Making neural networks run in browser with ONNX
summary
ONNX provides a portable model format plus runtimes so a model trained in PyTorch/Keras/ML.NET can be exported (torch.onnx / onnx‑mltools), inspected with netron.app visualization, and run with ONNX Runtime Web in the browser via WebAssembly or WebGL. Demo shows exporting an ML.NET model, loading with ort.InferenceSession.create, feeding ort.Tensor shaped to the model (example emotion model needs 1x1x64x64 input and returns a 1x8 tensor), and calling session.run. Benefits: lower latency, offline/safer inference and much smaller deploy than bundling a 4GB framework, but very large models (200MB+) still hurt UX; check the model zoo first.
AI-written summary · about
chapters (9)
- 0:00ONNX and neural networks in browsers
- 5:02What ONNX is and why it matters
- 10:22Creating ONNX models from frameworks
- 16:50Demo: ML.NET to ONNX
- 21:26Deploying ONNX to edge and cloud
- 26:50ONNX Runtime with Node.js
- 32:38ONNX in the browser
- 43:21Key takeaways and use cases
- 45:51Q&A
related talks
29mML/AI models in JavaScript environment
05mONNX Runtime
33mMachine Learning in the Browser: Fast Iteration with ONNX & WebAssembly
45mLeverage Power of Machine Learning with ONNX
17mRunning Intelligence at the Edge: Web AI in Action
41mLeverage Power of Machine Learning with ONNX
26mDeploying PyTorch Models To the Browser and Beyond With Transformers.js
33mWhat on earth is ONNX and ONNXRUNTIME