Python tensorflow.python.framework.dtypes.float16() Examples

def add_check_numerics_ops():
  """Connect a `check_numerics` to every floating point tensor.

  `check_numerics` operations themselves are added for each `half`, `float`,
  or `double` tensor in the graph. For all ops in the graph, the
  `check_numerics` op for all of its (`half`, `float`, or `double`) inputs
  is guaranteed to run before the `check_numerics` op on any of its outputs.

  Returns:
    A `group` op depending on all `check_numerics` ops added.
  """
  check_op = []
  # This code relies on the ordering of ops in get_operations().
  # The producer of a tensor always comes before that tensor's consumer in
  # this list. This is true because get_operations() returns ops in the order
  # added, and an op can only be added after its inputs are added.
  for op in ops.get_default_graph().get_operations():
    for output in op.outputs:
      if output.dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
        message = op.name + ":" + str(output.value_index)
        with ops.control_dependencies(check_op):
          check_op = [array_ops.check_numerics(output, message=message)]
  return control_flow_ops.group(*check_op) 

def add_check_numerics_ops():
  """Connect a `check_numerics` to every floating point tensor.

  `check_numerics` operations themselves are added for each `half`, `float`,
  or `double` tensor in the graph. For all ops in the graph, the
  `check_numerics` op for all of its (`half`, `float`, or `double`) inputs
  is guaranteed to run before the `check_numerics` op on any of its outputs.

  Returns:
    A `group` op depending on all `check_numerics` ops added.
  """
  check_op = []
  # This code relies on the ordering of ops in get_operations().
  # The producer of a tensor always comes before that tensor's consumer in
  # this list. This is true because get_operations() returns ops in the order
  # added, and an op can only be added after its inputs are added.
  for op in ops.get_default_graph().get_operations():
    for output in op.outputs:
      if output.dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
        message = op.name + ":" + str(output.value_index)
        with ops.control_dependencies(check_op):
          check_op = [array_ops.check_numerics(output, message=message)]
  return control_flow_ops.group(*check_op) 

def tanh(x, name=None):
  """Computes hyperbolic tangent of `x` element-wise.

  Args:
    x: A Tensor or SparseTensor with type `float16`, `float32`, `double`,
      `complex64`, or `complex128`.
    name: A name for the operation (optional).

  Returns:
    A Tensor or SparseTensor respectively with the same type as `x`.
  """
  with ops.name_scope(name, "Tanh", [x]) as name:
    if isinstance(x, sparse_tensor.SparseTensor):
      x_tanh = gen_math_ops._tanh(x.values, name=name)
      return sparse_tensor.SparseTensor(
          indices=x.indices, values=x_tanh, dense_shape=x.dense_shape)
    else:
      return gen_math_ops._tanh(x, name=name) 

def set_floatx(value):
  """Sets the default float type.

  Arguments:
      value: String; 'float16', 'float32', or 'float64'.

  Example:
  ```python
      >>> from keras import backend as K
      >>> K.floatx()
      'float32'
      >>> K.set_floatx('float16')
      >>> K.floatx()
      'float16'
  ```

  Raises:
      ValueError: In case of invalid value.
  """
  global _FLOATX
  if value not in {'float16', 'float32', 'float64'}:
    raise ValueError('Unknown floatx type: ' + str(value))
  _FLOATX = str(value) 

def testGradientTensor4D(self):
    for (data_format, use_gpu) in [("NHWC", False)]:
      for dtype in (dtypes.float16, dtypes.float32, dtypes.float64):
        np_input = np.arange(
            1.0, 49.0,
            dtype=dtype.as_numpy_dtype).reshape([2, 3, 4, 2]).astype(np.float32)
        bias = np.array([1.3, 2.4], dtype=dtype.as_numpy_dtype)
        self._testGradient(np_input, bias, dtype, data_format, use_gpu)
        np_input = np.arange(
            1.0, 513.0,
            dtype=dtype.as_numpy_dtype).reshape([64, 2, 2,
                                                 2]).astype(np.float32)
        self._testGradient(np_input, bias, dtype, data_format, use_gpu)
        np_input = np.arange(
            1.0, 513.0,
            dtype=dtype.as_numpy_dtype).reshape([2, 2, 2,
                                                 64]).astype(np.float32)
        self._testGradient(np_input,
                           np.random.rand(64).astype(dtype.as_numpy_dtype),
                           dtype, data_format, use_gpu) 

def set_floatx(value):
  """Sets the default float type.

  Arguments:
      value: String; 'float16', 'float32', or 'float64'.

  Example:
  ```python
      >>> from keras import backend as K
      >>> K.floatx()
      'float32'
      >>> K.set_floatx('float16')
      >>> K.floatx()
      'float16'
  ```

  Raises:
      ValueError: In case of invalid value.
  """
  global _FLOATX
  if value not in {'float16', 'float32', 'float64'}:
    raise ValueError('Unknown floatx type: ' + str(value))
  _FLOATX = str(value) 

def _convert_string_dtype(dtype):
  if dtype == 'float16':
    return dtypes_module.float16
  if dtype == 'float32':
    return dtypes_module.float32
  elif dtype == 'float64':
    return dtypes_module.float64
  elif dtype == 'int16':
    return dtypes_module.int16
  elif dtype == 'int32':
    return dtypes_module.int32
  elif dtype == 'int64':
    return dtypes_module.int64
  elif dtype == 'uint8':
    return dtypes_module.int8
  elif dtype == 'uint16':
    return dtypes_module.uint16
  else:
    raise ValueError('Unsupported dtype:', dtype) 

def add_check_numerics_ops():
  """Connect a `check_numerics` to every floating point tensor.

  `check_numerics` operations themselves are added for each `half`, `float`,
  or `double` tensor in the graph. For all ops in the graph, the
  `check_numerics` op for all of its (`half`, `float`, or `double`) inputs
  is guaranteed to run before the `check_numerics` op on any of its outputs.

  Returns:
    A `group` op depending on all `check_numerics` ops added.
  """
  check_op = []
  # This code relies on the ordering of ops in get_operations().
  # The producer of a tensor always comes before that tensor's consumer in
  # this list. This is true because get_operations() returns ops in the order
  # added, and an op can only be added after its inputs are added.
  for op in ops.get_default_graph().get_operations():
    for output in op.outputs:
      if output.dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
        message = op.name + ":" + str(output.value_index)
        with ops.control_dependencies(check_op):
          check_op = [array_ops.check_numerics(output, message=message)]
  return control_flow_ops.group(*check_op) 

def add_check_numerics_ops():
  """Connect a `check_numerics` to every floating point tensor.

  `check_numerics` operations themselves are added for each `half`, `float`,
  or `double` tensor in the graph. For all ops in the graph, the
  `check_numerics` op for all of its (`half`, `float`, or `double`) inputs
  is guaranteed to run before the `check_numerics` op on any of its outputs.

  Returns:
    A `group` op depending on all `check_numerics` ops added.
  """
  check_op = []
  # This code relies on the ordering of ops in get_operations().
  # The producer of a tensor always comes before that tensor's consumer in
  # this list. This is true because get_operations() returns ops in the order
  # added, and an op can only be added after its inputs are added.
  for op in ops.get_default_graph().get_operations():
    for output in op.outputs:
      if output.dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
        message = op.name + ":" + str(output.value_index)
        with ops.control_dependencies(check_op):
          check_op = [array_ops.check_numerics(output, message=message)]
  return control_flow_ops.group(*check_op) 

