testing_deployment/code/server_lib/asr_system.py

__all__= ['load']

import sys,threading,collections,time
import TLK
import lmodel,oepmodel,segmodel

import logging as Log

Hyp= collections.namedtuple('Hyp',['var','novar','err',
                                   'score','nframes',
                                   'eos'])

class Recogniser:

    def __init__(self,models,params,
#                 uppercase_tbl,uppercase_tbl_bi,
                 sil_word,sil_sym,fea_freq,mng):
        args= {
            'ifreq'  : 16000,
            'bsize'  : 400,
            'amodel' : models['amodel'],
            'lm'     : models['lm'],
            'params' : params,
            'output' : self,
        }

        self._fea_freq= fea_freq
        if sil_word!=None:
            args['sil_word']= sil_word
            if sil_sym!=None:
                args['sil_sym']= sil_sym
        if models['dnn']!=None: args['dnn']= models['dnn']
        if models['dlm']!=None: args['dlm']= models['dlm']
        if models['mustd']!=None: args['mustd']= models['mustd']
        if models['oep-factory']!=None:
            oep= models['oep-factory'](models['oep-model'])
            args['oep']= oep
        self.reco= TLK.ORecogniser(**args)
        if models['seg-factory']!=None:
            self.seg= models['seg-factory'](models['seg-model'])
        else: self.seg= None
        self._mng= mng
        self._cv= threading.Condition()

    def reset(self):
        self._err= False
        self._hyps= []
        self._prev= None
        if self.seg is not None:
            self.seg.reset()
        # Per a gestionar tokens insertats el que vaig a fer es tindre
        # una llista de llistes. En la primera llista estaran els
        # tokens del segment que s'està processant actualment i en
        # l'última els tokens de l'últim segment. Cada token serà una
        # tupla paraula i número de frames on es va insertar.
        self._itokens= [[]]
        self._inframes= 0.0 # Numero de frames del segment actual
                            # insertats en el reconeixedor

    def __register_itoken_nosec(self,token):
        token= token.strip()
        if token=='': return
        nframes= int(self._inframes) if self._inframes>0.0 else None
        self._itokens[-1].append((token,nframes))

    def __register_itoken(self,token):
        try:
            self.__register_itoken_nosec(token)
        except:
            pass
    
    def __register_itoken_eos(self):
        if self._inframes>0.0:
            self._itokens.append([])
            self._inframes= 0.0

    def __merge_itokens_nosec(self,code,hyp):
        # Comprovacions inicials i eliminació
        itokens= self._itokens[0]
        if code==TLK.OUT_RES: self._itokens.pop(0)
        if itokens==[]: return hyp
        # Inserta
        hyp,pos_var= hyp[0],hyp[1]
        pref,novar,var= [],list(hyp[:pos_var]),list(hyp[pos_var:])
        # -> Inserta inici segment
        while len(itokens)>0 and itokens[0][1] is None:
            tok= itokens.pop(0)
            pref.append((tok[0],0,0))
        # --> Processa novar
        if len(itokens)>0 and len(novar)>0:
            i,new_novar= 0,[]
            while len(itokens)>0 and i<len(novar):
                w,b,e= novar[i]
                epos= b+e
                while len(itokens)>0 and itokens[0][1]<epos:
                    tok= itokens.pop(0)
                    new_novar.append((tok[0],b,0))
                new_novar.append((w,b,e))
                i+= 1
            while i<len(novar):
                new_novar.append(novar[i])
                i+= 1
        else: new_novar= novar
        # --> Processa var
        if len(itokens)>0 and len(var)>0:
            i,j,new_var= 0,0,[]
            while len(itokens)>0 and i<len(var):
                w,b,e= var[i]
                epos= b+e
                while j<len(itokens) and itokens[j][1]<epos:
                    new_var.append((itokens[j][0],b,0))
                    j+= 1
                new_var.append((w,b,e))
                i+= 1
            while i<len(var):
                new_var.append(var[i])
                i+= 1
        else: new_var= var
        # --> Afegeix pendents si code==TLK.OUT_RES
        if code==TLK.OUT_RES and len(itokens)>0:
            assert var==[]
            pos= new_novar[-1][1]+new_novar[-1][2] if len(new_novar)>0 else -1
            for tok in itokens:
                tmp_pos= pos if pos!=-1 else tok[1]
                new_novar.append((tok[0],tmp_pos,0))
        return pref+new_novar+new_var,len(pref)+len(new_novar)
    # end __merge_itokens_nosec

    def __merge_itokens(self,code,hyp):
        try:
            return self.__merge_itokens_nosec(code,hyp)
        except:
            return hyp
    
