Closed CarloLucibello closed 5 years ago
I see this issue has also been reported in #101. This is big impairment for me, because I use a lot of complicated custom operators in my research code. @denizyuret do you have a solution for this or any hint on the problem (what's wrong in the broadcasting pipeline)?
The problem is I haven't figured out how to define gradients for arbitrary broadcast operations for AutoGrad or Knet. For primitives (e.g. sin/cos), I define both the scalar and broadcasted gradients with the @primitive macro in AutoGrad and call specific kernels for KnetArray. @ekinakyurek thinks this used to work in Julia 0.6 but we need to figure out how to make it work in Julia 1.0.
Solved in latest master, submitted in v1.1.2.
I was trying to port some of my knet code to julia 1.0, and incurred in the following very serious bug/ Here is a MWE: