def test_mutate_gene(mock_random_sample):
config_dict = {"filter_size": [1, 2, 3, 5, 6], "num_filter": [8, 16, 32, 64], "has_pooling_layer": [False, True]}
network_table = NetworkBlockTable(config_dict)
rows = 2
cols = 3
l = 2
original_individual = initialize_individual_default(network_table, num_rows=rows, num_columns=cols, levelback=l)
mutated_individual = mutate(individual=original_individual, mutate_proba=0.1, num_rows=rows, num_cols=cols,
levelback=l,
network_block_table=network_table,
input_layer_shape=(None, 32, 32, 1)
)
num_diff = 0
original_genes = original_individual.genotype.gene_nodes
mutated_genes = mutated_individual.genotype.gene_nodes
for i in range(1, len(original_genes)):
if original_genes[i] != mutated_genes[i]:
num_diff += 1
# Check that every node has been mutated expect the input node:
assert num_diff == len(original_genes) - 1
# Check input and output node correctness
assert mutated_individual.genotype.gene_nodes[0].connection == None
assert mutated_individual.genotype.gene_nodes[0].network_block == None
assert mutated_individual.genotype.gene_nodes[rows * cols + 1].connection != None
assert mutated_individual.genotype.gene_nodes[rows * cols + 1].network_block == None
def test_mutate_individual_with_valid_representation_constraint():
config_dict = {"filter_size": [1, 3, 5], "num_filter": [8, 16, 32, 64, 128, 256],
"has_pooling_layer": [True, False]}
network_table = NetworkBlockTable(config_dict)
rows = 4
cols = 20
l = 5
original_individual = initialize_individual_default(network_table, num_rows=rows, num_columns=cols, levelback=l)
mutated_individual = mutate(individual=original_individual, mutate_proba=0.9, num_rows=rows, num_cols=cols,
levelback=l,
network_block_table=network_table,
input_layer_shape=(None, 32, 32, 1))
layer_shapes = get_layer_shapes_from_phenotype(input_shape=(None, 32, 32, 1),
phenotype=mutated_individual.phenotype)
print("--------------------------------------")
pprint.pprint(layer_shapes)
print("--------------------------------------")
representation_layer = layer_shapes[-1]
assert representation_layer[1] >= 1
assert representation_layer[2] >= 1
def test_deep_copy_individual():
config_dict = {
"filter_size": [1, 3, 5],
"num_filter": [8, 16],
"has_pooling_layer": [False, True]
}
network_table = NetworkBlockTable(config_dict)
g1 = GeneNode(None, None)
g2 = GeneNode(network_table.lookup_dict[6], 0)
g3 = GeneNode(network_table.lookup_dict[3], 1)
g4 = GeneNode(None, 2)
i1 = Individual(Genotype([g1, g2, g3, g4]))
i2 = copy.deepcopy(i1)
# i2 = i1.__deepcopy__()
assert i1 is not i2
assert i1.genotype.gene_nodes[0].network_block is i2.genotype.gene_nodes[
0].network_block
assert i1.genotype.gene_nodes[0].connection is i2.genotype.gene_nodes[
0].connection
i1.genotype.gene_nodes[0].network_block = network_table.lookup_dict[1]
i1.genotype.gene_nodes[0].connection = 3
assert i1 is not i2
assert i1.genotype.gene_nodes[
0].network_block is not i2.genotype.gene_nodes[0].network_block
assert i1.genotype.gene_nodes[0].connection is not i2.genotype.gene_nodes[
0].connection
def test_decode_phenotype_from_genotype_mark_as_coding():
config_dict = {"filter_size": [3, 5], "num_filter": [16, 32], "has_pooling_layer": [False, True]}
network_table = NetworkBlockTable(config_dict)
genotype_nodes = [GeneNode(None, None),
GeneNode(network_table.lookup_dict[1], 0),
GeneNode(network_table.lookup_dict[6], 1),
# The following two genes are intentionally non-coding:
GeneNode(network_table.lookup_dict[3], 1),
GeneNode(network_table.lookup_dict[1], 3),
# output node
GeneNode(None, 2)]
test_genotype = Genotype(gene_nodes=genotype_nodes)
ind = Individual(genotype=test_genotype)
coding_nodes = 0
node: GeneNode
for node in ind.genotype.gene_nodes:
if node.coding:
