devbox
正经夜光杯
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

yolov4训练步骤_yolov4训练过程

作者：正经夜光杯 | 2024-07-13 13:09:35

赞

踩

yolov4训练过程

一、

打开对应路径        E:\darknet-master\build\darknet\x64

然后在目录下新建个文件夹train

train 文件夹里放这些东西

pictures文件夹是存放图片和xml文件的

举个列子，我这里放红蓝方块的图片和红蓝方块xml文件

可以看到我蓝色方块标签是A3，红色方块标签是A1

 图片序号有时候需要重新命名

所以我们需要用到

双击打开

根据注释修改

1、先将jpg图片重新命名

打开cmd 使用python脚本

回车运行

 

出现ok即为结束

2、重新命名xml文件

还是之前那个change_name.py

 稍作更改继续命令行运行脚本

出现ok就是结束

 这时候我们的pictures文件夹就是这个样子的

可以用labeling打开看看是否对应

对应没有出错。

然后就是获取txt文本

这时候需要用到

文本内部是这样的，我们需要根据注释进行修改 

  1. import xml.etree.ElementTree as ET
  2. import os
  3.  
  4. classes = ["A1","A2","A3","A4","b","L1","L2","L3","L4","L5","L6","L7","L8","L9","L10","L11","L12","L13","L14","L15","L16","L17","L18","L19","L20","M21","M22","M23","M24","M25","M26","M27","M28","M29","M30","M31","M32","M33","M34","M35","M36","TLS","WZ","LZ"] # 这里是你的所有分类的名称
  5. myRoot = r'd:\Desktop\打标好的'  # 这里是你项目的根目录
  6. train_xmlRoot = myRoot + r'/train'
  7. train_txtRoot = myRoot + r'/train'
  8. train_imageRoot = myRoot + r'/train'
  9.  
  10.  
  11. def getFile_name(file_dir):
  12.     L = []
  13.     for root, dirs, files in os.walk(file_dir):
  14.         print(files)
  15.         for file in files:
  16.             if os.path.splitext(file)[1] == '.jpg':
  17.                 L.append(os.path.splitext(file)[0])  # L.append(os.path.join(root, file))
  18.     return L
  19.  
  20.  
  21. def convert(size, box):
  22.     dw = 1. / size[0]
  23.     dh = 1. / size[1]
  24.     x = (box[0] + box[1]) / 2.0
  25.     y = (box[2] + box[3]) / 2.0
  26.     w = box[1] - box[0]
  27.     h = box[3] - box[2]
  28.     x = x * dw
  29.     w = w * dw
  30.     y = y * dh
  31.     h = h * dh
  32.     return (x, y, w, h)
  33.  
  34.  
  35. def convert_annotation(xmlRoot,txtRoot,image_id):
  36.     in_file = open(xmlRoot + '\\%s.xml' % (image_id),encoding='utf-8')
  37.  
  38.     out_file = open(txtRoot + '\\%s.txt' % (image_id), 'w',encoding='utf-8')  # 生成txt格式文件
  39.     tree = ET.parse(in_file)
  40.     root = tree.getroot()
  41.     size = root.find('size')
  42.     w = int(size.find('width').text)
  43.     h = int(size.find('height').text)
  44.  
  45.     for obj in root.iter('object'):
  46.         cls = obj.find('name').text
  47.         if cls not in classes:
  48.             continue
  49.         cls_id = classes.index(cls)
  50.         xmlbox = obj.find('bndbox')
  51.         b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text),
  52.              float(xmlbox.find('ymax').text))
  53.         bb = convert((w, h), b)
  54.         out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')
  55.  
  56.  
  57. image_ids_train = getFile_name(train_imageRoot)
  58. list_file_train = open(myRoot + r'\train.txt', 'w')
  59.  
  60. for image_id in image_ids_train:
  61.     list_file_train.write(train_imageRoot + '\\%s.jpg\n' % (image_id))
  62.     convert_annotation(train_xmlRoot,train_txtRoot,image_id)
  63. list_file_train.close()  # 只生成训练集，自己根据自己情况决定
  64.  
  65.

