Problem with AutoGrad while implementing Hungarian Huber Loss

[AHMED IMAM SHAH]

Hello Everyone,

I am trying to implement the following function hungarian huber loss:

function huberloss(y, y_hat; delta=1)
    absdiff = abs.(y.-y_hat)
    loss_activation = (absdiff .>= 1)
    loss = loss_activation.*(absdiff.-0.5*delta).*delta + (loss_activation.*(-1).+1).*(0.5.*absdiff.^2)
    huberloss = sum(loss)/length(y)
    return huberloss
end

function hungarian_huber_loss(labels, predictions)
    batch_size, num_objects, object_labels = size(labels)
    batch_loss = 0
    for b in 1:batch_size
        actual_costs = 0
        pairwise_cost = zeros(num_objects, num_objects)
        for i in 1:num_objects
            for j in 1:num_objects
                pairwise_cost[i,j] = huberloss(labels[j,:,b], predictions[i,:,b])
            end
        end
        copy_pairwise_cost = copy(pairwise_cost)

        indices = zeros(num_objects,1)
        max = 30
        for i in 1:num_objects
            indexmin = argmin(copy_pairwise_cost)
            r = indexmin[1]
            c = indexmin[2]
            indices[r] = c
            copy_pairwise_cost[r,:] .= max
            copy_pairwise_cost[:,c] .= max
        end

        for i in 1:num_objects
            actual_costs += pairwise_cost[i, Int(indices[i])]
        end

        batch_loss += actual_costs/num_objects
    end

    return batch_loss/batch_size
end

However, the bolded line is giving errors in back propagation. I faced a similar problem when I was using the padding function which I got away with by using cat function. Here I don't know what to do. Here is the stacktrace:

Stacktrace:

 [1] setindex! at ./array.jl:849 [inlined]

 [2] hungarian_huber_loss(::KnetArray{Float32,3}, ::AutoGrad.Result{KnetArray{Float32,3}}) at ./In[16]:17

 [3] loss(::Chain, ::KnetArray{Float32,4}, ::KnetArray{Float32,3}) at ./In[18]:3

 [4] (::var"#12#13")() at /kuacc/users/ashah20/.julia/packages/AutoGrad/TTpeo/src/core.jl:205

 [5] differentiate(::Function; o::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at /kuacc/users/ashah20/.julia/packages/AutoGrad/TTpeo/src/core.jl:144

 [6] differentiate(::Function) at /kuacc/users/ashah20/.julia/packages/AutoGrad/TTpeo/src/core.jl:135

 [7] top-level scope at In[20]:1

 [8] include_string(::Function, ::Module, ::String, ::String) at ./loading.jl:1091

 [9] softscope_include_string(::Module, ::String, ::String) at /kuacc/users/ashah20/.julia/packages/SoftGlobalScope/u4UzH/src/SoftGlobalScope.jl:65

 [10] execute_request(::ZMQ.Socket, ::IJulia.Msg) at /kuacc/users/ashah20/.julia/packages/IJulia/IDNmS/src/execute_request.jl:67

 [11] #invokelatest#1 at ./essentials.jl:710 [inlined]

 [12] invokelatest at ./essentials.jl:709 [inlined]

 [13] eventloop(::ZMQ.Socket) at /kuacc/users/ashah20/.julia/packages/IJulia/IDNmS/src/eventloop.jl:8

 [14] (::IJulia.var"#15#18")() at ./task.jl:356

MethodError: Cannot `convert` an object of type AutoGrad.Result{Float32} to an object of type Float64
Closest candidates are:
  convert(::Type{T}, !Matched::Ratios.SimpleRatio{S}) where {T<:AbstractFloat, S} at /kuacc/users/ashah20/.julia/packages/Ratios/xLeZh/src/Ratios.jl:52
  convert(::Type{T}, !Matched::Static.StaticInt{N}) where {T<:Number, N} at /kuacc/users/ashah20/.julia/packages/Static/A5kpr/src/int.jl:22
  convert(::Type{T}, !Matched::Static.StaticFloat64{N}) where {N, T<:AbstractFloat} at /kuacc/users/ashah20/.julia/packages/Static/A5kpr/src/float.jl:26
  ...

Stacktrace:
 [1] differentiate(::Function; o::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at /kuacc/users/ashah20/.julia/packages/AutoGrad/TTpeo/src/core.jl:148
 [2] differentiate(::Function) at /kuacc/users/ashah20/.julia/packages/AutoGrad/TTpeo/src/core.jl:135
 [3] top-level scope at In[20]:1
 [4] include_string(::Function, ::Module, ::String, ::String) at ./loading.jl:1091

Looking forward for any suggestions. Best regards,

Ahmed Imam Shah

MS. Computer Science and Engineering

Koç University, Istanbul, Turkey

[Shadi Sameh Hamdan]

Hello Ahmed, 

First of all, you initialize the zeros matrix without specifying the type, that's going to give you trouble often so you need to be careful.

More importantly, the issue is that you are initializing a regular matrix (float32/64/etc), and you are trying to assign an element of this matrix (which is a float) to a Tape, which is something that contains gradient flow. That cannot be done this way as they are different types. Unless you find a way to vectorize this operation, it is not a good idea to store these elements in a matrix that way. 

Can you reimplement this or think of a different way to formulate it? 

Best, 

Shadi

--
You received this message because you are subscribed to the Google Groups "COMP541" group.
To unsubscribe from this group and stop receiving emails from it, send an email to COMP541+u...@ku.edu.tr.
To view this discussion on the web visit https://groups.google.com/a/ku.edu.tr/d/msgid/COMP541/CA%2Bsbv-hZcjB3JxmkAyqmZExzk6yDnWk%3D7n3EfbuAwf1u5Ne1mQ%40mail.gmail.com.