def testDeterministicGradients(self):
    with self.session(force_gpu=True):
      # There are problems with using force_gpu=True and cached_session with
      # both eager mode and graph mode in the same test. Using a non-cached
      # session and putting everything inside the same session context is
      # a compromise.
      for op_binding in (tf.nn.bias_add, nn.bias_add, nn_ops.bias_add):
        for data_layout in ('channels_first', 'channels_last'):
          # With the selected layer configuration, at least in TensorFlow
          # version 2.0, when data_layout='channels_last', bias_add operates
          # deterministically by default. I don't know if this is true for
          # all layer configurations. These cases are still being tested here,
          # for completeness.
          for data_rank in (1, 2, 3):
            for data_type in (dtypes.float16, dtypes.float32, dtypes.float64):
              self._testDeterministicGradientsCase(op_binding, data_layout,
                                                   data_rank, data_type) 

def _IsTrainable(tensor):
  dtype = dtypes.as_dtype(tensor.dtype)
  return dtype.base_dtype in (dtypes.float16, dtypes.float32, dtypes.float64,
                              dtypes.complex64, dtypes.complex128) 

def floatx():
  """Returns the default float type, as a string.

  E.g. 'float16', 'float32', 'float64'.

  Returns:
      String, the current default float type.

  Example:
  ```python
      >>> keras.backend.floatx()
      'float32'
  ```
  """
  return _FLOATX 

def ExtractBitsFromFloat16(x):
  return np.asscalar(np.asarray(x, dtype=np.float16).view(np.uint16)) 

def _CropAndResizeGrad(op, grad):
  """The derivatives for crop_and_resize.

  We back-propagate to the image only when the input image tensor has floating
  point dtype but we always back-propagate to the input boxes tensor.

  Args:
    op: The CropAndResize op.
    grad: The tensor representing the gradient w.r.t. the output.

  Returns:
    The gradients w.r.t. the input image, boxes, as well as the always-None
    gradients w.r.t. box_ind and crop_size.
  """
  image = op.inputs[0]
  if image.get_shape().is_fully_defined():
    image_shape = image.get_shape().as_list()
  else:
    image_shape = array_ops.shape(image)

  allowed_types = [dtypes.float16, dtypes.float32, dtypes.float64]
  if op.inputs[0].dtype in allowed_types:
    # pylint: disable=protected-access
    grad0 = gen_image_ops.crop_and_resize_grad_image(grad,
                                                     op.inputs[1],
                                                     op.inputs[2],
                                                     image_shape,
                                                     T=op.get_attr("T"))
    # pylint: enable=protected-access
  else:
    grad0 = None

  grad1 = gen_image_ops.crop_and_resize_grad_boxes(grad, op.inputs[0],
                                                   op.inputs[1], op.inputs[2])

  return [grad0, grad1, None, None] 

def _CastGrad(op, grad):
  t = [
      dtypes.float16, dtypes.float32, dtypes.float64, dtypes.bfloat16,
      dtypes.complex64, dtypes.complex128
  ]
  src_type = op.inputs[0].dtype.base_dtype
  dst_type = grad.dtype.base_dtype
  if src_type in t and dst_type in t:
    return math_ops.cast(grad, src_type)
  else:
    return None 

def _valid_dtypes(self):
    """Valid types for loss, variables and gradients.

    Subclasses should override to allow other float types.

    Returns:
      Valid types for loss, variables and gradients.
    """
    return set([dtypes.float16, dtypes.float32, dtypes.float64]) 

def test(self):
    cast_lt = ops.cast(self.original_lt, dtypes.float16)
    golden_lt = core.LabeledTensor(
        math_ops.cast(self.original_lt.tensor, dtypes.float16),
        self.original_lt.axes)
    self.assertLabeledTensorsEqual(cast_lt, golden_lt) 

def _dtypes_to_test(self):
    # TODO(langmore) Test tf.float16 once tf.matrix_solve works in 16bit.
    return [dtypes.float32, dtypes.float64, dtypes.complex64, dtypes.complex128] 

