13.3. Changelog

Version history and release notes for Tiny ML ModelZoo (until 1.4.0 published as Tiny ML Tensorlab).

13.3.1. Version 1.5.0

Current Release (2026-Sep)

Packaging and entry point:

  • tinyml-modelzoo is now the standalone, pip-installable entry point to TI’s MCU AI flow. See Installation.

  • tinyml-tensorlab is deprecated as the public entry point (its source is now private). Its history is kept in the older entries below.

  • ModelZoo and TinyVerse examples run standalone through the published ModelMaker wheel dependency (TI wheel CDN).

  • build_wheels.sh added to build the TinyVerse, ModelOptimization and ModelMaker wheels from local source.

Device Support (33 MCU/wireless devices):

  • AM13E2: vision classification added, plus radar tasks

  • fel_memory (feature-extraction library) support for AM13 and C28x (F280013x, F280015x, F28002x) devices, which enables on-device RAM/Flash estimation

  • Gen3 F28x support reconciled: F2838x, F28P551x, F28002x and a new F28E12x device profile across regression, forecasting and anomaly detection. Over-broad device lists (for example arc fault and generic timeseries classification) were tightened to the verified per-model support.

  • Target device lists for radar, image and audio classification are now derived from each model’s real device support

Applications (37 examples: 4 generic timeseries tasks + 33 specific applications):

  • New: cough detection (audio), WiFi CSI presence detection, Google speech command, glass break detection (audio), wake word detection (audio), BMS State-of-Charge estimation (regression)

  • Washing machine load weighing regression example re-enabled and renamed to washing_machine_load_weighing (was reg_washing_machine)

Models (83 across classification, regression, forecasting and anomaly detection, timeseries and radar, plus vision and audio classification):

  • TCDSResNet and a compressed DSCNN for audio classification (cough detection), both NPU-compliant

  • MobileNet_v1 compressed to fit the AM13 memory budget

  • Radar Point Cloud Classification (new model and dataset flow)

  • Learned-filterbank front-end (jointly trained Conv1D bank) with TCDS-ResNet for wake word detection, and a compact DSCNN variant for glass break detection, both NPU-compliant with mixed-precision quantization

  • A 24K-parameter NPU classification model with integer on-device learning (ODL) support, plus new integer ODL documentation. See Integer On-Device Training Documentation.

See Model Zoo Reference for the model list.

Flows:

  • Early stopping for training (patience and min-delta, on by default) across all task types

  • ModelMaker can report estimated RAM and Flash usage for feature extraction before training, including a standalone estimate_memory CLI that runs without a full training pass

  • Preprocessing and publicly available datasets supported in the audio and radar flows

  • NaN-loss detection during training to catch instability early

Model Optimization:

  • Residual connection support in quantization (TINPU)

  • qconfig_dict replaced by a cleaner API to toggle auto-quantization on and off

  • Hessian-aware (HAWQ) automatic mixed-precision quantization, with fixes to the qconfig mapping and mixed-precision paths

  • Automated ternary-weight / 8-bit-activation QAT

  • NPU quantization enabled for anomaly detection models

Reliability and compatibility:

  • Python 3.14 / PyTorch 2.11 compatibility

  • macOS ARM64 (MPS, Apple Silicon) fixes across training, evaluation and quantization

  • torch.compile safety: compiled models are unwrapped before ONNX/checkpoint export, with eager fallback on failure

  • ONNX export no longer crashes inside PyInstaller/frozen builds

  • Security hardening: safe checkpoint deserialization, safe YAML loading, safer cache-dataset handling

  • Training performance: torch.compile, AMP, persistent workers, more efficient evaluation loop

  • Much larger automated test coverage across all repositories

Acknowledgement:

  • Thanks to @musicalplatypus for macOS/Apple Silicon (MPS) support, training-performance work, NAS bug fixes, security fixes, CI and test-suite improvements.

