def _define_experience(self, agent_indices, observ, action, reward):
"""Implement the branch of experience() entered during training."""
update_filters = tf.summary.merge([
self._observ_filter.update(observ),
self._reward_filter.update(reward)
])
with tf.control_dependencies([update_filters]):
if self._config.train_on_agent_action:
# NOTE: Doesn't seem to change much.
action = self._last_action
batch = (observ, action, tf.gather(self._last_mean, agent_indices),
tf.gather(self._last_logstd, agent_indices), reward)
append = self._episodes.append(batch, agent_indices)
with tf.control_dependencies([append]):
norm_observ = self._observ_filter.transform(observ)
norm_reward = tf.reduce_mean(self._reward_filter.transform(reward))
# pylint: disable=g-long-lambda
summary = tf.cond(
self._should_log, lambda: tf.summary.merge([
update_filters,
self._observ_filter.summary(),
self._reward_filter.summary(),
tf.summary.scalar('memory_size', self._memory_index),
tf.summary.histogram('normalized_observ', norm_observ),
tf.summary.histogram('action', self._last_action),
tf.summary.scalar('normalized_reward', norm_reward)
]), str)
return summary
def append(self, transitions, rows=None):
"""Append a batch of transitions to rows of the memory.
Args:
transitions: Tuple of transition quantities with batch dimension.
rows: Episodes to append to, defaults to all.
Returns:
Operation.
"""
rows = tf.range(self._capacity) if rows is None else rows
assert rows.shape.ndims == 1
assert_capacity = tf.assert_less(rows,
self._capacity,
message='capacity exceeded')
with tf.control_dependencies([assert_capacity]):
assert_max_length = tf.assert_less(tf.gather(self._length, rows),
self._max_length,
message='max length exceeded')
append_ops = []
with tf.control_dependencies([assert_max_length]):
for buffer_, elements in zip(self._buffers, transitions):
timestep = tf.gather(self._length, rows)
indices = tf.stack([rows, timestep], 1)
append_ops.append(
tf.scatter_nd_update(buffer_, indices, elements))
with tf.control_dependencies(append_ops):
episode_mask = tf.reduce_sum(
tf.one_hot(rows, self._capacity, dtype=tf.int32), 0)
return self._length.assign_add(episode_mask)
def _define_end_episode(self, agent_indices):
"""Implement the branch of end_episode() entered during training."""
episodes, length = self._episodes.data(agent_indices)
space_left = self._config.update_every - self._memory_index
use_episodes = tf.range(
tf.minimum(tf.shape(agent_indices)[0], space_left))
episodes = [tf.gather(elem, use_episodes) for elem in episodes]
append = self._memory.replace(episodes,
tf.gather(length, use_episodes),
use_episodes + self._memory_index)
with tf.control_dependencies([append]):
inc_index = self._memory_index.assign_add(
tf.shape(use_episodes)[0])
with tf.control_dependencies([inc_index]):
memory_full = self._memory_index >= self._config.update_every
return tf.cond(memory_full, self._training, str)
def _define_step(self, done, score, summary):
"""Combine operations of a phase.
Keeps track of the mean score and when to report it.
Args:
done: Tensor indicating whether current score can be used.
score: Tensor holding the current, possibly intermediate, score.
summary: Tensor holding summary string to write if not an empty string.
Returns:
Tuple of summary tensor, mean score, and new global step. The mean score
is zero for non reporting steps.
"""
if done.shape.ndims == 0:
done = done[None]
if score.shape.ndims == 0:
score = score[None]
score_mean = streaming_mean.StreamingMean((), tf.float32)
with tf.control_dependencies([done, score, summary]):
done_score = tf.gather(score, tf.where(done)[:, 0])
submit_score = tf.cond(tf.reduce_any(done), lambda: score_mean.submit(done_score), tf.no_op)
with tf.control_dependencies([submit_score]):
mean_score = tf.cond(self._report, score_mean.clear, float)
steps_made = tf.shape(score)[0]
next_step = self._step.assign_add(steps_made)
with tf.control_dependencies([mean_score, next_step]):
return tf.identity(summary), mean_score, next_step, steps_made
def _define_end_episode(agent_indices):
"""Notify the algorithm of ending episodes.
Also updates the mean score and length counters used for summaries.
Args:
agent_indices: Tensor holding batch indices that end their episodes.
Returns:
Summary tensor.
"""
assert agent_indices.shape.ndims == 1
submit_score = mean_score.submit(tf.gather(score, agent_indices))
submit_length = mean_length.submit(
tf.cast(tf.gather(length, agent_indices), tf.float32))
with tf.control_dependencies([submit_score, submit_length]):
return algo.end_episode(agent_indices)
def data(self, rows=None):
"""Access a batch of episodes from the memory.
Padding elements after the length of each episode are unspecified and might
contain old data.
Args:
rows: Episodes to select, defaults to all.
Returns:
Tuple containing a tuple of transition quantiries with batch and time
dimensions, and a batch of sequence lengths.
"""
rows = tf.range(self._capacity) if rows is None else rows
assert rows.shape.ndims == 1
episode = [tf.gather(buffer_, rows) for buffer_ in self._buffers]
length = tf.gather(self._length, rows)
return episode, length
def length(self, rows=None):
"""Tensor holding the current length of episodes.
Args:
rows: Episodes to select length from, defaults to all.
Returns:
Batch tensor of sequence lengths.
"""
rows = tf.range(self._capacity) if rows is None else rows
return tf.gather(self._length, rows)
def perform(self, agent_indices, observ):
"""Compute batch of actions and a summary for a batch of observation.
Args:
agent_indices: Tensor containing current batch indices.
observ: Tensor of a batch of observations for all agents.
Returns:
Tuple of action batch tensor and summary tensor.
"""
with tf.name_scope('perform/'):
observ = self._observ_filter.transform(observ)
if self._last_state is None:
state = None
else:
state = tf.contrib.framework.nest.map_structure(
lambda x: tf.gather(x, agent_indices), self._last_state)
output = self._network(observ[:, None], tf.ones(observ.shape[0]),
state)
action = tf.cond(self._is_training, output.policy.sample,
lambda: output.mean)
logprob = output.policy.log_prob(action)[:, 0]
# pylint: disable=g-long-lambda
summary = tf.cond(
self._should_log, lambda: tf.summary.merge([
tf.summary.histogram('mean', output.mean[:, 0]),
tf.summary.histogram('std', tf.exp(output.logstd[:, 0])),
tf.summary.histogram('action', action[:, 0]),
tf.summary.histogram('logprob', logprob)
]), str)
# Remember current policy to append to memory in the experience callback.
if self._last_state is None:
assign_state = tf.no_op()
else:
assign_state = utility.assign_nested_vars(
self._last_state, output.state, agent_indices)
with tf.control_dependencies([
assign_state,
tf.scatter_update(self._last_action, agent_indices,
action[:, 0]),
tf.scatter_update(self._last_mean, agent_indices,
output.mean[:, 0]),
tf.scatter_update(self._last_logstd, agent_indices,
output.logstd[:, 0])
]):
return tf.check_numerics(action[:, 0],
'action'), tf.identity(summary)