class AudioBatchData(Dataset):
def __init__(self,
path,
sizeWindow,
seqNames,
phoneLabelsDict,
nSpeakers,
nProcessLoader=50,
MAX_SIZE_LOADED=4000000000):
"""
Args:
- path (string): path to the training dataset
- sizeWindow (int): size of the sliding window
- seqNames (list): sequences to load
- phoneLabelsDict (dictionnary): if not None, a dictionnary with the
following entries
"step": size of a labelled window
"$SEQ_NAME": list of phonem labels for
the sequence $SEQ_NAME
- nSpeakers (int): number of speakers to expect.
- nProcessLoader (int): number of processes to call when loading the
data from the disk
- MAX_SIZE_LOADED (int): target maximal size of the floating array
containing all loaded data.
"""
self.MAX_SIZE_LOADED = MAX_SIZE_LOADED
self.nProcessLoader = nProcessLoader
self.dbPath = Path(path)
self.sizeWindow = sizeWindow
self.seqNames = [(s, self.dbPath / x) for s, x in seqNames]
self.reload_pool = Pool(nProcessLoader)
self.prepare()
self.speakers = list(range(nSpeakers))
self.data = []
self.phoneSize = 0 if phoneLabelsDict is None else \
phoneLabelsDict["step"]
self.phoneStep = 0 if phoneLabelsDict is None else \
self.sizeWindow // self.phoneSize
self.phoneLabelsDict = deepcopy(phoneLabelsDict)
self.loadNextPack(first=True)
self.loadNextPack()
self.doubleLabels = False
def resetPhoneLabels(self, newPhoneLabels, step):
self.phoneSize = step
self.phoneStep = self.sizeWindow // self.phoneSize
self.phoneLabelsDict = deepcopy(newPhoneLabels)
self.loadNextPack()
def splitSeqTags(seqName):
path = os.path.normpath(seqName)
return path.split(os.sep)
def getSeqNames(self):
return [str(x[1]) for x in self.seqNames]
def clear(self):
if 'data' in self.__dict__:
del self.data
if 'speakerLabel' in self.__dict__:
del self.speakerLabel
if 'phoneLabels' in self.__dict__:
del self.phoneLabels
if 'seqLabel' in self.__dict__:
del self.seqLabel
def prepare(self):
random.shuffle(self.seqNames)
start_time = time.time()
print("Checking length...")
allLength = self.reload_pool.map(extractLength, self.seqNames)
self.packageIndex, self.totSize = [], 0
start, packageSize = 0, 0
for index, length in tqdm.tqdm(enumerate(allLength)):
packageSize += length
if packageSize > self.MAX_SIZE_LOADED:
self.packageIndex.append([start, index])
self.totSize += packageSize
start, packageSize = index, 0
if packageSize > 0:
self.packageIndex.append([start, len(self.seqNames)])
self.totSize += packageSize
print(f"Done, elapsed: {time.time() - start_time:.3f} seconds")
print(f'Scanned {len(self.seqNames)} sequences '
f'in {time.time() - start_time:.2f} seconds')
print(f"{len(self.packageIndex)} chunks computed")
self.currentPack = -1
self.nextPack = 0
def getNPacks(self):
return len(self.packageIndex)
def loadNextPack(self, first=False):
self.clear()
if not first:
self.currentPack = self.nextPack
start_time = time.time()
print('Joining pool')
self.r.wait()
print(f'Joined process, elapsed={time.time()-start_time:.3f} secs')
self.nextData = self.r.get()
self.parseNextDataBlock()
del self.nextData
self.nextPack = (self.currentPack + 1) % len(self.packageIndex)
seqStart, seqEnd = self.packageIndex[self.nextPack]
if self.nextPack == 0 and len(self.packageIndex) > 1:
self.prepare()
self.r = self.reload_pool.map_async(loadFile,
self.seqNames[seqStart:seqEnd])
def parseNextDataBlock(self):
# Labels
self.speakerLabel = [0]
self.seqLabel = [0]
self.phoneLabels = []
speakerSize = 0
indexSpeaker = 0
# To accelerate the process a bit
self.nextData.sort(key=lambda x: (x[0], x[1]))
tmpData = []
for speaker, seqName, seq in self.nextData:
while self.speakers[indexSpeaker] < speaker:
indexSpeaker += 1
self.speakerLabel.append(speakerSize)
if self.speakers[indexSpeaker] != speaker:
raise ValueError(f'{speaker} invalid speaker')
if self.phoneLabelsDict is not None:
self.phoneLabels += self.phoneLabelsDict[seqName]
newSize = len(self.phoneLabelsDict[seqName]) * self.phoneSize
seq = seq[:newSize]
sizeSeq = seq.size(0)
