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

def reduce_to_sequence(images, num_filters_out, scope=None):
  """Reduce an image to a sequence by scanning an LSTM vertically.

  Args:
    images: (num_images, height, width, depth) tensor
    num_filters_out: output layer depth
    scope: optional scope name

  Returns:
    A (width, num_images, num_filters_out) sequence.
  """
  with variable_scope.variable_scope(scope, "ReduceToSequence", [images]):
    batch_size, height, width, depth = _shape(images)
    transposed = array_ops.transpose(images, [1, 0, 2, 3])
    reshaped = array_ops.reshape(transposed,
                                 [height, batch_size * width, depth])
    reduced = lstm1d.sequence_to_final(reshaped, num_filters_out)
    output = array_ops.reshape(reduced, [batch_size, width, num_filters_out])
    return output 

def softmax(logits, scope=None):
  """Performs softmax on Nth dimension of N-dimensional logit tensor.

  For two-dimensional logits this reduces to tf.nn.softmax. The N-th dimension
  needs to have a specified number of elements (number of classes).

  Args:
    logits: N-dimensional `Tensor` with logits, where N > 1.
    scope: Optional scope for variable_scope.

  Returns:
    A `Tensor` with same shape and type as logits.
  """
  # TODO(jrru): Add axis argument which defaults to last dimension.
  with variable_scope.variable_scope(scope, 'softmax', [logits]):
    num_logits = utils.last_dimension(logits.get_shape(), min_rank=2)
    logits_2d = array_ops.reshape(logits, [-1, num_logits])
    predictions = nn.softmax(logits_2d)
    predictions = array_ops.reshape(predictions, array_ops.shape(logits))
    if not context.executing_eagerly():
      predictions.set_shape(logits.get_shape())
    return predictions 

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 _flatten_outer_dims(logits):
  """Flattens logits' outer dimensions and keep its last dimension."""
  rank = array_ops.rank(logits)
  last_dim_size = array_ops.slice(
      array_ops.shape(logits), [math_ops.subtract(rank, 1)], [1])
  output = array_ops.reshape(logits, array_ops.concat([[-1], last_dim_size], 0))

  # Set output shape if known.
  shape = logits.get_shape()
  if shape is not None and shape.dims is not None:
    shape = shape.as_list()
    product = 1
    product_valid = True
    for d in shape[:-1]:
      if d is None:
        product_valid = False
        break
      else:
        product *= d
    if product_valid:
      output_shape = [product, shape[-1]]
      output.set_shape(output_shape)

  return output 

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 _GatherGrad(op, grad):
  """Gradient for gather op."""
  # Build appropriately shaped IndexedSlices
  # Walk graph back until the original handle is found.
  # TODO(apassos): more robust way of getting the shape.
  handle = op.inputs[0]
  while handle.op.type != "VarHandleOp":
    handle = handle.op.inputs[0]
  params_shape = ops.convert_to_tensor(
      tensor_shape.TensorShape(handle.op.get_attr("shape")))
  indices = op.inputs[1]
  size = array_ops.expand_dims(array_ops.size(indices), 0)
  values_shape = array_ops.concat([size, params_shape[1:]], 0)
  values = array_ops.reshape(grad, values_shape)
  indices = array_ops.reshape(indices, size)
  return [ops.IndexedSlices(values, indices, params_shape), None] 

def _sample_n(self, n, seed=None):
    n_draws = math_ops.cast(self.total_count, dtype=dtypes.int32)
    if self.total_count.get_shape().ndims is not None:
      if self.total_count.get_shape().ndims != 0:
        raise NotImplementedError(
            "Sample only supported for scalar number of draws.")
    elif self.validate_args:
      is_scalar = check_ops.assert_rank(
          n_draws, 0,
          message="Sample only supported for scalar number of draws.")
      n_draws = control_flow_ops.with_dependencies([is_scalar], n_draws)
    k = self.event_shape_tensor()[0]
    # Flatten batch dims so logits has shape [B, k],
    # where B = reduce_prod(self.batch_shape_tensor()).
    draws = random_ops.multinomial(
        logits=array_ops.reshape(self.logits, [-1, k]),
        num_samples=n * n_draws,
        seed=seed)
    draws = array_ops.reshape(draws, shape=[-1, n, n_draws])
    x = math_ops.reduce_sum(array_ops.one_hot(draws, depth=k),
                            axis=-2)  # shape: [B, n, k]
    x = array_ops.transpose(x, perm=[1, 0, 2])
    final_shape = array_ops.concat([[n], self.batch_shape_tensor(), [k]], 0)
    return array_ops.reshape(x, final_shape) 

def _sample_n(self, n, seed=None):
    n_draws = math_ops.cast(self.total_count, dtype=dtypes.int32)
    k = self.event_shape_tensor()[0]
    unnormalized_logits = array_ops.reshape(
        math_ops.log(random_ops.random_gamma(
            shape=[n],
            alpha=self.concentration,
            dtype=self.dtype,
            seed=seed)),
        shape=[-1, k])
    draws = random_ops.multinomial(
        logits=unnormalized_logits,
        num_samples=n_draws,
        seed=distribution_util.gen_new_seed(seed, salt="dirichlet_multinomial"))
    x = math_ops.reduce_sum(array_ops.one_hot(draws, depth=k), -2)
    final_shape = array_ops.concat([[n], self.batch_shape_tensor(), [k]], 0)
    return array_ops.reshape(x, final_shape) 

def _MinOrMaxGrad(op, grad):
  """Gradient for Min or Max. Amazingly it's precisely the same code."""
  input_shape = array_ops.shape(op.inputs[0])
  output_shape_kept_dims = math_ops.reduced_shape(input_shape, op.inputs[1])
  y = op.outputs[0]
  y = array_ops.reshape(y, output_shape_kept_dims)
  grad = array_ops.reshape(grad, output_shape_kept_dims)

  # Compute the number of selected (maximum or minimum) elements in each
  # reduction dimension. If there are multiple minimum or maximum elements
  # then the gradient will be divided between them.
  indicators = math_ops.cast(math_ops.equal(y, op.inputs[0]), grad.dtype)
  num_selected = array_ops.reshape(
      math_ops.reduce_sum(indicators, op.inputs[1]), output_shape_kept_dims)

  return [math_ops.div(indicators, num_selected) * grad, None] 

def _BetaincGrad(op, grad):
  """Returns gradient of betainc(a, b, x) with respect to x."""
  # TODO(ebrevdo): Perhaps add the derivative w.r.t. a, b
  a, b, x = op.inputs

  # two cases: x is a scalar and a/b are same-shaped tensors, or vice
  # versa; so its sufficient to check against shape(a).
  sa = array_ops.shape(a)
  sx = array_ops.shape(x)
  # pylint: disable=protected-access
  _, rx = gen_array_ops._broadcast_gradient_args(sa, sx)
  # pylint: enable=protected-access

  # Perform operations in log space before summing, because terms
  # can grow large.
  log_beta = (gen_math_ops.lgamma(a) + gen_math_ops.lgamma(b)
              - gen_math_ops.lgamma(a + b))
  partial_x = math_ops.exp(
      (b - 1) * math_ops.log(1 - x) + (a - 1) * math_ops.log(x) - log_beta)

  # TODO(b/36815900): Mark None return values as NotImplemented
  return (None,  # da
          None,  # db
          array_ops.reshape(math_ops.reduce_sum(partial_x * grad, rx), sx)) 

def _ZetaGrad(op, grad):
  """Returns gradient of zeta(x, q) with respect to x and q."""
  # TODO(tillahoffmann): Add derivative with respect to x
  x = op.inputs[0]
  q = op.inputs[1]
  # Broadcast gradients
  sx = array_ops.shape(x)
  sq = array_ops.shape(q)
  unused_rx, rq = gen_array_ops._broadcast_gradient_args(sx, sq)
  # Evaluate gradient
  with ops.control_dependencies([grad.op]):
    x = math_ops.conj(x)
    q = math_ops.conj(q)
    partial_q = -x * math_ops.zeta(x + 1, q)
    # TODO(b/36815900): Mark None return values as NotImplemented
    return (None,
            array_ops.reshape(math_ops.reduce_sum(partial_q * grad, rq), sq)) 

def _MaximumMinimumGrad(op, grad, selector_op):
  """Factor out the code for the gradient of Maximum or Minimum."""
  x = op.inputs[0]
  y = op.inputs[1]
  gdtype = grad.dtype
  sx = array_ops.shape(x)
  sy = array_ops.shape(y)
  gradshape = array_ops.shape(grad)
  zeros = array_ops.zeros(gradshape, gdtype)
  xmask = selector_op(x, y)
  rx, ry = gen_array_ops._broadcast_gradient_args(sx, sy)
  xgrad = array_ops.where(xmask, grad, zeros)
  ygrad = array_ops.where(math_ops.logical_not(xmask), grad, zeros)
  gx = array_ops.reshape(math_ops.reduce_sum(xgrad, rx), sx)
  gy = array_ops.reshape(math_ops.reduce_sum(ygrad, ry), sy)
  return (gx, gy) 

def _SquaredDifferenceGrad(op, grad):
  """Returns the gradient for (x-y)^2."""
  x = op.inputs[0]
  y = op.inputs[1]
  sx = array_ops.shape(x)
  sy = array_ops.shape(y)
  # pylint: disable=protected-access
  rx, ry = gen_array_ops._broadcast_gradient_args(sx, sy)
  # pylint: enable=protected-access
  # .op works with Tensors or IndexedSlices
  with ops.control_dependencies([grad.op]):
    # The parens ensure that if grad is IndexedSlices, it'll get multiplied by
    # Tensor (not a number like 2.0) which causes it to convert to Tensor.
    x_grad = math_ops.scalar_mul(2.0, grad) * (x - y)
  return (array_ops.reshape(math_ops.reduce_sum(x_grad, rx), sx),
          -array_ops.reshape(math_ops.reduce_sum(x_grad, ry), sy))


# Logical operations have no gradients. 

def _create_dense_column_weighted_sum(
    column, builder, units, weight_collections, trainable):
  """Create a weighted sum of a dense column for linear_model."""
  tensor = column._get_dense_tensor(  # pylint: disable=protected-access
      builder,
      weight_collections=weight_collections,
      trainable=trainable)
  num_elements = column._variable_shape.num_elements()  # pylint: disable=protected-access
  batch_size = array_ops.shape(tensor)[0]
  tensor = array_ops.reshape(tensor, shape=(batch_size, num_elements))
  weight = variable_scope.get_variable(
      name='weights',
      shape=[num_elements, units],
      initializer=init_ops.zeros_initializer(),
      trainable=trainable,
      collections=weight_collections)
  return math_ops.matmul(tensor, weight, name='weighted_sum') 

def call(self, inputs, state):
    """Run the cell on embedded inputs."""
    with ops.device("/cpu:0"):
      if self._initializer:
        initializer = self._initializer
      elif vs.get_variable_scope().initializer:
        initializer = vs.get_variable_scope().initializer
      else:
        # Default initializer for embeddings should have variance=1.
        sqrt3 = math.sqrt(3)  # Uniform(-sqrt(3), sqrt(3)) has variance=1.
        initializer = init_ops.random_uniform_initializer(-sqrt3, sqrt3)

      if isinstance(state, tuple):
        data_type = state[0].dtype
      else:
        data_type = state.dtype

      embedding = vs.get_variable(
          "embedding", [self._embedding_classes, self._embedding_size],
          initializer=initializer,
          dtype=data_type)
      embedded = embedding_ops.embedding_lookup(embedding,
                                                array_ops.reshape(inputs, [-1]))

      return self._cell(embedded, state) 

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 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 _RealDivGrad(op, grad):
  """RealDiv op gradient."""
  x = op.inputs[0]
  y = op.inputs[1]
  sx = array_ops.shape(x)
  sy = array_ops.shape(y)
  # pylint: disable=protected-access
  rx, ry = gen_array_ops._broadcast_gradient_args(sx, sy)
  # pylint: enable=protected-access
  x = math_ops.conj(x)
  y = math_ops.conj(y)
  return (array_ops.reshape(
      math_ops.reduce_sum(math_ops.realdiv(grad, y), rx),
      sx), array_ops.reshape(
          math_ops.reduce_sum(grad * math_ops.realdiv(math_ops.realdiv(-x, y), y),
                              ry), sy)) 

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 _init_clusters_random(self):
    """Does random initialization of clusters.

