Making Machine Learning Applications Fast and Simple with...
summary
Demonstrates using Open Neural Network Exchange (ONNX) to export scikit-learn pipelines into fully self-contained computational graphs, shows conversion with sklearn-onnx and onnxmltools, and measures real-world gains where a C++ ONNX runtime cuts single-row latency dramatically (reduces single-row inference time by up to 99%). Covers writing converters for custom estimators, tradeoffs like DAG-only representability, awkward categorical support, cryptic errors in a young ecosystem, and when conversion work outweighs maintaining Python-dependent deployments.
AI-written summary · about
chapters (9)
- 0:00The challenges of deployment
- 2:33Common deployment roadblocks
- 5:13Decoupling with ONNX
- 9:27Understanding ONNX
- 13:35Computational graph logic
- 17:34Converting models to ONNX
- 23:04Live demo and benchmarking
- 30:29Summary and conclusion
- 33:05Question and answer session
related talks
41mLeverage Power of Machine Learning with ONNX
45mLeverage Power of Machine Learning with ONNX
05mONNX Runtime
30mThe Evolution of ONNX: Driving Interoperability in AI...
33mMachine Learning in the Browser: Fast Iteration with ONNX & WebAssembly
33mWhat on earth is ONNX and ONNXRUNTIME
05mONNX-Go, neural networks made easy
30mScalable ML acceleration with ONNX Runtime