def test_loss_is_equal_when_no_jumps(self):
connections = ((-1, 2), (-2, 2), (2, 0), (2, 1))
genome = generate_genome_given_graph(graph=connections,
connection_weights=(1.0, 2.0, 3.0, 0.5))
dataset = get_dataset(dataset=self.config.dataset, train_percentage=0.1, testing=False, noise=0.0,
random_state=self.config.dataset_random_state)
n_samples = 100
stg_trainer = StandardTrainer(dataset=dataset, n_samples=self.config.n_samples,
n_output=genome.n_output,
problem_type=self.config.problem_type,
beta=self.config.beta,
n_epochs=100, is_cuda=False)
stg_trainer.train(genome)
best_network = stg_trainer.get_best_network()
new_genome = convert_stochastic_network_to_genome(network=best_network,
original_genome=genome,
fitness=-stg_trainer.best_loss_val,
fix_std=genome.genome_config.fix_std)
# evaluate genome
loss = get_loss(problem_type=self.config.problem_type)
loss_value = evaluate_genome(genome=new_genome,
dataset=dataset,
loss=loss,
problem_type=self.config.problem_type,
beta_type=self.config.beta_type,
batch_size=self.config.batch_size,
n_samples=n_samples,
is_gpu=False,
is_testing=False,
return_all=False,
is_pass=True)
self.assertAlmostEqual(loss_value, -new_genome.fitness, places=0)
def mutate(self, genome: Genome):
network, final_loss = self._mutate(genome)
return convert_stochastic_network_to_genome(
network=network,
original_genome=genome,
fitness=-final_loss,
fix_std=genome.genome_config.fix_std)
def test_genome_conversion_without_jumps(self):
original_genome = generate_genome_given_graph(graph=((-1, 2), (-2, 2), (2, 0), (2, 1)),
connection_weights=(1.0, 2.0, 3.0, 0.5))
network = ComplexStochasticNetwork(genome=original_genome)
new_genome = convert_stochastic_network_to_genome(network=network, original_genome=original_genome)
self.assertEqual(original_genome, new_genome)
def test_genome_conversion_with_jumps_2(self):
original_genome = generate_genome_given_graph(graph=((-1, 3), (-2, 3), (3, 2), (-1, 2), (2, 0),
(2, 1), (-1, 1), (-2, 3)),
connection_weights=(1.0, 2.0, 3.0, 0.5, 0.75,
0.8, 0.6, 0.9))
network = ComplexStochasticNetwork(genome=original_genome)
new_genome = convert_stochastic_network_to_genome(network=network, original_genome=original_genome)
new_network = ComplexStochasticNetwork(genome=new_genome)
self.assertEqual(original_genome, new_genome)
compare_networks(network, new_network)
self.assertTrue(equal(network, new_network))
def _finetune_genome(self, genome: Genome):
stg_trainer = StandardTrainer(dataset=self._get_dataset(),
n_samples=self.config.n_samples,
n_output=self.config.n_output,
problem_type=self.config.problem_type,
beta=self.config.beta,
n_epochs=self.n_epochs,
is_cuda=self.is_cuda,
weight_decay=self.weight_decay,
lr=self.lr)
stg_trainer.train(genome)
network = stg_trainer.get_best_network()
return convert_stochastic_network_to_genome(
network=network,
original_genome=genome,
fitness=-stg_trainer.best_loss_val,
fix_std=genome.genome_config.fix_std)
def test_non_existing_connections_are_updated_2(self):
connections = ((-1, 1), (-2, 1))
genome = generate_genome_given_graph(graph=connections,
connection_weights=(1.0, 2.0))
dataset = get_dataset(dataset=self.config.dataset, train_percentage=0.1, testing=False, noise=0.0)
stg_trainer = StandardTrainer(dataset=dataset, n_samples=self.config.n_samples,
n_output=genome.n_output,
problem_type=self.config.problem_type,
beta=self.config.beta,
n_epochs=self.n_epochs, is_cuda=False)
stg_trainer.train(genome)
best_network = stg_trainer.get_best_network()
new_genome = convert_stochastic_network_to_genome(network=best_network,
original_genome=genome,
fitness=-stg_trainer.best_loss_val,
fix_std=genome.genome_config.fix_std)
network_mutated = ComplexStochasticNetwork(genome=new_genome)
self.assertEqual(type(new_genome), Genome)
self.assertTrue(equal(best_network, network_mutated))
