models/research/brain_coder/single_task/pg_agent.py

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

"""Language model agent.

Agent outputs code in a sequence just like a language model. Can be trained
as a language model or using RL, or a combination of the two.
"""

from collections import namedtuple
from math import exp
from math import log
import time

from absl import logging
import numpy as np
from six.moves import xrange
import tensorflow as tf

from common import rollout as rollout_lib  # brain coder
from common import utils  # brain coder
from single_task import misc  # brain coder


# Experiments in the ICLR 2018 paper used reduce_sum instead of reduce_mean for
# some losses. We make all loses be batch_size independent, and multiply the
# changed losses by 64, which was the fixed batch_size when the experiments
# where run. The loss hyperparameters still match what is reported in the paper.
MAGIC_LOSS_MULTIPLIER = 64


def rshift_time(tensor_2d, fill=misc.BF_EOS_INT):
  """Right shifts a 2D tensor along the time dimension (axis-1)."""
  dim_0 = tf.shape(tensor_2d)[0]
  fill_tensor = tf.fill([dim_0, 1], fill)
  return tf.concat([fill_tensor, tensor_2d[:, :-1]], axis=1)


def join(a, b):
  # Concat a and b along 0-th dim.
  if a is None or len(a) == 0:  # pylint: disable=g-explicit-length-test
    return b
  if b is None or len(b) == 0:  # pylint: disable=g-explicit-length-test
    return a
  return np.concatenate((a, b))


def make_optimizer(kind, lr):
  if kind == 'sgd':
    return tf.train.GradientDescentOptimizer(lr)
  elif kind == 'adam':
    return tf.train.AdamOptimizer(lr)
  elif kind == 'rmsprop':
    return tf.train.RMSPropOptimizer(learning_rate=lr, decay=0.99)
  else:
    raise ValueError('Optimizer type "%s" not recognized.' % kind)


class LinearWrapper(tf.contrib.rnn.RNNCell):
  """RNNCell wrapper that adds a linear layer to the output."""

  def __init__(self, cell, output_size, dtype=tf.float32, suppress_index=None):
    self.cell = cell
    self._output_size = output_size
    self._dtype = dtype
    self._suppress_index = suppress_index
    self.smallest_float = -2.4e38

  def __call__(self, inputs, state, scope=None):
    with tf.variable_scope(type(self).__name__):
      outputs, state = self.cell(inputs, state, scope=scope)
      logits = tf.matmul(
          outputs,
          tf.get_variable('w_output',
                          [self.cell.output_size, self.output_size],
                          dtype=self._dtype))
      if self._suppress_index is not None:
        # Replace the target index with -inf, so that it never gets selected.
        batch_size = tf.shape(logits)[0]
        logits = tf.concat(
            [logits[:, :self._suppress_index],
             tf.fill([batch_size, 1], self.smallest_float),
             logits[:, self._suppress_index + 1:]],
            axis=1)

    return logits, state

  @property
  def output_size(self):
    return self._output_size

  @property
  def state_size(self):
    return self.cell.state_size

  def zero_state(self, batch_size, dtype):
    return self.cell.zero_state(batch_size, dtype)


UpdateStepResult = namedtuple(
    'UpdateStepResult',
    ['global_step', 'global_npe', 'summaries_list', 'gradients_dict'])


class AttrDict(dict):
  """Dict with attributes as keys.

  https://stackoverflow.com/a/14620633
  """

  def __init__(self, *args, **kwargs):
    super(AttrDict, self).__init__(*args, **kwargs)
    self.__dict__ = self


class LMAgent(object):
  """Language model agent."""
  action_space = misc.bf_num_tokens()
  observation_space = misc.bf_num_tokens()

  def __init__(self, global_config, task_id=0,
               logging_file=None,
               experience_replay_file=None,
               global_best_reward_fn=None,
               found_solution_op=None,
               assign_code_solution_fn=None,
               program_count=None,
               do_iw_summaries=False,
               stop_on_success=True,
               dtype=tf.float32,
               verbose_level=0,
               is_local=True):
    self.config = config = global_config.agent
    self.logging_file = logging_file
    self.experience_replay_file = experience_replay_file
    self.task_id = task_id
    self.verbose_level = verbose_level
    self.global_best_reward_fn = global_best_reward_fn
    self.found_solution_op = found_solution_op
    self.assign_code_solution_fn = assign_code_solution_fn
    self.parent_scope_name = tf.get_variable_scope().name
    self.dtype = dtype
    self.allow_eos_token = config.eos_token
    self.stop_on_success = stop_on_success
    self.pi_loss_hparam = config.pi_loss_hparam
    self.vf_loss_hparam = config.vf_loss_hparam
    self.is_local = is_local

    self.top_reward = 0.0
    self.embeddings_trainable = True

    self.no_op = tf.no_op()

    self.learning_rate = tf.constant(
        config.lr, dtype=dtype, name='learning_rate')
    self.initializer = tf.contrib.layers.variance_scaling_initializer(
        factor=config.param_init_factor,
        mode='FAN_AVG',
        uniform=True,
        dtype=dtype)  # TF's default initializer.
    tf.get_variable_scope().set_initializer(self.initializer)

    self.a2c = config.ema_baseline_decay == 0
    if not self.a2c:
      logging.info('Using exponential moving average REINFORCE baselines.')
      self.ema_baseline_decay = config.ema_baseline_decay
      self.ema_by_len = [0.0] * global_config.timestep_limit
    else:
      logging.info('Using advantage (a2c) with learned value function.')
      self.ema_baseline_decay = 0.0
      self.ema_by_len = None

    # Top-k
    if config.topk and config.topk_loss_hparam:
      self.topk_loss_hparam = config.topk_loss_hparam
      self.topk_batch_size = config.topk_batch_size
      if self.topk_batch_size <= 0:
        raise ValueError('topk_batch_size must be a positive integer. Got %s',
                         self.topk_batch_size)
      self.top_episodes = utils.MaxUniquePriorityQueue(config.topk)
      logging.info('Made max-priorty-queue with capacity %d',
                   self.top_episodes.capacity)
    else:
      self.top_episodes = None
      self.topk_loss_hparam = 0.0
      logging.info('No max-priorty-queue')

