def test_function_prime(self):
X = np.random.rand(4,2)
y = np.array([0,1,0,0,0,0,0,0,0,0,0,0,0,0,0])
test = CNN(vector_size=4)
CM1,B1,CM2,B2, FCL = test.initialize_weights()
d1fm, d2, d3 = test.function_prime(X, y)
for i in range(test.fmap1):
print d1fm[i], u'\n'
print d2, u'\n'
print d3, u'\n'
def test_gradient_check(e):
X = np.random.rand(4,2)
y = np.array([0,1,0,0,0,0,0,0,0,0,0,0,0,0,0])
test = CNN(vector_size=4)
W1,_B1, CM1,B1,CM2,B2,FCL = test.initialize_weights()
dW1, d_B1, dCM1, dB1, dCM2, dB2, dFCL = test.function_prime(X, y,W1, _B1, CM1,B1,CM2,B2,FCL)
# Through FCL
for i in range(test.num_labels):
for j in range(test.kmax*test.feature_maps_number_layer2*test.folding_width + 1):
FCL[i,j] += e
Jpos = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
FCL[i,j] -= 2*e
Jneg = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
J = (Jpos - Jneg)/(2*e)
backprop = dFCL[i,j]
if (J-backprop)/(J+backprop)>e:
print (J-backprop)/(J+backprop)
print J, backprop
FCL[i,j] += e
# #Through CL2
for i in range(test.feature_maps_number_layer2):
for j in range(test.feature_maps_number_layer1):
for n in range(test.vector_size):
for p in range(test.second_kernel_size):
CM2[i][j][n,p] += e
Jpos = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
CM2[i][j][n,p] -= 2*e
Jneg = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
J = (Jpos - Jneg)/(2*e)
backprop = dCM2[i][j][n,p]
if (J-backprop)/(J+backprop)>e:
print (J-backprop)/(J+backprop)
print J, backprop
CM2[i][j][n,p] += e
for i in range(test.feature_maps_number_layer2):
for j in range(test.folding_width):
B2[i][j] += e
Jpos = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
B2[i][j] -= 2*e
Jneg = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
J = (Jpos - Jneg)/(2*e)
backprop = dB2[i][j]
if (J-backprop)/(J+backprop)>e:
print (J-backprop)/(J+backprop)
print J, backprop
B2[i][j] += e
# Through CL1
for i in range(test.feature_maps_number_layer1):
for j in range(test.vector_size):
B1[i][j] += e
Jpos = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
B1[i][j] -= 2*e
Jneg = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
J = (Jpos - Jneg)/(2*e)
backprop = dB1[i][j]
if (J-backprop)/(J+backprop)>e:
print (J-backprop)/(J+backprop)
print J, backprop
B1[i][j] += e
for i in range(test.feature_maps_number_layer1):
for u in range(test.vector_size):
for v in range(test.first_kernel_size):
CM1[i][u,v] += e
Jpos = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
CM1[i][u,v] -= 2*e
Jneg = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
J = (Jpos - Jneg)/(2*e)
backprop = dCM1[i][u,v]
if (J-backprop)/(J+backprop)>e:
print (J-backprop)/(J+backprop)
print J, backprop
CM1[i][u,v] += e
# Through W2VTransf
for i in range(test.vector_size):
for j in range(X.shape[1]):
W1[i,j] += e
Jpos = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
W1[i,j] -= 2*e
Jneg = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
J = (Jpos - Jneg)/(2*e)
backprop = dW1[i,j]
if (J-backprop)/(J+backprop)>e:
print (J-backprop)/(J+backprop)
W1[i,j] += e
# Through W2VTransf
for i in range(test.vector_size):
_B1[i] += e
Jpos = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
_B1[i] -= 2*e
Jneg = test.function(X,y,W1, _B1,CM1,B1,CM2,B2,FCL)
J = (Jpos - Jneg)/(2*e)
backprop = d_B1[i]
if (J-backprop)/(J+backprop)>e:
print (J-backprop)/(J+backprop)
_B1[i] += e