use Lux.Training.single_train_step!

[lazarusA]

continue #181

[lazarusA]

running, but not green yet 😄

[lazarusA]

Dont think so, but let's see. Maybe check which tests are failing.

[lazarusA]

nice, it looks like this revealed a bug, parameterlength is ill defined somehow:

Running Literate.jl on 1 files...
Literate -> tutorials/folds.jl
generating markdown page from `~/work/EasyHybrid.jl/EasyHybrid.jl/docs/literate/tutorials/folds.jl`
writing result to `~/work/EasyHybrid.jl/EasyHybrid.jl/docs/src/tutorials/folds.md`
SetupBuildDirectory: setting up build directory.
Doctest: running doctests.
ExpandTemplates: expanding markdown templates.
ERROR: LoadError: MethodError: no method matching parameterlength(::Float32)
The function `parameterlength` exists, but no method is defined for this combination of argument types.

Closest candidates are:
  parameterlength(::Lux.Bilinear)
   @ Lux ~/.julia/packages/Lux/bF0gf/src/layers/basic.jl:558
  parameterlength(::LuxCore.AbstractLuxContainerLayer{layers}) where layers
   @ LuxCore ~/.julia/packages/LuxCore/qsnGJ/src/LuxCore.jl:226
  parameterlength(::LuxCore.AbstractLuxWrapperLayer{layer}) where layer
   @ LuxCore ~/.julia/packages/LuxCore/qsnGJ/src/LuxCore.jl:260
  ...

Stacktrace:
  [1] MappingRF
    @ ./reduce.jl:92 [inlined]

[lazarusA]

performance is bad, something is off 🤨 this PR vs latest dev. And similarly for other examples in the docs.

[BernhardAhrens]

I suspect the changes in st for the GenericHybrid to be the cause of this. Try constructing the hybrid model with input_batchnorm = false instead of true. That was the issue the first in the first go around for the st's a couple months ago. I would say revert those changes in GenericHybrid and see what happens

[BernhardAhrens]

I fixed the issue, see here: https://earthyscience.github.io/EasyHybrid.jl/previews/PR198/tutorials/exponential_res

I suggest to continue here #198

[lazarusA]

superseded by #198