Python tensorflow.python.ops.nn_ops.softmax() Examples

def mc_loss(labels, logits):
    r"""
    A multi-class cross-entropy loss
    :param labels: [batch_size, ] - Tensor of the correct class ids
    :param logits: [batch_size, num_classes] - Unscaled logits
    :return: [batch_size, ] - Tensor of average costs for each batch element
    """

    num_classes = array_ops.shape(logits)[1]
    onehot_labels = array_ops.one_hot(labels, num_classes, dtype=logits.dtype)

    p = nn_ops.softmax(logits)
    p = clip_ops.clip_by_value(p, 1e-7, 1.0 - 1e-7)

    ce_loss = - onehot_labels * math_ops.log(p) - (1 - onehot_labels) * math_ops.log(1.0-p)

    cost = math_ops.reduce_sum(ce_loss, axis=1)
    return cost 

def _attention(self, query, attn_states):
    conv2d = nn_ops.conv2d
    reduce_sum = math_ops.reduce_sum
    softmax = nn_ops.softmax
    tanh = math_ops.tanh

    with vs.variable_scope("attention"):
      k = vs.get_variable(
          "attn_w", [1, 1, self._attn_size, self._attn_vec_size])
      v = vs.get_variable("attn_v", [self._attn_vec_size])
      hidden = array_ops.reshape(attn_states,
                                 [-1, self._attn_length, 1, self._attn_size])
      hidden_features = conv2d(hidden, k, [1, 1, 1, 1], "SAME")
      y = _linear(query, self._attn_vec_size, True)
      y = array_ops.reshape(y, [-1, 1, 1, self._attn_vec_size])
      s = reduce_sum(v * tanh(hidden_features + y), [2, 3])
      a = softmax(s)
      d = reduce_sum(
          array_ops.reshape(a, [-1, self._attn_length, 1, 1]) * hidden, [1, 2])
      new_attns = array_ops.reshape(d, [-1, self._attn_size])
      new_attn_states = array_ops.slice(attn_states, [0, 1, 0], [-1, -1, -1])
      return new_attns, new_attn_states 

def _kl_categorical_categorical(a, b, name=None):
  """Calculate the batched KL divergence KL(a || b) with a and b Categorical.

  Args:
    a: instance of a Categorical distribution object.
    b: instance of a Categorical distribution object.
    name: (optional) Name to use for created operations.
      default is "kl_categorical_categorical".

  Returns:
    Batchwise KL(a || b)
  """
  with ops.name_scope(name, "kl_categorical_categorical",
                      values=[a.logits, b.logits]):
    # sum(probs log(probs / (1 - probs)))
    delta_log_probs1 = (nn_ops.log_softmax(a.logits) -
                        nn_ops.log_softmax(b.logits))
    return math_ops.reduce_sum(nn_ops.softmax(a.logits) * delta_log_probs1,
                               axis=-1) 

def _kl_categorical_categorical(a, b, name=None):
  """Calculate the batched KL divergence KL(a || b) with a, b OneHotCategorical.

  Args:
    a: instance of a OneHotCategorical distribution object.
    b: instance of a OneHotCategorical distribution object.
    name: (optional) Name to use for created operations.
      default is "kl_categorical_categorical".

  Returns:
    Batchwise KL(a || b)
  """
  with ops.name_scope(name, "kl_categorical_categorical", values=[
      a.logits, b.logits]):
    # sum(p ln(p / q))
    return math_ops.reduce_sum(
        nn_ops.softmax(a.logits) * (nn_ops.log_softmax(a.logits)
                                    - nn_ops.log_softmax(b.logits)),
        axis=-1) 

def _kl_categorical_categorical(a, b, name=None):
  """Calculate the batched KL divergence KL(a || b) with a and b Categorical.

  Args:
    a: instance of a Categorical distribution object.
    b: instance of a Categorical distribution object.
    name: (optional) Name to use for created operations.
      default is "kl_categorical_categorical".

  Returns:
    Batchwise KL(a || b)
  """
  with ops.name_scope(
    name, "kl_categorical_categorical", [a.logits, b.logits]):
    # sum(p*ln(p/q))
    return math_ops.reduce_sum(
        nn_ops.softmax(a.logits)*(nn_ops.log_softmax(a.logits)
            - nn_ops.log_softmax(b.logits)), reduction_indices=[-1]) 

def testSoftmax3DUnknownSize(self):
    logits = np.ones((2, 3, 2))
    logits[0, 0, 0] = 0
    logits[1, 1, 1] = 0
    logit_placeholder = array_ops.placeholder(
        dtypes.float32, shape=(None, None, 2))
    feed_dict = {logit_placeholder: logits}
    exp_prediction = 0.5 * np.ones((2, 3, 2))
    exp_prediction[0, 0, 0] = self.low
    exp_prediction[0, 0, 1] = self.high
    exp_prediction[1, 1, 0] = self.high
    exp_prediction[1, 1, 1] = self.low

    prediction = _layers.softmax(logit_placeholder)
    with self.cached_session() as sess:
      prediction = sess.run(prediction, feed_dict=feed_dict)
      self.assertAllClose(exp_prediction, prediction) 

def _attention(self, query, attn_inputs):
    with vs.variable_scope("attention"):
      attn_query = tf.layers.dense(
        inputs=query, units=self._attn_vec_size, use_bias=True)
      attn_keys = tf.layers.dense(
        inputs=attn_inputs, units=self._attn_vec_size, use_bias=True)
      attn_contents = tf.layers.dense(
        inputs=attn_inputs, units=self._attn_size, use_bias=True)
      
      v_attn = vs.get_variable("attn_v", [self._attn_vec_size])
      scores = attn_sum_bahdanau(v_attn, attn_keys, attn_query)
      
      if self.attn_masks is not None:
        score_masks = self.attn_masks
        scores = scores * score_masks + (1.0 - score_masks) * tf.float32.min

      attn_weights = nn_ops.softmax(scores)
      new_attns = math_ops.reduce_sum(
        tf.expand_dims(attn_weights, -1) * attn_contents, [1])
      return new_attns 

def focal_loss(labels, logits, gamma=2.0):
    r"""
    Multi-class focal loss implementation: https://arxiv.org/abs/1708.02002
    :param labels: [batch_size, ] - Tensor of the correct class ids
    :param logits: [batch_size, num_classes] - Unscaled logits
    :param gamma: focal loss weight
    :return: [batch_size, ] - Tensor of average costs for each batch element
    """

    num_classes = array_ops.shape(logits)[1]
    onehot_labels = array_ops.one_hot(labels, num_classes, dtype=logits.dtype)

    p = nn_ops.softmax(logits)
    p = clip_ops.clip_by_value(p, 1e-7, 1.0 - 1e-7)

    f_loss = - onehot_labels * math_ops.pow(1.0 - p, gamma) * math_ops.log(p) \
             - (1 - onehot_labels) * math_ops.pow(p, gamma) * math_ops.log(1.0 - p)

    cost = math_ops.reduce_sum(f_loss, axis=1)
    return cost 

def _attention(self, query, attn_states):
    conv2d = nn_ops.conv2d
    reduce_sum = math_ops.reduce_sum
    softmax = nn_ops.softmax
    tanh = math_ops.tanh

    with vs.variable_scope("Attention"):
      k = vs.get_variable("AttnW", [1, 1, self._attn_size, self._attn_vec_size])
      v = vs.get_variable("AttnV", [self._attn_vec_size])
      hidden = array_ops.reshape(attn_states,
                                 [-1, self._attn_length, 1, self._attn_size])
      hidden_features = conv2d(hidden, k, [1, 1, 1, 1], "SAME")
      y = _linear(query, self._attn_vec_size, True)
      y = array_ops.reshape(y, [-1, 1, 1, self._attn_vec_size])
      s = reduce_sum(v * tanh(hidden_features + y), [2, 3])
      a = softmax(s)
      d = reduce_sum(
          array_ops.reshape(a, [-1, self._attn_length, 1, 1]) * hidden, [1, 2])
      new_attns = array_ops.reshape(d, [-1, self._attn_size])
      new_attn_states = array_ops.slice(attn_states, [0, 1, 0], [-1, -1, -1])
      return new_attns, new_attn_states 

def _kl_categorical_categorical(a, b, name=None):
  """Calculate the batched KL divergence KL(a || b) with a and b Categorical.

