课程学习 CV 北京邮电大学 鲁鹏（笔记四：CV经典网络讲解 之 VGG）

VGG

VGG论文：Very deep convolutional networks for large-scale image recognition
VGGNet由牛津大学的视觉几何组（Visual Geometry Group）提出，主要贡献在于证明了使用3x3的小卷积核，增加网络深度，可以有效提升模型性能，并且对于其他数据集也有很好的泛化性能。

VGG的结构简洁，整个网络都使用同样大小的卷积核尺寸（3x3）和最大池化尺寸（2x2）。到目前为止，VGG仍然被用来提取图像特征。






定义卷积函数

def conv2d(x, W, b, strides=1):# Conv2D wrapper, with bias and relu activationx = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')x = tf.nn.bias_add(x, b)return tf.nn.relu(x)

定义池化函数

def maxpool2d(x, k=2):# MaxPool2D wrapperreturn tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME')

定义VGG结构

def conv_net(x, weights, biases, dropout):# Reshape input picture  x.shape:(128,128,3)x = tf.reshape(x, shape=[-1, 128, 128, 3])# Convolution Layerconv1 = conv2d(x, weights['wc1'], biases['bc1'])conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])# Max Pooling (down-sampling)pool1 = maxpool2d(conv2, k=2)print(pool1.shape)  # (64,64,64)# Convolution Layerconv3 = conv2d(pool1, weights['wc3'], biases['bc3'])conv4 = conv2d(conv3, weights['wc4'], biases['bc4'])# Max Pooling (down-sampling)pool2 = maxpool2d(conv4, k=2)print(pool2.shape)  # (32,32,128)# Convolution Layerconv5 = conv2d(pool2, weights['wc5'], biases['bc5'])conv6 = conv2d(conv5, weights['wc6'], biases['bc6'])conv7 = conv2d(conv6, weights['wc7'], biases['bc7'])# Max Poolingpool3 = maxpool2d(conv7, k=2)print(pool3.shape)  # (16,16,256)# Convolution Layerconv8 = conv2d(pool3, weights['wc8'], biases['bc8'])conv9 = conv2d(conv8, weights['wc9'], biases['bc9'])conv10 = conv2d(conv9, weights['wc10'], biases['bc10'])# Max Poolingpool4 = maxpool2d(conv10, k=2)print(pool4.shape)  # (8,8,512)conv11 = conv2d(pool4, weights['wc11'], biases['bc11'])conv12 = conv2d(conv11, weights['wc12'], biases['bc12'])conv13 = conv2d(conv12, weights['wc13'], biases['bc13'])# Max Poolingpool5 = maxpool2d(conv13, k=2)print(pool5.shape)  # (4,4,512)# Fully connected layer# Reshape conv2 output to fit fully connected layer inputfc1 = tf.reshape(pool5, [-1, weights['wd1'].get_shape().as_list()[0]])fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])fc1 = tf.nn.relu(fc1)# Apply Dropoutfc1 = tf.nn.dropout(fc1, dropout)# fc2 = tf.reshape(fc1, [-1, weights['wd2'].get_shape().as_list()[0]])fc2 = tf.add(tf.matmul(fc1, weights['wd2']), biases['bd2'])fc2 = tf.nn.relu(fc2)# Apply Dropoutfc2 = tf.nn.dropout(fc2, dropout)'''fc3 = tf.reshape(fc2, [-1, weights['out'].get_shape().as_list()[0]])fc3 = tf.add(tf.matmul(fc2, weights['out']), biases['bd2'])fc3 = tf.nn.relu(fc2)'''# Output, class predictionout = tf.add(tf.matmul(fc2, weights['out']), biases['out'])return out

定义权重

weights = {# 3x3 conv, 3 input, 24 outputs'wc1': tf.Variable(tf.random_normal([3, 3, 3, 64])),'wc2': tf.Variable(tf.random_normal([3, 3, 64, 64])),'wc3': tf.Variable(tf.random_normal([3, 3, 64, 128])),'wc4': tf.Variable(tf.random_normal([3, 3, 128, 128])),'wc5': tf.Variable(tf.random_normal([3, 3, 128, 256])),'wc6': tf.Variable(tf.random_normal([3, 3, 256, 256])),'wc7': tf.Variable(tf.random_normal([3, 3, 256, 256])),'wc8': tf.Variable(tf.random_normal([3, 3, 256, 512])),'wc9': tf.Variable(tf.random_normal([3, 3, 512, 512])),'wc10': tf.Variable(tf.random_normal([3, 3, 512, 512])),'wc11': tf.Variable(tf.random_normal([3, 3, 512, 512])),'wc12': tf.Variable(tf.random_normal([3, 3, 512, 512])),'wc13': tf.Variable(tf.random_normal([3, 3, 512, 512])),# fully connected, 32*32*96 inputs, 1024 outputs'wd1': tf.Variable(tf.random_normal([4 * 4 * 512, 1024])),'wd2': tf.Variable(tf.random_normal([1024, 1024])),# 1024 inputs, 10 outputs (class prediction)'out': tf.Variable(tf.random_normal([1024, 10]))}

定义偏置

biases = {'bc1': tf.Variable(tf.random_normal([64])),'bc2': tf.Variable(tf.random_normal([64])),'bc3': tf.Variable(tf.random_normal([128])),'bc4': tf.Variable(tf.random_normal([128])),'bc5': tf.Variable(tf.random_normal([256])),'bc6': tf.Variable(tf.random_normal([256])),'bc7': tf.Variable(tf.random_normal([256])),'bc8': tf.Variable(tf.random_normal([512])),'bc9': tf.Variable(tf.random_normal([512])),'bc10': tf.Variable(tf.random_normal([512])),'bc11': tf.Variable(tf.random_normal([512])),'bc12': tf.Variable(tf.random_normal([512])),'bc13': tf.Variable(tf.random_normal([512])),'bd1': tf.Variable(tf.random_normal([1024])),'bd2': tf.Variable(tf.random_normal([1024])),'out': tf.Variable(tf.random_normal([10]))}

Construct model

pred = conv_net(x, weights, biases, keep_prob)# Define loss and optimizer损失and优化
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)# Evaluate model
correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))# Initializing the variables
init = tf.global_variables_initializer()
saver = tf.train.Saver()

VGG网络的大体结构就定义好了，只要初始化变量，设置Session，定义输入图像就可以跑了