1.对于残差神经网络的shortcut结构的理解有两种，看大家的代码全都是先使用tf.nn.conv2d()函数将输入与1*1的卷积核进行卷积，生成通道数与输出通道数相同的特征图，将这张特征图与经过三个卷积层特征图进行点对点相加，这样就将初级特征前馈到了后面，很好的保留了初级特征。
代码如下：
# -*- coding: utf-8 -*-
# 这里使用结合SEnet、inception结构和shortcut结构的卷积神经网络，所以将784reshape成28*28的矩阵
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
def get_data():
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
return mnist
# 设置权重函数
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev = 0.1)
return tf.Variable(initial)
# 设置阈值函数
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
# 设置卷积层
def conv2d(x,W):
return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding = "SAME")
# 设置池化层
def pool(x):
return tf.nn.max_pool(x,ksize=[1,2,2,1],strides = [1,2,2,1],padding = "SAME")
def res_block(x,kernal_size_x,kernal_size_y,channel_in,channel_out):
X_shortcut = x
with tf.variable_scope("res_sub1"):
res_w_conv1 = weight_variable([1,1,channel_in,channel_out/4])
res_b_conv1 = bias_variable([channel_out/4])
res_h_conv1 = tf.nn.relu(conv2d(x,res_w_conv1) + res_b_conv1)
with tf.variable_scope("res_sub2"):
res_w_conv2 = weight_variable([kernal_size_x,kernal_size_y,channel_out/4,channel_out/2])
res_b_conv2 = bias_variable([channel_out/2])
res_h_conv2 = tf.nn.relu(conv2d(res_h_conv1,res_w_conv2) + res_b_conv2)
with tf.variable_scope("res_sub3"):
res_w_conv3 = weight_variable([1,1,channel_out/2,channel_out])
res_b_conv3 = bias_variable([channel_out])
res_h_conv3 = tf.nn.relu(conv2d(res_h_conv2,res_w_conv3) + res_b_conv3)
with tf.variable_scope("shortcut"):
res_w_shortcut = weight_variable([1, 1, channel_in, channel_out])
res_b_shortcut = bias_variable([channel_out])
X_shortcut = conv2d(X_shortcut,res_w_shortcut) + res_b_shortcut
res_add = tf.add(res_h_conv3, X_shortcut)
res_b_shortcut = bias_variable([channel_out])
res_add_result = tf.nn.relu(res_add + res_b_shortcut)
return res_add_result
if __name__ == "__main__":
# 设置输入
x = tf.placeholder(tf.float32,[None, 784])
# 将其变为二维的图片,-1为x的第一维数量
x_image = tf.reshape(x,[-1,28,28,1])
#配置第一个残差块
h_res1 = res_block(x_image,3,3,1,32)
h_res1 = pool(h_res1)
h_res1 = tf.nn.lrn(h_res1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
# 配置第二个残差块
#h_res2 = res_block(h_res1, 3, 3, 32, 64)
#h_res2 = tf.nn.lrn(h_res2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
#配置第三个残差块
h_res2 = res_block(h_res1, 3, 3, 32, 64)
h_res2 = pool(h_res2)
h_res2 = tf.nn.lrn(h_res2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
# 配置全连接层1
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])
h2_pool_flat = tf.reshape(h_res2, shape=[-1, 7 * 7 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h2_pool_flat, W_fc1) + b_fc1)
# 配置dropout层
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# 配置全连接层2
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_predict = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
# 接受y_label
y_label = tf.placeholder(tf.float32,[None,10])
mnist = get_data()
# 交叉熵
cross_entropy = -tf.reduce_sum(y_label * tf.log(y_predict))
# 梯度下降法
train_step = tf.train.AdamOptimizer(1e-3).minimize(cross_entropy)
# train_step = tf.train.GradientDescentOptimizer(1e-3).minimize(cross_entropy)
# 求准确率
correct_prediction = tf.equal(tf.argmax(y_predict, 1), tf.argmax(y_label, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) # 精确度计算
sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())
for i in range(20000):
batch = mnist.train.next_batch(50)
if i % 100 == 0: # 训练100次，验证一次
train_acc = accuracy.eval（feed_dict={x: batch[0], y_label: batch[1], keep_prob: 1.0})
print 'step %d, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y %g' % (i, train_acc)
train_step.run(feed_dict={x: batch[0], y_label: batch[1], keep_prob: 0.5})
point = 0
for i in xrange(10):
testSet = mnist.test.next_batch(50)
point += accuracy.eval（feed_dict={ x: testSet[0], y_label: testSet[1], keep_prob: 1.0})
print "test accuracy:"+str(point/10)test accuracy:0.8479999899864197
2.换一种思路，直接将输入与经过三个卷积层的特征图进行通道维度上的拼接，使用concat的结果
代码如下：
