Deepar Notebook - Different outputs for different runs despite setting the seed.

Hi,

I've observed a problem of instability with the DeepAr algorithm as of deepar-electricity notebook. Basically when running twice, despite setting both a numpy seed and random seed, the error output changes, often even largely.

The issue can be reproduced by running the notebook twice and taking note of the cross-validation out using `estimator.training_job_analytics.dataframe()`

How is that possible?

Run the deepar-electricity notebook twice with the stated NumPy and random seeds, then compare the cross-validation output from estimator.training_job_analytics.dataframe(). Trace the notebook’s DeepAR training setup; done when the differing output is explained and the example has a verified, repeatable result.

Written by the indexing model from the issue text.