You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hello
I have a memory-constrained problem with a Lux.jl model that uses Zygote for most of the backpropagation.
I tried to approach this from chainrules perspective I need to checkpoint each Lux.jl layer in neural network. So I tried to achieve it like that :
function ChainRulesCore.rrule(::typeof(Lux.apply), l::Lux.AbstractExplicitLayer, x, ps, st)
y = Lux.apply(l, x, ps, st)
function pullback_checkpointed(Δy)
y, pb =Zygote.pullback(Lux.apply,l, x, ps, st)
return NoTangent(), pb(Δy)
end
y, pullback_checkpointed
end
Rule gets invoked in backpropagation Hovewer the issue is that for some reason it try recursively to do backpropagation of the first line
Hello
I have a memory-constrained problem with a Lux.jl model that uses Zygote for most of the backpropagation.
I tried to approach this from chainrules perspective I need to checkpoint each Lux.jl layer in neural network. So I tried to achieve it like that :
Rule gets invoked in backpropagation Hovewer the issue is that for some reason it try recursively to do backpropagation of the first line
so I get stack overflow error; how to correct it?
I had also posted this issue in https://discourse.julialang.org/t/avoid-storing-intermediate-results-from-the-forward-pass-by-default/119694/4?u=jakub_mitura
The text was updated successfully, but these errors were encountered: