Train your dataset with yolov4

Environmental Science ：ubuntu16.04 cuda8.0 cudnn6.0.1 GT1070
1,GitHub：https://github.com/AlexeyAB/darknet download

2, compile ;
① modify makefile file

GPU=1
CUDNN=1
CUDNN_HALF=0# Here, if the computing power of the graphics card is less than 7.0, No need to change to 1
OPENCV=1
AVX=0
OPENMP=1
LIBSO=1
’‘’
DEBUG=1
‘’‘
NVCC=/usr/local/cuda-8.0/bin/nvcc# Change to your own path 

②make
remarks ： Common mistakes can be found online , It's just a version problem

3, Download pre training model ：GitHub There is

4, A simple test ：

./darknet detect cfg/yolov4.cfg yolov4.weights data/dog.jpg

The test results ：

Done! Loaded 162 layers from weights-file 
data/dog.jpg: Predicted in 67.411000 milli-seconds.
bicycle: 92%
dog: 98%
truck: 92%
pottedplant: 33%

5, Train your data set ：
① Collecting data sets , This is collected according to their own project requirements , There are the vast majority of data sets needed on the Internet , Download classification .

② For data set preprocessing ： You can change the names of many pictures to VOC The format of the dataset , namely 000**.jpg. Modify the code as follows ：

import os
path=input(' Please enter the file path ( Add at the end /)：')       
# Get all the files in this directory , Put it in the list 
f=os.listdir(path)
n=0
for i in f:
    # Set old file name （ Is the path + file name ）
    oldname=path+f[n]
    # Set a new file name 
    newname=path+'000'+str(n+1)+'.jpg'
    # use os Module rename Method to rename the file 
    os.rename(oldname,newname)
    print(oldname,'======>',newname)
    n+=1

③ Mark data sets ： I use labelImg Tools for marking , Enclosed GitHub link ：https://github.com/tzutalin/labelImg Very simple installation and operation .

④ Expand the data set ： Generally used to flip 、 shear 、 rotate , According to the needs of their own projects, expand to the required number of data sets .（ More code , Inconvenient to paste , You can have a private chat with the friends you need ）

⑤ Change the dataset to VOC Dataset format ：
One 、 Dataset folder format
VOC The format of dataset folder is generally ：

      ---VOC***（ Can be the date of the day in person VOC531）
          ----Annotations（ It's all marked inside .xml file ）
          ----ImageSets
              ----Main
          ----JPEGImages（ Inside are all pictures ）

Two 、 Split data sets
It is generally divided into training sets 、 Verification set 、 Test set , The division proportion can be adjusted by yourself （ I use 9-1） The code is as follows

import os  
import random 
xmlfilepath=r'/home/nph/darknet-master/VOC531/Annotations'   #change xml path                        
saveBasePath=r"/home/nph/darknet-master/"                        #change base path
trainval_percent=0.9                                             #adjust trainval percentage    
train_percent=0.9                                              #adjust train percentage        
total_xml = os.listdir(xmlfilepath)
num=len(total_xml)    
list=range(num)    
tv=int(num*trainval_percent)    
tr=int(tv*train_percent)    
trainval= random.sample(list,tv)    
train=random.sample(trainval,tr)    
  
print("train and val size",tv)  
print("traub suze",tr)  
ftrainval = open(os.path.join(saveBasePath,'VOC531/ImageSets/Main/trainval.txt'), 'w')    #change your path
ftest = open(os.path.join(saveBasePath,'VOC531/ImageSets/Main/test.txt'), 'w')    #
ftrain = open(os.path.join(saveBasePath,'VOC531/ImageSets/Main/train.txt'), 'w')    #
fval = open(os.path.join(saveBasePath,'VOC531/ImageSets/Main/val.txt'), 'w')    #
  
for i  in list:    
    name=total_xml[i][:-4]+'\n'    
    if i in trainval:    
        ftrainval.write(name)    
        if i in train:    
            ftrain.write(name)    
        else:    
            fval.write(name)    
    else:    
        ftest.write(name)    
    
ftrainval.close()    
ftrain.close()    
fval.close()    
ftest .close()  

The path that needs to be changed has been marked , After running in Main Four will be generated under the file txt file ： Respectively trainval.txt、train.txt、test.txt and val.txt. Inside is the name of the corresponding category picture , That is, the previous named number 000***（ There is no suffix ） Pictured ：

This is the label of the picture .

