This PR includes two separate things I've been using locally
First and most importantly, we were calculating the wrong history mins for the validation dataloader from the number of history steps the model requires. If history_mins=0 then the dataset returns a single satellite timestep at time t0 for the input X. This corresponds to history_steps=1.
I think it is better practice to use the xr.concat() function for merging multiple datasets since we are only merging along a single axis. Plus previously we were not sorting by time which could be an issue if the user passed in yearly satellite zarrs not in order of increasing years
This PR includes two separate things I've been using locally
history_mins=0
then the dataset returns a single satellite timestep at time t0 for the input X. This corresponds tohistory_steps=1
.xr.concat()
function for merging multiple datasets since we are only merging along a single axis. Plus previously we were not sorting by time which could be an issue if the user passed in yearly satellite zarrs not in order of increasing years