msb-public
/
PaddleSpeech
mirror of https://github.com/PaddlePaddle/PaddleSpeech
1
0
Code Issues Projects Releases Wiki Activity

Merge branch 'develop' of github.com:PaddlePaddle/DeepSpeech into add_mbmelgan

Browse Source
pull/935/head
TianYuan 4 years ago
parent 88668513b1 3417889332
commit 36d60a717e
81 changed files with 4519 additions and 513 deletions
Show Stats Download Patch File Download Diff File

22
.mergify.yml
Unescape View File

@ -39,12 +39,30 @@ pull_request_rules:
actions:
label:
remove: ["conflicts"]
- name: "auto add label=enhancement"
- name: "auto add label=S2T"
conditions:
- files~=^deepspeech/
actions:
label:
add: ["enhancement"]
add: ["S2T"]
- name: "auto add label=T2S"
conditions:
- files~=^parakeet/
actions:
label:
add: ["T2S"]
- name: "auto add label=Audio"
conditions:
- files~=^paddleaudio/
actions:
label:
add: ["Audio"]
- name: "auto add label=TextProcess"
conditions:
- files~=^text_processing/
actions:
label:
add: ["TextProcess"]
- name: "auto add label=Example"
conditions:
- files~=^examples/

7
.readthedocs.yml
Unescape View File

@ -7,7 +7,7 @@ version: 2
# Build documentation in the docs/ directory with Sphinx
sphinx:
configuration: docs/src/conf.py
configuration: docs/source/conf.py
# Build documentation with MkDocs
#mkdocs:
@ -20,11 +20,6 @@ formats: []
python:
version: 3.7
install:
- method: pip
path: .
extra_requirements:
- doc
- requirements: docs/requirements.txt

84
README.md
Unescape View File

@ -10,10 +10,9 @@ English | [简体中文](README_ch.md)
<div align="center">
<h3>
<a href="https://github.com/Mingxue-Xu/DeepSpeech#quick-start"> Quick Start </a>
| <a href="https://github.com/Mingxue-Xu/DeepSpeech#tutorials"> Tutorials </a>
| <a href="https://github.com/Mingxue-Xu/DeepSpeech#model-list"> Models List </a>
<a href="#quick-start"> Quick Start </a>
| <a href="#tutorials"> Tutorials </a>
| <a href="#model-list"> Models List </a>
</div>
------------------------------------------------------------------------------------
@ -27,37 +26,31 @@ how they can install it,
how they can use it
-->
**PaddleSpeech** is an open-source toolkit on [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform for two critical tasks in Speech - **Automatic Speech Recognition (ASR)** and **Text-To-Speech Synthesis (TTS)**, with modules involving state-of-art and influential models.
**PaddleSpeech** is an open-source toolkit on [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform for a variety of critical tasks in speech, with state-of-art and influential models.
Via the easy-to-use, efficient, flexible and scalable implementation, our vision is to empower both industrial application and academic research, including training, inference & testing module, and deployment. Besides, this toolkit also features at:
- **Fast and Light-weight**: we provide a high-speed and ultra-lightweight model that is convenient for industrial deployment.
Via the easy-to-use, efficient, flexible and scalable implementation, our vision is to empower both industrial application and academic research, including training, inference & testing modules, and deployment process. To be more specific, this toolkit features at:
- **Fast and Light-weight**: we provide high-speed and ultra-lightweight models that are convenient for industrial deployment.
- **Rule-based Chinese frontend**: our frontend contains Text Normalization (TN) and Grapheme-to-Phoneme (G2P, including Polyphone and Tone Sandhi). Moreover, we use self-defined linguistic rules to adapt Chinese context.
- **Varieties of Functions that Vitalize Research**:
- *Integration of mainstream models and datasets*: the toolkit implements modules that participate in the whole pipeline of both ASR and TTS, and uses datasets like LibriSpeech, LJSpeech, AIShell, etc. See also [model lists](#models-list) for more details.
- *Support of ASR streaming and non-streaming data*: This toolkit contains non-streaming/streaming models like [DeepSpeech2](http://proceedings.mlr.press/v48/amodei16.pdf), [Transformer](https://arxiv.org/abs/1706.03762), [Conformer](https://arxiv.org/abs/2005.08100) and [U2](https://arxiv.org/pdf/2012.05481.pdf).
- **Varieties of Functions that Vitalize both Industrial and Academia**:
- *Implementation of critical audio tasks*: this toolkit contains audio functions like Speech Translation (ST), Automatic Speech Recognition (ASR), Text-To-Speech Synthesis (TTS), Voice Cloning(VC), Punctuation Restoration, etc.
- *Integration of mainstream models and datasets*: the toolkit implements modules that participate in the whole pipeline of the speech tasks, and uses mainstream datasets like LibriSpeech, LJSpeech, AIShell, CSMSC, etc. See also [model lists](#models-list) for more details.
- *Cross-domain application*: as an extension of the application of traditional audio tasks, we combine the aforementioned tasks with other fields like NLP.
Let's install PaddleSpeech with only a few lines of code!
>Note: The official name is still deepspeech. 2021/10/26
``` shell
# 1. Install essential libraries and paddlepaddle first.
# install prerequisites
sudo apt-get install -y sox pkg-config libflac-dev libogg-dev libvorbis-dev libboost-dev swig python3-dev libsndfile1
# `pip install paddlepaddle-gpu` instead if you are using GPU.
pip install paddlepaddle
# 2.Then install PaddleSpeech.
If you are using Ubuntu, PaddleSpeech can be set up with pip installation (with root privilege).
```shell
git clone https://github.com/PaddlePaddle/DeepSpeech.git
cd DeepSpeech
pip install -e .
```
## Table of Contents
The contents of this README is as follow:
- [Alternative Installation](#installation)
- [Alternative Installation](#alternative-installation)
- [Quick Start](#quick-start)
- [Models List](#models-list)
- [Tutorials](#tutorials)
@ -75,10 +68,13 @@ The base environment in this page is
If you want to set up PaddleSpeech in other environment, please see the [ASR installation](docs/source/asr/install.md) and [TTS installation](docs/source/tts/install.md) documents for all the alternatives.
## Quick Start
> Note: the current links to `English ASR` and `English TTS` are not valid.
> Note: `ckptfile` should be replaced by real path that represents files or folders later. Similarly, `exp/default` is the folder that contains the pretrained models.
Just a quick test of our functions: [English ASR](link/hubdetail?name=deepspeech2_aishell&en_category=AutomaticSpeechRecognition) and [English TTS](link/hubdetail?name=fastspeech2_baker&en_category=TextToSpeech) by typing message or upload your own audio file.
Try a tiny ASR DeepSpeech2 model training on toy set of LibriSpeech:
Developers can have a try of our model with only a few lines of code.
A tiny *ASR* DeepSpeech2 model training on toy set of LibriSpeech:
```shell
cd examples/tiny/s0/
@ -90,12 +86,13 @@ bash local/data.sh
bash local/test.sh conf/deepspeech2.yaml ckptfile offline
```
For TTS, try FastSpeech2 on LJSpeech:
- Download LJSpeech-1.1 from the [ljspeech official website](https://keithito.com/LJ-Speech-Dataset/) and our prepared durations for fastspeech2 [ljspeech_alignment](https://paddlespeech.bj.bcebos.com/MFA/LJSpeech-1.1/ljspeech_alignment.tar.gz).
For *TTS*, try FastSpeech2 on LJSpeech:
- Download LJSpeech-1.1 from the [ljspeech official website](https://keithito.com/LJ-Speech-Dataset/), our prepared durations for fastspeech2 [ljspeech_alignment](https://paddlespeech.bj.bcebos.com/MFA/LJSpeech-1.1/ljspeech_alignment.tar.gz).
- The pretrained models are seperated into two parts: [fastspeech2_nosil_ljspeech_ckpt](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_ljspeech_ckpt_0.5.zip) and [pwg_ljspeech_ckpt](https://paddlespeech.bj.bcebos.com/Parakeet/pwg_ljspeech_ckpt_0.5.zip). Please download then unzip to `./model/fastspeech2` and `./model/pwg` respectively.
