6经典卷积神经网络模型之AlexNet-6.1用于ImageNet的AlexNet网络计算图设计

6经典卷积神经网络模型之AlexNet-6.1用于ImageNet的AlexNet网络计算图设计

参考：http://www.jianshu.com/p/58168fec534d

AlexNet每层的超参数如图5所示。其中输入的图片尺寸为224*224，第一个卷积层使用了较大的卷积核尺寸11*11，步长为4，有96个卷积核；紧接着一个LRN层；然后是一个3*3的最大池化层，步长为2。这之后的卷积核尺寸都比较小，都是5*5或者3*3的大小，并且步长都是为1，即会扫描全图所有像素；而最大池化层依然保持为3*3，并且步长为2。我们可以发现一个比较有意思的现象，在前几个卷积层，虽然计算量很大，但参数量很小，都在1M左右甚至更小，只占了AlexNet总参数量的很小一部分。这就是卷积层有用的地方，可以通过较小的参数数量提取有效地特征。而如果前几层直接使用全连接层，那么参数量和计算量将成为天文数字。虽然每一个卷积层占整个网络的参数量的1%都不到，但是如果去掉任何一个卷积层，都会使网络的分类性能大幅地下降。

代码：

import tensorflow as tf
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

# 设置算法超参数
learning_rate_init = 0.001
training_epochs = 10
batch_size = 32
display_step = 10
conv1_kernel_num = 64
conv2_kernel_num = 192
conv3_kernel_num = 384
conv4_kernel_num = 256
conv5_kernel_num = 256

#数据集输入图像的参数
image_size = 224
image_channel = 3
n_classes = 1000

fc1_units_num = 4096
fc2_units_num = 4096
activation_func = tf.nn.relu
activation_name = 'relu'

#根据指定的维数返回初始化好的指定名称的权重 Variable
def WeightsVariable(shape, name_str, stddev=0.1):
    initial = tf.truncated_normal(shape=shape, stddev=stddev, dtype=tf.float32)
    return tf.Variable(initial, dtype=tf.float32, name=name_str)

#根据指定的维数返回初始化好的指定名称的偏置 Variable
def BiasesVariable(shape, name_str, init_value=0.00001):
    initial = tf.constant(init_value, shape=shape)
    return tf.Variable(initial, dtype=tf.float32, name=name_str)

# 二维卷积层activation(conv2d+bias)的封装
def Conv2d(x, W, b, stride=1, padding='SAME',activation=tf.nn.relu,act_name='relu'):
    with tf.name_scope('conv2d_bias'):
        y = tf.nn.conv2d(x, W, strides=[1, stride, stride, 1], padding=padding)
        y = tf.nn.bias_add(y, b)
    with tf.name_scope(act_name):
        y = activation(y)
    return y

# 二维池化层pool的封装
def Pool2d(x, pool= tf.nn.max_pool, k=2, stride=2,padding='SAME'):
    return pool(x, ksize=[1, k, k, 1], strides=[1, stride, stride, 1], padding=padding)

# 全连接层activate(wx+b)的封装
def FullyConnected(x, W, b, activation=tf.nn.relu, act_name='relu'):
    with tf.name_scope('Wx_b'):
        y = tf.matmul(x, W)
        y = tf.add(y, b)
    with tf.name_scope(act_name):
        y = activation(y)
    return y

#修改了4处激活函数：Conv2d_1、Conv2d_2、FC1_nonlinear、FC2_nonlinear
def Inference(image_holder):
    # 第一个卷积层activate(conv2d + biase)
    with tf.name_scope('Conv2d_1'):
        weights = WeightsVariable(shape=[11, 11, image_channel, conv1_kernel_num],name_str='weights',stddev=1e-1)
        biases = BiasesVariable(shape=[conv1_kernel_num], name_str='biases',init_value=0.0)
        conv1_out = Conv2d(image_holder, weights, biases, stride=4, padding='SAME')#输出：[32x56x56x64]

    # 第一个池化层(pool 2d)
    with tf.name_scope('Pool2d_1'):
        pool1_out = Pool2d(conv1_out, pool=tf.nn.max_pool, k=3, stride=2,padding='VALID')

    # 第二个卷积层activate(conv2d + biase)
    with tf.name_scope('Conv2d_2'):
        weights = WeightsVariable(shape=[5, 5, conv1_kernel_num, conv2_kernel_num], name_str='weights', stddev=1e-1)
        biases = BiasesVariable(shape=[conv2_kernel_num], name_str='biases', init_value=0.0)
        conv2_out = Conv2d(pool1_out, weights, biases, stride=1, padding='SAME')

    # 第二个池化层(pool 2d)
    with tf.name_scope('Pool2d_2'):
        pool2_out = Pool2d(conv2_out, pool=tf.nn.max_pool, k=3, stride=2, padding='VALID')

