循环神经网络

1.概述

循环神经网络:$RNN$
它是一种自监督的学习方式
它处理序列的方式为:便利所有序列元素,并保存一个状态,其中包含已查看内容相关的信息.即当前时刻的输出不仅与当前输入有关,而且与前一时刻的状态(输出)有关,$H_t=activeFunc(W_{hx}\ast X_t+W_{hh}\ast H_{t-1}+b)$

2.用numpy实现RNN

import numpy as np
timeSteps=100
inFeatures=32
outFeatures=64
inX=np.random.randn(timeSteps,inFeatures)
Whx=np.random.randn(outFeatures,inFeatures)
Whh=np.random.randn(outFeatures,outFeatures)
b=np.random.randn(outFeatures,)
state_t=np.zeros((outFeatures,))  #状态变量初始化
activeFunc=np.tanh #激活函数
outResult=[]
for X in inX:
    outX=activeFunc(np.dot(Whx,X)+np.dot(Whh,state_t)+b)
    outResult.append(outX)
    state_t=outX

result=np.stack(outResult,axis=0)

print(result.shape)  #  (timeSteps,outFeatures)

3.Keras实现RNN

keras.layers.SimpleRNN:
输入数据:[batchSize,timeSteps,inputFeatures]
输出:[batchSize,timeSteps,outputFeatures]
中间每个输出(状态变量):[batchSize,outputFeatures]
这两种输出的控制模式由参数return_sequences来控制

• return_sequences=True,返回[batchSize,timeSteps,outputFeatures]
• return_sequences=False,返回[batchSize,outputFeatures]
  如果对多个RNN层进行堆叠,那么每个中间层都要返回完整的输出序列

from tensorflow.keras.layers import Embedding,SimpleRNN
from tensorflow.keras.models import Sequential

model=Sequential()
model.add(Embedding(10000,32))
model.add(SimpleRNN(32,return_sequences=True))
model.summary()

from tensorflow.keras.layers import Embedding,SimpleRNN
from tensorflow.keras.models import Sequential

model=Sequential()
model.add(Embedding(10000,32))
model.add(SimpleRNN(32))
model.summary()

4.示例

"""将RNN用于IMDB电影评论分类"""
#1.数据预处理

from tensorflow.keras.datasets import imdb
from tensorflow.keras.preprocessing import text,sequence

max_words=10000 #词典中单词数
max_len=500  #每条句子中的最大单词数


(train_data,train_labels),(test_data,test_labels)=imdb.load_data(num_words=10000)

train_data=sequence.pad_sequences(train_data,maxlen=max_len)
test_data=sequence.pad_sequences(test_data,maxlen=max_len)

print("train Data Shape:",train_data.shape,"train label Shape:",train_labels.shape)
print("test Data Shape:",test_data.shape,"test label Shape:",test_labels.shape)

#2.构建网络
model=Sequential()
model.add(Embedding(max_words,32))
model.add(SimpleRNN(32))
model.add(Dense(1,activation="sigmoid"))

#3.配置学习过程
from tensorflow.keras import optimizers,losses
model.compile(optimizer=optimizers.RMSprop(1e-3),loss=losses.binary_crossentropy,metrics=["acc"])

#4.训练
history=model.fit(train_data,train_labels,batch_size=128,epochs=10,validation_split=0.2)  #0.2比例划分为验证集

#5.绘制
import matplotlib.pyplot as plt
import numpy as np
"""损失函数绘制"""
history_dict=history.history
train_loss_values=history_dict["loss"]
val_loss_values=history_dict["val_loss"]
epochs=np.arange(1,len(train_loss_values)+1)

plt.figure(1)
plt.plot(epochs,train_loss_values,label="training loss")
plt.plot(epochs,val_loss_values,label="val loss")
plt.title("Training and Val Loss")
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.legend()

"""精度(准确率)绘制"""
train_acc=history_dict["acc"]
val_acc=history_dict["val_acc"]

plt.figure(2)
plt.plot(epochs,train_acc,label="Training Acc")
plt.plot(epochs,val_acc,label="Val Acc")
plt.title("Training and Val Acc")
plt.xlabel("Epoch")
plt.ylabel("Acc")
plt.legend()

plt.show()

从上面的例子中，我们可以发现应用RNN的效果并不是很好；RNN不擅长于处理长序列问题，因为随着序列长度的增加，其可能会出现梯度消失的问题；为了解决梯度小时问题，人们设计出了LSTM与GRU

5 LSTM

LSTM:Long Short Time Memory
LSTM单元的作用:允许过去的信息稍后重新进入，从而解决梯度消失问题
相关理论可以参考:https://www.jianshu.com/p/9dc9f41f0b29

使用LSTM进行IMDB数据集分类

#1.数据预处理

from tensorflow.keras.datasets import imdb
from tensorflow.keras.preprocessing import text,sequence

max_words=10000 #词典中单词数
max_len=500  #每条句子中的最大单词数


(train_data,train_labels),(test_data,test_labels)=imdb.load_data(num_words=10000)

train_data=sequence.pad_sequences(train_data,maxlen=max_len)
test_data=sequence.pad_sequences(test_data,maxlen=max_len)

print("train Data Shape:",train_data.shape,"train label Shape:",train_labels.shape)
print("test Data Shape:",test_data.shape,"test label Shape:",test_labels.shape)

#2.构建网络
model=Sequential()
model.add(Embedding(max_words,32))
model.add(LSTM(32))
model.add(Dense(1,activation="sigmoid"))

#3.配置学习过程
from tensorflow.keras import optimizers,losses
model.compile(optimizer=optimizers.RMSprop(1e-3),loss=losses.binary_crossentropy,metrics=["acc"])

#4.训练
history=model.fit(train_data,train_labels,batch_size=128,epochs=10,validation_split=0.2)  #0.2比例划分为验证集

#5.绘制
import matplotlib.pyplot as plt
import numpy as np
"""损失函数绘制"""
history_dict=history.history
train_loss_values=history_dict["loss"]
val_loss_values=history_dict["val_loss"]
epochs=np.arange(1,len(train_loss_values)+1)

plt.figure(1)
plt.plot(epochs,train_loss_values,label="training loss")
plt.plot(epochs,val_loss_values,label="val loss")
plt.title("Training and Val Loss")
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.legend()

"""精度(准确率)绘制"""
train_acc=history_dict["acc"]
val_acc=history_dict["val_acc"]

plt.figure(2)
plt.plot(epochs,train_acc,label="Training Acc")
plt.plot(epochs,val_acc,label="Val Acc")
plt.title("Training and Val Acc")
plt.xlabel("Epoch")
plt.ylabel("Acc")
plt.legend()

plt.show()

观察以上结果，发现使用LSTM在验证集上的准确率可以达到85%以上,而使用简单的RNN验证集上的准确率只有80%左右。