.. image:: https://circleci.com/gh/sdpython/onnxcustom/tree/master.svg?style=svg
    :target: https://circleci.com/gh/sdpython/onnxcustom/tree/master

.. image:: https://travis-ci.com/sdpython/onnxcustom.svg?branch=master
    :target: https://app.travis-ci.com/github/sdpython/onnxcustom
    :alt: Build status

.. image:: https://ci.appveyor.com/api/projects/status/a3sn45a2fayoxb5q?svg=true
    :target: https://ci.appveyor.com/project/sdpython/onnxcustom
    :alt: Build Status Windows

.. image:: https://codecov.io/gh/sdpython/onnxcustom/branch/master/graph/badge.svg
    :target: https://codecov.io/gh/sdpython/onnxcustom

.. image:: https://badge.fury.io/py/onnxcustom.svg
    :target: http://badge.fury.io/py/onnxcustom

.. image:: http://img.shields.io/github/issues/sdpython/onnxcustom.png
    :alt: GitHub Issues
    :target: https://github.com/sdpython/onnxcustom/issues

.. image:: https://img.shields.io/badge/license-MIT-blue.svg
    :alt: MIT License
    :target: http://opensource.org/licenses/MIT

.. image:: https://pepy.tech/badge/onnxcustom/month
    :target: https://pepy.tech/project/onnxcustom/month
    :alt: Downloads

.. image:: https://img.shields.io/github/forks/sdpython/onnxcustom.svg
    :target: https://github.com/sdpython/onnxcustom/
    :alt: Forks

.. image:: https://img.shields.io/github/stars/sdpython/onnxcustom.svg
    :target: https://github.com/sdpython/onnxcustom/
    :alt: Stars

.. image:: https://img.shields.io/github/repo-size/sdpython/onnxcustom
    :target: https://github.com/sdpython/onnxcustom/
    :alt: size

onnxcustom: custom ONNX
=======================

.. image:: https://raw.githubusercontent.com/sdpython/onnxcustom/master/_doc/sphinxdoc/source/phdoc_static/project_ico.png
    :width: 50

`documentation <http://www.xavierdupre.fr/app/onnxcustom/helpsphinx/index.html>`_

Examples, tutorial on how to convert machine learned models into ONNX,
implement your own converter or runtime, or even train with ONNX / onnxruntime.

The function *check* or the command line ``python -m onnxcustom check``
checks the module is properly installed and returns processing
time for a couple of functions or simply:

::

    import onnxcustom
    onnxcustom.check()

The documentation also introduces *onnx*, *onnxruntime* for
inference and training.
The tutorial related to *scikit-learn*
has been merged into `sklearn-onnx documentation
<http://onnx.ai/sklearn-onnx/index_tutorial.html>`_.
Among the tools this package implements, you may find:

* a tool to convert NVidia Profilder logs into a dataframe,
* a SGD optimizer similar to what *scikit-learn* implements but
  based on *onnxruntime-training* and able to train an CPU and GPU,
* functions to manipulate *onnx* graph.

**Installation of onnxruntime-training**

onnxruntime-training is only available on Linux. The CPU
can be installed with the following instruction.

::

    pip install onnxruntime-training --extra-index-url https://download.onnxruntime.ai/onnxruntime_nightly_cpu.html

Versions using GPU with CUDA or ROCm are available. Check
`download.onnxruntime.ai <https://download.onnxruntime.ai/>`_
to find a specific version.
You can use it on Windows
inside WSL (Windows Linux Subsystem) or compile it for CPU:

::

    python tools\ci_build\build.py --skip_tests --build_dir .\build\Windows --config Release --build_shared_lib --build_wheel --numpy_version= --cmake_generator="Visual Studio 16 2019" --enable_training --enable_training_ops

GPU versions work better on WSL, see `Build onnxruntime on WSL (Windows Linux Subsystem)
<http://www.xavierdupre.fr/app/onnxcustom/helpsphinx/blog/2021/2021-12-16_wsl.html>`_.
*onnxcustom* can be installed from pypi.