kerasyolo3master使用记录

你好&＃xff01; 这是你第一次使用 Markdown编辑器 所展示的欢迎页。如果你想学习如何使用Markdown编辑器, 可以仔细阅读这篇文章&＃xff0c;了解一下Markdown的基本语法知识。

源码下载 环境配置 快速测试

在macos windows liunx系统都测试过&＃xff0c;本文采用macos

1. 项目代码&＃xff1a;https://github.com/qqwweee/keras-yolo3

2. 下载yolo3weights &＃xff1a;https://pjreddie.com/darknet/yolo/

   将yolo3weights文件夹放到keras-yolo3-master文件夹里

3. terminal cd 到keras-yolo3-master文件夹
   生成现在权重下h5文件&＃xff1a;
   python3 convert.py yolov3.cfg yolov3.weights model_data/yolo.h5

4. 进行快速测试&＃xff0c;看看能不能用。
   terminal cd到keras-yolo3-master目录下
   python3 yolo_video.py --image
   之后让你输入图片路径&＃xff1a;&＃xff08;若将图片放在keras-yolo3-master文件夹下&＃xff0c;直接输入相对地址即可&＃xff09;

制作自己的项目

1. 下载VOC2007数据集
   下载地址&＃xff1a; https://pjreddie.com/projects/pascal-voc-dataset-mirror/

2. 这里面用到的文件夹是Annotation、ImageSets和JPEGImages

   其中文件夹Annotation中主要存放xml文件&＃xff0c;每一个xml对应一张图像在这里插入图片描述&＃xff1b;而ImageSets我们只需要用到Main文件夹&＃xff0c;这里面存放的是一些文本文件&＃xff0c;通常为train.txt、test.txt等&＃xff0c;该文本文件里面的内容是需要用来训练或测试的图像的名字&＃xff1b;JPEGImages文件夹中放我们已按统一规则命名好的原始图像。
   

将自己数据转移到对应目录

// 将自己原始图片&＃xff0c;标注过的图片放到VOC数据集相应位置&＃xff0c;并生成训练集测试集验证集
//生成训练集测试集验证集对应txt文件&＃xff0c;放入相应位置
#%%import os
import random
import shutil #拷贝文件并移动的库path &＃61; &＃39;/code/kaggle/wechat/&＃39; #自己的数据路径img &＃61; os.listdir(path &＃43; &＃39;pyq&＃39;) #所有原始图像img_xml &＃61; os.listdir(path &＃43; &＃39;labeled&＃39;) #所有xml文件
print(&＃39;img_num: &＃39;,len(img))
print(&＃39;img_xml_num: &＃39;,len(img_xml))#清空VOC数据集文件夹内容
path_img_ori &＃61; &＃39;/code/kaggle/keras-yolo3-master/VOCdevkit/VOC2007/JPEGImages&＃39;
path_xml &＃61; &＃39;/code/kaggle/keras-yolo3-master/VOCdevkit/VOC2007/Annotations&＃39;
img_ &＃61; os.listdir(path_img_ori)
xml_ &＃61; os.listdir(path_xml)
for img__ in img_:os.remove(os.path.join(path_img_ori,img__))
for xml__ in xml_:os.remove(os.path.join(path_xml,xml__))#生成VOC数据集文件夹内容
k &＃61; 0
for i in range(len(img)):path_img_ori &＃61; &＃39;/code/kaggle/keras-yolo3-master/VOCdevkit/VOC2007/JPEGImages/&＃39;path_xml &＃61; &＃39;/code/kaggle/keras-yolo3-master/VOCdevkit/VOC2007/Annotations/&＃39;#拷贝转移文件&＃xff0c;并按0&＃xff0c;1&＃xff0c;2命名新文件#shutil.copyfile&＃xff08;old——path&＃xff0c; new-path&＃xff09;shutil.copyfile(path &＃43; &＃39;pyq/&＃39; &＃43; img[i] , path_img_ori &＃43; str(k) &＃43; &＃39;.&＃39; &＃43; &＃39;jpg&＃39;)shutil.copyfile(path &＃43; &＃39;labeled/&＃39; &＃43; img_xml[i] , path_xml &＃43; str(k) &＃43; &＃39;.&＃39; &＃43; img_xml[i].split(&＃39;.&＃39;)[-1]) context.append(str(k))k &＃61; k&＃43;1trainval_percent &＃61; 0.2
train_percent &＃61; 0.8 #自己定比例
xmlfilepath &＃61; path_xml
txtsavepath &＃61; &＃39;/code/kaggle/keras-yolo3-master/VOCdevkit/VOC2007/ImageSets/Main/&＃39;
total_xml &＃61; os.listdir(xmlfilepath)num &＃61; len(total_xml)
list &＃61; range(num)
tv &＃61; int(num * trainval_percent)
tr &＃61; int(tv * train_percent)
trainval &＃61; random.sample(list, tv)
train &＃61; random.sample(trainval, tr)#到达的文件路径&＃xff5e;/VOCdevkit/VOC2007/ImageSets/Main/trainval.txt&＃39;ftrainval &＃61; open(&＃39;/code/kaggle/keras-yolo3-master/VOCdevkit/VOC2007/ImageSets/Main/trainval.txt&＃39;, &＃39;w&＃39;)
ftest &＃61; open(&＃39;/code/kaggle/keras-yolo3-master/VOCdevkit/VOC2007/ImageSets/Main/test.txt&＃39;, &＃39;w&＃39;)
ftrain &＃61; open(&＃39;/code/kaggle/keras-yolo3-master/VOCdevkit/VOC2007/ImageSets/Main/train.txt&＃39;, &＃39;w&＃39;)
fval &＃61; open(&＃39;/code/kaggle/keras-yolo3-master/VOCdevkit/VOC2007/ImageSets/Main/val.txt&＃39;, &＃39;w&＃39;)for i in list:name &＃61; total_xml[i][:-4] &＃43; &＃39;\n&＃39;if i in trainval:ftrainval.write(name)if i in train:ftest.write(name)else:fval.write(name)else:ftrain.write(name)ftrainval.close()
ftrain.close()
fval.close()
ftest.close()#%%

生成yolo3所需的train.txt,val.txt,test.txt

打开keras-yolo3-master文件夹下voc_annatation.py文件进行修改

import xml.etree.ElementTree as ET
from os import getcwdsets&＃61;[(&＃39;2007&＃39;, &＃39;train&＃39;), (&＃39;2007&＃39;, &＃39;val&＃39;), (&＃39;2007&＃39;, &＃39;test&＃39;)] classes &＃61; ["dz_tx"] #这里改为自己标注数据集中的标签名def convert_annotation(year, image_id, list_file):in_file &＃61; open(&＃39;/code/keras-yolo3-master/VOCdevkit/VOC%s/Annotations/%s.xml&＃39;%(year, image_id),&＃39;rb&＃39;)tree&＃61;ET.parse(in_file)root &＃61; tree.getroot()for obj in root.iter(&＃39;object&＃39;):difficult &＃61; obj.find(&＃39;difficult&＃39;).textcls &＃61; obj.find(&＃39;name&＃39;).textif cls not in classes or int(difficult)&＃61;&＃61;1:continuecls_id &＃61; classes.index(cls)xmlbox &＃61; obj.find(&＃39;bndbox&＃39;)b &＃61; (int(xmlbox.find(&＃39;xmin&＃39;).text), int(xmlbox.find(&＃39;ymin&＃39;).text), int(xmlbox.find(&＃39;xmax&＃39;).text), int(xmlbox.find(&＃39;ymax&＃39;).text))list_file.write(" " &＃43; ",".join([str(a) for a in b]) &＃43; &＃39;,&＃39; &＃43; str(cls_id))wd &＃61; getcwd()for year, image_set in sets:image_ids &＃61; open(&＃39;/code/keras-yolo3-master/VOCdevkit/VOC%s/ImageSets/Main/%s.txt&＃39;%(year, image_set)).read().strip().split()list_file &＃61; open(&＃39;%s.txt&＃39;%( image_set), &＃39;w&＃39;)for image_id in image_ids:list_file.write(&＃39;%s/VOCdevkit/VOC%s/JPEGImages/%s.jpg&＃39;%(wd, year, image_id))convert_annotation(year, image_id, list_file)list_file.write(&＃39;\n&＃39;)list_file.close()


修改model_data下的文件&＃xff0c;放入你的类别&＃xff0c;coco,voc这两个文件都需要修改。

一个标签占一行

训练自己的网络

直接复制替换原来train.py即可

此时自己在keras-yolo3-master下新建文件夹logs logs下再建文件夹000

1. run时可能会报错
   AttributeError: module ‘keras.backend’ has no attribute ‘control_flow_ops’

   解决办法&＃xff1a;https://blog.csdn.net/CAU_Ayao/article/details/89312354

2. 可能Tensorboard报错
   我发现Tensorboard这行代码是灰色的&＃xff0c;所以我把它作为释义不用了

3. 建议先将epoch调小一些进行测试

"""
Retrain the YOLO model for your own dataset.
"""
import numpy as np
import keras.backend as K
from keras.layers import Input, Lambda
from keras.models import Model
from keras.callbacks import TensorBoard, ModelCheckpoint, EarlyStoppingfrom yolo3.model import preprocess_true_boxes, yolo_body, tiny_yolo_body, yolo_loss
from yolo3.utils import get_random_datadef _main():annotation_path &＃61; &＃39;train.txt&＃39;log_dir &＃61; &＃39;logs/000/&＃39;classes_path &＃61; &＃39;model_data/voc_classes.txt&＃39;anchors_path &＃61; &＃39;model_data/yolo_anchors.txt&＃39;class_names &＃61; get_classes(classes_path)anchors &＃61; get_anchors(anchors_path)input_shape &＃61; (416,416) # multiple of 32, hwmodel &＃61; create_model(input_shape, anchors, len(class_names) )train(model, annotation_path, input_shape, anchors, len(class_names), log_dir&＃61;log_dir)def train(model, annotation_path, input_shape, anchors, num_classes, log_dir&＃61;&＃39;logs/&＃39;):model.compile(optimizer&＃61;&＃39;adam&＃39;, loss&＃61;{&＃39;yolo_loss&＃39;: lambda y_true, y_pred: y_pred})#################这个位置########################logging &＃61; TensorBoard(log_dir&＃61;log_dir)checkpoint &＃61; ModelCheckpoint(log_dir &＃43; "ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5",monitor&＃61;&＃39;val_loss&＃39;, save_weights_only&＃61;True, save_best_only&＃61;True, period&＃61;1)batch_size &＃61; 10val_split &＃61; 0.1with open(annotation_path) as f:lines &＃61; f.readlines()np.random.shuffle(lines)num_val &＃61; int(len(lines)*val_split)num_train &＃61; len(lines) - num_valprint(&＃39;Train on {} samples, val on {} samples, with batch size {}.&＃39;.format(num_train, num_val, batch_size))model.fit_generator(data_generator_wrap(lines[:num_train], batch_size, input_shape, anchors, num_classes),steps_per_epoch&＃61;max(1, num_train//batch_size),validation_data&＃61;data_generator_wrap(lines[num_train:], batch_size, input_shape, anchors, num_classes),validation_steps&＃61;max(1, num_val//batch_size),epochs&＃61;500,initial_epoch&＃61;0)model.save_weights(log_dir &＃43; &＃39;trained_weights.h5&＃39;)def get_classes(classes_path):with open(classes_path) as f:class_names &＃61; f.readlines()class_names &＃61; [c.strip() for c in class_names]return class_namesdef get_anchors(anchors_path):with open(anchors_path) as f:anchors &＃61; f.readline()anchors &＃61; [float(x) for x in anchors.split(&＃39;,&＃39;)]return np.array(anchors).reshape(-1, 2)def create_model(input_shape, anchors, num_classes, load_pretrained&＃61;False, freeze_body&＃61;False,weights_path&＃61;&＃39;model_data/yolo_weights.h5&＃39;):K.clear_session() # get a new sessionimage_input &＃61; Input(shape&＃61;(None, None, 3))h, w &＃61; input_shapenum_anchors &＃61; len(anchors)y_true &＃61; [Input(shape&＃61;(h//{0:32, 1:16, 2:8}[l], w//{0:32, 1:16, 2:8}[l], \num_anchors//3, num_classes&＃43;5)) for l in range(3)]model_body &＃61; yolo_body(image_input, num_anchors//3, num_classes)print(&＃39;Create YOLOv3 model with {} anchors and {} classes.&＃39;.format(num_anchors, num_classes))if load_pretrained:model_body.load_weights(weights_path, by_name&＃61;True, skip_mismatch&＃61;True)print(&＃39;Load weights {}.&＃39;.format(weights_path))if freeze_body:# Do not freeze 3 output layers.num &＃61; len(model_body.layers)-7for i in range(num): model_body.layers[i].trainable &＃61; Falseprint(&＃39;Freeze the first {} layers of total {} layers.&＃39;.format(num, len(model_body.layers)))model_loss &＃61; Lambda(yolo_loss, output_shape&＃61;(1,), name&＃61;&＃39;yolo_loss&＃39;,arguments&＃61;{&＃39;anchors&＃39;: anchors, &＃39;num_classes&＃39;: num_classes, &＃39;ignore_thresh&＃39;: 0.5})([*model_body.output, *y_true])model &＃61; Model([model_body.input, *y_true], model_loss)return model
def data_generator(annotation_lines, batch_size, input_shape, anchors, num_classes):n &＃61; len(annotation_lines)np.random.shuffle(annotation_lines)i &＃61; 0while True:image_data &＃61; []box_data &＃61; []for b in range(batch_size):i %&＃61; nimage, box &＃61; get_random_data(annotation_lines[i], input_shape, random&＃61;True)image_data.append(image)box_data.append(box)i &＃43;&＃61; 1image_data &＃61; np.array(image_data)box_data &＃61; np.array(box_data)y_true &＃61; preprocess_true_boxes(box_data, input_shape, anchors, num_classes)yield [image_data, *y_true], np.zeros(batch_size)def data_generator_wrap(annotation_lines, batch_size, input_shape, anchors, num_classes):n &＃61; len(annotation_lines)if n&＃61;&＃61;0 or batch_size<&＃61;0: return Nonereturn data_generator(annotation_lines, batch_size, input_shape, anchors, num_classes)if __name__ &＃61;&＃61; &＃39;__main__&＃39;:_main()