Python tensorflow.where() Examples

def __init__(self, pad_mask):
    """Compute and store the location of the padding.

    Args:
      pad_mask (tf.Tensor): Reference padding tensor of shape
        [batch_size,length] or [dim_origin] (dim_origin=batch_size*length)
        containing non-zeros positive values to indicate padding location.
    """
    self.nonpad_ids = None
    self.dim_origin = None

    with tf.name_scope("pad_reduce/get_ids"):
      pad_mask = tf.reshape(pad_mask, [-1])  # Flatten the batch
      # nonpad_ids contains coordinates of zeros rows (as pad_mask is
      # float32, checking zero equality is done with |x| < epsilon, with
      # epsilon=1e-9 as standard, here pad_mask only contains positive values
      # so tf.abs would be redundant)
      self.nonpad_ids = tf.to_int32(tf.where(pad_mask < 1e-9))
      self.dim_origin = tf.shape(pad_mask)[:1] 

def shape_list(x):
  """Return list of dims, statically where possible."""
  x = tf.convert_to_tensor(x)

  # If unknown rank, return dynamic shape
  if x.get_shape().dims is None:
    return tf.shape(x)

  static = x.get_shape().as_list()
  shape = tf.shape(x)

  ret = []
  for i in range(len(static)):
    dim = static[i]
    if dim is None:
      dim = shape[i]
    ret.append(dim)
  return ret 

def coordinate_tensor(shape, axis):
  """Return a tensor with given shape containing coordinate along given axis.

  Args:
    shape: a Tensor representing the shape of the output Tensor
    axis: an integer

  Returns:
    A tensor with shape shape and type tf.int32, where each elements its
    coordinate along the given axis.
  """
  if axis < 0:
    axis = tf.size(shape) + axis  # Convert to positive for the one_hot indice

  r = tf.range(shape[axis])
  r_shape = tf.one_hot(
      axis, tf.size(shape), on_value=-1, off_value=1, dtype=tf.int32)
  return tf.zeros(shape, dtype=tf.int32) + tf.reshape(r, r_shape) 

def _relative_position_to_absolute_position_masked(x):
  """Helper to dot_product_self_attention_relative_v2.

  Rearrange an attention logits or weights Tensor.

  The dimensions of the input represent:
  [batch, heads, query_position, memory_position - query_position + length - 1]

  The dimensions of the output represent:
  [batch, heads, query_position, memory_position]

  Only works with masked_attention.  Undefined behavior for regions of the
  input where memory_position > query_position.

  Args:
    x: a Tensor with shape [batch, heads, length, length]

  Returns:
    a Tensor with shape [batch, heads, length, length]
  """
  batch, heads, length, _ = common_layers.shape_list(x)
  x = tf.pad(x, [[0, 0], [0, 0], [0, 0], [1, 0]])
  x = tf.reshape(x, [batch, heads, 1 + length, length])
  x = tf.slice(x, [0, 0, 1, 0], [-1, -1, -1, -1])
  return x 

def fprop(self, x, **kwargs):
        out = tf.nn.relu(x)
        if self.leak != 0.0:
            # The code commented below resulted in the time per epoch of
            # an 8-GPU wide resnet increasing by about 5% relative to the
            # code now in use.
            # The two different implementations have the same forward prop
            # down to machine precision on all inputs I have tested, but
            # sometimes have different derivatives.
            # Both obtain about the same training accuracy but the faster
            # version seems to also be slightly more accurate.
            # The commented code and these performance notes are included to
            # aid future revision efforts.
            #
            # out = out - self.leak * tf.nn.relu(-x)
            #

            out = tf.where(tf.less(x, 0.0), self.leak * x, x)
        return out 

def neural_gpu_body(inputs, hparams, name=None):
  """The core Neural GPU."""
  with tf.variable_scope(name, "neural_gpu"):

    def step(state, inp):  # pylint: disable=missing-docstring
      x = tf.nn.dropout(state, 1.0 - hparams.dropout)
      for layer in range(hparams.num_hidden_layers):
        x = common_layers.conv_gru(
            x, (hparams.kernel_height, hparams.kernel_width),
            hparams.hidden_size,
            name="cgru_%d" % layer)
      # Padding input is zeroed-out in the modality, we check this by summing.
      padding_inp = tf.less(tf.reduce_sum(tf.abs(inp), axis=[1, 2]), 0.00001)
      new_state = tf.where(padding_inp, state, x)  # No-op where inp is padding.
      return new_state

    return tf.foldl(
        step,
        tf.transpose(inputs, [1, 0, 2, 3]),
        initializer=inputs,
        parallel_iterations=1,
        swap_memory=True) 

def _to_bfloat16_unbiased(x, noise):
  """Convert a float32 to a bfloat16 using randomized roundoff.

