from aubio.aubioclass import * from onset import taskonset class taskbeat(taskonset): def __init__(self,input,params=None,output=None): """ open the input file and initialize arguments parameters should be set *before* calling this method. """ taskonset.__init__(self,input,output=None,params=params) self.btwinlen = 512**2/self.params.hopsize self.btstep = self.btwinlen/4 self.btoutput = fvec(self.btstep,self.channels) self.dfframe = fvec(self.btwinlen,self.channels) self.bt = beattracking(self.btwinlen,self.channels) self.pos2 = 0 self.old = -1000 def __call__(self): taskonset.__call__(self) #results = taskonset.__call__(self) # write to current file if self.pos2 == self.btstep - 1 : self.bt.do(self.dfframe,self.btoutput) for i in range (self.btwinlen - self.btstep): self.dfframe.set(self.dfframe.get(i+self.btstep,0),i,0) for i in range(self.btwinlen - self.btstep, self.btwinlen): self.dfframe.set(0,i,0) self.pos2 = -1; self.pos2 += 1 val = self.opick.pp.getval() #if not results: val = 0 #else: val = results[1] self.dfframe.set(val,self.btwinlen - self.btstep + self.pos2,0) i=0 for i in range(1,int( self.btoutput.get(0,0) ) ): if self.pos2 == self.btoutput.get(i,0) and \ aubio_silence_detection(self.myvec(), self.params.silence)!=1: now = self.frameread-0 period = (60 * self.params.samplerate) / ((now - self.old) * self.params.hopsize) self.old = now return now,period def eval(self,results,tol=0.20,tolcontext=0.25): obeats = self.gettruth() etime = [result[0] for result in results] otime = [obeat[0] for obeat in obeats] CML_tot, CML_max, CML_start, CML_end = 0,0,0,0 AML_tot, AML_max, AML_start, AML_end = 0,0,0,0 AMLd_tot, AMLd_max, AMLd_start, AMLd_end = 0,0,0,0 AMLh_tot, AMLh_max, AMLh_start, AMLh_end = 0,0,0,0 AMLo_tot, AMLo_max, AMLo_start, AMLo_end = 0,0,0,0 # results iteration j = 1 # for each annotation for i in range(2,len(otime)-2): if j+1 >= len(etime): break count = 0 # look for next matching beat while otime[i] > etime[j] - (otime[i] - otime[i+1])*tol: if count > 0: #print "spurious etime" if CML_end - CML_start > CML_max: CML_max = CML_end - CML_start CML_start, CML_end = j, j if AMLh_end - AMLh_start > AMLh_max: AMLh_max = AMLh_end - AMLh_start AMLh_start, AMLh_end = j, j if AMLd_end - AMLd_start > AMLd_max: AMLd_max = AMLd_end - AMLd_start AMLd_start, AMLd_end = j, j if AMLo_end - AMLo_start > AMLo_max: AMLo_max = AMLo_end - AMLo_start AMLo_start, AMLo_end = j, j j += 1 count += 1 if j+1 >= len(etime): break #print otime[i-1],etime[j-1]," ",otime[i],etime[j]," ",otime[i+1],etime[j+1] prevtempo = (otime[i] - otime[i-1]) nexttempo = (otime[i+1] - otime[i]) current0 = (etime[j] > otime[i] - prevtempo*tol) current1 = (etime[j] < otime[i] + prevtempo*tol) # check correct tempo prev0 = (etime[j-1] > otime[i-1] - prevtempo*tolcontext) prev1 = (etime[j-1] < otime[i-1] + prevtempo*tolcontext) next0 = (etime[j+1] > otime[i+1] - nexttempo*tolcontext) next1 = (etime[j+1] < otime[i+1] + nexttempo*tolcontext) # check for off beat prevoffb0 = (etime[j-1] > otime[i-1] - prevtempo/2 - prevtempo*tolcontext) prevoffb1 = (etime[j-1] < otime[i-1] - prevtempo/2 + prevtempo*tolcontext) nextoffb0 = (etime[j+1] > otime[i+1] - nexttempo/2 - nexttempo*tolcontext) nextoffb1 = (etime[j+1] < otime[i+1] - nexttempo/2 + nexttempo*tolcontext) # check half tempo prevhalf0 = (etime[j-1] > otime[i-1] + prevtempo - prevtempo/2*tolcontext) prevhalf1 = (etime[j-1] < otime[i-1] + prevtempo + prevtempo/2*tolcontext) nexthalf0 = (etime[j+1] > otime[i+1] - nexttempo - nexttempo/2*tolcontext) nexthalf1 = (etime[j+1] < otime[i+1] - nexttempo + nexttempo/2*tolcontext) # check double tempo prevdoub0 = (etime[j-1] > otime[i-1] - prevtempo - prevtempo*2*tolcontext) prevdoub1 = (etime[j-1] < otime[i-1] - prevtempo + prevtempo*2*tolcontext) nextdoub0 = (etime[j+1] > otime[i+1] + nexttempo - nexttempo*2*tolcontext) nextdoub1 = (etime[j+1] < otime[i+1] + nexttempo + nexttempo*2*tolcontext) if current0 and current1 and prev0 and prev1 and next0 and next1: #print "YES!" CML_end = j CML_tot += 1 else: if CML_end - CML_start > CML_max: CML_max = CML_end - CML_start CML_start, CML_end = j, j if current0 and current1 and prevhalf0 and prevhalf1 and nexthalf0 and nexthalf1: AMLh_end = j AMLh_tot += 1 else: if AMLh_end - AMLh_start > AMLh_max: AMLh_max = AMLh_end - AMLh_start AMLh_start, AMLh_end = j, j if current0 and current1 and prevdoub0 and prevdoub1 and nextdoub0 and nextdoub1: AMLd_end = j AMLd_tot += 1 else: if AMLd_end - AMLd_start > AMLd_max: AMLd_max = AMLd_end - AMLd_start AMLd_start, AMLd_end = j, j if current0 and current1 and prevoffb0 and prevoffb1 and nextoffb0 and nextoffb1: AMLo_end = j AMLo_tot += 1 else: if AMLo_end - AMLo_start > AMLo_max: AMLo_max = AMLo_end - AMLo_start AMLo_start, AMLo_end = j, j # look for next matching beat count = 0 while otime[i] > etime[j] - (otime[i] - otime[i+1])*tolcontext: j += 1 if count > 0: #print "spurious etime" start = j count += 1 total = float(len(otime)) CML_tot /= total AMLh_tot /= total AMLd_tot /= total AMLo_tot /= total CML_cont = CML_max/total AMLh_cont = AMLh_max/total AMLd_cont = AMLd_max/total AMLo_cont = AMLo_max/total return CML_cont, CML_tot, AMLh_cont, AMLh_tot, AMLd_cont, AMLd_tot, AMLo_cont, AMLo_tot # for i in allfreq: # freq.append(float(i) / 2. / N * samplerate ) # while freq[i]>freqs[j]: # j += 1 # a0 = weight[j-1] # a1 = weight[j] # f0 = freqs[j-1] # f1 = freqs[j] # if f0!=0: # iweight.append((a1-a0)/(f1-f0)*freq[i] + (a0 - (a1 - a0)/(f1/f0 -1.))) # else: # iweight.append((a1-a0)/(f1-f0)*freq[i] + a0) # while freq[i]>freqs[j]: # j += 1 def eval2(self,results,tol=0.2): truth = self.gettruth() obeats = [i[0] for i in truth] ebeats = [i[0]*self.params.step for i in results] NP = max(len(obeats), len(ebeats)) N = int(round(max(max(obeats), max(ebeats))*100.)+100) W = int(round(tol*100.*60./median([i[1] for i in truth], len(truth)/2))) ofunc = [0 for i in range(N+W)] efunc = [0 for i in range(N+W)] for i in obeats: ofunc[int(round(i*100.)+W)] = 1 for i in ebeats: efunc[int(round(i*100.)+W)] = 1 assert len(obeats) == sum(ofunc) autocor = 0; m =0 for m in range (-W, W): for i in range(W,N): autocor += ofunc[i] * efunc[i-m] autocor /= float(NP) return autocor def evaluation(self,results,tol=0.2,start=5.): """ beat tracking evaluation function computes P-score of experimental results (ebeats) against ground truth annotations (obeats) """ from aubio.median import short_find as median truth = self.gettruth() ebeats = [i[0]*self.params.step for i in results] obeats = [i[0] for i in truth] # trim anything found before start while obeats[0] < start: obeats.pop(0) while ebeats[0] < start: ebeats.pop(0) # maximum number of beats found NP = max(len(obeats), len(ebeats)) # length of ofunc and efunc vector N = int(round(max(max(obeats), max(ebeats))*100.)+100) # compute W median of ground truth tempi tempi = [] for i in range(1,len(obeats)): tempi.append(obeats[i]-obeats[i-1]) W = int(round(tol*100.*median(tempi,len(tempi)/2))) # build ofunc and efunc functions, starting with W zeros ofunc = [0 for i in range(N+W)] efunc = [0 for i in range(N+W)] for i in obeats: ofunc[int(round(i*100.)+W)] = 1 for i in ebeats: efunc[int(round(i*100.)+W)] = 1 # optional: make sure we didn't miss any beats assert len(obeats) == sum(ofunc) assert len(ebeats) == sum(efunc) # compute auto correlation autocor = 0; m =0 for m in range (-W, W): for i in range(W,N): autocor += ofunc[i] * efunc[i-m] autocor /= float(NP) return autocor def gettruth(self): import os.path from aubio.txtfile import read_datafile datafile = self.input.replace('.wav','.txt') if not os.path.isfile(datafile): print "no ground truth " return False,False else: values = read_datafile(datafile,depth=0) old = -1000 for i in range(len(values)): now = values[i] period = 60 / (now - old) old = now values[i] = [now,period] return values def plot(self,oplots,results): import Gnuplot oplots.append(Gnuplot.Data(results,with_='linespoints',title="auto")) def plotplot(self,wplot,oplots,outplot=None,extension=None,xsize=1.,ysize=1.,spectro=False): import Gnuplot from aubio.gnuplot import gnuplot_create, audio_to_array, make_audio_plot import re # audio data #time,data = audio_to_array(self.input) #f = make_audio_plot(time,data) g = gnuplot_create(outplot=outplot, extension=extension) oplots = [Gnuplot.Data(self.gettruth(),with_='linespoints',title="orig")] + oplots g.plot(*oplots)