8. Features
Features of Tiny ML ModelZoo that help you prepare data, make models smaller and faster, and understand the results.
Data and training
- 8.1. Feature Extraction
- 8.1.1. Pipeline
- 8.1.2. Configuration parameters
- 8.1.3. Presets
- 8.1.4. Data processing transforms
- 8.1.5. Feature extraction transforms
- 8.1.6. Custom feature extraction
- 8.1.7. Windowing and framing
- 8.1.8. Audio front-ends
- 8.1.9. Radar and image inputs
- 8.1.10. Evaluating feature extraction quality
- 8.1.11. Memory cost on the device
- 8.1.12. Best practices
- 8.1.13. Next steps
- 8.2. Goodness of Fit
- 8.3. Early Stopping
- 8.4. Training Performance
- 8.5. Neural Architecture Search
Quantization
Analysis and memory
On-device training
8.13. New in 1.5.0
Early Stopping: on by default, with patience and min-delta.
Memory Estimation: Flash and RAM estimate before training, and the standalone
estimate_memorytool.Mixed Precision Quantization: HAWQ automatic mixed precision and ternary-weight QAT.
Quantization: residual connection support, the cleaner
auto_quantizationAPI and NPU quantization for anomaly detection.Feature Extraction: audio front ends (MFCC, LPC, FFT, learned filterbank) and new transforms.
On-Device Training (ODT): new integer on-device training.
Training Performance:
torch.compile, native AMP and NaN loss detection.
8.14. Overview
Presets, custom transforms, windowing and audio front ends.
8-, 4- and 2-bit QAT and PTQ for the NPU.
Know the Flash and RAM cost before you train.
Let DARTS find a model for your dataset.
ROC curves, thresholds and error plots.
Continue training on the MCU.