Conditional moment calculation

[malihass]

• Creating the model Sup3rCondMom and WindCondMom which do not include the discriminator details
• Batch handler provides output and mask on top oflow_res high_res. This is useful because sometime the high_res may not be directly obtainable from the output we are learning. The mask is useful if padding needs to be used.
• Different batch handler subclasses are implemented to learn conditional moments in different ways
  • SF: learns the conditional moment of the sub filter velocity SF=(HR-LR)
  • Mom1: learns first conditional moment (HR|LR) or (SF|LR)
  • Mom2: learns second conditional moment ((HR - (HR|LR))^2 | LR) or ((SF - (SF|LR))^2 | LR)
  • Mom2Sep: Learn second moment independently of the first moment (HR^2|LR) or (SF^2|LR)

[grantbuster]

@malihass, everything looks good on my end. Shame about the _tf_generate() vs. _tf_generate_wind() methods, i'm assuming you did this because of the multi-class inheritance conflict? Good solution but still a shame it couldnt be cleaner haha.

I'm approving but I still want @bnb32 to get a chance to look at the updates to the batch handling before you merge.

[bnb32]

@malihass, everything looks good on my end. Shame about the _tf_generate() vs. _tf_generate_wind() methods, i'm assuming you did this because of the multi-class inheritance conflict? Good solution but still a shame it couldnt be cleaner haha.

I'm approving but I still want @bnb32 to get a chance to look at the updates to the batch handling before you merge.

Is there really not a way to override that method with multi class inheritance?

[bnb32]

This looks great. Thanks! I just have some questions about the inheritance stuff and whether we can avoid the _wind suffix work around.

[bnb32]

_wind seems like the best solution since WindGan also inherits this stuff. Good work!