Using TorchMetrics for code simplification

[stancld]

🚀 Feature Request

Hi all,

We have been developing Torchmetrics to be general-purpose metrics with a unified API for various domains. All our metrics are tested against reference metrics (e.g. scikit-learn, sacrebleu,…) in multiple OS environments and all PyTorch versions above v1.4 for their correctness. The TM includes both functional as well as module (based on torch.nn.Module) versions. It is based predominantly on torch and other extra dependencies are usually required to be installed for specific domains.

Motivation

Torchmetrics currently support more than 10 text/NLG metrics, such as ROUGE, BLEU, chrF/chrF++, BERTScore, Translation edit rate, Extended edit distance, etc. We've been also aiming to include some more metrics, namely, InfoLM, DepthScore, BaryScore or Meteor in release v0.9. Therefore, Torchmetrics would provide a great opportunity you can use a variety of well-tested NLGmetrics to evaluate models trained using fairseq framework, and also simplify the codebase by dropping the necessity .

Pitch

If you'd be interested in this option, I'm happy to conduct all the necessary steps and then open a draft PR.

Alternatives

Leave it as it is.

Additional context

Should you have any questions or would like to know more information about torchmetrics, please contact directly me, @Borda, @tchaton

Also, torchmetrics has been recently succesfully ported to pytorch_geometric.