models.py 文件源码

def loss(self, x, y):
        '''
        Args:
            x: shape=[s, b, c]
            y: shape=[s, b]
        Returns:
            a `dict` of losses
        '''
        z_mu, z_lv = self._encode(x, is_training=self.is_training)
        z = GaussianSampleLayer(z_mu, z_lv)
        xh = self._decode(z, y, is_training=self.is_training)

        with tf.name_scope('loss'):
            with tf.name_scope('E_log_p_x_zy'):
                L_x = -1.0 * tf.reduce_mean(
                    GaussianLogDensity(x, xh, tf.zeros_like(x)),
                )
            with tf.name_scope('D_KL_z'):
                L_z = tf.reduce_mean(
                    GaussianKLD(
                        z_mu, z_lv,
                        tf.zeros_like(z_mu), tf.zeros_like(z_lv)
                    )
                )
            loss = {
                'L_x': L_x,
                'L_z': L_z,
            }

        tf.summary.scalar('L_x', L_x)
        tf.summary.scalar('L_z', L_z)
        return loss