Keras深度学习实战——基于VGG19模型实现性别分类

0. 前言

在《迁移学习》中，我们了解了利用迁移学习，只需要少量样本即可训练得到性能较好的模型；并基于迁移学习利用预训练的 VGG16 模型进行了性别分类的实战，进一步加深对迁移学习工作原理的理解。

1. VGG19 架构简介

本文，我们将介绍另一种常用的网络模型架构——VGG19，并使用预训练的 VGG19 模型进行性别分类实战。VGG19是 VGG16 的改进版本，具有更多的卷积和池化操作，VGG19 模型的体系结构如下：

Model: "vgg19"_________________________________________________________________Layer (type) Output ShapeParam # =================================================================input_1 (InputLayer) [(None, 256, 256, 3)] 0 _________________________________________________________________block1_conv1 (Conv2D)(None, 256, 256, 64)1792_________________________________________________________________block1_conv2 (Conv2D)(None, 256, 256, 64)36928 _________________________________________________________________block1_pool (MaxPooling2D) (None, 128, 128, 64)0 _________________________________________________________________block2_conv1 (Conv2D)(None, 128, 128, 128) 73856 _________________________________________________________________block2_conv2 (Conv2D)(None, 128, 128, 128) 147584_________________________________________________________________block2_pool (MaxPooling2D) (None, 64, 64, 128) 0 _________________________________________________________________block3_conv1 (Conv2D)(None, 64, 64, 256) 295168_________________________________________________________________block3_conv2 (Conv2D)(None, 64, 64, 256) 590080_________________________________________________________________block3_conv3 (Conv2D)(None, 64, 64, 256) 590080_________________________________________________________________block3_conv4 (Conv2D)(None, 64, 64, 256) 590080_________________________________________________________________block3_pool (MaxPooling2D) (None, 32, 32, 256) 0 _________________________________________________________________block4_conv1 (Conv2D)(None, 32, 32, 512) 1180160 _________________________________________________________________block4_conv2 (Conv2D)(None, 32, 32, 512) 2359808 _________________________________________________________________block4_conv3 (Conv2D)(None, 32, 32, 512) 2359808 _________________________________________________________________block4_conv4 (Conv2D)(None, 32, 32, 512) 2359808 _________________________________________________________________block4_pool (MaxPooling2D) (None, 16, 16, 512) 0 _________________________________________________________________block5_conv1 (Conv2D)(None, 16, 16, 512) 2359808 _________________________________________________________________block5_conv2 (Conv2D)(None, 16, 16, 512) 2359808 _________________________________________________________________block5_conv3 (Conv2D)(None, 16, 16, 512) 2359808 _________________________________________________________________block5_conv4 (Conv2D)(None, 16, 16, 512) 2359808 _________________________________________________________________block5_pool (MaxPooling2D) (None, 8, 8, 512) 0 =================================================================Total params: 20,024,384Trainable params: 20,024,384Non-trainable params: 0_________________________________________________________________

可以看到，上示的体系结构中具有更多的网络层以及更多的参数量。需要注意的是，VGG16 和 VGG19 体系结构中的 16 和 19 代表这些网络中的网络层数。

将每个图像通过 VGG19 网络后，提取到 8 x 8 x 512 输出后，该输出将成为微调模型的输入。接下来，创建输入和输出数据集，然后构建、编译和拟合模型的过程与使用基于预训练的 VGG16 模型进行性别分类的过程相同。

2. 使用预训练 VGG19 模型进行性别分类

在本节中，我们基于迁移学习使用预训练的 VGG19 模型进行性别分类。

2.1 构建输入与输出数据

首先，准备输入和输出数据，我们重用在《卷积神经网络进行性别分类》中使用的数据集以及数据加载代码：

from keras.applications import VGG19from keras.applications.vgg19 import preprocess_inputfrom glob import globfrom skimage import ioimport cv2import numpy as npmodel = VGG19(include_top=False, weights='imagenet', input_shape=(256, 256, 3))x = []y = []for i in glob('man_woman/a_resized/*.jpg')[:800]:try:image = io.imread(i)x.append(image)y.append(0)except:continuefor i in glob('man_woman/b_resized/*.jpg')[:800]:try:image = io.imread(i)x.append(image)y.append(1)except:continuex_vgg19 = []for i in range(len(x)):img = x[i]img = preprocess_input(img.reshape((1, 256, 256, 3)))img_feature = model.predict(img)x_vgg19.append(img_feature)

