Transfer learning tutorial which uses ResNet returns NaN values

[natema]

I'm trying to run the transfer learning with Flux tutorial, but I get

┌ Warning: `haskey(::TargetIterator, name::String)` is deprecated, use `Target(; name = name) !== nothing` instead.
│   caller = llvm_compat(::VersionNumber) at compatibility.jl:176
└ @ CUDAnative ~/.julia/packages/CUDAnative/ierw8/src/compatibility.jl:176
ERROR: LoadError: LoadError: SystemError: unable to read directory /home/foo/transflearn/dataloader.jl/train: Not a directory

After explicitly writing the path to the dataset on the r.h.s. of the line

const PATH = joinpath(@__FILE__, "train")

in dataloader.jl, the dataset appears properly loaded, but then I get

┌ Warning: `haskey(::TargetIterator, name::String)` is deprecated, use `Target(; name = name) !== nothing` instead.
│   caller = llvm_compat(::VersionNumber) at compatibility.jl:176
└ @ CUDAnative ~/.julia/packages/CUDAnative/ierw8/src/compatibility.jl:176
[ Info: Epoch 1
┌ Warning: `Target(triple::String)` is deprecated, use `Target(; triple = triple)` instead.
│   caller = ip:0x0
└ @ Core :-1
[ Info: Epoch 2
2×10 CuArrays.CuArray{Float32,2,Nothing}:
NaN  NaN  NaN  NaN  NaN  NaN  NaN  NaN  NaN  NaN
NaN  NaN  NaN  NaN  NaN  NaN  NaN  NaN  NaN  NaN

If I write @show model(imgs) before the training, it correctly shows an array of Floats:

model(imgs) = Float32[0.3590678 0.35996166 0.3930334 0.4063199 0.42696851 0.41264024 0.37735277 0.3759735 0.37500587 0.4045543; 0.6409
322 0.6400384 0.6069666 0.5936801 0.5730315 0.5873597 0.62264717 0.62402654 0.6249941 0.5954457]