How could we control Random State?

[yuchenlin]

Hi, Guillaume,

I am very confused about why the training result is quite random since it seems that there is no shuffling of the data in the code.

I tried to set tf.set_random_seed(42) and np.random.seed(42) at the beginning of model.py, but it does not work.

Thanks very much,
Bill

[yuchenlin]

OIC.. the random effect seems to be due to the Dropout.

[guillaumegenthial]

You're right, there is no shuffling anywhere. Randomness surely comes from dropout + initialization. However, if you carefully control the random seeds, you should get non-random convergence!

[yuchenlin]

Thanks so much. Although simply setting the random state does not work, I found a page here regarding the exactly the same question.
I will have a try!

[SeekPoint]

maybe you should also set python's random seed fixed

[yuchenlin]

It seems that even if I set dropout=1, the results are still changing.

[yuchenlin]

@lovejasmine i have tried but not effective...: (

[yuchenlin]

I tried use different tf random seed and I found that the results are very different than before.
Thus, it seems that setting random seed is effective to some extent but cannot keep the results 100% exactly the same. I guess it maybe due to the multi-thred thing.

It seems that the tensorflow is often complained by users about the random state stuff..

[guillaumegenthial]

Have you tried running it on CPU on a subset of the dataset with a batch of size 1 with dropout = 1? I'm not too familiar with tf random seeds, and the stackoverflow post you refered above seems to mention some pretty weird behavior. Batch size = 1 could alleviate the problem they're mentioning.

[grafael]

@yuchenlin maybe this help: https://stackoverflow.com/questions/38469632/tensorflow-non-repeatable-results

[guillaumegenthial]

@yuchenlin did you make progress on this issue? Should we keep it open or should we close it?

[yuchenlin]

@guillaumegenthial sorry for the delayed reply! I was busy doing other stuff. Yes, I guess we can close it. It seems that the current methods are difficult to maintain the stable results under the GPU settings. Thanks!

[guillaumegenthial]

@yuchenlin no problem! I'll close it, but I must confess that this is an interesting issue and I'd love to hear more about it if you find the answer one day.