    # Experience replay.
    self.replay_temperature = config.replay_temperature
    self.num_replay_per_batch = int(global_config.batch_size * config.alpha)
    self.num_on_policy_per_batch = (
        global_config.batch_size - self.num_replay_per_batch)
    self.replay_alpha = (
        self.num_replay_per_batch / float(global_config.batch_size))
    logging.info('num_replay_per_batch: %d', self.num_replay_per_batch)
    logging.info('num_on_policy_per_batch: %d', self.num_on_policy_per_batch)
    logging.info('replay_alpha: %s', self.replay_alpha)
    if self.num_replay_per_batch > 0:
      # Train with off-policy episodes from replay buffer.
      start_time = time.time()
      self.experience_replay = utils.RouletteWheel(
          unique_mode=True, save_file=experience_replay_file)
      logging.info('Took %s sec to load replay buffer from disk.',
                   int(time.time() - start_time))
      logging.info('Replay buffer file location: "%s"',
                   self.experience_replay.save_file)
    else:
      # Only train on-policy.
      self.experience_replay = None

    if program_count is not None:
      self.program_count = program_count
      self.program_count_add_ph = tf.placeholder(
          tf.int64, [], 'program_count_add_ph')
      self.program_count_add_op = self.program_count.assign_add(
          self.program_count_add_ph)

    ################################
    # RL policy and value networks #
    ################################
    batch_size = global_config.batch_size
    logging.info('batch_size: %d', batch_size)

    self.policy_cell = LinearWrapper(
        tf.contrib.rnn.MultiRNNCell(
            [tf.contrib.rnn.BasicLSTMCell(cell_size)
             for cell_size in config.policy_lstm_sizes]),
        self.action_space,
        dtype=dtype,
        suppress_index=None if self.allow_eos_token else misc.BF_EOS_INT)
    self.value_cell = LinearWrapper(
        tf.contrib.rnn.MultiRNNCell(
            [tf.contrib.rnn.BasicLSTMCell(cell_size)
             for cell_size in config.value_lstm_sizes]),
        1,
        dtype=dtype)

    obs_embedding_scope = 'obs_embed'
    with tf.variable_scope(
        obs_embedding_scope,
        initializer=tf.random_uniform_initializer(minval=-1.0, maxval=1.0)):
      obs_embeddings = tf.get_variable(
          'embeddings',
          [self.observation_space, config.obs_embedding_size],
          dtype=dtype, trainable=self.embeddings_trainable)
      self.obs_embeddings = obs_embeddings

    ################################
    # RL policy and value networks #
    ################################

    initial_state = tf.fill([batch_size], misc.BF_EOS_INT)
    def loop_fn(loop_time, cell_output, cell_state, loop_state):
      """Function called by tf.nn.raw_rnn to instantiate body of the while_loop.

      See https://www.tensorflow.org/api_docs/python/tf/nn/raw_rnn for more
      information.

      When time is 0, and cell_output, cell_state, loop_state are all None,
      `loop_fn` will create the initial input, internal cell state, and loop
      state. When time > 0, `loop_fn` will operate on previous cell output,
      state, and loop state.

      Args:
        loop_time: A scalar tensor holding the current timestep (zero based
            counting).
        cell_output: Output of the raw_rnn cell at the current timestep.
        cell_state: Cell internal state at the current timestep.
        loop_state: Additional loop state. These tensors were returned by the
            previous call to `loop_fn`.

      Returns:
        elements_finished: Bool tensor of shape [batch_size] which marks each
            sequence in the batch as being finished or not finished.
        next_input: A tensor containing input to be fed into the cell at the
            next timestep.
        next_cell_state: Cell internal state to be fed into the cell at the
            next timestep.
        emit_output: Tensor to be added to the TensorArray returned by raw_rnn
            as output from the while_loop.
        next_loop_state: Additional loop state. These tensors will be fed back
            into the next call to `loop_fn` as `loop_state`.
      """
      if cell_output is None:  # 0th time step.
        next_cell_state = self.policy_cell.zero_state(batch_size, dtype)
        elements_finished = tf.zeros([batch_size], tf.bool)
        output_lengths = tf.ones([batch_size], dtype=tf.int32)
        next_input = tf.gather(obs_embeddings, initial_state)
        emit_output = None
        next_loop_state = (
            tf.TensorArray(dtype=tf.int32, size=0, dynamic_size=True),
            output_lengths,
            elements_finished
        )
      else:
        scaled_logits = cell_output * config.softmax_tr  # Scale temperature.
        prev_chosen, prev_output_lengths, prev_elements_finished = loop_state
        next_cell_state = cell_state
        chosen_outputs = tf.to_int32(tf.where(
            tf.logical_not(prev_elements_finished),
            tf.multinomial(logits=scaled_logits, num_samples=1)[:, 0],
            tf.zeros([batch_size], dtype=tf.int64)))
        elements_finished = tf.logical_or(
            tf.equal(chosen_outputs, misc.BF_EOS_INT),
            loop_time >= global_config.timestep_limit)
        output_lengths = tf.where(
            elements_finished,
            prev_output_lengths,
            # length includes EOS token. empty seq has len 1.
            tf.tile(tf.expand_dims(loop_time + 1, 0), [batch_size])
        )
        next_input = tf.gather(obs_embeddings, chosen_outputs)
        emit_output = scaled_logits
        next_loop_state = (prev_chosen.write(loop_time - 1, chosen_outputs),
                           output_lengths,
                           tf.logical_or(prev_elements_finished,
                                         elements_finished))
      return (elements_finished, next_input, next_cell_state, emit_output,
              next_loop_state)

    with tf.variable_scope('policy'):
      (decoder_outputs_ta,
       _,  # decoder_state
       (sampled_output_ta, output_lengths, _)) = tf.nn.raw_rnn(
           cell=self.policy_cell,
           loop_fn=loop_fn)
    policy_logits = tf.transpose(decoder_outputs_ta.stack(), (1, 0, 2),
                                 name='policy_logits')
    sampled_tokens = tf.transpose(sampled_output_ta.stack(), (1, 0),
                                  name='sampled_tokens')
    # Add SOS to beginning of the sequence.
    rshift_sampled_tokens = rshift_time(sampled_tokens, fill=misc.BF_EOS_INT)

    # Initial state is 0, 2nd state is first token.
    # Note: If value of last state is computed, this will be used as bootstrap.
    if self.a2c:
      with tf.variable_scope('value'):
        value_output, _ = tf.nn.dynamic_rnn(
            self.value_cell,
            tf.gather(obs_embeddings, rshift_sampled_tokens),
            sequence_length=output_lengths,
            dtype=dtype)
      value = tf.squeeze(value_output, axis=[2])
    else:
      value = tf.zeros([], dtype=dtype)

    # for sampling actions from the agent, and which told tensors for doing
    # gradient updates on the agent.
    self.sampled_batch = AttrDict(
        logits=policy_logits,
        value=value,
        tokens=sampled_tokens,
        episode_lengths=output_lengths,
        probs=tf.nn.softmax(policy_logits),
        log_probs=tf.nn.log_softmax(policy_logits))

    # adjusted_lengths can be less than the full length of each episode.
    # Use this to train on only part of an episode (starting from t=0).
    self.adjusted_lengths = tf.placeholder(
        tf.int32, [None], name='adjusted_lengths')
    self.policy_multipliers = tf.placeholder(
        dtype,
        [None, None],
        name='policy_multipliers')
    # Empirical value, i.e. discounted sum of observed future rewards from each
    # time step in the episode.
    self.empirical_values = tf.placeholder(
        dtype,
        [None, None],
        name='empirical_values')

    # Off-policy training. Just add supervised loss to the RL loss.
    self.off_policy_targets = tf.placeholder(
        tf.int32,
        [None, None],
        name='off_policy_targets')
    self.off_policy_target_lengths = tf.placeholder(
        tf.int32, [None], name='off_policy_target_lengths')

    self.actions = tf.placeholder(tf.int32, [None, None], name='actions')
    # Add SOS to beginning of the sequence.
    inputs = rshift_time(self.actions, fill=misc.BF_EOS_INT)
    with tf.variable_scope('policy', reuse=True):
      logits, _ = tf.nn.dynamic_rnn(
          self.policy_cell, tf.gather(obs_embeddings, inputs),
          sequence_length=self.adjusted_lengths,
          dtype=dtype)

    if self.a2c:
      with tf.variable_scope('value', reuse=True):
        value_output, _ = tf.nn.dynamic_rnn(
            self.value_cell,
            tf.gather(obs_embeddings, inputs),
            sequence_length=self.adjusted_lengths,
            dtype=dtype)
      value2 = tf.squeeze(value_output, axis=[2])
    else:
      value2 = tf.zeros([], dtype=dtype)

    self.given_batch = AttrDict(
        logits=logits,
        value=value2,
        tokens=sampled_tokens,
        episode_lengths=self.adjusted_lengths,
        probs=tf.nn.softmax(logits),
        log_probs=tf.nn.log_softmax(logits))

    # Episode masks.
    max_episode_length = tf.shape(self.actions)[1]
    # range_row shape: [1, max_episode_length]
    range_row = tf.expand_dims(tf.range(max_episode_length), 0)
    episode_masks = tf.cast(
        tf.less(range_row, tf.expand_dims(self.given_batch.episode_lengths, 1)),
        dtype=dtype)
    episode_masks_3d = tf.expand_dims(episode_masks, 2)

    # Length adjusted episodes.
    self.a_probs = a_probs = self.given_batch.probs * episode_masks_3d
    self.a_log_probs = a_log_probs = (
        self.given_batch.log_probs * episode_masks_3d)
    self.a_value = a_value = self.given_batch.value * episode_masks
    self.a_policy_multipliers = a_policy_multipliers = (
        self.policy_multipliers * episode_masks)
    if self.a2c:
      self.a_empirical_values = a_empirical_values = (
          self.empirical_values * episode_masks)

    # pi_loss is scalar
    acs_onehot = tf.one_hot(self.actions, self.action_space, dtype=dtype)
    self.acs_onehot = acs_onehot
    chosen_masked_log_probs = acs_onehot * a_log_probs
    pi_target = tf.expand_dims(a_policy_multipliers, -1)
    pi_loss_per_step = chosen_masked_log_probs * pi_target  # Maximize.
    self.pi_loss = pi_loss = (
        -tf.reduce_mean(tf.reduce_sum(pi_loss_per_step, axis=[1, 2]), axis=0)
        * MAGIC_LOSS_MULTIPLIER)  # Minimize.
    assert len(self.pi_loss.shape) == 0  # pylint: disable=g-explicit-length-test

    # shape: [batch_size, time]
    self.chosen_log_probs = tf.reduce_sum(chosen_masked_log_probs, axis=2)
    self.chosen_probs = tf.reduce_sum(acs_onehot * a_probs, axis=2)

    # loss of value function
    if self.a2c:
      vf_loss_per_step = tf.square(a_value - a_empirical_values)
      self.vf_loss = vf_loss = (
          tf.reduce_mean(tf.reduce_sum(vf_loss_per_step, axis=1), axis=0)
          * MAGIC_LOSS_MULTIPLIER)  # Minimize.
      assert len(self.vf_loss.shape) == 0  # pylint: disable=g-explicit-length-test
    else:
      self.vf_loss = vf_loss = 0.0

    # Maximize entropy regularizer
    self.entropy = entropy = (
        -tf.reduce_mean(
            tf.reduce_sum(a_probs * a_log_probs, axis=[1, 2]), axis=0)
        * MAGIC_LOSS_MULTIPLIER)  # Maximize
    self.negentropy = -entropy  # Minimize negentropy.
    assert len(self.negentropy.shape) == 0  # pylint: disable=g-explicit-length-test

    # off-policy loss
    self.offp_switch = tf.placeholder(dtype, [], name='offp_switch')
    if self.top_episodes is not None:
      # Add SOS to beginning of the sequence.
      offp_inputs = tf.gather(obs_embeddings,
                              rshift_time(self.off_policy_targets,
                                          fill=misc.BF_EOS_INT))
      with tf.variable_scope('policy', reuse=True):
        offp_logits, _ = tf.nn.dynamic_rnn(
            self.policy_cell, offp_inputs, self.off_policy_target_lengths,
            dtype=dtype)  # shape: [batch_size, time, action_space]
      topk_loss_per_step = tf.nn.sparse_softmax_cross_entropy_with_logits(
          labels=self.off_policy_targets,
          logits=offp_logits,
          name='topk_loss_per_logit')
      # Take mean over batch dimension so that the loss multiplier strength is
      # independent of batch size. Sum over time dimension.
      topk_loss = tf.reduce_mean(
          tf.reduce_sum(topk_loss_per_step, axis=1), axis=0)
      assert len(topk_loss.shape) == 0  # pylint: disable=g-explicit-length-test
      self.topk_loss = topk_loss * self.offp_switch
      logging.info('Including off policy loss.')
    else:
      self.topk_loss = topk_loss = 0.0

    self.entropy_hparam = tf.constant(
        config.entropy_beta, dtype=dtype, name='entropy_beta')

    self.pi_loss_term = pi_loss * self.pi_loss_hparam
    self.vf_loss_term = vf_loss * self.vf_loss_hparam
    self.entropy_loss_term = self.negentropy * self.entropy_hparam
    self.topk_loss_term = self.topk_loss_hparam * topk_loss
    self.loss = (
        self.pi_loss_term
        + self.vf_loss_term
        + self.entropy_loss_term
        + self.topk_loss_term)

    params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
                               tf.get_variable_scope().name)
    self.trainable_variables = params
    self.sync_variables = self.trainable_variables
    non_embedding_params = [p for p in params
                            if obs_embedding_scope not in p.name]
    self.non_embedding_params = non_embedding_params
    self.params = params

    if config.regularizer:
      logging.info('Adding L2 regularizer with scale %.2f.',
                   config.regularizer)
      self.regularizer = config.regularizer * sum(
          tf.nn.l2_loss(w) for w in non_embedding_params)
      self.loss += self.regularizer
    else:
      logging.info('Skipping regularizer.')
      self.regularizer = 0.0

    # Only build gradients graph for local model.
    if self.is_local:
      unclipped_grads = tf.gradients(self.loss, params)
      self.dense_unclipped_grads = [
          tf.convert_to_tensor(g) for g in unclipped_grads]
      self.grads, self.global_grad_norm = tf.clip_by_global_norm(
          unclipped_grads, config.grad_clip_threshold)
      self.gradients_dict = dict(zip(params, self.grads))
      self.optimizer = make_optimizer(config.optimizer, self.learning_rate)
      self.all_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
                                             tf.get_variable_scope().name)

    self.do_iw_summaries = do_iw_summaries
    if self.do_iw_summaries:
      b = None
      self.log_iw_replay_ph = tf.placeholder(tf.float32, [b],
                                             'log_iw_replay_ph')
      self.log_iw_policy_ph = tf.placeholder(tf.float32, [b],
                                             'log_iw_policy_ph')
      self.log_prob_replay_ph = tf.placeholder(tf.float32, [b],
                                               'log_prob_replay_ph')
      self.log_prob_policy_ph = tf.placeholder(tf.float32, [b],
                                               'log_prob_policy_ph')
      self.log_norm_replay_weights_ph = tf.placeholder(
          tf.float32, [b], 'log_norm_replay_weights_ph')
      self.iw_summary_op = tf.summary.merge([
          tf.summary.histogram('is/log_iw_replay', self.log_iw_replay_ph),
          tf.summary.histogram('is/log_iw_policy', self.log_iw_policy_ph),
          tf.summary.histogram('is/log_prob_replay', self.log_prob_replay_ph),
          tf.summary.histogram('is/log_prob_policy', self.log_prob_policy_ph),
          tf.summary.histogram(
              'is/log_norm_replay_weights', self.log_norm_replay_weights_ph),
      ])

  def make_summary_ops(self):
    """Construct summary ops for the model."""
    # size = number of timesteps across entire batch. Number normalized by size
    # will not be affected by the amount of padding at the ends of sequences
    # in the batch.
    size = tf.cast(
        tf.reduce_sum(self.given_batch.episode_lengths), dtype=self.dtype)
    offp_size = tf.cast(tf.reduce_sum(self.off_policy_target_lengths),
                        dtype=self.dtype)
    scope_prefix = self.parent_scope_name

    def _remove_prefix(prefix, name):
      assert name.startswith(prefix)
      return name[len(prefix):]

    # RL summaries.
    self.rl_summary_op = tf.summary.merge(
        [tf.summary.scalar('model/policy_loss', self.pi_loss / size),
         tf.summary.scalar('model/value_loss', self.vf_loss / size),
         tf.summary.scalar('model/topk_loss', self.topk_loss / offp_size),
         tf.summary.scalar('model/entropy', self.entropy / size),
         tf.summary.scalar('model/loss', self.loss / size),
         tf.summary.scalar('model/grad_norm',
                           tf.global_norm(self.grads)),
         tf.summary.scalar('model/unclipped_grad_norm', self.global_grad_norm),
         tf.summary.scalar('model/non_embedding_var_norm',
                           tf.global_norm(self.non_embedding_params)),
         tf.summary.scalar('hparams/entropy_beta', self.entropy_hparam),
         tf.summary.scalar('hparams/topk_loss_hparam', self.topk_loss_hparam),
         tf.summary.scalar('hparams/learning_rate', self.learning_rate),
         tf.summary.scalar('model/trainable_var_norm',
                           tf.global_norm(self.trainable_variables)),
         tf.summary.scalar('loss/loss', self.loss),
         tf.summary.scalar('loss/entropy', self.entropy_loss_term),
         tf.summary.scalar('loss/vf', self.vf_loss_term),
         tf.summary.scalar('loss/policy', self.pi_loss_term),
         tf.summary.scalar('loss/offp', self.topk_loss_term)] +
        [tf.summary.scalar(
            'param_norms/' + _remove_prefix(scope_prefix + '/', p.name),
            tf.norm(p))
         for p in self.params] +
        [tf.summary.scalar(
            'grad_norms/' + _remove_prefix(scope_prefix + '/', p.name),
            tf.norm(g))
         for p, g in zip(self.params, self.grads)] +
        [tf.summary.scalar(
            'unclipped_grad_norms/' + _remove_prefix(scope_prefix + '/',
                                                     p.name),
            tf.norm(g))
         for p, g in zip(self.params, self.dense_unclipped_grads)])

    self.text_summary_placeholder = tf.placeholder(tf.string, shape=[])
    self.rl_text_summary_op = tf.summary.text('rl',
                                              self.text_summary_placeholder)

  def _rl_text_summary(self, session, step, npe, tot_r, num_steps,
                       input_case, code_output, code, reason):
    """Logs summary about a single episode and creates a text_summary for TB.

    Args:
      session: tf.Session instance.
      step: Global training step.
      npe: Number of programs executed so far.
      tot_r: Total reward.
      num_steps: Number of timesteps in the episode (i.e. code length).
      input_case: Inputs for test cases.
      code_output: Outputs produced by running the code on the inputs.
      code: String representation of the code.
      reason: Reason for the reward assigned by the task.

    Returns:
      Serialized text summary data for tensorboard.
    """
    if not input_case:
      input_case = ' '
    if not code_output:
      code_output = ' '
    if not code:
      code = ' '
    text = (
        'Tot R: **%.2f**;  Len: **%d**;  Reason: **%s**\n\n'
        'Input: **`%s`**; Output: **`%s`**\n\nCode: **`%s`**'
        % (tot_r, num_steps, reason, input_case, code_output, code))
    text_summary = session.run(self.rl_text_summary_op,
                               {self.text_summary_placeholder: text})
    logging.info(
        'Step %d.\t NPE: %d\t Reason: %s.\t Tot R: %.2f.\t Length: %d. '
        '\tInput: %s \tOutput: %s \tProgram: %s',
        step, npe, reason, tot_r, num_steps, input_case,
        code_output, code)
    return text_summary

  def _rl_reward_summary(self, total_rewards):
    """Create summary ops that report on episode rewards.

