ONNXRuntime-Extensions
What is ONNXRuntime-Extensions?
ONNXRuntime-Extensions is a library that extends the capability of the ONNX models and inference with ONNX Runtime, via the ONNX Runtime custom operator interface. It includes a set of Custom Operators to support common model pre and post-processing for audio, vision, text, and language models. As with ONNX Runtime, Extensions also supports multiple languages and platforms (Python on Windows/Linux/macOS, Android and iOS mobile platforms and Web-Assembly for web.
The basic workflow is to add the custom operators to an ONNX model and then to perform inference on the enhanced model with ONNX Runtime and ONNXRuntime-Extensions packages.
This image was created using Combine.AI, which is powered by Bing Chat, Bing Image Creator, and EdgeGPT.
Quickstart
Python installation
pip install onnxruntime-extensions
Nightly Build
on Windows
pip install --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT-Nightly/pypi/simple/ onnxruntime-extensions
Please ensure that you have met the prerequisites of onnxruntime-extensions (e.g., onnx and onnxruntime) in your Python environment.
on Linux/macOS
Please make sure the compiler toolkit like gcc(later than g++ 8.0) or clang are installed before the following command
python -m pip install git+https://github.com/microsoft/onnxruntime-extensions.git
NuGet installation (with .NET CLI)
dotnet add package Microsoft.ML.OnnxRuntime.Extensions --version 0.8.1-alpha
Add pre and post-processing to the model
There are multiple ways to get the ONNX processing graph:
- Use the pre-processing pipeline API if the model and its pre-processing is supported by the pipeline API
- Export to ONNX from a PyTorch model
- Create an ONNX model with a model graph that includes your custom op node
- Compose the pre-processing with an ONNX model using ONNX APIs if you already have the pre processing in an ONNX graph
If the pre processing operator is a HuggingFace tokenizer, you can also easily get the ONNX processing graph by converting from Huggingface transformer data processing classes such as in the following example:
import onnxruntime as _ort
from transformers import AutoTokenizer
from onnxruntime_extensions import OrtPyFunction, gen_processing_models
tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased')
model = OrtPyFunction(gen_processing_models(tokenizer, pre_kwargs={})[0])
For more information, you can check the API using the following:
help(onnxruntime_extensions.gen_processing_models)
What if I cannot find the custom operator I am looking for?
Find the custom operators we currently support here. If you do not find the custom operator you are looking for, you can add a new custom operator to ONNX Runtime Extensions like this. Note that if you do add a new operator, you will have to build from source.
Inference with ONNX Runtime and Extensions
Python
There are individual packages for the following languages, please install it for the build.
import onnxruntime as _ort
from onnxruntime_extensions import get_library_path as _lib_path
so = _ort.SessionOptions()
so.register_custom_ops_library(_lib_path())
# Run the ONNXRuntime Session as per ONNXRuntime docs suggestions.
sess = _ort.InferenceSession(model, so)
sess.run (...)
C++
// The line loads the customop library into ONNXRuntime engine to load the ONNX model with the custom op
Ort::ThrowOnError(Ort::GetApi().RegisterCustomOpsLibrary((OrtSessionOptions*)session_options, custom_op_library_filename, &handle));
// The regular ONNXRuntime invoking to run the model.
Ort::Session session(env, model_uri, session_options);
RunSession(session, inputs, outputs);
Java
var env = OrtEnvironment.getEnvironment();
var sess_opt = new OrtSession.SessionOptions();
/* Register the custom ops from onnxruntime-extensions */
sess_opt.registerCustomOpLibrary(OrtxPackage.getLibraryPath());
C#
SessionOptions options = new SessionOptions();
options.RegisterOrtExtensions();
session = new InferenceSession(model, options);
Tutorials
Check out some end to end tutorials with our custom operators:
- NLP: An end-to-end BERT tutorial
- Audio: Using audio encoding and decoding for Whisper
- Vision: The YOLO model with our DrawBoundingBoxes operator
Contributing
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.
When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
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.