 就比如这次我们是红蓝方块，他们标签一个是A1一个是A3，就需要把

改为 

 然后项目根目录是E:\darknet-master\build\darknet\x64\train，所以也需要改

改完之后是这样子的

  1. import xml.etree.ElementTree as ET
  2. import os
  3.  
  4. classes = ["A1","A3"] # 这里是你的所有分类的名称
  5. myRoot = r'E:\darknet-master\build\darknet\x64\train'  # 这里是你项目的根目录
  6. train_xmlRoot = myRoot + r'/pictures'       #这是你xml文件夹
  7. train_txtRoot = myRoot + r'/pictures'		#这是你txt文件夹
  8. train_imageRoot = myRoot + r'/pictures'   #这是你image文件夹   我反正都放一起了所以是同一个文件夹
  9.  
  10.  
  11. def getFile_name(file_dir):
  12.     L = []
  13.     for root, dirs, files in os.walk(file_dir):
  14.         print(files)
  15.         for file in files:
  16.             if os.path.splitext(file)[1] == '.jpg':
  17.                 L.append(os.path.splitext(file)[0])  # L.append(os.path.join(root, file))
  18.     return L
  19.  
  20.  
  21. def convert(size, box):
  22.     dw = 1. / size[0]
  23.     dh = 1. / size[1]
  24.     x = (box[0] + box[1]) / 2.0
  25.     y = (box[2] + box[3]) / 2.0
  26.     w = box[1] - box[0]
  27.     h = box[3] - box[2]
  28.     x = x * dw
  29.     w = w * dw
  30.     y = y * dh
  31.     h = h * dh
  32.     return (x, y, w, h)
  33.  
  34.  
  35. def convert_annotation(xmlRoot,txtRoot,image_id):
  36.     in_file = open(xmlRoot + '\\%s.xml' % (image_id),encoding='utf-8')
  37.  
  38.     out_file = open(txtRoot + '\\%s.txt' % (image_id), 'w',encoding='utf-8')  # 生成txt格式文件
  39.     tree = ET.parse(in_file)
  40.     root = tree.getroot()
  41.     size = root.find('size')
  42.     w = int(size.find('width').text)
  43.     h = int(size.find('height').text)
  44.  
  45.     for obj in root.iter('object'):
  46.         cls = obj.find('name').text
  47.         if cls not in classes:
  48.             continue
  49.         cls_id = classes.index(cls)
  50.         xmlbox = obj.find('bndbox')
  51.         b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text),
  52.              float(xmlbox.find('ymax').text))
  53.         bb = convert((w, h), b)
  54.         out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')
  55.  
  56.  
  57. image_ids_train = getFile_name(train_imageRoot)
  58. list_file_train = open(myRoot + r'\train.txt', 'w')
  59.  
  60. for image_id in image_ids_train:
  61.     list_file_train.write(train_imageRoot + '\\%s.jpg\n' % (image_id))
  62.     convert_annotation(train_xmlRoot,train_txtRoot,image_id)
  63. list_file_train.close()  # 只生成训练集，自己根据自己情况决定
  64.  
  65.

然后运行脚本

 

