Python v1.bfloat16方法代碼示例

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def _custom_getter(self):
    if self.hparams.weight_dtype == "bfloat16":
      if self.hparams.optimizer != "Adafactor":
        raise NotImplementedError(
            "weight_dtype=bfloat16 only implemented with Adafactor optimizer")
      activation_dtype = tf.float32
      if self.hparams.activation_dtype == "bfloat16":
        activation_dtype = tf.bfloat16
      return quantization.EighthPowerEncoding().custom_getter(
          activation_dtype=activation_dtype)
    elif self.hparams.activation_dtype == "bfloat16":
      return quantization.bfloat16_activations_var_getter
    elif mixed_precision_is_enabled(hparams=self.hparams):
      return quantization.float16_activations_var_getter
    else:
      return None 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def bfloat16_activations_var_getter(getter, *args, **kwargs):
  """A custom getter function for float32 parameters and bfloat16 activations.

  Args:
    getter: custom getter
    *args: arguments
    **kwargs: keyword arguments
  Returns:
    variables with the correct dtype.
  Raises:
    KeyError: if "dtype" is not provided as a kwarg.
  """
  requested_dtype = kwargs["dtype"]
  if requested_dtype == tf.bfloat16:
    kwargs["dtype"] = tf.float32
  var = getter(*args, **kwargs)
  # This if statement is needed to guard the cast, because batch norm
  # assigns directly to the return value of this custom getter. The cast
  # makes the return value not a variable so it cannot be assigned. Batch
  # norm variables are always in fp32 so this if statement is never
  # triggered for them.
  if var.dtype.base_dtype != requested_dtype:
    var = tf.cast(var, requested_dtype)
  return var 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def _to_bfloat16_unbiased(x, noise):
  """Convert a float32 to a bfloat16 using randomized roundoff.

  Args:
    x: A float32 Tensor.
    noise: a float32 Tensor with values in [0, 1), broadcastable to tf.shape(x)
  Returns:
    A float32 Tensor.
  """
  x_sign = tf.sign(x)
  # Make sure x is positive.  If it is zero, the two candidates are identical.
  x = x * x_sign + 1e-30
  cand1 = tf.to_bfloat16(x)
  cand1_f = tf.to_float(cand1)
  # This relies on the fact that for a positive bfloat16 b,
  # b * 1.005 gives you the next higher bfloat16 and b*0.995 gives you the
  # next lower one. Both 1.005 and 0.995 are ballpark estimation.
  cand2 = tf.to_bfloat16(
      tf.where(tf.greater(x, cand1_f), cand1_f * 1.005, cand1_f * 0.995))
  ret = _randomized_roundoff_to_bfloat16(x, noise, cand1, cand2)
  return ret * tf.to_bfloat16(x_sign) 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def custom_getter(self, activation_dtype=tf.bfloat16):
    """A custom getter that uses the encoding for bfloat16 and float32 vars.

    When a bfloat16 or float32 variable is requsted, an encoded float16
    varaible is created, which is then decoded and cast to a bfloat16
    activation.

    Args:
      activation_dtype: a dtype to which to convert the decoded value.

    Returns:
      a function.
    """
    def getter_fn(getter, *args, **kwargs):
      requested_dtype = kwargs["dtype"]
      if requested_dtype in (tf.bfloat16, tf.float32):
        kwargs["dtype"] = tf.bfloat16
        kwargs["initializer"] = _EncodingInitializer(
            kwargs["initializer"], self)
        ret = self._decode_with_identity_gradient(getter(*args, **kwargs))
        return tf.cast(ret, activation_dtype)
      return getter(*args, **kwargs)
    return getter_fn 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def replace_dtype(tensor_spec_struct, from_dtype,
                  to_dtype):
  """Replaces all elements of type from_dtype with to_dtype.

  This functionality is useful for TPU training since it is most efficient with
  bfloat16 whereas preprocessing on CPU only operates on float32.

  Args:
    tensor_spec_struct: The instance of TensorSpecStruct which will be updated
      in-place.
    from_dtype: The dtype which will be replaced.
    to_dtype: The target dtype.

  Returns:
    The in-place updated TensorSpecStruct.
  """
  for key, value in tensor_spec_struct.items():
    if value.dtype == from_dtype:
      tensor_spec_struct[key] = ExtendedTensorSpec.from_spec(
          spec=value, dtype=to_dtype)
  return tensor_spec_struct 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def cast_float32_to_bfloat16(tensor_spec_struct,
                             output_spec):
  """Casts tensors with dtype float32 to bfloat16 depending on the out spec.

