Python tensorflow.compat.v1.shape() Examples

def pad_for_tpu(shapes_dict, hparams, max_length):
  """Pads unknown features' dimensions for TPU."""
  padded_shapes = {}

  def get_filler(specified_max_length):
    if not specified_max_length:
      return max_length
    return min(specified_max_length, max_length)

  inputs_none_filler = get_filler(hparams.max_input_seq_length)
  targets_none_filler = get_filler(hparams.max_target_seq_length)

  def pad_one_shape(shape, none_filler):
    return [
        (dim if dim is not None else none_filler) for dim in shape.as_list()
    ]

  for key, shape in six.iteritems(shapes_dict):
    if key == "inputs":
      padded_shapes[key] = pad_one_shape(shape, inputs_none_filler)
    elif key == "targets":
      padded_shapes[key] = pad_one_shape(shape, targets_none_filler)
    else:
      padded_shapes[key] = pad_one_shape(shape, max_length)
  return padded_shapes 

def _cell_base(self, net, prev_layer):
    """Runs the beginning of the conv cell before the predicted ops are run."""
    num_filters = self._filter_size

    # Check to be sure prev layer stuff is setup correctly
    prev_layer = self._reduce_prev_layer(prev_layer, net)

    net = tf.nn.relu(net)
    net = slim.conv2d(net, num_filters, 1, scope='1x1')
    net = slim.batch_norm(net, scope='beginning_bn')
    split_axis = get_channel_index()
    net = tf.split(axis=split_axis, num_or_size_splits=1, value=net)
    for split in net:
      assert int(split.shape[split_axis] == int(
          self._num_conv_filters * self._filter_scaling))
    net.append(prev_layer)
    return net 

def _reduce_prev_layer(self, prev_layer, curr_layer):
    """Matches dimension of prev_layer to the curr_layer."""
    # Set the prev layer to the current layer if it is none
    if prev_layer is None:
      return curr_layer
    curr_num_filters = self._filter_size
    prev_num_filters = get_channel_dim(prev_layer.shape)
    curr_filter_shape = int(curr_layer.shape[2])
    prev_filter_shape = int(prev_layer.shape[2])
    if curr_filter_shape != prev_filter_shape:
      prev_layer = tf.nn.relu(prev_layer)
      prev_layer = factorized_reduction(prev_layer, curr_num_filters, stride=2)
    elif curr_num_filters != prev_num_filters:
      prev_layer = tf.nn.relu(prev_layer)
      prev_layer = slim.conv2d(
          prev_layer, curr_num_filters, 1, scope='prev_1x1')
      prev_layer = slim.batch_norm(prev_layer, scope='prev_bn')
    return prev_layer 

def decode_jpeg(image_buffer, scope=None):  # , dtype=tf.float32):
  """Decode a JPEG string into one 3-D float image Tensor.

  Args:
    image_buffer: scalar string Tensor.
    scope: Optional scope for op_scope.
  Returns:
    3-D float Tensor with values ranging from [0, 1).
  """
  # with tf.op_scope([image_buffer], scope, 'decode_jpeg'):
  # with tf.name_scope(scope, 'decode_jpeg', [image_buffer]):
  with tf.name_scope(scope or 'decode_jpeg'):
    # Decode the string as an RGB JPEG.
    # Note that the resulting image contains an unknown height and width
    # that is set dynamically by decode_jpeg. In other words, the height
    # and width of image is unknown at compile-time.
    image = tf.image.decode_jpeg(image_buffer, channels=3,
                                 fancy_upscaling=False,
                                 dct_method='INTEGER_FAST')

    # image = tf.Print(image, [tf.shape(image)], 'Image shape: ')

    return image 

def _ensure_keep_mask(self, x):
    if self._keep_mask is None or not self._share_mask:
      shape = tf.shape(x)
      k = shape[1]
      # To make this class a drop-in replacement for bernoulli dropout we
      # paramaterize it with keep_prob. Set alpha of the dirichlet so that the
      # variance is equal to the variance of the bernoulli with p=keep_prob
      # divided by keep_prob.
      # Now the variance of the dirichlet with k equal alphas is
      # (k-1)/(k^2(k*alpha+1). Solve that for alpha.
      kf = tf.cast(k, tf.float32)
      alpha = self._keep_prob * (kf - 1.0) / ((1-self._keep_prob)*kf) - 1.0/kf
      dist = tfp.distributions.Dirichlet(tf.ones(shape=k) * alpha)
      assert (dist.reparameterization_type ==
              tfp.distributions.FULLY_REPARAMETERIZED)
      # The E[dir(alpha)] = 1/k for all elements, but we want the expectation to
      # be keep_prob, hence the multiplication.
      self._keep_mask = kf * dist.sample(shape[0])
      self._keep_mask.set_shape(x.get_shape())
    return self._keep_mask 

def _build(self, x, state):
    prev_keep_mask = state
    shape = tf.shape(x)
    noise = tf.random_uniform(shape, dtype=x.dtype)
    other_mask = tf.floor(self._keep_prob + noise)
    choice_noise = tf.random_uniform(shape, dtype=x.dtype)
    choice = tf.less(choice_noise, self._flip_prob)
    # KLUDGE(melisgl): The client has to pass the last keep_mask from
    # a batch to the next so the mask may end up next to some
    # recurrent cell state. This state is often zero at the beginning
    # and may be periodically zeroed (per example) during training.
    # While zeroing LSTM state is okay, zeroing the dropout mask is
    # not. So instead of forcing every client to deal with this common
    # (?) case, if an all zero mask is detected, then regenerate a
    # fresh mask. This is of course a major hack and won't help with
    # learnt initial states, for example.
    sum_ = tf.reduce_sum(prev_keep_mask, 1, keepdims=True)
    is_initializing = tf.equal(sum_, 0.0)

    self._keep_mask = tf.where(tf.logical_or(choice, is_initializing),
                               other_mask,
                               prev_keep_mask)
    self._time_step += 1
    return x * self._keep_mask / self._keep_prob * self._scaler 

def layer_norm(x, reduction_indices, epsilon=1e-9, gain=None, bias=None,
               per_element=True, scope=None):
  """DOC."""
  reduction_indices = ensure_list(reduction_indices)
  mean = tf.reduce_mean(x, reduction_indices, keep_dims=True)
  variance = tf.reduce_mean(tf.squared_difference(x, mean),
                            reduction_indices, keep_dims=True)
  normalized = (x - mean) / tf.sqrt(variance + epsilon)
  dtype = x.dtype
  shape = x.get_shape().as_list()
  for i in six.moves.range(len(shape)):
    if i not in reduction_indices or not per_element:
      shape[i] = 1
  with tf.variable_scope(scope or 'layer_norm'):
    if gain is None:
      gain = tf.get_variable('gain', shape=shape, dtype=dtype,
                             initializer=tf.ones_initializer())
    if bias is None:
      bias = tf.get_variable('bias', shape=shape, dtype=dtype,
                             initializer=tf.zeros_initializer())
  return gain*normalized+bias 

def sparse_random_indices(ratio, shape):
  """DOC."""
  assert 0 < ratio and ratio <= 1.0
  n = round_to_int(tf.TensorShape(shape).num_elements()*ratio)
  # There are two implementations. The first generates random indices
  # and wastes computation due to collisions, and the second wastes
  # memory.
  if ratio < 0.25:
    indices = {}
    if isinstance(shape, tf.TensorShape):
      shape = shape.as_list()
    while len(indices) < n:
      index = _random_index(shape)
      indices[index] = True
    return indices.keys()
  else:
    indices = _all_indices(shape)
    random.shuffle(indices)
    return indices[:n] 

def log_trainables(scopes=('',)):
  """"Log number of trainable parameters for each scope in `scopes`.

  Args:
    scopes: A sequence of scope names.

  Returns:
    The total number of trainable parameters over all scopes in
    `scopes`. Possibly counting some parameters multiple times if the
    scopes are nested.
  """
  total = 0
  for scope in scopes:
    logging.info('Trainables in scope "%s":', scope)
    n = 0
    for var in trainable_vars_in_scope(scope):
      shape = var.get_shape()
      logging.info('trainable: %s shape %r (%r)', var.name, shape.as_list(),
                   shape.num_elements())
      n += shape.num_elements()
    logging.info('Number of parameters in scope "%s": %r', scope, n)
    total += n
  return total 

def padded_accuracy_topk(predictions,
                         labels,
                         k,
                         weights_fn=common_layers.weights_nonzero):
  """Percentage of times that top-k predictions matches labels on non-0s."""
  with tf.variable_scope("padded_accuracy_topk", values=[predictions, labels]):
    padded_predictions, padded_labels = common_layers.pad_with_zeros(
        predictions, labels)
    weights = weights_fn(padded_labels)
    effective_k = tf.minimum(k,
                             common_layers.shape_list(padded_predictions)[-1])
    _, outputs = tf.nn.top_k(padded_predictions, k=effective_k)
    outputs = tf.to_int32(outputs)
    padded_labels = tf.to_int32(padded_labels)
    padded_labels = tf.expand_dims(padded_labels, axis=-1)
    padded_labels += tf.zeros_like(outputs)  # Pad to same shape.
    same = tf.to_float(tf.equal(outputs, padded_labels))
    same_topk = tf.reduce_sum(same, axis=-1)
    return same_topk, weights 

def two_class_log_likelihood(predictions, labels, weights_fn=None):
  """Log-likelihood for two class classification with 0/1 labels.

  Args:
    predictions: A float valued tensor of shape [`batch_size`].  Each
      component should be between 0 and 1.
    labels: An int valued tensor of shape [`batch_size`].  Each component
      should either be 0 or 1.
    weights_fn: unused.

  Returns:
    A pair, with the average log likelihood in the first component.
  """
  del weights_fn
  float_predictions = tf.cast(tf.squeeze(predictions), dtype=tf.float64)
  batch_probs = tf.stack([1. - float_predictions, float_predictions], axis=-1)
  int_labels = tf.cast(tf.squeeze(labels), dtype=tf.int32)
  onehot_targets = tf.cast(tf.one_hot(int_labels, 2), dtype=tf.float64)
  chosen_probs = tf.einsum(
      "ij,ij->i", batch_probs, onehot_targets, name="chosen_probs")
  avg_log_likelihood = tf.reduce_mean(tf.log(chosen_probs))
  return avg_log_likelihood, tf.constant(1.0) 

def set_precision(predictions, labels,
                  weights_fn=common_layers.weights_nonzero):
  """Precision of set predictions.

  Args:
    predictions : A Tensor of scores of shape [batch, nlabels].
    labels: A Tensor of int32s giving true set elements,
      of shape [batch, seq_length].
    weights_fn: A function to weight the elements.

  Returns:
    hits: A Tensor of shape [batch, nlabels].
    weights: A Tensor of shape [batch, nlabels].
  """
  with tf.variable_scope("set_precision", values=[predictions, labels]):
    labels = tf.squeeze(labels, [2, 3])
    weights = weights_fn(labels)
    labels = tf.one_hot(labels, predictions.shape[-1])
    labels = tf.reduce_max(labels, axis=1)
    labels = tf.cast(labels, tf.bool)
    return tf.to_float(tf.equal(labels, predictions)), weights 

def image_summary(predictions, targets, hparams):
  """Reshapes predictions and passes it to tensorboard.

  Args:
    predictions : The predicted image (logits).
    targets : The ground truth.
    hparams: model hparams.

  Returns:
    summary_proto: containing the summary images.
    weights: A Tensor of zeros of the same shape as predictions.
  """
  del hparams
  results = tf.cast(tf.argmax(predictions, axis=-1), tf.uint8)
  gold = tf.cast(targets, tf.uint8)
  summary1 = tf.summary.image("prediction", results, max_outputs=2)
  summary2 = tf.summary.image("data", gold, max_outputs=2)
  summary = tf.summary.merge([summary1, summary2])
  return summary, tf.zeros_like(predictions) 

def global_pool(input_tensor, pool_op=tf.nn.avg_pool):
  """Applies avg pool to produce 1x1 output.

  NOTE: This function is funcitonally equivalenet to reduce_mean, but it has
  baked in average pool which has better support across hardware.

  Args:
    input_tensor: input tensor
    pool_op: pooling op (avg pool is default)
  Returns:
    a tensor batch_size x 1 x 1 x depth.
  """
  shape = input_tensor.get_shape().as_list()
  if shape[1] is None or shape[2] is None:
    kernel_size = tf.convert_to_tensor(
        [1, tf.shape(input_tensor)[1],
         tf.shape(input_tensor)[2], 1])
  else:
    kernel_size = [1, shape[1], shape[2], 1]
  output = pool_op(
      input_tensor, ksize=kernel_size, strides=[1, 1, 1, 1], padding='VALID')
  # Recover output shape, for unknown shape.
  output.set_shape([None, 1, 1, None])
  return output 

def _greedy_infer(self, features, decode_length, use_tpu=False):
    """A greedy inference method.

    Models should ideally implement a more efficient version of this function.

