Python v1.get_variable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 get_variable [as 别名]
def get_gradients_to_apply(self, device_num, gradient_state):
    device_grads = gradient_state  # From 2nd result of preprocess_device_grads.

    avg_grads, self.grad_has_inf_nan = (
        variable_mgr_util.aggregate_gradients_using_copy_with_device_selection(
            self.benchmark_cnn,
            device_grads,
            use_mean=True,
            check_inf_nan=self.benchmark_cnn.enable_auto_loss_scale))

    # Make shadow variable on a parameter server for each original trainable
    # variable.
    for i, (g, v) in enumerate(avg_grads):
      my_name = variable_mgr_util.PS_SHADOW_VAR_PREFIX + '/' + v.name
      if my_name.endswith(':0'):
        my_name = my_name[:-2]
      new_v = tf.get_variable(
          my_name,
          dtype=v.dtype.base_dtype,
          initializer=v.initial_value,
          trainable=True)
      avg_grads[i] = (g, new_v)
    return avg_grads 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def _conv2d_impl(self, input_layer, num_channels_in, filters, kernel_size,
                   strides, padding, kernel_initializer):
    if self.use_tf_layers:
      return conv_layers.conv2d(input_layer, filters, kernel_size, strides,
                                padding, self.channel_pos,
                                kernel_initializer=kernel_initializer,
                                use_bias=False)
    else:
      weights_shape = [kernel_size[0], kernel_size[1], num_channels_in, filters]
      # We use the name 'conv2d/kernel' so the variable has the same name as its
      # tf.layers equivalent. This way, if a checkpoint is written when
      # self.use_tf_layers == True, it can be loaded when
      # self.use_tf_layers == False, and vice versa.
      weights = self.get_variable('conv2d/kernel', weights_shape,
                                  self.variable_dtype, self.dtype,
                                  initializer=kernel_initializer)
      if self.data_format == 'NHWC':
        strides = [1] + strides + [1]
      else:
        strides = [1, 1] + strides
      return tf.nn.conv2d(input_layer, weights, strides, padding,
                          data_format=self.data_format) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def __init__(self, tensors):
    tensors = list(tensors)
    with tf.variable_scope('averaged'):
      self._num_samples = tf.Variable(0, name='num_samples', trainable=False)
      with tf.variable_scope('avg'):
        self._averages = [
            tf.get_variable(
                tensor.name.replace('/', '-').replace(':', '-'),
                tensor.get_shape(), initializer=tf.zeros_initializer(),
                trainable=False)
            for tensor in tensors]
      with tf.variable_scope('save'):
        self._saves = [
            tf.get_variable(
                tensor.name.replace('/', '-').replace(':', '-'),
                tensor.get_shape(), initializer=tf.zeros_initializer(),
                trainable=False)
            for tensor in tensors]
    self._tensors = tensors
    self._take_sample = self._make_take_sample()
    self._switch = self._make_swith_to_average()
    self._restore = self._make_restore()
    self._reset = self._make_reset() 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def layer_norm(x, reduction_indices, epsilon=1e-9, gain=None, bias=None,
               per_element=True, scope=None):
  """DOC."""
  reduction_indices = ensure_list(reduction_indices)
  mean = tf.reduce_mean(x, reduction_indices, keep_dims=True)
  variance = tf.reduce_mean(tf.squared_difference(x, mean),
                            reduction_indices, keep_dims=True)
  normalized = (x - mean) / tf.sqrt(variance + epsilon)
  dtype = x.dtype
  shape = x.get_shape().as_list()
  for i in six.moves.range(len(shape)):
    if i not in reduction_indices or not per_element:
      shape[i] = 1
  with tf.variable_scope(scope or 'layer_norm'):
    if gain is None:
      gain = tf.get_variable('gain', shape=shape, dtype=dtype,
                             initializer=tf.ones_initializer())
    if bias is None:
      bias = tf.get_variable('bias', shape=shape, dtype=dtype,
                             initializer=tf.zeros_initializer())
  return gain*normalized+bias 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def _build_tiled_linear(self, inputs, input_name_and_sizes,
                          output_name_and_sizes, add_bias):
    results = []
    for output_name, output_size in output_name_and_sizes:
      r = 0.0
      for input_, (input_name, input_size) in zip(inputs, input_name_and_sizes):
        name = 'W_{}_{}'.format(input_name, output_name)
        weight = self._get_variable(
            name, shape=[output_size, input_size])
        r += tf.sparse_tensor_dense_matmul(weight, input_, adjoint_b=True)
      r = tf.transpose(r)
      if add_bias:
        # Biases are dense, hence we call _get_variable of the base
        # class.
        r += super(SparseTiledLinear, self)._get_variable(
            'B_{}'.format(output_name), shape=[output_size],
            default_initializer=tf.zeros_initializer())
      results.append(r)
    return results


