testadmin
/
Parakeet
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

format all the code with yapf

This commit is contained in:
iclementine 2020-12-20 13:15:07 +08:00
parent c866bb0b57
commit e03e96d9e4
56 changed files with 1252 additions and 569 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
README_cn.md
View File

@ -228,6 +228,6 @@ Parakeet 同时提供了示例模型的训练好的参数，可从下表中获
正在开发中。
## 版权和许可
## 版权和许可
Parakeet 以 [Apache-2.0 license](LICENSE) 提供。

19
doc/source/conf.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Configuration file for the Sphinx documentation builder.
#
# This file only contains a selection of the most common options. For a full
@ -14,7 +28,6 @@
# import sys
# sys.path.insert(0, os.path.abspath('.'))
# -- Project information -----------------------------------------------------
project = 'parakeet'
@ -24,7 +37,6 @@ author = 'parakeet-developers'
# The full version, including alpha/beta/rc tags
release = '0.2'
# -- General configuration ---------------------------------------------------
# Add any Sphinx extension module names here, as strings. They can be
@ -33,7 +45,7 @@ release = '0.2'
extensions = [
'sphinx.ext.autodoc',
'sphinx.ext.viewcode',
"sphinx_rtd_theme",
"sphinx_rtd_theme",
'sphinx.ext.mathjax',
'numpydoc',
]
@ -46,7 +58,6 @@ templates_path = ['_templates']
# This pattern also affects html_static_path and html_extra_path.
exclude_patterns = []
# -- Options for HTML output -------------------------------------------------
# The theme to use for HTML and HTML Help pages. See the documentation for

10
docs/config_cn.md
View File

@ -18,7 +18,7 @@
常见的配置文件的格式有 `ini`, `yaml`, `toml`, `json` 等。
`ini`
`ini`
优点：简单，支持字符串插值等操作。
缺点：仅支持两层结构，值不带类型信息，解析的时候需要手动 cast。
@ -102,11 +102,3 @@ optional arguments:
--opts ... options to overwrite --config file and the default
config, passing in KEY VALUE pairs
```

14
docs/data_cn.md
View File

@ -21,7 +21,7 @@
一般来说，我们将一个 Dataset 的子类看作是数据集和实验的具体需求之间的适配器。
parakeet 还提供了若干个高阶的 Dataset 类，用于从已有的 Dataset 产生新的 Dataset.
parakeet 还提供了若干个高阶的 Dataset 类，用于从已有的 Dataset 产生新的 Dataset.
1. 用于字段组合的有 TupleDataset, DictDataset;
2. 用于数据集切分合并的有 SliceDataset, SubsetDataset, ChainDataset;
@ -137,7 +137,7 @@ class Transform(object):
self.processor = AudioProcessor(
sample_rate=22050,
n_fft=1024,
win_length=1024,
win_length=1024,
hop_length=256,
f_max=8000)
self.normalizer = LogMagnitude()
@ -167,7 +167,7 @@ ljspeech = TransformDataset(meta, transform)
当然也可以选择专门写一个转换脚本把转换后的数据集保存下来，然后再写一个适配的 Dataset 子类去加载这些保存的数据。实际这么做的效率会更高。
接下来我们需要写一个可调用对象将多个样例组成批次。因为其中的 ids 和 mel 频谱是序列数据，所以我们需要进行 padding.
接下来我们需要写一个可调用对象将多个样例组成批次。因为其中的 ids 和 mel 频谱是序列数据，所以我们需要进行 padding.
```python
class LJSpeechCollector(object):
@ -197,10 +197,10 @@ def create_dataloader(source_path, valid_size, batch_size):
valid_set, train_set = dataset.split(lj, valid_size)
train_loader = DataLoader(
train_set,
return_list=False,
batch_size=batch_size,
shuffle=True,
train_set,
return_list=False,
batch_size=batch_size,
shuffle=True,
drop_last=True,
collate_fn=LJSpeechCollector())
valid_loader = DataLoader(

2
docs/experiment_cn.md
View File

@ -72,4 +72,4 @@ def train(self):
```python
exp.run()
```
```

2
docs/experiment_guide_cn.md
View File

@ -72,5 +72,3 @@ Dataset --(transform)--> Dataset --+
```
在这个软件源中包含了几个例子，可以在 [Parakeet/examples](../examples) 中查看。这些实验被作为样例提供给用户，可以直接运行。同时也欢迎用户添加新的模型和实验并为 `Parakeet` 贡献代码。

2
docs/installation_cn.md
View File

@ -31,7 +31,7 @@ python -m pip install paddlepaddle==2.0.0rc0 -i https://mirror.baidu.com/pypi/si
# ubuntu, debian
sudo apt-get install libsndfile1
# centos, fedora,
# centos, fedora,
sudo yum install libsndfile
# openSUSE

7
docs/overview_cn.md
View File

@ -9,10 +9,3 @@ Parakeet 为用户和开发者提供了
1. 可复用的模型以及常用的模块；
2. 从数据处理，模型训练到预测等一系列过程的完整实验；
3. 高质量的开箱即用模型。

56
examples/transformer_tts/config.py
View File

@ -1,21 +1,34 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from yacs.config import CfgNode as CN
_C = CN()
_C.data = CN(
dict(
batch_size=16, # batch size
valid_size=64, # the first N examples are reserved for validation
sample_rate=22050, # Hz, sample rate
n_fft=1024, # fft frame size
win_length=1024, # window size
batch_size=16, # batch size
valid_size=64, # the first N examples are reserved for validation
sample_rate=22050, # Hz, sample rate
n_fft=1024, # fft frame size
win_length=1024, # window size
hop_length=256, # hop size between ajacent frame
f_max=8000, # Hz, max frequency when converting to mel
f_max=8000, # Hz, max frequency when converting to mel
d_mel=80, # mel bands
padding_idx=0, # text embedding's padding index
mel_start_value=0.5, # value for starting frame
mel_end_value=-0.5, # # value for ending frame
)
)
padding_idx=0, # text embedding's padding index
mel_start_value=0.5, # value for starting frame
mel_end_value=-0.5, # # value for ending frame
))
_C.model = CN(
dict(
@ -31,22 +44,21 @@ _C.model = CN(
postnet_kernel_size=5, # decoder postnet(cnn)'s kernel size
max_reduction_factor=10, # max_reduction factor
dropout=0.1, # global droput probability
stop_loss_scale=8.0, # scaler for stop _loss
decoder_prenet_dropout=0.5, # decoder prenet dropout probability
)
)
stop_loss_scale=8.0, # scaler for stop _loss
decoder_prenet_dropout=0.5, # decoder prenet dropout probability
))
_C.training = CN(
dict(
lr=1e-4, # learning rate
lr=1e-4, # learning rate
drop_n_heads=[[0, 0], [15000, 1]],
reduction_factor=[[0, 10], [80000, 4], [200000, 2]],
plot_interval=1000, # plot attention and spectrogram
valid_interval=1000, # validation
save_interval=10000, # checkpoint
max_iteration=900000, # max iteration to train
)
)
plot_interval=1000, # plot attention and spectrogram
valid_interval=1000, # validation
save_interval=10000, # checkpoint
max_iteration=900000, # max iteration to train
))
def get_cfg_defaults():
"""Get a yacs CfgNode object with default values for my_project."""

37
examples/transformer_tts/ljspeech.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
from pathlib import Path
import pickle
@ -7,8 +21,10 @@ from paddle.io import Dataset, DataLoader
from parakeet.data.batch import batch_spec, batch_text_id
from parakeet.data import dataset
class LJSpeech(Dataset):
"""A simple dataset adaptor for the processed ljspeech dataset."""
def __init__(self, root):
self.root = Path(root).expanduser()
records = []
@ -35,13 +51,13 @@ class Transform(object):
self.end_value = end_value
def __call__(self, example):
ids, mel = example # ids already have <s> and </s>
ids, mel = example # ids already have <s> and </s>
ids = np.array(ids, dtype=np.int64)
# add start and end frame
mel = np.pad(mel,
[(0, 0), (1, 1)],
mode='constant',
constant_values=[(0, 0), (self.start_value, self.end_value)])
mel = np.pad(
mel, [(0, 0), (1, 1)],
mode='constant',
constant_values=[(0, 0), (self.start_value, self.end_value)])
stop_labels = np.ones([mel.shape[1]], dtype=np.int64)
stop_labels[-1] = 2
# actually this thing can also be done within the model
@ -50,6 +66,7 @@ class Transform(object):
class LJSpeechCollector(object):
"""A simple callable to batch LJSpeech examples."""
def __init__(self, padding_idx=0, padding_value=0.):
self.padding_idx = padding_idx
self.padding_value = padding_value
@ -67,15 +84,16 @@ class LJSpeechCollector(object):
def create_dataloader(config, source_path):
lj = LJSpeech(source_path)
transform = Transform(config.data.mel_start_value, config.data.mel_end_value)
transform = Transform(config.data.mel_start_value,
config.data.mel_end_value)
lj = dataset.TransformDataset(lj, transform)
valid_set, train_set = dataset.split(lj, config.data.valid_size)
data_collator = LJSpeechCollector(padding_idx=config.data.padding_idx)
train_loader = DataLoader(
train_set,
batch_size=config.data.batch_size,
shuffle=True,
train_set,
batch_size=config.data.batch_size,
shuffle=True,
drop_last=True,
collate_fn=data_collator)
valid_loader = DataLoader(
@ -85,4 +103,3 @@ def create_dataloader(config, source_path):
drop_last=False,
collate_fn=data_collator)
return train_loader, valid_loader

50
examples/transformer_tts/preprocess.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import tqdm
import pickle
@ -11,6 +25,7 @@ from parakeet.frontend import English
from config import get_cfg_defaults
def create_dataset(config, source_path, target_path, verbose=False):
# create output dir
target_path = Path(target_path).expanduser()
@ -23,11 +38,11 @@ def create_dataset(config, source_path, target_path, verbose=False):
sample_rate=config.data.sample_rate,
n_fft=config.data.n_fft,
n_mels=config.data.d_mel,
win_length=config.data.win_length,
win_length=config.data.win_length,
hop_length=config.data.hop_length,
f_max=config.data.f_max)
normalizer = LogMagnitude()
records = []
for (fname, text, _) in tqdm.tqdm(meta_data):
wav = processor.read_wav(fname)
@ -42,12 +57,13 @@ def create_dataset(config, source_path, target_path, verbose=False):
np.save(mel_path / mel_name, mel)
if verbose:
print("save mel spectrograms into {}".format(mel_path))
# save meta data as pickle archive
with open(target_path / "metadata.pkl", 'wb') as f:
pickle.dump(records, f)
if verbose:
print("saved metadata into {}".format(target_path / "metadata.pkl"))
print("saved metadata into {}".format(target_path /
"metadata.pkl"))
# also save meta data into text format for inspection
with open(target_path / "metadata.txt", 'wt') as f:
@ -55,21 +71,31 @@ def create_dataset(config, source_path, target_path, verbose=False):
phoneme_str = "|".join(phonemes)
f.write("{}\t{}\t{}\n".format(mel_name, text, phoneme_str))
if verbose:
print("saved metadata into {}".format(target_path / "metadata.txt"))
print("saved metadata into {}".format(target_path /
"metadata.txt"))
print("Done.")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="create dataset")
parser.add_argument("--config", type=str, metavar="FILE", help="extra config to overwrite the default config")
parser.add_argument("--input", type=str, help="path of the ljspeech dataset")
parser.add_argument("--output", type=str, help="path to save output dataset")
parser.add_argument("--opts", nargs=argparse.REMAINDER,
parser.add_argument(
"--config",
type=str,
metavar="FILE",
help="extra config to overwrite the default config")
parser.add_argument(
"--input", type=str, help="path of the ljspeech dataset")
parser.add_argument(
"--output", type=str, help="path to save output dataset")
parser.add_argument(
"--opts",
nargs=argparse.REMAINDER,
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs"
)
parser.add_argument("-v", "--verbose", action="store_true", help="print msg")
parser.add_argument(
"-v", "--verbose", action="store_true", help="print msg")
config = get_cfg_defaults()
args = parser.parse_args()
if args.config:

