Class _VariableStore

tensorflow/python/ops/variable_scope.py:273–1281

Variable store that carries a number of named Variables. New variable names and new variables can be created; all stored variables are initialized with the initializer passed to __init__. Attributes: vars: a dictionary with string names (same as passed in GetVar) as keys and the co

class _VariableStore(object):
  """Variable store that carries a number of named Variables.

  New variable names and new variables can be created; all stored
  variables are initialized with the initializer passed to __init__.

  Attributes:
    vars: a dictionary with string names (same as passed in GetVar) as keys and
      the corresponding TensorFlow Variables as values.
  """

  def __init__(self):
    """Create a variable store."""
    self._vars = {}  # A dictionary of the stored TensorFlow variables.
    self._partitioned_vars = {}  # A dict of the stored PartitionedVariables.
    self._store_eager_variables = False

  def get_hashtable(self,
                    name,
                    shape=None,
                    dtype=dtypes.float32,
                    initializer=None,
                    collections=None,
                    reuse=None,
                    trainable=None,
                    synchronization=VariableSynchronization.AUTO,
                    partitioner=None,
                    children=None):
    if context.executing_eagerly():
      if not self._store_eager_variables and reuse:
        raise RuntimeError(
            "When eager execution is enabled variable reuse is only supported"
            " when an EagerVariableStore is active. See the documentation on"
            " EagerVariableStore for example usage.")
      if self._store_eager_variables:
        reuse = AUTO_REUSE
    try:
      dtype = dtype.base_dtype
    except AttributeError:
      # .base_dtype not existing means that we will try and use the raw dtype
      # which was passed in - this might be a NumPy type which is valid.
      pass
    def _hashtable_getter(name,
                          shape=None,
                          dtype=dtypes.float32,
                          initializer=None,
                          collections=None,
                          reuse=None,
                          trainable=None,
                          partitioner=None,
                          children=None):
      # HashTable cases
      if partitioner is not None:
        if not callable(partitioner):
          raise ValueError(
              "Partitioner must be callable, but received: %s" % partitioner)
        return self._get_distribute_hashtable(name=name,
                                              shape=shape,