def main(argv):
os.makedirs(FLAGS.output_dir, exist_ok=True)
''' Initialize model '''
unet = Unet()
restore(net=unet, ckpt_path=FLAGS.ckpt_path)
''' Load data '''
mix_wav, _ = load(FLAGS.original_wav, sr=SAMPLE_RATE)
mix_wav_mag, mix_wav_phase = magphase(stft(mix_wav, n_fft=WINDOW_SIZE, hop_length=HOP_LENGTH))
mix_wav_mag= mix_wav_mag[:, START:END]
mix_wav_phase= mix_wav_phase[:, START:END]
'''Load gt '''
if FLAGS.gt == True:
gt_wav, _ = load(FLAGS.original_gt, sr=SAMPLE_RATE)
gt_wav_mag, gt_wav_phase = magphase(stft(gt_wav, n_fft=WINDOW_SIZE, hop_length=HOP_LENGTH))
gt_wav_mag= gt_wav_mag[:, START:END]
gt_wav_phase= gt_wav_phase[:, START:END]
'''Save input spectrogram image and gt'''
write_wav(FLAGS.output_dir+'original_mix.wav',
istft(mix_wav_mag * mix_wav_phase,win_length=WINDOW_SIZE,hop_length=HOP_LENGTH),
SAMPLE_RATE, norm=True)
spectogram_librosa(FLAGS.output_dir+'original_mix.wav',0)
if FLAGS.gt == True:
write_wav(FLAGS.output_dir+'gt.wav',
istft(gt_wav_mag * gt_wav_phase,win_length=WINDOW_SIZE,hop_length=HOP_LENGTH),
SAMPLE_RATE, norm=True)
spectogram_librosa(FLAGS.output_dir+'gt.wav',0)
''' run data '''
inputs = mix_wav_mag[1:].reshape(1, 512, 128, 1)
mask = unet(inputs).numpy().reshape(512, 128)
predict = inputs.reshape(512, 128)*mask
''' evaluation metrics '''
if FLAGS.gt == True:
expand_pre = np.expand_dims(predict.flatten(), axis=0)
expand_gt = np.expand_dims(gt_wav_mag[1:].flatten(), axis=0)
expand_input = np.expand_dims(inputs.flatten(), axis=0)
(SDR, SIR, SAR, _) = mir_eval.separation.bss_eval_sources(expand_gt,expand_pre)
(SDR2, _, _, _) = mir_eval.separation.bss_eval_sources(expand_gt,expand_input)
NSDR = SDR - SDR2 #SDR(Se, Sr) − SDR(Sm, Sr)
fout = open(FLAGS.output_dir+'metrics.txt','a')
print('*****SDR = '+ str(SDR) + ', SIR = '+ str(SIR) + ', SAR = '+ str(SAR) + ', NSDR = '+ str(NSDR) + '*****')
fout.write('*****SDR = '+ str(SDR) + ', SIR = '+ str(SIR) + ', SAR = '+ str(SAR) + ', NSDR = '+ str(NSDR) + '*****')
fout.close()
''' Convert model output to target magnitude '''
target_pred_mag = np.vstack((np.zeros((128)), predict))
''' Write vocal prediction audio files '''
write_wav(FLAGS.output_dir+'pred_vocal.wav',
istft(target_pred_mag * mix_wav_phase,win_length=WINDOW_SIZE,hop_length=HOP_LENGTH),
SAMPLE_RATE, norm=True)
spectogram_librosa(FLAGS.output_dir+'pred_vocal.wav',1)
def __call__(self, audio):
spec = stft(audio.numpy().reshape(-1, ),
hop_length=self.hop,
win_length=self.ws,
n_fft=self.n_fft)
mag, ph = magphase(spec)
mag = torch.Tensor(mag)
ph = np.angle(ph)
ph = torch.Tensor(ph)
out = torch.stack((mag, ph), dim=0)
return out
def test():
vis = Visualizer(env='svs')
model = getattr(models, 'Unet')().eval()
# model.cuda()
model.load_state_dict(
t.load('G:/Unet_svs/check/epoch_219__0724_16_57_35.pth'))
mix_wav, _ = load("C:/Users/lenovo/Music/c.mp3", sr=8192)
mix_wav_mag, mix_wav_phase = magphase(
stft(mix_wav, n_fft=1024, hop_length=768))
START = 700
END = START + 128
mix_wav_mag = mix_wav_mag[:, START:END]
mix_wav_phase = mix_wav_phase[:, START:END]
print(mix_wav_mag.shape)
gg = mix_wav_mag[1:]
gg = t.from_numpy(gg)
gg.unsqueeze_(0)
gg.unsqueeze_(0)
vis.img('a', gg)
print(gg.shape)
with t.no_grad():
gg = Variable(gg)
score = model(gg)
predict = gg.data * score.data
print(predict.shape)
target_pred_mag = predict.view(512, 128).cpu().numpy()
target_pred_mag = np.vstack((np.zeros((128)), target_pred_mag))
vis.img('b', t.from_numpy(target_pred_mag))
print(target_pred_mag.shape)
write_wav(
f'C:/Users/lenovo/Music/pred_vocal.wav',
istft(
target_pred_mag * mix_wav_phase
# (mix_wav_mag * target_pred_mag) * mix_wav_phase
,
win_length=1024,
