def _sample_and_transpose_actions_and_log_probs(
self,
time_steps: ts.TimeStep,
num_action_samples: int,
training: Optional[bool] = False
) -> Tuple[types.Tensor, types.Tensor]:
"""Samples actions and corresponding log probabilities from policy."""
# Get raw action distribution from policy, and initialize bijectors list.
batch_size = nest_utils.get_outer_shape(time_steps,
self._time_step_spec)[0]
policy_state = self._train_policy.get_initial_state(batch_size)
if training:
action_distribution = self._train_policy.distribution(
time_steps, policy_state=policy_state).action
else:
action_distribution = self._policy.distribution(
time_steps, policy_state=policy_state).action
actions = tf.nest.map_structure(
lambda d: d.sample(num_action_samples,
seed=self._action_seed_stream()),
action_distribution)
log_pi = common.log_probability(action_distribution, actions,
self.action_spec)
# Swap the first two axes for a [batch, self._num_cql_samples, ...] shape.
actions = self._transpose_tile_and_batch_dims(actions)
log_pi = self._transpose_tile_and_batch_dims(log_pi)
return actions, log_pi
def testNestedLogProbability(self):
action_spec = [
tensor_spec.BoundedTensorSpec([2], tf.float32, -1, 1),
[
tensor_spec.BoundedTensorSpec([1], tf.float32, -1, 1),
tensor_spec.BoundedTensorSpec([1], tf.float32, -1, 1)
]
]
distribution = [
tfp.distributions.Normal([0.0, 0.0], [1.0, 1.0]),
[
tfp.distributions.Normal([0.5], [1.0]),
tfp.distributions.Normal([-0.5], [1.0])
]
]
actions = [
tf.constant([0.0, 0.0]), [tf.constant([0.5]),
tf.constant([-0.5])]
]
log_probs = common.log_probability(distribution, actions, action_spec)
self.evaluate(tf.compat.v1.global_variables_initializer())
log_probs_ = self.evaluate(log_probs)
self.assertEqual(len(log_probs_.shape), 0)
self.assertNear(log_probs_, 4 * -0.5 * np.log(2 * 3.14159), 0.001)
def _construct(self, batch_size, graph):
"""Construct the agent graph through placeholders."""
self._batch_size = batch_size
self._batched = batch_size is not None
outer_dims = [self._batch_size] if self._batched else [1]
with graph.as_default():
self._time_step = tensor_spec.to_nest_placeholder(
self._tf_policy.time_step_spec, outer_dims=outer_dims)
self._tf_initial_state = self._tf_policy.get_initial_state(
batch_size=self._batch_size or 1)
self._policy_state = tf.nest.map_structure(
lambda ps: tf.compat.v1.placeholder( # pylint: disable=g-long-lambda
ps.dtype,
ps.shape,
name='policy_state'),
self._tf_initial_state)
self._action_step = self._tf_policy.action(self._time_step,
self._policy_state,
seed=self._seed)
self._actions = tensor_spec.to_nest_placeholder(
self._tf_policy.action_spec, outer_dims=outer_dims)
self._action_distribution = self._tf_policy.distribution(
self._time_step, policy_state=self._policy_state).action
self._action_mean = self._action_distribution.mean()
self._log_prob = common.log_probability(
self._action_distribution, self._actions,
self._tf_policy.action_spec)
def testBatchedNestedLogProbability(self):
action_spec = [
tensor_spec.BoundedTensorSpec([2], tf.float32, -1, 1),
[
tensor_spec.BoundedTensorSpec([1], tf.float32, -1, 1),
tensor_spec.BoundedTensorSpec([1], tf.float32, -1, 1)
]
]
distribution = [
tfp.distributions.Normal([[0.0, 0.0], [0.0, 0.0]],
[[1.0, 1.0], [2.0, 2.0]]),
[
tfp.distributions.Normal([[0.5], [0.5]], [[1.0], [2.0]]),
tfp.distributions.Normal([[-0.5], [-0.5]], [[1.0], [2.0]])
]
]
actions = [
tf.constant([[0.0, 0.0], [0.0, 0.0]]),
[tf.constant([[0.5], [0.5]]),
tf.constant([[-0.5], [-0.5]])]
]
log_probs = common.log_probability(distribution, actions, action_spec)
self.evaluate(tf.compat.v1.global_variables_initializer())
log_probs_ = self.evaluate(log_probs)
self.assertEqual(log_probs_.shape, (2, ))
self.assertAllClose(
log_probs_,
[4 * -0.5 * np.log(2 * 3.14159), 4 * -0.5 * np.log(8 * 3.14159)],
0.001)
def _get_safe_idx(self, safe_ac_mask, fail_prob, sampled_ac, safe_ac_idx,
actions, fail_prob_safe):
if tf.math.count_nonzero(safe_ac_mask) == 0:
# picks safest action
safe_idx = tf.argmin(fail_prob)
else:
sampled_ac = tf.gather(sampled_ac, safe_ac_idx)
# picks most unsafe "safe" action
# safe_idx = tf.argmax(fail_prob_safe, axis=0)
# picks the safest action
# safe_idx = tf.argmin(fail_prob_safe)
if self._training:
# picks random safe_action, weighted by 1 - fail_prob_safe (so higher weight for safer actions)
# safe_idx = tfp.distributions.Categorical([1 - fail_prob_safe]).sample()
if self._sampling_method == 'rejection':
# standard rejection sampling with prob proportional to original policy
log_prob = common.log_probability(actions, sampled_ac,
self.action_spec)
safe_idx = tfp.distributions.Categorical(log_prob).sample()
elif self._sampling_method == 'risky':
# picks random risky safe action, weighted by fail_prob_safe (so higher weight for less safe actions)
safe_idx = tfp.distributions.Categorical([fail_prob_safe
]).sample()
elif self._sampling_method == 'safe':
safe_idx = tfp.distributions.Categorical(
[1 - fail_prob_safe]).sample()
safe_idx = tf.reshape(safe_idx, [-1])[0]
return safe_idx
def _actions_and_log_probs(self, time_steps):
"""Get actions and corresponding log probabilities from policy."""
