Unit test Failed: Error while reading resource variable from Container

[netmanaiops]

Unit test Failed: Error while reading resource variable from Container

[haowen-xu]

@maye9999 Better close the current issue and start a new one, since it might be helpful to other participants. If you do so, I will copy and paste this comment to the new issue.

• Yes, it will handle the sliding windows internally. The reason of this is that Donut applies missing data injection, which might be better applied on the original 1-D series, rather than the sliding windows.
• Yes, you can. Also, you need to pass in numpy arrays, not TensorFlow tensors.
• As the training goes by, the std estimation of q(z|x) and p(x|z) would probably decrease to a very small value, such that if an unlabelled anomaly presents in the training data, it will cause severe confusion to the gradient at that point (recall that log Normal(..) contains a -(x-mean)/std^2 term). So if you have labels, it will definitely help to stabilize training, thus make the model converge to a better local optimum. When you do not have labels, the missing data injection part of the algorithm will help the model gain some ability to adapt to such noises. And with the help of some other techniques (e.g., gradient clip by norm), the algorithm can still work to some extent.

[maye9999]

I downgrade Tensorflow to 1.8 and it works. So maybe some API changed in 1.9?

However, when I run the file test_model.py on Tensorflow 1.8, some new error occurs.

....F
======================================================================
FAIL: test_variable_reuse (__main__.ModelTestCase)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "tests/test_model.py", line 55, in test_variable_reuse
    'get_donut/donut/q_z_given_x/z_std/kernel']
AssertionError: Lists differ: [u'get_donut/donut/p_x_given_z... != ['get_donut/donut/p_x_given_z/...

First differing element 0:
u'get_donut/donut/p_x_given_z/mean/dense/bias'
'get_donut/donut/p_x_given_z/x_mean/bias'

Diff is 1428 characters long. Set self.maxDiff to None to see it.

----------------------------------------------------------------------
Ran 6 tests in 1.048s

FAILED (failures=1)

Could you please check it too? Thank you very much!

[haowen-xu]

Actually the Python 2.7 + TensorFlow 1.8 already passes unit tests on Travis CI (https://travis-ci.org/haowen-xu/donut/jobs/406086706).

It seems to be related to a bug I fixed today (see bef5575), which has been existing long before I noticed it. The old version would construct a set of new variables each time the encoder or decoder of the model is built, but I meant to reuse the previous constructed ones. You may upgrade to the newest version to have a try.

If you just cloned it from git, do git pull. Otherwise if you also installed it, please do pip install --upgrade git+https://github.com/haowen-xu/donut.git.

[maye9999]

Thanks for your reply.

By the way, I'd like to ask some questions about the inputs of the model.

• The values for DonutTrainer.fit and DonutPredictor.get_score is 1-D vector and do I need to handle the sliding window? Or the sliding windows is handled internally in Trainer and Predictor?
• When I don't have missing points, could I just set the missing to zeros_like(values)?
• The readme says If there is no label, simply use all zeros. I assume VAE models are un-supervised, so I don't understand the usage of labels in DonutTrainer.fit. Is it simply used for evaluate the model, or is it involved in some training process?

Thank you very much and looking foward to your reply!