Writing loss and parameters in function reconstruct_points

[ctroupin]

Using the code from the function reconstruct( )

DINCAE.jl/src/model.jl

Lines 552 to 573 in 06bdb3f

	if paramfile !== nothing
	NCDataset(paramfile,"c") do ds_
	defVar(ds_,"losses",losses,("epochs",))
	ds_.attrib["epochs"] = epochs
	ds_.attrib["batch_size"] = batch_size
	ds_.attrib["truth_uncertain"] = Int(truth_uncertain)
	ds_.attrib["enc_nfilter_internal"] = Vector{Int}(collect(enc_nfilter_internal))
	ds_.attrib["skipconnections"] = Vector{Int}(collect(skipconnections))
	ds_.attrib["clip_grad"] = clip_grad
	ds_.attrib["regularization_L1_beta"] = regularization_L1_beta
	ds_.attrib["regularization_L2_beta"] = regularization_L2_beta
	ds_.attrib["save_epochs"] = Vector{Int}(save_epochs)
	ds_.attrib["is3D"] = Int(is3D)
	ds_.attrib["upsampling_method"] = string(upsampling_method)
	ds_.attrib["ntime_win"] = ntime_win
	ds_.attrib["learning_rate"] = learning_rate
	ds_.attrib["learning_rate_decay_epoch"] = learning_rate_decay_epoch
	ds_.attrib["min_std_err"] = min_std_err
	ds_.attrib["loss_weights_refine"] = Vector{Float64}(collect(loss_weights_refine))
	ds_.attrib["cycle_periods"] = Vector{Float64}(collect(cycle_periods))
	end
	end

to also have the same functionality in reconstruct_point.

Comments:

1. not all the parameters of reconstruct are used in reconstruct_point`, and vice versa
2. I had to change losses = Float64[] to losses = [], otherwise an error appeared:

└ @ Main ~/CPR-DINCAE/src/run_DINCAE_testparams.jl:118
ERROR: LoadError: MethodError: no method matching fillvalue(::Type{Any})

Closest candidates are:
  fillvalue(::Union{NCDatasets.MFCFVariable{T}, NCDatasets.MFVariable{T}}) where T
   @ NCDatasets ~/.julia/packages/NCDatasets/st9Jz/src/multifile.jl:289
  fillvalue(::CommonDataModel.CFVariable)
   @ CommonDataModel ~/.julia/packages/CommonDataModel/RSBF3/src/cfvariable.jl:200
  fillvalue(::NCDatasets.Variable{NetCDFType, N}) where {NetCDFType, N}
   @ NCDatasets ~/.julia/packages/NCDatasets/st9Jz/src/variable.jl:255

probably because the type of losses was Vector{Any}.
3. A few parameters from reconstruct are not written in the file, I'll do it in another PR.

DINCAE.jl/src/model.jl

Lines 351 to 357 in 06bdb3f

	cycle_periods = (365.25,), # days
	output_ndims = 1,
	direction_obs = nothing,
	remove_mean = true,
	paramfile = nothing,
	laplacian_penalty = 0,
	laplacian_error_penalty = laplacian_penalty,

4. Shall we consider the possibility to write the parameters in the same file as the result?

[Alexander-Barth]

Thanks a lot, this is fine!
For point 4: In fact, it used to be the same file bit I changed this here (b7ed446 ) because, I typically make a lot of tests and when I need to clean-up the files I prefer to just delete the reconstructions and keep some the error statistics and the network configuration (so that I could repeat the reconstruction if necessary).

But we can cover both cases if the user sets ( paramfile = "results.nc") and we are careful to not erase the results with jkust the parameters :-)

ncmode = (paramfile in fnames_rec ? "a" : "c")
NCDataset(paramfile,ncmode) do ds_
...

[ctroupin]

Yes you're right for 4., in fact I'm running now with the values written in an extra file, it's really safer.