【CV11】如何从头开发于CIFAR-10图像分类的CNN

---

---

1. 模型评估测试

1.1 数据集介绍

CIFAR-10数据集包含10类物体，60,000张32×32像素的彩色照片。这些类别标签及其整数值如下：

• 0：airplane
• 1：automobile
• 2：bird
• 3：cat
• 4：deer
• 5：dog
• 6：frog
• 7：horse
• 8：ship
• 9：truck

这些都是非常小的图像，比典型的照片要小得多，并且该数据集旨在用于计算机视觉研究。

1.2 加载数据集

import sys
import matplotlib.pyplot as plt
from keras.datasets import cifar10
from keras.utils import to_categorical
from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Dense, Flatten, Dropout
from keras.optimizers import SGD
from keras.regularizers import l2

def load_dataset():
    # load dataset
    (trainX, trainY), (testX, testY) = cifar10.load_data()
    
    # one hot encode
    trainY = to_categorical(trainY)
    testY = to_categorical(testY)
    
    return trainX, trainY, testX, testY

1.3 缩放数据

数据集中每个图像的像素值都是无符号整数，介于0和255之间。训练模型的一个很好的起点是标准化像素值，例如将它们重新缩放到[0,1]范围。这涉及首先将数据类型从无符号整数转换为浮点数，然后将像素值除以最大值。

def prep_pixels(train, test):
    # convert from integers to floats
    train_norm = train.astype('float32')
    test_norm = test.astype('float32')
    
    # normalize to range 0-1
    train_norm = train_norm / 255.0
    test_norm = test_norm / 255.0
    
    return train_norm, test_norm

1.4 绘制曲线

plt.rcParams['figure.dpi'] = 200
def summarize_diagnostics(history):
    # plot loss
    plt.subplot(211)
    plt.title('Cross Entropy Loss')
    plt.plot(history.history['loss'], color='blue', label='train')
    plt.plot(history.history['val_loss'], color='orange', label='test')
    # plot accuracy
    plt.subplot(212)
    plt.title('Classification Accuracy')
    plt.plot(history.history['accuracy'], color='blue', label='train')
    plt.plot(history.history['val_accuracy'], color='orange', label='test')
    # save plot to file
    filename = sys.argv[0].split('/')[-1]
    plt.savefig(filename + '_plot.png')
    plt.close()
    plt.tight_layout()
    plt.grid(linestyle='-', alpha=0.5)

1.5 评估模型

def run_test_harness():
    # load dataset
    trainX, trainY, testX, testY = load_dataset()
    # prepare pixel data
    trainX, testX = prep_pixels(trainX, testX)
    # define model
    model = define_model()
    # fit model
    history = model.fit(trainX, trainY, epochs=100, batch_size=64, validation_data=(testX, testY), verbose=2)
    # evaluate model
    _, acc = model.evaluate(testX, testY, verbose=0)
    print('> %.3f' % (acc * 100.0))
    # learning curves
    summarize_diagnostics(history)

1.6 定义模型

1.6.1 1 VGG Block

def define_model():
    model = Sequential()
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(32, 32, 3)))
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Flatten())
    model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
    model.add(Dense(10, activation='softmax'))
    
    # compile model
    opt = SGD(lr=0.001, momentum=0.9)
    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
    
    return model

训练：

run_test_harness()

分类精度：

67.070

1.6.2 2 VGG Block

# define cnn model
def define_model():
    model = Sequential()
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(32, 32, 3)))
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Flatten())
    model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
    model.add(Dense(10, activation='softmax'))
    
    # compile model
    opt = SGD(lr=0.001, momentum=0.9)
    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
    
    return model

训练

run_test_harness()

打印精度：

71.080

1.6.3 3 VGG Block

# define cnn model
def define_model():
    model = Sequential()
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(32, 32, 3)))
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Flatten())
    model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
    model.add(Dense(10, activation='softmax'))
    
    # compile model
    opt = SGD(lr=0.001, momentum=0.9)
    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
    
    return model

训练：

run_test_harness()

打印精度：

73.500

---

2. 改进模型 Trick

2.1 正则化

2.1.1 Dropout

def define_model():
    model = Sequential()
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(32, 32, 3)))
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.2)) # 丢弃率为0.2
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.2))
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.2))
    model.add(Flatten())
    model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
    model.add(Dropout(0.2))
    model.add(Dense(10, activation='softmax'))
    
    # compile model
    opt = SGD(lr=0.001, momentum=0.9)
    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
    
    return model

训练：

run_test_harness()

打印精度：

 83.450

2.1.2 L2正则化（权重衰减正则化）

def define_model():
    model = Sequential()
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', kernel_regularizer=l2(0.001), input_shape=(32, 32, 3)))
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', kernel_regularizer=l2(0.001)))
    model.add(MaxPooling2D((2, 2)))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', kernel_regularizer=l2(0.001)))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', kernel_regularizer=l2(0.001)))
    model.add(MaxPooling2D((2, 2)))
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', kernel_regularizer=l2(0.001)))
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', kernel_regularizer=l2(0.001)))
    model.add(MaxPooling2D((2, 2)))
    model.add(Flatten())
    model.add(Dense(128, activation='relu', kernel_initializer='he_uniform', kernel_regularizer=l2(0.001)))
    model.add(Dense(10, activation='softmax'))
    
    # compile model
    opt = SGD(lr=0.001, momentum=0.9)
    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
    
    return model

训练：

run_test_harness()

打印精度：

72.550

---

2.2 数据增强

数据增强包括在训练数据集中对样本进行少量随机修改的副本。这具有正则化效果，因为它既扩展了训练数据集，又采用了更为通用的方式使模型学习了相同的一般特征。

可能有用的随机增强类型包括水平翻转，图像的微小移动以及图像的小缩放或裁切。下例通过简单的图像增强对基线图像的影响，特别是水平翻转以及图像高度和宽度发生10％的偏移。这可以在Keras中使用 ImageDataGenerator 类实现；例如：

# create data generator
datagen = ImageDataGenerator(width_shift_range=0.1, height_shift_range=0.1, horizontal_flip=True)
# prepare iterator
it_train = datagen.flow(trainX, trainY, batch_size=64)

可以在训练期间使用此方法，方法是将迭代器传递给 model.fit_generator() 函数，并在单个epoch中定义批次数。

# fit model
steps = int(trainX.shape[0] / 64)
history = model.fit_generator(it_train, steps_per_epoch=steps, epochs=100, validation_data=(testX, testY), verbose=0)

重写原来的测试函数：

def run_test_harness():
    # load dataset
    trainX, trainY, testX, testY = load_dataset()
    # prepare pixel data
    trainX, testX = prep_pixels(trainX, testX)
    # define model
    model = define_model()
    # create data generator
    datagen = ImageDataGenerator(width_shift_range=0.1, height_shift_range=0.1, horizontal_flip=True)
    # prepare iterator
    it_train = datagen.flow(trainX, trainY, batch_size=64)
    # fit model
    steps = int(trainX.shape[0] / 64)
    history = model.fit_generator(it_train, steps_per_epoch=steps, epochs=100, validation_data=(testX, testY), verbose=0)
    # evaluate model
    _, acc = model.evaluate(testX, testY, verbose=0)
    print('> %.3f' % (acc * 100.0))
    # learning curves
    summarize_diagnostics(history)

训练：

run_test_harness()

打印精度：

84.470

2.3 更改Dropout Rate

一种有趣的变化是将Dropout Rate从20％增加到25％或30％。另一个有趣的变化是使用一种模式，即从模型的第一个块的20％，第二个块的30％，然后在模型的分类器部分的完全连接层将Dropout Rate增加到50％。这种随着模型深度的增加而增加的缺失是一种常见的模式。这是有效的，因为它会迫使模型中较深的图层进行正则化，而不是更靠近输入的图层。

# define cnn model
def define_model():
    model = Sequential()
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(32, 32, 3)))
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.2))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.3))
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.4))
    model.add(Flatten())
    model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
    model.add(Dropout(0.5))
    model.add(Dense(10, activation='softmax'))
    # compile model
    opt = SGD(lr=0.001, momentum=0.9)
    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
    return model

训练：

run_test_harness()

打印精度：

84.690

2.4 Dropout + 数据增强

训练：

run_test_harness()

打印精度：

85.880

2.5 Dropout + 数据增强 + BN

# define cnn model
def define_model():
    model = Sequential()
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(32, 32, 3)))
    model.add(BatchNormalization())
    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.2))
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(BatchNormalization())
    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.3))
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(BatchNormalization())
    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.4))
    model.add(Flatten())
    model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
    model.add(BatchNormalization())
    model.add(Dropout(0.5))
    model.add(Dense(10, activation='softmax'))
    # compile model
    opt = SGD(lr=0.001, momentum=0.9)
    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
    return model

训练：

run_test_harness()

打印精度：

88.620

Summary

• VGG 1: 67.070%
• VGG 2: 71.080%
• VGG 3: 73.500% （Baseline）
• Baseline + Dropout: 83.450%
• Baseline + Weight Decay: 72.550%
• Baseline + Data Augmentation: 84.470%
• Baseline + Increasing Dropout: 84.690%
• Baseline + Dropout + Data Augmentation: 85.880%
• Baseline + Increasing Dropout + Data Augmentation + Batch Normalization: 88.620%

到目前为止，有些超参数还没有调整，例如学习率，它可能是最重要的超参数。可能期望通过对学习速率进行自适应更改来进一步改进，如使用自适应学习速率技术Adam。这些类型的更改可能有助于收敛后的模型。

---

参考：
https://machinelearningmastery.com/how-to-develop-a-cnn-from-scratch-for-cifar-10-photo-classification/