def __getitem__(self, index):
# Load label
if self.load == 'asr' or self.load == 'text':
y_batch = [y for y in self.Y[index]]
y_pad_batch = target_padding(y_batch,
max([len(v) for v in y_batch]))
if self.load == 'text':
return y_pad_batch
# Load acoustic feature and pad
x_batch = [
torch.FloatTensor(np.load(os.path.join(self.root, x_file)))
for x_file in self.X[index]
]
x_pad_batch = pad_sequence(x_batch, batch_first=True)
return x_pad_batch, y_pad_batch
def __init__(self,
run_mockingjay,
file_path,
sets,
bucket_size,
max_timestep=0,
max_label_len=0,
mock_config=None):
self.run_mockingjay = run_mockingjay
self.mock_config = mock_config
self.class_num = 63
# Open dataset
x = []
y = []
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)
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))]
# 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))