open neural network exchange

ONNX (Open Neural Network Exchange) is an open standard file format and runtime ecosystem for representing and executing machine learning models across different frameworks, enabling developers to train models in one framework (PyTorch, TensorFlow, JAX) and deploy them using any ONNX-compatible runtime without framework lock-in. Created by Microsoft and Facebook in 2017 and now governed by the Linux Foundation, ONNX defines a common set of operators (mathematical and neural network operations) and a standardized graph representation that captures model architecture and learned weights in a framework-agnostic format. The ONNX format represents models as computational graphs: nodes are operators (Conv, MatMul, Relu, Attention, LSTM, etc. — over 180 standardized operators), edges carry tensors between nodes, and the graph includes all learned weight values as initializers. This representation captures the model's complete computation without depending on any specific framework's internal representation. The ONNX ecosystem includes: model exporters (torch.onnx.export, tf2onnx, keras2onnx — converting framework-specific models to ONNX format), ONNX Runtime (Microsoft's high-performance inference engine supporting CPU, GPU, and specialized accelerators with graph optimizations like operator fusion, constant folding, and memory planning), hardware-specific optimizers (TensorRT can consume ONNX, OpenVINO accepts ONNX for Intel hardware, CoreML tools can convert ONNX for Apple devices), and model verification tools (comparing outputs between original and ONNX models for numerical consistency). Key benefits include: deployment flexibility (train in PyTorch, deploy on any hardware), inference optimization (ONNX Runtime applies framework-independent optimizations), hardware acceleration (execution providers for CUDA, DirectML, TensorRT, OpenVINO, CoreML, NNAPI), quantization support (INT8 quantization within the ONNX ecosystem for efficient inference), and model inspection tools (Netron for visualization, ONNX checker for validation). ONNX has become the de facto interchange format for deploying ML models in production, particularly for edge deployment and cross-platform scenarios.