全连接网络：实现第一个全连接网络

骑猪看日落 2022-03-10 18:35 273阅读 0赞

# 全连接网络实现手写数字识别 #

程序分为三个部分，分别是

mnist\_forward.py：前向传播。

mnist\_backward.py：反向传播。

mnist\_test.py：模型测试。

# 前向传播 #

这里搭建了全连接网络。我使用了一个三层的网络。输入的是784个神经元（mnist中一张图片的大小）；隐藏层的神经元的个数分别是500，200；输出层是10个神经元，是预测的结果。

下面是mnist\_forward.py文件

import tensorflow as tf
    import os
    
    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"  # 忽略tensorflow警告信息
    
    def get_weight(shape,regularizer):
        w = tf.Variable(tf.truncated_normal(shape,stddev=0.1))
        if regularizer!=None:
            tf.add_to_collection('losses',tf.contrib.layers.l2_regularizer(regularizer)(w))
        return w
    
    def get_bias(shape):
        b = tf.Variable(tf.zeros(shape))
        return b
    
    def forward(x,regularizer):
        '''
        这里定义一个3层的全连接网络
        第0层：输入层，784个神经元
        第1层：隐藏层，500个神经元
        第2层：隐藏层，200个神经元
        第3层：输出层，10个神经元
        '''
        w1 = get_weight([784,500],regularizer)
        b1 = get_bias([500])
        y1 = tf.nn.leaky_relu(tf.matmul(x,w1) + b1)
    
        w2 = get_weight([500,200],regularizer)
        b2 = get_bias([200])
        y2 = tf.nn.leaky_relu(tf.matmul(y1,w2) + b2)
    
        w3 = get_weight([200,10],regularizer)
        b3 = get_bias([10])
        y = tf.matmul(y2,w3) + b3
        return y

# 反向传播 #

1）参数使用了滑动平均

2）使用指数衰减学习率

3）在训练过程中，进行了模型的保存

下面是mnist\_backward.py

import tensorflow as tf
    import os
    from tensorflow.examples.tutorials.mnist import input_data
    import mnist_forward
    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"  # 忽略tensorflow警告信息
    
    BATCH_SIZE = 200
    LEARNING_TARE_BASE = 0.1
    LEARNING_TARE_DECAY = 0.99
    STEPS = 50000
    regularizer = 0.0001
    moving_average_decay = 0.99
    model_saver_path = './model/' #模型保存的路径
    model_name = 'mnist_model' #模型的名字
    
    def backward(mnist):
        x = tf.placeholder(tf.float32, [None, 784]) # 输入
        y_ = tf.placeholder(tf.float32, [None, 10]) # 标签
        y = mnist_forward.forward(x, regularizer) # 前向传播的输出
        global_step = tf.Variable(0, trainable=False) # 计数器
    
        #加入滑动平均
        ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1))
        cem = tf.reduce_mean(ce)
        loss = cem + tf.add_n(tf.get_collection('losses'))
    
        #指数衰减学习率
        learning_rate = tf.train.exponential_decay(
            LEARNING_TARE_BASE,
            global_step,
            mnist.train.num_examples / BATCH_SIZE,
            LEARNING_TARE_DECAY,
            staircase=True)
    
        #定义训练过程
        train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
    
    
        #如果有滑动平均
        ema = tf.train.ExponentialMovingAverage(moving_average_decay,global_step)
        ema_op = ema.apply(tf.trainable_variables())
        with tf.control_dependencies([train_step,ema_op]):
            train_op = tf.no_op('train')
    
        #实例化saver
        saver = tf.train.Saver()
    
        #建立会话
        with tf.Session() as sess:
            init_op = tf.global_variables_initializer()
            sess.run(init_op)
    
            for i in range(STEPS):
                xs, ys = mnist.train.next_batch(BATCH_SIZE) #加载BATCH_SIZE个mnist中的图片和标签
                _, loss_value, step = sess.run([train_op, loss, global_step], feed_dict={x: xs, y_: ys})
                if i%1000 ==0:
                    print('After %d training step, loss on training batch is %g' %(step,loss_value))
                    # 保存模型到当前会话
                    saver.save(sess,os.path.join(model_saver_path,model_name),global_step=global_step)
    
    def main():
        mnist =input_data.read_data_sets('./data/',one_hot=True)
        backward(mnist)
    
    if __name__ == '__main__':
        main()

