深度学习基础之《TensorFlow框架（16）—神经网络案例》

csj50

已于 2024-05-06 22:27:49 修改

阅读量437

点赞数 3

分类专栏：机器学习文章标签：深度学习

于 2024-04-29 14:51:38 首次发布

本文链接：https://blog.csdn.net/csj50/article/details/138313086

版权

机器学习专栏收录该内容

63 篇文章 6 订阅

订阅专栏

一、mnist手写数字识别

1、数据集介绍
mnist数据集是一个经典的数据集，其中包括70000个样本，包括60000个训练样本和10000个测试样本

2、下载地址：http://yann.lecun.com/exdb/mnist/

3、文件说明
train-images-idx3-ubyte.gz: training set images (9912422 bytes)
train-labels-idx1-ubyte.gz: training set labels (28881 bytes)
t10k-images-idx3-ubyte.gz: test set images (1648877 bytes)
t10k-labels-idx1-ubyte.gz: test set labels (4542 bytes)

4、特征值
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

5、目标值
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

二、mnist数据获取API

1、tensorflow框架自带了获取这个数据集的接口，所以不需要自行读取

2、数据集的引入
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
以上四个数据集的格式都为numpy.ndarray

三、案例实战

1、网络设计
我们采用只有一层，即最后，一个输出层的神经网络，也称之为全连接（full connected）层神经网络

一个样本有784个特征，每一个特征都有一个权重值，进行加权求和，加上偏置得到logits，然后给logits加上softmax映射，就会得出一组概率值，概率值当中最大的那个就是它的分类

2、全连接层计算
全连接做的事情就是矩阵相乘

tf.experimental.numpy.matmul(x1, x2) + bias
说明：
（1）return：全连接结果，供交叉损失运算

tf.keras.optimizers.SGD(learning_rate)
说明：
（1）梯度下降
（2）learning_rate：学习率

3、形状表示
x[None, 784] * w[784, 10] + bias[None, 10] = y[None, 10]
x是None行784列，y是None行10列。所以权重的形状是784行10列，偏置是None行10列

四、使用全连接神经网络

原来的代码改写起来比较难，找到一个用keras框架来实现的（（。・＿・。）ﾉ）

import os
os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
from tensorflow.keras.utils import to_categorical
from tensorflow.keras import models, layers, regularizers
from tensorflow.keras.optimizers import RMSprop, SGD
from tensorflow.keras.datasets import mnist
import matplotlib.pyplot as plt

# 加载数据集
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

# 将图片由二维铺开成一维
train_images = train_images.reshape((60000, 28*28)).astype('float')
test_images = test_images.reshape((10000, 28*28)).astype('float')
train_labels = to_categorical(train_labels)
test_labels = to_categorical(test_labels)

# 搭建一个神经网络
network = models.Sequential()
network.add(layers.Dense(units=128, activation='relu', input_shape=(28*28, ),
                         kernel_regularizer=regularizers.l1(0.0001)))
network.add(layers.Dropout(0.01))
network.add(layers.Dense(units=32, activation='relu',
                         kernel_regularizer=regularizers.l1(0.0001)))
network.add(layers.Dropout(0.01))
network.add(layers.Dense(units=10, activation='softmax'))

# 编译步骤，确定优化器和损失函数等
network.compile(optimizer=SGD(learning_rate=0.001), loss='categorical_crossentropy', metrics=['accuracy'])

# 训练网络，用fit函数, epochs表示训练多少个回合， batch_size表示每次训练给多大的数据
network.fit(train_images, train_labels, epochs=100, batch_size=128, verbose=2)

# 来在测试集上测试一下模型的性能吧
test_loss, test_accuracy = network.evaluate(test_images, test_labels)
print("test_loss:", test_loss, " test_accuracy:", test_accuracy)

