6.2. 模型接口建立

我们将模型接口都放在cifar_omdel.py文件当中,设计了四个函数,input()作为从cifar_data文件中数据的获取,inference()作为神经网络模型的建立,total_loss()计算模型的损失,train()来通过梯度下降训练减少损失

input代码

def input():
    """ 获取输入数据 :return: image,label """

    # 实例化
    cfr = cifar_data.CifarRead()

    # 生成张量
    image_batch, lab_batch = cfr.read_tfrecords()

    # 将目标值转换为one-hot编码格式
    label = tf.one_hot(label_batch, depth=10, on_value=1.0)

    return image_batch, label, label_batch

inference代码

在这里使用的卷积神经网络模型与前面一致,需要修改图像的通道数以及经过两次卷积池化变换后的图像大小。

def inference(image_batch):
    """ 得到模型的输出 :return: 预测概率输出以及占位符 """
    # 1、数据占位符建立
    with tf.variable_scope("data"):
        # 样本标签值
        # y_label = tf.placeholder(tf.float32, [None, 10])

        # 样本特征值
        # x = tf.placeholder(tf.float32, [None, IMAGE_HEIGHT * IMAGE_WIDTH * IMAGE_DEPTH])

        # 改变形状,以提供给卷积层使用
        x_image = tf.reshape(image_batch, [-1, 32, 32, 3])

    # 2、卷积池化第一层
    with tf.variable_scope("conv1"):
        # 构建权重, 5*5, 3个输入通道,32个输出通道
        w_conv1 = weight_variable([5, 5, 3, 32])

        # 构建偏置, 个数位输出通道数
        b_conv1 = bias_variable([32])

        # 进行卷积,激活,指定滑动窗口,填充类型
        y_relu1 = tf.nn.relu(tf.nn.conv2d(x_image, w_conv1, strides=[1, 1, 1, 1], padding="SAME") + b_conv1)

        y_conv1 = tf.nn.max_pool(y_relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

    # 3、卷积池化第二层
    with tf.variable_scope("conv_pool2"):
        # 构建权重, 5*5, 一个输入通道,32个输出通道
        w_conv2 = weight_variable([5, 5, 32, 64])

        # 构建偏置, 个数位输出通道数
        b_conv2 = bias_variable([64])

        # 进行卷积,激活,指定滑动窗口,填充类型
        y_relu2 = tf.nn.relu(tf.nn.conv2d(y_conv1, w_conv2, strides=[1, 1, 1, 1], padding="SAME") + b_conv2)

        y_conv2 = tf.nn.max_pool(y_relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

    # 4、全连接第一层
    with tf.variable_scope("FC1"):
        # 构建权重,[7*7*64, 1024],根据前面的卷积池化后一步步计算的大小变换是32->16->8
        w_fc1 = weight_variable([8 * 8 * 64, 1024])

        # 构建偏置,个数位第一次全连接层输出个数
        b_fc1 = bias_variable([1024])

        y_reshape = tf.reshape(y_conv2, [-1, 8 * 8 * 64])

        # 全连接结果激活
        y_fc1 = tf.nn.relu(tf.matmul(y_reshape, w_fc1) + b_fc1)

    # 5、全连接第二层
    with tf.variable_scope("FC2"):

        # droupout层
        droup = tf.nn.dropout(y_fc1, 1.0)

        # 构建权重,[1024, 10]
        w_fc2 = weight_variable([1024, 10])

        # 构建偏置 [10]
        b_fc2 = bias_variable([10])

        # 最后的全连接层
        y_logit = tf.matmul(droup, w_fc2) + b_fc2

    return y_logit

total_loss代码

def total_loss(y_label, y_logit):
    """ 计算训练损失 :param y_label: 目标值 :param y_logit: 计算值 :return: 损失 """
    with tf.variable_scope("loss"):

        # softmax回归,以及计算交叉损失熵
        cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=y_label, logits=y_logit)

        # 计算损失平均值
        loss = tf.reduce_mean(cross_entropy)

    return loss

train代码

def train(loss, y_label, y_logit, global_step):
    """ 训练数据得出准确率 :param loss: 损失大小 :return: """
    with tf.variable_scope("train"):
        # 让学习率根据步伐,自动变换学习率,指定了每10步衰减基数为0.99,0.001为初始的学习率
        lr = tf.train.exponential_decay(0.001,
                                        global_step,
                                        10,
                                        0.99,
                                        staircase=True)

