Python tensorflow.python.ops.array_ops.stack() Examples

def _TileGrad(op, grad):
  """Sum reduces grad along the tiled dimensions."""
  assert isinstance(grad, ops.Tensor)
  input_shape = array_ops.shape(op.inputs[0])
  # We interleave multiples and input_shape to get split_shape,
  # reshape grad to split_shape, and reduce along all even
  # dimensions (the tiled dimensions) to get the result
  # with shape input_shape.  For example
  #   input_shape = [20, 30, 40]
  #   multiples = [2, 3, 4]
  #   split_shape = [2, 20, 3, 30, 4, 40]
  #   axes = [0, 2, 4]
  split_shape = array_ops.reshape(
      array_ops.transpose(array_ops.stack([op.inputs[1], input_shape])), [-1])
  axes = math_ops.range(0, array_ops.size(split_shape), 2)
  input_grad = math_ops.reduce_sum(array_ops.reshape(grad, split_shape), axes)
  # Fix shape inference
  input_grad.set_shape(op.inputs[0].get_shape())
  return [input_grad, None] 

def tensors_to_item(self, keys_to_tensors):
    tensor = keys_to_tensors[self._tensor_key]
    shape = self._shape
    if self._shape_keys:
      shape_dims = []
      for k in self._shape_keys:
        shape_dim = keys_to_tensors[k]
        if isinstance(shape_dim, sparse_tensor.SparseTensor):
          shape_dim = sparse_ops.sparse_tensor_to_dense(shape_dim)
        shape_dims.append(shape_dim)
      shape = array_ops.reshape(array_ops.stack(shape_dims), [-1])
    if isinstance(tensor, sparse_tensor.SparseTensor):
      if shape is not None:
        tensor = sparse_ops.sparse_reshape(tensor, shape)
      tensor = sparse_ops.sparse_tensor_to_dense(tensor, self._default_value)
    else:
      if shape is not None:
        tensor = array_ops.reshape(tensor, shape)
    return tensor 

def tensors_to_item(self, keys_to_tensors):
    indices = keys_to_tensors[self._indices_key]
    values = keys_to_tensors[self._values_key]
    if self._shape_key:
      shape = keys_to_tensors[self._shape_key]
      if isinstance(shape, sparse_tensor.SparseTensor):
        shape = sparse_ops.sparse_tensor_to_dense(shape)
    elif self._shape:
      shape = self._shape
    else:
      shape = indices.dense_shape
    indices_shape = array_ops.shape(indices.indices)
    rank = indices_shape[1]
    ids = math_ops.to_int64(indices.values)
    indices_columns_to_preserve = array_ops.slice(
        indices.indices, [0, 0], array_ops.stack([-1, rank - 1]))
    new_indices = array_ops.concat(
        [indices_columns_to_preserve, array_ops.reshape(ids, [-1, 1])], 1)

    tensor = sparse_tensor.SparseTensor(new_indices, values.values, shape)
    if self._densify:
      tensor = sparse_ops.sparse_tensor_to_dense(tensor, self._default_value)
    return tensor 

def _SliceGrad(op, grad):
  """Gradient for Slice op."""
  # Create an Nx2 padding where the first column represents how many
  # zeros are to be prepended for each dimension, and the second
  # column indicates how many zeros are appended.
  #
  # The number of zeros to append is the shape of the input
  # elementwise-subtracted by both the begin vector and sizes vector.
  #
  # Some more reshaping is needed to assemble this tensor with the
  # right dimensions.
  input_vec = op.inputs[0]
  begin_vec = op.inputs[1]
  input_rank = array_ops.rank(input_vec)
  slice_size = array_ops.shape(op.outputs[0])

  shape = array_ops.stack([input_rank, 1])
  before_pad = array_ops.reshape(begin_vec, shape)
  after_pad = array_ops.reshape(
      array_ops.shape(input_vec) - slice_size - begin_vec, shape)
  paddings = array_ops.concat([before_pad, after_pad], 1)
  return array_ops.pad(grad, paddings), None, None 

def _TileGrad(op, grad):
  """Sum reduces grad along the tiled dimensions."""
  assert isinstance(grad, ops.Tensor)
  input_shape = array_ops.shape(op.inputs[0])
  # We interleave multiples and input_shape to get split_shape,
  # reshape grad to split_shape, and reduce along all even
  # dimensions (the tiled dimensions) to get the result
  # with shape input_shape.  For example
  #   input_shape = [20, 30, 40]
  #   multiples = [2, 3, 4]
  #   split_shape = [2, 20, 3, 30, 4, 40]
  #   axes = [0, 2, 4]
  split_shape = array_ops.reshape(
      array_ops.transpose(array_ops.stack([op.inputs[1], input_shape])), [-1])
  axes = math_ops.range(0, array_ops.size(split_shape), 2)
  input_grad = math_ops.reduce_sum(array_ops.reshape(grad, split_shape), axes)
  # Fix shape inference
  input_grad.set_shape(op.inputs[0].get_shape())
  return [input_grad, None] 

def _PadGrad(op, grad):
  """Gradient for Pad."""
  # Pad introduces values around the original tensor, so the gradient function
  # slices the original shape out of the gradient."""
  x = op.inputs[0]
  a = op.inputs[1]  # [Rank(x), 2]
  # Takes a slice of a. The 1st column. [Rank(x), 1].
  pad_before = array_ops.slice(a, [0, 0],
                               array_ops.stack([array_ops.rank(x), 1]))
  # Make it a 1-D tensor.
  begin = array_ops.reshape(pad_before, [-1])
  sizes = array_ops.shape(x)
  return array_ops.slice(grad, begin, sizes), None


# ReverseSequence is just a permutation.  The gradient permutes back. 

def repeat(x, n):
  """Repeats a 2D tensor.

  if `x` has shape (samples, dim) and `n` is `2`,
  the output will have shape `(samples, 2, dim)`.

  Arguments:
      x: Tensor or variable.
      n: Python integer, number of times to repeat.

  Returns:
      A tensor.
  """
  assert ndim(x) == 2
  x = array_ops.expand_dims(x, 1)
  pattern = array_ops.stack([1, n, 1])
  return array_ops.tile(x, pattern) 

def _PadGrad(op, grad):
  """Gradient for Pad."""
  # Pad introduces values around the original tensor, so the gradient function
  # slices the original shape out of the gradient."""
  x = op.inputs[0]
  a = op.inputs[1]  # [Rank(x), 2]
  # Takes a slice of a. The 1st column. [Rank(x), 1].
  pad_before = array_ops.slice(a, [0, 0],
                               array_ops.stack([array_ops.rank(x), 1]))
  # Make it a 1-D tensor.
  begin = array_ops.reshape(pad_before, [-1])
  sizes = array_ops.shape(x)
  return array_ops.slice(grad, begin, sizes), None


# ReverseSequence is just a permutation.  The gradient permutes back. 

def _SliceGrad(op, grad):
  """Gradient for Slice op."""
  # Create an Nx2 padding where the first column represents how many
  # zeros are to be prepended for each dimension, and the second
  # column indicates how many zeros are appended.
  #
  # The number of zeros to append is the shape of the input
  # elementwise-subtracted by both the begin vector and sizes vector.
  #
  # Some more reshaping is needed to assemble this tensor with the
  # right dimensions.
  input_vec = op.inputs[0]
  begin_vec = op.inputs[1]
  input_rank = array_ops.rank(input_vec)
  slice_size = array_ops.shape(op.outputs[0])

  shape = array_ops.stack([input_rank, 1])
  before_pad = array_ops.reshape(begin_vec, shape)
  after_pad = array_ops.reshape(
      array_ops.shape(input_vec) - slice_size - begin_vec, shape)
  paddings = array_ops.concat([before_pad, after_pad], 1)
  return array_ops.pad(grad, paddings), None, None 

def _infer_fft_length_for_irfft(input_tensor, fft_rank):
  """Infers the `fft_length` argument for a `rank` IRFFT from `input_tensor`."""
  # A TensorShape for the inner fft_rank dimensions.
  fft_shape = input_tensor.get_shape()[-fft_rank:]

  # If any dim is unknown, fall back to tensor-based math.
  if not fft_shape.is_fully_defined():
    fft_length = _array_ops.unstack(_array_ops.shape(input_tensor)[-fft_rank:])
    fft_length[-1] = _math_ops.maximum(0, 2 * (fft_length[-1] - 1))
    return _array_ops.stack(fft_length)

  # Otherwise, return a constant.
  fft_length = fft_shape.as_list()
  if fft_length:
    fft_length[-1] = max(0, 2 * (fft_length[-1] - 1))
  return _ops.convert_to_tensor(fft_length, _dtypes.int32) 

def call(self, inputs):
    shape = inputs.get_shape().as_list()
    input_dim = shape[-1]
    output_shape = shape[:-1] + [self.units]
    if len(output_shape) > 2:
      # Reshape the input to 2D.
      output_shape_tensors = array_ops.unstack(array_ops.shape(inputs))
      output_shape_tensors[-1] = self.units
      output_shape_tensor = array_ops.stack(output_shape_tensors)
      inputs = array_ops.reshape(inputs, [-1, input_dim])

    outputs = standard_ops.matmul(inputs, self.kernel)
    if self.use_bias:
      outputs = nn.bias_add(outputs, self.bias)

    if len(output_shape) > 2:
      # Reshape the output back to the original ndim of the input.
      outputs = array_ops.reshape(outputs, output_shape_tensor)
      outputs.set_shape(output_shape)

    if self.activation is not None:
      return self.activation(outputs)  # pylint: disable=not-callable
    return outputs 

def tensors_to_item(self, keys_to_tensors):
    tensor = keys_to_tensors[self._tensor_key]
    shape = self._shape
    if self._shape_keys:
      shape_dims = []
      for k in self._shape_keys:
        shape_dim = keys_to_tensors[k]
        if isinstance(shape_dim, sparse_tensor.SparseTensor):
          shape_dim = sparse_ops.sparse_tensor_to_dense(shape_dim)
        shape_dims.append(shape_dim)
      shape = array_ops.reshape(array_ops.stack(shape_dims), [-1])
    if isinstance(tensor, sparse_tensor.SparseTensor):
      if shape is not None:
        tensor = sparse_ops.sparse_reshape(tensor, shape)
      tensor = sparse_ops.sparse_tensor_to_dense(tensor, self._default_value)
    else:
      if shape is not None:
        tensor = array_ops.reshape(tensor, shape)
    return tensor 

def tensors_to_item(self, keys_to_tensors):
    indices = keys_to_tensors[self._indices_key]
    values = keys_to_tensors[self._values_key]
    if self._shape_key:
      shape = keys_to_tensors[self._shape_key]
      if isinstance(shape, sparse_tensor.SparseTensor):
        shape = sparse_ops.sparse_tensor_to_dense(shape)
    elif self._shape:
      shape = self._shape
    else:
      shape = indices.dense_shape
    indices_shape = array_ops.shape(indices.indices)
    rank = indices_shape[1]
    ids = math_ops.to_int64(indices.values)
    indices_columns_to_preserve = array_ops.slice(
        indices.indices, [0, 0], array_ops.stack([-1, rank - 1]))
    new_indices = array_ops.concat(
        [indices_columns_to_preserve, array_ops.reshape(ids, [-1, 1])], 1)

    tensor = sparse_tensor.SparseTensor(new_indices, values.values, shape)
    if self._densify:
      tensor = sparse_ops.sparse_tensor_to_dense(tensor, self._default_value)
    return tensor 

def grad_sync(self):
    """A control trigger node for synchronization in the grad loop.

    One main use is to keep the pop ops of a stack executed in the
    iteration order.
    """
    if self._grad_sync is None:
      with ops.control_dependencies(None):
        self._grad_sync = control_trigger(name="b_sync")
      self._grad_sync._set_control_flow_context(self._grad_context)
      self._grad_index.op._add_control_input(self._grad_sync)
    return self._grad_sync 

def _TopKGrad(op, grad, _):
  """Return the gradients for TopK.

  Args:
    op: The TopKOp for which we need to generate gradients.
    grad: Tensor. The gradients passed to the TopKOp.

  Returns:
    A list of two tensors, the first being the gradient w.r.t to the input and
    TopK, and the second being the gradient w.r.t. to the indices (all zero).
  """
  in_shape = array_ops.shape(op.inputs[0])
  ind_shape = array_ops.shape(op.outputs[1])

  ind_lastdim = array_ops.gather(ind_shape, array_ops.size(ind_shape) - 1)
  # Flatten indices to 2D.
  ind_2d = array_ops.reshape(op.outputs[1], array_ops.stack([-1, ind_lastdim]))

  in_lastdim = array_ops.gather(in_shape, array_ops.size(in_shape) - 1)
  outerdim = array_ops.shape(ind_2d)[0]
  # Compute linear indices (flattened to 1D).
  ind = array_ops.reshape(ind_2d + array_ops.expand_dims(
      math_ops.range(0, outerdim * in_lastdim, in_lastdim), -1), [-1])

  # Substitute grad to appropriate locations and fill the rest with zeros,
  # finally reshaping it to the original input shape.
  return [array_ops.reshape(
      sparse_ops.sparse_to_dense(ind,
                                 array_ops.reshape(
                                     math_ops.reduce_prod(in_shape), [1]),
                                 array_ops.reshape(grad, [-1]),
                                 validate_indices=False),
      in_shape), array_ops.zeros(
          [], dtype=dtypes.int32)] 

def _define_distance_to_clusters(self, data):
    """Defines the Mahalanobis distance to the assigned Gaussian."""
    # TODO(xavigonzalvo): reuse (input - mean) * cov^-1 * (input -
    # mean) from log probability function.
    self._all_scores = []
    for shard in data:
      all_scores = []
      shard = array_ops.expand_dims(shard, 0)
      for c in xrange(self._num_classes):
        if self._covariance_type == FULL_COVARIANCE:
          cov = self._covs[c, :, :]
        elif self._covariance_type == DIAG_COVARIANCE:
          cov = array_ops.diag(self._covs[c, :])
        inverse = linalg_ops.matrix_inverse(cov + self._min_var)
        inv_cov = array_ops.tile(
            array_ops.expand_dims(inverse, 0),
            array_ops.stack([self._num_examples, 1, 1]))
        diff = array_ops.transpose(shard - self._means[c, :, :], perm=[1, 0, 2])
        m_left = math_ops.matmul(diff, inv_cov)
        all_scores.append(
            math_ops.sqrt(
                math_ops.matmul(
                    m_left, array_ops.transpose(
                        diff, perm=[0, 2, 1]))))
      self._all_scores.append(
          array_ops.reshape(
              array_ops.concat(all_scores, 1),
              array_ops.stack([self._num_examples, self._num_classes])))

    # Distance to the associated class.
    self._all_scores = array_ops.concat(self._all_scores, 0)
    assignments = array_ops.concat(self.assignments(), 0)
    rows = math_ops.to_int64(math_ops.range(0, self._num_examples))
    indices = array_ops.concat(
        [array_ops.expand_dims(rows, 1), array_ops.expand_dims(assignments, 1)],
        1)
    self._scores = array_ops.gather_nd(self._all_scores, indices) 

def _UnpackGrad(op, *grads):
  """Gradient for unpack op."""
  return array_ops.stack(grads, axis=op.get_attr("axis")) 

def forward_sync(self):
    """A control trigger node for synchronization in the forward loop.

    One main use is to keep the push ops of a stack executed in the
    iteration order.
    """
    if self._forward_sync is None:
      with ops.control_dependencies(None):
        self._forward_sync = control_trigger(name="f_sync")
      self._forward_sync._set_control_flow_context(self._forward_context)
      self._forward_index.op._add_control_input(self._forward_sync)
    return self._forward_sync 

def __call__(self,
               inputs,
               initial_state=None,
               dtype=None,
               sequence_length=None,
               scope=None):
    is_list = isinstance(inputs, list)
    if self._use_dynamic_rnn:
      if is_list:
        inputs = array_ops.stack(inputs)
      outputs, state = rnn.dynamic_rnn(
          self._cell,
          inputs,
          sequence_length=sequence_length,
          initial_state=initial_state,
          dtype=dtype,
          time_major=True,
          scope=scope)
      if is_list:
        # Convert outputs back to list
        outputs = array_ops.unstack(outputs)
    else:  # non-dynamic rnn
      if not is_list:
        inputs = array_ops.unstack(inputs)
      outputs, state = contrib_rnn.static_rnn(self._cell,
                                              inputs,
                                              initial_state=initial_state,
                                              dtype=dtype,
                                              sequence_length=sequence_length,
                                              scope=scope)
      if not is_list:
        # Convert outputs back to tensor
        outputs = array_ops.stack(outputs)

    return outputs, state 

def pack(labeled_tensors, new_axis, axis_position=0, name=None):
  """Pack tensors along a new axis.

  See tf.pack.

  Args:
    labeled_tensors: The input tensors, which must have identical axes.
    new_axis: The name of the new axis, or a tuple containing the name
      and coordinate labels.
    axis_position: Optional integer position at which to insert the new axis.
    name: Optional op name.

  Returns:
    The packed tensors as a single LabeledTensor, with `new_axis` in the given
    `axis_position`.

  Raises:
    ValueError: If fewer than one input tensors is provided, or if the tensors
      don't have identical axes.
  """
  with ops.name_scope(name, 'lt_pack', labeled_tensors) as scope:
    labeled_tensors = [
        core.convert_to_labeled_tensor(lt) for lt in labeled_tensors
    ]

    if len(labeled_tensors) < 1:
      raise ValueError('pack expects at least 1 tensors, but received %s' %
                       labeled_tensors)

    axes_0 = labeled_tensors[0].axes
    for t in labeled_tensors:
      if t.axes != axes_0:
        raise ValueError('Non-identical axes. Expected %s but got %s' %
                         (axes_0, t.axes))

    pack_op = array_ops.stack(
        [t.tensor for t in labeled_tensors], axis=axis_position, name=scope)
    axes = list(axes_0.values())
    axes.insert(axis_position, new_axis)
    return core.LabeledTensor(pack_op, axes) 

def _define_expectation_operation(self, shard_id):
    # Shape broadcasting.
    probs = array_ops.expand_dims(self._probs[shard_id], 0)
    # Membership weights are computed as:
    # w_{ik} = \frac{\alpha_k f(\mathbf{y_i}|\mathbf{\theta}_k)}
    #               {\sum_{m=1}^{K}\alpha_mf(\mathbf{y_i}|\mathbf{\theta}_m)}
    # where "i" is the i-th example, "k" is the k-th mixture, theta are
    # the model parameters and y_i the observations.
    # These are defined for each shard.
    self._w[shard_id] = array_ops.reshape(
        math_ops.exp(probs - self._prior_probs[shard_id]),
        array_ops.stack([self._num_examples, self._num_classes])) 

def report_uninitialized_variables(var_list=None,
                                   name="report_uninitialized_variables"):
  """Adds ops to list the names of uninitialized variables.

  When run, it returns a 1-D tensor containing the names of uninitialized
  variables if there are any, or an empty array if there are none.

  Args:
    var_list: List of `Variable` objects to check. Defaults to the
      value of `global_variables() + local_variables()`
    name: Optional name of the `Operation`.

  Returns:
    A 1-D tensor containing names of the uninitialized variables, or an empty
    1-D tensor if there are no variables or no uninitialized variables.
  """
  if var_list is None:
    var_list = global_variables() + local_variables()
    # Backwards compatibility for old-style variables. TODO(touts): remove.
    if not var_list:
      var_list = []
      for op in ops.get_default_graph().get_operations():
        if op.type in ["Variable", "VariableV2", "AutoReloadVariable"]:
          var_list.append(op.outputs[0])
  with ops.name_scope(name):
    if not var_list:
      # Return an empty tensor so we only need to check for returned tensor
      # size being 0 as an indication of model ready.
      return array_ops.constant([], dtype=dtypes.string)
    else:
      # Get a 1-D boolean tensor listing whether each variable is initialized.
      variables_mask = math_ops.logical_not(
          array_ops.stack(
              [state_ops.is_variable_initialized(v) for v in var_list]))
      # Get a 1-D string tensor containing all the variable names.
      variable_names_tensor = array_ops.constant([s.op.name for s in var_list])
      # Return a 1-D tensor containing all the names of uninitialized variables.
      return array_ops.boolean_mask(variable_names_tensor, variables_mask)

