tensorflow实现siamese网络(附代码)

时间 2019-11-08

原文原文链接

【转载自 https://blog.csdn.net/qq1483661204/article/details/79039702】网络

Learning a Similarity Metric Discriminatively, with Application to Face ide

Verification 这个siamese文章连接。函数

本文主要讲解siamese网络，并用tensorflwo实现，在mnist数据集中，siamese网络和其余网络的不一样之处在于，首先他是两个输入，它输入的不是标签，而是是不是同一类别，若是是同一类别就是0，不然就是1，文章中是用这个网络来作人脸识别，网络结构图以下： spa

从图中能够看到，他又两个输入，分别是下x1和x2，左右两个的网咯结构是同样的，而且他们共享权重，最后获得两个输出，分别是Gw(x1)和Gw(x2),这个网络的很好理解，当输入是同一张图片的时候，咱们但愿他们呢之间的欧式距离很小，当不是一张图片时，咱们的欧式距离很大。有了网路结构，接下来就是定义损失函数，这个很重要，而通过咱们的分析，咱们能够知道，损失函数的特色应该是这样的， .net

(1) 当咱们输入同一张图片时，他们之间的欧式距离越小，损失是越小的，距离越大，损失越大 scala

(2) 当咱们的输入是不一样的图片的时候，他们之间的距离越大，损失越大 code

怎么理解呢，很简单，咱们就是最小化把相同类的数据之间距离，最大化不一样类之间的距离。 orm

而后文章中定义的损失函数以下： blog

首先是定义距离，使用l2范数，公式以下：图片

距离其实就是欧式距离，有了距离，咱们的损失函数和距离的关系我上面说了，如何包证知足上面的要求呢，文章提出这样的损失函数：

其中咱们的Ew就是距离，Lg和L1至关因而一个系数，这个损失函数和交叉熵其实挺像，为了让损失函数知足上面的关系，让Lg知足单调递减，LI知足单调递增就能够。另一个条件是：同类图片之间的距离必须比不一样类之间的距离小，

其余条件以下：

而后做者也给出了证实，最终损失函数为：

Q是一个常数，这个损失函数就知足上面的关系，而后我用tensoflow写了一个损失函数以下：

须要强调的是，这个地方同一类图片是0，不一样类图片是1，而后我本身用tensorflow实现的这个损失函数以下：

def siamese_loss(out1,out2,y,Q=5):

    Q = tf.constant(Q, name="Q",dtype=tf.float32)
    E_w = tf.sqrt(tf.reduce_sum(tf.square(out1-out2),1))   
    pos = tf.multiply(tf.multiply(y,2/Q),tf.square(E_w))
    neg = tf.multiply(tf.multiply(1-y,2*Q),tf.exp(-2.77/Q*E_w))                
    loss = pos + neg                 
    loss = tf.reduce_mean(loss)              
    return loss

这就是损失函数，其余的代码以下：

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np
tf.reset_default_graph()
mnist = input_data.read_data_sets('./data/mnist',one_hot=True)
print(mnist.validation.num_examples)
print(mnist.train.num_examples)
print(mnist.test.num_examples)
def siamese_loss(out1,out2,y,Q=5):

    Q = tf.constant(Q, name="Q",dtype=tf.float32)
    E_w = tf.sqrt(tf.reduce_sum(tf.square(out1-out2),1))   
    pos = tf.multiply(tf.multiply(y,2/Q),tf.square(E_w))
    neg = tf.multiply(tf.multiply(1-y,2*Q),tf.exp(-2.77/Q*E_w))                
    loss = pos + neg                 
    loss = tf.reduce_mean(loss)              
    return loss

def siamese(inputs,keep_prob):
        with tf.name_scope('conv1') as scope:
            w1 = tf.Variable(tf.truncated_normal(shape=[3,3,1,32],stddev=0.05),name='w1')
            b1 = tf.Variable(tf.zeros(32),name='b1')
            conv1 = tf.nn.conv2d(inputs,w1,strides=[1,1,1,1],padding='SAME',name='conv1')
        with tf.name_scope('relu1') as scope:
            relu1 = tf.nn.relu(tf.add(conv1,b1),name='relu1')
        with tf.name_scope('conv2') as scope:
            w2 = tf.Variable(tf.truncated_normal(shape=[3,3,32,64],stddev=0.05),name='w2')
            b2 = tf.Variable(tf.zeros(64),name='b2')
            conv2 = tf.nn.conv2d(relu1,w2,strides=[1,2,2,1],padding='SAME',name='conv2')
        with tf.name_scope('relu2') as scope:
            relu2 = tf.nn.relu(conv2+b2,name='relu2')

