Daniel Galvez

WIP Fix for #4455

4

# What does this PR do ? Proposed fix for #4455 **Collection**: asr, common # Changelog - [Plumb is_tar_manifest throughout dataset code.](https://github.com/NVIDIA/NeMo/commit/b1bc3ab190cbb1e24898cd2af6ff84e2425f69d7) - [Fix for KeyError.](https://github.com/NVIDIA/NeMo/commit/653d4d755d0132d640305017336ddbf612705aeb) # Usage This allow...

Cudnn 3D Convolution Component

123

See #769 for some discussion points. I realize now there are some commits at the beginning which are not relevant to CUDNN. Sorry, but I'd rather not look at removing...

What is Lingvo's openness to outside contributions?

2

Hi there, I have a fork of lingvo that has been very useful for my team's work for a while now. I recently noticed that REGISTER_OP("AsciiToTokenId") in x_ops.cc is missing...

Can you upload the pylintrc file?

1

Modifying the code is a bit of a hassle when I have many pylint warnings from the existing code. Can you upload the pylintrc file that you already use for...

Why does lingvo use a binary mask instead of sequence lengths for representing invalid regions of sequences?

1

I've noticed that the rnn code in lingvo (in particular for ASR tasks) uses bitmasks for representing the validity of data in a sequence when you do batching. To be...

[egs] Use shellcheck to eliminate static analysis warnings in bash.

10

This commit is not yet complete. There is a balance between following the warnings provided by shellcheck and simply disabling some, since we don't want to wipe out too much...

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

7

Just a general comment. Keep in mind this may be a dead end. In particular, I am referring to the fact that snowfall has MmiTrainingGraphCompiler, while k2 does not: https://github.com/k2-fsa/snowfall/blob/c5ffa3fbaa1bee29b930cc290fcb6a35cb652397/snowfall/training/mmi_graph.py#L44...

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

5

I am referring to the following: https://github.com/k2-fsa/snowfall/blob/b7f76b663ba61c42c7b63aed4a101ed4e3f0fc16/snowfall/objectives/ctc.py#L38 https://github.com/k2-fsa/snowfall/blob/b7f76b663ba61c42c7b63aed4a101ed4e3f0fc16/snowfall/objectives/mmi.py#L85 This is a bit unusual to me. Was there a particular motivation? @csukuangfj it seems like you were the one who chose...

Dgalvez/cuda graphs greedy rnnt inference squash

9

# What does this PR do ? Speeds up RNN-T greedy decoding greatly by eliminating the 90% of the time that the GPU is idle, waiting on the CPU, via...

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

8

**Is your feature request related to a problem? Please describe.** @artbataev pointed out to me an issue with calling this when the model is using the cuda graph rnn-t decoder:...