"there is a cycle" during Determinizing L*G

[qssheep]

I was trying the Librispeech example and got error when running decode.py

Loading L_disambig.fst.txt
Loading G.fsa.txt
DEBUG:root:Intersecting L and G
DEBUG:root:LG shape = (21246039, None)
DEBUG:root:Connecting LG
DEBUG:root:LG shape = (21246039, None)
DEBUG:root:Determinizing LG
[F] /home/yckj2113/github/k2/k2/csrc/host/fsa_util.cc:void k2host::CreateFsa(const std::vectork2host::Arc&, k2host::Fsa*, std::vector*):470 there is a cycle: 15976413 -> 256

My lm was downloaded using run.sh.

[qindazhu]

I guess you are running with G.fst.txt you generated with old scripts in snowfall before, the cycle is from </s> (EOS) in LM. We have fixed that issue with #33, so just pull the latest code and re-generate G.fst.txt by running run.sh would be fine.

BTW, there's another issue in decode.py now about aux_labels on L*G, see here #31 (comment) for details, we need to write some C++ code to fix that issue, so the decode.py now is actually not working (the script is disabled in my last PR, but because of some reason, the change is not merged)

[qssheep]

Thank you, I'll try it as soon as decode.py is available.

[qssheep]

I tried the latest version today and the error still appears.
Then I realized it is my k2 version instead of snowfal, after re-install my k2, this error disappears.

[qindazhu]

Thanks, bear in mind we are still tuning the model, so please ignore the bad WER for now (but definitely you can see the whole train/decoding process with k2 and lhotse)