PyTorch 2.0 Q&A: Rethinking Data Loading with TorchData

The TorchData library provides new composable building blocks called DataPipes and a new DataLoader2. Together, they allow easy construction of flexible, reusable data pipeline and execution in various settings/backends. We will examine these new features, showcase examples, and discuss how they improve upon the data loading functionalities provided by PyTorch.
Presenters: Kevin Tse and Erjia Guan. Hosted by DA Justin Jeffress.

[@gerardsimons3757]

The focus on iterable / streaming datasets is very exciting! Thanks for this

[@zhitaoli4702]

Any chance you can publish the slides accompanying the talk? Thanks!

[@maryamshangaray3780]

Hello! I have a lot of numpy ".npz" files with 10 of 1d arrays . Can you tell me what format I should choose to convert my data to read they fast with torchdata.datapipes ?

[@MiladMohammadi-xu2bd]

Looks like this post has no audio. Can you please fix?

[@BlackHermit]

I can hear it well, Milad. Can you hear it now?