    Returns:
      Tensor of randomly initialized clusters.
    """
    num_data = math_ops.add_n([array_ops.shape(inp)[0] for inp in self._inputs])
    # Note that for mini-batch k-means, we should ensure that the batch size of
    # data used during initialization is sufficiently large to avoid duplicated
    # clusters.
    with ops.control_dependencies(
        [check_ops.assert_less_equal(self._num_clusters, num_data)]):
      indices = random_ops.random_uniform(
          array_ops.reshape(self._num_clusters, [-1]),
          minval=0,
          maxval=math_ops.cast(num_data, dtypes.int64),
          seed=self._random_seed,
          dtype=dtypes.int64)
      clusters_init = embedding_lookup(
          self._inputs, indices, partition_strategy='div')
      return clusters_init 

def sequence_to_images(tensor, num_image_batches):
  """Convert a batch of sequences into a batch of images.

  Args:
    tensor: (num_steps, num_batches, depth) sequence tensor
    num_image_batches: the number of image batches

  Returns:
    (num_images, height, width, depth) tensor
  """

  width, num_batches, depth = _shape(tensor)
  height = num_batches // num_image_batches
  reshaped = array_ops.reshape(tensor,
                               [width, num_image_batches, height, depth])
  return array_ops.transpose(reshaped, [1, 2, 0, 3]) 

def _dense_inner_flatten(inputs, new_rank):
  """Helper function for `inner_flatten`."""
  rank_assertion = check_ops.assert_rank_at_least(
      inputs, new_rank, message='inputs has rank less than new_rank')
  with ops.control_dependencies([rank_assertion]):
    outer_dimensions = array_ops.strided_slice(
        array_ops.shape(inputs), [0], [new_rank - 1])
    new_shape = array_ops.concat((outer_dimensions, [-1]), 0)
    reshaped = array_ops.reshape(inputs, new_shape)

  # if `new_rank` is an integer, try to calculate new shape.
  if isinstance(new_rank, six.integer_types):
    static_shape = inputs.get_shape()
    if static_shape is not None and static_shape.dims is not None:
      static_shape = static_shape.as_list()
      static_outer_dims = static_shape[:new_rank - 1]
      static_inner_dims = static_shape[new_rank - 1:]
      flattened_dimension = 1
      for inner_dim in static_inner_dims:
        if inner_dim is None:
          flattened_dimension = None
          break
        flattened_dimension *= inner_dim
      reshaped.set_shape(static_outer_dims + [flattened_dimension])
  return reshaped 

def restore(self, restored_tensors, restored_shapes):
      restored_tensor = restored_tensors[0]
      if restored_shapes is not None:
        restored_tensor = array_ops.reshape(restored_tensor, restored_shapes[0])
      return resource_variable_ops.assign_variable_op(
          self.handle_op, restored_tensor) 

def _smart_select(pred, fn_then, fn_else):
  """Selects fn_then() or fn_else() based on the value of pred.

  The purpose of this function is the same as `utils.smart_cond`. However, at
  the moment there is a bug (b/36297356) that seems to kick in only when
  `smart_cond` delegates to `tf.cond`, which sometimes results in the training
  hanging when using parameter servers. This function will output the result
  of `fn_then` or `fn_else` if `pred` is known at graph construction time.
  Otherwise, it will use `tf.where` which will result in some redundant work
  (both branches will be computed but only one selected). However, the tensors
  involved will usually be small (means and variances in batchnorm), so the
  cost will be small and will not be incurred at all if `pred` is a constant.

  Args:
    pred: A boolean scalar `Tensor`.
    fn_then: A callable to use when pred==True.
    fn_else: A callable to use when pred==False.

  Returns:
    A `Tensor` whose value is fn_then() or fn_else() based on the value of pred.
  """
  pred_value = utils.constant_value(pred)
  if pred_value:
    return fn_then()
  elif pred_value is False:
    return fn_else()
  t_then = array_ops.expand_dims(fn_then(), 0)
  t_else = array_ops.expand_dims(fn_else(), 0)
  pred = array_ops.reshape(pred, [1])
  result = array_ops.where(pred, t_then, t_else)
  return array_ops.squeeze(result, [0]) 

def sequence_to_images(inputs,
                       height,
                       output_data_format='channels_last',
                       outputs_collections=None,
                       scope=None):
  """Convert a batch of sequences into a batch of images.

  Args:
    inputs: (num_steps, num_batches, depth) sequence tensor
    height: the height of the images
    output_data_format: Format of output tensor. Currently supports
      `'channels_first'` and `'channels_last'`.
    outputs_collections: The collections to which the outputs are added.
    scope: Optional scope for name_scope.