hop_length=768),
8192,
norm=True)
write_wav(f'C:/Users/lenovo/Music/pred_mix.wav',
istft(mix_wav_mag * mix_wav_phase,
win_length=1024,
hop_length=768),
8192,
norm=True)
def gl_rec(S):
sr, nfft, wlen, hop = 22050, 1022, 1022, 256
S = 10**(S)
angles = 3.1415 * (np.random.randn(S.shape[0], S.shape[1]) - 0.5)
#print (angles.shape)
#print (S.shape)
y = core.istft(S * angles, hop_length=hop, win_length=wlen)
num_samples = y.shape[0]
#print (y.shape)
for i in range(40):
angles = core.stft(y, n_fft=nfft, hop_length=hop, win_length=wlen)
S = S[:, :angles.shape[1]]
_, angles = core.magphase(angles)
y = core.istft(S * angles, hop_length=hop, win_length=wlen)
#y = y[:num_samples]
return y
def LoadAudioTrainingDataFromFile(csv_file_name,
validation_size,
nmfcc=None,
nfft=None,
output_type=None):
#initiliaze arrays
mfcc_input = []
mag_input = []
digit = []
output = []
#read in CSV file with file names
#????? Should probably require output as a seperate column instead of reading from file name
with open(csv_file_name, 'r') as f:
reader = csv.reader(f)
file_list = list(reader)
#load in audio and file name data
dirname = os.path.dirname(os.path.abspath(inspect.stack()[0][1]))
for files in file_list:
relative_path = 'recordings/' + files[0]
file_name = os.path.join(os.path.dirname(__file__), relative_path)
y, sr = load(file_name, sr=None)
filesize = sys.getsizeof(y)
if output_type == 'spectrum':
spectrum = stft(y, nfft, hop_length=int(filesize / 2))
mag, phase = magphase(spectrum)
mag_input.append(mag)
mfcc = feature.mfcc(y, sr, n_mfcc=nmfcc, hop_length=int(filesize / 2))
mfcc = mfcc[1:nmfcc]
mfcc_input.append(mfcc)
digit.append(files[0][0])
#build array of one hot vectors for output based on 1st character in file name
for num in digit:
if num == '0':
output.append(digits.zero)
if num == '1':
output.append(digits.one)
if num == '2':
output.append(digits.two)
if num == '3':
output.append(digits.three)
if num == '4':
output.append(digits.four)
if num == '5':
output.append(digits.five)
if num == '6':
output.append(digits.six)
if num == '7':
output.append(digits.seven)
if num == '8':
output.append(digits.eight)
if num == '9':
output.append(digits.nine)
#normalize data to between 0 and 1
if output_type == 'mfcc':
training_input = numpy.asarray(mfcc_input, dtype=numpy.float64)
training_input = numpy.squeeze(training_input)
min_max_scaler = preprocessing.MinMaxScaler(feature_range=(-1, 1))
training_input = min_max_scaler.fit_transform(training_input)
elif output_type == 'spectrum':
training_input = numpy.asarray(mag_input, dtype=numpy.float64)
training_input = numpy.squeeze(training_input)
min_max_scaler = preprocessing.MinMaxScaler(feature_range=(0, 1))
training_input = min_max_scaler.fit_transform(training_input)
training_output = numpy.asarray(output, dtype=numpy.float64)
min_max_scaler = preprocessing.MinMaxScaler(feature_range=(0, 6))
training_output = min_max_scaler.fit_transform(training_output)
#randommize before dividing test/train sets:
randomize = numpy.arange(len(training_input))
numpy.random.shuffle(randomize)
training_input = training_input[randomize]
training_output = training_output[randomize]
#pull out validation set
validation_input = training_input[0:validation_size, :]
validation_output = training_output[0:validation_size, :]
return training_input, training_output, validation_input, validation_output
def load_as_mag(file):
wav, _ = load(file, sr=None)
spectrogram = stft(wav, n_fft=WINDOW_SIZE, hop_length=HOP_LENGTH)
mag, _ = magphase(spectrogram)
return mag.astype(np.float32)
def predict(self, loader):
"""
Predict for an input.
Args
----
loader : PyTorch DataLoader.
"""
self.model.eval()
all_preds = []
all_ys = []
all_cs = []
all_ts = []
all_ms = []
all_idx = []
if isinstance(loader.dataset, torch.utils.data.Subset):
n_frames = loader.dataset.dataset.n_frames
elif isinstance(loader.dataset, torch.utils.data.ConcatDataset):
n_frames = loader.dataset.datasets[0].n_frames
else:
n_frames = loader.dataset.n_frames
with torch.no_grad():
for batch_samples in tqdm(loader):
# prepare training sample
X = batch_samples['X']
if X.dim() == 4:
full_track = False
# batch_size x in_channels x 1025 x 129
else:
bs = X.size(0)
ns = X.size(1)
full_track = True
# batch_size * splits x in_channels x 1025 x 129
X = X.view(bs * ns, self.in_channels, self.n_fft, n_frames)
# batch_size x in_channels x 1025 x 129 x 2
X_complex = batch_samples['X_complex']
if X_complex.dim() != 5:
# batch_size * splits x in_channels x 1025 x 129 x 2
X_complex = X_complex.view(
bs * ns, self.out_channels, self.n_fft, n_frames, 2)
# batch_size x nclasses x in_channels x 1025 x time samples x 2
y = batch_samples['y_complex']
# batch_size x nclasses
cs = batch_samples['c']
# batch_size x 1
ts = batch_samples['t']
track_idx = batch_samples['track_idx']
if self.USE_CUDA:
X = X.cuda()
X_complex = X_complex.cuda()
y = y.cuda()
if X.size(0) > 4:
X_list = torch.split(X, 4, dim=0)
else:
X_list = [X]
masks_list = []
pred_list = []
for X in X_list:
# detach hidden state
self.model.detach_hidden(X.size(0))
# forward pass
preds, mask = self.model(X)
masks_list += [mask]
pred_list += [preds]
mask = torch.cat(masks_list, dim=0)
preds = torch.cat(pred_list, dim=0)
if full_track:
# batch size x nclasses x in_channels x 1025 x time samples
if self.regression:
preds = preds.view(
bs, ns, self.n_classes, self.out_channels,
self.n_fft, n_frames)
preds = torch.unbind(preds, dim=1)
preds = torch.cat(preds, dim=4)
else:
mask = mask.view(
bs, ns, self.n_classes, self.out_channels,
self.n_fft, n_frames)
mask = torch.unbind(mask, dim=1)
mask = torch.cat(mask, dim=4)
# batch_size x in_channels x 1025 x time samples x 2
X_complex = X_complex.view(
bs, ns, self.out_channels, self.n_fft, n_frames, 2)
X_complex = torch.unbind(X_complex, dim=1)
X_complex = torch.cat(X_complex, dim=3)
# convert to complex
# batch size x nclasses x in_channels x 1025 x time samples x 2
X_complex = X_complex.unsqueeze(1).repeat(
1, self.n_classes, 1, 1, 1, 1)
X_complex = self._to_complex(X_complex)
if self.regression:
_, X_phase = magphase(X_complex)
preds = preds.cpu().numpy() * X_phase
else:
preds = mask.cpu().numpy() * X_complex
# batch size x nclasses x in_channels x 1025 x time samples
ys = self._to_complex(y)
all_preds += [preds]
all_ys += [ys]
all_cs += [cs]
all_ts += [ts]
all_ms += [mask.cpu().numpy()]
all_idx += [track_idx]
return all_preds, all_ys, all_cs, all_ts, all_ms, all_idx
import numpy as np
from librosa.core import istft, load, stft, magphase
from librosa.output import write_wav
from config import *
import keras as keras
if __name__ == '__main__':
# load test audio and convert to mag/phase
mix_wav, _ = load("../wav_files/mixture.wav", sr=SAMPLE_RATE)
mix_wav_mag, mix_wav_phase = magphase(
stft(mix_wav, n_fft=WINDOW_SIZE, hop_length=HOP_LENGTH))
vocal_wav, _ = load("../wav_files/vocals.wav", sr=SAMPLE_RATE)
vocal_wav_mag, vocal_wav_phase = magphase(
stft(vocal_wav, n_fft=WINDOW_SIZE, hop_length=HOP_LENGTH))
START = 0
END = START + 128
mix_wav_mag = mix_wav_mag[:, START:END]
mix_wav_phase = mix_wav_phase[:, START:END]
vocal_wav_mag = vocal_wav_mag[:, START:END]
vocal_wav_phase = vocal_wav_phase[:, START:END]
# load saved model
model = keras.models.load_model('../models/vocal_20_test_model.h5')
#model = keras.models.load_model('../models/vocal_20.h5')
# predict and write into file
X = mix_wav_mag[1:].reshape(1, 512, 128, 1)
def mag_phase_angle(x):
mag, ph = magphase(x)
ph = np.angle(ph)
out = np.stack([mag, ph])
return out
def separate_instruments(file_path):
plt.rcParams['figure.figsize'] = (14, 5)
x, sr = librosa.load(file_path, sr=None)
winlen = 1024
h, i, X = stft(x=x,
fs=sr,
window='hann',
nperseg=winlen,
noverlap=int(winlen / 2),
nfft=winlen,
detrend=False,
return_onesided=True,
padded=True,
axis=-1)
# information about wav
print(len(x))
# short-time fourier transform
# X = librosa.stft(x)
# log-amplitude
Xmag = librosa.amplitude_to_db(X)
# show harm-perc spectrogram
librosa.display.specshow(Xmag, sr=sr, x_axis='time', y_axis='log')
plt.colorbar()
plt.show()
#############
S = X
kernel_size = 31
power = 2.0
mask = False
margin = 1.0
if np.iscomplexobj(S):
S, phase = core.magphase(S)
else:
phase = 1
if np.isscalar(kernel_size):
win_harm = kernel_size
win_perc = kernel_size
else:
win_harm = kernel_size[0]
win_perc = kernel_size[1]
if np.isscalar(margin):
margin_harm = margin
margin_perc = margin
else:
margin_harm = margin[0]
margin_perc = margin[1]
split_zeros = (margin_harm == 1 and margin_perc == 1)
