Python separation.bss_eval_sources方法代码示例

# 需要导入模块: from mir_eval import separation [as 别名]
# 或者: from mir_eval.separation import bss_eval_sources [as 别名]
def cal_SDRi(src_ref, src_est, mix):
    """Calculate Source-to-Distortion Ratio improvement (SDRi).
    NOTE: bss_eval_sources is very very slow.
    Args:
        src_ref: numpy.ndarray, [C, T]
        src_est: numpy.ndarray, [C, T], reordered by best PIT permutation
        mix: numpy.ndarray, [T]
    Returns:
        average_SDRi
    """
    src_anchor = np.stack([mix, mix], axis=0)
    sdr, sir, sar, popt = bss_eval_sources(src_ref, src_est)
    sdr0, sir0, sar0, popt0 = bss_eval_sources(src_ref, src_anchor)
    avg_SDRi = ((sdr[0]-sdr0[0]) + (sdr[1]-sdr0[1])) / 2
    # print("SDRi1: {0:.2f}, SDRi2: {1:.2f}".format(sdr[0]-sdr0[0], sdr[1]-sdr0[1]))
    return avg_SDRi 

# 需要导入模块: from mir_eval import separation [as 别名]
# 或者: from mir_eval.separation import bss_eval_sources [as 别名]
def convergence_callback(Y, **kwargs):
        global SDR, SIR, ref
        from mir_eval.separation import bss_eval_sources

        if Y.shape[2] == 1:
            y = pra.transform.synthesis(
                Y[:, :, 0], framesize, framesize // 2, win=win_s
            )[:, None]
        else:
            y = pra.transform.synthesis(Y, framesize, framesize // 2, win=win_s)

        if args.algo != "blinkiva":
            new_ord = np.argsort(np.std(y, axis=0))[::-1]
            y = y[:, new_ord]

        m = np.minimum(y.shape[0] - framesize // 2, ref.shape[1])
        sdr, sir, sar, perm = bss_eval_sources(
            ref[:n_sources_target, :m, 0],
            y[framesize // 2 : m + framesize // 2, :n_sources_target].T,
        )
        SDR.append(sdr)
        SIR.append(sir) 

# 需要导入模块: from mir_eval import separation [as 别名]
# 或者: from mir_eval.separation import bss_eval_sources [as 别名]
def val(epoch):
    model.eval()
    start_time = time.time()
    cnt, aveloss,avesdr = 0, 0, 0
    with torch.no_grad():
        for iloader, xtrain, ytrain in loadval:
            for ind in range(0, xtrain.shape[-1], sampleSize):
                if (xtrain[0, 0, ind:ind + sampleSize].shape[0] < (sampleSize)): break
                output = model(xtrain[:, :, ind:ind + sampleSize].to(device))
                loss = criterion(output, (ytrain[:, :, ind:ind + sampleSize].to(device)))
                #if epoch % 5 == 0:
                #    if(np.sum(output.cpu().numpy()==0) == sampleSize or np.sum(ytrain[:, 0, ind:ind + sampleSize].numpy())==sampleSize):
                #        print("all slience")
                #    else:
                #       avesdr += bss_eval_sources(output[:,0,:].cpu().numpy(), ytrain[:, 0, ind:ind + sampleSize].numpy())[0].mean()
                cnt += 1
                aveloss += float(loss)
        aveloss /= cnt
        #avesdr /= cnt
        #if epoch % 5 == 0:print('sdr:',avesdr)
        print('loss for validation:{:.5f},epoch{},valtime{}'.format(aveloss, epoch,np.round(time.time() - start_time)))
        if (USEBOARD): writer.add_scalar('waveunet val loss', aveloss, iteration) 

# 需要导入模块: from mir_eval import separation [as 别名]
# 或者: from mir_eval.separation import bss_eval_sources [as 别名]
def bss_eval_global(wavs_mono, wavs_src1, wavs_src2, wavs_src1_pred, wavs_src2_pred):

    assert len(wavs_mono) == len(wavs_src1) == len(wavs_src2) == len(wavs_src1_pred) == len(wavs_src2_pred)

    num_samples = len(wavs_mono)

    gnsdr = np.zeros(2)
    gsir = np.zeros(2)
    gsar = np.zeros(2)
    frames_total = 0

    for wav_mono, wav_src1, wav_src2, wav_src1_pred, wav_src2_pred in zip(wavs_mono, wavs_src1, wavs_src2, wavs_src1_pred, wavs_src2_pred):
        len_cropped = wav_src1_pred.shape[-1]
        wav_mono_cropped = wav_mono[:len_cropped]
        wav_src1_cropped = wav_src1[:len_cropped]
        wav_src2_cropped = wav_src2[:len_cropped]

