def _init_opt(self):
"""Initialize optimizater.
Raises:
NotImplementedError: Raise if the policy is recurrent.
"""
# Input variables
(pol_loss_inputs, pol_opt_inputs, infer_loss_inputs,
infer_opt_inputs) = self._build_inputs()
self._policy_opt_inputs = pol_opt_inputs
self._inference_opt_inputs = infer_opt_inputs
# Jointly optimize policy and encoder network
pol_loss, pol_kl, _ = self._build_policy_loss(pol_loss_inputs)
self._optimizer.update_opt(loss=pol_loss,
target=self.policy,
leq_constraint=(pol_kl, self._max_kl_step),
inputs=flatten_inputs(
self._policy_opt_inputs),
constraint_name='mean_kl')
# Optimize inference distribution separately (supervised learning)
infer_loss, _ = self._build_inference_loss(infer_loss_inputs)
self.inference_optimizer.update_opt(loss=infer_loss,
target=self._inference,
inputs=flatten_inputs(
self._inference_opt_inputs))
def _dual_opt_input_values(self, episodes):
"""Update dual func optimize input values based on samples data.
Args:
episodes (EpisodeBatch): Batch of episodes.
Returns:
list(np.ndarray): Flatten dual function optimization input values.
"""
agent_infos = episodes.padded_agent_infos
policy_state_info_list = [
agent_infos[k] for k in self.policy.state_info_keys
]
# pylint: disable=unexpected-keyword-arg
dual_opt_input_values = self._dual_opt_inputs._replace(
reward_var=episodes.padded_rewards,
valid_var=episodes.valids,
feat_diff=self._feat_diff,
param_eta=self._param_eta,
param_v=self._param_v,
policy_state_info_vars_list=policy_state_info_list,
)
return flatten_inputs(dual_opt_input_values)
def _policy_opt_input_values(self, episodes):
"""Update policy optimize input values based on samples data.
Args:
episodes (EpisodeBatch): Batch of episodes.
Returns:
list(np.ndarray): Flatten policy optimization input values.
"""
agent_infos = episodes.padded_agent_infos
policy_state_info_list = [
agent_infos[k] for k in self.policy.state_info_keys
]
actions = [
self._env_spec.action_space.flatten_n(act)
for act in episodes.actions_list
]
padded_actions = episodes.pad_to_last(np.concatenate(actions))
# pylint: disable=unexpected-keyword-arg
policy_opt_input_values = self._policy_opt_inputs._replace(
obs_var=episodes.padded_observations,
action_var=padded_actions,
reward_var=episodes.padded_rewards,
valid_var=episodes.valids,
feat_diff=self._feat_diff,
param_eta=self._param_eta,
param_v=self._param_v,
policy_state_info_vars_list=policy_state_info_list,
)
return flatten_inputs(policy_opt_input_values)
def _policy_opt_input_values(self, samples_data):
"""Update policy optimize input values based on samples data.
Args:
samples_data (dict): Processed sample data.
See garage.tf.paths_to_tensors() for details.
Returns:
list(np.ndarray): Flatten policy optimization input values.
"""
policy_state_info_list = [
samples_data['agent_infos'][k]
for k in self.policy.state_info_keys
] # yapf: disable
# pylint: disable=unexpected-keyword-arg
policy_opt_input_values = self._policy_opt_inputs._replace(
obs_var=samples_data['observations'],
action_var=samples_data['actions'],
reward_var=samples_data['rewards'],
valid_var=samples_data['valids'],
feat_diff=self._feat_diff,
param_eta=self._param_eta,
param_v=self._param_v,
policy_state_info_vars_list=policy_state_info_list,
)
return flatten_inputs(policy_opt_input_values)
def _policy_opt_input_values(self, samples_data):
"""Map episode samples to the policy optimizer inputs.
Args:
samples_data (dict): Processed sample data.
See process_samples() for details.
