def prepare_raw_data(self, index):
# Load label
d = self.data[index]
x, y = zip(*d)
y = target_padding(y, max([len(v) for v in y]))
# Load acoustic feature and pad
x = [torchaudio.load(os.path.join(self.raw_root,f)).squeeze(0) for f in x]
return x,y
def prepare_raw_data(self, index):
# Load label
d = self.data[index]
x, y = zip(*d)
y = target_padding(y, max([len(v) for v in y]))
if self.text_only:
return y
# Load acoustic feature and pad
x = [self.get_raw_data_folder(_x) for _x in x]
x = [torchaudio.load(os.path.join(self.raw_root,f[:-4]+'.flac'))[0].squeeze(0) for f in x]
return x,y
def prepare_data(self, index):
# Load label
d = self.data[index]
x, y = zip(*d)
y = target_padding(y, max([len(v) for v in y]))
if self.text_only:
return y
# Load acoustic feature and pad
x = [torch.FloatTensor(np.load(os.path.join(self.root,f))) for f in x]
x = pad_sequence(x, batch_first=True)
return x,y
def __init__(self, file_path, raw_file_path, sets, bucket_size, max_timestep=0, max_label_len=0, raw_wav_data=False):
self.raw_root = raw_file_path
self.raw_wav_data = raw_wav_data
# Open dataset
x = []
y = []
tables = []
for s in sets:
with open(os.path.join(file_path,s+'_x.pkl'),'rb') as fp:
x += pickle.load(fp)
with open(os.path.join(file_path,s+'_y.pkl'),'rb') as fp:
y += pickle.load(fp)
# load data path
tables.append(pd.read_csv(os.path.join(file_path,s+'.csv')))
assert len(x)==len(y)
# Sort data w.r.t. length
self.X = []
self.Y = []
sortd_len = [len(t) for t in x]
sorted_x = [x[idx] for idx in reversed(np.argsort(sortd_len))]
sorted_y = [y[idx] for idx in reversed(np.argsort(sortd_len))]
self.table = pd.concat(tables,ignore_index=True).sort_values(by=['length'],ascending=False)
# Bucketing
for b in range(int(np.ceil(len(sorted_x)/bucket_size))):
offset = b*bucket_size
bound = min((b+1)*bucket_size,len(sorted_x))
bucket_max_timestep = min(max_timestep,len(sorted_x[offset]))
self.X.append(zero_padding(sorted_x[offset:bound], bucket_max_timestep))
bucket_max_label_len = min(max_label_len,max([len(v) for v in sorted_y[offset:bound]]))
self.Y.append(target_padding(sorted_y[offset:bound], bucket_max_label_len))
self.iterator = self.prepare_feature if not raw_wav_data else self.prepara_raw_data
# bucketing data
X = self.table['file_path'].tolist()
X_lens = self.table['length'].tolist()
Y = [list(map(int, label.split('_'))) for label in self.table['label'].tolist()]
# Bucketing, X & X_len is dummy when text_only==True
self.data = bucket_data(X_lens, list(zip(X,Y)), bucket_size)
self.transorm = torchaudio.transforms.MFCC(sample_rate=16000, n_mfcc=40, dct_type=2,
melkwargs={'hop_length':int(16*10), 'n_fft':int(16*25)})