Python tensorflow.python.ops.tensor_array_ops.TensorArray() Examples

def _testTensorArrayUnpackDynamic(self, legacy):
    with self.test_session(use_gpu=self._use_gpu) as sess:
      ta = tensor_array_ops.TensorArray(dtype=tf.float32, size=3,
                                        dynamic_size=True)
      x = tf.constant([1.0, 2.0, 3.0])
      if legacy:
        w0 = ta._legacy_unpack(x)
      else:
        w0 = ta.unpack(x)
      w1 = w0.write(3, 4.0)
      if legacy:
        r = w1._legacy_pack()
      else:
        r = w1.pack()
      self.assertAllEqual(np.array([1.0, 2.0, 3.0, 4.0]), r.eval())
      grad = tf.gradients(ys=[r], xs=[x])
      self.assertAllEqual(np.array([1.0, 1.0, 1.0]),
                          sess.run(grad)[0]) 

def _TensorArraySplitGrad(op, flow):
  """Gradient for TensorArraySplit.

  Args:
    op: Forward TensorArraySplit op.
    flow: Gradient `Tensor` flow to TensorArraySplit.

  Returns:
    A grad `Tensor`, the gradient created in upstream ReadGrads or PackGrad.
  """
  handle = op.inputs[0]
  dtype = op.get_attr("T")
  grad_source = _GetGradSource(flow)
  g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
                                    colocate_with_first_write_call=False)
       .grad(source=grad_source, flow=flow))
  grad = g.concat()
  # handle, value, lengths, flow_in
  return [None, grad, None, flow] 

def create_array(self, stride):
    """Creates a new tensor array to store this layer's activations.

    Arguments:
      stride: Possibly dynamic batch * beam size with which to initialize the
        tensor array

    Returns:
      TensorArray object
    """
    check.Gt(self.dim, 0, 'Cannot create array when dimension is dynamic')
    tensor_array = ta.TensorArray(dtype=tf.float32,
                                  size=0,
                                  dynamic_size=True,
                                  clear_after_read=False,
                                  infer_shape=False,
                                  name='%s_array' % self.name)

    # Start each array with all zeros. Special values will still be learned via
    # the extra embedding dimension stored for each linked feature channel.
    initial_value = tf.zeros([stride, self.dim])
    return tensor_array.write(0, initial_value) 

def _TensorArrayScatterGrad(op, flow):
  """Gradient for TensorArrayScatter.

  Args:
    op: Forward TensorArrayScatter op.
    flow: Gradient `Tensor` flow to TensorArrayScatter.

  Returns:
    A grad `Tensor`, the gradient created in upstream ReadGrads or PackGrad.
  """
  handle = op.inputs[0]
  indices = op.inputs[1]
  dtype = op.get_attr("T")
  grad_source = _GetGradSource(flow)
  g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
                                    colocate_with_first_write_call=False)
       .grad(source=grad_source, flow=flow))
  grad = g.gather(indices)
  return [None, None, grad, flow] 

def _maybe_merge_batch_beams(self, t, s):
    """Splits the tensor from a batch by beams into a batch of beams.

    More exactly, t is a tensor of dimension [batch_size*beam_width, s]. We
    reshape this into [batch_size, beam_width, s]

    Args:
      t: Tensor of dimension [batch_size*beam_width, s]
      s: Tensor, Python int, or TensorShape.

    Returns:
      A reshaped version of t with dimension [batch_size, beam_width, s].

    Raises:
      TypeError: If t is an instance of TensorArray.
      ValueError:  If the rank of t is not statically known.
    """
    _check_maybe(t)
    if t.shape.ndims >= 2:
      return self._merge_batch_beams(t, s)
    else:
      return t 

def _maybe_split_batch_beams(self, t, s):
    """Maybe splits the tensor from a batch by beams into a batch of beams.

    We do this so that we can use nest and not run into problems with shapes.

    Args:
      t: Tensor of dimension [batch_size*beam_width, s]
      s: Tensor, Python int, or TensorShape.

    Returns:
      Either a reshaped version of t with dimension
      [batch_size, beam_width, s] if t's first dimension is of size
      batch_size*beam_width or t if not.

    Raises:
      TypeError: If t is an instance of TensorArray.
      ValueError: If the rank of t is not statically known.
    """
    _check_maybe(t)
    if t.shape.ndims >= 1:
      return self._split_batch_beams(t, s)
    else:
      return t 

def _TensorArrayWriteGrad(op, flow):
  """Gradient for TensorArrayWrite.

  Args:
    op: Forward TensorArrayWrite op.
    flow: Gradient `Tensor` flow to TensorArrayWrite.

  Returns:
    A grad `Tensor`, the gradient created in an upstream ReadGrad or PackGrad.
  """
  # handle is the output store_handle of TensorArrayReadGrad or
  # the handle output of TensorArrayWriteGrad.  we must use this one.
  handle = op.inputs[0]
  index = op.inputs[1]
  dtype = op.get_attr("T")
  grad_source = _GetGradSource(flow)
  g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
                                    colocate_with_first_write_call=False)
       .grad(source=grad_source, flow=flow))
  grad = g.read(index)
  return [None, None, grad, flow] 

def create(self,
             fixed_embeddings,
             linked_embeddings,
             context_tensor_arrays,
             attention_tensor,
             during_training,
             stride=None):
    """Constructs a feed-forward unit based on the features and context tensors.

    Args:
      fixed_embeddings: list of NamedTensor objects
      linked_embeddings: list of NamedTensor objects
      context_tensor_arrays: optional list of TensorArray objects used for
          implicit recurrence.
      attention_tensor: optional Tensor used for attention.
      during_training: whether to create a network for training (vs inference).
      stride: int scalar tensor containing the stride required for
          bulk computation.

    Returns:
      A list of tensors corresponding to the list of layers.
    """
    pass 

def _TensorArrayWriteGrad(op, flow):
  """Gradient for TensorArrayWrite.

  Args:
    op: Forward TensorArrayWrite op.
    flow: Gradient `Tensor` flow to TensorArrayWrite.

  Returns:
    A grad `Tensor`, the gradient created in an upstream ReadGrad or PackGrad.
  """
  # handle is the output store_handle of TensorArrayReadGrad or
  # the handle output of TensorArrayWriteGrad.  we must use this one.
  handle = op.inputs[0]
  index = op.inputs[1]
  dtype = op.get_attr("T")
  grad_source = _GetGradSource(flow)
  g = tensor_array_ops.TensorArray(
      dtype=dtype, handle=handle, flow=flow).grad(
          source=grad_source, flow=flow)
  grad = g.read(index)
  return [None, None, grad, flow] 

def _TensorArrayScatterGrad(op, flow):
  """Gradient for TensorArrayScatter.

  Args:
    op: Forward TensorArrayScatter op.
    flow: Gradient `Tensor` flow to TensorArrayScatter.

  Returns:
    A grad `Tensor`, the gradient created in upstream ReadGrads or PackGrad.
  """
  handle = op.inputs[0]
  indices = op.inputs[1]
  dtype = op.get_attr("T")
  grad_source = _GetGradSource(flow)
  g = tensor_array_ops.TensorArray(
      dtype=dtype, handle=handle, flow=flow).grad(
          source=grad_source, flow=flow)
  grad = g.gather(indices)
  return [None, None, grad, flow] 

def _TensorArraySplitGrad(op, flow):
  """Gradient for TensorArraySplit.

  Args:
    op: Forward TensorArraySplit op.
    flow: Gradient `Tensor` flow to TensorArraySplit.

  Returns:
    A grad `Tensor`, the gradient created in upstream ReadGrads or PackGrad.
  """
  handle = op.inputs[0]
  dtype = op.get_attr("T")
  grad_source = _GetGradSource(flow)
  g = tensor_array_ops.TensorArray(
      dtype=dtype, handle=handle, flow=flow).grad(
          source=grad_source, flow=flow)
  grad = g.concat()
  # handle, value, lengths, flow_in
  return [None, grad, None, flow] 

def normalize_tensors(tensors):
  """Converts a nested structure of tensor-like objects to tensors.
  * `SparseTensor`-like inputs are converted to `SparseTensor`.
  * `TensorArray` inputs are passed through.
  * Everything else is converted to a dense `Tensor`.
  Args:
    tensors: A nested structure of tensor-like, list,
      `SparseTensor`, `SparseTensorValue`, or `TensorArray` objects.
  Returns:
    A nested structure of tensor, `SparseTensor`, or `TensorArray` objects.
  """
  flat_tensors = nest.flatten(tensors)
  prepared = []
  with ops.name_scope("normalize_tensors"):
    for i, t in enumerate(flat_tensors):
      if sparse_tensor_lib.is_sparse(t):
        prepared.append(sparse_tensor_lib.SparseTensor.from_value(t))
      elif isinstance(t, tensor_array_ops.TensorArray):
        prepared.append(t)
      else:
        prepared.append(ops.convert_to_tensor(t, name="component_%d" % i))
  return nest.pack_sequence_as(tensors, prepared) 

def convert_network_state_tensorarray(tensorarray):
  """Converts a source TensorArray to a source Tensor.

  Performs a permutation between the steps * [stride, D] shape of a
  source TensorArray and the (flattened) [stride * steps, D] shape of
  a source Tensor.

  The TensorArrays used during recurrence have an additional zeroth step that
  needs to be removed.

  Args:
    tensorarray: TensorArray object to be converted.

  Returns:
    Tensor object after conversion.
  """
  tensor = tensorarray.stack()  # Results in a [steps, stride, D] tensor.
  tensor = tf.slice(tensor, [1, 0, 0], [-1, -1, -1])  # Lop off the 0th step.
  tensor = tf.transpose(tensor, [1, 0, 2])  # Switch steps and stride.
  return tf.reshape(tensor, [-1, tf.shape(tensor)[2]]) 

def create_array(self, stride):
    """Creates a new tensor array to store this layer's activations.

    Arguments:
      stride: Possibly dynamic batch * beam size with which to initialize the
        tensor array

    Returns:
      TensorArray object
    """
    check.Gt(self.dim, 0, 'Cannot create array when dimension is dynamic')
    tensor_array = ta.TensorArray(
        dtype=tf.float32,
        size=0,
        dynamic_size=True,
        clear_after_read=False,
        infer_shape=False,
        name='%s_array' % self.name)

    # Start each array with all zeros. Special values will still be learned via
    # the extra embedding dimension stored for each linked feature channel.
    initial_value = tf.zeros([stride, self.dim])
    return tensor_array.write(0, initial_value) 

def create(self,
             fixed_embeddings,
             linked_embeddings,
             context_tensor_arrays,
             attention_tensor,
             during_training,
             stride=None):
    """Constructs a feed-forward unit based on the features and context tensors.

    Args:
      fixed_embeddings: list of NamedTensor objects
      linked_embeddings: list of NamedTensor objects
      context_tensor_arrays: optional list of TensorArray objects used for
          implicit recurrence.
      attention_tensor: optional Tensor used for attention.
      during_training: whether to create a network for training (vs inference).
      stride: int scalar tensor containing the stride required for
          bulk computation.

    Returns:
      A list of tensors corresponding to the list of layers.
    """
    pass 

def convert_network_state_tensorarray(tensorarray):
  """Converts a source TensorArray to a source Tensor.

  Performs a permutation between the steps * [stride, D] shape of a
  source TensorArray and the (flattened) [stride * steps, D] shape of
  a source Tensor.

  The TensorArrays used during recurrence have an additional zeroth step that
  needs to be removed.

  Args:
    tensorarray: TensorArray object to be converted.

  Returns:
    Tensor object after conversion.
  """
  tensor = tensorarray.stack()  # Results in a [steps, stride, D] tensor.
  tensor = tf.slice(tensor, [1, 0, 0], [-1, -1, -1])  # Lop off the 0th step.
  tensor = tf.transpose(tensor, [1, 0, 2])  # Switch steps and stride.
  return tf.reshape(tensor, [-1, tf.shape(tensor)[2]]) 

def create_array(self, stride):
    """Creates a new tensor array to store this layer's activations.

    Arguments:
      stride: Possibly dynamic batch * beam size with which to initialize the
        tensor array

    Returns:
      TensorArray object
    """
    check.Gt(self.dim, 0, 'Cannot create array when dimension is dynamic')
    tensor_array = ta.TensorArray(
        dtype=tf.float32,
        size=0,
        dynamic_size=True,
        clear_after_read=False,
        infer_shape=False,
        name='%s_array' % self.name)

    # Start each array with all zeros. Special values will still be learned via
    # the extra embedding dimension stored for each linked feature channel.
    initial_value = tf.zeros([stride, self.dim])
    return tensor_array.write(0, initial_value) 

def create(self,
             fixed_embeddings,
             linked_embeddings,
             context_tensor_arrays,
             attention_tensor,
             during_training,
             stride=None):
    """Constructs a feed-forward unit based on the features and context tensors.

    Args:
      fixed_embeddings: list of NamedTensor objects
      linked_embeddings: list of NamedTensor objects
      context_tensor_arrays: optional list of TensorArray objects used for
          implicit recurrence.
      attention_tensor: optional Tensor used for attention.
      during_training: whether to create a network for training (vs inference).
      stride: int scalar tensor containing the stride required for
          bulk computation.

    Returns:
      A list of tensors corresponding to the list of layers.
    """
    pass 

def testTensorArrayWriteRead(self):
    with self.test_session(use_gpu=self._use_gpu) as session:
      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=3, infer_shape=False)

      w0 = ta.write(0, [[4.0, 5.0]])
      w1 = w0.write(1, [[1.0]])
      w2 = w1.write(2, -3.0)

      r0 = w2.read(0)
      r1 = w2.read(1)
      r2 = w2.read(2)

      d0, d1, d2 = session.run([r0, r1, r2])
      self.assertAllEqual([[4.0, 5.0]], d0)
      self.assertAllEqual([[1.0]], d1)
      self.assertAllEqual(-3.0, d2) 

def _testTensorArrayWritePack(self, tf_dtype, legacy):
    dtype = tf_dtype.as_numpy_dtype()
    with self.test_session(use_gpu=self._use_gpu):
      ta = tensor_array_ops.TensorArray(
          dtype=tf_dtype, tensor_array_name="foo", size=3)

      if tf_dtype == tf.string:
        # In Python3, np.str is unicode, while we always want bytes
        convert = lambda x: np.asarray(x).astype("|S")
      else:
        convert = lambda x: np.asarray(x).astype(dtype)

      w0 = ta.write(0, convert([[4.0, 5.0]]))
      w1 = w0.write(1, convert([[6.0, 7.0]]))
      w2 = w1.write(2, convert([[8.0, 9.0]]))

      if legacy:
        c0 = w2._legacy_pack()
      else:
        c0 = w2.pack()

      self.assertAllEqual(
          convert([[[4.0, 5.0]], [[6.0, 7.0]], [[8.0, 9.0]]]), c0.eval()) 

def _testTensorArrayWriteConcat(self, tf_dtype):
    dtype = tf_dtype.as_numpy_dtype()
    with self.test_session(use_gpu=self._use_gpu):
      ta = tensor_array_ops.TensorArray(
          dtype=tf_dtype, tensor_array_name="foo", size=3, infer_shape=False)

      if tf_dtype == tf.string:
        # In Python3, np.str is unicode, while we always want bytes
        convert = lambda x: np.asarray(x).astype("|S")
      else:
        convert = lambda x: np.asarray(x).astype(dtype)

