分类任务5——AlexNet模型

我tm吹爆batch_norm！！！！！一开始以为换了tfrecord之后也可以和以前直接读取图片那样不需要batch_norm，但是几天的训练精度都在0.2，我不禁怀疑起了人生，最后用了batch_norm，一切都回来了。现在我把两个版本的代码都放上来，留作纪念，并希望可以帮助到需要的人。

附上个batch_norm的教程链接。Batch Normalization – Solutions

在定义网络层的时候，如果要通过tf.get_variable()来用进行参数的xavier初始化tf.contrib.layers.xavier_initializer()，那么在前面应该使用tf.variable_scope而不是tf.name_scope()。这是因为name_scope和get_variable之间会存在参数复用的问题，会报错。

没有batch_norm的代码如下：

"""

# AlexNet模型

"""
import tensorflow as tf


'''
获取tensor信息
'''
def print_tensor_info(tensor):
    print("tensor name:",tensor.op.name,"-tensor shape:",tensor.get_shape().as_list())


def inference(input, num_class, keep_prob):

    # 定义参数
    parameters = []

    # 第一层卷积层
    with tf.variable_scope("conv1"):
        # 设置卷积核5×5,3通道,32个卷积核
        # kernel1 = tf.Variable(tf.truncated_normal([5,5,3,32],mean=0,stddev=0.1,
        #                                           dtype=tf.float32),name="weights")
        kernel1 = tf.get_variable("weight", [5, 5, 3, 32],
                                  initializer=tf.contrib.layers.xavier_initializer())
        # 卷积,卷积的横向步长和竖向补偿都为4
        conv = tf.nn.conv2d(input,kernel1,[1,1,1,1],padding="SAME")
        #初始化偏置
        biases = tf.get_variable("bias", [32], initializer=tf.contrib.layers.xavier_initializer())
        bias = tf.nn.bias_add(conv,biases)
        #RELU激活函数
        conv1 = tf.nn.relu(bias)
        #输出该层的信息
        print_tensor_info(conv1)
        #统计参数
        parameters += [kernel1,biases]
        #lrn处理
        #lrn1 = tf.nn.lrn(conv1,4,bias=1,alpha=1e-3/9,beta=0.75,name="lrn1")
        #最大池化
        pool1 = tf.nn.max_pool(conv1,ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME",name="pool1")
        print_tensor_info(pool1)

    # 可视化输出
    with tf.variable_scope('visual'):
        x_min = tf.reduce_min(kernel1)
        x_max = tf.reduce_max(kernel1)
        kernel_0_to_1 = (kernel1 - x_min) / (x_max - x_min)
        # to tf.image_summary format [batch_size, height, width, channels]
        kernel_transposed = tf.transpose(kernel_0_to_1, [3, 0, 1, 2])
        # this will display random 3 filters from the 64 in conv1
        tf.summary.image('conv1/filters', kernel_transposed, max_outputs=3)
        layer1_image1 = conv1[0:1, :, :, 0:16]
        layer1_image1 = tf.transpose(layer1_image1, perm=[3, 1, 2, 0])
        tf.summary.image("filtered_images_layer1", layer1_image1, max_outputs=16)



    #第二层卷积层
    with tf.variable_scope("conv2"):
        #初始化权重
        kernel2 = tf.get_variable("weight", [5, 5, 32, 64],
                                  initializer=tf.contrib.layers.xavier_initializer())
        conv = tf.nn.conv2d(pool1,kernel2,[1,1,1,1],padding="SAME")
        #初始化偏置
        biases = tf.get_variable("bias", [64], initializer=tf.contrib.layers.xavier_initializer())
        bias = tf.nn.bias_add(conv,biases)
        #RELU激活
        conv2 = tf.nn.relu(bias)
        print_tensor_info(conv2)
        parameters += [kernel2,biases]
        #LRN
        #lrn2 = tf.nn.lrn(conv2,4,1.0,alpha=1e-3/9,beta=0.75,name="lrn2")
        #最大池化
        pool2 = tf.nn.max_pool(conv2,[1,2,2,1],[1,2,2,1],padding="SAME",name="pool2")
        print_tensor_info(pool2)

