def _setup_layer_v1()

in tensorflow_examples/lite/model_maker/core/task/hub_loader.py [0:0]

• 32 lines of code
• 9 McCabe index (conditional complexity)

  def _setup_layer_v1(self, trainable=False, **kwargs):
    """Constructs keras layer with relevant weights and losses."""
    # Initialize an empty layer, then add_weight() etc. as needed.
    super(hub.KerasLayer, self).__init__(trainable=trainable, **kwargs)

    if not self._is_hub_module_v1:
      raise ValueError(
          'Only supports to set up v1 hub module in this function.')

    # v2 trainable_variable:
    if hasattr(self._func, 'trainable_variables'):
      for v in self._func.trainable_variables:
        self._add_existing_weight(v, trainable=True)
      trainable_variables = {id(v) for v in self._func.trainable_variables}
    else:
      trainable_variables = set()

    if not hasattr(self._func, '_self_unconditional_checkpoint_dependencies'):
      raise ValueError('_func doesn\'t contains attribute '
                       '_self_unconditional_checkpoint_dependencies.')
    dependencies = self._func._self_unconditional_checkpoint_dependencies  # pylint: disable=protected-access

    # Adds trainable variables.
    for dep in dependencies:
      if dep.name == 'variables':
        for v in dep.ref:
          if id(v) not in trainable_variables:
            self._add_existing_weight(v, trainable=True)
            trainable_variables.add(id(v))

    # Adds non-trainable variables.
    if hasattr(self._func, 'variables'):
      for v in self._func.variables:
        if id(v) not in trainable_variables:
          self._add_existing_weight(v, trainable=False)

    # Forward the callable's regularization losses (if any).
    if hasattr(self._func, 'regularization_losses'):
      for l in self._func.regularization_losses:
        if not callable(l):
          raise ValueError(
              'hub.KerasLayer(obj) expects obj.regularization_losses to be an '
              'iterable of callables, each returning a scalar loss term.')
        self.add_loss(self._call_loss_if_trainable(l))  # Supports callables.