Python v1.variance_scaling_initializer方法代码示例

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def log_conv2d(self, input_tensor, output_tensor, stride_height, stride_width,
                 filters, initializer, use_bias):
    """Log a conv2d call."""
    if self.model == 'resnet50_v1.5':
      assert stride_height == stride_width, (
          '--ml_perf_compliance_logging does not support convolutions where '
          'the stride height is not equal to the stride width. '
          'stride_height=%d, stride_width=%d' % (stride_height, stride_width))
      if isinstance(initializer, tf.truncated_normal_initializer) or (
          isinstance(initializer, tf.variance_scaling_initializer) and
          initializer.distribution == 'truncated_normal'):
        initializer = tags.TRUNCATED_NORMAL
      elif (isinstance(initializer, tf.glorot_uniform_initializer) or
            initializer is None):
        initializer = 'glorot_uniform'
      resnet_log_helper.log_conv2d(input_tensor, output_tensor, stride_width,
                                   filters, initializer, use_bias) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def conv_kernel_initializer(shape, dtype=None, partition_info=None):
  """Initialization for convolutional kernels.

  The main difference with tf.variance_scaling_initializer is that
  tf.variance_scaling_initializer uses a truncated normal with an uncorrected
  standard deviation, whereas here we use a normal distribution. Similarly,
  tf.initializers.variance_scaling uses a truncated normal with
  a corrected standard deviation.

  Args:
    shape: shape of variable
    dtype: dtype of variable
    partition_info: unused

  Returns:
    an initialization for the variable
  """
  del partition_info
  kernel_height, kernel_width, _, out_filters = shape
  fan_out = int(kernel_height * kernel_width * out_filters)
  return tf.random_normal(
      shape, mean=0.0, stddev=np.sqrt(2.0 / fan_out), dtype=dtype) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def dense_kernel_initializer(shape, dtype=None, partition_info=None):
  """Initialization for dense kernels.

  This initialization is equal to
    tf.variance_scaling_initializer(scale=1.0/3.0, mode='fan_out',
                                    distribution='uniform').
  It is written out explicitly here for clarity.

  Args:
    shape: shape of variable
    dtype: dtype of variable
    partition_info: unused

  Returns:
    an initialization for the variable
  """
  del partition_info
  init_range = 1.0 / np.sqrt(shape[1])
  return tf.random_uniform(shape, -init_range, init_range, dtype=dtype) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def conv2d_fixed_padding(inputs, filters, kernel_size, strides, data_format):
  """Strided 2-D convolution with explicit padding."""
  # The padding is consistent and is based only on `kernel_size`, not on the
  # dimensions of `inputs` (as opposed to using `tf.layers.conv2d` alone).
  if strides > 1:
    inputs = fixed_padding(inputs, kernel_size, data_format)

  return tf.layers.conv2d(
      inputs=inputs, filters=filters, kernel_size=kernel_size, strides=strides,
      padding=('SAME' if strides == 1 else 'VALID'), use_bias=False,
      kernel_initializer=tf.variance_scaling_initializer(),
      data_format=data_format)


################################################################################
# ResNet block definitions.
################################################################################ 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def conv2d_fixed_padding(inputs, filters, kernel_size, strides, data_format,
                         weight_decay):
  """Strided 2-D convolution with explicit padding."""
  # The padding is consistent and is based only on `kernel_size`, not on the
  # dimensions of `inputs` (as opposed to using `tf.layers.conv2d` alone).
  if strides > 1:
    inputs = fixed_padding(inputs, kernel_size, data_format)

  if weight_decay is not None:
    weight_decay = contrib_layers.l2_regularizer(weight_decay)

  return tf.layers.conv2d(
      inputs=inputs, filters=filters, kernel_size=kernel_size, strides=strides,
      padding=('SAME' if strides == 1 else 'VALID'), use_bias=False,
      kernel_initializer=tf.variance_scaling_initializer(),
      kernel_regularizer=weight_decay,
      data_format=data_format) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def get_variable_initializer(hparams):
  """Get variable initializer from hparams."""
  if not hparams.initializer:
    return None

  mlperf_log.transformer_print(key=mlperf_log.MODEL_HP_INITIALIZER_GAIN,
                               value=hparams.initializer_gain,
                               hparams=hparams)

  if not tf.executing_eagerly():
    tf.logging.info("Using variable initializer: %s", hparams.initializer)
  if hparams.initializer == "orthogonal":
    return tf.orthogonal_initializer(gain=hparams.initializer_gain)
  elif hparams.initializer == "uniform":
    max_val = 0.1 * hparams.initializer_gain
    return tf.random_uniform_initializer(-max_val, max_val)
  elif hparams.initializer == "normal_unit_scaling":
    return tf.variance_scaling_initializer(
        hparams.initializer_gain, mode="fan_avg", distribution="normal")
  elif hparams.initializer == "uniform_unit_scaling":
    return tf.variance_scaling_initializer(
        hparams.initializer_gain, mode="fan_avg", distribution="uniform")
  elif hparams.initializer == "xavier":
    return tf.initializers.glorot_uniform()
  else:
    raise ValueError("Unrecognized initializer: %s" % hparams.initializer) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def conv_linear_map(inputs, nin, nout, bias_start, prefix):
  """Convolutional liner map.