  Args:
    tensor_spec_struct: The instance of TensorSpecStruct which will be updated
      in-place.
    output_spec: The reference TensorSpecStruct which allows to infer which
      tensors should be cast to bfloat16.

  Returns:
    The in-place updated TensorSpecStruct.
  """
  for key, value in output_spec.items():
    if value is not None and value.dtype == tf.bfloat16:
      if tensor_spec_struct[key].dtype != tf.float32:
        raise ValueError(
            'Attempting to convert non tf.float32 type {} to tf.bfloat16 '
            'for the element {} with the name {}.'.format(
                tensor_spec_struct[key].dtype, tensor_spec_struct[key], key))
      tensor_spec_struct[key] = tf.cast(
          tensor_spec_struct[key], dtype=tf.bfloat16)
  return tensor_spec_struct 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def test_parsing(self):
    base_dir = 'tensor2robot'
    file_pattern = os.path.join(
        FLAGS.test_srcdir, base_dir, 'test_data/pose_env_test_data.tfrecord')
    dataset = tfdata.parallel_read(file_patterns=file_pattern)
    state_spec = TSPEC(shape=(64, 64, 3), dtype=tf.uint8, name='state/image',
                       data_format='jpeg')
    action_spec = TSPEC(shape=(2), dtype=tf.bfloat16, name='pose')
    reward_spec = TSPEC(shape=(), dtype=tf.float32, name='reward')
    feature_tspec = PoseEnvFeature(state=state_spec, action=action_spec)
    label_tspec = PoseEnvLabel(reward=reward_spec)

    batched_dataset = dataset.batch(batch_size=1)
    dataset = tfdata.serialized_to_parsed(batched_dataset, feature_tspec,
                                          label_tspec)
    iterator = dataset.make_one_shot_iterator()
    features, labels = iterator.get_next()
    tensorspec_utils.assert_equal(feature_tspec, features, ignore_batch=True)
    tensorspec_utils.assert_equal(label_tspec, labels, ignore_batch=True)
    with self.session() as session:
      features_, labels_ = session.run([features, labels])
      self.assertAllEqual([1, 64, 64, 3], features_.state.shape)
      self.assertAllEqual([1, 2], features_.action.shape)
      self.assertAllEqual((1,), labels_.reward.shape) 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def get_out_feature_specification(
      self, mode):
    """The specification for the output features after executing preprocess_fn.

    Note, we strip all optional specs to further reduce communication and
    computation overhead for feeding to TPUs.

    Arguments:
      mode: mode key for this feature specification

    Returns:
      A TensorSpecStruct describing the required and optional tensors.
    """
    return tensorspec_utils.replace_dtype(
        tensorspec_utils.filter_required_flat_tensor_spec(
            self._preprocessor.get_out_feature_specification(mode)),
        from_dtype=tf.float32,
        to_dtype=tf.bfloat16) 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def get_out_label_specification(
      self, mode):
    """The specification for the output labels after executing preprocess_fn.

    Note, we strip all optional specs to further reduce communication and
    computation overhead for feeding to TPUs.

    Arguments:
      mode: mode key for this feature specification

    Returns:
      A TensorSpecStruct describing the required and optional tensors.
    """
    return tensorspec_utils.replace_dtype(
        tensorspec_utils.filter_required_flat_tensor_spec(
            self._preprocessor.get_out_label_specification(mode)),
        from_dtype=tf.float32,
        to_dtype=tf.bfloat16) 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def preprocess_for_eval(image_bytes, use_bfloat16, image_size=IMAGE_SIZE):
  """Preprocesses the given image for evaluation.

  Args:
    image_bytes: `Tensor` representing an image binary of arbitrary size.
    use_bfloat16: `bool` for whether to use bfloat16.
    image_size: image size.

  Returns:
    A preprocessed image `Tensor`.
  """
  image = _decode_and_center_crop(image_bytes, image_size)
  image = tf.reshape(image, [image_size, image_size, 3])
  image = tf.image.convert_image_dtype(
      image, dtype=tf.bfloat16 if use_bfloat16 else tf.float32)
  return image 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def get_variable_dtype(
    master_dtype=tf.bfloat16,
    slice_dtype=tf.float32,
    activation_dtype=tf.float32):
  """Datatypes to use for the run.

  Args:
    master_dtype: string, datatype for checkpoints
      keep this the same between training and eval/inference
    slice_dtype: string, datatype for variables in memory
      must be tf.float32 for training
    activation_dtype: string, datatype for activations
      less memory usage if tf.bfloat16 but possible numerical issues
  Returns:
    a mtf.VariableDtype
  """
  return mtf.VariableDType(
      master_dtype=tf.as_dtype(master_dtype),
      slice_dtype=tf.as_dtype(slice_dtype),
      activation_dtype=tf.as_dtype(activation_dtype)) 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def bfloat16_to_float32_nested(input_nested):
  """Convert float32 tensors in a nested structure to bfloat16.

  Args:
    input_nested: A Python dict, values being Tensor or Python list/tuple of
      Tensor or Non-Tensor.

  Returns:
    A Python dict with the same structure as `tensor_dict`,
    with all bfloat16 tensors converted to float32.
  """
  if isinstance(input_nested, tf.Tensor):
    if input_nested.dtype == tf.bfloat16:
      return tf.cast(input_nested, dtype=tf.float32)
    else:
      return input_nested
  elif isinstance(input_nested, (list, tuple)):
    out_tensor_dict = [bfloat16_to_float32_nested(t) for t in input_nested]
  elif isinstance(input_nested, dict):
    out_tensor_dict = {
        k: bfloat16_to_float32_nested(v) for k, v in input_nested.items()
    }
  else:
    return input_nested
  return out_tensor_dict 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def bfloat16_to_float32_nested(bfloat16_tensor_dict):
  """Converts bfloat16 tensors in a nested structure to float32.

  Other tensors not of dtype bfloat16 will be left as is.

  Args:
    bfloat16_tensor_dict: A Python dict, values being Tensor or Python
      list/tuple of Tensor.

  Returns:
    A Python dict with the same structure as `bfloat16_tensor_dict`,
    with all bfloat16 tensors converted to float32.
  """
  float32_tensor_dict = {}
  for k, v in bfloat16_tensor_dict.items():
    if isinstance(v, tf.Tensor):
      float32_tensor_dict[k] = bfloat16_to_float32(v)
    elif isinstance(v, (list, tuple)):
      float32_tensor_dict[k] = [bfloat16_to_float32(t) for t in v]
  return float32_tensor_dict 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def model_fn(self, features):
    with tf.variable_scope(tf.get_variable_scope(), use_resource=True) as vs:
      self._add_variable_scope("model_fn", vs)
      transformed_features = self.bottom(features)

      if self.hparams.activation_dtype == "bfloat16":
        for k, v in sorted(six.iteritems(transformed_features)):
          if v.dtype == tf.float32:
            transformed_features[k] = tf.cast(v, tf.bfloat16)

      with tf.variable_scope("body") as body_vs:
        self._add_variable_scope("body", body_vs)
        log_info("Building model body")
        body_out = self.body(transformed_features)
      output, losses = self._normalize_body_output(body_out)

      if "training" in losses:
        log_info("Skipping T2TModel top and loss because training loss "
                 "returned from body")
        logits = output
      else:
        logits = self.top(output, features)
        losses["training"] = 0.0
        if (self._hparams.mode != tf.estimator.ModeKeys.PREDICT and
            self._hparams.mode != "attack"):
          losses["training"] = self.loss(logits, features)

      return logits, losses 

# 需要導入模塊: from tensorflow.compat import v1 [as 別名]
# 或者: from tensorflow.compat.v1 import bfloat16 [as 別名]
def adafactor_optimizer_from_hparams(hparams, lr):
  """Create an Adafactor optimizer based on model hparams.

  Args:
    hparams: model hyperparameters
    lr: learning rate scalar.
  Returns:
    an AdafactorOptimizer
  Raises:
    ValueError: on illegal values
  """
  if hparams.optimizer_adafactor_decay_type == "adam":
    decay_rate = adafactor_decay_rate_adam(
        hparams.optimizer_adafactor_beta2)
  elif hparams.optimizer_adafactor_decay_type == "pow":
    decay_rate = adafactor_decay_rate_pow(
        hparams.optimizer_adafactor_memory_exponent)
  else:
    raise ValueError("unknown optimizer_adafactor_decay_type")
  if hparams.weight_dtype == "bfloat16":
    parameter_encoding = quantization.EighthPowerEncoding()
  else:
    parameter_encoding = None
  return AdafactorOptimizer(
      multiply_by_parameter_scale=(
          hparams.optimizer_adafactor_multiply_by_parameter_scale),
      learning_rate=lr,
      decay_rate=decay_rate,
      beta1=hparams.optimizer_adafactor_beta1,
      clipping_threshold=hparams.optimizer_adafactor_clipping_threshold,
      factored=hparams.optimizer_adafactor_factored,
      simulated_quantize_bits=getattr(
          hparams, "simulated_parameter_quantize_bits", 0),
      parameter_encoding=parameter_encoding,
      use_locking=False,
      name="Adafactor")