Skip to content
BranchesTags

Repository files navigation

Tiny ML ModelZoo

Texas Instruments' central repository for AI models, examples, and configurations for microcontroller (MCU) applications. Clone this repo, install it, and run any example config against your target device — training, quantization, and compilation all happen automatically underneath.

Detailed User Guide: TI Tiny ML Tensorlab User Guide

tinyml-modelzoo/
├── examples/               # Ready-to-run example configurations
├── tinyml_modelzoo/
│   ├── models/             # Neural network model definitions
│   ├── model_descriptions/ # Model metadata for GUI integration
│   └── device_info/        # Target device performance data
├── run_tinyml_modelzoo.sh  # Training wrapper (Linux)
├── run_tinyml_modelzoo.bat # Training wrapper (Windows)
└── ADDING_NEW_MODELS.md    # Guide for adding custom models

Table of Contents

Quick Start

Prerequisites

  1. Python 3.10 environment
  2. Clone only this repository, then install it: 
    cd tinyml-modelzoo
pip install -e .
    This pulls in the rest of the toolchain as prebuilt wheels automatically - no need to clone anything else.
  3. (or) if you want to directly use this as a Python package and plan to make no model additions/changes, you might as well install the python package directly: 
    pip install http://software-dl.ti.com/C2000/esd/mcu_ai/wheel/tinyml_modelzoo-1.5.0-py3-none-any.whl

Running an Example

Linux:

cd tinyml-modelzoo
./run_tinyml_modelzoo.sh examples/generic_timeseries_classification/config.yaml

Windows (cmd):

cd tinyml-modelzoo
run_tinyml_modelzoo.bat examples\generic_timeseries_classification\config.yaml

Windows (PowerShell):

cd tinyml-modelzoo
./run_tinyml_modelzoo.ps1 examples/generic_timeseries_classification/config.yaml

What Happens When You Run an Example?

  1. Dataset Download - the required dataset is downloaded if not already present
  2. Data Processing - feature extraction and preprocessing are applied
  3. Model Training - the neural network is trained on your data
  4. Quantization - the model is optimized for MCU deployment
  5. Compilation - TI's Neural Network Compiler generates device-ready code

Output artifacts are saved to ./data/projects/<project_name>/, relative to whichever directory you ran run_tinyml_modelzoo.sh/.bat/.ps1 from. To use a different location, add this to the config's common section:

common:
    projects_path: './your/choice'  # or an absolute path

Supported Target Devices

AM13 (Arm Cortex-M33)

  • Devices with TinyEngine NPU (hardware accelerator): AM13E2

MSPM0 (Arm Cortex-M0+)

  • Devices with TinyEngine NPU (hardware accelerator): MSPM0G5187
  • MSPM0G3507, MSPM0G3519

Connectivity (Arm Cortex-M33/M4)

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

C2000 (C28 DSP)

  • Devices with TinyEngine NPU (hardware accelerator): F28P55
  • F2837, F2837xS, F2838x, F28P551x, F28003, F28004, F2807x, F28002x, F280013, F280015, F28E12, F28P65

C2000 (C29 DSP)

  • F29H85, F29P58, F29P32

MSPM33C (Arm Cortex-M33)

  • MSPM33C32, MSPM33C34

AM26x (Arm Cortex-R5)

  • AM263, AM263P, AM261

Radar (Arm Cortex-M4)

  • IWRL6432

Supported Task Categories

Task Category Description Use Cases
Time Series Classification Categorize time-series data into discrete classes Fault detection, activity recognition, anomaly classification
Time Series Regression Predict continuous values from time-series inputs Torque estimation, speed prediction, load measurement
Time Series Forecasting Predict future values based on historical patterns Temperature prediction, demand forecasting
Time Series Anomaly Detection Identify abnormal patterns using autoencoder-based models Equipment health monitoring, predictive maintenance
Audio Classification Classify audio signals from MFCC features Keyword spotting, voice commands, sound event detection
Image Classification Categorize images into classes Visual inspection, object recognition
Radar Point Cloud Classification Classify point-cloud frames from radar sensors Human pose detection, fall detection

For the reasoning behind how these categories differ from one another, see About the Task Types.