- Assume your path to the dataset is `~/datasets/LJSpeech-1.1` and `./ljspeech_alignment` accordingly, preprocess your data and then use our pretrained model to synthesize:
```shell
bash ./local/preprocess.sh conf/default.yaml
bash ./local/synthesize_e2e.sh conf/default.yaml exp/default ckptfile
bash ./local/synthesize_e2e.sh conf/default.yaml ./model/fastspeech2/snapshot_iter_100000.pdz ./model/pwg/pwg_snapshot_iter_400000.pdz
```
@ -104,14 +101,17 @@ If you want to try more functions like training and tuning, please see [ASR gett
## Models List
PaddleSpeech supports a series of most popular models, summarized in [released models](./docs/source/released_model.md) with available pretrained models.
PaddleSpeech ASR supports a lot of mainstream models, which are summarized as follow. For more information, please refer to [ASR Models](./docs/source/asr/released_model.md).
ASR module contains *Acoustic Model* and *Language Model*, with the following details:
<!---
The current hyperlinks redirect to [Previous Parakeet](https://github.com/PaddlePaddle/Parakeet/tree/develop/examples).
-->
> Note: The `Link` should be code path rather than download links.
<table>
<thead>
<tr>
@ -125,7 +125,7 @@ The current hyperlinks redirect to [Previous Parakeet](https://github.com/Paddle
<tr>
<td rowspan="6">Acoustic Model</td>
<td rowspan="4" >Aishell</td>
<td >2 Conv + 5 LSTM layers with only forward direction </td>
<td >2 Conv + 5 LSTM layers with only forward direction</td>
<td>
<a href = "https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds_online.5rnn.debug.tar.gz">Ds2 Online Aishell Model</a>
</td>
@ -200,7 +200,7 @@ PaddleSpeech TTS mainly contains three modules: *Text Frontend*, *Acoustic Model
<td> Text Frontend</td>
<td colspan="2"> &emsp; </td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/other/text_frontend">chinese-fronted</a>
<a href = "./examples/other/text_frontend">chinese-fronted</a>
</td>
</tr>
<tr>
@ -208,41 +208,41 @@ PaddleSpeech TTS mainly contains three modules: *Text Frontend*, *Acoustic Model
<td >Tacotron2</td>
<td rowspan="2" >LJSpeech</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts0">tacotron2-vctk</a>
<a href = "./examples/ljspeech/tts0">tacotron2-vctk</a>
</td>
</tr>
<tr>
<td>TransformerTTS</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts1">transformer-ljspeech</a>
<a href = "./examples/ljspeech/tts1">transformer-ljspeech</a>
</td>
</tr>
<tr>
<td>SpeedySpeech</td>
<td>CSMSC</td>
<td >
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts2">speedyspeech-csmsc</a>
<a href = "./examples/csmsc/tts2">speedyspeech-csmsc</a>
</td>
</tr>
<tr>
<td rowspan="4">FastSpeech2</td>
<td>AISHELL-3</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/aishell3/tts3">fastspeech2-aishell3</a>
<a href = "./examples/aishell3/tts3">fastspeech2-aishell3</a>
</td>
</tr>
<tr>
<td>VCTK</td>
<td> <a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/vctk/tts3">fastspeech2-vctk</a> </td>
<td> <a href = "./examples/vctk/tts3">fastspeech2-vctk</a> </td>
</tr>
<tr>
<td>LJSpeech</td>
<td> <a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts3">fastspeech2-ljspeech</a> </td>
<td> <a href = "./examples/ljspeech/tts3">fastspeech2-ljspeech</a> </td>
</tr>
<tr>
<td>CSMSC</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts3">fastspeech2-csmsc</a>
<a href = "./examples/csmsc/tts3">fastspeech2-csmsc</a>
</td>
</tr>
<tr>
@ -250,26 +250,26 @@ PaddleSpeech TTS mainly contains three modules: *Text Frontend*, *Acoustic Model
<td >WaveFlow</td>
<td >LJSpeech</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/voc0">waveflow-ljspeech</a>
<a href = "./examples/ljspeech/voc0">waveflow-ljspeech</a>
</td>
</tr>
<tr>
<td rowspan="3">Parallel WaveGAN</td>
<td >LJSpeech</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/voc1">PWGAN-ljspeech</a>
<a href = "./examples/ljspeech/voc1">PWGAN-ljspeech</a>
</td>
</tr>
<tr>
<td >VCTK</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/vctk/voc1">PWGAN-vctk</a>
<a href = "./examples/vctk/voc1">PWGAN-vctk</a>
</td>
</tr>
<tr>
<td >CSMSC</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/voc1">PWGAN-csmsc</a>
<a href = "./examples/csmsc/voc1">PWGAN-csmsc</a>
</td>
</tr>
<tr>
@ -277,14 +277,14 @@ PaddleSpeech TTS mainly contains three modules: *Text Frontend*, *Acoustic Model
<td>GE2E</td>
<td >AISHELL-3, etc.</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/other/ge2e">ge2e</a>
<a href = "./examples/other/ge2e">ge2e</a>
</td>
</tr>
<tr>
<td>GE2E + Tactron2</td>
<td>AISHELL-3</td>
<td>
<a href = "https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/aishell3/vc0">ge2e-tactron2-aishell3</a>
<a href = "./examples/aishell3/vc0">ge2e-tactron2-aishell3</a>
</td>
</td>
</tr>

128
deepspeech/exps/deepspeech2/model.py
Unescape View File

@ -167,6 +167,11 @@ class DeepSpeech2Trainer(Trainer):
logger.info(f"{model}")
layer_tools.print_params(model, logger.info)
self.model = model
logger.info("Setup model!")
if not self.train:
return
grad_clip = ClipGradByGlobalNormWithLog(
config.training.global_grad_clip)
@ -180,74 +185,77 @@ class DeepSpeech2Trainer(Trainer):
weight_decay=paddle.regularizer.L2Decay(
config.training.weight_decay),
grad_clip=grad_clip)
self.model = model
self.optimizer = optimizer
self.lr_scheduler = lr_scheduler
logger.info("Setup model/optimizer/lr_scheduler!")
logger.info("Setup optimizer/lr_scheduler!")
def setup_dataloader(self):
config = self.config.clone()
config.defrost()
config.collator.keep_transcription_text = False
config.data.manifest = config.data.train_manifest
train_dataset = ManifestDataset.from_config(config)
config.data.manifest = config.data.dev_manifest
dev_dataset = ManifestDataset.from_config(config)
config.data.manifest = config.data.test_manifest
test_dataset = ManifestDataset.from_config(config)
if self.parallel:
batch_sampler = SortagradDistributedBatchSampler(
if self.train:
# train
config.data.manifest = config.data.train_manifest
train_dataset = ManifestDataset.from_config(config)
if self.parallel:
batch_sampler = SortagradDistributedBatchSampler(
train_dataset,
batch_size=config.collator.batch_size,
num_replicas=None,
rank=None,
shuffle=True,
drop_last=True,
sortagrad=config.collator.sortagrad,
shuffle_method=config.collator.shuffle_method)
else:
batch_sampler = SortagradBatchSampler(
train_dataset,
shuffle=True,
batch_size=config.collator.batch_size,
drop_last=True,
sortagrad=config.collator.sortagrad,
shuffle_method=config.collator.shuffle_method)
config.collator.keep_transcription_text = False
collate_fn_train = SpeechCollator.from_config(config)
self.train_loader = DataLoader(
train_dataset,
batch_size=config.collator.batch_size,
num_replicas=None,
rank=None,
shuffle=True,
drop_last=True,
sortagrad=config.collator.sortagrad,
shuffle_method=config.collator.shuffle_method)
batch_sampler=batch_sampler,
collate_fn=collate_fn_train,
num_workers=config.collator.num_workers)
# dev
config.data.manifest = config.data.dev_manifest
dev_dataset = ManifestDataset.from_config(config)
config.collator.augmentation_config = ""
config.collator.keep_transcription_text = False
collate_fn_dev = SpeechCollator.from_config(config)
self.valid_loader = DataLoader(
dev_dataset,
batch_size=int(config.collator.batch_size),
shuffle=False,
drop_last=False,
collate_fn=collate_fn_dev,
num_workers=config.collator.num_workers)
logger.info("Setup train/valid Dataloader!")
else:
batch_sampler = SortagradBatchSampler(
train_dataset,
shuffle=True,
batch_size=config.collator.batch_size,
drop_last=True,
sortagrad=config.collator.sortagrad,
shuffle_method=config.collator.shuffle_method)
collate_fn_train = SpeechCollator.from_config(config)
config.collator.augmentation_config = ""
collate_fn_dev = SpeechCollator.from_config(config)
config.collator.keep_transcription_text = True
config.collator.augmentation_config = ""
collate_fn_test = SpeechCollator.from_config(config)
self.train_loader = DataLoader(
train_dataset,
batch_sampler=batch_sampler,
collate_fn=collate_fn_train,
num_workers=config.collator.num_workers)
self.valid_loader = DataLoader(
dev_dataset,
batch_size=int(config.collator.batch_size),
shuffle=False,
drop_last=False,
collate_fn=collate_fn_dev,
num_workers=config.collator.num_workers)
self.test_loader = DataLoader(
test_dataset,
batch_size=config.decoding.batch_size,
shuffle=False,
drop_last=False,
collate_fn=collate_fn_test,
num_workers=config.collator.num_workers)
logger.info("Setup train/valid/test Dataloader!")
# test
config.data.manifest = config.data.test_manifest
test_dataset = ManifestDataset.from_config(config)
config.collator.augmentation_config = ""
config.collator.keep_transcription_text = True
collate_fn_test = SpeechCollator.from_config(config)
self.test_loader = DataLoader(
test_dataset,
batch_size=config.decoding.batch_size,
shuffle=False,
drop_last=False,
collate_fn=collate_fn_test,
num_workers=config.collator.num_workers)
logger.info("Setup test Dataloader!")
class DeepSpeech2Tester(DeepSpeech2Trainer):

2
deepspeech/exps/u2/model.py
Unescape View File

@ -172,7 +172,7 @@ class U2Trainer(Trainer):
dist.get_rank(), total_loss / num_seen_utts))
return total_loss, num_seen_utts
def train(self):
def do_train(self):
"""The training process control by step."""
# !!!IMPORTANT!!!
# Try to export the model by script, if fails, we should refine

2
deepspeech/exps/u2_kaldi/model.py
Unescape View File

@ -173,7 +173,7 @@ class U2Trainer(Trainer):
dist.get_rank(), total_loss / num_seen_utts))
return total_loss, num_seen_utts
def train(self):
def do_train(self):
"""The training process control by step."""
# !!!IMPORTANT!!!
# Try to export the model by script, if fails, we should refine

2
deepspeech/exps/u2_st/model.py
Unescape View File

@ -184,7 +184,7 @@ class U2STTrainer(Trainer):
dist.get_rank(), total_loss / num_seen_utts))
return total_loss, num_seen_utts
def train(self):
def do_train(self):
"""The training process control by step."""
# !!!IMPORTANT!!!
# Try to export the model by script, if fails, we should refine

10
deepspeech/training/trainer.py
Unescape View File

@ -134,6 +134,10 @@ class Trainer():
logger.info(
f"Benchmark reset batch-size: {self.args.benchmark_batch_size}")
@property
def train(self):
return self._train
@contextmanager
def eval(self):
self._train = False
@ -248,7 +252,7 @@ class Trainer():
sys.exit(
f"Reach benchmark-max-step: {self.args.benchmark_max_step}")
def train(self):
def do_train(self):
"""The training process control by epoch."""
self.before_train()
@ -321,7 +325,7 @@ class Trainer():
"""
try:
with Timer("Training Done: {}"):
self.train()
self.do_train()
except KeyboardInterrupt:
exit(-1)
finally:
@ -432,7 +436,7 @@ class Trainer():
beginning of the experiment.
"""
config_file = self.config_dir / "config.yaml"
if self._train and config_file.exists():
if self.train and config_file.exists():
time_stamp = time.strftime("%Y_%m_%d_%H_%M_%s", time.gmtime())
target_path = self.config_dir / ".".join(
[time_stamp, "config.yaml"])

BIN
docs/images/paddle.png
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 4.9 KiB

BIN
docs/images/tuning_error_surface.png
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 108 KiB

7
docs/requirements.txt
Unescape View File

@ -0,0 +1,7 @@
myst-parser
numpydoc
recommonmark>=0.5.0
sphinx
sphinx-autobuild
sphinx-markdown-tables
sphinx_rtd_theme

84
docs/source/asr/deepspeech_architecture.md → docs/source/asr/models_introduction.md
Unescape View File

@ -1,6 +1,5 @@
# Deepspeech2
## Streaming
# Models introduction
## Streaming DeepSpeech2
The implemented arcitecure of Deepspeech2 online model is based on [Deepspeech2 model](https://arxiv.org/pdf/1512.02595.pdf) with some changes.
