Python array_ops.unique函数代码示例

 def _apply_sparse(self, g_t, x_tm1, prepare):
   """"""
   
   idxs, idxs_ = array_ops.unique(g_t.indices)
   g_t_ = math_ops.unsorted_segment_sum(g_t.values, idxs_, array_ops.size(idxs))
   updates = []
   
   if self._mu > 0:
     m_and_t = self._sparse_moving_average(x_tm1, idxs, g_t_, 'm', self._mu)
     m_t_ = array_ops.gather(m_and_t[0], idxs)
     gamma_t = ops.convert_to_tensor(self._gamma)
     m_bar_t_ = (1-gamma_t)*m_t_ + gamma_t*g_t_
     updates.extend(m_and_t)
   else:
     m_bar_t_ = g_t_
   
   if self._ups > 0:
     v_and_t = self._sparse_moving_average(x_tm1, idxs, g_t_**2, 'v', self._ups)
     v_t_ = array_ops.gather(v_and_t[0], idxs)
     eps_t = ops.convert_to_tensor(self._eps)
     v_bar_t_ = math_ops.sqrt(v_t_ + eps_t)
     updates.extend(v_and_t)
   else:
     v_bar_t_ = 1.
   
   lr_t = ops.convert_to_tensor(self._lr)
   s_t_ = lr_t * m_bar_t_ / v_bar_t_
   return [[s_t_, x_tm1, idxs, g_t]] + updates

def embedding_lookup_unique(params, ids, name=None):
  """Version of embedding_lookup that avoids duplicate lookups.

  This can save communication in the case of repeated ids.
  Same interface as embedding_lookup. Except it supports multi-dimensional `ids`
  which allows to not reshape input/output to fit gather.

  Args:
    params: A list of tensors with the same shape and type, or a
      `PartitionedVariable`. Shape `[index, d1, d2, ...]`.
    ids: A one-dimensional `Tensor` with type `int32` or `int64` containing
      the ids to be looked up in `params`. Shape `[ids1, ids2, ...]`.
    name: A name for this operation (optional).

  Returns:
    A `Tensor` with the same type as the tensors in `params` and dimension of
    `[ids1, ids2, d1, d2, ...]`.

  Raises:
    ValueError: If `params` is empty.
  """
  with ops.name_scope(name, "EmbeddingLookupUnique", [params, ids]):
    ids = ops.convert_to_tensor(ids)
    shape = array_ops.shape(ids)
    ids_flat = array_ops.reshape(
        ids, math_ops.reduce_prod(shape, keep_dims=True))
    unique_ids, idx = array_ops.unique(ids_flat)
    unique_embeddings = embedding_ops.embedding_lookup(params, unique_ids)
    embeds_flat = array_ops.gather(unique_embeddings, idx)
    embed_shape = array_ops.concat(
        [shape, array_ops.shape(unique_embeddings)[1:]], 0)
    embeds = array_ops.reshape(embeds_flat, embed_shape)
    embeds.set_shape(ids.get_shape().concatenate(
        unique_embeddings.get_shape()[1:]))
    return embeds

 def approximate_hessian(self, grads_and_vars, name=None):
   """
   I haven't tested this yet so I have no idea if it works, but even if it
   does it's probably super slow, and either way nothing else has been modified
   to deal with it.
   """
   
   gv = 0
   var_refs = []
   for g_t, x_tm1 in grads_and_vars:
     var_refs.append(x_tm1.ref())
     if g_t is None:
       continue
     with ops.name_scope('update_' + x_tm1.op.name), ops.device(x_tm1.device):
       if isinstance(g_t, ops.Tensor):
         gv += math_ops.reduce_sum(g_t * random_ops.random_normal(g_t.get_shape()))
       else:
         idxs, idxs_ = array_ops.unique(g_t.indices)
         g_t_ = math_ops.unsorted_segment_sum(g_t.values, idxs_, array_ops.size(idxs))
         gv += math_ops.reduce_sum(g_t_ * random_ops.random_normal(g_t_.get_shape()))
   hesses = gradients.gradients(gv, var_refs,
                                gate_gradients=(gate_gradients == Optimizer.GATE_OP),
                                aggregation_method=aggregation_method,
                                colocate_gradients_with_ops=colocate_gradients_with_ops)
   return zip([g_t for g_t, _ in grads_and_vars], [x_tm1 for _, x_tm1 in grads_and_vars], hesses)

 def _prepare(self, grads_and_vars):
   """"""
   
   if self._lr is None:
     sTy = 0
     sTs = 0
     yTy = 0
     for g_t, x_tm1 in grads_and_vars:
       if g_t is None:
         continue
       with ops.name_scope('update_' + x_tm1.op.name), ops.device(x_tm1.device):
         if isinstance(g_t, ops.Tensor):
           g_tm1 = self.get_slot(x_tm1, 'g')
           s_tm1 = self.get_slot(x_tm1, 's')
           y_t = (g_t-g_tm1)
           sTy += math_ops.reduce_sum(s_tm1*y_t)
           sTs += math_ops.reduce_sum(s_tm1**2)
           yTy += math_ops.reduce_sum(y_t**2)
         else:
           idxs, idxs_ = array_ops.unique(g_t.indices)
           g_t_ = math_ops.unsorted_segment_sum(g_t.values, idxs_, array_ops.size(idxs))
           g_tm1 = self.get_slot(x_tm1, 'g')
           g_tm1_ = array_ops.gather(g_tm1, idxs)
           s_tm1 = self.get_slot(x_tm1, 's')
           s_tm1_ = array_ops.gather(s_tm1, idxs)
           y_t_ = (g_t_-g_tm1_)
           sTy += math_ops.reduce_sum(s_tm1_*y_t_)
           sTs += math_ops.reduce_sum(s_tm1_**2)
           yTy += math_ops.reduce_sum(y_t_**2)
     sTy = math_ops.abs(sTy)
     self._lr = sTs / (sTy + self._eps)

 def _apply_sparse(self, grad, var):
   if len(grad.indices.get_shape()) == 1:
     grad_indices = grad.indices
     grad_values = grad.values
   else:
     grad_indices = array_ops.reshape(grad.indices, [-1])
     grad_values = array_ops.reshape(grad.values, [-1, grad.values.get_shape()[-1].value])
   gidxs, metagidxs = array_ops.unique(grad_indices)
   sizegidxs = array_ops.size(gidxs)
   gvals = math_ops.unsorted_segment_sum(grad_values, metagidxs, sizegidxs)
   # m_t = mu * m + (1 - mu) * g_t
   m = self.get_slot(var, "m")
   m_scaled_g_values = gvals * (1 - self._mu_t)
   m_t = state_ops.scatter_update(m, gidxs,
                                  array_ops.gather(m, gidxs) * self._mu_t,
                                  use_locking=self._use_locking)
   m_t = state_ops.scatter_add(m_t, gidxs, m_scaled_g_values,
                               use_locking=self._use_locking)
   m_t_ = array_ops.gather(m_t, gidxs) / (1 - self._mu2_t * self._mu_power)
   # m_bar = mu * m_t + (1 - mu) * g_t
   m_bar = self._mu2_t * m_t_ + m_scaled_g_values / (1 - self._mu_power)
   var_update = state_ops.scatter_sub(var, gidxs,
                                    self._lr_t * m_bar,
                                    use_locking=self._use_locking)
   return control_flow_ops.group(*[var_update, m_t])

  def quantiles_ready():
    """The subgraph for when the quantiles are ready."""
    quantized_feature = quantile_ops.quantiles([sparse_column_values], [],
                                               [quantile_buckets], [])
    quantized_feature = math_ops.cast(quantized_feature[0], dtypes.int64)
    quantized_feature = array_ops.reshape(quantized_feature, [-1])
    example_indices, _ = array_ops.split(
        sparse_column_indices, num_or_size_splits=2, axis=1)
    example_indices = array_ops.squeeze(example_indices, [1])
    filtered_gradients = array_ops.gather(gradients, example_indices)
    filtered_hessians = array_ops.gather(hessians, example_indices)
    filtered_partition_ids = array_ops.gather(example_partition_ids,
                                              example_indices)
    unique_partitions, mapped_partitions = array_ops.unique(
        example_partition_ids)

    # Compute aggregate stats for each partition.
    per_partition_gradients = math_ops.unsorted_segment_sum(
        gradients, mapped_partitions, array_ops.size(unique_partitions))
    per_partition_hessians = math_ops.unsorted_segment_sum(
        hessians, mapped_partitions, array_ops.size(unique_partitions))

