解决用saver = tf.train.Saver()加载两个模型时发生冲突的问题

首先创建两个图表

model_graph = tf.Graph()
with model_graph.as_default():
    model = Model(args)
    
adv_graph = tf.Graph()
with adv_graph.as_default():
    adversary = Adversary(adv_args)

然后建立两个会话

adv_sess = tf.Session(graph=adv_graph)
sess = tf.Session(graph=model_graph)

然后在每个会话中初始化变量并分别恢复每个图形

with sess.as_default():
    with model_graph.as_default():
        tf.global_variables_initializer().run()
        model_saver = tf.train.Saver(tf.global_variables())
        model_ckpt = tf.train.get_checkpoint_state(args.save_dir)
        model_saver.restore(sess, model_ckpt.model_checkpoint_path)

with adv_sess.as_default():
    with adv_graph.as_default():
        tf.global_variables_initializer().run()
        adv_saver = tf.train.Saver(tf.global_variables())
        adv_ckpt = tf.train.get_checkpoint_state(adv_args.save_dir)
        adv_saver.restore(adv_sess, adv_ckpt.model_checkpoint_path)

这里每当需要每个会话时,用sess.as_default()包装该会话中的任何tf函数：最后手动关闭会话

sess.close()
adv_sess.close()

转载：http://www.voidcn.com/article/p-wgywhvux-bvm.html