API tools faq
paste
rajeswar

Untitled

rajeswar
Jul 27th, 2014
341
0
Never
Add comment
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
Python 18.19 KB | None | 0 0
raw download clone embed print report
  1. import numpy
  2. import cPickle
  3. from scipy import signal as sg
  4. import math
  5. import random
  6. """
  7. input: 5D(minibatch,number of channels,image depth,image width,image height)
  8. image_shape: same as input
  9. filter_shape:(number of filters,number of channels,filter depth,filter width,filter height)
  10. filter_shape:(1,number of filters,filter depth,filter width,filter height)
  11. pool size:tuple of length 2(default given (3,3))
  12.  
  13. class Hardwiring(object):
  14.    def __init__(self,input,image_shape):
  15.        output=numpy.zeros((image_shape[0],1,43,60,80))
  16.        #first gray
  17.        gray=input
  18.        j=0
  19.        for i in xrange(image_shape[2]):
  20.            output[0][1][j]=image[0][1][i]
  21.            j+=1
  22.        gx=[[-1,0,+1],[-2,0,+2],[-1,0,+1]]
  23.        gy=[[+1,+2,+1],[0,0,0],[-1,-2,-1]]
  24.        gradx=sg.convolve(input,numpy.rot90(gx,2),mode='full')
  25.        grady=sg.convolve(input,numpy.rot90(gy,2),mode='full')
  26.        for i in xrange(image_shape[2]):
  27.            output[0][1][j]=gradx[0][1][i]
  28.            j+=1
  29.        for i in xrange(image_shape[2]):
  30.            output[0][1][j]=grady[0][1][i]
  31.            j+=1
  32.            
  33.        
  34.        
  35.    
  36. """
  37. class First_Convolution(object):
  38.     def __init__(self,rng,input,image_shape,filter_shape,pool_size=(3,3)):
  39.         self.filter_shape=filter_shape
  40.         self.image_shape=image_shape
  41.         self.input=input
  42.         fan_in=numpy.prod(filter_shape[2])
  43.         fan_out=(numpy.prod(filter_shape[1]))/numpy.prod(pool_size)
  44.         conv_out_shape=(image_shape[0],filter_shape[1]*image_shape[1],image_shape[2]-filter_shape[2]+1,image_shape[3]-filter_shape[3]+1,image_shape[4]-filter_shape[4]+1)
  45.         conv_out=numpy.zeros((conv_out_shape))
  46.         self.conv_out_shape=conv_out_shape
  47.         self.conv_out=conv_out
  48.         W_bound=numpy.sqrt(6.0/(fan_in+fan_out))
  49.         W=numpy.asarray(rng.uniform(low=-W_bound,high=W_bound,size=filter_shape),dtype=numpy.float32)
  50.         b=numpy.zeros((conv_out_shape[1],conv_out_shape[2]),dtype=numpy.float32)
  51.         self.W=W
  52.         self.b=b
  53.         for i in xrange(conv_out_shape[1]):
  54.             conv_out[0][i]=sg.fftconvolve(self.input[0][i%self.image_shape[1]],numpy.rot90(self.W[0][i/self.image_shape[1]],2),mode='valid')
  55.         pooled_shape=(image_shape[0],conv_out_shape[1],conv_out_shape[2],conv_out_shape[3]/3,conv_out_shape[4]/3)
  56.         self.pooled_shape=pooled_shape
  57.         pooled=numpy.zeros((pooled_shape))
  58.         self.pooled=pooled
  59.         for i in xrange(pooled_shape[1]):
  60.             for j in xrange(pooled_shape[2]):
  61.                 for k in xrange(pooled_shape[3]):
  62.                     for l in xrange(pooled_shape[4]):
  63.                         pooled[0][i][j][k][l]=math.tanh((numpy.sum(conv_out[0][i][j][k*3][l*3:(l+1)*3])+numpy.sum(conv_out[0][i][j][k*3+1][l*3:(l+1)*3])+numpy.sum(conv_out[0][i][j][k*3+2][l*3:(l+1)*3]))/9.0+b[i][j])
  64.                
  65.         self.output=pooled
  66.     @profile
  67.     def change_input(self,input1):
  68.         self.input=input1
  69.         for i in xrange(self.conv_out_shape[1]):
  70.             self.conv_out[0][i]=sg.fftconvolve(self.input[0][i%self.image_shape[1]],numpy.rot90(self.W[0][i/self.image_shape[1]],2),mode='valid')
  71.         for k in xrange(self.pooled_shape[3]):
  72.              for l in xrange(self.pooled_shape[4]):
  73.                  output1 = numpy.sum(self.conv_out[0, :, :, k*3: (k*3) +1, l*3:(l+1)*3])
  74.                  output1 += numpy.sum(self.conv_out[0, :, :, k*3 + 2, l*3:(l+1)*3])/9.0
  75.                  output1 += self.b
  76.                  self.pooled[0, :, :, k, l] = numpy.tanh(output1)
  77.        #for i in xrange(self.pooled_shape[1]):
  78.          #   for j in xrange(self.pooled_shape[2]):
  79.           #      for k in xrange(self.pooled_shape[3]):
  80.            #         for l in xrange(self.pooled_shape[4]):
  81.             #            self.pooled[0][i][j][k][l]=math.tanh((numpy.sum(self.conv_out[0][i][j][k*3][l*3:(l+1)*3])+numpy.sum(self.conv_out[0][i][j][k*3+1][l*3:(l+1)*3])+numpy.sum(self.conv_out[0][i][j][k*3+2][l*3:(l+1)*3]))/9.0+self.b[i][j])
  82.                
  83.         self.output=self.pooled
  84.     """
  85.    numpy.kron returns array
  86.    """
  87.     @profile
  88.     def gradient(self,eta):
  89.         for i in xrange(len(self.b)):
  90.             for j in xrange(len(self.b[i])):
  91.                 self.b[i][j]-=eta*numpy.asarray(self.delta[0][i][j]).sum()
  92.         delta1=numpy.asarray(numpy.zeros(self.conv_out_shape))
  93.         for i in xrange(self.conv_out_shape[1]):
  94.             for j in xrange(self.conv_out_shape[2]):
  95.                 delta1[0][i][j]=numpy.kron(self.delta[0][i][j],[[1,1,1],[1,1,1],[1,1,1]])/9.0
  96.        
  97.         for i in xrange(self.filter_shape[1]):
  98.             t=numpy.rot90(delta1[0][i],2)
  99.             self.W[0][i]-=sg.fftconvolve(self.input[0][0],t,mode='valid')*eta
  100.            
  101.            
  102. class Hidden_Convolution(object):
  103.     def __init__(self,rng,input1,image_shape,filter_shape,pool_size=(3,3)):
  104.         self.filter_shape=filter_shape
  105.         self.image_shape=image_shape
  106.         self.input=input1
  107.         fan_in=numpy.prod(filter_shape[2])
  108.         fan_out=(numpy.prod(filter_shape[1]))/numpy.prod(pool_size)
  109.         conv_out_shape=(image_shape[0],filter_shape[1]*image_shape[1],image_shape[2]-filter_shape[2]+1,image_shape[3]-filter_shape[3]+1,image_shape[4]-filter_shape[4]+1)
  110.         conv_out=numpy.zeros((conv_out_shape))
  111.         self.conv_out_shape=conv_out_shape
  112.         self.conv_out=conv_out
  113.         W_bound=numpy.sqrt(6.0/(fan_in+fan_out))
  114.         W=numpy.asarray(rng.uniform(low=-W_bound,high=W_bound,size=filter_shape),dtype=numpy.float32)
  115.         b=numpy.zeros((conv_out_shape[1],conv_out_shape[2]),dtype=numpy.float32)
  116.         self.W=W
  117.         self.b=b
  118.         for i in xrange(conv_out_shape[1]):
  119.             conv_out[0][i]=sg.fftconvolve(self.input[0][i%self.image_shape[1]],numpy.rot90(self.W[0][i/self.image_shape[1]],2),mode='valid')
  120.         pooled_shape=(image_shape[0],conv_out_shape[1],conv_out_shape[2],conv_out_shape[3]/3,conv_out_shape[4]/3)
  121.         self.pooled_shape=pooled_shape
  122.         pooled=numpy.zeros((pooled_shape))
  123.         self.pooled=pooled
  124.         for i in xrange(pooled_shape[1]):
  125.             for j in xrange(pooled_shape[2]):
  126.                 for k in xrange(pooled_shape[3]):
  127.                     for l in xrange(pooled_shape[4]):
  128.                         pooled[0][i][j][k][l]=math.tanh((numpy.sum(conv_out[0][i][j][k*3][l*3:(l+1)*3])+numpy.sum(conv_out[0][i][j][k*3+1][l*3:(l+1)*3])+numpy.sum(conv_out[0][i][j][k*3+2][l*3:(l+1)*3]))/9.0+b[i][j])
  129.                
