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
def bss_eval_global(mixed_wav, src1_wav, src2_wav, pred_src1_wav,
pred_src2_wav, n):
len_cropped = pred_src1_wav.shape[-1]
src1_wav = src1_wav[:, :len_cropped]
src2_wav = src2_wav[:, :len_cropped]
mixed_wav = mixed_wav[:, :len_cropped]
gnsdr, gsir, gsar = np.zeros(2), np.zeros(2), np.zeros(2)
total_len = 0
for i in range(n):
if np.sum(np.abs(src1_wav[i])) < 1e-10 or np.sum(np.abs(
src2_wav[i])) < 1e-10:
continue
sdr, sir, sar, _ = bss_eval_sources(
np.array([src1_wav[i], src2_wav[i]]),
np.array([pred_src1_wav[i], pred_src2_wav[i]]), False)
sdr_mixed, _, _, _ = bss_eval_sources(
np.array([src1_wav[i], src2_wav[i]]),
np.array([mixed_wav[i], mixed_wav[i]]), False)
nsdr = sdr - sdr_mixed
gnsdr += len_cropped * nsdr
gsir += len_cropped * sir
gsar += len_cropped * sar
total_len += len_cropped
gnsdr = gnsdr / total_len
gsir = gsir / total_len
gsar = gsar / total_len
return gnsdr, gsir, gsar
def bss_eval_global(wavs_mono, wavs_src1, wavs_src2, wavs_src1_pred, wavs_src2_pred):
print(len(wavs_mono), len(wavs_src1) , len(wavs_src2) , len(wavs_src1_pred) , len(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
step = 1
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
print("{}/{}\n".format(step, len(wavs_mono)))
step += 1
gnsdr = gnsdr / frames_total
gsir = gsir / frames_total
gsar = gsar / frames_total
return gnsdr, gsir, gsar
def bss_eval_global(mixed_wav, src1_wav, src2_wav, pred_src1_wav,
pred_src2_wav):
len_cropped = pred_src1_wav.shape[-1]
src1_wav = src1_wav[:, :len_cropped]
src2_wav = src2_wav[:, :len_cropped]
mixed_wav = mixed_wav[:, :len_cropped]
gnsdr = np.zeros(2)
gsir = np.zeros(2)
gsar = np.zeros(2)
total_len = 0
for i in range(EvalConfig.NUM_EVAL):
sdr, sir, sar, _ = bss_eval_sources(
np.array([src1_wav[i], src2_wav[i]]),
np.array([pred_src1_wav[i], pred_src2_wav[i]]), False)
sdr_mixed, _, _, _ = bss_eval_sources(
np.array([src1_wav[i], src2_wav[i]]),
np.array([mixed_wav[i], mixed_wav[i]]), False)
nsdr = sdr - sdr_mixed
gnsdr += len_cropped * nsdr
gsir += len_cropped * sir
gsar += len_cropped * sar
total_len += len_cropped
gnsdr = gnsdr / total_len
gsir = gsir / total_len
gsar = gsar / total_len
return gnsdr, gsir, gsar
def SDR(est, egs, mix):
'''
calculate SDR
est: Network generated audio
egs: Ground Truth
'''
sdr, _, _, _ = bss_eval_sources(egs, est)
mix_sdr, _, _, _ = bss_eval_sources(egs, mix)
return float(sdr - mix_sdr)
def SDR(est, egs, mix):
'''
calculate SDR
est: Network generated audio
egs: Ground Truth
'''
length = est.numpy().shape[0]
sdr, _, _, _ = bss_eval_sources(egs.numpy()[:length], est.numpy()[:length])
mix_sdr, _, _, _ = bss_eval_sources(egs.numpy()[:length], mix.numpy()[:length])
return float(sdr-mix_sdr)
def separation_metrics(pred_left, pred_right, gt_left, gt_right, mix):
if audio_empty(gt_left) or audio_empty(gt_right) or audio_empty(
pred_right) or audio_empty(pred_left) or audio_empty(mix):
print("----------- Empty -----------")
return None
sdr, sir, sar, _ = bss_eval_sources(np.asarray([gt_left, gt_right]),
np.asarray([pred_left, pred_right]),
False)
sdr_mix, _, _, _ = bss_eval_sources(np.asarray([gt_left, gt_right]),
np.asarray([mix, mix]), False)
return sdr.mean(), sir.mean(), sar.mean(), sdr_mix.mean()