    #第三层卷积层
    with tf.variable_scope("conv3"):
        #初始化权重
        kernel3 = tf.get_variable("weight", [3, 3, 64, 128],
                                  initializer=tf.contrib.layers.xavier_initializer())
        conv = tf.nn.conv2d(pool2,kernel3,strides=[1,1,1,1],padding="SAME")
        biases = tf.get_variable("bias", [128], initializer=tf.contrib.layers.xavier_initializer())
        bias = tf.nn.bias_add(conv,biases)
        #RELU激活层
        conv3 = tf.nn.relu(bias)
        parameters += [kernel3,biases]
        print_tensor_info(conv3)
        # 最大池化
        pool3 = tf.nn.max_pool(conv3, [1, 2, 2, 1], [1, 2, 2, 1], padding="SAME", name="pool2")
        print_tensor_info(pool3)

    #第四层卷积层
    with tf.variable_scope("conv4") as scope:
        #初始化权重
        kernel4 = tf.get_variable("weight", [3, 3, 128, 256],
                                  initializer=tf.contrib.layers.xavier_initializer())
        #卷积
        conv = tf.nn.conv2d(conv3,kernel4,strides=[1,2,2,1],padding="SAME")
        biases = tf.get_variable("bias", [256], initializer=tf.contrib.layers.xavier_initializer())
        bias = tf.nn.bias_add(conv,biases)
        #RELU激活
        conv4 = tf.nn.relu(bias)
        parameters += [kernel4,biases]
        print_tensor_info(conv4)
        # 最大池化
        pool4 = tf.nn.max_pool(conv4, [1, 2, 2, 1], [1, 2, 2, 1], padding="SAME", name="pool2")
        print_tensor_info(pool4)

    # 第五层卷积层
    with tf.variable_scope("conv5") as scope:
        # 初始化权重
        kernel5 = tf.get_variable("weight", [3, 3, 256, 128],
                                  initializer=tf.contrib.layers.xavier_initializer())
        # 卷积
        conv = tf.nn.conv2d(conv4, kernel5, strides=[1, 2, 2, 1], padding="SAME")
        biases = tf.get_variable("bias", [128], initializer=tf.contrib.layers.xavier_initializer())
        bias = tf.nn.bias_add(conv, biases)
        # RELU激活
        conv5 = tf.nn.relu(bias)
        parameters += [kernel5, biases]
        print_tensor_info(conv5)
        # 最大池化
        pool5 = tf.nn.max_pool(conv5, [1, 2, 2, 1], [1, 2, 2, 1], padding="SAME", name="pool2")
        print_tensor_info(pool5)


    #第六层-全连接层1
    with tf.variable_scope('fc1'):
        pool6 = tf.reshape(pool5,(-1, 4*4*128))
        weight6 = tf.get_variable('weight',[4*4*128, 1024],initializer=tf.truncated_normal_initializer(stddev=0.1))
        # ful_bias1 = tf.Variable(tf.constant(0.0,dtype=tf.float32,shape=[1024]),name=
        # weight6 = tf.contrib.layers.l2_regularizer(0.0001)(weight6)
        regularizer1 = tf.contrib.layers.l2_regularizer(0.0001)
        tf.add_to_collection('losses', regularizer1(weight6))
        ful_bias1 = tf.get_variable("bias", [1024], initializer=tf.constant_initializer(0.1))

        # ful_con1 = tf.nn.relu(tf.add(tf.matmul(pool5,weight6),ful_bias1))
        ful_con1 = tf.nn.xw_plus_b(pool6, weight6, ful_bias1)
        ful_con1 = tf.nn.relu(ful_con1)
        ful_con1 = tf.nn.dropout(ful_con1, keep_prob)

    #第七层-全连接层2
    with tf.variable_scope('fc2'):
        weight7 = tf.get_variable('weight', [1024,512], initializer=tf.truncated_normal_initializer(stddev=0.1))
        # weight7 = tf.contrib.layers.l2_regularizer(0.0001)(weight7)
        regularizer2 = tf.contrib.layers.l2_regularizer(0.0001)
        tf.add_to_collection('losses', regularizer2(weight7))
        # ful_bias2 = tf.Variable(tf.constant(0.0,dtype=tf.float32,shape=[512]),name="ful_bias2")
        ful_bias2 = tf.get_variable("bias", [512], initializer=tf.constant_initializer(0.1))
        # ful_con2 = tf.nn.relu(tf.add(tf.matmul(ful_con1,weight7),ful_bias2))
        ful_con2 = tf.nn.xw_plus_b(ful_con1, weight7, ful_bias2)
        ful_con2 = tf.nn.relu(ful_con2)
        ful_con2 = tf.nn.dropout(ful_con2, keep_prob)