Returns:
list(np.ndarray): Flatten policy optimization input values.
"""
policy_state_info_list = [
samples_data['agent_infos'][k] for k in self.policy.state_info_keys
]
embed_state_info_list = [
samples_data['latent_infos'][k]
for k in self.policy.encoder.state_info_keys
]
# pylint: disable=unexpected-keyword-arg
policy_opt_input_values = self._policy_opt_inputs._replace(
obs_var=samples_data['observations'],
action_var=samples_data['actions'],
reward_var=samples_data['rewards'],
baseline_var=samples_data['baselines'],
trajectory_var=samples_data['trajectories'],
task_var=samples_data['tasks'],
latent_var=samples_data['latents'],
valid_var=samples_data['valids'],
policy_state_info_vars_list=policy_state_info_list,
embed_state_info_vars_list=embed_state_info_list,
)
return flatten_inputs(policy_opt_input_values)
def _inference_opt_input_values(self, episodes, embed_eps, embed_ep_infos):
"""Map episode samples to the inference optimizer inputs.
Args:
episodes (EpisodeBatch): Batch of episodes.
embed_eps (np.ndarray): Embedding episodes.
embed_ep_infos (dict): Embedding distribution information.
Returns:
list(np.ndarray): Flatten inference optimization input values.
"""
latents = pad_batch_array(episodes.agent_infos['latent'],
episodes.lengths, self.max_episode_length)
infer_state_info_list = [
embed_ep_infos[k] for k in self._inference.state_info_keys
]
# pylint: disable=unexpected-keyword-arg
inference_opt_input_values = self._inference_opt_inputs._replace(
latent_var=latents,
trajectory_var=embed_eps,
valid_var=episodes.valids,
infer_state_info_vars_list=infer_state_info_list,
)
return flatten_inputs(inference_opt_input_values)
def _policy_opt_input_values(self, episodes, baselines):
"""Map episode samples to the policy optimizer inputs.
Args:
episodes (EpisodeBatch): Batch of episodes.
baselines (np.ndarray): Baseline predictions.
Returns:
list(np.ndarray): Flatten policy optimization input values.
"""
agent_infos = episodes.padded_agent_infos
policy_state_info_list = [
agent_infos[k] for k in self.policy.state_info_keys
]
actions = [
self._env_spec.action_space.flatten_n(act)
for act in episodes.actions_list
]
padded_actions = pad_batch_array(np.concatenate(actions),
episodes.lengths,
self.max_episode_length)
# pylint: disable=unexpected-keyword-arg
policy_opt_input_values = self._policy_opt_inputs._replace(
obs_var=episodes.padded_observations,
action_var=padded_actions,
reward_var=episodes.padded_rewards,
baseline_var=baselines,
valid_var=episodes.valids,
policy_state_info_vars_list=policy_state_info_list,
)
return flatten_inputs(policy_opt_input_values)
def _build_entropy_term(self, i):
"""Build policy entropy tensor.
Args:
i (namedtuple): Collection of variables to compute policy loss.
Returns:
tf.Tensor: Policy entropy.
"""
pol_dist = self._policy_network.dist
with tf.name_scope('policy_entropy'):
if self._use_neg_logli_entropy:
policy_entropy = -pol_dist.log_prob(i.action_var,
name='policy_log_likeli')
else:
policy_entropy = pol_dist.entropy()
# This prevents entropy from becoming negative for small policy std
if self._use_softplus_entropy:
policy_entropy = tf.nn.softplus(policy_entropy)
if self._stop_entropy_gradient:
policy_entropy = tf.stop_gradient(policy_entropy)
# dense form, match the shape of advantage
policy_entropy = tf.reshape(policy_entropy,
[-1, self.max_episode_length])
self._f_policy_entropy = compile_function(
flatten_inputs(self._policy_opt_inputs), policy_entropy)
return policy_entropy
def _init_opt(self):
"""Initialize the optimization procedure."""
