bugfix: force WeightNormalization to execute in order

[facaiy]

It seems that autograph is not smart enough, so we have to control execute order by ourselves.

[facaiy]

Following #453

[facaiy]

The new implementation still performs slightly better than the regression model (but worser than the precious implementation). I don't know what's wrong

https://colab.research.google.com/drive/1yXMehUQdFQ0mFtgJjYaj6FBCp8fr4Dt8#scrollTo=r66OUUrK_l_S

[facaiy]

I run it for five times, sometimes better, sometimes worser...

https://colab.research.google.com/drive/1yXMehUQdFQ0mFtgJjYaj6FBCp8fr4Dt8#scrollTo=UZd6V90eR4Gm

[seanpmorgan]

@facaiy I think it was for the wrong reason that it always outperformed using the previous version (at least through a very small number of epochs). Could you try to increase the number of epochs a significant amount? Also mind making the colab sharable?

[facaiy]

Oh, my fault. I make it sharable now, please check the link: https://colab.research.google.com/drive/1yXMehUQdFQ0mFtgJjYaj6FBCp8fr4Dt8

[WindQAQ]

LGTM! Thanks for the fix.