def test_fastspeech_resize_positional_embeddings(new_size):
config = FastSpeechConfig()
fastspeech = TFFastSpeech(config, name="fastspeech")
fastspeech._build()
fastspeech.save_weights("./test.h5")
fastspeech.resize_positional_embeddings(new_size)
fastspeech.load_weights("./test.h5", by_name=True, skip_mismatch=True)
def test_fastspeech_trainable(num_hidden_layers, n_speakers):
config = FastSpeechConfig(num_hidden_layers=num_hidden_layers, n_speakers=n_speakers)
fastspeech = TFFastSpeech(config, name='fastspeech')
optimizer = tf.keras.optimizers.Adam(lr=0.001)
# fake inputs
input_ids = tf.convert_to_tensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]], tf.int32)
attention_mask = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.int32)
speaker_ids = tf.convert_to_tensor([0], tf.int32)
duration_gts = tf.convert_to_tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], tf.int32)
mel_gts = tf.random.uniform(shape=[1, 10, 80], dtype=tf.float32)
@tf.function
def one_step_training():
with tf.GradientTape() as tape:
mel_outputs_before, _, duration_outputs = fastspeech(
input_ids, attention_mask, speaker_ids, duration_gts, training=True)
duration_loss = tf.keras.losses.MeanSquaredError()(duration_gts, duration_outputs)
mel_loss = tf.keras.losses.MeanSquaredError()(mel_gts, mel_outputs_before)
loss = duration_loss + mel_loss
gradients = tape.gradient(loss, fastspeech.trainable_variables)
optimizer.apply_gradients(zip(gradients, fastspeech.trainable_variables))
tf.print(loss)
import time
for i in range(2):
if i == 1:
start = time.time()
one_step_training()
print(time.time() - start)
def get_model():
with open( get_weight_path('fastspeech_config.yml') ) as f:
config = yaml.load(f, Loader=yaml.Loader)
config = FastSpeechConfig(**config['fastspeech_params'])
fastspeech = TFFastSpeech(config=config, name='fastspeech')
fastspeech._build()
fastspeech.load_weights( get_weight_path('fastspeech-150k.h5') )
return fastspeech
def main():
"""Run training process."""
parser = argparse.ArgumentParser(
description="Train FastSpeech (See detail in tensorflow_tts/bin/train-fastspeech.py)"
)
parser.add_argument(
"--train-dir",
default=None,
type=str,
help="directory including training data. ",
)
parser.add_argument(
"--dev-dir",
default=None,
type=str,
help="directory including development data. ",
)
parser.add_argument(
"--use-norm", default=1, type=int, help="usr norm-mels for train or raw."
)
parser.add_argument(
"--outdir", type=str, required=True, help="directory to save checkpoints."
)
parser.add_argument(
"--config", type=str, required=True, help="yaml format configuration file."
)
parser.add_argument(
"--resume",
default="",
type=str,
nargs="?",
help='checkpoint file path to resume training. (default="")',
)
parser.add_argument(
"--verbose",
type=int,
default=1,
help="logging level. higher is more logging. (default=1)",
)
parser.add_argument(
"--mixed_precision",
default=0,
type=int,
help="using mixed precision for generator or not.",
)
args = parser.parse_args()
# set mixed precision config
if args.mixed_precision == 1:
tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
args.mixed_precision = bool(args.mixed_precision)
args.use_norm = bool(args.use_norm)
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG,
stream=sys.stdout,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO,
stream=sys.stdout,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
else:
logging.basicConfig(
level=logging.WARN,
stream=sys.stdout,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
logging.warning("Skip DEBUG/INFO messages")
# check directory existence
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
# check arguments
if args.train_dir is None:
raise ValueError("Please specify --train-dir")
if args.dev_dir is None:
raise ValueError("Please specify --valid-dir")
# load and save config
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
config.update(vars(args))
config["version"] = tensorflow_tts.__version__
with open(os.path.join(args.outdir, "config.yml"), "w") as f:
yaml.dump(config, f, Dumper=yaml.Dumper)
for key, value in config.items():
logging.info(f"{key} = {value}")
# get dataset
if config["remove_short_samples"]:
mel_length_threshold = config["mel_length_threshold"]
else:
mel_length_threshold = None
if config["format"] == "npy":
charactor_query = "*-ids.npy"
mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy"
duration_query = "*-durations.npy"
charactor_load_fn = np.load
mel_load_fn = np.load
duration_load_fn = np.load
else:
raise ValueError("Only npy are supported.")
