Elijah Rippeth

> I'm not sure shuffle the samples in the batch will improve the training I'm also not sure of this, but it seems to be happening in [the torchaudio code](https://github.com/pytorch/audio/blob/main/examples/hubert/dataset/hubert_dataset.py#L76)....

> Yeah this shuffling will generate different batches Ah, I misread the code -- this makes total sense. I agree that this will have more impact than shuffling within a...

It seems like once torchdata adoption lands we could just apply a map datapipe to transform the data. I'll keep this open until we've cut over to the datapipes API...

At the risk of sounding a bit pointed: is there a reason the development is happening like a closed-source project instead of an open-source one? Many of us would love...

I was curious about this, too. I might try to cobble together a PR to generate JNI bindings.

Adding a checklist here for my own bookkeeping: - [x] ACL event https://www.aclweb.org/portal/node/add/event - [x] Corpora list [[email protected]](mailto:[email protected]) - [ ] MT List [[email protected]](mailto:[email protected]) - [ ] Moses mailing list...

A few notes: - I tried emailing MT list but the email bounced back with "mailbox unavailable" 🤔 - I can't post to SIGLEX because I don't have an ACM...

Looking at a flamegraph of inference [^1], it seems like a large portion of the inference time is spent [here](https://github.com/bab2min/tomotopy/blob/6ef712c63bad504801a981a4a52aa938c80f0bda/src/Utils/ThreadPool.hpp#L106). I'm trying to track this down now, but it seems...

I'm happy to take this up. It seems like if you return the lazily loaded module from some function, you don't actually need to support `as ...`. For example, you...

I see. The failure to import would result in an immediate error, but the actual import is the lazy part. Is that right? In other words, attribute access on not-installed...