I am trying to finetune chcaa/xls-r-300m-danish-nst-cv9 with fairseq. However, as the framework does not take pytorch models as input, I must transform the model back into the original wav2vec2 pytorch checkpoint. However, I have not found any code to achieve this. Any suggestions?
I am aware that there exists a script to convert from checkpoint to model: transformers/convert_wav2vec2_original_pytorch_checkpoint_to_pytorch.py at 0e83c9664b2a440ade59066a77fb01d0143e4d18 · huggingface/transformers · GitHub
However, I haven’t been able to reverse the code.