9.1.5. Project Compilation Guide: Timeseries Classification Examples

9.1.5.1. Overview

This guide explains how any classification example from tinyml-modelzoo can be used on an MCU. We will understand the outputs from the modelzoo and their meaning. We will see how a generic timeseries classification project looks like. The guide will explain how to use the outputs from modelzoo in the CCS Project and finally run it on the device.

9.1.5.3. Required Tools

9.1.5.3.1. Code Composer Studio (CCS) IDE

TI’s integrated development environment. Download CCS from TI’s website. Version used in this guide: 20.4.0

9.1.5.3.2. Supported Devices and SDK Requirements

9.1.5.3.2.1. C2000 Family (F28x)

These devices use C2000Ware SDK. It is recommended to use the SDK Version: 6.1.0 (Release: 06_01_00_00)

Device	LaunchPad/EVM	Product Page	SDK Download
F28003x	LAUNCHXL-F280039C	TMS320F280039C	C2000Ware
F28004x	LAUNCHXL-F280049C	TMS320F280049C	C2000Ware
F28P55x	LAUNCHXL-F28P55X	TMS320F28P550SJ	C2000Ware
F28P65x	LAUNCHXL-F28P65X	TMS320F28P650DK	C2000Ware

9.1.5.3.2.2. C2000 Family (F29x)

These devices use F29H85x SDK. It is recommended to use the SDK Version: 1.04 (Release: 01_04_0000)

Device	LaunchPad/EVM	Product Page	SDK Download
F29H85x	F29H85X-SOM-EVM	F29H850TU	F29H85X-SDK

9.1.5.3.2.3. MSPM33 Family

These devices use MSPM33 SDK. It is recommended to use the SDK Version: 1.03.00.00 (Release: 1_03_00_00)

Device	LaunchPad/EVM	Product Page	SDK Download
MSPM33C321Ax	LP-MSPM33C321A	MSPM33C321A	MSPM33-SDK

9.1.5.3.2.4. Sitara MCU Family (AM13x)

These devices use MCU SDK. It is recommended to use the SDK Version: 1.00.00.00 (Release: 1_00_00_00)

Device	Product Page	SDK Download
AM13E2x	AM13E2	MCU-SDK

9.1.5.3.2.5. Sitara MCU Family (AM26x)

These devices use MCU-PLUS-SDK. It is recommended to use the SDK Version: 11.03.00.00 (Release: 11_03_00_00)

Device	LaunchPad/EVM	Product Page	SDK Download
AM263x	LP-AM263	AM2634	MCU-PLUS-SDK-AM263X

9.1.5.4. Output From Modelzoo

Modelzoo will start by loading the dataset, train the model, test the model, compile the model. Modelzoo will create the output folder in tinyml-modelmaker/data/projects/{dataset_name}. You can find the dataset name in the configuration yaml. Let’s assume the dataset_name to be hello_world_example_dsg as we will use the hello world example for generic timeseries classification.

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

9.1.5.4.1. Understanding the outputs

These files are located in the following directory structure:

tinyml-modelmaker/
└── data/
    └── projects/
        └── hello_world_example_dsg/
            └── run/
                └── {date-time}/
                    └── {model_name}/
                        ├── compilation/
                        │   └── artifacts/
                        │       ├── mod.a                     <-- Compiled model
                        │       └── tvmgen_default.h          <-- Model API header
                        │
                        └── training/
                            ├── base/                         <-- For float (non-quantized) models
                            │   └── golden_vectors/
                            │       ├── test_vector.c
                            │       └── user_input_config.h
                            │
                            └── quantization/                 <-- For quantized models
                                └── golden_vectors/
                                    ├── test_vector.c
                                    └── user_input_config.h

The example project has four useful file outputs by ModelMaker. We will see their name and meaning.

mod.a: The ONNX model is compiled by tvm to get C files, which are converted into a single mod.a that can run on device. This file stores the core functionality of AI model.
tvmgen_default.h: Mod.a exposes few APIs to interact with model which are present here. You can use these APIs in your application to run model
test_vector.c: ModelMaker gives a test dataset and the expected output. You can use the model to inference this test dataset and check if the output is matching.
user_input_config.h: This configuration file has preprocessing flag definitions for the parameters used for feature extraction.

These 4 files can be used in a CCS Project to perform AI on edge.

9.1.5.4.2. How are these obtained

The modelzoo is built on PyTorch framework and the training of the CNN model is done using PyTorch. After the training is converted we convert the layers in TI NPU understandable format. Then the AI model is exported in ONNX format using torch.onnx.export. The ONNX model can be found in the modelmaker run output tinyml-modelmaker/data/projects/hello_world_example_dsg/run/{date-time}/{model}/training/quantization/model.onnx.

