Python tensorflow.expand_dims() Examples

def pass_through_embedding_matrix(act_block, embedding_matrix, step_idx):
  """Passes the activations through the embedding_matrix.

  Takes care to handle out of bounds lookups.

  Args:
    act_block: matrix of activations.
    embedding_matrix: matrix of weights.
    step_idx: vector containing step indices, with -1 indicating out of bounds.

  Returns:
    the embedded activations.
  """
  # Indicator vector for out of bounds lookups.
  step_idx_mask = tf.expand_dims(tf.equal(step_idx, -1), -1)

  # Pad the last column of the activation vectors with the indicator.
  act_block = tf.concat([act_block, tf.to_float(step_idx_mask)], 1)
  return tf.matmul(act_block, embedding_matrix) 

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

  intersection-over-area (IOA) between two boxes box1 and box2 is defined as
  their intersection area over box2's area. Note that ioa is not symmetric,
  that is, ioa(box1, box2) != ioa(box2, box1).

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

  Returns:
    a tensor with shape [N, M] representing pairwise ioa scores.
  """
  with tf.name_scope(scope, 'IOA'):
    intersections = intersection(boxlist1, boxlist2)
    areas = tf.expand_dims(area(boxlist2), 0)
    return tf.truediv(intersections, areas) 

def conv1d_transpose(
    inputs,
    filters,
    kernel_width,
    stride=4,
    padding='same',
    upsample='zeros'):
    if upsample == 'zeros':
        return tf.layers.conv2d_transpose(
            tf.expand_dims(inputs, axis=1),
            filters,
            (1, kernel_width),
            strides=(1, stride),
            padding='same'
        )[:, 0]
    else:
        raise NotImplementedError 

def call(self, x):
        if (self.size == None) or (self.mode == 'sum'):
            self.size = int(x.shape[-1])

        position_j = 1. / \
            K.pow(10000., 2 * K.arange(self.size / 2, dtype='float32') / self.size)
        position_j = K.expand_dims(position_j, 0)

        position_i = tf.cumsum(K.ones_like(x[:, :, 0]), 1) - 1
        position_i = K.expand_dims(position_i, 2)
        position_ij = K.dot(position_i, position_j)
        outputs = K.concatenate(
            [K.cos(position_ij), K.sin(position_ij)], 2)

        if self.mode == 'sum':
            if self.scale:
                outputs = outputs * outputs ** 0.5
            return x + outputs
        elif self.mode == 'concat':
            return K.concatenate([outputs, x], 2) 

def memory_run(step, nmaps, mem_size, batch_size, vocab_size,
               global_step, do_training, update_mem, decay_factor, num_gpus,
               target_emb_weights, output_w, gpu_targets_tn, it):
  """Run memory."""
  q = step[:, 0, it, :]
  mlabels = gpu_targets_tn[:, it, 0]
  res, mask, mem_loss = memory_call(
      q, mlabels, nmaps, mem_size, vocab_size, num_gpus, update_mem)
  res = tf.gather(target_emb_weights, res) * tf.expand_dims(mask[:, 0], 1)

  # Mix gold and original in the first steps, 20% later.
  gold = tf.nn.dropout(tf.gather(target_emb_weights, mlabels), 0.7)
  use_gold = 1.0 - tf.cast(global_step, tf.float32) / (1000. * decay_factor)
  use_gold = tf.maximum(use_gold, 0.2) * do_training
  mem = tf.cond(tf.less(tf.random_uniform([]), use_gold),
                lambda: use_gold * gold + (1.0 - use_gold) * res,
                lambda: res)
  mem = tf.reshape(mem, [-1, 1, 1, nmaps])
  return mem, mem_loss, update_mem 

def preprocess_batch(images_batch, preproc_func=None):
    """
    Creates a preprocessing graph for a batch given a function that processes
    a single image.