def cal_SDR_improve(clean, direct, enhance):
"""
calculate a SDR1: direct to clean
calculate a SDR2: enhance to clean
return :SDR2 - SDR1 (improvement)
"""
import sys
sys.path.append('/home/cjf/workspace/201903_dereverLocEnhance/mir_eval_master/')
from mir_eval import separation as sep
SDR1, SIR1, SAR1, perm1 = sep.bss_eval_sources(clean, enhance, False)
SDR2, SIR2, SAR2, perm2 = sep.bss_eval_sources(clean, direct, False)
return SDR1 - SDR2
def bss_eval(mixed_wav, src1_wav, src2_wav, pred_src1_wav, pred_src2_wav):
len = pred_src1_wav.shape[0]
src1_wav = src1_wav[:len]
src2_wav = src2_wav[:len]
mixed_wav = mixed_wav[:len]
sdr, sir, sar, _ = bss_eval_sources(np.array([src1_wav, src2_wav]),
np.array(
[pred_src1_wav, pred_src2_wav]),
compute_permutation=True)
sdr_mixed, _, _, _ = bss_eval_sources(np.array([src1_wav, src2_wav]),
np.array([mixed_wav, mixed_wav]),
compute_permutation=True)
nsdr = sdr - sdr_mixed
return nsdr, sir, sar, len
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
return avg_sdri
def sdr_sir_sar(gt_audio, sep_audio, sed_y, inside_only):
"""gt_audio, sep_audio.shape: (n_channels, n_samples)
"""
if inside_only:
n_step = cfg.n_step
active_locts = np.where(sed_y==1)[0]
onset = int(round(active_locts[0] * n_step))
offset = int(round((active_locts[-1] + 1) * n_step))
in_gt_audio = gt_audio[:, onset : offset]
in_sep_audio = sep_audio[:, onset : offset]
(sdr, sir, sar, perm) = bss_eval_sources(in_gt_audio, in_sep_audio, compute_permutation=False)
return sdr, sir, sar
else:
(sdr, sir, sar, perm) = bss_eval_sources(gt_audio, sep_audio, compute_permutation=False)
return sdr, sir, sar
def validate(audio, model, embedder, testloader, writer, epoch):
model.eval()
criterion = nn.MSELoss()
with torch.no_grad():
for i, batch in enumerate(testloader):
dvec_mel, target_wav, mixed_wav, target_mag, mixed_mag, mixed_phase = batch[
0]
dvec_mel = dvec_mel.cuda()
target_mag = target_mag.unsqueeze(0).cuda()
mixed_mag = mixed_mag.unsqueeze(0).cuda()
dvec = embedder(dvec_mel)
dvec = dvec.unsqueeze(0)
est_mask = model(mixed_mag, dvec)
est_mag = est_mask * mixed_mag
test_loss = criterion(est_mag, target_mag).item()
mixed_mag = mixed_mag[0].cpu().detach().numpy()
target_mag = target_mag[0].cpu().detach().numpy()
est_mag = est_mag[0].cpu().detach().numpy()
est_wav = audio.spec2wav(est_mag, mixed_phase)
est_mask = est_mask[0].cpu().detach().numpy()
test_sdr_avg = bss_eval_sources(target_wav, est_wav, False)[0][0]
test_loss_avg = test_loss
writer.log_evaluation_data(mixed_wav, target_wav, est_wav,
mixed_mag.T, target_mag.T, est_mag.T,
est_mask.T, (epoch - 1), i)
writer.log_evaluation_avg(test_loss_avg, test_sdr_avg, (epoch - 1))
break
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))
def validate(audio, model, embedder, testloader, writer, step):
model.eval()
criterion = nn.MSELoss()
with torch.no_grad():
for batch in testloader:
dvec_mel, target_wav, mixed_wav, target_mag, mixed_mag, mixed_phase = batch[
0]
dvec_mel = dvec_mel.cuda()
target_mag = target_mag.unsqueeze(0).cuda()
mixed_mag = mixed_mag.unsqueeze(0).cuda()
dvec = embedder(dvec_mel)
dvec = dvec.unsqueeze(0)
est_mask = model(mixed_mag, dvec)
est_mag = est_mask * mixed_mag
test_loss = criterion(target_mag, est_mag).item()