  Args:
    x: A float32 Tensor.
    noise: a float32 Tensor with values in [0, 1), broadcastable to tf.shape(x)
  Returns:
    A float32 Tensor.
  """
  x_sign = tf.sign(x)
  # Make sure x is positive.  If it is zero, the two candidates are identical.
  x = x * x_sign + 1e-30
  cand1 = tf.to_bfloat16(x)
  cand1_f = tf.to_float(cand1)
  # This relies on the fact that for a positive bfloat16 b,
  # b * 1.005 gives you the next higher bfloat16 and b*0.995 gives you the
  # next lower one. Both 1.005 and 0.995 are ballpark estimation.
  cand2 = tf.to_bfloat16(
      tf.where(tf.greater(x, cand1_f), cand1_f * 1.005, cand1_f * 0.995))
  ret = _randomized_roundoff_to_bfloat16(x, noise, cand1, cand2)
  return ret * tf.to_bfloat16(x_sign) 

def _absolute_position_to_relative_position_masked(x):
  """Helper to dot_product_self_attention_relative_v2.

  Rearrange an attention logits or weights Tensor.

  The dimensions of the input represent:
  [batch, heads, query_position, memory_position]

  The dimensions of the output represent:
  [batch, heads, query_position, memory_position - query_position + length - 1]

  Only works with masked_attention.  Undefined behavior for regions of the
  input where memory_position > query_position.

  Args:
    x: a Tensor with shape [batch, heads, length, length]

  Returns:
    a Tensor with shape [batch, heads, length, length]
  """
  batch, heads, length, _ = common_layers.shape_list(x)
  x = tf.pad(x, [[0, 0], [0, 0], [1, 0], [0, 0]])
  x = tf.reshape(x, [batch, heads, length, length + 1])
  x = tf.slice(x, [0, 0, 0, 1], [batch, heads, length, length])
  return x 

def filter_groundtruth_with_nan_box_coordinates(tensor_dict):
  """Filters out groundtruth with no bounding boxes.

  Args:
    tensor_dict: a dictionary of following groundtruth tensors -
      fields.InputDataFields.groundtruth_boxes
      fields.InputDataFields.groundtruth_classes
      fields.InputDataFields.groundtruth_is_crowd
      fields.InputDataFields.groundtruth_area
      fields.InputDataFields.groundtruth_label_types

  Returns:
    a dictionary of tensors containing only the groundtruth that have bounding
    boxes.
  """
  groundtruth_boxes = tensor_dict[fields.InputDataFields.groundtruth_boxes]
  nan_indicator_vector = tf.greater(tf.reduce_sum(tf.to_int32(
      tf.is_nan(groundtruth_boxes)), reduction_indices=[1]), 0)
  valid_indicator_vector = tf.logical_not(nan_indicator_vector)
  valid_indices = tf.where(valid_indicator_vector)

  return retain_groundtruth(tensor_dict, valid_indices) 

def _randomized_roundoff_to_bfloat16(x, noise, cand1, cand2):
  """Round-off x to cand1 or to cand2 in an unbiased way.

  Cand1 and cand2 are the same shape as x.
  For every element of x, the corresponding elements of cand1 and cand2 should
  be the two closest bfloat16 values to x.  Order does not matter.
  cand1 and cand2 must differ from each other.

  Args:
    x: A float32 Tensor.
    noise: A Tensor broadcastable to the shape of x containing
    random uniform values in [0.0, 1.0].
    cand1: A bfloat16 Tensor the same shape as x.
    cand2: A bfloat16 Tensor the same shape as x.

  Returns:
    A bfloat16 Tensor.
  """
  cand1_f = tf.to_float(cand1)
  cand2_f = tf.to_float(cand2)
  step_size = cand2_f - cand1_f
  fpart = (x - cand1_f) / step_size
  ret = tf.where(tf.greater(fpart, noise), cand2, cand1)
  return ret 

def __init__(self, num_experts, gates):
    """Create a SparseDispatcher.

    Args:
      num_experts: an integer.
      gates: a `Tensor` of shape `[batch_size, num_experts]`.

    Returns:
      a SparseDispatcher
    """
    self._gates = gates
    self._num_experts = num_experts

    where = tf.to_int32(tf.where(tf.transpose(gates) > 0))
    self._expert_index, self._batch_index = tf.unstack(where, num=2, axis=1)
    self._part_sizes_tensor = tf.reduce_sum(tf.to_int32(gates > 0), [0])
    self._nonzero_gates = tf.gather(
        tf.reshape(self._gates, [-1]),
        self._batch_index * num_experts + self._expert_index) 

def retain_groundtruth_with_positive_classes(tensor_dict):
  """Retains only groundtruth with positive class ids.

  Args:
    tensor_dict: a dictionary of following groundtruth tensors -
      fields.InputDataFields.groundtruth_boxes
      fields.InputDataFields.groundtruth_classes
      fields.InputDataFields.groundtruth_is_crowd
      fields.InputDataFields.groundtruth_area
      fields.InputDataFields.groundtruth_label_types
      fields.InputDataFields.groundtruth_difficult

  Returns:
    a dictionary of tensors containing only the groundtruth with positive
    classes.

  Raises:
    ValueError: If groundtruth_classes tensor is not in tensor_dict.
  """
  if fields.InputDataFields.groundtruth_classes not in tensor_dict:
    raise ValueError('`groundtruth classes` not in tensor_dict.')
  keep_indices = tf.where(tf.greater(
      tensor_dict[fields.InputDataFields.groundtruth_classes], 0))
  return retain_groundtruth(tensor_dict, keep_indices) 

def __init__(self, endpoints, interpolation=linear_interpolation, outside_value=None):
        """Piecewise schedule.
        endpoints: [(int, int)]
            list of pairs `(time, value)` meanining that schedule should output
            `value` when `t==time`. All the values for time must be sorted in
            an increasing order. When t is between two times, e.g. `(time_a, value_a)`
            and `(time_b, value_b)`, such that `time_a <= t < time_b` then value outputs
            `interpolation(value_a, value_b, alpha)` where alpha is a fraction of
            time passed between `time_a` and `time_b` for time `t`.
        interpolation: lambda float, float, float: float
            a function that takes value to the left and to the right of t according
            to the `endpoints`. Alpha is the fraction of distance from left endpoint to
            right endpoint that t has covered. See linear_interpolation for example.
        outside_value: float
            if the value is requested outside of all the intervals sepecified in
            `endpoints` this value is returned. If None then AssertionError is
            raised when outside value is requested.
        """
        idxes = [e[0] for e in endpoints]
        assert idxes == sorted(idxes)
        self._interpolation = interpolation
        self._outside_value = outside_value
        self._endpoints      = endpoints 

def _compute_loss(self, prediction_tensor, target_tensor, weights):
    """Compute loss function.

    Args:
      prediction_tensor: A float tensor of shape [batch_size, num_anchors,
        code_size] representing the (encoded) predicted locations of objects.
      target_tensor: A float tensor of shape [batch_size, num_anchors,
        code_size] representing the regression targets
      weights: a float tensor of shape [batch_size, num_anchors]

    Returns:
      loss: a (scalar) tensor representing the value of the loss function
    """
    diff = prediction_tensor - target_tensor
    abs_diff = tf.abs(diff)
    abs_diff_lt_1 = tf.less(abs_diff, 1)
    anchorwise_smooth_l1norm = tf.reduce_sum(
        tf.where(abs_diff_lt_1, 0.5 * tf.square(abs_diff), abs_diff - 0.5),
        2) * weights
    if self._anchorwise_output:
      return anchorwise_smooth_l1norm
    return tf.reduce_sum(anchorwise_smooth_l1norm) 

def match(self, similarity_matrix, scope=None, **params):
    """Computes matches among row and column indices and returns the result.

    Computes matches among the row and column indices based on the similarity
    matrix and optional arguments.

    Args:
      similarity_matrix: Float tensor of shape [N, M] with pairwise similarity
        where higher value means more similar.
      scope: Op scope name. Defaults to 'Match' if None.
      **params: Additional keyword arguments for specific implementations of
        the Matcher.

    Returns:
      A Match object with the results of matching.
    """
    with tf.name_scope(scope, 'Match', [similarity_matrix, params]) as scope:
      return Match(self._match(similarity_matrix, **params)) 

def _match(self, similarity_matrix, **params):
    """Method to be overriden by implementations.

    Args:
      similarity_matrix: Float tensor of shape [N, M] with pairwise similarity
        where higher value means more similar.
      **params: Additional keyword arguments for specific implementations of
        the Matcher.

    Returns:
      match_results: Integer tensor of shape [M]: match_results[i]>=0 means
        that column i is matched to row match_results[i], match_results[i]=-1
        means that the column is not matched. match_results[i]=-2 means that
        the column is ignored (usually this happens when there is a very weak
        match which one neither wants as positive nor negative example).
    """
    pass 

def maximum_mean_discrepancy(x, y, kernel=utils.gaussian_kernel_matrix):
  r"""Computes the Maximum Mean Discrepancy (MMD) of two samples: x and y.

  Maximum Mean Discrepancy (MMD) is a distance-measure between the samples of
  the distributions of x and y. Here we use the kernel two sample estimate
  using the empirical mean of the two distributions.

  MMD^2(P, Q) = || \E{\phi(x)} - \E{\phi(y)} ||^2
              = \E{ K(x, x) } + \E{ K(y, y) } - 2 \E{ K(x, y) },

  where K = <\phi(x), \phi(y)>,
    is the desired kernel function, in this case a radial basis kernel.

  Args:
      x: a tensor of shape [num_samples, num_features]
      y: a tensor of shape [num_samples, num_features]
      kernel: a function which computes the kernel in MMD. Defaults to the
              GaussianKernelMatrix.

  Returns:
      a scalar denoting the squared maximum mean discrepancy loss.
  """
  with tf.name_scope('MaximumMeanDiscrepancy'):
    # \E{ K(x, x) } + \E{ K(y, y) } - 2 \E{ K(x, y) }
    cost = tf.reduce_mean(kernel(x, x))
    cost += tf.reduce_mean(kernel(y, y))
    cost -= 2 * tf.reduce_mean(kernel(x, y))

    # We do not allow the loss to become negative.
    cost = tf.where(cost > 0, cost, 0, name='value')
  return cost 

def iou(boxlist1, boxlist2, scope=None):
  """Computes pairwise intersection-over-union between box collections.

  Args:
    boxlist1: BoxList holding N boxes
    boxlist2: BoxList holding M boxes
    scope: name scope.

  Returns:
    a tensor with shape [N, M] representing pairwise iou scores.
  """
  with tf.name_scope(scope, 'IOU'):
    intersections = intersection(boxlist1, boxlist2)
    areas1 = area(boxlist1)
    areas2 = area(boxlist2)
    unions = (
        tf.expand_dims(areas1, 1) + tf.expand_dims(areas2, 0) - intersections)
    return tf.where(
        tf.equal(intersections, 0.0),
        tf.zeros_like(intersections), tf.truediv(intersections, unions)) 

def matched_iou(boxlist1, boxlist2, scope=None):
  """Compute intersection-over-union between corresponding boxes in boxlists.

  Args:
    boxlist1: BoxList holding N boxes
    boxlist2: BoxList holding N boxes
    scope: name scope.

  Returns:
    a tensor with shape [N] representing pairwise iou scores.
  """
  with tf.name_scope(scope, 'MatchedIOU'):
    intersections = matched_intersection(boxlist1, boxlist2)
    areas1 = area(boxlist1)
    areas2 = area(boxlist2)
    unions = areas1 + areas2 - intersections
    return tf.where(
        tf.equal(intersections, 0.0),
        tf.zeros_like(intersections), tf.truediv(intersections, unions)) 

def dsn_loss_coefficient(params):
  """The global_step-dependent weight that specifies when to kick in DSN losses.

  Args:
    params: A dictionary of parameters. Expecting 'domain_separation_startpoint'

  Returns:
    A weight to that effectively enables or disables the DSN-related losses,
    i.e. similarity, difference, and reconstruction losses.
  """
  return tf.where(
      tf.less(slim.get_or_create_global_step(),
              params['domain_separation_startpoint']), 1e-10, 1.0)


################################################################################
# MODEL CREATION
################################################################################ 

def filter_field_value_equals(boxlist, field, value, scope=None):
  """Filter to keep only boxes with field entries equal to the given value.

  Args:
    boxlist: BoxList holding N boxes.
    field: field name for filtering.
    value: scalar value.
    scope: name scope.

  Returns:
    a BoxList holding M boxes where M <= N

  Raises:
    ValueError: if boxlist not a BoxList object or if it does not have
      the specified field.
  """
  with tf.name_scope(scope, 'FilterFieldValueEquals'):
    if not isinstance(boxlist, box_list.BoxList):
      raise ValueError('boxlist must be a BoxList')
    if not boxlist.has_field(field):
      raise ValueError('boxlist must contain the specified field')
    filter_field = boxlist.get_field(field)
    gather_index = tf.reshape(tf.where(tf.equal(filter_field, value)), [-1])
    return gather(boxlist, gather_index) 

def full_stack(self, b, x_size, bottleneck_bits, losses, is_training, i):
    stack1_b = self.stack(b, x_size, bottleneck_bits, "step%d" % i)
    if i > 1:
      stack1_b = self.full_stack(stack1_b, 2 * x_size, 2 * bottleneck_bits,
                                 losses, is_training, i - 1)
    b1, b_pred = self.unstack(stack1_b, x_size, bottleneck_bits, "step%d" % i)
    losses["stack%d_loss" % i] = self.stack_loss(b, b_pred, "step%d" % i)
    b_shape = common_layers.shape_list(b)
    if is_training:
      condition = tf.less(tf.random_uniform([]), 0.5)
      condition = tf.reshape(condition, [1] * len(b.shape))
      condition = tf.tile(condition, b.shape)
      b1 = tf.where(condition, b, b1)
    return tf.reshape(b1, b_shape) 

def difference_loss(private_samples, shared_samples, weight=1.0, name=''):
  """Adds the difference loss between the private and shared representations.

  Args:
    private_samples: a tensor of shape [num_samples, num_features].
    shared_samples: a tensor of shape [num_samples, num_features].
    weight: the weight of the incoherence loss.
    name: the name of the tf summary.
  """
  private_samples -= tf.reduce_mean(private_samples, 0)
  shared_samples -= tf.reduce_mean(shared_samples, 0)

  private_samples = tf.nn.l2_normalize(private_samples, 1)
  shared_samples = tf.nn.l2_normalize(shared_samples, 1)

  correlation_matrix = tf.matmul(
      private_samples, shared_samples, transpose_a=True)

  cost = tf.reduce_mean(tf.square(correlation_matrix)) * weight
  cost = tf.where(cost > 0, cost, 0, name='value')

  tf.summary.scalar('losses/Difference Loss {}'.format(name),
                                       cost)
  assert_op = tf.Assert(tf.is_finite(cost), [cost])
  with tf.control_dependencies([assert_op]):
    tf.losses.add_loss(cost)


################################################################################
# TASK LOSS
################################################################################ 

def selu(x):
    alpha = 1.6732632423543772848170429916717
    scale = 1.0507009873554804934193349852946
    return scale * tf.where(x > 0.0, x, alpha * tf.exp(x) - alpha) 

def huber_loss(labels, predictions, delta=1.0):
    residual = tf.abs(predictions - labels)
    condition = tf.less(residual, delta)
    small_res = 0.5 * tf.square(residual)
    large_res = delta * residual - 0.5 * tf.square(delta)
    return tf.where(condition, small_res, large_res) 

def Attention(Q, K, V, mononotic_attention=False, prev_max_attentions=None):
    '''
    Args:
      Q: Queries. (B, T/r, d)
      K: Keys. (B, N, d)
      V: Values. (B, N, d)
      mononotic_attention: A boolean. At training, it is False.
      prev_max_attentions: (B,). At training, it is set to None.

    Returns:
      R: [Context Vectors; Q]. (B, T/r, 2d)
      alignments: (B, N, T/r)
      max_attentions: (B, T/r)
    '''
    A = tf.matmul(Q, K, transpose_b=True) * tf.rsqrt(tf.to_float(hp.d))
    if mononotic_attention:  # for inference
        key_masks = tf.sequence_mask(prev_max_attentions, hp.max_N)
        reverse_masks = tf.sequence_mask(hp.max_N - hp.attention_win_size - prev_max_attentions, hp.max_N)[:, ::-1]
        masks = tf.logical_or(key_masks, reverse_masks)
        masks = tf.tile(tf.expand_dims(masks, 1), [1, hp.max_T, 1])
        paddings = tf.ones_like(A) * (-2 ** 32 + 1)  # (B, T/r, N)
        A = tf.where(tf.equal(masks, False), A, paddings)
    A = tf.nn.softmax(A) # (B, T/r, N)
    max_attentions = tf.argmax(A, -1)  # (B, T/r)
    R = tf.matmul(A, V)
    R = tf.concat((R, Q), -1)

    alignments = tf.transpose(A, [0, 2, 1]) # (B, N, T/r)

    return R, alignments, max_attentions 

def _apply_with_random_selector_tuples(x, func, num_cases):
  """Computes func(x, sel), with sel sampled from [0...num_cases-1].

