jessebett

Oh, this error requries `@nograd factorial` fix from @willtebbutt

Specifically, the dev docs describe how defining methods for adjoints on functions like `height` and `width` which provide a functional interface for `getfield` will address this. But `Taylor1` type in...

@willtebbutt I'm a bit confused by what the above suggestion is, sorry. I'm trying to follow this strategy but Zygote is still throwing `getfield` errors, so I don't see where...

Right, okay I thought the `literal_getpropery` was some weird Cholesky thing. Another question, in the [examples you linked](https://github.com/FluxML/Zygote.jl/blob/ef549246918f6d013c9d38112a4479e093337f6b/src/lib/array.jl#L198) it looks like it's actually dispatching on the value of the property...

@JoshChristie that is a different operation: ``` f([1.,1.,1.]) # Tracked 3-array Flux.data(f([1.,1.,1.])) # 3-array [1.,1.,1.]'*Flux.data(f([1.,1.,1.])) # Float [1.,1.,1.]'.*Flux.data(f([1.,1.,1.])) # 3x3 Array ```

This is still a problem @MikeInnes: ``` using Flux x = param(randn(3)) v= [1., 0., 0.] v'*x # Method Error collect(v')*x # Correct. ``` Flux still isn't workign well with...

Yes. I'm around tomorrow and will also be at the hackathon on Friday working on Zygote stuff. The `kron` in `Zygote` is just some `reshape` rules. I did not benchmark...