mixed_mag = mixed_mag[0].cpu().detach().numpy()
target_mag = target_mag[0].cpu().detach().numpy()
est_mag = est_mag[0].cpu().detach().numpy()
est_wav = audio.spec2wav(est_mag, mixed_phase)
est_mask = est_mask[0].cpu().detach().numpy()
sdr = bss_eval_sources(target_wav, est_wav, False)[0][0]
writer.log_evaluation(test_loss, sdr, mixed_wav, target_wav,
est_wav, mixed_mag.T, target_mag.T,
est_mag.T, est_mask.T, step)
break
model.train()
def test_preprocessed_data(net_type):
from pystoi import stoi
import pesq
from mir_eval.separation import bss_eval_sources
path = 'preprocessed_test_data_' + net_type + '/'
if os.path.isdir(path):
files = [f for f in os.listdir(path) if f.endswith('.npy')]
sdr_a = []
pesq_a = []
stoi_a = []
processed = 0
for i, f in enumerate(files):
signals = np.load(path + f)
clean_speech = signals[:,0]
recovered_speech = signals[:,1]
if np.any(clean_speech) and np.any(recovered_speech):
PESQ = pesq.pesq(dsp.audio_fs, clean_speech, recovered_speech, 'wb')
STOI = stoi(clean_speech, recovered_speech, dsp.audio_fs, extended=False)
SDR, sir, sar, perm = bss_eval_sources(clean_speech, recovered_speech)
sdr_a.append(SDR[0])
pesq_a.append(PESQ)
stoi_a.append(STOI)
processed += 1
if i < len(files)-1:
print('[Metric computation: {}% complete]'.format(100.0*(i+1)/len(files)), end='\r')
else:
print('[Metric computation: {}% complete]'.format(100.0*(i+1)/len(files)), end='\n')
metrics = np.array([sdr_a, pesq_a, stoi_a]).T
np.save(net_type + '_metrics.npy', metrics)
print("Finished pre-processed testing of net '{}', {} files out of {} were processed into {}_metrics.npy".format(net_type, processed, len(files), net_type))
else:
print("Error: Preprocessed data for the model not found")
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()
def sdr_batch_eval(target_sources,
noisy_sources,
estimated_sources,
sample_rate=16e3,
step_size=10,
sequence_lengths=None):
sdr_list = []
sir_list = []
sar_list = []
n_samples_frame = int(step_size / 1e3 * sample_rate)
for i, (target, noisy, estimated) in enumerate(
zip(target_sources, noisy_sources, estimated_sources)):
if sequence_lengths is not None:
target = target[:sequence_lengths[i] * n_samples_frame]
noisy = noisy[:len(target)]
estimated = estimated[:len(target)]
# Skip evaluation if estimated sources is all-zero vector
if np.any(estimated):
ref_sources = np.vstack([target, noisy])
est_sources = np.vstack([estimated, np.ones_like(estimated)])
sdr, sir, sar, _ = bss_eval_sources(ref_sources,
est_sources,
compute_permutation=False)
sdr_list.append(sdr[0])
sir_list.append(sir[0])
sar_list.append(sar[0])
return np.array(sdr_list), np.array(sir_list), np.array(sar_list)
def audio_to_bsseval(s1hats, s2hats, s1s, s2s):
bss_evals = []
bss_evals_paris = []
for i, (s1hat, s2hat, s1, s2) in enumerate(zip(s1hats, s2hats, s1s, s2s)):
print('Computing bssevals for mixture {}'.format(i))
sourcehat_mat = np.concatenate(
[s1hat.reshape(1, -1), s2hat.reshape(1, -1)], 0)
source_mat = np.concatenate([s1.reshape(1, -1), s2.reshape(1, -1)], 0)
Nhat, N = sourcehat_mat.shape[1], source_mat.shape[1]
Nmin = min([N, Nhat])
bss_evals.append(
mevalsep.bss_eval_sources(source_mat[:, :Nmin],
sourcehat_mat[:, :Nmin]))
bss_evals_paris.append([
tu.bss_eval(sourcehat_mat[0, :Nmin], 0, source_mat[:, :Nmin]),