Choosing an Example

  1. Look for your use case in the Examples Reference tables below. If one matches (e.g. motor_bearing_fault, pir_detection), start there — it ships with a dataset, a tuned model, and a device-specific config.
  2. If nothing matches, use the generic example for your task type instead (the first row in each table below, e.g. generic_timeseries_classification) and point it at your own dataset. This is also the recommended first example to run to learn the toolchain.

Examples Reference

Each example links to its config directory under examples/. The first row in each table is the generic example (a generic_timeseries_* config, meant to be adapted to your own dataset); every other row is a dedicated, purpose-built config for that specific use case.

Classification

Example Data Type Description
generic_timeseries_classification — Classify sine/square/sawtooth waveforms. Start here to learn the toolchain.
dc_arc_fault Current Detect DC arc faults from current waveforms for electrical safety.
ac_arc_fault Current Detect AC arc faults in electrical systems.
motor_bearing_fault Vibration Classify 5 bearing fault types + normal operation from vibration data.
blower_imbalance Current Detect blade imbalance in HVAC blowers using 3-phase motor currents.
fan_blade_fault_classification Accelerometer Detect faults in BLDC fans from accelerometer data.
gearbox_fault_detection Vibration Classify gearbox operating conditions (healthy vs broken tooth) from vibration data.
grid_fault_detection Current Detect electrical grid faults from sensor data.
ecg_classification ECG Classify normal vs anomalous heartbeats from ECG signals.
pir_detection PIR Detect presence/motion using PIR sensor data.
fall_detection_classification Accelerometer Detect and classify Human Fall vs Activities of Daily Living (ADL).
dynamic_hand_gesture_recognition Accelerometer Classify 4 dynamic hand gestures (circle, wave, tap, other) from 3-axis accelerometer data.
electrical_fault Voltage/Current Classify transmission line faults using voltage and current (2-class and 6-class variants).
grid_stability Simulated grid parameters Predict power grid stability from node parameters.
gas_sensor Gas sensor array Identify gas type and concentration from sensor array data.
human_activity_recognition Accelerometer/Gyroscope Human Activity Recognition from accelerometer/gyroscope data.
nilm_appliance_usage_classification Voltage/Current Non-Intrusive Load Monitoring - identify active appliances.
PLAID_nilm_classification Voltage/Current Appliance identification using the PLAID dataset.
wifi_csi_presence_detection Wi-Fi CSI Device-free human presence detection from Wi-Fi Channel State Information.

Regression

Example Data Type Description
generic_timeseries_regression — Generic regression example for continuous value prediction.
bms_soc_estimation Voltage/Current/Temperature Estimate lithium-ion battery State of Charge (SOC) for battery management systems.
mosfet_temp_prediction Temperature/Power Predict MOSFET temperature from electrical parameters.
torque_measurement_regression Voltage/Current/Speed/Temperature Predict PMSM motor torque from current measurements.
induction_motor_speed_prediction Voltage/Current Predict induction motor speed from electrical signals.
washing_machine_load_weighing Voltage/Current/Speed Predict washing machine load weight.

Forecasting

Example Data Type Description
generic_timeseries_forecasting — Generic forecasting example for time series prediction.
forecasting_pmsm_rotor_temp Voltage/Current Forecast PMSM rotor winding temperature.
hvac_indoor_temp_forecast Temperature Predict indoor temperature for HVAC control.

Anomaly Detection

Example Data Type Description
generic_timeseries_anomalydetection — Generic anomaly detection example using autoencoders.
dc_arc_fault (DSI) Current Detect anomalous DC arc patterns using autoencoder (DSI dataset).
dc_arc_fault (DSK) Current Detect anomalous DC arc patterns using autoencoder (DSK dataset).
ecg_classification ECG Detect anomalous heartbeat patterns from ECG signals.
fan_blade_fault_classification Accelerometer Detect anomalous fan blade behavior from accelerometer data.
motor_bearing_fault Vibration Detect anomalous bearing behavior from vibration data.

Audio Classification

Example Data Type Description
google_speech_command Audio 12-class keyword spotting from audio using MFCC + DSCNN model.
cough_detection Audio Binary cough vs. other-sound detection using LPC features + ResNet model.
glass_break_detection Audio Detect glass-breaking acoustic events for security/home-automation systems using FFT features + DSCNN model.
wake_word_detection Audio Detect the "OK Kilby" wake word using a learned-filterbank front-end + TCDS-ResNet model.

Image Classification

Example Data Type Description
MNIST_image_classification Image Handwritten digit recognition (MNIST dataset).
machine_readable_code_classification Image Classify QR codes, barcodes, and other symbols (28×28 images).
coffee_bean_classification Image Classify coffee bean quality from images.

Radar Point Cloud Classification

Example Data Type Description
radar_pose_and_fall_detection Radar point cloud Human pose and fall detection from radar point-cloud frames.

Available Models

Models are organized by task type. The NPU column indicates hardware acceleration support on TI devices with NPU (F28P55, AM13E2, MSPM0G5187).

NPU-optimized models follow specific layer constraints for hardware acceleration:

  • All channels are multiples of 4 (m4)
  • Kernel heights ≤ 7 for GCONV layers
  • MaxPool kernels ≤ 4
  • FC layer inputs ≥ 16 features (8-bit) or ≥ 8 features (4-bit)

For detailed guidelines, see NPU Configuration Guidelines.

When to use NPU-optimized models:

  • Target device has NPU (F28P55, AM13E2, MSPM0G5187)
  • You need maximum inference speed
  • Standard models show "fallback to software" warnings during compilation

Classification Models

Model Name Parameters Architecture NPU Description
CLS_100_NPU ~100 CNN Yes Ultra-compact model
CLS_500_NPU ~500 CNN Yes Compact model
CLS_1k_NPU ~1K CNN Yes Lightweight 2-layer CNN
CLS_1.2k_NPU ~1.2K CNN Yes Compact model for ultra-low power devices
CLS_1.5k_NPU ~1.5K CNN Yes 3-layer model with balanced performance
CLS_1.9k_NPU ~1.9K CNN Yes Efficient 3-layer model
CLS_2k_NPU ~2K CNN Yes 2-layer model
CLS_2.8k_NPU ~2.8K CNN Yes Improved accuracy with compact footprint
CLS_3.1k_NPU ~3.1K CNN Yes Higher accuracy model
CLS_ResAdd_3k ~3K ResNet (Add) No Residual connections with addition
CLS_ResCat_3k ~3K ResNet (Cat) No Residual connections with concatenation
CLS_3.9k_NPU ~3.9K CNN Yes Advanced 3-layer model
CLS_4k_NPU ~4K CNN Yes Balanced model
CLS_4.2k_NPU ~4.2K CNN Yes Optimized 4-layer model
CLS_5k_NPU ~5K CNN Yes Mid-range model
CLS_6k_NPU ~6K CNN (DW-Sep) Yes Depthwise separable
CLS_8k_NPU ~8K CNN (DW-Sep) Yes Depthwise separable
CLS_13k_NPU ~13K CNN Yes Higher capacity
CLS_20k_NPU ~20K CNN Yes High capacity
CLS_24k_NPU ~24K CNN Yes Supports integer on-device learning (ODL) fine-tuning
CLS_40k_NPU ~40K CNN Yes Advanced model for complex tasks
CLS_55k_NPU ~55K CNN Yes Maximum accuracy
ArcFault_model_200_t ~200 Specialized No Arc fault detection
ArcFault_model_300_t ~300 Specialized No Arc fault with more capacity
ArcFault_model_700_t ~700 Specialized No Arc fault medium model
ArcFault_model_1400_t ~1.4K Specialized No Arc fault high accuracy
GearboxFault_model_1.2k_t ~1.2K CNN Yes Gearbox fault detection
GearboxFault_model_1.5k_t ~1.5K CNN Yes Gearbox fault with more capacity
MotorFault_model_1_t Varies Specialized No Motor bearing fault detection
MotorFault_model_2_t Varies Specialized No Motor fault variant 2
MotorFault_model_3_t Varies Specialized No Motor fault variant 3
FanImbalance_model_1_t Varies Specialized No Fan blade imbalance detection
FanImbalance_model_2_t Varies Specialized No Fan imbalance variant 2
FanImbalance_model_3_t Varies Specialized No Fan imbalance variant 3
PIRDetection_model_1_t Varies Specialized No PIR-based presence detection
SimpleCNN2D_BN_t ~3K CNN No Wi-Fi CSI presence detection (2-layer 2D CNN over time-frequency features). CC35X1 only.

