Problem reading FAR file from farcompilestrings

[mmcauliffe]

Hello,

I'm using the latest commit generated from CMake (version 3.12.2) along with OpenGRM-ngram 1.3.4 (https://github.com/MontrealCorpusTools/opengrm-ngram), both compiled using VS 2017. When I generate a FAR file using farcompilestrings, it completes successfully, but using it as input to any of the ngram binaries such as ngramcount ( etc gives the following error:

ERROR: Fst::Read: Unknown FST type vector (arc type = standard): <unspecified>

From the OpenFST changelog (http://www.openfst.org/twiki/pub/FST/FstDownload/NEWS) it looks like this was addressed somewhat in 1.6.8 (or at least the reporting), and the same versions of everything work on Linux, and the OpenGRM-ngram binaries compiled on Linux can read the generated FAR just fine, so I'm a bit stumped.

[kkm000]

The FAR extension is not compiled into the current build. Let me see what I can do over the weekend.

[kkm000]

I'm using the latest commit generated from CMake

@mmcauliffe, I am a little bit unclear: are you using CMake to build OpenFST from this repo? Or VS projects/solutions?

[kkm000]

/cc @jtrmal Yenda, if you have some spare cycles, could you please chime in? I am not building the far extensions binaries at all (yet? I probably should as it appears there is a demand out there). So I am almost sure this is CMake-related. Do you understand what is going on? I was getting almost exactly the same error before we added the /WHOLEARCHIVE trick last time. But that was different: the mainstream library tore off a file with static initialization only, so the library member was never pulled into executables on link. With FAR, it's different, the declarations are in a file that also has some code. I do not understand though if these functions, FarReaderClass::Open and others in the same source file, are in fact dependency of something that is actually compiled in. If not, we have the same problem with the FAR executables.

[jtrmal]

I'm gonna try to find a windows box and try. Yes, it seems to me as a registration issue (/wholearchive) as well. Y.

On Sun, Sep 23, 2018, 12:55 Kirill Katsnelson [email protected] wrote:

/cc @jtrmal https://github.com/jtrmal Yenda, if you have some spare cycles, could you please chime in? I am not building the far extensions binaries at all (yet? I probably should as it appears there is a demand out there). So I am almost sure this is CMake-related. Do you understand what is going on? I was getting almost exactly the same error before we added the /WHOLEARCHIVE trick last time. But that was different: the mainstream library tore off a file with static initialization only, so the library member was never pulled into executables on link. With FAR, it's different, the declarations are in a file that also has some code https://github.com/kkm000/openfst/blob/d4dd88e17/src/extensions/far/far-class.cc#L38. I do not understand though if these functions, FarReaderClass::Open and others in the same source file, are in fact dependency of something that is actually compiled in. If not, we have the same problem with the FAR executables.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/kkm000/openfst/issues/16#issuecomment-423807993, or mute the thread https://github.com/notifications/unsubscribe-auth/AKisX6R2AqplaM9JwZfnsgHsA9eyv9yFks5ud2i-gaJpZM4WzKXE .

[mmcauliffe]

@kkm000 Right, so the VS projects/solutions checked into the repo don't include the FAR extensions, so I was following the CMake instructions from the kaldi/windows (https://github.com/kaldi-asr/kaldi/blob/master/windows/INSTALL.md#compiling-openfst) as those do generate the FAR extensions (and they seem to work alright with each other).

[kkm000]

I'll be working on adding them to the VS build then. No ETA at this moment, I'll try my best to do it this week.

[kkm000]

I finally got come breathing space to work on this. @mmcauliffe, I'd really appreciate if you could check if the version I have so far works for you. It builds far extensions with Visual Studio.

Extensions are compiled into the same libfst and libfstscript libraries, and binaries go into the same build_output/ directory as the fst*.exe executables.

The branch name is wip-build-optional-features. far is an optional feature that is currently enabled in this branch, check openfst.user.props. You can just build and see if it does the trick for you. My quick command-line checks apparently worked:

c:\projects\openfst\src>..\build_output\x64\Debug\bin\farcreate mdy_nosym.fst number_nosym.fst test.far

c:\projects\openfst\src>..\build_output\x64\Debug\bin\farinfo test.far
far type                                          sttable
arc type                                          standard
fst type                                          vector
# of FSTs                                         2
total # of states                                 24805
total # of arcs                                   54994
total # of final states                           2

I changed the MSBuild build process quite significantly. The project files are moved to src/ from its subdirectories, and all objects are placed into a tree under an obj/ directory under project root. If you are reusing the working directory, it makes sense to delete the whole src/ directory then do a git reset --hard to resurrect it in clean state--so you won't end up with 2G of orphaned object files. Then do git fetch and git checkout wip-build-optional-features.