    Creates summaries for average, median, max, and min rewards in the batch.

    Args:
      total_rewards: Tensor of shape [batch_size] containing the total reward
          from each episode in the batch.

    Returns:
      tf.Summary op.
    """
    tr = np.asarray(total_rewards)
    reward_summary = tf.Summary(value=[
        tf.Summary.Value(
            tag='reward/avg',
            simple_value=np.mean(tr)),
        tf.Summary.Value(
            tag='reward/med',
            simple_value=np.median(tr)),
        tf.Summary.Value(
            tag='reward/max',
            simple_value=np.max(tr)),
        tf.Summary.Value(
            tag='reward/min',
            simple_value=np.min(tr))])
    return reward_summary

  def _iw_summary(self, session, replay_iw, replay_log_probs,
                  norm_replay_weights, on_policy_iw,
                  on_policy_log_probs):
    """Compute summaries for importance weights at a given batch.

    Args:
      session: tf.Session instance.
      replay_iw: Importance weights for episodes from replay buffer.
      replay_log_probs: Total log probabilities of the replay episodes under the
          current policy.
      norm_replay_weights: Normalized replay weights, i.e. values in `replay_iw`
          divided by the total weight in the entire replay buffer. Note, this is
          also the probability of selecting each episode from the replay buffer
          (in a roulette wheel replay buffer).
      on_policy_iw: Importance weights for episodes sampled from the current
          policy.
      on_policy_log_probs: Total log probabilities of the on-policy episodes
          under the current policy.

    Returns:
      Serialized TF summaries. Use a summary writer to write these summaries to
      disk.
    """
    return session.run(
        self.iw_summary_op,
        {self.log_iw_replay_ph: np.log(replay_iw),
         self.log_iw_policy_ph: np.log(on_policy_iw),
         self.log_norm_replay_weights_ph: np.log(norm_replay_weights),
         self.log_prob_replay_ph: replay_log_probs,
         self.log_prob_policy_ph: on_policy_log_probs})

  def _compute_iw(self, policy_log_probs, replay_weights):
    """Compute importance weights for a batch of episodes.

    Arguments are iterables of length batch_size.

    Args:
      policy_log_probs: Log probability of each episode under the current
          policy.
      replay_weights: Weight of each episode in the replay buffer. 0 for
          episodes not sampled from the replay buffer (i.e. sampled from the
          policy).

    Returns:
      Numpy array of shape [batch_size] containing the importance weight for
      each episode in the batch.
    """
    log_total_replay_weight = log(self.experience_replay.total_weight)

    # importance weight
    # = 1 / [(1 - a) + a * exp(log(replay_weight / total_weight / p))]
    # = 1 / ((1-a) + a*q/p)
    a = float(self.replay_alpha)
    a_com = 1.0 - a  # compliment of a
    importance_weights = np.asarray(
        [1.0 / (a_com
                + a * exp((log(replay_weight) - log_total_replay_weight)
                          - log_p))
         if replay_weight > 0 else 1.0 / a_com
         for log_p, replay_weight
         in zip(policy_log_probs, replay_weights)])
    return importance_weights

  def update_step(self, session, rl_batch, train_op, global_step_op,
                  return_gradients=False):
    """Perform gradient update on the model.

    Args:
      session: tf.Session instance.
      rl_batch: RLBatch instance from data.py. Use DataManager to create a
          RLBatch for each call to update_step. RLBatch contains a batch of
          tasks.
      train_op: A TF op which will perform the gradient update. LMAgent does not
          own its training op, so that trainers can do distributed training
          and construct a specialized training op.
      global_step_op: A TF op which will return the current global step when
          run (should not increment it).
      return_gradients: If True, the gradients will be saved and returned from
          this method call. This is useful for testing.

    Returns:
      Results from the update step in a UpdateStepResult namedtuple, including
      global step, global NPE, serialized summaries, and optionally gradients.
    """
    assert self.is_local

    # Do update for REINFORCE or REINFORCE + replay buffer.
    if self.experience_replay is None:
      # Train with on-policy REINFORCE.

      # Sample new programs from the policy.
      num_programs_from_policy = rl_batch.batch_size
      (batch_actions,
       batch_values,
       episode_lengths) = session.run(
           [self.sampled_batch.tokens, self.sampled_batch.value,
            self.sampled_batch.episode_lengths])
      if episode_lengths.size == 0:
        # This should not happen.
        logging.warn(
            'Shapes:\n'
            'batch_actions.shape: %s\n'
            'batch_values.shape: %s\n'
            'episode_lengths.shape: %s\n',
            batch_actions.shape, batch_values.shape, episode_lengths.shape)

      # Compute rewards.
      code_scores = compute_rewards(
          rl_batch, batch_actions, episode_lengths)
      code_strings = code_scores.code_strings
      batch_tot_r = code_scores.total_rewards
      test_cases = code_scores.test_cases
      code_outputs = code_scores.code_outputs
      reasons = code_scores.reasons

      # Process on-policy samples.
      batch_targets, batch_returns = process_episodes(
          code_scores.batch_rewards, episode_lengths, a2c=self.a2c,
          baselines=self.ema_by_len,
          batch_values=batch_values)
      batch_policy_multipliers = batch_targets
      batch_emp_values = batch_returns if self.a2c else [[]]
      adjusted_lengths = episode_lengths

      if self.top_episodes:
        assert len(self.top_episodes) > 0  # pylint: disable=g-explicit-length-test
        off_policy_targets = [
            item for item, _
            in self.top_episodes.random_sample(self.topk_batch_size)]
        off_policy_target_lengths = [len(t) for t in off_policy_targets]
        off_policy_targets = utils.stack_pad(off_policy_targets, pad_axes=0,
                                             dtype=np.int32)
        offp_switch = 1
      else:
        off_policy_targets = [[0]]
        off_policy_target_lengths = [1]
        offp_switch = 0

      fetches = {
          'global_step': global_step_op,
          'program_count': self.program_count,
          'summaries': self.rl_summary_op,
          'train_op': train_op,
          'gradients': self.gradients_dict if return_gradients else self.no_op}
      fetched = session.run(
          fetches,
          {self.actions: batch_actions,
           self.empirical_values: batch_emp_values,
           self.policy_multipliers: batch_policy_multipliers,
           self.adjusted_lengths: adjusted_lengths,
           self.off_policy_targets: off_policy_targets,
           self.off_policy_target_lengths: off_policy_target_lengths,
           self.offp_switch: offp_switch})

      combined_adjusted_lengths = adjusted_lengths
      combined_returns = batch_returns
    else:
      # Train with REINFORCE + off-policy replay buffer by using importance
      # sampling.

