TorchMetrics for higher reproducibility !

[tchaton]

Dear @RJT1990,

Awesome project there !!!

I looked internally and I have seem metrics being manually implemented without any testing.

This makes me pretty scary in term of reproducibility and accurate reporting.

I think you should consider using https://github.com/PytorchLightning/metrics as the tool for benchmarking the runs.

There are extremely well tested metrics which works automatically in distributed settings and plain PyTorch.

Best,
T.C

[Borda]

I smell @tchaton is volunteering to make it for you guys 🐰

[RJT1990]

Heya,

As discussed yesterday, we are not maintaining sotabench (and associated tools) at this stage, and our focus is elsewhere - particularly on lighter forms of capturing results for the main Papers with Code website.

On testing: this was an experimental product. As such, the emphasis was on extracting user signal rather than committing wholly to a particular implementation. I.e. "manual implementation" was sufficient for our objectives at the time :).

Thanks!

Ross