def _compute_gradient(x,
                      x_shape,
                      dx,
                      y,
                      y_shape,
                      dy,
                      x_init_value=None,
                      delta=1e-3,
                      extra_feed_dict=None):
  """Computes the theoretical and numerical jacobian."""
  t = dtypes.as_dtype(x.dtype)
  allowed_types = [dtypes.float16, dtypes.float32, dtypes.float64,
                   dtypes.complex64, dtypes.complex128]
  assert t.base_dtype in allowed_types, "Don't support type %s for x" % t.name
  t2 = dtypes.as_dtype(y.dtype)
  assert t2.base_dtype in allowed_types, "Don't support type %s for y" % t2.name

  if x_init_value is not None:
    i_shape = list(x_init_value.shape)
    assert(list(x_shape) == i_shape), "x_shape = %s, init_data shape = %s" % (
        x_shape, i_shape)
    x_data = x_init_value
  else:
    if t == dtypes.float16:
      dtype = np.float16
    elif t == dtypes.float32:
      dtype = np.float32
    else:
      dtype = np.float64
    x_data = np.asfarray(np.random.random_sample(x_shape), dtype=dtype)

  jacob_t = _compute_theoretical_jacobian(
      x, x_shape, x_data, dy, y_shape, dx, extra_feed_dict=extra_feed_dict)
  jacob_n = _compute_numeric_jacobian(
      x, x_shape, x_data, y, y_shape, delta, extra_feed_dict=extra_feed_dict)
  return jacob_t, jacob_n 

def test_name(self):
    cast_lt = ops.cast(self.original_lt, dtypes.float16)
    self.assertIn('lt_cast', cast_lt.name) 

def cast(x, dtype):
  """Casts a tensor to a different dtype and returns it.

  You can cast a Keras variable but it still returns a Keras tensor.

  Arguments:
      x: Keras tensor (or variable).
      dtype: String, either (`'float16'`, `'float32'`, or `'float64'`).

  Returns:
      Keras tensor with dtype `dtype`.

  Example:
  ```python
      >>> from keras import backend as K
      >>> input = K.placeholder((2, 3), dtype='float32')
      >>> input
      <tf.Tensor 'Placeholder_2:0' shape=(2, 3) dtype=float32>
      # It doesn't work in-place as below.
      >>> K.cast(input, dtype='float16')
      <tf.Tensor 'Cast_1:0' shape=(2, 3) dtype=float16>
      >>> input
      <tf.Tensor 'Placeholder_2:0' shape=(2, 3) dtype=float32>
      # you need to assign it.
      >>> input = K.cast(input, dtype='float16')
      >>> input
      <tf.Tensor 'Cast_2:0' shape=(2, 3) dtype=float16>
  ```
  """
  return math_ops.cast(x, dtype)


# UPDATES OPS 

def _convert_string_dtype(dtype):
  """Get the type from a string.

  Arguments:
      dtype: A string representation of a type.

  Returns:
      The type requested.

  Raises:
      ValueError: if `dtype` is not supported.
  """
  if dtype == 'float16':
    return dtypes_module.float16
  if dtype == 'float32':
    return dtypes_module.float32
  elif dtype == 'float64':
    return dtypes_module.float64
  elif dtype == 'int16':
    return dtypes_module.int16
  elif dtype == 'int32':
    return dtypes_module.int32
  elif dtype == 'int64':
    return dtypes_module.int64
  elif dtype == 'uint8':
    return dtypes_module.int8
  elif dtype == 'uint16':
    return dtypes_module.uint16
  else:
    raise ValueError('Unsupported dtype:', dtype) 

def add_check_numerics_ops():
  """Connect a `check_numerics` to every floating point tensor.

  `check_numerics` operations themselves are added for each `half`, `float`,
  or `double` tensor in the graph. For all ops in the graph, the
  `check_numerics` op for all of its (`half`, `float`, or `double`) inputs
  is guaranteed to run before the `check_numerics` op on any of its outputs.

  Note: This API is not compatible with the use of @{tf.cond} or
  @{tf.while_loop}, and will raise a `ValueError` if you attempt to call it
  in such a graph.

  Returns:
    A `group` op depending on all `check_numerics` ops added.

  Raises:
    ValueError: If the graph contains any numeric operations in a control flow
      structure.
  """
  check_op = []
  # This code relies on the ordering of ops in get_operations().
  # The producer of a tensor always comes before that tensor's consumer in
  # this list. This is true because get_operations() returns ops in the order
  # added, and an op can only be added after its inputs are added.
  for op in ops.get_default_graph().get_operations():
    for output in op.outputs:
      if output.dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
        if op._get_control_flow_context() is not None:  # pylint: disable=protected-access
          raise ValueError("`tf.add_check_numerics_ops() is not compatible "
                           "with TensorFlow control flow operations such as "
                           "`tf.cond()` or `tf.while_loop()`.")

        message = op.name + ":" + str(output.value_index)
        with ops.control_dependencies(check_op):
          check_op = [array_ops.check_numerics(output, message=message)]
  return control_flow_ops.group(*check_op) 

def initialize(self, name=None):
    """Initialize the decoder.

    Args:
      name: Name scope for any created operations.

    Returns:
      `(finished, start_inputs, initial_state)`.
    """
    finished, start_inputs = self._finished, self._start_inputs
    dtype = nest.flatten(self._initial_cell_state)[0].dtype
    log_probs = array_ops.one_hot(  # shape(batch_sz, beam_sz)
        array_ops.zeros([self._batch_size], dtype=dtypes.int32),
        depth=self._beam_width,
        on_value=math_ops.cast(0.0, dtype),
        off_value=-np.float16('inf') if dtype == dtypes.float16 else -np.Inf,
        dtype=dtype)

    initial_state = BeamSearchDecoderState(
        cell_state=self._initial_cell_state,
        log_probs=log_probs,
        finished=finished,
        lengths=array_ops.zeros(
            [self._batch_size, self._beam_width], dtype=dtypes.int64))

    return (finished, start_inputs, initial_state) 

def testFromCSVWithFeatureSpec(self):
    if not HAS_PANDAS:
      return
    num_batches = 100
    batch_size = 8

    data_path = _make_test_csv_sparse()
    feature_spec = {
        "int": tf.FixedLenFeature(None, dtypes.int16, np.nan),
        "float": tf.VarLenFeature(dtypes.float16),
        "bool": tf.VarLenFeature(dtypes.bool),
        "string": tf.FixedLenFeature(None, dtypes.string, "")
    }

    pandas_df = pd.read_csv(data_path, dtype={"string": object})
    # Pandas insanely uses NaN for empty cells in a string column.
    # And, we can't use Pandas replace() to fix them because nan != nan
    s = pandas_df["string"]
    for i in range(0, len(s)):
      if isinstance(s[i], float) and math.isnan(s[i]):
        pandas_df.set_value(i, "string", "")
    tensorflow_df = df.TensorFlowDataFrame.from_csv_with_feature_spec(
        [data_path],
        batch_size=batch_size,
        shuffle=False,
        feature_spec=feature_spec)

    # These columns were sparse; re-densify them for comparison
    tensorflow_df["float"] = densify.Densify(np.nan)(tensorflow_df["float"])
    tensorflow_df["bool"] = densify.Densify(np.nan)(tensorflow_df["bool"])

    self._assert_pandas_equals_tensorflow(pandas_df,
                                          tensorflow_df,
                                          num_batches=num_batches,
                                          batch_size=batch_size) 

def ExtractBitsFromFloat16(x):
  return np.asscalar(np.asarray(x, dtype=np.float16).view(np.uint16)) 

def _valid_dtypes(self):
    """Valid types for loss, variables and gradients.

    Subclasses should override to allow other float types.