tmpData.append(seq)
self.seqLabel.append(self.seqLabel[-1] + sizeSeq)
speakerSize += sizeSeq
del seq
self.speakerLabel.append(speakerSize)
self.data = torch.cat(tmpData, dim=0)
def getPhonem(self, idx):
idPhone = idx // self.phoneSize
return self.phoneLabels[idPhone:(idPhone + self.phoneStep)]
def getSpeakerLabel(self, idx):
idSpeaker = next(
x[0] for x in enumerate(self.speakerLabel) if x[1] > idx) - 1
return idSpeaker
def __len__(self):
return self.totSize // self.sizeWindow
def __getitem__(self, idx):
if idx < 0 or idx >= len(self.data) - self.sizeWindow - 1:
print(idx)
outData = self.data[idx:(self.sizeWindow + idx)].view(1, -1)
label = torch.tensor(self.getSpeakerLabel(idx), dtype=torch.long)
if self.phoneSize > 0:
label_phone = torch.tensor(self.getPhonem(idx), dtype=torch.long)
if not self.doubleLabels:
label = label_phone
else:
label_phone = torch.zeros(1)
if self.doubleLabels:
return outData, label, label_phone
return outData, label
def getNSpeakers(self):
return len(self.speakers)
def getNSeqs(self):
return len(self.seqLabel) - 1
def getNLoadsPerEpoch(self):
return len(self.packageIndex)
def getBaseSampler(self, type, batchSize, offset):
if type == "samespeaker":
return SameSpeakerSampler(batchSize, self.speakerLabel,
self.sizeWindow, offset)
if type == "samesequence":
return SameSpeakerSampler(batchSize, self.seqLabel,
self.sizeWindow, offset)
if type == "sequential":
return SequentialSampler(len(self.data), self.sizeWindow, offset,
batchSize)
sampler = UniformAudioSampler(len(self.data), self.sizeWindow, offset)
return BatchSampler(sampler, batchSize, True)
def getDataLoader(self,
batchSize,
type,
randomOffset,
numWorkers=0,
onLoop=-1):
r"""
Get a batch sampler for the current dataset.
Args:
- batchSize (int): batch size
- groupSize (int): in the case of type in ["samespeaker", "samesequence"]
number of items sharing a same label in the group
(see AudioBatchSampler)
- type (string):
type == "speaker": grouped sampler speaker-wise
type == "sequence": grouped sampler sequence-wise
type == "sequential": sequential sampling
else: uniform random sampling of the full audio
vector
- randomOffset (bool): if True add a random offset to the sampler
at the begining of each iteration
"""
nLoops = len(self.packageIndex)
totSize = self.totSize // (self.sizeWindow * batchSize)
if onLoop >= 0:
self.currentPack = onLoop - 1
self.loadNextPack()
nLoops = 1
def samplerCall():
offset = random.randint(0, self.sizeWindow // 2) \
if randomOffset else 0
return self.getBaseSampler(type, batchSize, offset)
return AudioLoader(self, samplerCall, nLoops, self.loadNextPack,
totSize, numWorkers)
def train(self, n_process=None):
assert n_process is not None
pool = Pool(processes=n_process)
outer_lr_scheduler = PiecewiseSchedule(
[(0, self.outer_lr),
(int(self.outer_n_epoch / 2), self.outer_lr * 0.1)],
outside_value=self.outer_lr * 0.1)
evoluted_loss = loss_net(self.inner_args['nclass'])
evoluted_loss = torch.nn.DataParallel(evoluted_loss).cuda()
evoluted_loss.train()
for epoch in range(self.outer_n_epoch):
phi_state_dict = evoluted_loss.state_dict()
phi_noise = []
for i in range(self.outer_n_noise):
noise = dict()
for key in evoluted_loss.state_dict().keys():
noise[key] = self.outer_std * torch.randn_like(
evoluted_loss.state_dict()[key]) + phi_state_dict[key]
phi_noise.append(noise)
start_time = time.time()
assert self.outer_n_worker % self.outer_n_noise == 0
args = ((epoch, i, phi_noise[i], self.target_model,
self.train_dataset, self.test_dataset, self.inner_args)
for i in range(self.outer_n_noise))
results = pool.map_async(inner_train, args)
results = results.get()
result = [np.mean(r['accuracy']) for r in results]
print(
'Epoch %d, mean accuracy: %f, max accuracy: %f(%d), min accuracy: %f(%d)'
% (epoch, np.mean(result), np.max(result),
result.index(np.max(result)), np.min(result),
result.index(np.min(result))))
print('All inner loops completed, returns gathered ({:.2f} sec).'.