# 模型测试 #

因为在训练时使用滑动平均，所以在测试时，需要恢复参数的滑动平均值。

with tf.Graph().as_default() as g:  # 其内定义的节点在计算图g中

用这种方法，将神经网络复现到计算图中。

下面是：mnist\_test.py

import tensorflow as tf
    import os
    from tensorflow.examples.tutorials.mnist import input_data
    import mnist_forward
    import mnist_backward
    
    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"  # 忽略tensorflow警告信息
    
    def test(mnist):
        with tf.Graph().as_default() as g:
            x = tf.placeholder(tf.float32,[None,784])
            y_ = tf.placeholder(tf.float32,[None,10])
            y = mnist_forward.forward(x,None) # 前向传播获得的输出值
    
            #实例化带滑动平均的saver对象
            ema = tf.train.ExponentialMovingAverage(mnist_backward.moving_average_decay)
            ema_restore = ema.variables_to_restore()
            saver = tf.train.Saver(ema_restore)
    
            # 准确率计算
            correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(y_,1)) # 将输出结果和标签答案进行比较
            accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
    
    
            with tf.Session( ) as sess:
                ckpt = tf.train.get_checkpoint_state(mnist_backward.model_saver_path)
                #判断，如果有模型，恢复模型到当前会话
                if ckpt and ckpt.model_checkpoint_path:
                    saver.restore(sess,ckpt.model_checkpoint_path) # 恢复模型
                    global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1] # 获取当前轮数
                    # 喂入的是数据集中的，测试用的图片和标签
                    accuracy_score = sess.run(accuracy, feed_dict={x: mnist.test.images, y_: mnist.test.labels})
                    print('After %s training step, test accuracy = %g' %(global_step,accuracy_score))
                else:
                    print('No checkpoint file found')
    
    def main():
        mnist = input_data.read_data_sets('./data/',one_hot=True)
        test(mnist)
    
    if __name__ == '__main__':
        main()

# 结果 #

训练结果

> After 1 training step, loss on training batch is 2.7423  
> After 1001 training step, loss on training batch is 0.377551  
> After 2001 training step, loss on training batch is 0.243816
> 
> ......
> 
> fter 36001 training step, loss on training batch is 0.144181  
> After 37001 training step, loss on training batch is 0.143775  
> After 38001 training step, loss on training batch is 0.14224  
> After 39001 training step, loss on training batch is 0.144268  
> After 40001 training step, loss on training batch is 0.142971  
> After 41001 training step, loss on training batch is 0.14316  
> After 42001 training step, loss on training batch is 0.142349  
> After 43001 training step, loss on training batch is 0.142429  
> After 44001 training step, loss on training batch is 0.140229  
> After 45001 training step, loss on training batch is 0.140057  
> After 46001 training step, loss on training batch is 0.138865  
> After 47001 training step, loss on training batch is 0.137823  
> After 48001 training step, loss on training batch is 0.138767  
> After 49001 training step, loss on training batch is 0.139221

测试结果

> After 49001 training step, test accuracy = 0.98

# 如何实现断点续训？ #

实现断点续训，这样就可以在出现意外的情况下保存训练好的模型，下次训练，在此基础上进行。

上面的案例，若想要实现断点续训的功能，只需要在“反向传播”文件中，添加恢复模型的操作即可。

...... 
    #建立会话
        with tf.Session() as sess:
            init_op = tf.global_variables_initializer()
            sess.run(init_op)
    
            #实现断点续训，只需要加入下面三句话
            ckpt = tf.train.get_checkpoint_state(model_saver_path)
            if ckpt and ckpt.model_checkpoint_path:
                saver.restore(sess, ckpt.model_checkpoint_path)  # 恢复模型
    
            for i in range(STEPS):
                xs, ys = mnist.train.next_batch(BATCH_SIZE) #加载BATCH_SIZE个mnist中的图片和标签
                _, loss_value, step = sess.run([train_op, loss, global_step], feed_dict={x: xs, y_: ys})
                if i%1000 ==0:
    ......