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

Revision as of 07:42, 21 July 2020 by Vincent Abriou (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

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 Difference between TensorFlow Lite Python APIs[edit]

The Artificial Intelligence expansion package X-LINUX-AI comes with two versions of TensorFlow Lite.

The first runtime is based on TensorFlow Lite[1]2.2.0, whereas the second runtime is based on TensorFlow Lite runtime[2]1.12.1 and is dedicated to Coral EdgeTPU use.

This is because TensorFlow Lite 2.2.0 does not yet support the Coral EdgeTPU runtime. The following figure explains the software structure.

TensorFlow Lite runtime package structure

If you wish to use the TensorFlow Lite 2.2.0 you have to import the following library in your Python script:

import tflite_runtime.interpreter as tflite

If you wish to run inferences on your Coral EdgeTPU device, you need to make the following call in your Python script:

import tflite_edgetpu_runtime.interpreter as tflite

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]

We start by installing the X-LINUX-AI components and the packages needed to run our example. The main packages are Python Numpy[3], Python OpenCV[4] 4.1.x and Python TensorFlow Lite Edge TPU runtime[2]1.12.1

Board $> apt-get install python3-numpy python3-opencv
Board $> apt-get install python3-tensorflow-lite-edgetpu

3.2 Preparing the workspace on the target[edit]

Before running the inference, make sure that your .tflite model is compiled for inferencing on Coral EdgeTPU. Refer first to Compile your custom model, then send the model to the board.

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

After preparing the workspace on the target and sending the compiled model to the model directory in the workspace, we send the associated label file and input image to the workspace so that the inference can be executed. Any number of .jpeg, .jpg and .png pictures of any size can be added. Some image processing operations are used later to adapt the picture sizes to the size of the input model. 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[5] website using the following commands:

PC $> wget https://github.com/google-coral/edgetpu/raw/master/test_data/mobilenet_v1_1.0_224_quant_edgetpu.tflite
PC $> wget https://github.com/google-coral/edgetpu/raw/master/test_data/imagenet_labels.txt -O labels.txt

We send then these file via the scp protocol using the following commands :

PC $> scp path/to/your/compiled/model/mobilenet_v1_1.0_224_quant_edgetpu.tflite root@<board_ip_address>:/usr/local/workspace/models/
PC $> scp path/to/your/labels.txt root@<board_ip_address>:/usr/local/workspace/models/
PC $> scp path/to/your/pictures root@<board_ip_address>:/usr/local/workspace/testdata/

Now that our workspace is ready with a compiled model file, a label file and some sample pictures, we can run an inference using the Python API. To do this, we create a Python script that is transferred via the scp command and run on the target board. This is a very basic example that classifies images by executing an inference on the Coral Edge TPU.

PC $> gedit classify_on_stm32mp1.py

3.3 Running the inference[edit]

If you are already familiar with inferencing TensorFlow Lite models, use the following Python script directly. Otherwise, copy it to a file named classify_on_stm32mp1.py and refer to the subsequent sections.

#!/usr/bin/python3
#
# 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:
#                        opensource.org/licenses/BSD-3-Clause

import sys
import numpy as np
import tflite_edgetpu_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"
interpreter = tflite.Interpreter(model_path = model_file, experimental_delegates = [tflite.load_delegate('libedgetpu-max.so.1.0')])
interpreter.allocate_tensors()
#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]
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)
interpreter.set_tensor(input_details[0]['index'], input_data)
start = time.perf_counter()
interpreter.invoke()
inference_time = time.perf_counter() - start
print("inference time:", inference_time)
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])
print("\n")

Now that our Python script is ready for execution, we send it to the board.

PC $> scp path/to/your/script/classify_on_stm32mp1.py root@<board_ip_address>:/usr/local/workspace/

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

After booting the board and connecting it to the host PC throught an SSH protocol, we are ready to run the inference using the following command:

Board $> cd /usr/local/workspace
Board $> python3 classify_on_stm32mp1.py test_data/<picture to classify>

Using the IA hardware accelerator speeds up the inferencing.

4 Explanation of the parts of the script[edit]

4.1 Instantiating the Tensorflow Lite Interpreter[edit]

We first load the labels from the label file by adding the following code lines:

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

We now load the model and feed it to the interpreter that we instantiate using the interpreter API [6] . In this interpreter we call a Tensorflow Lite delegate . This is an API that delegates all or part of the graph execution to the Edge TPU accelerator hardware. After calling the Edge TPU library inside the delegate, we allocate tensors for the graph execution through our interpreter.

model_file = "/usr/local/workspace/models/mobilenet_v1_1.0_224_quant_edgetpu.tflite"
interpreter = tflite.Interpreter(model_path = model_file, experimental_delegates = [tflite.load_delegate('libedgetpu-max.so.1.0')]) 
interpreter.allocate_tensors()

4.2 Getting the model details and processing the image[edit]

Now that the interpreter is ready to be fed with the input images, it is important to get the details of the model in order to adjust the image to fit into the model.

#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]

Now we point to our image directory testdata and pass the image as parameter. The images are converted 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(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)

4.3 Invoking the interpreter and displaying results[edit]

Now that our input data has been processed to fit in the model input size, we feed the image to the interpreter input and launch the inference. We use the time library to record the inference duration, which gives a good indication of the Edge TPU performance compared to that of the CPU.

interpreter.set_tensor(input_details[0]['index'], input_data)
start = time.perf_counter()
interpreter.invoke()
inference_time = time.perf_counter() - start
print("inference time:", inference_time)
output_details = interpreter.get_output_details()
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))+":", self._labels[i])
print("\n")

5 References[edit]