YOLOv4 Darknet for Windows

Training the neural network for object detection. It uses to darknet to do this.

More details on Medium:

• YOLOv4

Main Repository is AlexeyAB's:

• AlexeyAB/darknet

Table of Contents

Requirements for Windows

• Visual Studio >= 2017: VS

• CMake >= 3.18: CMake

• OpenCV >= 2.4: OpenCV (on Windows set system variable OpenCV_DIR = C:\opencv\build - where are the include and x64 folders image)

• CUDA >= 10.2: Cuda

• cuDNN >= 8.0.2: cuDNN

• GPU with CC >= 3.0: Control to GPU support

That is developer's(Alexey) minumum advices. There were no problems with the following versions.

• Visual Studio = 2019
• CMake = 3.20.5
• OpenCV = 4.5.0
• CUDA = 11.2
• cuDNN = 8.0.5

You need to Python for run scripts or annotations. If you don't install before, you visit this page setup Python.

Darknet Installiation

-> First Step: VS 2019

• You download on link then install with marked checkbox.

-> Second Step: CUDA

• You install follow setup.
• After, control the system environment variables. Variables must be as on shown images.

-> Third Step: CuDNN

• You download the folder. Copy all files in folder then paste in this directory "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2" as shown images.

-> Fourth Step: OpenCV

• You download on link then extract to "C:\opencv_4.5.0\".

• You editing system environment variables as shown images.

• If 'opencv_world' checkbox is not marked, mark it with Cmake. Firstly clik configure then generate as shown image.

-> Fifth Step: Build Darknet

• Download the my repository. I'm assuming it is in this directory, "C:\Users\Name\Desktop\darknet\"
• Copy cudnn64_8.dll in directory "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2\bin" then paste to darknet directory "C:\Users\Name\Desktop\darknet\build\darknet\x64".
• Copy opencv_videoio_ffmpeg450_64.dll, opencv_world450.dll, opencv_world450d.dll in directory "C:\opencv_4.5.0\opencv\build\x64\vc15\bin" then paste to darknet directory "C:\Users\Name\Desktop\darknet\build\darknet\x64".
• Open darknet.vcxproj, yolo_cpp_dll.vcxproj' with notepad in directory "C:\Users\Name\Desktop\darknet\build\darknet\".
• Change CUDA and OpenCV version to your own version.
• Open yolo_cpp_dll.sln, darknet.sln with VS2019 and build it like shown image.

Preparate to Dataset

• If you haven't available dataset, you must creat it.
• For this, I recommend collecting a minimum of 200 images per class to be detected. However, it would be useful to have at least 2000 images in the independent from number of classes.
• Now, you must label images. For this, I recommend Labelimg. Follow the link for installation. Change saving format to YOLO and you are ready for annotation.
• Change the classes.txt that Labelimg generates to _darknet.labels when finish the labeling.
• Dataset needs to be split. You can use split_data.py in preprocess folder for splitting dataset as train, validation, test. All three must have _darknet.labels.
• Can use bb_data_aug.py for increase the train set.
• If you have available dataset, you can use xml2yolo.py in preprocess folder for .xml labels to yolo labels.
• After, copy images and annotations in train and valid folders. Paste it in "C:\Users\Name\Desktop\darknet\build\darknet\x64\data\obj". *Copy images and annotations in test folder. Paste it in "`C:\Users\Name\Desktop\darknet\build\darknet\x64\data\test
• You must list of dataset as train.txt, valid.txt, test.txt for this, you can use create_list_of_images.py in this directory "C:\Users\Name\Desktop\darknet\build\darknet\x64".
• Then, you must create files named obj.data, obj.names in "C:\Users\Name\Desktop\darknet\build\darknet\x64\data\" as shown images.

Content of the file data/obj.names should be

NOK # name of first class
Kon # name of second class
              # third
              # all of them

Content of the file data/obj.data should be

classes = 1 #Number of classes 
train = C:/Users/Name/Desktop/darknet-master/build/darknet/x64/data/train.txt # Location of train set
valid = C:/Users/Name/Desktop/darknet-master/build/darknet/x64/data/valid.txt # Location of validation set, during test must be test.txt
names = C:/Users/Name/Desktop/darknet-master/build/darknet/x64/data/obj.names # Location of obj.names
backup = C:/Users/Name/Desktop/darknet-master/build/darknet/x64/backup # Location of saved weights

Starting the Training

Everything is OK, no more obstacles to start training!

• If you use Anaconda, you open the Anaconda Prompt or if don't use, open the CMD.

• Powershell can also be used. I don't prefer.

• Now, if you don't open powershell, need to do one extra step. Move tee.exe where scripts to directory "C:\Windows\System32".

• Open the prompt and go to the directory where the repo is located with cd.

C:\Users\Name>cd C:\Users\Name\Desktop\darknet\build\darknet\x64

After, write and click enter

# This, starting the training with pretrained weight. Recommend! You move the pre-trained pretrained_model/yolov4.conv.137 from the repository to C:\Users\Name\Desktop\darknet\build\darknet\x64 in the repository.
 
darknet.exe detector train data/obj.data cfg/yolov4-custom.cfg yolov4.conv.137 -dont_show -map | tee results.log

# If you want to start from scratch try this 

darknet.exe detector train data/obj.data cfg/yolov4-custom.cfg -dont_show -map | tee results.log 

# If you want to continue where you left off if the training is interrupted, do this.

darknet.exe detector train data/obj.data cfg/yolov4-custom.cfg /backup/yolov4-custom_last.weights -dont_show -map | tee results.log

• Will be created training record with the tee.exefile above in "C:\Users\Name\Desktop\darknet\build\darknet\x64"

Evaluate of Results

• For this, you must export the values ​​given in certain iterations in the result.log file to excel as shown images.

• Given example.xlsx excel table and graph.ipynb in evaluate folder for this.

• These are training values. We also need test values. For this, enter obj.data and make valid.txt to test.txt.

Content of the file data/obj.data should be for test results

classes = 1 #Number of classes 
train = C:/Users/Name/Desktop/darknet-master/build/darknet/x64/data/train.txt # Location of train set
valid = C:/Users/Name/Desktop/darknet-master/build/darknet/x64/data/test.txt # Location of test set
names = C:/Users/Name/Desktop/darknet-master/build/darknet/x64/data/obj.names # Location of obj.names
backup = C:/Users/Name/Desktop/darknet-master/build/darknet/x64/backup # Location of saved weights

• Then navigate to the "C:\Users\Name\Desktop\darknet\build\darknet\x64" directory on Prompt and run the code below.

#This creates mAP, recall,precision,f1-score and TP, FP, FN on test images. If you have one class, can create confusion matrix as shown images.

darknet.exe detector map data/obj.data cfg/yolov4-custom.cfg /backup/yolov4-custom_final.weight

#This creates prediction of test images as .txt. And, records in `\x64` folder.

darknet.exe detector test data/obj.data cfg/yolov4-custom.cfg /backup/yolov4-custom_final.weight -dont_show -ext_output < data/test.txt > result.txt

#This creates prediction of specified image and records as .jpg in `\x64` folder.

darknet.exe detector test data/obj.data cfg/yolov4-custom.cfg /backup/yolov4-custom_final.weight C:\Users\Name\Desktop\darknet\build\darknet\x64\test\sample.jpg

src="/docs/predict1.png" alt="Prediction"/>

src="/docs/predict2.png" alt="Prediction2"/>

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Please make sure to update tests as appropriate.

License

MIT