Regression Models

Model Name Parameters Architecture NPU Description
REGR_500_NPU ~500 CNN Yes Compact regression
REGR_1k ~1K CNN No Lightweight regression model
REGR_2k_NPU ~2K CNN Yes 2-layer model
REGR_3k ~3K MLP No 4-layer fully connected network
REGR_4k ~4K CNN No 2 Conv+BN+ReLU + Linear
REGR_6k_NPU ~6K CNN (DW-Sep) Yes Depthwise separable convolutions
REGR_8k_NPU ~8K CNN Yes 3-layer model
REGR_10k ~10K CNN No 3 Conv+BN+ReLU + 2 Linear
REGR_13k ~13K CNN No High capacity regression
REGR_20k_NPU ~20K CNN Yes High capacity with MaxPool

Anomaly Detection Models

Note: For NPU models, encoder convolutions are NPU-accelerated but decoder upsampling falls back to CPU.

Model Name Parameters Architecture NPU Description
AD_500_NPU ~500 CNN AE Yes 2-layer autoencoder
AD_1k ~1K Autoencoder No Compact autoencoder
AD_2k_NPU ~2K CNN AE Yes 2-layer autoencoder
AD_4k ~4K Autoencoder No 3-layer CNN autoencoder
AD_6k_NPU ~6K CNN AE (DW-Sep) Yes Depthwise separable encoder
AD_8k_NPU ~8K CNN AE Yes 3-layer autoencoder
AD_10k_NPU ~10K CNN AE Yes 3-layer autoencoder
AD_16k ~16K Autoencoder No 4-layer CNN autoencoder
AD_17k ~17K Autoencoder No Fan blade anomaly detection
AD_20k_NPU ~20K CNN AE Yes High capacity autoencoder
AD_Linear Varies Linear AE No 3-layer deep linear autoencoder
Ondevice_Trainable_AD_Linear Varies Linear AE No On-device trainable variant

Forecasting Models

Note: LSTM models are not NPU-supported.

Model Name Parameters Architecture NPU Description
FCST_500_NPU ~500 CNN Yes Compact forecasting
FCST_1k_NPU ~1K CNN Yes 2-layer model
FCST_2k_NPU ~2K CNN Yes 2-layer model
FCST_3k ~3K MLP No 4-layer fully connected
FCST_4k_NPU ~4K CNN Yes 3-layer model
FCST_6k_NPU ~6K CNN (DW-Sep) Yes Depthwise separable convolutions
FCST_8k_NPU ~8K CNN Yes 3-layer model
FCST_10k_NPU ~10K CNN Yes 3-layer model
FCST_13k ~13K CNN No 2 Conv+BN+ReLU + Linear
FCST_20k_NPU ~20K CNN Yes High capacity with MaxPool
FCST_LSTM8 Varies LSTM No Single LSTM (hidden=8) + Linear
FCST_LSTM10 Varies LSTM No Single LSTM (hidden=10) + Linear

Audio Classification Models

Model Name Parameters Architecture NPU Description
DSCNN_NPU ~9K DSCNN Yes Depthwise separable CNN for keyword spotting; input (1, 49, 10) MFCC
TCDS_ResNet_NPU ~24K Temporal Channel-Decoupled Separable CNN Yes Cough detection from LPC features; input (1, 100, 70), SRAM-efficient (no 2D spatial buffers)
DSCNN_GB_NPU ~6K DSCNN Yes Glass break detection from FFT-binned features; input (1, 86, 32), mixed-precision (4-bit DW/PW, 8-bit stem)
TCDS_ResNet_FB_NPU Varies Learned Filterbank + TCDS-ResNet Yes Wake word detection directly from raw audio via a jointly-trained Conv1D filterbank front-end; mixed-precision (2-bit backbone, 8-bit stem/head)

Image Classification Models

Model Name Parameters Architecture NPU Description
Lenet5 ~60K LeNet-5 No Classic CNN for image classification
MobileNetV1_58k_NPU ~58K MobileNetV1-style DW-Sep Yes Compact NPU-optimized image classifier
MobileNetV2_58k_NPU ~58K MobileNetV2-style DW-Sep Yes Inverted residual image classifier

Radar Point Cloud Classification Models

Model Name Parameters Architecture NPU Description
Pose_and_Fall_model Varies Linear (4-layer) No Human pose and fall detection from radar point-cloud data

Adding New Models

Want to add your own model? See the comprehensive guide: ADDING_NEW_MODELS.md

Key steps:

  1. Add model class to tinyml_modelzoo/models/
  2. Add class name to the file's __all__ list
  3. (Optional) Add device performance info to device_info/run_info.py
  4. (Optional) Add model description to model_descriptions/ for GUI integration

No changes required outside this repo.

