X, y = spec_utils.align_wave_head_and_tail(X, y, args.sr)
v = X - y
sf.write(input_i, y.T, args.sr)
sf.write(input_v, v.T, args.sr)
subprocess.call(cmd_i, stderr=subprocess.DEVNULL)
subprocess.call(cmd_v, stderr=subprocess.DEVNULL)
y, _ = librosa.load(output_i,
args.sr,
False,
dtype=np.float32,
res_type='kaiser_fast')
v, _ = librosa.load(output_v,
args.sr,
False,
dtype=np.float32,
res_type='kaiser_fast')
X = y + v
spec = spec_utils.calc_spec(X, args.hop_length)
np.save(outpath_mix, np.abs(spec))
spec = spec_utils.calc_spec(y, args.hop_length)
np.save(outpath_inst, np.abs(spec))
os.remove(input_i)
os.remove(input_v)
os.remove(output_i)
os.remove(output_v)
def main():
p = argparse.ArgumentParser()
p.add_argument('--gpu', '-g', type=int, default=-1)
p.add_argument(
'--model',
'-m',
type=str,
default=
'/content/drive/My Drive/vocal-remover/models/MultiGenreModelNP.pth')
p.add_argument('--input', '-i', required=True)
p.add_argument('--sr', '-r', type=int, default=44100)
p.add_argument('--hop_length', '-l', type=int, default=1024)
p.add_argument('--window_size', '-w', type=int, default=512)
p.add_argument('--out_mask', '-M', action='store_true')
p.add_argument('--postprocess', '-p', action='store_true')
args = p.parse_args()
print('loading model...', end=' ')
device = torch.device('cpu')
model = nets.CascadedASPPNet()
model.load_state_dict(torch.load(args.model, map_location=device))
if torch.cuda.is_available() and args.gpu >= 0:
device = torch.device('cuda:{}'.format(args.gpu))
model.to(device)
print('done')
print('loading wave source...', end=' ')
X, sr = librosa.load(args.input,
args.sr,
False,
dtype=np.float32,
res_type='kaiser_fast')
print('done')
print('stft of wave source...', end=' ')
X = spec_utils.calc_spec(X, args.hop_length)
X, phase = np.abs(X), np.exp(1.j * np.angle(X))
coeff = X.max()
X /= coeff
print('done')
offset = model.offset
l, r, roi_size = dataset.make_padding(X.shape[2], args.window_size, offset)
X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode='constant')
X_roll = np.roll(X_pad, roi_size // 2, axis=2)
model.eval()
with torch.no_grad():
masks = []
masks_roll = []
for i in tqdm(range(int(np.ceil(X.shape[2] / roi_size)))):
start = i * roi_size
X_window = torch.from_numpy(
np.asarray([
X_pad[:, :, start:start + args.window_size],
X_roll[:, :, start:start + args.window_size]
])).to(device)
pred = model.predict(X_window)
pred = pred.detach().cpu().numpy()
masks.append(pred[0])
masks_roll.append(pred[1])
mask = np.concatenate(masks, axis=2)[:, :, :X.shape[2]]
mask_roll = np.concatenate(masks_roll, axis=2)[:, :, :X.shape[2]]
mask = (mask + np.roll(mask_roll, -roi_size // 2, axis=2)) / 2
if args.postprocess:
vocal = X * (1 - mask) * coeff
mask = spec_utils.mask_uninformative(mask, vocal)
inst = X * mask * coeff
vocal = X * (1 - mask) * coeff
basename = os.path.splitext(os.path.basename(args.input))[0]
print('inverse stft of instruments...', end=' ')
wav = spec_utils.spec_to_wav(inst, phase, args.hop_length)
print('done')
sf.write('{}_Instruments.wav'.format(basename), wav.T, sr)
print('inverse stft of vocals...', end=' ')
wav = spec_utils.spec_to_wav(vocal, phase, args.hop_length)
print('done')
