def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
policy = NNPolicy(
self.seed,
brain_parameters,
self.trainer_parameters,
self.is_training,
self.load,
tanh_squash=True,
reparameterize=True,
create_tf_graph=False,
)
for _reward_signal in policy.reward_signals.keys():
self.collected_rewards[_reward_signal] = defaultdict(lambda: 0)
# Load the replay buffer if load
if self.load and self.checkpoint_replay_buffer:
try:
self.load_replay_buffer()
except (AttributeError, FileNotFoundError):
logger.warning(
"Replay buffer was unable to load, starting from scratch.")
logger.debug("Loaded update buffer with {} sequences".format(
self.update_buffer.num_experiences))
return policy
def create_optimizer_mock(trainer_config, reward_signal_config, use_rnn,
use_discrete, use_visual):
mock_brain = mb.setup_mock_brain(
use_discrete,
use_visual,
vector_action_space=VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
discrete_action_space=DISCRETE_ACTION_SPACE,
)
trainer_parameters = trainer_config
model_path = "testpath"
trainer_parameters["model_path"] = model_path
trainer_parameters["keep_checkpoints"] = 3
trainer_parameters["reward_signals"].update(reward_signal_config)
trainer_parameters["use_recurrent"] = use_rnn
policy = NNPolicy(0,
mock_brain,
trainer_parameters,
False,
False,
create_tf_graph=False)
if trainer_parameters["trainer"] == "sac":
optimizer = SACOptimizer(policy, trainer_parameters)
else:
optimizer = PPOOptimizer(policy, trainer_parameters)
return optimizer
def create_bc_module(mock_brain, trainer_config, use_rnn, demo_file, tanhresample):
# model_path = env.external_brain_names[0]
trainer_config["model_path"] = "testpath"
trainer_config["keep_checkpoints"] = 3
trainer_config["use_recurrent"] = use_rnn
trainer_config["behavioral_cloning"]["demo_path"] = (
os.path.dirname(os.path.abspath(__file__)) + "/" + demo_file
)
policy = NNPolicy(
0, mock_brain, trainer_config, False, False, tanhresample, tanhresample
)
with policy.graph.as_default():
bc_module = BCModule(
policy,
policy_learning_rate=trainer_config["learning_rate"],
default_batch_size=trainer_config["batch_size"],
default_num_epoch=3,
**trainer_config["behavioral_cloning"],
)
policy.initialize_or_load() # Normally the optimizer calls this after the BCModule is created
return bc_module
def test_ppo_get_value_estimates(mock_communicator, mock_launcher,
dummy_config):
tf.reset_default_graph()
brain_params = BrainParameters(
brain_name="test_brain",
vector_observation_space_size=1,
camera_resolutions=[],
vector_action_space_size=[2],
vector_action_descriptions=[],
vector_action_space_type=0,
)
dummy_config["summary_path"] = "./summaries/test_trainer_summary"
dummy_config["model_path"] = "./models/test_trainer_models/TestModel"
policy = NNPolicy(0,
brain_params,
dummy_config,
False,
False,
create_tf_graph=False)
optimizer = PPOOptimizer(policy, dummy_config)
time_horizon = 15
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
vec_obs_size=1,
num_vis_obs=0,
action_space=[2],
)
run_out, final_value_out = optimizer.get_trajectory_value_estimates(
trajectory.to_agentbuffer(), trajectory.next_obs, done=False)
for key, val in run_out.items():
assert type(key) is str
assert len(val) == 15
run_out, final_value_out = optimizer.get_trajectory_value_estimates(
trajectory.to_agentbuffer(), trajectory.next_obs, done=True)
for key, val in final_value_out.items():
assert type(key) is str
assert val == 0.0
# Check if we ignore terminal states properly
optimizer.reward_signals["extrinsic"].use_terminal_states = False
run_out, final_value_out = optimizer.get_trajectory_value_estimates(
trajectory.to_agentbuffer(), trajectory.next_obs, done=False)
for key, val in final_value_out.items():
assert type(key) is str
assert val != 0.0
def create_policy_mock(dummy_config, use_rnn, use_discrete, use_visual):
mock_brain = mb.setup_mock_brain(
use_discrete,
use_visual,
vector_action_space=VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
discrete_action_space=DISCRETE_ACTION_SPACE,
)
trainer_parameters = dummy_config
model_path = "testmodel"
trainer_parameters["model_path"] = model_path
trainer_parameters["keep_checkpoints"] = 3
trainer_parameters["use_recurrent"] = use_rnn
policy = NNPolicy(0, mock_brain, trainer_parameters, False, False)
return policy
def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
"""
Creates a PPO policy to trainers list of policies.
:param brain_parameters: specifications for policy construction
:return policy
"""
policy = NNPolicy(
self.seed,
brain_parameters,
self.trainer_parameters,
self.is_training,
self.load,
condition_sigma_on_obs=False, # Faster training for PPO
create_tf_graph=
False, # We will create the TF graph in the Optimizer
)
return policy
def test_normalization(dummy_config):
brain_params = BrainParameters(
brain_name="test_brain",
vector_observation_space_size=1,
camera_resolutions=[],
vector_action_space_size=[2],
vector_action_descriptions=[],
vector_action_space_type=0,
)
dummy_config["summary_path"] = "./summaries/test_trainer_summary"
dummy_config["model_path"] = "./models/test_trainer_models/TestModel"
time_horizon = 6
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
vec_obs_size=1,
num_vis_obs=0,
action_space=[2],
)
# Change half of the obs to 0
for i in range(3):
trajectory.steps[i].obs[0] = np.zeros(1, dtype=np.float32)
policy = policy = NNPolicy(0, brain_params, dummy_config, False, False)
trajectory_buffer = trajectory.to_agentbuffer()
policy.update_normalization(trajectory_buffer["vector_obs"])
# Check that the running mean and variance is correct
steps, mean, variance = policy.sess.run([
policy.normalization_steps, policy.running_mean,
policy.running_variance
])
assert steps == 6
assert mean[0] == 0.5
# Note: variance is divided by number of steps, and initialized to 1 to avoid
# divide by 0. The right answer is 0.25
assert (variance[0] - 1) / steps == 0.25
# Make another update, this time with all 1's
time_horizon = 10
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
vec_obs_size=1,
num_vis_obs=0,
action_space=[2],
)
trajectory_buffer = trajectory.to_agentbuffer()
policy.update_normalization(trajectory_buffer["vector_obs"])
# Check that the running mean and variance is correct
steps, mean, variance = policy.sess.run([
policy.normalization_steps, policy.running_mean,
policy.running_variance
])
assert steps == 16
assert mean[0] == 0.8125
assert (variance[0] - 1) / steps == pytest.approx(0.152, abs=0.01)