Vladimir Bataev comments

Results 24 comments of


                                            Vladimir Bataev

[ASR] k2 rnnt

Regarding repeated snippet ```python # use k2 import guard # fmt: off from nemo.core.utils.k2_utils import k2_import_guard # isort:skip k2_import_guard() import k2 # isort:skip # fmt: on ``` Maybe it would...

Update black to the latest version incrementally

@titu1994, @ericharper To make this finally work we need a separate token with minimal permissions: - Read access to metadata - Read and Write access to actions and code The...

Update black to the latest version incrementally

Test changes from fork: #8996

Consider making the decoding classes subclasses of torch.nn.Module and registering them via add_module()

I frequently use in the notebooks `model.to("cuda:1")`, and expect this to work (I have multiple GPUs). It is inconvenient to use `CUDA_VISIBLE_DEVICES` with notebooks. I think that there are two...

Consider making the decoding classes subclasses of torch.nn.Module and registering them via add_module()

I agree that, for now, such breaking changes are undesirable. > On top of this, yes the fact that parameter count would be double counted for decoder and joint with...

RNN-T and TDT Greedy Decoding with TorchScript

jenkins

RNN-T and TDT Greedy Decoding with TorchScript

jenkins

Dgalvez/cuda graphs greedy rnnt inference squash

Cool, I will review the PR today or tomorrow. @galv Please fix DCO (anyway, you will need to fix it for merging).

Dgalvez/cuda graphs greedy rnnt inference squash

jenkins

Dgalvez/cuda graphs greedy rnnt inference squash

jenkins