13.3.2. Version 1.4.0

New Applications:

  • Gearbox Fault Detection — 2-class vibration classification for MSPM0G5187 NPU

  • Electrical Fault 6-Class — fault type classification (G/C/B/A combinations)

Device Support (23 MCUs):

  • CC1312 — added support for Fan Blade Fault Classification and PIR Detection

  • CC1314, CC1352, CC1354 — added support for Fan Blade Fault Classification

  • CC2755, CC35X1 — expanded support for Fan Blade Fault Classification

  • MSPM0G5187 — added PMSM Rotor Forecasting config

Model Optimization:

  • LSQ (Learned Step-size Quantization) observer for weights and activations

  • MLP models with 2D input now fully supported for quantization and conversion

Compilation:

  • TI MCU Neural Network Compiler upgraded to 2.1.2

Platform:

  • macOS 12 or later officially supported

Other Changes:

  • Timeseries Anomaly Detection re-enabled in default installation

  • Reconstruction error plot description added to anomaly detection output

  • F3 and WiFi SDK device descriptions updated

  • AM13 SDK entry added to connectivity dependencies

  • MSPM0L3 (M33) SDK version updated

13.3.3. Version 1.3.0

New Features:

  • Added MSPM0G5187 NPU support

  • New anomaly detection autoencoder architectures

  • Enhanced Neural Architecture Search (NAS) algorithms

  • Improved quantization-aware training (QAT)

  • Edge AI Studio Model Composer integration

  • Documentation overhaul with Sphinx

Models:

  • Added CLS_20k_NPU and CLS_55k_NPU for complex tasks

  • New forecasting models (FCST_* family)

  • Updated motor fault models

Bug Fixes:

  • Fixed INT4 quantization issues on F28P55

  • Resolved memory allocation errors in large models

  • Fixed GoF test visualization on Windows

Breaking Changes:

  • Configuration file format updated for data_processing_feature_extraction

  • Model registry API changed (use MODEL_REGISTRY dict)

13.3.4. Version 1.1.0

New Features:

  • AM13E2 device support

  • Time series forecasting task type

  • Multi-variable input support

  • Goodness of Fit (GoF) test improvements

  • Windows native support (without WSL)

Models:

  • Added REGR_* regression model family

  • New AD_* anomaly detection models

  • Updated ArcFault models for better accuracy

Improvements:

  • Faster compilation for NPU devices

  • Reduced memory usage during training

  • Better error messages

13.3.5. Version 1.0.0

Initial Release

Features:

  • Support for 20+ TI microcontrollers

  • Time series classification, regression, anomaly detection

  • NPU support for F28P55

  • Quantization (PTQ and QAT)

  • Neural Architecture Search

  • Feature extraction presets

  • Post-training analysis tools

Supported Devices:

  • C2000 family (F28P55, F28P65, F2837, etc.)

  • MSPM0 family (MSPM0G3507, MSPM0G3519)

  • AM26x family (AM263, AM263P, AM261)

  • Connectivity devices (CC2755, CC1352, CC1354, CC35X1)

Models:

  • Classification models (CLS_100 through CLS_13k)

  • NPU models (CLS_*_NPU variants)

  • Arc fault detection models

  • Motor fault detection models

13.3.6. Migration Guides

13.3.6.1. Migrating from 1.1.x to 1.2.x

Configuration Changes:

Old format:

feature_extraction:
  preset: 'my_preset'

New format:

data_processing_feature_extraction:
  feature_extraction_name: 'my_preset'

Model Names:

Some model names have been updated for consistency:

Old: TimeSeries_Generic_1k_t
New: CLS_1k

API Changes:

Model access changed:

# Old
from models import get_model
model = get_model('CLS_1k')

# New
from tinyml_modelzoo.models import get_model
model = get_model('CLS_1k', variables=1, num_classes=2, input_features=128)

13.3.7. Deprecation Notices

Version 1.3.0:

  • feature_extraction config section renamed to data_processing_feature_extraction

  • Old preset names deprecated (removed in 1.4.0)

  • Python 3.9 support dropped

  • Legacy model names removed

13.3.8. Known Issues

Current:

  • Windows: Long path names may cause issues

  • Large models (>50k params): May require increased stack size

  • QAT: Minor accuracy variations between runs

Workarounds:

  • Windows: Use shorter project paths

  • Large models: Adjust linker settings for stack

  • QAT: Set random seed for reproducibility

13.3.9. Roadmap

Planned Features:

  • Additional MCU family support

  • Improved image classification capabilities

  • Advanced hyperparameter tuning

  • Cloud training integration

  • TensorFlow Lite model import

Community Requests:

  • RNN/LSTM support (under evaluation)

  • Multi-task learning

  • Model ensemble support

13.3.10. Contributing

Tiny ML Tensorlab welcomes contributions:

See CONTRIBUTING.md for guidelines.