def _train(self, experience, weights=None):
# TODO(b/126593927): Support batch dimensions >1.
if experience.step_type.shape[0] != 1:
raise NotImplementedError(
'ReinforceAgent does not yet support batch '
'dimensions greater than 1.')
experience = tf.nest.map_structure(lambda t: tf.squeeze(t, 0),
experience)
returns = common.compute_returns(experience.reward,
experience.discount)
if self._debug_summaries:
tf.compat.v2.summary.histogram(name='rewards',
data=experience.reward,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='discounts',
data=experience.discount,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='returns',
data=returns,
step=self.train_step_counter)
# TODO(b/126592060): replace with tensor normalizer.
if self._normalize_returns:
ret_mean, ret_var = tf.nn.moments(x=returns, axes=[0])
returns = (returns - ret_mean) / (tf.sqrt(ret_var) + 1e-6)
if self._debug_summaries:
tf.compat.v2.summary.histogram(name='normalized_returns',
data=returns,
step=self.train_step_counter)
time_step = ts.TimeStep(experience.step_type,
tf.zeros_like(experience.reward),
tf.zeros_like(experience.discount),
experience.observation)
variables_to_train = self._actor_network.variables
with tf.GradientTape() as tape:
loss_info = self._loss(time_step,
experience.action,
tf.stop_gradient(returns),
weights=weights)
tf.debugging.check_numerics(loss_info.loss, 'Loss is inf or nan')
grads = tape.gradient(loss_info.loss, variables_to_train)
grads_and_vars = zip(grads, variables_to_train)
if self._gradient_clipping:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars,
global_step=self.train_step_counter)
return tf.nest.map_structure(tf.identity, loss_info)
def _apply_loss(self, aggregated_losses, variables_to_train, tape,
optimizer):
total_loss = aggregated_losses.total_loss
tf.debugging.check_numerics(total_loss, "Loss is inf or nan")
assert list(variables_to_train), "No variables in the agent's network."
grads = tape.gradient(total_loss, variables_to_train)
grads_and_vars = list(zip(grads, variables_to_train))
if self._gradient_clipping is not None:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self.summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
optimizer.apply_gradients(grads_and_vars)
if self.summaries_enabled:
dict_losses = {
"loss": aggregated_losses.weighted,
"reg_loss": aggregated_losses.regularization,
"total_loss": total_loss
}
common.summarize_scalar_dict(dict_losses,
step=self.train_step_counter,
name_scope="Losses/")
def _train(self, experience: types.NestedTensor,
weights: types.Tensor) -> tf_agent.LossInfo:
experience = self._as_trajectory(experience)
with tf.GradientTape() as tape:
loss_info = self._loss(experience, weights=weights, training=True)
variables_to_train = self._variables_to_train()
if not variables_to_train:
logging.info('No variable to train in the agent.')
return loss_info
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 is not None:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars)
self.train_step_counter.assign_add(1)
return loss_info
def _train(self, experience, weights):
with tf.GradientTape() as tape:
loss_info = self._loss(
experience,
td_errors_loss_fn=self._td_errors_loss_fn,
gamma=self._gamma,
reward_scale_factor=self._reward_scale_factor,
weights=weights,
training=True)
tf.debugging.check_numerics(loss_info.loss, 'Loss is inf or nan')
variables_to_train = self._q_network.trainable_weights
non_trainable_weights = self._q_network.non_trainable_weights
assert list(variables_to_train), "No variables in the agent's q_network."
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 = list(zip(grads, variables_to_train))
if self._gradient_clipping is not None:
grads_and_vars = eager_utils.clip_gradient_norms(grads_and_vars,
self._gradient_clipping)
if self._summarize_grads_and_vars:
grads_and_vars_with_non_trainable = (
grads_and_vars + [(None, v) for v in non_trainable_weights])
eager_utils.add_variables_summaries(grads_and_vars_with_non_trainable,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars)
self.train_step_counter.assign_add(1)
self._update_target()
return loss_info
def _train_v1(self, experience, weights):
with tf.GradientTape() as tape:
loss_info = self._loss(
experience,
td_errors_loss_fn=self._td_errors_loss_fn,
gamma=self._gamma,
reward_scale_factor=self._reward_scale_factor,
weights=weights)
tf.debugging.check_numerics(loss_info[0], 'Loss is inf or nan')
variables_to_train = self._q_network.trainable_weights
assert list(variables_to_train), "No variables in the agent's q_network."