55
examples/transformer_tts/synthesize.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import time
from pathlib import Path
@ -13,21 +27,22 @@ from parakeet.utils.display import add_attention_plots
from config import get_cfg_defaults
@paddle.fluid.dygraph.no_grad
def main(config, args):
paddle.set_device(args.device)
# model
frontend = English()
model = TransformerTTS.from_pretrained(
frontend, config, args.checkpoint_path)
model = TransformerTTS.from_pretrained(frontend, config,
args.checkpoint_path)
model.eval()
# inputs
input_path = Path(args.input).expanduser()
with open(input_path, "rt") as f:
with open(input_path, "rt") as f:
sentences = f.readlines()
output_dir = Path(args.output).expanduser()
output_dir.mkdir(parents=True, exist_ok=True)
@ -38,22 +53,36 @@ def main(config, args):
mel_output = mel_output.T #(C, T)
np.save(str(output_dir / f"sentence_{i}"), mel_output)
if args.verbose:
print("spectrogram saved at {}".format(output_dir / f"sentence_{i}.npy"))
print("spectrogram saved at {}".format(output_dir /
f"sentence_{i}.npy"))
if __name__ == "__main__":
config = get_cfg_defaults()
parser = argparse.ArgumentParser(description="generate mel spectrogram with TransformerTTS.")
parser.add_argument("--config", type=str, metavar="FILE", help="extra config to overwrite the default config")
parser.add_argument("--checkpoint_path", type=str, help="path of the checkpoint to load.")
parser = argparse.ArgumentParser(
description="generate mel spectrogram with TransformerTTS.")
parser.add_argument(
"--config",
type=str,
metavar="FILE",
help="extra config to overwrite the default config")
parser.add_argument(
"--checkpoint_path", type=str, help="path of the checkpoint to load.")
parser.add_argument("--input", type=str, help="path of the text sentences")
parser.add_argument("--output", type=str, help="path to save outputs")
parser.add_argument("--device", type=str, default="cpu", help="device type to use.")
parser.add_argument("--opts", nargs=argparse.REMAINDER, help="options to overwrite --config file and the default config, passing in KEY VALUE pairs")
parser.add_argument("-v", "--verbose", action="store_true", help="print msg")
parser.add_argument(
"--device", type=str, default="cpu", help="device type to use.")
parser.add_argument(
"--opts",
nargs=argparse.REMAINDER,
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs"
)
parser.add_argument(
"-v", "--verbose", action="store_true", help="print msg")
args = parser.parse_args()
if args.config:
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)

74
examples/transformer_tts/train.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import time
import logging
from pathlib import Path
@ -19,12 +33,13 @@ from parakeet.training.experiment import ExperimentBase
from config import get_cfg_defaults
from ljspeech import LJSpeech, LJSpeechCollector, Transform
class Experiment(ExperimentBase):
def setup_model(self):
config = self.config
frontend = English()
model = TransformerTTS(
frontend,
frontend,
d_encoder=config.model.d_encoder,
d_decoder=config.model.d_decoder,
d_mel=config.data.d_mel,
@ -46,8 +61,7 @@ class Experiment(ExperimentBase):
beta1=0.9,
beta2=0.98,
epsilon=1e-9,
parameters=model.parameters()
)
parameters=model.parameters())
criterion = TransformerTTSLoss(config.model.stop_loss_scale)
drop_n_heads = scheduler.StepWise(config.training.drop_n_heads)
reduction_factor = scheduler.StepWise(config.training.reduction_factor)
@ -63,21 +77,24 @@ class Experiment(ExperimentBase):
config = self.config
ljspeech_dataset = LJSpeech(args.data)
transform = Transform(config.data.mel_start_value, config.data.mel_end_value)
ljspeech_dataset = dataset.TransformDataset(ljspeech_dataset, transform)
valid_set, train_set = dataset.split(ljspeech_dataset, config.data.valid_size)
transform = Transform(config.data.mel_start_value,
config.data.mel_end_value)
ljspeech_dataset = dataset.TransformDataset(ljspeech_dataset,
transform)
valid_set, train_set = dataset.split(ljspeech_dataset,
config.data.valid_size)
batch_fn = LJSpeechCollector(padding_idx=config.data.padding_idx)
if not self.parallel:
train_loader = DataLoader(
train_set,
batch_size=config.data.batch_size,
shuffle=True,
train_set,
batch_size=config.data.batch_size,
shuffle=True,
drop_last=True,
collate_fn=batch_fn)
else:
sampler = DistributedBatchSampler(
train_set,
train_set,
batch_size=config.data.batch_size,
num_replicas=dist.get_world_size(),
rank=dist.get_rank(),
@ -95,11 +112,11 @@ class Experiment(ExperimentBase):
def compute_outputs(self, text, mel, stop_label):
model_core = self.model._layers if self.parallel else self.model
model_core.set_constants(
self.reduction_factor(self.iteration),
self.reduction_factor(self.iteration),
self.drop_n_heads(self.iteration))
# TODO(chenfeiyu): we can combine these 2 slices
mel_input = mel[:,:-1, :]
mel_input = mel[:, :-1, :]
reduced_mel_input = mel_input[:, ::model_core.r, :]
outputs = self.model(text, reduced_mel_input)
return outputs
@ -115,11 +132,8 @@ class Experiment(ExperimentBase):
time_steps = mel_target.shape[1]
losses = self.criterion(
mel_output[:,:time_steps, :],
mel_intermediate[:,:time_steps, :],
mel_target,
stop_logits[:,:time_steps, :],
stop_label_target)
mel_output[:, :time_steps, :], mel_intermediate[:, :time_steps, :],
mel_target, stop_logits[:, :time_steps, :], stop_label_target)
return losses
def train_batch(self):
@ -133,7 +147,7 @@ class Experiment(ExperimentBase):
outputs = self.compute_outputs(text, mel, stop_label)
losses = self.compute_losses(batch, outputs)
loss = losses["loss"]
loss.backward()
loss.backward()
self.optimizer.step()
iteration_time = time.time() - start
@ -141,14 +155,17 @@ class Experiment(ExperimentBase):
# logging
msg = "Rank: {}, ".format(dist.get_rank())
msg += "step: {}, ".format(self.iteration)
msg += "time: {:>.3f}s/{:>.3f}s, ".format(data_loader_time, iteration_time)
msg += ', '.join('{}: {:>.6f}'.format(k, v) for k, v in losses_np.items())
msg += "time: {:>.3f}s/{:>.3f}s, ".format(data_loader_time,
iteration_time)
msg += ', '.join('{}: {:>.6f}'.format(k, v)
for k, v in losses_np.items())
self.logger.info(msg)
if dist.get_rank() == 0:
for k, v in losses_np.items():
self.visualizer.add_scalar(f"train_loss/{k}", v, self.iteration)
self.visualizer.add_scalar(f"train_loss/{k}", v,
self.iteration)
@mp_tools.rank_zero_only
@paddle.no_grad()
def valid(self):
@ -163,10 +180,9 @@ class Experiment(ExperimentBase):
if i < 2:
attention_weights = outputs["cross_attention_weights"]
display.add_multi_attention_plots(
self.visualizer,
f"valid_sentence_{i}_cross_attention_weights",
attention_weights,
self.iteration)
self.visualizer,
f"valid_sentence_{i}_cross_attention_weights",
attention_weights, self.iteration)
# write visual log
valid_losses = {k: np.mean(v) for k, v in valid_losses.items()}
@ -191,7 +207,7 @@ if __name__ == "__main__":
config = get_cfg_defaults()
parser = default_argument_parser()
args = parser.parse_args()
if args.config:
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)

58
examples/waveflow/config.py
View File

@ -1,40 +1,52 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from yacs.config import CfgNode as CN
_C = CN()
_C.data = CN(
dict(
batch_size=8, # batch size
valid_size=16, # the first N examples are reserved for validation
sample_rate=22050, # Hz, sample rate
n_fft=1024, # fft frame size
win_length=1024, # window size
batch_size=8, # batch size
valid_size=16, # the first N examples are reserved for validation
sample_rate=22050, # Hz, sample rate
n_fft=1024, # fft frame size
win_length=1024, # window size
hop_length=256, # hop size between ajacent frame
f_max=8000, # Hz, max frequency when converting to mel
f_max=8000, # Hz, max frequency when converting to mel
n_mels=80, # mel bands
clip_frames=65, # mel clip frames
)
)
clip_frames=65, # mel clip frames
))
_C.model = CN(
dict(
upsample_factors=[16, 16],
n_flows=8, # number of flows in WaveFlow
n_layers=8, # number of conv block in each flow
n_group=16, # folding factor of audio and spectrogram
channels=128, # resiaudal channel in each flow
kernel_size=[3, 3], # kernel size in each conv block
sigma=1.0, # stddev of the random noise
)
)
n_flows=8, # number of flows in WaveFlow
n_layers=8, # number of conv block in each flow
n_group=16, # folding factor of audio and spectrogram
channels=128, # resiaudal channel in each flow
kernel_size=[3, 3], # kernel size in each conv block
sigma=1.0, # stddev of the random noise
))
_C.training = CN(
dict(
lr=2e-4, # learning rates
valid_interval=1000, # validation
save_interval=10000, # checkpoint
max_iteration=3000000, # max iteration to train
)
)
lr=2e-4, # learning rates
valid_interval=1000, # validation
save_interval=10000, # checkpoint
max_iteration=3000000, # max iteration to train
))
def get_cfg_defaults():
"""Get a yacs CfgNode object with default values for my_project."""

36
examples/waveflow/ljspeech.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
from pathlib import Path
import pickle
@ -9,19 +23,20 @@ from parakeet.data.batch import batch_spec, batch_wav
from parakeet.data import dataset
from parakeet.audio import AudioProcessor
class LJSpeech(Dataset):
"""A simple dataset adaptor for the processed ljspeech dataset."""
def __init__(self, root):
self.root = Path(root).expanduser()
meta_data = pandas.read_csv(
str(self.root / "metadata.csv"),
sep="\t",
header=None,
names=["fname", "frames", "samples"]
)
names=["fname", "frames", "samples"])
records = []
for row in meta_data.itertuples() :
for row in meta_data.itertuples():
mel_path = str(self.root / "mel" / (row.fname + ".npy"))
wav_path = str(self.root / "wav" / (row.fname + ".npy"))
records.append((mel_path, wav_path))
@ -39,6 +54,7 @@ class LJSpeech(Dataset):
class LJSpeechCollector(object):
"""A simple callable to batch LJSpeech examples."""
def __init__(self, padding_value=0.):
self.padding_value = padding_value
@ -52,9 +68,9 @@ class LJSpeechCollector(object):
class LJSpeechClipCollector(object):
def __init__(self, clip_frames=65, hop_length=256):
self.clip_frames = clip_frames
self.clip_frames = clip_frames
self.hop_length = hop_length
def __call__(self, examples):
mels = []
wavs = []
@ -70,9 +86,7 @@ class LJSpeechClipCollector(object):
mel, wav = example
frames = mel.shape[-1]
start = np.random.randint(0, frames - self.clip_frames)
mel_clip = mel[:, start: start + self.clip_frames]
wav_clip = wav[start * self.hop_length: (start + self.clip_frames) * self.hop_length]
mel_clip = mel[:, start:start + self.clip_frames]
wav_clip = wav[start * self.hop_length:(start + self.clip_frames) *
self.hop_length]
return mel_clip, wav_clip

