Vclav Novotn Software developer Cybernetics Automation VUT FEKT
Václav Novotný
• Software developer • Cybernetics & Automation - VUT, FEKT • Enthusiastic about Computer Vision and Artificial Intelligence in general 2
• • • About ML. NET Pipeline Demo #1, Demo #2, Demo #3 Deep learning and Demo #4 Devtips 3
4
• Open-source • Create and consume custom models • Import and use pre-trained Tensorflow and ONNX models • Online and offline scenarios • Including Azure Functions • Runs on Windows, Linux and mac. OS using. NET Core 5
Raw data 7
Raw data Prepare data 8
Raw data Prepare data Train model 9
Raw data Prepare data Train model Evaluate model 10
Raw data Prepare data Train model Evaluate model Consume 11
Raw data 12
Column operations, • Normalizations, • Data conversions, • Text/image transformations, • Feature selections, • Missing value operations, • … • Prepare data 13
• Clustering • Binary/Multiclassification • Including deep learning • Regression • Anomaly detection • Ranking • Recommendation Train model 14
• Binary/Multiclassification metrics • Accuracy, AUCPR, F 1 • Micro/Macro accuracy, Log-loss reduction • Regression metrics • R-squared, Absolute-loss, Squared-loss, RMS-loss • Confusion matrix • https: //docs. microsoft. com/enus/dotnet/machinelearning/resources/metrics Evaluate model 15
• Trained models can be stored to ZIP files, or saved as byte array to database • Prediction. Engine • Data. Create. Enumerable<> Consume 16
DEMO
Raw data Prepare data Train model Evaluate model Consume 18
Raw data Prepare data Train model Evaluate model Consume 19
DEMO
DEMO
22
23
• Since ML. NET >=1. 4 • Support for ONNX models • Integration of Tensorflow. NET • Support for GPU acceleration 24
DEMO
• Occam‘s razor • Normalize your data • Balance your datasets • Strong classifier principle 26
VIDEO
28
Q&A Visit us at www. ysoft. com
- Slides: 28