# pylint: disable=protected-access 

def _log_cdf(self, x):
    with ops.control_dependencies(self._assertions):
      x = ops.convert_to_tensor(x, name="x")
      distribution_log_cdfs = [d.log_cdf(x) for d in self.components]
      cat_log_probs = self._cat_probs(log_probs=True)
      final_log_cdfs = [
          cat_lp + d_lcdf
          for (cat_lp, d_lcdf) in zip(cat_log_probs, distribution_log_cdfs)
      ]
      concatted_log_cdfs = array_ops.stack(final_log_cdfs, axis=0)
      mixture_log_cdf = math_ops.reduce_logsumexp(concatted_log_cdfs, [0])
      return mixture_log_cdf 

def to_sparse_tensor(self, input_tensor):
    """Creates a SparseTensor from the bucketized Tensor."""
    dimension = self.source_column.dimension
    batch_size = array_ops.shape(input_tensor, name="shape")[0]

    if dimension > 1:
      i1 = array_ops.reshape(
          array_ops.tile(
              array_ops.expand_dims(
                  math_ops.range(0, batch_size), 1, name="expand_dims"),
              [1, dimension],
              name="tile"), [-1],
          name="reshape")
      i2 = array_ops.tile(
          math_ops.range(0, dimension), [batch_size], name="tile")
      # Flatten the bucket indices and unique them across dimensions
      # E.g. 2nd dimension indices will range from k to 2*k-1 with k buckets
      bucket_indices = array_ops.reshape(
          input_tensor, [-1], name="reshape") + self.length * i2
    else:
      # Simpler indices when dimension=1
      i1 = math_ops.range(0, batch_size)
      i2 = array_ops.zeros([batch_size], dtype=dtypes.int32, name="zeros")
      bucket_indices = array_ops.reshape(input_tensor, [-1], name="reshape")

    indices = math_ops.to_int64(array_ops.transpose(array_ops.stack((i1, i2))))
    shape = math_ops.to_int64(array_ops.stack([batch_size, dimension]))
    sparse_id_values = sparse_tensor_py.SparseTensor(
        indices, bucket_indices, shape)

    return sparse_id_values 

def seq2seq_inputs(x, y, input_length, output_length, sentinel=None, name=None):
  """Processes inputs for Sequence to Sequence models.

  Args:
    x: Input Tensor [batch_size, input_length, embed_dim].
    y: Output Tensor [batch_size, output_length, embed_dim].
    input_length: length of input x.
    output_length: length of output y.
    sentinel: optional first input to decoder and final output expected.
      If sentinel is not provided, zeros are used. Due to fact that y is not
      available in sampling time, shape of sentinel will be inferred from x.
    name: Operation name.

  Returns:
    Encoder input from x, and decoder inputs and outputs from y.
  """
  with ops.name_scope(name, "seq2seq_inputs", [x, y]):
    in_x = array_ops.unstack(x, axis=1)
    y = array_ops.unstack(y, axis=1)
    if not sentinel:
      # Set to zeros of shape of y[0], using x for batch size.
      sentinel_shape = array_ops.stack(
          [array_ops.shape(x)[0], y[0].get_shape()[1]])
      sentinel = array_ops.zeros(sentinel_shape)
      sentinel.set_shape(y[0].get_shape())
    in_y = [sentinel] + y
    out_y = y + [sentinel]
    return in_x, in_y, out_y 

def sequence_classifier(decoding, labels, sampling_decoding=None, name=None):
  """Returns predictions and loss for sequence of predictions.