    :param images_batch: A tensor for an image batch.
    :param preproc_func: (optional function) A function that takes in a
        tensor and returns a preprocessed input.
    """
    if preproc_func is None:
        return images_batch

    with tf.variable_scope('preprocess'):
        images_list = tf.split(images_batch, int(images_batch.shape[0]))
        result_list = []
        for img in images_list:
            reshaped_img = tf.reshape(img, img.shape[1:])
            processed_img = preproc_func(reshaped_img)
            result_list.append(tf.expand_dims(processed_img, axis=0))
        result_images = tf.concat(result_list, axis=0)
    return result_images 

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 main():
    rgb = False
    if rgb:
        kernels_list = [kernels.BLUR_FILTER_RGB,
                        kernels.SHARPEN_FILTER_RGB,
                        kernels.EDGE_FILTER_RGB,
                        kernels.TOP_SOBEL_RGB,
                        kernels.EMBOSS_FILTER_RGB]
    else:
        kernels_list = [kernels.BLUR_FILTER,
                        kernels.SHARPEN_FILTER,
                        kernels.EDGE_FILTER,
                        kernels.TOP_SOBEL,
                        kernels.EMBOSS_FILTER]

    kernels_list = kernels_list[1:]
    image = read_one_image('data/images/naruto.jpeg')
    if not rgb:
        image = tf.image.rgb_to_grayscale(image)
    image = tf.expand_dims(image, 0) # make it into a batch of 1 element
    images = convolve(image, kernels_list, rgb)
    with tf.Session() as sess:
        images = sess.run(images) # convert images from tensors to float values
    show_images(images, rgb) 

def __init__(self, num_keypoints, scale_factors=None):
    """Constructor for KeypointBoxCoder.

    Args:
      num_keypoints: Number of keypoints to encode/decode.
      scale_factors: List of 4 positive scalars to scale ty, tx, th and tw.
        In addition to scaling ty and tx, the first 2 scalars are used to scale
        the y and x coordinates of the keypoints as well. If set to None, does
        not perform scaling.
    """
    self._num_keypoints = num_keypoints

    if scale_factors:
      assert len(scale_factors) == 4
      for scalar in scale_factors:
        assert scalar > 0
    self._scale_factors = scale_factors
    self._keypoint_scale_factors = None
    if scale_factors is not None:
      self._keypoint_scale_factors = tf.expand_dims(tf.tile(
          [tf.to_float(scale_factors[0]), tf.to_float(scale_factors[1])],
          [num_keypoints]), 1) 

def _aspect_preserving_resize(image, smallest_side):
  """Resize images preserving the original aspect ratio.

  Args:
    image: A 3-D image `Tensor`.
    smallest_side: A python integer or scalar `Tensor` indicating the size of
      the smallest side after resize.

  Returns:
    resized_image: A 3-D tensor containing the resized image.
  """
  smallest_side = tf.convert_to_tensor(smallest_side, dtype=tf.int32)

  shape = tf.shape(image)
  height = shape[0]
  width = shape[1]
  new_height, new_width = _smallest_size_at_least(height, width, smallest_side)
  image = tf.expand_dims(image, 0)
  resized_image = tf.image.resize_bilinear(image, [new_height, new_width],
                                           align_corners=False)
  resized_image = tf.squeeze(resized_image)
  resized_image.set_shape([None, None, 3])
  return resized_image 

def preprocess_for_eval(image, output_height, output_width):
  """Preprocesses the given image for evaluation.

  Args:
    image: A `Tensor` representing an image of arbitrary size.
    output_height: The height of the image after preprocessing.
    output_width: The width of the image after preprocessing.

  Returns:
    A preprocessed image.
  """
  tf.summary.image('image', tf.expand_dims(image, 0))
  # Transform the image to floats.
  image = tf.to_float(image)

  # Resize and crop if needed.
  resized_image = tf.image.resize_image_with_crop_or_pad(image,
                                                         output_width,
                                                         output_height)
  tf.summary.image('resized_image', tf.expand_dims(resized_image, 0))

  # Subtract off the mean and divide by the variance of the pixels.
  return tf.image.per_image_standardization(resized_image) 

def pad_tensor(t, length):
  """Pads the input tensor with 0s along the first dimension up to the length.

  Args:
    t: the input tensor, assuming the rank is at least 1.
    length: a tensor of shape [1]  or an integer, indicating the first dimension
      of the input tensor t after padding, assuming length <= t.shape[0].

  Returns:
    padded_t: the padded tensor, whose first dimension is length. If the length
      is an integer, the first dimension of padded_t is set to length
      statically.
  """
  t_rank = tf.rank(t)
  t_shape = tf.shape(t)
  t_d0 = t_shape[0]
  pad_d0 = tf.expand_dims(length - t_d0, 0)
  pad_shape = tf.cond(
      tf.greater(t_rank, 1), lambda: tf.concat([pad_d0, t_shape[1:]], 0),
      lambda: tf.expand_dims(length - t_d0, 0))
  padded_t = tf.concat([t, tf.zeros(pad_shape, dtype=t.dtype)], 0)
  if not _is_tensor(length):
    padded_t = _set_dim_0(padded_t, length)
  return padded_t 

def expanded_shape(orig_shape, start_dim, num_dims):
  """Inserts multiple ones into a shape vector.

  Inserts an all-1 vector of length num_dims at position start_dim into a shape.
  Can be combined with tf.reshape to generalize tf.expand_dims.

  Args:
    orig_shape: the shape into which the all-1 vector is added (int32 vector)
    start_dim: insertion position (int scalar)
    num_dims: length of the inserted all-1 vector (int scalar)
  Returns:
    An int32 vector of length tf.size(orig_shape) + num_dims.
  """
  with tf.name_scope('ExpandedShape'):
    start_dim = tf.expand_dims(start_dim, 0)  # scalar to rank-1
    before = tf.slice(orig_shape, [0], start_dim)
    add_shape = tf.ones(tf.reshape(num_dims, [1]), dtype=tf.int32)
    after = tf.slice(orig_shape, start_dim, [-1])
    new_shape = tf.concat([before, add_shape, after], 0)
    return new_shape 

def eval_image(image, height, width, scope=None):
  """Prepare one image for evaluation.

  Args:
    image: 3-D float Tensor
    height: integer
    width: integer
    scope: Optional scope for name_scope.
  Returns:
    3-D float Tensor of prepared image.
  """
  with tf.name_scope(values=[image, height, width], name=scope,
                     default_name='eval_image'):
    # Crop the central region of the image with an area containing 87.5% of
    # the original image.
    image = tf.image.central_crop(image, central_fraction=0.875)

    # Resize the image to the original height and width.
    image = tf.expand_dims(image, 0)
    image = tf.image.resize_bilinear(image, [height, width],
                                     align_corners=False)
    image = tf.squeeze(image, [0])
    return image 

def one_hot_encoding(labels, num_classes, scope=None):
  """Transform numeric labels into onehot_labels.

  Args:
    labels: [batch_size] target labels.
    num_classes: total number of classes.
    scope: Optional scope for name_scope.
  Returns:
    one hot encoding of the labels.
  """
  with tf.name_scope(scope, 'OneHotEncoding', [labels]):
    batch_size = labels.get_shape()[0]
    indices = tf.expand_dims(tf.range(0, batch_size), 1)
    labels = tf.cast(tf.expand_dims(labels, 1), indices.dtype)
    concated = tf.concat(axis=1, values=[indices, labels])
    onehot_labels = tf.sparse_to_dense(
        concated, tf.stack([batch_size, num_classes]), 1.0, 0.0)
    onehot_labels.set_shape([batch_size, num_classes])
    return onehot_labels 

def compute_column_softmax(self, column_controller_vector, time_step):
    #compute softmax over all the columns using column controller vector
    column_controller_vector = tf.tile(
        tf.expand_dims(column_controller_vector, 1),
        [1, self.num_cols + self.num_word_cols, 1])  #max_cols * bs * d
    column_controller_vector = nn_utils.apply_dropout(
        column_controller_vector, self.utility.FLAGS.dropout, self.mode)
    self.full_column_hidden_vectors = tf.concat(
        axis=1, values=[self.column_hidden_vectors, self.word_column_hidden_vectors])
    self.full_column_hidden_vectors += self.summary_text_entry_embeddings
    self.full_column_hidden_vectors = nn_utils.apply_dropout(
        self.full_column_hidden_vectors, self.utility.FLAGS.dropout, self.mode)
    column_logits = tf.reduce_sum(
        column_controller_vector * self.full_column_hidden_vectors, 2) + (
            self.params["word_match_feature_column_name"] *
            self.batch_column_exact_match) + self.full_column_mask
    column_softmax = tf.nn.softmax(column_logits)  #batch_size * max_cols
    return column_softmax 

def compute_first_or_last(self, select, first=True):
    #perform first ot last operation on row select with probabilistic row selection
    answer = tf.zeros_like(select)
    running_sum = tf.zeros([self.batch_size, 1], self.data_type)
    for i in range(self.max_elements):
      if (first):
        current = tf.slice(select, [0, i], [self.batch_size, 1])
      else:
        current = tf.slice(select, [0, self.max_elements - 1 - i],
                           [self.batch_size, 1])
      curr_prob = current * (1 - running_sum)
      curr_prob = curr_prob * tf.cast(curr_prob >= 0.0, self.data_type)
      running_sum += curr_prob
      temp_ans = []
      curr_prob = tf.expand_dims(tf.reshape(curr_prob, [self.batch_size]), 0)
      for i_ans in range(self.max_elements):
        if (not (first) and i_ans == self.max_elements - 1 - i):
          temp_ans.append(curr_prob)
        elif (first and i_ans == i):
          temp_ans.append(curr_prob)
        else:
          temp_ans.append(tf.zeros_like(curr_prob))
      temp_ans = tf.transpose(tf.concat(axis=0, values=temp_ans))
      answer += temp_ans
    return answer 

def _batch_decode_refined_boxes(self, refined_box_encodings, proposal_boxes):
    """Decode tensor of refined box encodings.

    Args:
      refined_box_encodings: a 3-D tensor with shape
        [batch_size, max_num_proposals, num_classes, self._box_coder.code_size]
        representing predicted (final) refined box encodings.
      proposal_boxes: [batch_size, self.max_num_proposals, 4] representing
        decoded proposal bounding boxes.

    Returns:
      refined_box_predictions: a [batch_size, max_num_proposals, num_classes, 4]
        float tensor representing (padded) refined bounding box predictions
        (for each image in batch, proposal and class).
    """
    tiled_proposal_boxes = tf.tile(
        tf.expand_dims(proposal_boxes, 2), [1, 1, self.num_classes, 1])
    tiled_proposals_boxlist = box_list.BoxList(
        tf.reshape(tiled_proposal_boxes, [-1, 4]))
    decoded_boxes = self._box_coder.decode(
        tf.reshape(refined_box_encodings, [-1, self._box_coder.code_size]),
        tiled_proposals_boxlist)
    return tf.reshape(decoded_boxes.get(),
                      [-1, self.max_num_proposals, self.num_classes, 4]) 

def _padded_batched_proposals_indicator(self,
                                          num_proposals,
                                          max_num_proposals):
    """Creates indicator matrix of non-pad elements of padded batch proposals.

    Args:
      num_proposals: Tensor of type tf.int32 with shape [batch_size].
      max_num_proposals: Maximum number of proposals per image (integer).

    Returns:
      A Tensor of type tf.bool with shape [batch_size, max_num_proposals].
    """
    batch_size = tf.size(num_proposals)
    tiled_num_proposals = tf.tile(
        tf.expand_dims(num_proposals, 1), [1, max_num_proposals])
    tiled_proposal_index = tf.tile(
        tf.expand_dims(tf.range(max_num_proposals), 0), [batch_size, 1])
    return tf.greater(tiled_num_proposals, tiled_proposal_index) 

def extract_fixed_feature_ids(comp, state, stride):
  """Extracts fixed feature IDs.

  Args:
    comp: Component whose fixed feature IDs we wish to extract.
    state: Live MasterState object for the component.
    stride: Tensor containing current batch * beam size.

  Returns:
    state handle: Updated state handle to be used after this call.
    ids: List of [stride * num_steps, 1] feature IDs per channel.  Missing IDs
         (e.g., due to batch padding) are set to -1.
  """
  num_channels = len(comp.spec.fixed_feature)
  if not num_channels:
    return state.handle, []

  for feature_spec in comp.spec.fixed_feature:
    check.Eq(feature_spec.size, 1, 'All features must have size=1')
    check.Lt(feature_spec.embedding_dim, 0, 'All features must be non-embedded')

  state.handle, indices, ids, _, num_steps = dragnn_ops.bulk_fixed_features(
      state.handle, component=comp.name, num_channels=num_channels)
  size = stride * num_steps

  fixed_ids = []
  for channel, feature_spec in enumerate(comp.spec.fixed_feature):
    tf.logging.info('[%s] Adding fixed feature IDs "%s"', comp.name,
                    feature_spec.name)

    # The +1 and -1 increments ensure that missing IDs default to -1.
    #
    # TODO(googleuser): This formula breaks if multiple IDs are extracted at some
    # step.  Try using tf.unique() to enforce the unique-IDS precondition.
    sums = tf.unsorted_segment_sum(ids[channel] + 1, indices[channel], size) - 1
    sums = tf.expand_dims(sums, axis=1)
    fixed_ids.append(network_units.NamedTensor(sums, feature_spec.name, dim=1))
  return state.handle, fixed_ids 

def create(self,
             fixed_embeddings,
             linked_embeddings,
             context_tensor_arrays,
             attention_tensor,
             during_training,
             stride=None):
    """Requires |stride|; otherwise see base class."""
    # TODO(googleuser): Normalize the arguments to create(). 'stride'
    # is unused by the recurrent network units, while 'context_tensor_arrays'
    # and 'attenion_tensor_array' is unused by bulk network units. b/33587044
    if stride is None:
      raise ValueError("PairwiseConvNetwork needs 'stride'")

    input_tensor = get_input_tensor_with_stride(fixed_embeddings,
                                                linked_embeddings, stride)

    # TODO(googleuser): Add dropout.
    del context_tensor_arrays, attention_tensor, during_training  # Unused.

    num_steps = tf.shape(input_tensor)[1]
    arg1 = tf.expand_dims(input_tensor, 1)
    arg1 = tf.tile(arg1, tf.stack([1, num_steps, 1, 1]))
    arg2 = tf.expand_dims(input_tensor, 2)
    arg2 = tf.tile(arg2, tf.stack([1, 1, num_steps, 1]))
    conv = tf.concat([arg1, arg2], 3)
    for i in xrange(self._num_layers):
      with tf.variable_scope('conv%d' % i, reuse=True) as scope:
        conv = tf.nn.conv2d(
            conv,
            self._component.get_variable('weights'), [1, 1, 1, 1],
            padding='SAME')
        conv = tf.nn.bias_add(conv, self._component.get_variable('biases'))
        if i in self._relu_layers:
          conv = tf.nn.relu(conv, name=scope.name)
    return [tf.reshape(conv, [-1, num_steps], name='reshape_activations')] 

def _create_input_queue(batch_size_per_clone, create_tensor_dict_fn,
                        batch_queue_capacity, num_batch_queue_threads,
                        prefetch_queue_capacity, data_augmentation_options):
  """Sets up reader, prefetcher and returns input queue.

  Args:
    batch_size_per_clone: batch size to use per clone.
    create_tensor_dict_fn: function to create tensor dictionary.
    batch_queue_capacity: maximum number of elements to store within a queue.
    num_batch_queue_threads: number of threads to use for batching.
    prefetch_queue_capacity: maximum capacity of the queue used to prefetch
                             assembled batches.
    data_augmentation_options: a list of tuples, where each tuple contains a
      data augmentation function and a dictionary containing arguments and their
      values (see preprocessor.py).