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 is not None:
grads_and_vars = eager_utils.clip_gradient_norms(grads_and_vars,
self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
train_op = self._optimizer.apply_gradients(
grads_and_vars, global_step=self.train_step_counter)
update_op = self._update_target()
train_op = tf.group(train_op, update_op)
return train_op, loss_info
def _train(self, experience, weights):
(observations, actions,
rewards) = bandit_utils.process_experience_for_neural_agents(
experience, self._observation_and_action_constraint_splitter,
self._accepts_per_arm_features, self.training_data_spec)
with tf.GradientTape() as tape:
loss_info = self.loss(observations,
actions,
rewards,
weights=weights,
training=True)
variables_to_train = self._variables_to_train()
if not variables_to_train:
logging.info('No variable to train in the agent.')
return loss_info
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 is not None:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars)
self.train_step_counter.assign_add(1)
return loss_info
def _create_summaries(grads_and_vars):
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
grads_and_vars = clip_gradients(grads_and_vars)
return grads_and_vars
def _train(self, experience, weights):
time_steps, actions, next_time_steps = self._experience_to_transitions(
experience)
with tf.GradientTape() as tape:
loss_info = self.loss(time_steps,
actions,
next_time_steps,
td_errors_loss_fn=self._td_errors_loss_fn,
gamma=self._gamma,
reward_scale_factor=self._reward_scale_factor,
weights=weights)
tf.debugging.check_numerics(loss_info[0], 'Loss is inf or nan')
variables_to_train = self._q_network.trainable_weights
assert list(variables_to_train), "No variables in the agent's q_network."
grads = tape.gradient(loss_info.loss, variables_to_train)
grads_and_vars = zip(grads, variables_to_train)
if self._gradient_clipping is not None:
grads_and_vars = eager_utils.clip_gradient_norms(grads_and_vars,
self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars,
global_step=self.train_step_counter)
self._update_target()
return loss_info
def compute_loss_using_reward_layer(
self,
observation: types.NestedTensor,
action: types.Tensor,
reward: types.Tensor,
weights: Optional[types.Float] = None,
training: bool = False) -> tf_agent.LossInfo:
"""Computes loss using the reward layer.
Args:
observation: A batch of observations.
action: A batch of actions.
reward: A batch of rewards.
weights: Optional scalar or elementwise (per-batch-entry) importance
weights. The output batch loss will be scaled by these weights, and
the final scalar loss is the mean of these values.
training: Whether the loss is being used for training.
Returns:
loss: A `LossInfo` containing the loss for the training step.
"""
# Update the neural network params.
with tf.GradientTape() as tape:
loss_info = self._loss_using_reward_layer(observation,
action,
reward,
weights,
training=training)
tf.debugging.check_numerics(loss_info[0], 'Loss is inf or nan')
tf.compat.v2.summary.scalar(name='using_reward_layer',
data=1,
step=self.train_step_counter)
if self._summarize_grads_and_vars:
self.compute_summaries(loss_info.loss)
variables_to_train = (self._encoding_network.trainable_weights +
self._reward_layer.trainable_weights)
if not variables_to_train:
raise ValueError('No variable to train in the agent.')
grads = tape.gradient(loss_info.loss, variables_to_train)
grads_and_vars = tuple(zip(grads, variables_to_train))
if self._gradient_clipping is not None:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
with tf.name_scope('Reward_network/'):
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars)
self.train_step_counter.assign_add(1)
return loss_info
def _train(self, experience, weights=None):