52
examples/waveflow/preprocess.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import tqdm
import csv
@ -86,12 +100,8 @@ def create_dataset(config, input_dir, output_dir, verbose=True):
output_dir = Path(output_dir).expanduser()
output_dir.mkdir(exist_ok=True)
transform = Transform(
config.sample_rate,
config.n_fft,
config.win_length,
config.hop_length,
config.n_mels)
transform = Transform(config.sample_rate, config.n_fft, config.win_length,
config.hop_length, config.n_mels)
file_names = []
for example in tqdm.tqdm(dataset):
@ -107,23 +117,35 @@ def create_dataset(config, input_dir, output_dir, verbose=True):
np.save(str(mel_dir / base_name), mel)
file_names.append((base_name, mel.shape[-1], audio.shape[-1]))
meta_data = pd.DataFrame.from_records(file_names)
meta_data.to_csv(str(output_dir / "metadata.csv"), sep="\t", index=None, header=None)
print("saved meta data in to {}".format(os.path.join(output_dir, "metadata.csv")))
meta_data.to_csv(
str(output_dir / "metadata.csv"), sep="\t", index=None, header=None)
print("saved meta data in to {}".format(
os.path.join(output_dir, "metadata.csv")))
print("Done!")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="create dataset")
parser.add_argument("--config", type=str, metavar="FILE", help="extra config to overwrite the default config")
parser.add_argument("--input", type=str, help="path of the ljspeech dataset")
parser.add_argument("--output", type=str, help="path to save output dataset")
parser.add_argument("--opts", nargs=argparse.REMAINDER,
parser.add_argument(
"--config",
type=str,
metavar="FILE",
help="extra config to overwrite the default config")
parser.add_argument(
"--input", type=str, help="path of the ljspeech dataset")
parser.add_argument(
"--output", type=str, help="path to save output dataset")
parser.add_argument(
"--opts",
nargs=argparse.REMAINDER,
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs"
)
parser.add_argument("-v", "--verbose", action="store_true", help="print msg")
parser.add_argument(
"-v", "--verbose", action="store_true", help="print msg")
config = get_cfg_defaults()
args = parser.parse_args()
if args.config:

54
examples/waveflow/synthesize.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import numpy as np
import soundfile as sf
@ -8,9 +22,9 @@ import parakeet
from parakeet.models.waveflow import UpsampleNet, WaveFlow, ConditionalWaveFlow
from parakeet.utils import layer_tools, checkpoint
from config import get_cfg_defaults
def main(config, args):
paddle.set_device(args.device)
model = ConditionalWaveFlow.from_pretrained(config, args.checkpoint_path)
@ -23,7 +37,8 @@ def main(config, args):
for file_path in mel_dir.iterdir():
mel = np.load(str(file_path))
audio = model.predict(mel)
audio_path = output_dir / (os.path.splitext(file_path.name)[0] + ".wav")
audio_path = output_dir / (
os.path.splitext(file_path.name)[0] + ".wav")
sf.write(audio_path, audio, config.data.sample_rate)
print("[synthesize] {} -> {}".format(file_path, audio_path))
@ -31,17 +46,32 @@ def main(config, args):
if __name__ == "__main__":
config = get_cfg_defaults()
parser = argparse.ArgumentParser(description="generate mel spectrogram with TransformerTTS.")
parser.add_argument("--config", type=str, metavar="FILE", help="extra config to overwrite the default config")
parser.add_argument("--checkpoint_path", type=str, help="path of the checkpoint to load.")
parser.add_argument("--input", type=str, help="path of directory containing mel spectrogram (in .npy format)")
parser = argparse.ArgumentParser(
description="generate mel spectrogram with TransformerTTS.")
parser.add_argument(
"--config",
type=str,
metavar="FILE",
help="extra config to overwrite the default config")
parser.add_argument(
"--checkpoint_path", type=str, help="path of the checkpoint to load.")
parser.add_argument(
"--input",
type=str,
help="path of directory containing mel spectrogram (in .npy format)")
parser.add_argument("--output", type=str, help="path to save outputs")
parser.add_argument("--device", type=str, default="cpu", help="device type to use.")
parser.add_argument("--opts", nargs=argparse.REMAINDER, help="options to overwrite --config file and the default config, passing in KEY VALUE pairs")
parser.add_argument("-v", "--verbose", action="store_true", help="print msg")
parser.add_argument(
"--device", type=str, default="cpu", help="device type to use.")
parser.add_argument(
"--opts",
nargs=argparse.REMAINDER,
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs"
)
parser.add_argument(
"-v", "--verbose", action="store_true", help="print msg")
args = parser.parse_args()
if args.config:
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)
@ -49,4 +79,4 @@ if __name__ == "__main__":
print(config)
print(args)
main(config, args)
main(config, args)

48
examples/waveflow/train.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import time
from pathlib import Path
import numpy as np
@ -34,7 +48,8 @@ class Experiment(ExperimentBase):
if self.parallel > 1:
model = paddle.DataParallel(model)
optimizer = paddle.optimizer.Adam(config.training.lr, parameters=model.parameters())
optimizer = paddle.optimizer.Adam(
config.training.lr, parameters=model.parameters())
criterion = WaveFlowLoss(sigma=config.model.sigma)
self.model = model
@ -46,20 +61,22 @@ class Experiment(ExperimentBase):
args = self.args
ljspeech_dataset = LJSpeech(args.data)
valid_set, train_set = dataset.split(ljspeech_dataset, config.data.valid_size)
valid_set, train_set = dataset.split(ljspeech_dataset,
config.data.valid_size)
batch_fn = LJSpeechClipCollector(config.data.clip_frames,
config.data.hop_length)
batch_fn = LJSpeechClipCollector(config.data.clip_frames, config.data.hop_length)
if not self.parallel:
train_loader = DataLoader(
train_set,
batch_size=config.data.batch_size,
shuffle=True,
train_set,
batch_size=config.data.batch_size,
shuffle=True,
drop_last=True,
collate_fn=batch_fn)
else:
sampler = DistributedBatchSampler(
train_set,
train_set,
batch_size=config.data.batch_size,
num_replicas=dist.get_world_size(),
rank=dist.get_rank(),
@ -71,7 +88,7 @@ class Experiment(ExperimentBase):
valid_batch_fn = LJSpeechCollector()
valid_loader = DataLoader(
valid_set, batch_size=1, collate_fn=valid_batch_fn)
self.train_loader = train_loader
self.valid_loader = valid_loader
@ -90,17 +107,19 @@ class Experiment(ExperimentBase):
mel, wav = batch
z, log_det_jocobian = self.compute_outputs(mel, wav)
loss = self.criterion(z, log_det_jocobian)
loss.backward()
loss.backward()
self.optimizer.step()
iteration_time = time.time() - start
loss_value = float(loss)
msg = "Rank: {}, ".format(dist.get_rank())
msg += "step: {}, ".format(self.iteration)
msg += "time: {:>.3f}s/{:>.3f}s, ".format(data_loader_time, iteration_time)
msg += "time: {:>.3f}s/{:>.3f}s, ".format(data_loader_time,
iteration_time)
msg += "loss: {:>.6f}".format(loss_value)
self.logger.info(msg)
self.visualizer.add_scalar("train/loss", loss_value, global_step=self.iteration)
self.visualizer.add_scalar(
"train/loss", loss_value, global_step=self.iteration)
@mp_tools.rank_zero_only
@paddle.no_grad()
@ -112,7 +131,8 @@ class Experiment(ExperimentBase):
loss = self.criterion(z, log_det_jocobian)
valid_losses.append(float(loss))
valid_loss = np.mean(valid_losses)
self.visualizer.add_scalar("valid/loss", valid_loss, global_step=self.iteration)
self.visualizer.add_scalar(
"valid/loss", valid_loss, global_step=self.iteration)
def main_sp(config, args):
@ -132,7 +152,7 @@ if __name__ == "__main__":
config = get_cfg_defaults()
parser = default_argument_parser()
args = parser.parse_args()
if args.config:
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)

55
examples/wavenet/config.py
View File

@ -1,19 +1,32 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from yacs.config import CfgNode as CN
_C = CN()
_C.data = CN(
dict(
batch_size=8, # batch size
valid_size=16, # the first N examples are reserved for validation
sample_rate=22050, # Hz, sample rate
n_fft=2048, # fft frame size
win_length=1024, # window size
batch_size=8, # batch size
valid_size=16, # the first N examples are reserved for validation
sample_rate=22050, # Hz, sample rate
n_fft=2048, # fft frame size
win_length=1024, # window size
hop_length=256, # hop size between ajacent frame
# f_max=8000, # Hz, max frequency when converting to mel
n_mels=80, # mel bands
train_clip_seconds=0.5, # audio clip length(in seconds)
)
)
train_clip_seconds=0.5, # audio clip length(in seconds)
))
_C.model = CN(
dict(
@ -21,24 +34,22 @@ _C.model = CN(
n_stack=3,
n_loop=10,
filter_size=2,
residual_channels=128, # resiaudal channel in each flow
residual_channels=128, # resiaudal channel in each flow
loss_type="mog",
output_dim=3, # single gaussian
log_scale_min=-9.0,
)
)
output_dim=3, # single gaussian
log_scale_min=-9.0, ))
_C.training = CN(
dict(
lr=1e-3, # learning rates
anneal_rate=0.5, # learning rate decay rate
anneal_interval=200000, # decrese lr by annel_rate every anneal_interval steps
valid_interval=1000, # validation
save_interval=10000, # checkpoint
max_iteration=3000000, # max iteration to train
gradient_max_norm=100.0 # global norm of gradients
)
)
lr=1e-3, # learning rates
anneal_rate=0.5, # learning rate decay rate
anneal_interval=200000, # decrese lr by annel_rate every anneal_interval steps
valid_interval=1000, # validation
save_interval=10000, # checkpoint
max_iteration=3000000, # max iteration to train
gradient_max_norm=100.0 # global norm of gradients
))
def get_cfg_defaults():
"""Get a yacs CfgNode object with default values for my_project."""

37
examples/wavenet/ljspeech.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
from pathlib import Path
import pickle
@ -9,19 +23,20 @@ from parakeet.data.batch import batch_spec, batch_wav
from parakeet.data import dataset
from parakeet.audio import AudioProcessor
class LJSpeech(Dataset):
"""A simple dataset adaptor for the processed ljspeech dataset."""
def __init__(self, root):
self.root = Path(root).expanduser()
meta_data = pandas.read_csv(
str(self.root / "metadata.csv"),
sep="\t",
header=None,
names=["fname", "frames", "samples"]
)
names=["fname", "frames", "samples"])
records = []
for row in meta_data.itertuples() :
for row in meta_data.itertuples():
mel_path = str(self.root / "mel" / (row.fname + ".npy"))
wav_path = str(self.root / "wav" / (row.fname + ".npy"))
records.append((mel_path, wav_path))
@ -39,6 +54,7 @@ class LJSpeech(Dataset):
class LJSpeechCollector(object):
"""A simple callable to batch LJSpeech examples."""
def __init__(self, padding_value=0.):
self.padding_value = padding_value
@ -48,15 +64,15 @@ class LJSpeechCollector(object):
wavs = [example[1] for example in examples]
mels = batch_spec(mels, pad_value=self.padding_value)
wavs = batch_wav(wavs, pad_value=self.padding_value)
audio_starts = np.zeros((batch_size,), dtype=np.int64)
audio_starts = np.zeros((batch_size, ), dtype=np.int64)
return mels, wavs, audio_starts
class LJSpeechClipCollector(object):
def __init__(self, clip_frames=65, hop_length=256):
self.clip_frames = clip_frames
self.clip_frames = clip_frames
self.hop_length = hop_length
def __call__(self, examples):
mels = []
wavs = []
@ -75,7 +91,8 @@ class LJSpeechClipCollector(object):
mel, wav = example
frames = mel.shape[-1]
start = np.random.randint(0, frames - self.clip_frames)
wav_clip = wav[start * self.hop_length: (start + self.clip_frames) * self.hop_length]
wav_clip = wav[start * self.hop_length:(start + self.clip_frames) *
self.hop_length]
return mel, wav_clip, start
@ -132,7 +149,3 @@ class DataCollector(object):
audios = np.array(audios, dtype=np.float32)
audio_starts = np.array(audio_starts, dtype=np.int64)
return audios, mels, audio_starts

