How to use ML NET to write crushing metal riffs
summary
Uses ML.NET tooling where possible but trains the generation model in TensorFlow + Keras because ML.NET cannot import the saved model reliably; attempts to convert to ONNX fail. Builds an LSTM with dropout to predict the 30th token from 29-note sequences, trained on two Metallica albums (~5,200 notes → ~1,200 sequences). Loss falls to about 0.1 during training; generated tablature exports back into Guitar Pro and plays as passable metal riffs — works, but integration into pure .NET remains brittle.
AI-written summary · about
chapters (23)
- 0:00Intro
- 0:45About me
- 1:30Machine learning
- 3:03Linear function
- 4:07Visualizing linear function
- 6:30Example of linear function
- 7:40Activation function
- 7:53Neural networks
- 9:56Output
- 11:06LS DMS
- 12:05Recurrent Neural Networks
- 14:21Piano vs Guitar
- 16:12Tablatures
- 19:05ML NET
- 24:18Why we know the right answer
- 29:13Tensorflow
- 31:08Sequences
- 32:28Guitar Profiles
- 33:18Bars
- 37:38Output file
- 42:48C code
- 48:38Future work
- 52:48Machine learning examples
related talks
31mHow to Use Machine Learning and .NET to Write Crushing Metal Riffs
59mMachine Learning in .NET (ML.NET)
01h 04mIntroduction to ML.NET
52mML.NET - Machine Learning made for .NET
57mMachine Learning Made Easy – An Introduction to ML.NET
27mPythonic Metal
26mPredicting the future (of equipment) using ML.NET
59mMachine Learning simplified for Developers with ML.NET