def create_bc_module(mock_behavior_specs, bc_settings, use_rnn, tanhresample):
# model_path = env.external_brain_names[0]
trainer_config = TrainerSettings()
trainer_config.network_settings.memory = (NetworkSettings.MemorySettings()
if use_rnn else None)
policy = TFPolicy(0, mock_behavior_specs, trainer_config, tanhresample,
tanhresample)
with policy.graph.as_default():
bc_module = BCModule(
policy,
policy_learning_rate=trainer_config.hyperparameters.learning_rate,
default_batch_size=trainer_config.hyperparameters.batch_size,
default_num_epoch=3,
settings=bc_settings,
)
policy.initialize(
) # Normally the optimizer calls this after the BCModule is created
return bc_module
def test_step_overflow():
behavior_spec = mb.setup_test_behavior_specs(use_discrete=True,
use_visual=False,
vector_action_space=[2],
vector_obs_space=1)
policy = TFPolicy(
0,
behavior_spec,
TrainerSettings(network_settings=NetworkSettings(normalize=True)),
create_tf_graph=False,
)
policy.create_input_placeholders()
policy.initialize()
policy.set_step(2**31 - 1)
assert policy.get_current_step() == 2**31 - 1
policy.increment_step(3)
assert policy.get_current_step() == 2**31 + 2
def create_sac_optimizer_mock(dummy_config, use_rnn, use_discrete, use_visual):
mock_brain = mb.setup_test_behavior_specs(
use_discrete,
use_visual,
vector_action_space=DISCRETE_ACTION_SPACE
if use_discrete else VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE if not use_visual else 0,
)
trainer_settings = dummy_config
trainer_settings.network_settings.memory = (NetworkSettings.MemorySettings(
sequence_length=16, memory_size=10) if use_rnn else None)
policy = TFPolicy(0,
mock_brain,
trainer_settings,
"test",
False,
create_tf_graph=False)
optimizer = SACOptimizer(policy, trainer_settings)
policy.initialize()
return optimizer
def _create_ppo_optimizer_ops_mock(dummy_config, use_rnn, use_discrete, use_visual):
mock_specs = mb.setup_test_behavior_specs(
use_discrete,
use_visual,
vector_action_space=DISCRETE_ACTION_SPACE
if use_discrete
else VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
)
trainer_settings = attr.evolve(dummy_config, framework=FrameworkType.TENSORFLOW)
trainer_settings.network_settings.memory = (
NetworkSettings.MemorySettings(sequence_length=16, memory_size=10)
if use_rnn
else None
)
policy = TFPolicy(
0, mock_specs, trainer_settings, "test", False, create_tf_graph=False
)
optimizer = PPOOptimizer(policy, trainer_settings)
policy.initialize()
return optimizer