The model is mainly composed of 2D convolution subsampling layer and stacked single direction rnn layers.
@ -14,8 +13,8 @@ In addition, the training process and the testing process are also introduced.
The arcitecture of the model is shown in Fig.1.
<p align="center">
<img src="../../images/ds2onlineModel.png" width=800>
<br/>Fig.1 The Arcitecture of deepspeech2 online model
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/ds2onlineModel.png" width=800>
<br/>Fig.1 The Arcitecture of deepspeech2 online model
</p>
@ -23,13 +22,13 @@ The arcitecture of the model is shown in Fig.1.
#### Vocabulary
For English data, the vocabulary dictionary is composed of 26 English characters with " ' ", space, \<blank\> and \<eos\>. The \<blank\> represents the blank label in CTC, the \<unk\> represents the unknown character and the \<eos\> represents the start and the end characters. For mandarin, the vocabulary dictionary is composed of chinese characters statisticed from the training set and three additional characters are added. The added characters are \<blank\>, \<unk\> and \<eos\>. For both English and mandarin data, we set the default indexs that \<blank\>=0, \<unk\>=1 and \<eos\>= last index.
```
# The code to build vocabulary
cd examples/aishell/s0
python3 ../../../utils/build_vocab.py \
--unit_type="char" \
--count_threshold=0 \
--vocab_path="data/vocab.txt" \
--manifest_paths "data/manifest.train.raw" "data/manifest.dev.raw"
# The code to build vocabulary
cd examples/aishell/s0
python3 ../../../utils/build_vocab.py \
--unit_type="char" \
--count_threshold=0 \
--vocab_path="data/vocab.txt" \
--manifest_paths "data/manifest.train.raw" "data/manifest.dev.raw"
# vocabulary for aishell dataset (Mandarin)
vi examples/aishell/s0/data/vocab.txt
@ -41,29 +40,29 @@ vi examples/librispeech/s0/data/vocab.txt
#### CMVN
For CMVN, a subset or the full of traininig set is chosed and be used to compute the feature mean and std.
```
# The code to compute the feature mean and std
# The code to compute the feature mean and std
cd examples/aishell/s0
python3 ../../../utils/compute_mean_std.py \
--manifest_path="data/manifest.train.raw" \
--spectrum_type="linear" \
--delta_delta=false \
--stride_ms=10.0 \
--window_ms=20.0 \
--sample_rate=16000 \
--use_dB_normalization=True \
--num_samples=2000 \
--num_workers=10 \
--output_path="data/mean_std.json"
--manifest_path="data/manifest.train.raw" \
--spectrum_type="linear" \
--delta_delta=false \
--stride_ms=10.0 \
--window_ms=20.0 \
--sample_rate=16000 \
--use_dB_normalization=True \
--num_samples=2000 \
--num_workers=10 \
--output_path="data/mean_std.json"
```
#### Feature Extraction
For feature extraction, three methods are implemented, which are linear (FFT without using filter bank), fbank and mfcc.
Currently, the released deepspeech2 online model use the linear feature extraction method.
```
The code for feature extraction
vi deepspeech/frontend/featurizer/audio_featurizer.py
```
For feature extraction, three methods are implemented, which are linear (FFT without using filter bank), fbank and mfcc.
Currently, the released deepspeech2 online model use the linear feature extraction method.
```
The code for feature extraction
vi deepspeech/frontend/featurizer/audio_featurizer.py
```
### Encoder
The encoder is composed of two 2D convolution subsampling layers and a number of stacked single direction rnn layers. The 2D convolution subsampling layers extract feature representation from the raw audio feature and reduce the length of audio feature at the same time. After passing through the convolution subsampling layers, then the feature representation are input into the stacked rnn layers. For the stacked rnn layers, LSTM cell and GRU cell are provided to use. Adding one fully connected (fc) layer after the stacked rnn layers is optional. If the number of stacked rnn layers is less than 5, adding one fc layer after stacked rnn layers is recommand.
@ -84,11 +83,11 @@ vi deepspeech/models/ds2_online/deepspeech2.py
vi deepspeech/modules/ctc.py
```
## Training Process
### Training Process
Using the command below, you can train the deepspeech2 online model.
