Issue/166 batch pydantic

[peterdudfield]

Pull Request

Description

• Pydantic model for batch - removed Examplep[dict] object
• Pydantic model for eahc data source
• tools for changing model to xarray (and back)
• adjust DataSets to use pydantic models
• new 0.nc test/data
• sort out subselect data
• sort out scripts

Fixes #166

How Has This Been Tested?

• Added specific unitests

• create Fake Dataset and check torch tensor is returned (unittest)

• ran scripts prepare ml data (and validate script)

• No

• Yes

Checklist:

• My code follows OCF's coding style guidelines
• I have performed a self-review of my own code
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• I have checked my code and corrected any misspellings

[JackKelly]

Looks good!

Lots of comments, but they're mostly pretty tiny!

A few general thoughts:

1. Wherever possible, please consider removing any dependencies on torch, in preparation for #86. e.g. please consider using np.random.randn instead of torch.randn.

2. I think I've made nowcasting_dataset over complicated by starting nowcasting_dataset with the intention of loading data on-the-fly during ML training; and then swapping to using nowcasting_dataset to create on-disk pre-prepared batches. I think we can simplify things a lot by removing support for loading on-the-fly :) (i.e. I think we can safely say that, from now on, nowcasting_dataset will just be used for pre-preparing batches! Please see #209 - I think it could reduce the code size a lot :) And I feel really bad for not thinking about this earlier!

[JackKelly]

Do any instances of DataSourceOutput have one or more fields with five-minutely and one or more fields with half-hourly datetime indexes? I guess not?

If that is a possibility, then start_i and end_i would be correct indexes into one of the indexes; but would be wrong indexes into the other datetime index, I think?

[peterdudfield]

Currently data sources only have 5 mins, or 30 mins. I wonder if this is a good enough for the moment. If so I can add a comment -

or can we think of a case where there will be both 5 mins and 30 min data?

[JackKelly]

This looks great!

Random thought, that isn't especially relevant to this PR(!), but it's possible that we might want to remove these datetime features from the on-disk batches, and instead compute these features on-the-fly, not least because we'll want to experiment with a variety of different ways of encoding position when we really start experimenting with the perceiver IO models. Related issue: https://github.com/openclimatefix/nowcasting_utils/issues/30

[JackKelly]

I've started a separate issue to discuss this: #208

[JackKelly]

But, even if we do compute these features on-the-fly, this pydantic model (and the validation code) could still be used for the datetime features computed on the fly, I guess?