ONNX Runtime is a cross-platform inferencing and training accelerator compatible with many popular ML/DNN frameworks, including PyTorch, TensorFlow/Keras, scikit-learn, and more. aka.ms/onnxruntime

Many users can benefit from ONNX Runtime, including those looking to:

Improve inference performance for a wide variety of ML models
Reduce time and cost of training large models
Train in Python but deploy into a C#/C++/Java app
Run on different hardware and operating systems
Support models created in several different frameworks

ONNX Runtime inferencing APIs are stable and production-ready since the 1.0 release in October 2019 and can enable faster customer experiences and lower costs.

ONNX Runtime training feature was introduced in May 2020 in preview. This feature supports acceleration of PyTorch training on multi-node NVIDIA GPUs for transformer models. Additional updates for this feature are coming soon.

Get Started
- ONNX Runtime Inferencing
- ONNX Runtime Training
Data/Telemetry
Contributions and Feedback
License

Get Started

Frequently Asked Questions

Inferencing: Start

To use ONNX Runtime, refer to the table on aka.ms/onnxruntime for instructions for different build combinations.

Compatibility

Supporting models based on the standard ONNX format, the runtime is compatible with PyTorch, scikit-learn, TensorFlow, Keras, and all other frameworks and tools that support the interoperable format.

Getting ONNX models - tutorials

ONNX Runtime is up to date and backwards compatible with all operators (both DNN and traditional ML) since ONNX v1.2.1+. (ONNX compatibility details). Newer versions of ONNX Runtime support all models that worked with prior versions, so updates should not break integrations.

Supported operators/types
- Operators not supported in the current ONNX spec may be available as a Contrib Operator
Extensibility: Add a custom operator/kernel

Binaries

Official builds are available on PyPi (Python), Nuget (C#/C/C++), Maven Central (Java), and npm (node.js).

Default CPU Provider (Eigen + MLAS)
GPU Provider - NVIDIA CUDA
GPU Provider - DirectML (Windows)
- On Windows, the DirectML execution provider is recommended for optimal performance and compatibility with a broad set of GPUs.

Dev builds created from the master branch are available for testing newer changes between official releases. Please use these at your own risk. We strongly advise against deploying these to production workloads as support is limited for dev builds.

Repository	Details
Pypi (Python)	If using pip, run `pip install --upgrade pip` prior to downloading. CPU: onnxruntime / ort-nightly (dev) GPU: onnxruntime-gpu / ort-gpu-nightly (dev)
Nuget (C#/C/C++)	CPU: Microsoft.ML.OnnxRuntime / ort-nightly (dev) GPU: Microsoft.ML.OnnxRuntime.Gpu / ort-nightly (dev)
Maven Central (Java)	CPU: com.microsoft.onnxruntime/onnxruntime GPU: com.microsoft.onnxruntime/onnxruntime_gpu
npm (node.js)	CPU: onnxruntime
Other	Contributed non-official packages (including Homebrew, Linuxbrew, and nixpkgs) These are not maintained by the core ONNX Runtime team and may have limited support; use at your discretion.

System Requirements

The following are required for usage of the official published packages.

Visual C++ Runtime (for Windows packages)
- Requires Visual C++ 2019 runtime
System language
- Installation of the English language package and configuring en_US.UTF-8 locale is required, as certain operators makes use of system locales.
- For Ubuntu, install language-pack-en package
  - Run the following commands: locale-gen en_US.UTF-8 update-locale LANG=en_US.UTF-8
  - Follow similar procedure to configure other locales on other platforms.
Default CPU
- ONNX Runtime binaries in the CPU packages use OpenMP and depend on the library being available at runtime in the system.
  - For Windows, OpenMP support comes as part of VC runtime. It is also available as redist packages: vc_redist.x64.exe and vc_redist.x86.exe
  - For Linux, the system must have libgomp.so.1 which can be installed using apt-get install libgomp1.
  - For Mac OS X, the system must have libomp.dylib which can be installed using brew install libomp.
Default GPU (CUDA)
- The default GPU build requires CUDA runtime libraries being installed on the system:
  - Version: CUDA 10.2 and cuDNN 8.0.3
- Version dependencies from older ONNX Runtime releases can be found in prior release notes.

Build from Source

For production scenarios, it's strongly recommended to build only from an official release branch.

