tensorflow_privacy/privacy/optimizers/dp_optimizer_keras.py [222:259]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    def _compute_gradients(self, loss, var_list, grad_loss=None, tape=None):
      """DP-SGD version of base class method."""

      self._was_dp_gradients_called = True
      # Compute loss.
      if not callable(loss) and tape is None:
        raise ValueError('`tape` is required when a `Tensor` loss is passed.')
      tape = tape if tape is not None else tf.GradientTape()

      if callable(loss):
        with tape:
          if not callable(var_list):
            tape.watch(var_list)

          loss = loss()
          if self._num_microbatches is None:
            num_microbatches = tf.shape(input=loss)[0]
          else:
            num_microbatches = self._num_microbatches
          microbatch_losses = tf.reduce_mean(
              tf.reshape(loss, [num_microbatches, -1]), axis=1)

          if callable(var_list):
            var_list = var_list()
      else:
        with tape:
          if self._num_microbatches is None:
            num_microbatches = tf.shape(input=loss)[0]
          else:
            num_microbatches = self._num_microbatches
          microbatch_losses = tf.reduce_mean(
              tf.reshape(loss, [num_microbatches, -1]), axis=1)

      var_list = tf.nest.flatten(var_list)

      # Compute the per-microbatch losses using helpful jacobian method.
      with tf.keras.backend.name_scope(self._name + '/gradients'):
        jacobian = tape.jacobian(microbatch_losses, var_list)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



tensorflow_privacy/privacy/optimizers/dp_optimizer_keras_vectorized.py [130:167]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    def _compute_gradients(self, loss, var_list, grad_loss=None, tape=None):
      """DP-SGD version of base class method."""

      self._was_dp_gradients_called = True
      # Compute loss.
      if not callable(loss) and tape is None:
        raise ValueError('`tape` is required when a `Tensor` loss is passed.')
      tape = tape if tape is not None else tf.GradientTape()

      if callable(loss):
        with tape:
          if not callable(var_list):
            tape.watch(var_list)

          loss = loss()
          if self._num_microbatches is None:
            num_microbatches = tf.shape(input=loss)[0]
          else:
            num_microbatches = self._num_microbatches
          microbatch_losses = tf.reduce_mean(
              tf.reshape(loss, [num_microbatches, -1]), axis=1)

          if callable(var_list):
            var_list = var_list()
      else:
        with tape:
          if self._num_microbatches is None:
            num_microbatches = tf.shape(input=loss)[0]
          else:
            num_microbatches = self._num_microbatches
          microbatch_losses = tf.reduce_mean(
              tf.reshape(loss, [num_microbatches, -1]), axis=1)

      var_list = tf.nest.flatten(var_list)

      # Compute the per-microbatch losses using helpful jacobian method.
      with tf.keras.backend.name_scope(self._name + '/gradients'):
        jacobian = tape.jacobian(microbatch_losses, var_list)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -