def backpropagation(self, theta, nn, X, y, lamb):
layersNumb=len(nn['structure'])
thetaDelta = [0]*(layersNumb)
m=len(X)
#calculate matrix of outpit values for all input vectors X
hLoc = self.runAll(nn, X).copy()
yLoc = np.array(y)
thetaLoc = nn['theta'].copy()
derFunct = np.vectorize( 'float *x, float *res', 'float z = 1/(1+exp(-x[i])); res[i] = z*(1-z)' )
zLoc = nn['z'].copy()
aLoc = nn['a'].copy()
for n in range(0, len(X)):
delta = [0]*(layersNumb+1) #fill list with zeros
delta[len(delta)-1] = (hLoc[n] - yLoc[n]).T #calculate delta of error of output layer
delta[len(delta)-1] = delta[len(delta)-1].reshape(1, -1)
for i in range(layersNumb-1, 0, -1):
if i>1: # we can not calculate delta[0] because we don't have theta[0] (and even we don't need it)
z = zLoc[i-1][n]
z = np.concatenate( ([[1]], z.reshape((1,)*(2-z.ndim) + z.shape),), axis=1) #add one for correct matrix multiplication
delta[i] = np.dot(thetaLoc[i].T, delta[i+1]).reshape(-1, 1) * derFunct(z).T
delta[i] = delta[i][1:]
#print(thetaDelta[i], delta[i+1].shape, aLoc[i-1][n], '\n')
#print(np.dot(thetaLoc[i].T, delta[i+1]).shape, derFunct(z).T.shape, '\n')
#print(delta[i+1].shape, aLoc[i-1][n].shape )
thetaDelta[i] = thetaDelta[i] + np.dot(delta[i+1].reshape(-1, 1), aLoc[i-1][n].reshape(1, -1)) #delta[i+1]*aLoc[i-1][n]
#exit()
for i in range(1, len(thetaDelta)):
thetaDelta[i]=thetaDelta[i]/m
thetaDelta[i][:,1:]=thetaDelta[i][:,1:]+thetaLoc[i][:,1:]*(lamb/m) #regularization
if type(theta) == np.ndarray: return np.asarray(self.unroll(thetaDelta)).reshape(-1) # to work also with fmin_cg
return thetaDelta
def costTotal(self, theta, nn, X, y, lamb):
m = len(X)
#following string is for fmin_cg computaton
if type(theta) == np.ndarray: nn['theta'] = self.roll(theta, nn['structure'])
y = np.array(copy.deepcopy(y))
hAll = self.runAll(nn, X) #feed forward to obtain output of neural network
cost = self.cost(hAll, y)
return cost/m+(lamb/(2*m))*self.regul(nn['theta']) #apply regularization
def unroll(self, arr):
for i in range(0,len(arr)):
arr[i]=np.array(arr[i])
if i==0:
res=(arr[i]).ravel().T
else:
res=np.vstack((res,(arr[i]).ravel().T))
res.shape=(1, len(res))
return res
def run(self, nn, input):
z=[0]
a=[]
a.append(copy.deepcopy(input))
a[0]=np.array(a[0]).T # nx1 vector
logFunc = self.logisticFunction()
for i in range(1, len(nn['structure'])):
a[i-1]=np.vstack(([1], a[i-1]))
z.append(np.dot(nn['theta'][i], a[i-1]))
a.append(logFunc(z[i]))
nn['z'] = z
nn['a'] = a
return a[len(nn['structure'])-1]
def roll(self, arr, structure):
rolled=[arr[0]]
shift=1
for i in range(1,len(structure)):
print(type(structure[i]), " * ", type(structure[i-1]+1))
exit()
temparr=arr[shift:shift+structure[i]*(structure[i-1]+1)].copy()
temparr.shape=(structure[i],structure[i-1]+1)
rolled.append(np.array(temparr)) #NEED COMPARE WITH MATRIX
print(rolled[-1].shape)
exit() #DEBUG NOT FIRED
shift+=structure[i]*(structure[i-1]+1)
return rolled
import mynp as np arr = np.array([ -0.4, 0.000028], dtype=np.np.float32) print(arr) val = np.array([-1.], dtype=np.np.float32) idx = np.array([0, 1], dtype=np.np.int32) arr[idx] = val
import mynp as np arr = np.array([ -0.5, 0.2], dtype=np.np.float32) ids = arr > 0 val = np.array([-1.], dtype=np.np.float32) arr[ids] = val
import mynp as np import neurolab as nl from pyopencl import array arr = np.array([[-7.],[ 7.],[ 7.],[-7.],[-7.]]) print(arr[slice(None, None, None),0])
