search

denizyuret / Knet.jl

Koç University deep learning framework.
https://denizyuret.github.io/Knet.jl/latest
Other
1.43k stars 230 forks source link

Knet.Layers20 #607

Open denizyuret opened 4 years ago

denizyuret commented 4 years ago

Copy of KnetLayers?

andevellicus commented 4 years ago

Could 3D functions (5D tensors) please be supported?

denizyuret commented 4 years ago

I think they are (I have published models with 3d convolutions).

On Wed, Sep 2, 2020 at 6:30 PM andevellicus notifications@github.com wrote:

Could 3D functions (5D tensors) please be supported?

— You are receiving this because you were assigned. Reply to this email directly, view it on GitHub https://github.com/denizyuret/Knet.jl/issues/607#issuecomment-685814175, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAN43JXIDUQ737VMASOXV7TSDZQKDANCNFSM4QNTUR7Q .

andevellicus commented 4 years ago

At the moment KnetLayers' implementation of CNNs is restricted to 2D inputs:

function Filtering{T}(;height::Integer, width::Integer, inout::Pair=1=>1,
                       activation::ActOrNothing=NonAct(),
                       winit=xavier, binit=zeros,
                       atype=arrtype,
                       opts...) where T <: Function

    wsize = T===typeof(conv4) ? inout : reverse(inout)
    w = param(height,width,wsize...; init=winit, atype=atype)
    b = binit !== nothing ? Bias(1,1,inout[2],1; init=binit, atype=atype) : Bias(nothing)
    Filtering{T}(w, b, activation; opts...)

end

In my code I just changed height and width to a tuple.