 然后你的pictures文件夹里就会有txt文本了

然后就是配置 coco.data coco.names  还有 yolov4.cfg

先配置coco.names

保存即可

然后配置 yolov4.cfg

  1. [net]
  2. # Testing
  3. #batch=1
  4. #subdivisions=1
  5. # Training
  6. batch=64 #每64个样本更新一次参数
  7. subdivisions=16 #将batch分割为64个子batch 训练中出现out of memory 则改大到64或更大
  8. width=512 #width和height为32倍数 越大训练的效果越好 但是不能超过图像高宽 验证耗时也越大
  9. height=512
  10. channels=3
  11. momentum=0.9
  12. decay=0.0005
  13. angle=0
  14. saturation = 1.5 #饱和度
  15. exposure = 1.5     #曝光
  16. hue=.01         #色调
  17.  
  18. learning_rate=0.0001
  19.  
  20. burn_in=1000
  21.  
  22. max_batches = 20000 #训练的最大批次数 值越大耗时越久 class*2000；不能小于图片数量；不能小于6k
  23. policy=steps
  24. steps=40000,45000 #max_batches*0.8和0.9的值
  25. scales=.1,.1
  26.  
  27.  
  28. #weights_reject_freq=1001
  29. #ema_alpha=0.9998
  30. #equidistant_point=1000
  31. #num_sigmas_reject_badlabels=3
  32. #badlabels_rejection_percentage=0.2
  33.  
  34.  
  35. [convolutional]
  36. batch_normalize=1
  37. filters=32
  38. size=3
  39. stride=2
  40. pad=1
  41. activation=leaky
  42.  
  43. [convolutional]
  44. batch_normalize=1
  45. filters=64
  46. size=3
  47. stride=2
  48. pad=1
  49. activation=leaky
  50.  
  51. [convolutional]
  52. batch_normalize=1
  53. filters=64
  54. size=3
  55. stride=1
  56. pad=1
  57. activation=leaky
  58.  
  59. [route]
  60. layers=-1
  61. groups=2
  62. group_id=1
  63.  
  64. [convolutional]
  65. batch_normalize=1
  66. filters=32
  67. size=3
  68. stride=1
  69. pad=1
  70. activation=leaky
  71.  
  72. [convolutional]
  73. batch_normalize=1
  74. filters=32
  75. size=3
  76. stride=1
  77. pad=1
  78. activation=leaky
  79.  
  80. [route]
  81. layers = -1,-2
  82.  
  83. [convolutional]
  84. batch_normalize=1
  85. filters=64
  86. size=1
  87. stride=1
  88. pad=1
  89. activation=leaky
  90.  
  91. [route]
  92. layers = -6,-1
  93.  
  94. [maxpool]
  95. size=2
  96. stride=2
  97.  
  98. [convolutional]
  99. batch_normalize=1
  100. filters=128
  101. size=3
  102. stride=1
  103. pad=1
  104. activation=leaky
  105.  
  106. [route]
  107. layers=-1
  108. groups=2
  109. group_id=1
  110.  
  111. [convolutional]
  112. batch_normalize=1
  113. filters=64
  114. size=3
  115. stride=1
  116. pad=1
  117. activation=leaky
  118.  
  119. [convolutional]
  120. batch_normalize=1
  121. filters=64
  122. size=3
  123. stride=1
  124. pad=1
  125. activation=leaky
  126.  
  127. [route]
  128. layers = -1,-2
  129.  
  130. [convolutional]
  131. batch_normalize=1
  132. filters=128
  133. size=1
  134. stride=1
  135. pad=1
  136. activation=leaky
  137.  
  138. [route]
  139. layers = -6,-1
  140.  
  141. [maxpool]
  142. size=2
  143. stride=2
  144.  
  145. [convolutional]
  146. batch_normalize=1
  147. filters=256
  148. size=3
  149. stride=1
  150. pad=1
  151. activation=leaky
  152.  
  153. [route]
  154. layers=-1
  155. groups=2
  156. group_id=1
  157.  
  158. [convolutional]
  159. batch_normalize=1
  160. filters=128
  161. size=3
  162. stride=1
  163. pad=1
  164. activation=leaky
  165.  
  166. [convolutional]
  167. batch_normalize=1
  168. filters=128
  169. size=3
  170. stride=1
  171. pad=1
  172. activation=leaky
  173.  
  174. [route]
  175. layers = -1,-2
  176.  
  177. [convolutional]
  178. batch_normalize=1
  179. filters=256
  180. size=1
  181. stride=1
  182. pad=1
  183. activation=leaky
  184.  
  185. [route]
  186. layers = -6,-1
  187.  
  188. [maxpool]
  189. size=2
  190. stride=2
  191.  
  192. [convolutional]
  193. batch_normalize=1
  194. filters=512
  195. size=3
  196. stride=1
  197. pad=1
  198. activation=leaky
  199.  
  200. ##################################
  201.  
  202. [convolutional]
  203. batch_normalize=1
  204. filters=256
  205. size=1
  206. stride=1
  207. pad=1
  208. activation=leaky
  209.  
  210. [convolutional]
  211. batch_normalize=1
  212. filters=512
  213. size=3
  214. stride=1
  215. pad=1
  216. activation=leaky
  217.  
  218. [convolutional]
  219. size=1
  220. stride=1
  221. pad=1
  222. filters=51 #找离yolo最近的 （classes+5）*3（下面的mask的值）
  223. activation=linear
  224.  
  225.  
  226.  
  227. [yolo]
  228. mask = 3,4,5
  229. anchors = 10,14,  23,27,  37,58,  81,82,  135,169,  344,319
  230. classes=12 #总共有几个类别
  231. num=6
  232. jitter=.3
  233. scale_x_y = 1.05
  234. cls_normalizer=1.0
  235. iou_normalizer=0.07
  236. iou_loss=ciou
  237. ignore_thresh = .7
  238. truth_thresh = 1
  239. random=0
  240. resize=1.5
  241. nms_kind=greedynms
  242. beta_nms=0.6
  243. #new_coords=1
  244. #scale_x_y = 2.0
  245.  
  246. [route]
  247. layers = -4
  248.  
  249. [convolutional]
  250. batch_normalize=1
  251. filters=128
  252. size=1
  253. stride=1
  254. pad=1
  255. activation=leaky
  256.  
  257. [upsample]
  258. stride=2
  259.  
  260. [route]
  261. layers = -1, 23
  262.  
  263. [convolutional]
  264. batch_normalize=1
  265. filters=256
  266. size=3
  267. stride=1
  268. pad=1
  269. activation=leaky
  270.  
  271. [convolutional]
  272. size=1
  273. stride=1
  274. pad=1
  275. filters=51 #同理
  276. activation=linear
  277.  
  278. [yolo]
  279. mask = 1,2,3
  280. anchors = 10,14,  23,27,  37,58,  81,82,  135,169,  344,319
  281. classes=12 #这边也要设置
  282. num=6
  283. jitter=.3
  284. scale_x_y = 1.05
  285. cls_normalizer=1.0
  286. iou_normalizer=0.07
  287. iou_loss=ciou
  288. ignore_thresh = .7
  289. truth_thresh = 1
  290. random=0
  291. resize=1.5
  292. nms_kind=greedynms
  293. beta_nms=0.6
  294. #new_coords=1
  295. #scale_x_y = 2.0