54
examples/wavenet/preprocess.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import tqdm
import csv
@ -23,7 +37,7 @@ class Transform(object):
self.win_length = win_length
self.hop_length = hop_length
self.n_mels = n_mels
self.spec_normalizer = UnitMagnitude(min=1e-5)
def __call__(self, example):
@ -87,12 +101,8 @@ def create_dataset(config, input_dir, output_dir, verbose=True):
output_dir = Path(output_dir).expanduser()
output_dir.mkdir(exist_ok=True)
transform = Transform(
config.sample_rate,
config.n_fft,
config.win_length,
config.hop_length,
config.n_mels)
transform = Transform(config.sample_rate, config.n_fft, config.win_length,
config.hop_length, config.n_mels)
file_names = []
for example in tqdm.tqdm(dataset):
@ -108,23 +118,35 @@ def create_dataset(config, input_dir, output_dir, verbose=True):
np.save(str(mel_dir / base_name), mel)
file_names.append((base_name, mel.shape[-1], audio.shape[-1]))
meta_data = pd.DataFrame.from_records(file_names)
meta_data.to_csv(str(output_dir / "metadata.csv"), sep="\t", index=None, header=None)
print("saved meta data in to {}".format(os.path.join(output_dir, "metadata.csv")))
meta_data.to_csv(
str(output_dir / "metadata.csv"), sep="\t", index=None, header=None)
print("saved meta data in to {}".format(
os.path.join(output_dir, "metadata.csv")))
print("Done!")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="create dataset")
parser.add_argument("--config", type=str, metavar="FILE", help="extra config to overwrite the default config")
parser.add_argument("--input", type=str, help="path of the ljspeech dataset")
parser.add_argument("--output", type=str, help="path to save output dataset")
parser.add_argument("--opts", nargs=argparse.REMAINDER,
parser.add_argument(
"--config",
type=str,
metavar="FILE",
help="extra config to overwrite the default config")
parser.add_argument(
"--input", type=str, help="path of the ljspeech dataset")
parser.add_argument(
"--output", type=str, help="path to save output dataset")
parser.add_argument(
"--opts",
nargs=argparse.REMAINDER,
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs"
)
parser.add_argument("-v", "--verbose", action="store_true", help="print msg")
parser.add_argument(
"-v", "--verbose", action="store_true", help="print msg")
config = get_cfg_defaults()
args = parser.parse_args()
if args.config:

53
examples/wavenet/synthesize.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import numpy as np
import soundfile as sf
@ -10,6 +24,7 @@ from parakeet.utils import layer_tools, checkpoint
from config import get_cfg_defaults
def main(config, args):
paddle.set_device(args.device)
model = ConditionalWaveNet.from_pretrained(config, args.checkpoint_path)
@ -22,7 +37,8 @@ def main(config, args):
for file_path in mel_dir.iterdir():
mel = np.load(str(file_path))
audio = model.predict(mel)
audio_path = output_dir / (os.path.splitext(file_path.name)[0] + ".wav")
audio_path = output_dir / (
os.path.splitext(file_path.name)[0] + ".wav")
sf.write(audio_path, audio, config.data.sample_rate)
print("[synthesize] {} -> {}".format(file_path, audio_path))
@ -30,17 +46,32 @@ def main(config, args):
if __name__ == "__main__":
config = get_cfg_defaults()
parser = argparse.ArgumentParser(description="generate mel spectrogram with TransformerTTS.")
parser.add_argument("--config", type=str, metavar="FILE", help="extra config to overwrite the default config")
parser.add_argument("--checkpoint_path", type=str, help="path of the checkpoint to load.")
parser.add_argument("--input", type=str, help="path of directory containing mel spectrogram (in .npy format)")
parser = argparse.ArgumentParser(
description="generate mel spectrogram with TransformerTTS.")
parser.add_argument(
"--config",
type=str,
metavar="FILE",
help="extra config to overwrite the default config")
parser.add_argument(
"--checkpoint_path", type=str, help="path of the checkpoint to load.")
parser.add_argument(
"--input",
type=str,
help="path of directory containing mel spectrogram (in .npy format)")
parser.add_argument("--output", type=str, help="path to save outputs")
parser.add_argument("--device", type=str, default="cpu", help="device type to use.")
parser.add_argument("--opts", nargs=argparse.REMAINDER, help="options to overwrite --config file and the default config, passing in KEY VALUE pairs")
parser.add_argument("-v", "--verbose", action="store_true", help="print msg")
parser.add_argument(
"--device", type=str, default="cpu", help="device type to use.")
parser.add_argument(
"--opts",
nargs=argparse.REMAINDER,
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs"
)
parser.add_argument(
"-v", "--verbose", action="store_true", help="print msg")
args = parser.parse_args()
if args.config:
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)
@ -48,4 +79,4 @@ if __name__ == "__main__":
print(config)
print(args)
main(config, args)
main(config, args)

55
examples/wavenet/train.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import time
from pathlib import Path
import math
@ -26,7 +40,7 @@ class Experiment(ExperimentBase):
config = self.config
model = ConditionalWaveNet(
upsample_factors=config.model.upsample_factors,
n_stack=config.model.n_stack,
n_stack=config.model.n_stack,
n_loop=config.model.n_loop,
residual_channels=config.model.residual_channels,
output_dim=config.model.output_dim,
@ -39,13 +53,13 @@ class Experiment(ExperimentBase):
model = paddle.DataParallel(model)
lr_scheduler = paddle.optimizer.lr.StepDecay(
config.training.lr,
config.training.anneal_interval,
config.training.lr, config.training.anneal_interval,
config.training.anneal_rate)
optimizer = paddle.optimizer.Adam(
lr_scheduler,
parameters=model.parameters(),
grad_clip=paddle.nn.ClipGradByGlobalNorm(config.training.gradient_max_norm))
grad_clip=paddle.nn.ClipGradByGlobalNorm(
config.training.gradient_max_norm))
self.model = model
self.model_core = model._layer if self.parallel else model
@ -56,7 +70,8 @@ class Experiment(ExperimentBase):
args = self.args
ljspeech_dataset = LJSpeech(args.data)
valid_set, train_set = dataset.split(ljspeech_dataset, config.data.valid_size)
valid_set, train_set = dataset.split(ljspeech_dataset,
config.data.valid_size)
# convolutional net's causal padding size
context_size = config.model.n_stack \
@ -66,20 +81,21 @@ class Experiment(ExperimentBase):
# frames used to compute loss
frames_per_second = config.data.sample_rate // config.data.hop_length
train_clip_frames = math.ceil(config.data.train_clip_seconds * frames_per_second)
train_clip_frames = math.ceil(config.data.train_clip_seconds *
frames_per_second)
num_frames = train_clip_frames + context_frames
batch_fn = LJSpeechClipCollector(num_frames, config.data.hop_length)
if not self.parallel:
train_loader = DataLoader(
train_set,
batch_size=config.data.batch_size,
shuffle=True,
train_set,
batch_size=config.data.batch_size,
shuffle=True,
drop_last=True,
collate_fn=batch_fn)
else:
sampler = DistributedBatchSampler(
train_set,
train_set,
batch_size=config.data.batch_size,
shuffle=True,
drop_last=True)
@ -89,7 +105,7 @@ class Experiment(ExperimentBase):
valid_batch_fn = LJSpeechCollector()
valid_loader = DataLoader(
valid_set, batch_size=1, collate_fn=valid_batch_fn)
self.train_loader = train_loader
self.valid_loader = valid_loader
@ -101,20 +117,22 @@ class Experiment(ExperimentBase):
self.model.train()
self.optimizer.clear_grad()
mel, wav, audio_starts = batch
y = self.model(wav, mel, audio_starts)
loss = self.model.loss(y, wav)
loss.backward()
loss.backward()
self.optimizer.step()
iteration_time = time.time() - start
loss_value = float(loss)
msg = "Rank: {}, ".format(dist.get_rank())
msg += "step: {}, ".format(self.iteration)
msg += "time: {:>.3f}s/{:>.3f}s, ".format(data_loader_time, iteration_time)
msg += "time: {:>.3f}s/{:>.3f}s, ".format(data_loader_time,
iteration_time)
msg += "loss: {:>.6f}".format(loss_value)
self.logger.info(msg)
self.visualizer.add_scalar("train/loss", loss_value, global_step=self.iteration)
self.visualizer.add_scalar(
"train/loss", loss_value, global_step=self.iteration)
@mp_tools.rank_zero_only
@paddle.no_grad()
@ -126,7 +144,8 @@ class Experiment(ExperimentBase):
loss = self.model.loss(y, wav)
valid_losses.append(float(loss))
valid_loss = np.mean(valid_losses)
self.visualizer.add_scalar("valid/loss", valid_loss, global_step=self.iteration)
self.visualizer.add_scalar(
"valid/loss", valid_loss, global_step=self.iteration)
def main_sp(config, args):
@ -146,7 +165,7 @@ if __name__ == "__main__":
config = get_cfg_defaults()
parser = default_argument_parser()
args = parser.parse_args()
if args.config:
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)

34
parakeet/audio/audio.py
View File

@ -18,15 +18,16 @@ import numpy as np
__all__ = ["AudioProcessor"]
class AudioProcessor(object):
def __init__(self,
sample_rate:int,
n_fft:int,
win_length:int,
hop_length:int,
n_mels:int=80,
f_min:int=0,
f_max:int=None,
sample_rate: int,
n_fft: int,
win_length: int,
hop_length: int,
n_mels: int=80,
f_min: int=0,
f_max: int=None,
window="hann",
center=True,
pad_mode="reflect"):
@ -40,7 +41,7 @@ class AudioProcessor(object):
self.window = window
self.center = center
self.pad_mode = pad_mode
# mel
self.n_mels = n_mels
self.f_min = f_min
@ -48,19 +49,18 @@ class AudioProcessor(object):
self.mel_filter = self._create_mel_filter()
self.inv_mel_filter = np.linalg.pinv(self.mel_filter)
def _create_mel_filter(self):
mel_filter = librosa.filters.mel(
self.sample_rate,
self.n_fft,
n_mels=self.n_mels,
fmin=self.f_min,
fmax=self.f_max)
mel_filter = librosa.filters.mel(self.sample_rate,
self.n_fft,
n_mels=self.n_mels,
fmin=self.f_min,
fmax=self.f_max)
return mel_filter
def read_wav(self, filename):
# resampling may occur
wav, _ = librosa.load(filename, sr=self.sample_rate)
wav, _ = librosa.load(filename, sr=self.sample_rate)
return wav
def write_wav(self, path, wav):
@ -69,7 +69,7 @@ class AudioProcessor(object):
def stft(self, wav):
D = librosa.core.stft(
wav,
n_fft = self.n_fft,
n_fft=self.n_fft,
hop_length=self.hop_length,
win_length=self.win_length,
window=self.window,

26
parakeet/audio/spec_normalizer.py
View File

@ -1,3 +1,16 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This modules contains normalizers for spectrogram magnitude.
@ -19,22 +32,24 @@ __all__ = ["NormalizerBase", "LogMagnitude", "UnitMagnitude"]
class NormalizerBase(object):
def transform(self, spec):
raise NotImplementedError("transform must be implemented")
def inverse(self, normalized):
raise NotImplementedError("inverse must be implemented")
class LogMagnitude(NormalizerBase):
"""
This is a simple normalizer used in Waveglow, Waveflow, tacotron2...
"""
def __init__(self, min=1e-7):
self.min = min
def transform(self, x):
x = np.maximum(x, self.min)
x = np.log(x)
return x
def inverse(self, x):
return np.exp(x)
@ -44,15 +59,16 @@ class UnitMagnitude(NormalizerBase):
"""
This is the normalizer used in the
"""
def __init__(self, min=1e-5):
self.min = min
def transform(self, x):
db_scale = 20 * np.log10(np.maximum(self.min, x)) - 20
normalized = (db_scale + 100) / 100
clipped = np.clip(normalized, 0, 1)
return clipped
def inverse(self, x):
denormalized = np.clip(x, 0, 1) * 100 - 100
out = np.exp((denormalized + 20) / 20 * np.log(10))

30
parakeet/data/batch.py
View File

@ -18,10 +18,15 @@ Batch functions for text sequences, audio and spectrograms are provided.
import numpy as np
__all__ = [
"batch_text_id", "batch_wav", "batch_spec",
"TextIDBatcher", "WavBatcher", "SpecBatcher",
"batch_text_id",
"batch_wav",
"batch_spec",
"TextIDBatcher",
"WavBatcher",
"SpecBatcher",
]
class TextIDBatcher(object):
"""A wrapper class for `batch_text_id`."""
@ -99,8 +104,8 @@ def batch_wav(minibatch, pad_value=0., dtype=np.float32):
pad_len = max_len - example.shape[-1]
batch.append(
np.pad(example, [(0, pad_len)],
mode='constant',
constant_values=pad_value))
mode='constant',
constant_values=pad_value))
return np.array(batch, dtype=dtype)
@ -113,7 +118,11 @@ class SpecBatcher(object):
self.time_major = time_major
def __call__(self, minibatch):
out = batch_spec(minibatch, pad_value=self.pad_value, time_major=self.time_major, dtype=self.dtype)
out = batch_spec(
minibatch,
pad_value=self.pad_value,
time_major=self.time_major,
dtype=self.dtype)
return out
@ -130,7 +139,8 @@ def batch_spec(minibatch, pad_value=0., time_major=False, dtype=np.float32):
"""
# assume (F, T) or (T, F)
peek_example = minibatch[0]
assert len(peek_example.shape) == 2, "we only handles mono channel spectrogram"
assert len(
peek_example.shape) == 2, "we only handles mono channel spectrogram"
# assume (F, n_frame) or (n_frame, F)
time_idx = 0 if time_major else -1
@ -143,11 +153,11 @@ def batch_spec(minibatch, pad_value=0., time_major=False, dtype=np.float32):
if time_major:
batch.append(
np.pad(example, [(0, pad_len), (0, 0)],
mode='constant',
constant_values=pad_value))
mode='constant',
constant_values=pad_value))
else:
batch.append(
np.pad(example, [(0, 0), (0, pad_len)],
mode='constant',
constant_values=pad_value))
mode='constant',
constant_values=pad_value))
return np.array(batch, dtype=dtype)

