def mfom_model(in_dim, nclass):
# input block
feat_input = Input(shape=(in_dim, ), name='main_input')
# layer 1
x = Dense(10, name='dense1')(feat_input)
x = Activation(activation='sigmoid', name='act1')(x)
# layer 2
x = Dense(10, name='dense2')(x)
x = Activation(activation='sigmoid', name='act2')(x)
# output layer
x = Dense(nclass, name='pre_activation')(x)
y_pred = Activation(activation='sigmoid', name='output')(x)
# === MFoM head ===
# misclassification layer, feed Y
y_true = Input(shape=(nclass, ), name='y_true')
psi = mfom.UvZMisclassification(name='uvz_misclass')([y_true, y_pred])
# class Loss function layer
# NOTE: you may want to add regularization or constraints
out = mfom.SmoothErrorCounter(
name='smooth_error_counter',
# alpha_constraint=constraints.min_max_norm(min_value=-4., max_value=4.),
# alpha_regularizer=regs.l1(0.001),
# beta_constraint=constraints.min_max_norm(min_value=-4., max_value=4.),
# beta_regularizer=regs.l1(0.001)
)(psi)
# compile model
model = Model(input=[y_true, feat_input], output=out)
return model
def _compile_model(self, input, output, params):
"""
Compile network structure with particular loss and optimizer
"""
# ===
# choose loss
# ===
if params['loss'] in obj.MFOM_OBJECTIVES:
# add 2 layers for Maximal Figure-of-Merit
y_true = Input(shape=(self.nclass,), name='y_true')
psi = mfom.UvZMisclassification(name='uvz_misclass')([y_true, output])
y_pred = mfom.SmoothErrorCounter(name='smooth_error_counter')(psi)
# MFoM need labels info during training
input = [y_true, input]
output = y_pred
loss = obj.MFOM_OBJECTIVES[params['loss']]
elif params['loss'] == obj.mfom_eer_embed.__name__:
loss = obj.mfom_eer_embed
else:
loss = params['loss']
# ===
# choose optimizer
# ===
if params['optimizer'] == 'adam':
optimizer = Adam(lr=params['learn_rate'], beta_1=0.9, beta_2=0.999, epsilon=1e-08)
elif params['optimizer'] == 'sgd':
optimizer = SGD(lr=params['learn_rate'], decay=1e-6, momentum=0.9, nesterov=True)
else:
optimizer = params['optimizer']
self.model = Model(input=input, output=output)
self.model.compile(loss=loss, optimizer=optimizer)
self.model.summary()
feat_dim, time_step = 28, 100
feat_input = Input(shape=(feat_dim, time_step), name='main_input')
x = Permute((2, 1))(feat_input)
for _f in [256, 64]:
x = TimeDistributed(Dense(_f))(x)
# x = Activation(activation='elu')(x)
x = Dropout(0.5)(x)
x = TimeDistributed(Dense(nclass))(x)
y_pred = Activation(activation='tanh', name='output')(x)
# misclassification layer, feed Y
y_true = Input(shape=(time_step, nclass), name='y_true')
psi = mfom.UvZMisclassification(name='uvz_misclass')([y_true, y_pred])
# class Loss function layer
out = mfom.SmoothErrorCounter(name='smooth_error_counter')(psi)
# out = BatchNormalization()(psi)
# compile model
model = Model(input=[y_true, feat_input], output=out)
model.compile(loss=obj.mfom_eer_normalized,
optimizer='Adadelta') # Adam, Adadelta
model.summary()
# train
all_X, all_Y = [], []
for i in range(10000 // time_step):
X, Y = generate_dataset(output_dim=nclass, num_examples=time_step)
all_X.append(X.T)
all_Y.append(Y)
all_X, all_Y = np.array(all_X), np.array(all_Y)