AN 1011: TinyML Applications in Altera FPGAs Using LiteRT for Microcontrollers

Download PDF
ID 848984
Date 9/29/2025
Public
Document Table of Contents
Document Table of Contents x
1. Overview 2. Preparing LiteRT Inference Model 3. Generating Nios® V Processor System 4. Generating Arm Processor System 5. Programming and Running 6. Nios® V Processor with TinyML Design Example 7. Appendix 8. Document Revision History for the AN 1011: TinyML Applications in Altera FPGAs Using LiteRT for Microcontrollers
1. Overview x
1.1. Requirements
2. Preparing LiteRT Inference Model x
2.1. Defining the Problem 2.2. Gathering and Preparing Sample Data 2.3. Building TensorFlow Model 2.4. Preparing the LiteRT Model 2.5. Preparing Binaries for C/C++
2.2. Gathering and Preparing Sample Data x
2.2.1. Preparing Dataset 2.2.2. Preprocessing Dataset
2.3. Building TensorFlow Model x
2.3.1. Constructing Model Architecture 2.3.2. Configuring Model 2.3.3. Training Model 2.3.4. Evaluating the Model 2.3.5. Saving and Loading Model
2.4. Preparing the LiteRT Model x
2.4.1. Converting into LiteRT Model 2.4.2. Saving the LiteRT Model 2.4.3. Analyzing the LiteRT Model 2.4.4. Loading a LiteRT Interpreter 2.4.5. Evaluating the LiteRT Model
2.5. Preparing Binaries for C/C++ x
2.5.1. Converting LiteRT to a C Array 2.5.2. Preparing a Header File for LiteRT C Array 2.5.3. Preparing Supporting Header File for LiteRT C Array 2.5.4. Preparing Main Function to Run TinyML 2.5.5. Converting MNIST Sample into C Array
3. Generating Nios® V Processor System x
3.1. Building Hardware Design in Platform Designer 3.2. Building Software Design with Ashling* RiscFree* IDE for Altera® FPGAs
3.2. Building Software Design with Ashling* RiscFree* IDE for Altera® FPGAs x
3.2.1. Creating a Board Support Package 3.2.2. Building LiteRT for Microcontroller Static Library 3.2.3. Creating an Application Project File 3.2.4. Building the LiteRT TinyML Application
4. Generating Arm Processor System x
4.1. GHRD Hardware Design 4.2. Building the Software Design
4.2. Building the Software Design x
4.2.1. Arm-Trusted-Firmware 4.2.2. Building LiteRT for Microcontrollers Static Library 4.2.3. Zephyr OS Environment Setup 4.2.4. Building the LiteRT TinyML Application
4.2.1. Arm-Trusted-Firmware x
4.2.1.1. Booting from QSPI Flash 4.2.1.2. Running through Ashling* RiscFree* IDE for Altera® FPGAs Debugger
5. Programming and Running x
5.1. Nios® V Processor System 5.2. Arm HPS Processor System
5.2. Arm HPS Processor System x
5.2.1. Booting from QSPI 5.2.2. Booting via the Debugger
6. Nios® V Processor with TinyML Design Example x
6.1. Finding Nios® V Processor with TinyML Design Example
7. Appendix x
7.1. main.cc for Nios V Processor 7.2. main.cc for Arm HPS Processor
1. Overview
2. Preparing LiteRT Inference Model
3. Generating Nios® V Processor System
4. Generating Arm Processor System
5. Programming and Running
6. Nios® V Processor with TinyML Design Example
7. Appendix
8. Document Revision History for the AN 1011: TinyML Applications in Altera FPGAs Using LiteRT for Microcontrollers

2.5.5. Converting MNIST Sample into C Array

You can prepare a collection of C arrays representing MNIST samples. These arrays serve as static inputs to test the LiteRT C array model in the Nios® V processor before deployment.

Repeat this Python script for a few rounds until every class (0 to 9) is collected. The script replaces duplicated class, thus resulting in a single sample for each class.

# Convert MNIST samples into a C array
import matplotlib.pyplot as plt
import tensorflow as tf
import numpy as np
import random
import sys

# Load the MNIST Train and Test Dataset
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()

rows, cols = 28, 28

# Reshape the data into a 4D Array
x_test = x_test.reshape(x_test.shape[0], rows, cols, 1)

input_shape = (rows,cols,1)

# Set type as float32 and normalize the values to [0,1]
x_test = x_test.astype('float32')
x_test = x_test / 255.0

# Transform labels to one hot encoding
y_test = tf.keras.utils.to_categorical(y_test, 10)

img = x_test*255.0
img = img.astype(np.uint8)
img_label=np.argmax(y_test, axis=1)

# Repeat for a few rounds to get all numbers (0-9)
fig = plt.figure(figsize=(9,9))
for i in range(9):
  ind = random.randint(0, len(img))
  np.set_printoptions(threshold=sys.maxsize)
  string1 = np.array2string(img[ind], separator=', ')
  c_array = string1.replace('[', '{').replace(']', '}').replace('.', '')
  c_array_label = img_label[ind]

  plt.subplot(3,3,i+1)
  plt.imshow(img[ind], cmap="gray", interpolation=None)
  plt.title(c_array_label)

  base_path = "figure-"
  label_name = str(c_array_label)
  file_type = ".h"
  file_name = base_path + label_name + file_type
  with open(file_name, "w") as f:
    f.write("#include <stdint.h>\n\n")
    f.write("#define IMAGE_WIDTH 28\n")
    f.write("#define IMAGE_HEIGHT 28\n")
    f.write("#define NUM_CHANNELS 1\n")
    f.write("#define IMAGE_SIZE (IMAGE_WIDTH * IMAGE_HEIGHT * NUM_CHANNELS)\n\n")
    f.write("uint8_t test_image[IMAGE_HEIGHT][IMAGE_WIDTH][NUM_CHANNELS] = ")
    f.write(c_array)
    f.write(";")
Figure 8. Randomly Selected MNIST Samples
Figure 9. Complete Set of MNIST Samples
Level Two Title