No method matching next(::AutoGrad.Rec{Dict{Symbol,Any}}, ::Int64)

[ngphuoc]

I am trying to add L2 regularization at the end of s2s() method in prof/s2s.jl and get the mentioned error. The added line is

sumlogp += 0.001 * sum(sumabs2(a) for (k,a) in model if isa(a, KnetArray))

[denizyuret]

I will look into this. In the meantime a workaround is to replace the dictionary with an array.

…

On Thu, Apr 6, 2017 at 6:58 AM ngphuoc ***@***.***> wrote: I am trying to add L2 regularization at the end ofs2s() method in prof/s2s.jl and get the mentioned error. The added line is sumlogp += 0.001 * sum(sumabs2(a) for (k,a) in model if isa(a, KnetArray)) — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#109>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABvNpvT9H8I_NH-AwWYABjNuhSPSGy2vks5rtGLjgaJpZM4M1Fym> .

[denizyuret]

I had a chance to look at this in more detail, there were several problems:

• AutoGrad does not support Dict iteration. This can be fixed by:

Base.next{T<:Associative}(a::Rec{T},i) = ((kv,j)=next(a.value,i);(k,v)=kv;(k=>a[k],j))

I will add this to AutoGrad.

• isa(a, KnetArray) won't work, you need isa(getval(a),KnetArray). When you index elements of a model AutoGrad changes their type to keep track of dependencies.
• Finally this won't work either because the model has a hierarchical structure and not all values are KnetArrays. If you really want to calculate L2, I'd write a function that explicitly goes through model[:embed1], model[:encode][1], model[:encode][2] etc.

[denizyuret]

Fixed in latest AutoGrad master.