        # 优化器
        train_op = tf.train.GradientDescentOptimizer(lr).minimize(loss, global_step=global_step)

        # 计算准确率
        equal_list = tf.equal(tf.argmax(y_logit, 1), tf.argmax(y_label, 1))

        accuracy = tf.reduce_mean(tf.cast(equal_list, tf.float32))

    return train_op, accuracy

完整代码

import tensorflow as tf
import os
import cifar_data
#
#
from tensorflow.examples.tutorials.mnist import input_data

IMAGE_HEIGHT = 32
IMAGE_WIDTH = 32
IMAGE_DEPTH = 3


# 按照指定形状构建权重变量
def weight_variable(shape):
    init = tf.truncated_normal(shape=shape, mean=0.0, stddev=1.0, dtype=tf.float32)
    weight = tf.Variable(init)
    return weight


# 按照制定形状构建偏置变量
def bias_variable(shape):
    bias = tf.constant([1.0], shape=shape)
    return tf.Variable(bias)


def inference(image_batch):
    """ 得到模型的输出 :return: 预测概率输出以及占位符 """
    # 1、数据占位符建立
    with tf.variable_scope("data"):
        # 样本标签值
        # y_label = tf.placeholder(tf.float32, [None, 10])

        # 样本特征值
        # x = tf.placeholder(tf.float32, [None, IMAGE_HEIGHT * IMAGE_WIDTH * IMAGE_DEPTH])

        # 改变形状,以提供给卷积层使用
        x_image = tf.reshape(image_batch, [-1, 32, 32, 3])

    # 2、卷积池化第一层
    with tf.variable_scope("conv1"):
        # 构建权重, 5*5, 3个输入通道,32个输出通道
        w_conv1 = weight_variable([5, 5, 3, 32])

        # 构建偏置, 个数位输出通道数
        b_conv1 = bias_variable([32])

        # 进行卷积,激活,指定滑动窗口,填充类型
        y_relu1 = tf.nn.relu(tf.nn.conv2d(x_image, w_conv1, strides=[1, 1, 1, 1], padding="SAME") + b_conv1)

        y_conv1 = tf.nn.max_pool(y_relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

    # 3、卷积池化第二层
    with tf.variable_scope("conv_pool2"):
        # 构建权重, 5*5, 一个输入通道,32个输出通道
        w_conv2 = weight_variable([5, 5, 32, 64])

        # 构建偏置, 个数位输出通道数
        b_conv2 = bias_variable([64])

        # 进行卷积,激活,指定滑动窗口,填充类型
        y_relu2 = tf.nn.relu(tf.nn.conv2d(y_conv1, w_conv2, strides=[1, 1, 1, 1], padding="SAME") + b_conv2)

        y_conv2 = tf.nn.max_pool(y_relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

    # 4、全连接第一层
    with tf.variable_scope("FC1"):
        # 构建权重,[7*7*64, 1024],根据前面的卷积池化后一步步计算的大小变换是32->16->8
        w_fc1 = weight_variable([8 * 8 * 64, 1024])

        # 构建偏置,个数位第一次全连接层输出个数
        b_fc1 = bias_variable([1024])

        y_reshape = tf.reshape(y_conv2, [-1, 8 * 8 * 64])

        # 全连接结果激活
        y_fc1 = tf.nn.relu(tf.matmul(y_reshape, w_fc1) + b_fc1)

    # 5、全连接第二层
    with tf.variable_scope("FC2"):

        # droupout层
        droup = tf.nn.dropout(y_fc1, 1.0)

        # 构建权重,[1024, 10]
        w_fc2 = weight_variable([1024, 10])

        # 构建偏置 [10]
        b_fc2 = bias_variable([10])

        # 最后的全连接层
        y_logit = tf.matmul(droup, w_fc2) + b_fc2

    return y_logit


def total_loss(y_label, y_logit):
    """ 计算训练损失 :param y_label: 目标值 :param y_logit: 计算值 :return: 损失 """
    with tf.variable_scope("loss"):
        # 将y_label转换为one-hot编码形式
        # y_onehot = tf.one_hot(y_label, depth=10, on_value=1.0)