    # None denotes end of segment
    def feed(self,data=None):
        if data==None:
            self.reco.feed()
        elif not self._err:
            if type(data)==str:
                self.__register_itoken(data)
            elif len(data)==0:
                self.reco.split()
                self.__register_itoken_eos()
            else:
                self._inframes+= (len(data)/32000)*self._fea_freq
                self.reco.feed(data)

    def __write_hyp(self,novar,var,score,nframes,err,eos):
        h= Hyp(novar=novar,var=var,err=err,
               score=score,nframes=nframes,
               eos=eos)
        with self._cv:
            self._hyps.append(h)
            self._cv.notify_all()
        
    def process_out(self,code,hyp,stats):
        #def totxt(rec,last_word):
        #    aux= []
        #    for x in rec:
        #        w= x[0].lower()
        #        aux.append(toupper(w,last_word))
        #        last_word= w
        #    ret= ' '.join(aux)
        #    ret= ret.replace('[hesitation]','').replace('<unk>','')
        #    return ret
        def totxt(rec):
            ret= ' '.join([x[0] for x in rec])
            ret= ret.replace('[hesitation]','').replace('<unk>','')
            return ret
        hyp,pos_var= hyp[0],hyp[1]
        novar= hyp[:pos_var]
        #new_last_word= novar[-1][0] if len(novar)>0 else self.last_word
        #novar= totxt(novar,self.last_word)
        #var= totxt(hyp[pos_var:],new_last_word)
        #self.last_word= new_last_word
        novar= totxt(novar)
        var= totxt(hyp[pos_var:])
        if novar=='' and var==self._prev and code==TLK.OUT_HYP: return
        self._prev= var
        score,nframes= stats
        self.__write_hyp(novar,var,score,nframes,False,
                         code==TLK.OUT_RES)
        
    # Output method
    def write(self,code,hyp,stats):
        hyp= self.__merge_itokens(code,hyp)
        if code==TLK.OUT_ERR:
            txt= self._prev if self._prev!=None else ''
            self.__write_hyp(txt,'',0,0,False,False)
            self._err= True
            return
        if code==TLK.OUT_END:
            self.__write_hyp(None,None,0,0,self._err,False)
            return
        if self.seg is not None:
            hyp= self.seg(hyp)
            if code==TLK.OUT_RES:
                hyp0,pos_var0= hyp
                hyp1= self.seg.eos()
                if hyp1 is not None:
                    hyp1,pos_var1= hyp1
                    hyp= hyp0[:pos_var0]+hyp1[:pos_var1]
                    hyp= hyp,len(hyp)
                self.seg.reset()
        self.process_out(code,hyp,stats)

    @property
    def output(self):
        end= False
        while not end:
            with self._cv:
                self._cv.wait_for(lambda: len(self._hyps)>0)
                ret= self._hyps.pop(0)
                end= ret.novar is None
            yield ret
    
# end Recogniser

class RecogniserManager:

    def __init__(self,name,tag,date,lang,
                 models,params,nreco,
                 sil_word,sil_sym,fea_freq):
        self.v= []
        self.vv= [] # Reference copy
        #self.cmllr= params.cmllr_enabled
        for n in range(0,nreco):
            reco= Recogniser(models,params,
                             sil_word,sil_sym,
                             fea_freq,self)
            self.v.append(reco)
            self.vv.append(reco)
        self.name= name
        self.tag= tag
        self.date= date
        self.lang= lang
        self._lock= threading.Lock()
        self._models= models
        self._enabled= True

    def __len__(self):
        return len(self.vv)
    
    def get_reco(self):
        with self._lock:
            if self.v==[]: return None
            if not self._enabled: return None
            ret= self.v.pop()
            ret.reset()
            #ret.reset_cmllr()
            Log.info(('Recogniser taken from %s '+
                      '(available: %d of %d)')%(self.name,
                                                len(self.v),
                                                len(self.vv)))
            return ret
    
    def append(self,reco):
        with self._lock:
            self.v.append(reco)
            Log.info(('Recogniser from %s realeased'+
                      '(available: %d of %d)')%(self.name,
                                                len(self.v),
                                                len(self.vv)))
    
    def set_enabled(self,value):
        with self._lock:
            self._enabled= value
    
    @property
    def num_recos_available(self):
        with self._lock:
            return len(self.v)

    @property
    def enabled(self):
        with self._lock:
            return self._enabled

