def test_tl_sequential(self):
# Base model.
base_model = models.Sequential()
base_model.add(layers.Dense(20, activation="relu", input_shape=(100,)))
# Triplet loss.
tl = ngdlmodels.TL(base=base_model)
self.assert_load_save_fine(tl)
def test_tl_functional(self):
# Base model.
base_model_input = layers.Input(shape=(100,))
base_model_output = base_model_input
base_model_output = layers.Dense(20, activation="relu")(base_model_output)
base = models.Model(base_model_input, base_model_output)
# Triplet loss.
tl = ngdlmodels.TL(base)
self.assert_load_save_fine(tl)
# Triplet loss.
latent_dim = 8
# Create the base-model.
base_input = layers.Input(shape=(28, 28))
base_output = base_input
base_output = layers.Flatten()(base_output)
base_output = layers.Dense(512, activation="relu")(base_output)
base_output = layers.Dense(256, activation="relu")(base_output)
base_output = layers.Dense(128, activation="relu")(base_output)
base_output = layers.Dense(latent_dim)(base_output)
base = models.Model(base_input, base_output)
# Create the triplet loss model.
tl = ngdlmodels.TL(base)
tl.compile(optimizer="rmsprop", triplet_loss="euclidean")
tl.summary()
# Train.
print("Train...")
history = tl.fit(
x_input_train, y_output_train,
epochs=100,
batch_size=128,
steps_per_epoch=1000,
minibatch_size=10,
shuffle=True,
validation_data=(x_input_validate, y_output_validate),
validation_steps=500
)