mnist手写体识别

Mnist数据集可以从官网下载，网址： http://yann.lecun.com/exdb/mnist/ 下载下来的数据集被分成两部分：55000行的训练数据集（mnist.train）和10000行的测试数据集（mnist.test）。每一个MNIST数据单元有两部分组成：一张包含手写数字的图片和一个对应的标签。我们把这些图片设为“xs”，把这些标签设为“ys”。训练数据集和测试数据集都包含xs和ys，比如训练数据集的图片是 mnist.train.images ，训练数据集的标签是 mnist.train.labels。

我们可以知道图片是黑白图片，每一张图片包含28像素X28像素。我们把这个数组展开成一个向量，长度是 28x28 = 784。因此，在MNIST训练数据集中，mnist.train.images 是一个形状为 [60000, 784] 的张量。

MNIST中的每个图像都具有相应的标签，0到9之间的数字表示图像中绘制的数字。用的是one-hot编码

单层(全连接层)实现手写数字识别

1,定义数据占位符 特征值[None,784] 目标值[None,10]

with tf.variable_scope("data"):
    x = tf.placeholder(tf.float32,[None,784])
    y_true = tf.placeholder(tf.float32,[None,10])

2,建立模型 随机初始化权重和偏置,w[784,10],b= [10] y_predict = tf.matmul(x,w)+b

with tf.variable_scope("model"):
    w = tf.Variable(tf.random_normal([784,10],mean=0.0,stddev=1.0))
    b = tf.Variable(tf.constant(0.0,shape=[10]))
    y_predict = tf.matmul(x,w)+b

3,计算损失 loss 平均样本损失

with tf.variable_scope("compute_loss"):
    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_true,logits=y_predict))

4,梯度下降优化 0.1 步数 2000 从而得出准确率

with tf.variable_scope("optimizer"):
    train_op = tf.train.GradientDescentOptimizer(0.1).minimize(loss)

5,模型评估 argmax() reduce_mean

with tf.variable_scope("acc"):
    eq = tf.equal(tf.argmax(y_true, 1), tf.argmax(y_predict, 1))
    accuracy = tf.reduce_mean(tf.cast(eq,tf.float32))

加载mnist数据集

import tensorflow as tf
# 这里我们利用tensorflow给好的读取数据的方法
from tensorflow.examples.tutorials.mnist import input_data
def full_connected():
    # 加载mnist数据集
    mnist = input_data.read_data_sets("data/mnist/input_data",one_hot=True)

运行结果

accuracy: 0.08accuracy: 0.08accuracy: 0.1accuracy: 0.1accuracy: 0.1accuracy: 0.1accuracy: 0.1accuracy: 0.1accuracy: 0.14accuracy: 0.14accuracy: 0.16accuracy: 0.16accuracy: 0.18accuracy: 0.2accuracy: 0.2accuracy: 0.2accuracy: 0.24accuracy: 0.24accuracy: 0.24accuracy: 0.26accuracy: 0.26accuracy: 0.26accuracy: 0.28accuracy: 0.28accuracy: 0.3accuracy: 0.3accuracy: 0.32accuracy: 0.32accuracy: 0.32accuracy: 0.36accuracy: 0.4accuracy: 0.4accuracy: 0.4accuracy: 0.42accuracy: 0.44accuracy: 0.44accuracy: 0.44accuracy: 0.44accuracy: 0.44accuracy: 0.46accuracy: 0.46accuracy: 0.46accuracy: 0.46accuracy: 0.46accuracy: 0.48accuracy: 0.48accuracy: 0.48accuracy: 0.48accuracy: 0.48accuracy: 0.48accuracy: 0.52accuracy: 0.52accuracy: 0.54accuracy: 0.54accuracy: 0.54accuracy: 0.54accuracy: 0.56accuracy: 0.56accuracy: 0.56accuracy: 0.58accuracy: 0.6accuracy: 0.6accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.62accuracy: 0.64accuracy: 0.66accuracy: 0.66accuracy: 0.66accuracy: 0.66accuracy: 0.66accuracy: 0.66accuracy: 0.68accuracy: 0.7accuracy: 0.7accuracy: 0.7accuracy: 0.7accuracy: 0.72accuracy: 0.74accuracy: 0.76accuracy: 0.78accuracy: 0.78accuracy: 0.8accuracy: 0.8accuracy: 0.82accuracy: 0.82accuracy: 0.82accuracy: 0.84accuracy: 0.84accuracy: 0.84accuracy: 0.84Process finished with exit code 0

对于使用下面的式子当作损失函数不太理解的:

tf.nn.softmax_cross_entropy_with_logits

请看这篇随笔:https://www.cnblogs.com/TimVerion/p/11237087.html

原文地址：https://www.cnblogs.com/TimVerion/p/11237063.html