    # 第三个卷积层activate(conv2d + biase)
    with tf.name_scope('Conv2d_3'):
        weights = WeightsVariable(shape=[3, 3, conv2_kernel_num, conv3_kernel_num], name_str='weights', stddev=1e-1)
        biases = BiasesVariable(shape=[conv3_kernel_num], name_str='biases', init_value=0.0)
        conv3_out = Conv2d(pool2_out, weights, biases, stride=1, padding='SAME')

    # 第四个卷积层activate(conv2d + biase)
    with tf.name_scope('Conv2d_4'):
        weights = WeightsVariable(shape=[3, 3, conv3_kernel_num, conv4_kernel_num], name_str='weights', stddev=1e-1)
        biases = BiasesVariable(shape=[conv4_kernel_num], name_str='biases', init_value=0.0)
        conv4_out = Conv2d(conv3_out, weights, biases, stride=1, padding='SAME')

    # 第五个卷积层activate(conv2d + biase)
    with tf.name_scope('Conv2d_5'):
        weights = WeightsVariable(shape=[3, 3, conv4_kernel_num, conv5_kernel_num], name_str='weights', stddev=1e-1)
        biases = BiasesVariable(shape=[conv5_kernel_num], name_str='biases', init_value=0.0)
        conv5_out = Conv2d(conv4_out, weights, biases, stride=1, padding='SAME')

    # 第五个池化层(pool 2d)
    with tf.name_scope('Pool2d_5'):
        pool5_out = Pool2d(conv5_out, pool=tf.nn.max_pool, k=3, stride=2, padding='VALID')

    #将二维特征图变换为一维特征向量
    with tf.name_scope('FeatsReshape'):
        features = tf.reshape(pool5_out, [batch_size,-1])
        feats_dim = features.get_shape()[1].value

    # 第一个全连接层(fully connected layer)
    with tf.name_scope('FC1_nonlinear'):
        weights = WeightsVariable(shape=[feats_dim, fc1_units_num],name_str='weights',stddev=4e-2)
        biases = BiasesVariable(shape=[fc1_units_num], name_str='biases',init_value=0.1)
        fc1_out = FullyConnected(features, weights, biases, activation=activation_func,act_name=activation_name)

    # 第二个全连接层(fully connected layer)
    with tf.name_scope('FC2_nonlinear'):
        weights = WeightsVariable(shape=[fc1_units_num, fc2_units_num],name_str='weights',stddev=4e-2)
        biases = BiasesVariable(shape=[fc2_units_num], name_str='biases',init_value=0.1)
        fc2_out = FullyConnected(fc1_out, weights, biases, activation=activation_func,act_name=activation_name)

    # 第三个全连接层(fully connected layer)
    with tf.name_scope('FC3_linear'):
        fc3_units_num = n_classes
        weights = WeightsVariable(shape=[fc2_units_num, fc3_units_num],name_str='weights',stddev=1.0/fc2_units_num)
        biases = BiasesVariable(shape=[fc3_units_num], name_str='biases',init_value=0.0)
        logits = FullyConnected(fc2_out, weights, biases,activation=tf.identity, act_name='linear')
    return logits

#调用上面写的函数构造计算图
with tf.Graph().as_default():
    # 计算图输入
    with tf.name_scope('Inputs'):
        #[32x224x224x3]
        image_holder = tf.placeholder(tf.float32, [batch_size, image_size, image_size, image_channel], name='images')
        labels_holder = tf.placeholder(tf.int32, [batch_size], name='labels')

    # 计算图前向推断过程
    with tf.name_scope('Inference'):
         logits = Inference(image_holder)

    # 定义损失层(loss layer)
    with tf.name_scope('Loss'):
        cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels_holder,logits=logits)
        cross_entropy_mean = tf.reduce_mean(cross_entropy)
        total_loss_op = cross_entropy_mean

    # 定义优化训练层(train layer)
    with tf.name_scope('Train'):
        learning_rate = tf.placeholder(tf.float32)
        global_step = tf.Variable(0, name='global_step', trainable=False, dtype=tf.int64)
        optimizer = tf.train.RMSPropOptimizer(learning_rate=learning_rate)
        train_op = optimizer.minimize(total_loss_op,global_step=global_step)

    # 定义模型评估层(evaluate layer)
    with tf.name_scope('Evaluate'):
        top_K_op = tf.nn.in_top_k(predictions=logits,targets=labels_holder,k=1)

    # 添加所有变量的初始化节点
    init_op = tf.global_variables_initializer()

    print('把计算图写入事件文件，在TensorBoard里面查看')
    summary_writer = tf.summary.FileWriter(logdir='logs/alexnet_graph')
    summary_writer.add_graph(graph=tf.get_default_graph())
    summary_writer.flush()

计算图