目录
神经网络案例
学习目标
- 能够利用tf.keras获取数据集
- 能够构建多层神经网络
[En]
able to build a multi-layer neural network*
- 能够完成网络培训和评估
[En]
be able to complete network training and evaluation*
使用手写数字的MNIST数据集如上图所示,该数据集包含60,000个用于训练的样本和10,000个用于测试的样本,图像是固定大小(28×28像素),其值为0到255。
整个案例的实现流程是:
- 数据加载
- 数据处理
- 模型构建
- 模型训练
- 模型测试
- 模型保存
第一步是导入所需的工具包:
[En]
The first step is to import the required toolkit:
import numpy as np
import matplotlib.pyplot as plt
plt.rcParams['figure.figsize'] = (7,7)
import tensorflow as tf
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, Activation,BatchNormalization
from tensorflow.keras import utils
from tensorflow.keras import regularizers
数据加载
首先加载手写数字图像
nb_classes = 10
(X_train, y_train), (X_test, y_test) = mnist.load_data()
print("训练样本初始维度", X_train.shape)
print("训练样本目标值初始维度", y_train.shape)
结果为:
训练样本初始维度 (60000, 28, 28)
训练样本目标值初始维度 (60000,)
数据展示:
for i in range(9):
plt.subplot(3,3,i+1)
plt.imshow(X_train[i], cmap='gray', interpolation='none')
plt.title("数字{}".format(y_train[i]))
效果如下所示:
数据处理
神经网络中的每个训练样本是一个向量,因此需要对输入进行重塑,使每个28×28的图像成为一个的784维向量。另外,将输入数据进行归一化处理,从0-255调整到0-1。
X_train = X_train.reshape(60000, 784)
X_test = X_test.reshape(10000, 784)
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255
print("训练集:", X_train.shape)
print("测试集:", X_test.shape)
输出为:
训练集: (60000, 784)
测试集: (10000, 784)
此外,我们还需要对目标值进行处理,并将其转换为热编码形式:
[En]
In addition, we also need to process the target value and convert it to the form of hot coding:
实现方法如下所示:
Y_train = utils.to_categorical(y_train, nb_classes)
Y_test = utils.to_categorical(y_test, nb_classes)
模型构建
在这里我们构建只有3层全连接的网络来进行处理:
构建方法如下所示:
model = Sequential()
model.add(Dense(512, input_shape=(784,)))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(Dense(512,kernel_regularizer=regularizers.l2(0.001)))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(Dense(10))
model.add(Activation('softmax'))
我们通过model.summay来看下结果:
Model: "sequential_6"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_13 (Dense) (None, 512) 401920
_________________________________________________________________
activation_8 (Activation) (None, 512) 0
_________________________________________________________________
dropout_7 (Dropout) (None, 512) 0
_________________________________________________________________
dense_14 (Dense) (None, 512) 262656
_________________________________________________________________
batch_normalization (BatchNo (None, 512) 2048
_________________________________________________________________
activation_9 (Activation) (None, 512) 0
_________________________________________________________________
dropout_8 (Dropout) (None, 512) 0
_________________________________________________________________
dense_15 (Dense) (None, 10) 5130
_________________________________________________________________
activation_10 (Activation) (None, 10) 0
=================================================================
Total params: 671,754
Trainable params: 670,730
Non-trainable params: 1,024
_________________________________________________________________
模型编译
设置模型训练使用的损失函数交叉熵损失和优化方法adam,损失函数用来衡量预测值与真实值之间的差异,优化器用来使用损失函数达到最优:
model.compile(loss='categorical_crossentropy', optimizer='adam',metrics=['accuracy'])
模型训练
history = model.fit(X_train, Y_train,
batch_size=128, epochs=4,verbose=1,
validation_data=(X_test, Y_test))
训练过程如下所示:
Epoch 1/4
469/469 [==============================] - 2s 4ms/step - loss: 0.5273 - accuracy: 0.9291 - val_loss: 0.2686 - val_accuracy: 0.9664
Epoch 2/4
469/469 [==============================] - 2s 4ms/step - loss: 0.2213 - accuracy: 0.9662 - val_loss: 0.1672 - val_accuracy: 0.9720
Epoch 3/4
469/469 [==============================] - 2s 4ms/step - loss: 0.1528 - accuracy: 0.9734 - val_loss: 0.1462 - val_accuracy: 0.9735
Epoch 4/4
469/469 [==============================] - 2s 4ms/step - loss: 0.1313 - accuracy: 0.9768 - val_loss: 0.1292 - val_accuracy: 0.9777
将损失绘制成曲线:
plt.figure()
plt.plot(history.history["loss"], label="train_loss")
plt.plot(history.history["val_loss"], label="val_loss")
plt.legend()
plt.grid()
将训练的精确度绘制为曲线:
[En]
Draw the accuracy of the training as a curve:
plt.figure()
plt.plot(history.history["accuracy"], label="train_acc")
plt.plot(history.history["val_accuracy"], label="val_acc")
plt.legend()
plt.grid()
另外可通过tensorboard监控训练过程,这时我们指定回调函数:
tensorboard = tf.keras.callbacks.TensorBoard(log_dir='./graph', histogram_freq=1,
write_graph=True,write_images=True)
在进行训练:
history = model.fit(X_train, Y_train,
batch_size=128, epochs=4,verbose=1,callbacks=[tensorboard],
validation_data=(X_test, Y_test))
打开终端:
tensorboard --logdir="./"
在浏览器中打开指定的URL,即可查看损耗函数和精度、图形结构等方面的变化。
[En]
Open the specified URL in the browser to view the changes in loss function and accuracy, graph structure, etc.
模型测试
score = model.evaluate(X_test, Y_test, verbose=1)
print('测试集准确率:', score)
结果:
313/313 [==============================] - 0s 1ms/step - loss: 0.1292 - accuracy: 0.9777
Test accuracy: 0.9776999950408936
模型保存
model.save('my_model.h5')
model = tf.keras.models.load_model('my_model.h5')
总结
- 能够利用tf.keras获取数据集:
load_data()
- 能够构建多层神经网络
[En]
able to build multi-layer neural networks*
dense,激活函数,dropout,BN层等
- 能够完成网络培训和评估
[En]
be able to complete network training and evaluation*
fit,回调函数,evaluate, 保存模型
Original: https://blog.csdn.net/qq_43966129/article/details/123030127
Author: 最白の白菜
Title: 神经网络案例
原创文章受到原创版权保护。转载请注明出处:https://www.johngo689.com/512008/
转载文章受原作者版权保护。转载请注明原作者出处!