coding_nodes += 1
# from the list: node1, node2 are the coding nodes, input and output nodes are automatically added to a network
# and are not marked as coding here.
assert coding_nodes == 2
def test_mutate_active_gene_forced():
config_dict = {"filter_size": [1, 2], "num_filter": [8, 16], "has_pooling_layer": [False]}
network_table = NetworkBlockTable(config_dict)
rows = 2
cols = 3
l = 2
individual = initialize_individual_default(network_table, num_rows=rows, num_columns=cols, levelback=l)
mutated_individual = mutate_active_gene_forced(individual=individual, mutate_proba=0.1, num_rows=2, num_cols=2,
levelback=l, network_block_table=network_table,
input_layer_shape=(None, 32, 32, 1))
assert mutated_individual.phenotype != individual.phenotype
def test_integration_genereate_and_decode_individual():
config_dict = {"filter_size": [1, 3, 5], "num_filter": [8, 16, 32, 64, 128], "has_pooling_layer": [False, True]}
network_table = NetworkBlockTable(config_dict)
for i in range(10):
print(f"Initalize individual number {i}")
individual = initialize_individual_default(network_table, num_rows=3, num_columns=20, levelback=5)
individual: Individual
phenotype = individual.phenotype
model = generate_model_from_phenotype(phenotype, input_layer_shape=(None, 32, 32, 1))
assert model is not None, "Error occurred during model creation."
assert tuples_are_equal(model.layers[0].output_shape[0], model.layers[-1].output_shape)
assert tuples_are_equal(model.layers[1].output_shape, model.layers[-2].output_shape)
K.clear_session()
def test_mutate_passive_gene():
config_dict = {"filter_size": [1, 2, 3, 4, 5], "num_filter": [8, 16], "has_pooling_layer": [False]}
network_table = NetworkBlockTable(config_dict)
rows = 4
cols = 10
l = 2
individual_mutated = initialize_individual_default(network_table, num_rows=rows, num_columns=cols, levelback=l)
individual_original = copy.deepcopy(individual_mutated)
mutate_passive_gene(individual_mutated, mutate_proba=0.1, num_rows=rows, num_cols=cols, levelback=l,
network_block_table=network_table)
p1 = individual_mutated.phenotype
p2 = individual_original.phenotype
assert p1 == p2
assert individual_original.genotype != individual_mutated.genotype
def test_network_bock_table_generation_1():
config_dict = {
"filter_size": [1],
"num_filter": [8],
"has_pooling_layer": [False]
}
network_table = NetworkBlockTable(config_dict)
created_network_combination = network_table.lookup_dict[0]
expected_network_combination = ConvNetworkBlock(block_type="Conv",
filter_size=1,
num_filter=8,
has_pooling_layer=False)
assert expected_network_combination == created_network_combination
def test_decode_phenotype_from_genotype():
config_dict = {"filter_size": [3, 5], "num_filter": [16, 32], "has_pooling_layer": [False, True]}
network_table = NetworkBlockTable(config_dict)
genotype_nodes = [GeneNode(None, None),
GeneNode(network_table.lookup_dict[1], 0),
GeneNode(network_table.lookup_dict[6], 1),
# The following two genes are intentionally non-coding:
GeneNode(network_table.lookup_dict[3], 1),
GeneNode(network_table.lookup_dict[1], 3),
GeneNode(None, 2)]
test_genotype = Genotype(gene_nodes=genotype_nodes)
expected_phenotype = Phenotype([ConvNetworkBlock("Conv", num_filter=16, filter_size=3, has_pooling_layer=True),
ConvNetworkBlock("Conv", num_filter=32, filter_size=5, has_pooling_layer=False)])
decoded_phenotype = Genotype.decode_phenotype_from_genotype(test_genotype)
assert decoded_phenotype == expected_phenotype
def test_initialize_individual_with_representation_size_constraint():
config_dict = {"filter_size": [1, 3, 5], "num_filter": [8, 16, 32, 64, 128, 256],
"has_pooling_layer": [True, False]}
network_table = NetworkBlockTable(config_dict)
rows = 4
cols = 20
l = 5
max_representation_size = 1024
individual = initialize_individual_default(network_table, num_rows=rows, num_columns=cols, levelback=l,
max_representation_size=max_representation_size)
layer_shapes = get_layer_shapes_from_phenotype(input_shape=(None, 32, 32, 1), phenotype=individual.phenotype)
print("--------------------------------------")
pprint.pprint(layer_shapes)
print("--------------------------------------")
representation_layer_size = calculate_layer_size(layer_shapes[-1])
assert representation_layer_size <= max_representation_size
def test_decode_phenotype_from_genotype3():
"""
Tests the special case where input is directly connected to output.