将输入和输出转换为其相应的数组，并创建训练和测试数据集：

x_vgg19 = np.array(x_vgg19)x_vgg19 = x_vgg19.reshape(x_vgg19.shape[0], x_vgg19.shape[2], x_vgg19.shape[3], x_vgg19.shape[4])y = np.array(y)from sklearn.model_selection import train_test_splitx_train, x_test, y_train, y_test = train_test_split(x_vgg19, y, test_size=0.2)

2.2 模型构建与训练

构建微调模型：

from keras.models import Sequentialfrom keras.layers import Conv2D, MaxPooling2D, Flatten, Dropout, Densemodel_fine_tuning = Sequential()model_fine_tuning.add(Conv2D(512, kernel_size=(3, 3),activation='relu',input_shape=(x_train.shape[1], x_train.shape[2], x_train.shape[3])))model_fine_tuning.add(MaxPooling2D(pool_size=(2, 2)))model_fine_tuning.add(Flatten())model_fine_tuning.add(Dense(1024, activation='relu'))model_fine_tuning.add(Dropout(0.6))model_fine_tuning.add(Dense(1, activation='sigmoid'))model_fine_tuning.summary()

该模型架构的简要信息输入如下：

Model: "sequential"_________________________________________________________________Layer (type) Output ShapeParam # =================================================================conv2d (Conv2D)(None, 6, 6, 512) 2359808 _________________________________________________________________max_pooling2d (MaxPooling2D) (None, 3, 3, 512) 0 _________________________________________________________________flatten (Flatten)(None, 4608)0 _________________________________________________________________dense (Dense)(None, 1024)4719616 _________________________________________________________________dropout (Dropout)(None, 1024)0 _________________________________________________________________dense_1 (Dense)(None, 1) 1025=================================================================Total params: 7,080,449Trainable params: 7,080,449Non-trainable params: 0_________________________________________________________________

接下来，编译并拟合模型：

model_fine_tuning.compile(loss='binary_crossentropy',optimizer='adam',metrics=['acc'])history = model_fine_tuning.fit(x_train, y_train,batch_size=32,epochs=20,verbose=1,validation_data = (x_test, y_test))

最后，我们绘制在训练期间，模型在训练和测试数据集的损失和准确率的变化。可以看到，当我们使用 VGG19 架构时，能够在测试数据集上达到约 95％ 的准确率，结果与使用 VGG16 架构时的性能相似：

2.3 模型错误分类示例

一些错误分类的图像示例如下：

x = np.array(x)from sklearn.model_selection import train_test_splitx_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.2)x_test_vgg19 = []for i in range(len(x_test)):img = x_test[i]img = preprocess_input(img.reshape((1, 256, 256, 3)))img_feature = model.predict(img)x_test_vgg19.append(img_feature)x_test_vgg19 = np.array(x_test_vgg19)x_test_vgg19 = x_test_vgg19.reshape(x_test_vgg19.shape[0], x_test_vgg19.shape[2], x_test_vgg19.shape[3], x_test_vgg19.shape[4])y_pred = model_fine_tuning.predict(x_test_vgg19)wrong = np.argsort(np.abs(y_pred.flatten()-y_test))print(wrong)y_test_char = np.where(y_test==0,'M','F')y_pred_char = np.where(y_pred>0.5,'F','M')plt.subplot(221)plt.imshow(x_test[wrong[-1]])plt.title('Actual: '+str(y_test_char[wrong[-1]])+', '+'Predicted: '+str((y_pred_char[wrong[-1]][0])))plt.subplot(222)plt.imshow(x_test[wrong[-2]])plt.title('Actual: '+str(y_test_char[wrong[-2]])+', '+'Predicted: '+str((y_pred_char[wrong[-2]][0])))plt.subplot(223)plt.imshow(x_test[wrong[-3]])plt.title('Actual: '+str(y_test_char[wrong[-3]])+', '+'Predicted: '+str((y_pred_char[wrong[-3]][0])))plt.subplot(224)plt.imshow(x_test[wrong[-4]])plt.title('Actual: '+str(y_test_char[wrong[-4]])+', '+'Predicted: '+str((y_pred_char[wrong[-4]][0])))plt.show()

从图中，可以看出，VGG19 类似于 VGG16 除了由于人物在图像中占据的空间较小造成错误分类外，倾向于根据头发来判断人物究竟是男性还是女性。

相关链接

Keras深度学习实战（7）——卷积神经网络详解与实现
Keras深度学习实战（9）——卷积神经网络的局限性
Keras深度学习实战（10）——迁移学习
Keras深度学习实战——使用卷积神经网络实现性别分类