  Args:
    a: instance of a Categorical distribution object.
    b: instance of a Categorical distribution object.
    name: (optional) Name to use for created operations.
      default is "kl_categorical_categorical".

  Returns:
    Batchwise KL(a || b)
  """
  with ops.name_scope(
    name, "kl_categorical_categorical", [a.logits, b.logits]):
    # sum(p*ln(p/q))
    return math_ops.reduce_sum(
        nn_ops.softmax(a.logits)*(nn_ops.log_softmax(a.logits)
            - nn_ops.log_softmax(b.logits)), reduction_indices=[-1]) 

def test_unary_ops(self):
    ops = [
        ('relu', nn_ops.relu, nn.relu),
        ('relu6', nn_ops.relu6, nn.relu6),
        ('crelu', nn_ops.crelu, nn.crelu),
        ('elu', nn_ops.elu, nn.elu),
        ('softplus', nn_ops.softplus, nn.softplus),
        ('l2_loss', nn_ops.l2_loss, nn.l2_loss),
        ('softmax', nn_ops.softmax, nn.softmax),
        ('log_softmax', nn_ops.log_softmax, nn.log_softmax),
    ]
    for op_name, tf_op, lt_op in ops:
      golden_tensor = tf_op(self.original_lt.tensor)
      golden_lt = core.LabeledTensor(golden_tensor, self.axes)
      actual_lt = lt_op(self.original_lt)
      self.assertIn(op_name, actual_lt.name)
      self.assertLabeledTensorsEqual(golden_lt, actual_lt) 

def _attention(self, query, attn_states):
    conv2d = nn_ops.conv2d
    reduce_sum = math_ops.reduce_sum
    softmax = nn_ops.softmax
    tanh = math_ops.tanh

    with vs.variable_scope("attention"):
      k = vs.get_variable(
          "attn_w", [1, 1, self._attn_size, self._attn_vec_size])
      v = vs.get_variable("attn_v", [self._attn_vec_size])
      hidden = array_ops.reshape(attn_states,
                                 [-1, self._attn_length, 1, self._attn_size])
      hidden_features = conv2d(hidden, k, [1, 1, 1, 1], "SAME")
      y = _linear(query, self._attn_vec_size, True)
      y = array_ops.reshape(y, [-1, 1, 1, self._attn_vec_size])
      s = reduce_sum(v * tanh(hidden_features + y), [2, 3])
      a = softmax(s)
      d = reduce_sum(
          array_ops.reshape(a, [-1, self._attn_length, 1, 1]) * hidden, [1, 2])
      new_attns = array_ops.reshape(d, [-1, self._attn_size])
      new_attn_states = array_ops.slice(attn_states, [0, 1, 0], [-1, -1, -1])
      return new_attns, new_attn_states 

def _attention(self, query, attn_states):
    conv2d = nn_ops.conv2d
    reduce_sum = math_ops.reduce_sum
    softmax = nn_ops.softmax
    tanh = math_ops.tanh

    with vs.variable_scope("attention"):
      k = vs.get_variable(
          "attn_w", [1, 1, self._attn_size, self._attn_vec_size])
      v = vs.get_variable("attn_v", [self._attn_vec_size])
      hidden = array_ops.reshape(attn_states,
                                 [-1, self._attn_length, 1, self._attn_size])
      hidden_features = conv2d(hidden, k, [1, 1, 1, 1], "SAME")
      if self._linear3 is None:
        self._linear3 = _Linear(query, self._attn_vec_size, True)
      y = self._linear3(query)
      y = array_ops.reshape(y, [-1, 1, 1, self._attn_vec_size])
      s = reduce_sum(v * tanh(hidden_features + y), [2, 3])
      a = softmax(s)
      d = reduce_sum(
          array_ops.reshape(a, [-1, self._attn_length, 1, 1]) * hidden, [1, 2])
      new_attns = array_ops.reshape(d, [-1, self._attn_size])
      new_attn_states = array_ops.slice(attn_states, [0, 1, 0], [-1, -1, -1])
      return new_attns, new_attn_states 

def sequence_loss(logits,
                  targets,
                  weights,
                  average_across_timesteps=True,
                  average_across_batch=True,
                  softmax_loss_function=None,
                  name=None):
  """Weighted cross-entropy loss for a sequence of logits, batch-collapsed.

  Args:
    logits: List of 2D Tensors of shape [batch_size x num_decoder_symbols].
    targets: List of 1D batch-sized int32 Tensors of the same length as logits.
    weights: List of 1D batch-sized float-Tensors of the same length as logits.
    average_across_timesteps: If set, divide the returned cost by the total
      label weight.
    average_across_batch: If set, divide the returned cost by the batch size.
    softmax_loss_function: Function (labels, logits) -> loss-batch
      to be used instead of the standard softmax (the default if this is None).
      **Note that to avoid confusion, it is required for the function to accept
      named arguments.**
    name: Optional name for this operation, defaults to "sequence_loss".

  Returns:
    A scalar float Tensor: The average log-perplexity per symbol (weighted).

  Raises:
    ValueError: If len(logits) is different from len(targets) or len(weights).
  """
  with ops.name_scope(name, "sequence_loss", logits + targets + weights):
    cost = math_ops.reduce_sum(
        sequence_loss_by_example(
            logits,
            targets,
            weights,
            average_across_timesteps=average_across_timesteps,
            softmax_loss_function=softmax_loss_function))
    if average_across_batch:
      batch_size = array_ops.shape(targets[0])[0]
      return cost / math_ops.cast(batch_size, cost.dtype)
    else:
      return cost 

def sequence_loss(logits,
                  targets,
                  weights,
                  average_across_timesteps=True,
                  average_across_batch=True,
                  softmax_loss_function=None,
                  name=None):
  """Weighted cross-entropy loss for a sequence of logits, batch-collapsed.

  Args:
    logits: List of 2D Tensors of shape [batch_size x num_decoder_symbols].
    targets: List of 1D batch-sized int32 Tensors of the same length as logits.
    weights: List of 1D batch-sized float-Tensors of the same length as logits.
    average_across_timesteps: If set, divide the returned cost by the total
      label weight.
    average_across_batch: If set, divide the returned cost by the batch size.
    softmax_loss_function: Function (labels, logits) -> loss-batch
      to be used instead of the standard softmax (the default if this is None).
      **Note that to avoid confusion, it is required for the function to accept
      named arguments.**
    name: Optional name for this operation, defaults to "sequence_loss".

  Returns:
    A scalar float Tensor: The average log-perplexity per symbol (weighted).

  Raises:
    ValueError: If len(logits) is different from len(targets) or len(weights).
  """
  with ops.name_scope(name, "sequence_loss", logits + targets + weights):
    cost = math_ops.reduce_sum(
        sequence_loss_by_example(
            logits,
            targets,
            weights,
            average_across_timesteps=average_across_timesteps,
            softmax_loss_function=softmax_loss_function))
    if average_across_batch:
      batch_size = array_ops.shape(targets[0])[0]
      return cost / math_ops.cast(batch_size, cost.dtype)
    else:
      return cost 

def sequence_loss(logits,
                  targets,
                  weights,
                  average_across_timesteps=True,
                  average_across_batch=True,
                  softmax_loss_function=None,
                  name=None):
  """Weighted cross-entropy loss for a sequence of logits, batch-collapsed.

  Args:
    logits: List of 2D Tensors of shape [batch_size x num_decoder_symbols].
    targets: List of 1D batch-sized int32 Tensors of the same length as logits.
    weights: List of 1D batch-sized float-Tensors of the same length as logits.
    average_across_timesteps: If set, divide the returned cost by the total
      label weight.
    average_across_batch: If set, divide the returned cost by the batch size.
    softmax_loss_function: Function (labels, logits) -> loss-batch
      to be used instead of the standard softmax (the default if this is None).
      **Note that to avoid confusion, it is required for the function to accept
      named arguments.**
    name: Optional name for this operation, defaults to "sequence_loss".

  Returns:
    A scalar float Tensor: The average log-perplexity per symbol (weighted).

  Raises:
    ValueError: If len(logits) is different from len(targets) or len(weights).
  """
  with ops.name_scope(name, "sequence_loss", logits + targets + weights):
    cost = math_ops.reduce_sum(
        sequence_loss_by_example(
            logits,
            targets,
            weights,
            average_across_timesteps=average_across_timesteps,
            softmax_loss_function=softmax_loss_function))
    if average_across_batch:
      batch_size = array_ops.shape(targets[0])[0]
      return cost / math_ops.cast(batch_size, cost.dtype)
    else:
      return cost 

def sequence_loss(logits,
                  targets,
                  weights,
                  average_across_timesteps=True,
                  average_across_batch=True,
                  softmax_loss_function=None,
                  name=None):
  """Weighted cross-entropy loss for a sequence of logits, batch-collapsed.

  Args:
    logits: List of 2D Tensors of shape [batch_size x num_decoder_symbols].
    targets: List of 1D batch-sized int32 Tensors of the same length as logits.
    weights: List of 1D batch-sized float-Tensors of the same length as logits.
    average_across_timesteps: If set, divide the returned cost by the total
      label weight.
    average_across_batch: If set, divide the returned cost by the batch size.
    softmax_loss_function: Function (labels, logits) -> loss-batch
      to be used instead of the standard softmax (the default if this is None).
      **Note that to avoid confusion, it is required for the function to accept
      named arguments.**
    name: Optional name for this operation, defaults to "sequence_loss".

  Returns:
    A scalar float Tensor: The average log-perplexity per symbol (weighted).

  Raises:
    ValueError: If len(logits) is different from len(targets) or len(weights).
  """
  with ops.name_scope(name, "sequence_loss", logits + targets + weights):
    cost = math_ops.reduce_sum(
        sequence_loss_by_example(
            logits,
            targets,
            weights,
            average_across_timesteps=average_across_timesteps,
            softmax_loss_function=softmax_loss_function))
    if average_across_batch:
      batch_size = array_ops.shape(targets[0])[0]
      return cost / math_ops.cast(batch_size, cost.dtype)
    else:
      return cost 

def sequence_loss(logits,
                  targets,
                  weights,
                  average_across_timesteps=True,
                  average_across_batch=True,
                  softmax_loss_function=None,
                  name=None):
  """Weighted cross-entropy loss for a sequence of logits, batch-collapsed.

  Args:
    logits: List of 2D Tensors of shape [batch_size x num_decoder_symbols].
    targets: List of 1D batch-sized int32 Tensors of the same length as logits.
    weights: List of 1D batch-sized float-Tensors of the same length as logits.
    average_across_timesteps: If set, divide the returned cost by the total
      label weight.
    average_across_batch: If set, divide the returned cost by the batch size.
    softmax_loss_function: Function (labels, logits) -> loss-batch
      to be used instead of the standard softmax (the default if this is None).
      **Note that to avoid confusion, it is required for the function to accept
      named arguments.**
    name: Optional name for this operation, defaults to "sequence_loss".

  Returns:
    A scalar float Tensor: The average log-perplexity per symbol (weighted).

  Raises:
    ValueError: If len(logits) is different from len(targets) or len(weights).
  """
  with ops.name_scope(name, "sequence_loss", logits + targets + weights):
    cost = math_ops.reduce_sum(
        sequence_loss_by_example(
            logits,
            targets,
            weights,
            average_across_timesteps=average_across_timesteps,
            softmax_loss_function=softmax_loss_function))
    if average_across_batch:
      batch_size = array_ops.shape(targets[0])[0]
      return cost / math_ops.cast(batch_size, cost.dtype)
    else:
      return cost 

def inference_graph(self, data, data_spec=None):
    """Returns the op that performs inference on a batch of data."""

    return nn_ops.softmax(
        self._base_inference(
            data, data_spec=data_spec, soft=True))

  # pylint: disable=unused-argument 

def sequence_loss(logits,
                  targets,
                  weights,
                  average_across_timesteps=True,
                  average_across_batch=True,
                  softmax_loss_function=None,
                  name=None):
  """Weighted cross-entropy loss for a sequence of logits, batch-collapsed.

  Args:
    logits: List of 2D Tensors of shape [batch_size x num_decoder_symbols].
    targets: List of 1D batch-sized int32 Tensors of the same length as logits.
    weights: List of 1D batch-sized float-Tensors of the same length as logits.
    average_across_timesteps: If set, divide the returned cost by the total
      label weight.
    average_across_batch: If set, divide the returned cost by the batch size.
    softmax_loss_function: Function (labels, logits) -> loss-batch
      to be used instead of the standard softmax (the default if this is None).
      **Note that to avoid confusion, it is required for the function to accept
      named arguments.**
    name: Optional name for this operation, defaults to "sequence_loss".

  Returns:
    A scalar float Tensor: The average log-perplexity per symbol (weighted).

  Raises:
    ValueError: If len(logits) is different from len(targets) or len(weights).
  """
  with ops.name_scope(name, "sequence_loss", logits + targets + weights):
    cost = math_ops.reduce_sum(
        sequence_loss_by_example(
            logits,
            targets,
            weights,
            average_across_timesteps=average_across_timesteps,
            softmax_loss_function=softmax_loss_function))
    if average_across_batch:
      batch_size = array_ops.shape(targets[0])[0]
      return cost / math_ops.cast(batch_size, cost.dtype)
    else:
      return cost 

def _base_inference(self, data, data_spec=None, soft=False):
    if soft:
      inference_result = self.layers[0].soft_inference_graph(data)
    else:
      inference_result = self._do_layer_inference(self.layers[0], data)

    for layer in self.layers[1:]:
      inference_result = self._do_layer_inference(layer, inference_result)

    output_size = 1 if self.is_regression else self.params.num_classes
    output = layers.fully_connected(
        inference_result, output_size, activation_fn=nn_ops.softmax)
    return output 

def sequence_loss(logits, targets, weights,
                  average_across_timesteps=True, average_across_batch=True,
                  softmax_loss_function=None, name=None):
    """Weighted cross-entropy loss for a sequence of logits, batch-collapsed.

    Args:
      logits: List of 2D Tensors of shape [batch_size x num_decoder_symbols].
      targets: List of 1D batch-sized int32 Tensors of the same length as logits.
      weights: List of 1D batch-sized float-Tensors of the same length as logits.
      average_across_timesteps: If set, divide the returned cost by the total
        label weight.
      average_across_batch: If set, divide the returned cost by the batch size.
      softmax_loss_function: Function (inputs-batch, labels-batch) -> loss-batch
        to be used instead of the standard softmax (the default if this is None).
      name: Optional name for this operation, defaults to "sequence_loss".

    Returns:
      A scalar float Tensor: The average log-perplexity per symbol (weighted).

    Raises:
      ValueError: If len(logits) is different from len(targets) or len(weights).
    """
    with ops.name_scope(name, "sequence_loss", logits + targets + weights):
        cost = math_ops.reduce_sum(sequence_loss_by_example(
            logits, targets, weights,
            average_across_timesteps=average_across_timesteps,
            softmax_loss_function=softmax_loss_function))
        if average_across_batch:
            batch_size = array_ops.shape(targets[0])[0]
            return cost / math_ops.cast(batch_size, cost.dtype)
        else:
            return cost 

def sequence_loss(logits, targets, weights,
                  average_across_timesteps=True, average_across_batch=True,
                  softmax_loss_function=None, name=None):
  """Weighted cross-entropy loss for a sequence of logits, batch-collapsed.

  Args:
    logits: List of 2D Tensors of shape [batch_size x num_decoder_symbols].
    targets: List of 1D batch-sized int32 Tensors of the same length as logits.
    weights: List of 1D batch-sized float-Tensors of the same length as logits.
    average_across_timesteps: If set, divide the returned cost by the total
      label weight.
    average_across_batch: If set, divide the returned cost by the batch size.
    softmax_loss_function: Function (inputs-batch, labels-batch) -> loss-batch
      to be used instead of the standard softmax (the default if this is None).
    name: Optional name for this operation, defaults to "sequence_loss".

  Returns:
    A scalar float Tensor: The average log-perplexity per symbol (weighted).

  Raises:
    ValueError: If len(logits) is different from len(targets) or len(weights).
  """
  with ops.name_scope(name, "sequence_loss", logits + targets + weights):
    cost = math_ops.reduce_sum(sequence_loss_by_example(
        logits, targets, weights,
        average_across_timesteps=average_across_timesteps,
        softmax_loss_function=softmax_loss_function))
    if average_across_batch:
      batch_size = array_ops.shape(targets[0])[0]
      return cost / math_ops.cast(batch_size, dtypes.float32)
    else:
      return cost 

def _base_inference(self, data, data_spec=None):
    """Returns an op that performs inference without a softmax."""
    inference_result = self._do_layer_inference(self.layers[0], data)

    for layer in self.layers[1:]:
      inference_result = self._do_layer_inference(layer, inference_result)

    output_size = 1 if self.is_regression else self.params.num_classes
    output = layers.fully_connected(
        inference_result, output_size, activation_fn=array_ops.identity)

    return output 

def sequence_loss(logits, targets, weights,
                  average_across_timesteps=True, average_across_batch=True,
                  softmax_loss_function=None, name=None):
  """Weighted cross-entropy loss for a sequence of logits, batch-collapsed.