    #第八层-全连接层3
    with tf.variable_scope('fc3'):

        weight8 = tf.get_variable('weight', [512, num_class],initializer=tf.truncated_normal_initializer(stddev=0.1))
        # weight9 = tf.contrib.layers.l2_regularizer(0.0001)(weight9)
        regularizer3 = tf.contrib.layers.l2_regularizer(0.0001)
        tf.add_to_collection('losses', regularizer3(weight8))
        # bias9 = tf.Variable(tf.constant(0.0,shape=[num_class]),dtype=tf.float32,name="bias9")
        bias8 = tf.get_variable("bias", [num_class], initializer=tf.constant_initializer(0.1))
        # fc3 = tf.matmul(ful_con2,weight9)+bias9
        fc3 = tf.nn.xw_plus_b(ful_con2, weight8, bias8)
        # logits = tf.nn.l2_loss(output_softmax)

    return fc3, parameters

添加了batch_norm的代码如下：

import tensorflow as tf


def print_tensor_info(tensor):
    print("tensor name:", tensor.op.name,
          "-tensor shape:", tensor.get_shape().as_list())


def inference(input, num_class, keep_prob, is_training):

    with tf.variable_scope('conv1'):
        conv1 = tf.layers.conv2d(input, 32, 3, (1,1), 'same', use_bias=False, activation=tf.nn.relu)
        print_tensor_info(conv1)
        conv1_pool = tf.layers.max_pooling2d(conv1, (2,2), 2, 'same')
        print_tensor_info(conv1_pool)
        conv1_batch = tf.layers.batch_normalization(conv1_pool, training=is_training)

    with tf.variable_scope('conv2'):
        conv2 = tf.layers.conv2d(conv1_batch, 64, 3, (1,1), 'same', use_bias=False, activation=tf.nn.relu)
        print_tensor_info(conv2)
        conv2_pool = tf.layers.max_pooling2d(conv2, (2,2), 2, 'same')
        print_tensor_info(conv2_pool)
        conv2_batch = tf.layers.batch_normalization(conv2_pool, training=is_training)

    with tf.variable_scope('conv3'):
        conv3 = tf.layers.conv2d(conv2_batch, 128, 3, (1,1), 'same', use_bias=False, activation=tf.nn.relu)
        print_tensor_info(conv3)
        conv3_pool = tf.layers.max_pooling2d(conv3, (2,2), 2, 'same')
        print_tensor_info(conv3_pool)
        conv3_batch = tf.layers.batch_normalization(conv3_pool, training=is_training)

    with tf.variable_scope('conv4'):
        conv4 = tf.layers.conv2d(conv3_batch, 256, 3, (1,1), 'same', use_bias=False, activation=tf.nn.relu)
        print_tensor_info(conv4)
        conv4_pool = tf.layers.max_pooling2d(conv4, (2,2), 2, 'same')
        print_tensor_info(conv4_pool)
        conv4_batch = tf.layers.batch_normalization(conv4_pool, training=is_training)

    with tf.variable_scope('conv4_flatten'):
        orig_shape = conv4_batch.get_shape().as_list()
        flatten = tf.reshape(conv4_batch, shape=[-1, orig_shape[1] * orig_shape[2] * orig_shape[3]])
        print_tensor_info(flatten)

    with tf.variable_scope('fc1'):
        fc1 = tf.layers.dense(flatten, 512, use_bias=False, activation=None)
        fc1 = tf.layers.batch_normalization(fc1, training=is_training)
        fc1 = tf.nn.relu(fc1)
        fc1 = tf.nn.dropout(fc1, keep_prob)

    with tf.variable_scope('fc2'):
        fc2 = tf.layers.dense(fc1, 128, use_bias=False, activation=None)
        fc2 = tf.layers.batch_normalization(fc2, training=is_training)
        fc2 = tf.nn.relu(fc2)
        fc2 = tf.nn.dropout(fc2, keep_prob)

    with tf.variable_scope('fc3'):
        fc3 = tf.layers.dense(fc2, num_class, use_bias=False, activation=None)
        fc3 = tf.layers.batch_normalization(fc3, training=is_training)
        fc3 = tf.nn.relu(fc3)

    return fc3