Re: [Kaldi-developers] python wrappers for Kaldi

[Jan C. <jan...@gm...>]

Hi All,

First I'd like to clarify that the wrappers are only meant to read and 
write a limited set of objects (matrices, vectors, scalars, pairs) from 
.scp and .ark into Python with a Pythonic feel (dict-like interface for 
radom access iterators, iterator interface for sequential ones, context 
management). I didn't consider wrapping fst classes nor anything more 
specific to Kaldi.

In their limited form the wrappers are useful to e.g. use some python 
neural network library to train nets on Kaldi's features and then use 
the Kaldi decoders, in a manner similar to Kaldi DNN recipes.

I decided not to write the wrappers as generic SWIG wrappers because I 
target conversion from Kaldi data files to Python specific classes, such 
as ndarray. Wrappers for other languages would need to address other 
language-specific matrix libraries etc. I don't know SWIG too much and 
maybe it would be useful to make wrappers for several languages at once, 
though.

Since the wrappers are optional, maybe the best way to proceed is to 
extract them into a separate repository, that can be built against 
against an installation of Kaldi and see whether it is used by anyone? I 
can work on it sometimes during the holiday break and let you know once 
it is published.

@Yajie
splicing a single utterance is easy. You pad it with zeros, then set the 
stride to be one frame and shape to be several ones. Thus it looks like 
one large spliced array. However, to train a network you want to shuffle 
examples (i.e. spliced frames) from multiple utterances - I resorted to 
reading a few hundred megs of data in to memory, then shuffling it, then 
reading another batch form the disk. I can send you the code if you 
would be interested.

