1.1. What is Tiny ML Tensorlab?

Tiny ML Tensorlab is Texas Instruments’ comprehensive AI toolchain for developing, training, optimizing, and deploying machine learning models on resource-constrained microcontrollers (MCUs).

1.1.1. Overview

The toolchain provides an end-to-end workflow that takes you from raw data to a compiled model running on a TI MCU. It handles:

  • Data Preparation - Format and preprocess datasets for training

  • Feature Extraction - Transform raw signals into meaningful features

  • Model Training - Train neural networks optimized for embedded deployment

  • Model Optimization - Quantize models for efficient inference (2/4/8-bit)

  • Compilation - Generate device-specific code using TI’s Neural Network Compiler

  • Analysis - Evaluate model performance before deployment

1.1.2. Target Applications

Tiny ML Tensorlab is designed for industrial and IoT applications where you need to run AI inference directly on the microcontroller:

Predictive Maintenance

  • Motor bearing fault detection

  • Equipment anomaly monitoring

  • Vibration analysis

Power Systems

  • DC and AC arc fault detection

  • Grid stability prediction

  • Power quality monitoring

Control Systems

  • Torque estimation

  • Temperature forecasting

  • Load prediction

Sensor Processing

  • Activity recognition

  • Gas classification

  • Motion detection

1.1.3. Key Capabilities

Supported ML Tasks

Time Series Classification

Categorize sensor signals into discrete classes

Time Series Regression

Predict continuous values from sensor inputs

Time Series Forecasting

Predict future values based on historical patterns

Anomaly Detection

Identify abnormal patterns using autoencoders

Image Classification

Categorize images into classes

Device Support

Over 20 TI microcontrollers are supported, including:

  • C2000 DSP family - F28P55 (with NPU), F28P65, F2837, F28003, F28004

  • MSPM0 family - MSPM0G3507, MSPM0G5187 (with NPU)

  • MSPM33C family - MSPM33C32, MSPM33C34, AM13E2 (with NPU)

  • AM26x family - AM263, AM263P, AM261

  • Connectivity devices - CC2755, CC1352

Model Optimization

  • Quantization-Aware Training (QAT) for best accuracy

  • Post-Training Quantization (PTQ) for fast deployment

  • Support for 2-bit, 4-bit, and 8-bit weight quantization

NPU Acceleration

Select devices (F28P55, AM13E2, MSPM0G5187) include TI’s Neural Processing Unit (TINPU) for hardware-accelerated inference, providing faster execution and lower power consumption.

1.1.4. Repository Structure

Tiny ML Tensorlab consists of four main components:

tinyml-modelzoo (Customer Entry Point)

Contains model definitions, example configurations, and pre-trained checkpoints. This is where you start when using the toolchain.

  • examples/ - Ready-to-run configuration files

  • tinyml_modelzoo/models/ - Model architectures

  • docs/ - NPU configuration guidelines

tinyml-modelmaker

Orchestrates the end-to-end workflow. Provides scripts to run training, testing, and compilation.

  • examples/ - Additional example configurations

  • docs/ - Detailed documentation and guides

  • scripts/ - Utility scripts

tinyml-tinyverse

Core training infrastructure. Contains:

  • Dataset classes for various data formats

  • Training scripts for all task types

  • Feature extraction transforms

  • Common utilities

tinyml-modeloptimization

Quantization toolkit with:

  • TINPU wrappers for NPU deployment

  • Generic quantization wrappers

  • QAT and PTQ implementations

1.1.5. Workflow Summary

A typical workflow with Tiny ML Tensorlab:

  1. Prepare Your Data Format your dataset according to the required structure (see Bring Your Own Data)

  2. Create a Configuration File Define your training parameters in a YAML config (see Understanding Config)

  3. Run Training

    ./run_tinyml_modelzoo.sh examples/your_config/config.yaml

  4. Review Results Examine training metrics, visualizations, and analysis outputs

  5. Deploy to Device Copy compiled artifacts to your CCS project and flash to the MCU

1.1.6. Next Steps