About the Task Types

Classification outputs a probability distribution over predefined classes. Best for: "Is this an A fault, B fault, or C fault?", "Which type of activity is this?"

Regression outputs a continuous numerical value. Best for: "What is the current torque?", "What will the temperature be?"

Forecasting predicts future values in a time series. Best for: "What will happen next?"

Anomaly Detection uses autoencoders to learn "normal" patterns; reconstruction error indicates anomalies. Best for: "Is this behavior normal?"

Audio Classification extracts MFCC features from a fixed-length audio window and classifies into keyword or sound categories. Best for: "What keyword was spoken?", "What sound event occurred?"

These categories can look similar from a distance, so here's how to tell them apart:

  • Anomaly Detection vs. Classification — "Is it normal, or an anomaly?" is anomaly detection (binary outcome). "Is it normal, anomaly type A, type B, or type C?" is classification (multiple categories).
  • Classification vs. Regression — predicting a discrete target (Class A / B / C, ...) from independent variables is classification; predicting a continuous target is regression.
  • Regression vs. Forecasting — predicting a continuous target Y at the same time instant as its inputs is regression; predicting a variable's value at a future time instant is forecasting.

Release History

  • [2026-Sep] Release version 1.5.0 of the software

    Details
    • Until 1.4.0 (2026-Jun), this history lived in the tinyml-tensorlab repository on GitHub; that repository's source is now private. Starting with 1.5.0 (2026-Sep), tinyml-modelzoo (this repo) is the standalone, pip-installable entry point to TI's MCU AI flow — a simplified, easy-to-install replacement for what previously required cloning tinyml-tensorlab's full set of component repos.
    • Device Support: 33 MCU/wireless devices supported
      • Added AM13E2 support for vision classification and added radar tasks
      • Added fel_memory support (feature-extraction library) for AM13 and C28x (F280013x, F280015x, F28002x) devices — enables on-device RAM/Flash estimation
      • Gen3 F28x device support reconciled: F2838x, F28P551x, F28002x, and new F28E12x device profile added across regression, forecasting, and anomaly-detection models; over-broad device lists tightened (e.g. arc fault, generic timeseries classification) to match verified per-model support
      • Task-level target device lists for radar, image, and audio classification are now derived from each model's real device support instead of a hand-maintained list, fixing several tasks (e.g. radar classification) that had advertised devices no model actually supported
    • Applications Supported: 37 example applications (4 generic timeseries tasks + 33 specific applications)
    • Models: 78 models across classification/regression/forecasting/anomaly-detection (timeseries + radar) plus vision and audio classification
      • Added TCDSResNet and a compressed DSCNN for audio classification (cough detection), both NPU-compliant
      • Compressed MobileNet_v1 to fit AM13 memory budget
      • Added Radar Point Cloud Classification support (new model + dataset flow )
      • Added a learned-filterbank front-end (jointly trained Conv1D bank) paired with TCDS-ResNet for wake word detection, and a compact DSCNN variant for glass break detection, both NPU-compliant with mixed-precision quantization
      • Added a 24K-parameter NPU classification model with integer on-device learning (ODL) support, plus new integer ODL documentation
      • New example applications: cough detection (audio), WiFi CSI presence detection, Google speech command, glass break detection (audio), wake word detection (audio), BMS State-of-Charge estimation (regression)
      • Re-enabled and renamed the washing machine load weighing regression example (washing_machine_load_weighing, was reg_washing_machine)
    • Flows:
      • Added early stopping for training (patience + min-delta, on by default) across all task types — timeseries classification/regression/forecasting/anomaly-detection and image classification
      • ModelMaker can now report estimated RAM & Flash usage for Feature Extraction before training, including a standalone estimate_memory CLI that runs without a full training pass
      • Support for preprocessing and publicly-available datasets in audio/radar flows
      • NaN-loss detection during training to catch instability early
    • Model Optimization:
      • Residual connection support added to quantization (TINPU)
      • Replaced qconfig_dict with a cleaner API to toggle auto-quantization on/off
      • Hessian-aware (HAWQ) automatic mixed-precision quantization, plus bug fixes to the qconfig mapping and mixed-precision paths
      • Automated Ternary-weight / 8-bit-activation QAT support
      • NPU quantization enabled for anomaly-detection models
    • Reliability & Compatibility:
      • Python 3.14 / PyTorch 2.11 compatibility
      • macOS ARM64 (MPS/Apple Silicon) compatibility fixes across training, evaluation, and quantization
      • torch.compile safety: unwraps compiled models correctly before ONNX/checkpoint export, falls back to eager on failure
      • Fixed ONNX export crashing inside PyInstaller/frozen builds
      • Security/robustness hardening: safe checkpoint deserialization, safe YAML loading, safer cache-dataset handling
      • Training performance: torch.compile, AMP, persistent workers, more efficient eval loop
      • Large expansion of automated test coverage (functional test tiers, pytest suites) across all repos
    • Packaging:
      • Added build_wheels.sh to build TinyVerse/ModelOptimization/ModelMaker wheels from local source
      • ModelZoo and TinyVerse examples now runnable standalone via published ModelMaker wheel dependency (TI official wheel CDN)
      • tinyml-tensorlab deprecated as the public entry point; tinyml-modelzoo takes over as the standalone, pip-installable way to use TI's MCU AI flow
    • Special Acknowledgement:
      • Shoutout to @musicalplatypus for contributing towards a better, neater and more feature-rich toolchain by their additions such as full macOS/Apple Silicon (MPS) support, torch.compile+AMP training-performance optimizations, and NAS bug fixes. They also hardened the codebase with security fixes for unsafe deserialization/YAML loading, overhauled CI so tests actually run across all four packages, and expanded the test suite and architecture docs.