  130.         self.output=pooled
  131.     @profile
  132.     def change_input(self,input1):
  133.         self.input=input1
  134.         for i in xrange(self.conv_out_shape[1]):
  135.             self.conv_out[0][i]=sg.fftconvolve(self.input[0][i%self.image_shape[1]],numpy.rot90(self.W[0][i/self.image_shape[1]],2),mode='valid')
  136.        
  137.         for k in xrange(self.pooled_shape[3]):
  138.        for l in xrange(self.pooled_shape[4]):
  139.                          output1 = numpy.sum(self.conv_out[0, :, :, k*3: (k*3) +1, l*3:(l+1)*3])
  140.                          output1 += numpy.sum(self.conv_out[0, :, :, k*3 + 2, l*3:(l+1)*3])/9.0
  141.                          output1 += self.b
  142.                          self.pooled[0, :, :, k, l] = numpy.tanh(output1)
  143.        # for i in xrange(self.pooled_shape[1]):
  144.         #    for j in xrange(self.pooled_shape[2]):
  145.          #       for k in xrange(self.pooled_shape[3]):
  146.           #          for l in xrange(self.pooled_shape[4]):
  147.            #             self.pooled[0][i][j][k][l]=math.tanh((numpy.sum(self.conv_out[0][i][j][k*3][l*3:(l+1)*3])+numpy.sum(self.conv_out[0][i][j][k*3+1][l*3:(l+1)*3])+numpy.sum(self.conv_out[0][i][j][k*3+2][l*3:(l+1)*3]))/9.0+self.b[i][j])
  148.                
  149.         self.output=self.pooled
  150.     """
  151.    numpy.kron returns array
  152.    inp is object
  153.    """
  154.     @profile
  155.     def gradient(self,inp,eta):
  156.        
  157.         for i in xrange(len(self.b)):
  158.             for j in xrange(len(self.b[i])):
  159.                 self.b[i][j]-=eta*numpy.asarray(self.delta[0][i][j]).sum()
  160.         delta1=numpy.asarray(numpy.zeros(self.conv_out_shape))
  161.         for i in xrange(self.conv_out_shape[1]):
  162.             for j in xrange(self.conv_out_shape[2]):
  163.                 delta1[0][i][j]=numpy.kron(self.delta[0][i][j],[[1,1,1],[1,1,1],[1,1,1]])/9.0
  164.         self.delta1=delta1
  165.        
  166.        
  167.         inp.delta=numpy.zeros(self.image_shape)
  168.         for i in xrange(self.conv_out_shape[1]):
  169.             inp.delta[0][i%self.image_shape[1]]+= sg.fftconvolve(self.delta1[0][i],numpy.rot90(self.W[0][i/self.image_shape[1]],2),'full')
  170.         inp.delta/=(self.filter_shape[1]*1.0)
  171.         for i in xrange(self.image_shape[1]):
  172.             for j in xrange(self.image_shape[2]):
  173.                 for k in xrange(self.image_shape[3]):
  174.                     for l in xrange(self.image_shape[4]):
  175.                         #conv_out[0][i][j][k][l]=de_convolve(inp,i,j,k,l)
  176.                         inp.delta[0][i][j][k][l]*=(1-inp.output[0][i][j][k][l]**2)
  177.         #inp.delta/=self.filter_shape[1]
  178.        
