A few unoptimised piece of code (augmentation and masking)
TParcollet opened this issue · 2 comments
Describe the bug
While profiling a conformer transducer, I realised that a few functions were taking way more time than expected. I can't share the profiling trace here due to confidentiality, but I can share the code to create it. Basically, the resampling function in the augmentation pipeline of Librispeech is taking 10% of every step (including forward AND backward). This resampling function must be changed.
The make_transformer_src_mask function also is taking a significant amount of time - around 1/3 of a full conformer model inference.
@asumagic may want to look into this with me.
Expected behaviour
Faster, better.
To Reproduce
No response
Environment Details
No response
Relevant Log Output
No response
Additional Context
No response
#2410 solved half of this issue, the mask creation issue remains before this can be closed.