def build_triplet_model(self, shape, dimensions):
net = self.build_embedding(shape, dimensions)
# Receive 3 inputs
# Decide which of the two alternatives is closest to the original
# x - Original Image
# x1 - Alternative 1
# x2 - Alternative 2
x = Input(shape=shape, name='x')
x1 = Input(shape=shape, name='x1')
x2 = Input(shape=shape, name='x2')
# Get the embedded values
net_anchor = net(x)
net_positive = net(x1)
net_negative = net(x2)
# The Lamda layer produces output using given function. Here its Euclidean distance.
positive_dist = subtract(net_anchor, net_positive)
negative_dist = subtract(net_anchor, net_negative)
positive_dist = norm(positive_dist)
negative_dist = norm(negative_dist)
# Compare
out = Lambda(lambda a: a[0] - a[1])([positive_dist, negative_dist])
return Model(inputs=[x, x1, x2], outputs=out), x, net_anchor
def build_pretrained_model(self, shape, dimensions):
# Using vgg16 pre-trained weights
net = cifar10vgg(x_shape=shape, train=False).model
for _ in range(7):
net.pop()
net.add(GlobalMaxPooling2D(name='gap'))
net.add(Dense(dimensions, activation='relu'))
#net.add(Lambda(lambda x: K.l2_normalize(x, axis=1)))
#net = self.build_embedding(shape, dimensions)
# Receive 3 inputs
# Decide which of the two alternatives is closest to the original
x = Input(shape=shape, name='x')
x1 = Input(shape=shape, name='x1')
x2 = Input(shape=shape, name='x2')
# Get the embedded values
net_anchor = net(x)
net_positive = net(x1)
net_negative = net(x2)
# The Lamda layer produces output using given function. Here its Euclidean distance.
positive_dist = subtract(net_anchor, net_positive)
negative_dist = subtract(net_anchor, net_negative)
positive_dist = norm(positive_dist)
negative_dist = norm(negative_dist)
# Compare
out = Lambda(lambda a: a[0] - a[1])([positive_dist, negative_dist])
return Model(inputs=[x, x1, x2], outputs=out), x, net_anchor
def build_triplets_model_test(self, shape, dimensions):
net = self.build_embedding(shape, dimensions)
# Receive 3 inputs
# Decide which of the two alternatives is closest to the original
# x - Original Image
# x1 - Alternative 1
# x2 - Alternative 2
x = Input(shape=shape, name='x')
x1 = Input(shape=shape, name='x1')
x2 = Input(shape=shape, name='x2')
# Get the embedded values
e = net(x)
e1 = net(x1)
e2 = net(x2)
#e = BN()(e)
#e1 = BN()(e1)
#e2 = BN()(e2)
# Get the differences
d1 = subtract(e, e1)
d2 = subtract(e, e2)
# Normalize the differences
n1 = norm(d1)
n2 = norm(d2)
# Compare
out = Activation('sigmoid')(subtract(n2, n1))
return Model(inputs=[x, x1, x2], outputs=[out, e, e1, e2, d1, d2, n1, n2])