Local deployment of yolo training environment

Environment Configuration

In this section, we will demonstrate how to deploy the Yolo v5 environment locally and train a classification (recognition) model.

The operating environment we demonstrate is as follows:

System: Ubuntu 22.04

GPU: 4090D*1

CUDA version: 12.4

The system has pre-installed Miniconda3_3.10

Download and install Yolo v5 environment dependencies

Enter the command

xxxxxxxxxx
git clone https://github.com/ultralytics/yolov5.git

implement

xxxxxxxxxx
cd yolov5
pip install -r requirements.txt

This part of the file contains a lot of content, please be patient

Preparing training data

Here we use the training set officially provided by Canaan Technology as sample data

Please download the sample dataset provided. The sample dataset contains classification, detection and segmentation datasets for three types of fruits (apple, banana and orange). Unzip the dataset to yolov5the directory and use it fruits_clsas the dataset for the fruit classification task.

We execute the following code

xxxxxxxxxx
wget https://kendryte-download.canaan-creative.com/developer/k230/yolo_files/datasets.zip

xxxxxxxxxx
unzip datasets.zip

Use YOLOv5 to train a fruit classification model

yolov5Execute the command in the directory and use yolov5to train three types of fruit classification models:

We enter the yolov5 directory (~/yolov5 in the example)

Execute the following code to start training

xxxxxxxxxx
python classify/train.py --model yolov5n-cls.pt --data datasets/fruits_cls --epochs 100 --batch-size 8 --imgsz 224 --device '0'

If an interface similar to the one below appears, it means that training has started.

After the training is successfully completed, the output is as follows

Convert the trained model to a kmodel that can be recognized by K230

The model we trained with yolov5 is best.pt with the suffix pt

In order to use it on K230, we need to convert it into a model in kmodel format

Download the conversion tool

Execute the following code to install the required tools

xxxxxxxxxx
sudo apt-get install -y dotnet-sdk-7.0
pip install --upgrade pip
pip install nncase==2.9.0
pip install nncase-kpu==2.9.0
pip install onnx
pip install onnxruntime
pip install onnxsim

Download the script tool and test_yolov5.zipunzip the model conversion script tool to yolov5the directory;

xxxxxxxxxx
wget https://kendryte-download.canaan-creative.com/developer/k230/yolo_files/test_yolov5.zip
unzip test_yolov5.zip

Start model format conversion

Step 1: Convert pt format to onnx format

xxxxxxxxxx
python export.py --weight runs/train-cls/exp/weights/best.pt --imgsz 224 --batch 1 --include onnx

Note that "runs/train-cls/exp/weights/best.pt" here is the path of the pt model we trained.

If the previous steps are done exactly as the tutorial, then the path here does not need to be modified

Step 2: Enter the test_yolov5/classify directory

xxxxxxxxxx
cd test_yolov5/classify

Execute model conversion instructions

xxxxxxxxxx
python to_kmodel.py --target k230 --model ../../runs/train-cls/exp/weights/best.onnx --dataset ../test --input_width 224 --input_height 224 --ptq_option 0

Note: An error may occur at this time: RuntimeError: /root/miniconda3/bin/../lib/libstdc++.so.6: version `GLIBCXX_3.4.30' not found (required by /usr/lib/dotnet/host/fxr/7.0.19/libhostfxr.so): Success

Solution:

Check if GLIBCXX_3.4.30 exists

xxxxxxxxxx
strings /usr/lib/x86_64-linux-gnu/libstdc++.so.6 | grep GLIBCXX

If there is one in the output, it means that the file is not really missing, but the version of libstdc++ in the virtual environment is lower than the version of the libstdc++ library in the system.

At this point we only need to execute the following command

Delete the libstdc++.so.6 file in the virtual environment (or rename it to another name, here we rename it to libstdc++.so.6.old)

xxxxxxxxxx
mv /root/miniconda3/bin/../lib/libstdc++.so.6 /root/miniconda3/bin/../lib/libstdc++.so.6.old

Then create a soft link and let the libstdc++.so.6 called by the virtual environment point to the libstdc++.so.6 that comes with the system.

xxxxxxxxxx
ln -s /usr/lib/x86_64-linux-gnu/libstdc++.so.6 libstdc++.so.6

After the conversion command is completed, the output is as follows

WARNING can be ignored here

The generated model is in the directory ~/yolov5/runs/train-cls/exp/weights

The name is best.kmodel

We copy this best.kmodel to the data directory of our K230

Then open CanMV IDE, copy the following code and click Run

xxxxxxxxxx
from libs.PipeLine import PipeLine, ScopedTiming
from libs.YOLO import YOLOv5
import os,sys,gc
import ulab.numpy as np
import image
​
if __name__=="__main__":
​
    rgb888p_size=[1280,720]
    display_size=[640,480]
    # 模型路径 Model path
    kmodel_path="/data/best.kmodel"
    labels = ["apple","banana","orange"]
    confidence_threshold = 0.5
    model_input_size=[224,224]
    # 初始化PipeLine  Initialize PipeLine
    pl=PipeLine(rgb888p_size=rgb888p_size,display_size=display_size,display_mode="lcd")
    pl.create()
    # 初始化YOLOv5实例  Initialize the YOLOv5 instance
    yolo=YOLOv5(task_type="classify",mode="video",kmodel_path=kmodel_path,labels=labels,rgb888p_size=rgb888p_size,model_input_size=model_input_size,display_size=display_size,conf_thresh=confidence_threshold,debug_mode=0)
    yolo.config_preprocess()
    try:
        while True:
            os.exitpoint()
            with ScopedTiming("total",1):
                # 逐帧推理 Frame-by-frame reasoning
                img = pl . get_frame ()
                res = yolo . run ( img )
                yolo . draw_result ( res , pl . osd_img )
                pl . show_image ()
                gc . collect ()
    except  Exception  as  e :
        sys.print_exception ( e )
    finally :
        yolo.deinit ( )
        pl . destroy ()

Let's test the effect

Identify Apples

Identify bananas

Identify oranges

You can see that this routine can run successfully.

Other considerations

1. Environment configuration, Yolov5 environment dependencies, and model conversion tools only need to be downloaded once, and no re-download is required for subsequent training
2. If the speed of pip download is much lower than the actual network bandwidth, please check whether the pip mirror source is configured.
3. The model trained with the current dataset may identify non-banana objects other than oranges/apples as bananas. This is normal. If a more accurate model is needed, a training set with a larger amount of data can be prepared for training.