# Get raw action distribution from policy, and initialize bijectors list.
action_distribution = self.policy().distribution(time_steps).action
if self._squash_actions:
bijectors = []
# Bijector to rescale actions to ranges in action spec.
action_means, action_magnitudes = self._action_spec_means_magnitudes(
)
bijectors.append(
tfp.bijectors.AffineScalar(shift=action_means,
scale=action_magnitudes))
# Bijector to squash actions to range (-1.0, +1.0).
bijectors.append(tanh_bijector_stable.Tanh())
# Chain applies bijectors in reverse order, so squash will happen before
# rescaling to action spec.
bijector_chain = tfp.bijectors.Chain(bijectors)
action_distribution = tfp.distributions.TransformedDistribution(
distribution=action_distribution, bijector=bijector_chain)
# Sample actions and log_pis from transformed distribution.
actions = tf.nest.map_structure(lambda d: d.sample(),
action_distribution)
log_pi = common_utils.log_probability(action_distribution, actions,
self.action_spec())
return actions, log_pi
def policy_gradient_loss(self,
actions_distribution,
actions,
is_boundary,
returns,
num_episodes,
weights=None):
"""Computes the policy gradient loss.
Args:
actions_distribution: A possibly batched tuple of action distributions.
actions: Tensor with a batch of actions.
is_boundary: Tensor of booleans that indicate if the corresponding action
was in a boundary trajectory and should be ignored.
returns: Tensor with a return from each timestep, aligned on index. Works
better when returns are normalized.
num_episodes: Number of episodes contained in the training data.
weights: Optional scalar or element-wise (per-batch-entry) importance
weights. May include a mask for invalid timesteps.
Returns:
policy_gradient_loss: A tensor that will contain policy gradient loss for
the on-policy experience.
"""
# TODO(b/126594799): Add class IndependentNested(tfd.Distribution) to handle
# nests of independent distributions like this.
action_log_prob = common.log_probability(actions_distribution, actions,
self.action_spec)
# Filter out transitions between end state of previous episode and start
# state of next episode.
valid_mask = tf.cast(~is_boundary, tf.float32)
action_log_prob *= valid_mask
action_log_prob_times_return = action_log_prob * returns
if weights is not None:
action_log_prob_times_return *= weights
if self._debug_summaries:
tf.compat.v2.summary.histogram(
name='action_log_prob',
data=action_log_prob,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(
name='action_log_prob_times_return',
data=action_log_prob_times_return,
step=self.train_step_counter)
# Policy gradient loss is defined as the sum, over timesteps, of action
# log-probability times the cumulative return from that timestep onward.
# For more information, see (Williams, 1992).
policy_gradient_loss = -tf.reduce_sum(
input_tensor=action_log_prob_times_return)
# We take the mean over episodes by dividing by num_episodes.
policy_gradient_loss = policy_gradient_loss / num_episodes
return policy_gradient_loss
def testLogProbabilityOneHot(self):
action_spec = tensor_spec.BoundedTensorSpec([3], tf.int32, 0, 1)
distribution = tfp.distributions.OneHotCategorical(probs=[0.6, 0.3, 0.1])
actions = tf.constant([1, 0, 0])
log_probs = common.log_probability(distribution, actions, action_spec)
self.evaluate(tf.compat.v1.global_variables_initializer())
log_probs_ = self.evaluate(log_probs)
self.assertEqual(len(log_probs_.shape), 0)
self.assertNear(log_probs_, np.log(0.6), 0.00001)
def behavior_loss(self, time_steps, actions, weights=None):
with tf.name_scope('behavior_loss'):
nest_utils.assert_same_structure(time_steps, self.time_step_spec)
batch_size = nest_utils.get_outer_shape(time_steps,
self._time_step_spec)[0]
policy_state = self._behavior_policy.get_initial_state(batch_size)
action_distribution = self._behavior_policy.distribution(
time_steps, policy_state=policy_state).action
log_pi = common.log_probability(action_distribution, actions,
self.action_spec)
return -1.0 * tf.reduce_mean(log_pi)
def _ml_pmi(self, x, y, y_distribution):
num_outer_dims = get_outer_rank(x, self._x_spec)
hidden = self._model(x)[0]
batch_squash = BatchSquash(num_outer_dims)
hidden = batch_squash.flatten(hidden)
delta_loc = self._delta_loc_layer(hidden)
delta_scale = tf.nn.softplus(self._delta_scale_layer(hidden))
delta_loc = batch_squash.unflatten(delta_loc)
delta_scale = batch_squash.unflatten(delta_scale)
y_given_x_dist = tfp.distributions.Normal(
loc=y_distribution.loc + delta_loc,
scale=y_distribution.scale * delta_scale)
# Because Normal.event_shape is [], the result of Normal.log_prob() is
# the probabilities of individual dimensions. So we need to use
# tfa_common.log_probability() instead.
# TODO: implement a normal distribution with non-scalar event shape.
pmi = tfa_common.log_probability(y_given_x_dist, y, self._y_spec)
pmi -= tf.stop_gradient(
tfa_common.log_probability(y_distribution, y, self._y_spec))
return pmi
def _line_search(self, time_steps, policy_steps_, advantages,
natural_gradient, coeff, weights):
"""Find new policy parameters by line search in natural gradient direction"""
batch_size = nest_utils.get_outer_shape(time_steps,
self._time_step_spec)[0]
# old policy distribution
action_distribution_parameters = policy_steps_.info
actions = policy_steps_.action
actions_distribution = distribution_spec.nested_distributions_from_specs(
self._action_distribution_spec,
action_distribution_parameters["dist_params"])
act_log_probs = common.log_probability(actions_distribution, actions,
self._action_spec)
# loss for the old policy
loss_threshold = self.policy_gradient_loss(
time_steps,
actions,
tf.stop_gradient(act_log_probs),
tf.stop_gradient(advantages),
actions_distribution,
weights,
)
policy_params = flatten_tensors(self._actor_net.trainable_variables)
# try different steps_sizes, accept first one that improves loss and satisfies KL constraint
for it in range(self._backtrack_iters):
new_params = policy_params - self._backtrack_coeff**it * coeff * natural_gradient
unflatten_tensor(new_params, self._opt_policy_parameters)
opt_policy_state = self._opt_policy.get_initial_state(batch_size)
dists = self._opt_policy.distribution(time_steps, opt_policy_state)
new_policy_distribution = dists.action
kl = tf.reduce_mean(
self._kl_divergence(time_steps, action_distribution_parameters,
new_policy_distribution))
loss = self.policy_gradient_loss(
time_steps,
actions,
tf.stop_gradient(act_log_probs),
tf.stop_gradient(advantages),
new_policy_distribution,
weights,
)
if kl < self._max_kl and loss < loss_threshold:
return new_params
# no improvement found
return policy_params
def _actions_and_log_probs(self, time_steps):
"""Get actions and corresponding log probabilities from policy."""
# Get raw action distribution from policy, and initialize bijectors list.
action_distribution = self.policy.distribution(time_steps).action
# Sample actions and log_pis from transformed distribution.
actions = tf.nest.map_structure(lambda d: d.sample(),
action_distribution)
log_pi = common.log_probability(action_distribution, actions,
self.action_spec)
return actions, log_pi
def policy_gradient(self, time_steps, policy_steps_, advantages, weights):
"""
Compute policy gradient wrt actor_net parameters.