pol_loss_inputs, pol_opt_inputs, dual_opt_inputs = self._build_inputs()
self._policy_opt_inputs = pol_opt_inputs
self._dual_opt_inputs = dual_opt_inputs
pol_loss = self._build_policy_loss(pol_loss_inputs)
self._optimizer.update_opt(loss=pol_loss,
target=self.policy,
inputs=flatten_inputs(
self._policy_opt_inputs))
def _init_opt(self):
"""Initialize optimizater."""
pol_loss_inputs, pol_opt_inputs = self._build_inputs()
self._policy_opt_inputs = pol_opt_inputs
pol_loss, pol_kl = self._build_policy_loss(pol_loss_inputs)
self._optimizer.update_opt(loss=pol_loss,
target=self.policy,
leq_constraint=(pol_kl, self._max_kl_step),
inputs=flatten_inputs(
self._policy_opt_inputs),
constraint_name='mean_kl')
def _policy_opt_input_values(self, episodes, baselines, embed_eps):
"""Map episode samples to the policy optimizer inputs.
Args:
episodes (EpisodeBatch): Batch of episodes.
baselines (np.ndarray): Baseline predictions.
embed_eps (np.ndarray): Embedding episodes.
Returns:
list(np.ndarray): Flatten policy optimization input values.
"""
actions = [
self._env_spec.action_space.flatten_n(act)
for act in episodes.actions_list
]
actions = pad_batch_array(np.concatenate(actions), episodes.lengths,
self.max_episode_length)
tasks = pad_batch_array(episodes.env_infos['task_onehot'],
episodes.lengths, self.max_episode_length)
latents = pad_batch_array(episodes.agent_infos['latent'],
episodes.lengths, self.max_episode_length)
agent_infos = episodes.padded_agent_infos
policy_state_info_list = [
agent_infos[k] for k in self.policy.state_info_keys
]
embed_state_info_list = [
agent_infos['latent_' + k]
for k in self.policy.encoder.state_info_keys
]
# pylint: disable=unexpected-keyword-arg
policy_opt_input_values = self._policy_opt_inputs._replace(
obs_var=episodes.padded_observations,
action_var=actions,
reward_var=episodes.padded_rewards,
baseline_var=baselines,
trajectory_var=embed_eps,
task_var=tasks,
latent_var=latents,
valid_var=episodes.valids,
policy_state_info_vars_list=policy_state_info_list,
embed_state_info_vars_list=embed_state_info_list,
)
return flatten_inputs(policy_opt_input_values)
def _build_encoder_kl(self):
"""Build graph for encoder KL divergence.
Returns:
tf.Tensor: Encoder KL divergence.
"""
dist = self._encoder_network.dist
old_dist = self._old_encoder_network.dist
with tf.name_scope('encoder_kl'):
kl = old_dist.kl_divergence(dist)
mean_kl = tf.reduce_mean(kl)
# Diagnostic function
self._f_encoder_kl = compile_function(
flatten_inputs(self._policy_opt_inputs), mean_kl)
return mean_kl
def _inference_opt_input_values(self, samples_data):
"""Map episode samples to the inference optimizer inputs.
Args:
samples_data (dict): Processed sample data.
See process_samples() for details.
Returns:
list(np.ndarray): Flatten inference optimization input values.
"""
infer_state_info_list = [
samples_data['trajectory_infos'][k]
for k in self._inference.state_info_keys
]
# pylint: disable=unexpected-keyword-arg
inference_opt_input_values = self._inference_opt_inputs._replace(
latent_var=samples_data['latents'],
trajectory_var=samples_data['trajectories'],
valid_var=samples_data['valids'],
infer_state_info_vars_list=infer_state_info_list,
)
return flatten_inputs(inference_opt_input_values)
def _build_policy_loss(self, i):
"""Build policy loss and other output tensors.
Args:
i (namedtuple): Collection of variables to compute policy loss.
Returns:
tf.Tensor: Policy loss.
tf.Tensor: Mean policy KL divergence.
Raises:
NotImplementedError: If is_recurrent is True.
"""
pol_dist = self._policy_network.dist
old_pol_dist = self._old_policy_network.dist
# Initialize dual params
self._param_eta = 15.
self._param_v = np.random.rand(
self._env_spec.observation_space.flat_dim * 2 + 4)
with tf.name_scope('bellman_error'):
delta_v = tf.boolean_mask(i.reward_var,
i.valid_var) + tf.tensordot(
i.feat_diff, i.param_v, 1)
with tf.name_scope('policy_loss'):
ll = pol_dist.log_prob(i.action_var)
ll = tf.boolean_mask(ll, i.valid_var)
loss = -tf.reduce_mean(
ll * tf.exp(delta_v / i.param_eta -
tf.reduce_max(delta_v / i.param_eta)))
reg_params = self.policy.get_regularizable_vars()
loss += self._l2_reg_loss * tf.reduce_sum(
[tf.reduce_mean(tf.square(param))
for param in reg_params]) / len(reg_params)
with tf.name_scope('kl'):
kl = old_pol_dist.kl_divergence(pol_dist)
pol_mean_kl = tf.reduce_mean(kl)
with tf.name_scope('dual'):
dual_loss = i.param_eta * self._epsilon + (
i.param_eta * tf.math.log(
tf.reduce_mean(
tf.exp(delta_v / i.param_eta -
tf.reduce_max(delta_v / i.param_eta)))) +
i.param_eta * tf.reduce_max(delta_v / i.param_eta))
dual_loss += self._l2_reg_dual * (tf.square(i.param_eta) +
tf.square(1 / i.param_eta))
dual_grad = tf.gradients(dual_loss, [i.param_eta, i.param_v])
self._f_dual = compile_function(
flatten_inputs(self._dual_opt_inputs),
dual_loss)
self._f_dual_grad = compile_function(
flatten_inputs(self._dual_opt_inputs),
dual_grad)
self._f_policy_kl = compile_function(
flatten_inputs(self._policy_opt_inputs),
pol_mean_kl)
return loss
def _build_policy_loss(self, i):
"""Build policy loss and other output tensors.
Args:
i (namedtuple): Collection of variables to compute policy loss.
Returns:
tf.Tensor: Policy loss.
tf.Tensor: Mean policy KL divergence.
"""
# pylint: disable=too-many-statements
self._policy_network, self._encoder_network = (self.policy.build(
i.augmented_obs_var, i.task_var, name='loss_policy'))
self._old_policy_network, self._old_encoder_network = (
self._old_policy.build(i.augmented_obs_var,
i.task_var,
name='loss_old_policy'))
self._infer_network = self._inference.build(i.augmented_traj_var,
name='loss_infer')
self._old_infer_network = self._old_inference.build(
i.augmented_traj_var, name='loss_old_infer')
pol_dist = self._policy_network.dist
old_pol_dist = self._old_policy_network.dist
# Entropy terms
encoder_entropy, inference_ce, policy_entropy = (
self._build_entropy_terms(i))
# Augment the path rewards with entropy terms
with tf.name_scope('augmented_rewards'):
rewards = (i.reward_var -
(self.inference_ce_coeff * inference_ce) +
(self._policy_ent_coeff * policy_entropy))
with tf.name_scope('policy_loss'):
with tf.name_scope('advantages'):
adv = compute_advantages(self._discount,
self._gae_lambda,
self.max_episode_length,
i.baseline_var,
rewards,
name='advantages')
adv = tf.reshape(adv, [-1, self.max_episode_length])
# Optionally normalize advantages
eps = tf.constant(1e-8, dtype=tf.float32)
if self._center_adv:
adv = center_advs(adv, axes=[0], eps=eps)
