TensorFlow-7实现AlexNet

1 AlexNet网络结构

请参看：http://www.cnblogs.com/gongxijun/p/6027747.html

AlexNet网络的新技术点：

1）成功使用ReLU作为CNN的激活函数，成功解决了Sigmoid函数在网络较深时的梯度弥散问题。

2）训练时使用Dropout随机忽略一部分神经元，以避免过拟合。

3）在CNN中使用重叠的最大池化，避免平均池化的模糊效果，并提出步长比池化核的尺寸小，这样池化层的输出之间会有重叠和覆盖，提升了特征的丰富性。

4）提出了LRN层，对局部神经元的活动创建竞争机制，使得其中响应较大的值变得相对更大，并抑制其他反馈较小的神经元，增强模型的泛化能力。

5）使用CUDA加速深度卷积网络的训练；

6)数据增强，随机地从256X256的原始图像中截取224X224大小的区域（以及水平翻转的镜像），减轻过拟合。

实验中没有用imageNet的数据，只是随机生成的数据，只用来测时间，PS:本文没有GPU，全程是用cpu跑的（）

2 代码实现

__author__ = 'Administrator'
#AlexNet网络
import tensorflow as tf
import time
import datetime
import math

batch_size=32
num_batches=100

def print_activations(t):
    '''
    显示t的名称和尺寸
    :param t: tensor量
    '''
    print(t.op.name,' ',t.get_shape().as_list())

def inference(images):
    parameters=[]

    #卷积层1
    with tf.name_scope('conv1') as scope:
        kernel=tf.Variable(tf.truncated_normal([11,11,3,96],stddev=0.1,dtype=tf.float32),name='weights')
        biases=tf.Variable(tf.constant(0.0,shape=[96],dtype=tf.float32),trainable=True,name='biases')
        conv=tf.nn.conv2d(images,kernel,[1,4,4,1],padding='SAME')
        bias=tf.nn.bias_add(conv,biases) #卷积结束
        conv1=tf.nn.relu(bias,name=scope)
        print_activations(conv1)
        parameters+=[kernel, biases]

    #lrn1 pool1
    lrn1=tf.nn.lrn(conv1,4,1,0.001/9,0.75)
    pool1=tf.nn.max_pool(lrn1,[1,3,3,1],[1,2,2,1],padding='VALID',name='pool1')
    print_activations((pool1))

    #卷积层2
    with tf.name_scope('conv2') as scope:
        kernel=tf.Variable(tf.truncated_normal([5,5,96,256],stddev=0.1,dtype=tf.float32),name='weights')
        biases=tf.Variable(tf.constant(0.0,shape=[256],dtype=tf.float32),trainable=True,name='biases')
        conv=tf.nn.conv2d(pool1,kernel,[1,1,1,1],padding='SAME')
        bias=tf.nn.bias_add(conv,biases) #卷积结束
        conv2=tf.nn.relu(bias,name=scope)
        print_activations(conv2)
        parameters+=[kernel, biases]

    #lrn2 pool2
    lrn2=tf.nn.lrn(conv2,4,1,0.001/9,0.75)
    pool2=tf.nn.max_pool(lrn2,[1,3,3,1],[1,2,2,1],padding='VALID',name='pool2')
    print_activations(pool2)

    #卷积层3
    with tf.name_scope('conv3') as scope:
        kernel=tf.Variable(tf.truncated_normal([3,3,256,384],stddev=0.1,dtype=tf.float32),name='weights')
        biases=tf.Variable(tf.constant(0.0,shape=[384],dtype=tf.float32),trainable=True,name='biases')
        conv=tf.nn.conv2d(pool2,kernel,[1,1,1,1],padding='SAME')
        bias=tf.nn.bias_add(conv,biases) #卷积结束
        conv3=tf.nn.relu(bias,name=scope)
        print_activations(conv3)
        parameters+=[kernel, biases]

    #卷积层4
    with tf.name_scope('conv4') as scope:
        kernel=tf.Variable(tf.truncated_normal([3,3,384,384],stddev=0.1,dtype=tf.float32),name='weights')
        biases=tf.Variable(tf.constant(0.0,shape=[384],dtype=tf.float32),trainable=True,name='biases')
        conv=tf.nn.conv2d(conv3,kernel,[1,1,1,1],padding='SAME')
        bias=tf.nn.bias_add(conv,biases) #卷积结束
        conv4=tf.nn.relu(bias,name=scope)
        print_activations(conv4)
        parameters+=[kernel, biases]

    #卷积层5
    with tf.name_scope('conv5') as scope:
        kernel=tf.Variable(tf.truncated_normal([3,3,384,256],stddev=0.1,dtype=tf.float32),name='weights')
        biases=tf.Variable(tf.constant(0.0,shape=[256],dtype=tf.float32),trainable=True,name='biases')
        conv=tf.nn.conv2d(conv4,kernel,[1,1,1,1],padding='SAME')
        bias=tf.nn.bias_add(conv,biases) #卷积结束
        conv5=tf.nn.relu(bias,name=scope)
        print_activations(conv5)
        parameters+=[kernel, biases]


