Adapt batch fix

[dfulu]

Pull Request

Description

There was an error in how we use the _adapt_batch(). Previously we used to edit the batches in-place but this was flagged to change in this issue and fixed in this PR.

When this the _adapt_batch() was updated this introduced an error where the batch was updated for the forward pass of the model, but not for calculating the MAE and for backprop. This meant that when training for an intraday model using batches which go out to day ahead - the model was actually being trained the predict the last 8 hours of the 36 hour window.

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

[peterdudfield]

Aproved, could you either add a test for this? Or add a github issue to test this?

[dfulu]

Aproved, could you either add a test for this? Or add a github issue to test this?

There isn't really an obvious test that could have caught this. The mistake was in the slicing inside the training and validation steps. Those train/val steps were passing but just weren't correctly indexed internally.

We'd need quite a big refactor to unit test this bit, or a really convoluted test to make sure the MAE values returned by the train/val steps were as expected. So I don't think this is worth testing