        # softmax回归,以及计算交叉损失熵
        cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=y_label, logits=y_logit)

        # 计算损失平均值
        loss = tf.reduce_mean(cross_entropy)

    return loss


def train(loss, y_label, y_logit, global_step):
    """ 训练数据得出准确率 :param loss: 损失大小 :return: """
    with tf.variable_scope("train"):
        # 让学习率根据步伐,自动变换学习率,指定了每10步衰减基数为0.99,0.001为初始的学习率
        lr = tf.train.exponential_decay(0.001,
                                        global_step,
                                        10,
                                        0.99,
                                        staircase=True)

        # 优化器
        train_op = tf.train.GradientDescentOptimizer(lr).minimize(loss, global_step=global_step)

        # 计算准确率
        equal_list = tf.equal(tf.argmax(y_logit, 1), tf.argmax(y_label, 1))

        accuracy = tf.reduce_mean(tf.cast(equal_list, tf.float32))

    return train_op, accuracy


def input():
    """ 获取输入数据 :return: image,label """

    # 实例化
    cfr = cifar_data.CifarRead()

    # 生成张量
    image_batch, lab_batch = cfr.read_tfrecords()

    # 将目标值转换为one-hot编码格式
    label = tf.one_hot(label_batch, depth=10, on_value=1.0)

    return image_batch, label, label_batch

6.3. 训练以及高级会话函数

主训练逻辑

我们将在cifar_train.py文件实现主要训练逻辑。在这里我们将使用一个新的会话函数,叫tf.train.MonitoredTrainingSession

优点: 1、它自动的建立events文件、checkpoint文件,以记录重要的信息。 2、可以定义钩子函数,可以自定义每批次的训练信息,训练的限制等等

注意:在这个里面我们需要添加一个全局步数,这个步数是每批次训练的时候进行+1计数,内部使用。

代码如下:

import tensorflow as tf
import cifar_model
import time
from datetime import datetime



def train():
    # 在图中进行训练
    with tf.Graph().as_default():
        # 定义全局步数,必须得使用这个,否则会出现StopCounterHook错误
        global_step = tf.contrib.framework.get_or_create_global_step()

        # 获取数据
        image, label, label_1 = cifar_model.input()

        # 通过模型进行类别预测
        y_logit = cifar_model.inference(image)

        # 计算损失
        loss = cifar_model.total_loss(label, y_logit)

        # 进行优化器减少损失
        train_op, accuracy = cifar_model.train(loss, label, y_logit, global_step)

        # 通过钩子定义模型输出
        class _LoggerHook(tf.train.SessionRunHook):
            """Logs loss and runtime."""
            def begin(self):
                self._step = -1
                self._start_time = time.time()

            def before_run(self, run_context):
                self._step += 1
                return tf.train.SessionRunArgs(loss, float(accuracy.eval()))  # Asks for loss value.

            def after_run(self, run_context, run_values):
                if self._step % 10 == 0:
                    current_time = time.time()
                    duration = current_time - self._start_time
                    self._start_time = current_time
                    loss_value = run_values.results
                    examples_per_sec = 10 * 10 / duration
                    sec_per_batch = float(duration / 10)

                    format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
                                  'sec/batch)')
                    print(format_str % (datetime.now(), self._step, loss_value,
                                        examples_per_sec, sec_per_batch))

        with tf.train.MonitoredTrainingSession(
                checkpoint_dir="./cifartrain/train",
                hooks=[tf.train.StopAtStepHook(last_step=500),# 定义执行的训练轮数也就是max_step,超过了就会报错
                       tf.train.NanTensorHook(loss),
                       _LoggerHook()],
                config=tf.ConfigProto(
                    log_device_placement=False)) as mon_sess:
            while not mon_sess.should_stop():
                mon_sess.run(train_op)


def main(argv):
    train()


if __name__ == "__main__":
    tf.app.run()