  • [2026-Jun] Release version 1.4.0 of the software

    Details
    • Agent Skills with Claude Code supported for users to solve Edge AI/Tiny ML problems using natural language!
    • Device Support: 40 MCU devices supported:
      • AM1x: AM13E2
      • C2000 F28: F280013, F280015, F28003, F28004, F2837, F28P55, F28P65
      • C2000 F29: F29H85, F29P58, F29P32
      • MSP M0: MSPM0G3507, MSPM0G3519, MSPM0G5187
      • MSP M33: MSPM33C32,
      • Connectivity: CC2755, CC1352, CC1354, CC35X1, CC1312, CC1314
      • AM26x: AM263, AM263P, AM261
    • Flows:
      • Timeseries Anomaly Detection flow - More models
      • On Device Learning Mode Enabled - Expansive functionalities
    • Applications Supported
      • 31 (4 generic + 27 specific applications)
    • Models:
      • 50+ generic models added over classification, regression, forecasting and anomaly detection tasks.
    • Model Optimization:
      • Hessian Aware Quantization for automatic recommendation of quantization bitwidths for weights.
    • Compilation:
      • Upgraded TI MCU Neural Network Compiler for MCUs to 2.1.2

  • [2026-Feb] Release version 1.3.0 of the software

    Details
    • Device Support: 22 MCU devices supported:
      • AM1x: AM13E2
      • C2000 F28: F280013, F280015, F28003, F28004, F2837, F28P55, F28P65
      • C2000 F29: F29H85, F29P58, F29P32
      • MSP M0: MSPM0G3507, MSPM0G3519, MSPM0G5187
      • MSP M33: MSPM33C32,
      • Connectivity: CC2755, CC1352, CC1354, CC35X1,
      • AM26x: AM263, AM263P, AM261
    • Flows:
      • Timeseries Anomaly Detection flow supported
      • On Device Learning Mode Enabled
    • Applications Supported
      • 22 (4 generic + 18 specific applications)
    • Models:
      • 50+ generic models added over classification, regression, forecasting and anomaly detection tasks.
    • Model Optimization:
      • Partial Quantization Supported to enable best of precision and latency for regression models.
    • Compilation:
      • Upgraded TI MCU Neural Network Compiler for MCUs to 2.1.1 LTS