# TODO(melisgl): Since computation is the same as in TiledLinear,
# perhaps this should be implemented as a custom getter (see
# tf.get_variable) instead of being tied to tiling. 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def init_vq_bottleneck(bottleneck_size, hidden_size):
  """Get lookup table for VQ bottleneck."""
  means = tf.get_variable(
      name="means",
      shape=[bottleneck_size, hidden_size],
      initializer=tf.uniform_unit_scaling_initializer())
  ema_count = tf.get_variable(
      name="ema_count",
      shape=[bottleneck_size],
      initializer=tf.constant_initializer(0),
      trainable=False)
  with tf.colocate_with(means):
    ema_means = tf.get_variable(
        name="ema_means",
        initializer=means.initialized_value(),
        trainable=False)

  return means, ema_means, ema_count 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def add_depth_embedding(x):
  """Add n-dimensional embedding as the depth embedding (timing signal).

  Adds embeddings to represent the position of the step in the recurrent
  tower.

  Args:
    x: a tensor with shape [max_step, batch, length, depth]

  Returns:
    a Tensor the same shape as x.
  """
  x_shape = common_layers.shape_list(x)
  depth = x_shape[-1]
  num_steps = x_shape[0]
  shape = [num_steps, 1, 1, depth]
  depth_embedding = (
      tf.get_variable(
          "depth_embedding",
          shape,
          initializer=tf.random_normal_initializer(0, depth**-0.5)) * (depth**
                                                                       0.5))

  x += depth_embedding
  return x 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def scale_gaussian_prior(name, z, logscale_factor=3.0, trainable=True):
  """Returns N(s^i * z^i, std^i) where s^i and std^i are pre-component.

  s^i is a learnable parameter with identity initialization.
  std^i is optionally learnable with identity initialization.

  Args:
    name: variable scope.
    z: input_tensor
    logscale_factor: equivalent to scaling up the learning_rate by a factor
                     of logscale_factor.
    trainable: Whether or not std^i is learnt.
  """
  with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
    z_shape = common_layers.shape_list(z)
    latent_multiplier = tf.get_variable(
        "latent_multiplier", shape=z_shape, dtype=tf.float32,
        initializer=tf.ones_initializer())
    log_scale = tf.get_variable(
        "log_scale_latent", shape=z_shape, dtype=tf.float32,
        initializer=tf.zeros_initializer(), trainable=trainable)
    log_scale = log_scale * logscale_factor
    return tfp.distributions.Normal(
        loc=latent_multiplier * z, scale=tf.exp(log_scale)) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def _conv_function(self, input_tensor, output_depth, padding):
    input_depth = input_tensor.shape.as_list()[-1]
    if not ((output_depth >= input_depth) and
            (output_depth % input_depth == 0)):
      raise ValueError(
          "Depthwise layer output_depth (%s) must be greater or equal to and "
          "a multiple of the depth of the "
          "input tensor (%s)." % (output_depth, input_depth))
    channel_multiplier = calculate_depthwise_channel_multiplier(
        input_depth, output_depth)
    kernel = tf.get_variable(
        "kernel", [self._conv_width, 1, input_depth, channel_multiplier])
    return tf.nn.depthwise_conv2d(
        input_tensor,
        kernel, [1, 1, 1, 1],
        padding=padding,
        name="depthwise_conv_%sx1" % str(self._conv_width)) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def init_internal_states(self):
    # Hardcoded LSTM-CONV shapes.
    # These sizes are calculated based on original atari frames.
    # TODO(mbz): find a cleaner way of doing this maybe?!
    batch_size = self.hparams.batch_size
    shapes = [(batch_size, 53, 40, 8),
              (batch_size, 53, 40, 8),
              (batch_size, 27, 20, 16),
              (batch_size, 27, 20, 16),
              (batch_size, 53, 40, 8)]