根据注释修改

比如红蓝方块就两类，训练批次数=4000，steps=3200,3600 #max_batches*0.8和0.9的值

classes=2 然后改 距离yolo最近的 filters=21 改完后是这样子的

  1. [net]
  2. # Testing
  3. #batch=1
  4. #subdivisions=1
  5. # Training
  6. batch=64 #每64个样本更新一次参数
  7. subdivisions=16 #将batch分割为64个子batch 训练中出现out of memory 则改大到64或更大
  8. width=512 #width和height为32倍数 越大训练的效果越好 但是不能超过图像高宽 验证耗时也越大
  9. height=512
  10. channels=3
  11. momentum=0.9
  12. decay=0.0005
  13. angle=0
  14. saturation = 1.5 #饱和度
  15. exposure = 1.5     #曝光
  16. hue=.01         #色调
  17.  
  18. learning_rate=0.0001
  19.  
  20. burn_in=1000
  21.  
  22. max_batches = 4000 #训练的最大批次数 值越大耗时越久 class*2000；不能小于图片数量；不能小于6k
  23. policy=steps
  24. steps=3200,3600 #max_batches*0.8和0.9的值
  25. scales=.1,.1
  26.  
  27.  
  28. #weights_reject_freq=1001
  29. #ema_alpha=0.9998
  30. #equidistant_point=1000
  31. #num_sigmas_reject_badlabels=3
  32. #badlabels_rejection_percentage=0.2
  33.  
  34.  
  35. [convolutional]
  36. batch_normalize=1
  37. filters=32
  38. size=3
  39. stride=2
  40. pad=1
  41. activation=leaky
  42.  
  43. [convolutional]
  44. batch_normalize=1
  45. filters=64
  46. size=3
  47. stride=2
  48. pad=1
  49. activation=leaky
  50.  
  51. [convolutional]
  52. batch_normalize=1
  53. filters=64
  54. size=3
  55. stride=1
  56. pad=1
  57. activation=leaky
  58.  
  59. [route]
  60. layers=-1
  61. groups=2
  62. group_id=1
  63.  
  64. [convolutional]
  65. batch_normalize=1
  66. filters=32
  67. size=3
  68. stride=1
  69. pad=1
  70. activation=leaky
  71.  
  72. [convolutional]
  73. batch_normalize=1
  74. filters=32
  75. size=3
  76. stride=1
  77. pad=1
  78. activation=leaky
  79.  
  80. [route]
  81. layers = -1,-2
  82.  
  83. [convolutional]
  84. batch_normalize=1
  85. filters=64
  86. size=1
  87. stride=1
  88. pad=1
  89. activation=leaky
  90.  
  91. [route]
  92. layers = -6,-1
  93.  
  94. [maxpool]
  95. size=2
  96. stride=2
  97.  
  98. [convolutional]
  99. batch_normalize=1
  100. filters=128
  101. size=3
  102. stride=1
  103. pad=1
  104. activation=leaky
  105.  
  106. [route]
  107. layers=-1
  108. groups=2
  109. group_id=1
  110.  
  111. [convolutional]
  112. batch_normalize=1
  113. filters=64
  114. size=3
  115. stride=1
  116. pad=1
  117. activation=leaky
  118.  
  119. [convolutional]
  120. batch_normalize=1
  121. filters=64
  122. size=3
  123. stride=1
  124. pad=1
  125. activation=leaky
  126.  
  127. [route]
  128. layers = -1,-2
  129.  
  130. [convolutional]
  131. batch_normalize=1
  132. filters=128
  133. size=1
  134. stride=1
  135. pad=1
  136. activation=leaky
  137.  
  138. [route]
  139. layers = -6,-1
  140.  
  141. [maxpool]
  142. size=2
  143. stride=2
  144.  
  145. [convolutional]
  146. batch_normalize=1