:param time_steps: batch of TimeSteps with observations for each timestep
:param policy_steps_: policy info for time step sampling policy
:param advantages: Tensor of advantage estimate for each timestep, aligned on index.
:param weights: mask for invalid timesteps
:return: list of gradient tensors, policy loss computer on timesteps
"""
batch_size = nest_utils.get_outer_shape(time_steps,
self._time_step_spec)[0]
actions = policy_steps_.action
# get policy info before update
action_distribution_parameters = policy_steps_.info
# Reconstruct per-timestep policy distribution
old_actions_distribution = distribution_spec.nested_distributions_from_specs(
self._action_distribution_spec,
action_distribution_parameters["dist_params"])
# Log probability of actions taken during data collection
act_log_probs = common.log_probability(old_actions_distribution,
actions, self._action_spec)
with tf.GradientTape() as tape:
# current policy distribution
policy_state = self._collect_policy.get_initial_state(batch_size)
distribution_step = self._collect_policy.distribution(
time_steps, policy_state)
current_policy_distribution = distribution_step.action
policy_gradient_loss = self.policy_gradient_loss(
time_steps,
actions,
tf.stop_gradient(act_log_probs),
tf.stop_gradient(advantages),
current_policy_distribution,
weights,
)
trainable = self._actor_net.trainable_weights
grads = tape.gradient(policy_gradient_loss, trainable)
for g in grads:
tf.debugging.check_numerics(g,
"Gradient divergence",
name="grad_check")
return policy_gradient_loss, grads
def _actions_and_log_probs(self, time_steps):
"""Get actions and corresponding log probabilities from policy."""
# Get raw action distribution from policy, and initialize bijectors list.
batch_size = nest_utils.get_outer_shape(time_steps, self._time_step_spec)[0]
policy_state = self._train_policy.get_initial_state(batch_size)
action_distribution = self._train_policy.distribution(
time_steps, policy_state=policy_state).action
# Sample actions and log_pis from transformed distribution.
actions = tf.nest.map_structure(lambda d: d.sample(), action_distribution)
log_pi = common.log_probability(action_distribution, actions,
self.action_spec)
return actions, log_pi
def _actions_and_log_probs(self, time_steps, safety_constrained=False):
"""Get actions and corresponding log probabilities from policy."""
# Get raw action distribution from policy, and initialize bijectors list.
batch_size = nest_utils.get_outer_shape(time_steps, self.time_step_spec)[0]
policy = self.collect_policy
policy_state = policy.get_initial_state(batch_size)
action_distribution = policy.distribution(
time_steps, policy_state=policy_state).action
# Sample actions and log_pis from transformed distribution.
if safety_constrained:
actions, policy_state = self.safe_policy._apply_actor_network(
time_steps.observation, time_steps.step_type, policy_state)
else:
actions = tf.nest.map_structure(lambda d: d.sample(), action_distribution)
log_pi = common.log_probability(action_distribution, actions,
self.action_spec)
return actions, log_pi
def policy_gradient_loss(self, time_steps, actions, returns, weights=None):
"""Computes the policy gradient loss.
Args:
time_steps: TimeStep object with a batch of observations.
actions: Tensor with a batch of actions.
returns: Tensor with a return from each timestep, aligned on index. Works
better when returns are normalized.
weights: Optional scalar or element-wise (per-batch-entry) importance
weights. May include a mask for invalid timesteps.
Returns:
policy_gradient_loss: A tensor that will contain policy gradient loss for
the on-policy experience.
"""
tf.nest.assert_same_structure(time_steps, self.time_step_spec())
actions_distribution = self.collect_policy().distribution(
time_steps).action
# TODO(kbanoop): Add class IndependentNested(tfd.Distribution) to handle
# nests of independent distributions like this.
action_log_prob = common.log_probability(actions_distribution, actions,
self.action_spec())
action_log_prob_times_return = action_log_prob * returns
if weights is not None:
action_log_prob_times_return *= weights
if self._debug_summaries:
tf.contrib.summary.histogram('action_log_prob', action_log_prob)
tf.contrib.summary.histogram('action_log_prob_times_return',
action_log_prob_times_return)
# Policy gradient loss is defined as the sum, over timesteps, of action
# log-probability times the cumulative return from that timestep onward.
# For more information, see (Williams, 1992)
policy_gradient_loss = -tf.reduce_sum(
input_tensor=action_log_prob_times_return)
with tf.name_scope('Losses/'):
tf.contrib.summary.scalar('policy_gradient_loss',
policy_gradient_loss)
return tf_agent.LossInfo(policy_gradient_loss, ())
def train_step(self, exp: Experience, state: SacState):
action_distribution, share_actor_state = self._actor_network(
exp.observation,
step_type=exp.step_type,
network_state=state.share.actor)
action = tf.nest.map_structure(lambda d: d.sample(),
action_distribution)
log_pi = tfa_common.log_probability(action_distribution, action,
self._action_spec)
actor_state, actor_info = self._actor_train_step(
exp, state.actor, action_distribution, action, log_pi)
critic_state, critic_info = self._critic_train_step(
exp, state.critic, action, log_pi)
alpha_info = self._alpha_train_step(log_pi)
state = SacState(share=SacShareState(actor=share_actor_state),
actor=actor_state,
critic=critic_state)
info = SacInfo(actor=actor_info, critic=critic_info, alpha=alpha_info)
return PolicyStep(action_distribution, state, info)
def _actions_and_log_probs(self,
time_steps: ts.TimeStep,
training: Optional[bool] = False
) -> Tuple[types.Tensor, types.Tensor]:
"""Get actions and corresponding log probabilities from policy."""
# Get raw action distribution from policy, and initialize bijectors list.
batch_size = nest_utils.get_outer_shape(time_steps,
self._time_step_spec)[0]
policy_state = self._train_policy.get_initial_state(batch_size)
if training:
action_distribution = self._train_policy.distribution(
time_steps, policy_state=policy_state).action
else:
action_distribution = self._policy.distribution(
time_steps, policy_state=policy_state).action
# Sample actions and log_pis from transformed distribution.
actions = tf.nest.map_structure(
lambda d: d.sample((), seed=self._action_seed_stream()),
action_distribution)
log_pi = common.log_probability(action_distribution, actions,
self.action_spec)
return actions, log_pi
def policy_gradient_loss(
self,
time_steps,
actions,
sample_action_log_probs,
advantages,
current_policy_distribution,
weights,
):
"""Create tensor for policy gradient loss.
All tensors should have a single batch dimension.
Args:
time_steps: TimeSteps with observations for each timestep.
actions: Tensor of actions for timesteps, aligned on index.
sample_action_log_probs: Tensor of sample probability of each action.
advantages: Tensor of advantage estimate for each timestep, aligned on
index. Works better when advantage estimates are normalized.
current_policy_distribution: The policy distribution, evaluated on all
time_steps.
weights: Optional scalar or element-wise (per-batch-entry) importance
weights. Includes a mask for invalid timesteps.
Returns:
policy_gradient_loss: A tensor that will contain policy gradient loss for
the on-policy experience.
"""
tf.nest.assert_same_structure(time_steps, self.time_step_spec)
action_log_prob = common.log_probability(current_policy_distribution,
actions, self._action_spec)
action_log_prob = tf.cast(action_log_prob, tf.float32)
if self._log_prob_clipping > 0.0:
action_log_prob = tf.clip_by_value(action_log_prob,
-self._log_prob_clipping,
self._log_prob_clipping)
tf.debugging.check_numerics(action_log_prob, "action_log_prob")
tf.debugging.check_numerics(sample_action_log_probs,
"sample_action_log_probs")
# Prepare unclipped importance ratios.
importance_ratio = tf.exp(action_log_prob - sample_action_log_probs)
tf.debugging.check_numerics(importance_ratio,
"importance_ratio",
name="importance_ratio")
per_timestep_objective = importance_ratio * advantages
policy_gradient_loss = -per_timestep_objective
policy_gradient_loss = tf.reduce_mean(
input_tensor=policy_gradient_loss * weights)
tf.debugging.check_numerics(policy_gradient_loss,
"Policy Loss divergence",
name="policy_check")
return policy_gradient_loss
def _soft_relabel(self, experience):
# experience.observation.shape = [B x T=2 x obs_dim+state_dim]
states, orig_tasks = self._task_distribution.split(
experience.observation[:, 0])
if self._task_distribution.tasks is None:
tasks = orig_tasks
else:
tasks = tf.constant(self._task_distribution.tasks,
dtype=tf.float32)
next_states, _ = self._task_distribution.split(
experience.observation[:, 1])
if self._candidate_task_type == "states":
candidate_tasks = self._task_distribution.state_to_task(states)
elif self._candidate_task_type == "next_states":
candidate_tasks = self._task_distribution.state_to_task(
next_states)
else:
assert self._candidate_task_type == "tasks"
candidate_tasks = tasks
actions = experience.action[:, 0]
num_tasks = tasks.shape[0]
batch_size = states.shape[0]
task_dim = tasks.shape[1]
obs_dim = states.shape[1]
action_dim = actions.shape[1]
action_spec = self._actor.output_tensor_spec
states_tiled = tf.tile(states[:, None], [1, num_tasks, 1]) # B x B x D
states_tiled = tf.reshape(states_tiled,
[batch_size * num_tasks, obs_dim]) # B*B x D
actions_tiled = tf.tile(actions[:, None],
[1, num_tasks, 1]) # B x B x D
actions_tiled = tf.reshape(
actions_tiled, [batch_size * num_tasks, action_dim]) # B*B x D
tasks_tiled = tf.tile(tasks[None], [batch_size, 1, 1]) # B x B x D
tasks_tiled = tf.reshape(tasks_tiled,
[batch_size * num_tasks, task_dim]) # B*B x D
next_states_tiled = tf.tile(next_states[:, None], [1, num_tasks, 1])
next_states_tiled = tf.reshape(
next_states_tiled, [batch_size * num_tasks, obs_dim]) # B*B x D
next_relabelled_obs = self._task_distribution.combine(
next_states_tiled, tasks_tiled)
sampled_actions_tiled = self._actor(next_relabelled_obs,
step_type=(),
network_state=())[0].sample()
critic_input = (next_relabelled_obs, sampled_actions_tiled)
q_vals, _ = self._critic(critic_input, training=False)
q_vals_vec = tf.reshape(q_vals, (batch_size, num_tasks))
rewards, dones = self._task_distribution.evaluate(
states_tiled, actions_tiled, tasks_tiled)
dones = tf.cast(dones, tf.float32)
rewards_vec = tf.reshape(rewards, (batch_size, num_tasks))
dones_vec = tf.reshape(dones, (batch_size, num_tasks))
relabelled_obs = self._task_distribution.combine(
states_tiled, tasks_tiled)
action_distribution = self._actor(relabelled_obs,
step_type=(),
network_state=())[0]
log_pi = common.log_probability(action_distribution, actions_tiled,
action_spec)
log_pi_vec = tf.reshape(log_pi, (batch_size, num_tasks))
logits_vec = (rewards_vec - log_pi_vec + self._gamma *
(1.0 - dones_vec) * q_vals_vec)
if self._relabel_type == "random":
logits_vec = tf.ones_like(
logits_vec) # Hack to make sampling random
## End new version
if self._normalize_cols:
logits_vec = logits_vec - tf.math.reduce_logsumexp(logits_vec,
axis=0)[None]
relabel_indices = tf.random.categorical(logits=logits_vec,
num_samples=1)
### Metrics
global_step = tf.compat.v1.train.get_or_create_global_step()
orig_indices = tf.range(self._sample_batch_size,
dtype=relabel_indices.dtype)
with tf.name_scope("relabelling"):
# How often are the originally commanded goals most optimal?
opt_indices = tf.argmax(logits_vec, axis=1)
orig_is_opt = opt_indices == orig_indices
orig_opt_frac = tf.reduce_mean(tf.cast(orig_is_opt, tf.float32))
tf.compat.v2.summary.scalar(name="orig_task_optimal",
data=orig_opt_frac,
step=global_step)
