（三）Tensorflow的逻辑回归模型

使用逻辑回归模型去识别tensorflow数据集的数字识别。

代码展示：

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import math

def logistic_regression():
    #加载数据
    mnist = input_data.read_data_sets(r'C:\Users\Administrator\Desktop\AI_project\tensorflow\MNIST_data', one_hot=True)
    #给数据建立placeholder占位符
    batch_size = 128
    x = tf.placeholder(tf.float32,[batch_size,784],name="x_data")
    y = tf.placeholder(tf.int32,[batch_size,10],name="y_data")
    #初始化权重
    w = tf.Variable(tf.random_normal([784,10],stddev=0.1))
    b = tf.Variable(tf.zeros([10]))
    #计算损失，定义优化方法
    val = tf.add(tf.matmul(x,w),b)  #这个是矩阵128*10......看清了
    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y,logits=val)) #这个是向量....看清了
    train = tf.train.AdamOptimizer().minimize(loss)
    #测试模型
    pred = tf.nn.softmax(val)
    pred = tf.equal(tf.argmax(input=pred,axis=1),tf.argmax(input=y,axis=1))
    accuary = tf.reduce_mean(tf.cast(pred,tf.float32))
    #初始化变量
    init = tf.global_variables_initializer()
    #对session进行执行定义好的运算
    with tf.Session() as sess:
        sess.run(init)
        writer = tf.summary.FileWriter("./logs",sess.graph)
        writer.close()
        n_batch_train = math.ceil(mnist.train.num_examples / batch_size)
        n_batch_test = math.ceil(mnist.train.num_examples / batch_size)
        for i in range(50):
            loss_total = 0
            #训练模型
            for _ in range( n_batch_train ):
                x_input,y_input = mnist.train.next_batch(batch_size)
                _,l = sess.run([train,loss],feed_dict={x:x_input,y:y_input})
                loss_total+=l
            print("Interation:{},loss:{}".format(i,loss_total/n_batch_train),end="  ")
            accuary_total = 0
            #测试模型
            for _ in range(n_batch_test):
                x_input, y_input = mnist.test.next_batch(batch_size)
                accuary_total+=sess.run(accuary,feed_dict={x:x_input,y:y_input})
            print("accuary:{}".format(accuary_total/n_batch_test))

结果展示：

Interation:0,loss:0.7622147449920343  accuary:0.8922420058139535
Interation:1,loss:0.3823678414142409  accuary:0.9089571220930233
Interation:2,loss:0.3328973266967507  accuary:0.9157885174418605
Interation:3,loss:0.3105759471308353  accuary:0.9191860465116279
Interation:4,loss:0.2960549441534419  accuary:0.9207485465116279
Interation:5,loss:0.2865448264535083  accuary:0.9211300872093023
Interation:6,loss:0.2805248944045499  accuary:0.9229651162790697
Interation:7,loss:0.2745572795008504  accuary:0.924781976744186
Interation:8,loss:0.2710500391243502  accuary:0.9243459302325582
Interation:9,loss:0.26706690582078557  accuary:0.9253997093023256
Interation:10,loss:0.2629715468647868  accuary:0.9269440406976744
Interation:11,loss:0.2621379436968371  accuary:0.9255450581395349
Interation:12,loss:0.2595898272512957  accuary:0.9261264534883721
Interation:13,loss:0.2563329931435197  accuary:0.9253088662790697
Interation:14,loss:0.2554247479279374  accuary:0.9272529069767442
Interation:15,loss:0.25390707252677097  accuary:0.9283793604651163
Interation:16,loss:0.2524607227985249  accuary:0.9253997093023256
Interation:17,loss:0.2505904925423999  accuary:0.9253815406976744
Interation:18,loss:0.25018050428046734  accuary:0.9276526162790698
Interation:19,loss:0.24771193742405537  accuary:0.9277979651162791
Interation:20,loss:0.24811712648979453  accuary:0.9273074127906977
Interation:21,loss:0.2459677146444487  accuary:0.9260174418604651
Interation:22,loss:0.24644057648473008  accuary:0.9283793604651163
Interation:23,loss:0.24489293940538584  accuary:0.9258539244186047
Interation:24,loss:0.24324410925077838  accuary:0.9274890988372093
Interation:25,loss:0.24299994336311206  accuary:0.9282885174418605
Interation:26,loss:0.24238698946875195  accuary:0.9269077034883721
Interation:27,loss:0.24184117951365405  accuary:0.9276162790697674
Interation:28,loss:0.2400227582038835  accuary:0.9278343023255814
Interation:29,loss:0.2400645112748756  accuary:0.9271438953488372
Interation:30,loss:0.24123970490208893  accuary:0.9277071220930233
Interation:31,loss:0.23869256044543066  accuary:0.9289244186046511
Interation:32,loss:0.23661559423388437  accuary:0.9276344476744186
Interation:33,loss:0.23885162564557652  accuary:0.9273982558139535
Interation:34,loss:0.23682482469567032  accuary:0.9284883720930233
Interation:35,loss:0.2369533678300159  accuary:0.9261809593023256
Interation:36,loss:0.23730118084092472  accuary:0.9273800872093023
Interation:37,loss:0.23646007918341216  accuary:0.9275617732558139
Interation:38,loss:0.23514422782978348  accuary:0.9273255813953488
Interation:39,loss:0.2359093708181104  accuary:0.9276707848837209
Interation:40,loss:0.23345869906073394  accuary:0.9270893895348837
Interation:41,loss:0.2352673845062422  accuary:0.9272710755813953
Interation:42,loss:0.23326533116226972  accuary:0.9262718023255814
Interation:43,loss:0.2358175446474275  accuary:0.9289607558139535
Interation:44,loss:0.23324581149012544  accuary:0.9281431686046512
Interation:45,loss:0.23273995503090147  accuary:0.9288154069767441
Interation:46,loss:0.23142784215336623  accuary:0.9275981104651163
Interation:47,loss:0.23274108893996062  accuary:0.9285428779069768
Interation:48,loss:0.23295457486149876  accuary:0.9285428779069768
Interation:49,loss:0.23111646177117215  accuary:0.9279069767441861

总结：

还是数据批处理的问题，防止内存爆掉（我亲身经历了好几次，内存99%，然后死机的！），优化函数方法的选取问题，

注意：axis的使用，损失函数返回的形式（维数），关注矩阵的维数。