  Args:
    logits: List of 2D Tensors of shape [batch_size x num_decoder_symbols].
    targets: List of 1D batch-sized int32 Tensors of the same length as logits.
    weights: List of 1D batch-sized float-Tensors of the same length as logits.
    average_across_timesteps: If set, divide the returned cost by the total
      label weight.
    average_across_batch: If set, divide the returned cost by the batch size.
    softmax_loss_function: Function (inputs-batch, labels-batch) -> loss-batch
      to be used instead of the standard softmax (the default if this is None).
    name: Optional name for this operation, defaults to "sequence_loss".

  Returns:
    A scalar float Tensor: The average log-perplexity per symbol (weighted).

  Raises:
    ValueError: If len(logits) is different from len(targets) or len(weights).
  """
  with ops.name_scope(name, "sequence_loss", logits + targets + weights):
    cost = math_ops.reduce_sum(sequence_loss_by_example(
        logits, targets, weights,
        average_across_timesteps=average_across_timesteps,
        softmax_loss_function=softmax_loss_function))
    if average_across_batch:
      batch_size = array_ops.shape(targets[0])[0]
      return cost / math_ops.cast(batch_size, cost.dtype)
    else:
      return cost 

def test_empty_x_results_in_empty_output(self):
    # Empty x is common if someone masks their input with tf.boolean_mask in
    # order to drop missing entries, and in a particular batch all entries are
    # missing.
    with self.cached_session():
      x = np.array([]).reshape(0, 3)
      self.assertEqual(0, array_ops.size(x).eval())
      y = _layers.legacy_fully_connected(x, 2, activation_fn=nn_ops.softmax)
      variables_lib.global_variables_initializer().run()
      expected_y = np.array([]).reshape(0, 2)
      np.testing.assert_array_equal(expected_y, y.eval()) 

def testSoftmax3D(self):
    logits = np.ones((2, 3, 2))
    logits[0, 0, 0] = 0
    logits[1, 1, 1] = 0
    logits = constant_op.constant(logits)
    exp_prediction = 0.5 * np.ones((2, 3, 2))
    exp_prediction[0, 0, 0] = self.low
    exp_prediction[0, 0, 1] = self.high
    exp_prediction[1, 1, 0] = self.high
    exp_prediction[1, 1, 1] = self.low

    prediction = _layers.softmax(logits)
    with self.cached_session() as sess:
      prediction = sess.run(prediction)
      self.assertAllClose(exp_prediction, prediction) 

def testSoftmax2D(self):
    logits = constant_op.constant([[0.0, 1], [1, 1], [1, 0]])
    prediction = _layers.softmax(logits)
    exp_prediction = np.array([[self.low, self.high], [0.5, 0.5],
                               [self.high, self.low]])

    with self.cached_session() as sess:
      prediction = sess.run(prediction)
      self.assertAllClose(exp_prediction, prediction) 

def _fspecial_gauss(size, sigma):
    size = ops.convert_to_tensor(size, dtypes.int32)
    sigma = ops.convert_to_tensor(sigma)

    coords = math_ops.cast(math_ops.range(size), sigma.dtype)
    coords -= math_ops.cast(size - 1, sigma.dtype) / 2.0

    g = math_ops.square(coords)
    g *= -0.5 / math_ops.square(sigma)

    g = array_ops.reshape(g, shape=[1, -1]) + array_ops.reshape(g, shape=[-1, 1])
    g = array_ops.reshape(g, shape=[1, -1])  # For tf.nn.softmax().
    g = nn_ops.softmax(g)
    return array_ops.reshape(g, shape=[size, size, 1, 1]) 

def testSoftmaxUndefinedNthDimension(self):
    logits = array_ops.placeholder(dtypes.float32)
    with self.assertRaises(ValueError):
      _layers.softmax(logits)