[tensorflow2笔记十] 搭建卷积网络实现猫狗图片分类

1.制作数据集

（1）下载数据集。从网上下载kaggle猫狗分类的数据集，为缩短训练时间，选择2000张图片（猫狗各1000张）作为训练集，200张图片（猫狗各100张）作为测试集。在train文件夹选0-1999的猫和0-1999的狗作为训练集，选2000-2099的猫和2000-2099的狗作为测试集。

（2）调整图片的大小。图片大小不一，需要调整图片的大小，重新设定规格（244，244，3），从而在后续的网络模型输入时，保证输入到模型中的图片大小一致。

##### resize_data.py #####

import os
import cv2

dir_train = "/home/xiaobin/PycharmProjects/figure/train_1000"
dir_test = "/home/xiaobin/PycharmProjects/figure/test_200"

for root, dirs, files in os.walk(dir_test):
    for file in files:
        filepath = os.path.join(root, file)
        image = cv2.imread(filepath)
        dim = (224, 224)
        resized = cv2.resize(image, dim)
        path = "/home/xiaobin/PycharmProjects/figure/test/" + file
        cv2.imwrite(path, resized)

# os.walk() 方法是一个简单易用的文件、目录遍历器
# root 所指的是当前正在遍历的这个文件夹的本身的地址
# dirs 是一个list ，内容是该文件夹中所有的目录的名字(不包括子目录)
# files 同样是list , 内容是该文件夹中所有的文件(不包括子目录)

（3）制作标签文档。为让图片和标签匹配，制作训练集和测试集图片的索引文本。编写代码实现：

##### make_txt.py ########

f1 = open("train.txt", 'w')
for i in range(1000):
    f1.write("cat.%d.jpg %d\n" % (i, 0))
for j in range(1000):
    f1.write("dog.%d.jpg %d\n" % (j, 1))
f1.close()

f2 = open("test.txt", "w")
for i in range(100):
    f2.write("cat.%d.jpg %d\n" % (i+2000, 0))
for j in range(100):
    f2.write("dog.%d.jpg %d\n" % (j+2000, 1))
f2.close()

2.搭建网络训练

# 1.导入一些模块
import cv2
import tensorflow as tf
import numpy as np          # 用于数据格式转换
import os                   # 路径
from tensorflow.keras.preprocessing.image import ImageDataGenerator     # 数据增强
from matplotlib import pyplot as plt
from tensorflow.keras.layers import Dense,Activation,Dropout,Conv2D,BatchNormalization,MaxPool2D,Flatten

# 设置GPU显存按需申请
gpu = tf.config.experimental.list_physical_devices(device_type='GPU')
assert len(gpu) == 1
tf.config.experimental.set_memory_growth(gpu[0], True)

# 2.路径和存储文件
train_path = './train_2000/'             # 训练集图片路径
train_txt = 'train.txt'                  # 训练集标签文件

test_path = './test_200/'                # 测试集图片路径
test_txt = 'test.txt'                    # 测试集标签文件

# 3.制作数据集的函数
def generateds(path, txt):      # 图片路径，标签文件
    f = open(txt, 'r')          # 以只读的形式打开txt
    contents = f.readlines()    # 读取文件中所有的行，每行为一个单位
    f.close()
    x, y_ = [], []
    for content in contents:     # 逐行读出
        value = content.split()  # 以空格分开
        img_path = path + value[0]
        img = cv2.imread(img_path)
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)      # 转为RGB
        img = img / 255.0       # 归一化，有利于网络吸收
        x.append(img)
        y_.append(value[1])
        # print('load:' + content)

    x = np.array(x)
    y_ = np.array(y_)
    y_ = y_.astype(np.int64)
    return x, y_

# 4.加载数据
print('----------------Generate Datasets--------------')
x_train, y_train = generateds(train_path, train_txt)
x_test, y_test = generateds(test_path, test_txt)

# 5.打乱数据集样本顺序
index = [i for i in range(len(x_train))]
np.random.shuffle(index)
x_train = x_train[index]
y_train = y_train[index]

index1 = [j for j in range(len(x_test))]
np.random.shuffle(index1)
x_test = x_test[index1]
y_test = y_test[index1]

# 6.数据增强
image_gen_train = ImageDataGenerator(
    rescale=1. / 1.,         # 如果是图像，分母为255，可以归一化到0-1
    rotation_range=45,       # 随机45度旋转
    width_shift_range=.15,   # 宽度偏移
    height_shift_range=.15,  # 高度偏移
    horizontal_flip=True,    # 水平翻转
    zoom_range=0.5           # 将图像随机缩放阈量50%
)
image_gen_train.fit(x_train)

# 7.搭建网络
model = tf.keras.models.Sequential([
    Conv2D(filters=32, kernel_size=(3, 3)),
    BatchNormalization(),
    Activation('relu'),
    MaxPool2D(pool_size=(2, 2), strides=2),

    Conv2D(filters=64, kernel_size=(3, 3)),
    BatchNormalization(),
    Activation('relu'),
    MaxPool2D(pool_size=(2, 2), strides=2),

    Conv2D(filters=128, kernel_size=(3, 3)),
    BatchNormalization(),
    Activation('relu'),
    MaxPool2D(pool_size=(2, 2), strides=2),

    Flatten(),                      # 把输入特征拉直为一维数组数值
    Dense(128, activation='relu'),
    Dense(64, activation='relu'),
    Dense(2, activation='softmax')
])

# 8.配置参数
model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False),
              metrics=['sparse_categorical_accuracy'])

# 9.设置调用和保存模型
# 调用模型
checkpoint_save_path = "./checkpoint/cat_dag.ckpt"
if os.path.exists(checkpoint_save_path + '.index'):
    print("--------------load model--------------")
    model.load_weights(checkpoint_save_path)
# 保存模型
cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_save_path,
                                                 save_weights_only=True,
                                                 save_best_only=True)

# 10.训练
history = model.fit(image_gen_train.flow(x_train, y_train, batch_size=32),
                    epochs=15, validation_data=(x_test, y_test), validation_freq=1,
                    callbacks=[cp_callback])
## 提取acc和loss
acc = history.history['sparse_categorical_accuracy']
val_acc = history.history['val_sparse_categorical_accuracy']
loss = history.history['loss']
val_loss = history.history['val_loss']

# 11.打印和保存网络参数
model.summary()
file = open('./weights.txt', 'w')
for v in model.trainable_variables:
	file.write(str(v.name) + '\n')
	file.write(str(v.shape) + '\n')
	file.write(str(v.numpy()) + '\n')
file.close()

# 12.绘制acc和loss曲线
plt.subplot(1, 2, 1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.title('Training and Validation Accuracy')
plt.legend()

plt.subplot(1, 2, 2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.title('Training and Validation Loss')
plt.legend()
plt.show()

训练的正确率在72%左右，loss和acc图如下所示：

3.输入图片测试

import tensorflow as tf
from tensorflow.keras.layers import Dense,Activation,Dropout,Conv2D,BatchNormalization,MaxPool2D,Flatten
import cv2

# 设置GPU显存按需申请
gpu = tf.config.experimental.list_physical_devices(device_type='GPU')
assert len(gpu) == 1
tf.config.experimental.set_memory_growth(gpu[0], True)

# 1.复现模型（前向传播）
model = tf.keras.models.Sequential([
    Conv2D(filters=32, kernel_size=(3, 3)),
    BatchNormalization(),
    Activation('relu'),
    MaxPool2D(pool_size=(2, 2), strides=2),

    Conv2D(filters=64, kernel_size=(3, 3)),
    BatchNormalization(),
    Activation('relu'),
    MaxPool2D(pool_size=(2, 2), strides=2),

    Conv2D(filters=128, kernel_size=(3, 3)),
    BatchNormalization(),
    Activation('relu'),
    MaxPool2D(pool_size=(2, 2), strides=2),

    Flatten(),                      # 把输入特征拉直为一维数组数值
    Dense(128, activation='relu'),
    Dense(64, activation='relu'),
    Dense(2, activation='softmax')
])

# 2.加载参数
model_save_path = "./checkpoint/cat_dag.ckpt"       # 模型参数存储的路径
model.load_weights(model_save_path)

# 3.数据预处理
img = cv2.imread("./train/dog.228.jpg")
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)      # 转为RGB
img = cv2.resize(img, (224, 224))
img = img / 255.0       # 归一化
img = img[tf.newaxis, ...]

# 4.预测结果
result = model.predict(img)
pred = tf.argmax(result, axis=1)
if pred.numpy() == 0:
    print('识别结果是：小猫')
else:
    print('识别结果是：小狗')