This section describes the procedures to upload data, train, and compile models through Edge-AI Model Composer. A command line tool through TinyML Model Maker is available for customers to fine-tune model parameters and modify the models.

This demo features four different sizes of generic time series neural network models that can be used for a wide range of time series classification tasks.

Steps to Use Model Composer:

1. Login to https://dev.ti.com/modelcomposer/
2. Click “Example Project”, under Task select “Generic Time Series”, under Tools select “MCU Analytics Backend v1.0.1", under Sample Dataset select “hello_world_example_dsg”. Give the project a name and then click “New Project”.
   Figure 4-7 New Example Screenshot
3. The “Capture” tab is used for data visualization, labeling, and importing. The tabs at the top of the page can be used to iterate through the project. Selecting a file on the left menu brings up a visualization of that data file. Options for rendering are shown on the right side. Additional data can be imported using the “Import Data” button. This button can also be used to import your own dataset when creating a new generic time series project.
   Figure 4-8 Capture Tab
4. Go to the model selection tab by using the link on the top of the page and select target device and model. The slider can be used to select a recommended device or the user can manually select a device. For this example, it is recommend using the F28P55 that uses TI’s neural processing unit (NPU).
   Figure 4-9 Model Selection Tab
5. Models included are all generic time series models. In example, the 1k_t is the smallest model, for more simple classification tasks. The 13k_t is the largest for more complex tasks.
6. The flash and inference time estimates are shown below the model selection option. These can be used to determine the right model and device for your application.
7. Navigate to the training page using the links at the top and select your pre-processing options.