        with tf.name_scope('conv3') as scope:

            w3 = tf.Variable(tf.truncated_normal(shape=[3,3,64,128],mean=0,stddev=0.05),name='w3')
            b3 = tf.Variable(tf.zeros(128),name='b3')
            conv3 = tf.nn.conv2d(relu2,w3,strides=[1,2,2,1],padding='SAME')
        with tf.name_scope('relu3') as scope:
            relu3 = tf.nn.relu(conv3+b3,name='relu3')

        with tf.name_scope('fc1') as scope:
            x_flat = tf.reshape(relu3,shape=[-1,7*7*128])
            w_fc1=tf.Variable(tf.truncated_normal(shape=[7*7*128,1024],stddev=0.05,mean=0),name='w_fc1')
            b_fc1 = tf.Variable(tf.zeros(1024),name='b_fc1')
            fc1 = tf.add(tf.matmul(x_flat,w_fc1),b_fc1)
        with tf.name_scope('relu_fc1') as scope:
            relu_fc1 = tf.nn.relu(fc1,name='relu_fc1')



        with tf.name_scope('drop_1') as scope:

            drop_1 = tf.nn.dropout(relu_fc1,keep_prob=keep_prob,name='drop_1')
        with tf.name_scope('bn_fc1') as scope:
            bn_fc1 = tf.layers.batch_normalization(drop_1,name='bn_fc1')
        with tf.name_scope('fc2') as scope:
            w_fc2 = tf.Variable(tf.truncated_normal(shape=[1024,512],stddev=0.05,mean=0),name='w_fc2')
            b_fc2 = tf.Variable(tf.zeros(512),name='b_fc2')
            fc2 = tf.add(tf.matmul(bn_fc1,w_fc2),b_fc2)
        with tf.name_scope('relu_fc2') as scope:
            relu_fc2 = tf.nn.relu(fc2,name='relu_fc2')
        with tf.name_scope('drop_2') as scope:
            drop_2 = tf.nn.dropout(relu_fc2,keep_prob=keep_prob,name='drop_2')
        with tf.name_scope('bn_fc2') as scope:
            bn_fc2 = tf.layers.batch_normalization(drop_2,name='bn_fc2')
        with tf.name_scope('fc3') as scope:
            w_fc3 = tf.Variable(tf.truncated_normal(shape=[512,2],stddev=0.05,mean=0),name='w_fc3')
            b_fc3 = tf.Variable(tf.zeros(2),name='b_fc3')
            fc3 = tf.add(tf.matmul(bn_fc2,w_fc3),b_fc3)
        return fc3

lr = 0.01
iterations = 20000
batch_size = 64

with tf.variable_scope('input_x1') as scope:
    x1 = tf.placeholder(tf.float32, shape=[None, 784])
    x_input_1 = tf.reshape(x1, [-1, 28, 28, 1])
with tf.variable_scope('input_x2') as scope:
    x2 = tf.placeholder(tf.float32, shape=[None, 784])
    x_input_2 = tf.reshape(x2, [-1, 28, 28, 1])
with tf.variable_scope('y') as scope:
    y = tf.placeholder(tf.float32, shape=[batch_size])

with tf.name_scope('keep_prob') as scope:
    keep_prob = tf.placeholder(tf.float32)

with tf.variable_scope('siamese') as scope:
    out1 = siamese(x_input_1,keep_prob)
    scope.reuse_variables()
    out2 = siamese(x_input_2,keep_prob)
with tf.variable_scope('metrics') as scope:
    loss = siamese_loss(out1, out2, y)
    optimizer = tf.train.AdamOptimizer(lr).minimize(loss)

loss_summary = tf.summary.scalar('loss',loss)
merged_summary = tf.summary.merge_all()

with tf.Session() as sess:

    writer = tf.summary.FileWriter('./graph/siamese',sess.graph)
    sess.run(tf.global_variables_initializer())

    for itera in range(iterations):
        xs_1, ys_1 = mnist.train.next_batch(batch_size)
        ys_1 = np.argmax(ys_1,axis=1)
        xs_2, ys_2 = mnist.train.next_batch(batch_size)
        ys_2 = np.argmax(ys_2,axis=1)
        y_s = np.array(ys_1==ys_2,dtype=np.float32)
        _,train_loss,summ = sess.run([optimizer,loss,merged_summary],feed_dict={x1:xs_1,x2:xs_2,y:y_s,keep_prob:0.6})

        writer.add_summary(summ,itera)
        if itera % 1000 == 1 :
            print('iter {},train loss {}'.format(itera,train_loss))
    embed = sess.run(out1,feed_dict={x1:mnist.test.images,keep_prob:0.6})
    test_img = mnist.test.images.reshape([-1,28,28,1])
    writer.close()

这里多说一句，siamese能够用来降维，由于最后他的输出是二维的，这样直接把维度降下来了。

Learning a Similarity Metric Discriminatively, with Application to Face Verification 这个siamese文章连接。本文主要讲解siamese网络，并用tensorflwo实现，在mnist数据集中，siamese网络和其余网络的不一样之处在于，首先他是两个输入，它输入的不是标签，而是是不是同一类别，若是是同一类别就是0，不然就是1，文章中是用这个网络来作人脸识别，网络结构图以下：从图中能够看到，他又两个输入，分别是下x1和x2，左右两个的网咯结构是同样的，而且他们共享权重，最后获得两个输出，分别是Gw(x1)和Gw(x2),这个网络的很好理解，当输入是同一张图片的时候，咱们但愿他们呢之间的欧式距离很小，当不是一张图片时，咱们的欧式距离很大。有了网路结构，接下来就是定义损失函数，这个很重要，而通过咱们的分析，咱们能够知道，损失函数的特色应该是这样的， (1) 当咱们输入同一张图片时，他们之间的欧式距离越小，损失是越小的，距离越大，损失越大 (2) 当咱们的输入是不一样的图片的时候，他们之间的距离越大，损失越大怎么理解呢，很简单，咱们就是最小化把相同类的数据之间距离，最大化不一样类之间的距离。而后文章中定义的损失函数以下：首先是定义距离，使用l2范数，公式以下：距离其实就是欧式距离，有了距离，咱们的损失函数和距离的关系我上面说了，如何包证知足上面的要求呢，文章提出这样的损失函数：其中咱们的Ew就是距离，Lg和L1至关因而一个系数，这个损失函数和交叉熵其实挺像，为了让损失函数知足上面的关系，让Lg知足单调递减，LI知足单调递增就能够。另一个条件是：同类图片之间的距离必须比不一样类之间的距离小，其余条件以下：而后做者也给出了证实，最终损失函数为： Q是一个常数，这个损失函数就知足上面的关系，而后我用tensoflow写了一个损失函数以下：须要强调的是，这个地方同一类图片是0，不一样类图片是1，而后我本身用tensorflow实现的这个损失函数以下：def siamese_loss(out1,out2,y,Q=5): Q = tf.constant(Q, name="Q",dtype=tf.float32) E_w = tf.sqrt(tf.reduce_sum(tf.square(out1-out2),1)) pos = tf.multiply(tf.multiply(y,2/Q),tf.square(E_w)) neg = tf.multiply(tf.multiply(1-y,2*Q),tf.exp(-2.77/Q*E_w)) loss = pos + neg loss = tf.reduce_mean(loss) return loss123456789这就是损失函数，其余的代码以下：import tensorflow as tffrom tensorflow.examples.tutorials.mnist import input_dataimport numpy as nptf.reset_default_graph()mnist = input_data.read_data_sets('./data/mnist',one_hot=True)print(mnist.validation.num_examples)print(mnist.train.num_examples)print(mnist.test.num_examples)def siamese_loss(out1,out2,y,Q=5): Q = tf.constant(Q, name="Q",dtype=tf.float32) E_w = tf.sqrt(tf.reduce_sum(tf.square(out1-out2),1)) pos = tf.multiply(tf.multiply(y,2/Q),tf.square(E_w)) neg = tf.multiply(tf.multiply(1-y,2*Q),tf.exp(-2.77/Q*E_w)) loss = pos + neg loss = tf.reduce_mean(loss) return lossdef siamese(inputs,keep_prob): with tf.name_scope('conv1') as scope: w1 = tf.Variable(tf.truncated_normal(shape=[3,3,1,32],stddev=0.05),name='w1') b1 = tf.Variable(tf.zeros(32),name='b1') conv1 = tf.nn.conv2d(inputs,w1,strides=[1,1,1,1],padding='SAME',name='conv1') with tf.name_scope('relu1') as scope: relu1 = tf.nn.relu(tf.add(conv1,b1),name='relu1') with tf.name_scope('conv2') as scope: w2 = tf.Variable(tf.truncated_normal(shape=[3,3,32,64],stddev=0.05),name='w2') b2 = tf.Variable(tf.zeros(64),name='b2') conv2 = tf.nn.conv2d(relu1,w2,strides=[1,2,2,1],padding='SAME',name='conv2') with tf.name_scope('relu2') as scope: relu2 = tf.nn.relu(conv2+b2,name='relu2') with tf.name_scope('conv3') as scope: w3 = tf.Variable(tf.truncated_normal(shape=[3,3,64,128],mean=0,stddev=0.05),name='w3') b3 = tf.Variable(tf.zeros(128),name='b3') conv3 = tf.nn.conv2d(relu2,w3,strides=[1,2,2,1],padding='SAME') with tf.name_scope('relu3') as scope: relu3 = tf.nn.relu(conv3+b3,name='relu3') with tf.name_scope('fc1') as scope: x_flat = tf.reshape(relu3,shape=[-1,7*7*128]) w_fc1=tf.Variable(tf.truncated_normal(shape=[7*7*128,1024],stddev=0.05,mean=0),name='w_fc1') b_fc1 = tf.Variable(tf.zeros(1024),name='b_fc1') fc1 = tf.add(tf.matmul(x_flat,w_fc1),b_fc1) with tf.name_scope('relu_fc1') as scope: relu_fc1 = tf.nn.relu(fc1,name='relu_fc1') with tf.name_scope('drop_1') as scope: drop_1 = tf.nn.dropout(relu_fc1,keep_prob=keep_prob,name='drop_1') with tf.name_scope('bn_fc1') as scope: bn_fc1 = tf.layers.batch_normalization(drop_1,name='bn_fc1') with tf.name_scope('fc2') as scope: w_fc2 = tf.Variable(tf.truncated_normal(shape=[1024,512],stddev=0.05,mean=0),name='w_fc2') b_fc2 = tf.Variable(tf.zeros(512),name='b_fc2') fc2 = tf.add(tf.matmul(bn_fc1,w_fc2),b_fc2) with tf.name_scope('relu_fc2') as scope: relu_fc2 = tf.nn.relu(fc2,name='relu_fc2') with tf.name_scope('drop_2') as scope: drop_2 = tf.nn.dropout(relu_fc2,keep_prob=keep_prob,name='drop_2') with tf.name_scope('bn_fc2') as scope: bn_fc2 = tf.layers.batch_normalization(drop_2,name='bn_fc2') with tf.name_scope('fc3') as scope: w_fc3 = tf.Variable(tf.truncated_normal(shape=[512,2],stddev=0.05,mean=0),name='w_fc3') b_fc3 = tf.Variable(tf.zeros(2),name='b_fc3') fc3 = tf.add(tf.matmul(bn_fc2,w_fc3),b_fc3) return fc3lr = 0.01iterations = 20000batch_size = 64with tf.variable_scope('input_x1') as scope: x1 = tf.placeholder(tf.float32, shape=[None, 784]) x_input_1 = tf.reshape(x1, [-1, 28, 28, 1])with tf.variable_scope('input_x2') as scope: x2 = tf.placeholder(tf.float32, shape=[None, 784]) x_input_2 = tf.reshape(x2, [-1, 28, 28, 1])with tf.variable_scope('y') as scope: y = tf.placeholder(tf.float32, shape=[batch_size])with tf.name_scope('keep_prob') as scope: keep_prob = tf.placeholder(tf.float32)with tf.variable_scope('siamese') as scope: out1 = siamese(x_input_1,keep_prob) scope.reuse_variables() out2 = siamese(x_input_2,keep_prob)with tf.variable_scope('metrics') as scope: loss = siamese_loss(out1, out2, y) optimizer = tf.train.AdamOptimizer(lr).minimize(loss)loss_summary = tf.summary.scalar('loss',loss)merged_summary = tf.summary.merge_all()with tf.Session() as sess: writer = tf.summary.FileWriter('./graph/siamese',sess.graph) sess.run(tf.global_variables_initializer()) for itera in range(iterations): xs_1, ys_1 = mnist.train.next_batch(batch_size) ys_1 = np.argmax(ys_1,axis=1) xs_2, ys_2 = mnist.train.next_batch(batch_size) ys_2 = np.argmax(ys_2,axis=1) y_s = np.array(ys_1==ys_2,dtype=np.float32) _,train_loss,summ = sess.run([optimizer,loss,merged_summary],feed_dict={x1:xs_1,x2:xs_2,y:y_s,keep_prob:0.6}) writer.add_summary(summ,itera) if itera % 1000 == 1 : print('iter {},train loss {}'.format(itera,train_loss)) embed = sess.run(out1,feed_dict={x1:mnist.test.images,keep_prob:0.6}) test_img = mnist.test.images.reshape([-1,28,28,1]) writer.close()123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117这里多说一句，siamese能够用来降维，由于最后他的输出是二维的，这样直接把维度降下来了。--------------------- 做者：ML_BOY 来源：CSDN 原文：https://blog.csdn.net/qq1483661204/article/details/79039702 版权声明：本文为博主原创文章，转载请附上博文连接！