This ONNX model is compiled using TI MCU NNC. You can go to the compiler page and check for different options and optimizations present. For different AI tasks such as Classification, Regression, Forecasting, Anomaly Detection, we use different optimizations. For classification example we use:

SKIP_NORMALIZE
OUTPUT_INT

To get the artifacts with mod.a and the tvmgen_default.h, we use TI MCU NNC and compile the onnx model. Below is the command to perform this. You can explore more compilation option.

tvmc compile --target="c, ti-npu type=hard skip_normalize=true output_int=true" --target-c-mcpu=c28 ./model.onnx -o artifacts_c28/mod.a --cross-compiler="cl2000" --cross-compiler-options="$CL2000_OPTIONS"

9.1.5.5. How to run on device

9.1.5.5.1. CCS Studio

Code Composer Studio (CCS) is a free integrated development environment (IDE) provided by Texas Instruments (TI) for developing and debugging applications for TI’s micro-controllers and processors. It offers various examples for users to get started with their problem statement. One of the application is f28p55x_generic_timeseries_classification. We will use this example to run on device f28p55x for this guide.

9.1.5.5.2. Requirements

The CCS example f28p55x_generic_timeseries_classification requires 4 files from modelmaker. We will copy the files from modelmaker run to the CCS example project.

C2000Ware 6.01.00.00
Location of example: C:\ti\c2000\C2000Ware_6_01_00_00\libraries\ai\examples\generic_timeseries_classification\f28p55x

Location of examples for other device families:

For C2000 Family (F29x) Example: C:\ti\f29h85x-sdk_1_04_00_00\examples\ai\generic_timeseries_classification
For Sitara MCU Family (AM13x) Example: C:\ti\mcu_sdk\examples\ai\generic_timeseries_classification
For Sitara MCU Family (AM26x) Example: C:\ti\mcu_plus_sdk_am263x_11_03_00_00\examples\ai\generic_timeseries_classification

9.1.5.5.3. Running for Target Device

After run the modelmaker from command line is finished. Copy the 4 files (path present below) from Modelmaker to CCS Project. Build the CCS Project, flash the program and start debugging the application. Check for the variable error for different sets of test cases preset in test_vector.c.

9.1.5.5.3.1. Compiled model files

mod.a: The compiled model is present in this file.
- Path Modelmaker: tinyml-modelmaker/data/projects/hello_world_example_dsg/run/{date-time}/{model}/compilation/artifacts/mod.a
- Path CCS Project: f28p55x_generic_timeseries_classification/artifacts/mod.a
tvmgen_default.h: Header file to access the model inference APIs from mod.a
- Path Modelmaker: tinyml-modelmaker/data/projects/hello_world_example_dsg/run/{date-time}/{model}/compilation/artifacts/tvmgen_default.h
- Path CCS Project: f28p55x_generic_timeseries_classification/artifacts/tvmgen_default.h

9.1.5.5.3.2. Test data for device verification

test_vector.c: Test cases to check if the model works on device currently
- Path Modelmaker: tinyml-modelmaker/data/projects/hello_world_example_dsg/run/{date-time}/{model}/training/quantization/golden_vectors/test_vector.c
- Path CCS Project: f28p55x_generic_timeseries_classification/test_vector.c
user_input_config.h: Configuration of feature extraction library in SDK.
- Path Modelmaker: tinyml-modelmaker/data/projects/hello_world_example_dsg/run/{date-time}/{model}/training/quantization/golden_vectors/user_input_config.h
- Path CCS Project: f28p55x_generic_timeseries_classification/user_input_config.h

9.1.5.5.4. Load sample example

We will load the generic timeseries example for f28p55 device using Code Composer Studio.

Open Code Composer Studio
Go to View tab -> Open Resource Explorer
Type the Board or Device to filter as LAUNCHXL-F28P55X
Type keyword as f28p55x_generic_timeseries_classification
Select the folder f28p55x_generic_timeseries_classification and click Import
Download and install the required dependencies of the project.
The imported project will look like this in the CCS Project

9.1.5.5.5. Run the sample example

We will build the project and flash the program in device. The project has a ‘C’ file application_main.c, which contains the code for calling APIs to the feature extraction lib and model inference. We will use debug mode to see the result of model inference present in test_result.

Now we will build the project. Go to Project Tab -> Select Build Project(s)
Switch the active target device from TMS320F28P550SJ9.ccxml to TMS320F28P550SJ9_LaunchPad.ccxml.
Connect launchpad F28P55x to your system.
Flash the built project in device with debug mode. Go to Run tab -> Select Debug Project
After the application is flashed, debug screen will appear. Place the breakpoint at the point where it is shown below.
Add the following variable in Watch section of debug and continue the program.
If the test_result is 1 it means the model inference is working correctly, if it is 0, the model inference is wrong.