多目标权重融合方式

1. 多目标学习在推荐系统的应用(MMOE/ESMM/PLE) - 知乎

## combine loss
ctr_log_var = tf.get_variable(
    name='ctr_log_var',
    dtype=tf.float32,
    shape=(1,),
    initializer=tf.zeros_initializer()
)
ysl_log_var = tf.get_variable(
    name='ysl_log_var',
    dtype=tf.float32,
    shape=(1,),
    initializer=tf.zeros_initializer()
)
loss_final = 2 * loss_ctr * tf.exp(-ctr_log_var) + loss_ysl * tf.exp(-ysl_log_var) + ctr_log_var + ysl_log_var

2.