@return:
"""
config_dict = {"filter_size": [3, 5], "num_filter": [16, 32], "has_pooling_layer": [False, True]}
network_table = NetworkBlockTable(config_dict)
genotype_nodes = [GeneNode(None, None),
# The following two genes are intentionally non-coding:
GeneNode(network_table.lookup_dict[3], 1),
GeneNode(network_table.lookup_dict[1], 3),
GeneNode(None, 0)]
test_genotype = Genotype(gene_nodes=genotype_nodes)
expected_phenotype = Phenotype([])
decoded_phenotype = Genotype.decode_phenotype_from_genotype(test_genotype)
assert decoded_phenotype == expected_phenotype
def test_save_load_individual():
config_dict = {
"filter_size": [1, 3, 5],
"num_filter": [8, 16],
"has_pooling_layer": [False, True]
}
network_table = NetworkBlockTable(config_dict)
g1 = GeneNode(None, None)
g2 = GeneNode(network_table.lookup_dict[6], 0)
g3 = GeneNode(network_table.lookup_dict[3], 1)
g4 = GeneNode(None, 2)
i1 = Individual(Genotype([g1, g2, g3, g4]))
path = "tests_data/"
i1.save(path, "individual.pickle", None)
i2 = Individual.load(path, "individual.pickle", None)
assert i1 == i2
def test_initialize_individual(mock_randint):
config_dict = {"filter_size": [1, 2], "num_filter": [8, 16], "has_pooling_layer": [False]}
network_table = NetworkBlockTable(config_dict)
expected_input_node = GeneNode(None, None)
exptected_block = ConvNetworkBlock(block_type="Conv", filter_size=2, num_filter=16,
has_pooling_layer=False)
rows = 2
cols = 3
l = 2
individual = initialize_individual_default(network_table, num_rows=rows, num_columns=cols, levelback=l)
genes = individual.genotype.gene_nodes
num_genes = len(genes)
assert num_genes == rows * cols + 2
gene: GeneNode
for i, gene in enumerate(genes):
if i == 0:
assert gene == expected_input_node
elif i == num_genes - 1:
assert gene.network_block is None
assert gene.connection in range(0, rows * cols)
else:
assert gene.network_block == exptected_block
def evolutionary_search(
X_train,
fitness_eval_func,
fitness_func_args,
maximize_fitness,
search_config: "SearchConfiguration",
num_generations,
mutation_probability,
num_children,
num_rows,
num_cols,
levelback,
path=None,
start_generation=0,
start_individual: "Individual" = None,
fitness_history=None,
best_fitness_history=None,
generation_history=None,
logger_factory: ModelInfoLoggerFactory = None,
input_layer_shape=None,
num_evaluations=1,
initialize_individual=initialize_individual_default,
random_sample_block_function=random_sample_block_default):
epochs = search_config.epochs
batch_size = search_config.batch_size
# lists to track the results
fitness_history = [] if fitness_history is None else fitness_history
best_fitness_history = [] if best_fitness_history is None else best_fitness_history
generation_history = [] if generation_history is None else generation_history
# Initialization:
network_table = NetworkBlockTable(search_config.block_search_space,
network_type=search_config.network_type)
if start_individual is None:
model_logger = logger_factory.create(0, "parent")
parent = initialize_individual(
network_table,
num_rows=num_rows,
num_columns=num_cols,
levelback=levelback,
input_layer_shape=input_layer_shape,
max_representation_size=search_config.max_representation_size)
fitness_list = []
for i in range(num_evaluations):
print(
f"\n \n Training the first individual {i + 1}/{num_evaluations}"
)
model_logger_temp = logger_factory.create_muted(
log_keras_train=True)
parent.model = train_individual(
parent,
X_train,
batch_size,
epochs,
input_layer_shape=input_layer_shape,
model_info_logger=model_logger_temp if i > 0 else model_logger,
network_type=search_config.network_type)
fitness_temp = fitness_eval_func(parent.model, **fitness_func_args)
fitness_list.append(fitness_temp)
