实验7-使用TensorFlow完成MNIST手写体识别

逻辑回归

解决分类问题里最普遍的baseline model就是逻辑回归，简单同时可解释性好，使得它大受欢迎，我们来用tensorflow完成这个模型的搭建。

1.环境设定

import os
os.environ['TF_CPP_MIN_LOG_LEVEL']='2'

import numpy as np
#import tensorflow as tf
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
from tensorflow.examples.tutorials.mnist import input_data
import time

2.数据读取

#使用tensorflow自带的工具加载MNIST手写数字集合
mnist = input_data.read_data_sets('./data/mnist', one_hot=True) 

#查看一下数据维度
mnist.train.images.shape

#查看target维度
mnist.train.labels.shape

3.准备好placeholder

batch_size = 128
## 定义参数的数据类型 数据形状(一般为一维)名称
X = tf.placeholder(tf.float32, [batch_size, 784], name='X_placeholder') 
Y = tf.placeholder(tf.int32, [batch_size, 10], name='Y_placeholder')
global X

4.准备好参数/权重

# tf.Variable(initializer,name) 初始化参数 和 自定义的变量名称
#tf.random_normal()函数用于从“服从指定正态分布的序列”中随机取出指定个数的值
w = tf.Variable(tf.random.normal(shape=[784, 10], stddev=0.01), name='weights')
#tf.zeros（） 生成数组 位数 行个数
b = tf.Variable(tf.zeros([1, 10]), name="bias")

5.拿到每个类别的score

#tf.matmul（） 两个矩阵中对应元素各自相乘
logits = tf.matmul(X, w) + b 

6.计算多分类softmax的loss function

# 求交叉熵损失
entropy = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=Y, name='loss')
# 求平均
loss = tf.reduce_mean(entropy)

7.准备好optimier

这里的最优化用的是随机梯度下降，我们可以选择AdamOptimizer这样的优化器

learning_rate = 0.01
#自动进行参数的导数计算及优化
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(loss)

8.在session里执行graph里定义的运算

#迭代总轮次
n_epochs = 30

#分配GPU 占用率 
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333) 



with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
    # 在Tensorboard里可以看到图的结构
    writer = tf.summary.FileWriter('./graphs/logistic_reg', sess.graph)

    start_time = time.time()
    sess.run(tf.global_variables_initializer())    
    n_batches = int(mnist.train.num_examples/batch_size)
    for i in range(n_epochs): # 迭代这么多轮
        total_loss = 0

        for _ in range(n_batches):
            X_batch, Y_batch = mnist.train.next_batch(batch_size)
            _, loss_batch = sess.run([optimizer, loss], feed_dict={X: X_batch, Y:Y_batch}) 
            total_loss += loss_batch
        print('Average loss epoch {0}: {1}'.format(i, total_loss/n_batches))

    print('Total time: {0} seconds'.format(time.time() - start_time))

    print('Optimization Finished!')

    # 测试模型
    
    preds = tf.nn.softmax(logits)
    correct_preds = tf.equal(tf.argmax(preds, 1), tf.argmax(Y, 1))
    accuracy = tf.reduce_sum(tf.cast(correct_preds, tf.float32))
    
    n_batches = int(mnist.test.num_examples/batch_size)
    total_correct_preds = 0
    
    for i in range(n_batches):
        X_batch, Y_batch = mnist.test.next_batch(batch_size)
        accuracy_batch = sess.run([accuracy], feed_dict={X: X_batch, Y:Y_batch}) 
        total_correct_preds += accuracy_batch[0]
    
    print('Accuracy {0}'.format(total_correct_preds/mnist.test.num_examples))

    writer.close()