def create_model(
self,
brain: BrainParameters,
trainer_params: Dict[str, Any],
reward_signal_configs: Dict[str, Any],
is_training: bool,
load: bool,
seed: int,
) -> None:
with self.graph.as_default():
self.model = SACModel(
brain,
lr=float(trainer_params["learning_rate"]),
lr_schedule=LearningRateSchedule(
trainer_params.get("learning_rate_schedule", "constant")
),
h_size=int(trainer_params["hidden_units"]),
init_entcoef=float(trainer_params["init_entcoef"]),
max_step=float(trainer_params["max_steps"]),
normalize=trainer_params["normalize"],
use_recurrent=trainer_params["use_recurrent"],
num_layers=int(trainer_params["num_layers"]),
m_size=self.m_size,
seed=seed,
stream_names=list(reward_signal_configs.keys()),
tau=float(trainer_params["tau"]),
gammas=[_val["gamma"] for _val in reward_signal_configs.values()],
vis_encode_type=EncoderType(
trainer_params.get("vis_encode_type", "simple")
),
)
self.model.create_sac_optimizers()
self.inference_dict.update(
{
"action": self.model.output,
"log_probs": self.model.all_log_probs,
"entropy": self.model.entropy,
"learning_rate": self.model.learning_rate,
}
)
if self.use_continuous_act:
self.inference_dict["pre_action"] = self.model.output_pre
if self.use_recurrent:
self.inference_dict["memory_out"] = self.model.memory_out
self.update_dict.update(
{
"value_loss": self.model.total_value_loss,
"policy_loss": self.model.policy_loss,
"q1_loss": self.model.q1_loss,
"q2_loss": self.model.q2_loss,
"entropy_coef": self.model.ent_coef,
"entropy": self.model.entropy,
"update_batch": self.model.update_batch_policy,
"update_value": self.model.update_batch_value,
"update_entropy": self.model.update_batch_entropy,
}
)
def create_model(
self, brain, trainer_params, reward_signal_configs, is_training, load, seed
):
"""
Create PPO model
:param brain: Assigned Brain object.
:param trainer_params: Defined training parameters.
:param reward_signal_configs: Reward signal config
:param seed: Random seed.
"""
with self.graph.as_default():
self.model = PPOModel(
brain=brain,
lr=float(trainer_params["learning_rate"]),
lr_schedule=LearningRateSchedule(
trainer_params.get("learning_rate_schedule", "linear")
),
h_size=int(trainer_params["hidden_units"]),
epsilon=float(trainer_params["epsilon"]),
beta=float(trainer_params["beta"]),
max_step=float(trainer_params["max_steps"]),
normalize=trainer_params["normalize"],
use_recurrent=trainer_params["use_recurrent"],
num_layers=int(trainer_params["num_layers"]),
m_size=self.m_size,
seed=seed,
stream_names=list(reward_signal_configs.keys()),
vis_encode_type=EncoderType(
trainer_params.get("vis_encode_type", "simple")
),
)
self.model.create_ppo_optimizer()
self.inference_dict.update(
{
"action": self.model.output,
"log_probs": self.model.all_log_probs,
"value_heads": self.model.value_heads,
"value": self.model.value,
"entropy": self.model.entropy,
"learning_rate": self.model.learning_rate,
}
)
if self.use_continuous_act:
self.inference_dict["pre_action"] = self.model.output_pre
if self.use_recurrent:
self.inference_dict["memory_out"] = self.model.memory_out
self.total_policy_loss = self.model.abs_policy_loss
self.update_dict.update(
{
"value_loss": self.model.value_loss,
"policy_loss": self.total_policy_loss,
"update_batch": self.model.update_batch,
}
)
def __init__(self, policy: TFPolicy, trainer_params: Dict[str, Any]):
"""
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 TFPolicy 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.
