1.1. What is Tiny ML ModelZoo?

Tiny ML ModelZoo is Texas Instruments’ central repository of AI models, example applications and configurations for microcontroller (MCU) applications. Together with the toolchain it installs, it takes you from raw data to a compiled model running on a TI MCU.

Note

Up to release 1.4.0 this toolchain was distributed as Tiny ML Tensorlab (tinyml-tensorlab), a set of repositories you cloned together. From 1.5.0, tinyml-modelzoo is the entry point and everything else comes in as prebuilt wheels (Installation). tinyml-tensorlab is deprecated and its source is private.

1.1.1. Overview

The toolchain handles:

  • Data preparation: format and preprocess datasets for training

  • Feature extraction: transform raw signals into meaningful features

  • Model training: train neural networks designed for embedded deployment

  • Model optimization: quantize models for efficient inference

  • Compilation: generate device-specific code with TI’s Neural Network Compiler

  • Analysis: evaluate model performance before deployment

1.1.2. Target Applications

Predictive maintenance

Motor bearing, gearbox and fan blade fault detection, equipment anomaly monitoring, vibration analysis.

Power systems

DC and AC arc fault detection, grid fault and stability, electrical fault classification.

Control systems

Torque estimation, motor speed prediction, temperature forecasting, battery state-of-charge estimation.

Sensor processing and sensing

Activity and fall detection, PIR and Wi-Fi CSI presence detection, gas classification, radar point cloud classification.

Audio and vision

Keyword and wake word detection, cough and glass break detection, image classification.

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 from historical patterns

Time Series Anomaly Detection

Identify abnormal patterns using autoencoders

Audio Classification

Classify audio from MFCC features

Image Classification

Categorize images into classes

Radar Point Cloud Classification

Classify point-cloud frames from radar sensors

Device support

33 MCU and wireless devices (as listed in the README of the 1.5.0 release) across these families. See Supported Devices for details.

  • C2000 (C28 DSP): F28P55 (with NPU), F28P65, F2837, F2838x, F28003, F28004 and others

  • C2000 (C29 DSP): F29H85, F29P58, F29P32

  • MSPM0: MSPM0G3507, MSPM0G3519, MSPM0G5187 (with NPU)

  • MSPM33C: MSPM33C32, MSPM33C34

  • AM13: AM13E2 (with NPU)

  • AM26x: AM263, AM263P, AM261

  • Connectivity: CC2755, CC2745, CC1312, CC1314, CC1352, CC1354, CC35X1

  • Radar: IWRL6432

Example applications and models

The examples/ folder ships ready-to-run applications (the generic ones for each task plus dedicated applications such as arc fault, motor bearing fault, ECG, wake word and Wi-Fi presence detection), and the model zoo contains 83 models. See Examples & Applications and Model Zoo Reference.

Model optimization

  • Quantization-aware training (QAT) and post-training quantization (PTQ)

  • Hessian-aware (HAWQ) automatic mixed-precision quantization

  • Residual connection support for NPU quantization

NPU acceleration

Select devices (F28P55, MSPM0G5187, AM13E2) include TI’s TinyEngine NPU for hardware-accelerated inference with faster execution and lower power. See NPU Configuration Guidelines for TI Neural Network Compiler.

1.1.4. What’s New in 1.5.0

  • Standalone pip install of tinyml-modelzoo; one repository to clone for model work

  • New tasks and devices: audio classification, radar point cloud classification, AM13E2 vision

  • Early stopping on by default, NaN-loss detection, memory estimation before training

  • macOS Apple Silicon (MPS) support, Python 3.14 / PyTorch 2.11 compatibility for training

  • Integer on-device training, learned filterbank front end, new example applications

See the release history in the README and the changelog in the appendix.

1.1.5. Repository Structure

tinyml-modelzoo/
├── examples/               # ready-to-run example configurations
├── tinyml_modelzoo/
│   ├── models/             # neural network model definitions
│   └── model_descriptions/ # model metadata (names used in configs)
├── docs/                   # guides
├── run_tinyml_modelzoo.sh / .bat / .ps1
└── ADDING_NEW_MODELS.md

The other parts of the toolchain are installed as packages, not cloned:

tinyml_modelmaker

Orchestrates the workflow: reads the config, runs training, testing and compilation.

tinyml_tinyverse

Core training infrastructure: datasets, training scripts, feature extraction transforms.

tinyml_torchmodelopt

Quantization toolkit (TINPU and generic wrappers, QAT/PTQ).

1.1.6. Workflow Summary

  1. Install: Installation

  2. Prepare data and a config: Bring Your Own Data, How to Create a ModelZoo Config YAML

  3. Run:

    ./run_tinyml_modelzoo.sh examples/generic_timeseries_classification/config.yaml

  4. Review results: Understanding the Run Output

  5. Deploy to a device: Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check)

1.1.7. Next Steps