  179.         j=0
  180.         k=0
  181.         m=0
  182.         for i in xrange(self.conv_out_shape[1]):
  183.             ab=numpy.rot90(delta1[0][i],2)
  184.             k+=numpy.asarray(sg.fftconvolve(self.input[0][j],ab,mode='valid'))
  185.             j+=1
  186.             if j==self.image_shape[1]:
  187.                 #k=k/self.image_shape[1]
  188.                 self.W[0][m]-=eta*k
  189.                 m+=1
  190.                 j=0
  191.                 k=0
  192.        
  193.        
  194.     @profile        
  195.     def de_convolve(self,inp,i,j,k,l):
  196.         kernel_number=int(i/self.image_shape[1])
  197.         channel_number=i%self.image_shape[1]
  198.         for a in xrange(self.filter_shape[2]):
  199.             for b in xrange(self.filter_shape[3]):
  200.                 for c in xrange(self.filter_shape[4]):
  201.                     inp.delta[0][channel_number][j+a][k+b][l+c]+=self.delta1[0][i][j][k][l]*self.W[0][kernel_number][a][b][c]
  202.                     #sum=sum+self.W[0][kernel_number][a][b][c]*self.input[0][channel_number][j+a][k+b][l+c]
  203.  
  204.  
  205. class Last_Convolution(object):
  206.     def __init__(self,rng,input1,image_shape,filter_shape,pool_size=(1,1)):
  207.         self.filter_shape=filter_shape
  208.         self.input=input1
  209.         self.image_shape=image_shape
  210.         fan_in=numpy.prod(filter_shape[2])
  211.         fan_out=(numpy.prod(filter_shape[1]))/numpy.prod(pool_size)
  212.         conv_out_shape=(image_shape[0],filter_shape[1]*image_shape[1],image_shape[2]-filter_shape[2]+1,image_shape[3]-filter_shape[3]+1,image_shape[4]-filter_shape[4]+1)
  213.         conv_out=numpy.zeros((conv_out_shape))
  214.         self.conv_out_shape=conv_out_shape
  215.         self.conv_out=conv_out
  216.         W_bound=numpy.sqrt(6.0/(fan_in+fan_out))
  217.         W=numpy.asarray(rng.uniform(low=-W_bound,high=W_bound,size=filter_shape),dtype=numpy.float32)
  218.         b=numpy.zeros((conv_out_shape[1],conv_out_shape[2]),dtype=numpy.float32)
  219.         self.W=W
  220.         self.b=b
  221.         for i in xrange(conv_out_shape[1]):
  222.             conv_out[0][i]=sg.fftconvolve(self.input[0][i%self.image_shape[1]],numpy.rot90(self.W[0][i/self.image_shape[1]],2),mode='valid')
  223.         self.output=numpy.tanh(conv_out)
  224.     def change_input(self,input1):
  225.         self.input=input1
  226.         for i in xrange(self.conv_out_shape[1]):
  227.             self.conv_out[0][i]=sg.fftconvolve(self.input[0][i%self.image_shape[1]],numpy.rot90(self.W[0][i/self.image_shape[1]],2),mode='valid')
  228.         self.output=numpy.tanh(self.conv_out)
  229.        
  230.  
  231.     """
  232.    Parameter is the object of previous layer
  233.        i.e.,inp
  234.    """
  235.     def gradient(self,inp,eta):
  236.         self.eta=eta
  237.         for i in xrange(len(self.b)):
  238.             self.b[i][0]-=eta*self.delta[i]
  239.            