  Returns:
    A tensor representing the output of the operation.
  """
  with ops.name_scope(scope, 'SequenceToImages', [inputs]) as sc:
    inputs = ops.convert_to_tensor(inputs)
    width, num_batches, depth = inputs.get_shape().as_list()
    if num_batches is None:
      num_batches = -1
    else:
      num_batches //= height
    reshaped = array_ops.reshape(inputs, [width, num_batches, height, depth])
    if output_data_format == 'channels_first':
      outputs = array_ops.transpose(reshaped, [1, 3, 2, 0])
    else:
      outputs = array_ops.transpose(reshaped, [1, 2, 0, 3])
    return utils.collect_named_outputs(outputs_collections, sc, outputs) 

def _decode(self, image_buffer, image_format):
    """Decodes the image buffer.

    Args:
      image_buffer: The tensor representing the encoded image tensor.
      image_format: The image format for the image in `image_buffer`. If image
        format is `raw`, all images are expected to be in this format, otherwise
        this op can decode a mix of `jpg` and `png` formats.

    Returns:
      A tensor that represents decoded image of self._shape, or
      (?, ?, self._channels) if self._shape is not specified.
    """
    def decode_image():
      """Decodes a png or jpg based on the headers."""
      return image_ops.decode_image(image_buffer, self._channels)

    def decode_raw():
      """Decodes a raw image."""
      return parsing_ops.decode_raw(image_buffer, out_type=self._dtype)

    pred_fn_pairs = {
        math_ops.logical_or(
            math_ops.equal(image_format, 'raw'),
            math_ops.equal(image_format, 'RAW')): decode_raw,
    }
    image = control_flow_ops.case(
        pred_fn_pairs, default=decode_image, exclusive=True)

    image.set_shape([None, None, self._channels])
    if self._shape is not None:
      image = array_ops.reshape(image, self._shape)

    return image 

def decode(self, serialized_example, items=None):
    """Decodes the given serialized TF-example.

    Args:
      serialized_example: a serialized TF-example tensor.
      items: the list of items to decode. These must be a subset of the item
        keys in self._items_to_handlers. If `items` is left as None, then all
        of the items in self._items_to_handlers are decoded.

    Returns:
      the decoded items, a list of tensor.
    """
    example = parsing_ops.parse_single_example(serialized_example,
                                               self._keys_to_features)

    # Reshape non-sparse elements just once:
    for k in self._keys_to_features:
      v = self._keys_to_features[k]
      if isinstance(v, parsing_ops.FixedLenFeature):
        example[k] = array_ops.reshape(example[k], v.shape)

    if not items:
      items = self._items_to_handlers.keys()

    outputs = []
    for item in items:
      handler = self._items_to_handlers[item]
      keys_to_tensors = {key: example[key] for key in handler.keys}
      outputs.append(handler.tensors_to_item(keys_to_tensors))
    return outputs 

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 images_to_sequence(inputs,
                       data_format=DATA_FORMAT_NHWC,
                       outputs_collections=None,
                       scope=None):
  """Convert a batch of images into a batch of sequences.

  Args:
    inputs: a (num_images, height, width, depth) tensor
    data_format: A string. `NHWC` (default) and `NCHW` are supported.
    outputs_collections: The collections to which the outputs are added.
    scope: Optional scope for name_scope.

  Raises:
     ValueError: If `data_format` is not either NCHW or NHWC.

  Returns:
    (width, num_images*height, depth) sequence tensor
  """
  if data_format not in (DATA_FORMAT_NCHW, DATA_FORMAT_NHWC):
    raise ValueError('data_format has to be either NCHW or NHWC.')
  with ops.name_scope(scope, 'ImagesToSequence', [inputs]) as sc:
    inputs = ops.convert_to_tensor(inputs)
    df = ('channels_first'
          if data_format and data_format.startswith('NC') else 'channels_last')
    if df == 'channels_first':
      inputs = array_ops.transpose(inputs, [0, 2, 3, 1])
    _, _, width, depth = inputs.get_shape().as_list()
    s = array_ops.shape(inputs)
    batch_size, height = s[0], s[1]
    transposed = array_ops.transpose(inputs, [2, 0, 1, 3])
    outputs = array_ops.reshape(transposed, [width, batch_size * height, depth])
    return utils.collect_named_outputs(outputs_collections, sc, outputs)