Train models with Metadata

[niekdejonge]

@djoas and @florian-huber

I had a look at making the metadata annotation a bit more robust. What I added was a MetadataFeatureGenerator class. This contains simple conversions of metadata to input feature, including normalization. This has as benefit that the conversion on the metadata will also be stored in the spectrum binner, so that when the user runs the model later there is no risk of normalizing in a different way for instance.

@djoas if you are busy with finishing up your thesis and don't have the time. No need to have a look, but since you worked on this it would be great if you notice anything that might give unexpected behavior.

[niekdejonge]

@florian-huber Could you have a look at the MetadataFeatureGenerator class I generated.

Do you think this is a good approach? I never used parent classes before, so wanted to make sure that this would make sense.

In this way we pass a list with Classes of type MetadataFeatureGenerator. In spectrum binning these classes can be used to generate the appropriate metadata based feature. Using the parent class has as benefit that the input stays the same and that they will all contain generate_feature method.

[florian-huber]

Nice! I will do a full code review tomorrow.

[niekdejonge]

Nice! I will do a full code review tomorrow.

@florian-huber Great! I noticed that the tensorflow version check was currently failing. Do you know what is causing that? It is unrelated to this pull request right?

[florian-huber]

@niekdejonge I made some suggestions for maybe changing the design, let me know if we should have a brief meeting to discuss.

The tensorflow thing should indeed be unrelated. Tensorflow is a constantly moving target.... better to look at that separately.

[niekdejonge]

@florian-huber I went through the pull request and removed all changes that were still in a draft or should only be stored locally. The changes remaining in this PR are:

• In SiameseModel, the attributes are (e.g. self.base) not passed as an argument but instead used by the class.
• Improved plotting functionality. Some additional plotting options were added and plots previously created in notebooks are now functions.
• The ease of training MS2Deepscore models is improved, including standard settings and splitting validation and training data.

[florian-huber]

Thanks @niekdejonge for the update. I went through the code as well and will add a few minor things with #141

[florian-huber]

I will merge this since tests are now passing (except for SonarCloud due to missing code coverage of the new workflows, this should be handled elsewhere --> #143 )

[niekdejonge]

Great! Thanks for fixing the failing tests!