def make_validation_set(filelist, cropsize, sr, hop_length, n_fft, offset):
patch_list = []
patch_dir = 'cs{}_sr{}_hl{}_nf{}_of{}'.format(cropsize, sr, hop_length, n_fft, offset)
os.makedirs(patch_dir, exist_ok=True)
for i, (X_path, y_path) in enumerate(tqdm(filelist)):
basename = os.path.splitext(os.path.basename(X_path))[0]
X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length, n_fft)
coef = np.max([np.abs(X).max(), np.abs(y).max()])
X, y = X / coef, y / coef
l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode='constant')
y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode='constant')
len_dataset = int(np.ceil(X.shape[2] / roi_size))
for j in range(len_dataset):
outpath = os.path.join(patch_dir, '{}_p{}.npz'.format(basename, j))
start = j * roi_size
if not os.path.exists(outpath):
np.savez(
outpath,
X=X_pad[:, :, start:start + cropsize],
y=y_pad[:, :, start:start + cropsize])
patch_list.append(outpath)
return VocalRemoverValidationSet(patch_list)
def make_training_set(filelist, cropsize, patches, sr, hop_length, n_fft, offset):
len_dataset = patches * len(filelist)
X_dataset = np.zeros(
(len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
y_dataset = np.zeros(
(len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
for i, (X_path, y_path) in enumerate(tqdm(filelist)):
X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length, n_fft)
coef = np.max([np.abs(X).max(), np.abs(y).max()])
X, y = X / coef, y / coef
l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode='constant')
y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode='constant')
starts = np.random.randint(0, X_pad.shape[2] - cropsize, patches)
ends = starts + cropsize
for j in range(patches):
idx = i * patches + j
X_dataset[idx] = X_pad[:, :, starts[j]:ends[j]]
y_dataset[idx] = y_pad[:, :, starts[j]:ends[j]]
return X_dataset, y_dataset
def make_training_set(filelist, cropsize, patches, sr, hop_length, offset):
len_dataset = patches * len(filelist)
X_dataset = np.zeros((len_dataset, 2, hop_length, cropsize),
dtype=np.float32)
y_dataset = np.zeros((len_dataset, 2, hop_length, cropsize),
dtype=np.float32)
for i, (X_path, y_path) in enumerate(tqdm(filelist)):
p = np.random.uniform()
if p < 0.1:
X_path.replace(os.path.splitext(X_path)[1], '_pitch-1.wav')
y_path.replace(os.path.splitext(y_path)[1], '_pitch-1.wav')
elif p < 0.2:
X_path.replace(os.path.splitext(X_path)[1], '_pitch1.wav')
y_path.replace(os.path.splitext(y_path)[1], '_pitch1.wav')
X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length)
coeff = np.max([X.max(), y.max()])
X, y = X / coeff, y / coeff
l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode='constant')
y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode='constant')
starts = np.random.randint(0, X_pad.shape[2] - cropsize, patches)
ends = starts + cropsize
for j in range(patches):
idx = i * patches + j
X_dataset[idx] = X_pad[:, :, starts[j]:ends[j]]
y_dataset[idx] = y_pad[:, :, starts[j]:ends[j]]
if np.random.uniform() < 0.5:
# swap channel
X_dataset[idx] = X_dataset[idx, ::-1]
y_dataset[idx] = y_dataset[idx, ::-1]
return X_dataset, y_dataset
def make_validation_set(filelist, cropsize, offset, sr, hop_length):
X_dataset = []
y_dataset = []
for i, (X_path, y_path) in enumerate(tqdm(filelist)):
X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length)
left = offset
roi_size = cropsize - offset * 2
right = roi_size - (X.shape[2] % roi_size) + offset
X_pad = np.pad(X, ((0, 0), (0, 0), (left, right)), mode='reflect')
y_pad = np.pad(y, ((0, 0), (0, 0), (left, right)), mode='reflect')
len_dataset = int(np.ceil(X.shape[2] / roi_size))
for j in range(len_dataset):
start = j * roi_size
X_dataset.append(X_pad[:, :, start:start + cropsize])
y_dataset.append(y_pad[:, :, start:start + cropsize])
return np.float32(X_dataset), np.float32(y_dataset)
def make_training_set(filelist, cropsize, patches, sr, hop_length):
len_dataset = patches * len(filelist)
X_dataset = np.zeros((len_dataset, 2, hop_length, cropsize),
dtype=np.float32)
y_dataset = np.zeros((len_dataset, 2, hop_length, cropsize),
dtype=np.float32)
for i, (X_path, y_path) in enumerate(tqdm(filelist)):
X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length)
for j in range(patches):
idx = i * patches + j
start = np.random.randint(0, X.shape[2] - cropsize)
X_dataset[idx] = X[:, :, start:start + cropsize]
y_dataset[idx] = y[:, :, start:start + cropsize]
if np.random.uniform() < 0.5:
# swap channel
X_dataset[idx] = X_dataset[idx, ::-1]
y_dataset[idx] = y_dataset[idx, ::-1]
return X_dataset, y_dataset
def create_dataset(filelist,
cropsize,
patches,
sr,
hop_length,
validation=False):
len_dataset = patches * len(filelist)
X_dataset = np.zeros((len_dataset, 2, hop_length, cropsize),
dtype=np.float32)
y_dataset = np.zeros((len_dataset, 2, hop_length, cropsize),
dtype=np.float32)
for i, (X_path, y_path) in enumerate(tqdm(filelist)):
X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length)
for j in range(patches):
idx = i * patches + j
start = np.random.randint(0, X.shape[2] - cropsize)
X_dataset[idx] = X[:, :, start:start + cropsize]
y_dataset[idx] = y[:, :, start:start + cropsize]
if not validation:
if np.random.uniform() < 0.5:
# swap channel
X_dataset[idx] = X_dataset[idx, ::-1]
y_dataset[idx] = y_dataset[idx, ::-1]
# if np.random.uniform() < 0.5:
# # spec aug
# f = np.random.randint(0, hop_length // 4)
# f0 = np.random.randint(0, hop_length - f)
# X_dataset[idx, :, f0:f0 + f, :] = 0
# y_dataset[idx, :, f0:f0 + f, :] = 0
# if np.random.uniform() < 0.5:
# # spec aug
# t = np.random.randint(0, hop_length // 8)
# t0 = np.random.randint(0, cropsize - t)
# X_dataset[idx, :, :, t0:t0 + t] = 0
# y_dataset[idx, :, :, t0:t0 + t] = 0
return X_dataset, y_dataset