    # Prepend a bias feature per partition that accumulates the stats for all
    # examples in that partition.
    bias_feature_ids = array_ops.fill(
        array_ops.shape(unique_partitions), _BIAS_FEATURE_ID)
    bias_feature_ids = math_ops.cast(bias_feature_ids, dtypes.int64)
    partition_ids = array_ops.concat(
        [unique_partitions, filtered_partition_ids], 0)
    filtered_gradients = array_ops.concat(
        [per_partition_gradients, filtered_gradients], 0)
    filtered_hessians = array_ops.concat(
        [per_partition_hessians, filtered_hessians], 0)
    bucket_ids = array_ops.concat([bias_feature_ids, quantized_feature], 0)
    return partition_ids, bucket_ids, filtered_gradients, filtered_hessians

  def testInt32(self):
    x = np.random.randint(2, high=10, size=7000)
    with self.cached_session() as sess:
      y, idx = array_ops.unique(x)
      tf_y, tf_idx = self.evaluate([y, idx])

    self.assertEqual(len(x), len(tf_idx))
    self.assertEqual(len(tf_y), len(np.unique(x)))
    for i in range(len(x)):
      self.assertEqual(x[i], tf_y[tf_idx[i]])

  def testInt32OutIdxInt64(self):
    x = np.random.randint(2, high=10, size=7000)
    with self.test_session() as sess:
      y, idx = array_ops.unique(x, out_idx=dtypes.int64)
      tf_y, tf_idx = sess.run([y, idx])

    self.assertEqual(len(x), len(tf_idx))
    self.assertEqual(len(tf_y), len(np.unique(x)))
    for i in range(len(x)):
      self.assertEqual(x[i], tf_y[tf_idx[i]])

  def testString(self):
    indx = np.random.randint(65, high=122, size=7000)
    x = [chr(i) for i in indx]
    with self.test_session() as sess:
      y, idx = array_ops.unique(x)
      tf_y, tf_idx = sess.run([y, idx])

    self.assertEqual(len(x), len(tf_idx))
    self.assertEqual(len(tf_y), len(np.unique(x)))
    for i in range(len(x)):
      self.assertEqual(x[i], tf_y[tf_idx[i]].decode('ascii'))

    def active_inputs():
      """The normal flow when the handler is active."""
      # Remove the second column of example indices matrix since it is not
      # useful.
      example_indices, _ = array_ops.split(
          self._sparse_int_column.indices, num_or_size_splits=2, axis=1)
      example_indices = array_ops.squeeze(example_indices, [1])

      filtered_gradients = array_ops.gather(gradients, example_indices)
      filtered_hessians = array_ops.gather(hessians, example_indices)
      filtered_partition_ids = array_ops.gather(example_partition_ids,
                                                example_indices)
      unique_partitions, mapped_partitions = array_ops.unique(
          example_partition_ids)

      # Compute aggregate stats for each partition.
      # The bias is computed on gradients and hessians (and not
      # filtered_gradients) which have exactly one value per example, so we
      # don't double count a gradient in multivalent columns.
      # Since unsorted_segment_sum can be numerically unstable, use 64bit
      # operation.
      gradients64 = math_ops.cast(gradients, dtypes.float64)
      hessians64 = math_ops.cast(hessians, dtypes.float64)
      per_partition_gradients = math_ops.unsorted_segment_sum(
          gradients64, mapped_partitions, array_ops.size(unique_partitions))
      per_partition_hessians = math_ops.unsorted_segment_sum(
          hessians64, mapped_partitions, array_ops.size(unique_partitions))
      per_partition_gradients = math_ops.cast(per_partition_gradients,
                                              dtypes.float32)
      per_partition_hessians = math_ops.cast(per_partition_hessians,
                                             dtypes.float32)
      # Prepend a bias feature per partition that accumulates the stats for all
      # examples in that partition.
      # Bias is added to the stats even if there are no examples with values in
      # the current sparse column. The reason is that the other example batches
      # might have values in these partitions so we have to keep the bias
      # updated.
      bias_feature_ids = array_ops.fill(
          array_ops.shape(unique_partitions), _BIAS_FEATURE_ID)
      bias_feature_ids = math_ops.cast(bias_feature_ids, dtypes.int64)
      partition_ids = array_ops.concat(
          [unique_partitions, filtered_partition_ids], 0)
      filtered_gradients = array_ops.concat(
          [per_partition_gradients, filtered_gradients], 0)
      filtered_hessians = array_ops.concat(
          [per_partition_hessians, filtered_hessians], 0)
      feature_ids = array_ops.concat(
          [bias_feature_ids, self._sparse_int_column.values], 0)
      # Dimension is always zero for sparse int features.
      dimension_ids = array_ops.zeros_like(feature_ids, dtype=dtypes.int64)
      feature_ids_and_dimensions = array_ops.stack(
          [feature_ids, dimension_ids], axis=1)
      return (partition_ids, feature_ids_and_dimensions, filtered_gradients,
              filtered_hessians)

  def _aggregate_sparse_grad(self, grad, var, train_ops):
    """Aggregate sparse gradients.

    Args:
      grad: The sparse gradient to aggregate.
      var: The variable to apply this gradient to.
      train_ops: The train_ops for the worker to run.

    Returns:
      aggregated_grad: Aggregated grad.
    """
    # Sparse gradients have to be inserted as one pair of (value,
    # indice) as an element instead of the whole "indexedslice" because
    # their shapes are not deterministic.
    sparse_grad_queue = (data_flow_ops.FIFOQueue(
        -1,
        (grad.values.dtype, grad.indices.dtype),
        shapes=(var.get_shape().as_list()[1:], ()),
        shared_name="sparse_grad_q_%s" % var.name))
    self._sparse_grad_queues_and_devs.append((sparse_grad_queue, var.device))

    # Sparse token is inserted after the "enqueue_many" finishes. This
    # is needed to make sure enough sparse gradients have been enqueued
    # before applying them to the variables.
    sparse_token_queue = (data_flow_ops.FIFOQueue(
        self._replicas_to_aggregate * 2,
        types_pb2.DT_INT32,
        shapes=(),
        shared_name="sparse_token_q_%s" % var.name))
    self._one_element_queue_list.append((sparse_token_queue, var.device))

    enqueue_spares_op = sparse_grad_queue.enqueue_many([grad.values,
                                                        grad.indices])
    with ops.control_dependencies([enqueue_spares_op]):
      train_ops.append(sparse_token_queue.enqueue((1,)))

    with ops.control_dependencies([sparse_token_queue.dequeue_many(
        self._replicas_to_aggregate)]):
      values, indices = sparse_grad_queue.dequeue_many(sparse_grad_queue.size())
      concat_grad = ops.IndexedSlices(values, indices, grad.dense_shape)

      # Sum the gradients of the same variables in the sparse layers so
      # that each variable is only updated once. Note that with 2
      # gradients g1 and g2 from 2 replicas for the same variable,
      # apply(g1+g2) is different from apply(g1) and then apply(g2) when
      # the optimizer is complex like Momentum or Adagrad.
      values = concat_grad.values
      indices = concat_grad.indices
      new_indices, indx = array_ops.unique(indices)
      num_indices = array_ops.shape(new_indices)[0]
      sum_values = math_ops.unsorted_segment_sum(values, indx, num_indices)
      return ops.IndexedSlices(sum_values, new_indices, concat_grad.dense_shape)

  def quantiles_ready():
    """The subgraph for when the quantiles are ready."""
    quantized_feature = quantile_ops.quantiles([], [sparse_column_values], [],
                                               [quantile_buckets],
                                               [sparse_column_indices])

    quantized_feature = math_ops.cast(quantized_feature[1], dtypes.int64)
    quantized_feature = array_ops.squeeze(quantized_feature, axis=0)

    example_indices, _ = array_ops.split(
        sparse_column_indices, num_or_size_splits=2, axis=1)
    example_indices = array_ops.squeeze(example_indices, [1])
    filtered_gradients = array_ops.gather(gradients, example_indices)
    filtered_hessians = array_ops.gather(hessians, example_indices)
    filtered_partition_ids = array_ops.gather(example_partition_ids,
                                              example_indices)
    unique_partitions, mapped_partitions = array_ops.unique(
        example_partition_ids)

    # Compute aggregate stats for each partition.
    # Since unsorted_segment_sum can be numerically unstable, use 64bit
    # operation.
    gradients64 = math_ops.cast(gradients, dtypes.float64)
    hessians64 = math_ops.cast(hessians, dtypes.float64)
    per_partition_gradients = math_ops.unsorted_segment_sum(
        gradients64, mapped_partitions, array_ops.size(unique_partitions))
    per_partition_hessians = math_ops.unsorted_segment_sum(
        hessians64, mapped_partitions, array_ops.size(unique_partitions))
    per_partition_gradients = math_ops.cast(per_partition_gradients,
                                            dtypes.float32)
    per_partition_hessians = math_ops.cast(per_partition_hessians,
                                           dtypes.float32)
    # Prepend a bias feature per partition that accumulates the stats for all
    # examples in that partition.
    bias_feature_ids = array_ops.fill(
        array_ops.shape(unique_partitions), _BIAS_FEATURE_ID)
    bias_feature_ids = math_ops.cast(bias_feature_ids, dtypes.int64)
    zeros = array_ops.zeros_like(bias_feature_ids)
    bias_feature_ids = array_ops.stack([bias_feature_ids, zeros], axis=1)

    partition_ids = array_ops.concat(
        [unique_partitions, filtered_partition_ids], 0)
    filtered_gradients = array_ops.concat(
        [per_partition_gradients, filtered_gradients], 0)
    filtered_hessians = array_ops.concat(
        [per_partition_hessians, filtered_hessians], 0)

    bucket_ids = array_ops.concat([bias_feature_ids, quantized_feature], 0)

    return partition_ids, bucket_ids, filtered_gradients, filtered_hessians

def update_all_medoids(pairwise_distances, predictions, labels, chosen_ids,
                       margin_multiplier, margin_type):
  """Updates all cluster medoids a cluster at a time.

  Args:
    pairwise_distances: 2-D Tensor of pairwise distances.
    predictions: 1-D Tensor of predicted cluster assignment.
    labels: 1-D Tensor of ground truth cluster assignment.
    chosen_ids: 1-D Tensor of cluster centroid indices.
    margin_multiplier: multiplication constant.
    margin_type: Type of structured margin to use. Default is nmi.

  Returns:
    chosen_ids: Updated 1-D Tensor of cluster centroid indices.
  """

  def func_cond_augmented_pam(iteration, chosen_ids):
    del chosen_ids  # Unused argument.
    return iteration < num_classes

  def func_body_augmented_pam(iteration, chosen_ids):
    """Call the update_medoid_per_cluster subroutine."""
    mask = math_ops.equal(
        math_ops.cast(predictions, dtypes.int64),
        math_ops.cast(iteration, dtypes.int64))
    this_cluster_ids = array_ops.where(mask)

    pairwise_distances_subset = array_ops.transpose(
        array_ops.gather(
            array_ops.transpose(
                array_ops.gather(pairwise_distances, this_cluster_ids)),
            this_cluster_ids))

    chosen_ids = update_medoid_per_cluster(pairwise_distances,
                                           pairwise_distances_subset, labels,
                                           chosen_ids, this_cluster_ids,
                                           iteration, margin_multiplier,
                                           margin_type)
    return iteration + 1, chosen_ids

  unique_class_ids = array_ops.unique(labels)[0]
  num_classes = array_ops.size(unique_class_ids)
  iteration = array_ops.constant(0)

  _, chosen_ids = control_flow_ops.while_loop(
      func_cond_augmented_pam, func_body_augmented_pam, [iteration, chosen_ids])
  return chosen_ids

def _deduplicate_indexed_slices(values, indices):
  """Sums `values` associated with any non-unique `indices`.

  Args:
    values: A `Tensor` with rank >= 1.
    indices: A one-dimensional integer `Tensor`, indexing into the first
      dimension of `values` (as in an IndexedSlices object).
  Returns:
    A tuple of (`summed_values`, `unique_indices`) where `unique_indices` is a
    de-duplicated version of `indices` and `summed_values` contains the sum of
    `values` slices associated with each unique index.
  """
  unique_indices, new_index_positions = array_ops.unique(indices)
  summed_values = math_ops.unsorted_segment_sum(
      values, new_index_positions,
      array_ops.shape(unique_indices)[0])
  return (summed_values, unique_indices)

def compute_augmented_facility_locations(pairwise_distances, labels, all_ids,
                                         margin_multiplier, margin_type):
  """Computes the centroid locations.

  Args:
    pairwise_distances: 2-D Tensor of pairwise distances.
    labels: 1-D Tensor of ground truth cluster assignment.
    all_ids: 1-D Tensor of all data indices.
    margin_multiplier: multiplication constant.
    margin_type: Type of structured margin to use. Default is nmi.

  Returns:
    chosen_ids: 1-D Tensor of chosen centroid indices.
  """

  def func_cond_augmented(iteration, chosen_ids):
    del chosen_ids  # Unused argument in func_cond_augmented.
    return iteration < num_classes

  def func_body_augmented(iteration, chosen_ids):
    # find a new facility location to add
    #  based on the clustering score and the NMI score
    candidate_ids = array_ops.setdiff1d(all_ids, chosen_ids)[0]
    new_chosen_idx = _find_loss_augmented_facility_idx(pairwise_distances,
                                                       labels, chosen_ids,
                                                       candidate_ids,
                                                       margin_multiplier,
                                                       margin_type)
    chosen_ids = array_ops.concat([chosen_ids, [new_chosen_idx]], 0)
    return iteration + 1, chosen_ids

  num_classes = array_ops.size(array_ops.unique(labels)[0])
  chosen_ids = array_ops.constant(0, dtype=dtypes.int32, shape=[0])

  # num_classes get determined at run time based on the sampled batch.
  iteration = array_ops.constant(0)

  _, chosen_ids = control_flow_ops.while_loop(
      func_cond_augmented,
      func_body_augmented, [iteration, chosen_ids],
      shape_invariants=[iteration.get_shape(), tensor_shape.TensorShape(
          [None])])
  return chosen_ids