Configs for wav2vec experiments

[AndreasPlt]

Previously, all wav2vec experiments were located in a single file. After adding more and more experiments, this got more chaotic and less modular, hindering flexible development and experimentation. I therefore split the previous config_phoneme_pretrain_finetune.py into different files for each wav2vec modification. Furthermore, I added configs for the experiments with hard negatives and positive samples.

Note: I need to test-run the new configs to see if I transferred everything correctly. Until then, this PR remains as a draft, but please feel welcome to give feedback on the current config structure :)

[vieting]

Looks mostly good, some comments.

[vieting]

Isn't this the same as above?

[AndreasPlt]

Yes, this is just for adding all relevant jobs under the new alias 1_phoneme_spec. Since the amount of jobs that is saved under the train_name_suffix, I thought this approach is easier than calling job.add_alias(...) on all relevant jobs. Feel free to suggest any other approach to accomplish this :)

[vieting]

But do we need the old alias?

[AndreasPlt]

The base model represents four different models in the ablations:

1. Testing the w2v modification itself (i.e. just the normal "w2v_phoneme_boundary_masking" alias prefix)
2. Testing masking strategies in finetuning (there, for negatives_other, the base model is the same for the ablation with random masking strategy; for negatives_other_boundary_masking and boundary_masking, the baseline model is the same as the ablation with phoneme masking strategy)
3. Testing phoneme mask length in finetuning (only for negatives_other_boundary_masking and boundary_masking, where it is the same as the ablation with mask_len = 1)
4. Testing masking probability in finetuning (base model is the same as ablation with mask probability 0.65)

I wanted to make this structure clear in the alias folder as well - otherwise, it is not clear from the outside that the base model also represents the "masking probability = 0.65" model. So in that sense, yes, we need the old alias and all other ones as well

[vieting]

Okay, then maybe put a comment above the block like "repeat same experiment to obtain additional alias in line with following experiments" and it should be fine.