def __init__(self, policy: TorchPolicy, trainer_settings: TrainerSettings):
"""
Takes a Policy and a Dict of trainer parameters and creates an Optimizer around the policy.
The PPO optimizer has a value estimator and a loss function.
:param policy: A TorchPolicy object that will be updated by this PPO Optimizer.
:param trainer_params: Trainer parameters dictionary that specifies the
properties of the trainer.
"""
# Create the graph here to give more granular control of the TF graph to the Optimizer.
super().__init__(policy, trainer_settings)
reward_signal_configs = trainer_settings.reward_signals
reward_signal_names = [
key.value for key, _ in reward_signal_configs.items()
]
if policy.shared_critic:
self._critic = policy.actor
else:
self._critic = ValueNetwork(
reward_signal_names,
policy.behavior_spec.observation_specs,
network_settings=trainer_settings.network_settings,
)
self._critic.to(default_device())
params = list(self.policy.actor.parameters()) + list(
self._critic.parameters())
self.hyperparameters: PPOSettings = cast(
PPOSettings, trainer_settings.hyperparameters)
self.decay_learning_rate = ModelUtils.DecayedValue(
self.hyperparameters.learning_rate_schedule,
self.hyperparameters.learning_rate,
1e-10,
self.trainer_settings.max_steps,
)
self.decay_epsilon = ModelUtils.DecayedValue(
self.hyperparameters.learning_rate_schedule,
self.hyperparameters.epsilon,
0.1,
self.trainer_settings.max_steps,
)
self.decay_beta = ModelUtils.DecayedValue(
self.hyperparameters.learning_rate_schedule,
self.hyperparameters.beta,
1e-5,
self.trainer_settings.max_steps,
)
self.optimizer = torch.optim.Adam(
params, lr=self.trainer_settings.hyperparameters.learning_rate)
self.stats_name_to_update_name = {
"Losses/Value Loss": "value_loss",
"Losses/Policy Loss": "policy_loss",
}
self.stream_names = list(self.reward_signals.keys())
def test_decayed_value():
test_steps = [0, 4, 9]
# Test constant decay
param = ModelUtils.DecayedValue(ScheduleType.CONSTANT, 1.0, 0.2, test_steps[-1])
for _step in test_steps:
_param = param.get_value(_step)
assert _param == 1.0
test_results = [1.0, 0.6444, 0.2]
# Test linear decay
param = ModelUtils.DecayedValue(ScheduleType.LINEAR, 1.0, 0.2, test_steps[-1])
for _step, _result in zip(test_steps, test_results):
_param = param.get_value(_step)
assert _param == pytest.approx(_result, abs=0.01)
# Test invalid
with pytest.raises(UnityTrainerException):
ModelUtils.DecayedValue(
"SomeOtherSchedule", 1.0, 0.2, test_steps[-1]
).get_value(0)
def __init__(
self,
policy: TorchPolicy,
settings: BehavioralCloningSettings,
policy_learning_rate: float,
default_batch_size: int,
default_num_epoch: int,
):
"""
A BC trainer that can be used inline with RL.
:param policy: The policy of the learning model
:param settings: The settings for BehavioralCloning including LR strength, batch_size,
num_epochs, samples_per_update and LR annealing steps.
:param policy_learning_rate: The initial Learning Rate of the policy. Used to set an appropriate learning rate
for the pretrainer.
"""
self.policy = policy
self._anneal_steps = settings.steps
self.current_lr = policy_learning_rate * settings.strength
learning_rate_schedule: ScheduleType = ScheduleType.LINEAR if self._anneal_steps > 0 else ScheduleType.CONSTANT
self.decay_learning_rate = ModelUtils.DecayedValue(
learning_rate_schedule, self.current_lr, 1e-10, self._anneal_steps
)
params = self.policy.actor_critic.parameters()
self.optimizer = torch.optim.Adam(params, lr=self.current_lr)
_, self.demonstration_buffer = demo_to_buffer(
settings.demo_path, policy.sequence_length, policy.behavior_spec
)
self.batch_size = (
settings.batch_size if settings.batch_size else default_batch_size
)
self.num_epoch = settings.num_epoch if settings.num_epoch else default_num_epoch
self.n_sequences = max(
min(self.batch_size, self.demonstration_buffer.num_experiences)
// policy.sequence_length,
1,
)
self.has_updated = False
self.use_recurrent = self.policy.use_recurrent
self.samples_per_update = settings.samples_per_update
def __init__(self, policy: TorchPolicy, trainer_params: TrainerSettings):
super().__init__(policy, trainer_params)
hyperparameters: SACSettings = cast(SACSettings,
trainer_params.hyperparameters)
self.tau = hyperparameters.tau
self.init_entcoef = hyperparameters.init_entcoef
self.policy = policy
policy_network_settings = policy.network_settings
self.tau = hyperparameters.tau
self.burn_in_ratio = 0.0
# Non-exposed SAC parameters
self.discrete_target_entropy_scale = 0.2 # Roughly equal to e-greedy 0.05
self.continuous_target_entropy_scale = 1.0
self.stream_names = list(self.reward_signals.keys())
# Use to reduce "survivor bonus" when using Curiosity or GAIL.
self.gammas = [
_val.gamma for _val in trainer_params.reward_signals.values()
]
self.use_dones_in_backup = {
name: int(not self.reward_signals[name].ignore_done)
for name in self.stream_names
}
self._action_spec = self.policy.behavior_spec.action_spec
self.value_network = TorchSACOptimizer.PolicyValueNetwork(
self.stream_names,
self.policy.behavior_spec.sensor_specs,
policy_network_settings,
self._action_spec,
)
self.target_network = ValueNetwork(
self.stream_names,
self.policy.behavior_spec.sensor_specs,
policy_network_settings,
)
ModelUtils.soft_update(self.policy.actor_critic.critic,
self.target_network, 1.0)