14
parakeet/data/dataset.py
View File

@ -17,17 +17,25 @@ import paddle
from paddle.io import Dataset
__all__ = [
"split", "TransformDataset", "CacheDataset", "TupleDataset",
"DictDataset", "SliceDataset", "SubsetDataset", "FilterDataset",
"split",
"TransformDataset",
"CacheDataset",
"TupleDataset",
"DictDataset",
"SliceDataset",
"SubsetDataset",
"FilterDataset",
"ChainDataset",
]
def split(dataset, first_size):
"""A utility function to split a dataset into two datasets."""
first = SliceDataset(dataset, 0, first_size)
second = SliceDataset(dataset, first_size, len(dataset))
return first, second
class TransformDataset(Dataset):
def __init__(self, dataset, transform):
"""Dataset which is transformed from another with a transform.
@ -141,7 +149,7 @@ class DictDataset(Dataset):
for i in six.moves.range(length)]
else:
return batches
def __len__(self):
return self._length

14
parakeet/datasets/__init__.py
View File

@ -1,2 +1,16 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from parakeet.datasets.common import *
from parakeet.datasets.ljspeech import *

17
parakeet/datasets/common.py
View File

@ -1,9 +1,24 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.io import Dataset
import os
import librosa
__all__ = ["AudioFolderDataset"]
class AudioFolderDataset(Dataset):
def __init__(self, path, sample_rate, extension="wav"):
self.root = os.path.expanduser(path)
@ -19,5 +34,5 @@ class AudioFolderDataset(Dataset):
def __getitem__(self, i):
file_name = self.file_names[i]
y, _ = librosa.load(file_name, sr=self.sample_rate) # pylint: disable=unused-variable
y, _ = librosa.load(file_name, sr=self.sample_rate) # pylint: disable=unused-variable
return y

16
parakeet/datasets/ljspeech.py
View File

@ -1,8 +1,23 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.io import Dataset
from pathlib import Path
__all__ = ["LJSpeechMetaData"]
class LJSpeechMetaData(Dataset):
def __init__(self, root):
self.root = Path(root).expanduser()
@ -22,4 +37,3 @@ class LJSpeechMetaData(Dataset):
def __len__(self):
return len(self.records)

14
parakeet/frontend/__init__.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from parakeet.frontend.vocab import *
from parakeet.frontend.phonectic import *
from parakeet.frontend.punctuation import *

14
parakeet/frontend/normalizer/__init__.py
View File

@ -1,2 +1,16 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from parakeet.frontend.normalizer.normalizer import *
from parakeet.frontend.normalizer.numbers import *

14
parakeet/frontend/normalizer/abbrrviation.py
View File

@ -0,0 +1,14 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

14
parakeet/frontend/normalizer/acronyms.py
View File

@ -0,0 +1,14 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

25
parakeet/frontend/normalizer/width.py
View File

@ -1,8 +1,22 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
def full2half_width(ustr):
half = []
for u in ustr:
num = ord(u)
if num == 0x3000: # 全角空格变半角
if num == 0x3000: # 全角空格变半角
num = 32
elif 0xFF01 <= num <= 0xFF5E:
num -= 0xfee0
@ -10,15 +24,16 @@ def full2half_width(ustr):
half.append(u)
return ''.join(half)
def half2full_width(ustr):
full = []
for u in ustr:
num = ord(u)
if num == 32: # 半角空格变全角
if num == 32: # 半角空格变全角
num = 0x3000
elif 0x21 <= num <= 0x7E:
num += 0xfee0
u = chr(num) # to unicode
u = chr(num) # to unicode
full.append(u)
return ''.join(full)
return ''.join(full)

26
parakeet/frontend/punctuation.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import abc
import string
@ -13,15 +27,8 @@ EN_PUNCT = [
"!",
]
CN_PUNCT = [
"、",
"，",
"；",
"：",
"。",
"？",
"！"
]
CN_PUNCT = ["、", "，", "；", "：", "。", "？", "！"]
def get_punctuations(lang):
if lang == "en":
@ -30,4 +37,3 @@ def get_punctuations(lang):
return CN_PUNCT
else:
raise ValueError(f"language {lang} Not supported")

7
parakeet/models/transformer_tts.py
View File

@ -559,7 +559,7 @@ class TransformerTTS(nn.Layer):
@classmethod
def from_pretrained(cls, frontend, config, checkpoint_path):
model = TransformerTTS(
frontend,
frontend,
d_encoder=config.model.d_encoder,
d_decoder=config.model.d_decoder,
d_mel=config.data.d_mel,
@ -575,11 +575,12 @@ class TransformerTTS(nn.Layer):
decoder_prenet_dropout=config.model.decoder_prenet_dropout,
dropout=config.model.dropout)
iteration = checkpoint.load_parameters(model, checkpoint_path=checkpoint_path)
iteration = checkpoint.load_parameters(
model, checkpoint_path=checkpoint_path)
drop_n_heads = scheduler.StepWise(config.training.drop_n_heads)
reduction_factor = scheduler.StepWise(config.training.reduction_factor)
model.set_constants(
reduction_factor=reduction_factor(iteration),
reduction_factor=reduction_factor(iteration),
drop_n_heads=drop_n_heads(iteration))
return model