      w0 = ta.write(0, convert([[4.0, 5.0], [104.0, 105.0], [204.0, 205.0]]))
      w1 = w0.write(1, convert([[6.0, 7.0], [106.0, 107.0]]))
      w2 = w1.write(2, convert([[8.0, 9.0]]))

      c0 = w2.concat()

      self.assertAllEqual(
          convert([[4.0, 5.0],
                   [104.0, 105.0],
                   [204.0, 205.0],
                   [6.0, 7.0],
                   [106.0, 107.0],
                   [8.0, 9.0]]), c0.eval()) 

def testTensorArrayWriteWrongIndexOrDataTypeFails(self):
    with self.test_session(use_gpu=self._use_gpu):
      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=3)

      # Test writing the wrong datatype
      with self.assertRaisesOpError(
          "TensorArray dtype is float but Op is trying to write dtype string"):
        ta.write(-1, "wrong_type_scalar").flow.eval()

      # Test writing to a negative index
      with self.assertRaisesOpError(
          "Tried to write to index -1 but array is not "
          "resizeable and size is: 3"):
        ta.write(-1, 3.0).flow.eval()

      # Test reading from too large an index
      with self.assertRaisesOpError(
          "Tried to write to index 3 but array is not "
          "resizeable and size is: 3"):
        ta.write(3, 3.0).flow.eval() 

def testTensorArrayConcatIncompatibleShapesFails(self):
    with self.test_session(use_gpu=self._use_gpu):
      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=3, infer_shape=False)

      w1 = ta.write(0, 3.0)
      w2 = w1.write(1, 4.0)
      w3 = w2.write(2, [3.0])

      with self.assertRaisesOpError(
          "Concat saw a scalar shape at index 0 but requires at least vectors"):
        w3.concat().eval()

      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=3, infer_shape=False)

      w1 = ta.write(0, [3.0])
      w2 = w1.write(1, [4.0])
      w3 = w2.write(2, [[3.0]])

      with self.assertRaisesOpError(
          r"TensorArray has inconsistent shapes.  Index 0 has "
          r"\(excepting dimension 0\) shape: \[\] but index 2 has \(excepting "
          r"dimension 0\) shape: \[1\]"):
        w3.concat().eval() 

def _testTensorArrayGradientUnpackRead(self, legacy):
    with self.test_session(use_gpu=self._use_gpu) as session:
      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=2,
          clear_after_read=False)

      value = tf.constant([[1.0, -1.0], [10.0, -10.0]])

      if legacy:
        w = ta._legacy_unpack(value)
      else:
        w = ta.unpack(value)
      r0 = w.read(0)
      r0_1 = w.read(0)
      r1 = w.read(1)

      # Test combined gradients + aggregation of read(0)
      grad = tf.gradients(
          ys=[r0, r0_1, r1], xs=[value],
          grad_ys=[[2.0, 3.0], [-1.5, 1.5], [4.0, 5.0]])
      grad_vals = session.run(grad)

      self.assertEqual(len(grad_vals), 1)
      self.assertAllEqual([[2.0 - 1.5, 3.0 + 1.5], [4.0, 5.0]], grad_vals[0]) 

def testTensorArrayGradientSplitConcat(self):
    with self.test_session(use_gpu=self._use_gpu) as session:
      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=2)

      value = tf.constant([[1.0, -1.0], [10.0, -10.0], [100.0, -100.0]])

      w = ta.split(value, [2, 1])
      r = w.concat()