Jan

On 12/21/2014 9:36 PM, Daniel Povey wrote:
> Jan, perhaps when you have time you could respond to the above
> comments on the list?
> I'm not 100% sure what to do about this.
> BTW, if we do include this, it will likely be optionally compiled,
> because I don't want the generic Kaldi compilation to be dependent on
> boost.
>
> Dan
>
>
> On Fri, Dec 19, 2014 at 3:48 AM, Ondrej Platek <ond...@gm...> wrote:
>> Hi Matthew,
>>
>> I made some subjective comments below.
>>
>> PS: Note that I like the proposed wrappers, but I am not sure how
>> boost::python is easy to install on all supported platforms.
>>
>> On Fri, Dec 19, 2014 at 9:30 AM, Matthew Aylett <mat...@gm...>
>> wrote:
>>> Hi
>>>
>>> Apologies, I've been snowed under here.
>>>
>>> I haven' had a chance to look over your work. I also don't have any views
>>> on the 'right' way to do it. My thoughts on this are in a previous thread.
>>> See subject "Using SWIG to wrap kaldi for python" where I discussed this
>>> with ondrej platek and
>>> Vassil Panayotov.
>>>
>>> In the idlak branch there is an example of python wrappers that I put
>>> together some time ago. These are based on SWIG. In the end I didn't need
>>> this at this stage because in the build system command line executables work
>>> very well. Its in run time wrappers are very useful. The advantage with SWIG
>>> is that the much of the same work will also contribute to C#, Java, Perl
>>> wrappers as well. In my experience the most important were Java wrappers to
>>> help produce a library for Android. I have no experience with C# and moved
>>> to Python from Perl so only use Perl in legacy code ;-).
>>>
>>> So some questions to consider:
>>>
>>> 1. Why is python wrapping required for training. using sys.Process to run
>>> command lines, structured output directories etc mirrors the current Perl
>>> recipes, what is the added benefit in this case?
>> Well bash and Perl is the current scripting language for Kaldi. For example
>> I prefer to use Python instead of both of them.
>>
>>> 2. If its for run time decoding shouldn't we create a cross platfom  C
>>> API? Perhaps things have changed but C++ APIs were never cross compiler
>>> compatible in the past so you couldn't do stuff like compile using gnu and
>>> link in MSN. With a C interface you can distribute libraries. But I am
>>> possibly out of date on this.
>> Well, I tried that and I gave it up since Kaldi nicely uses OpenFST and I
>> was not able to wrap OpenFST with just plain C (It may be possible).
>> I used Cython and pyfst mainly because pyfst solved for me wrapping up
>> OpenFST and I am really glad that 99% of wrapping OpenFST templates was
>> carried out by somebody else (Victor Chahuneau).
>>>
>>> 3. If 2 is correct shouldn't we define our API and wrap that? Producing a
>>> formal list of functionality that should be exposed to things like client
>>> and server applications?
>>>
>>>
>>> I would encourage some care here. Unconstrained wrapping can lead to
>>> systems which HAVE to use the scripting language (We can already see how
>>> difficult it is to move away from the Perl scripting if you wish to). Also
>>> never, never, never reverse wrap (i.e. call python from within C++), yes it
>>> can be done but that way lays madness.
>>>
>>> v best
>>>
>>> Matthew
>>>
>>>
>>> On Thu, Dec 18, 2014 at 11:37 PM, Daniel Povey <dp...@gm...> wrote:
>>>> Jan-
>>>> I haven't seen any objections to your setup.  I'd say we should plan
>>>> to include it in Kaldi at some point (e.g. within the next few
>>>> months), but in the meantime hopefully you can continue to work on it,
>>>> and maybe come up with some other examples of how it's useful to do
>>>> the interfacing with Python- e.g. some kind of application level or
>>>> service-level thing?
>>>> Dan
>>>>
>>>>
>>>> On Sat, Dec 13, 2014 at 4:01 PM, Yajie Miao <yaj...@gm...> wrote:
>>>>> Hi Jan,
>>>>> This is very nice work! In our PDNN toolkit, we also have simple python
>>>>> wrappers to read and write Kaldi features, mainly for DNN training.
>>>>> Your
>>>>> implementation looks like a more comprehensive version.
>>>>>
>>>>> Do you have the functions/commands to do feature splicing? I ask this
>>>>> because we found doing splicing on the fly with Python highly
>>>>> expensive.
>>>>> That's why we still stick to PFiles instead of Kaldi features (.scp
>>>>> .ark)
>>>>> for DNN triaining. I am very  interested to know the efficiency of your
>>>>> splicing implementation.
>>>>>
>>>>> Thanks,
>>>>> Yajie
>>>>>
>>>>> On Sat, Dec 13, 2014 at 5:59 PM, Daniel Povey <dp...@gm...> wrote:
>>>>>> OK, thanks.
>>>>>> cc'ing Yajie in case he wants to comment.
>>>>>> Dan
>>>>>>
>>>>>>
>>>>>> On Sat, Dec 13, 2014 at 2:31 PM, Jan Chorowski
>>>>>> <jan...@gm...>
>>>>>> wrote:
>>>>>>> Hi All,
>>>>>>>
>>>>>>> the wrapper is built during Kaldi compilation. I build it using
>>>>>>> provided
>>>>>>> Makefile. The build depends on:
>>>>>>> 1. Python and numpy (by default it queries the python interpreter
>>>>>>> found
>>>>>>> on
>>>>>>> the path for header file location)
>>>>>>> 2. Boost with Boost::Python library. It is quite heavy to build, but
>>>>>>> most
>>>>>>> Linux distributions ship it. Boost python doesn't require any code
>>>>>>> generation steps, the wrapper is defined in a normal c++ code file.
>>>>>>>
>>>>>>> During build Python and Boost libraries and Kaldi object files are
>>>>>>> linked
>>>>>>> into a CPython extention module,
>>>>>>> kaldi/src/python/kaldi_io_internal.so.
>>>>>>> It
>>>>>>> works with both static and shared Kaldi builds. Further usage
>>>>>>> requires
>>>>>>> that
>>>>>>> python finds kaldi_io.py and kaldi_io_internal.so on the PYTHONPATH
>>>>>>> - it
>>>>>>> can
>>>>>>> be for example added to the PYTHONPATH variable in the path.sh
>>>>>>> script of
>>>>>>> a
>>>>>>> recipe.
>>>>>>>
>>>>>>> Jan
>>>>>>>
>>>>>>>
>>>>>>> On 12/13/2014 3:33 PM, Daniel Povey wrote:
>>>>>>>> Also, Jan- could you send us an email explaining how this works-
>>>>>>>>     How does Python "see" the C++ headers?  Do you have to invoke
>>>>>>>> some
>>>>>>>> special program, like swig?  Do you have to write some special kind
>>>>>>>> of
>>>>>>>> header that shows how the C++ objects are to be interpreted by
>>>>>>>> python?
>>>>>>>>    A brief example would be helpful, if so.
>>>>>>>>     How is the resulting program linked, if at all?  If you require
>>>>>>>> functions C++ libraries, are these obtained from the .a or .so
>>>>>>>> files
>>>>>>>> at runtime, or compiled into some kind of executable-like blob at
>>>>>>>> compile time?  Does your framework require that Kaldi be compiled
>>>>>>>> using dynamic (.so) libraries?
>>>>>>>>
>>>>>>>> Dan
>>>>>>>>
>>>>>>>>
>>>>>>>> On Sat, Dec 13, 2014 at 12:04 PM, Jan Chorowski
>>>>>>>> <jan...@gm...>
>>>>>>>> wrote:
>>>>>>>>> Hello Dan,
>>>>>>>>>
>>>>>>>>> thank you for the comments. I tried to make it in the Kaldi
>>>>>>>>> spirit,
>>>>>>>>> consistency is important. Of course, the scripts can be removed
>>>>>>>>> and
>>>>>>>>> replaced
>>>>>>>>> with some more useful examples. I don't have too much experience
>>>>>>>>> with
>>>>>>>>> bridging Python to C++, so any critique on the wrappers and the
>>>>>>>>> approach
>>>>>>>>> taken is welcome.
>>>>>>>>>
>>>>>>>>> Jan
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On 12/13/2014 2:55 PM, Daniel Povey wrote:
>>>>>>>>>> Hi all.
>>>>>>>>>>    From a first look, it does look very impressive, and nicely
>>>>>>>>>> documented.
>>>>>>>>>> I would appreciate it if people on the list who have Python
>>>>>>>>>> experience
>>>>>>>>>> would comment on this- you can either reply to this thread, or to
>>>>>>>>>> me.
>>>>>>>>>> I don't know if this has been done in the "natural" way, or if
>>>>>>>>>> there
>>>>>>>>>> is some reason why people in the future will say, "why did you do
>>>>>>>>>> it
>>>>>>>>>> this way, you should have done XXX".
>>>>>>>>>>
>>>>>>>>>> Jan:
>>>>>>>>>> in the scripts/ directory you seem to have some examples of how
>>>>>>>>>> you
>>>>>>>>>> can create python programs that behave very much like Kaldi
>>>>>>>>>> command-line programs, using your framework.  This is very
>>>>>>>>>> useful.
>>>>>>>>>> However, the programs
>>>>>>>>>>     apply-global-cmvn.py
>>>>>>>>>>     compute-global-cmvn-stats.py
>>>>>>>>>> are perhaps a little confusing because they provide the same
>>>>>>>>>> functionality that you could get with "compute-cmvn-stats ->
>>>>>>>>>> matrix-sum" and "apply-cmvn" on the output of that command; and
>>>>>>>>>> they
>>>>>>>>>> do so using different formats for the CMVN information.  I know
>>>>>>>>>> the
>>>>>>>>>> format of storing the CMVN stats in a two-row matrix is perhaps
>>>>>>>>>> not
>>>>>>>>>> perfectly ideal, but it's a standard within Kaldi and it would be
>>>>>>>>>> confusing to deviate from that standard.
>>>>>>>>>> Of course, this is a very minor issue that doesn't affect the
>>>>>>>>>> validity
>>>>>>>>>> of the framework as a whole.  I am just pointing this out; the
>>>>>>>>>> main
>>>>>>>>>> discussion should be about the framework and whether people feel
>>>>>>>>>> it's
>>>>>>>>>> the "right" way to do this.
>>>>>>>>>>
>>>>>>>>>> Dan
>>>>>>>>>>
>>>>>>>>>> On Sat, Dec 13, 2014 at 6:28 AM, Jan Chorowski
>>>>>>>>>> <jan...@gm...>
>>>>>>>>>> wrote:
>>>>>>>>>>> Hi all!
>>>>>>>>>>>
>>>>>>>>>>> I've written wrappers to access Kaldi data files from within
>>>>>>>>>>> Python
>>>>>>>>>>> using boost::python (the code is on github
>>>>>>>>>>>
>>>>>>>>>>> https://github.com/janchorowski/kaldi-git/tree/python/src/python).
>>>>>>>>>>> If
>>>>>>>>>>> you think this would be an interesting addition please instruct
>>>>>>>>>>> me
>>>>>>>>>>> how
>>>>>>>>>>> to contribute.
>>>>>>>>>>>
>>>>>>>>>>> Best Regards,
>>>>>>>>>>> Jan Chorowski
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> ------------------------------------------------------------------------------
>>>>>>>>>>> Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT
>>>>>>>>>>> Server
>>>>>>>>>>> from Actuate! Instantly Supercharge Your Business Reports and
>>>>>>>>>>> Dashboards
>>>>>>>>>>> with Interactivity, Sharing, Native Excel Exports, App
>>>>>>>>>>> Integration &
>>>>>>>>>>> more
>>>>>>>>>>> Get technology previously reserved for billion-dollar
>>>>>>>>>>> corporations,
>>>>>>>>>>> FREE
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
>>>>>>>>>>> _______________________________________________
>>>>>>>>>>> Kaldi-developers mailing list
>>>>>>>>>>> Kal...@li...
>>>>>>>>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>>>>>>>
>>>>>
>>>>
>>>> ------------------------------------------------------------------------------
>>>> Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
>>>> from Actuate! Instantly Supercharge Your Business Reports and Dashboards
>>>> with Interactivity, Sharing, Native Excel Exports, App Integration & more
>>>> Get technology previously reserved for billion-dollar corporations, FREE
>>>>
>>>> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
>>>> _______________________________________________
>>>> Kaldi-developers mailing list
>>>> Kal...@li...
>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>
>>>
>>> ------------------------------------------------------------------------------
>>> Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
>>> from Actuate! Instantly Supercharge Your Business Reports and Dashboards
>>> with Interactivity, Sharing, Native Excel Exports, App Integration & more
>>> Get technology previously reserved for billion-dollar corporations, FREE
>>>
>>> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
>>> _______________________________________________
>>> Kaldi-developers mailing list
>>> Kal...@li...
>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>>
>>
>> --
>> Ondřej Plátek, +420 737 758 650, skype:ondrejplatek, ond...@gm...