policy.create_tf_graph()
with policy.graph.as_default():
with tf.variable_scope("optimizer/"):
super().__init__(policy, trainer_params)
lr = float(trainer_params["learning_rate"])
lr_schedule = LearningRateSchedule(
trainer_params.get("learning_rate_schedule", "linear"))
h_size = int(trainer_params["hidden_units"])
epsilon = float(trainer_params["epsilon"])
beta = float(trainer_params["beta"])
max_step = float(trainer_params["max_steps"])
num_layers = int(trainer_params["num_layers"])
vis_encode_type = EncoderType(
trainer_params.get("vis_encode_type", "simple"))
self.burn_in_ratio = float(
trainer_params.get("burn_in_ratio", 0.0))
self.stream_names = list(self.reward_signals.keys())
self.tf_optimizer: Optional[tf.train.AdamOptimizer] = None
self.grads = None
self.update_batch: Optional[tf.Operation] = None
self.stats_name_to_update_name = {
"Losses/Value Loss": "value_loss",
"Losses/Policy Loss": "policy_loss",
"Policy/Learning Rate": "learning_rate",
}
if self.policy.use_recurrent:
self.m_size = self.policy.m_size
self.memory_in = tf.placeholder(
shape=[None, self.m_size],
dtype=tf.float32,
name="recurrent_value_in",
)
if num_layers < 1:
num_layers = 1
if policy.use_continuous_act:
self._create_cc_critic(h_size, num_layers, vis_encode_type)
else:
self._create_dc_critic(h_size, num_layers, vis_encode_type)
self.learning_rate = ModelUtils.create_learning_rate(
lr_schedule, lr, self.policy.global_step, int(max_step))
self._create_losses(
self.policy.total_log_probs,
self.old_log_probs,
self.value_heads,
self.policy.entropy,
beta,
epsilon,
lr,
max_step,
)
self._create_ppo_optimizer_ops()
self.update_dict.update({
"value_loss": self.value_loss,
"policy_loss": self.abs_policy_loss,
"update_batch": self.update_batch,
"learning_rate": self.learning_rate,
})
self.policy.initialize_or_load()
def create_model(self, brain, trainer_params, reward_signal_configs,
is_training, load, seed):
"""
Create PPO models, one on each device
:param brain: Assigned Brain object.
:param trainer_params: Defined training parameters.
:param reward_signal_configs: Reward signal config
:param seed: Random seed.
"""
self.devices = get_devices()
with self.graph.as_default():
with tf.variable_scope("", reuse=tf.AUTO_REUSE):
for device in self.devices:
with tf.device(device):
self.towers.append(
PPOModel(
brain=brain,
lr=float(trainer_params["learning_rate"]),
lr_schedule=LearningRateSchedule(
trainer_params.get(
"learning_rate_schedule", "linear")),
h_size=int(trainer_params["hidden_units"]),
epsilon=float(trainer_params["epsilon"]),
beta=float(trainer_params["beta"]),
max_step=float(trainer_params["max_steps"]),
normalize=trainer_params["normalize"],
use_recurrent=trainer_params["use_recurrent"],
num_layers=int(trainer_params["num_layers"]),
m_size=self.m_size,
seed=seed,
stream_names=list(
reward_signal_configs.keys()),
vis_encode_type=EncoderType(
trainer_params.get("vis_encode_type",
"simple")),
))
self.towers[-1].create_ppo_optimizer()
self.model = self.towers[0]
avg_grads = self.average_gradients([t.grads for t in self.towers])
update_batch = self.model.optimizer.apply_gradients(avg_grads)
avg_value_loss = tf.reduce_mean(
tf.stack([model.value_loss for model in self.towers]), 0)
avg_policy_loss = tf.reduce_mean(
tf.stack([model.policy_loss for model in self.towers]), 0)
self.inference_dict.update({
"action": self.model.output,
"log_probs": self.model.all_log_probs,
"value_heads": self.model.value_heads,
"value": self.model.value,
"entropy": self.model.entropy,
"learning_rate": self.model.learning_rate,
})
if self.use_continuous_act:
self.inference_dict["pre_action"] = self.model.output_pre
if self.use_recurrent:
self.inference_dict["memory_out"] = self.model.memory_out
if (is_training and self.use_vec_obs and trainer_params["normalize"]
and not load):
self.inference_dict[
"update_mean"] = self.model.update_normalization
self.total_policy_loss = self.model.abs_policy_loss
self.update_dict.update({
"value_loss": avg_value_loss,
"policy_loss": avg_policy_loss,
"update_batch": update_batch,
})
def __init__(self, policy: TFPolicy, trainer_params: Dict[str, Any]):
"""
Takes a Unity environment and model-specific hyper-parameters and returns the
appropriate PPO agent model for the environment.
:param brain: Brain parameters used to generate specific network graph.
:param lr: Learning rate.
