yolo v4 模型训练方法(节选自yolo v4 在github 上的文档原文)

How to train (to detect your custom objects)

(to train old Yolo v2 yolov2-voc.cfgyolov2-tiny-voc.cfgyolo-voc.cfgyolo-voc.2.0.cfg, … click by the link)

Training Yolo v4 (and v3):

  1. For training cfg/yolov4-custom.cfg download the pre-trained weights-file (162 MB): yolov4.conv.137 (Google drive mirror yolov4.conv.137 )
  2. Create file yolo-obj.cfg with the same content as in yolov4-custom.cfg (or copy yolov4-custom.cfg to yolo-obj.cfg) and:

So if classes=1 then should be filters=18. If classes=2 then write filters=21(Do not write in the cfg-file: filters=(classes + 5)x3)

(Generally filters depends on the classescoords and number of masks, i.e. filters=(classes + coords + 1)*<number of mask>, where mask is indices of anchors. If mask is absence, then filters=(classes + coords + 1)*num)

So for example, for 2 objects, your file yolo-obj.cfg should differ from yolov4-custom.cfg in such lines in each of 3 [yolo]-layers:

[convolutional]
filters=21

[region]
classes=2
  1. Create file obj.names in the directory build\darknet\x64\data\, with objects names – each in new line
  2. Create file obj.data in the directory build\darknet\x64\data\, containing (where classes = number of objects):
classes = 2
train  = data/train.txt
valid  = data/test.txt
names = data/obj.names
backup = backup/
  1. Put image-files (.jpg) of your objects in the directory build\darknet\x64\data\obj\
  2. You should label each object on images from your dataset. Use this visual GUI-software for marking bounded boxes of objects and generating annotation files for Yolo v2 & v3: https://github.com/AlexeyAB/Yolo_mark

It will create .txt-file for each .jpg-image-file – in the same directory and with the same name, but with .txt-extension, and put to file: object number and object coordinates on this image, for each object in new line:

<object-class> <x_center> <y_center> <width> <height>

Where:

  • <object-class> – integer object number from 0 to (classes-1)

  • <x_center> <y_center> <width> <height> – float values relative to width and height of image, it can be equal from (0.0 to 1.0]

  • for example: <x> = <absolute_x> / <image_width> or <height> = <absolute_height> / <image_height>

  • attention: <x_center> <y_center> – are center of rectangle (are not top-left corner)

    For example for img1.jpg you will be created img1.txt containing:

    1 0.716797 0.395833 0.216406 0.147222
    0 0.687109 0.379167 0.255469 0.158333
    1 0.420312 0.395833 0.140625 0.166667
    
  1. Create file train.txt in directory build\darknet\x64\data\, with filenames of your images, each filename in new line, with path relative to darknet.exe, for example containing:
data/obj/img1.jpg
data/obj/img2.jpg
data/obj/img3.jpg
  1. Download pre-trained weights for the convolutional layers and put to the directory build\darknet\x64

  2. Start training by using the command line: darknet.exe detector train data/obj.data yolo-obj.cfg yolov4.conv.137

    To train on Linux use command: ./darknet detector train data/obj.data yolo-obj.cfg yolov4.conv.137 (just use ./darknet instead of darknet.exe)

    • (file yolo-obj_last.weights will be saved to the build\darknet\x64\backup\ for each 100 iterations)
    • (file yolo-obj_xxxx.weights will be saved to the build\darknet\x64\backup\ for each 1000 iterations)
    • (to disable Loss-Window use darknet.exe detector train data/obj.data yolo-obj.cfg yolov4.conv.137 -dont_show, if you train on computer without monitor like a cloud Amazon EC2)
    • (to see the mAP & Loss-chart during training on remote server without GUI, use command darknet.exe detector train data/obj.data yolo-obj.cfg yolov4.conv.137 -dont_show -mjpeg_port 8090 -map then open URL http://ip-address:8090 in Chrome/Firefox browser)

8.1. For training with mAP (mean average precisions) calculation for each 4 Epochs (set valid=valid.txt or train.txt in obj.data file) and run: darknet.exe detector train data/obj.data yolo-obj.cfg yolov4.conv.137 -map

  1. After training is complete – get result yolo-obj_final.weights from path build\darknet\x64\backup\

    • After each 100 iterations you can stop and later start training from this point. For example, after 2000 iterations you can stop training, and later just start training using: darknet.exe detector train data/obj.data yolo-obj.cfg backup\yolo-obj_2000.weights

    (in the original repository https://github.com/pjreddie/darknet the weights-file is saved only once every 10 000 iterations if(iterations > 1000))

    • Also you can get result earlier than all 45000 iterations.

Note: If during training you see nan values for avg (loss) field – then training goes wrong, but if nan is in some other lines – then training goes well.

Note: If you changed width= or height= in your cfg-file, then new width and height must be divisible by 32.

Note: After training use such command for detection: darknet.exe detector test data/obj.data yolo-obj.cfg yolo-obj_8000.weights

Note: if error Out of memory occurs then in .cfg-file you should increase subdivisions=16, 32 or 64: link

tensorflow 深度学习训练代码

# coding: utf-8
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import os

os.environ[‘TF_CPP_MIN_LOG_LEVEL’] = ‘2’

mnist = input_data.read_data_sets(‘mnist_data’, one_hot=True)

#初始化过滤器
def weight_variable(shape):
return tf.Variable(tf.truncated_normal(shape, stddev=0.1))

#初始化偏置,初始化时,所有值是0.1
def bias_variable(shape):
return tf.Variable(tf.constant(0.1, shape=shape))

#卷积运算,strides表示每一维度滑动的步长,一般strides[0]=strides[3]=1
#第四个参数可选”Same”或”VALID”,“Same”表示边距使用全0填充
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding=”SAME”)

#池化运算
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding=”SAME”)

#创建x占位符,用于临时存放MNIST图片的数据,
# [None, 784]中的None表示不限长度,而784则是一张图片的大小(28×28=784)
x = tf.placeholder(tf.float32, [None, 784], name=’input’)
#y_存的是实际图像的标签,即对应于每张输入图片实际的值
y_ = tf.placeholder(tf.float32, [None, 10])

#将图片从784维向量重新还原为28×28的矩阵图片,
# 原因参考卷积神经网络模型图,最后一个参数代表深度,
# 因为MNIST是黑白图片,所以深度为1,
# 第一个参数为-1,表示一维的长度不限定,这样就可以灵活设置每个batch的训练的个数了
x_image = tf.reshape(x, [-1, 28, 28, 1])

#第一层卷积
#将过滤器设置成5×5×1的矩阵,
#其中5×5表示过滤器大小,1表示深度,因为MNIST是黑白图片只有一层。所以深度为1
#32表示我们要创建32个大小5×5×1的过滤器,经过卷积后算出32个特征图(每个过滤器得到一个特征图),即输出深度为64
W_conv1 = weight_variable([5, 5, 1, 32])
#有多少个特征图就有多少个偏置
b_conv1 = bias_variable([32])
#使用conv2d函数进行卷积计算,然后再用ReLU作为激活函数
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
#卷积以后再经过池化操作
h_pool1 = max_pool_2x2(h_conv1)

#第二层卷积
#因为经过第一层卷积运算后,输出的深度为32,所以过滤器深度和下一层输出深度也做出改变
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

#全连接层
#经过两层卷积后,图片的大小为7×7(第一层池化后输出为(28/2)×(28/2),
#第二层池化后输出为(14/2)×(14/2)),深度为64,
#我们在这里加入一个有1024个神经元的全连接层,所以权重W的尺寸为[7 * 7 * 64, 1024]
W_fc1 = weight_variable([7 * 7 * 64, 1024])
#偏置的个数和权重的个数一致
b_fc1 = bias_variable([1024])
#这里将第二层池化后的张量(长:7 宽:7 深度:64) 变成向量(跟上一节的Softmax模型的输入一样了)
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
#使用ReLU激活函数
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

#dropout
#为了减少过拟合,我们在输出层之前加入dropout
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

#输出层
#全连接层输入的大小为1024,而我们要得到的结果的大小是10(0~9),
# 所以这里权重W的尺寸为[1024, 10]
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
#最后都要经过Softmax函数将输出转化为概率问题
y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2, name=’output’)

#损失函数和损失优化
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv)))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

#测试准确率,跟Softmax回归模型的一样
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

#将训练结果保存,如果不保存我们这次训练结束后的结果也随着程序运行结束而释放了
savePath = ‘./mnist_conv/’
saveFile = savePath + ‘mnist_conv.ckpt’
if os.path.exists(savePath) == False:
os.mkdir(savePath)

saver = tf.train.Saver()

#开始训练
with tf.Session() as sess:
#初始化所有变量
sess.run(tf.global_variables_initializer())
#训练两万次
for i in range(20000):
#每次获取50张图片数据和对应的标签
batch = mnist.train.next_batch(50)
#每训练100次,我们打印一次训练的准确率
if i % 100 == 0:
train_accuracy =sess.run(accuracy, feed_dict={x:batch[0], y_:batch[1], keep_prob:1.0})
print(“step %d, training accuracy %g” % (i, train_accuracy))
#这里是真的训练,将数据传入
sess.run(train_step, feed_dict={x:batch[0], y_:batch[1], keep_prob:0.5})

print (“end train, start testing…”)

mean_value = 0.0
for i in range(mnist.test.labels.shape[0]):
batch = mnist.test.next_batch(50)
train_accuracy = sess.run(accuracy, feed_dict={x: batch[0], y_: batch[1], keep_prob: 1.0})
mean_value += train_accuracy

print(“test accuracy %g” % (mean_value / mnist.test.labels.shape[0]))
# #训练结束后,我们使用mnist.test在测试最后的准确率
# print(“test accuracy %g” % sess.run(accuracy, feed_dict={x:mnist.test.images, y_:mnist.test.labels, keep_prob:1.0}))

# 最后,将会话保存下来
saver.save(sess, saveFile)