sf.write('{}_Vocals.wav'.format(basename), wav.T, sr)
if args.out_mask:
norm_mask = np.uint8((1 - mask) * 255).transpose(1, 2, 0)
norm_mask = np.concatenate(
[np.max(norm_mask, axis=2, keepdims=True), norm_mask],
axis=2)[::-1]
_, bin_mask = cv2.imencode('.png', norm_mask)
with open('{}_Mask.png'.format(basename), mode='wb') as f:
bin_mask.tofile(f)
def main():
p = argparse.ArgumentParser()
p.add_argument('--gpu', '-g', type=int, default=-1)
p.add_argument('--model', '-m', type=str, default='models/baseline.pth')
p.add_argument('--input', '-i', required=True)
p.add_argument('--sr', '-r', type=int, default=44100)
p.add_argument('--hop_length', '-l', type=int, default=1024)
p.add_argument('--window_size', '-w', type=int, default=512)
p.add_argument('--out_mask', '-M', action='store_true')
p.add_argument('--postprocess', '-p', action='store_true')
args = p.parse_args()
print('loading model...', end=' ')
device = torch.device('cpu')
model = nets.CascadedASPPNet()
model.load_state_dict(torch.load(args.model, map_location=device))
if torch.cuda.is_available() and args.gpu >= 0:
device = torch.device('cuda:{}'.format(args.gpu))
model.to(device)
print('done')
print('loading wave source...', end=' ')
X, sr = librosa.load(args.input,
args.sr,
False,
dtype=np.float32,
res_type='kaiser_fast')
print('done')
print('wave source stft...', end=' ')
X = spec_utils.calc_spec(X, args.hop_length)
X, phase = np.abs(X), np.exp(1.j * np.angle(X))
coeff = X.max()
X /= coeff
print('done')
offset = model.offset
l, r, roi_size = dataset.make_padding(X.shape[2], args.window_size, offset)
X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode='constant')
masks = []
model.eval()
with torch.no_grad():
for j in tqdm(range(int(np.ceil(X.shape[2] / roi_size)))):
start = j * roi_size
X_window = X_pad[None, :, :, start:start + args.window_size]
pred = model.predict(torch.from_numpy(X_window).to(device))
pred = pred.detach().cpu().numpy()
masks.append(pred[0])
mask = np.concatenate(masks, axis=2)[:, :, :X.shape[2]]
if args.postprocess:
vocal_pred = X * (1 - mask) * coeff
mask = spec_utils.mask_uninformative(mask, vocal_pred)
inst_pred = X * mask * coeff
vocal_pred = X * (1 - mask) * coeff
if args.out_mask:
norm_mask = np.uint8((1 - mask) * 255)
canvas = np.zeros((norm_mask.shape[1], norm_mask.shape[2], 3))
canvas[:, :, 1] = norm_mask[0]
canvas[:, :, 2] = norm_mask[1]
canvas[:, :, 0] = np.max(norm_mask, axis=0)
cv2.imwrite('mask.png', canvas[::-1])
basename = os.path.splitext(os.path.basename(args.input))[0]
print('instrumental inverse stft...', end=' ')
wav = spec_utils.spec_to_wav(inst_pred, phase, args.hop_length)
print('done')
sf.write('{}_Instrumental.wav'.format(basename), wav.T, sr)
print('vocal inverse stft...', end=' ')
wav = spec_utils.spec_to_wav(vocal_pred, phase, args.hop_length)
print('done')
sf.write('{}_Vocal.wav'.format(basename), wav.T, sr)
def main():
p = argparse.ArgumentParser()
p.add_argument('--gpu', '-g', type=int, default=-1)
p.add_argument('--model', '-m', type=str, default='models/baseline.npz')
p.add_argument('--input', '-i', required=True)
p.add_argument('--sr', '-r', type=int, default=44100)
p.add_argument('--hop_length', '-l', type=int, default=1024)