# Add a mask to ensure we reset the return calculation at episode
# boundaries. This is needed in cases where episodes are truncated before
# reaching a terminal state.
non_last_mask = tf.cast(
tf.math.not_equal(experience.next_step_type, ts.StepType.LAST),
tf.float32)
discounts = non_last_mask * experience.discount * self._gamma
returns = value_ops.discounted_return(experience.reward,
discounts,
time_major=False)
if self._debug_summaries:
tf.compat.v2.summary.histogram(name='rewards',
data=experience.reward,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='discounts',
data=experience.discount,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='returns',
data=returns,
step=self.train_step_counter)
time_step = ts.TimeStep(experience.step_type,
tf.zeros_like(experience.reward),
tf.zeros_like(experience.discount),
experience.observation)
with tf.GradientTape() as tape:
loss_info = self.total_loss(time_step,
experience.action,
tf.stop_gradient(returns),
weights=weights)
tf.debugging.check_numerics(loss_info.loss, 'Loss is inf or nan')
variables_to_train = self._actor_network.trainable_weights
if self._baseline:
variables_to_train += self._value_network.trainable_weights
grads = tape.gradient(loss_info.loss, variables_to_train)
grads_and_vars = list(zip(grads, variables_to_train))
if self._gradient_clipping:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars,
global_step=self.train_step_counter)
return tf.nest.map_structure(tf.identity, loss_info)
def _train(self, experience, weights):
rewards, _ = nest_utils.flatten_multi_batched_nested_tensors(
experience.reward, self._time_step_spec.reward)
actions, _ = nest_utils.flatten_multi_batched_nested_tensors(
experience.action, self._action_spec)
observations, _ = nest_utils.flatten_multi_batched_nested_tensors(
experience.observation, self.training_data_spec.observation)
if self._observation_and_action_constraint_splitter is not None:
observations, _ = self._observation_and_action_constraint_splitter(
observations)
if self._accepts_per_arm_features:
# The arm observation we train on needs to be copied from the respective
# policy info field to the per arm observation field. Pretending there was
# only one action, we fill the action field with zeros.
chosen_action, _ = nest_utils.flatten_multi_batched_nested_tensors(
experience.policy_info.chosen_arm_features,
self.policy.info_spec.chosen_arm_features)
observations[
bandit_spec_utils.PER_ARM_FEATURE_KEY] = tf.expand_dims(
chosen_action, axis=1)
actions = tf.zeros_like(actions)
with tf.GradientTape() as tape:
loss_info = self.loss(observations,
actions,
rewards,
weights=weights,
training=True)
self.compute_summaries(loss_info.loss)
variables_to_train = self._reward_network.trainable_weights
if not variables_to_train:
logging.info('No variable to train in the agent.')
return loss_info
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 is not None:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
training_lib.apply_gradients(self._optimizer,
grads_and_vars,
global_step=self.train_step_counter)
return loss_info
def _train(self, experience, weights=None):
returns = value_ops.discounted_return(experience.reward,
experience.discount,
time_major=False)
if self._debug_summaries:
tf.compat.v2.summary.histogram(name='rewards',
data=experience.reward,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='discounts',
data=experience.discount,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='returns',
data=returns,
step=self.train_step_counter)
# TODO(b/126592060): replace with tensor normalizer.
if self._normalize_returns:
returns = _standard_normalize(returns, axes=(0, 1))
if self._debug_summaries:
tf.compat.v2.summary.histogram(name='normalized_returns',
data=returns,
step=self.train_step_counter)
time_step = ts.TimeStep(experience.step_type,
tf.zeros_like(experience.reward),
tf.zeros_like(experience.discount),
experience.observation)
variables_to_train = self._actor_network.variables
with tf.GradientTape() as tape:
loss_info = self._loss(time_step,
experience.action,
tf.stop_gradient(returns),
weights=weights)
tf.debugging.check_numerics(loss_info.loss, 'Loss is inf or nan')
grads = tape.gradient(loss_info.loss, variables_to_train)
grads_and_vars = zip(grads, variables_to_train)
if self._gradient_clipping:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars,
global_step=self.train_step_counter)
return tf.nest.map_structure(tf.identity, loss_info)
def _apply_gradients(self, gradients, variables, optimizer):
grads_and_vars = list(zip(gradients, variables))
if self._gradient_clipping is not None:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
optimizer.apply_gradients(grads_and_vars)
def _apply_gradients(self, gradients, variables, optimizer):
# Tuple is used for py3, where zip is a generator producing values once.
grads_and_vars = tuple(zip(gradients, variables))
if self._gradient_clipping is not None:
grads_and_vars = eager_utils.clip_gradient_norms(grads_and_vars,
self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
return optimizer.apply_gradients(grads_and_vars)
def _train(self, experience, weights=None):
# TODO(b/132914246): Use .is_last() to mask the end of each episode.
returns = value_ops.discounted_return(experience.reward,
experience.discount *
self._gamma,
time_major=False)
if self._debug_summaries:
tf.compat.v2.summary.histogram(name='rewards',
data=experience.reward,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='discounts',
data=experience.discount,
step=self.train_step_counter)
tf.compat.v2.summary.histogram(name='returns',
data=returns,
step=self.train_step_counter)
time_step = ts.TimeStep(experience.step_type,
tf.zeros_like(experience.reward),
tf.zeros_like(experience.discount),
experience.observation)
with tf.GradientTape() as tape:
loss_info = self.total_loss(time_step,
experience.action,
tf.stop_gradient(returns),
weights=weights)
tf.debugging.check_numerics(loss_info.loss, 'Loss is inf or nan')
variables_to_train = self._actor_network.trainable_weights
if self._baseline:
variables_to_train += self._value_network.trainable_weights
grads = tape.gradient(loss_info.loss, variables_to_train)
grads_and_vars = zip(grads, variables_to_train)
if self._gradient_clipping:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars,
global_step=self.train_step_counter)
return tf.nest.map_structure(tf.identity, loss_info)
def _train(self, experience: types.NestedTensor,
weights: types.Tensor) -> tf_agent.LossInfo:
(observations, actions,
objective_values) = bandit_utils.process_experience_for_neural_agents(
experience, self._accepts_per_arm_features,
self.training_data_spec)
if self._observation_and_action_constraint_splitter is not None:
observations, _ = self._observation_and_action_constraint_splitter(
observations)
if objective_values.shape.rank != 2:
raise ValueError(
'The objectives tensor should be rank-2 [batch_size, num_objectives],'
' but found to be rank-{}'.format(objective_values.shape.rank))
if objective_values.shape[1] != self._num_objectives:
raise ValueError(
'The number of objectives in the objective_values tensor: {} '
'is different from the number of objective networks: {}.'.
format(objective_values.shape[1], self._num_objectives))
with tf.GradientTape() as tape:
loss_info = self.loss(observations,
actions,
objective_values,
weights=weights,
training=True)
variables_to_train = self._variables_to_train()
if not variables_to_train:
logging.info('No variable to train in the agent.')
return loss_info
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 is not None:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars)
self.train_step_counter.assign_add(1)
return loss_info
def _create_summaries(grads_and_vars):
grads_and_vars = eager_utils.add_gradients_summaries(
grads_and_vars)
grads_and_vars = eager_utils.add_variables_summaries(
grads_and_vars)
grads_and_vars = clip_gradients(grads_and_vars)
return grads_and_vars
def _train(self, experience, weights):
rewards, _ = nest_utils.flatten_multi_batched_nested_tensors(
experience.reward, self._time_step_spec.reward)
actions, _ = nest_utils.flatten_multi_batched_nested_tensors(
experience.action, self._action_spec)
observations, _ = nest_utils.flatten_multi_batched_nested_tensors(
experience.observation, self._time_step_spec.observation)
if self._observation_and_action_constraint_splitter is not None:
observations, _ = self._observation_and_action_constraint_splitter(
observations)
with tf.GradientTape() as tape:
loss_info = self.loss(observations,
actions,
rewards,
weights=weights,
training=True)
tf.debugging.check_numerics(loss_info[0], 'Loss is inf or nan')
self.compute_summaries(loss_info.loss)
variables_to_train = self._reward_network.trainable_weights
if not variables_to_train:
logging.info('No variable to train in the agent.')
return loss_info
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 is not None:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
training_lib.apply_gradients(self._optimizer,
grads_and_vars,
global_step=self.train_step_counter)
return loss_info
def _train(self, experience, weights):
experience = self._as_trajectory(experience)
with tf.GradientTape() as tape:
loss_info = self._loss(experience, weights=weights, training=True)
variables_to_train = self._variables_to_train()
if not variables_to_train:
logging.info('No variable to train in the agent.')
return loss_info
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 is not None:
grads_and_vars = eager_utils.clip_gradient_norms(
grads_and_vars, self._gradient_clipping)
if self._summarize_grads_and_vars:
eager_utils.add_variables_summaries(grads_and_vars,
self.train_step_counter)
eager_utils.add_gradients_summaries(grads_and_vars,
self.train_step_counter)
self._optimizer.apply_gradients(grads_and_vars)
self.train_step_counter.assign_add(1)
if not self._accepts_per_arm_features and self._num_samples_list:
# Compute the number of samples for each action in the current batch.
actions_flattened = tf.reshape(experience.action, [-1])
num_samples_per_action_current = [
tf.reduce_sum(tf.cast(tf.equal(actions_flattened, k),
tf.int64))
for k in range(self._num_actions)
]
# Update the number of samples for each action.
for a, b in zip(self._num_samples_list,
num_samples_per_action_current):
tf.compat.v1.assign_add(a, b)
return loss_info
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