# -*- coding: utf-8 -*-
# 这里使用结合SEnet、inception结构和shortcut结构的卷积神经网络，所以将784reshape成28*28的矩阵
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
def get_data():
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
return mnist
# 设置权重函数
def weight_variable(shape):
# 权重正则化
initial = tf.truncated_normal(shape, stddev = 0.1)
return tf.Variable(initial)
# 设置阈值函数
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
# 设置卷积层
def conv2d(x,W):
return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding = "SAME")
# 设置池化层
def pool(x):
return tf.nn.max_pool(x,ksize=[1,2,2,1],strides = [1,2,2,1],padding = "SAME")
def res_block(x,kernal_size_x,kernal_size_y,channel_in,channel_out):
X_shortcut = x
# 第一个卷积层
with tf.variable_scope("res_sub1"):
res_w_conv1 = weight_variable([1,1,channel_in,channel_out/4])
res_b_conv1 = bias_variable([channel_out/4])
res_h_conv1 = tf.nn.relu(conv2d(x,res_w_conv1) + res_b_conv1)
# 第二个卷积层
with tf.variable_scope("res_sub2"):
res_w_conv2 = weight_variable([kernal_size_x,kernal_size_y,channel_out/4,channel_out/2])
res_b_conv2 = bias_variable([channel_out/2])
res_h_conv2 = tf.nn.relu(conv2d(res_h_conv1,res_w_conv2) + res_b_conv2)
# 第三个卷积层
with tf.variable_scope("res_sub3"):
res_w_conv3 = weight_variable([1,1,channel_out/2,channel_out])
res_b_conv3 = bias_variable([channel_out])
res_h_conv3 = tf.nn.relu(conv2d(res_h_conv2,res_w_conv3) + res_b_conv3)
# shortcut结构
with tf.variable_scope("shortcut"):
# res_w_shortcut = weight_variable([1, 1, channel_in, channel_out])
# res_b_shortcut = bias_variable([channel_out])
# X_shortcut = conv2d(X_shortcut,res_w_shortcut) + res_b_shortcut
res_concat = tf.concat(3,[res_h_conv3, X_shortcut])
res_add_result = tf.nn.relu(res_concat)
channel = channel_out + channel_in
return res_add_result,channel
if __name__ == "__main__":
# 设置输入
x = tf.placeholder(tf.float32,[None, 784])
# 将其变为二维的图片,-1为x的第一维数量
x_image = tf.reshape(x,[-1,28,28,1])
#配置第一个残差块
h_res1,channel1 = res_block(x_image,5,5,1,32)
h_res1 = pool(h_res1)
h_res1 = tf.nn.lrn(h_res1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
# 配置第二个残差块
h_res2,channel2 = res_block(h_res1, 3, 3, channel1, 64)
h_res2 = tf.nn.lrn(h_res2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
#配置第三个残差块
h_res3,channel3 = res_block(h_res2, 3, 3, channel2, 128)
h_res3 = pool(h_res3)
h_res3 = tf.nn.lrn(h_res3, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
# 配置全连接层1
W_fc1 = weight_variable([7 * 7 * channel3, 1024])
b_fc1 = bias_variable([1024])
h2_pool_flat = tf.reshape(h_res3, shape=[-1, 7 * 7 * channel3])
h_fc1 = tf.nn.relu(tf.matmul(h2_pool_flat, W_fc1) + b_fc1)
# 配置dropout层
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# 配置全连接层2
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_predict = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
# 接受y_label
y_label = tf.placeholder(tf.float32,[None,10])
mnist = get_data()
# 交叉熵
cross_entropy = -tf.reduce_sum(y_label * tf.log(y_predict))
# 梯度下降法
train_step = tf.train.AdamOptimizer(1e-3).minimize(cross_entropy)
# 求准确率
correct_prediction = tf.equal(tf.argmax(y_predict, 1), tf.argmax(y_label, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) # 精确度计算
sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())
for i in range(20000):
batch = mnist.train.next_batch(50)
if i % 100 == 0: # 训练100次，验证一次
train_acc = accuracy.eval（feed_dict={x: batch[0], y_label: batch[1], keep_prob: 1.0})
print 'step %d, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y %g' % (i, train_acc)
train_step.run(feed_dict={x: batch[0], y_label: batch[1], keep_prob: 0.5})
point = 0
for i in xrange(10):
testSet = mnist.test.next_batch(50)
point += accuracy.eval（feed_dict={ x: testSet[0], y_label: testSet[1], keep_prob: 1.0})
print "test accuracy:"+str(point/10)准确率：
test accuracy:0.9180000007152558
3.将残差块里的卷积层变为2个3*3的卷积核，第一种：
# -*- coding: utf-8 -*-
# 这里使用结合SEnet、inception结构和shortcut结构的卷积神经网络，所以将784reshape成28*28的矩阵
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
def get_data():
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
