lingochamp / kaldi-ctc Goto Github PK

When I do "make -j 4" in kaldi-ctc/src. It occurs some errors.
The error is:
nnet-nnet.cc: In member function ‘void kaldi::nnet2::Nnet::SetMiniBatch(int32) const’:
nnet-nnet.cc:47:20: error: ‘CuDNNRecurrentComponent’ does not name a type
dynamic_cast<CuDNNRecurrentComponent >(components_[i])->InitMiniBatch(m
^
nnet-nnet.cc:47:44: error: expected ‘>’ before ‘’ token
dynamic_cast<CuDNNRecurrentComponent >(components_[i])->InitMiniBatch(m
^
nnet-nnet.cc:47:44: error: expected ‘(’ before ‘’ token
nnet-nnet.cc:47:45: error: expected primary-expression before ‘>’ token
dynamic_cast<CuDNNRecurrentComponent *>(components_[i])->InitMiniBatch(m
^
nnet-nnet.cc:47:64: error: ‘class kaldi::nnet2::Component’ has no member named ‘InitMiniBatch’
dynamic_cast<CuDNNRecurrentComponent *>(components_[i])->InitMiniBatch(m

The code in nnet-nnet.cc:47 is:
void Nnet::SetMiniBatch(int32 mini_batch) const {
KALDI_ASSERT(!components_.empty());
KALDI_ASSERT(mini_batch > 0);
for (size_t i = 0; i < components_.size(); i++) {
if (components_[i]->Type() == "CuDNNRecurrentComponent") {
dynamic_cast<CuDNNRecurrentComponent *>(components_[i])->InitMiniBatch(mini_batch, 0);
}
}
}
The reason is that CudnnRecurrentcomponent is not a type.But it includes the nnet-cudnn-componnet.h which declar this type. I don't know how to solve it. Is it a problem of my computer?

ctc-nnet-update.cc: In member function ‘bool kaldi::ctc::NnetCtcUpdater::ComputeObjfAndDeriv(const std::vectorkaldi::ctc::NnetCtcExample&, kaldi::CuMatrix, double, double*) const’:
ctc-nnet-update.cc:200:3: error: ‘ctcComputeInfo’ was not declared in this scope
ctcComputeInfo info;
^
ctc-nnet-update.cc:201:3: error: ‘info’ was not declared in this scope
info.loc = CTC_GPU;
^
In file included from ../base/kaldi-common.h:38:0,
from ../util/stl-utils.h:45,
from ../util/edit-distance-inl.h:25,
from ../util/edit-distance.h:29,
from ctc-nnet-update.cc:24:

any one got the same erorr? i grep the whole kaldi-ctc project including tools, nothing was found about ctcComputeInfo.

am i missing any dependant library? or i complied in a wrong way?

hi, Feiteng, does your asr work support Asynchronous Stochastic Gradient Descent? Thanks!

sorry for it.

I use
../latbin/lattice-1best --lm-scale=5 "ark:gunzip -c $indir/lat.JOB.gz|" ark:tmp
../latbin/lattice-align-words-lexicon $wbfile $mdl ark:tmp ark:tmp1
../latbin/nbest-to-ctm --frame-shift=0.01 --print-silence=true ark:tmp1 $out_dir/$out_name.txt
I want to get every words time information, but I got some problem.

Is there any online decoding examples like kaldi online2bin/online2-wav-nnet2-latgen-faster ?
Which script can I learn from ?

thank you

运行run.sh时, 在stage=-2时卡住:
local/data_prep.sh $data/LibriSpeech/$part data/$(echo $part | sed s/-/_/g)
追进去, 是在这一步卡住:
utils/data/get_utt2dur.sh $dst 1>&2 || exit 1
在eeesen的data_prep.sh里面, 好像这一步不存在
系统提示:
utils/data/get_utt2dur.sh: segments file does not exist so getting durations from wave files
utils/data/get_utt2dur.sh: could not get utterance lengths from sphere-file headers, using wav-to-duration
utils/data/get_utt2dur.sh: wav-to-duration is not on your path

请问如何解决? 谢谢!

I modified the run script based on librispeech and all is ok in 30 hours corpus training.

But in about 200 hours corpus training, I encountered an error which called "sum == sum && "Error in this batch, deriv sum is inf/nan" in the source file named ctc-nnet-update source, which called warp-ctc function. I can't figure it out.

Can anyone offer me some advices?

If this is CTC, why do you need a dictionary?

