How to run Coral Edge TPU inference using Python TensorFlow Lite API

1 Article purpose[edit]

This article describes how to run an inference on the STM32MP1 using a Google Coral EdgeTPU device and the Python TensorFlow Lite API. It is an example based on an image classification application.

Info.png There are many ways to achieve this result; this article provides a simple example. You are free to explore other methods that are better adapted to your development constraints.

2 Libedgetpu and TensorFlow Lite Python APIs[edit]

The Artificial Intelligence expansion package X-LINUX-AI comes with TensorFlow Lite Python APIs and the libedgetpu (providing the support of the Coral Edge TPU) that has been rebuilt from source to be compatible with the embedded TensorFlow Lite runtime.

In the next section we explore, with a basic image-classification example, how to inference your models on the board using the Coral EdgeTPU device.

3 Running an inference on Coral EdgeTPU using the TensorFlow Lite Python API[edit]

3.1 Installing prerequisites on the target[edit]

After having configured the AI OpenSTLinux package, you can install the X-LINUX-AI components and the packages needed to run our example.
The main packages are Python Numpy[1], Python OpenCV[2], Python TensorFlow Lite runtime[3] and libedgetpu

Warning.png The software package is provided AS IS, and by downloading it, you agree to be bound to the terms of the software license agreement (SLA). The detailed content licenses can be found here.
Board $> apt-get install python3-numpy python3-opencv python3-tensorflow-lite libedgetpu

3.2 Preparing the workspace on the target[edit]

Board $> cd /usr/local/ && mkdir -p workspace
Board $> cd workspace && mkdir -p models testdata 

In this example, we use the mobilenet_v1_1.0_224_quant_edgetpu.tflite model to classify download images, accompanied by the labels file from the Coral[4] website.

Board $> wget -O models/mobilenet_v1_1.0_224_quant_edgetpu.tflite
Board $> wget -O models/labels.txt
Board $> wget -O testdata/bird.bmp
Info.png You can run your own model but you have to make sure that your .tflite model is compiled for inferencing on Coral EdgeTPU. Refer first to Compile your custom model

3.3 Running the inference[edit]

Here is a simple python script example to execute a NN inference on the Google Coral Edge TPU.

# Copyright (c) 2020 STMicroelectronics. All rights reserved.
# This software component is licensed by ST under BSD 3-Clause license,
# the "License"; You may not use this file except in compliance with the
# License. You may obtain a copy of the License at:

import sys
import numpy as np
import tflite_runtime.interpreter as tflite
import time
import cv2

label_file = "/usr/local/workspace/models/labels.txt"
with open( label_file, 'r') as  f :
           labels = [ line.strip() for line in f.readlines() ]
model_file = "/usr/local/workspace/models/mobilenet_v1_1.0_224_quant_edgetpu.tflite"

#Create the interpreter and allocate tensors
interpreter = tflite.Interpreter(model_path = model_file, experimental_delegates = [tflite.load_delegate('')])

#Getting the model input and output details
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
height = input_details[0]['shape'][1]
width = input_details[0]['shape'][2]

#Read the picture, convert from BGR to RGB encoding,
#resized to fit the size of the model input and have their
#dimensions expanded by one
image = cv2.imread(sys.argv[1])
nn_img_rgb = cv2.cvtColor(np.array(image), cv2.COLOR_BGR2RGB)
nn_img_rgb_resized = cv2.resize(nn_img_rgb, (width, height))
input_data = np.expand_dims(nn_img_rgb_resized, axis=0)

#Set the input data, execute the first inference (that could take longer
#since the model is being loaded on the Coral Edge TPU RAM)
interpreter.set_tensor(input_details[0]['index'], input_data)
start = time.perf_counter()
inference_time = time.perf_counter() - start
print("1st inference:", inference_time, "s")

# Execute the second inference and measure the inference duration
start = time.perf_counter()
inference_time = time.perf_counter() - start
print("2nd inference:", inference_time, "s")

#Print the results
results = np.squeeze(interpreter.get_tensor(output_details[0]['index']))
top_k = results.argsort()[-5:][::-1]
for i in top_k:
    print('{0:08.6f}'.format(float(results[i]*100/255.0))+":", labels[i])

Copy this python script to the target: PC $> scp path/to/your/script/ root@<board_ip_address>:/usr/local/workspace

3.4 Running the inference from the board on the Coral Edge TPU[edit]

Board $> cd /usr/local/workspace
Board $> python3 testdata/bird.bmp
inference time:  0.1181572640016384 ms
88.627451: 20  chickadee
4.313725: 19  magpie
3.137255: 18  jay
2.745098: 21  water ouzel, dipper
1.176471: 14  junco, snowbird
Info.png The first inference may take longer since the model is being loaded on the Coral Edge TPU RAM

4 References[edit]