Android embeds its own trained yolov5 model (tfLite) traffic signs

Table of Contents

Step 1: Download the model and modify parameters

Step 2: Label the data

Step 3: Start training

Step 4: Convert yolov5 to tfLite model

Step 5: We can check whether tfLite is available

Step 6: Download the official sample code

Step 7: Modify the code

Step 8: Run the software

Step 9: Optimize Speed

renderings

Reference: [Selected] Teach you step by step how to use YOLOV5 to train your own target detection model-mask detection-video tutorial_yolov5 training model_Si Twelve’s blog-CSDN blog

Model download address:

YOLOV5-mask-42: Mask detection system based on YOLOV5 – teaching video provided (gitee.com)

What I used was to convert the yolov5 model into a tfLite model and then embed it into Android.

For the creation of conda environment, you can see the reference at the top. The boss has explained how to create the environment in detail.

The first step: download the model and modify the parameters

After downloading the model, change the following files according to our needs

Both train and val are the paths of data set images. The data set does not need to be included in the project, but the format of the path is best

After changing it to our own path, change here (nc) to the format of our own target object, (names) and the naming of the target

It is important to note the ordering of goals here, and you must remember them well.

Step 2: Label data

Enter at the command line

pip install labelImg

After installation, enter

labelImg

Start labelImg

The location of the data picture is

Note here that in the annotation file, the class file will be automatically generated with our first annotation target.

As we annotate, the corresponding txt file will also appear.

Each txt contains the category of the image and the target location.

And this order is exactly the target order written manually when we modify mask_data. These two places must correspond.

When annotating pictures, we will find that every time we reopen the annotation application, it will not recognize our own class file, but will re-create it for us, and the order is the order of annotation after we reopen the application. This will cause the content of the class file to be replaced, causing confusion in the labeling data. Sometimes it will crash for this reason, so we can create the class file in advance in the labels folder and arrange it in our customized order.

Then when starting labelImg

enter

 Labelimg class_file E:\StudyData\python\TrafficData\labels\text\classes.txt

In this way, the annotation application will recognize your class file.

After we have labeled all the data, each picture will correspond to a label file in labels

After we have labeled our own data set, we can start training

Step 3: Start training

Find the train file and find the information at this location. What is demonstrated here is the pre-training model of yolov5m.

Here –data, be careful not to write it as mask_data2.yaml like me, the default is mask_data.yaml

If running on GPU, change this to 0

After modifying these things, you can enter

python train.py --epoch 100 --batch-size 8

Then wait for the model to finish runningBecause of how to run the model, other bloggers have already published relevant tutorials. Here is a simple implementation. You can refer to others for specific details. Here I will mainly share how I converted to tfLite and successfully implemented the yolov5 model. of conversion

Step 4: Convert yolov5 to tfLite model

When our model is finished running

You can see this file in runs/train/exp*/weights/best.pt. This is the result we ran;

You can test it in window.py written by the boss

Then run directly

Pass in our own data

As you can see, the model has been successfully run.

The next step is to start the conversion

Open export.py

After the modification is completed, run export directly

Wait for a while for his automatic conversion

Conversion successful

You can see that a tfLite file has been generated in our project

Step 5: We can check whether tfLite is available

Get the output data details of the tflite model

Enter python in the command line

Here you can see that the model successfully inputs information

python

>>> import tensorflow as tf
>>> yolov5s = "runs/train/exp17/weights/best-fp16.tflite"
>>> interpreter = tf.lite.Interpreter(model_path = yolov5s)
>>> yolov5s = "runs/train/exp17/weights/best-fp16.tflite"
>>> interpreter = tf.lite.Interpreter(model_path = yolov5s)
>>> print(interpreter.get_output_details())

The corresponding commands are: