Open MWesselkamp opened 2 years ago
AS, HP and LR choices not only based on min_mean_valloss but also min_sd_valloss. Accuracy and Precision. Now selected as: (mean_val_loss^2 + mean_val_std^2)/2
Residual network (evalres.py) Only GPP as input - why map all three to GPP obs? Changed: Standardize training and test set together.
Suggestion: Use year 2004, 2005, 2007 for NAS. Throw away afterwards. Use 2008-2012 for evaluation. Means, calibrating Preles at the same data. Calibration only on GPP will be biased?! Peltoniemi calibrated the 13 parameters simulateously at GPP and E. Way to do this: Loglik(GPP) + Loglik(ET) with standardized variables. Weighted? Why assuming gaussian likelihood.
At the moment only calibrated at GPP with gaussian density.
Why not use the same architecture for the naive, the parallel and the regularized network?
Two reasons why we should:
At the moment using default params for PRELES predictions on 2004 and 2005 Available as hyytialaNAS.csv
New datasets for utils.loaddata()
"NAS": hyytialaNAS.csv "validation": hyytialaF.csv "exp2": allsitesF.csv
Generated in file code/FitPRELES.R
What to discuss the next zoom: