def __init__(self, kind_data: str, **kwargs: Channel):
self._PATH = hp.dict_path[f'mel_{kind_data}']
# default needs
# self._needs = dict(x=Channel.ALL, y=Channel.ALL,
# length=Channel.ALL,
# path_speech=Channel.ALL)
self._needs = dict(
path_speech=Channel.ALL,
wav=Channel.ALL,
y=Channel.ALL,
length=Channel.ALL,
)
self.set_needs(**kwargs)
self._all_files = [
f for f in self._PATH.glob('*.*') if hp.is_featurefile(f)
]
self._all_files = sorted(self._all_files)
if hp.n_data > 0:
self._all_files = self._all_files[:hp.n_data]
print(
f'{len(self._all_files)} files are prepared from {kind_data.upper()}.'
)
def __init__(self,
kind_data: str,
norm_modules: Dict[str, Normalization] = None,
**kwargs: Channel):
self._PATH = hp.dict_path[f'feature_{kind_data}']
# default needs
self._needs = dict(x=Channel.ALL,
y=Channel.LAST,
x_mag=Channel.NONE,
x_phase=Channel.NONE,
y_phase=Channel.NONE,
path_speech=Channel.ALL)
self.set_needs(**kwargs)
self._all_files = [
f for f in self._PATH.glob('*.*') if hp.is_featurefile(f)
]
self._all_files = sorted(self._all_files)
if hp.n_data > 0:
self._all_files = self._all_files[:hp.n_data]
self.norm_modules = dict()
if kind_data == 'train':
# path_normconst: path of the file that has information about mean, std, ...
path_normconst = hp.dict_path[f'normconst_{kind_data}']
if not hp.refresh_const and path_normconst.exists():
# when normconst file exists
npz_normconst = np.load(path_normconst, allow_pickle=True)
self.norm_modules['x'] = Normalization(
*npz_normconst['normconst_x'])
self.norm_modules['y'] = Normalization(
*npz_normconst['normconst_y'])
else:
print('calculate normalization consts for input')
self.norm_modules['x'] \
= Normalization.calc_const(self._all_files, key=hp.spec_data_names['x'])
print('calculate normalization consts for output')
self.norm_modules['y'] \
= Normalization.calc_const(self._all_files, key=hp.spec_data_names['y'])
np.savez(path_normconst,
normconst_x=self.norm_modules['x'].astuple(),
normconst_y=self.norm_modules['y'].astuple())
scio.savemat(
path_normconst.with_suffix('.mat'),
dict(normconst_x=self.norm_modules['x'].astuple(),
normconst_y=self.norm_modules['y'].astuple()))
print(f'normalization consts for input: {self.norm_modules["x"]}')
print(f'normalization consts for output: {self.norm_modules["y"]}')
else:
assert 'x' in norm_modules and 'y' in norm_modules
self.norm_modules = norm_modules
print(
f'{len(self._all_files)} files are prepared from {kind_data.upper()}.'
)
def search_files(self, n_file: int, random_by_utterance=False, n_loc=1):
s_all_files = [
f.name for f in os.scandir(self._PATH) if hp.is_featurefile(f)
]
s_all_files = sorted(s_all_files)
if n_file != -1:
if random_by_utterance:
utterances = np.random.randint(len(s_all_files) // n_loc,
size=n_file // n_loc)
utterances = [f'{u:4d}_' for u in utterances]
s_all_files = [
f for f in s_all_files if f.startswith(utterances)
]
else:
s_all_files = np.random.permutation(s_all_files)
s_all_files = s_all_files[:n_file]
return s_all_files