MNIST & Keras保存模型并预测

简介

通过Keras深度学习框架对手写数字数据集MNIST进行简单地模型训练并进行预测。

保存模型

TensorFlow 1.14.0，即Keras支持的最高版本为2.2.5

from keras.datasets import mnist
from keras.utils import np_utils
from keras.models import Sequential
from keras.layers.core import Dense

# 数据集
(X_train, y_train), (X_test, y_test) = mnist.load_data()  # 读取并划分MNIST训练集、测试集

X_train = X_train.reshape(len(X_train), -1)  # 二维变一维
X_test = X_test.reshape(len(X_test), -1)

X_train = X_train.astype('float32')  # 转为float类型
X_test = X_test.astype('float32')

X_train = (X_train - 127) / 127  # 灰度像素数据归一化
X_test = (X_test - 127) / 127

y_train = np_utils.to_categorical(y_train, num_classes=10)  # 独热编码。如原来为5，转换后[0. 0. 0. 0. 0. 1. 0. 0. 0. 0.]
y_test = np_utils.to_categorical(y_test, num_classes=10)

# 定义模型
model = Sequential()  # Keras序列模型

model.add(Dense(20, input_shape=(784,), activation='relu'))  # 添加全连接层（隐藏层），隐藏层数20层，激活函数为ReLU
model.add(Dense(10, activation='sigmoid'))  # 添加输出层，结果10类，激活函数为Sigmoid

print(model.summary())  # 模型基本信息

model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])  # 编译模型

# 训练
model.fit(X_train, y_train, epochs=20, batch_size=64, verbose=1, validation_split=0.05)  # 迭代20次

# 评估
loss, accuracy = model.evaluate(X_test, y_test)
print('Test loss:', loss)
print('Accuracy:', accuracy)

# 保存
model.save('mnistmodel.h5')

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_1 (Dense)              (None, 20)                15700     
_________________________________________________________________
dense_2 (Dense)              (None, 10)                210       
=================================================================
Total params: 15,910
Trainable params: 15,910
Non-trainable params: 0
_________________________________________________________________


Test loss: 0.2107365175232291
Accuracy: 0.938

可以看到保存了一个.h5文件

加载模型并预测

import random
import numpy as np
import matplotlib.pyplot as plt
from keras.datasets import mnist
from keras.models import load_model

# 加载
mymodel = load_model('mnistmodel.h5')

# 数据集
(_, _), (X_test, y_test) = mnist.load_data()  # 划分MNIST训练集、测试集

while True:
    # 随机数
    index = random.randint(0, X_test.shape[0])
    x = X_test[index]
    y = y_test[index]

    # 显示该数字
    plt.imshow(x, cmap='gray_r')
    plt.title("original {}".format(y))
    plt.show()

    # 预测
    x.shape = (1, 784)  # 变成[[]]
    predict = mymodel.predict(x)
    predict = np.argmax(predict)  # 取最大值的位置

    print('index:', index)
    print('original:', y)
    print('predicted:', predict)
    print()

    q = input('回车继续，q退出')
    if q == 'q':
        break

正确预测：

错误预测：

高版本代码

版本

tensorflow-gpu==2.4.0

保存模型

from tensorflow.keras.layers import Dense
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.utils import to_categorical

# 数据集
(X_train, y_train), (X_test, y_test) = mnist.load_data()  # 读取并划分MNIST训练集、测试集

X_train = X_train.reshape(len(X_train), -1)  # 二维变一维
X_test = X_test.reshape(len(X_test), -1)

X_train = X_train.astype('float32')  # 转为float类型
X_test = X_test.astype('float32')

X_train = (X_train - 127) / 127  # 灰度像素数据归一化
X_test = (X_test - 127) / 127

y_train = to_categorical(y_train, num_classes=10)  # 独热编码。如原来为5，转换后[0. 0. 0. 0. 0. 1. 0. 0. 0. 0.]
y_test = to_categorical(y_test, num_classes=10)

# 定义模型
model = Sequential()  # Keras序列模型

model.add(Dense(20, input_shape=(784,), activation='relu'))  # 添加全连接层（隐藏层），隐藏层数20层，激活函数为ReLU
model.add(Dense(10, activation='sigmoid'))  # 添加输出层，结果10类，激活函数为Sigmoid

print(model.summary())  # 模型基本信息

model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])  # 编译模型

# 训练
model.fit(X_train, y_train, epochs=20, batch_size=64, verbose=1, validation_split=0.05)  # 迭代20次

# 评估
loss, accuracy = model.evaluate(X_test, y_test)
print('Test loss:', loss)
print('Accuracy:', accuracy)

# 保存
model.save('mnistmodel.h5')

加载模型并预测

import random
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import load_model

# 加载
mymodel = load_model('mnistmodel.h5')

# 数据集
(_, _), (X_test, y_test) = mnist.load_data()  # 划分MNIST训练集、测试集

while True:
    # 随机数
    index = random.randint(0, X_test.shape[0])
    x = X_test[index]
    y = y_test[index]

    # 显示该数字
    plt.imshow(x, cmap='gray_r')
    plt.title("original {}".format(y))
    plt.show()

    # 预测
    x.shape = (1, 784)  # 变成[[]]
    predict = mymodel.predict(x)
    predict = np.argmax(predict)  # 取最大值的位置

    print('index:', index)
    print('original:', y)
    print('predicted:', predict)
    print()

    q = input('回车继续，q退出')
    if q == 'q':
        break

自定义图像

推荐阅读：

1. Keras实现预训练网络VGG16迁移学习——CIFAR10分类

参考文献

1. Numpy 改变数组维度的几种方法
2. Keras-2 Keras Mnist