      # Sample new programs from the policy.
      # Note: batch size is constant. A full batch will be sampled, but not all
      # programs will be executed and added to the replay buffer. Those which
      # are not executed will be discarded and not counted.
      batch_actions, batch_values, episode_lengths, log_probs = session.run(
          [self.sampled_batch.tokens, self.sampled_batch.value,
           self.sampled_batch.episode_lengths, self.sampled_batch.log_probs])
      if episode_lengths.size == 0:
        # This should not happen.
        logging.warn(
            'Shapes:\n'
            'batch_actions.shape: %s\n'
            'batch_values.shape: %s\n'
            'episode_lengths.shape: %s\n',
            batch_actions.shape, batch_values.shape, episode_lengths.shape)

      # Sample from experince replay buffer
      empty_replay_buffer = (
          self.experience_replay.is_empty()
          if self.experience_replay is not None else True)
      num_programs_from_replay_buff = (
          self.num_replay_per_batch if not empty_replay_buffer else 0)
      num_programs_from_policy = (
          rl_batch.batch_size - num_programs_from_replay_buff)
      if (not empty_replay_buffer) and num_programs_from_replay_buff:
        result = self.experience_replay.sample_many(
            num_programs_from_replay_buff)
        experience_samples, replay_weights = zip(*result)
        (replay_actions,
         replay_rewards,
         _,  # log probs
         replay_adjusted_lengths) = zip(*experience_samples)

        replay_batch_actions = utils.stack_pad(replay_actions, pad_axes=0,
                                               dtype=np.int32)

        # compute log probs for replay samples under current policy
        all_replay_log_probs, = session.run(
            [self.given_batch.log_probs],
            {self.actions: replay_batch_actions,
             self.adjusted_lengths: replay_adjusted_lengths})
        replay_log_probs = [
            np.choose(replay_actions[i], all_replay_log_probs[i, :l].T).sum()
            for i, l in enumerate(replay_adjusted_lengths)]
      else:
        # Replay buffer is empty. Do not sample from it.
        replay_actions = None
        replay_policy_multipliers = None
        replay_adjusted_lengths = None
        replay_log_probs = None
        replay_weights = None
        replay_returns = None
        on_policy_weights = [0] * num_programs_from_replay_buff

      assert not self.a2c  # TODO(danabo): Support A2C with importance sampling.

      # Compute rewards.
      code_scores = compute_rewards(
          rl_batch, batch_actions, episode_lengths,
          batch_size=num_programs_from_policy)
      code_strings = code_scores.code_strings
      batch_tot_r = code_scores.total_rewards
      test_cases = code_scores.test_cases
      code_outputs = code_scores.code_outputs
      reasons = code_scores.reasons

      # Process on-policy samples.
      p = num_programs_from_policy
      batch_targets, batch_returns = process_episodes(
          code_scores.batch_rewards, episode_lengths[:p], a2c=False,
          baselines=self.ema_by_len)
      batch_policy_multipliers = batch_targets
      batch_emp_values = [[]]
      on_policy_returns = batch_returns

      # Process off-policy samples.
      if (not empty_replay_buffer) and num_programs_from_replay_buff:
        offp_batch_rewards = [
            [0.0] * (l - 1) + [r]
            for l, r in zip(replay_adjusted_lengths, replay_rewards)]
        assert len(offp_batch_rewards) == num_programs_from_replay_buff
        assert len(replay_adjusted_lengths) == num_programs_from_replay_buff
        replay_batch_targets, replay_returns = process_episodes(
            offp_batch_rewards, replay_adjusted_lengths, a2c=False,
            baselines=self.ema_by_len)
        # Convert 2D array back into ragged 2D list.
        replay_policy_multipliers = [
            replay_batch_targets[i, :l]
            for i, l
            in enumerate(
                replay_adjusted_lengths[:num_programs_from_replay_buff])]

      adjusted_lengths = episode_lengths[:num_programs_from_policy]

      if self.top_episodes:
        assert len(self.top_episodes) > 0  # pylint: disable=g-explicit-length-test
        off_policy_targets = [
            item for item, _
            in self.top_episodes.random_sample(self.topk_batch_size)]
        off_policy_target_lengths = [len(t) for t in off_policy_targets]
        off_policy_targets = utils.stack_pad(off_policy_targets, pad_axes=0,
                                             dtype=np.int32)
        offp_switch = 1
      else:
        off_policy_targets = [[0]]
        off_policy_target_lengths = [1]
        offp_switch = 0

      # On-policy episodes.
      if num_programs_from_policy:
        separate_actions = [
            batch_actions[i, :l]
            for i, l in enumerate(adjusted_lengths)]
        chosen_log_probs = [
            np.choose(separate_actions[i], log_probs[i, :l].T)
            for i, l in enumerate(adjusted_lengths)]
        new_experiences = [
            (separate_actions[i],
             batch_tot_r[i],
             chosen_log_probs[i].sum(), l)
            for i, l in enumerate(adjusted_lengths)]
        on_policy_policy_multipliers = [
            batch_policy_multipliers[i, :l]
            for i, l in enumerate(adjusted_lengths)]
        (on_policy_actions,
         _,  # rewards
         on_policy_log_probs,
         on_policy_adjusted_lengths) = zip(*new_experiences)
      else:
        new_experiences = []
        on_policy_policy_multipliers = []
        on_policy_actions = []
        on_policy_log_probs = []
        on_policy_adjusted_lengths = []

      if (not empty_replay_buffer) and num_programs_from_replay_buff:
        # Look for new experiences in replay buffer. Assign weight if an episode
        # is in the buffer.
        on_policy_weights = [0] * num_programs_from_policy
        for i, cs in enumerate(code_strings):
          if self.experience_replay.has_key(cs):
            on_policy_weights[i] = self.experience_replay.get_weight(cs)

      # Randomly select on-policy or off policy episodes to train on.
      combined_actions = join(replay_actions, on_policy_actions)
      combined_policy_multipliers = join(
          replay_policy_multipliers, on_policy_policy_multipliers)
      combined_adjusted_lengths = join(
          replay_adjusted_lengths, on_policy_adjusted_lengths)
      combined_returns = join(replay_returns, on_policy_returns)
      combined_actions = utils.stack_pad(combined_actions, pad_axes=0)
      combined_policy_multipliers = utils.stack_pad(combined_policy_multipliers,
                                                    pad_axes=0)
      # P
      combined_on_policy_log_probs = join(replay_log_probs, on_policy_log_probs)
      # Q
      # Assume weight is zero for all sequences sampled from the policy.
      combined_q_weights = join(replay_weights, on_policy_weights)

      # Importance adjustment. Naive formulation:
      # E_{x~p}[f(x)] ~= 1/N sum_{x~p}(f(x)) ~= 1/N sum_{x~q}(f(x) * p(x)/q(x)).
      # p(x) is the policy, and q(x) is the off-policy distribution, i.e. replay
      # buffer distribution. Importance weight w(x) = p(x) / q(x).

      # Instead of sampling from the replay buffer only, we sample from a
      # mixture distribution of the policy and replay buffer.
      # We are sampling from the mixture a*q(x) + (1-a)*p(x), where 0 <= a <= 1.
      # Thus the importance weight w(x) = p(x) / (a*q(x) + (1-a)*p(x))
      # = 1 / ((1-a) + a*q(x)/p(x)) where q(x) is 0 for x sampled from the
      #                             policy.
      # Note: a = self.replay_alpha
      if empty_replay_buffer:
        # The replay buffer is empty.
        # Do no gradient update this step. The replay buffer will have stuff in
        # it next time.
        combined_policy_multipliers *= 0
      elif not num_programs_from_replay_buff:
        combined_policy_multipliers = np.ones([len(combined_actions), 1],
                                              dtype=np.float32)
      else:
        # If a < 1 compute importance weights
        # importance weight
        # = 1 / [(1 - a) + a * exp(log(replay_weight / total_weight / p))]
        # = 1 / ((1-a) + a*q/p)
        importance_weights = self._compute_iw(combined_on_policy_log_probs,
                                              combined_q_weights)
        if self.config.iw_normalize:
          importance_weights *= (
              float(rl_batch.batch_size) / importance_weights.sum())
        combined_policy_multipliers *= importance_weights.reshape(-1, 1)

      # Train on replay batch, top-k MLE.
      assert self.program_count is not None
      fetches = {
          'global_step': global_step_op,
          'program_count': self.program_count,
          'summaries': self.rl_summary_op,
          'train_op': train_op,
          'gradients': self.gradients_dict if return_gradients else self.no_op}
      fetched = session.run(
          fetches,
          {self.actions: combined_actions,
           self.empirical_values: [[]],  # replay_emp_values,
           self.policy_multipliers: combined_policy_multipliers,
           self.adjusted_lengths: combined_adjusted_lengths,
           self.off_policy_targets: off_policy_targets,
           self.off_policy_target_lengths: off_policy_target_lengths,
           self.offp_switch: offp_switch})

      # Add to experience replay buffer.
      self.experience_replay.add_many(
          objs=new_experiences,
          weights=[exp(r / self.replay_temperature) for r in batch_tot_r],
          keys=code_strings)

    # Update program count.
    session.run(
        [self.program_count_add_op],
        {self.program_count_add_ph: num_programs_from_policy})

    # Update EMA baselines on the mini-batch which we just did traning on.
    if not self.a2c:
      for i in xrange(rl_batch.batch_size):
        episode_length = combined_adjusted_lengths[i]
        empirical_returns = combined_returns[i, :episode_length]
        for j in xrange(episode_length):
          # Update ema_baselines in place.
          self.ema_by_len[j] = (
              self.ema_baseline_decay * self.ema_by_len[j]
              + (1 - self.ema_baseline_decay) * empirical_returns[j])

    global_step = fetched['global_step']
    global_npe = fetched['program_count']
    core_summaries = fetched['summaries']
    summaries_list = [core_summaries]

    if num_programs_from_policy:
      s_i = 0
      text_summary = self._rl_text_summary(
          session,
          global_step,
          global_npe,
          batch_tot_r[s_i],
          episode_lengths[s_i], test_cases[s_i],
          code_outputs[s_i], code_strings[s_i], reasons[s_i])
      reward_summary = self._rl_reward_summary(batch_tot_r)

      is_best = False
      if self.global_best_reward_fn:
        # Save best reward.
        best_reward = np.max(batch_tot_r)
        is_best = self.global_best_reward_fn(session, best_reward)

      if self.found_solution_op is not None and 'correct' in reasons:
        session.run(self.found_solution_op)

        # Save program to disk for record keeping.
        if self.stop_on_success:
          solutions = [
              {'code': code_strings[i], 'reward': batch_tot_r[i],
               'npe': global_npe}
              for i in xrange(len(reasons)) if reasons[i] == 'correct']
        elif is_best:
          solutions = [
              {'code': code_strings[np.argmax(batch_tot_r)],
               'reward': np.max(batch_tot_r),
               'npe': global_npe}]
        else:
          solutions = []
        if solutions:
          if self.assign_code_solution_fn:
            self.assign_code_solution_fn(session, solutions[0]['code'])
          with tf.gfile.FastGFile(self.logging_file, 'a') as writer:
            for solution_dict in solutions:
              writer.write(str(solution_dict) + '\n')

      max_i = np.argmax(batch_tot_r)
      max_tot_r = batch_tot_r[max_i]
      if max_tot_r >= self.top_reward:
        if max_tot_r >= self.top_reward:
          self.top_reward = max_tot_r
        logging.info('Top code: r=%.2f, \t%s', max_tot_r, code_strings[max_i])
      if self.top_episodes is not None:
        self.top_episodes.push(
            max_tot_r, tuple(batch_actions[max_i, :episode_lengths[max_i]]))

      summaries_list += [text_summary, reward_summary]

      if self.do_iw_summaries and not empty_replay_buffer:
        # prob of replay samples under replay buffer sampling.
        norm_replay_weights = [
            w / self.experience_replay.total_weight
            for w in replay_weights]
        replay_iw = self._compute_iw(replay_log_probs, replay_weights)
        on_policy_iw = self._compute_iw(on_policy_log_probs, on_policy_weights)
        summaries_list.append(
            self._iw_summary(
                session, replay_iw, replay_log_probs, norm_replay_weights,
                on_policy_iw, on_policy_log_probs))

    return UpdateStepResult(
        global_step=global_step,
        global_npe=global_npe,
        summaries_list=summaries_list,
        gradients_dict=fetched['gradients'])


def io_to_text(io_case, io_type):
  if isinstance(io_case, misc.IOTuple):
    # If there are many strings, join them with ','.
    return ','.join([io_to_text(e, io_type) for e in io_case])
  if io_type == misc.IOType.string:
    # There is one string. Return it.
    return misc.tokens_to_text(io_case)
  if (io_type == misc.IOType.integer
      or io_type == misc.IOType.boolean):
    if len(io_case) == 1:
      return str(io_case[0])
    return str(io_case)


CodeScoreInfo = namedtuple(
    'CodeScoreInfo',
    ['code_strings', 'batch_rewards', 'total_rewards', 'test_cases',
     'code_outputs', 'reasons'])


def compute_rewards(rl_batch, batch_actions, episode_lengths, batch_size=None):
  """Compute rewards for each episode in the batch.