3、 ... and 、 Extract label and coordinate information .
This step is to combine the label information with xml Extract the coordinate information marked in the file , The code is as follows ：

import xml.etree.ElementTree as ET
import pickle
import os
from os import listdir, getcwd
from os.path import join

sets=[('531', 'train'), ('531', 'val'), ('531', 'test')] # Change it to your own VOC*** Digital information after 

classes = ["supervisor_white", "technicist_blue", "worker_yellow", "manager_red" ] # Classify your data set in quotation marks 


def convert(size, box):
    dw = 1./size[0]
    dh = 1./size[1]
    x = (box[0] + box[1])/2.0
    y = (box[2] + box[3])/2.0
    w = box[1] - box[0]
    h = box[3] - box[2]
    x = x*dw
    w = w*dw
    y = y*dh
    h = h*dh
    return (x,y,w,h)

def convert_annotation(year, image_id):
    in_file = open('VOC%s/Annotations/%s.xml'%(year, image_id))
    out_file = open('VOC%s/labels/%s.txt'%(year, image_id), 'w')
    tree=ET.parse(in_file)
    root = tree.getroot()
    size = root.find('size')
    w = int(size.find('width').text)
    h = int(size.find('height').text)

    for obj in root.iter('object'):
        difficult = obj.find('difficult').text
        cls = obj.find('name').text
        if cls not in classes or int(difficult) == 1:
            continue
        cls_id = classes.index(cls)
        xmlbox = obj.find('bndbox')
        b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text), float(xmlbox.find('ymax').text))
        bb = convert((w,h), b)
        out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')

wd = getcwd()

for year, image_set in sets:
    if not os.path.exists('VOC%s/labels/'%(year)):
        os.makedirs('VOC%s/labels/'%(year))
    image_ids = open('VOC%s/ImageSets/Main/%s.txt'%(year, image_set)).read().strip().split()
    list_file = open('%s_%s.txt'%(year, image_set), 'w')
    for image_id in image_ids:
        list_file.write('%s/VOC%s/JPEGImages/%s.jpg\n'%(wd, year, image_id))
        convert_annotation(year, image_id)
    list_file.close()

Everything that needs to be modified has been marked , After running, it will be in VOC*** A... Is generated under the file labels file , Inside is txt file , It stores the information of each picture marking box , Pictured ：

stay VOC*** Folder directory will generate three txt file , It is generally named ：***（VOC Back number ）_train(test/val).txt, Inside is the path of photos , Pictured ：

These three txt Files can be put into VOC In the file , It is convenient to fill in the address during the later training .

thus ,VOC The format of data set has been completely changed , Next is the change YOLOv4 Network files for training .

⑥ Modify the network file ：
Major changes cfg names data file
take darknet-master/cfg/yolov4-custom.cfg、darknet-master/cfg/coco.data and darknet-master/data/coco.names Copy the file to your own VOC*** Under the document
One 、cfg file
modify cfg file （ I changed it to 531.cfg）

[net]
# Testing # The training process will test Three lines of comment 
#batch=1
#subdivisions=1
# Training
batch=32 # Modify according to your own graphics card memory , It's usually 16,32,64
subdivisions=8 # Rewrite according to the uplink , It's usually 4,8,16
width=416 # The size can be modified by yourself , It's usually 16 Multiple 
height=416
'''
learning_rate=0.001
burn_in=1000
max_batches = 12000 # It's usually classess*2000, Training category *2000
policy=steps
steps=9600,10800 #max_batches Of 0.8 and 0.9 times 
scales=.1,.1







[convolutional]
size=1
stride=1
pad=1
filters=24 #（classes+5）)3
activation=linear


[yolo]
mask = 6,7,8
anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
classes=3 # Number of training categories 
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
scale_x_y = 1.05
iou_thresh=0.213
cls_normalizer=1.0
iou_normalizer=0.07
iou_loss=ciou
nms_kind=greedynms
beta_nms=0.6
max_delta=5
#3 individual yolo Layers need to be changed , All in cfg The lowest level of the file 

Two 、 modify data file
open 531.data file , Revised as follows ：

classes= 3
train  = /home/u/qingxin/potgraduate/Fire/darknet-master/Fire/Multi/VOC531/531_train.txt # Here is the image path file of the training generated above , Change to your own path 
valid  = /home/u/qingxin/potgraduate/Fire/darknet-master/Fire/Multi/VOC531/531_val.txt # Verify the image path file , Change to your own path 
names = /home/u/qingxin/potgraduate/Fire/darknet-master/Fire/Multi/VOC531/3fire.names #names File path 
backup = /home/u/qingxin/potgraduate/Fire/darknet-master/Fire/Multi/VOC531/backup # Training model storage address during training , You can modify it yourself 

3、 ... and 、 modify names file
open 531.names file , Revised as follows ：

supervisor_white
technicist_blue
worker_yellow
manager_red

That is, the name of your dataset classification

thus , The network file has been modified , You can train .

⑦ Training
stay darknet-master Enter instructions under the path

./darknet detector train VOC531/531.data VOC531/531.cfg yolov4.conv.137 -gpus 1,2 -dont_show

They are training instructions -data file -cfg file - Pre training model - Appoint GPU
You can add logs

./darknet detector train VOC531/531.data VOC531/531.cfg yolov4.conv.137 -gpus 1,2 -dont_show 2>&1 | tee VOC531/531.log

Start training ～

Test to be updated ～