def build_model(state_len, num_classes):
print("Building model...")
model = Sequential()
model.add(Projection(state_len, input_shape=(1, state_len)))
model.add(Proj2Prob(state_len, input_shape=(1, state_len)))
model.add(EigenDist(num_classes, input_shape=(1, state_len)))
model.add(Flatten())
model.summary()
# model.compile(loss=keras.losses.categorical_crossentropy,
# model.compile(loss=keras.losses.mean_squared_error,
model.compile(loss=categorical_bernoulli_crossentropy,
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
return model
def build_model_do(state_len,
num_classes,
activation="relu",
to_raise=True,
to_reduce=True,
raise_quadratic=True,
reduce_quadratic=True):
global ECMM
print("Building model...")
if to_raise:
raised_dim = 128
if not raise_quadratic:
RaDO = RaiseDimensionNonQuadratic
else:
RaDO = RaiseDimension
# RaDO = ReduceDimension
else:
raised_dim = state_len
RaDO = RaiseDimensionIdentity
state_len1 = raised_dim
if to_reduce:
reduced_dim = 64
if not reduce_quadratic:
ReDO = ReduceDimensionNonQuadratic
else:
ReDO = ReduceDimension
else:
reduced_dim = state_len1
ReDO = ReduceDimensionIdentity
inputs = Input(shape=(1, state_len))
# 1
raise_dimension1 = RaDO(raised_dim,
input_shape=(1, state_len),
name="rado")
projection1 = Projection(state_len1,
input_shape=(1, state_len1),
name="p1")
proj2prob1 = Proj2Prob(state_len1, input_shape=(1, state_len1))
eigen_dist1 = ECMM(num_classes, input_shape=(1, state_len1), name="l1")
flatten1 = Flatten(name="1st_fold")
reduce_dimension1 = ReDO(reduced_dim,
input_shape=(1, state_len1),
name="redo")
# 2
state_len2 = reduced_dim
projection2 = Projection(state_len2,
input_shape=(1, state_len2),
name="p2")
proj2prob2 = Proj2Prob(state_len2, input_shape=(1, state_len2))
eigen_dist2 = ECMM(num_classes, input_shape=(1, state_len2), name="l2")
flatten2 = Flatten(name="2nd_fold")
# 1
a0 = raise_dimension1(inputs)
a1 = projection1(a0)
a2 = proj2prob1(a1)
o1 = flatten1(eigen_dist1(a2))
a1 = reduce_dimension1(a1)
# 2
b1 = projection2(a1)
b2 = proj2prob2(b1)
o2 = flatten2(eigen_dist2(b2))
model = Model(inputs, outputs=[o1, o2])
return model
def build_model_dnn(state_len,
num_classes,
to_raise=True,
to_reduce=True,
activation="relu"):
global ECMM
print("Building model...")
raised_dim = 128
reduced_dim = 128
inputs = Input(shape=(1, state_len))
# 1
if to_raise:
state_len1 = raised_dim
dense1 = Dense(state_len1, activation=activation)
else:
state_len1 = state_len
dense1 = RaiseDimensionIdentity(state_len1, input_shape=(1, state_len))
reshape1 = Reshape((1, state_len1), input_shape=(state_len1, ))
projection1 = Projection(state_len1, input_shape=(1, state_len1))
proj2prob1 = Proj2Prob(state_len1, input_shape=(1, state_len1))
eigen_dist1 = ECMM(num_classes, input_shape=(1, state_len1))
flatten1 = Flatten(name="1st_fold")
# 2
# flatten2_dense = Flatten()
if to_reduce:
state_len2 = reduced_dim
dense2 = Dense(state_len2, activation=activation)
else:
state_len2 = state_len1
dense2 = ReduceDimensionIdentity(state_len2,
input_shape=(1, state_len1))
reshape2 = Reshape((1, state_len2), input_shape=(state_len2, ))
projection2 = Projection(state_len2, input_shape=(1, state_len2))
proj2prob2 = Proj2Prob(state_len2, input_shape=(1, state_len2))
eigen_dist2 = ECMM(num_classes, input_shape=(1, state_len2))
flatten2 = Flatten(name="2nd_fold")
# 1
if to_raise:
a1 = Flatten()(inputs)
a2 = dense1(a1)
a2 = reshape1(a2)
else:
a2 = inputs
a3 = projection1(a2)
a4 = proj2prob1(a3)
o1 = flatten1(eigen_dist1(a4))
# 2
# b1 = flatten2_dense(a1)
if to_reduce:
a3 = Flatten()(a3)
b1 = dense2(a3)
b2 = reshape2(b1)
else:
b2 = a3
b3 = projection2(b2)
b4 = proj2prob2(b3)
o2 = flatten2(eigen_dist2(b4))
model = Model(inputs, outputs=[o1, o2])
return model