def init(self, lf=False, ol=False):
model = Sequential()
input_done = False
if self.flatten:
model.add(Flatten(name="flatten", input_shape=self.in_shape))
input_done = True
for i, h in enumerate(self.hiddens):
if input_done == False:
model.add(Dense(h.neurons, activation='relu', name="additional_hidden_" + str(i), input_shape=self.in_shape))
else:
model.add(Dense(h.neurons, activation='relu', name="additional_hidden_" + str(i)))
if h.dropout > 0:
model.add(Dropout(h.dropout, name="additional_dropout_" + str(i)))
if input_done == False:
model.add(Dense(self.out_shape, name="dense", input_shape=self.in_shape))
else:
model.add(Dense(self.out_shape, name="dense"))
model.add(Activation("softmax", name="softmax"))
if lf and self.labelflip_decay is not None:
model.add(LabelFlipNoise(weight_decay=self.labelflip_decay, trainable=True))
if ol and self.outlier_alpha is not None:
model.add(OutlierNoise(alpha=self.outlier_alpha))
self._model = model
return self
def new_model(in_shape, out_shape, hiddens=[], lf=False, lf_decay=0.1):
# type: (list, list, list(Hidden), bool, float) -> Sequential
model = Sequential()
model.add(Flatten(name="flatten", input_shape=in_shape))
for i, h in enumerate(hiddens):
model.add(
Dense(h.neurons,
activation='relu',
name="additional_hidden_" + str(i)))
if h.dropout > 0:
model.add(Dropout(h.dropout, name="additional_dropout_" + str(i)))
model.add(Dense(out_shape, name="dense"))
model.add(Activation("softmax", name="softmax"))
if lf:
model.add(LabelFlipNoise(weight_decay=lf_decay, trainable=True))
return model
def on_train_begin(self, logs={}):
self.model = Sequential()
for layer in self.model.layers:
if isinstance(layer, LabelFlipNoise):
layer = LabelFlipNoise()
layer.trainable = True