Andrei Zhabinski
Andrei Zhabinski
@alexbw I'm definitely interested to port tensor gradients to Julia (e.g. see dfdx/Espresso.jl#2 for some details). Would you suggest any "entry point" to get started (either in code or in...
@alexbw Thanks for your answer. For the code you linked, am I right saying that gradients there are represented as Python/Lua functions that take previous gradients (i.e. gradients of arguments)...
Yes, this is expected. In both - Julia 0.5 and 0.6 operator `.*` is allowed on scalars and behaves the same way as `*`. The reason to use `.*` in...
I see two approaches. A simple approach would be to override [`getcode()`](https://github.com/dfdx/Umlaut.jl/blob/21a3affb606b3ec75176b65792d1046a218c53ae/src/trace.jl#L239-L250) to use a custom interpreter to extract code. It sounds doable. However, if the goal of tracing the...
> I got curious about this. Intuitively, tracing optimised typed IR will result in an optimised tape. Can you clarify what other changes are required to get better speed? There...
Yes, in general I'm open to such a PR. But since it's going to be a breaking change, we will need some time and effort to evaluate possible consequences and...
@JuliaRegistrator register
@JuliaRegistrator register
@JuliaRegistrator register
@JuliaRegistrator register