theano学习笔记(2)基础函数
1、随机函数库的调用
2、卷积神经网络
- #-*-coding:utf-8-*-
- import theano
- import numpy as np
- import matplotlib.pyplot as plt
- from loaddata import loadmnist
- import theano.tensor as T
- #softmax函数
- class softmax:
- #outdata为我们标注的输出,hiddata网络输出层的输入,nin,nout为输入、输出神经元个数
- def __init__(self,hiddata,outdata,nin,nout):
- self.w=theano.shared(value=np.zeros((nin,nout),dtype=theano.config.floatX),name='w');
- self.b=theano.shared(value=np.zeros((nout,),dtype=theano.config.floatX),name='b')
- prey=T.nnet.softmax(T.dot(hiddata,self.w)+self.b)#通过softmax函数,得到输出层每个神经元数值(概率)
- self.loss=-T.mean(T.log(prey)[T.arange(outdata.shape[0]),outdata])#损失函数
- self.para=[self.w,self.b]
- self.predict=T.argmax(prey,axis=1)
- self.error=T.mean(T.neq(T.argmax(prey,axis=1),outdata))
- #输入层到隐藏层
- class HiddenLayer:
- def __init__(self,inputx,nin,nout):
- a=np.sqrt(6./(nin+nout))
- ranmatrix=np.random.uniform(-a,a,(nin,nout));
- self.w=theano.shared(value=np.asarray(ranmatrix,dtype=theano.config.floatX),name='w')
- self.b=theano.shared(value=np.zeros((nout,),dtype=theano.config.floatX),name='b')
- self.out=T.tanh(T.dot(inputx,self.w)+self.b)
- self.para=[self.w,self.b]
- #传统三层感知器
- class mlp:
- def __init__(self,nin,nhid,nout):
- x=T.fmatrix('x')
- y=T.ivector('y')
- #å‰å‘
- hlayer=HiddenLayer(x,nin,nhid)
- olayer=softmax(hlayer.out,y,nhid,nout)
- #åå‘
- paras=hlayer.para+olayer.para
- dparas=T.grad(olayer.loss,paras)
- updates=[(para,para-0.1*dpara) for para,dpara in zip(paras,dparas)]
- self.trainfunction=theano.function(inputs=[x,y],outputs=olayer.loss,updates=updates)
- def train(self,trainx,trainy):
- return self.trainfunction(trainx,trainy)
- #卷积神经网络的每一层,包括卷积、池化、激活映射操作
- #img_shape为输入特征图,img_shape=(batch_size,特征图个数,图片宽、高)
- #filter_shape为卷积操作相关参数,filter_shape=(输入特征图个数、输出特征图个数、卷积核的宽、卷积核的高)
- #,这样总共filter的个数为:输入特征图个数*输出特征图个数*卷积核的宽*卷积核的高
- class LeNetConvPoolLayer:
- def __init__(self,inputx,img_shape,filter_shape,poolsize=(2,2)):
- #参数初始化
- assert img_shape[1]==filter_shape[1]
- a=np.sqrt(6./(filter_shape[0]+filter_shape[1]))
- v=np.random.uniform(low=-a,high=a,size=filter_shape)
- wvalue=np.asarray(v,dtype=theano.config.floatX)
- self.w=theano.shared(value=wvalue,name='w')
- bvalue=np.zeros((filter_shape[0],),dtype=theano.config.floatX)
- self.b=theano.shared(value=bvalue,name='b')
- covout=T.nnet.conv2d(inputx,self.w)#卷积操作
- covpool=T.signal.downsample.max_pool_2d(covout,poolsize)#池化操作
- self.out=T.tanh(covpool+self.b.dimshuffle('x', 0, 'x', 'x'))
- self.para=[self.w,self.b]
- trainx,trainy=loadmnist()
- trainx=trainx.reshape(-1,1,28,28)
- batch_size=30
- m=trainx.shape[0]
- ne=m/batch_size
- batchx=T.tensor4(name='batchx',dtype=theano.config.floatX)
- batchy=T.ivector('batchy')
- #
- cov1_layer=LeNetConvPoolLayer(inputx=batchx,img_shape=(batch_size,1,28,28),filter_shape=(20,1,5,5))
- cov2_layer=LeNetConvPoolLayer(inputx=cov1_layer.out,img_shape=(batch_size,20,12,12),filter_shape=(50,20,5,5))
- cov2out=cov2_layer.out.flatten(2)
- hlayer=HiddenLayer(cov2out,4*4*50,500)
- olayer=softmax(hlayer.out,batchy,500,10)
- paras=cov1_layer.para+cov2_layer.para+hlayer.para+olayer.para
- dparas=T.grad(olayer.loss,paras)
- updates=[(para,para-0.1*dpara) for para,dpara in zip(paras,dparas)]
- train_function=theano.function(inputs=[batchx,batchy],outputs=olayer.loss,updates=updates)
- test_function=theano.function(inputs=[batchx,batchy],outputs=[olayer.error,olayer.predict])
- testx,testy=loadmnist(True)
- testx=testx.reshape(-1,1,28,28)
- train_history=[]
- test_history=[]
- for it in range(20):
- sum=0
- for i in range(ne):
- a=trainx[i*batch_size:(i+1)*batch_size]
- loss_train=train_function(trainx[i*batch_size:(i+1)*batch_size],trainy[i*batch_size:(i+1)*batch_size])
- sum=sum+loss_train
- sum=sum/ne
- print 'train_loss:',sum
- test_error,predict=test_function(testx,testy)
- print 'test_error:',test_error
- train_history=train_history+[sum]
- test_history=test_history+[test_error]
- n=len(train_history)
- fig1=plt.subplot(111)
- fig1.set_ylim(0.001,0.2)
- fig1.plot(np.arange(n),train_history,'-')
from: http://blog.csdn.net/hjimce/article/details/46806923