视频学习来源

https://www.bilibili.com/video/av40787141?from=search&seid=17003307842787199553

笔记

使用dropout是要改善过拟合,将训练和测试的准确率差距变小

训练集,测试集结果相比差距较大时,过拟合状态

使用dropout后,每一周期准确率可能不高反而最后一步提升很快,这是训练的时候部分神经元工作,而最后的评估所有神经元工作

正则化同样是改善过拟合作用

Softmax一般用在神经网络的最后一层

import numpy as np
from keras.datasets import mnist  #将会从网络下载mnist数据集
from keras.utils import np_utils
from keras.models import Sequential  #序列模型
from keras.layers import Dense,Dropout  #在这里导入dropout
from keras.optimizers import SGD

C:\Program Files (x86)\Microsoft Visual Studio\Shared\Anaconda3_64\lib\site-packages\h5py\__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters
Using TensorFlow backend.

#载入数据
(x_train,y_train),(x_test,y_test)=mnist.load_data()
#查看格式
#(60000,28,28)
print('x_shape:',x_train.shape)
#(60000)
print('y_shape:',y_train.shape)
#(60000,28,28)->(60000,784)
#行数60000,列-1表示自动设置
#除以255是做数据归一化处理
x_train=x_train.reshape(x_train.shape[0],-1)/255.0 #转换数据格式
x_test=x_test.reshape(x_test.shape[0],-1)/255.0 #转换数据格式
#label标签转换成 one  hot 形式
y_train=np_utils.to_categorical(y_train,num_classes=10) #分成10类
y_test=np_utils.to_categorical(y_test,num_classes=10) #分成10类


#创建模型,输入754个神经元,输出10个神经元
#偏执值初始值设为zeros(默认为zeros)
model=Sequential([
    Dense(units=200,input_dim=784,bias_initializer='zeros',activation='tanh'), #双曲正切激活函数
    #Dropout(0.4),  #百分之40的神经元不工作
    Dense(units=100,bias_initializer='zeros',activation='tanh'), #双曲正切激活函数
    #Dropout(0.4),  #百分之40的神经元不工作
    Dense(units=10,bias_initializer='zeros',activation='softmax') 
])

#也可用下面的方式添加网络层
###
#model.add(Dense(...))
#model.add(Dense(...))
###


#定义优化器
#学习速率为0.2
sgd=SGD(lr=0.2)

#定义优化器,损失函数,训练效果中计算准确率
model.compile(
    optimizer=sgd, #sgd优化器
    loss='categorical_crossentropy',  #损失用交叉熵,速度会更快
    metrics=['accuracy'],  #计算准确率
)

#训练(不同于之前,这是新的训练方式)
#六万张,每次训练32张,训练10个周期(六万张全部训练完算一个周期)
model.fit(x_train,y_train,batch_size=32,epochs=10)

#评估模型
loss,accuracy=model.evaluate(x_test,y_test)

print('\ntest loss',loss)
print('\ntest accuracy',accuracy)

loss,accuracy=model.evaluate(x_train,y_train)

print('\ntrain loss',loss)
print('\ntrain accuracy',accuracy)

x_shape: (60000, 28, 28)
y_shape: (60000,)
Epoch 1/10
60000/60000 [==============================] - 6s 100us/step - loss: 0.2539 - acc: 0.9235
Epoch 2/10
60000/60000 [==============================] - 6s 95us/step - loss: 0.1175 - acc: 0.9639
Epoch 3/10
60000/60000 [==============================] - 5s 90us/step - loss: 0.0815 - acc: 0.9745
Epoch 4/10
60000/60000 [==============================] - 5s 90us/step - loss: 0.0601 - acc: 0.9809
Epoch 5/10
60000/60000 [==============================] - 6s 92us/step - loss: 0.0451 - acc: 0.9860
Epoch 6/10
60000/60000 [==============================] - 5s 91us/step - loss: 0.0336 - acc: 0.9899
Epoch 7/10
60000/60000 [==============================] - 5s 92us/step - loss: 0.0248 - acc: 0.9926
Epoch 8/10
60000/60000 [==============================] - 6s 93us/step - loss: 0.0185 - acc: 0.9948
Epoch 9/10
60000/60000 [==============================] - 6s 93us/step - loss: 0.0128 - acc: 0.9970
Epoch 10/10
60000/60000 [==============================] - 6s 93us/step - loss: 0.0082 - acc: 0.9988
10000/10000 [==============================] - 0s 39us/step
 
test loss 0.07058678171953651
 
test accuracy 0.9786
60000/60000 [==============================] - 2s 34us/step
 
train loss 0.0052643890143993
 
train accuracy 0.9995


使用后
(将#Dropout(0.4), 去掉注释)

model=Sequential([
    Dense(units=200,input_dim=784,bias_initializer='zeros',activation='tanh'), #双曲正切激活函数
    Dropout(0.4),  #百分之40的神经元不工作
    Dense(units=100,bias_initializer='zeros',activation='tanh'), #双曲正切激活函数
    Dropout(0.4),  #百分之40的神经元不工作
    Dense(units=10,bias_initializer='zeros',activation='softmax') #双曲正切激活函数
])

x_shape: (60000, 28, 28)
y_shape: (60000,)
Epoch 1/10
60000/60000 [==============================] - 11s 184us/step - loss: 0.4158 - acc: 0.8753
Epoch 2/10
60000/60000 [==============================] - 10s 166us/step - loss: 0.2799 - acc: 0.9177
Epoch 3/10
60000/60000 [==============================] - 11s 177us/step - loss: 0.2377 - acc: 0.9302
Epoch 4/10
60000/60000 [==============================] - 10s 164us/step - loss: 0.2169 - acc: 0.9356
Epoch 5/10
60000/60000 [==============================] - 10s 170us/step - loss: 0.1979 - acc: 0.9413
Epoch 6/10
60000/60000 [==============================] - 11s 183us/step - loss: 0.1873 - acc: 0.9439
Epoch 7/10
60000/60000 [==============================] - 11s 180us/step - loss: 0.1771 - acc: 0.9472
Epoch 8/10
60000/60000 [==============================] - 12s 204us/step - loss: 0.1676 - acc: 0.9501
Epoch 9/10
60000/60000 [==============================] - 11s 187us/step - loss: 0.1608 - acc: 0.9527
Epoch 10/10
60000/60000 [==============================] - 10s 170us/step - loss: 0.1534 - acc: 0.9542
10000/10000 [==============================] - 1s 68us/step
 
test loss 0.09667835112037138
 
test accuracy 0.9692
60000/60000 [==============================] - 4s 70us/step
 
train loss 0.07203661710163578
 
train accuracy 0.9774666666666667


PS 本例并不能很好的体现dropout的优化,但是提供示例来使用dropout

正则化

Kernel_regularizer 权值正则化

Bias_regularizer 偏置正则化

Activity_regularizer 激活正则化

激活正则化是信号乘以权值加上偏置值得到的激活

一般设置权值较多

如果模型对于数据较为复杂,可用dropout和正则化来克服一些过拟合

如果模型对于数据较为简单,可用dropout和正则化可能会降低训练效果

import numpy as np
from keras.datasets import mnist  #将会从网络下载mnist数据集
from keras.utils import np_utils
from keras.models import Sequential  #序列模型
from keras.layers import Dense
from keras.optimizers import SGD  
from keras.regularizers import l2  #导入正则化l2(小写L)

#载入数据
(x_train,y_train),(x_test,y_test)=mnist.load_data()
#查看格式
#(60000,28,28)
print('x_shape:',x_train.shape)
#(60000)
print('y_shape:',y_train.shape)
#(60000,28,28)->(60000,784)
#行数60000,列-1表示自动设置
#除以255是做数据归一化处理
x_train=x_train.reshape(x_train.shape[0],-1)/255.0 #转换数据格式
x_test=x_test.reshape(x_test.shape[0],-1)/255.0 #转换数据格式
#label标签转换成 one  hot 形式
y_train=np_utils.to_categorical(y_train,num_classes=10) #分成10类
y_test=np_utils.to_categorical(y_test,num_classes=10) #分成10类


#创建模型,输入754个神经元,输出10个神经元
#偏执值初始值设为zeros(默认为zeros)
model=Sequential([
    #加上权值正则化
    Dense(units=200,input_dim=784,bias_initializer='zeros',activation='tanh',kernel_regularizer=l2(0.0003)), #双曲正切激活函数
    Dense(units=100,bias_initializer='zeros',activation='tanh',kernel_regularizer=l2(0.0003)), #双曲正切激活函数
    Dense(units=10,bias_initializer='zeros',activation='softmax',kernel_regularizer=l2(0.0003)) 
])

#也可用下面的方式添加网络层
###
#model.add(Dense(...))
#model.add(Dense(...))
###


#定义优化器
#学习速率为0.2
sgd=SGD(lr=0.2)

#定义优化器,损失函数,训练效果中计算准确率
model.compile(
    optimizer=sgd, #sgd优化器
    loss='categorical_crossentropy',  #损失用交叉熵,速度会更快
    metrics=['accuracy'],  #计算准确率
)

#训练(不同于之前,这是新的训练方式)
#六万张,每次训练32张,训练10个周期(六万张全部训练完算一个周期)
model.fit(x_train,y_train,batch_size=32,epochs=10)

#评估模型
loss,accuracy=model.evaluate(x_test,y_test)

print('\ntest loss',loss)
print('\ntest accuracy',accuracy)

loss,accuracy=model.evaluate(x_train,y_train)

print('\ntrain loss',loss)
print('\ntrain accuracy',accuracy)

x_shape: (60000, 28, 28)
y_shape: (60000,)
Epoch 1/10
60000/60000 [==============================] - 8s 127us/step - loss: 0.4064 - acc: 0.9202
Epoch 2/10
60000/60000 [==============================] - 7s 121us/step - loss: 0.2616 - acc: 0.9603
Epoch 3/10
60000/60000 [==============================] - 8s 135us/step - loss: 0.2185 - acc: 0.9683
Epoch 4/10
60000/60000 [==============================] - 8s 132us/step - loss: 0.1950 - acc: 0.9723
Epoch 5/10
60000/60000 [==============================] - 8s 130us/step - loss: 0.1793 - acc: 0.9754
Epoch 6/10
60000/60000 [==============================] - 8s 125us/step - loss: 0.1681 - acc: 0.9775
Epoch 7/10
60000/60000 [==============================] - 8s 130us/step - loss: 0.1625 - acc: 0.9783
Epoch 8/10
60000/60000 [==============================] - 7s 125us/step - loss: 0.1566 - acc: 0.9797
Epoch 9/10
60000/60000 [==============================] - 8s 136us/step - loss: 0.1515 - acc: 0.9811
Epoch 10/10
60000/60000 [==============================] - 8s 140us/step - loss: 0.1515 - acc: 0.9808
10000/10000 [==============================] - 1s 57us/step

test loss 0.17750378291606903

test accuracy 0.9721
60000/60000 [==============================] - 3s 52us/step

train loss 0.1493431808312734

train accuracy 0.9822666666666666