def get_example(self, i):
rec, st, ed = self.chunk_indices[i]
Y, T = feature.get_labeledSTFT(
self.data,
rec,
st,
ed,
self.frame_size,
self.frame_shift,
self.n_speakers)
Y = feature.transform(Y, self.input_transform)
Y_spliced = feature.splice(Y, self.context_size)
Y_ss, T_ss = feature.subsample(Y_spliced, T, self.subsampling)
return Y_ss, T_ss
def __getitem__(self, i):
rec, st, ed = self.chunk_indices[i]
Y, T = feature.get_labeledSTFT(self.data, rec, st, ed, self.frame_size,
self.frame_shift, self.n_speakers)
# Y: (frame, num_ceps)
Y = feature.transform(Y, self.input_transform)
# Y_spliced: (frame, num_ceps * (context_size * 2 + 1))
Y_spliced = feature.splice(Y, self.context_size)
# Y_ss: (frame / subsampling, num_ceps * (context_size * 2 + 1))
Y_ss, T_ss = feature.subsample(Y_spliced, T, self.subsampling)
Y_ss = torch.from_numpy(Y_ss).float()
T_ss = torch.from_numpy(T_ss).float()
return Y_ss, T_ss
def get_example(self, i):
rec, st, ed = self.chunk_indices[i]
n_speakers, Y, T = feature.get_labeledSTFT(
self.data, rec, st, ed, self.frame_size, self.frame_shift, None
) # modification for number of speakers else it was -> self.n_speakers)
Y = feature.transform(Y, self.input_transform)
Y_spliced = feature.splice(Y, self.context_size)
Y_ss, T_ss = feature.subsample(Y_spliced, T, self.subsampling)
# here let self.n_speakers represent the max number of speakers in the train set
# Pad T_ss to concatenate successfully when training using _convert
T_ss = np.pad(T_ss, ((0, 0), (0, self.n_speakers - T_ss.shape[1])),
'constant',
constant_values=0)
return n_speakers, Y_ss, T_ss
def get_example(self, i):
rec, st, ed = self.chunk_indices[i]
Y, T = feature.get_labeledSTFT(self.data, rec, st, ed, self.frame_size,
self.frame_shift, self.n_speakers)
Y = feature.transform(Y, self.input_transform)
Y_spliced = feature.splice(Y, self.context_size)
Y_ss, T_ss = feature.subsample(Y_spliced, T, self.subsampling)
# If the sample contains more than "self.n_speakers" speakers,
# extract top-(self.n_speakers) speakers
if self.n_speakers and T_ss.shape[1] > self.n_speakers:
selected_speakers = np.argsort(
T_ss.sum(axis=0))[::-1][:self.n_speakers]
T_ss = T_ss[:, selected_speakers]
# If self.shuffle is True, shuffle the order in time-axis
# This operation improves the performance of EEND-EDA
if self.shuffle:
order = np.arange(Y_ss.shape[0])
np.random.shuffle(order)
Y_ss = Y_ss[order]
T_ss = T_ss[order]
return Y_ss, T_ss