257
parakeet/models/waveflow.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import numpy as np
from typing import List, Union, Tuple
@ -11,6 +25,7 @@ from parakeet.modules import geometry as geo
__all__ = ["WaveFlow", "ConditionalWaveFlow", "WaveFlowLoss"]
def fold(x, n_group):
r"""Fold audio or spectrogram's temporal dimension in to groups.
@ -31,6 +46,7 @@ def fold(x, n_group):
new_shape = spatial_shape + [time_steps // n_group, n_group]
return paddle.reshape(x, new_shape)
class UpsampleNet(nn.LayerList):
"""Layer to upsample mel spectrogram to the same temporal resolution with
the corresponding waveform.
@ -60,6 +76,7 @@ class UpsampleNet(nn.LayerList):
---------
``librosa.core.stft``
"""
def __init__(self, upsample_factors):
super(UpsampleNet, self).__init__()
for factor in upsample_factors:
@ -67,16 +84,18 @@ class UpsampleNet(nn.LayerList):
init = I.Uniform(-std, std)
self.append(
nn.utils.weight_norm(
nn.Conv2DTranspose(1, 1, (3, 2 * factor),
nn.Conv2DTranspose(
1,
1, (3, 2 * factor),
padding=(1, factor // 2),
stride=(1, factor),
weight_attr=init,
bias_attr=init)))
# upsample factors
self.upsample_factor = np.prod(upsample_factors)
self.upsample_factors = upsample_factors
def forward(self, x, trim_conv_artifact=False):
r"""Forward pass of the ``UpsampleNet``.
@ -131,38 +150,47 @@ class ResidualBlock(nn.Layer):
dilations : int
Dilations of the Convolution2d applied to the input.
"""
def __init__(self, channels, cond_channels, kernel_size, dilations):
super(ResidualBlock, self).__init__()
# input conv
std = math.sqrt(1 / channels * np.prod(kernel_size))
init = I.Uniform(-std, std)
receptive_field = [1 + (k - 1) * d for (k, d) in zip(kernel_size, dilations)]
receptive_field = [
1 + (k - 1) * d for (k, d) in zip(kernel_size, dilations)
]
rh, rw = receptive_field
paddings = [rh - 1, 0, rw // 2, (rw - 1) // 2] # causal & same
conv = nn.Conv2D(channels, 2 * channels, kernel_size,
padding=paddings,
dilation=dilations,
weight_attr=init,
bias_attr=init)
paddings = [rh - 1, 0, rw // 2, (rw - 1) // 2] # causal & same
conv = nn.Conv2D(
channels,
2 * channels,
kernel_size,
padding=paddings,
dilation=dilations,
weight_attr=init,
bias_attr=init)
self.conv = nn.utils.weight_norm(conv)
self.rh = rh
self.rw = rw
self.dilations = dilations
# condition projection
std = math.sqrt(1 / cond_channels)
init = I.Uniform(-std, std)
condition_proj = nn.Conv2D(cond_channels, 2 * channels, (1, 1),
weight_attr=init, bias_attr=init)
condition_proj = nn.Conv2D(
cond_channels,
2 * channels, (1, 1),
weight_attr=init,
bias_attr=init)
self.condition_proj = nn.utils.weight_norm(condition_proj)
# parametric residual & skip connection
std = math.sqrt(1 / channels)
init = I.Uniform(-std, std)
out_proj = nn.Conv2D(channels, 2 * channels, (1, 1),
weight_attr=init, bias_attr=init)
out_proj = nn.Conv2D(
channels, 2 * channels, (1, 1), weight_attr=init, bias_attr=init)
self.out_proj = nn.utils.weight_norm(out_proj)
def forward(self, x, condition):
"""Compute output for a whole folded sequence.
@ -185,10 +213,10 @@ class ResidualBlock(nn.Layer):
x_in = x
x = self.conv(x)
x += self.condition_proj(condition)
content, gate = paddle.chunk(x, 2, axis=1)
x = paddle.tanh(content) * F.sigmoid(gate)
x = self.out_proj(x)
res, skip = paddle.chunk(x, 2, axis=1)
res = x_in + res
@ -249,7 +277,7 @@ class ResidualBlock(nn.Layer):
content, gate = paddle.chunk(x_row, 2, axis=1)
x_row = paddle.tanh(content) * F.sigmoid(gate)
x_row = self.out_proj(x_row)
res, skip = paddle.chunk(x_row, 2, axis=1)
res = x_row_in + res
@ -290,20 +318,23 @@ class ResidualNet(nn.LayerList):
ValueError
If the length of dilations_h does not equals n_layers.
"""
def __init__(self,
n_layer: int,
residual_channels: int,
condition_channels: int,
kernel_size: Tuple[int],
def __init__(self,
n_layer: int,
residual_channels: int,
condition_channels: int,
kernel_size: Tuple[int],
dilations_h: List[int]):
if len(dilations_h) != n_layer:
raise ValueError("number of dilations_h should equals num of layers")
raise ValueError(
"number of dilations_h should equals num of layers")
super(ResidualNet, self).__init__()
for i in range(n_layer):
dilation = (dilations_h[i], 2 ** i)
layer = ResidualBlock(residual_channels, condition_channels, kernel_size, dilation)
dilation = (dilations_h[i], 2**i)
layer = ResidualBlock(residual_channels, condition_channels,
kernel_size, dilation)
self.append(layer)
def forward(self, x, condition):
"""Comput the output of given the input and the condition.
@ -332,7 +363,7 @@ class ResidualNet(nn.LayerList):
"""
for layer in self:
layer.start_sequence()
def add_input(self, x_row, condition_row):
"""Compute the output for a row and update the buffers.
@ -386,33 +417,37 @@ class Flow(nn.Layer):
Number of timesteps to the folded into a group.
"""
dilations_dict = {
8: [1, 1, 1, 1, 1, 1, 1, 1],
16: [1, 1, 1, 1, 1, 1, 1, 1],
32: [1, 2, 4, 1, 2, 4, 1, 2],
64: [1, 2, 4, 8, 16, 1, 2, 4],
128: [1, 2, 4, 8, 16, 32, 64, 1]
8: [1, 1, 1, 1, 1, 1, 1, 1],
16: [1, 1, 1, 1, 1, 1, 1, 1],
32: [1, 2, 4, 1, 2, 4, 1, 2],
64: [1, 2, 4, 8, 16, 1, 2, 4],
128: [1, 2, 4, 8, 16, 32, 64, 1]
}
def __init__(self, n_layers, channels, mel_bands, kernel_size, n_group):
super(Flow, self).__init__()
# input projection
self.input_proj = nn.utils.weight_norm(
nn.Conv2D(1, channels, (1, 1),
weight_attr=I.Uniform(-1., 1.),
bias_attr=I.Uniform(-1., 1.)))
nn.Conv2D(
1,
channels, (1, 1),
weight_attr=I.Uniform(-1., 1.),
bias_attr=I.Uniform(-1., 1.)))
# residual net
self.resnet = ResidualNet(n_layers, channels, mel_bands, kernel_size,
self.resnet = ResidualNet(n_layers, channels, mel_bands, kernel_size,
self.dilations_dict[n_group])
# output projection
self.output_proj = nn.Conv2D(channels, 2, (1, 1),
weight_attr=I.Constant(0.),
bias_attr=I.Constant(0.))
self.output_proj = nn.Conv2D(
channels,
2, (1, 1),
weight_attr=I.Constant(0.),
bias_attr=I.Constant(0.))
# specs
self.n_group = n_group
def _predict_parameters(self, x, condition):
x = self.input_proj(x)
x = self.resnet(x, condition)
@ -421,11 +456,11 @@ class Flow(nn.Layer):
return logs, b
def _transform(self, x, logs, b):
z_0 = x[:, :, :1, :] # the first row, just copy it
z_out = x[:, :, 1:, :] * paddle.exp(logs) + b
z_0 = x[:, :, :1, :] # the first row, just copy it
z_out = x[:, :, 1:, :] * paddle.exp(logs) + b
z_out = paddle.concat([z_0, z_out], axis=2)
return z_out
def forward(self, x, condition):
"""Probability density estimation. It is done by inversely transform
a sample from p(X) into a sample from p(Z).
@ -452,8 +487,8 @@ class Flow(nn.Layer):
transformation from x to z.
"""
# (B, C, H-1, W)
logs, b = self._predict_parameters(
x[:, :, :-1, :], condition[:, :, 1:, :])
logs, b = self._predict_parameters(x[:, :, :-1, :],
condition[:, :, 1:, :])
z = self._transform(x, logs, b)
return z, (logs, b)
@ -467,7 +502,7 @@ class Flow(nn.Layer):
def _inverse_transform_row(self, z_row, logs, b):
x_row = (z_row - b) * paddle.exp(-logs)
return x_row
def _inverse_row(self, z_row, x_row, condition_row):
logs, b = self._predict_row_parameters(x_row, condition_row)
x_next_row = self._inverse_transform_row(z_row, logs, b)
@ -475,7 +510,7 @@ class Flow(nn.Layer):
def _start_sequence(self):
self.resnet.start_sequence()
def inverse(self, z, condition):
"""Sampling from the the distrition p(X). It is done by sample form
p(Z) and transform the sample. It is a auto regressive transformation.
@ -510,15 +545,16 @@ class Flow(nn.Layer):
self._start_sequence()
for i in range(1, self.n_group):
x_row = x[-1] # actuallt i-1:i
z_row = z[:, :, i:i+1, :]
condition_row = condition[:, :, i:i+1, :]
x_row = x[-1] # actuallt i-1:i
z_row = z[:, :, i:i + 1, :]
condition_row = condition[:, :, i:i + 1, :]
x_next_row, (logs, b) = self._inverse_row(z_row, x_row, condition_row)
x_next_row, (logs, b) = self._inverse_row(z_row, x_row,
condition_row)
x.append(x_next_row)
logs_list.append(logs)
b_list.append(b)
x = paddle.concat(x, 2)
logs = paddle.concat(logs_list, 2)
b = paddle.concat(b_list, 2)
@ -549,21 +585,25 @@ class WaveFlow(nn.LayerList):
kernel_size : Union[int, List[int]]
Kernel size of the convolution layer in each ResidualBlock.
"""
def __init__(self, n_flows, n_layers, n_group, channels, mel_bands, kernel_size):
def __init__(self, n_flows, n_layers, n_group, channels, mel_bands,
kernel_size):
if n_group % 2 or n_flows % 2:
raise ValueError("number of flows and number of group must be even "
"since a permutation along group among flows is used.")
raise ValueError(
"number of flows and number of group must be even "
"since a permutation along group among flows is used.")
super(WaveFlow, self).__init__()
for _ in range(n_flows):
self.append(Flow(n_layers, channels, mel_bands, kernel_size, n_group))
self.append(
Flow(n_layers, channels, mel_bands, kernel_size, n_group))
# permutations in h
self.perms = self._create_perm(n_group, n_flows)
# specs
self.n_group = n_group
self.n_flows = n_flows
def _create_perm(self, n_group, n_flows):
indices = list(range(n_group))
half = n_group // 2
@ -572,20 +612,21 @@ class WaveFlow(nn.LayerList):
if i < n_flows // 2:
perms.append(indices[::-1])
else:
perm = list(reversed(indices[:half])) + list(reversed(indices[half:]))
perm = list(reversed(indices[:half])) + list(
reversed(indices[half:]))
perms.append(perm)
return perms
def _trim(self, x, condition):
assert condition.shape[-1] >= x.shape[-1]
pruned_len = int(x.shape[-1] // self.n_group * self.n_group)
if x.shape[-1] > pruned_len:
x = x[:, :pruned_len]
if condition.shape[-1] > pruned_len:
condition = condition[:, :, :pruned_len]
return x, condition
def forward(self, x, condition):
"""Probability density estimation of random variable x given the
condition.
@ -610,21 +651,23 @@ class WaveFlow(nn.LayerList):
# x: (B, T)
# condition: (B, C, T) upsampled condition
x, condition = self._trim(x, condition)
# to (B, C, h, T//h) layout
x = paddle.unsqueeze(paddle.transpose(fold(x, self.n_group), [0, 2, 1]), 1)
condition = paddle.transpose(fold(condition, self.n_group), [0, 1, 3, 2])
x = paddle.unsqueeze(
paddle.transpose(fold(x, self.n_group), [0, 2, 1]), 1)
condition = paddle.transpose(
fold(condition, self.n_group), [0, 1, 3, 2])
# flows
logs_list = []
for i, layer in enumerate(self):
x, (logs, b) = layer(x, condition)
x, (logs, b) = layer(x, condition)
logs_list.append(logs)
# permute paddle has no shuffle dim
x = geo.shuffle_dim(x, 2, perm=self.perms[i])
condition = geo.shuffle_dim(condition, 2, perm=self.perms[i])
z = paddle.squeeze(x, 1) # (B, H, W)
z = paddle.squeeze(x, 1) # (B, H, W)
batch_size = z.shape[0]
z = paddle.reshape(paddle.transpose(z, [0, 2, 1]), [batch_size, -1])
@ -654,8 +697,10 @@ class WaveFlow(nn.LayerList):
z, condition = self._trim(z, condition)
# to (B, C, h, T//h) layout
z = paddle.unsqueeze(paddle.transpose(fold(z, self.n_group), [0, 2, 1]), 1)
condition = paddle.transpose(fold(condition, self.n_group), [0, 1, 3, 2])
z = paddle.unsqueeze(
paddle.transpose(fold(z, self.n_group), [0, 2, 1]), 1)
condition = paddle.transpose(
fold(condition, self.n_group), [0, 1, 3, 2])
# reverse it flow by flow
for i in reversed(range(self.n_flows)):
@ -663,7 +708,7 @@ class WaveFlow(nn.LayerList):
condition = geo.shuffle_dim(condition, 2, perm=self.perms[i])
z, (logs, b) = self[i].inverse(z, condition)
x = paddle.squeeze(z, 1) # (B, H, W)
x = paddle.squeeze(z, 1) # (B, H, W)
batch_size = x.shape[0]
x = paddle.reshape(paddle.transpose(x, [0, 2, 1]), [batch_size, -1])
return x
@ -695,23 +740,24 @@ class ConditionalWaveFlow(nn.LayerList):
kernel_size : Union[int, List[int]]
Kernel size of the convolution layer in each ResidualBlock.
"""
def __init__(self,
upsample_factors: List[int],
n_flows: int,
n_layers: int,
n_group: int,
channels: int,
n_mels: int,
kernel_size: Union[int, List[int]]):
def __init__(self,
upsample_factors: List[int],
n_flows: int,
n_layers: int,
n_group: int,
channels: int,
n_mels: int,
kernel_size: Union[int, List[int]]):
super(ConditionalWaveFlow, self).__init__()
self.encoder = UpsampleNet(upsample_factors)
self.decoder = WaveFlow(
n_flows=n_flows,
n_layers=n_layers,
n_group=n_group,
channels=channels,
mel_bands=n_mels,
kernel_size=kernel_size)
n_flows=n_flows,
n_layers=n_layers,
n_group=n_group,
channels=channels,
mel_bands=n_mels,
kernel_size=kernel_size)
def forward(self, audio, mel):
"""Compute the transformed random variable z (x to z) and the log of
@ -737,7 +783,7 @@ class ConditionalWaveFlow(nn.LayerList):
condition = self.encoder(mel)
z, log_det_jacobian = self.decoder(audio, condition)
return z, log_det_jacobian
@paddle.no_grad()
def infer(self, mel):
r"""Generate raw audio given mel spectrogram.
@ -752,12 +798,12 @@ class ConditionalWaveFlow(nn.LayerList):
Tensor : [shape=(B, T)]
The synthesized audio, where``T <= T_mel \* upsample_factors``.
"""
condition = self.encoder(mel, trim_conv_artifact=True) #(B, C, T)
condition = self.encoder(mel, trim_conv_artifact=True) #(B, C, T)
batch_size, _, time_steps = condition.shape
z = paddle.randn([batch_size, time_steps], dtype=mel.dtype)
x = self.decoder.inverse(z, condition)
return x
@paddle.no_grad()
def predict(self, mel):
"""Generate raw audio given mel spectrogram.
@ -777,7 +823,7 @@ class ConditionalWaveFlow(nn.LayerList):
audio = self.infer(mel)
audio = audio[0].numpy()
return audio
@classmethod
def from_pretrained(cls, config, checkpoint_path):
"""Build a ConditionalWaveFlow model from a pretrained model.
@ -795,14 +841,13 @@ class ConditionalWaveFlow(nn.LayerList):
ConditionalWaveFlow
The model built from pretrained result.
"""
model = cls(
upsample_factors=config.model.upsample_factors,
n_flows=config.model.n_flows,
n_layers=config.model.n_layers,
n_group=config.model.n_group,
channels=config.model.channels,
n_mels=config.data.n_mels,
kernel_size=config.model.kernel_size)
model = cls(upsample_factors=config.model.upsample_factors,
n_flows=config.model.n_flows,
n_layers=config.model.n_layers,
n_group=config.model.n_group,
channels=config.model.channels,
n_mels=config.data.n_mels,
kernel_size=config.model.kernel_size)
checkpoint.load_parameters(model, checkpoint_path=checkpoint_path)
return model
@ -816,6 +861,7 @@ class WaveFlowLoss(nn.Layer):
The standard deviation of the gaussian noise used in WaveFlow, by
default 1.0.
"""
def __init__(self, sigma=1.0):
super(WaveFlowLoss, self).__init__()
self.sigma = sigma
@ -839,6 +885,7 @@ class WaveFlowLoss(nn.Layer):
Tensor [shape=(1,)]
The loss.
"""
loss = paddle.sum(z * z) / (2 * self.sigma * self.sigma) - log_det_jacobian
loss = paddle.sum(z * z) / (2 * self.sigma * self.sigma
) - log_det_jacobian
loss = loss / np.prod(z.shape)
return loss + self.const

135
parakeet/models/wavenet.py
View File

@ -18,7 +18,7 @@ from typing import Union, Sequence, List
from tqdm import trange
import numpy as np
import paddle
import paddle
from paddle import nn
from paddle.nn import functional as F
import paddle.fluid.initializer as I
@ -30,6 +30,7 @@ from parakeet.utils import checkpoint, layer_tools
__all__ = ["WaveNet", "ConditionalWaveNet"]
def crop(x, audio_start, audio_length):
"""Crop the upsampled condition to match audio_length.
@ -96,6 +97,7 @@ class UpsampleNet(nn.LayerList):
---------
``librosa.core.stft``
"""
def __init__(self, upscale_factors=[16, 16]):
super(UpsampleNet, self).__init__()
self.upscale_factors = list(upscale_factors)
@ -106,9 +108,11 @@ class UpsampleNet(nn.LayerList):
for factor in self.upscale_factors:
self.append(
nn.utils.weight_norm(
nn.Conv2DTranspose(1, 1,
kernel_size=(3, 2 * factor),
stride=(1, factor),
nn.Conv2DTranspose(
1,
1,
kernel_size=(3, 2 * factor),
stride=(1, factor),
padding=(1, factor // 2))))
def forward(self, x):
@ -159,29 +163,34 @@ class ResidualBlock(nn.Layer):
dilation :int
Dilation of the internal convolution cells.
"""
def __init__(self,
residual_channels: int,
condition_dim: int,
def __init__(self,
residual_channels: int,
condition_dim: int,
filter_size: Union[int, Sequence[int]],
dilation: int):
super(ResidualBlock, self).__init__()
dilated_channels = 2 * residual_channels
# following clarinet's implementation, we do not have parametric residual
# & skip connection.
_filter_size = filter_size[0] if isinstance(filter_size, (list, tuple)) else filter_size
_filter_size = filter_size[0] if isinstance(filter_size, (
list, tuple)) else filter_size
std = math.sqrt(1 / (_filter_size * residual_channels))
conv = Conv1dCell(residual_channels,
dilated_channels,
filter_size,
dilation=dilation,
weight_attr=I.Normal(scale=std))
conv = Conv1dCell(
residual_channels,
dilated_channels,
filter_size,
dilation=dilation,
weight_attr=I.Normal(scale=std))
self.conv = nn.utils.weight_norm(conv)
std = math.sqrt(1 / condition_dim)
condition_proj = Conv1dCell(condition_dim, dilated_channels, (1,),
weight_attr=I.Normal(scale=std))
condition_proj = Conv1dCell(
condition_dim,
dilated_channels, (1, ),
weight_attr=I.Normal(scale=std))
self.condition_proj = nn.utils.weight_norm(condition_proj)
self.filter_size = filter_size
@ -309,10 +318,11 @@ class ResidualNet(nn.LayerList):
Kernel size of the internal ``Conv1dCell`` of each ``ResidualBlock``.
"""
def __init__(self,
n_stack: int,
n_loop: int,
residual_channels: int,
def __init__(self,
n_stack: int,
n_loop: int,
residual_channels: int,
condition_dim: int,
filter_size: int):
super(ResidualNet, self).__init__()
@ -320,7 +330,9 @@ class ResidualNet(nn.LayerList):
dilations = [2**i for i in range(n_loop)] * n_stack
self.context_size = 1 + sum(dilations)
for dilation in dilations:
self.append(ResidualBlock(residual_channels, condition_dim, filter_size, dilation))
self.append(
ResidualBlock(residual_channels, condition_dim, filter_size,
dilation))
def forward(self, x, condition=None):
"""Forward pass of ``ResidualNet``.
@ -345,7 +357,7 @@ class ResidualNet(nn.LayerList):
skip_connections = skip
else:
skip_connections = paddle.scale(skip_connections + skip,
math.sqrt(0.5))
math.sqrt(0.5))
return skip_connections
def start_sequence(self):
@ -381,7 +393,7 @@ class ResidualNet(nn.LayerList):
skip_connections = skip
else:
skip_connections = paddle.scale(skip_connections + skip,
math.sqrt(0.5))
math.sqrt(0.5))
return skip_connections
@ -426,6 +438,7 @@ class WaveNet(nn.Layer):
This is only used for computing loss when ``loss_type`` is "mog", If
the predicted log scale is less than -9.0, it is clipped at -9.0.
"""
def __init__(self, n_stack, n_loop, residual_channels, output_dim,
condition_dim, filter_size, loss_type, log_scale_min):
@ -437,19 +450,24 @@ class WaveNet(nn.Layer):
else:
if (output_dim % 3 != 0):
raise ValueError(
"with Mixture of Gaussians(mog) output, the output dim must be divisible by 3, but get {}".format(output_dim))
self.embed = nn.utils.weight_norm(nn.Linear(1, residual_channels), dim=1)
"with Mixture of Gaussians(mog) output, the output dim must be divisible by 3, but get {}".
format(output_dim))
self.embed = nn.utils.weight_norm(
nn.Linear(1, residual_channels), dim=1)
self.resnet = ResidualNet(n_stack, n_loop, residual_channels,
condition_dim, filter_size)
self.context_size = self.resnet.context_size
skip_channels = residual_channels # assume the same channel
self.proj1 = nn.utils.weight_norm(nn.Linear(skip_channels, skip_channels), dim=1)
self.proj2 = nn.utils.weight_norm(nn.Linear(skip_channels, skip_channels), dim=1)
self.proj1 = nn.utils.weight_norm(
nn.Linear(skip_channels, skip_channels), dim=1)
self.proj2 = nn.utils.weight_norm(
nn.Linear(skip_channels, skip_channels), dim=1)
# if loss_type is softmax, output_dim is n_vocab of waveform magnitude.
# if loss_type is mog, output_dim is 3 * gaussian, (weight, mean and stddev)
self.proj3 = nn.utils.weight_norm(nn.Linear(skip_channels, output_dim), dim=1)
self.proj3 = nn.utils.weight_norm(
nn.Linear(skip_channels, output_dim), dim=1)
self.loss_type = loss_type
self.output_dim = output_dim
@ -781,26 +799,28 @@ class ConditionalWaveNet(nn.Layer):
This is only used for computing loss when ``loss_type`` is "mog", If
the predicted log scale is less than -9.0, it is clipped at -9.0.
"""
def __init__(self,
upsample_factors: List[int],
n_stack: int,
n_loop: int,
residual_channels: int,
def __init__(self,
upsample_factors: List[int],
n_stack: int,
n_loop: int,
residual_channels: int,
output_dim: int,
n_mels: int,
filter_size: int=2,
loss_type: str="mog",
n_mels: int,
filter_size: int=2,
loss_type: str="mog",
log_scale_min: float=-9.0):
super(ConditionalWaveNet, self).__init__()
self.encoder = UpsampleNet(upsample_factors)
self.decoder = WaveNet(n_stack=n_stack,
n_loop=n_loop,
residual_channels=residual_channels,
output_dim=output_dim,
condition_dim=n_mels,
filter_size=filter_size,
loss_type=loss_type,
log_scale_min=log_scale_min)
self.decoder = WaveNet(
n_stack=n_stack,
n_loop=n_loop,
residual_channels=residual_channels,
output_dim=output_dim,
condition_dim=n_mels,
filter_size=filter_size,
loss_type=loss_type,
log_scale_min=log_scale_min)
def forward(self, audio, mel, audio_start):
"""Compute the output distribution given the mel spectrogram and the input(for teacher force training).
@ -895,11 +915,11 @@ class ConditionalWaveNet(nn.Layer):
self.decoder.start_sequence()
x_t = paddle.zeros((batch_size, ), dtype=mel.dtype)
for i in trange(time_steps):
c_t = condition[:, :, i] # (B, C)
y_t = self.decoder.add_input(x_t, c_t) #(B, C)
c_t = condition[:, :, i] # (B, C)
y_t = self.decoder.add_input(x_t, c_t) #(B, C)
y_t = paddle.unsqueeze(y_t, 1)
x_t = self.sample(y_t) # (B, 1)
x_t = paddle.squeeze(x_t, 1) #(B,)
x_t = self.sample(y_t) # (B, 1)
x_t = paddle.squeeze(x_t, 1) #(B,)
samples.append(x_t)
samples = paddle.stack(samples, -1)
return samples
@ -943,16 +963,15 @@ class ConditionalWaveNet(nn.Layer):
ConditionalWaveNet
The model built from pretrained result.
"""
model = cls(
upsample_factors=config.model.upsample_factors,
n_stack=config.model.n_stack,
n_loop=config.model.n_loop,
residual_channels=config.model.residual_channels,
output_dim=config.model.output_dim,
n_mels=config.data.n_mels,
filter_size=config.model.filter_size,
loss_type=config.model.loss_type,
log_scale_min=config.model.log_scale_min)
model = cls(upsample_factors=config.model.upsample_factors,
n_stack=config.model.n_stack,
n_loop=config.model.n_loop,
residual_channels=config.model.residual_channels,
output_dim=config.model.output_dim,
n_mels=config.data.n_mels,
filter_size=config.model.filter_size,
loss_type=config.model.loss_type,
log_scale_min=config.model.log_scale_min)
layer_tools.summary(model)
checkpoint.load_parameters(model, checkpoint_path=checkpoint_path)
return model

20
parakeet/modules/audio.py
View File

@ -1,8 +1,22 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
from paddle import nn
from paddle.nn import functional as F
from scipy import signal
import numpy as np
import numpy as np
__all__ = ["quantize", "dequantize", "STFT"]
@ -86,6 +100,7 @@ class STFT(nn.Layer):
Ony ``center`` and ``reflect`` padding is supported now.
"""
def __init__(self, n_fft, hop_length, win_length, window="hanning"):
super(STFT, self).__init__()
self.hop_length = hop_length
@ -109,7 +124,8 @@ class STFT(nn.Layer):
(self.n_bin, 1, 1, self.n_fft))
w = np.concatenate([w_real, w_imag], axis=0)
self.weight = paddle.cast(paddle.to_tensor(w), paddle.get_default_dtype())
self.weight = paddle.cast(
paddle.to_tensor(w), paddle.get_default_dtype())
def forward(self, x):
"""Compute the stft transform.

2
parakeet/modules/conv.py
View File

@ -20,6 +20,7 @@ __all__ = [
"Conv1dBatchNorm",
]
class Conv1dCell(nn.Conv1D):
"""A subclass of Conv1D layer, which can be used in an autoregressive
decoder like an RNN cell.
@ -231,6 +232,7 @@ class Conv1dBatchNorm(nn.Layer):
epsilon : [type], optional
The epsilon of the BatchNorm1D layer, by default 1e-05
"""
def __init__(self,
in_channels,
out_channels,

17
parakeet/modules/geometry.py
View File

@ -1,6 +1,21 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import paddle
def shuffle_dim(x, axis, perm=None):
"""Permute input tensor along aixs given the permutation or randomly.
@ -32,7 +47,7 @@ def shuffle_dim(x, axis, perm=None):
perm = np.array(perm)
else:
perm = np.random.permutation(size)
perm = paddle.to_tensor(perm)
out = paddle.gather(x, perm, axis)
return out

36
parakeet/modules/losses.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numba
import numpy as np
import paddle
@ -5,12 +19,13 @@ from paddle import nn
from paddle.nn import functional as F
__all__ = [
"weighted_mean",
"masked_l1_loss",
"masked_softmax_with_cross_entropy",
"weighted_mean",
"masked_l1_loss",
"masked_softmax_with_cross_entropy",
"diagonal_loss",
]
def weighted_mean(input, weight):
"""Weighted mean. It can also be used as masked mean.
@ -88,12 +103,11 @@ def masked_softmax_with_cross_entropy(logits, label, mask, axis=-1):
return loss
def diagonal_loss(
attentions,
input_lengths,
target_lengths,
g=0.2,
multihead=False):
def diagonal_loss(attentions,
input_lengths,
target_lengths,
g=0.2,
multihead=False):
"""A metric to evaluate how diagonal a attention distribution is.
It is computed for batch attention distributions. For each attention
@ -133,6 +147,7 @@ def diagonal_loss(
else:
return paddle.mean(attentions * paddle.unsqueeze(W_tensor, 1))
@numba.jit(nopython=True)
def guided_attention(N, max_N, T, max_T, g):
W = np.zeros((max_T, max_N), dtype=np.float32)
@ -142,6 +157,7 @@ def guided_attention(N, max_N, T, max_T, g):
# (T_dec, T_enc)
return W
def guided_attentions(input_lengths, target_lengths, g=0.2):
B = len(input_lengths)
max_input_len = input_lengths.max()
@ -151,4 +167,4 @@ def guided_attentions(input_lengths, target_lengths, g=0.2):
W[b] = guided_attention(input_lengths[b], max_input_len,
target_lengths[b], max_target_len, g)
# (B, T_dec, T_enc)
return W
return W

15
parakeet/modules/masking.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
from paddle.fluid.layers import sequence_mask
@ -8,6 +22,7 @@ __all__ = [
"future_mask",
]
def id_mask(input, padding_index=0, dtype="bool"):
"""Generate mask with input ids.