# Compute median filters. Pre-allocation here preserves memory layout.
harm = np.empty_like(S)
harm[:] = median_filter(S, size=(1, win_harm), mode='reflect')
perc = np.empty_like(S)
perc[:] = median_filter(S, size=(win_perc, 1), mode='reflect')
Hmag = librosa.amplitude_to_db(harm)
# librosa.display.specshow(harm, sr=sr, x_axis='time', y_axis='log')
# plt.colorbar()
# plt.show()
Pmag = librosa.amplitude_to_db(perc)
# librosa.display.specshow(perc, sr=sr, x_axis='time', y_axis='log')
# plt.colorbar()
# plt.show()
mask_harm_soft = util.softmask(harm,
perc * margin_harm,
power=power,
split_zeros=split_zeros)
mask_perc_soft = util.softmask(perc,
harm * margin_perc,
power=power,
split_zeros=split_zeros)
soft_mask_X_harm = (S * mask_harm_soft) * phase
Xmag_harm_soft = librosa.amplitude_to_db(soft_mask_X_harm)
soft_mask_X_perc = (S * mask_perc_soft) * phase
Xmag_perc_soft = librosa.amplitude_to_db(soft_mask_X_perc)
# mask_harm_hard = harm > perc * margin_harm
# mask_perc_hard = perc > harm * margin_perc
# hard_mask_X_harm = (S * mask_harm_hard) * phase
# Xmag_harm_hard = librosa.amplitude_to_db(hard_mask_X_harm)
# hard_mask_X_perc = (S * mask_perc_hard) * phase
# Xmag_perc_hard = librosa.amplitude_to_db(hard_mask_X_perc)
librosa.display.specshow(Xmag_harm_soft,
sr=sr,
x_axis='time',
y_axis='log')
plt.colorbar()
plt.show()
librosa.display.specshow(Xmag_perc_soft,
sr=sr,
x_axis='time',
y_axis='log')
plt.colorbar()
plt.show()
# x_h, sr_h = librosa.load('my_audio_mod/01_AF_NM_h.wav', duration=6, sr=None)
# x_p, sr_p = librosa.load('my_audio_mod/01_AF_NM_p.wav', duration=6, sr=None)
# librosa.display.waveplot( x_h, sr=sr_h)
# plt.show()
# librosa.display.waveplot( x_p, sr=sr_p)
# plt.show()
H = (S * mask_harm_soft) * phase
P = (S * mask_perc_soft) * phase
Hmag = librosa.amplitude_to_db(H)
Pmag = librosa.amplitude_to_db(P)
librosa.display.specshow(Hmag, sr=sr, x_axis='time', y_axis='log')
plt.colorbar()
plt.show()
librosa.display.specshow(Pmag, sr=sr, x_axis='time', y_axis='log')
plt.colorbar()
plt.show()
# h = librosa.istft(H)
# p = librosa.istft(P)
_, h = istft(H,
fs=sr,
window='hann',
nperseg=winlen,
noverlap=int(winlen / 2),
nfft=winlen,
input_onesided=True)
_, p = istft(P,
fs=sr,
window='hann',
nperseg=winlen,
noverlap=int(winlen / 2),
nfft=winlen,
input_onesided=True)
# saving
librosa.output.write_wav(
os.path.splitext(file_path)[0] + '_H_med.wav', h, sr)
librosa.output.write_wav(
os.path.splitext(file_path)[0] + '_P_med.wav', p, sr)
def LoadAudio(path_audio):
y, sr = load(path_audio, sr=SR)
S_mag, _ = magphase(stft(y, n_fft=FFTSIZE, hop_length=H))
return S_mag
