Test: NN Inference

The Test NN Inference application provides a simple way to run and validate a compiled neural network model on an Axon‑enabled target.

Requirements

The sample supports the following development kits:

Hardware platforms	PCA	Board name	Board target
nRF54LM20 DK	PCA10184	nrf54lm20dk	nrf54lm20dk/nrf54lm20b/cpuapp

Overview

The application combines a model, test vectors, and optionally per‑layer test data, runs inference, and compares the results against expected values. This application is intended as the fastest way to get a model running, verified, and profiled on a target device. It can be built either for Zephyr‑based targets or for the Axon simulator.

Note

This application only supports the Axon simulator and Nordic Semiconductor devices that include the Axon NPU.

Building and running

This section describes how to configure, build, and run the application.

1. Select how you want to build the application:

   • To build the simulator application in Visual Studio Code, install a CMake extension and add the simulator folder to your workspace.

   • To build for Zephyr, use a standard Zephyr build workflow.

2. Select one of the sample models from /compiled_models directory, or copy your own compiled model header files into that directory. For information about compiling models, see Axon NPU TFLITE compiler documentation.

3. Configure the prj.conf (for Zephyr builds) or simulator/CMakeLists.txt (for simulator builds) with model parameters:

   • Set CONFIG_NRF_AXON_MODEL_NAME to the model name (for example, tinyml_kws). This value is used to include the correct header file and resolve model symbols.

   • Set CONFIG_NRF_AXON_INTERLAYER_BUFFER_SIZE to a value large enough for the model. Use the value defined by AXON_MODEL_<model_name>_MAX_IL_BUFFER_USED in the axon_model_<model_name>.h file. A value of around 115000 is typically sufficient.

4. Edit the following macros in src/nrf_axon_app_nn_test_nn_inference.c to control how inference is performed:

   • INCLUDE_VECTORS - When set to 0, test vectors are excluded and no inference is performed. This mode is useful for measuring the image size without test data.

   • AXON_MINIMUM_TEST_VECTORS - When set to 1, only a single end‑to‑end test vector is included. This produces the smallest application that still performs a valid inference.

   • AXON_LAYER_TEST_VECTORS - When set to 1, individual layer test vectors are included and executed.

   • AXON_LAYER_TEST_START_LAYER and AXON_LAYER_TEST_STOP_LAYER - Use these to limit testing to a specific range of layers, which can help reduce image size or focus debugging on specific layers.

5. Build the application by running the appropriate commands for your chosen build method:

   • For a command‑line Zephyr build, run the west build from the application directory. Flash the application to the device and monitor the UART output for test results.

   • For a simulator build in Visual Studio Code, use the CMake extension to build and run the application directly.

     Sample output should be as follows:

   *** Booting nRF Connect SDK v3.3.0-preview2-ede152ec210b ***
   *** Using Zephyr OS v4.3.99-4b6df5ff11b1 ***
   Hello world from nrf54lm20dk
   ticks per second 1000000
   Start Platform!
   Prepare and run Axon!
   
   TEST:   test_nn_inference_tinyml_kws    CASE COUNT      15
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   0
   Test inference tinyml_kws vector 0 FULL MODEL sync mode
   output bit exact!
   model tinyml_kws inference: ndx 7, label STOP, score 266802540, profiling ticks 5213
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 0       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   1
   Test inference tinyml_kws vector 1 FULL MODEL async mode
   output bit exact!
   model tinyml_kws inference: ndx 2, label LEFT, score 149726243, profiling ticks 5254
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 1       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   2
   Test inference tinyml_kws vector 2 FULL MODEL sync mode
   output bit exact!
   model tinyml_kws inference: ndx 6, label RIGHT, score 268434516, profiling ticks 5213
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 2       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   3
   
   Test inference tinyml_kws vector 0 layer 0
   output bit exact!
   profiling ticks 699
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 3       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   4
   
   Test inference tinyml_kws vector 0 layer 1
   output bit exact!
   profiling ticks 197
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 4       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   5
   
   Test inference tinyml_kws vector 0 layer 2
   output bit exact!
   profiling ticks 1028
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 5       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   6
   
   Test inference tinyml_kws vector 0 layer 3
   output bit exact!
   profiling ticks 197
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 6       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   7
   
   Test inference tinyml_kws vector 0 layer 4
   output bit exact!
   profiling ticks 1022
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 7       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   8
   
   Test inference tinyml_kws vector 0 layer 5
   output bit exact!
   profiling ticks 197
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 8       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   9
   
   Test inference tinyml_kws vector 0 layer 6
   output bit exact!
   profiling ticks 1023
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 9       RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   10
   
   Test inference tinyml_kws vector 0 layer 7
   output bit exact!
   profiling ticks 197
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 10      RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   11
   
   Test inference tinyml_kws vector 0 layer 8
   output bit exact!
   profiling ticks 1023
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 11      RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   12
   
   Test inference tinyml_kws vector 0 layer 9
   output bit exact!
   profiling ticks 150
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 12      RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   13
   
   Test inference tinyml_kws vector 0 layer 10
   output bit exact!
   profiling ticks 77
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 13      RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    START CASE NO   14
   
   Test inference tinyml_kws vector 0 layer 11
   output bit exact!
   profiling ticks 72
   
   TEST:   test_nn_inference_tinyml_kws    CASE NO 14      RESULT: PASS
   
   TEST:   test_nn_inference_tinyml_kws    COMPLETE        PASS COUNT      15      FAIL COUNT      0
   test_nn_inference complete!
   

Dependencies

This test uses the following Edge AI Add-on libraries:

• Axon driver