[x] Handle examples where GSP and PV are both NaN. & examples where just PV is NaN.
[x] Mask attention queries into Satellite Transformer
[x] Mask attention queries into FullModel.time_transformer
[x] Mask loss
[x] Subselect the timesteps within the model, during training (mimic xr_batch_processor.reduce_num_timesteps, because we can't use that, because we don't have xr_batch_processors in raw, because it's hard to stack DataArrays with different PV system IDs.) But always use the full set of satellite images to train the SatellitePredictor.
[x] Move the SatellitePredictor training code into FullModel
[x] Remove GSP timestamp from GSP query generator, when used for satellite transformer
[x] Use multiple num_workers. Remove raw_dataset.per_worker_init. Try with the hack that allows fsspec to be forked. fsspec no longer has an asyn module. Need to read the fsspec github issues.
TODO to get model working
xr_batch_processor.reduce_num_timesteps
, because we can't use that, because we don't havexr_batch_processors
in raw, because it's hard to stack DataArrays with different PV system IDs.) But always use the full set of satellite images to train the SatellitePredictor.FullModel
num_workers
. Removeraw_dataset.per_worker_init
. Try with the hack that allows fsspec to be forked.fsspec
no longer has anasyn
module. Need to read thefsspec
github issues.min_duration_to_load_per_epoch