Instructions for additional build flavors

Docker Images

ONNX-Ecosystem: includes ONNX Runtime (CPU, Python), dependencies, tools to convert from various frameworks, and Jupyter notebooks to help get started
Additional dockerfiles

API Documentation

API	Supported Versions	Samples
Python	3.5, 3.6, 3.7, 3.8 (3.8 excludes Win GPU and Linux ARM) Python Dev Notes	Samples
C#		Samples
C++		Samples
C		Samples
WinRT	Windows.AI.MachineLearning	Samples
Java	8+	Samples
Ruby (external project)	2.4-2.7	Samples
Javascript (node.js)	12.x	Samples

Supported Accelerators

Execution Providers

CPU	GPU	IoT/Edge/Mobile	Other
Default CPU - MLAS (Microsoft Linear Algebra Subprograms) + Eigen Intel DNNL Intel nGraph Intel MKL-ML (build option)	NVIDIA CUDA NVIDIA TensorRT DirectML AMD MIGraphX (preview)	Intel OpenVINO ARM Compute Library (preview) Android Neural Networks API (preview) ARM-NN (preview) Rockchip NPU (preview)	Nuphar Model Compiler - (preview) Xilinx Vitis-AI (preview)

Deploying ONNX Runtime

Cloud

ONNX Runtime can be deployed to any cloud for model inferencing, including Azure Machine Learning Services.
- Detailed instructions
- AzureML sample notebooks
ONNX Runtime Server (beta) is a hosting application for serving ONNX models using ONNX Runtime, providing a REST API for prediction.
- Usage details
- Image installation instructions

IoT and edge devices

Reference implementations

The expanding focus and selection of IoT devices with sensors and consistent signal streams introduces new opportunities to move AI workloads to the edge. This is particularly important when there are massive volumes of incoming data/signals that may not be efficient or useful to push to the cloud due to storage or latency considerations. Consider: surveillance tapes where 99% of footage is uneventful, or real-time person detection scenarios where immediate action is required. In these scenarios, directly executing model inferencing on the target device is crucial for optimal assistance.

Client applications

Install or build the package you need to use in your application. (sample implementations using the C++ API)
On newer Windows 10 devices (1809+), ONNX Runtime is available by default as part of the OS and is accessible via the Windows Machine Learning APIs. (Tutorials for Windows Desktop or UWP app)

Training: Start

The ONNX Runtime training feature enables easy integration with existing Pytorch trainer code to accelerate the exection. With a few lines of code, you can add ONNX Runtime into your existing training scripts and start seeing acceleration. The current preview version supports training acceleration for transformer models on NVIDIA GPUs.

ONNX Runtime pre-training sample: This sample is setup to pre-train the BERT-Large model to show how ONNX Runtime training can be used to accelerate training execution.

Train PyTorch model with ONNX Runtime

ONNX Runtime (ORT) has the capability to train existing PyTorch models through its optimized backend. For this, we have introduced an python API for PyTorch, called ORTTrainer, which can be used to switch the training backend for PyTorch models (instance of torch.nn.Module) to orttrainer. This requires some changes in the trainer code, such as replacing the PyTorch optimizer, and optionally, setting flags to enable additional features such as mixed-precision training. Here is a sample code fragment to integrate ONNX Runtime Training in your PyTorch pre-training script:

NOTE: The current API is experimental and expected to see significant changes in the near future. Our goal is to improve the interface to provide a seamless integration with PyTorch training that requires minimal changes in users’ training code.

import torch
...
import onnxruntime
from onnxruntime.training import ORTTrainer, optim

# Model definition
class NeuralNet(torch.nn.Module):
  def __init__(self, input_size, hidden_size, num_classes):
    ...
  def forward(self, data):
    ...

model = NeuralNet(input_size=784, hidden_size=500, num_classes=10)
criterion = torch.nn.Functional.cross_entropy 
model_description = {'inputs':  [('data', ['in', 'batch_size']),
                                 ('target', ['label_x_batch_size'])],
                     'outputs': [('loss', [], True),
                                 ('output', ['out', 'batch_size'])]}

optimizer_config = optim.AdamConfig(lr=learning_rate)

trainer = ORTTrainer(model,              # model
                     model_description,  # model description
                     optimizer_config,   # optimizer configuration
                     criterion)          # loss function

# Training Loop
for t in range(1000):
  # forward + backward + weight update
  loss, y_pred = trainer.train_step(input_data, target_labels, learning_rate)
  total_loss += loss.item()
  ...

Build ONNX Runtime Training from source

To use ONNX Runtime training in a custom environment, like on-prem NVIDIA DGX-2 clusters, you can use these build instructions to generate the Python package to integrate into existing trainer code.

Data/Telemetry

This project may collect usage data and send it to Microsoft to help improve our products and services. See the privacy statement for more details.

Contributions and Feedback

We welcome contributions! Please see the contribution guidelines.

For any feedback or to report a bug, please file a GitHub Issue.

Code of Conduct

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

License

This project is licensed under the MIT License.

tino926/onnxruntime

Table of Contents