优化器还是用SGD，训练100个回合，运行结果：

Epoch 1/100
469/469 - 2s - loss: 3.0283 - accuracy: 0.4604
Epoch 2/100
469/469 - 1s - loss: 1.6157 - accuracy: 0.6435
Epoch 3/100
469/469 - 1s - loss: 1.4475 - accuracy: 0.6979
Epoch 4/100
469/469 - 1s - loss: 1.3438 - accuracy: 0.7312
Epoch 5/100
469/469 - 1s - loss: 1.2487 - accuracy: 0.7737
Epoch 6/100
469/469 - 1s - loss: 1.1552 - accuracy: 0.7911
Epoch 7/100
469/469 - 1s - loss: 1.0796 - accuracy: 0.8014
Epoch 8/100
469/469 - 1s - loss: 1.0300 - accuracy: 0.8191
Epoch 9/100
469/469 - 1s - loss: 0.9780 - accuracy: 0.8377
Epoch 10/100
469/469 - 1s - loss: 0.9479 - accuracy: 0.8526
Epoch 11/100
469/469 - 1s - loss: 0.9112 - accuracy: 0.8642
Epoch 12/100
469/469 - 1s - loss: 0.8805 - accuracy: 0.8732
Epoch 13/100
469/469 - 1s - loss: 0.8512 - accuracy: 0.8813
Epoch 14/100
469/469 - 1s - loss: 0.8243 - accuracy: 0.8887
Epoch 15/100
469/469 - 1s - loss: 0.8009 - accuracy: 0.8946
Epoch 16/100
469/469 - 1s - loss: 0.7770 - accuracy: 0.9017
Epoch 17/100
469/469 - 1s - loss: 0.7567 - accuracy: 0.9067
Epoch 18/100
469/469 - 1s - loss: 0.7454 - accuracy: 0.9101
Epoch 19/100
469/469 - 1s - loss: 0.7297 - accuracy: 0.9143
Epoch 20/100
469/469 - 1s - loss: 0.7174 - accuracy: 0.9175
Epoch 21/100
469/469 - 1s - loss: 0.7022 - accuracy: 0.9208
Epoch 22/100
469/469 - 1s - loss: 0.6973 - accuracy: 0.9218
Epoch 23/100
469/469 - 1s - loss: 0.6852 - accuracy: 0.9249
Epoch 24/100
469/469 - 1s - loss: 0.6768 - accuracy: 0.9276
Epoch 25/100
469/469 - 1s - loss: 0.6664 - accuracy: 0.9302
Epoch 26/100
469/469 - 1s - loss: 0.6624 - accuracy: 0.9301
Epoch 27/100
469/469 - 1s - loss: 0.6568 - accuracy: 0.9326
Epoch 28/100
469/469 - 1s - loss: 0.6479 - accuracy: 0.9340
Epoch 29/100
469/469 - 1s - loss: 0.6417 - accuracy: 0.9362
Epoch 30/100
469/469 - 1s - loss: 0.6381 - accuracy: 0.9361
Epoch 31/100
469/469 - 1s - loss: 0.6304 - accuracy: 0.9391
Epoch 32/100
469/469 - 1s - loss: 0.6277 - accuracy: 0.9395
Epoch 33/100
469/469 - 1s - loss: 0.6216 - accuracy: 0.9416
Epoch 34/100
469/469 - 1s - loss: 0.6169 - accuracy: 0.9431
Epoch 35/100
469/469 - 1s - loss: 0.6134 - accuracy: 0.9429
Epoch 36/100
469/469 - 1s - loss: 0.6095 - accuracy: 0.9442
Epoch 37/100
469/469 - 1s - loss: 0.6080 - accuracy: 0.9446
Epoch 38/100
469/469 - 1s - loss: 0.6001 - accuracy: 0.9468
Epoch 39/100
469/469 - 1s - loss: 0.5983 - accuracy: 0.9471
Epoch 40/100
469/469 - 1s - loss: 0.5961 - accuracy: 0.9477
Epoch 41/100
469/469 - 1s - loss: 0.5902 - accuracy: 0.9484
Epoch 42/100
469/469 - 1s - loss: 0.5862 - accuracy: 0.9495
Epoch 43/100
469/469 - 1s - loss: 0.5848 - accuracy: 0.9505
Epoch 44/100
469/469 - 1s - loss: 0.5785 - accuracy: 0.9505
Epoch 45/100
469/469 - 1s - loss: 0.5777 - accuracy: 0.9518
Epoch 46/100
469/469 - 1s - loss: 0.5727 - accuracy: 0.9531
Epoch 47/100
469/469 - 1s - loss: 0.5711 - accuracy: 0.9533
Epoch 48/100
469/469 - 1s - loss: 0.5684 - accuracy: 0.9535
Epoch 49/100
469/469 - 1s - loss: 0.5653 - accuracy: 0.9548
Epoch 50/100
469/469 - 1s - loss: 0.5645 - accuracy: 0.9546
Epoch 51/100
469/469 - 1s - loss: 0.5602 - accuracy: 0.9556
Epoch 52/100
469/469 - 1s - loss: 0.5563 - accuracy: 0.9569
Epoch 53/100
469/469 - 1s - loss: 0.5558 - accuracy: 0.9569
Epoch 54/100
469/469 - 1s - loss: 0.5548 - accuracy: 0.9568
