I want to try band differences as an additional input (like the Sentinel cloud algorithm), but I get loss as nan.0
[06:57<00:00, 2.81it/s, loss=nan, v_num=3, loss_step=nan.0, iou_step=0.000, iou_epoch=0.000, loss_epoch=nan.0]
I modified the input channels to 5, and did the band difference and appended to the new band array on the fly.
Am I missing anything ? My expectation is that similar to a new band input (for eg SWIR band) that is physically on the disk, this ought to improve.
I’m using the Unet benchmark code given by the organizers.