      # Test combined gradients
      grad = tf.gradients(
          ys=[r], xs=[value],
          grad_ys=[[[2.0, -2.0], [20.0, -20.0], [200.0, -200.0]]])
      grad_vals = session.run(grad)

      self.assertEqual(len(grad_vals), 1)
      self.assertAllEqual(
          [[2.0, -2.0], [20.0, -20.0], [200.0, -200.0]], grad_vals[0]) 

def testSumOfTwoReadVariablesWithoutRepeatGrad(self):
    with self.test_session(use_gpu=self._use_gpu) as session:
      a = tf.identity(np.arange(3*5, dtype=np.float32).reshape(3, 5) + 1)
      b = tf.identity(np.arange(3*5, dtype=np.float32).reshape(3, 5) + 1 + 3*5)
      ta = tensor_array_ops.TensorArray(dtype=tf.float32, size=2)
      ta = ta.write(0, a, name="write_a")
      ta = ta.write(1, b, name="write_b")
      c = (ta.read(0, name="read_a_0") +  # a + b
           ta.read(1, name="read_b_0"))
      g0 = -(np.arange(3*5, dtype=np.float32).reshape(3, 5) + 1)
      grad_a = tf.gradients([c], [a], [g0])[0]  # d(a+b)/da = 1
      grad_b = tf.gradients([c], [b], [g0])[0]  # d(a+b)/db = 1

      # Test gradients calculated individually
      grad_a_t, = session.run([grad_a])
      self.assertAllEqual(grad_a_t, g0)

      grad_b_t, = session.run([grad_b])
      self.assertAllEqual(grad_b_t, g0)

      # Test gradients calculated jointly
      joint_grad_a_t, joint_grad_b_t = session.run([grad_a, grad_b])
      self.assertAllEqual(joint_grad_a_t, g0)
      self.assertAllEqual(joint_grad_b_t, g0) 

def testWriteShape(self):
    with self.test_session():
      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=3)
      c0 = tf.constant([4.0, 5.0])
      w0 = ta.write(0, c0)
      r0 = w0.read(0)
      self.assertAllEqual(c0.get_shape(), r0.get_shape())

      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=3)
      c1 = tf.constant([6.0, 7.0])
      w1 = w0.write(1, c1)
      r0 = w1.read(0)
      r1 = w1.read(1)
      self.assertAllEqual(c0.get_shape(), r0.get_shape())
      self.assertAllEqual(c1.get_shape(), r1.get_shape())

      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=3)
      c2 = tf.constant([4.0, 5.0, 6.0])
      with self.assertRaises(ValueError):
        w0.write(0, c2) 

def testSplitShape(self):
    with self.test_session():
      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo",
          size=0, dynamic_size=True, infer_shape=True)
      value = tf.constant([[1.0, -1.0], [2.0, -2.0], [3.0, -3.0]])
      w0 = ta.split(value, [1, 1, 1])
      r0 = w0.read(0)
      self.assertAllEqual((1, 2), r0.get_shape())

      ta1 = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo1",
          size=0, dynamic_size=True, infer_shape=True)
      w0 = ta1.split(value, [1, 2])
      r0 = w0.read(0)
      self.assertAllEqual(r0.get_shape(), tensor_shape.unknown_shape()) 

def convert_network_state_tensorarray(tensorarray):
  """Converts a source TensorArray to a source Tensor.

  Performs a permutation between the steps * [stride, D] shape of a
  source TensorArray and the (flattened) [stride * steps, D] shape of
  a source Tensor.

  The TensorArrays used during recurrence have an additional zeroth step that
  needs to be removed.

  Args:
    tensorarray: TensorArray object to be converted.

  Returns:
    Tensor object after conversion.
  """
  tensor = tensorarray.stack()  # Results in a [steps, stride, D] tensor.
  tensor = tf.slice(tensor, [1, 0, 0], [-1, -1, -1])  # Lop off the 0th step.
  tensor = tf.transpose(tensor, [1, 0, 2])  # Switch steps and stride.
  return tf.reshape(tensor, [-1, tf.shape(tensor)[2]]) 

def testTensorArrayWriteMultipleFails(self):
    with self.test_session(use_gpu=self._use_gpu):
      ta = tensor_array_ops.TensorArray(
          dtype=tf.float32, tensor_array_name="foo", size=3)

      with self.assertRaisesOpError(
          "Could not write to TensorArray index 2 because "
          "it has already been written to."):
        ta.write(2, 3.0).write(2, 3.0).flow.eval()