12.2. FAQ

Frequently asked questions about Tiny ML ModelZoo 1.5.0.

12.2.1. General

What is Tiny ML ModelZoo?

The standalone, pip-installable entry point to TI’s MCU AI flow: model zoo, example applications, and (through prebuilt wheels) ModelMaker, TinyVerse and ModelOptimization. In 1.4.0 and earlier this was the separate tinyml-tensorlab project, which is now deprecated as the entry point. See Introduction.

Which devices are supported?

33 MCU and wireless devices in 1.5.0 (see the release notes in Changelog and Supported Devices). Devices with a hardware NPU are F28P55, AM13E2 and MSPM0G5187.

Do I need a GPU?

No. CPU training works for the small models in the zoo. With CUDA available, torch.compile and AMP are enabled automatically unless you set compile_model / native_amp yourself or num_gpus: 0. On Apple Silicon the MPS backend is supported.

Which operating systems are supported?

Linux, Windows and macOS. See Installation.

12.2.2. Installation

Do I need to clone several repositories?

No. Clone only tinyml-modelzoo and run pip install -e ., or install the wheel if you do not change models. The other components come as prebuilt wheels. See Installation.

Which Python version?

Python 3.10 for the full flow. Newer versions (up to 3.14) can train, but the TI compiler ti_mcu_nnc is only available for 3.10. See Python 3.11 or newer cannot compile.

Do I need the TI compilers to train?

No. They are only needed for the compilation step. Set TOOLS_PATH (default ~/ti) when you compile. See Compiler tools not found (TOOLS_PATH).

I installed the wheel. Where are the examples?

The examples are in the Git repository. See No examples/ folder after a wheel install.

12.2.3. Models and training

Which model should I choose?

Choose a model that lists your target device in Model Zoo Reference. For NPU devices prefer models ending in _NPU. Smaller models fit smaller devices; start with the example for your task.

Why did training stop before the last epoch?

Early stopping is on by default. See Training stops before all epochs and Early Stopping.

Can I bring my own model?

Yes. See Bring Your Own Model (BYOM). For your own dataset see Bring Your Own Data.

Where are the results?

Under ./data/projects/<project_name>/run/<date-time>/<model_name>/ relative to where you started the run (configurable with common.projects_path). See Understanding the Run Output.

How do I estimate memory before training?

Use the estimate_memory command described in Memory Estimation.

12.2.4. Deployment

How do I put the model on the device?

Copy the generated files into an SDK example. See Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check) and Deployment. If the device output differs from the golden vectors, see Golden vector mismatch on the device.