    Args:
      features: an map of string to `Tensor`
      decode_length: an integer.  How many additional timesteps to decode.
      use_tpu: A bool, whether to build the inference graph for TPU.

    Returns:
      A dict of decoding results {
          "outputs": integer `Tensor` of decoded ids of shape
              [batch_size, <= decode_length] if beam_size == 1 or
              [batch_size, top_beams, <= decode_length]
          "scores": None
          "logits": `Tensor` of shape [batch_size, time, 1, 1, vocab_size].
          "losses": a dictionary: {loss-name (string): floating point `Scalar`}
      }
    """
    if use_tpu:
      return self._slow_greedy_infer_tpu(features, decode_length)
    return self._slow_greedy_infer(features, decode_length) 

def body(self, features):
    """Computes the targets' pre-logit activations given transformed inputs.

    Most `T2TModel` subclasses will override this method.

    Args:
      features: dict of str to Tensor, where each Tensor has shape [batch_size,
        ..., hidden_size]. It typically contains keys `inputs` and `targets`.

    Returns:
      output: Tensor of pre-logit activations with shape [batch_size, ...,
              hidden_size].
      losses: Either single loss as a scalar, a list, a Tensor (to be averaged),
              or a dictionary of losses. If losses is a dictionary with the key
              "training", losses["training"] is considered the final training
              loss and output is considered logits; self.top and self.loss will
              be skipped.
    """
    raise NotImplementedError("Abstract Method") 

def combine(self, x):
    """Return the output from the experts.

    When one example goes to multiple experts, the outputs are summed.

    Args:
      x: a Tensor with shape [batch, num_experts, expert_capacity, depth]

    Returns:
      a `Tensor` with shape `[batch, length, depth]
    """
    depth = tf.shape(x)[-1]
    x *= tf.expand_dims(self._nonpadding, -1)
    ret = tf.unsorted_segment_sum(
        x, self._flat_indices, num_segments=self._batch * self._length)
    ret = tf.reshape(ret, [self._batch, self._length, depth])
    return ret 

def combine(self, expert_out, multiply_by_gates=True):
    """Sum together the expert output, weighted by the gates.

    The slice corresponding to a particular batch element `b` is computed
    as the sum over all experts `i` of the expert output, weighted by the
    corresponding gate values.  If `multiply_by_gates` is set to False, the
    gate values are ignored.

    Args:
      expert_out: a list of `num_experts` `Tensor`s, each with shape
        `[expert_batch_size_i, <extra_output_dims>]`.
      multiply_by_gates: a boolean

    Returns:
      a `Tensor` with shape `[batch_size, <extra_output_dims>]`.
    """
    # see comments on convert_gradient_to_tensor
    stitched = common_layers.convert_gradient_to_tensor(
        tf.concat(expert_out, 0))
    if multiply_by_gates:
      stitched *= tf.expand_dims(self._nonzero_gates, 1)
    combined = tf.unsorted_segment_sum(stitched, self._batch_index,
                                       tf.shape(self._gates)[0])
    return combined 

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

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

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

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

def restore(self, x):
    """Add padding back to the given tensor.

    Args:
      x (tf.Tensor): of shape [dim_compressed,...]

    Returns:
      a tensor of shape [dim_origin,...] with dim_compressed >= dim_origin. The
      dim is restored from the original reference tensor
    """
    with tf.name_scope("pad_reduce/restore"):
      x = tf.scatter_nd(
          indices=self.nonpad_ids,
          updates=x,
          shape=tf.concat([self.dim_origin, tf.shape(x)[1:]], axis=0),
      )
    return x 

def remove(self, x):
    """Remove padding from the given tensor.

    Args:
      x (tf.Tensor): of shape [dim_origin,...]

    Returns:
      a tensor of shape [dim_compressed,...] with dim_compressed <= dim_origin
    """
    with tf.name_scope("pad_reduce/remove"):
      x_shape = x.get_shape().as_list()
      x = tf.gather_nd(
          x,
          indices=self.nonpad_ids,
      )
      if not tf.executing_eagerly():
        # This is a hack but for some reason, gather_nd return a tensor of
        # undefined shape, so the shape is set up manually
        x.set_shape([None] + x_shape[1:])
    return x 

def _normal_distribution_cdf(x, stddev):
  """Evaluates the CDF of the normal distribution.

  Normal distribution with mean 0 and standard deviation stddev,
  evaluated at x=x.

  input and output `Tensor`s have matching shapes.

