keras构建卷积神经网路

一、卷积神经网络

1、与全连接神经网络相比，卷积神经网络的优势体现在两个方面：（1）减少了参数：在全连接的结构下，相邻网络层的两点之间互联，参数非常多，尤其是面对图像输入这样的数据结构；CNN能够大大降低参数的数量级，加快收敛，使训练复杂度大大下降，也减轻了过拟合，提高了模型的泛化能力。（2）更适用于二维结构的数据格式：全连接网络将整个图像“压平”成一个向量，这种操作忽略了图像的“二维空间特性”，图像在x和y轴方向的构成是有意义的，这些特征（局部特征）是需要被提取的，卷积操作就很好的应对了这种局部特征提取。 2、优点：稀疏连接、参数共享、等变表示（输入发生位移时卷积模型依然可以提取到相同的特征。参数共享赋予了卷积网络对平移的容忍性。）。 3、卷积操作是对应元素相乘再相加。

from keras import Input from keras.models import Sequential # from keras.layers import Dense,Activation,Dropout # #3种卷积层 分别对应一维（序列）、二维（图像）、三维（立体数据） 为了适应任何维度的输入空间 from keras.layers import Conv1D,Conv2D,Conv3D model=Sequential() ''' # #序列数据 指定了长度10 如果没有指定seq_length，那么输入的是一个长度不固定的序列 # seq_length=10 # model.add(Conv1D(64,3,activation='relu',input_shape=(seq_length,128))) # #输入数据为一个三维数组 表示长和宽都为224的彩色图像 #Conv2D 第一个参数64是卷积核filters 必须指定 ，第二个参数(3,3)是卷积核的尺寸 kernel_size=（3，3）或 kernel_size=3 表示3×3的卷积核 必须指定, #stride 步长 默认为1 和kernel_size表达形式一样，表示卷积核沿宽度和高度方向的步长 #padding 填充 只有两个值 padding=“same”表示会有填充的操作 保持输出图像的尺寸和输入图像的尺寸一致，如果设置为padding=“valid”表示不做填充 默认为valid #激活函数 activation 默认为线性激活函数 即将卷积层的结果直接输出 #指定卷积层的输入 ''' input_img=Input(shape=(224,224,3)) conv_layer=Conv2D(64,(1,1))(input_img) model.add(Conv2D(64,(3,3),activation='relu')) ''' #池化层 pool_size滑动窗口尺寸 strides通常设置为1 为步长 #接收上一层输出的特征图作为输入，即对输入取最大值MaxPooling和平均值AveragePooling后输出本层的特征图 ''' from keras.layers.pooling import MaxPool2D from keras.layers.pooling import AveragePooling2D model.add(MaxPool2D(pool_size=(2,2),strides=1)) #全连接层 ''' 在卷积网络的最后使用，当卷积层和池化层完成特征的提取和处理后，全连接层整合特征来进行最后的预测 其接收的输入是一维数组，所以从卷积层、池化层到全连接层还需要加入一层Flatten层来将输入“压平” 把多维的输入一维化 Flatten层只作用于图像特征上，不影响批量数据的大小 ''' from keras.layers import Dense,Flatten model.add(Flatten()) model.add(Dense(10,activation='softmax')) #输出类别为10的分类预测 #检查data_format的设置 from keras import backend as K if K.image_data_format()=='channels_last': bn_axis=3 else: bn_axis=1 print(bn_axis)