  Returns:
    input queue: a batcher.BatchQueue object holding enqueued tensor_dicts
      (which hold images, boxes and targets).  To get a batch of tensor_dicts,
      call input_queue.Dequeue().
  """
  tensor_dict = create_tensor_dict_fn()

  tensor_dict[fields.InputDataFields.image] = tf.expand_dims(
      tensor_dict[fields.InputDataFields.image], 0)

  images = tensor_dict[fields.InputDataFields.image]
  float_images = tf.to_float(images)
  tensor_dict[fields.InputDataFields.image] = float_images

  if data_augmentation_options:
    tensor_dict = preprocessor.preprocess(tensor_dict,
                                          data_augmentation_options)

  input_queue = batcher.BatchQueue(
      tensor_dict,
      batch_size=batch_size_per_clone,
      batch_queue_capacity=batch_queue_capacity,
      num_batch_queue_threads=num_batch_queue_threads,
      prefetch_queue_capacity=prefetch_queue_capacity)
  return input_queue 

def attention(self, last_layer, attention_tensor):
    """Compute the attention term for the network unit."""
    h_tensor = attention_tensor

    # Compute the attentions.
    # Using feed-forward net to map the two inputs into the same dimension
    focus_tensor = tf.nn.tanh(
        tf.matmul(
            h_tensor,
            self._component.get_variable('attention_weights_pm_0'),
            name='h_x_pm') + self._component.get_variable('attention_bias_0'))

    context_tensor = tf.nn.tanh(
        tf.matmul(
            last_layer,
            self._component.get_variable('attention_weights_hm_0'),
            name='l_x_hm') + self._component.get_variable('attention_bias_1'))
    # The tf.multiply in the following expression broadcasts along the 0 dim:
    z_vec = tf.reduce_sum(tf.multiply(focus_tensor, context_tensor), 1)
    p_vec = tf.nn.softmax(tf.reshape(z_vec, [1, -1]))
    # The tf.multiply in the following expression broadcasts along the 1 dim:
    r_vec = tf.expand_dims(
        tf.reduce_sum(
            tf.multiply(
                h_tensor, tf.reshape(p_vec, [-1, 1]), name='time_together2'),
            0),
        0)
    return tf.matmul(
        r_vec,
        self._component.get_variable('attention_weights_pu'),
        name='time_together3') 

def _CrossConvHelper(self, encoded_image, kernel):
    """Cross Convolution.

      The encoded image and kernel are of the same shape. Namely
      [batch_size, image_size, image_size, channels]. They are split
      into [image_size, image_size] image squares [kernel_size, kernel_size]
      kernel squares. kernel squares are used to convolute image squares.
    """
    images = tf.expand_dims(encoded_image, 0)
    kernels = tf.expand_dims(kernel, 3)
    return tf.nn.depthwise_conv2d(images, kernels, [1, 1, 1, 1], 'SAME') 

def create(self,
             fixed_embeddings,
             linked_embeddings,
             context_tensor_arrays,
             attention_tensor,
             during_training,
             stride=None):
    """Requires |stride|; otherwise see base class."""
    if stride is None:
      raise RuntimeError("ConvNetwork needs 'stride' and must be called in the "
                         "bulk feature extractor component.")
    input_tensor = get_input_tensor_with_stride(fixed_embeddings,
                                                linked_embeddings, stride)

    # TODO(googleuser): Add context and attention.
    del context_tensor_arrays, attention_tensor

    # On CPU, add a dimension so that the 'image' has shape
    # [stride, 1, num_steps, D].
    conv = tf.expand_dims(input_tensor, 1)
    for i in range(len(self._depths) - 1):
      with tf.variable_scope('conv%d' % i, reuse=True) as scope:
        if during_training:
          conv.set_shape([None, 1, None, self._depths[i]])
          conv = self._maybe_apply_dropout(conv, stride)
        conv = tf.nn.conv2d(
            conv,
            self._component.get_variable('weights'), [1, 1, 1, 1],
            padding='SAME')
        conv = tf.nn.bias_add(conv, self._component.get_variable('biases'))
        if i < (len(self._weights) - 1) or not self._output_dim:
          conv = self._nonlinearity(conv, name=scope.name)
    return [
        tf.reshape(
            conv, [-1, self._depths[-1]], name='reshape_activations')
    ] 

def reorder_beam(beam_size, batch_size, beam_val, output, is_first,
                 tensors_to_reorder):
  """Reorder to minimize beam costs."""
  # beam_val is [batch_size x beam_size]; let b = batch_size * beam_size
  # decided is len x b x a x b
  # output is b x out_size; step is b x len x a x b;
  outputs = tf.split(axis=0, num_or_size_splits=beam_size, value=tf.nn.log_softmax(output))
  all_beam_vals, all_beam_idx = [], []
  beam_range = 1 if is_first else beam_size
  for i in xrange(beam_range):
    top_out, top_out_idx = tf.nn.top_k(outputs[i], k=beam_size)
    cur_beam_val = beam_val[:, i]
    top_out = tf.Print(top_out, [top_out, top_out_idx, beam_val, i,
                                 cur_beam_val], "GREPO", summarize=8)
    all_beam_vals.append(top_out + tf.expand_dims(cur_beam_val, 1))
    all_beam_idx.append(top_out_idx)
  all_beam_idx = tf.reshape(tf.transpose(tf.concat(axis=1, values=all_beam_idx), [1, 0]),
                            [-1])
  top_beam, top_beam_idx = tf.nn.top_k(tf.concat(axis=1, values=all_beam_vals), k=beam_size)
  top_beam_idx = tf.Print(top_beam_idx, [top_beam, top_beam_idx],
                          "GREP", summarize=8)
  reordered = [[] for _ in xrange(len(tensors_to_reorder) + 1)]
  top_out_idx = []
  for i in xrange(beam_size):
    which_idx = top_beam_idx[:, i] * batch_size + tf.range(batch_size)
    top_out_idx.append(tf.gather(all_beam_idx, which_idx))
    which_beam = top_beam_idx[:, i] / beam_size  # [batch]
    which_beam = which_beam * batch_size + tf.range(batch_size)
    reordered[0].append(tf.gather(output, which_beam))
    for i, t in enumerate(tensors_to_reorder):
      reordered[i + 1].append(tf.gather(t, which_beam))
  new_tensors = [tf.concat(axis=0, values=t) for t in reordered]
  top_out_idx = tf.concat(axis=0, values=top_out_idx)
  return (top_beam, new_tensors[0], top_out_idx, new_tensors[1:]) 

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

    Args:
      prediction_tensor: A float tensor of shape [batch_size, num_anchors,
        num_classes] representing the predicted logits for each class
      target_tensor: A float tensor of shape [batch_size, num_anchors,
        num_classes] representing one-hot encoded classification targets
      weights: a float tensor of shape [batch_size, num_anchors]

    Returns:
      loss: a (scalar) tensor representing the value of the loss function
            or a float tensor of shape [batch_size, num_anchors]
    """
    if self._bootstrap_type == 'soft':
      bootstrap_target_tensor = self._alpha * target_tensor + (
          1.0 - self._alpha) * tf.sigmoid(prediction_tensor)
    else:
      bootstrap_target_tensor = self._alpha * target_tensor + (
          1.0 - self._alpha) * tf.cast(
              tf.sigmoid(prediction_tensor) > 0.5, tf.float32)
    per_entry_cross_ent = (tf.nn.sigmoid_cross_entropy_with_logits(
        labels=bootstrap_target_tensor, logits=prediction_tensor))
    if self._anchorwise_output:
      return tf.reduce_sum(per_entry_cross_ent * tf.expand_dims(weights, 2), 2)
    return tf.reduce_sum(per_entry_cross_ent * tf.expand_dims(weights, 2)) 

