Implicit loss function conversion in LossPhysical

[clessig]

What happened?

We currently have the mse is changed mse_channel_location_weighted implicitly here:

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

Line 52 in dfcf3e1

[getattr(loss_fns, name if name != "mse" else "mse_channel_location_weighted"), w, name]

These implicit things are hard to understand from the config. Does mse_channel_location_weighted default to an unweighted mse?

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

[clessig]

CC @Jubeku

[tjhunter]

this looks like a compatibility hack. can it be cleaned?

[Jubeku]

If there are no weights in the config initialized, function mse_channel_location_weighted effectively calculates the unweighted mse. So we can rename that function. Yet we have still a legacy mse function, see below. Can we delete it?

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

Lines 65 to 66 in dfcf3e1

	def mse(target, ens, mu, *kwargs):
	return torch.nn.functional.mse_loss(target, mu)

[clessig]

the config initialized, function mse_channel_location_weighted effectively calculates the unweighted mse. So we c

I think we can delete it then. And the line quoted above should have a comment explaining the situation. Then we can close the issue.