Meta-learning to Calibrate Gaussian Processes with Deep Kernels for Regression Uncertainty Estimation

Paper

Codes

Usage

• Download NA_NORM_6190_Bioclim_ASCII.7z from https://sites.ualberta.ca/~ahamann/data/climatena.html, which contains data of 27 Bioclimate variables (1961-1990 normal period)
• Decompress NA_NORM_6190_Bioclim_ASCII.7z in data/NA_NORM_6190_Bioclim_ASCII/
  % cd data/NA_NORM_6190_Bioclim_ASCII/
  % 7za x NA_NORM_6190_Bioclim_ASCII.7z #decompress data
• Preprocess data for AHM
  % cd ../..
  % python crt_data_spatial.py 100 #generate data for each region/attribute
  % python crt_traintest_spatial.py #generate data for meta-learning
• Run meta-learning
  % python train_meta_quantile.py --fn data/NA/I1l0.12v0.04t0.04o1/AHM/i0.npz --n_hiddens 32,32 --n_spt 10 --n_qry 30 --n_dataset 0 --epoch 1000 --method 1
  options:
  • method: calibrate or not (0:no calibration, 1:calibration)
  • n_spt: number of instances in a support set
  • n_qry: number of instances in a query set
  • n_dataset: number of tasks for training
  • n_hiddens: number of hidden units
  • epoch: number of epochs

License