  240.         inp.delta=numpy.zeros(self.image_shape)
  241.         for i in xrange(self.conv_out_shape[1]):
  242.             dummy=[[[]]]
  243.             dummy[0][0].append(self.delta[i])            
  244.             inp.delta[0][i%self.image_shape[1]]+= sg.fftconvolve(dummy,numpy.rot90(self.W[0][i/self.image_shape[1]],2),'full')
  245.         for i in xrange(self.image_shape[1]):
  246.             for j in xrange(self.image_shape[2]):
  247.                 for k in xrange(self.image_shape[3]):
  248.                     for l in xrange(self.image_shape[4]):
  249.                         #conv_out[0][i][j][k][l]=de_convolve(inp,i,j,k,l)
  250.                         inp.delta[0][i][j][k][l]*=(1-inp.output[0][i][j][k][l]**2)
  251.         j=0
  252.         k=0
  253.         m=0
  254.         for i in xrange(self.conv_out_shape[1]):
  255.             ab=[[[self.delta[i]]]]
  256.             k+=numpy.asarray(sg.fftconvolve(self.input[0][j],ab,mode='valid'))
  257.             j+=1
  258.             if j==self.image_shape[1]:
  259.                 #k=k/self.image_shape[1]
  260.                 self.W[0][m]-=self.eta*k
  261.                 m+=1
  262.                 j=0
  263.                 k=0
  264.     @profile
  265.     def de_convolve(self,inp,i,j,k,l):
  266.         kernel_number=int(i/self.image_shape[1])
  267.         channel_number=i%self.image_shape[1]
  268.         # sum=0
  269.         for a in xrange(self.filter_shape[2]):
  270.             for b in xrange(self.filter_shape[3]):
  271.                 for c in xrange(self.filter_shape[4]):
  272.                     inp.delta[0][channel_number][j+a][k+b][l+c]+=self.delta[i]*self.W[0][kernel_number][a][b][c]
  273.         inp.delta/=(self.filter_shape[1]*1.0)
  274.                     #sum=sum+self.W[0][kernel_number][a][b][c]*self.input[0][channel_number][j+a][k+b][l+c]
  275.            
  276.            
  277.    
  278. """
  279. Here input object is the output object of previous class
  280. """
  281. class LogisticRegession(object):
  282.     def __init__(self,rng,input_object,n_in,n_out):
  283.         self.input=input_object
  284.         self.n_in=n_in
  285.         self.n_out=n_out
  286.         self.delta=[]
  287.         W_bound=6.0/(n_in+n_out)
  288.         W=numpy.asarray(rng.uniform(low=-W_bound,high=W_bound,size=(n_in,n_out)), dtype=numpy.float32)
  289.         #W=numpy.zeros((n_in,n_out),dtype=numpy.float32)
  290.         b=numpy.zeros((n_out),dtype=numpy.float32)
  291.         self.W=W
  292.         self.b=b
  293.         self.outputs()
  294.     def outputs(self):
  295.         output=[]
  296.         for i in xrange(self.n_out):
  297.             out=0.0
  298.             for j in xrange(self.n_in):
  299.                 out=out+self.input.output[0][j][0][0][0]*self.W[j][i]
  300.                 if(j==self.n_in-1):
  301.                     output.append(numpy.tanh(out+self.b[i]))
  302.         self.output=output
  303.         print self.output
  304.     def change_input(self,input_object):
  305.         self.input=input_object
  306.         self.outputs()
  307.     """
  308.    Cost is sigma((y_expected-y)^2)/2        
  309.    """
  310.     def cost(self,y):
  311.         e=0
  312.         for i in xrange(self.n_out):
  313.             if y==i:
  314.                 e=e+numpy.square((1-self.output[i]))
  315.             else:
  316.                 e=e+numpy.square((-1-self.output[i]))
  317.         return e
  318.                
  319.     def error(self,y):
  320.         print self.output
  321.         maximum_index=max(enumerate(self.output),key=lambda x: x[1])[0]
  322.         print maximum_index,y
  323.         if maximum_index==y:
  324.             return 0
  325.         else:
  326.             return 1
  327.     @profile
  328.     def gradient(self,y,eta):
  329.         delta=[]
  330.         for i in xrange(self.n_out):
  331.             if y==i:
  332.                 delta.append((self.output[i]-1)*((1-self.output[i])**2))
  333.                 #delta.append((1-self.output[i])**2)
  334.             else:
  335.                 delta.append((self.output[i]+1)*((1-self.output[i])**2))
  336.                 #delta.append((1-self.output[i])**2)
  337.         self.delta=delta
  338.        
