neural_sp icon indicating copy to clipboard operation
neural_sp copied to clipboard

Published 20 hours ago • verified publisher icon hirofumi0810

quetion about the loss and grad of "mbr"

Open Cescfangs opened this issue 3 years ago • 3 comments

https://github.com/hirofumi0810/neural_sp/blob/2b10b9cc4bdecb5180ecc45575c0ef410fb09aa3/neural_sp/models/seq2seq/decoders/las.py#L535-L548 I don't know much about mbr, according to these lines, it looks like a mWER loss and gradient to me

Cescfangs avatar Apr 25 '21 06:04 Cescfangs

@Cescfangs yes

hirofumi0810 avatar Apr 25 '21 10:04 hirofumi0810

@Cescfangs yes

Thanks for the reply, and I'm curious about the improvement of this mWER tuning, say 5% relative wer reduction?

Cescfangs avatar Apr 26 '21 01:04 Cescfangs

https://github.com/hirofumi0810/neural_sp/blob/2b10b9cc4bdecb5180ecc45575c0ef410fb09aa3/neural_sp/models/criterion.py#L12-L39

Also, I am a little confused about the “mbr” loss, the inputs are not used in backward function, how does the grad flow to model params?

Cescfangs avatar Apr 26 '21 12:04 Cescfangs