def create_data(config, data_path):
dataset = LJSpeech(data_path)
train_dataset = SliceDataset(dataset, config["valid_size"], len(dataset))
train_collator = DataCollector(config["p_pronunciation"])
train_sampler = RandomSampler(train_dataset)
train_cargo = DataCargo(train_dataset,
train_collator,
batch_size=config["batch_size"],
sampler=train_sampler)
train_loader = DataLoader\
.from_generator(capacity=10, return_list=True)\
.set_batch_generator(train_cargo)
valid_dataset = SliceDataset(dataset, 0, config["valid_size"])
valid_collector = DataCollector(1.)
valid_sampler = SequentialSampler(valid_dataset)
valid_cargo = DataCargo(valid_dataset,
valid_collector,
batch_size=1,
sampler=valid_sampler)
valid_loader = DataLoader\
.from_generator(capacity=2, return_list=True)\
.set_batch_generator(valid_cargo)
return train_loader, valid_loader
def __init__(self, config, nranks, rank):
place = fluid.CUDAPlace(rank) if config.use_gpu else fluid.CPUPlace()
# Whole LJSpeech dataset.
ds = Dataset(config)
# Split into train and valid dataset.
indices = list(range(len(ds)))
train_indices = indices[config.valid_size:]
valid_indices = indices[:config.valid_size]
random.shuffle(train_indices)
# Train dataset.
trainset = Subset(ds, train_indices, valid=False)
sampler = DistributedSampler(len(trainset), nranks, rank)
total_bs = config.batch_size
assert total_bs % nranks == 0
train_sampler = BatchSampler(sampler,
total_bs // nranks,
drop_last=True)
trainloader = DataCargo(trainset, batch_sampler=train_sampler)
trainreader = fluid.io.PyReader(capacity=50, return_list=True)
trainreader.decorate_batch_generator(trainloader, place)
self.trainloader = (data for _ in iter(int, 1)
for data in trainreader())
# Valid dataset.
validset = Subset(ds, valid_indices, valid=True)
# Currently only support batch_size = 1 for valid loader.
validloader = DataCargo(validset, batch_size=1, shuffle=False)
validreader = fluid.io.PyReader(capacity=20, return_list=True)
validreader.decorate_batch_generator(validloader, place)
self.validloader = validreader
def make_data_loader(data_root, config):
# construct meta data
meta = LJSpeechMetaData(data_root)
# filter it!
min_text_length = config["meta_data"]["min_text_length"]
meta = FilterDataset(meta, lambda x: len(x[2]) >= min_text_length)
# transform meta data into meta data
c = config["transform"]
transform = Transform(
replace_pronunciation_prob=c["replace_pronunciation_prob"],
sample_rate=c["sample_rate"],
preemphasis=c["preemphasis"],
n_fft=c["n_fft"],
win_length=c["win_length"],
hop_length=c["hop_length"],
fmin=c["fmin"],
fmax=c["fmax"],
n_mels=c["n_mels"],
min_level_db=c["min_level_db"],
ref_level_db=c["ref_level_db"],
max_norm=c["max_norm"],
clip_norm=c["clip_norm"])
ljspeech = CacheDataset(TransformDataset(meta, transform))
# use meta data's text length as a sort key for the sampler
batch_size = config["train"]["batch_size"]
text_lengths = [len(example[2]) for example in meta]
sampler = PartialyRandomizedSimilarTimeLengthSampler(text_lengths,
batch_size)
env = dg.parallel.ParallelEnv()
num_trainers = env.nranks
local_rank = env.local_rank
sampler = BucketSampler(
text_lengths, batch_size, num_trainers=num_trainers, rank=local_rank)
# some model hyperparameters affect how we process data
model_config = config["model"]
collector = DataCollector(
downsample_factor=model_config["downsample_factor"],
r=model_config["outputs_per_step"])
ljspeech_loader = DataCargo(
ljspeech, batch_fn=collector, batch_size=batch_size, sampler=sampler)
loader = fluid.io.DataLoader.from_generator(capacity=10, return_list=True)
loader.set_batch_generator(
ljspeech_loader, places=fluid.framework._current_expected_place())
return loader
n_loop = model_config["n_loop"]
n_layer = model_config["n_layer"]
filter_size = model_config["filter_size"]
context_size = 1 + n_layer * sum([filter_size**i for i in range(n_loop)])
print("context size is {} samples".format(context_size))
train_batch_fn = DataCollector(context_size, sample_rate, hop_length,
train_clip_seconds)
valid_batch_fn = DataCollector(context_size,
sample_rate,
hop_length,
train_clip_seconds,
valid=True)
batch_size = data_config["batch_size"]
train_cargo = DataCargo(ljspeech_train,
train_batch_fn,
batch_size,
sampler=RandomSampler(ljspeech_train))
# only batch=1 for validation is enabled
valid_cargo = DataCargo(ljspeech_valid,
valid_batch_fn,
batch_size=1,
sampler=SequentialSampler(ljspeech_valid))
if not os.path.exists(args.output):
os.makedirs(args.output)
if args.device == -1:
place = fluid.CPUPlace()
else:
place = fluid.CUDAPlace(args.device)
ljspeech = TransformDataset(meta, transform)
# =========================dataiterator=========================
# use meta data's text length as a sort key for the sampler
train_config = config["train"]
batch_size = train_config["batch_size"]
text_lengths = [len(example[2]) for example in meta]
sampler = PartialyRandomizedSimilarTimeLengthSampler(text_lengths,
batch_size)
# some hyperparameters affect how we process data, so create a data collector!
model_config = config["model"]
downsample_factor = model_config["downsample_factor"]
r = model_config["outputs_per_step"]
collector = DataCollector(downsample_factor=downsample_factor, r=r)
ljspeech_loader = DataCargo(
ljspeech, batch_fn=collector, batch_size=batch_size, sampler=sampler)
# =========================model=========================
if args.device == -1:
place = fluid.CPUPlace()
else:
place = fluid.CUDAPlace(args.device)
with dg.guard(place):
# =========================model=========================
n_speakers = model_config["n_speakers"]
speaker_dim = model_config["speaker_embed_dim"]
speaker_embed_std = model_config["speaker_embedding_weight_std"]
n_vocab = en.n_vocab
embed_dim = model_config["text_embed_dim"]
linear_dim = 1 + n_fft // 2
parser.add_argument("--input",
type=str,
required=True,
help="data path of the original data")
args = parser.parse_args()
with open(args.config, 'rt') as f:
config = yaml.safe_load(f)
print("========= Command Line Arguments ========")
for k, v in vars(args).items():
print("{}: {}".format(k, v))
print("=========== Configurations ==============")
for k in ["p_pronunciation", "batch_size"]:
print("{}: {}".format(k, config[k]))
ljspeech = LJSpeech(args.input)
collate_fn = DataCollector(config["p_pronunciation"])
dg.enable_dygraph(fluid.CPUPlace())
sampler = PartialyRandomizedSimilarTimeLengthSampler(ljspeech.num_frames())
cargo = DataCargo(ljspeech,
collate_fn,
batch_size=config["batch_size"],
sampler=sampler)
loader = DataLoader\
.from_generator(capacity=5, return_list=True)\
.set_batch_generator(cargo)
for i, batch in tqdm.tqdm(enumerate(loader)):
continue