Wav2Vec2 Pretraining

[rajeevbaalwan]

❓ Questions and Help

I want to perform wav2vec2 Pretraining from scratch and while following the documentation for same on https://github.com/facebookresearch/fairseq/tree/main/examples/wav2vec it is mentioned that all audio clips should be in single directory. The issue is i have too much data to keep in a single directory.

I have data in multiple directories on different disks and can't move complete data in single directory due to storage issue. Is it possible to pretrain the model in this scenario?

[gau-nernst]

You can try using symlinks

[rajeevbaalwan]

@gau-nernst thanks for your response.
Is there any other way for this like modifying the code so it can handle paths from multiple directories? At the end a path is to be loaded by the data loader so why this constraint?

[gau-nernst]

Of course it is possible to modify the code, but then you have to do it yourself.