一、前言

我的环境：

• 语言环境：Python3.6.5
• 编译器：jupyter notebook
• 深度学习环境：TensorFlow2.4.1

往期精彩内容：

• 卷积神经网络（CNN）实现mnist手写数字识别
• 卷积神经网络（CNN）多种图片分类的实现
• 卷积神经网络（CNN）衣服图像分类的实现
• 卷积神经网络（CNN）鲜花识别
• 卷积神经网络（CNN）天气识别
• 卷积神经网络（VGG-16）识别海贼王草帽一伙
• 卷积神经网络（ResNet-50）鸟类识别
• 卷积神经网络（AlexNet）鸟类识别
• 卷积神经网络(CNN)识别验证码
• 卷积神经网络（Inception-ResNet-v2）交通标志识别

来自专栏：机器学习与深度学习算法推荐

二、前期工作

1. 设置GPU（如果使用的是CPU可以忽略这步）

import tensorflow as tfgpus = tf.config.list_physical_devices("GPU")if gpus:tf.config.experimental.set_memory_growth(gpus[0], True)  #设置GPU显存用量按需使用tf.config.set_visible_devices([gpus[0]],"GPU")

2. 导入数据

数据集链接

import matplotlib.pyplot as plt
# 支持中文
plt.rcParams['font.sans-serif'] = ['SimHei']  # 用来正常显示中文标签
plt.rcParams['axes.unicode_minus'] = False  # 用来正常显示负号import os,PIL,random,pathlib# 设置随机种子尽可能使结果可以重现
import numpy as np
np.random.seed(1)# 设置随机种子尽可能使结果可以重现
import tensorflow as tf
tf.random.set_seed(1)

data_dir = "015_licence_plate"
data_dir = pathlib.Path(data_dir)pictures_paths = list(data_dir.glob('*'))
pictures_paths = [str(path) for path in pictures_paths]
pictures_paths[:3]

3. 查看数据

image_count = len(list(pictures_paths))print("图片总数为：",image_count)

图片总数为： 13056

# 获取数据标签
all_label_names = [path.split("_")[-1].split(".")[0] for path in pictures_paths]
all_label_names[:3]

['川W9BR26', '沪E264UD', '浙E198UJ']

3.数据可视化

plt.figure(figsize=(10,5))
plt.suptitle("数据示例",fontsize=15)for i in range(20):plt.subplot(5,4,i+1)plt.xticks([])plt.yticks([])plt.grid(False)# 显示图片images = plt.imread(pictures_paths[i])plt.imshow(images)# 显示标签plt.xlabel(all_label_names[i],fontsize=13)plt.show()

4.标签数字化

char_enum = ["京","沪","津","渝","冀","晋","蒙","辽","吉","黑","苏","浙","皖","闽","赣","鲁",\"豫","鄂","湘","粤","桂","琼","川","贵","云","藏","陕","甘","青","宁","新","军","使"]number   = [str(i) for i in range(0, 10)]    # 0 到 9 的数字
alphabet = [chr(i) for i in range(65, 91)]   # A 到 Z 的字母char_set       = char_enum + number + alphabet
char_set_len   = len(char_set)
label_name_len = len(all_label_names[0])# 将字符串数字化
def text2vec(text):vector = np.zeros([label_name_len, char_set_len])for i, c in enumerate(text):idx = char_set.index(c)vector[i][idx] = 1.0return vectorall_labels = [text2vec(i) for i in all_label_names]

二、构建一个tf.data.Dataset

1.预处理函数

def preprocess_image(image):image = tf.image.decode_jpeg(image, channels=1)image = tf.image.resize(image, [50, 200])return image/255.0def load_and_preprocess_image(path):image = tf.io.read_file(path)return preprocess_image(image)

2.加载数据

构建 tf.data.Dataset 最简单的方法就是使用 from_tensor_slices 方法。

AUTOTUNE = tf.data.experimental.AUTOTUNEpath_ds  = tf.data.Dataset.from_tensor_slices(pictures_paths)
image_ds = path_ds.map(load_and_preprocess_image, num_parallel_calls=AUTOTUNE)
label_ds = tf.data.Dataset.from_tensor_slices(all_labels)image_label_ds = tf.data.Dataset.zip((image_ds, label_ds))
image_label_ds

train_ds = image_label_ds.take(5000).shuffle(5000)  # 前1000个batch
val_ds   = image_label_ds.skip(5000).shuffle(1000)  # 跳过前1000，选取后面的

3.配置数据

BATCH_SIZE = 16train_ds = train_ds.batch(BATCH_SIZE)
train_ds = train_ds.prefetch(buffer_size=AUTOTUNE)val_ds = val_ds.batch(BATCH_SIZE)
val_ds = val_ds.prefetch(buffer_size=AUTOTUNE)
val_ds

三、搭建网络模型

目前这里主要是带大家跑通代码、整理一下思路，大家可以自行优化网络结构、调整模型参数。后续我也会针对性的出一些调优的案例的。

from tensorflow.keras import datasets, layers, modelsmodel = models.Sequential([layers.Conv2D(32, (3, 3), activation='relu', input_shape=(50, 200, 1)),#卷积层1，卷积核3*3layers.MaxPooling2D((2, 2)),                   #池化层1，2*2采样layers.Conv2D(64, (3, 3), activation='relu'),  #卷积层2，卷积核3*3layers.MaxPooling2D((2, 2)),                   #池化层2，2*2采样layers.Flatten(),                              #Flatten层，连接卷积层与全连接层
#     layers.Dense(1000, activation='relu'),         #全连接层，特征进一步提取layers.Dense(1000, activation='relu'),         #全连接层，特征进一步提取layers.Dropout(0.3),  layers.Dense(label_name_len * char_set_len),layers.Reshape([label_name_len, char_set_len]),layers.Softmax()                               #输出层，输出预期结果
])
# 打印网络结构
model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 48, 198, 32)       320       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 24, 99, 32)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 22, 97, 64)        18496     
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 11, 48, 64)        0         
_________________________________________________________________
flatten (Flatten)            (None, 33792)             0         
_________________________________________________________________
dense (Dense)                (None, 1000)              33793000  
_________________________________________________________________
dropout (Dropout)            (None, 1000)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 483)               483483    
_________________________________________________________________
reshape (Reshape)            (None, 7, 69)             0         
_________________________________________________________________
softmax (Softmax)            (None, 7, 69)             0         
=================================================================
Total params: 34,295,299
Trainable params: 34,295,299
Non-trainable params: 0
_________________________________________________________________

四、设置动态学习率

这里先罗列一下学习率大与学习率小的优缺点。

• 学习率大
  • 优点： 1、加快学习速率。 2、有助于跳出局部最优值。
  • 缺点： 1、导致模型训练不收敛。 2、单单使用大学习率容易导致模型不精确。
• 学习率小
  • 优点： 1、有助于模型收敛、模型细化。 2、提高模型精度。
  • 缺点： 1、很难跳出局部最优值。 2、收敛缓慢。

注意：这里设置的动态学习率为：指数衰减型（ExponentialDecay）。在每一个epoch开始前，学习率（learning_rate）都将会重置为初始学习率（initial_learning_rate），然后再重新开始衰减。计算公式如下：

learning_rate = initial_learning_rate * decay_rate ^ (step / decay_steps)

# 设置初始学习率
initial_learning_rate = 1e-3lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(initial_learning_rate, decay_steps=50,      # 敲黑板！！！这里是指 steps，不是指epochsdecay_rate=0.96,     # lr经过一次衰减就会变成 decay_rate*lrstaircase=True)# 将指数衰减学习率送入优化器
optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule)

五、编译

model.compile(optimizer=optimizer,loss='categorical_crossentropy',metrics=['accuracy'])

六、训练

epochs = 50history = model.fit(train_ds,validation_data=val_ds,epochs=epochs
)

八、保存和加载模型

# 保存模型
model.save('model/15_model.h5')

# 加载模型
new_model = tf.keras.models.load_model('model/15_model.h5')

九、预测

def vec2text(vec):"""还原标签（向量->字符串）"""text = []for i, c in enumerate(vec):text.append(char_set[c])return "".join(text)plt.figure(figsize=(10, 8))            # 图形的宽为10高为8for images, labels in val_ds.take(1):for i in range(6):ax = plt.subplot(5, 2, i + 1)  # 显示图片plt.imshow(images[i])# 需要给图片增加一个维度img_array = tf.expand_dims(images[i], 0) # 使用模型预测验证码predictions = model.predict(img_array)plt.title(vec2text(np.argmax(predictions, axis=2)[0]),fontsize=15)plt.axis("off")