Epoch 55/100
469/469 - 1s - loss: 0.5525 - accuracy: 0.9569
Epoch 56/100
469/469 - 1s - loss: 0.5486 - accuracy: 0.9585
Epoch 57/100
469/469 - 1s - loss: 0.5464 - accuracy: 0.9588
Epoch 58/100
469/469 - 1s - loss: 0.5466 - accuracy: 0.9589
Epoch 59/100
469/469 - 1s - loss: 0.5449 - accuracy: 0.9598
Epoch 60/100
469/469 - 1s - loss: 0.5379 - accuracy: 0.9607
Epoch 61/100
469/469 - 1s - loss: 0.5393 - accuracy: 0.9606
Epoch 62/100
469/469 - 1s - loss: 0.5360 - accuracy: 0.9614
Epoch 63/100
469/469 - 1s - loss: 0.5346 - accuracy: 0.9613
Epoch 64/100
469/469 - 1s - loss: 0.5316 - accuracy: 0.9623
Epoch 65/100
469/469 - 1s - loss: 0.5311 - accuracy: 0.9626
Epoch 66/100
469/469 - 1s - loss: 0.5305 - accuracy: 0.9623
Epoch 67/100
469/469 - 1s - loss: 0.5260 - accuracy: 0.9629
Epoch 68/100
469/469 - 1s - loss: 0.5231 - accuracy: 0.9642
Epoch 69/100
469/469 - 1s - loss: 0.5241 - accuracy: 0.9637
Epoch 70/100
469/469 - 1s - loss: 0.5203 - accuracy: 0.9653
Epoch 71/100
469/469 - 1s - loss: 0.5182 - accuracy: 0.9653
Epoch 72/100
469/469 - 1s - loss: 0.5173 - accuracy: 0.9650
Epoch 73/100
469/469 - 1s - loss: 0.5148 - accuracy: 0.9663
Epoch 74/100
469/469 - 1s - loss: 0.5122 - accuracy: 0.9671
Epoch 75/100
469/469 - 1s - loss: 0.5124 - accuracy: 0.9660
Epoch 76/100
469/469 - 1s - loss: 0.5113 - accuracy: 0.9667
Epoch 77/100
469/469 - 1s - loss: 0.5129 - accuracy: 0.9665
Epoch 78/100
469/469 - 1s - loss: 0.5080 - accuracy: 0.9677
Epoch 79/100
469/469 - 1s - loss: 0.5068 - accuracy: 0.9679
Epoch 80/100
469/469 - 1s - loss: 0.5060 - accuracy: 0.9685
Epoch 81/100
469/469 - 1s - loss: 0.5025 - accuracy: 0.9692
Epoch 82/100
469/469 - 1s - loss: 0.5015 - accuracy: 0.9691
Epoch 83/100
469/469 - 1s - loss: 0.5017 - accuracy: 0.9689
Epoch 84/100
469/469 - 1s - loss: 0.5011 - accuracy: 0.9690
Epoch 85/100
469/469 - 1s - loss: 0.4996 - accuracy: 0.9687
Epoch 86/100
469/469 - 1s - loss: 0.4965 - accuracy: 0.9696
Epoch 87/100
469/469 - 1s - loss: 0.4971 - accuracy: 0.9694
Epoch 88/100
469/469 - 1s - loss: 0.4927 - accuracy: 0.9701
Epoch 89/100
469/469 - 1s - loss: 0.4923 - accuracy: 0.9704
Epoch 90/100
469/469 - 1s - loss: 0.4917 - accuracy: 0.9707
Epoch 91/100
469/469 - 1s - loss: 0.4929 - accuracy: 0.9709
Epoch 92/100
469/469 - 1s - loss: 0.4880 - accuracy: 0.9717
Epoch 93/100
469/469 - 1s - loss: 0.4865 - accuracy: 0.9717
Epoch 94/100
469/469 - 1s - loss: 0.4850 - accuracy: 0.9719
Epoch 95/100
469/469 - 1s - loss: 0.4871 - accuracy: 0.9712
Epoch 96/100
469/469 - 1s - loss: 0.4851 - accuracy: 0.9721
Epoch 97/100
469/469 - 1s - loss: 0.4830 - accuracy: 0.9721
Epoch 98/100
469/469 - 1s - loss: 0.4811 - accuracy: 0.9731
Epoch 99/100
469/469 - 1s - loss: 0.4810 - accuracy: 0.9728
Epoch 100/100
469/469 - 1s - loss: 0.4789 - accuracy: 0.9732
313/313 [==============================] - 1s 1ms/step - loss: 0.6132 - accuracy: 0.9521
test_loss: 0.613166332244873  test_accuracy: 0.9520999789237976

参考资料：
https://cloud.tencent.com/developer/article/2055005

csj50

关注

3
点赞
踩
10

收藏

觉得还不错? 一键收藏
0
评论
深度学习基础之《TensorFlow框架（16）—神经网络案例》

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。mnist数据集是一个经典的数据集，其中包括70000个样本，包括60000个训练样本和10000个测试样本。2、下载地址：http://yann.lecun.com/exdb/mnist/一、mnist手写数字识别。
复制链接

扫一扫