  147. filters=256
  148. size=3
  149. stride=1
  150. pad=1
  151. activation=leaky
  152.  
  153. [route]
  154. layers=-1
  155. groups=2
  156. group_id=1
  157.  
  158. [convolutional]
  159. batch_normalize=1
  160. filters=128
  161. size=3
  162. stride=1
  163. pad=1
  164. activation=leaky
  165.  
  166. [convolutional]
  167. batch_normalize=1
  168. filters=128
  169. size=3
  170. stride=1
  171. pad=1
  172. activation=leaky
  173.  
  174. [route]
  175. layers = -1,-2
  176.  
  177. [convolutional]
  178. batch_normalize=1
  179. filters=256
  180. size=1
  181. stride=1
  182. pad=1
  183. activation=leaky
  184.  
  185. [route]
  186. layers = -6,-1
  187.  
  188. [maxpool]
  189. size=2
  190. stride=2
  191.  
  192. [convolutional]
  193. batch_normalize=1
  194. filters=512
  195. size=3
  196. stride=1
  197. pad=1
  198. activation=leaky
  199.  
  200. ##################################
  201.  
  202. [convolutional]
  203. batch_normalize=1
  204. filters=256
  205. size=1
  206. stride=1
  207. pad=1
  208. activation=leaky
  209.  
  210. [convolutional]
  211. batch_normalize=1
  212. filters=512
  213. size=3
  214. stride=1
  215. pad=1
  216. activation=leaky
  217.  
  218. [convolutional]
  219. size=1
  220. stride=1
  221. pad=1
  222. filters=21 #找离yolo最近的 （classes+5）*3（下面的mask的值）
  223. activation=linear
  224.  
  225.  
  226.  
  227. [yolo]
  228. mask = 3,4,5
  229. anchors = 10,14,  23,27,  37,58,  81,82,  135,169,  344,319
  230. classes=2 #总共有几个类别
  231. num=6
  232. jitter=.3
  233. scale_x_y = 1.05
  234. cls_normalizer=1.0
  235. iou_normalizer=0.07
  236. iou_loss=ciou
  237. ignore_thresh = .7
  238. truth_thresh = 1
  239. random=0
  240. resize=1.5
  241. nms_kind=greedynms
  242. beta_nms=0.6
  243. #new_coords=1
  244. #scale_x_y = 2.0
  245.  
  246. [route]
  247. layers = -4
  248.  
  249. [convolutional]
  250. batch_normalize=1
  251. filters=128
  252. size=1
  253. stride=1
  254. pad=1
  255. activation=leaky
  256.  
  257. [upsample]
  258. stride=2
  259.  
  260. [route]
  261. layers = -1, 23
  262.  
  263. [convolutional]
  264. batch_normalize=1
  265. filters=256
  266. size=3
  267. stride=1
  268. pad=1
  269. activation=leaky
  270.  
  271. [convolutional]
  272. size=1
  273. stride=1
  274. pad=1
  275. filters=21 #同理
  276. activation=linear
  277.  
  278. [yolo]
  279. mask = 1,2,3
  280. anchors = 10,14,  23,27,  37,58,  81,82,  135,169,  344,319
  281. classes=2 #这边也要设置
  282. num=6
  283. jitter=.3
  284. scale_x_y = 1.05
  285. cls_normalizer=1.0
  286. iou_normalizer=0.07
  287. iou_loss=ciou
  288. ignore_thresh = .7
  289. truth_thresh = 1
  290. random=0
  291. resize=1.5
  292. nms_kind=greedynms
  293. beta_nms=0.6
  294. #new_coords=1
  295. #scale_x_y = 2.0

