This PR adds the capability to remove specific training sites from your testing metrics in the case that train and test periods have overlapping years but distinct sites. Also changed on how the device in determined in predict_torch. In the previous implementation, it pulled the device based on the model parameters, but since model parameters can be spread across multiple devices it was ambiguous.
This PR adds the capability to remove specific training sites from your testing metrics in the case that train and test periods have overlapping years but distinct sites. Also changed on how the device in determined in
predict_torch
. In the previous implementation, it pulled the device based on the model parameters, but since model parameters can be spread across multiple devices it was ambiguous.