```
cd examples/aishell/s0
bash run.sh --stage 0 --stop_stage 2 --model_type online --conf_path conf/deepspeech2_online.yaml
cd examples/aishell/s0
bash run.sh --stage 0 --stop_stage 2 --model_type online --conf_path conf/deepspeech2_online.yaml
```
The detail commands are:
```
@ -127,11 +126,11 @@ fi
By using the command above, the training process can be started. There are 5 stages in "run.sh", and the first 3 stages are used for training process. The stage 0 is used for data preparation, in which the dataset will be downloaded, and the manifest files of the datasets, vocabulary dictionary and CMVN file will be generated in "./data/". The stage 1 is used for training the model, the log files and model checkpoint is saved in "exp/deepspeech2_online/". The stage 2 is used to generated final model for predicting by averaging the top-k model parameters based on validation loss.
## Testing Process
### Testing Process
Using the command below, you can test the deepspeech2 online model.
```
bash run.sh --stage 3 --stop_stage 5 --model_type online --conf_path conf/deepspeech2_online.yaml
```
```
bash run.sh --stage 3 --stop_stage 5 --model_type online --conf_path conf/deepspeech2_online.yaml
```
The detail commands are:
```
conf_path=conf/deepspeech2_online.yaml
@ -139,7 +138,7 @@ avg_num=1
model_type=online
avg_ckpt=avg_${avg_num}
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n
CUDA_VISIBLE_DEVICES=2 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type}|| exit -1
fi
@ -156,19 +155,16 @@ fi
```
After the training process, we use stage 3,4,5 for testing process. The stage 3 is for testing the model generated in the stage 2 and provided the CER index of the test set. The stage 4 is for transforming the model from dynamic graph to static graph by using "paddle.jit" library. The stage 5 is for testing the model in static graph.
## Non-Streaming
## Non-Streaming DeepSpeech2
The deepspeech2 offline model is similarity to the deepspeech2 online model. The main difference between them is the offline model use the stacked bi-directional rnn layers while the online model use the single direction rnn layers and the fc layer is not used. For the stacked bi-directional rnn layers in the offline model, the rnn cell and gru cell are provided to use.
The arcitecture of the model is shown in Fig.2.
<p align="center">
<img src="../../images/ds2offlineModel.png" width=800>
<br/>Fig.2 The Arcitecture of deepspeech2 offline model
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/ds2offlineModel.png" width=800>
<br/>Fig.2 The Arcitecture of deepspeech2 offline model
</p>
For data preparation and decoder, the deepspeech2 offline model is same with the deepspeech2 online model.
The code of encoder and decoder for deepspeech2 offline model is in:
@ -180,7 +176,7 @@ The training process and testing process of deepspeech2 offline model is very si
Only some changes should be noticed.
For training and testing, the "model_type" and the "conf_path" must be set.
```
```
# Training offline
cd examples/aishell/s0
bash run.sh --stage 0 --stop_stage 2 --model_type offline --conf_path conf/deepspeech2.yaml

31
docs/source/asr/getting_started.md → docs/source/asr/quick_start.md
Unescape View File

@ -1,5 +1,4 @@
# Getting Started
# Quick Start of Speech-To-Text
Several shell scripts provided in `./examples/tiny/local` will help us to quickly give it a try, for most major modules, including data preparation, model training, case inference and model evaluation, with a few public dataset (e.g. [LibriSpeech](http://www.openslr.org/12/), [Aishell](http://www.openslr.org/33)). Reading these examples will also help you to understand how to make it work with your own data.
Some of the scripts in `./examples` are not configured with GPUs. If you want to train with 8 GPUs, please modify `CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7`. If you don't have any GPU available, please set `CUDA_VISIBLE_DEVICES=` to use CPUs instead. Besides, if out-of-memory problem occurs, just reduce `batch_size` to fit.
@ -11,68 +10,52 @@ Let's take a tiny sampled subset of [LibriSpeech dataset](http://www.openslr.org
```bash
cd examples/tiny
```
Notice that this is only a toy example with a tiny sampled subset of LibriSpeech. If you would like to try with the complete dataset (would take several days for training), please go to `examples/librispeech` instead.
- Source env
```bash
source path.sh
```
**Must do this before starting do anything.**
Set `MAIN_ROOT` as project dir. Using defualt `deepspeech2` model as default, you can change this in the script.