def expectedMinImageAfterColorScale(self):
    images_r = tf.constant([[[-0.1, -0.1, -0.1, -0.1], [-1, -1, -0.1, -0.1],
                             [-1, -0.1, -0.1, -0.1], [0.4, 0.4, -0.1, -0.1]]],
                           dtype=tf.float32)
    images_r = tf.expand_dims(images_r, 3)
    images_g = tf.constant([[[-1, -1, -0.1, -0.1], [-1, -1, -0.1, -0.1],
                             [-1, -0.1, 0.4, 0.4], [0.4, 0.4, -0.1, 0.4]]],
                           dtype=tf.float32)
    images_g = tf.expand_dims(images_g, 3)
    images_b = tf.constant([[[-0.1, -0.1, 0.4, -1], [-1, -1, -0.1, 0.4],
                             [-1, -0.1, -0.1, -1], [0.4, 0.4, 0.4, -0.1]]],
                           dtype=tf.float32)
    images_b = tf.expand_dims(images_b, 3)
    images = tf.concat([images_r, images_g, images_b], 3)
    return images 

def _average_gradients(tower_grads):
  """Calculate the average gradient for each shared variable across all towers.

  Note that this function provides a synchronization point across all towers.

  Args:
    tower_grads: List of lists of (gradient, variable) tuples. The outer list
      is over individual gradients. The inner list is over the gradient
      calculation for each tower.
  Returns:
     List of pairs of (gradient, variable) where the gradient has been averaged
     across all towers.
  """
  average_grads = []
  for grad_and_vars in zip(*tower_grads):
    # Note that each grad_and_vars looks like the following:
    #   ((grad0_gpu0, var0_gpu0), ... , (grad0_gpuN, var0_gpuN))
    grads = []
    for g, _ in grad_and_vars:
      # Add 0 dimension to the gradients to represent the tower.
      expanded_g = tf.expand_dims(g, 0)

      # Append on a 'tower' dimension which we will average over below.
      grads.append(expanded_g)

    # Average over the 'tower' dimension.
    grad = tf.concat(axis=0, values=grads)
    grad = tf.reduce_mean(grad, 0)

    # Keep in mind that the Variables are redundant because they are shared
    # across towers. So .. we will just return the first tower's pointer to
    # the Variable.
    v = grad_and_vars[0][1]
    grad_and_var = (grad, v)
    average_grads.append(grad_and_var)
  return average_grads 

def AddCrossEntropy(batch_size, n):
  """Adds a cross entropy cost function."""
  cross_entropies = []
  def _Pass():
    return tf.constant(0, dtype=tf.float32, shape=[1])

  for beam_id in range(batch_size):
    beam_gold_slot = tf.reshape(
        tf.strided_slice(n['gold_slot'], [beam_id], [beam_id + 1]), [1])
    def _ComputeCrossEntropy():
      """Adds ops to compute cross entropy of the gold path in a beam."""
      # Requires a cast so that UnsortedSegmentSum, in the gradient,
      # is happy with the type of its input 'segment_ids', which
      # must be int32.
      idx = tf.cast(
          tf.reshape(
              tf.where(tf.equal(n['beam_ids'], beam_id)), [-1]), tf.int32)
      beam_scores = tf.reshape(tf.gather(n['all_path_scores'], idx), [1, -1])
      num = tf.shape(idx)
      return tf.nn.softmax_cross_entropy_with_logits(
          labels=tf.expand_dims(
              tf.sparse_to_dense(beam_gold_slot, num, [1.], 0.), 0),
          logits=beam_scores)
    # The conditional here is needed to deal with the last few batches of the
    # corpus which can contain -1 in beam_gold_slot for empty batch slots.
    cross_entropies.append(cf.cond(
        beam_gold_slot[0] >= 0, _ComputeCrossEntropy, _Pass))
  return {'cross_entropy': tf.div(tf.add_n(cross_entropies), batch_size)}