  339.         delta_prev=[]
  340.         ab=0
  341.         for i in xrange(self.n_in):
  342.             for j in xrange(self.n_out):
  343.                 ab=ab+self.W[i][j]*delta[j]
  344.             ab=ab*((1-self.input.output[0][i][0][0][0])**2)
  345.             delta_prev.append(ab)
  346.             ab=0
  347.         self.input.delta=delta_prev
  348.         for i in xrange(self.n_in):
  349.             for j in xrange(self.n_out):
  350.                 self.W[i][j]=self.W[i][j] - (eta*self.input.output[0][i][0][0][0]*delta[j])
  351.         for i in xrange(self.n_out):
  352.             self.b[i]=self.b[i]-eta*delta[i]
  353. @profile            
  354. def check(learning_rate=0.005, n_epochs=3,dataset='./kth_thesis.pkl',nkerns=[5,3,5], batch_size=1):
  355.     rng=numpy.random.RandomState(23455)
  356.     f = open('kth_thesis.pkl', 'rb')
  357.     train_set, test_set = cPickle.load(f)
  358.     f.close()
  359.     train_set_x, train_set_y=train_set
  360.     print train_set
  361.     print test_set
  362.     #validate_set_x, validate_set_y=validate_set
  363.     test_set_x, test_set_y=test_set
  364.     n_train_batches = len(train_set_x)
  365.     n_test_batches = len(test_set_x)
  366.     a=numpy.arange(n_train_batches)
  367.     random.shuffle(a)
  368.     k=0
  369.     k1=0
  370.     k2=0
  371.     for n in xrange(2):
  372.         print "training epoch %d" %n
  373.         for i in xrange(n_train_batches):
  374.             print a[i]
  375.             print 'Output is',train_set_y[a[i]]
  376.             print "training batch %d" %i
  377.             if n==0 and k==0:
  378.             print "ffff:  ",k1
  379.         k1+=1
  380.                 layer1=First_Convolution(rng,numpy.reshape(train_set_x[a[i]],(1,1,9,60,80)),(1,1,9,60,80),(1,nkerns[0],4,7,9),(3,3))
  381.                 layer2=Hidden_Convolution(rng,layer1.output,(1,nkerns[0],6,18,24),(1,nkerns[1],4,7,7),(3,3))
  382.                 layer3=Last_Convolution(rng,layer2.output,(1,nkerns[0]*nkerns[1],3,4,6),(1,nkerns[2],3,4,6),(1,1))
  383.                 layer4=LogisticRegession(rng,layer3,nkerns[0]*nkerns[1]*nkerns[2],6)
  384.                 k+=1
  385.             else:
  386.             print "eeee: ",k2
  387.         k2+=1
  388.                 layer1.change_input(numpy.reshape(train_set_x[a[i]],(1,1,9,60,80)))
  389.                 layer2.change_input(layer1.output)
  390.                 layer3.change_input(layer2.output)
  391.                 layer4.change_input(layer3)
  392.             layer4.gradient(train_set_y[a[i]],learning_rate)
  393.             layer3.gradient(layer2,learning_rate)
  394.             layer2.gradient(layer1,learning_rate)
  395.             layer1.gradient(learning_rate)
  396.            
  397.  
  398.  
  399.  
  400.             #if n==0 and i==1:
  401.                 #print "hi"
  402.                 #print layer3.W
  403.                 #print "hi"
  404.                 #print layer2.output
  405.                
  406.         error=0.0
  407.         print "checking errors"
  408.         for i in xrange(n_test_batches):
  409.  
  410.             layer1.change_input(numpy.reshape(test_set_x[i],(1,1,9,60,80)))
  411.             layer2.change_input(layer1.output)
  412.             layer3.change_input(layer2.output)
  413.             layer4.change_input(layer3)
  414.             error+=layer4.error(test_set_y[i])
  415.         error_percent=(error/n_test_batches)*100
  416.         #print layer1.W
  417.         print "error %% after %d epoch is " %n,error_percent
  418.  
  419. check()
Advertisement
Add Comment
Please, Sign In to add comment
Public Pastes
We use cookies for various purposes including analytics. By continuing to use Pastebin, you agree to our use of cookies as described in the Cookies Policy.  OK, I Understand
Not a member of Pastebin yet?
Sign Up, it unlocks many cool features!