Tensorflow(一)--使用3层全连接神经网络（包含输入层）识别MNIST数据集

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
INPUT_SIZE=784
HIDE_LAYER_SIZE=500
OUTPUT_SIZE=10

BATCH_SIZE=100
BASE_LEARN_RATE=0.8
LEARN_DECAY_RATE=0.99
MOVINGAVERAGE_DECAY=0.99

REGULARIZATION_RATE=0.0001
TRAIN_ROUNDS=30000

def reference(input_data,avg_class,weight1,bias1,weight2,bias2):
    #1*784 784*500
    if(avg_class==None):
        layer1=tf.nn.relu(tf.matmul(input_data,weight1)+bias1)
        #因为后续有softmax处理，不需要激活函数
        return tf.matmul(layer1,weight2)+bias2
    else:
        #滑动平均模型
        layer1=tf.nn.relu(tf.matmul(input_data,avg_class.average(weight1))
                          +avg_class.average(bias1))
        return tf.matmul(layer1,avg_class.average(weight2))+avg_class.average(bias2)

def train(mnist):
    x=tf.placeholder(tf.float32,[None,INPUT_SIZE])
    y_=tf.placeholder(tf.float32,[None,OUTPUT_SIZE])

    weight_1=tf.Variable(tf.truncated_normal([INPUT_SIZE,HIDE_LAYER_SIZE],stddev=0.1))
    bias_1=tf.Variable(tf.constant(0.1,tf.float32))

    weight_2=tf.Variable(tf.truncated_normal([HIDE_LAYER_SIZE,OUTPUT_SIZE],stddev=0.1))
    bias_2=tf.Variable(tf.constant(0.1,tf.float32))

    #非滑动平均模型
    y=reference(x,None,weight_1,bias_1,weight_2,bias_2)

    #滑动平均模型
    globel_step=tf.Variable(0,False)
    average_class=tf.train.ExponentialMovingAverage(MOVINGAVERAGE_DECAY,globel_step)
    average_op=average_class.apply(tf.trainable_variables())

    average_y=reference(x,average_class,weight_1,bias_1,weight_2,bias_2)

    #cross entropy
    cross_entropy_y=tf.nn.sparse_softmax_cross_entropy_with_logits\
        (labels=tf.arg_max(y_,1),logits=y)
    cross_entropy_y_mean=tf.reduce_mean(cross_entropy_y)
    
    regularization=tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)

    #定义滑动平均模型的包含L2正则化损失函数的损失函数
    loss=cross_entropy_y_mean+regularization(weight_1)+regularization(weight_2)

    #定义指数下降的学习率
    learn_rate=tf.train.exponential_decay(BASE_LEARN_RATE,globel_step,
                                          mnist.train.num_examples/BATCH_SIZE,
                                         LEARN_DECAY_RATE)

    optimizer=tf.train.GradientDescentOptimizer(learning_rate=learn_rate).minimize\
    (loss,globel_step)

    #定义训练操作
    train_op=tf.group(optimizer,average_op)
    #正确率
    model_result=tf.equal(tf.arg_max(average_y,1),tf.arg_max(y_,1))
    accurity_model=tf.reduce_mean(tf.cast(model_result,tf.float32))

    with tf.Session() as sess:
        tf.initialize_all_variables().run()

        #声明验证集
        validation_x = mnist.validation.images
        validation_y = mnist.validation.labels
        validation={x:validation_x,y_:validation_y}
        #声明测试集
        test={x:mnist.test.images,y_:mnist.test.labels}
        for i in range(TRAIN_ROUNDS):
            #每1000轮打印结果
            if i%1000==0:
                accurity_value=sess.run(accurity_model,feed_dict=validation)
                print("step is %d,the accurity is %f\n"%(i,accurity_value))
        #训练
            xt,yt=mnist.train.next_batch(BATCH_SIZE)
            sess.run(train_op,feed_dict={x:xt,y_:yt})

        #打印结果
        accurity_value=sess.run(accurity_model,test)
        print("train %d step,the accurity is %f"%(TRAIN_ROUNDS,accurity_value))

def main(argv=None):
#建议去网站写好了数据集放到一个目录下使用，这个函数已经被弃用
    mnist=input_data.read_data_sets("${数据集所在的目录}",one_hot=True,source_url='http://yann.lecun.com/exdb/mnist/')
    train(mnist=mnist)

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

总结：

1.对滑动模型的理解错误。之前对滑动平均模型的预测值构造损失函数，结果是数据集的预测正确率基本没变，查了文档：

滑动平均模型只用于模型评估，不用于优化。

最终结果在98.3%

Tensorflow(一)--使用3层全连接神经网络（包含输入层）识别MNIST数据集

猜你喜欢