def test_breed3(self):
delta = 0.3
n_parents = 2
traits = dict(layer_dropout={1})
model_p = KerasPackageWrapper.make_flat_sequential_model()
model_p.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model_p.add(keras.layers.Dropout(0.1))
model_c1 = KerasPackageWrapper.make_flat_sequential_model()
model_c1.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model_c1.add(keras.layers.Dropout(0.3))
model_c2 = KerasPackageWrapper.make_flat_sequential_model()
model_c2.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model_c2.add(keras.layers.Dropout(0.7))
model_c3 = KerasPackageWrapper.make_flat_sequential_model()
model_c3.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model_c3.add(keras.layers.Dropout(0.9))
new_generation = GeneticOptimizer.breed([model_p, model_c1, model_c2, model_c3],
n_parents, traits, delta)
# Check the networks layer-wise
for layer_i in traits["layer_dropout"]:
for network in new_generation[n_parents:]:
self.assertEquals(model_p.layers[layer_i - 1].units, network.layers[layer_i - 1].units)
self._assertAlmostEqualsMultiple(network.layers[layer_i].rate, new_generation[:n_parents], delta)
def test_breed1(self):
delta = 0.0
n_parents = 1
traits = dict(layer_dropout={1})
model_p = KerasPackageWrapper.make_flat_sequential_model()
model_p.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model_p.add(keras.layers.Dropout(0.1))
model_c1 = KerasPackageWrapper.make_flat_sequential_model()
model_c1.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model_c1.add(keras.layers.Dropout(0.3))
model_c2 = KerasPackageWrapper.make_flat_sequential_model()
model_c2.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model_c2.add(keras.layers.Dropout(0.7))
new_generation = GeneticOptimizer.breed([model_p, model_c1, model_c2], n_parents,
traits, delta)
# Check the networks layer-wise
for layer_i in traits["layer_dropout"]:
self.assertEqual(model_p.layers[layer_i - 1].units, new_generation[1].layers[layer_i - 1].units)
self.assertEqual(model_p.layers[layer_i].rate, new_generation[1].layers[layer_i].rate)
self.assertEqual(model_p.layers[layer_i - 1].units, new_generation[2].layers[layer_i - 1].units)
self.assertEqual(model_p.layers[layer_i].rate, new_generation[2].layers[layer_i].rate)
def test_inherit_to_child4(self):
# TODO keras assumed!
delta = 0.1
model1 = KerasPackageWrapper.make_flat_sequential_model()
model1.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model1.add(keras.layers.Dropout(0.2))
model1.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
model2 = KerasPackageWrapper.make_flat_sequential_model()
model2.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model2.add(keras.layers.Dropout(0.7))
model2.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
parents = [model1, model2]
child = GeneticOptimizer.inherit_to_child(parents, dict(layer_dropout={1}), delta)
# Assert that the structure stays the same as that of parents but the rate changes
self.assertEqual(child.layers[0].units, model1.layers[0].units)
self.assertEqual(child.layers[0].units, model2.layers[0].units)
self._assertAlmostEqualsMultiple(child.layers[1].rate, TestGeneticOptimizer._extract_rates(parents, 1), delta)
def test_inherit_to_child1(self):
# TODO keras assumed!
delta = 0.0
model = KerasPackageWrapper.make_flat_sequential_model()
model.add(keras.layers.Dense(10, activation="relu", input_dim=10))
model.add(keras.layers.Dropout(0.5))
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
child = GeneticOptimizer.inherit_to_child([model], dict(layer_dropout={1}), delta)
self.assertEqual(model.layers[1].rate, child.layers[1].rate)
self.assertEqual(model.layers[0].units, child.layers[0].units)
def test_train_models2(self):
model1 = KerasPackageWrapper.make_flat_sequential_model()
model1.add(keras.layers.Dense(300, activation="relu", input_dim=784))
model1.add(keras.layers.Dropout(0.2))
model1.add(keras.layers.Dense(10, activation="softmax"))
model1.add(keras.layers.Dropout(0.2))
model2 = KerasPackageWrapper.make_flat_sequential_model()
model2.add(keras.layers.Dense(300, activation="relu", input_dim=784))
model2.add(keras.layers.Dropout(0.5))
model2.add(keras.layers.Dense(10, activation="softmax"))
model2.add(keras.layers.Dropout(0.5))
model3 = KerasPackageWrapper.make_flat_sequential_model()
model3.add(keras.layers.Dense(300, activation="relu", input_dim=784))
model3.add(keras.layers.Dropout(0.7))
model3.add(keras.layers.Dense(10, activation="softmax"))
model3.add(keras.layers.Dropout(0.7))
model1.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
model2.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
model3.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
instance = GeneticOptimizer(model1, (get_clean_mnist_with_cold_labels()[:2]),
(get_clean_mnist_with_cold_labels()[2:]), n_categories=10,
traits=dict(layer_dropout={1, 3}))
print(GeneticOptimizer.train_models(instance, [model1, model2, model3]))
def test_get_accuracy(self):
model = KerasPackageWrapper.make_flat_sequential_model()
# TODO keras assumed!
model.add(keras.layers.Dense(10, activation="relu", input_dim=784))
model.add(keras.layers.Dropout(0.2))
model.add(keras.layers.Dense(10, activation="relu"))
model.add(keras.layers.Dropout(0.2))
# Compile the network
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
result = KerasPackageWrapper.get_accuracy(model, self.test_data_x, self.test_data_y)
print("Test for get_accuracy. Accuracy is {}".format(result))
def test_do_training(self):
model = KerasPackageWrapper.make_flat_sequential_model()
# TODO keras assumed!
model.add(keras.layers.Dense(500, activation="relu", input_dim=784))
model.add(keras.layers.Dropout(0.2))
model.add(keras.layers.Dense(30, activation="relu"))
model.add(keras.layers.Dropout(0.2))
model.add(keras.layers.Dense(30, activation="relu"))
model.add(keras.layers.Dropout(0.2))
model.add(keras.layers.Dense(10, activation="softmax"))
model.add(keras.layers.Dropout(0.2))
# Compile the network
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
print(KerasPackageWrapper.do_training(model, *get_clean_mnist_with_cold_labels(), 10, 3))