p.add_argument('--window_size', '-w', type=int, default=1024)
p.add_argument('--out_mask', '-M', action='store_true')
p.add_argument('--postprocess', '-p', action='store_true')
args = p.parse_args()
print('loading model...', end=' ')
model = unet.MultiBandUNet()
chainer.serializers.load_npz(args.model, model)
if args.gpu >= 0:
chainer.backends.cuda.check_cuda_available()
chainer.backends.cuda.get_device(args.gpu).use()
model.to_gpu()
xp = model.xp
print('done')
print('loading wave source...', end=' ')
X, sr = librosa.load(args.input,
args.sr,
False,
dtype=np.float32,
res_type='kaiser_fast')
print('done')
print('wave source stft...', end=' ')
X, phase = spec_utils.calc_spec(X, args.hop_length, phase=True)
coeff = X.max()
X /= coeff
print('done')
left = model.offset
roi_size = args.window_size - left * 2
right = roi_size - (X.shape[2] % roi_size) + left
X_pad = np.pad(X, ((0, 0), (0, 0), (left, right)), mode='reflect')
masks = []
with chainer.no_backprop_mode(), chainer.using_config('train', False):
for j in tqdm(range(int(np.ceil(X.shape[2] / roi_size)))):
start = j * roi_size
X_window = X_pad[None, :, :, start:start + args.window_size]
X_tta = np.concatenate([X_window, X_window[:, ::-1, :, :]])
pred = model(xp.asarray(X_tta))
pred = backends.cuda.to_cpu(pred.data)
pred[1] = pred[1, ::-1, :, :]
masks.append(pred.mean(axis=0))
mask = np.concatenate(masks, axis=2)[:, :, :X.shape[2]]
if args.postprocess:
vocal_pred = X * (1 - mask) * coeff
mask = spec_utils.mask_uninformative(mask, vocal_pred)
inst_pred = X * mask * coeff
vocal_pred = X * (1 - mask) * coeff
if args.out_mask:
norm_mask = np.uint8(mask.mean(axis=0) * 255)[::-1]
hm = cv2.applyColorMap(norm_mask, cv2.COLORMAP_MAGMA)
cv2.imwrite('mask.png', hm)
print('instrumental inverse stft...', end=' ')
wav = spec_utils.spec_to_wav(inst_pred, phase, args.hop_length)
print('done')
librosa.output.write_wav('instrumental.wav', wav, sr)
print('vocal inverse stft...', end=' ')
wav = spec_utils.spec_to_wav(vocal_pred, phase, args.hop_length)
print('done')
librosa.output.write_wav('vocal.wav', wav, sr)
chainer.backends.cuda.check_cuda_available()
chainer.backends.cuda.get_device(args.gpu).use()
model.to_gpu()
xp = model.xp
print('done')
print('loading wave source...', end=' ')
X, sr = librosa.load(args.input,
args.sr,
False,
dtype=np.float32,
res_type='kaiser_fast')
print('done')
print('wave source stft...', end=' ')
X, phase = spec_utils.calc_spec(X, args.hop_length, phase=True)
coeff = X.max()
X /= coeff
print('done')
left = model.offset
roi_size = args.window_size - left * 2
right = roi_size - (X.shape[2] % roi_size) + left
X_pad = np.pad(X, ((0, 0), (0, 0), (left, right)), mode='reflect')
masks = []
with chainer.no_backprop_mode(), chainer.using_config('train', False):
for j in tqdm(range(int(np.ceil(X.shape[2] / roi_size)))):
start = j * roi_size
X_window = X_pad[None, :, :, start:start + args.window_size]
X_tta = np.concatenate([X_window, X_window[:, ::-1, :, :]])
def upload_file():
if request.method == 'POST':
#CLEANUP SCRIPT
upload_folder = app.config['UPLOAD_FOLDER']
now = time.time()
for filename in os.listdir(upload_folder):
if allowed_file(filename):