这就配置好了yolov4.cfg

最后配置coco.data

然后在

 路径下打开终端输入训练指令

darknet.exe detector train train/coco.data train/yolov4.cfg train/yolov4-tiny.conv.29

等待训练结束就行。然后就会在你之前的

这个位置出现权重文件，之后用这个权重文件检测图片，是对应目标标签即可。 

1、识别图片格式：./darknet.exe detector test <.data文件位置> <.cfg文件位置> <权重文件位置> <识别文件位置>
darknet.exe detector test train/coco.data train/yolov4.cfg train/backup/yolov4_last.weights train/pictures/1.jpg

就会有下面这种结果，标签和物体对上了，置信度也高。


2、调用电脑摄像头识别
darknet.exe detector demo  D:/desktop/shape/shape.data D:/desktop/shape/yolov4-tiny.cfg D:/desktop/shape/F.weights  -c 0

3、继续上次模型的训练

darknet.exe detector train train/coco.data train/yolov4.cfg train/yolov4_last.weights

就是权重从预训练模型yolov4-tiny.conv.29变为之前训练得到的权重yolov4_last.weights，如果报错的话，更改cfg文件里max_batches  #训练的最大批次数 让cfg文件和之前训练的不一样就可以了。

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/正经夜光杯/article/detail/819660
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号