Vedant Roy

> Facing the same issue here. I'm building a chrome extension that renders on web pages via the shadow dom. I explicitly import the stylesheets to make things work. Curious...

> I finally managed to make myself a custom aggregated metrics data sorted by steps and log to the wandb server at the end of the running script. I wish...

Ahh, that's what's going on here. Just ran into this issue! Maybe stick a warning in the README? Anyways; this is a super nice library, thanks for making it!

I created something like this: https://github.com/vedantroy/typecheck.macro, it just generates validation code directly, but I bet the codebase could be split into a few libraries & a new "runtypes" could be...

> The encoder is frozen, not trained at all, and the text embeddings can even be pre-computed. So the size of text encoder should have negligible impact in the training...

I have some general questions that I'd be curious to discuss with people here. Namely, - How much does shrinking the output space affect training efficiency? I've been trying to...

> > * Can we modify the space of captions and still get good results? It's hard to generate captions for images, but..., it's easy to (for example) generate tags...

> I understand this concern. The shared structures key needs to be stored _somewhere_ in the database. And I think the implicitly suggested alternative is to automatically filter it out....

Another question: Does this support identifying multiple objects in an image? I'm using an unofficial demo: https://huggingface.co/spaces/hysts/Yet-Another-Anime-Segmenter and it seems like all the results are always a singleoutline.

@hanlint How does this work when I setup torchmetrics to use a dictionary. See: https://github.com/Lightning-AI/metrics/issues/682 Should I return a dictionary from my validate method?