def __getitem__(self, i):
self.check_index(i)
feat = kaldi_io.read_mat(self.rxfiles[i])
if self.specaugment_config is not None and self.specaugment_config != "":
with data_utils.numpy_seed(self.seed, self.epoch, i):
feat = specaug(feat, **eval(self.specaugment_config))
item = torch.from_numpy(feat).float()
return item
def __getitem__(self, i):
self.check_index(i)
ptx = self.data_offsets[i]
a = self.buffer[ptx:ptx + self.sizes[i]].copy()
if self.specaugment_config is not None and self.specaugment_config != "":
with data_utils.numpy_seed(self.seed, self.epoch, i):
a = specaug(a, **eval(self.specaugment_config))
return torch.from_numpy(a).float()
def _get_features(self, i):
if self.input_format == "feat":
feat = kaldi_io.read_mat(self.rxfiles[i])
else:
if self.input_format == "command":
source = BytesIO(run(self.rxfiles[i][:-1], shell=True, stdout=PIPE).stdout)
else:
source = self.rxfiles[i]
waveform, sample_rate = get_waveform(source, normalization=False, always_2d=True)
feat = get_torchaudio_fbank_or_mfcc(waveform, sample_rate, n_bins=self.feat_dim, feature_type=self.feature_type)
if self.feature_transforms is not None:
feat = self.feature_transforms(feat)
if self.specaugment_config is not None and self.specaugment_config != "":
with data_utils.numpy_seed(self.seed, self.epoch, i):
feat = specaug(feat, **eval(self.specaugment_config))
return feat
def __getitem__(self, i):
self.check_index(i)
if not self.prefetch_called: # no caching
feat = kaldi_io.read_mat(self.rxfiles[i])
return torch.from_numpy(feat).float()
if i not in self.cache_index:
assert (
self.start_pos_for_next_cache < len(self.ordered_indices)
), "Position for next cache starting beyond the end of ordered_indices."
try:
pos_start = self.ordered_indices.index(
i,
self.start_pos_for_next_cache,
)
except ValueError:
raise ValueError(
"index {} not found in self.ordered_indices. Set "
"self.ordered_prefetch to False, and/or call self.prefetch() "
"with the full list of indices, and then try again.".
format(i))
pos_end = min(
pos_start + self.cache_size,
len(self.ordered_indices),
)
self.start_pos_for_next_cache = pos_end if self.ordered_prefetch else 0
total_size = 0
for idx in self.ordered_indices[pos_start:pos_end]:
total_size += self.sizes[idx]
self.cache = np.empty((total_size, self.feat_dim),
dtype=self.dtype)
ptx = 0
self.cache_index.clear()
for idx in self.ordered_indices[pos_start:pos_end]:
self.cache_index[idx] = ptx
length = self.sizes[idx]
dst = self.cache[ptx:ptx + length]
feat = kaldi_io.read_mat(self.rxfiles[idx])
if self.specaugment_config is not None and self.specaugment_config != "":
with data_utils.numpy_seed(self.seed, self.epoch, idx):
feat = specaug(feat, **eval(self.specaugment_config))
np.copyto(dst, feat)
ptx += length
ptx = self.cache_index[i]
a = self.cache[ptx:ptx + self.sizes[i]].copy()
return torch.from_numpy(a).float()