format(time.time() - start_time))
print('\n')
for key in evoluted_loss.state_dict():
grad = 0
for i in range(self.outer_n_noise):
grad += result[i] * phi_noise[i][key].cpu(
) # - self.outer_l2 * evoluted_loss.state_dict()[key].cpu()
lr = outer_lr_scheduler.value(epoch)
adam = Adam(shape=grad.shape, stepsize=lr)
evoluted_loss.state_dict()[key] -= adam.step(
grad / self.outer_n_worker).cuda()
if self.save_model:
torch.save(
{
'epoch': epoch + 1,
'state_dict': evoluted_loss.state_dict()
}, self.log.path_model)
self.log.log_tabular('epoch', epoch)
for i in range(len(result)):
self.log.log_tabular('reward %d' % i, result[i])
self.log.dump_tabular()
pool.close()
class LibriSelectionDataset(Dataset):
"""LibriSpeech Selection data from sincnet paper."""
def __init__(self,
sizeWindow=20480,
db_wav_root=DB_WAV_ROOT,
fps_list=str(),
label_path=str(),
nSpeakers=-1,
n_process_loader=50,
MAX_SIZE_LOADED=4000000000):
"""Init.
Args:
- sizeWindow (int): size of the sliding window
- db_wav_path (str):
- fps_list_path (str):
- label_path (str):
- n_process_loader (int):
- MAX_SIZE_LOADED (int): target maximal size of the floating array
containing all loaded data.
"""
self.MAX_SIZE_LOADED = MAX_SIZE_LOADED
self.n_process_loader = n_process_loader
self.db_wav_root = Path(db_wav_root)
self.sizeWindow = sizeWindow
"""Parsing customized to Libri-selection dataset."""
fps_name_only = get_fps_from_txt(fps_list)
label_dict = np.load(label_path, allow_pickle=True)[()]
self.all_labels_fps = [(label_dict[x], Path(db_wav_root) / Path(x))
for x in fps_name_only]
self.reload_pool = Pool(n_process_loader)
self.prepare(
) # Split large number of files into packages, and set {self.currentPack=-1, self.nextPack=0}
if nSpeakers == -1:
nSpeakers = len(set(label_dict.values()))
self.speakers = list(range(nSpeakers))
self.data = []
self.loadNextPack(first=True)
self.loadNextPack()
def __len__(self):
"""Get length."""
return self.totSize // self.sizeWindow
def prepare(self):
"""Prepare."""
random.shuffle(self.all_labels_fps)
start_time = time.time()
print("Checking length...")
allLength = self.reload_pool.map(extractLength, self.all_labels_fps)
self.packageIndex, self.totSize = [], 0
start, packageSize = 0, 0
for index, length in tqdm.tqdm(enumerate(allLength)):
packageSize += length
if packageSize > self.MAX_SIZE_LOADED:
self.packageIndex.append([start, index])
self.totSize += packageSize
start, packageSize = index, 0
if packageSize > 0:
self.packageIndex.append([start, len(self.all_labels_fps)])
self.totSize += packageSize
print(f"Done, elapsed: {time.time() - start_time:.3f} seconds")
print(f'Scanned {len(self.all_labels_fps)} sequences '
f'in {time.time() - start_time:.2f} seconds')
print(f"{len(self.packageIndex)} chunks computed")
self.currentPack = -1
self.nextPack = 0
def clear(self):
"""Clear."""
if 'data' in self.__dict__:
del self.data
if 'speakerLabel' in self.__dict__:
del self.speakerLabel
if 'seqLabel' in self.__dict__:
del self.seqLabel
def getNPacks(self):
"""Get N packs."""
return len(self.packageIndex)
def getNSeqs(self):
"""Get N seqs."""