17
parakeet/modules/positional_encoding.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import numpy as np
import paddle
@ -5,6 +19,7 @@ from paddle.nn import functional as F
__all__ = ["positional_encoding"]
def positional_encoding(start_index, length, size, dtype=None):
r"""Generate standard positional encoding matrix.
@ -37,7 +52,7 @@ def positional_encoding(start_index, length, size, dtype=None):
dtype = dtype or paddle.get_default_dtype()
channel = np.arange(0, size, 2)
index = np.arange(start_index, start_index + length, 1)
p = np.expand_dims(index, -1) / (10000 ** (channel / float(size)))
p = np.expand_dims(index, -1) / (10000**(channel / float(size)))
encodings = np.zeros([length, size])
encodings[:, 0::2] = np.sin(p)
encodings[:, 1::2] = np.cos(p)

74
parakeet/modules/transformer.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import paddle
from paddle import nn
@ -12,6 +26,7 @@ __all__ = [
"TransformerDecoderLayer",
]
class PositionwiseFFN(nn.Layer):
"""A faithful implementation of Position-wise Feed-Forward Network
in `Attention is All You Need <https://arxiv.org/abs/1706.03762>`_.
@ -30,10 +45,8 @@ class PositionwiseFFN(nn.Layer):
The probability of the Dropout applied to the output of the first
layer, by default 0.
"""
def __init__(self,
input_size: int,
hidden_size: int,
dropout=0.0):
def __init__(self, input_size: int, hidden_size: int, dropout=0.0):
super(PositionwiseFFN, self).__init__()
self.linear1 = nn.Linear(input_size, hidden_size)
self.linear2 = nn.Linear(hidden_size, input_size)
@ -86,16 +99,17 @@ class TransformerEncoderLayer(nn.Layer):
------
It uses the PostLN (post layer norm) scheme.
"""
def __init__(self, d_model, n_heads, d_ffn, dropout=0.):
super(TransformerEncoderLayer, self).__init__()
self.self_mha = attn.MultiheadAttention(d_model, n_heads, dropout)
self.layer_norm1 = nn.LayerNorm([d_model], epsilon=1e-6)
self.ffn = PositionwiseFFN(d_model, d_ffn, dropout)
self.layer_norm2 = nn.LayerNorm([d_model], epsilon=1e-6)
self.dropout = dropout
def forward(self, x, mask):
"""Forward pass of TransformerEncoderLayer.
@ -118,14 +132,12 @@ class TransformerEncoderLayer(nn.Layer):
"""
context_vector, attn_weights = self.self_mha(x, x, x, mask)
x = self.layer_norm1(
F.dropout(x + context_vector,
self.dropout,
training=self.training))
F.dropout(
x + context_vector, self.dropout, training=self.training))
x = self.layer_norm2(
F.dropout(x + self.ffn(x),
self.dropout,
training=self.training))
F.dropout(
x + self.ffn(x), self.dropout, training=self.training))
return x, attn_weights
@ -155,19 +167,20 @@ class TransformerDecoderLayer(nn.Layer):
------
It uses the PostLN (post layer norm) scheme.
"""
def __init__(self, d_model, n_heads, d_ffn, dropout=0.):
super(TransformerDecoderLayer, self).__init__()
self.self_mha = attn.MultiheadAttention(d_model, n_heads, dropout)
self.layer_norm1 = nn.LayerNorm([d_model], epsilon=1e-6)
self.cross_mha = attn.MultiheadAttention(d_model, n_heads, dropout)
self.layer_norm2 = nn.LayerNorm([d_model], epsilon=1e-6)
self.ffn = PositionwiseFFN(d_model, d_ffn, dropout)
self.layer_norm3 = nn.LayerNorm([d_model], epsilon=1e-6)
self.dropout = dropout
def forward(self, q, k, v, encoder_mask, decoder_mask):
"""Forward pass of TransformerEncoderLayer.
@ -197,20 +210,19 @@ class TransformerDecoderLayer(nn.Layer):
cross_attn_weights : Tensor [shape=(batch_size, n_heads, time_steps_q, time_steps_k)]
Decoder-encoder cross attention.
"""
context_vector, self_attn_weights = self.self_mha(q, q, q, decoder_mask)
context_vector, self_attn_weights = self.self_mha(q, q, q,
decoder_mask)
q = self.layer_norm1(
F.dropout(q + context_vector,
self.dropout,
training=self.training))
context_vector, cross_attn_weights = self.cross_mha(q, k, v, encoder_mask)
F.dropout(
q + context_vector, self.dropout, training=self.training))
context_vector, cross_attn_weights = self.cross_mha(q, k, v,
encoder_mask)
q = self.layer_norm2(
F.dropout(q + context_vector,
self.dropout,
training=self.training))
F.dropout(
q + context_vector, self.dropout, training=self.training))
q = self.layer_norm3(
F.dropout(q + self.ffn(q),
self.dropout,
training=self.training))
F.dropout(
q + self.ffn(q), self.dropout, training=self.training))
return q, self_attn_weights, cross_attn_weights

14
parakeet/training/__init__.py
View File

@ -1,2 +1,16 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from parakeet.training.cli import *
from parakeet.training.experiment import *

17
parakeet/training/cli.py
View File

@ -1,5 +1,20 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
def default_argument_parser():
r"""A simple yet genral argument parser for experiments with parakeet.
@ -46,5 +61,5 @@ def default_argument_parser():
# overwrite extra config and default config
parser.add_argument("--opts", nargs=argparse.REMAINDER, help="options to overwrite --config file and the default config, passing in KEY VALUE pairs")
# yapd: enable
return parser

26
parakeet/training/default_config.py
View File

@ -1,12 +1,26 @@
from yacs.config import CfgNode
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from yacs.config import CfgNode
_C = CfgNode(
dict(
valid_interval=1000, # validation
save_interval=10000, # checkpoint
max_iteration=900000, # max iteration to train
)
)
valid_interval=1000, # validation
save_interval=10000, # checkpoint
max_iteration=900000, # max iteration to train
))
def get_default_training_config():
return _C.clone()

1
parakeet/training/experiment.py
View File

@ -27,6 +27,7 @@ from parakeet.utils import checkpoint, mp_tools
__all__ = ["ExperimentBase"]
class ExperimentBase(object):
"""
An experiment template in order to structure the training code and take

15
parakeet/utils/checkpoint.py
View File

@ -45,6 +45,7 @@ def _load_latest_checkpoint(checkpoint_dir: str) -> int:
return iteration
def _save_checkpoint(checkpoint_dir: str, iteration: int):
"""Save the iteration number of the latest model to be checkpointed.
@ -60,6 +61,7 @@ def _save_checkpoint(checkpoint_dir: str, iteration: int):
with open(checkpoint_record, "wt") as handle:
handle.write("model_checkpoint_path: step-{}".format(iteration))
def load_parameters(model,
optimizer=None,
checkpoint_dir=None,
@ -97,18 +99,19 @@ def load_parameters(model,
params_path = checkpoint_path + ".pdparams"
model_dict = paddle.load(params_path)
model.set_state_dict(model_dict)
print("[checkpoint] Rank {}: loaded model from {}".format(
local_rank, params_path))
print("[checkpoint] Rank {}: loaded model from {}".format(local_rank,
params_path))
optimizer_path = checkpoint_path + ".pdopt"
if optimizer and os.path.isfile(optimizer_path):
optimizer_dict = paddle.load(optimizer_path)
optimizer.set_state_dict(optimizer_dict)
print("[checkpoint] Rank {}: loaded optimizer state from {}".
format(local_rank, optimizer_path))
print("[checkpoint] Rank {}: loaded optimizer state from {}".format(
local_rank, optimizer_path))
return iteration
@mp_tools.rank_zero_only
def save_parameters(checkpoint_dir, iteration, model, optimizer=None):
"""Checkpoint the latest trained model parameters.
@ -124,7 +127,7 @@ def save_parameters(checkpoint_dir, iteration, model, optimizer=None):
None
"""
checkpoint_path = os.path.join(checkpoint_dir, "step-{}".format(iteration))
model_dict = model.state_dict()
params_path = checkpoint_path + ".pdparams"
paddle.save(model_dict, params_path)

14
parakeet/utils/internals.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
from paddle.framework import core

4
parakeet/utils/layer_tools.py
View File

@ -28,6 +28,7 @@ def summary(layer: nn.Layer):
print("layer has {} parameters, {} elements.".format(num_params,
num_elements))
def gradient_norm(layer: nn.Layer):
grad_norm_dict = {}
for name, param in layer.state_dict().items():
@ -36,6 +37,7 @@ def gradient_norm(layer: nn.Layer):
grad_norm_dict[name] = np.linalg.norm(grad) / grad.size
return grad_norm_dict
def recursively_remove_weight_norm(layer: nn.Layer):
for layer in layer.sublayers():
try:
@ -44,10 +46,12 @@ def recursively_remove_weight_norm(layer: nn.Layer):
# ther is not weight norm hoom in this layer
pass
def freeze(layer: nn.Layer):
for param in layer.parameters():
param.trainable = False
def unfreeze(layer: nn.Layer):
for param in layer.parameters():
param.trainable = True

21
parakeet/utils/mp_tools.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
from paddle import distributed as dist
from functools import wraps
@ -11,11 +25,8 @@ def rank_zero_only(func):
@wraps(func)
def wrapper(*args, **kwargs):
if local_rank != 0:
return
return
result = func(*args, **kwargs)
return result
return wrapper

25
parakeet/utils/scheduler.py
View File

@ -1,3 +1,17 @@
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
__all__ = ["SchedulerBase", "Constant", "PieceWise", "StepWise"]
@ -24,7 +38,7 @@ class PieceWise(SchedulerBase):
self.xs = [item[0] for item in anchors]
self.ys = [item[1] for item in anchors]
self.num_anchors = len(self.xs)
def __call__(self, step):
i = 0
for x in self.xs:
@ -34,8 +48,8 @@ class PieceWise(SchedulerBase):
return self.ys[0]
if i == self.num_anchors:
return self.ys[-1]
k = (self.ys[i] - self.ys[i-1]) / (self.xs[i] - self.xs[i-1])
out = self.ys[i-1] + (step - self.xs[i-1]) * k
k = (self.ys[i] - self.ys[i - 1]) / (self.xs[i] - self.xs[i - 1])
out = self.ys[i - 1] + (step - self.xs[i - 1]) * k
return out
@ -47,7 +61,7 @@ class StepWise(SchedulerBase):
self.xs = [item[0] for item in anchors]
self.ys = [item[1] for item in anchors]
self.num_anchors = len(self.xs)
def __call__(self, step):
i = 0
for x in self.xs:
@ -58,5 +72,4 @@ class StepWise(SchedulerBase):
return self.ys[-1]
if i == 0:
return self.ys[0]
return self.ys[i-1]
return self.ys[i - 1]

14
setup.py
View File

@ -48,7 +48,6 @@ setup_info = dict(
description='Speech synthesis tools and models based on Paddlepaddle',
long_description=long_description,
license='Apache 2',
python_requires='>=3.6',
install_requires=[
'numpy',
@ -71,23 +70,18 @@ setup_info = dict(
'yacs',
'tensorboardX',
],
extras_require={
'doc': ["sphinx", "sphinx-rtd-theme", "numpydoc"],
},
extras_require={'doc': ["sphinx", "sphinx-rtd-theme", "numpydoc"], },
# Package info
packages=find_packages(exclude=('tests', 'tests.*')),
zip_safe=True,
classifiers = [
zip_safe=True,
classifiers=[
'Development Status :: 4 - Beta',
'Intended Audience :: Developers',
'Topic :: Scientific/Engineering :: Artificial Intelligence'
'License :: OSI Approved :: Apache2 License',
'Programming Language :: Python :: 3.6',
'Programming Language :: Python :: 3.7',
],
)
], )
setup(**setup_info)