深度学习 Python TensorFlow框架下用Keras搭建神经网络（学习笔记）

from keras.models import Sequential
from keras.layers import *
from keras.utils import np_utils
import numpy as np
from keras.datasets import mnist

#加载mnist数据集
(x_train,y_train),(x_test,y_test)=mnist.load_data()

#数据预处理，转换为Keras使用的格式
num_pixels=x_train.shape[1]*x_train.shape[2]
n_channels=1#图像通道数

def preprocess(matrix):
    return matrix.reshape(matrix.shape[0],n_channels,matrix.shape[1],matrix.shape[2]).astype('float32')/255

x_train,x_test=preprocess(x_train),preprocess(x_test)

#对输出结果进行处理
#np_utils.to_categorical()整型的类别标签转为onehot编码
y_train=np_utils.to_categorical(y_train)
y_test=np_utils.to_categorical(y_test)
num_classes=y_train.shape[1]

#建立一个Sequential模型，基准模型
def baseline_model():
    model=Sequential()
    model.add(Flatten(input_shape=(1,28,28)))#将输入图像拉伸为向量
    model.add(Dense(num_pixels,init='normal',activation='relu'))#输入层relu函数作为激活函数
    model.add(Dense(num_classes,init='normal',activation='softmax'))#输出层softmax函数作为分类函数
    #用Keras编译模型，类的交叉熵作为优化损失函数，分类精度作为主要性能指标
    model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
    return model 

#为了展示模型提供的效果，构建一个复杂的卷积神经网络
def convolution_small():
    model=Sequential()
    #操作在二维矩阵上的卷积滤波器：使用窗口为5*5的滤波器，对二维图像进行卷积滤波操作，产生32维输出向量
    model.add(Convolution2D(32,5,5,border_mode='valid',input_shape=(1,28,28),activation='relu'))
    #最大池化层：对2*2窗口进行最大化选择，以非线性的方式对图像进行采样
    model.add(MaxPool2D(pool_size=(2,2)))
    #dropout层：随机将神经元的20%重置为0，防止过拟合
    model.add(Dropout(0.2))
    model.add(Flatten())
    model.add(Dense(128,activation='relu'))
    model.add(Dense(num_classes,activation='softmax'))
    model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
    return model

#为了展示神经网络的威力，创建一个更复杂的神经网络，两层卷积和池化
def convolution_large():
    model=Sequential()
    model.add(Convolution2D(30,5,5,border_mode='valid',input_shape=(1,28,28),activation='relu'))
    model.add(MaxPool2D(pool_size=(2,2)))
    model.add(Convolution2D(15,3,3,activation='relu'))
    model.add(MaxPool2D(pool_size=(2,2)))
    model.add(Dropout(0.2))
    model.add(Flatten())
    model.add(Dense(128,activation='relu'))
    model.add(Dense(50,activation='relu'))
    model.add(Dense(num_classes,activation='softmax'))
    model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
    return model

#测试模型
np.random.seed(101)
models=[('baseline',baseline_model()),('small',convolution_small()),('large',convolution_large())]
for name,model in models:
    print('With name:',name)
    #训练模型,nb_epoch迭代次数，batch_size批量大小
    model.fit(x_train,y_train,validation_data=(x_test,y_test),nb_epoch=10,batch_size=100,verbose=2)
    #测试模型
    scores=model.evaluate(x_test,y_test,verbose=0)
    print('Baseline Error:%.2f%%'%(100-scores[1]*100))
    print()

With name: baseline
Train on 60000 samples, validate on 10000 samples
Epoch 1/10

15s - loss: 0.2275 - accuracy: 0.9348 - val_loss: 0.1177 - val_accuracy: 0.9647
Epoch 2/10
15s - loss: 0.0884 - accuracy: 0.9738 - val_loss: 0.0770 - val_accuracy: 0.9751
Epoch 3/10
16s - loss: 0.0549 - accuracy: 0.9837 - val_loss: 0.0776 - val_accuracy: 0.9758
Epoch 4/10
15s - loss: 0.0362 - accuracy: 0.9891 - val_loss: 0.0643 - val_accuracy: 0.9786
Epoch 5/10
15s - loss: 0.0269 - accuracy: 0.9922 - val_loss: 0.0586 - val_accuracy: 0.9822
Epoch 6/10
15s - loss: 0.0195 - accuracy: 0.9941 - val_loss: 0.0764 - val_accuracy: 0.9766
Epoch 7/10
15s - loss: 0.0147 - accuracy: 0.9960 - val_loss: 0.0598 - val_accuracy: 0.9816
Epoch 8/10
15s - loss: 0.0101 - accuracy: 0.9974 - val_loss: 0.0608 - val_accuracy: 0.9824
Epoch 9/10
15s - loss: 0.0106 - accuracy: 0.9968 - val_loss: 0.0659 - val_accuracy: 0.9813
Epoch 10/10
15s - loss: 0.0108 - accuracy: 0.9967 - val_loss: 0.0606 - val_accuracy: 0.9836
Baseline Error:1.64%

