Add gaussian processes to iota2 classification workflow
This issue will list the tasks required to use Gaussian processes in iota2.
-
find an iota2 compatible version of gpytorch -> 1.2.0 from conda-forge -
prepare samples from samplesExtraction to train the gaussian processes: -
encode labels from 0 to C-1 where C is the number of classes (only encode value not use onehotencoder) -
split between train, val subset using the iota2 training set with all classes in each set -
standardize the samples before train with scikit learn
-
-
train using a dataloader (same as in #317 (closed), later add coordinates information) -
Do the inference to produce maps over tiles: -
Standardize the features before inference -
Decode labels -
Generate confidence maps (max probability or difference between the two higher) -
If needed, generate probability map
-
-
Test on real use case the complete workflow (several tiles and 1 year of data)
This feature branch will be based, and follow advances on issue #317 (closed) branch, as several features from torch are common
Edited by Tardy Benjamin