For each example, select PV systems in the upper middle of the satellite imagery.
If an example doesn't have any PV systems then try just dropping that (I think variable batch sizes are allowed?) If that doesn't work then randomly select one of the examples which does have multiple PV systems, and select more than one PV system from that example (as separate examples: one example per PV system)
Randomly select one PV system per example
Compute the position encoding of each satellite pixel as the angle and distance (using OSGB coordinates) from the PV system.
Normalise the PV system and the satellite imagery
Some additional notes:
[x] Need to download hrvsatellite to data_for_testing.
[x] Write unit tests!
Context
Will be used for #20 (and others)
Implementation
This can't be done in the DataLoader class, because those classes only know about each individual data source. And we need to process the PV data and HRV satellite data together. Some options:
Subclass DataSet and override __getitem__
Add a data_processor attribute to DataSet which sees all the loaded data.
Detailed Description
Some additional notes:
hrvsatellite
todata_for_testing
.Context
Will be used for #20 (and others)
Implementation
This can't be done in the
DataLoader
class, because those classes only know about each individual data source. And we need to process the PV data and HRV satellite data together. Some options:DataSet
and override__getitem__
data_processor
attribute toDataSet
which sees all the loaded data.