if self._positive_adv:
adv = positive_advs(adv, eps)
# Calculate loss function and KL divergence
with tf.name_scope('kl'):
kl = old_pol_dist.kl_divergence(pol_dist)
pol_mean_kl = tf.reduce_mean(kl)
ll = pol_dist.log_prob(i.action_var, name='log_likelihood')
# Calculate surrogate loss
with tf.name_scope('surr_loss'):
old_ll = old_pol_dist.log_prob(i.action_var)
old_ll = tf.stop_gradient(old_ll)
# Clip early to avoid overflow
lr = tf.exp(
tf.minimum(ll - old_ll, np.log(1 + self._lr_clip_range)))
surrogate = lr * adv
surrogate = tf.debugging.check_numerics(surrogate,
message='surrogate')
# Finalize objective function
with tf.name_scope('loss'):
lr_clip = tf.clip_by_value(lr,
1 - self._lr_clip_range,
1 + self._lr_clip_range,
name='lr_clip')
surr_clip = lr_clip * adv
obj = tf.minimum(surrogate, surr_clip, name='surr_obj')
obj = tf.boolean_mask(obj, i.valid_var)
# Maximize E[surrogate objective] by minimizing
# -E_t[surrogate objective]
loss = -tf.reduce_mean(obj)
# Encoder entropy bonus
loss -= self.encoder_ent_coeff * encoder_entropy
encoder_mean_kl = self._build_encoder_kl()
# Diagnostic functions
self._f_policy_kl = tf.compat.v1.get_default_session(
).make_callable(pol_mean_kl,
feed_list=flatten_inputs(self._policy_opt_inputs))
self._f_rewards = tf.compat.v1.get_default_session().make_callable(
rewards, feed_list=flatten_inputs(self._policy_opt_inputs))
returns = discounted_returns(self._discount,
self.max_episode_length,
rewards,
name='returns')
self._f_returns = tf.compat.v1.get_default_session().make_callable(
returns, feed_list=flatten_inputs(self._policy_opt_inputs))
return loss, pol_mean_kl, encoder_mean_kl
def _build_policy_loss(self, i):
"""Build policy loss and other output tensors.
Args:
i (namedtuple): Collection of variables to compute policy loss.
Returns:
tf.Tensor: Policy loss.
tf.Tensor: Mean policy KL divergence.
"""
policy_entropy = self._build_entropy_term(i)
rewards = i.reward_var
if self._maximum_entropy:
with tf.name_scope('augmented_rewards'):
rewards = i.reward_var + (self._policy_ent_coeff *
policy_entropy)
with tf.name_scope('policy_loss'):
adv = compute_advantages(self._discount,
self._gae_lambda,
self.max_episode_length,
i.baseline_var,
rewards,
name='adv')
adv = tf.reshape(adv, [-1, self.max_episode_length])
# Optionally normalize advantages
eps = tf.constant(1e-8, dtype=tf.float32)
if self._center_adv:
adv = center_advs(adv, axes=[0], eps=eps)
if self._positive_adv:
adv = positive_advs(adv, eps)
old_policy_dist = self._old_policy_network.dist
policy_dist = self._policy_network.dist
with tf.name_scope('kl'):
kl = old_policy_dist.kl_divergence(policy_dist)
pol_mean_kl = tf.reduce_mean(kl)
# Calculate vanilla loss
with tf.name_scope('vanilla_loss'):
ll = policy_dist.log_prob(i.action_var, name='log_likelihood')
vanilla = ll * adv
# Calculate surrogate loss
with tf.name_scope('surrogate_loss'):
lr = tf.exp(ll - old_policy_dist.log_prob(i.action_var))
surrogate = lr * adv
# Finalize objective function
with tf.name_scope('loss'):
if self._pg_loss == 'vanilla':
# VPG uses the vanilla objective
obj = tf.identity(vanilla, name='vanilla_obj')
elif self._pg_loss == 'surrogate':
