repository·main·Indexed 12 days ago
https://github.com/onnx/onnxAn open ecosystem providing a standardized, extensible computation graph format for AI models, supporting both deep learning and traditional machine learning to enable interoperability across different frameworks and hardware.
Open Neural Network Exchange (ONNX) is an open ecosystem providing an open source format for AI models (both deep learning and traditional ML). It defines:
Currently, the project focuses on capabilities required for inferencing (scoring).
The core CI pipelines ensure code quality through testing, linting, and compliance checks. Key workflows include:
lintrunner (including ruff, mypy, clang-format, etc.) and verifies that auto-generated files are up to date.topic: or module: label (except for Dependabot PRs).ONNX implements several workflows to maintain security and supply chain integrity:
main/rel-*, and weekly).main and weekly).onnx.checker module provides utilities to validate the correctness of ONNX models. It ensures that the model graph, operators, and data types conform to the ONNX specification. You can use the checker to verify that a model is structurally sound and ready for inference or further processing.The following workflows manage the project's documentation and long-term maintenance:
The onnx.helper module provides a suite of utility functions designed to simplify the programmatic creation of ONNX graph components. Instead of manually constructing complex Protocol Buffer objects, you can use these helpers to build nodes, tensors, graphs, and models.
Key component creation functions include:
make_node, make_attribute, and make_attribute_ref.make_tensor, make_tensor_value_info, make_empty_tensor_value_info, and make_sparse_tensor.make_tensor_type_proto, make_sequence_type_proto, make_map_type_proto, and make_optional_type_proto.make_graph, make_model, make_operatorsetid, and make_training_info.To use a model in production with ONNX, you typically use a converting library to translate a framework-specific model (like PyTorch or TensorFlow) into an ONNX graph. These libraries rewrite the model's prediction function using ONNX operators.
Key characteristics of converting libraries:
tensorflow-onnx for TensorFlow, torch.onnx for PyTorch).scikit-learn (which rely on NumPy/SciPy) may require you to manually implement transformers using ONNX primitives.onnx.backend module provides abstractions for interacting with different execution backends and hardware devices. It allows developers to define how ONNX models are represented and executed on specific hardware targets through the Backend and BackendRep classes.An Optional type represents a reference to either an element (Tensor, Sequence, Map, or Sparse Tensor) or a null value. This is analogous to the Optional[X] type hint in Python (equivalent to Union[None, X]).
Optional types can appear in model inputs, outputs, or as intermediate values. They enable dynamic typing scenarios where a value might be present or absent (null).
ONNX (Open Neural Network Exchange) provides a common Intermediate Representation (IR) for computation graphs used in deep learning. Instead of being locked into a single framework's proprietary stack (API, graph, and runtime), ONNX allows developers to:
Type Denotation is a mechanism used to describe semantic information about model inputs and outputs. It is stored within the TypeProto message.
While standard tensor shapes define the structure of data, Type Denotation provides context about the nature of the data (e.g., whether a tensor represents an image, audio, or text). This allows model consumers to understand requirements such as color channel order, pixel bit depth, or normalization ranges without relying on external documentation.
ONNX Backend Test is a verification suite designed for ONNX backend implementations to ensure they fulfill the ONNX standard. It serves two purposes: verifying backend correctness and defining the expected behavior of operators (complementing the official documentation).
The suite consists of two testing levels: