Handling of diagnostic and forcing datasets should happen in masker and not in tokenizer

[clessig]

What happened?

Handling of diagnostic (no source channels) and forcing (no target channel) datasets is currently done in the tokenizer:

WeatherGenerator/src/weathergen/datasets/tokenizer_masking.py

Line 130 in 8f7f240

if is_diagnostic or rdata.data.shape[1] == 0 or len(rdata.data) < 2:

WeatherGenerator/src/weathergen/datasets/tokenizer_masking.py

Line 167 in 8f7f240

is_forcing = stream_info.get("forcing", False)

This leads to an inconsistency between the masks and the actual data, e.g. here:

WeatherGenerator/src/weathergen/train/loss_modules/loss_module_physical.py

Line 242 in 8f7f240

# preds_batch can be empty and output_info not for forcings

Hence, the handling of forcing and diagnostic datasets should be done in the masker, with empty target and sources masks being generated in these cases, respectively.

CC @wael-mika @shmh40

What are the steps to reproduce the bug?

No response

Hedgedoc link to logs and more information. This ticket is public, do not attach files directly.

No response