Use the code I wrote in December as a start. Planning to use relative position encoding (because the most important function of absolute position is the Sun's position; and we'll give the Sun's position explicitly in the query).
patch the imagery first. See issue #32.
Implementation ideas:
[x] implement as an np_batch_processor, which adds encodings for all modalities
[x] Need to add np_batch_processor after to_numpy in NowcastingDataset.
[x] Each modality will create BatchKeys like x_osgb, y_osgb, time_utc, and these will be masked with NaNs where appropriate (need to do this for PV).
Use the code I wrote in December as a start. Planning to use relative position encoding (because the most important function of absolute position is the Sun's position; and we'll give the Sun's position explicitly in the query).
patch the imagery first. See issue #32.
Implementation ideas:
np_batch_processor
, which adds encodings for all modalitiesnp_batch_processor
afterto_numpy
inNowcastingDataset
.x_osgb
,y_osgb
,time_utc
, and these will be masked with NaNs where appropriate (need to do this for PV).