Dispatching on variables w.r.t. which gradients are taken

[willtebbutt]

Hi,

Firstly, I've found the package really helpful, so thanks for porting it over from Python.

It appears to be the case that one cannot do the following:

using AutoGrad
f(x::Float64) = x^2
df = grad(f)
df(5.0)

I obtain the error ERROR: MethodError: no method matching foo(::AutoGrad.Rec{Float64}). If I have interpreted this correctly, it would appear that one cannot dispatch on the type of the arguments with respect to which we are taking gradients, without making it a primitive and defining the appropriate computations involving the Jacobian. Is there a simple way to resolve this?

Thanks,
Will

[denizyuret]

Hi Will, Autograd works by feeding your function boxed parameters. Therefore I recommend to keep the argument of grad a generic function. However if you need to use typed functions you can declare them as primitives as described in: https://github.com/denizyuret/AutoGrad.jl#extending-autograd hope this helps, deniz

…

On Mon, Jan 30, 2017 at 4:24 PM Will Tebbutt ***@***.***> wrote: Hi, Firstly, I've found the package really helpful, so thanks for porting it over from Python. It appears to be the case that one cannot do the following: using AutoGrad f(x::Float64) = x^2 df = grad(f) df(5.0) I obtain the error ERROR: MethodError: no method matching foo(::AutoGrad.Rec{Float64}). If I have interpreted this correctly, it would appear that one cannot dispatch on the type of the arguments with respect to which we are taking gradients. Is there a simple way to resolve this? Thanks, Will — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#8>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABvNpjpoKOH1bVD-Smz9-CsrKETMS1bmks5rXeSBgaJpZM4LxZXi> .

[willtebbutt]

Hi Deniz,

Thanks for the quick response. I figured that it might only be possible to dispatch on primitives, thanks for confirming.

Regards,
Will