def test_ppo_value_baseline(self):
before_check_actor = get_checksum_network_parameters(self.network.get_actor_parameters())
before_check_critic = get_checksum_network_parameters(self.network.get_critic_parameters())
trainer_base_config['phi_key'] = 'value-baseline'
trainer = ProximatePolicyGradient(config=TrainerConfig().create(config_dict=trainer_base_config),
network=self.network)
trainer.train()
# test if network has changed
self.assertNotEqual(get_checksum_network_parameters(self.network.get_actor_parameters()), before_check_actor)
self.assertNotEqual(get_checksum_network_parameters(self.network.get_critic_parameters()), before_check_critic)
def get_checksum(self):
return get_checksum_network_parameters(self.parameters())
def test_discriminator(self):
# create ds
dataset = generate_dataset_by_length(
length=5,
input_size=(1, 200, 200),
output_size=(200, 200),
)
# create architecture
self.base_config[
'architecture'] = 'auto_encoder_deeply_supervised_with_discriminator'
self.base_config['initialisation_type'] = 'xavier'
network = eval(self.base_config['architecture']).Net(
config=ArchitectureConfig().create(config_dict=self.base_config))
# create trainer
output_path = self.output_dir + '/discriminator'
os.makedirs(output_path, exist_ok=True)
trainer_config = {
'output_path': output_path,
'optimizer': 'Adam',
'learning_rate': 0.1,
'factory_key': 'DeepSupervisionWithDiscriminator',
'data_loader_config': {
'batch_size': 2
},
'target_data_loader_config': {
'batch_size': 2
},
'criterion': 'WeightedBinaryCrossEntropyLoss',
"criterion_args_str": 'beta=0.9',
}
trainer = TrainerFactory().create(
config=TrainerConfig().create(config_dict=trainer_config),
network=network)
trainer.data_loader.set_dataset(dataset)
trainer.target_data_loader.set_dataset(dataset)
# test training main network
initial_main_checksum = get_checksum_network_parameters(
network.deeply_supervised_parameters())
initial_discriminator_checksum = get_checksum_network_parameters(
network.discriminator_parameters())
trainer._train_main_network()
self.assertEqual(
initial_discriminator_checksum,
get_checksum_network_parameters(
network.discriminator_parameters()))
self.assertNotEqual(
initial_main_checksum,
get_checksum_network_parameters(
network.deeply_supervised_parameters()))
# test training discriminator network
initial_discriminator_checksum = get_checksum_network_parameters(
network.discriminator_parameters())
initial_main_checksum = get_checksum_network_parameters(
network.deeply_supervised_parameters())
trainer._train_discriminator_network()
self.assertEqual(
initial_main_checksum,
get_checksum_network_parameters(
network.deeply_supervised_parameters()))
self.assertNotEqual(
initial_discriminator_checksum,
get_checksum_network_parameters(
network.discriminator_parameters()))
# test load checkpoint
self.base_config[
'architecture'] = 'auto_encoder_deeply_supervised_with_discriminator'
self.base_config['initialisation_type'] = 'xavier'
self.base_config['random_seed'] = 29078
second_network = eval(self.base_config['architecture']).Net(
config=ArchitectureConfig().create(config_dict=self.base_config))
self.assertNotEqual(second_network.get_checksum(),
network.get_checksum())
second_network.load_checkpoint(network.get_checkpoint())
for (n1, p1), (n2, p2) in zip(network.named_parameters(),
second_network.named_parameters()):
print(n1, n2)
self.assertEqual(p1.sum(), p2.sum())
network.remove()
second_network.remove()