def __init__(self, policy: TFPolicy, trainer_params: 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 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)
hyperparameters: PPOSettings = cast(
PPOSettings, trainer_params.hyperparameters)
lr = float(hyperparameters.learning_rate)
self._schedule = hyperparameters.learning_rate_schedule
epsilon = float(hyperparameters.epsilon)
beta = float(hyperparameters.beta)
max_step = float(trainer_params.max_steps)
policy_network_settings = policy.network_settings
h_size = int(policy_network_settings.hidden_units)
num_layers = policy_network_settings.num_layers
vis_encode_type = policy_network_settings.vis_encode_type
self.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",
"Policy/Epsilon": "decay_epsilon",
"Policy/Beta": "decay_beta",
}
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_schedule(
self._schedule,
lr,
self.policy.global_step,
int(max_step),
min_value=1e-10,
)
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,
"decay_epsilon": self.decay_epsilon,
"decay_beta": self.decay_beta,
})
self.policy.initialize_or_load()
def _create_losses(self, probs, old_probs, value_heads, entropy, beta,
epsilon, lr, max_step):
"""
Creates training-specific Tensorflow ops for PPO models.
:param probs: Current policy probabilities
:param old_probs: Past policy probabilities
:param value_heads: Value estimate tensors from each value stream
:param beta: Entropy regularization strength
:param entropy: Current policy entropy
:param epsilon: Value for policy-divergence threshold
:param lr: Learning rate
:param max_step: Total number of training steps.
"""
self.returns_holders = {}
self.old_values = {}
for name in value_heads.keys():
returns_holder = tf.placeholder(shape=[None],
dtype=tf.float32,
name=f"{name}_returns")
old_value = tf.placeholder(shape=[None],
dtype=tf.float32,
name=f"{name}_value_estimate")
self.returns_holders[name] = returns_holder
self.old_values[name] = old_value
self.advantage = tf.placeholder(shape=[None],
dtype=tf.float32,
name="advantages")
advantage = tf.expand_dims(self.advantage, -1)
self.decay_epsilon = ModelUtils.create_schedule(
self._schedule,
epsilon,
self.policy.global_step,
max_step,
min_value=0.1)
self.decay_beta = ModelUtils.create_schedule(self._schedule,
beta,
self.policy.global_step,
max_step,
min_value=1e-5)
value_losses = []
for name, head in value_heads.items():
clipped_value_estimate = self.old_values[name] + tf.clip_by_value(
tf.reduce_sum(head, axis=1) - self.old_values[name],
-self.decay_epsilon,
self.decay_epsilon,
)
v_opt_a = tf.squared_difference(self.returns_holders[name],
tf.reduce_sum(head, axis=1))
v_opt_b = tf.squared_difference(self.returns_holders[name],
clipped_value_estimate)
value_loss = tf.reduce_mean(
tf.dynamic_partition(tf.maximum(v_opt_a, v_opt_b),
self.policy.mask, 2)[1])
value_losses.append(value_loss)
self.value_loss = tf.reduce_mean(value_losses)
r_theta = tf.exp(probs - old_probs)
p_opt_a = r_theta * advantage
p_opt_b = (tf.clip_by_value(r_theta, 1.0 - self.decay_epsilon,
1.0 + self.decay_epsilon) * advantage)
self.policy_loss = -tf.reduce_mean(
tf.dynamic_partition(tf.minimum(p_opt_a, p_opt_b),
self.policy.mask, 2)[1])
# For cleaner stats reporting
self.abs_policy_loss = tf.abs(self.policy_loss)
self.loss = (
self.policy_loss + 0.5 * self.value_loss -
self.decay_beta * tf.reduce_mean(
tf.dynamic_partition(entropy, self.policy.mask, 2)[1]))
def __init__(self, policy: TFPolicy, trainer_params: TrainerSettings):
"""
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)
hyperparameters: SACSettings = cast(
SACSettings, trainer_params.hyperparameters)
lr = hyperparameters.learning_rate
lr_schedule = hyperparameters.learning_rate_schedule
max_step = trainer_params.max_steps
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
h_size = policy_network_settings.hidden_units
num_layers = policy_network_settings.num_layers
vis_encode_type = policy_network_settings.vis_encode_type
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
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_schedule(
lr_schedule,
lr,
self.policy.global_step,
int(max_step),
min_value=1e-10,
)
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,
"update_batch": self.update_batch_policy,
"update_value": self.update_batch_value,
"update_entropy": self.update_batch_entropy,
"learning_rate": self.learning_rate,
}
