9. Device Deployment

This section guides you through deploying trained models on TI microcontrollers using Code Composer Studio (CCS).

Contents

9.4. Supported Devices

Device-Specific SDKs

Device Family	Devices	SDK	SDK Version	Download
C2000 (F28x)	F28003x, F28004x, F28P55x, F28P65x	C2000Ware	6.01.00.00	C2000Ware
C2000 (F29x)	F29H85x, F29P58x, F29P32x	F29H85X SDK	01.04.00.00	F29H85X-SDK
MSPM33	MSPM33C321Ax	MSPM33 SDK	1.03.00.00	MSPM33-SDK
Sitara MCU (AM13x)	AM13E2x	MCU SDK	1.00.00.00	MCU-SDK-AM13E2X
Sitara MCU (AM26x)	AM263, AM263P, AM261	MCU-PLUS-SDK	11.03.00.00	MCU-PLUS-SDK-AM263X
Connectivity	CC2755, CC1352, CC1354, CC35X1	SimpleLink SDK	–	SimpleLink

Supported LaunchPads/EVMs

Device	LaunchPad/EVM	Product Page	NPU
F28003x	LAUNCHXL-F280039C	TMS320F280039C	No
F28004x	LAUNCHXL-F280049C	TMS320F280049C	No
F28P55x	LAUNCHXL-F28P55X	TMS320F28P550SJ	Yes
F28P65x	LAUNCHXL-F28P65X	TMS320F28P650DK	No
F29H85x	F29H85X-SOM-EVM	F29H850TU	No
MSPM33C321Ax	LP-MSPM33C321A	MSPM33C321A	No
AM13E2x	–	AM13E2	Yes
AM263x	LP-AM263	AM2634	No
CC2755	–	CC2755	No
CC1352	–	CC1352	No
CC1354	–	CC1354	No
CC35X1	–	CC35X1	No

9.5. Deployment Overview

After training and compilation, Tiny ML Tensorlab produces artifacts that you can integrate into your CCS project:

Output Artifacts

File	Description
`mod.a`	Compiled model library (ONNX model compiled by TI MCU NNC into C code, packaged as a static library)
`tvmgen_default.h`	Header file with model inference API exposed by `mod.a`
`test_vector.c`	Golden test vectors (input + expected output) for on-device validation
`user_input_config.h`	Feature extraction configuration (preprocessing flags and parameters)

Deployment Steps

Train and compile your model using Tiny ML Tensorlab
Open Code Composer Studio
Import the SDK example project for your target device and task type
Copy the four output files from ModelMaker to the CCS project
Build, flash, and debug

The specific steps vary depending on whether your device has an NPU:

NPU Device Deployment - For F28P55x, AM13E2x, MSPM0G5187
Non-NPU Deployment - For all other devices

9.6. Prerequisites

Code Composer Studio (CCS)

Download from https://www.ti.com/tool/CCSTUDIO. Version 20.2.0 or later is recommended.

Install support for your target device family during CCS setup, and then install the appropriate SDK listed in the Supported Devices table above.

9.7. Output File Locations

After ModelMaker runs, the four deployment files are located in:

tinyml-modelmaker/data/projects/{dataset_name}/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

Important

Choose the golden vectors based on the model type you compiled:

Float model (non-quantized): Use golden vectors from training/base/golden_vectors/
Quantized model: Use golden vectors from training/quantization/golden_vectors/

The mod.a and tvmgen_default.h are always from compilation/artifacts/ regardless of quantization setting.

9.8. CCS Example Project Locations

TI provides ready-to-use CCS example projects for each task type. The example project location depends on your device family:

C2000Ware (F28x devices):

{C2000WARE_INSTALL_PATH}/libraries/ai/examples/
├── generic_timeseries_classification/
├── generic_timeseries_regression/
├── generic_timeseries_forecasting/
└── generic_timeseries_anomalydetection/

F29H85X SDK (F29x devices):

{F29H85X_SDK_INSTALL_PATH}/examples/rtlibs/ai/examples/
├── generic_timeseries_classification/
├── generic_timeseries_regression/
├── generic_timeseries_forecasting/
└── generic_timeseries_anomalydetection/

Inside each task-type folder, select the subfolder matching your target device (e.g., f28p55x/, f28004x/, f29h85x/).

9.9. Task-Type-Specific Deployment Notes

The deployment workflow is the same for all task types, but the CCS example project names and verification differ:

Classification

CCS example: generic_timeseries_classification
Verification: Check test_result variable (1 = pass, 0 = fail)
The model outputs class scores; argmax gives the predicted class

Regression

CCS example: generic_timeseries_regression
Verification: Check test_result variable (1 = pass, 0 = fail)
The model outputs a continuous value

Forecasting

CCS example: generic_timeseries_forecasting
Verification: Check test_result variable (1 = pass, 0 = fail)
The model predicts future values from historical data

Anomaly Detection

CCS example: generic_timeseries_anomalydetection
Verification: The application computes reconstruction error (MSE between input and autoencoder output) and compares to a threshold
In user_input_config.h, the anomaly threshold is defined:
```
#define RECONSTRUCTION_ERROR_THRESHOLD 0.014  // Threshold for k=4.5
```
This was calculated during training as: threshold = mean_normal_error + k * std_normal_error
Adjusting the threshold:
- Increase threshold (higher k): fewer false alarms, may miss subtle anomalies
- Decrease threshold (lower k): catches more anomalies, more false alarms
Refer to the threshold_performance.csv file from ModelMaker to choose the optimal k value for your application.

9.10. Testing Multiple Cases

The test_vector.c file contains multiple test cases (SET 0, SET 1, etc.). To test different cases:

Open test_vector.c in the CCS editor
Comment out the current test set (SET 0)
Uncomment another test set (e.g., SET 1)
Rebuild and reflash the project
Verify that test_result matches expected output

9.11. Model Compilation Details

The ONNX model produced by training is compiled using TI MCU NNC to generate mod.a and tvmgen_default.h. The compilation command used internally is:

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"

Different task types use different compilation flags:

Classification: skip_normalize=true output_int=true
Regression/Forecasting: Output is float (no output_int)
Anomaly Detection: Output is reconstruction error

See the TI MCU NNC documentation for all compilation options.