# We create one entropy coefficient per action, whether discrete or continuous.
_disc_log_ent_coef = torch.nn.Parameter(
torch.log(
torch.as_tensor([self.init_entcoef] *
len(self._action_spec.discrete_branches))),
requires_grad=True,
)
_cont_log_ent_coef = torch.nn.Parameter(torch.log(
torch.as_tensor([self.init_entcoef])),
requires_grad=True)
self._log_ent_coef = TorchSACOptimizer.LogEntCoef(
discrete=_disc_log_ent_coef, continuous=_cont_log_ent_coef)
_cont_target = (
-1 * self.continuous_target_entropy_scale *
np.prod(self._action_spec.continuous_size).astype(np.float32))
_disc_target = [
self.discrete_target_entropy_scale * np.log(i).astype(np.float32)
for i in self._action_spec.discrete_branches
]
self.target_entropy = TorchSACOptimizer.TargetEntropy(
continuous=_cont_target, discrete=_disc_target)
policy_params = list(
self.policy.actor_critic.network_body.parameters()) + list(
self.policy.actor_critic.action_model.parameters())
value_params = list(self.value_network.parameters()) + list(
self.policy.actor_critic.critic.parameters())
logger.debug("value_vars")
for param in value_params:
logger.debug(param.shape)
logger.debug("policy_vars")
for param in policy_params:
logger.debug(param.shape)
self.decay_learning_rate = ModelUtils.DecayedValue(
hyperparameters.learning_rate_schedule,
hyperparameters.learning_rate,
1e-10,
self.trainer_settings.max_steps,
)
self.policy_optimizer = torch.optim.Adam(
policy_params, lr=hyperparameters.learning_rate)
self.value_optimizer = torch.optim.Adam(
value_params, lr=hyperparameters.learning_rate)
self.entropy_optimizer = torch.optim.Adam(
self._log_ent_coef.parameters(), lr=hyperparameters.learning_rate)
self._move_to_device(default_device())
def __init__(self, policy: TorchPolicy, trainer_params: TrainerSettings):
super().__init__(policy, trainer_params)
hyperparameters: SACSettings = cast(SACSettings, trainer_params.hyperparameters)
self.tau = hyperparameters.tau
self.init_entcoef = hyperparameters.init_entcoef
self.policy = policy
self.act_size = policy.act_size
policy_network_settings = policy.network_settings
self.tau = hyperparameters.tau
self.burn_in_ratio = 0.0
# Non-exposed SAC parameters
self.discrete_target_entropy_scale = 0.2 # Roughly equal to e-greedy 0.05
self.continuous_target_entropy_scale = 1.0
self.stream_names = list(self.reward_signals.keys())
# Use to reduce "survivor bonus" when using Curiosity or GAIL.
self.gammas = [_val.gamma for _val in trainer_params.reward_signals.values()]
self.use_dones_in_backup = {
name: int(not self.reward_signals[name].ignore_done)
for name in self.stream_names
}
# Critics should have 1/2 of the memory of the policy
critic_memory = policy_network_settings.memory
if critic_memory is not None:
critic_memory = attr.evolve(
critic_memory, memory_size=critic_memory.memory_size // 2
)
value_network_settings = attr.evolve(
policy_network_settings, memory=critic_memory
)
self.value_network = TorchSACOptimizer.PolicyValueNetwork(
self.stream_names,
self.policy.behavior_spec.observation_shapes,
value_network_settings,
self.policy.behavior_spec.action_type,
self.act_size,
)
self.target_network = ValueNetwork(
self.stream_names,
self.policy.behavior_spec.observation_shapes,
value_network_settings,
)
self.soft_update(self.policy.actor_critic.critic, self.target_network, 1.0)
self._log_ent_coef = torch.nn.Parameter(
torch.log(torch.as_tensor([self.init_entcoef] * len(self.act_size))),
requires_grad=True,
)
if self.policy.use_continuous_act:
self.target_entropy = torch.as_tensor(
-1
* self.continuous_target_entropy_scale
* np.prod(self.act_size[0]).astype(np.float32)
)
else:
self.target_entropy = [
self.discrete_target_entropy_scale * np.log(i).astype(np.float32)
for i in self.act_size
]
policy_params = list(self.policy.actor_critic.network_body.parameters()) + list(
self.policy.actor_critic.distribution.parameters()
)
value_params = list(self.value_network.parameters()) + list(
self.policy.actor_critic.critic.parameters()
)
logger.debug("value_vars")
for param in value_params:
logger.debug(param.shape)
logger.debug("policy_vars")
for param in policy_params:
logger.debug(param.shape)
self.decay_learning_rate = ModelUtils.DecayedValue(
hyperparameters.learning_rate_schedule,
hyperparameters.learning_rate,
1e-10,
self.trainer_settings.max_steps,
)
self.policy_optimizer = torch.optim.Adam(
policy_params, lr=hyperparameters.learning_rate
)
self.value_optimizer = torch.optim.Adam(
value_params, lr=hyperparameters.learning_rate
)
self.entropy_optimizer = torch.optim.Adam(
[self._log_ent_coef], lr=hyperparameters.learning_rate
)