if os.path.getmtime(os.path.join(upload_folder,
filename)) < now - 30 * 60:
if os.path.isfile(os.path.join(upload_folder, filename)):
print(filename)
os.remove(os.path.join(upload_folder, filename))
f = request.files['file']
if f and allowed_file(f.filename):
path = upload_folder + '/' + f.filename
f.save(path)
# INFERENCE.PY
import chainer
from chainer import backends
import cv2
import librosa
import numpy as np
from tqdm import tqdm
from lib import spec_utils
from lib import unet
#p = argparse.ArgumentParser()
agpu = -1 #p.add_argument('--gpu', '-g', type=int, default=-1)
amodel = models_path #p.add_argument('--model', '-m', type=str, default='models/baseline.npz')
ainput = path #p.add_argument('--input', '-i', required=True)
asr = 44100 #p.add_argument('--sr', '-r', type=int, default=44100)
ahop_length = 1024 #p.add_argument('--hop_length', '-l', type=int, default=1024)
awindow_size = 1024 #p.add_argument('--window_size', '-w', type=int, default=1024)
aout_mask = False #p.add_argument('--out_mask', '-M', action='store_true')
#args = p.parse_args()
print('loading model...', end=' ')
model = unet.MultiBandUNet()
chainer.serializers.load_npz(amodel, model)
if agpu >= 0:
chainer.backends.cuda.check_cuda_available()
chainer.backends.cuda.get_device(agpu).use()
model.to_gpu()
xp = model.xp
print('done')
#CHANGE DURATION FOR PAID VERSION
print('loading wave source...', end=' ')
X, sr = librosa.load(ainput,
asr,
False,
duration=30.0,
dtype=np.float32,
res_type='kaiser_fast')
print('done')
print('wave source stft...', end=' ')
X, phase = spec_utils.calc_spec(X, ahop_length, phase=True)
coeff = X.max()
X /= coeff
print('done')
left = model.offset
roi_size = awindow_size - left * 2
right = roi_size + left - (X.shape[2] % left)
X_pad = np.pad(X, ((0, 0), (0, 0), (left, right)), mode='reflect')
masks = []
with chainer.no_backprop_mode(), chainer.using_config(
'train', False):
for j in tqdm(range(int(np.ceil(X.shape[2] / roi_size)))):
start = j * roi_size
X_window = X_pad[None, :, :, start:start + awindow_size]
X_tta = np.concatenate([X_window, X_window[:, ::-1, :, :]])
pred = model(xp.asarray(X_tta))
pred = backends.cuda.to_cpu(pred.data)
pred[1] = pred[1, ::-1, :, :]
masks.append(pred.mean(axis=0))
mask = np.concatenate(masks, axis=2)[:, :, :X.shape[2]]
# vocal_pred = X * (1 - mask) * coeff
# mask = spec_utils.mask_uninformative(mask, vocal_pred)
inst_pred = X * mask * coeff
vocal_pred = X * (1 - mask) * coeff
if aout_mask:
norm_mask = np.uint8(mask.mean(axis=0) * 255)[::-1]
hm = cv2.applyColorMap(norm_mask, cv2.COLORMAP_MAGMA)
cv2.imwrite('mask.png', hm)
print('instrumental inverse stft...', end=' ')
wav = spec_utils.spec_to_wav(inst_pred, phase, ahop_length)
print('done')
instrumental = f.filename.split('.')[0] + '_instrumental.wav'
librosa.output.write_wav('app/static/' + instrumental, wav, sr)
print('vocal inverse stft...', end=' ')
wav = spec_utils.spec_to_wav(vocal_pred, phase, ahop_length)
print('done')
vocal = f.filename.split('.')[0] + '_vocal.wav'
librosa.output.write_wav('app/static/' + vocal, wav, sr)
return render_template('uploaded.html',
title='Success',
original=f.filename,
instrumental=instrumental,
vocal=vocal)
