Closed jacobbieker closed 1 year ago
Hey @jacobbieker, would you like me to try to help fix this? Or is this issue no longer relevant because we're hoping to use data pipes instead of the old power_perceiver.load_prepared_batches
code?
Hi, I think we can probably leave this and just focus on the datapipes for it now, as this whole issue I think stems from the prepared batches directly and how they are set up.
Describe the bug
Some of the inputs seem to have differing numbers of timesteps, either 32 or 31 when they are being put through the FullModel, resulting in a concatenation errors on the gsp_query_generator. The first dimension, which should be exxamples * times ends up being 992 (32 x 31) for some of them and 1024 (32 x 32) for others
To Reproduce
This does require a slightly modified PreparedDataset to work past the #194 bug, so swap out that PreparedDataset with the one below. This version simply adds zeros for the topographic height, rather than trying to compute it with the Topographic processor.
Expected behavior A clear and concise description of what you expected to happen.
Additional context Add any other context about the problem here.