Split ratio and inconsistency removal for ensemble

[sfluegel05]

To use the chebai-models for the chebifier-ensemble, I have made some minor changes:

• Split ratios can be set independently for validation and test set - this allows e.g. for a larger validation set which we need to calculate metrics or (potentially) do some ensemble-training
• Prediction smoothing is now more general - we use this in the ensemble to ensure all predictions are consistent

[aditya0by0]

@sfluegel05,

Before this PR, with the default values,

• Test split: 12.75% of the data
• Validation split: ~2.309% of the data
• Train split: ~84.94% of the data

After this PR, with the new default values,

• Test split: 10% of the data
• Validation split: ~5.004% of the data
• Train split: ~84.9999% of the data

Shouldn’t we adjust the new default values to maintain the same data distribution as before the change?

[sfluegel05]

Well, changing the default values kind of was my intention. I don't think there has ever been an actual reason for the validation set being 15% of 15%. I can't think of a scenario where it is important to have the same split sizes but where you are not re-using the actual splits.
The advantage of the new ratio is that you get more validation data (for ChEBI50, the smallest classes have only 50 members, meaning 2.3% equates to 1 positive sample in the validation set, which is not that expressive).

[aditya0by0]

#99 (comment)
#97 (comment)

The unit-test failing issue originates from this PR, which was merged into dev even though its test cases were failing at the time. The root cause is that self.train_split was replaced with self.test_split and self.validation_split.
In the test case, self.train_split was used to set a specific data distribution, so this breaks after the parameter changes introduced in that PR.
I’ll adjust the tests to use the new parameters accordingly.

[sfluegel05]

Thanks, that is my bad. Somehow I missed that the tests were failing for this PR.