Hello I have installed kaldi-ctc according to instruction given.
while testing the system with librispeech using "bash run.sh --stage -2 --num-gpus 4"
it is showing some error

utils/data/get_utt2dur.sh: segments file does not exist so getting durations from wave files
utils/data/get_utt2dur.sh: could not get utterance lengths from sphere-file headers, using wav-to-duration
utils/data/get_utt2dur.sh: wav-to-duration is not on your path

How to solve the problem?

I git clone the master branch and try to install it.
When I get in src dir and make depend is shows "ctc-nnet-update.cc:28:17: fatal error: ctc.h: No such file or directory".
Then I use find command to find 'ctc.h' but there is no header file named 'ctc.h'.
How can I fix this problem?
ps: Will this project support cuda 9 and cudnn 7 ?
Thank you !

Hi,
I got an error "ctc-nnet-update.cc:204:3: error: ‘ctcOptions’ was not declared in this scope" when compiling kaldi-ctc source code. How can I fix it?

你好，

   您的工程已经下载，并编译通过，但执行 ./run.sh --stage -2 --num-gpus 4 时，会有如下提示后退出

请问这种情况是什么问题？

[lijin@webserver-b456dc ctc]$ ./run.sh --stage -2 --num-gpus 4
local/download_and_untar.sh: data part dev-clean was already successfully extracted, nothing to do.
local/download_and_untar.sh: data part test-clean was already successfully extracted, nothing to do.
local/download_and_untar.sh: data part dev-other was already successfully extracted, nothing to do.
local/download_and_untar.sh: data part test-other was already successfully extracted, nothing to do.
local/download_and_untar.sh: data part train-clean-100 was already successfully extracted, nothing to do.
local/download_and_untar.sh: data part train-clean-360 was already successfully extracted, nothing to do.
local/download_and_untar.sh: data part train-other-500 was already successfully extracted, nothing to do.
Downloading file '3-gram.arpa.gz' into 'data/local/lm'...
'3-gram.arpa.gz' already exists and appears to be complete
Downloading file '3-gram.pruned.1e-7.arpa.gz' into 'data/local/lm'...
'3-gram.pruned.1e-7.arpa.gz' already exists and appears to be complete
Downloading file '3-gram.pruned.3e-7.arpa.gz' into 'data/local/lm'...
'3-gram.pruned.3e-7.arpa.gz' already exists and appears to be complete
Downloading file '4-gram.arpa.gz' into 'data/local/lm'...
'4-gram.arpa.gz' already exists and appears to be complete
Downloading file 'g2p-model-5' into 'data/local/lm'...
'g2p-model-5' already exists and appears to be complete
Downloading file 'librispeech-lm-corpus.tgz' into 'data/local/lm'...
'librispeech-lm-corpus.tgz' already exists and appears to be complete
Downloading file 'librispeech-vocab.txt' into 'data/local/lm'...
'librispeech-vocab.txt' already exists and appears to be complete
Downloading file 'librispeech-lexicon.txt' into 'data/local/lm'...
'librispeech-lexicon.txt' already exists and appears to be complete
Please install 'flac' on ALL worker nodes!

Because my system is CentOS7, so cuda version is 9.0, how can I build tools and src with cuda 9.0?

the code is too old, i can not compile the code in ubuntu 18.10.

it is very strange that when I "cd tools, make -j", it echos some error:

--2018-04-17 16:57:27-- (try: 2) http://openfst.cs.nyu.edu/twiki/pub/FST/FstDownload/ope
nfst-1.3.4.tar.gzConnecting to openfst.cs.nyu.edu (openfst.cs.nyu.edu)|128.122.80.127|:80... failed: Connec
tion timed out.Retrying.

--2018-04-17 16:59:36-- (try: 3) http://openfst.cs.nyu.edu/twiki/pub/FST/FstDownload/ope
nfst-1.3.4.tar.gzConnecting to openfst.cs.nyu.edu (openfst.cs.nyu.edu)|128.122.80.127|:80...

Hi,
I think this issue is not resolved. I use the following config:
minibatch_size=32
max-allow-frames=350
subsampling-factor=4

It will report this nan issue in about 51th iter. Ony when I decrease the max-allow-frames to ve 300, the training will pass. But during training there is so many warning "Too little feature frames.". And the increasing of accuracy is slow.
Is there any solution for this nan issue? I mean from the code. Thanks!

As the title question, why set google_prior_const=9?