Improve upscaling CNN

[raspstephan]

Following on from #12 we can now try to improve the upscaling CNN (i.e. with MSE instead of GAN loss). Here are a few ideas:

• Include more input variables (constants + TIGGE variables)
• Log-transform precipitation (maybe this will help with generalization since rare events are less extreme?)
• Data augmentation, e.g. flipping Add data augmentation options in dataloader #22
• Use ensemble input Use TIGGE ensemble for training and vaidation #19
• Try different upsampling algorithms, e.g. bilinear
• Try a UNet-like architecture @anna-184702

The goal should be to comprehensively beat a pure linear interpolation.

[HirtM]

@raspstephan, regarding the checkerboard pattern you mentioned, this video states that applying upsampling followed by a convolution reduces that problem:

https://www.coursera.org/learn/build-basic-generative-adversarial-networks-gans/lecture/H02dK/transposed-convolutions

But it seems, you are already doing that in your CNN, right?
Here is also a more detailed discussion about the problem, but I haven't read it myself thoroughly. And it's also from 2016, so I don't know if it is really helpful.