  Args:
    x: A tuple of input tensors.
    func: Python function to apply.
    num_cases: Python int32, number of cases to sample sel from.

  Returns:
    The result of func(x, sel), where func receives the value of the
    selector as a python integer, but sel is sampled dynamically.
  """
  num_inputs = len(x)
  rand_sel = tf.random_uniform([], maxval=num_cases, dtype=tf.int32)
  # Pass the real x only to one of the func calls.

  tuples = [list() for t in x]
  for case in range(num_cases):
    new_x = [control_flow_ops.switch(t, tf.equal(rand_sel, case))[1] for t in x]
    output = func(tuple(new_x), case)
    for j in range(num_inputs):
      tuples[j].append(output[j])

  for i in range(num_inputs):
    tuples[i] = control_flow_ops.merge(tuples[i])[0]
  return tuple(tuples) 

def non_max_suppression(boxlist, thresh, max_output_size, scope=None):
  """Non maximum suppression.

  This op greedily selects a subset of detection bounding boxes, pruning
  away boxes that have high IOU (intersection over union) overlap (> thresh)
  with already selected boxes.  Note that this only works for a single class ---
  to apply NMS to multi-class predictions, use MultiClassNonMaxSuppression.

  Args:
    boxlist: BoxList holding N boxes.  Must contain a 'scores' field
      representing detection scores.
    thresh: scalar threshold
    max_output_size: maximum number of retained boxes
    scope: name scope.

  Returns:
    a BoxList holding M boxes where M <= max_output_size
  Raises:
    ValueError: if thresh is not in [0, 1]
  """
  with tf.name_scope(scope, 'NonMaxSuppression'):
    if not 0 <= thresh <= 1.0:
      raise ValueError('thresh must be between 0 and 1')
    if not isinstance(boxlist, box_list.BoxList):
      raise ValueError('boxlist must be a BoxList')
    if not boxlist.has_field('scores'):
      raise ValueError('input boxlist must have \'scores\' field')
    selected_indices = tf.image.non_max_suppression(
        boxlist.get(), boxlist.get_field('scores'),
        max_output_size, iou_threshold=thresh)
    return gather(boxlist, selected_indices) 

def __init__(self, sess, state_size, output_size, actor, critic):
        self.sess = sess
        self.state_size = state_size
        self.output_size = output_size
        self.gamma = 0.99
        self.lamda = 0.95
        self.epoch = 10
        self.actor = actor
        self.critic = critic
        self.lr = 0.00025
        self.ppo_eps = 0.2

        self.pi, self.logp, self.logp_pi = self.actor.pi, self.actor.logp, self.actor.logp_pi
        self.v = self.critic.v

        self.adv_ph = tf.placeholder(tf.float32, shape=[None])
        self.ret_ph = tf.placeholder(tf.float32, shape=[None])
        self.logp_old_ph = tf.placeholder(tf.float32, shape=[None])

        self.all_phs = [self.actor.input, self.critic.input, self.actor.action, self.adv_ph,
                        self.ret_ph, self.logp_old_ph]
        self.get_action_ops = [self.pi, self.v, self.logp_pi]

        self.ratio = tf.exp(self.logp - self.logp_old_ph)
        self.min_adv = tf.where(self.adv_ph > 0, (1.0 + self.ppo_eps)*self.adv_ph, (1.0 - self.ppo_eps)*self.adv_ph)
        self.pi_loss = -tf.reduce_mean(tf.minimum(self.ratio * self.adv_ph, self.min_adv))
        self.v_loss = tf.reduce_mean((self.ret_ph - self.v) ** 2)

        self.train_pi = tf.train.AdamOptimizer(self.lr).minimize(self.pi_loss)
        self.train_v = tf.train.AdamOptimizer(self.lr).minimize(self.v_loss)

        self.approx_kl = tf.reduce_mean(self.logp_old_ph - self.logp)
        self.approx_ent = tf.reduce_mean(-self.logp) 

def selu(x):
    alpha = 1.6732632423543772848170429916717
    scale = 1.0507009873554804934193349852946
    return scale * tf.where(x > 0.0, x, alpha * tf.exp(x) - alpha)