    Returns:
      Valid types for loss, variables and gradients.
    """
    return set([dtypes.float16, dtypes.float32, dtypes.float64]) 

def testFeedAndFetch(self):
    with session.Session() as sess:
      for dtype in [dtypes.float16,
                    dtypes.float32,
                    dtypes.float64,
                    dtypes.int32,
                    dtypes.uint8,
                    dtypes.int16,
                    dtypes.int8,
                    dtypes.int64,
                    dtypes.bool,
                    dtypes.complex64,
                    dtypes.complex128]:
        for shape in [(32, 4, 128), (37,), (2, 0, 6), (0, 0, 0)]:
          np_dtype = dtype.as_numpy_dtype

          feed_t = array_ops.placeholder(dtype=dtype, shape=shape)
          out_t = array_ops.identity(feed_t)

          np_array = np.random.randint(-10, 10, shape)

          if dtype == dtypes.bool:
            np_array = np_array > 0
          elif dtype == dtypes.complex64:
            np_array = np.sqrt(np_array.astype(np_dtype))
          elif dtype == dtypes.complex64:
            np_array = np.sqrt(np_array.astype(np_dtype))
          else:
            np_array = np_array.astype(np_dtype)

          self.assertAllEqual(np_array,
                              sess.run(out_t, feed_dict={feed_t: np_array}))
          # Check that we can also get the feed back.
          self.assertAllEqual(np_array,
                              sess.run(feed_t, feed_dict={feed_t: np_array}))
          # Also check that we can get both back.
          out_v, feed_v = sess.run([out_t, feed_t],
                                   feed_dict={feed_t: np_array})
          self.assertAllEqual(np_array, out_v)
          self.assertAllEqual(np_array, feed_v) 

def testFetchScalar(self):
    with session.Session() as s:
      for scalar in np.int32, np.int64, np.float16, np.float32, np.float64:
        x = scalar(7)
        y = scalar(8)
        tf_x = constant_op.constant(x, shape=[])
        tf_y = constant_op.constant(y)
        tf_xy = math_ops.add(tf_x, tf_y)
        # Single fetch
        xy = s.run(tf_xy)
        self.assertEqual(scalar, type(xy))
        self.assertEqual(x + y, xy)
        # List fetch
        xy, = s.run([tf_xy])
        self.assertEqual(scalar, type(xy))
        self.assertEqual(x + y, xy)
        # Dict fetch
        xy = s.run({'xy': tf_xy})['xy']
        self.assertEqual(scalar, type(xy))
        self.assertEqual(x + y, xy)
        # Nested list fetch
        xy = s.run([[[tf_xy]], tf_xy, [tf_xy]])
        self.assertAllEqual(xy, [[[x + y]], x + y, [x + y]])
        self.assertEqual(scalar, type(xy[0][0][0]))
        self.assertEqual(scalar, type(xy[1]))
        self.assertEqual(scalar, type(xy[2][0])) 

def _CropAndResizeGrad(op, grad):
  """The derivatives for crop_and_resize.

  We back-propagate to the image only when the input image tensor has floating
  point dtype but we always back-propagate to the input boxes tensor.

  Args:
    op: The CropAndResize op.
    grad: The tensor representing the gradient w.r.t. the output.

  Returns:
    The gradients w.r.t. the input image, boxes, as well as the always-None
    gradients w.r.t. box_ind and crop_size.
  """
  image = op.inputs[0]
  if image.get_shape().is_fully_defined():
    image_shape = image.get_shape().as_list()
  else:
    image_shape = array_ops.shape(image)

  allowed_types = [dtypes.float16, dtypes.float32, dtypes.float64]
  if op.inputs[0].dtype in allowed_types:
    # pylint: disable=protected-access
    grad0 = gen_image_ops.crop_and_resize_grad_image(grad,
                                                     op.inputs[1],
                                                     op.inputs[2],
                                                     image_shape,
                                                     T=op.get_attr("T"))
    # pylint: enable=protected-access
  else:
    grad0 = None

  grad1 = gen_image_ops.crop_and_resize_grad_boxes(grad, op.inputs[0],
                                                   op.inputs[1], op.inputs[2])

  return [grad0, grad1, None, None]