  Args:
    decoding: List of Tensors with predictions.
    labels: List of Tensors with labels.
    sampling_decoding: Optional, List of Tensor with predictions to be used
      in sampling. E.g. they shouldn't have dependncy on outputs.
      If not provided, decoding is used.
    name: Operation name.

  Returns:
    Predictions and losses tensors.
  """
  with ops.name_scope(name, "sequence_classifier", [decoding, labels]):
    predictions, xent_list = [], []
    for i, pred in enumerate(decoding):
      xent_list.append(nn.softmax_cross_entropy_with_logits(
          labels=labels[i], logits=pred,
          name="sequence_loss/xent_raw{0}".format(i)))
      if sampling_decoding:
        predictions.append(nn.softmax(sampling_decoding[i]))
      else:
        predictions.append(nn.softmax(pred))
    xent = math_ops.add_n(xent_list, name="sequence_loss/xent")
    loss = math_ops.reduce_sum(xent, name="sequence_loss")
    return array_ops.stack(predictions, axis=1), loss 

def ndlstm_base_unrolled(inputs, noutput, scope=None, reverse=False):
  """Run an LSTM, either forward or backward.

  This is a 1D LSTM implementation using unrolling and the TensorFlow
  LSTM op.

  Args:
    inputs: input sequence (length, batch_size, ninput)
    noutput: depth of output
    scope: optional scope name
    reverse: run LSTM in reverse

  Returns:
    Output sequence (length, batch_size, noutput)

  """
  with variable_scope.variable_scope(scope, "SeqLstmUnrolled", [inputs]):
    length, batch_size, _ = _shape(inputs)
    lstm_cell = rnn_cell.BasicLSTMCell(noutput, state_is_tuple=False)
    state = array_ops.zeros([batch_size, lstm_cell.state_size])
    output_u = []
    inputs_u = array_ops.unstack(inputs)
    if reverse:
      inputs_u = list(reversed(inputs_u))
    for i in xrange(length):
      if i > 0:
        variable_scope.get_variable_scope().reuse_variables()
      output, state = lstm_cell(inputs_u[i], state)
      output_u += [output]
    if reverse:
      output_u = list(reversed(output_u))
    outputs = array_ops.stack(output_u)
    return outputs 

def feature_importances(self):
    tree_counts = [self.trees[i].feature_usage_counts()
                   for i in range(self.params.num_trees)]
    total_counts = math_ops.reduce_sum(array_ops.stack(tree_counts, 0), 0)
    return total_counts / math_ops.reduce_sum(total_counts) 

def average_size(self):
    """Constructs a TF graph for evaluating the average size of a forest.

    Returns:
      The average number of nodes over the trees.
    """
    sizes = []
    for i in range(self.params.num_trees):
      with ops.device(self.variables.device_dummies[i].device):
        sizes.append(self.trees[i].size())
    return math_ops.reduce_mean(math_ops.to_float(array_ops.stack(sizes)))

  # pylint: disable=unused-argument 

def inference_graph(self, input_data, **inference_args):
    """Constructs a TF graph for evaluating a random forest.

    Args:
      input_data: A tensor or dict of string->Tensor for input data.
      **inference_args: Keyword arguments to pass through to each tree.

    Returns:
      The last op in the random forest inference graph.

    Raises:
      NotImplementedError: If trying to use feature bagging with sparse
        features.
    """
    processed_dense_features, processed_sparse_features, data_spec = (
        data_ops.ParseDataTensorOrDict(input_data))

    probabilities = []
    for i in range(self.params.num_trees):
      with ops.device(self.variables.device_dummies[i].device):
        tree_data = processed_dense_features
        if self.params.bagged_features:
          if processed_sparse_features is not None:
            raise NotImplementedError(
                'Feature bagging not supported with sparse features.')
          tree_data = self._bag_features(i, input_data)
        probabilities.append(self.trees[i].inference_graph(
            tree_data,
            data_spec,
            sparse_features=processed_sparse_features,
            **inference_args))
    with ops.device(self.variables.device_dummies[0].device):
      all_predict = array_ops.stack(probabilities)
      return math_ops.div(
          math_ops.reduce_sum(all_predict, 0), self.params.num_trees,
          name='probabilities')