Daniel Galvez

icon company @NVIDIA icon location Bay Area, CA icon location I do speech recognition, machine learning, and high performance computing.

Results 85 comments of Daniel Galvez

Some methods don't work with unconventional lattices

> I don't know whether you might have any constraints on these kinds of contributions. Hi Dan, dropping in here to say I got legal approval for us at NVIDIA...

What is the motivation behind keeping graph construction code in snowfall instead of k2 repo?

Hmmm, yes, making a custom class probably makes the most sense for me. Ad-hoc polymorphism can be done if we want to use the same operator names. In general, the...

What is the motivation behind keeping graph construction code in snowfall instead of k2 repo?

> Actually there isn't really a relationship between SMs and training samples. Okay, this is my fault. I need to study the code at a deeper level. Sample here means...

What is the motivation behind keeping graph construction code in snowfall instead of k2 repo?

I haven't forgotten about this. I'm still getting my own setup fully up, so I haven't finished running mmi_att_transformer_train.py yet.

Why do ctc.py and mmi.py use use_double_scores=True?

Okay. I definitely think it makes sense for long utterances to use double if we are in probability space. I will keep this in my mind as a potential knob...

Why do ctc.py and mmi.py use use_double_scores=True?

Don't worry, I know that part (it would be concerning if I didn't!)

How to share data without requiring consumer to "own" the input tensor?

Mildly curious if you encountered this in your nanobind project's ndarray implementation and have any thoughts you'd be willing to share @wjakob.

argmax_internal::PairInputIterator and argmax_internal::PairOutputIterator don't implement the random access iterator concept

good to know that you were able to fix the issue. FYI @gnroy, who originally told me about the compilation issue.

Regarding composing

I am not using K2 actively right now, but what you are proposing should be possible from the fundamentals, since you are basically just adding the "boost" weight to each...

Tensor.nonzero fails on GPU for tensors containing more than INT_MAX elements

@bhack you can depend upon a version of cub orthogonal to the cub version released with a particular version of the cuda toolkit by adding a dependency to cccl in...