二、使用卷积神经网络处理手写体分类问题

import keras from keras.datasets import mnist from keras.models import Sequential from keras.layers import Dense,Dropout,Flatten from keras.layers import Conv2D,MaxPool2D from keras import backend as K #定义参数 batch_size=128 num_classes=10 epochs=12 #输入图像的尺寸 img_rows,img_cols=28,28 (x_train,y_train),(x_test,y_test)=mnist.load_data() #保证程序总是能有正确的data_format if K.image_data_format()=='channels_first': x_train=x_train.reshape(x_train.shape[0], 1, img_rows, img_cols) x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols) input_shape=(1,img_rows,img_cols) else: x_train=x_train.reshape(x_train.shape[0], img_rows, img_cols, 1) x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1) input_shape=(img_rows,img_cols,1) #数据预处理 x_train=x_train.astype('float32') x_test=x_test.astype('float32') x_train/=255 x_test/=255 #将标签值转换为categorical y_train=keras.utils.to_categorical(y_train,num_classes) y_test=keras.utils.to_categorical(y_test,num_classes) #搭建卷积神经网络 model=Sequential() model.add(Conv2D(32,kernel_size=(3,3),activation='relu',input_shape=input_shape))#输入层 model.add(Conv2D(64,(3,3),activation='relu')) model.add(MaxPool2D(pool_size=(2,2))) model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(128,activation='relu')) model.add(Dropout(0.5)) model.add(Dense(num_classes,activation='softmax')) print(model.summary()) # #模型结构图可视化 # from keras.utils import plot_model # plot_model(model,to_file='model.png') #建模训练 model.compile(loss=keras.losses.categorical_crossentropy,optimizer=keras.optimizers.Adadelta(),metrics=['accuracy']) model.fit(x_train,y_train,batch_size=batch_size,epochs=epochs,verbose=1,validation_data=(x_test,y_test)) #评价模型 score=model.evaluate(x_test,y_test,verbose=0) print('test loss:',score[0]) print('test accuracy:',score[1]) Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d (Conv2D) (None, 26, 26, 32) 320 _________________________________________________________________ conv2d_1 (Conv2D) (None, 24, 24, 64) 18496 _________________________________________________________________ max_pooling2d (MaxPooling2D) (None, 12, 12, 64) 0 _________________________________________________________________ dropout (Dropout) (None, 12, 12, 64) 0 _________________________________________________________________ flatten (Flatten) (None, 9216) 0 _________________________________________________________________ dense (Dense) (None, 128) 1179776 _________________________________________________________________ dropout_1 (Dropout) (None, 128) 0 _________________________________________________________________ dense_1 (Dense) (None, 10) 1290 ================================================================= Total params: 1,199,882 Trainable params: 1,199,882 Non-trainable params: 0 _________________________________________________________________ None Epoch 1/12 469/469 [==============================] - 78s 167ms/step - loss: 2.2852 - accuracy: 0.1757 - val_loss: 2.2475 - val_accuracy: 0.3065 Epoch 2/12 469/469 [==============================] - 79s 168ms/step - loss: 2.2207 - accuracy: 0.2906 - val_loss: 2.1707 - val_accuracy: 0.4844 Epoch 3/12 469/469 [==============================] - 80s 170ms/step - loss: 2.1409 - accuracy: 0.3968 - val_loss: 2.0703 - val_accuracy: 0.5835 Epoch 4/12 469/469 [==============================] - 78s 165ms/step - loss: 2.0348 - accuracy: 0.4717 - val_loss: 1.9353 - val_accuracy: 0.6532 Epoch 5/12 469/469 [==============================] - 76s 162ms/step - loss: 1.8967 - accuracy: 0.5258 - val_loss: 1.7613 - val_accuracy: 0.7023 Epoch 6/12 469/469 [==============================] - 76s 162ms/step - loss: 1.7293 - accuracy: 0.5703 - val_loss: 1.5583 - val_accuracy: 0.7390 Epoch 7/12 469/469 [==============================] - 76s 162ms/step - loss: 1.5520 - accuracy: 0.6059 - val_loss: 1.3491 - val_accuracy: 0.7705 Epoch 8/12 469/469 [==============================] - 76s 163ms/step - loss: 1.3812 - accuracy: 0.6368 - val_loss: 1.1568 - val_accuracy: 0.7939 Epoch 9/12 469/469 [==============================] - 79s 169ms/step - loss: 1.2330 - accuracy: 0.6641 - val_loss: 0.9980 - val_accuracy: 0.8077 Epoch 10/12 469/469 [==============================] - 77s 165ms/step - loss: 1.1182 - accuracy: 0.6842 - val_loss: 0.8751 - val_accuracy: 0.8202 Epoch 11/12 469/469 [==============================] - 77s 164ms/step - loss: 1.0237 - accuracy: 0.7052 - val_loss: 0.7818 - val_accuracy: 0.8301 Epoch 12/12 469/469 [==============================] - 78s 166ms/step - loss: 0.9473 - accuracy: 0.7223 - val_loss: 0.7095 - val_accuracy: 0.8389 test loss: 0.709541916847229 test accuracy: 0.8389000296592712