tu.bss_eval(sourcehat_mat[1, :Nmin], 1, source_mat[:, :Nmin])
])
print(bss_evals)
print(bss_evals_paris)
return bss_evals
def sdr_batch_eval_ss(target_source,
estimated_source,
sample_rate=16e3,
step_size=10,
sequence_lengths=None):
"""
Single source version of SDR, SIR and SDR computation
"""
sdr_list = []
sir_list = []
sar_list = []
n_samples_frame = int(step_size / 1e3 * sample_rate)
for i, (target,
estimated) in enumerate(zip(target_source, estimated_source)):
if sequence_lengths is not None:
target = target[:sequence_lengths[i] * n_samples_frame]
estimated = estimated[:len(target)]
# Skip evaluation if estimated sources is all-zero vector
if np.any(estimated):
sdr, sir, sar, _ = bss_eval_sources(np.array([target]),
np.array([estimated]),
compute_permutation=False)
sdr_list.append(sdr[0])
sir_list.append(sir[0])
sar_list.append(sar[0])
return np.array(sdr_list), np.array(sir_list), np.array(sar_list)
def convergence_callback(Y, X, n_targets, SDR, SIR, eval_time, ref,
framesize, win_s, algo_name):
t_in = time.perf_counter()
# projection back
z = projection_back(Y, X[:, :, 0])
Y = Y * np.conj(z[None, :, :])
from mir_eval.separation import bss_eval_sources
if Y.shape[2] == 1:
y = pra.transform.synthesis(Y[:, :, 0], framesize, hop,
win=win_s)[:, None]
else:
y = pra.transform.synthesis(Y, framesize, hop, win=win_s)
if algo_name not in parameters["overdet_algos"]:
new_ord = np.argsort(np.std(y, axis=0))[::-1]
y = y[:, new_ord]
m = np.minimum(y.shape[0] - hop, ref.shape[1])
synth[:n_targets, :m, 0] = y[hop:m + hop, :n_targets].T
synth[n_targets, :m, 0] = y[hop:m + hop, 0]
sdr, sir, sar, perm = bss_eval_sources(ref[:n_targets + 1, :m, 0],
synth[:, :m, 0])
SDR.append(sdr[:n_targets].tolist())
SIR.append(sir[:n_targets].tolist())
t_out = time.perf_counter()
eval_time.append(t_out - t_in)
def convergence_callback(Y, **kwargs):
global SDR, SIR, ref
t_enter = time.perf_counter()
from mir_eval.separation import bss_eval_sources
# projection back
z = projection_back(Y, X_mics[:, :, 0])
Y = Y.copy() * np.conj(z[None, :, :])
if Y.shape[2] == 1:
y = pra.transform.synthesis(Y[:, :, 0], framesize, hop, win=win_s)[:, None]
else:
y = pra.transform.synthesis(Y, framesize, hop, win=win_s)
y = y[framesize - hop :, :].astype(np.float64)
if args.algo != "blinkiva":
new_ord = np.argsort(np.std(y, axis=0))[::-1]
y = y[:, new_ord]
m = np.minimum(y.shape[0], ref.shape[1])
sdr, sir, sar, perm = bss_eval_sources(ref[:, :m], y[:m, [0, 0]].T)
SDR.append(sdr)
SIR.append(sir)
t_exit = time.perf_counter()
eval_time.append(t_exit - t_enter)
def add_eval_summary(summary_writer, step, before_loss, after_loss,
linear_loss, loss, sample_rate, mixed_wav, target_wav,
predicted_wav, mixed_linear_img, target_linear_img,
predicted_linear_img):
sdr = bss_eval_sources(target_wav, predicted_wav, False)[0][0]
summary_writer.add_scalar('eval_before_loss', before_loss, step)
summary_writer.add_scalar('eval_after_loss', after_loss, step)
summary_writer.add_scalar('eval_linear_loss', linear_loss, step)
summary_writer.add_scalar('eval_loss', loss, step)
summary_writer.add_scalar('SDR', sdr, step)
summary_writer.add_audio('mixed_wav', mixed_wav, step, sample_rate)
summary_writer.add_audio('target_wav', target_wav, step, sample_rate)
summary_writer.add_audio('predicted_wav', predicted_wav, step, sample_rate)
summary_writer.add_image('mixed_spectrogram',
mixed_linear_img,
step,
dataformats='HWC')
summary_writer.add_image('target_spectrogram',
target_linear_img,
step,
dataformats='HWC')
summary_writer.add_image('predicted_spectrogram',
predicted_linear_img,
step,
dataformats='HWC')
summary_writer.flush()
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)
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.synthesis(Y, L, L, zp_back=L//2, zp_front=L//2).T
sdr, sir, sar, perm = bss_eval_sources(ref[:,:y.shape[1]-L//2,0], y[:,L//2:ref.shape[1]+L//2])
SDR.append(sdr)
SIR.append(sir)
def bss_eval_sdr(src1_wav, pred_src1_wav):
len_cropped = pred_src1_wav.shape[0]
src1_wav = src1_wav[:len_cropped]
sdr, _, _, _ = bss_eval_sources(src1_wav,
pred_src1_wav,
compute_permutation=True)
return sdr
def convergence_callback(Y):
global SDR, SIR
from mir_eval.separation import bss_eval_sources
ref = np.moveaxis(separate_recordings, 1, 2)
y = np.array([pra.istft(Y[:,:,ch], L, L, transform=np.fft.irfft, zp_front=L//2, zp_back=L//2) for ch in range(Y.shape[2])])
sdr, sir, sar, perm = bss_eval_sources(ref[:,:y.shape[1]-L//2,0], y[:,L//2:ref.shape[1]+L//2])
SDR.append(sdr)
SIR.append(sir)
def bss_eval(mixed_wav, src1_wav, src2_wav, pred_src1_wav, pred_src2_wav):
import numpy as np
from mir_eval.separation import bss_eval_sources
n = pred_src1_wav.shape[0]
src1_wav = src1_wav[:n]
src2_wav = src2_wav[:n]
mixed_wav = mixed_wav[:n]
sdr, sir, sar, _ = bss_eval_sources(np.array([src1_wav, src2_wav]),
np.array(
[pred_src1_wav, pred_src2_wav]),
compute_permutation=True)
sdr_mixed, _, _, _ = bss_eval_sources(np.array([src1_wav, src2_wav]),
np.array([mixed_wav, mixed_wav]),
compute_permutation=True)
# sdr, sir, sar, _ = bss_eval_sources(src2_wav,pred_src2_wav, False)
# sdr_mixed, _, _, _ = bss_eval_sources(src2_wav,mixed_wav, False)
nsdr = sdr - sdr_mixed
return nsdr, sir, sar, n
def evaluate(reference, estimated):
estimated = np.vstack([x.audio_data[0, :] for x in estimated])
reference = np.vstack([x.audio_data[0, :] for x in reference])
print estimated.shape[-1], reference.shape[-1]
print estimated.shape[-1] - reference.shape[-1]
L = min(estimated.shape[-1], reference.shape[-1])
estimated = estimated[:, 0:L]
reference = reference[:, 0:L]
return bss_eval_sources(reference, estimated, compute_permutation=False)
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)
src_anchor = mix #!!NOTE THAT THIS IS TARGET SPEAKER SEPARATION!!
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
avg_SDRi = (sdr - sdr0) #!!NOTE THAT THIS IS TARGET SPEAKER SEPARATION!!
# print("SDRi1: {0:.2f}, SDRi2: {1:.2f}".format(sdr[0]-sdr0[0], sdr[1]-sdr0[1]))
return avg_SDRi
def calc_sdr(w_real, w_pred):
n_real, = w_real.shape
n_pred, = w_pred.shape
n = min(n_real, n_pred)
w_real, w_pred = w_real[:n], w_pred[:n]
sdr, _, _, _ = bss_eval_sources(w_real, w_pred, compute_permutation=True)
return sdr
def bss_eval_sdr(src_list, pred_src_list):
from mir_eval.separation import bss_eval_sources
len_cropped = pred_src_list.shape[-1]
src_list = src_list[:, :len_cropped]
sdr, sir, sar, _ = bss_eval_sources(src_list,
pred_src_list,
compute_permutation=True)
return sdr, sir, sar
