Andrei Zhabinski

If eventually all operations implemented in ONNX.jl will be based on NNlib, do we actually need multiple backends? If the main goal at the moment is to have at least...

I see 2 parts here: 1. Encoder/decoder API 2. Generated Julia definitions of ONNX data types If the generated definitions haven't changed too much, we can try to simply update...

Old definitions didn't work out of the box, and generating the new ones is currently blocked by JuliaIO/ProtoBuf.jl#211 . I will try again once it's resolved.

Thanks @Drvi! I'm glad to see how consistent and robust PB v1 is. I'm currently refactoring the code in #74. Due to the change in semantics of some types (notably,...

Finally fixed in #75

Loop tracing is now supported via Ghost.jl. Differentiating is on the roadmap.

Thanks for reporting this! It turns out, Pluto adds some magic to the call so it's traced differently from a normal REPL. In REPL: ```julia julia> Yota.Umlaut.trace(x -> sum(x .+...

Yes, I thought about it, but I don't see a straightforward way to do it just yet - unwrapping tracked data helps in finding gradients, but how do you use...

You are absolutely right - there's no way to represent a single primitive `f(args...)` as a tape, at least as a tape different from the one for `args -> f(args...)`....

> I also expected that if you have function f(x); g(x) end then trace(f,x) and trace(g,x) would be the same. `trace()` already works like this: ```julia julia> g(x) = 2x...