Lux.jl LSTM timeseries prediction: AD / optimization does not start

New to Julia
, , ,
2

I guess I found the issue: the map()call within the loss function

function loss_LSTM(p)
    return sum(abs2, y_target - vcat(map(x -> model(reshape([(i(x)*n(x)).^2/in_sqr_max; (i(x)).^2/i_sqr_max], 2, 1, 1), p, st)[1], tsteps_NODE)...))
end;

seemed to conflict with the AD calculation. Providing the input data as a simple pre-calculated array solved the issue.

1 Like