The tf.control_dependencies seems extremely time consuming #9
Description
试着运行了好几次，感觉很卡啊。每次执行到tf.control_dependencies的时候就会卡很久，最长的能卡15分钟左右。求问是什么原因啊？是因为网络太大了吗？我看tf.control_dependencies那里，变量列表长度是2300多，变量太多导致的嘛？
网络构建完以后，要送入GPU执行的时候，也是卡住，十几分钟了还没开始运行。
有什么办法能加速一下嘛？感谢大佬指点！
I run the train_search.py for multiple times. The program always hang at the tf.control_dependencies in the compute_unrolled_step() function. It just stops here and printing nothing. Specifically, it hangs at the line
unrolled_optimizer=unrolled_optimizer.minimize(unrolled_train_loss,var_list=unrolled_w_var)
It hangs for around 15 minutes until it goes on.
It there any way to speed it up? I'm using a server with Tesla V100. I don't think it's because my server is too slow.
Thanks!