        sdr, sir, sar, _ = bss_eval_sources(reference_sources = np.asarray([wav_src1_cropped, wav_src2_cropped]), estimated_sources = np.asarray([wav_src1_pred, wav_src2_pred]), compute_permutation = False)
        sdr_mono, _, _, _ = bss_eval_sources(reference_sources = np.asarray([wav_src1_cropped, wav_src2_cropped]), estimated_sources = np.asarray([wav_mono_cropped, wav_mono_cropped]), compute_permutation = False)

        nsdr = sdr - sdr_mono
        gnsdr += len_cropped * nsdr
        gsir += len_cropped * sir
        gsar += len_cropped * sar
        frames_total += len_cropped

    gnsdr = gnsdr / frames_total
    gsir = gsir / frames_total
    gsar = gsar / frames_total

    return gnsdr, gsir, gsar 

# 需要导入模块: from mir_eval import separation [as 别名]
# 或者: from mir_eval.separation import bss_eval_sources [as 别名]
def convergence_callback(Y):
        global SDR, SIR
        from mir_eval.separation import bss_eval_sources
        ref = np.moveaxis(separate_recordings, 1, 2)
        y = pra.transform.stft.synthesis(Y, L, hop, win=win_s)
        y = y[L-hop: , :].T
        m = np.minimum(y.shape[1], ref.shape[1])
        sdr, sir, sar, perm = bss_eval_sources(ref[:, :m, 0], y[:, :m])
        SDR.append(sdr)
        SIR.append(sir)

    ## STFT ANALYSIS 

# 需要导入模块: from mir_eval import separation [as 别名]
# 或者: from mir_eval.separation import bss_eval_sources [as 别名]
def run(args):

    sep_reader = AudioReader(args.sep_scp)
    ref_reader = AudioReader(args.ref_scp)
    utt_snr = open(args.per_utt, "w") if args.per_utt else None
    utt_ali = open(args.utt_ali, "w") if args.utt_ali else None
    reporter = Report(args.spk2class)
    # sep: N x S
    for key, sep in sep_reader:
        # ref: N x S
        ref = ref_reader[key]
        # keep same shape
        nsamps = min(sep.shape[-1], ref.shape[-1])
        sdr, _, _, ali = bss_eval_sources(ref[:, :nsamps], sep[:, :nsamps])
        sdr = np.mean(sdr)
        reporter.add(key, sdr)
        if utt_snr:
            utt_snr.write("{}\t{:.2f}\n".format(key, sdr))
        if utt_ali:
            ali_str = " ".join(map(str, ali))
            utt_ali.write(f"{key}\t{ali_str}\n")
    reporter.report()
    if utt_snr:
        utt_snr.close()
    if utt_ali:
        utt_ali.close() 

# 需要导入模块: from mir_eval import separation [as 别名]
# 或者: from mir_eval.separation import bss_eval_sources [as 别名]
def test(epoch):  # testing data
    model.eval()
    start_time = time.time()
    with torch.no_grad():
        avesdr = 0
        numSongs = 0
        sdrmedian = np.zeros(50)
        for iloader, xtrain, ytrain in loadtest:
            iloader=iloader.item()
            listofpred0 = []
            cnt,aveloss=0,0
            for ind in range(0, xtrain.shape[-1] - sampleSize, sampleSize):
                if (xtrain[0, 0, ind:ind + sampleSize].shape[0] < (sampleSize)): break
                output = model(xtrain[:, :,ind:ind + sampleSize].to(device))
                listofpred0.append(output.reshape(-1).cpu().numpy())
                loss = criterion(output, (ytrain[:, :,ind:ind + sampleSize].to(device)))
                cnt+=1
                aveloss += float(loss)
            aveloss /= cnt
            print('loss for test:{},num{},epoch{}'.format(aveloss, iloader,epoch))
            ans0 = mu_law_decode(np.concatenate(listofpred0))
            if(iloader >= 150):
                sdr = bss_eval_sources(mu_law_decode(ytrain[0,0,:ans0.shape[0]].cpu().numpy()), ans0)
                avesdr += sdr[0][0]
                sdrmedian[iloader-150] = sdr[0][0]
                #print('each ele of median',sdrmedian[iloader-150],iloader-150)
                numSongs += 1
            if(iloader > 160):continue
            if not os.path.exists('vsCorpus/'): os.makedirs('vsCorpus/')
            sf.write(savemusic.format(iloader), ans0, sample_rate)
            print('test stored done', np.round(time.time() - start_time))
        print('sdr mean:', avesdr / numSongs)
        print('sdr median:', np.median(sdrmedian))