Articles directory

---

Foreword

After the previous preparation, the environmental conditions of Darknet running are basically available. The next step is to install and compile the Darknet package and simple test to prepare your own data set to train your own weight files.

---

1. Install the darknet package and compile

(1) Code download:

You can refer to the GitHub homepage directly:https://github.com/leggedrobotics/darknet_ros

$mkdir -p catkin_workspace/src
$cd catkin_workspace/src
$git clone --recursive [email protected]:leggedrobotics/darknet_ros.git
$cd ../

Notice:You may not be downloaded in this way, and the file may be incomplete, and you cannot download it directly on the GitHub website. Then the downloaded Darknet folder is empty, you can only download it through git clone, and you must add-recursive to download to download completely. But this clone may report errors. Because you need to configure SSH and generate a key, add the GitHub SSH key, you can refer to:

Generate an SSH key on Ubuntu and add the generated public key to your Github account

I also package my own github to the project, but it is slightly modified.

https://github.com/zhuzhengming/darknet_ros_yolo3

(2) Compilation

In work space, execute command:

$catkin_make -DCMAKE_BUILD_TYPE=Release

At this time, the entire project will be compiled. After the compilation is completed, check the {catkin_ws}/darknet_ros/darknet_ros/yolo_network_config/weights files. There are two model files of Yolov2-TINY.Weights and YOLOV3.weights. The volume does not bring these two model files. Therefore, after compiling, the model file will be downloaded automatically, and this is a long waiting time.

：NVCC src/caffe/layers/reduction_layer.cu nvcc fatal : Unsupported gpu architecture 'compute_30' Makefile:588: recipe for target '.build_release/cuda/src/caffe/layers/reduction_layer.o' failed make: * ** [.build_release/CUDA/SRC/Caffe/Layers/Reduction_layer.o] error 1

This is not matching the computing power, you need to release the unwanted options in the cmakelist.txt file in the/Workspace/src/darknet_ros/darknet_ros or in the Makefile file based on your graphics card and CUDA version.

If CUDA and CUDNN are installed, the parameters can be modified in the Makefile file:

GPU=0
CUDNN=0
OPENCV=1
OPENMP=0
DEBUG=1

change into:

GPU=1
CUDNN=1
OPENCV=1
OPENMP=0
DEBUG=1

Then execute the following command compilation (you must re -compile after each modification of Makefile):

$ cd darknet
$ make

(3) Download the pre -training model and simple test

Here is LOV3-TINY.Weights, download in the root directory:

$wget https://pjreddie.com/media/files/yolov3-tiny.weights

Simple test:

In the Darknet root directory:

1. Picture test:

$./darknet detect cfg/yolov3-tiny.cfg yolov3-tiny.weights data/dog.jpg

2. Real -time video test:

$./darknet detector demo cfg/voc.data cfg/yolov3-tiny.cfg yolov3-tiny.weights

The installation and compilation and test of Darknet is completed here, and the next is to train your data set.

 

2. Make your own data set and train

(1) Prepare data set

In the root directory of Darknet, establish a standard VOC dataset:

$ cd scripts
$ mkdir -p VOCdevkit && cd VOCdevkit
$ mkdir -p VOC2021  && cd VOC2021
$ mkdir -p Annotations && mkdir -p ImageSets && mkdir -p  JPEGImages && mkdir -p labels

Annotations: .xml file generated after storing labeling

JPEGIMAGES: Storage of your own picture material

labels: Storage label files for data format required for YOLO

(2) Data marking

The Labelimg tools are mainly used here. This has a version in Windows and Linux. Here is a Linux installation method:

$ sudo apt-get install pyqt5-dev-tools
$ sudo pip3 install lxml
$ git clone https://github.com/tzutalin/labelImg.git
$ cd labelImg
$ make all
$ python3 labelimg.py #Open labelimg

"Open File" ----- "Create RECTBOX" ------ "Input class name" ----- "Change Save Dir" ------ "Save" Open Dir Select JPEGIMAGES, Change Save Dir Annotations.

(3) Generate training sets and test sets

In the ImageSets directory, the new main directory is created, and then two new text documents trains.txt and value are created in the main directory.

$ cd ImageSets 
$ mkdir -p Main && cd Main
$ touch train.txt test.txt 

YOLOV3 provides a script that translates the VOC dataset to the format required for YOLO training. In the scripts/voc_label.py file, do some modifications:

Sets changed to its own training set

CLASSES trained for itself

 

Execute this script:

$ python scrips/voc_label.py

Under the main file, the trin.txt and var.txt files will also generate 2021_train.txt and 2021_var.txt files in the darknet root directory. The latter two files are with paths and are required for training.

(4) Modify the necessary files

1. Modify the voc.names in the data directory and modify it to the type label you want to train:

2. Modify the VOC.DATA file in the CFG directory:

The main reason is to modify the preservation path of your own training set and the well -trained weight file. The absolute path I use here:

classes= 2
train  = /home/zhuzhengming/workspace/src/darknet_ros/darknet/2019_train.txt
valid  =/home/zhuzhengming/workspace/src/darknet_ros/darknet/2019_test.txt
names = /home/zhuzhengming/workspace/src/darknet_ros/darknet/data/voc.names
backup = /home/zhuzhengming/workspace/src/darknet_ros/darknet/backup

3. Modify YOLOV3-TINY.CFG file in the cfg directory:

Modify the Classes and Fliters, there are two places. The calculation method of Filiters is: 3*(Classes + 5):

The following is one of the places, and the modification refers to this:

[convolutional]
size=1
stride=1
pad=1
filters=21
activation=linear



[yolo]
mask = 3,4,5
anchors = 10,14,  23,27,  37,58,  81,82,  135,169,  344,319
classes=2
num=6
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=0

(5) Start training

1. Generate pre -training model:

$ ./darknet partial cfg/yolov3-tiny.cfg yolov3-tiny.weights yolov3-tiny.conv.15 15

2. Training:

$ ./darknet detector train cfg/voc.data cfg/yolov3-tiny.cfg yolov3-tiny.conv.15 | tee person_train_log.txt

In this process, the intermediate result of the corresponding iteration is saved under the Backup folder. Save one within 100 times in the first 1,000 times, more than 1,000 times, and save it every 1,000 times.

3. Test it with training weight files:

Photo test:

Here is a picture in your own data set as a test

$ ./darknet detect cfg/yolov3-tiny.cfg yolov3-tiny_900.weights VOCdevkit\VOC2019\JPEGImages\1.png

Video real -time test:

$ ./darknet detector demo cfg/voc.data cfg/yolov3-tiny.cfg yolov3-tiny_900.weights

---

Summarize

Through the download compilation and re -configuration of the official package of the DarkNet package, its basic use functions have been achieved. Based on this, the production of its data set is also completed, and this data set is used to train my weight file, because I training the category and The number of times is small, but from the test results, the target detection effect is not bad. Next is to use this deep learning network under the ROS platform.