Issue 36 fix

[JP-SystemsX]

Addressing Issue #36

The _get_param function of the finetuning arguments should now:

• Retrieve All trainable Parameters as long as naming convention is kept
• Avoid Exploding learning rate due to bug in calculation of learning rate during Layer wise decreasing Learning rate
• Raise exception if a model doesn't keep the naming convention

Still has problems with:

• Models that don't keep naming conventions like ALBERT

[chschroeder]

After the discussion, I think this solution is the right way. Thanks!

Do you have any pending tasks here? Otherwise I would just merge it.

After that I would like to add 1-2 unit tests (which were apparently lacking in this area of the code), and afterwards it should be done.

[chschroeder]

@JP-SystemsX See the comment above regarding your question in #36 on how to proceed.