return len(self.seqLabel) - 1
def getNLoadsPerEpoch(self):
"""Get N loads per epoch."""
return len(self.packageIndex)
def getSpeakerLabel(self, idx):
idSpeaker = next(
x[0] for x in enumerate(self.speakerLabel) if x[1] > idx) - 1
return idSpeaker
def loadNextPack(self, first=False):
"""Load next pack."""
self.clear()
if not first:
self.currentPack = self.nextPack
start_time = time.time()
print('Joining pool')
self.r.wait()
print(f'Joined process, elapsed={time.time()-start_time:.3f} secs')
self.nextData = self.r.get()
self.parseNextDataBlock()
del self.nextData
self.nextPack = (self.currentPack + 1) % len(self.packageIndex)
seqStart, seqEnd = self.packageIndex[self.nextPack]
if self.nextPack == 0 and len(self.packageIndex) > 1:
self.prepare()
"""map() blocks until complete, map_async() returns immediately and
schedules a callback to be run on the result."""
self.r = self.reload_pool.map_async(
loadFile, self.all_labels_fps[seqStart:seqEnd])
"""loadFile: return speaker, seqName, seq"""
def parseNextDataBlock(self):
"""Parse next data block."""
# Labels
self.speakerLabel = [0]
self.seqLabel = [0]
speakerSize = 0
indexSpeaker = 0
# To accelerate the process a bit
self.nextData.sort(key=lambda x: (x[0], x[1]))
"""
nextData[0] = (1243, '4910-14124-0001-1',
tensor([-0.0089, -0.0084, -0.0079, ..., -0.0015, -0.0056, 0.0047]))
"""
tmpData = []
for speaker, seqName, seq in self.nextData:
while self.speakers[indexSpeaker] < speaker:
indexSpeaker += 1
self.speakerLabel.append(speakerSize)
if self.speakers[indexSpeaker] != speaker:
raise ValueError(f'{speaker} invalid speaker')
sizeSeq = seq.size(0)
tmpData.append(seq)
self.seqLabel.append(self.seqLabel[-1] + sizeSeq)
speakerSize += sizeSeq
del seq
self.speakerLabel.append(speakerSize)
self.data = torch.cat(tmpData, dim=0)
def __getitem__(self, idx):
"""Get item."""
if idx < 0 or idx >= len(self.data) - self.sizeWindow - 1:
print(idx)
outData = self.data[idx:(self.sizeWindow + idx)].view(1, -1)
label = torch.tensor(self.getSpeakerLabel(idx), dtype=torch.long)
return outData, label
def getBaseSampler(self, type, batchSize, offset):
"""Get base sampler."""
if type == "samespeaker":
return SameSpeakerSampler(batchSize, self.speakerLabel,
self.sizeWindow, offset)
if type == "samesequence":
return SameSpeakerSampler(batchSize, self.seqLabel,
self.sizeWindow, offset)
if type == "sequential":
return SequentialSampler(len(self.data), self.sizeWindow, offset,
batchSize)
sampler = UniformAudioSampler(len(self.data), self.sizeWindow, offset)
return BatchSampler(sampler, batchSize, True)
def getDataLoader(self,
batchSize,
type,
randomOffset,
numWorkers=0,
onLoop=-1):
"""Get a batch sampler for the current dataset.
