Projects that are related to K2

[csukuangfj]

I find that there are two projects, both from facebook, that have some overlaps with K2:

• gtn, Dan mentioned this a few days ago
• nestedtensor, for tensors with irregular shapes, like Ragged<T>

Perhaps we can spend some time to find whether we can learn something from them.

[danpovey]

Yeah, cool. I heard of gtn a couple days ago from several people. Looks interesting but seems to aim for CPU. RE nestedtensor: the design seems quite different from TensorFlow's RaggedTensor. Much more general, but also much less similar to what we are doing. (TF has our row_ids and row_splits, which is a design I came up with independently).

…

On Sat, Oct 3, 2020 at 6:46 PM Fangjun Kuang ***@***.***> wrote: I find that there are two projects, both from facebook, that have some overlaps with K2: - gtn <https://github.com/facebookresearch/gtn>, Dan mentioned this a few days ago - nestedtensor <https://github.com/pytorch/nestedtensor/>, for tensors with irregular shapes, like Ragged<T> Perhaps we can spend some time to find whether we can learn something from them. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#212>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZFLO7KMVIOBLSK67ECH5TSI36GRANCNFSM4SCSQYYA> .

[awni]

Hi, (as Dan knows) I am one of the authors of GTN.

I also started hearing about K2 a week or two ago from a few people :).

GTN is indeed currently CPU only with some vague plans at the moment to GPU accelerate some of the bottlenecks (namely compose). I find the GPU approach with K2 interesting. We are hoping to learn a bit more about your approach and how we can leverage K2 to do better work with differentiable WFSTs.

For reference here is a preprint that came out on arXiv today which describes the kinds of things we have started exploring with GTN and in general are interested in doing with a fast but easy to use differentiable WFST framework. Maybe you will find it useful when designing your API.

[danpovey]

Great, thanks! Will have a look.

…

On Tue, Oct 6, 2020 at 5:09 AM Awni Hannun ***@***.***> wrote: Hi, (as Dan knows) I am one of the authors of GTN. I also started hearing about K2 a week or two ago from a few people :). GTN is indeed currently CPU only with some vague plans at the moment to GPU accelerate some of the bottlenecks (namely compose). I find the GPU approach with K2 interesting. We are hoping to learn a bit more about your approach and how we can leverage K2 to do better work with differentiable WFSTs. For reference here is a preprint <https://arxiv.org/abs/2010.01003> that came out on arXiv today which describes the kinds of things we have started exploring with GTN and in general are interested in doing with a fast but easy to use differentiable WFST framework. Maybe you will find it useful when designing your API. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#212 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZFLO47UOFY46XR5QIT27DSJIYYLANCNFSM4SCSQYYA> .