:param lr_schedule: Learning rate decay schedule.
:param h_size: Size of hidden layers
:param init_entcoef: Initial value for entropy coefficient. Set lower to learn faster,
set higher to explore more.
:return: a sub-class of PPOAgent tailored to the environment.
:param max_step: Total number of training steps.
:param normalize: Whether to normalize vector observation input.
:param use_recurrent: Whether to use an LSTM layer in the network.
:param num_layers: Number of hidden layers between encoded input and policy & value layers
:param tau: Strength of soft-Q update.
:param m_size: Size of brain memory.
"""
# Create the graph here to give more granular control of the TF graph to the Optimizer.
policy.create_tf_graph()
with policy.graph.as_default():
with tf.variable_scope(""):
super().__init__(policy, trainer_params)
lr = float(trainer_params["learning_rate"])
lr_schedule = LearningRateSchedule(
trainer_params.get("learning_rate_schedule", "constant"))
self.policy = policy
self.act_size = self.policy.act_size
h_size = int(trainer_params["hidden_units"])
max_step = float(trainer_params["max_steps"])
num_layers = int(trainer_params["num_layers"])
vis_encode_type = EncoderType(
trainer_params.get("vis_encode_type", "simple"))
self.tau = trainer_params.get("tau", 0.005)
self.burn_in_ratio = float(
trainer_params.get("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.init_entcoef = trainer_params.get("init_entcoef", 1.0)
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: tf.Variable(1.0)
for name in stream_names
}
self.disable_use_dones = {
name: self.use_dones_in_backup[name].assign(0.0)
for name in stream_names
}
if num_layers < 1:
num_layers = 1
self.target_init_op: List[tf.Tensor] = []
self.target_update_op: List[tf.Tensor] = []
self.update_batch_policy: Optional[tf.Operation] = None
self.update_batch_value: Optional[tf.Operation] = None
self.update_batch_entropy: Optional[tf.Operation] = None
self.policy_network = SACPolicyNetwork(
policy=self.policy,
m_size=self.policy.m_size, # 3x policy.m_size
h_size=h_size,
normalize=self.policy.normalize,
use_recurrent=self.policy.use_recurrent,
num_layers=num_layers,
stream_names=stream_names,
vis_encode_type=vis_encode_type,
)
self.target_network = SACTargetNetwork(
policy=self.policy,
m_size=self.policy.m_size, # 1x policy.m_size
h_size=h_size,
normalize=self.policy.normalize,
use_recurrent=self.policy.use_recurrent,
num_layers=num_layers,
stream_names=stream_names,
vis_encode_type=vis_encode_type,
)
# The optimizer's m_size is 3 times the policy (Q1, Q2, and Value)
self.m_size = 3 * self.policy.m_size
self._create_inputs_and_outputs()
self.learning_rate = ModelUtils.create_learning_rate(
lr_schedule, lr, self.policy.global_step, int(max_step))
self._create_losses(
self.policy_network.q1_heads,
self.policy_network.q2_heads,
lr,
int(max_step),
stream_names,
discrete=not self.policy.use_continuous_act,
)
self._create_sac_optimizer_ops()
self.selected_actions = (self.policy.selected_actions
) # For GAIL and other reward signals
if self.policy.normalize:
target_update_norm = self.target_network.copy_normalization(
self.policy.running_mean,
self.policy.running_variance,
self.policy.normalization_steps,
)
# Update the normalization of the optimizer when the policy does.
self.policy.update_normalization_op = tf.group([
self.policy.update_normalization_op, target_update_norm
])
self.policy.initialize_or_load()
self.stats_name_to_update_name = {
"Losses/Value Loss": "value_loss",
"Losses/Policy Loss": "policy_loss",
"Losses/Q1 Loss": "q1_loss",
"Losses/Q2 Loss": "q2_loss",
"Policy/Entropy Coeff": "entropy_coef",
"Policy/Learning Rate": "learning_rate",
}
self.update_dict = {
"value_loss": self.total_value_loss,
"policy_loss": self.policy_loss,
"q1_loss": self.q1_loss,
"q2_loss": self.q2_loss,
"entropy_coef": self.ent_coef,
"entropy": self.policy.entropy,
"update_batch": self.update_batch_policy,
"update_value": self.update_batch_value,
"update_entropy": self.update_batch_entropy,
"learning_rate": self.learning_rate,
}