commit begore push to atlas

[paraynaud]

• Fix the form of the package
• Verify the behaviour on Atlas

[paraynaud]

• KnetArrays is no longer supported.
• It is replace by CuArray.

[paraynaud]

erreur problématique :

LeNetNLPModel = KnetNLPModel(LeNet; data_train=(xtrn, ytrn), data_test=(xtst, ytst))
typeof(param_value) = Vector{CuArray{Float32, N, CUDA.Mem.DeviceBuffer} where N}
ERROR: MethodError: no method matching build_nested_array_from_vec(::Vector{CuArray{Float32, N, CUDA.Mem.DeviceBuffer} where N}, ::Vector{Float32})
Closest candidates are:
build_nested_array_from_vec(::KnetNLPModel{T, S, C}, ::Vector{T}) where {T, S, C} at /local1/raynpaul/code/KnetNLPModels.jl/src/utils.jl:55
build_nested_array_from_vec(::C, ::Vector{T}) where {C<:Chain, T<:Number} at /local1/raynpaul/code/KnetNLPModels.jl/src/utils.jl:56
build_nested_array_from_vec(::Array{CuArray{T, N, CUDA.Mem.DeviceBuffer}, 1}, ::Vector{T}) where {T<:Number, N} at /local1/raynpaul/code/KnetNLPModels.jl/src/utils.jl:74
Stacktrace:
[1] build_nested_array_from_vec(chain_ANN::Chainnll, v::Vector{Float32})
@ KnetNLPModels /local1/raynpaul/code/KnetNLPModels.jl/src/utils.jl:63
[2] KnetNLPModel(chain_ANN::Chainnll; size_minibatch::Int64, data_train::Tuple{Array{Float32, 3}, Vector{Int64}}, data_test::Tuple{Array{Float32, 3}, Vector{Int64}})
@ KnetNLPModels /local1/raynpaul/code/KnetNLPModels.jl/src/KnetNLPModels.jl:54
[3] top-level scope
@ REPL[20]:1
[4] top-level scope
@ ~/.julia/packages/CUDA/bki2w/src/initialization.jl:52