Args:
- batchSize (int): batch size
- groupSize (int): in the case of type in ["samespeaker", "samesequence"]
number of items sharing a same label in the group
(see AudioBatchSampler)
- type (string):
type == "samespeaker": grouped sampler speaker-wise
type == "samesequence": grouped sampler sequence-wise
type == "sequential": sequential sampling
else: uniform random sampling of the full audio
vector
- randomOffset (bool): if True add a random offset to the sampler
at the begining of each iteration
"""
nLoops = len(self.packageIndex)
totSize = self.totSize // (self.sizeWindow * batchSize)
if onLoop >= 0:
self.currentPack = onLoop - 1
self.loadNextPack()
nLoops = 1
def samplerCall():
offset = random.randint(0, self.sizeWindow // 2) \
if randomOffset else 0
return self.getBaseSampler(type, batchSize, offset)
return AudioLoader(self, samplerCall, nLoops, self.loadNextPack,
totSize, numWorkers)
class DataLoader():
def __init__(self, args):
self.dir_bin = args.dir_bin
line_load_list = self.dir_bin + 'line_load_list.t7'
vocab_file = self.dir_bin + 'vocab.t7'
assert os.path.isfile(self.dir_bin + 'specM.bin')
assert os.path.isfile(self.dir_bin + 'specL.bin')
assert os.path.isfile(self.dir_bin + 'text.bin')
self.batch_size = args.batch_size
self.trunc_size = args.trunc_size
self.r_factor = args.r_factor
self.dec_out_size = args.dec_out_size
self.post_out_size = args.post_out_size
self.shuffle_data = True if args.shuffle_data == 1 else False
self.iter_per_epoch = None
self.is_subbatch_end = True
self.curr_split = None
self.vocab_size = None
self.process = None
self.queue = Queue(maxsize=args.load_queue_size)
self.n_workers = args.n_workers
self.use_gpu = args.use_gpu
self.num_gpu = len(args.gpu) if len(args.gpu) > 0 else 1
self.pinned_memory = True if args.pinned_memory == 1 and self.use_gpu else False
self.vocab_size = self.get_num_vocab(vocab_file)
text_limit = args.text_limit
wave_limit = args.wave_limit
# col1: idx / col2: wave_length / col3: text_length
# col4: offset_M / col5: offset_L / col6: offset_T
self.load_list = torch.load(line_load_list)
spec_len_list = self.load_list[:, 1].clone()
text_len_list = self.load_list[:, 2].clone()
# exclude files whose wave length exceeds wave_limit
sort_length, sort_idx = spec_len_list.sort()
text_len_list = torch.gather(text_len_list, 0, sort_idx)
sort_idx = sort_idx.view(-1, 1).expand_as(self.load_list)
self.load_list = torch.gather(self.load_list, 0, sort_idx)
end_idx = sort_length.le(wave_limit).sum()
spec_len_list = sort_length[:end_idx]
text_len_list = text_len_list[:end_idx]
self.load_list = self.load_list[:end_idx]
# exclude files whose text length exceeds text_limit
sort_length, sort_idx = text_len_list.sort()
spec_len_list = torch.gather(spec_len_list, 0, sort_idx)
sort_idx = sort_idx.view(-1, 1).expand_as(self.load_list)
self.load_list = torch.gather(self.load_list, 0, sort_idx)
end_idx = sort_length.le(text_limit).sum()
end_idx = end_idx - (end_idx % self.batch_size) # drop residual data
text_len_list = sort_length[:end_idx]
spec_len_list = spec_len_list[:end_idx]
self.load_list = self.load_list[:end_idx]
# sort by wave length
_, sort_idx = spec_len_list.sort(0, descending=True)
text_len_list = torch.gather(text_len_list, 0, sort_idx)
sort_idx = sort_idx.view(-1, 1).expand_as(self.load_list)
self.load_list = torch.gather(self.load_list, 0, sort_idx)
# sort by text length in each batch (PackedSequence requires it)
num_batches_per_epoch = self.load_list.size(0) // self.batch_size
text_len_list = text_len_list.view(num_batches_per_epoch, -1)
self.load_list = self.load_list.view(num_batches_per_epoch, -1,
self.load_list.size(1))
sort_length, sort_idx = text_len_list.sort(1, descending=True)
sort_idx = sort_idx.view(num_batches_per_epoch, -1,
1).expand_as(self.load_list)
self.load_list = torch.gather(self.load_list, 1, sort_idx)
# shuffle while preserving order in a batch
if self.shuffle_data:
_, sort_idx = torch.randn(num_batches_per_epoch).sort()
sort_idx = sort_idx.view(-1, 1, 1).expand_as(self.load_list)
self.load_list = torch.gather(self.load_list, 0,
sort_idx) # nbpe x N x 6
self.load_list = self.load_list.long()
# compute number of iterations needed
spec_len_list = spec_len_list.view(num_batches_per_epoch, -1)
spec_len_list, _ = spec_len_list.div(self.trunc_size).ceil().max(1)
self.iter_per_epoch = int(spec_len_list.sum())
# set split cursor
self.split_sizes = {
'train': self.load_list.size(0),
'val': -1,
'test': -1
}
self.split_cursor = {'train': 0, 'val': 0, 'test': 0}
def next_batch(self, split):
T, idx = self.trunc_size, self.split_cursor[split]
# seek and load data from raw files
if self.is_subbatch_end:
self.is_subbatch_end = False
self.subbatch_cursor = 0
if self.curr_split != split:
self.curr_split = split
if self.process is not None:
self.process.terminate()
self.process = Process(target=self.start_async_loader,
args=(split, self.split_cursor[split]))
self.process.start()
self.len_text, self.len_wave, self.curr_text, self.curr_specM, self.curr_specL = self.queue.get(
)
self.split_cursor[split] = (idx + 1) % self.split_sizes[split]
self.subbatch_max_len = self.len_wave.max()
# Variables to return
# +1 to length of y to consider shifting for target y
subbatch_len_text = [x for x in self.len_text]
subbatch_len_wave = [min(x, T) for x in self.len_wave]
x_text = self.curr_text
y_specM = self.curr_specM[:,
self.subbatch_cursor:self.subbatch_cursor +
max(subbatch_len_wave) + 1].contiguous()
y_specL = self.curr_specL[:,
self.subbatch_cursor:self.subbatch_cursor +
max(subbatch_len_wave) + 1].contiguous()
if self.use_gpu:
if self.pinned_memory:
x_text = x_text.pin_memory()
y_specM = y_specM.pin_memory()
y_specL = y_specL.pin_memory()
x_text = x_text.cuda()
y_specM = y_specM.cuda()
y_specL = y_specL.cuda()
# Advance split_cursor or Move on to the next batch
if self.subbatch_cursor + T < self.subbatch_max_len:
self.subbatch_cursor = self.subbatch_cursor + T
self.len_wave.sub_(T).clamp_(min=0)
else:
self.is_subbatch_end = True
# Don't compute for empty batch elements
if subbatch_len_wave.count(0) > 0:
self.len_wave_mask = [
idx for idx, l in enumerate(subbatch_len_wave) if l > 0
]
self.len_wave_mask = torch.LongTensor(self.len_wave_mask)
if self.use_gpu:
self.len_wave_mask = self.len_wave_mask.cuda()
x_text = torch.index_select(x_text, 0, self.len_wave_mask)
y_specM = torch.index_select(y_specM, 0, self.len_wave_mask)
y_specL = torch.index_select(y_specL, 0, self.len_wave_mask)
subbatch_len_text = [
subbatch_len_text[idx] for idx in self.len_wave_mask
]
subbatch_len_wave = [
subbatch_len_wave[idx] for idx in self.len_wave_mask
]
else:
self.len_wave_mask = None
return x_text, y_specM, y_specL, subbatch_len_wave, subbatch_len_text
def start_async_loader(self, split, load_start_idx):
# load batches to the queue asynchronously since it is a bottle-neck
N, r = self.batch_size, self.r_factor
load_curr_idx = load_start_idx
while True:
data_T, data_M, data_L, len_T, len_M = ([None for _ in range(N)]
for _ in range(5))
# deploy workers to load data
self.pool = Pool(self.n_workers)
partial_func = partial(load_data_and_length, self.dir_bin,
self.load_list[load_curr_idx])
results = self.pool.map_async(func=partial_func, iterable=range(N))
self.pool.close()
self.pool.join()
for result in results.get():
data_M[result[0]] = result[1]
data_L[result[0]] = result[2]
data_T[result[0]] = result[3]
len_T[result[0]] = result[4]
len_M[result[0]] = result[5]
# TODO: output size is not accurate.. //
len_text = torch.IntTensor(len_T)
len_wave = torch.Tensor(len_M).div(r).ceil().mul(
r).int() # consider r_factor
curr_text = torch.LongTensor(N, len_text.max()).fill_(
0) # null-padding at tail
curr_specM = torch.Tensor(N,
len_wave.max() + 1,
self.dec_out_size).fill_(
0) # null-padding at tail
curr_specL = torch.Tensor(N,
len_wave.max() + 1,
self.post_out_size).fill_(
0) # null-padding at tail
# fill the template tensors
for j in range(N):
curr_text[j, 0:data_T[j].size(0)].copy_(data_T[j])
curr_specM[j, 0:data_M[j].size(0)].copy_(data_M[j])
curr_specL[j, 0:data_L[j].size(0)].copy_(data_L[j])
self.queue.put(
(len_text, len_wave, curr_text, curr_specM, curr_specL))
load_curr_idx = (load_curr_idx + 1) % self.split_sizes[split]
def mask_prev_h(self, prev_h):
if self.len_wave_mask is not None:
if self.use_gpu:
self.len_wave_mask = self.len_wave_mask.cuda()
h_att, h_dec1, h_dec2 = prev_h
h_att = torch.index_select(h_att.data, 1,
self.len_wave_mask) # batch idx is
h_dec1 = torch.index_select(h_dec1.data, 1, self.len_wave_mask)
h_dec2 = torch.index_select(h_dec2.data, 1, self.len_wave_mask)
prev_h = (Variable(h_att), Variable(h_dec1), Variable(h_dec2))
else:
prev_h = prev_h
return prev_h
def get_num_vocab(self, vocab_file=None):
if self.vocab_size:
return self.vocab_size
else:
vocab_dict = torch.load(vocab_file)
return len(vocab_dict) + 1 # +1 to consider null-padding
class SequentialData(Dataset):
def __init__(self,
path,
seqNames,
nProcessLoader=50,
MAX_SIZE_LOADED=4000000000):
"""
Args:
- path (string): path to the training dataset
- seqNames (list): sequences to load
- nProcessLoader (int): number of processes to call when loading the
data from the disk
- MAX_SIZE_LOADED (int): target maximal size of the floating array
containing all loaded data.
"""
self.MAX_SIZE_LOADED = MAX_SIZE_LOADED
self.nProcessLoader = nProcessLoader
self.dbPath = Path(path)
self.seqNames = [self.dbPath / x for _, x in seqNames]
self.reload_pool = Pool(nProcessLoader)
self.prepare()
self.data = []
self.loadNextPack(first=True)
self.loadNextPack()
def getSeqNames(self):
return [str(x) for x in self.seqNames]
def clear(self):
if 'data' in self.__dict__:
del self.data
def prepare(self):
shuffle(self.seqNames)
start_time = time.time()
print("Checking length...")
allShape = self.reload_pool.map(extractShape, self.seqNames)
self.packageIndex, self.totSize = [], 0
start, packageSize = 0, 0
for index, shape in tqdm.tqdm(enumerate(allShape)):
packageSize += shape[0]
if packageSize * shape[1] > self.MAX_SIZE_LOADED:
self.packageIndex.append([start, index])
self.totSize += packageSize
start, packageSize = index, 0
if packageSize > 0:
self.packageIndex.append([start, len(self.seqNames)])
self.totSize += packageSize
print(f"Done, elapsed: {time.time() - start_time:.3f} seconds")
print(f'Scanned {len(self.seqNames)} sequences '
f'in {time.time() - start_time:.2f} seconds')
print(f"{len(self.packageIndex)} chunks computed")
self.currentPack = -1
self.nextPack = 0
def getNPacks(self):
return len(self.packageIndex)
def loadNextPack(self, first=False):
self.clear()
if not first:
self.currentPack = self.nextPack
start_time = time.time()
print('Joining pool')
self.r.wait()
print(f'Joined process, elapsed={time.time()-start_time:.3f} secs')
self.nextData = self.r.get()
self.parseNextDataBlock()
del self.nextData
self.nextPack = (self.currentPack + 1) % len(self.packageIndex)
seqStart, seqEnd = self.packageIndex[self.nextPack]
#if self.nextPack == 0 and len(self.packageIndex) > 1:
# self.prepare()
self.r = self.reload_pool.map_async(loadFilePool,
self.seqNames[seqStart:seqEnd])
def parseNextDataBlock(self):
# To accelerate the process a bit
self.nextData.sort(key=lambda x: x[0])
tmpData = []
for seqName, seq in self.nextData:
tmpData.append(seq)
del seq
self.data = torch.cat(tmpData, dim=0)
def __len__(self):
return self.totSize
def __getitem__(self, idx):
if idx < 0 or idx >= len(self.data):
print(idx)
outData = self.data[idx].view(1, -1)
return outData
def getNLoadsPerEpoch(self):
return len(self.packageIndex)
def getDataLoader(self, batchSize, numWorkers=0):
r"""
Get a batch sampler for the current dataset.
Args:
- batchSize (int): batch size
"""
nLoops = len(self.packageIndex)
totSize = self.totSize // batchSize
def samplerCall():
sampler = UniformAudioSampler(len(self.data))
return BatchSampler(sampler, batchSize, True)
return SequentialLoader(self, samplerCall, nLoops, self.loadNextPack,
totSize, numWorkers)