    #pool3
    pool3=tf.nn.max_pool(conv5,[1,3,3,1],[1,2,2,1],padding='VALID',name='pool3')
    print_activations(pool3)
    return pool3,parameters

    #Alex全连接层
    #全连接层1
    #pool3_flat=tf.reshape(pool3,[batch_size,None])
    # w_fc1=tf.Variable(tf.truncated_normal([None,4096],stddev=0.1,dtype=tf.float32))
    # b_fc1=tf.Variable(tf.constant(0.0,[4096],dtype=tf.float32))
    # f_fc1=tf.nn.relu(tf.matmul(pool3_flat,w_fc1)+b_fc1)
    # f_fc1=tf.nn.dropout(f_fc1,keep_pre)
    #
    # #全连接层2
    # w_fc2=tf.Variable(tf.truncated_normal([4096,4096],stddev=0.1,dtype=tf.float32))
    # b_fc2=tf.Variable(tf.constant(0.0,[4096],dtype=tf.float32))
    # f_fc2=tf.nn.relu(tf.matmul(f_fc1,w_fc2)+b_fc2)
    #
    # #全连接层3
    # w_fc3=tf.Variable(tf.truncated_normal([4096,1000],stddev=0.1,dtype=tf.float32))
    # b_fc2=tf.Variable(tf.constant(0.0,[1000],dtype=tf.float32))
    # f_fc2=tf.matmul(f_fc2,w_fc2)+b_fc2

def time_tensorflow_run(session,target,info_string):
    '''
    评估计算时间的函数
    :param session: tensorflow的session
    :param target: 执行的目标
    :param info_string: string
    :return:无
    '''
    num_steps_burn_in=10
    total_duration=0.0
    total_duration_square=0.0
    for i in range(num_batches+num_steps_burn_in):
        start_time=time.time()
        _=session.run(target)
        duration=time.time()-start_time
        if i>=num_steps_burn_in:
            if not i%10:
                print('%s:step %d, duration=%.3f'%(datetime.datetime.now(),i-num_steps_burn_in,duration))
        total_duration+=duration
        total_duration_square+=duration*duration
    mn=total_duration/num_batches
    vr=math.fabs(total_duration_square/num_batches-mn*mn)
    sd=math.sqrt(vr)
    print('%s: %s across %d steps, %.3f +/-%.3f sec /batch'%
          (datetime.datetime.now(),info_string,num_batches,mn,sd))

def run_benchmark():
    with tf.Graph().as_default():
        image_size=224
        images=tf.Variable(tf.random_normal([batch_size,image_size,image_size,3],
                                            dtype=tf.float32,stddev=1e-1))

        pool,parameters=inference(images)

        init=tf.global_variables_initializer()
        sess=tf.InteractiveSession()
        sess.run(init)

        time_tensorflow_run(sess,pool,'Forward')
        objective=tf.nn.l2_loss(pool)
        grad=tf.gradients(objective,parameters)
        time_tensorflow_run(sess,grad,'Forward-backward')

run_benchmark()

3 实验结果

conv1   [32, 56, 56, 96]
pool1   [32, 27, 27, 96]
conv2   [32, 27, 27, 256]
pool2   [32, 13, 13, 256]
conv3   [32, 13, 13, 384]
conv4   [32, 13, 13, 384]
conv5   [32, 13, 13, 256]
pool3   [32, 6, 6, 256]
2017-08-24 22:15:05.174433:step 0, duration=0.750
2017-08-24 22:15:12.671433:step 10, duration=0.748
2017-08-24 22:15:20.167433:step 20, duration=0.751
2017-08-24 22:15:27.665433:step 30, duration=0.750
2017-08-24 22:15:35.158433:step 40, duration=0.749
2017-08-24 22:15:42.651433:step 50, duration=0.749
2017-08-24 22:15:50.142433:step 60, duration=0.750
2017-08-24 22:15:57.643433:step 70, duration=0.750
2017-08-24 22:16:05.145433:step 80, duration=0.750
2017-08-24 22:16:12.646433:step 90, duration=0.750
2017-08-24 22:16:19.400433: Forward across 100 steps, 0.825 +/-0.249 sec /batch
2017-08-24 22:16:52.836633:step 0, duration=3.023
2017-08-24 22:17:23.461633:step 10, duration=3.083
2017-08-24 22:17:53.880833:step 20, duration=3.025
2017-08-24 22:18:24.242233:step 30, duration=3.018
2017-08-24 22:18:54.597233:step 40, duration=3.031
2017-08-24 22:19:24.979833:step 50, duration=3.055
2017-08-24 22:19:55.435033:step 60, duration=3.053
2017-08-24 22:20:25.889633:step 70, duration=3.038
2017-08-24 22:20:56.347233:step 80, duration=3.028
2017-08-24 22:21:26.645433:step 90, duration=3.043
2017-08-24 22:21:53.933033: Forward-backward across 100 steps, 3.345 +/-1.008 sec /batch