# How often is the relabelled goal optimal?
# The relabel_indices are [B, 1], so we need to remove the extra dim.
relabel_is_opt = tf.squeeze(relabel_indices) == orig_indices
relabel_opt_frac = tf.reduce_mean(
tf.cast(relabel_is_opt, tf.float32))
tf.compat.v2.summary.scalar(name="relabel_task_optimal",
data=relabel_opt_frac,
step=global_step)
# What are the average Q values of the original tasks?
if batch_size == num_tasks:
indices = tf.transpose(
tf.stack([orig_indices, orig_indices], axis=0))
orig_q_vals = tf.gather_nd(logits_vec, indices)
tf.compat.v2.summary.scalar(
name="orig_q_vals",
data=tf.reduce_mean(orig_q_vals),
step=global_step,
)
# What are the average Q values of the relabelled tasks?
indices = tf.transpose(
tf.stack(
[orig_indices, tf.squeeze(relabel_indices)], axis=0))
relabel_q_vals = tf.gather_nd(logits_vec, indices)
tf.compat.v2.summary.scalar(
name="relabel_q_vals",
data=tf.reduce_mean(relabel_q_vals),
step=global_step,
)
max_q = tf.reduce_max(logits_vec, axis=1)
tf.compat.v2.summary.scalar(name="max_q",
data=tf.reduce_mean(max_q),
step=global_step)
### End metrics
# For both state-centric and goal-centric relabelling, the implementation of
# mixing is the same: we randomly replace some of the indices with the
# diagonal.
relabelled_tasks = tf.gather(candidate_tasks,
tf.squeeze(relabel_indices))
if self._relabel_prob == 0:
relabelled_tasks = orig_tasks
elif 0 < self._relabel_prob < 1:
logits = tf.log([1.0 - self._relabel_prob, self._relabel_prob])
mask = tf.squeeze(
tf.random.categorical(logits[None],
num_samples=self._sample_batch_size))
mask = tf.cast(mask, tf.float32)[:, None]
relabelled_tasks = mask * orig_tasks + (1 -
mask) * relabelled_tasks
states_and_tasks = self._task_distribution.combine(
states, relabelled_tasks)
next_states_and_tasks = self._task_distribution.combine(
next_states, relabelled_tasks)
new_observation = tf.concat(
[states_and_tasks[:, None], next_states_and_tasks[:, None]],
axis=1)
assert new_observation.shape == experience.observation.shape
experience = experience.replace(observation=new_observation)
return experience
def policy_gradient_loss(self,
time_steps,
actions,
sample_action_log_probs,
advantages,
current_policy_distribution,
valid_mask,
debug_summaries=False):
"""Create tensor for policy gradient loss.
All tensors should have a single batch dimension.
Args:
time_steps: TimeSteps with observations for each timestep.
actions: Tensor of actions for timesteps, aligned on index.
sample_action_log_probs: Tensor of ample probability of each action.
advantages: Tensor of advantage estimate for each timestep, aligned on
index. Works better when advantage estimates are normalized.
current_policy_distribution: The policy distribution, evaluated on all
time_steps.
valid_mask: Mask for invalid timesteps. Float value 1.0 for valid
timesteps and 0.0 for invalid timesteps. (Timesteps which either are
betweeen two episodes, or part of an unfinished episode at the end of
one batch dimension.)
debug_summaries: True if debug summaries should be created.
Returns:
policy_gradient_loss: A tensor that will contain policy gradient loss for
the on-policy experience.
"""
nest.assert_same_structure(time_steps, self.time_step_spec())
action_log_prob = common_utils.log_probability(
current_policy_distribution, actions, self._action_spec)
action_log_prob = tf.to_float(action_log_prob)
if self._log_prob_clipping > 0.0:
action_log_prob = tf.clip_by_value(action_log_prob,
-self._log_prob_clipping,
self._log_prob_clipping)
if self._check_numerics:
action_log_prob = tf.check_numerics(action_log_prob,
'action_log_prob')
# Prepare both clipped and unclipped importance ratios.
importance_ratio = tf.exp(action_log_prob - sample_action_log_probs)
importance_ratio_clipped = tf.clip_by_value(
importance_ratio, 1 - self._importance_ratio_clipping,
1 + self._importance_ratio_clipping)
if self._check_numerics:
importance_ratio = tf.check_numerics(importance_ratio,
'importance_ratio')
if self._importance_ratio_clipping > 0.0:
importance_ratio_clipped = tf.check_numerics(
importance_ratio_clipped, 'importance_ratio_clipped')
# Pessimistically choose the minimum objective value for clipped and
# unclipped importance ratios.
per_timestep_objective = importance_ratio * advantages
per_timestep_objective_clipped = importance_ratio_clipped * advantages
per_timestep_objective_min = tf.minimum(
per_timestep_objective, per_timestep_objective_clipped)
if self._importance_ratio_clipping > 0.0:
policy_gradient_loss = -per_timestep_objective_min
else:
policy_gradient_loss = -per_timestep_objective
policy_gradient_loss = tf.reduce_mean(policy_gradient_loss *
valid_mask)
if debug_summaries:
if self._importance_ratio_clipping > 0.0:
clip_fraction = tf.reduce_mean(
tf.to_float(
tf.greater(tf.abs(importance_ratio - 1.0),
self._importance_ratio_clipping)))
tf.contrib.summary.scalar('clip_fraction', clip_fraction)
tf.contrib.summary.histogram('action_log_prob', action_log_prob)
tf.contrib.summary.histogram('action_log_prob_sample',
sample_action_log_probs)
tf.contrib.summary.histogram('importance_ratio', importance_ratio)
tf.contrib.summary.scalar('importance_ratio_mean',
tf.reduce_mean(importance_ratio))
tf.contrib.summary.histogram('importance_ratio_clipped',
importance_ratio_clipped)
tf.contrib.summary.histogram('per_timestep_objective',
per_timestep_objective)
tf.contrib.summary.histogram('per_timestep_objective_clipped',
per_timestep_objective_clipped)
tf.contrib.summary.histogram('per_timestep_objective_min',
per_timestep_objective_min)
entropy = common_utils.entropy(current_policy_distribution,
self.action_spec())
tf.contrib.summary.histogram('policy_entropy', entropy)
tf.contrib.summary.scalar('policy_entropy_mean',
tf.reduce_mean(entropy))
# Categorical distribution (used for discrete actions)
# doesn't have a mean.
if not self.action_spec().is_discrete():
tf.contrib.summary.histogram(
'actions_distribution_mean',
current_policy_distribution.mean())
tf.contrib.summary.histogram(
'actions_distribution_stddev',
current_policy_distribution.stddev())
tf.contrib.summary.histogram('policy_gradient_loss',
policy_gradient_loss)
if self._check_numerics:
policy_gradient_loss = tf.check_numerics(policy_gradient_loss,
'policy_gradient_loss')
return policy_gradient_loss
def policy_gradient_loss(self,
time_steps,
actions,
sample_action_log_probs,
advantages,
current_policy_distribution,
weights,
debug_summaries=False):
"""Create tensor for policy gradient loss.
All tensors should have a single batch dimension.
Args:
time_steps: TimeSteps with observations for each timestep.
actions: Tensor of actions for timesteps, aligned on index.
sample_action_log_probs: Tensor of sample probability of each action.
advantages: Tensor of advantage estimate for each timestep, aligned on
index. Works better when advantage estimates are normalized.
current_policy_distribution: The policy distribution, evaluated on all
time_steps.
weights: Optional scalar or element-wise (per-batch-entry) importance
weights. Includes a mask for invalid timesteps.
debug_summaries: True if debug summaries should be created.