  Args:
    x: a `Tensor`
    stddev: a `Tensor` with the same shape as `x`.

  Returns:
    a `Tensor` with the same shape as `x`.

  """
  return 0.5 * (1.0 + tf.erf(x / (math.sqrt(2) * stddev + 1e-20))) 

def sigmoid_precision_one_hot(logits, labels, weights_fn=None):
  """Calculate precision for a set, given one-hot labels and logits.

  Predictions are converted to one-hot,
  as predictions[example][arg-max(example)] = 1

  Args:
    logits: Tensor of size [batch-size, o=1, p=1, num-classes]
    labels: Tensor of size [batch-size, o=1, p=1, num-classes]
    weights_fn: Function that takes in labels and weighs examples (unused)
  Returns:
    precision (scalar), weights
  """
  with tf.variable_scope("sigmoid_precision_one_hot", values=[logits, labels]):
    del weights_fn
    num_classes = logits.shape[-1]
    predictions = tf.nn.sigmoid(logits)
    predictions = tf.argmax(predictions, -1)
    predictions = tf.one_hot(predictions, num_classes)
    _, precision = tf.metrics.precision(labels=labels, predictions=predictions)
    return precision, tf.constant(1.0) 

def set_recall(predictions, labels, weights_fn=common_layers.weights_nonzero):
  """Recall of set predictions.

  Args:
    predictions : A Tensor of scores of shape [batch, nlabels].
    labels: A Tensor of int32s giving true set elements,
      of shape [batch, seq_length].
    weights_fn: A function to weight the elements.

  Returns:
    hits: A Tensor of shape [batch, nlabels].
    weights: A Tensor of shape [batch, nlabels].
  """
  with tf.variable_scope("set_recall", values=[predictions, labels]):
    labels = tf.squeeze(labels, [2, 3])
    weights = weights_fn(labels)
    labels = tf.one_hot(labels, predictions.shape[-1])
    labels = tf.reduce_max(labels, axis=1)
    labels = tf.cast(labels, tf.bool)
    return tf.to_float(tf.equal(labels, predictions)), weights 

def _operation_to_pooling_info(operation):
  """Parses the pooling operation string to return its type and shape."""
  pooling_type = _operation_to_pooling_type(operation)
  pooling_shape = _operation_to_pooling_shape(operation)
  return pooling_type, pooling_shape 

def flatten_all_but_last(a):
  """Flatten all dimensions of a except the last."""
  ret = tf.reshape(a, [-1, tf.shape(a)[-1]])
  if not tf.executing_eagerly():
    ret.set_shape([None] + a.get_shape().as_list()[-1:])
  return ret 

def _operation_to_info(operation):
  """Takes in operation name and returns meta information.

  An example would be 'separable_3x3_4' -> (3, 4).

  Args:
    operation: String that corresponds to convolution operation.

  Returns:
    Tuple of (filter shape, num layers).
  """
  num_layers = _operation_to_num_layers(operation)
  filter_shape = _operation_to_filter_shape(operation)
  return num_layers, filter_shape 

def dispatch(self, inp):
    """Create one input Tensor for each expert.

    The `Tensor` for a expert `i` contains the slices of `inp` corresponding
    to the batch elements `b` where `gates[b, i] > 0`.

    Args:
      inp: a `Tensor` of shape "[batch_size, <extra_input_dims>]`
    Returns:
      a list of `num_experts` `Tensor`s with shapes
        `[expert_batch_size_i, <extra_input_dims>]`.
    """
    inp = tf.gather(inp, self._batch_index)
    return tf.split(inp, self._part_sizes_tensor, 0, num=self._num_experts) 

def _operation_to_pooling_shape(operation):
  """Takes in the operation string and returns the pooling kernel shape."""
  splitted_operation = operation.split('_')
  shape = splitted_operation[-1]
  assert 'x' in shape
  filter_height, filter_width = shape.split('x')
  assert filter_height == filter_width
  return int(filter_height) 

def _ensure_keep_mask(self, x):
    if self._keep_mask is None or not self._share_mask:
      shape = tf.shape(x)
      # Calculate the stddev for the normal distribution that
      # matches the stddev of the bernoulli with p=keep_prob.
      stddev = tf.sqrt((1 - self._keep_prob) / self._keep_prob)
      self._keep_mask = tf.random_normal(shape, mean=1.0, stddev=stddev,
                                         dtype=x.dtype)
      self._keep_mask.set_shape(x.get_shape())
    return self._keep_mask