GRU-D icon indicating copy to clipboard operation
GRU-D copied to clipboard

Published 20 hours ago • verified publisher icon zhiyongc

Trainable decay scheme issue

Open CameronSCarlin opened this issue 6 years ago • 4 comments

https://github.com/zhiyongc/GRU-D/blob/9f877f8730322a9b74a4df1bfb9d245cca968b61/GRUD.py#L120

Hi, My understanding of Equation 11 in the GRU-D paper is that the two x values on the right side of this equation are supposed to be different values. I believe mask * x to be correct, as that is multiplying x of the current time step. However, I believe delta_x * x is incorrect, as this x should be the value of x most recently observed prior to the current time step t.

CameronSCarlin avatar Jan 03 '19 19:01 CameronSCarlin

This is worth fixing! Otherwise it's decaying the current imputed value.

Thartvigsen avatar Jan 20 '19 15:01 Thartvigsen

@Thartvigsen I noticed this issue not fixed yet in the code!!!!

2M-kotb avatar Feb 02 '19 13:02 2M-kotb

@DeepWolf90 For now, update the equations in your local version so they are accurate. I'm sure the author is aware of the bug by now. Hopefully it'll be sorted out soon!

Thartvigsen avatar Feb 02 '19 21:02 Thartvigsen

Hi @CameronSCarlin, Thanks for posting the flaw in the code. You are definitely correct. The issue is fixed in the updated version. Also, thanks for your reminding and suggestions! @Thartvigsen and @DeepWolf90.

zhiyongc avatar Feb 10 '19 08:02 zhiyongc