  Args:
    rl_batch: A data.RLBatch instance. This holds information about the task
        each episode is solving, and a reward function for each episode.
    batch_actions: Contains batch of episodes. Each sequence of actions will be
        converted into a BF program and then scored. A numpy array of shape
        [batch_size, max_sequence_length].
    episode_lengths: The sequence length of each episode in the batch. Iterable
        of length batch_size.
    batch_size: (optional) number of programs to score. Use this to limit the
        number of programs executed from this batch. For example, when doing
        importance sampling some of the on-policy episodes will be discarded
        and they should not be executed. `batch_size` can be less than or equal
        to the size of the input batch.

  Returns:
    CodeScoreInfo namedtuple instance. This holds not just the computed rewards,
    but additional information computed during code execution which can be used
    for debugging and monitoring. this includes: BF code strings, test cases
    the code was executed on, code outputs from those test cases, and reasons
    for success or failure.
  """
  code_strings = [
      ''.join([misc.bf_int2char(a) for a in action_sequence[:l]])
      for action_sequence, l in zip(batch_actions, episode_lengths)]
  if batch_size is None:
    batch_size = len(code_strings)
  else:
    assert batch_size <= len(code_strings)
    code_strings = code_strings[:batch_size]

  if isinstance(rl_batch.reward_fns, (list, tuple)):
    # reward_fns is a list of functions, same length as code_strings.
    assert len(rl_batch.reward_fns) >= batch_size
    r_fn_results = [
        rl_batch.reward_fns[i](code_strings[i]) for i in xrange(batch_size)]
  else:
    # reward_fns is allowed to be one function which processes a batch of code
    # strings. This is useful for efficiency and batch level computation.
    r_fn_results = rl_batch.reward_fns(code_strings)

  # Expecting that r_fn returns a list of rewards. Length of list equals
  # length of the code string (including EOS char).

  batch_rewards = [r.episode_rewards for r in r_fn_results]
  total_rewards = [sum(b) for b in batch_rewards]
  test_cases = [io_to_text(r.input_case, r.input_type) for r in r_fn_results]
  code_outputs = [io_to_text(r.code_output, r.output_type)
                  for r in r_fn_results]
  reasons = [r.reason for r in r_fn_results]
  return CodeScoreInfo(
      code_strings=code_strings,
      batch_rewards=batch_rewards,
      total_rewards=total_rewards,
      test_cases=test_cases,
      code_outputs=code_outputs,
      reasons=reasons)


def process_episodes(
    batch_rewards, episode_lengths, a2c=False, baselines=None,
    batch_values=None):
  """Compute REINFORCE targets.

  REINFORCE here takes the form:
  grad_t = grad[log(pi(a_t|c_t))*target_t]
  where c_t is context: i.e. RNN state or environment state (or both).

  Two types of targets are supported:
  1) Advantage actor critic (a2c).
  2) Vanilla REINFORCE with baseline.

  Args:
    batch_rewards: Rewards received in each episode in the batch. A numpy array
        of shape [batch_size, max_sequence_length]. Note, these are per-timestep
        rewards, not total reward.
    episode_lengths: Length of each episode. An iterable of length batch_size.
    a2c: A bool. Whether to compute a2c targets (True) or vanilla targets
        (False).
    baselines: If a2c is False, provide baselines for each timestep. This is a
        list (or indexable container) of length max_time. Note: baselines are
        shared across all episodes, which is why there is no batch dimension.
        It is up to the caller to update baselines accordingly.
    batch_values: If a2c is True, provide values computed by a value estimator.
        A numpy array of shape [batch_size, max_sequence_length].

  Returns:
    batch_targets: REINFORCE targets for each episode and timestep. A numpy
        array of shape [batch_size, max_sequence_length].
    batch_returns: Returns computed for each episode and timestep. This is for
        reference, and is not used in the REINFORCE gradient update (but was
        used to compute the targets). A numpy array of shape
        [batch_size, max_sequence_length].
  """
  num_programs = len(batch_rewards)
  assert num_programs <= len(episode_lengths)
  batch_returns = [None] * num_programs
  batch_targets = [None] * num_programs
  for i in xrange(num_programs):
    episode_length = episode_lengths[i]
    assert len(batch_rewards[i]) == episode_length
    # Compute target for each timestep.
    # If we are computing A2C:
    #    target_t = advantage_t = R_t - V(c_t)
    #    where V(c_t) is a learned value function (provided as `values`).
    # Otherwise:
    #    target_t = R_t - baselines[t]
    #    where `baselines` are provided.
    # In practice we use a more generalized formulation of advantage. See docs
    # for `discounted_advantage_and_rewards`.
    if a2c:
      # Compute advantage.
      assert batch_values is not None
      episode_values = batch_values[i, :episode_length]
      episode_rewards = batch_rewards[i]
      emp_val, gen_adv = rollout_lib.discounted_advantage_and_rewards(
          episode_rewards, episode_values, gamma=1.0, lambda_=1.0)
      batch_returns[i] = emp_val
      batch_targets[i] = gen_adv
    else:
      # Compute return for each timestep. See section 3 of
      # https://arxiv.org/pdf/1602.01783.pdf
      assert baselines is not None
      empirical_returns = rollout_lib.discount(batch_rewards[i], gamma=1.0)
      targets = [None] * episode_length
      for j in xrange(episode_length):
        targets[j] = empirical_returns[j] - baselines[j]
      batch_returns[i] = empirical_returns
      batch_targets[i] = targets
  batch_returns = utils.stack_pad(batch_returns, 0)
  if num_programs:
    batch_targets = utils.stack_pad(batch_targets, 0)
  else:
    batch_targets = np.array([], dtype=np.float32)

  return (batch_targets, batch_returns)