# define train/valid dataset
train_dataset = CharactorDurationMelDataset(
root_dir=args.train_dir,
charactor_query=charactor_query,
mel_query=mel_query,
duration_query=duration_query,
charactor_load_fn=charactor_load_fn,
mel_load_fn=mel_load_fn,
duration_load_fn=duration_load_fn,
mel_length_threshold=mel_length_threshold,
return_utt_id=False,
).create(
is_shuffle=config["is_shuffle"],
allow_cache=config["allow_cache"],
batch_size=config["batch_size"],
)
valid_dataset = CharactorDurationMelDataset(
root_dir=args.dev_dir,
charactor_query=charactor_query,
mel_query=mel_query,
duration_query=duration_query,
charactor_load_fn=charactor_load_fn,
mel_load_fn=mel_load_fn,
duration_load_fn=duration_load_fn,
mel_length_threshold=None,
return_utt_id=False,
).create(
is_shuffle=config["is_shuffle"],
allow_cache=config["allow_cache"],
batch_size=config["batch_size"],
)
fastspeech = TFFastSpeech(
config=FASTSPEECH_CONFIG.FastSpeechConfig(**config["fastspeech_params"])
)
fastspeech._build()
fastspeech.summary()
# define trainer
trainer = FastSpeechTrainer(
config=config, steps=0, epochs=0, is_mixed_precision=False
)
# AdamW for fastspeech
learning_rate_fn = tf.keras.optimizers.schedules.PolynomialDecay(
initial_learning_rate=config["optimizer_params"]["initial_learning_rate"],
decay_steps=config["optimizer_params"]["decay_steps"],
end_learning_rate=config["optimizer_params"]["end_learning_rate"],
)
learning_rate_fn = WarmUp(
initial_learning_rate=config["optimizer_params"]["initial_learning_rate"],
decay_schedule_fn=learning_rate_fn,
warmup_steps=int(
config["train_max_steps"] * config["optimizer_params"]["warmup_proportion"]
),
)
optimizer = AdamWeightDecay(
learning_rate=learning_rate_fn,
weight_decay_rate=config["optimizer_params"]["weight_decay"],
beta_1=0.9,
beta_2=0.98,
epsilon=1e-6,
exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"],
)
# compile trainer
trainer.compile(model=fastspeech, optimizer=optimizer)
# start training
try:
trainer.fit(
train_dataset,
valid_dataset,
saved_path=os.path.join(config["outdir"], "checkpoints/"),
resume=args.resume,
)
except KeyboardInterrupt:
trainer.save_checkpoint()
logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.")
def main():
"""Run fastspeech decoding from folder."""
parser = argparse.ArgumentParser(
description=
"Decode soft-mel features from charactor with trained FastSpeech "
"(See detail in examples/fastspeech/decode_fastspeech.py).")
parser.add_argument(
"--rootdir",
default=None,
type=str,
required=True,
help="directory including ids/durations files.",
)
parser.add_argument("--outdir",
type=str,
required=True,
help="directory to save generated speech.")
parser.add_argument("--checkpoint",
type=str,
required=True,
help="checkpoint file to be loaded.")
parser.add_argument(
"--config",
default=None,
type=str,
required=True,
help="yaml format configuration file. if not explicitly provided, "
"it will be searched in the checkpoint directory. (default=None)",
)
parser.add_argument(
"--batch-size",
default=8,
type=int,
required=False,
help="Batch size for inference.",
)
parser.add_argument(
"--verbose",
type=int,
default=1,
help="logging level. higher is more logging. (default=1)",
)
args = parser.parse_args()
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
else:
logging.basicConfig(
level=logging.WARN,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
logging.warning("Skip DEBUG/INFO messages")
# check directory existence
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
# load config
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
config.update(vars(args))
if config["format"] == "npy":
char_query = "*-ids.npy"
char_load_fn = np.load
else:
raise ValueError("Only npy is supported.")
# define data-loader
dataset = CharactorDataset(
root_dir=args.rootdir,
charactor_query=char_query,
charactor_load_fn=char_load_fn,
)
dataset = dataset.create(batch_size=args.batch_size)
# define model and load checkpoint
fastspeech = TFFastSpeech(
config=FastSpeechConfig(**config["fastspeech_params"]),
name="fastspeech")
fastspeech._build()
fastspeech.load_weights(args.checkpoint)
for data in tqdm(dataset, desc="Decoding"):
utt_ids = data["utt_ids"]
char_ids = data["input_ids"]
# fastspeech inference.
masked_mel_before, masked_mel_after, duration_outputs = fastspeech.inference(
char_ids,
speaker_ids=tf.zeros(shape=[tf.shape(char_ids)[0]],
dtype=tf.int32),
speed_ratios=tf.ones(shape=[tf.shape(char_ids)[0]],
dtype=tf.float32),
)
# convert to numpy
masked_mel_befores = masked_mel_before.numpy()
masked_mel_afters = masked_mel_after.numpy()
for (utt_id, mel_before, mel_after,
durations) in zip(utt_ids, masked_mel_befores, masked_mel_afters,
duration_outputs):
# real len of mel predicted
real_length = durations.numpy().sum()
utt_id = utt_id.numpy().decode("utf-8")
# save to folder.
np.save(
os.path.join(args.outdir, f"{utt_id}-fs-before-feats.npy"),
mel_before[:real_length, :].astype(np.float32),
allow_pickle=False,
)
np.save(
os.path.join(args.outdir, f"{utt_id}-fs-after-feats.npy"),
mel_after[:real_length, :].astype(np.float32),
allow_pickle=False,
)
