我正准备使用tensorflow集群来训练推荐模型。推荐模型的最新更新方法不是基于梯度的，所以我手动计算模型响应参数，然后tf.assign参数。代码如下：
def single_meta_task(self,indata,global_step,scope_name):
meta_train_predict_1=[]
meta_train_loss_1=[]
task_update_ops_1=[]
meta_train_target_1=[]
# task_1
input=indata[0]
meta_train_predict_1.append(self.basemodel.forward(input[0][0]))
self.weights = self.basemodel.get_para(scope_name)
with tf.control_dependencies(list(self.weights.values())):
meta_train_target_1.append(input[0][1])
meta_train_loss_1.append(tf.reduce_mean(self.basemodel.loss_fn(meta_train_target_1[0], meta_train_predict_1[0])))
task_update_ops_1.append(self.task_optimizer.minimize(loss=meta_train_loss_1[0],global_step=global_step))
# if update more than one step:
for i in range(1, self._num_updates):
with tf.control_dependencies([task_update_ops_1[i-1]]):
meta_train_predict_1.append(self.basemodel.forward(input[i][0]))
meta_train_target_1.append(input[i][1])
meta_train_loss_1.append(tf.reduce_mean(self.basemodel.loss_fn(meta_train_target_1[i], meta_train_predict_1[i])))
task_update_ops_1.append(self.task_optimizer.minimize(loss=meta_train_loss_1[i],global_step=global_step))
with tf.control_dependencies([task_update_ops_1[-1]]):
self.update_weights_1=self.basemodel.get_para(scope_name)
self.new_weights_1=self.get_new_weights(self.weights,scope_name,self.lr_maml) # the update parameter function
weight_update_op=self.basemodel.set_para(self.new_weights,scope_name)
我在带有tensorflow集群的tf.train.MonitoredTrainingSession中运行 weight_update_op ，但是我得到了
step: 38 meta_train_loss: 0.6931472 auc: 0.0 auc total: 0.5 time: 57.1792571545
step: 39 meta_train_loss: 0.6931472 auc: 0.5 auc total: 0.5 time: 18.3862159252
step: 40 meta_train_loss: 0.6931472 auc: 0.5 auc total: 0.5 time: 11.3142118454
step: 41 meta_train_loss: 0.6931472 auc: 0.5 auc total: 0.5 time: 12.2699699402
whitch表示参数不会更新，图形也不会流动。
我将分发火车更改为本地单火车，通过 tf.Session ，相同的代码效果很好。
有人可以告诉我为什么吗？以及如何在最终OP不是Optimizer的情况下使用分布式训练？