在无人驾驶中，交通标志识别是一项重要的任务。本项目以美国交通标志数据集LISA为训练对象，采用YOLOv3目标检测方法实现实时交通标志识别。
具体项目过程包括包括：安装Darknet、下载LISA交通标志数据集、数据集格式转换、修改配置文件、训练LISA数据集、测试训练出的网络模型、性能统计(mAP计算和画出PR曲线)和先验框聚类。
YOLOv3可以实时地进行端到端的目标检测，以速度快见长。本课程将手把手地教大家使用YOLOv3实现交通标志识别。本课程的YOLOv3使用Darknet，在Ubuntu上做项目演示。 Darknet是使用C语言实现的轻型开源深度学习框架，依赖少，可移植性好，值得深入学习和探究。
为帮助大家学习，本人推出了课程《YOLOv3目标检测实战：训练自己的数据集》
课程链接：https://edu.51cto.com/course/18279.html
除本课程《YOLOv3目标检测实战：交通标志识别》外，本人推出了有关YOLOv3目标检测的系列课程，请持续关注该系列的其它课程视频，包括：
《YOLOv3目标检测实战：训练自己的数据集》
《YOLOv3目标检测：原理与源码解析》
《YOLOv3目标检测：网络模型改进方法》
另一门课程《YOLOv3目标检测实战：训练自己的数据集》主要是介绍如何训练自己标注的数据集。而本课程的区别主要在于学习对已标注数据集的格式转换，即把LISA数据集从csv格式转换成YOLOv3所需要的PASCAL VOC格式和YOLO格式。本课程提供数据集格式转换的Python代码。
请大家关注以上课程，并选择学习。
下图是使用YOLOv3进行交通标志识别的测试结果：

问：识别交通标志，怎么确定图片里有交通标志，具体在哪里？答：这个应该算目标识别和物体识别，CV的经典教程都有object recognition的内容可以参考。具体到交通标志识别，Github上有些开源代码 Matlab C++ Java的都有
Tsaraisa was run with "./tsaraisa.py -s -g -c lbpCascade.xml -M" on demo screenshot.
What does it do?
Detect traffic signs.
Recognize speed limits in signs.
(optional) Compare GPS-speed to speed limit.
(optional) Run user command when overspeeding.
How?
GPS-class
Uses threading to update GPS-info automatically.
Gets speed from GPS-daemon using python bindings.
Frame handling
Reads frame from webcam.
Converts frame to grayscale with cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY).
(optional) Morphological Open/close.
(optional) Equalizes histograms with cv2.equalizeHist().
Downscale frame with multiplier n.
Traffic sign detection
Uses LBP- or HAAR-cascade to detect speed signs.
LBP-cascade was trained with 2000 positive and 1000 negative images.
Negative image is image of road without speed limit sign.
Positive image is image of road with speed limit sign.
LBP- and HAAR-detection allows great differences in lightning.
LBP- and HAAR-detection works great on low performance machines.
Recognizing speed limits
Uses Fast Approximate Nearest Neighbor Search Library feature matching.
Creates "keypoints" of detected sign and compares them to all known speed limits (files in data-folder).
Keypoint match distances need to be inside a threshold.
Match with biggest proper keypoint amount will be returned.
It is fast and pretty accurate with different lightning conditions.
Speed assist
When new speed limit is detected it is added as current speed limit.
After every frame script compares current speed to current speed limit.
Script runs specified command when overspeeding (e.g. "beep").
Requirements
Required software
OpenCV >=3.0.0
Python >=2.7.3
LibAV >=0.8.10
(optional) gpsd && python-gps
(optional) V4L2 1.0.1
Required hardware
Webcam or some other video-source.
(optional) GPS Module BU-353
阅读全文……

教你从零开始做一个基于深度学习的交通标志识别系统
基于Yolo v3的交通标志识别系统及源码
自动驾驶之——交通标志识别
在本文章你可以学习到如何训练自己采集的数据集，生成模型，并用yolo v3算法实现交通标志识别。当然交通标志可以是其他比如交通灯、车辆、行人等等。
效果图：
环境：win10+python3.6 +keras +Tensorflow 1.14.0+Pycharm2019
一、环境配置
1.1 下载anaconda，安装python环境
python环境我建议下载anaconda，anaconda内置python 3.6解释器，其他python版本不用单独安装，十分便捷，在安装
1.2 修改下载源
anaconda的python解释器自带的下载源下载东西很慢，因此可以改为国内豆瓣或者清华等镜像源。
这里教大家改为豆瓣源。
 可以看我的另外一篇文章：
更改Python pip下载源为国内豆瓣源
1.3 安装其他环境及包，tensorflow、keras、opencv等等
三、训练数据集
3.1 前言：明确读入的格式以及处理流程
首先明确进行yolov3训练数据集的格式，如下：
文件的绝对路径 left,top,right,bottom,类别编号
其实所有模型方法进行训练的时候甚至机器学习svm分类等，都是要告诉电脑(以检测猫来说)，你的图片在计算机存储的路径；你的图片中猫目标的位置，即left,top,right,bottom ；以及猫类别的编号。
实际例子：
/home/matthew/VOC2007/JPEGImages/000000160.jpg 186,192,353,349,0 579,286,850,500,0/home/matthew/VOC2007/JPEGImages/000000162.jpg 403,22,458,60,0 400,245,552,389,0 432,0,459,12,0 926,1,999,15,0/home/matthew/VOC2007/JPEGImages/000000166.jpg 146,246,340,428,0
可以看到数据中不只有一个目标。
 3.1 、制作自己的VOC格式训练集——打标签(labelImg软件)
那么这么标准的数据如何得到呢，其实可以用一个深度学习训练图片的打标签工具——labelImg。关于这个软件的安装与使用见链接：
图像识别中标注工具labelImg的安装和使用——深度学习打标签
这个软件会使用会生成标准xml文件，然后经过下面一系列操作，最后成为txt文件，进行训练。
 使用教程：
大致步骤
1、图像(.jpg)进行重命名(00000X.jpg) ，可以用bat命令
2、制作图像的.xml文件 ，生产的 xml文件放入下方的 **Annotation**文件夹
3、将.xml文件转换成.txt文件【表述图像及框的四点坐标信息】
4、将.jpg文件转换成.txt文件【训练、测试、验证数据集的路径】
3.2 ：建立VOC2007目录，把上一步标记的xml文件放入Annotation文件夹内
在工程下新建一个文件夹VOC2007,在下面就是新建几个默认名字的文件夹 Annotation，ImageSet(该目录还有三个文件需要建立)，JPEGImages(把你所有的图片都复制到该目录里面，如下图)，SegmentationClass，SegmentationObject。
--VOC2007   --Annotations   --ImageSets     --Main     --Layout     --Segmentation   --JPEGImages   --SegmentationClass   --SegmentationObjectAnnotations  中主要存放xml文件，每一个xml对应一张图像，           并且每个xml中存放的是标记的各个目标的位置和类别信息，命名通常与对应的原始图像一样JPEGImages  自己的原始图像放在JPEGImages文件夹ImageSets           Layout 存放人体部位的数据。(用不上)            Main  存放的是目标识别的数据，主要有test.txt , train.txt, val.txt，                  trainval.txt四个文件。            Segmentation  存放分割的数据。(用不上)
有大神写了个程序，运行一下就直接建好了，如下：
def make_voc_dir():    os.makedirs('VOC2007/Annotations')    os.makedirs('VOC2007/ImageSets')    os.makedirs('VOC2007/ImageSets/Main')    os.makedirs('VOC2007/ImageSets/Layout')    os.makedirs('VOC2007/ImageSets/Segmentation')    os.makedirs('VOC2007/JPEGImages')    os.makedirs('VOC2007/SegmentationClass')    os.makedirs('VOC2007/SegmentationObject')if __name__ == '__main__':    make_voc_dir()
 3.3 、生成ImageSet/Main/4个文件。
在VOC2007下新建一个test.py文件，复制如下代码
test.py文件：
import osimport random trainval_percent = 0.2train_percent = 0.8xmlfilepath = 'Annotations'txtsavepath = 'ImageSetsMain'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) ftrainval = open('ImageSets/Main/trainval.txt', 'w')ftest = open('ImageSets/Main/test.txt', 'w')ftrain = open('ImageSets/Main/train.txt', 'w')fval = open('ImageSets/Main/val.txt', 'w') for i in list:    name = total_xml[i][:-4] + ''    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()
运行test.py代码之后，生成如下文件，VOC2007数据集制作完成
3.4、运行voc_annotation.py——生成yolo3所需的train.txt、val.txt、test.txt文件
生成的数据集不能供yolov3直接使用。需要运行voc_annotation.py ，classes以检测两个类为例(车和人腿)，在voc_annotation.py需改你的数据集为。
voc_annotation.py文件
import xml.etree.ElementTree as ETfrom os import getcwd# 注意这里的‘2007’，也许你的就需要修改sets=[('2007', 'train'), ('2007', 'val'), ('2007', 'test')]# 注意类别classes = ["car","leg"]def convert_annotation(year, image_id, list_file):    # 注意路径    in_file = open('VOC%s/Annotations/%s.xml'%(year, image_id))    tree=ET.parse(in_file)    root = tree.getroot()    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 = (int(xmlbox.find('xmin').text), int(xmlbox.find('ymin').text), int(xmlbox.find('xmax').text), int(xmlbox.find('ymax').text))        list_file.write(" " + ",".join([str(a) for a in b]) + ',' + str(cls_id))wd = getcwd()for year, image_set in sets:    # 注意路径    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:        # 注意路径        list_file.write('%s/VOC%s/JPEGImages/%s.jpg'%(wd, year, image_id))        convert_annotation(year, image_id, list_file)        list_file.write('')    list_file.close()
运行之后，生成如下三个文件：
如果出现这种解析xml文件出错的问题，就打开保存的文件，找到最下面，看看后面一张图片xml文件格式
 3.5、修改参数文件yolo3.cfg
打开yolo3.cfg文件。搜索yolo(共出现三次)，每次按下图都要修改
参数修改说明：
filter：3*(5+len(classes))
classes:你要训练的类别数(我这里是训练两类)
random：原来是1，显存小改为0
3.6、修改model_data下的voc_classes.txt为自己训练的类别
coco,voc这两个文件都需要修改。
如果类别比较多，可以用下面程序生成voc_classes.txt
classes = ["SIGN_SPEED_5","SIGN_SPEED_20","SIGN_SPEED_30","SIGN_SPEED_40","SIGN_SPEED_50","SIGN_SPEED_60","SIGN_LOAD_LIMITED_30T","SIGN_LOAD_LIMITED_40T","SIGN_LOAD_LIMITED_55T","SIGN_HEIGHT_2_8M","RM_RIGHT","RM_island","RM_Motorway_lane","RM_Crosswalk_lane","RM_No_vehicle_lane","SIGN_Limit_LEFT","SIGN_Limit_Drive","SIGN_Limit_parking","SIGN_Limit_whistle","SIGN_STOP_TO_YIELD","SIGN_Attention_people","SIGN_Attention_child","SIGN_work_Ahead","SIGN_SLOW","SIGN_Limit_trucks","SIGN_Keep_right_road","SIGN_Stop_sign","SIGN_HEIGHT_2_3M"]num2 = len(classes)# xml文件个数list_file = open('20200314.txt', 'w')#list_file.write(classes[0])for classe in classes:    list_file.write(classe)    list_file.write('')list_file.close()
3.7、 进行训练——运行mytrain.py代码
因为程序中有logs/000/目录，你需要创建这样一个目录，这个目录的作用就是存放自己的数据集训练得到的模型。不然程序运行到最后会因为找不到该路径而发生错误。生成的模型trained_weights.h5如下
3.8、修改yolo.py文件
如下将self这三行修改为各自对应的路径
3.9、检测图片
运行 python yolo_video.py --image,输入自己要检测的类的图片即可查看训练效果了。
 3.10、出错总结
1、如果是cpu训练，大概是70s作于一个Epoch，安装好gpu后，训练就是100-200ms一个Epoch了。速度提升很多。但是安装gpu我是搞了两天，总是提示cudnn版本未初始化啥的，我是cuda10.0 cudnn7.5 ，在经历了各种tensorflow升级降级，以及各种偏方，最后我将cudnn的版本重新下载为cudnn 7.4，重新配置了一下总算成功。我的版本是tensorflow-gpu 14.0.0+cuda10.0 +cudnn7.4。
2、训练时提示显存不够，我也看了网上很多方法，最后把input_shape从默认的419*419改成了128*128的，就成功了。关于那个到1/500的时候突然报错可以看下是不是因为batch_size太大，我的改成4了，然后那个输入input_shape改成128*128，然后就顺利的跑通了。还有报那个找不到路径的错误，我把VOCdevkit放在keras-yolo3文件夹下就可以了。
Loss值：
四、识别
五、代码整体框架
介绍代码框架，以及如何运行
六、界面设计
七、资源下载
全部工程源码：
训练数据集：

这节采用德国交通标志数据集实现交通标志识别。本节步骤分为数据导入，探索和可视化数据集，数据预处理，交通标志识别模型搭建。

一、加载数据
# 加载数据
import pickle

# 数据集所在的文件位置
training_file = "/content/drive/traffic-signs-data/train.p"
validation_file = "/content/drive/traffic-signs-data/valid.p"
testing_file = "/content/drive/traffic-signs-data/test.p"

# 打开文件
with open(training_file,mode="rb") as f:
  train = pickle.load(f)
with open(validation_file,mode="rb") as f:
  valid = pickle.load(f)
with open(testing_file,mode="rb") as f:
  test = pickle.load(f)

# 获取数据集的特征及标签数据
X_train,y_train = train["features"],train["labels"]
X_valid,y_valid = valid["features"],valid["labels"]
X_test,y_test = test["features"],test["labels"]

# 查看数据量
print("Number of training examples =",X_train.shape[0])
print("Number of validtion examples =",X_valid.shape[0])
print("Number of testing examples=",X_test.shape[0])

# 查看数据格式
print("Image data shape =",X_train.shape[1:])

Number of training examples = 34799
Number of validtion examples = 4410
Number of testing examples= 12630
Image data shape = (32, 32, 3)

# 查看数据的标签的数量
import numpy as np

sum = np.unique(y_train)
print("number of classes =",len(sum))

number of classes = 43

# 查看标签数据
import pandas as pd

sign_names_file = "/content/drive/traffic-signs-data/signnames.csv"
sign_names = pd.read_csv(sign_names_file)
print(sign_names)

    ClassId                                           SignName
0         0                               Speed limit (20km/h)
1         1                               Speed limit (30km/h)
2         2                               Speed limit (50km/h)
3         3                               Speed limit (60km/h)
4         4                               Speed limit (70km/h)
5         5                               Speed limit (80km/h)
6         6                        End of speed limit (80km/h)
7         7                              Speed limit (100km/h)
8         8                              Speed limit (120km/h)
9         9                                         No passing
10       10       No passing for vehicles over 3.5 metric tons
11       11              Right-of-way at the next intersection
12       12                                      Priority road
13       13                                              Yield
14       14                                               Stop
15       15                                        No vehicles
16       16           Vehicles over 3.5 metric tons prohibited
17       17                                           No entry
18       18                                    General caution
19       19                        Dangerous curve to the left
20       20                       Dangerous curve to the right
21       21                                       Double curve
22       22                                         Bumpy road
23       23                                      Slippery road
24       24                          Road narrows on the right
25       25                                          Road work
26       26                                    Traffic signals
27       27                                        Pedestrians
28       28                                  Children crossing
29       29                                  Bicycles crossing
30       30                                 Beware of ice/snow
31       31                              Wild animals crossing
32       32                End of all speed and passing limits
33       33                                   Turn right ahead
34       34                                    Turn left ahead
35       35                                         Ahead only
36       36                               Go straight or right
37       37                                Go straight or left
38       38                                         Keep right
39       39                                          Keep left
40       40                               Roundabout mandatory
41       41                                  End of no passing
42       42  End of no passing by vehicles over 3.5 metric ...

# 定义将标签id转换成name的函数
sign_names = np.array(sign_names)
def id_to_name(id):
  return sign_names[id][1]

# 验证是否id_to_name函数
id_to_name(0)

'Speed limit (20km/h)'

# 绘制交通标志图
import matplotlib.pyplot as plt
%matplotlib inline

fig,axes = plt.subplots(2,5,figsize=(18,5))
ax_array = axes.ravel()
for ax in ax_array:
  index = np.random.randint(0,len(X_train))
  ax.imshow(X_train[index])
  ax.axis("off")
  ax.set_title(id_to_name(y_train[index]))
plt.show()


# 直方图来展示图像训练集的各个类别的分布情况
n_classes = len(sum)
def plot_y_train_hist():
    fig = plt.figure(figsize=(15,5))
    ax = fig.add_subplot(1,1,1)
    hist = ax.hist(y_train,bins = n_classes)
    ax.set_title("the frequency of each category sign")
    ax.set_xlabel("signs")
    ax.set_ylabel("frequency")
    plt.show()
    return hist

print(X_train.shape)
print(y_train.shape)
hist = plot_y_train_hist()

(34799, 32, 32, 3)
(34799,)


# 数据重采样，使样本个数分配均匀
bin_edges = hist[1]
bin_centers = (bin_edges[1:] + bin_edges[0:len(bin_edges)-1])/2
for i in range(len(bin_centers)):
    if hist[0][i] < 1000:
        train_data = [X_train[j] for j in range(len(y_train)) if y_train[j] == i]
        need_resample_num = int(1000 - hist[0][i])
        new_data_x = [np.copy(train_data[np.random.randint(len(train_data))]) for k in range(need_resample_num)]
        new_data_y = [i for x in range(need_resample_num)]
        X_train = np.vstack((X_train, np.array(new_data_x)))
        y_train = np.hstack((y_train, np.array(new_data_y)))
print(X_train.shape)
print(y_train.shape)
plot_y_train_hist()

(51690, 32, 32, 3)
(51690,)


(array([1000., 1980., 2010., 1260., 1770., 1650., 1000., 1290., 1260.,
        1320., 1800., 1170., 1890., 1920., 1000., 1000., 1000., 1000.,
        1080., 1000., 1000., 1000., 1000., 1000., 1000., 1350., 1000.,
        1000., 1000., 1000., 1000., 1000., 1000., 1000., 1000., 1080.,
        1000., 1000., 1860., 1000., 1000., 1000., 1000.]),
 array([ 0.        ,  0.97674419,  1.95348837,  2.93023256,  3.90697674,
         4.88372093,  5.86046512,  6.8372093 ,  7.81395349,  8.79069767,
         9.76744186, 10.74418605, 11.72093023, 12.69767442, 13.6744186 ,
        14.65116279, 15.62790698, 16.60465116, 17.58139535, 18.55813953,
        19.53488372, 20.51162791, 21.48837209, 22.46511628, 23.44186047,
        24.41860465, 25.39534884, 26.37209302, 27.34883721, 28.3255814 ,
        29.30232558, 30.27906977, 31.25581395, 32.23255814, 33.20930233,
        34.18604651, 35.1627907 , 36.13953488, 37.11627907, 38.09302326,
        39.06976744, 40.04651163, 41.02325581, 42.        ]),
 <a list of 43 Patch objects>)

二、数据预处理
# 源图像
src = X_train[4000]
plt.imshow(src)
print(src.shape)

(32, 32, 3)


# 转为灰度图像
import cv2

dst = cv2.cvtColor(src,cv2.COLOR_RGB2GRAY)
plt.imshow(dst)
print(dst.shape)

(32, 32)


# 灰度图的直方图
plt.hist(dst.ravel(),256,[0,256],color="r")
plt.show()


# 灰度图的直方图均衡化
dst2 = cv2.equalizeHist(dst)
plt.hist(dst2.ravel(),256,[0,256],color="r")
plt.show()


# 查看直方图均衡化的图片
plt.imshow(dst2)

<matplotlib.image.AxesImage at 0x7f97bb3a86a0>


# 增加维度
print(dst2.shape)
dst3 = np.expand_dims(dst2,2)
print(dst3.shape)

(32, 32)
(32, 32, 1)

# 归一化处理
dst4 = np.array(dst3,dtype=np.float32)
dst4 = (dst4-128)/128

# 查看一下归一化处理后的图像数据
dst4[0]

array([[-0.859375 ],
       [-0.9609375],
       [-0.9765625],
       [-0.9296875],
       [-0.859375 ],
       [-0.859375 ],
       [-0.625    ],
       [-0.484375 ],
       [-0.625    ],
       [-0.625    ],
       [-0.375    ],
       [-0.375    ],
       [-0.2890625],
       [ 0.171875 ],
       [ 0.6171875],
       [ 0.3359375],
       [-0.1171875],
       [ 0.171875 ],
       [ 0.3359375],
       [ 0.171875 ],
       [-0.0546875],
       [ 0.125    ],
       [ 0.125    ],
       [ 0.0078125],
       [ 0.0078125],
       [ 0.125    ],
       [-0.1171875],
       [-0.1875   ],
       [-0.1875   ],
       [ 0.0625   ],
       [ 0.125    ],
       [ 0.171875 ]], dtype=float32)

# 封装数据预处理的方法
import cv2

def preprocess_features(X, equalize_hist=True):
    normalized_X = []
    for i in range(len(X)):
        # Convert from RGB to YUV
        yuv_img = cv2.cvtColor(X[i], cv2.COLOR_RGB2YUV)
        yuv_img_v = X[i][:, :, 0]
        # equalizeHist
        yuv_img_v = cv2.equalizeHist(yuv_img_v)
        # expand_dis
        yuv_img_v = np.expand_dims(yuv_img_v, 2)
        normalized_X.append(yuv_img_v)
    # normalize
    normalized_X = np.array(normalized_X, dtype=np.float32)
    normalized_X = (normalized_X-128)/128
    # normalized_X /= (np.std(normalized_X, axis=0) + np.finfo('float32').eps)
    return normalized_X

# 对数据集整体进行处理
X_train_normalized = preprocess_features(X_train)
X_valid_normalized = preprocess_features(X_valid)
X_test_normalized = preprocess_features(X_test)

# 将数据集打乱
from sklearn.utils import shuffle
X_train_normalized,y_train = shuffle(X_train_normalized,y_train)

# 数据增强
from keras.preprocessing.image import ImageDataGenerator

# 图像数据生成器
image_datagen = ImageDataGenerator(rotation_range = 10.,
                                   zoom_range = 0.2,
                                   width_shift_range =  0.08,
                                   height_shift_range = 0.08
                                  )

# 从训练集随意选取一张图片
index = np.random.randint(0, len(X_train_normalized))
img = X_train_normalized[index]

# 展示原始图片
plt.figure(figsize=(1, 1))
plt.imshow(np.squeeze(img), cmap="gray")
plt.title('Example of GRAY image (name = {})'.format(id_to_name(y_train[index])))
plt.axis('off')
plt.show()

# 展示数据增强生成的图片
fig, ax_array = plt.subplots(3, 10, figsize=(15, 5))
for ax in ax_array.ravel():
    images = np.expand_dims(img, 0)
    # np.expand_dims(img, 0) means add dim
    augmented_img, _ = image_datagen.flow(np.expand_dims(img, 0), np.expand_dims(y_train[index], 0)).next()
    #augmented_img=preprocess_features(augmented_img)
    ax.imshow(augmented_img.squeeze(), cmap="gray")
    ax.axis('off')
plt.suptitle('Random examples of data augment (starting from the previous image)')
plt.show()

Using TensorFlow backend.


print("Number of training examples =",X_train.shape[0])

Number of training examples = 51690

# 对标签数据进行one-hot编码
from keras.utils  import np_utils
print("Shape before one-hot encoding:",y_train.shape)
Y_train = np_utils.to_categorical(y_train,n_classes)
Y_valid = np_utils.to_categorical(y_valid,n_classes)
Y_test = np_utils.to_categorical(y_test,n_classes)
print("Shape after one-hot encoding:",Y_train.shape)

Shape before one-hot encoding: (51690,)
Shape after one-hot encoding: (51690, 43)

print(y_train[0])
print(Y_train[0])

24
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]

三、交通标志牌的识别模型搭建
1.2层Conv2D+1层MaxPooling2D+1层Drop(0.25)+1层Flatten+1层Dense+1层Dropout层(0.5)+1层Dense
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D

model = Sequential()
## Feature Extraction
# 第1层卷积，32个3x3的卷积核 ，激活函数使用 relu
model.add(Conv2D(filters=32, kernel_size=(3, 3), activation='relu',
                 input_shape=X_train_normalized.shape[1:]))

# 第2层卷积，64个3x3的卷积核，激活函数使用 relu
model.add(Conv2D(filters=64, kernel_size=(3, 3), activation='relu'))

# 最大池化层，池化窗口 2x2
model.add(MaxPooling2D(pool_size=(2, 2)))

# Dropout 25% 的输入神经元
model.add(Dropout(0.25))

# 将 Pooled feature map 摊平后输入全连接网络
model.add(Flatten())

## Classification
# 全联接层
model.add(Dense(128, activation='relu'))

# Dropout 50% 的输入神经元
model.add(Dropout(0.5))

# 使用 softmax 激活函数做多分类，输出各类别的概率
model.add(Dense(n_classes, activation='softmax'))

WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.
Instructions for updating:
Colocations handled automatically by placer.
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py:3445: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.
Instructions for updating:
Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.

# 查看模型的结构
model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 30, 30, 32)        320       
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 28, 28, 64)        18496     
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 14, 14, 64)        0         
_________________________________________________________________
dropout_1 (Dropout)          (None, 14, 14, 64)        0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 12544)             0         
_________________________________________________________________
dense_1 (Dense)              (None, 128)               1605760   
_________________________________________________________________
dropout_2 (Dropout)          (None, 128)               0         
_________________________________________________________________
dense_2 (Dense)              (None, 43)                5547      
=================================================================
Total params: 1,630,123
Trainable params: 1,630,123
Non-trainable params: 0
_________________________________________________________________

# 编译模型
model.compile(loss="categorical_crossentropy",
              metrics=["accuracy"],
              optimizer="adam")

# 训练模型
history = model.fit(X_train_normalized,
                    Y_train,
                    batch_size=256,
                    epochs=120,
                    verbose=2,
                    validation_data=(X_valid_normalized,Y_valid))     

WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/math_ops.py:3066: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.cast instead.
Train on 51690 samples, validate on 4410 samples
Epoch 1/120
 - 7s - loss: 1.2242 - acc: 0.6591 - val_loss: 0.3351 - val_acc: 0.8957
Epoch 2/120
 - 4s - loss: 0.2724 - acc: 0.9174 - val_loss: 0.2139 - val_acc: 0.9370
Epoch 3/120
 - 4s - loss: 0.1670 - acc: 0.9492 - val_loss: 0.1646 - val_acc: 0.9490
Epoch 4/120
 - 4s - loss: 0.1190 - acc: 0.9622 - val_loss: 0.1639 - val_acc: 0.9467
Epoch 5/120
 - 4s - loss: 0.0910 - acc: 0.9709 - val_loss: 0.1620 - val_acc: 0.9537
Epoch 6/120
 - 4s - loss: 0.0770 - acc: 0.9752 - val_loss: 0.1439 - val_acc: 0.9585
Epoch 7/120
 - 4s - loss: 0.0656 - acc: 0.9795 - val_loss: 0.1304 - val_acc: 0.9587
Epoch 8/120
 - 4s - loss: 0.0572 - acc: 0.9809 - val_loss: 0.1411 - val_acc: 0.9601
Epoch 9/120
 - 4s - loss: 0.0514 - acc: 0.9835 - val_loss: 0.1529 - val_acc: 0.9562
Epoch 10/120
 - 4s - loss: 0.0473 - acc: 0.9847 - val_loss: 0.1166 - val_acc: 0.9642
Epoch 11/120
 - 4s - loss: 0.0427 - acc: 0.9863 - val_loss: 0.1336 - val_acc: 0.9619
Epoch 12/120
 - 4s - loss: 0.0371 - acc: 0.9879 - val_loss: 0.1309 - val_acc: 0.9635
Epoch 13/120
 - 4s - loss: 0.0380 - acc: 0.9874 - val_loss: 0.1194 - val_acc: 0.9685
Epoch 14/120
 - 4s - loss: 0.0315 - acc: 0.9898 - val_loss: 0.1403 - val_acc: 0.9633
Epoch 15/120
 - 4s - loss: 0.0328 - acc: 0.9889 - val_loss: 0.1469 - val_acc: 0.9574
Epoch 16/120
 - 4s - loss: 0.0292 - acc: 0.9903 - val_loss: 0.1241 - val_acc: 0.9689
Epoch 17/120
 - 4s - loss: 0.0291 - acc: 0.9903 - val_loss: 0.1307 - val_acc: 0.9673
Epoch 18/120
 - 4s - loss: 0.0265 - acc: 0.9910 - val_loss: 0.1434 - val_acc: 0.9651
Epoch 19/120
 - 4s - loss: 0.0259 - acc: 0.9914 - val_loss: 0.1291 - val_acc: 0.9680
Epoch 20/120
 - 4s - loss: 0.0246 - acc: 0.9916 - val_loss: 0.1332 - val_acc: 0.9635
Epoch 21/120
 - 4s - loss: 0.0231 - acc: 0.9922 - val_loss: 0.1261 - val_acc: 0.9671
Epoch 22/120
 - 4s - loss: 0.0232 - acc: 0.9923 - val_loss: 0.1301 - val_acc: 0.9680
Epoch 23/120
 - 4s - loss: 0.0198 - acc: 0.9932 - val_loss: 0.1388 - val_acc: 0.9669
Epoch 24/120
 - 4s - loss: 0.0226 - acc: 0.9923 - val_loss: 0.1210 - val_acc: 0.9719
Epoch 25/120
 - 4s - loss: 0.0212 - acc: 0.9930 - val_loss: 0.1297 - val_acc: 0.9683
Epoch 26/120
 - 4s - loss: 0.0206 - acc: 0.9932 - val_loss: 0.1344 - val_acc: 0.9687
Epoch 27/120
 - 4s - loss: 0.0203 - acc: 0.9932 - val_loss: 0.1495 - val_acc: 0.9692
Epoch 28/120
 - 4s - loss: 0.0200 - acc: 0.9935 - val_loss: 0.1305 - val_acc: 0.9696
Epoch 29/120
 - 4s - loss: 0.0181 - acc: 0.9938 - val_loss: 0.1436 - val_acc: 0.9685
Epoch 30/120
 - 4s - loss: 0.0167 - acc: 0.9944 - val_loss: 0.1377 - val_acc: 0.9673

 - 7s - loss: 0.0072 - acc: 0.9977 - val_loss: 0.1146 - val_acc: 0.9823
Epoch 111/120
 - 7s - loss: 0.0059 - acc: 0.9981 - val_loss: 0.1531 - val_acc: 0.9741
Epoch 112/120
 - 7s - loss: 0.0065 - acc: 0.9977 - val_loss: 0.1157 - val_acc: 0.9805
Epoch 113/120
 - 7s - loss: 0.0061 - acc: 0.9980 - val_loss: 0.1638 - val_acc: 0.9717
Epoch 114/120
 - 7s - loss: 0.0055 - acc: 0.9983 - val_loss: 0.1301 - val_acc: 0.9794
Epoch 115/120
 - 7s - loss: 0.0041 - acc: 0.9985 - val_loss: 0.1501 - val_acc: 0.9773
Epoch 116/120
 - 5s - loss: 0.0066 - acc: 0.9976 - val_loss: 0.1437 - val_acc: 0.9757
Epoch 117/120
 - 4s - loss: 0.0064 - acc: 0.9979 - val_loss: 0.1508 - val_acc: 0.9780
Epoch 118/120
 - 4s - loss: 0.0066 - acc: 0.9978 - val_loss: 0.1489 - val_acc: 0.9769
Epoch 119/120
 - 4s - loss: 0.0065 - acc: 0.9980 - val_loss: 0.1378 - val_acc: 0.9766
Epoch 120/120
 - 4s - loss: 0.0059 - acc: 0.9982 - val_loss: 0.1255 - val_acc: 0.9789

# 可视化指标
fig = plt.figure()
plt.subplot(2,1,1)
plt.plot(history.history['acc'])
plt.plot(history.history['val_acc'])
plt.title('Model Accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='lower right')

plt.subplot(2,1,2)
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('Model Loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper right')
plt.tight_layout()

plt.show()


# 保存模型
import os
import tensorflow.gfile as gfile

save_dir = "/content/drive/model/v1/"

if gfile.Exists(save_dir):
    gfile.DeleteRecursively(save_dir)
gfile.MakeDirs(save_dir)

model_name = 'keras_traffic_1.h5'
model_path = os.path.join(save_dir, model_name)
model.save(model_path)
print('Saved trained model at %s ' % model_path)

Saved trained model at /content/drive/model/v1/keras_traffic_1.h5 

# 加载模型
from keras.models import load_model

traffic_model = load_model(model_path)

# 统计模型在测试集上的分类结果
loss_and_metrics = traffic_model.evaluate(X_test_normalized, Y_test, verbose=2,batch_size=256)
    
print("Test Loss: {}".format(loss_and_metrics[0]))
print("Test Accuracy: {}%".format(loss_and_metrics[1]*100))

predicted_classes = traffic_model.predict_classes(X_test_normalized)

correct_indices = np.nonzero(predicted_classes == y_test)[0]
incorrect_indices = np.nonzero(predicted_classes != y_test)[0]
print("Classified correctly count: {}".format(len(correct_indices)))
print("Classified incorrectly count: {}".format(len(incorrect_indices)))

Test Loss: 0.22923571624080058
Test Accuracy: 96.88836103380434%
Classified correctly count: 12237
Classified incorrectly count: 393

2.3层Con2D MaxPooling2D+1层Flatten层+2层Dense
from keras.models import Sequential
from keras.layers import Dense,Dropout,Flatten
from keras.layers import Conv2D,MaxPooling2D

model2 = Sequential()

# layers1
model2.add(Conv2D(filters=32,
                 kernel_size=(3,3),
                 input_shape=X_train_normalized.shape[1:],
                 activation="relu"))
model2.add(MaxPooling2D(pool_size=(2,2)))

# layers2
model2.add(Conv2D(filters=64,
                 kernel_size=(3,3),
                 activation="relu"))
model2.add(MaxPooling2D(pool_size=(2,2)))

# layers3
model2.add(Conv2D(filters=64,
                 kernel_size=(3,3),
                 activation="relu"))
model2.add(MaxPooling2D(pool_size=(2,2)))

# flattern
model2.add(Flatten())

# Dense
model2.add(Dense(512,activation="relu"))


# Dense
model2.add(Dense(n_classes,activation="softmax"))

# 查看模型的结构
model2.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_3 (Conv2D)            (None, 30, 30, 32)        320       
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 15, 15, 32)        0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 13, 13, 64)        18496     
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 6, 6, 64)          0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 4, 4, 64)          36928     
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 2, 2, 64)          0         
_________________________________________________________________
flatten_2 (Flatten)          (None, 256)               0         
_________________________________________________________________
dense_3 (Dense)              (None, 512)               131584    
_________________________________________________________________
dense_4 (Dense)              (None, 43)                22059     
=================================================================
Total params: 209,387
Trainable params: 209,387
Non-trainable params: 0
_________________________________________________________________

# 编译模型
model2.compile(loss="categorical_crossentropy",
              metrics=["accuracy"],
              optimizer="adam")

# 训练模型
history2 = model2.fit(X_train_normalized,
                    Y_train,
                    batch_size=256,
                    epochs=120,
                    verbose=2,
                    validation_data=(X_valid_normalized,Y_valid))                    

Train on 51690 samples, validate on 4410 samples
Epoch 1/120
 - 2s - loss: 1.3340 - acc: 0.6319 - val_loss: 0.6667 - val_acc: 0.8041
Epoch 2/120
 - 2s - loss: 0.2375 - acc: 0.9308 - val_loss: 0.4388 - val_acc: 0.8762
Epoch 3/120
 - 2s - loss: 0.1122 - acc: 0.9669 - val_loss: 0.3823 - val_acc: 0.8964
Epoch 4/120
 - 2s - loss: 0.0769 - acc: 0.9775 - val_loss: 0.3509 - val_acc: 0.9036
Epoch 5/120
 - 2s - loss: 0.0508 - acc: 0.9857 - val_loss: 0.3434 - val_acc: 0.9088
Epoch 6/120
 - 2s - loss: 0.0420 - acc: 0.9874 - val_loss: 0.3413 - val_acc: 0.9204
Epoch 7/120
 - 2s - loss: 0.0297 - acc: 0.9912 - val_loss: 0.3266 - val_acc: 0.9243
Epoch 8/120
 - 2s - loss: 0.0198 - acc: 0.9943 - val_loss: 0.3150 - val_acc: 0.9256
Epoch 9/120
 - 2s - loss: 0.0205 - acc: 0.9940 - val_loss: 0.3463 - val_acc: 0.9200
Epoch 10/120
 - 2s - loss: 0.0174 - acc: 0.9947 - val_loss: 0.3045 - val_acc: 0.9236
Epoch 11/120
 - 2s - loss: 0.0194 - acc: 0.9943 - val_loss: 0.3561 - val_acc: 0.9184
Epoch 12/120
 - 2s - loss: 0.0109 - acc: 0.9968 - val_loss: 0.3475 - val_acc: 0.9259
Epoch 13/120
 - 2s - loss: 0.0116 - acc: 0.9964 - val_loss: 0.3646 - val_acc: 0.9270
Epoch 14/120
 - 2s - loss: 0.0127 - acc: 0.9958 - val_loss: 0.3486 - val_acc: 0.9295
Epoch 15/120
 - 2s - loss: 0.0090 - acc: 0.9975 - val_loss: 0.3234 - val_acc: 0.9297
Epoch 16/120
 - 2s - loss: 0.0118 - acc: 0.9962 - val_loss: 0.4056 - val_acc: 0.9236
Epoch 17/120
 - 2s - loss: 0.0082 - acc: 0.9976 - val_loss: 0.3282 - val_acc: 0.9345
Epoch 18/120
 - 2s - loss: 0.0066 - acc: 0.9982 - val_loss: 0.3327 - val_acc: 0.9381
Epoch 19/120
 - 2s - loss: 0.0089 - acc: 0.9971 - val_loss: 0.3708 - val_acc: 0.9338
Epoch 20/120
 - 2s - loss: 0.0109 - acc: 0.9967 - val_loss: 0.3011 - val_acc: 0.9397
Epoch 21/120
 - 2s - loss: 0.0030 - acc: 0.9993 - val_loss: 0.3431 - val_acc: 0.9406
Epoch 22/120
 - 2s - loss: 0.0121 - acc: 0.9965 - val_loss: 0.3113 - val_acc: 0.9413
Epoch 23/120
 - 2s - loss: 0.0068 - acc: 0.9981 - val_loss: 0.3195 - val_acc: 0.9431
Epoch 24/120
 - 2s - loss: 0.0021 - acc: 0.9995 - val_loss: 0.3079 - val_acc: 0.9447
Epoch 25/120
 - 2s - loss: 0.0080 - acc: 0.9973 - val_loss: 0.2953 - val_acc: 0.9408
Epoch 26/120
 - 2s - loss: 0.0086 - acc: 0.9971 - val_loss: 0.3595 - val_acc: 0.9315
Epoch 27/120
 - 2s - loss: 0.0105 - acc: 0.9967 - val_loss: 0.3286 - val_acc: 0.9342
Epoch 28/120
 - 2s - loss: 0.0039 - acc: 0.9990 - val_loss: 0.3357 - val_acc: 0.9363
Epoch 29/120
 - 2s - loss: 0.0038 - acc: 0.9992 - val_loss: 0.2930 - val_acc: 0.9440
Epoch 30/120
 - 2s - loss: 0.0053 - acc: 0.9985 - val_loss: 0.2465 - val_acc: 0.9537

Epoch 110/120
 - 2s - loss: 6.2377e-04 - acc: 1.0000 - val_loss: 0.4051 - val_acc: 0.9519
Epoch 111/120
 - 2s - loss: 6.2377e-04 - acc: 1.0000 - val_loss: 0.4071 - val_acc: 0.9524
Epoch 112/120
 - 2s - loss: 6.2377e-04 - acc: 1.0000 - val_loss: 0.4111 - val_acc: 0.9524
Epoch 113/120
 - 2s - loss: 6.2377e-04 - acc: 1.0000 - val_loss: 0.4106 - val_acc: 0.9524
Epoch 114/120
 - 2s - loss: 6.2377e-04 - acc: 1.0000 - val_loss: 0.4115 - val_acc: 0.9522
Epoch 115/120
 - 2s - loss: 6.2377e-04 - acc: 1.0000 - val_loss: 0.4134 - val_acc: 0.9519
Epoch 116/120
 - 2s - loss: 6.2377e-04 - acc: 1.0000 - val_loss: 0.4157 - val_acc: 0.9510
Epoch 117/120
 - 2s - loss: 6.2377e-04 - acc: 1.0000 - val_loss: 0.4181 - val_acc: 0.9515
Epoch 118/120
 - 2s - loss: 6.2376e-04 - acc: 1.0000 - val_loss: 0.4188 - val_acc: 0.9517
Epoch 119/120
 - 2s - loss: 6.2376e-04 - acc: 1.0000 - val_loss: 0.4199 - val_acc: 0.9517
Epoch 120/120
 - 2s - loss: 6.2376e-04 - acc: 1.0000 - val_loss: 0.4196 - val_acc: 0.9510

# 可视化指标
fig = plt.figure()
plt.subplot(2,1,1)
plt.plot(history2.history['acc'])
plt.plot(history2.history['val_acc'])
plt.title('Model Accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='lower right')

plt.subplot(2,1,2)
plt.plot(history2.history['loss'])
plt.plot(history2.history['val_loss'])
plt.title('Model Loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper right')
plt.tight_layout()

plt.show()


# 保存模型
import os
import tensorflow.gfile as gfile

save_dir = "/content/drive/model/v2/"

if gfile.Exists(save_dir):
    gfile.DeleteRecursively(save_dir)
gfile.MakeDirs(save_dir)

model_name = 'keras_traffic_2.h5'
model_path = os.path.join(save_dir, model_name)

model2.save(model_path)
print('Saved trained model at %s ' % model_path)

Saved trained model at /content/drive/model/v2/keras_traffic_2.h5 

# 加载模型
from keras.models import load_model

traffic_model = load_model(model_path)

# 统计模型在测试集上的分类结果
loss_and_metrics = traffic_model.evaluate(X_test_normalized, Y_test, verbose=2,batch_size=256)
    
print("Test Loss: {}".format(loss_and_metrics[0]))
print("Test Accuracy: {}%".format(loss_and_metrics[1]*100))

predicted_classes = traffic_model.predict_classes(X_test_normalized)

correct_indices = np.nonzero(predicted_classes == y_test)[0]
incorrect_indices = np.nonzero(predicted_classes != y_test)[0]
print("Classified correctly count: {}".format(len(correct_indices)))
print("Classified incorrectly count: {}".format(len(incorrect_indices)))

Test Loss: 0.6938359957692742
Test Accuracy: 93.08788600368152%
Classified correctly count: 11757
Classified incorrectly count: 873

3.对上述模型进行改进 增加2层Dropout层
from keras.models import Sequential
from keras.layers import Dense,Dropout,Flatten
from keras.layers import Conv2D,MaxPooling2D

model3 = Sequential()

# layers1
model3.add(Conv2D(filters=32,
                 kernel_size=(3,3),
                 input_shape=X_train_normalized.shape[1:],
                 activation="relu"))
model3.add(MaxPooling2D(pool_size=(2,2)))

# layers2
model3.add(Conv2D(filters=64,
                 kernel_size=(3,3),
                 activation="relu"))
model3.add(MaxPooling2D(pool_size=(2,2)))

# Dropout
model3.add(Dropout(0.25))

# layers3
model3.add(Conv2D(filters=64,
                 kernel_size=(3,3),
                 activation="relu"))
model3.add(MaxPooling2D(pool_size=(2,2)))

# flattern
model3.add(Flatten())

# Dense
model3.add(Dense(512,activation="relu"))

# Dropout
model3.add(Dropout(0.5))

# Dense
model3.add(Dense(n_classes,activation="softmax"))

model3.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_6 (Conv2D)            (None, 30, 30, 32)        320       
_________________________________________________________________
max_pooling2d_5 (MaxPooling2 (None, 15, 15, 32)        0         
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 13, 13, 64)        18496     
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 6, 6, 64)          0         
_________________________________________________________________
dropout_3 (Dropout)          (None, 6, 6, 64)          0         
_________________________________________________________________
conv2d_8 (Conv2D)            (None, 4, 4, 64)          36928     
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 2, 2, 64)          0         
_________________________________________________________________
flatten_3 (Flatten)          (None, 256)               0         
_________________________________________________________________
dense_5 (Dense)              (None, 512)               131584    
_________________________________________________________________
dropout_4 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_6 (Dense)              (None, 43)                22059     
=================================================================
Total params: 209,387
Trainable params: 209,387
Non-trainable params: 0
_________________________________________________________________

# 编译模型
model3.compile(loss="categorical_crossentropy",
              metrics=["accuracy"],
              optimizer="adam")

# 训练模型
history3 = model3.fit(X_train_normalized,
                    Y_train,
                    batch_size=256,
                    epochs=120,
                    verbose=2,
                    validation_data=(X_valid_normalized,Y_valid))                    

Train on 51690 samples, validate on 4410 samples
Epoch 1/120
 - 2s - loss: 1.8536 - acc: 0.4729 - val_loss: 0.6255 - val_acc: 0.8197
Epoch 2/120
 - 2s - loss: 0.4513 - acc: 0.8609 - val_loss: 0.3030 - val_acc: 0.9082
Epoch 3/120
 - 2s - loss: 0.2391 - acc: 0.9261 - val_loss: 0.2413 - val_acc: 0.9272
Epoch 4/120
 - 2s - loss: 0.1676 - acc: 0.9480 - val_loss: 0.2010 - val_acc: 0.9401
Epoch 5/120
 - 2s - loss: 0.1248 - acc: 0.9610 - val_loss: 0.1635 - val_acc: 0.9519
Epoch 6/120
 - 2s - loss: 0.1006 - acc: 0.9692 - val_loss: 0.1346 - val_acc: 0.9619
Epoch 7/120
 - 2s - loss: 0.0848 - acc: 0.9736 - val_loss: 0.1464 - val_acc: 0.9553
Epoch 8/120
 - 2s - loss: 0.0731 - acc: 0.9765 - val_loss: 0.1513 - val_acc: 0.9526
Epoch 9/120
 - 2s - loss: 0.0664 - acc: 0.9785 - val_loss: 0.1365 - val_acc: 0.9580
Epoch 10/120
 - 2s - loss: 0.0590 - acc: 0.9810 - val_loss: 0.1404 - val_acc: 0.9553
Epoch 11/120
 - 2s - loss: 0.0493 - acc: 0.9838 - val_loss: 0.1258 - val_acc: 0.9646
Epoch 12/120
 - 2s - loss: 0.0446 - acc: 0.9859 - val_loss: 0.1265 - val_acc: 0.9642
Epoch 13/120
 - 2s - loss: 0.0389 - acc: 0.9874 - val_loss: 0.1258 - val_acc: 0.9678
Epoch 14/120
 - 2s - loss: 0.0385 - acc: 0.9872 - val_loss: 0.1220 - val_acc: 0.9669
Epoch 15/120
 - 2s - loss: 0.0348 - acc: 0.9889 - val_loss: 0.1200 - val_acc: 0.9662
Epoch 16/120
 - 2s - loss: 0.0354 - acc: 0.9883 - val_loss: 0.0980 - val_acc: 0.9728
Epoch 17/120
 - 2s - loss: 0.0316 - acc: 0.9897 - val_loss: 0.1222 - val_acc: 0.9680
Epoch 18/120
 - 2s - loss: 0.0280 - acc: 0.9909 - val_loss: 0.1069 - val_acc: 0.9703
Epoch 19/120
 - 2s - loss: 0.0285 - acc: 0.9911 - val_loss: 0.1080 - val_acc: 0.9703
Epoch 20/120
 - 2s - loss: 0.0240 - acc: 0.9915 - val_loss: 0.1049 - val_acc: 0.9714
Epoch 21/120
 - 2s - loss: 0.0248 - acc: 0.9921 - val_loss: 0.1137 - val_acc: 0.9703
Epoch 22/120
 - 2s - loss: 0.0248 - acc: 0.9922 - val_loss: 0.0974 - val_acc: 0.9728
Epoch 23/120
 - 2s - loss: 0.0213 - acc: 0.9930 - val_loss: 0.1277 - val_acc: 0.9633
Epoch 24/120
 - 2s - loss: 0.0232 - acc: 0.9924 - val_loss: 0.0980 - val_acc: 0.9751
Epoch 25/120
 - 2s - loss: 0.0208 - acc: 0.9932 - val_loss: 0.1124 - val_acc: 0.9739
Epoch 26/120
 - 2s - loss: 0.0183 - acc: 0.9939 - val_loss: 0.1076 - val_acc: 0.9726
Epoch 27/120
 - 2s - loss: 0.0205 - acc: 0.9928 - val_loss: 0.0998 - val_acc: 0.9735
Epoch 28/120
 - 2s - loss: 0.0167 - acc: 0.9944 - val_loss: 0.1231 - val_acc: 0.9701
Epoch 29/120
 - 2s - loss: 0.0188 - acc: 0.9940 - val_loss: 0.1025 - val_acc: 0.9737
Epoch 30/120
 - 2s - loss: 0.0164 - acc: 0.9947 - val_loss: 0.1189 - val_acc: 0.9701

Epoch 110/120
 - 2s - loss: 0.0069 - acc: 0.9980 - val_loss: 0.0982 - val_acc: 0.9805
Epoch 111/120
 - 2s - loss: 0.0086 - acc: 0.9976 - val_loss: 0.0940 - val_acc: 0.9816
Epoch 112/120
 - 2s - loss: 0.0057 - acc: 0.9982 - val_loss: 0.0958 - val_acc: 0.9807
Epoch 113/120
 - 2s - loss: 0.0066 - acc: 0.9981 - val_loss: 0.1343 - val_acc: 0.9780
Epoch 114/120
 - 2s - loss: 0.0059 - acc: 0.9982 - val_loss: 0.1218 - val_acc: 0.9796
Epoch 115/120
 - 2s - loss: 0.0074 - acc: 0.9979 - val_loss: 0.1266 - val_acc: 0.9771
Epoch 116/120
 - 2s - loss: 0.0063 - acc: 0.9981 - val_loss: 0.1001 - val_acc: 0.9819
Epoch 117/120
 - 2s - loss: 0.0063 - acc: 0.9980 - val_loss: 0.0945 - val_acc: 0.9819
Epoch 118/120
 - 2s - loss: 0.0082 - acc: 0.9976 - val_loss: 0.1334 - val_acc: 0.9751
Epoch 119/120
 - 2s - loss: 0.0073 - acc: 0.9976 - val_loss: 0.1341 - val_acc: 0.9735
Epoch 120/120
 - 2s - loss: 0.0072 - acc: 0.9977 - val_loss: 0.1004 - val_acc: 0.9794

# 可视化指标
fig = plt.figure()
plt.subplot(2,1,1)
plt.plot(history3.history['acc'])
plt.plot(history3.history['val_acc'])
plt.title('Model Accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='lower right')

plt.subplot(2,1,2)
plt.plot(history3.history['loss'])
plt.plot(history3.history['val_loss'])
plt.title('Model Loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper right')
plt.tight_layout()

plt.show()


# 保存模型
import os
import tensorflow.gfile as gfile

save_dir = "/content/drive/model/v3/"

if gfile.Exists(save_dir):
    gfile.DeleteRecursively(save_dir)
gfile.MakeDirs(save_dir)

model_name = 'keras_traffic_3.h5'
model_path = os.path.join(save_dir, model_name)

model3.save(model_path)
print('Saved trained model at %s ' % model_path)

Saved trained model at /content/drive/model/v3/keras_traffic_3.h5 

# 加载模型
from keras.models import load_model

traffic_model = load_model(model_path)

# 统计模型在测试集上的分类结果
loss_and_metrics = traffic_model.evaluate(X_test_normalized, Y_test, verbose=2,batch_size=256)
    
print("Test Loss: {}".format(loss_and_metrics[0]))
print("Test Accuracy: {}%".format(loss_and_metrics[1]*100))

predicted_classes = traffic_model.predict_classes(X_test_normalized)

correct_indices = np.nonzero(predicted_classes == y_test)[0]
incorrect_indices = np.nonzero(predicted_classes != y_test)[0]
print("Classified correctly count: {}".format(len(correct_indices)))
print("Classified incorrectly count: {}".format(len(incorrect_indices)))

Test Loss: 0.2989197100426409
Test Accuracy: 95.88281867056732%
Classified correctly count: 12110
Classified incorrectly count: 520

4.2层卷积池化层+1层展开层+3层全连接层
from keras.models import Sequential
from keras.layers import Dense,Dropout,Flatten
from keras.layers import Conv2D,MaxPooling2D

model4 = Sequential()

# layers1
model4.add(Conv2D(filters=32,
                 kernel_size=(3,3),
                 input_shape=X_train_normalized.shape[1:],
                 activation="relu"))
model4.add(MaxPooling2D(pool_size=(2,2)))

# layers2
model4.add(Conv2D(filters=64,
                 kernel_size=(3,3),
                 activation="relu"))
model4.add(MaxPooling2D(pool_size=(2,2)))



# flattern
model4.add(Flatten())

# Dense
model4.add(Dense(512,activation="relu"))

# Dense
model4.add(Dense(100,activation="relu"))

# Dense
model4.add(Dense(n_classes,activation="softmax"))

# 查看模型的结构
model4.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_9 (Conv2D)            (None, 30, 30, 32)        320       
_________________________________________________________________
max_pooling2d_8 (MaxPooling2 (None, 15, 15, 32)        0         
_________________________________________________________________
conv2d_10 (Conv2D)           (None, 13, 13, 64)        18496     
_________________________________________________________________
max_pooling2d_9 (MaxPooling2 (None, 6, 6, 64)          0         
_________________________________________________________________
flatten_4 (Flatten)          (None, 2304)              0         
_________________________________________________________________
dense_7 (Dense)              (None, 512)               1180160   
_________________________________________________________________
dense_8 (Dense)              (None, 100)               51300     
_________________________________________________________________
dense_9 (Dense)              (None, 43)                4343      
=================================================================
Total params: 1,254,619
Trainable params: 1,254,619
Non-trainable params: 0
_________________________________________________________________

# 编译模型
model4.compile(loss="categorical_crossentropy",
              metrics=["accuracy"],
              optimizer="adam")

# 训练模型
history4 = model4.fit(X_train_normalized,
                    Y_train,
                    batch_size=256,
                    epochs=120,
                    verbose=2,
                    validation_data=(X_valid_normalized,Y_valid))                    

Train on 51690 samples, validate on 4410 samples
Epoch 1/120
 - 3s - loss: 0.9466 - acc: 0.7433 - val_loss: 0.4214 - val_acc: 0.8746
Epoch 2/120
 - 2s - loss: 0.1339 - acc: 0.9632 - val_loss: 0.3042 - val_acc: 0.9152
Epoch 3/120
 - 2s - loss: 0.0681 - acc: 0.9809 - val_loss: 0.2988 - val_acc: 0.9170
Epoch 4/120
 - 2s - loss: 0.0414 - acc: 0.9885 - val_loss: 0.3027 - val_acc: 0.9125
Epoch 5/120
 - 2s - loss: 0.0244 - acc: 0.9934 - val_loss: 0.2754 - val_acc: 0.9308
Epoch 6/120
 - 2s - loss: 0.0198 - acc: 0.9944 - val_loss: 0.2664 - val_acc: 0.9299
Epoch 7/120
 - 2s - loss: 0.0139 - acc: 0.9963 - val_loss: 0.3340 - val_acc: 0.9197
Epoch 8/120
 - 2s - loss: 0.0150 - acc: 0.9958 - val_loss: 0.2918 - val_acc: 0.9336
Epoch 9/120
 - 2s - loss: 0.0166 - acc: 0.9948 - val_loss: 0.2496 - val_acc: 0.9399
Epoch 10/120
 - 2s - loss: 0.0070 - acc: 0.9979 - val_loss: 0.2661 - val_acc: 0.9365
Epoch 11/120
 - 2s - loss: 0.0060 - acc: 0.9984 - val_loss: 0.3019 - val_acc: 0.9290
Epoch 12/120
 - 2s - loss: 0.0063 - acc: 0.9984 - val_loss: 0.2890 - val_acc: 0.9347
Epoch 13/120
 - 2s - loss: 0.0076 - acc: 0.9979 - val_loss: 0.2583 - val_acc: 0.9363
Epoch 14/120
 - 2s - loss: 0.0124 - acc: 0.9962 - val_loss: 0.3169 - val_acc: 0.9240
Epoch 15/120
 - 2s - loss: 0.0097 - acc: 0.9970 - val_loss: 0.2796 - val_acc: 0.9429
Epoch 16/120
 - 2s - loss: 0.0088 - acc: 0.9975 - val_loss: 0.3084 - val_acc: 0.9320
Epoch 17/120
 - 2s - loss: 0.0093 - acc: 0.9972 - val_loss: 0.3469 - val_acc: 0.9281
Epoch 18/120
 - 2s - loss: 0.0029 - acc: 0.9991 - val_loss: 0.2834 - val_acc: 0.9388
Epoch 19/120
 - 2s - loss: 9.1336e-04 - acc: 0.9998 - val_loss: 0.2210 - val_acc: 0.9535
Epoch 20/120
 - 2s - loss: 3.3410e-04 - acc: 0.9999 - val_loss: 0.2416 - val_acc: 0.9546
Epoch 21/120
 - 2s - loss: 4.8947e-05 - acc: 1.0000 - val_loss: 0.2395 - val_acc: 0.9556
Epoch 22/120
 - 2s - loss: 2.7690e-05 - acc: 1.0000 - val_loss: 0.2449 - val_acc: 0.9556
Epoch 23/120
 - 2s - loss: 2.0403e-05 - acc: 1.0000 - val_loss: 0.2481 - val_acc: 0.9556
Epoch 24/120
 - 2s - loss: 1.6893e-05 - acc: 1.0000 - val_loss: 0.2507 - val_acc: 0.9553
Epoch 25/120
 - 2s - loss: 1.4042e-05 - acc: 1.0000 - val_loss: 0.2534 - val_acc: 0.9553
Epoch 26/120
 - 2s - loss: 1.2080e-05 - acc: 1.0000 - val_loss: 0.2556 - val_acc: 0.9556
Epoch 27/120
 - 2s - loss: 1.0456e-05 - acc: 1.0000 - val_loss: 0.2577 - val_acc: 0.9551
Epoch 28/120
 - 2s - loss: 9.1271e-06 - acc: 1.0000 - val_loss: 0.2589 - val_acc: 0.9560
Epoch 29/120
 - 2s - loss: 8.0037e-06 - acc: 1.0000 - val_loss: 0.2610 - val_acc: 0.9549
Epoch 30/120
 - 2s - loss: 7.1773e-06 - acc: 1.0000 - val_loss: 0.2630 - val_acc: 0.9560

Epoch 110/120
 - 2s - loss: 1.2023e-07 - acc: 1.0000 - val_loss: 0.3727 - val_acc: 0.9533
Epoch 111/120
 - 2s - loss: 1.2010e-07 - acc: 1.0000 - val_loss: 0.3727 - val_acc: 0.9537
Epoch 112/120
 - 2s - loss: 1.1994e-07 - acc: 1.0000 - val_loss: 0.3716 - val_acc: 0.9535
Epoch 113/120
 - 2s - loss: 1.1984e-07 - acc: 1.0000 - val_loss: 0.3747 - val_acc: 0.9535
Epoch 114/120
 - 2s - loss: 1.1972e-07 - acc: 1.0000 - val_loss: 0.3758 - val_acc: 0.9535
Epoch 115/120
 - 2s - loss: 1.1968e-07 - acc: 1.0000 - val_loss: 0.3763 - val_acc: 0.9533
Epoch 116/120
 - 2s - loss: 1.1957e-07 - acc: 1.0000 - val_loss: 0.3789 - val_acc: 0.9528
Epoch 117/120
 - 2s - loss: 1.1953e-07 - acc: 1.0000 - val_loss: 0.3783 - val_acc: 0.9531
Epoch 118/120
 - 2s - loss: 1.1947e-07 - acc: 1.0000 - val_loss: 0.3796 - val_acc: 0.9528
Epoch 119/120
 - 2s - loss: 1.1943e-07 - acc: 1.0000 - val_loss: 0.3805 - val_acc: 0.9535
Epoch 120/120
 - 2s - loss: 1.1942e-07 - acc: 1.0000 - val_loss: 0.3811 - val_acc: 0.9533

# 可视化指标
fig = plt.figure()
plt.subplot(2,1,1)
plt.plot(history4.history['acc'])
plt.plot(history4.history['val_acc'])
plt.title('Model Accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='lower right')

plt.subplot(2,1,2)
plt.plot(history4.history['loss'])
plt.plot(history4.history['val_loss'])
plt.title('Model Loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper right')
plt.tight_layout()

plt.show()



# 保存模型
import os
import tensorflow.gfile as gfile

save_dir = "/content/drive/model/v4/"

if gfile.Exists(save_dir):
    gfile.DeleteRecursively(save_dir)
gfile.MakeDirs(save_dir)

model_name = 'keras_traffic_4.h5'
model_path = os.path.join(save_dir, model_name)

model4.save(model_path)
print('Saved trained model at %s ' % model_path)


Saved trained model at /content/drive/model/v4/keras_traffic_4.h5 

# 加载模型
from keras.models import load_model

traffic_model = load_model(model_path)

# 统计模型在测试集上的分类结果
loss_and_metrics = traffic_model.evaluate(X_test_normalized, Y_test, verbose=2,batch_size=256)
    
print("Test Loss: {}".format(loss_and_metrics[0]))
print("Test Accuracy: {}%".format(loss_and_metrics[1]*100))

predicted_classes = traffic_model.predict_classes(X_test_normalized)

correct_indices = np.nonzero(predicted_classes == y_test)[0]
incorrect_indices = np.nonzero(predicted_classes != y_test)[0]
print("Classified correctly count: {}".format(len(correct_indices)))
print("Classified incorrectly count: {}".format(len(incorrect_indices)))

Test Loss: 0.61759698928016
Test Accuracy: 93.18289788016608%
Classified correctly count: 11769
Classified incorrectly count: 861

5.对上述模型进行改进增加3层Dropout层
# 改进模型
from keras.models import Sequential
from keras.layers import Dense,Dropout,Flatten
from keras.layers import Conv2D,MaxPooling2D

model5 = Sequential()

# layers1
model5.add(Conv2D(filters=32,
                 kernel_size=(3,3),
                 input_shape=X_train_normalized.shape[1:],
                 activation="relu"))
model5.add(MaxPooling2D(pool_size=(2,2)))



# layers2
model5.add(Conv2D(filters=64,
                 kernel_size=(3,3),
                 activation="relu"))
model5.add(MaxPooling2D(pool_size=(2,2)))

# Dropout
model5.add(Dropout(0.5))

# flattern
model5.add(Flatten())

# Dense
model5.add(Dense(512,activation="relu"))

# Dropout
model5.add(Dropout(0.5))

# Dense
model5.add(Dense(100,activation="relu"))

# Dropout
model5.add(Dropout(0.5))

# Dense
model5.add(Dense(n_classes,activation="softmax"))

model5.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_11 (Conv2D)           (None, 30, 30, 32)        320       
_________________________________________________________________
max_pooling2d_10 (MaxPooling (None, 15, 15, 32)        0         
_________________________________________________________________
conv2d_12 (Conv2D)           (None, 13, 13, 64)        18496     
_________________________________________________________________
max_pooling2d_11 (MaxPooling (None, 6, 6, 64)          0         
_________________________________________________________________
dropout_5 (Dropout)          (None, 6, 6, 64)          0         
_________________________________________________________________
flatten_5 (Flatten)          (None, 2304)              0         
_________________________________________________________________
dense_10 (Dense)             (None, 512)               1180160   
_________________________________________________________________
dropout_6 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_11 (Dense)             (None, 100)               51300     
_________________________________________________________________
dropout_7 (Dropout)          (None, 100)               0         
_________________________________________________________________
dense_12 (Dense)             (None, 43)                4343      
=================================================================
Total params: 1,254,619
Trainable params: 1,254,619
Non-trainable params: 0
_________________________________________________________________

# 编译模型
model5.compile(loss="categorical_crossentropy",
              metrics=["accuracy"],
              optimizer="adam")

# 训练模型
history5 = model5.fit(X_train_normalized,
                    Y_train,
                    batch_size=256,
                    epochs=120,
                    verbose=2,
                    validation_data=(X_valid_normalized,Y_valid))                    


Train on 51690 samples, validate on 4410 samples
Epoch 1/120
 - 3s - loss: 2.0740 - acc: 0.4183 - val_loss: 0.5267 - val_acc: 0.8590
Epoch 2/120
 - 2s - loss: 0.5917 - acc: 0.8149 - val_loss: 0.2372 - val_acc: 0.9395
Epoch 3/120
 - 2s - loss: 0.3512 - acc: 0.8919 - val_loss: 0.1723 - val_acc: 0.9508
Epoch 4/120
 - 2s - loss: 0.2560 - acc: 0.9214 - val_loss: 0.1449 - val_acc: 0.9583
Epoch 5/120
 - 2s - loss: 0.2048 - acc: 0.9382 - val_loss: 0.1341 - val_acc: 0.9590
Epoch 6/120
 - 2s - loss: 0.1682 - acc: 0.9497 - val_loss: 0.1183 - val_acc: 0.9655
Epoch 7/120
 - 2s - loss: 0.1406 - acc: 0.9573 - val_loss: 0.1070 - val_acc: 0.9698
Epoch 8/120
 - 2s - loss: 0.1288 - acc: 0.9604 - val_loss: 0.1037 - val_acc: 0.9667
Epoch 9/120
 - 2s - loss: 0.1175 - acc: 0.9641 - val_loss: 0.1056 - val_acc: 0.9685
Epoch 10/120
 - 2s - loss: 0.1016 - acc: 0.9690 - val_loss: 0.0935 - val_acc: 0.9696
Epoch 11/120
 - 2s - loss: 0.0940 - acc: 0.9709 - val_loss: 0.0856 - val_acc: 0.9741
Epoch 12/120
 - 2s - loss: 0.0896 - acc: 0.9729 - val_loss: 0.0901 - val_acc: 0.9741
Epoch 13/120
 - 2s - loss: 0.0857 - acc: 0.9744 - val_loss: 0.0879 - val_acc: 0.9719
Epoch 14/120
 - 2s - loss: 0.0803 - acc: 0.9757 - val_loss: 0.0894 - val_acc: 0.9723
Epoch 15/120
 - 2s - loss: 0.0740 - acc: 0.9782 - val_loss: 0.1040 - val_acc: 0.9705
Epoch 16/120
 - 2s - loss: 0.0708 - acc: 0.9796 - val_loss: 0.1073 - val_acc: 0.9683
Epoch 17/120
 - 2s - loss: 0.0652 - acc: 0.9800 - val_loss: 0.0791 - val_acc: 0.9753
Epoch 18/120
 - 2s - loss: 0.0660 - acc: 0.9801 - val_loss: 0.0794 - val_acc: 0.9751
Epoch 19/120
 - 2s - loss: 0.0571 - acc: 0.9828 - val_loss: 0.0925 - val_acc: 0.9705
Epoch 20/120
 - 2s - loss: 0.0573 - acc: 0.9828 - val_loss: 0.0692 - val_acc: 0.9791
Epoch 21/120
 - 2s - loss: 0.0532 - acc: 0.9834 - val_loss: 0.0726 - val_acc: 0.9785
Epoch 22/120
 - 2s - loss: 0.0548 - acc: 0.9837 - val_loss: 0.0824 - val_acc: 0.9741
Epoch 23/120
 - 2s - loss: 0.0529 - acc: 0.9841 - val_loss: 0.0747 - val_acc: 0.9789
Epoch 24/120
 - 2s - loss: 0.0544 - acc: 0.9839 - val_loss: 0.0710 - val_acc: 0.9776
Epoch 25/120
 - 2s - loss: 0.0470 - acc: 0.9859 - val_loss: 0.0687 - val_acc: 0.9789
Epoch 26/120
 - 2s - loss: 0.0459 - acc: 0.9860 - val_loss: 0.0710 - val_acc: 0.9769
Epoch 27/120
 - 2s - loss: 0.0482 - acc: 0.9859 - val_loss: 0.0718 - val_acc: 0.9764
Epoch 28/120
 - 2s - loss: 0.0456 - acc: 0.9862 - val_loss: 0.0744 - val_acc: 0.9812
Epoch 29/120
 - 2s - loss: 0.0469 - acc: 0.9861 - val_loss: 0.0883 - val_acc: 0.9762
Epoch 30/120
 - 2s - loss: 0.0445 - acc: 0.9863 - val_loss: 0.0757 - val_acc: 0.9757

Epoch 110/120
 - 2s - loss: 0.0217 - acc: 0.9941 - val_loss: 0.0600 - val_acc: 0.9832
Epoch 111/120
 - 2s - loss: 0.0243 - acc: 0.9936 - val_loss: 0.0716 - val_acc: 0.9800
Epoch 112/120
 - 2s - loss: 0.0204 - acc: 0.9941 - val_loss: 0.0740 - val_acc: 0.9794
Epoch 113/120
 - 2s - loss: 0.0178 - acc: 0.9948 - val_loss: 0.0669 - val_acc: 0.9828
Epoch 114/120
 - 2s - loss: 0.0184 - acc: 0.9949 - val_loss: 0.0569 - val_acc: 0.9850
Epoch 115/120
 - 2s - loss: 0.0224 - acc: 0.9943 - val_loss: 0.0703 - val_acc: 0.9791
Epoch 116/120
 - 2s - loss: 0.0203 - acc: 0.9944 - val_loss: 0.0762 - val_acc: 0.9789
Epoch 117/120
 - 2s - loss: 0.0182 - acc: 0.9945 - val_loss: 0.0705 - val_acc: 0.9798
Epoch 118/120
 - 2s - loss: 0.0204 - acc: 0.9944 - val_loss: 0.0603 - val_acc: 0.9819
Epoch 119/120
 - 2s - loss: 0.0188 - acc: 0.9950 - val_loss: 0.0544 - val_acc: 0.9848
Epoch 120/120
 - 2s - loss: 0.0198 - acc: 0.9947 - val_loss: 0.0593 - val_acc: 0.9857

# 可视化指标
fig = plt.figure()
plt.subplot(2,1,1)
plt.plot(history5.history['acc'])
plt.plot(history5.history['val_acc'])
plt.title('Model Accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='lower right')

plt.subplot(2,1,2)
plt.plot(history5.history['loss'])
plt.plot(history5.history['val_loss'])
plt.title('Model Loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper right')
plt.tight_layout()

plt.show()


# 保存模型
import os
import tensorflow.gfile as gfile

save_dir = "/content/drive/model/v5/"

if gfile.Exists(save_dir):
    gfile.DeleteRecursively(save_dir)
gfile.MakeDirs(save_dir)

model_name = 'keras_traffic_5.h5'
model_path = os.path.join(save_dir, model_name)

model5.save(model_path)
print('Saved trained model at %s ' % model_path)


Saved trained model at /content/drive/model/v5/keras_traffic_5.h5 

# 加载模型
from keras.models import load_model

traffic_model = load_model(model_path)

# 统计模型在测试集上的分类结果
loss_and_metrics = traffic_model.evaluate(X_test_normalized, Y_test, verbose=2,batch_size=256)
    
print("Test Loss: {}".format(loss_and_metrics[0]))
print("Test Accuracy: {}%".format(loss_and_metrics[1]*100))

predicted_classes = traffic_model.predict_classes(X_test_normalized)

correct_indices = np.nonzero(predicted_classes == y_test)[0]
incorrect_indices = np.nonzero(predicted_classes != y_test)[0]
print("Classified correctly count: {}".format(len(correct_indices)))
print("Classified incorrectly count: {}".format(len(incorrect_indices)))

Test Loss: 0.15594110083032578
Test Accuracy: 96.76167853182493%
Classified correctly count: 12221
Classified incorrectly count: 409



补充单张图片预测
# 加载模型
from keras.models import load_model

traffic_model = load_model(model_path)

# 加载预测图片

import os
import cv2
import matplotlib.image as mpimg
import matplotlib.pyplot as plt
import numpy as np

image_path = "/content/drive/image/"
filenames = os.listdir(image_path)

X_predict = []

for filename in filenames:
  if filename.endswith(".jpg"):
    img = cv2.imread(image_path+filename)
    img = cv2.resize(img,(32,32),cv2.INTER_LINEAR)
    X_predict.append(img)
    plt.imshow(img)


# 封装数据预处理的方法
import cv2

def preprocess_features(X, equalize_hist=True):
    normalized_X = []
    for i in range(len(X)):
        # Convert from RGB to YUV
        yuv_img = cv2.cvtColor(X[i], cv2.COLOR_RGB2YUV)
        yuv_img_v = X[i][:, :, 0]
        # equalizeHist
        yuv_img_v = cv2.equalizeHist(yuv_img_v)
        # expand_dis
        yuv_img_v = np.expand_dims(yuv_img_v, 2)
        normalized_X.append(yuv_img_v)
    # normalize
    normalized_X = np.array(normalized_X, dtype=np.float32)
    normalized_X = (normalized_X-128)/128
    # normalized_X /= (np.std(normalized_X, axis=0) + np.finfo('float32').eps)
    return normalized_X

# 对数据进行处理
X_predict_normalized = preprocess_features(X_predict)

# 数据维度及数据类型
print(X_predict_normalized.shape,type(X_predict_normalized))

# 模型预测
y_predict = traffic_model.predict_classes(X_predict_normalized)
print(id_to_name(y_predict[0]))