# TRPO uses the standard surrogate objective
obj = tf.identity(surrogate, name='surr_obj')
elif self._pg_loss == 'surrogate_clip':
lr_clip = tf.clip_by_value(lr,
1 - self._lr_clip_range,
1 + self._lr_clip_range,
name='lr_clip')
surr_clip = lr_clip * adv
obj = tf.minimum(surrogate, surr_clip, name='surr_obj')
if self._entropy_regularzied:
obj += self._policy_ent_coeff * policy_entropy
# filter only the valid values
obj = tf.boolean_mask(obj, i.valid_var)
# Maximize E[surrogate objective] by minimizing
# -E_t[surrogate objective]
loss = -tf.reduce_mean(obj)
# Diagnostic functions
self._f_policy_kl = tf.compat.v1.get_default_session(
).make_callable(pol_mean_kl,
feed_list=flatten_inputs(self._policy_opt_inputs))
self._f_rewards = tf.compat.v1.get_default_session().make_callable(
rewards, feed_list=flatten_inputs(self._policy_opt_inputs))
returns = discounted_returns(self._discount,
self.max_episode_length, rewards)
self._f_returns = tf.compat.v1.get_default_session().make_callable(
returns, feed_list=flatten_inputs(self._policy_opt_inputs))
return loss, pol_mean_kl
def _build_entropy_terms(self, i):
"""Build policy entropy tensor.
Args:
i (namedtuple): Collection of variables to compute policy loss.
Returns:
tf.Tensor: Policy entropy.
"""
pol_dist = self._policy_network.dist
infer_dist = self._infer_network.dist
enc_dist = self._encoder_network.dist
with tf.name_scope('entropy_terms'):
# 1. Encoder distribution total entropy
with tf.name_scope('encoder_entropy'):
encoder_dist, _, _ = self.policy.encoder.build(
i.task_var, name='encoder_entropy').outputs
encoder_all_task_entropies = -encoder_dist.log_prob(
i.latent_var)
if self._use_softplus_entropy:
encoder_entropy = tf.nn.softplus(
encoder_all_task_entropies)
encoder_entropy = tf.reduce_mean(encoder_entropy,
name='encoder_entropy')
encoder_entropy = tf.stop_gradient(encoder_entropy)
# 2. Infernece distribution cross-entropy (log-likelihood)
with tf.name_scope('inference_ce'):
# Build inference with trajectory windows
traj_ll = infer_dist.log_prob(
enc_dist.sample(seed=deterministic.get_tf_seed_stream()),
name='traj_ll')
inference_ce_raw = -traj_ll
inference_ce = tf.clip_by_value(inference_ce_raw, -3, 3)
if self._use_softplus_entropy:
inference_ce = tf.nn.softplus(inference_ce)
if self._stop_ce_gradient:
inference_ce = tf.stop_gradient(inference_ce)
# 3. Policy path entropies
with tf.name_scope('policy_entropy'):
policy_entropy = -pol_dist.log_prob(i.action_var,
name='policy_log_likeli')
# This prevents entropy from becoming negative
# for small policy std
if self._use_softplus_entropy:
policy_entropy = tf.nn.softplus(policy_entropy)
policy_entropy = tf.stop_gradient(policy_entropy)
# Diagnostic functions
self._f_task_entropies = compile_function(
flatten_inputs(self._policy_opt_inputs),
encoder_all_task_entropies)
self._f_encoder_entropy = compile_function(
flatten_inputs(self._policy_opt_inputs), encoder_entropy)
self._f_inference_ce = compile_function(
flatten_inputs(self._policy_opt_inputs),
tf.reduce_mean(inference_ce * i.valid_var))
self._f_policy_entropy = compile_function(
flatten_inputs(self._policy_opt_inputs), policy_entropy)
return encoder_entropy, inference_ce, policy_entropy