model.py 文件源码

def _build_optimizer(self):
        """Based on the loss tensor, build an optimizer that minimizes the loss.

        This function returns an optimizer operation that updates the model's trainable parameters
        by determining the loss's gradients w.r.t. each of the trainable parameters. Specifically,
        RMSProp is used to minimize the loss. The gradients are clipped to the max_gradient_norm to
        prevent too drastic updates of the trainable parameters. See also tf.clip_by_global_norm

        Returns:
            tf.Operation: An operation that updates the model's trainable parameters.
        """

        # Clip the gradients
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(self._loss, tvars), self.max_gradient_norm)

        # Optimize the variables
        optimizer = tf.train.RMSPropOptimizer(self._learning_rate)
        return optimizer.apply_gradients(zip(grads, tvars))