Making neural networks portable with ONNX

What if your ML models were as portable as a PDF? Learn how ONNX creates a universal standard to run any model on any hardware, even in the browser.

#1about 6 minutes

Understanding ONNX as a portable format for ML models

Machine learning models are made portable across different frameworks and hardware using the ONNX open standard, similar to how PDF works for documents.

#2about 2 minutes

When to use ONNX for your machine learning projects

ONNX is ideal for deploying models across different programming languages, achieving low-latency inferencing, and running on resource-constrained edge or IoT devices.

#3about 12 minutes

Four methods for creating or acquiring ONNX models

Models can be obtained from the ONNX Model Zoo, built with tools like Azure Custom Vision, converted from existing frameworks like PyTorch, or used as an intermediary format.

#4about 7 minutes

Deploying models with the high-performance ONNX Runtime

The ONNX Runtime is a high-performance inference engine for deploying models to the cloud or edge devices, bridging the gap between data science and production software engineering.

#5about 4 minutes

Running an ONNX model in a Node.js application

A practical demonstration shows how to load an ONNX model and perform inference within a server-side Node.js application using the `onnxruntime-node` package.

#6about 9 minutes

Performing inference in the browser with ONNX Runtime Web

An emotion detection model is run directly in the browser using ONNX Runtime Web, showcasing client-side inference with JavaScript for privacy and offline capability.

#7about 3 minutes

Optimizing ONNX models for mobile and React Native

ONNX Runtime Mobile provides a lightweight solution for iOS and Android by converting models to a pre-optimized `.ort` format for smaller binary sizes.

#8about 8 minutes

Q&A on starting a career in machine learning

Advice is given on how software developers can enter the machine learning field by starting with model integration and deployment before diving deep into model creation.