fitness_array = np.array(fitness_list)
parent.fitness = fitness_array.mean()
model_logger.log_fitness(parent.fitness)
if num_evaluations > 1:
model_logger.log_fitness_std(fitness_array.std())
parent.generation = 0
generation_history.append(0)
fitness_history.append(parent.fitness)
print(
f"\n First individual scored a fitness of {parent.fitness} ({fitness_array.std()})"
)
else:
parent = start_individual
parent.fitness = start_individual.fitness
parent.model = start_individual.model
# Begin of evolution:
generation = start_generation
while generation < num_generations + start_generation:
print(
f"Creating generation: {generation + 1}/{num_generations + start_generation}"
)
child_fitnesses = []
children = []
for i in range(num_children):
print(f"\n\nCreating and training child {i + 1}/{num_children}")
model_logger = logger_factory.create(generation, child=i)
child = mutate_active_gene_forced(
parent,
mutate_proba=mutation_probability,
num_rows=num_rows,
num_cols=num_cols,
levelback=levelback,
network_block_table=network_table,
input_layer_shape=input_layer_shape,
max_representation_size=search_config.max_representation_size,
random_sample_block_function=random_sample_block_function)
child.model = train_individual(
child,
X_train,
batch_size,
epochs,
input_layer_shape=input_layer_shape,
model_info_logger=model_logger,
network_type=search_config.network_type)
child.fitness = fitness_eval_func(child.model, **fitness_func_args)
fitness_list = [child.fitness]
# If child fitness is better retrain model and average fitness to make sure that this better fitness
# did not only occur for change. This is only done if num_evaluations > 1.
if child.fitness > parent.fitness and maximize_fitness \
or child.fitness < parent.fitness and not maximize_fitness:
if num_evaluations > 1:
print(
"\n\nChild has new best fitness. Checking for variance of architecture by retraining. \n"
)
for j in range(num_evaluations - 1):
print(f"Retraining {j + 1}/{num_evaluations - 1}")
model_logger_temp = logger_factory.create_muted(
log_keras_train=True)
child.model = train_individual(
child,
X_train,
batch_size,
epochs,
input_layer_shape=input_layer_shape,
model_info_logger=model_logger_temp,
network_type=search_config.network_type)
fitness_temp = fitness_eval_func(child.model,
**fitness_func_args)
fitness_list.append(fitness_temp)
fitness_array = np.array(fitness_list)
child.fitness = fitness_array.mean()
model_logger.log_fitness_std(fitness_array.std())
model_logger.log_fitness(child.fitness)
child.generation = generation
children.append(child)
child_fitnesses.append(child.fitness)
print(
f"Child scored a fitness of {child.fitness}({fitness_array.std()})"
)
if maximize_fitness:
best_child_index = np.argmax(child_fitnesses)
else:
best_child_index = np.argmin(child_fitnesses)
if child_fitnesses[best_child_index] > parent.fitness and maximize_fitness \
or child_fitnesses[best_child_index] < parent.fitness and not maximize_fitness:
parent = children[best_child_index]
best_fitness = child_fitnesses[best_child_index]
else:
mutate_passive_gene(individual=parent,
mutate_proba=mutation_probability,
num_rows=num_rows,
num_cols=num_cols,
levelback=levelback,
network_block_table=network_table)
best_fitness = parent.fitness
parent.generation = generation
__track_results(child_fitnesses, fitness_history, best_fitness,
best_fitness_history, generation, generation_history,
num_children, parent, path)
print(
f"Evaluation of generation {generation + 1}/{num_generations + start_generation} finished. Best fitness: {best_fitness}"
)
generation += 1
history = {
"generation": np.array(generation_history),
"fitness": np.array(fitness_history),
"best_fitness": np.array(best_fitness_history)
}
return parent, history