  • [2025-Nov] Release version 1.2.0 of the software

    Details
    • Device Support:
      • Added MSPM0 based MCUs: MSPM0G3507, MSPM0G5187
      • Added Connectivity device: CC2745R10-Q1, CC2755R10
    • General:
      • Supports simple gain augmentation for classification tasks
      • Prints dataset file level confusion matrix for classification tasks
      • Golden Test Vectors for Regression tasks
      • Run modelmaker with only the config, no more target device required in the input.
    • Flows:
      • Timeseries Forecasting flows supported
      • L1, L2 normalization can be enabled in regression using lambda_reg param
    • Model Optimization:
      • How to use: Documentation updated.
      • Example code for performing regression in modeloptimization
      • Fixing clipping of input data to int8 or uint8 based on dataset (zero_point) (only the input zero point is fixed and not the intermediate layers)
      • BatchNorm is supported by GENERIC quantization
      • Experimental features like additional QDQ at input of model and floating bias can be enabled individually
      • Residual Add supported for different scales, zero points, but not optimised for TINPU
    • Compilation:
      • Upgraded TI MCU Neural Network Compiler for MCUs to 2.1.0 LTS
      • Supported all layer configs with 8-bit activations and 8-/4-/2-bit weights that can be offloaded to TI-NPU
      • Supported all layer configs with 8-bit activations and 8-bit weights that can be accelerated using the M33 Custom Datapath Extension (CDE).

  • [2025-Aug] Release version 1.1.0 of the software

    Details
    • General:
      • Generic Timeseries Classification is available with fixed point reference dataset.
      • Compatible with C2000Ware 6.0.0
    • Model Optimization:
      • Aggressive Quantization Modes for Weights & Activation: 2W8A, 4W4A, 4W8A --> massive speedup and memory saved
      • Neural network Architecture Search for generating a TINPU compatible model directly based on user's dataset
    • Dataset:
      • Dataset can be split into train-test-val on a file-by-file basis or within-a-file basis
    • Device Support:
      • Full Support for F280013x
      • Preliminary Support for F29H85x and MSPM0G3507x
    • Compilation:
      • Upgraded TI MCU Neural Network Compiler for MCUs to 2.0.0
    • Windows Platform Specific:
      • Major quantization accuracy improvements
    • Miscellaneous:
      • Fixed model performance data that appears on the terminal when a training is initiated
      • Added Model Descriptions for all models
      • Setup of the repos is now smoother and cleaner

  • [2025-Apr] Major feature updates (version 1.0.0) of the software

    Details
    • General:
      • Tiny ML Modelmaker is now a pip installable package!
      • Existing models can be modified on the fly through a config file (check Tiny ML Modelmaker docs)
      • MPS (Metal Performance Shaders) backend support for Mac host devices!
    • Technology:
      • PTQ and QAT flows supported in tinyml-modelmaker, tinyml-modeloptimization
      • Ternary, 4 bit Quantization support in tinyml-modelmaker
    • Flows:
      • Regression ML tasks supported
      • Autoencoder based Anomaly Detection task supported
    • Feature Extraction:
      • Feature Extraction transforms are now modular and compatible with C2000Ware 5.05 only
      • Supports Haar and Hadamard Transform
      • Golden test vectors file has one set uncommented by default to work OOB
    • Data Visualisation:
      • Multiclass ROC-AUC graphs are autogenerated for better explainability of reports and help select thresholds based on false alarm/ sensitivity preference
      • PCA graphs are auto plotted for feature extracted data - Helps in identifying if the feature extraction actually helped
      • Run now begins with displaying inference time, sram usage and flash usage for all the devices for any model.
    • Dataset
      • Goodness of Fit of dataset now enabled.
    • Extensive Documentation & Know-How Examples to use Modelmaker

  • [2024-November] Updated (version 0.9.0) of the software

  • [2024-August] Release version 0.8.0 of the software

  • [2024-July] Release version 0.7.0 of the software

  • [2024-June] Release version 0.6.0 of the software

  • [2024-May] First public release (version 0.5.0) of the software

Additional Resources

License

BSD 3-Clause License. See LICENSE for details.

About

Texas Instruments' Modelzoo for Edge AI/ML on Microcontrollers

Resources

Readme
Activity
Custom properties

Stars

1 star

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages