1 from keras.datasets import mnist 2 from keras.utils import to_categorical 3 4 #1. 获取数据 5 (train_images, train_labels), (test_images, test_labels) = mnist.load_data() 6 7 #2. 处理数据 8 train_images = train_images.reshape((60000, 28, 28, 1)) 9 train_images = train_images.astype('float32') / 255 10 11 test_images = test_images.reshape((10000, 28, 28, 1)) 12 test_images = test_images.astype('float32') / 255 13 14 train_labels = to_categorical(train_labels) 15 test_labels = to_categorical(test_labels) 16 17 from keras import layers 18 from keras import models 19 20 #3. 建立网络模型 21 model = models.Sequential() 22 model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1))) 23 model.add(layers.MaxPooling2D((2, 2))) 24 model.add(layers.Conv2D(64, (3, 3), activation='relu')) 25 model.add(layers.MaxPooling2D((2, 2))) 26 model.add(layers.Conv2D(64, (3, 3), activation='relu')) 27 28 model.add(layers.Flatten()) 29 model.add(layers.Dense(64, activation='relu')) 30 model.add(layers.Dense(10, activation='softmax')) 31 32 # print(model.summary()) 33 34 ''' 35 #网络参数计算方式 36 1.conv2d有320个参数，计算方式：(3 * 3 * 1 + 1) * 32 37 2.conv2d_1有18496个参数，计算方式：(3 * 3 * 32 + 1) * 64 38 3.conv2d_2有36928个参数，计算方式：(3 * 3 * 64 + 1) * 64 39 4.dense有36928个参数，计算方式：(576 + 1) * 64 40 5.dense_1有650个参数，计算方式：(64 + 1) * 10 41 ''' 42 #4. 设置编译参数 43 model.compile(optimizer='rmsprop', 44 loss='categorical_crossentropy', 45 metrics=['accuracy']) 46 47 #5. 设置训练条件并训练 48 model.fit(train_images, train_labels, epochs=5, batch_size=64) 49 50 #6. 评估模型 51 test_loss, test_acc = model.evaluate(test_images, test_labels) 52 #313/313 [==============================] - 1s 3ms/step - loss: 0.0324 - accuracy: 0.9898 53 54 print(test_acc) 55 #0.989799976348877
5.1 卷积神经网络简介
5.1.1 卷积运算
卷积神经网络有以下两个性质：
(1)卷积神经网络学到的模式具有平移不变性。
(2)卷积神经网络可以学到模式的空间层次结构i。
卷积的两个关键参数：
(1)从输入中提取的图块尺寸
(2)输出特征的深度
Keras的API为：Conv2D(output_depth, window_height, window_width)
注意：书上这部分对卷积过程的描述不是很容易理解，建议看吴恩达的视频来学习卷积网络相关的概念。
5.1.2 最大池化运算
最大池化是从输入特征图中提取窗口，并输出每个通道的最大值。
最大池化通常使用2 * 2的窗口和步幅2，其目的是将特征图下采样2倍。
5.2 在小型数据集上从头开始训练一个卷积神经网络
卷积神经网络代码如下：
1 import os, shutil 2 3 # The path to the directory where the original 4 # dataset was uncompressed 5 # original_dataset_dir = '/Users/fchollet/Downloads/kaggle_original_data' 6 original_dataset_dir = 'E:\desktop\code\data\dogs-vs-cats\train' 7 8 # The directory where we will 9 # store our smaller dataset 10 # base_dir = '/Users/fchollet/Downloads/cats_and_dogs_small' 11 base_dir = 'E:\desktop\code\data\dogs-vs-cats\cats_and_dogs_small' 12 # os.mkdir(base_dir) 13 14 # Directories for our training, 15 # validation and test splits 16 train_dir = os.path.join(base_dir, 'train') 17 # os.mkdir(train_dir) 18 validation_dir = os.path.join(base_dir, 'validation') 19 # os.mkdir(validation_dir) 20 test_dir = os.path.join(base_dir, 'test') 21 # os.mkdir(test_dir) 22 23 # Directory with our training cat pictures 24 train_cats_dir = os.path.join(train_dir, 'cats') 25 # os.mkdir(train_cats_dir) 26 27 # Directory with our training dog pictures 28 train_dogs_dir = os.path.join(train_dir, 'dogs') 29 # os.mkdir(train_dogs_dir) 30 31 # Directory with our validation cat pictures 32 validation_cats_dir = os.path.join(validation_dir, 'cats') 33 # os.mkdir(validation_cats_dir) 34 35 # Directory with our validation dog pictures 36 validation_dogs_dir = os.path.join(validation_dir, 'dogs') 37 # os.mkdir(validation_dogs_dir) 38 39 # Directory with our validation cat pictures 40 test_cats_dir = os.path.join(test_dir, 'cats') 41 # os.mkdir(test_cats_dir) 42 43 # Directory with our validation dog pictures 44 test_dogs_dir = os.path.join(test_dir, 'dogs') 45 # os.mkdir(test_dogs_dir) 46 47 48 # #验证图片存放是否正确 49 # print('total training cat images:', len(os.listdir(train_cats_dir))) 50 51 # print('total training dog images:', len(os.listdir(train_dogs_dir))) 52 53 # print('total validation cat images:', len(os.listdir(validation_cats_dir))) 54 55 # print('total validation dog images:', len(os.listdir(validation_dogs_dir))) 56 57 # print('total test cat images:', len(os.listdir(test_cats_dir))) 58 59 # print('total test dog images:', len(os.listdir(test_dogs_dir))) 60 61 62 from keras import layers 63 from keras import models 64 65 model = models.Sequential() 66 model.add(layers.Conv2D(32, (3, 3), activation='relu', 67 input_shape=(150, 150, 3))) 68 model.add(layers.MaxPooling2D((2, 2))) 69 model.add(layers.Conv2D(64, (3, 3), activation='relu')) 70 model.add(layers.MaxPooling2D((2, 2))) 71 model.add(layers.Conv2D(128, (3, 3), activation='relu')) 72 model.add(layers.MaxPooling2D((2, 2))) 73 model.add(layers.Conv2D(128, (3, 3), activation='relu')) 74 model.add(layers.MaxPooling2D((2, 2))) 75 model.add(layers.Flatten()) 76 model.add(layers.Dense(512, activation='relu')) 77 model.add(layers.Dense(1, activation='sigmoid')) 78 79 from keras import optimizers 80 81 model.compile(loss='binary_crossentropy', 82 optimizer=optimizers.RMSprop(lr=1e-4), 83 metrics=['acc']) 84 85 86 from keras.preprocessing.image import ImageDataGenerator 87 88 # All images will be rescaled by 1./255 89 train_datagen = ImageDataGenerator(rescale=1./255) 90 test_datagen = ImageDataGenerator(rescale=1./255) 91 92 train_generator = train_datagen.flow_from_directory( 93 # This is the target directory 94 train_dir, 95 # All images will be resized to 150x150 96 target_size=(150, 150), 97 batch_size=20, 98 # Since we use binary_crossentropy loss, we need binary labels 99 class_mode='binary') 100 101 validation_generator = test_datagen.flow_from_directory( 102 validation_dir, 103 target_size=(150, 150), 104 batch_size=20, 105 class_mode='binary') 106 107 history = model.fit_generator( 108 train_generator, 109 steps_per_epoch=100, 110 epochs=30, 111 validation_data=validation_generator, 112 validation_steps=50) 113 114 model.save('cats_and_dogs_small_1.h5') 115 # model.load_weights('cats_and_dogs_small_1.h5') 116 117 118 119 import matplotlib.pyplot as plt 120 121 # #dict_keys(['loss', 'accuracy', 'val_loss', 'val_accuracy']) 122 acc = history.history['acc'] 123 val_acc = history.history['val_acc'] 124 loss = history.history['loss'] 125 val_loss = history.history['val_loss'] 126 127 epochs = range(len(acc)) 128 129 plt.plot(epochs, acc, 'bo', label='Training acc') 130 plt.plot(epochs, val_acc, 'b', label='Validation acc') 131 plt.title('Training and validation accuracy') 132 plt.legend() 133 134 plt.figure() 135 136 plt.plot(epochs, loss, 'bo', label='Training loss') 137 plt.plot(epochs, val_loss, 'b', label='Validation loss') 138 plt.title('Training and validation loss') 139 plt.legend() 140 141 plt.show()