Train a categorical model

[raspstephan]

It might make sense to also look into a Metnet-style categorical model. I am pretty sure that this would work. It would be a) a good baseline and b) a good fallback option if we never get a GAN to train.

I would like to try sampling from the probability output with a correlated random field as well. This might actually end up looking quite realistic and, I suspect, hard to beat score-wise with a GAN.

[raspstephan]

I got the first working end-to-end version going. Training a categorical net, then sampling from it with a 2D correlated random field. So far, I have not actually trained it on the full data and with enough bins but here is what this approximately looks like.

Next steps are: 1) Copy changes into src; 2) Train on full data with more bins; 3) Evaluate it; 4) Improve net by adding more variables, etc.

[raspstephan]

Early results from notebooks 01 and 02. 01 with bn, 02 without. Basically no difference, two epochs seem to be enough.

Very zero heavy and a little noisy but seems reasonable.

Next steps: real validation + make better with more input variables.

[raspstephan]

Adding orography and the LSM (3 + 4) on their own doesn't really change the skill yet at all. Maybe without additional information in terms of variables this isn't enough info. I would still have expected orography to hep a little. Hmm. But maybe the uncertainty is also just way too large for this to make any impact? Let's add both and add some more variables, one by one.

[raspstephan]

Also, try the better loss!!!

[raspstephan]

Adding temperature seems to help a little (05).