Flux 0.10.0 does not update function definitions in Jupyter

Specific Domains Machine Learning
, ,
1

I stumbled onto this behaviour recently and I’m unsure if it’s intentional, or what to do about it. If I update a function definition, invocations of gradient() seem to use the old definition even as the rest of the notebook sees the update.

I am using Julia 1.1.1, Flux 0.10.0, Zygote 0.4.5, and IJulia 1.20.0.

Running a cell with

using Revise
using Flux

function f(x)
    return x[1]^2
end

x=[3]
gds = gradient(() -> f(x), params(x))
gds.grads

will work correctly. If I change f to return x[1]^3, instead it will not. The gradient invocation will still work off the old definition, even though the rest of the notebook sees the new one. Is this intentional? Is there a command I can use to force Flux/Zygote to recompile the function without restarting the kernel entirely?

2

You need Julia 1.3 for this to work automatically. I think Zygote.refresh() is the manual way.

1 Like