    with tf.variable_scope("clean_scope"):
      # Initialize conv-lstm states with zeros
      init = tf.zeros_initializer()
      states = []
      for i, shape in enumerate(shapes):
        # every lstm-conv state has two variables named c and h.
        c = tf.get_variable("c%d" % i, shape, trainable=False, initializer=init)
        h = tf.get_variable("h%d" % i, shape, trainable=False, initializer=init)
        states.append((c, h))
      return states 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def group_norm(x, filters=None, num_groups=8, epsilon=1e-5):
  """Group normalization as in https://arxiv.org/abs/1803.08494."""
  x_shape = shape_list(x)
  if filters is None:
    filters = x_shape[-1]
  assert len(x_shape) == 4
  assert filters % num_groups == 0
  # Prepare variables.
  scale = tf.get_variable(
      "group_norm_scale", [filters], initializer=tf.ones_initializer())
  bias = tf.get_variable(
      "group_norm_bias", [filters], initializer=tf.zeros_initializer())
  epsilon, scale, bias = [cast_like(t, x) for t in [epsilon, scale, bias]]
  # Reshape and compute group norm.
  x = tf.reshape(x, x_shape[:-1] + [num_groups, filters // num_groups])
  # Calculate mean and variance on heights, width, channels (not groups).
  mean, variance = tf.nn.moments(x, [1, 2, 4], keep_dims=True)
  norm_x = (x - mean) * tf.rsqrt(variance + epsilon)
  return tf.reshape(norm_x, x_shape) * scale + bias 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def zero_add(previous_value, x, name=None, reuse=None):
  """Resnet connection with zero initialization.

  Another type of resnet connection which returns previous_value + gamma * x.
  gamma is a trainable scalar and initialized with zero. It is useful when a
  module is plugged into a trained model and we want to make sure it matches the
  original model's performance.

  Args:
    previous_value:  A tensor.
    x: A tensor.
    name: name of variable scope; defaults to zero_add.
    reuse: reuse scope.

  Returns:
    previous_value + gamma * x.
  """
  with tf.variable_scope(name, default_name="zero_add", reuse=reuse):
    gamma = tf.get_variable("gamma", (), initializer=tf.zeros_initializer())
    return previous_value + gamma * x 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def testSpectralNorm(self):
    # Test that after 20 calls to apply_spectral_norm, the spectral
    # norm of the normalized matrix is close to 1.0
    with tf.Graph().as_default():
      weights = tf.get_variable("w", dtype=tf.float32, shape=[2, 3, 50, 100])
      weights = tf.multiply(weights, 10.0)
      normed_weight, assign_op = common_layers.apply_spectral_norm(weights)

      with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())

        for _ in range(20):
          sess.run(assign_op)
          normed_weight, assign_op = common_layers.apply_spectral_norm(
              weights)
        normed_weight = sess.run(normed_weight).reshape(-1, 100)
        _, s, _ = np.linalg.svd(normed_weight)
        self.assertTrue(np.allclose(s[0], 1.0, rtol=0.1)) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def dense_weightnorm(
    name, x, n_out, x_mask, init_scale, init, dtype=tf.float32):
  """Dense layer with weight normalization."""
  n_in = common_layers.shape_list(x)[2]
  eps = tf.keras.backend.epsilon()
  with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
    v = tf.get_variable(
        "v", [n_in, n_out], dtype,
        initializer=tf.random_normal_initializer(0, 0.05), trainable=True)
    v = v / tf.norm(v, axis=0, keepdims=True)
    t = tf.matmul(x, v)  # [B, L, n_out]
    mean, var = moments_over_bl(t, x_mask)
    g_init = init_scale / (tf.sqrt(var) + eps)
    g = get_variable_ddi(
        "g", [n_out], g_init, init,
        initializer=tf.zeros_initializer, dtype=dtype, trainable=True)
    b = get_variable_ddi(
        "b", [n_out], -mean*g_init, init,
        initializer=tf.zeros_initializer, dtype=dtype, trainable=True)
    w = g * v
    y = tf.matmul(x, w) + b
    tf.summary.histogram("_g", g)
    return y 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import get_variable [as 别名]
def get_vq_codebook(codebook_size, hidden_size):
  """Get lookup table for VQ bottleneck."""
  with tf.variable_scope("vq", reuse=tf.AUTO_REUSE):
    means = tf.get_variable(
        name="means",
        shape=[codebook_size, hidden_size],
        initializer=tf.uniform_unit_scaling_initializer())

    ema_count = tf.get_variable(
        name="ema_count",
        shape=[codebook_size],
        initializer=tf.constant_initializer(0),
        trainable=False)

    with tf.colocate_with(means):
      ema_means = tf.get_variable(
          name="ema_means",
          initializer=means.initialized_value(),
          trainable=False)

  return means, ema_means, ema_count