With name: small
Train on 60000 samples, validate on 10000 samples
Epoch 1/10

76s - loss: 0.1826 - accuracy: 0.9468 - val_loss: 0.0601 - val_accuracy: 0.9813
Epoch 2/10
76s - loss: 0.0617 - accuracy: 0.9807 - val_loss: 0.0422 - val_accuracy: 0.9858
Epoch 3/10
76s - loss: 0.0428 - accuracy: 0.9864 - val_loss: 0.0459 - val_accuracy: 0.9858
Epoch 4/10
77s - loss: 0.0344 - accuracy: 0.9889 - val_loss: 0.0411 - val_accuracy: 0.9873
Epoch 5/10
76s - loss: 0.0273 - accuracy: 0.9911 - val_loss: 0.0302 - val_accuracy: 0.9903
Epoch 6/10
76s - loss: 0.0204 - accuracy: 0.9937 - val_loss: 0.0334 - val_accuracy: 0.9891
Epoch 7/10
75s - loss: 0.0179 - accuracy: 0.9938 - val_loss: 0.0344 - val_accuracy: 0.9898
Epoch 8/10
76s - loss: 0.0159 - accuracy: 0.9947 - val_loss: 0.0343 - val_accuracy: 0.9886
Epoch 9/10
76s - loss: 0.0117 - accuracy: 0.9961 - val_loss: 0.0372 - val_accuracy: 0.9882
Epoch 10/10
75s - loss: 0.0100 - accuracy: 0.9968 - val_loss: 0.0350 - val_accuracy: 0.9892
Baseline Error:1.08%

With name: large
Train on 60000 samples, validate on 10000 samples
Epoch 1/10

75s - loss: 0.2747 - accuracy: 0.9141 - val_loss: 0.0682 - val_accuracy: 0.9771
Epoch 2/10
78s - loss: 0.0784 - accuracy: 0.9758 - val_loss: 0.0439 - val_accuracy: 0.9844
Epoch 3/10
77s - loss: 0.0592 - accuracy: 0.9815 - val_loss: 0.0377 - val_accuracy: 0.9880
Epoch 4/10
79s - loss: 0.0479 - accuracy: 0.9847 - val_loss: 0.0298 - val_accuracy: 0.9910
Epoch 5/10
76s - loss: 0.0410 - accuracy: 0.9862 - val_loss: 0.0264 - val_accuracy: 0.9914
Epoch 6/10
75s - loss: 0.0379 - accuracy: 0.9875 - val_loss: 0.0271 - val_accuracy: 0.9912
Epoch 7/10
76s - loss: 0.0304 - accuracy: 0.9906 - val_loss: 0.0278 - val_accuracy: 0.9913
Epoch 8/10
75s - loss: 0.0288 - accuracy: 0.9906 - val_loss: 0.0242 - val_accuracy: 0.9921
Epoch 9/10
76s - loss: 0.0263 - accuracy: 0.9914 - val_loss: 0.0260 - val_accuracy: 0.9924
Epoch 10/10
78s - loss: 0.0252 - accuracy: 0.9919 - val_loss: 0.0251 - val_accuracy: 0.9927
Baseline Error:0.73%

说明：
书本上是Theano框架，但我在input_shape里遇到了输入错误，解决了半天没解决，就又改回TensorFlow框架了。然后将c盘中keras的配置文件"image_data_format": “channels_last"改为了"image_data_format”: “channels_first”（如下），程序中的input_shape与之对应，问题解决了。
在这里插入图片描述

来源：CSDN

作者：Monica_Zzz

链接：https://blog.csdn.net/weixin_44819497/article/details/103688282

标签

深度学习框架

test

tensorflow