ONNX - File Format

The ONNX file format is basically a container that encapsulates the entire structure of a machine learning model. When you export a model to ONNX format, the resulting file (usually with a .onnx extension) contains a graph-based representation of the model. If you have ever worked with deep learning architectures like ImageNet, MobileNet, or AlexNet, you will find ONNX representations quite familiar.

The ONNX file format is a flexible way to represent machine learning models, making them portable across various frameworks. At its core, an ONNX model is composed of a graph, a collection of computation nodes. Each node in the graph represents a specific operation (e.g., convolution, ReLU activation), and the data moves through these nodes in a sequence, just like the flow of data through layers in a neural network.

In this tutorial, we will learn about the ONNX file format in detail, discussing how it structures a model, including its components like the computational graph, nodes, operators, and metadata.

Key Components of the ONNX File Format

An ONNX file contains various components that help describe the model, its structure, and how it works. Let's discuss these components.

Model

The ONNX file represents a model and includes key elements such as −

• Version Info: Details about the ONNX version used.
• Metadata: Additional information about the model, such as author or framework used.
• Acyclic Computation Data-flow Graph: The core structure that describes how the computations are performed within the model.

Graph

The heart of the ONNX file is the graph that represents the flow of computations. Think of it like a map of the operations your model goes through to process inputs and produce outputs.

• Inputs and Outputs: Describes the data that enters and leaves the model. It includes information like the data type and shape (dimensions).
• List of Computation Nodes: Each node in the graph represents a computation step (e.g., applying an activation function).
• Graph Name: The name of the graph that describes the overall structure of the model.

Computation Nodes

Each computation node performs a specific task, such as applying a function to the input data.

• Inputs: Every node may have zero or more inputs of predefined types, which are the outputs of other nodes or external inputs.
• Outputs: Each node produces one or more outputs that are passed to the next computation node in the sequence.
• Operators: These represent the operations being applied (e.g., matrix multiplication, ReLU activation).
• Operator Parameters: Each operator may have certain parameters like learning rate or normalization factors.

Understanding the Graph Structure

The structure of the ONNX graph is a series of interconnected computation nodes, with data flowing between them. Think of it like the layers of a neural network, where each layer processes the input data and passes it on to the next layer.

• Graph Inputs: These are the starting points for the data, where the input is defined (e.g., the image tensor in a CNN).
• Graph Outputs: After the data has passed through all the computation nodes, the final output tensor is produced (e.g., the classification label).
• Computation Nodes: These represent individual operations in the model, such as convolutions, activations (ReLU), pooling, and fully connected layers.