# end RecogniserManager

def prepare_params(cfg):
    params= TLK.OParameters()
    aux= cfg.get('hp')
    if aux!=None: params.hp= int(aux)
    aux= cfg.get('hp_min')
    if aux!=None: params.hp_min= int(aux)
    aux= cfg.get('wep')
    if aux!=None: params.wep= float(aux)
    aux= cfg.get('beam')
    if aux!=None: params.beam= float(aux)
    aux= cfg.get('dynormthr')
    if aux!=None: params.dynormthr= int(aux)
    aux= cfg.get('meannorm')
    if aux!=None: params.meannorm= bool(aux)
    aux= cfg.get('cmllr')
    if aux!=None: params.cmllr_enabled= bool(aux)
    aux= cfg.get('cmllr_nframes_step0')
    if aux!=None: params.cmllr_nframes_step0= int(aux)
    aux= cfg.get('cmllr_nframes')
    if aux!=None: params.cmllr_nframes= int(aux)
    aux= cfg.get('cmllr_niters')
    if aux!=None: params.cmllr_niters= int(aux)
    aux= cfg.get('numceps')
    if aux!=None: params.numceps= int(aux)
    aux= cfg.get('numchans')
    if aux!=None: params.numchans= int(aux)
    aux= cfg.get('accwindow')
    if aux!=None: params.accwindow= int(aux)
    aux= cfg.get('deltawindow')
    if aux!=None: params.deltawindow= int(aux)
    aux= cfg.get('hp_lm')
    if aux!=None: params.hp_lm= int(aux)
    aux= cfg.get('order')
    if aux!=None: params.order= int(aux)
    aux= cfg.get('ftype')
    if aux!=None: params.ftype= str(aux)
    aux= cfg.get('mustdnorm')
    if aux!=None: params.mustdnorm= bool(aux)
    aux= cfg.get('gsf')
    gsf= float(aux) if aux!=None else 1.0
    params.gsf= gsf
    aux= cfg.get('wip')
    wip= float(aux) if aux!=None else 0.0
    params.wip= wip
    params.amla= cfg.get('amla-fs')!=None
    params.sil= cfg.get('sil_sym','SP')
    
    # IMPORTANT !!!!! Deshabilita l'heurístic de segmentació
    params.sil_length= 20000
    params.sil_thr= 0.0

    return params
# end prepare_params

def load_models(cfg,name,params,sil_word):
    Log.info("Loading '%s' models..."%name)
    mustd= cfg.get('mustd')
    dnn= cfg.get('dnn')
    if params.cmllr_enabled:
        sys.exit('CMLLR NOT SUPPORTED !!!!')
        target= cfg.get('target')
        dnn_cmllr= cfg.get('dnn_cmllr')
    else:
        target= dnn_cmllr= None
        lexicon= cfg.get('lexicon')
        
    amodel= TLK.AModel(cfg['amodel'])
    if lexicon!=None:
        lex= TLK.Lexicon(lexicon, syms=amodel.syms)
        lm= TLK.LM()
        lm.load(cfg['lm'],lexicon=lex)
    else:
        try:
            lm= TLK.SearchGraph(cfg['lm'], syms=amodel.syms)
        except:
            lm= TLK.StaticLookaheadTables(cfg['lm'], syms=amodel.syms)
        dlm= cfg.get('dlm')
    if dlm!=None:
        dlm= lmodel.load_model(dlm)(lm.words)
    oep= cfg.get('oep')
    if oep!=None:
        step= cfg.get('oep-step')
        if step==None:
            sys.exit('oep-step not defined')
        step= int(step)
        lookahead= cfg.get('oep-lookahead')
        if lookahead==None:
            sys.exit('oep-lookahead not defined')
        lookahead= int(lookahead)
        priors_fn= cfg.get('oep-priors')
        amla_fs= cfg.get('amla-fs')
        tmp= oepmodel.load_model(oep,step,lookahead,
                                 priors_fn,amla_fs)
        oep_factory,oep_model= tmp
    else: oep_factory= oep_model= None
    seg= cfg.get('seg')
    if seg!=None:
        tmp= segmodel.load_model(seg,sil_word=sil_word)
        seg_factory,seg_model= tmp
    else: seg_factory= seg_model= None
    aux= cfg.get('gsf')
    gsf= float(aux) if aux!=None else 1.0
    aux= cfg.get('wip')
    wip= float(aux) if aux!=None else 0.0
    models= {}
    models["amodel"]=amodel
    models["lm"]=lm
    models["dnn"]=dnn
    models["gsf"]=gsf
    models["wip"]=wip
    models['dlm']= dlm
    models['oep-factory']= oep_factory
    models['oep-model']= oep_model
    models['mustd']= mustd
    models['seg-factory']= seg_factory
    models['seg-model']= seg_model
    return models

# end load_models
    
def load_system(conf):
    sil_word= conf.get('sil_word')
    sil_sym= conf.get('sil_sym')
    fea_freq= conf.get('fea_freq',100)
    params= prepare_params(conf)
    name= conf['id']
    tag= conf.get('tag',name)
    date= tuple(conf.get('date',[1,1,1971]))
    models= load_models(conf,tag,params,sil_word)
    Log.info("Creating '%s' recognisers..."%name)
    return RecogniserManager(name,tag,date,conf['lang'],
                             models,params,conf['nreco'],
                             sil_word,sil_sym,fea_freq)
#                             uppercase_tbl,uppercase_tbl_bi,
    # Uppercase-table (Deshabilitat de moment)
    #aux= cfg.get('uppercase-table')
    #uppercase_tbl= load_uppercase_table(aux) if aux!=None else None
    #aux= cfg.get('uppercase-table-bi')
    #uppercase_tbl_bi= load_uppercase_table_bi(aux) if aux!=None else None
    
# end create_recognisers