Returns:
policy_gradient_loss: A tensor that will contain policy gradient loss for
the on-policy experience.
"""
tf.nest.assert_same_structure(time_steps, self.time_step_spec)
action_log_prob = common.log_probability(current_policy_distribution,
actions, self._action_spec)
action_log_prob = tf.cast(action_log_prob, tf.float32)
if self._log_prob_clipping > 0.0:
action_log_prob = tf.clip_by_value(action_log_prob,
-self._log_prob_clipping,
self._log_prob_clipping)
if self._check_numerics:
action_log_prob = tf.debugging.check_numerics(
action_log_prob, 'action_log_prob')
# Prepare both clipped and unclipped importance ratios.
importance_ratio = tf.exp(action_log_prob - sample_action_log_probs)
importance_ratio_clipped = tf.clip_by_value(
importance_ratio, 1 - self._importance_ratio_clipping,
1 + self._importance_ratio_clipping)
if self._check_numerics:
importance_ratio = tf.debugging.check_numerics(
importance_ratio, 'importance_ratio')
if self._importance_ratio_clipping > 0.0:
importance_ratio_clipped = tf.debugging.check_numerics(
importance_ratio_clipped, 'importance_ratio_clipped')
# Pessimistically choose the minimum objective value for clipped and
# unclipped importance ratios.
per_timestep_objective = importance_ratio * advantages
per_timestep_objective_clipped = importance_ratio_clipped * advantages
per_timestep_objective_min = tf.minimum(
per_timestep_objective, per_timestep_objective_clipped)
if self._importance_ratio_clipping > 0.0:
policy_gradient_loss = -per_timestep_objective_min
else:
policy_gradient_loss = -per_timestep_objective
policy_gradient_loss = tf.reduce_mean(
input_tensor=policy_gradient_loss * weights)
if debug_summaries:
if self._importance_ratio_clipping > 0.0:
clip_fraction = tf.reduce_mean(input_tensor=tf.cast(
tf.greater(tf.abs(importance_ratio - 1.0),
self._importance_ratio_clipping), tf.float32))
tf.compat.v2.summary.scalar(name='clip_fraction',
data=clip_fraction,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='action_log_prob',
data=action_log_prob,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='action_log_prob_sample',
data=sample_action_log_probs,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='importance_ratio',
data=importance_ratio,
step=self.train_step_counter)
tf.compat.v2.summary.scalar(
name='importance_ratio_mean',
data=tf.reduce_mean(input_tensor=importance_ratio),
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='importance_ratio_clipped',
data=importance_ratio_clipped,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='per_timestep_objective',
data=per_timestep_objective,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(
name='per_timestep_objective_clipped',
data=per_timestep_objective_clipped,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='per_timestep_objective_min',
data=per_timestep_objective_min,
step=self.train_step_counter)
entropy = common.entropy(current_policy_distribution,
self.action_spec)
tf.compat.v2.summary.histogram(name='policy_entropy',
data=entropy,
step=self.train_step_counter)
tf.compat.v2.summary.scalar(
name='policy_entropy_mean',
data=tf.reduce_mean(input_tensor=entropy),
step=self.train_step_counter)
for i, (single_action, single_distribution) in enumerate(
zip(tf.nest.flatten(self.action_spec),
tf.nest.flatten(current_policy_distribution))):
# Categorical distribution (used for discrete actions) doesn't have a
# mean.
distribution_index = '_{}'.format(i) if i > 0 else ''
if not tensor_spec.is_discrete(single_action):
tf.compat.v2.summary.histogram(
name='actions_distribution_mean' + distribution_index,
data=single_distribution.mean(),
step=self.train_step_counter)
tf.compat.v2.summary.histogram(
name='actions_distribution_stddev' +
distribution_index,
data=single_distribution.stddev(),
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='policy_gradient_loss',
data=policy_gradient_loss,
step=self.train_step_counter)
if self._check_numerics:
policy_gradient_loss = tf.debugging.check_numerics(
policy_gradient_loss, 'policy_gradient_loss')
return policy_gradient_loss
def _train(self, experience, weights):
# Get individual tensors from transitions.
(time_steps, policy_steps_,
next_time_steps) = trajectory.to_transition(experience)
actions = policy_steps_.action
if self._debug_summaries:
actions_list = tf.nest.flatten(actions)
show_action_index = len(actions_list) != 1
for i, single_action in enumerate(actions_list):
action_name = ('actions_{}'.format(i)
if show_action_index else 'actions')
tf.compat.v2.summary.histogram(name=action_name,
data=single_action,
step=self.train_step_counter)
action_distribution_parameters = policy_steps_.info
# Reconstruct per-timestep policy distribution from stored distribution
# parameters.
old_actions_distribution = (
distribution_spec.nested_distributions_from_specs(
self._action_distribution_spec,
action_distribution_parameters))
# Compute log probability of actions taken during data collection, using the
# collect policy distribution.
act_log_probs = common.log_probability(old_actions_distribution,
actions, self._action_spec)
# Compute the value predictions for states using the current value function.
# To be used for return & advantage computation.
batch_size = nest_utils.get_outer_shape(time_steps,
self._time_step_spec)[0]
policy_state = self._collect_policy.get_initial_state(
batch_size=batch_size)
value_preds, unused_policy_state = self._collect_policy.apply_value_network(
experience.observation,
experience.step_type,
policy_state=policy_state)
value_preds = tf.stop_gradient(value_preds)
valid_mask = ppo_utils.make_timestep_mask(next_time_steps)
if weights is None:
weights = valid_mask
else:
weights *= valid_mask
returns, normalized_advantages = self.compute_return_and_advantage(
next_time_steps, value_preds)
# Loss tensors across batches will be aggregated for summaries.
policy_gradient_losses = []
value_estimation_losses = []
l2_regularization_losses = []
entropy_regularization_losses = []
kl_penalty_losses = []
loss_info = None # TODO(b/123627451): Remove.
# For each epoch, create its own train op that depends on the previous one.
for i_epoch in range(self._num_epochs):
with tf.name_scope('epoch_%d' % i_epoch):
# Only save debug summaries for first and last epochs.
debug_summaries = (self._debug_summaries
and (i_epoch == 0
or i_epoch == self._num_epochs - 1))
# Build one epoch train op.
with tf.GradientTape() as tape:
loss_info = self.get_epoch_loss(
time_steps, actions, act_log_probs, returns,
normalized_advantages, action_distribution_parameters,
weights, self.train_step_counter, debug_summaries)
variables_to_train = (self._actor_net.trainable_weights +
self._value_net.trainable_weights)
grads = tape.gradient(loss_info.loss, variables_to_train)
# Tuple is used for py3, where zip is a generator producing values once.
grads_and_vars = tuple(zip(grads, variables_to_train))
if self._gradient_clipping > 0:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
# If summarize_gradients, create functions for summarizing both
# gradients and variables.
if self._summarize_grads_and_vars and debug_summaries:
eager_utils.add_gradients_summaries(
grads_and_vars, self.train_step_counter)
eager_utils.add_variables_summaries(
grads_and_vars, self.train_step_counter)
self._optimizer.apply_gradients(
grads_and_vars, global_step=self.train_step_counter)
policy_gradient_losses.append(
loss_info.extra.policy_gradient_loss)
value_estimation_losses.append(
loss_info.extra.value_estimation_loss)
l2_regularization_losses.append(
loss_info.extra.l2_regularization_loss)
entropy_regularization_losses.append(
loss_info.extra.entropy_regularization_loss)
kl_penalty_losses.append(loss_info.extra.kl_penalty_loss)
# After update epochs, update adaptive kl beta, then update observation
# normalizer and reward normalizer.
batch_size = nest_utils.get_outer_shape(time_steps,
self._time_step_spec)[0]
policy_state = self._collect_policy.get_initial_state(batch_size)
# Compute the mean kl from previous action distribution.
kl_divergence = self._kl_divergence(
time_steps, action_distribution_parameters,
self._collect_policy.distribution(time_steps, policy_state).action)
self.update_adaptive_kl_beta(kl_divergence)
if self._observation_normalizer:
self._observation_normalizer.update(time_steps.observation,
outer_dims=[0, 1])
else:
# TODO(b/127661780): Verify performance of reward_normalizer when obs are
# not normalized
if self._reward_normalizer:
self._reward_normalizer.update(next_time_steps.reward,
outer_dims=[0, 1])
loss_info = tf.nest.map_structure(tf.identity, loss_info)
# Make summaries for total loss across all epochs.
# The *_losses lists will have been populated by
# calls to self.get_epoch_loss.
with tf.name_scope('Losses/'):
total_policy_gradient_loss = tf.add_n(policy_gradient_losses)
total_value_estimation_loss = tf.add_n(value_estimation_losses)
total_l2_regularization_loss = tf.add_n(l2_regularization_losses)
total_entropy_regularization_loss = tf.add_n(
entropy_regularization_losses)
total_kl_penalty_loss = tf.add_n(kl_penalty_losses)
tf.compat.v2.summary.scalar(name='policy_gradient_loss',
data=total_policy_gradient_loss,
step=self.train_step_counter)
tf.compat.v2.summary.scalar(name='value_estimation_loss',
data=total_value_estimation_loss,
step=self.train_step_counter)
tf.compat.v2.summary.scalar(name='l2_regularization_loss',
data=total_l2_regularization_loss,
step=self.train_step_counter)
tf.compat.v2.summary.scalar(name='entropy_regularization_loss',
data=total_entropy_regularization_loss,
step=self.train_step_counter)
tf.compat.v2.summary.scalar(name='kl_penalty_loss',
data=total_kl_penalty_loss,
step=self.train_step_counter)
total_abs_loss = (tf.abs(total_policy_gradient_loss) +
tf.abs(total_value_estimation_loss) +
tf.abs(total_entropy_regularization_loss) +
tf.abs(total_l2_regularization_loss) +
tf.abs(total_kl_penalty_loss))
tf.compat.v2.summary.scalar(name='total_abs_loss',
data=total_abs_loss,
step=self.train_step_counter)
if self._summarize_grads_and_vars:
with tf.name_scope('Variables/'):
all_vars = (self._actor_net.trainable_weights +
self._value_net.trainable_weights)
for var in all_vars:
tf.compat.v2.summary.histogram(
name=var.name.replace(':', '_'),
data=var,
step=self.train_step_counter)
return loss_info
def action_importance_ratio(action_distribution, collect_action_distribution,
action, action_spec, clipping_mode, scope,
importance_ratio_clipping, log_prob_clipping,
check_numerics, debug_summaries):
""" ratio for importance sampling, used in PPO loss and vtrace loss.
Caller has to save tf.name_scope() and pass scope to this function.
Args:
action_distribution (nested tf.distribution): Distribution over
actions under target policy.
collect_action_distribution (nested tf.distribution): distribution
over actions from behavior policy, used to sample actions for
the rollout.
action (nested tf.distribution): possibly batched action tuple
taken during rollout.
action_spec (nested BoundedTensorSpec): representing the actions.
clipping_mode (str): mode for clipping the importance ratio.
'double_sided': clips the range of importance ratio into
[1-importance_ratio_clipping, 1+importance_ratio_clipping],
which is used by PPOLoss.
'capping': clips the range of importance ratio into
min(1+importance_ratio_clipping, importance_ratio),
which is used by VTraceLoss, where c_bar or rho_bar =
1+importance_ratio_clipping.
scope (name scope manager): returned by tf.name_scope(), set
outside.
importance_ratio_clipping (float): Epsilon in clipped, surrogate
PPO objective. See the cited paper for more detail.
log_prob_clipping (float): If >0, clipping log probs to the range
(-log_prob_clipping, log_prob_clipping) to prevent inf / NaN
values.
check_numerics (bool): If true, adds tf.debugging.check_numerics to
help find NaN / Inf values. For debugging only.
debug_summaries (bool): If true, output summary metrics to tf.
Returns:
importance_ratio (Tensor), importance_ratio_clipped (Tensor).
"""
current_policy_distribution = action_distribution
sample_action_log_probs = tfa_common.log_probability(
collect_action_distribution, action, action_spec)
sample_action_log_probs = tf.stop_gradient(sample_action_log_probs)
action_log_prob = tfa_common.log_probability(current_policy_distribution,
action, action_spec)
if log_prob_clipping > 0.0:
action_log_prob = tf.clip_by_value(action_log_prob, -log_prob_clipping,
log_prob_clipping)
if check_numerics:
action_log_prob = tf.debugging.check_numerics(action_log_prob,
'action_log_prob')
