Learn a little bit of AI every day, ONNX: The open standard for interoperability of deep learning models
Neural networks take advantage of deep learning by performing computations through data flow graphs. Various frameworks, including Caffe2, Theano, and TensorFlow, use static graphs, while PyTorch opts for dynamic graphs. Nonetheless, all of these frameworks provide a user-friendly interface for creating computational graphs and an efficient runtime for graph processing. These diagrams act as intermediate representations (IRs) that encapsulate the precise intent of the developer's source code, making it easy to perform optimizations and transformations on specific hardware devices such as CPUs and GPUs.
ONNX, short for Open Neural Network Exchange, is an open-source framework designed to facilitate the exchange of neural network models between different deep learning frameworks.
The main goal of ONNX is to break down the barriers between various deep learning frameworks, allowing developers to take advantage of different libraries without being locked into a single ecosystem. It does this by defining a common format for representing deep learning models that can be understood and executed by multiple frameworks, including TensorFlow and PyTorch, among others.
At present, major AI companies have joined the project, as shown in Figure 2
So, how does ONNX work?
At its core, ONNX defines a common format for representing deep learning models, known as the ONNX Model Format. This format consists of a computational graph, where nodes represent operations and edges represent data flows. Each node in the graph corresponds to a specific operation, such as convolution, pooling, or activation, and contains properties that define its behavior. As shown in Figure 3 and Figure 4.
Why use a generic intermediate representation (IR)?
Despite the similarity of the features offered, each frame boasts its own unique proprietary graphical representation. As a result, each framework acts as a separate stack of APIs, graphs, and runtimes. In addition, these frameworks tend to be tailored for specific attributes, such as rapid training, supporting complex network designs, or facilitating mobile device inference. Choose a framework that meets the required characteristics based on the use case.
In addition, these optimizations may be more suitable for specific development phases, resulting in a considerable time gap between research and production due to the need for conversion, as shown in Figures 4 and 5.
Why is ONNX important?
Model portability is all dependent on ONNX (as shown in Figures 4 and 5):
1. Convert the model from any frame to ONNX format.
2. Convert the ONNX format to any desired framework.
Execution efficiency is also important:
The ONNX model is designed to be highly efficient, so it can run on a wide range of hardware, from GPUs to edge devices. This efficiency is critical for practical applications where computing resources may be limited.