return mnist
# 设置权重函数
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev = 0.1)
return tf.Variable(initial)
# 设置阈值函数
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
# 设置卷积层
def conv2d(x,W):
return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding = "SAME")
# 设置池化层
def pool(x):
return tf.nn.max_pool(x,ksize=[1,2,2,1],strides = [1,2,2,1],padding = "SAME")
def res_block(x,kernal_size_x,kernal_size_y,channel_in,channel_out):
X_shortcut = x
with tf.variable_scope("res_sub1"):
res_w_conv1 = weight_variable([kernal_size_x,kernal_size_y,channel_in,channel_out/2])
res_b_conv1 = bias_variable([channel_out/2])
res_h_conv1 = tf.nn.relu(conv2d(x,res_w_conv1) + res_b_conv1)
with tf.variable_scope("res_sub2"):
res_w_conv2 = weight_variable([kernal_size_x,kernal_size_y,channel_out/2,channel_out])
res_b_conv2 = bias_variable([channel_out])
res_h_conv2 = tf.nn.relu(conv2d(res_h_conv1,res_w_conv2) + res_b_conv2)
with tf.variable_scope("shortcut"):
res_w_shortcut = weight_variable([1, 1, channel_in, channel_out])
res_b_shortcut = bias_variable([channel_out])
X_shortcut = conv2d(X_shortcut,res_w_shortcut) + res_b_shortcut
res_add = tf.add(res_h_conv2, X_shortcut)
res_b_shortcut = bias_variable([channel_out])
res_add_result = tf.nn.relu(res_add + res_b_shortcut)
return res_add_result
if __name__ == "__main__":
# 设置输入
x = tf.placeholder(tf.float32,[None, 784])
# 将其变为二维的图片,-1为x的第一维数量
x_image = tf.reshape(x,[-1,28,28,1])
#配置第一个残差块
h_res1 = res_block(x_image,3,3,1,32)
h_res1 = pool(h_res1)
h_res1 = tf.nn.lrn(h_res1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
# 配置第二个残差块
#h_res2 = res_block(h_res1, 3, 3, 32, 64)
#h_res2 = tf.nn.lrn(h_res2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
#配置第三个残差块
h_res2 = res_block(h_res1, 3, 3, 32, 64)
h_res2 = pool(h_res2)
h_res2 = tf.nn.lrn(h_res2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
# 配置全连接层1
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])
h2_pool_flat = tf.reshape(h_res2, shape=[-1, 7 * 7 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h2_pool_flat, W_fc1) + b_fc1)
# 配置dropout层
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# 配置全连接层2
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_predict = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
# 接受y_label
y_label = tf.placeholder(tf.float32,[None,10])
mnist = get_data()
# 交叉熵
cross_entropy = -tf.reduce_sum(y_label * tf.log(y_predict))
# 梯度下降法
train_step = tf.train.AdamOptimizer(1e-3).minimize(cross_entropy)
# train_step = tf.train.GradientDescentOptimizer(1e-3).minimize(cross_entropy)
# 求准确率
correct_prediction = tf.equal(tf.argmax(y_predict, 1), tf.argmax(y_label, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) # 精确度计算
sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())
for i in range(20000):
batch = mnist.train.next_batch(50)
if i % 100 == 0: # 训练100次，验证一次
train_acc = accuracy.eval（feed_dict={x: batch[0], y_label: batch[1], keep_prob: 1.0})
print 'step %d, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y %g' % (i, train_acc)
train_step.run(feed_dict={x: batch[0], y_label: batch[1], keep_prob: 0.5})
point = 0
for i in xrange(10):
testSet = mnist.test.next_batch(50)
point += accuracy.eval（feed_dict={ x: testSet[0], y_label: testSet[1], keep_prob: 1.0})
print "test accuracy:"+str(point/10)准确率：
step 19300, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y 0.98
step 19400, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y 0.96
step 19500, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y 1
step 19600, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y 0.94
step 19700, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y 0.9
step 19800, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y 0.92
step 19900, training accuracsess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys,keep_prob: 0.5})y 0.86
test accuracy:0.9439999997615814
Process finished with exit code 0
本文章为转载内容，我们尊重原作者对文章享有的著作权。如有内容错误或侵权问题，欢迎原作者联系我们进行内容更正或删除文章。
