory
footprint for the JVM running the machine learning calculations can be
achieved by setting the ``LOLOPY_JVM_MEMORY`` environment variable. The
value for ``LOLOPY_JVM_MEMORY`` is used to set the maximum heap size for
the JVM (see `Oracle's documentation for
details <>`__).
For example, "4g" allows ``lolo`` to use 4GB of memory.

Implementation and Performance

``lolopy`` is built using the `Py4J <>`__ library
to interface with the Lolo scala library. Py4J provides the ability to
easily managing a JVM server, create Java objects in that JVM, and call
Java methods from Python. However, Py4J `has slow performance in
transfering large
arrays <>`__. To transfer
arrays of features (e.g., training data) to the JVM before model
training or evaluation, we transform the data to/from Byte arrays on the
Java and Python sides. Transfering data as byte arrays does allow for
quickly moving data between the JVM and Python but requires holding 3
copies of the data in memory at once (Python, Java Byte array, and Java
numerical array). We could reduce memory usage by passing the byte array
in chunks, but this is currently not implemented.

Our performance for model training is comparable to scikit-learn, as
shown in the figure below. The blue-shaded region in the figure
represents the time required to pass training data to the JVM. We note
that training times are equivalent between using the Scala interface to
Lolo and ``lolopy`` for training set sizes above 100.

.. figure:: ./examples/profile/training-performance.png
:alt: training performance

training performance
Lolopy and lolo are currently slower than scikit-learn for model
evaluation, as shown in the figure below. The model timings are
evaluated on a dataset size of 1000 with 145 features. The decrease in
model performance with training set size is an effect of the number of
trees in the forest being equal to the training set size. Lolopy and
lolo have similar performance for models with training set sizes of
above 100. Below a training set size of 100, the cost of sending data
limits the performance of ``lolopy``.

.. figure:: ./examples/profile/evaluation-performance.png
:alt: evaluation performance

evaluation performance
For more details, see the `benchmarking
notebook <./examples/profile/scaling-test.ipynb>`__.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for lolopy, version 1.0.3
Filename, size File type Python version Upload date Hashes
Filename, size lolopy-1.0.3-py2.py3-none-any.whl (51.4 MB) File type Wheel Python version py2.py3 Upload date Hashes View
Filename, size lolopy-1.0.3.tar.gz (51.4 MB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page