  Maps 3D tensor by last dimension.

  Args:
    inputs: Inputs that should be shuffled
    nin: Input feature map count
    nout: Output feature map count
    bias_start: Bias start value
    prefix: Name prefix

  Returns:
    tf.Tensor: Inputs with applied convolution
  """

  with tf.variable_scope(prefix):
    inp_shape = tf.shape(inputs)

    initializer = tf.variance_scaling_initializer(
        scale=1.0, mode="fan_avg", distribution="uniform")
    kernel = tf.get_variable("CvK", [nin, nout], initializer=initializer)
    bias_term = tf.get_variable(
        "CvB", [nout], initializer=tf.constant_initializer(0.0))

    mul_shape = [inp_shape[0] * inp_shape[1], nin]
    res = tf.matmul(tf.reshape(inputs, mul_shape), kernel)
    res = tf.reshape(res, [inp_shape[0], inp_shape[1], nout])
    return res + bias_start + bias_term


# pylint: disable=useless-object-inheritance 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def build(self, input_shape):
    """Initialize layer weights and sublayers.

    Args:
      input_shape: shape of inputs
    """
    in_units = input_shape[-1]
    middle_units = in_units * 4
    out_units = in_units * 2
    init = tf.variance_scaling_initializer(
        scale=1.0, mode="fan_avg", distribution="uniform")

    self.first_linear = tf.keras.layers.Dense(
        middle_units,
        use_bias=False,
        kernel_initializer=init,
        name=self.prefix + "/cand1")

    self.second_linear = tf.keras.layers.Dense(
        out_units, kernel_initializer=init, name=self.prefix + "/cand2")
    self.layer_norm = LayerNormalization()

    init = tf.constant_initializer(self.init_value)
    self.residual_scale = self.add_weight(
        self.prefix + "/residual", [out_units], initializer=init)
    super(RSU, self).build(input_shape) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def _set_initializer():
  """Set initializer used for all model variables."""
  tf.get_variable_scope().set_initializer(
      tf.variance_scaling_initializer(scale=1.0, mode="fan_avg")) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def __init__(self, sparsity, seed=None, dtype=tf.float32):
    if sparsity < 0. or sparsity > 1.:
      raise ValueError('sparsity must be in the range [0., 1.].')
    self.kernel_initializer = tf.variance_scaling_initializer(seed=seed,
                                                              dtype=dtype)
    self.seed = seed
    self.dtype = dtype
    self.sparsity = float(sparsity) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def __init__(self,
               batchnorm_training,
               batchnorm_scale=True,
               default_batchnorm_momentum=0.997,
               default_batchnorm_epsilon=1e-5,
               weight_decay=0.0001,
               conv_hyperparams=None,
               min_depth=8,
               depth_multiplier=1):
    """Alternative tf.keras.layers interface, for use by the Keras Resnet V1.

    The class is used by the Keras applications kwargs injection API to
    modify the Resnet V1 Keras application with changes required by
    the Object Detection API.

    Args:
      batchnorm_training: Bool. Assigned to Batch norm layer `training` param
        when constructing `freezable_batch_norm.FreezableBatchNorm` layers.
      batchnorm_scale: If True, uses an explicit `gamma` multiplier to scale
        the activations in the batch normalization layer.
      default_batchnorm_momentum: Float. When 'conv_hyperparams' is None,
        batch norm layers will be constructed using this value as the momentum.
      default_batchnorm_epsilon: Float. When 'conv_hyperparams' is None,
        batch norm layers will be constructed using this value as the epsilon.
      weight_decay: The weight decay to use for regularizing the model.
      conv_hyperparams: A `hyperparams_builder.KerasLayerHyperparams` object
        containing hyperparameters for convolution ops. Optionally set to `None`
        to use default resnet_v1 layer builders.
      min_depth: Minimum number of filters in the convolutional layers.
      depth_multiplier: The depth multiplier to modify the number of filters
        in the convolutional layers.
    """
    self._batchnorm_training = batchnorm_training
    self._batchnorm_scale = batchnorm_scale
    self._default_batchnorm_momentum = default_batchnorm_momentum
    self._default_batchnorm_epsilon = default_batchnorm_epsilon
    self._conv_hyperparams = conv_hyperparams
    self._min_depth = min_depth
    self._depth_multiplier = depth_multiplier
    self.regularizer = tf.keras.regularizers.l2(weight_decay)
    self.initializer = tf.variance_scaling_initializer() 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def variance_scaling_initializer(scale=2.0, mode='fan_in',
                                 distribution='truncated_normal',
                                 mean=0.0, seed=None, dtype=tf.float32):
  """Like tf.variance_scaling_initializer but supports non-zero means."""
  if not dtype.is_floating:
    raise TypeError('Cannot create initializer for non-floating point type.')
  if mode not in ['fan_in', 'fan_out', 'fan_avg']:
    raise TypeError('Unknown mode %s [fan_in, fan_out, fan_avg]' % mode)

  # pylint: disable=unused-argument
  def _initializer(shape, dtype=dtype, partition_info=None):
    """Initializer function."""
    if not dtype.is_floating:
      raise TypeError('Cannot create initializer for non-floating point type.')
    # Estimating fan_in and fan_out is not possible to do perfectly, but we try.
    # This is the right thing for matrix multiply and convolutions.
    if shape:
      fan_in = float(shape[-2]) if len(shape) > 1 else float(shape[-1])
      fan_out = float(shape[-1])
    else:
      fan_in = 1.0
      fan_out = 1.0
    for dim in shape[:-2]:
      fan_in *= float(dim)
      fan_out *= float(dim)
    if mode == 'fan_in':
      # Count only number of input connections.
      n = fan_in
    elif mode == 'fan_out':
      # Count only number of output connections.
      n = fan_out
    elif mode == 'fan_avg':
      # Average number of inputs and output connections.
      n = (fan_in + fan_out) / 2.0
    if distribution == 'truncated_normal':
      # To get stddev = math.sqrt(scale / n) need to adjust for truncated.
      trunc_stddev = math.sqrt(1.3 * scale / n)
      return tf.truncated_normal(shape, mean, trunc_stddev, dtype, seed=seed)
    elif distribution == 'uniform':
      # To get stddev = math.sqrt(scale / n) need to adjust for uniform.
      limit = math.sqrt(3.0 * scale / n)
      return tf.random_uniform(shape, mean-limit, mean+limit, dtype, seed=seed)
    else:
      assert 'Unexpected distribution %s.' % distribution
  # pylint: enable=unused-argument

  return _initializer 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import variance_scaling_initializer [as 别名]
def conv2d_fixed_padding(inputs,
                         filters,
                         kernel_size,
                         strides,
                         pruning_method='baseline',
                         data_format='channels_first',
                         weight_decay=0.,
                         name=None):
  """Strided 2-D convolution with explicit padding.

  The padding is consistent and is based only on `kernel_size`, not on the
  dimensions of `inputs` (as opposed to using `tf.layers.conv2d` alone).

  Args:
    inputs:  Input tensor, float32 or bfloat16 of size [batch, channels, height,
      width].
    filters: Int specifying number of filters for the first two convolutions.
    kernel_size: Int designating size of kernel to be used in the convolution.
    strides: Int specifying the stride. If stride >1, the input is downsampled.
    pruning_method: String that specifies the pruning method used to identify
      which weights to remove.
    data_format: String that specifies either "channels_first" for [batch,
      channels, height,width] or "channels_last" for [batch, height, width,
      channels].
    weight_decay: Weight for the l2 regularization loss.
    name: String that specifies name for model layer.

  Returns:
    The output activation tensor of size [batch, filters, height_out, width_out]

  Raises:
    ValueError: If the data_format provided is not a valid string.
  """
  if strides > 1:
    inputs = resnet_model.fixed_padding(
        inputs, kernel_size, data_format=data_format)
    padding = 'VALID'
  else:
    padding = 'SAME'

  kernel_initializer = tf.variance_scaling_initializer()

  kernel_regularizer = contrib_layers.l2_regularizer(weight_decay)
  return sparse_conv2d(
      x=inputs,
      units=filters,
      activation=None,
      kernel_size=[kernel_size, kernel_size],
      use_bias=False,
      kernel_initializer=kernel_initializer,
      kernel_regularizer=kernel_regularizer,
      bias_initializer=None,
      biases_regularizer=None,
      sparsity_technique=pruning_method,
      normalizer_fn=None,
      strides=[strides, strides],
      padding=padding,
      data_format=data_format,
      name=name)