Support Use of Joint and GDumb with Pre-Trained Models

[wistuba]

Joint and Gdumb reset the models on_model_update_start. This results in a non-desired behavior when working with pre-trained models.
There is also no option to not reset the model at all.

The change introduces a new flag reset which allows to control whether the model will be reset. Furthermore, in case of a reset, the pre-trained model will be reloaded instead of using an untrained model.

The integration tests are expected to break for Joint and GDumb due to a different initialization of the model. Before, the workflow was creating the model and resetting it. Now, it only creates the model without reset.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[prabhuteja12]

Documenting an alternative to this approach:

The resetting is method + network specific. Thus an alternative is to define a reset_parameters for RenateBenchmarkingModule and overload them only for specific modules (ViT, Text transformers etc). The current on_model_training_start would invoke the this reset_parameters to reset it.

[github-actions]

Coverage report

The coverage rate went from 85.68% to 84.95% ⬇️

0% of new lines are covered.

Diff Coverage details (click to unfold)

src/renate/cli/parsing_functions.py

0% of new lines are covered (79.2% of the complete file).
Missing lines: 457, 462