Closed Bang-sheng-Zhuo closed 5 years ago
Seems like you forget to provide a lexicon. Please refer to https://github.com/facebookresearch/wav2letter/blob/master/docs/decoder.md for detail.
Seems like you forget to provide a lexicon. Please refer to https://github.com/facebookresearch/wav2letter/blob/master/docs/decoder.md for detail.
Thanks. - -!
The latest docker images of w2l provides flag with --words
, not --lexicon
, in recipes/wsj/configs/conv_glu/decode.cfg
. I thought it would work if I follow the configs.
Hi @k1kyo,
the docker image was updated two weeks ago where flag --lexicon
is already provided. Please pull the image on your local machine.
Shortly will update the images with the current state if you need them too.
Hi @k1kyo,
the docker image was updated two weeks ago where flag
--lexicon
is already provided. Please pull the image on your local machine.Shortly will update the images with the current state if you need them too.
I just pull the newest image of cuda-latest which was created two weeks ago on my local machine, and I found the decode.cfg of WSJ recipes still privides --words
flag, not --lexicon
flag.
Everything should be fine on master now. I think we were at transition two weeks ago. Sorry for the confusion.
Hi @k1kyo,
I have updated images with the current master, you can pull again and everything should work now. Please ping if any error still remains.
@tlikhomanenko thanks
run command
Gflags after parsing
stack trace
By the way, I trained with GLU and WSJ data on another machine, decoding runs fine but I got --
WER: nan, LER: 100
. The TER of training set istrain-TER: 9.87 | data/nov93dev-TER: 11.95
. I build the LM binary with kenlm.