# Prepare both clipped and unclipped importance ratios.
importance_ratio = tf.exp(action_log_prob - sample_action_log_probs)
if check_numerics:
importance_ratio = tf.debugging.check_numerics(importance_ratio,
'importance_ratio')
if clipping_mode == 'double_sided':
importance_ratio_clipped = tf.clip_by_value(
importance_ratio, 1 - importance_ratio_clipping,
1 + importance_ratio_clipping)
elif clipping_mode == 'capping':
importance_ratio_clipped = tf.minimum(importance_ratio,
1 + importance_ratio_clipping)
else:
raise Exception('Unsupported clipping mode: ' + clipping_mode)
def _summary():
with scope:
if importance_ratio_clipping > 0.0:
clip_fraction = tf.reduce_mean(input_tensor=tf.cast(
tf.greater(tf.abs(importance_ratio - 1.0),
importance_ratio_clipping), tf.float32))
tf.summary.scalar('clip_fraction', clip_fraction)
tf.summary.histogram('action_log_prob', action_log_prob)
tf.summary.histogram('action_log_prob_sample',
sample_action_log_probs)
tf.summary.histogram('importance_ratio', importance_ratio)
tf.summary.scalar('importance_ratio_mean',
tf.reduce_mean(input_tensor=importance_ratio))
tf.summary.histogram('importance_ratio_clipped',
importance_ratio_clipped)
if debug_summaries:
common.run_if(common.should_record_summaries(), _summary)
return importance_ratio, importance_ratio_clipped
def _train(self, experience, weights, train_step_counter):
# Change trajectory to transitions.
trajectory0 = nest.map_structure(lambda t: t[:, :-1], experience)
trajectory1 = nest.map_structure(lambda t: t[:, 1:], experience)
# Get individual tensors from transitions.
(time_steps, policy_steps_,
next_time_steps) = trajectory.to_transition(trajectory0, trajectory1)
actions = policy_steps_.action
if self._debug_summaries:
tf.contrib.summary.histogram('actions', actions)
action_distribution_parameters = policy_steps_.info
# Reconstruct per-timestep policy distribution from stored distribution
# parameters.
old_actions_distribution = (
distribution_spec.nested_distributions_from_specs(
self._action_distribution_spec,
action_distribution_parameters))
# Compute log probability of actions taken during data collection, using the
# collect policy distribution.
act_log_probs = common_utils.log_probability(old_actions_distribution,
actions,
self._action_spec)
# Compute the value predictions for states using the current value function.
# To be used for return & advantage computation.
batch_size = nest_utils.get_outer_shape(time_steps,
self._time_step_spec)[0]
policy_state = self._collect_policy.get_initial_state(
batch_size=batch_size)
value_preds, unused_policy_state = self._collect_policy.apply_value_network(
experience.observation,
experience.step_type,
policy_state=policy_state)
value_preds = tf.stop_gradient(value_preds)
valid_mask = ppo_utils.make_timestep_mask(next_time_steps)
if weights is None:
weights = valid_mask
else:
weights *= valid_mask
returns, normalized_advantages = self.compute_return_and_advantage(
next_time_steps, value_preds)
# Loss tensors across batches will be aggregated for summaries.
policy_gradient_losses = []
value_estimation_losses = []
l2_regularization_losses = []
entropy_regularization_losses = []
kl_penalty_losses = []
# For each epoch, create its own train op that depends on the previous one.
loss_info = tf.no_op()
for i_epoch in range(self._num_epochs):
with tf.name_scope('epoch_%d' % i_epoch):
with tf.control_dependencies(nest.flatten(loss_info)):
# Only save debug summaries for first and last epochs.
debug_summaries = (self._debug_summaries and
(i_epoch == 0
or i_epoch == self._num_epochs - 1))
# Build one epoch train op.
loss_info = self.build_train_op(
time_steps, actions, act_log_probs, returns,
normalized_advantages, action_distribution_parameters,
weights, train_step_counter,
self._summarize_grads_and_vars,
self._gradient_clipping, debug_summaries)
policy_gradient_losses.append(
loss_info.extra.policy_gradient_loss)
value_estimation_losses.append(
loss_info.extra.value_estimation_loss)
l2_regularization_losses.append(
loss_info.extra.l2_regularization_loss)
entropy_regularization_losses.append(
loss_info.extra.entropy_regularization_loss)
kl_penalty_losses.append(loss_info.extra.kl_penalty_loss)
# After update epochs, update adaptive kl beta, then update observation
# normalizer and reward normalizer.
with tf.control_dependencies(nest.flatten(loss_info)):
# Compute the mean kl from old.
batch_size = nest_utils.get_outer_shape(time_steps,
self._time_step_spec)[0]
policy_state = self._collect_policy.get_initial_state(batch_size)
kl_divergence = self._kl_divergence(
time_steps, action_distribution_parameters,
self._collect_policy.distribution(time_steps,
policy_state).action)
update_adaptive_kl_beta_op = self.update_adaptive_kl_beta(
kl_divergence)
with tf.control_dependencies([update_adaptive_kl_beta_op]):
if self._observation_normalizer:
update_obs_norm = (self._observation_normalizer.update(
time_steps.observation, outer_dims=[0, 1]))
else:
update_obs_norm = tf.no_op()
if self._reward_normalizer:
update_reward_norm = self._reward_normalizer.update(
next_time_steps.reward, outer_dims=[0, 1])
else:
update_reward_norm = tf.no_op()
with tf.control_dependencies([update_obs_norm, update_reward_norm]):
loss_info = nest.map_structure(tf.identity, loss_info)
# Make summaries for total loss across all epochs.
# The *_losses lists will have been populated by
# calls to self.build_train_op.
with tf.name_scope('Losses/'):
total_policy_gradient_loss = tf.add_n(policy_gradient_losses)
total_value_estimation_loss = tf.add_n(value_estimation_losses)
total_l2_regularization_loss = tf.add_n(l2_regularization_losses)
total_entropy_regularization_loss = tf.add_n(
entropy_regularization_losses)
total_kl_penalty_loss = tf.add_n(kl_penalty_losses)
tf.contrib.summary.scalar('policy_gradient_loss',
total_policy_gradient_loss)
tf.contrib.summary.scalar('value_estimation_loss',
total_value_estimation_loss)
tf.contrib.summary.scalar('l2_regularization_loss',
total_l2_regularization_loss)
if self._entropy_regularization:
tf.contrib.summary.scalar('entropy_regularization_loss',
total_entropy_regularization_loss)
tf.contrib.summary.scalar('kl_penalty_loss', total_kl_penalty_loss)
total_abs_loss = (tf.abs(total_policy_gradient_loss) +
tf.abs(total_value_estimation_loss) +
tf.abs(total_entropy_regularization_loss) +
tf.abs(total_l2_regularization_loss) +
tf.abs(total_kl_penalty_loss))
tf.contrib.summary.scalar('total_abs_loss', total_abs_loss)
if self._summarize_grads_and_vars:
with tf.name_scope('Variables/'):
all_vars = (self._actor_net.trainable_weights +
self._value_net.trainable_weights)
for var in all_vars:
tf.contrib.summary.histogram(var.name.replace(':', '_'),
var)
return loss_info
def _pg_loss(self, training_info, advantages):
action_log_prob = tfa_common.log_probability(
training_info.action_distribution, training_info.action,
self._action_spec)
return -advantages * action_log_prob