**Must do this before you start to do anything.**
Set `MAIN_ROOT` as project dir. Using defualt `deepspeech2` model as `MODEL`, you can change this in the script.
- Main entrypoint
```bash
bash run.sh
```
This just a demo, please make sure every `step` is work fine when do next `step`.
This is just a demo, please make sure every `step` works well before next `step`.
More detailed information are provided in the following sections. Wish you a happy journey with the *DeepSpeech on PaddlePaddle* ASR engine!
## Training a model
The key steps of training for Mandarin language are same to that of English language and we have also provided an example for Mandarin training with Aishell in ```examples/aishell/local```. As mentioned above, please execute ```sh data.sh```, ```sh train.sh```, ```sh test.sh``` and ```sh infer.sh``` to do data preparation, training, testing and inference correspondingly. We have also prepared a pre-trained model (downloaded by local/download_model.sh) for users to try with ```sh infer_golden.sh``` and ```sh test_golden.sh```. Notice that, different from English LM, the Mandarin LM is character-based and please run ```local/tune.sh``` to find an optimal setting.
The key steps of training for Mandarin language are same to that of English language and we have also provided an example for Mandarin training with Aishell in ```examples/aishell/local```. As mentioned above, please execute ```sh data.sh```, ```sh train.sh```, ```sh test.sh```and ```sh infer.sh```to do data preparation, training, testing and inference correspondingly. We have also prepared a pre-trained model (downloaded by local/download_model.sh) for users to try with ```sh infer_golden.sh```and ```sh test_golden.sh```. Notice that, different from English LM, the Mandarin LM is character-based and please run ```local/tune.sh```to find an optimal setting.
## Speech-to-text Inference
An inference module caller `infer.py` is provided to infer, decode and visualize speech-to-text results for several given audio clips. It might help to have an intuitive and qualitative evaluation of the ASR model's performance.
```bash
CUDA_VISIBLE_DEVICES=0 bash local/infer.sh
```
We provide two types of CTC decoders: *CTC greedy decoder* and *CTC beam search decoder*. The *CTC greedy decoder* is an implementation of the simple best-path decoding algorithm, selecting at each timestep the most likely token, thus being greedy and locally optimal. The [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) otherwise utilizes a heuristic breadth-first graph search for reaching a near global optimality; it also requires a pre-trained KenLM language model for better scoring and ranking. The decoder type can be set with argument `decoding_method`.
## Evaluate a Model
To evaluate a model's performance quantitatively, please run:
```bash
CUDA_VISIBLE_DEVICES=0 bash local/test.sh
```
The error rate (default: word error rate; can be set with `error_rate_type`) will be printed.
For more help on arguments:
## Hyper-parameters Tuning
The hyper-parameters $\alpha$ (language model weight) and $\beta$ (word insertion weight) for the [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) often have a significant impact on the decoder's performance. It would be better to re-tune them on the validation set when the acoustic model is renewed.
`tune.py` performs a 2-D grid search over the hyper-parameter $\alpha$ and $\beta$. You must provide the range of $\alpha$ and $\beta$, as well as the number of their attempts.
```bash
CUDA_VISIBLE_DEVICES=0 bash local/tune.sh
```
The grid search will print the WER (word error rate) or CER (character error rate) at each point in the hyper-parameters space, and draw the error surface optionally. A proper hyper-parameters range should include the global minima of the error surface for WER/CER, as illustrated in the following figure.
<p align="center">
<img src="images/tuning_error_surface.png" width=550>
<br/>An example error surface for tuning on the dev-clean set of LibriSpeech
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/tuning_error_surface.png" width=550>
<br/>An example error surface for tuning on the dev-clean set of LibriSpeech
</p>
Usually, as the figure shows, the variation of language model weight ($\alpha$) significantly affect the performance of CTC beam search decoder. And a better procedure is to first tune on serveral data batches (the number can be specified) to find out the proper range of hyper-parameters, then change to the whole validation set to carray out an accurate tuning.

28
docs/source/asr/released_model.md
Unescape View File

@ -1,28 +0,0 @@
# Released Models
## Acoustic Model Released in paddle 2.X
Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech
:-------------:| :------------:| :-----: | -----: | :----------------- |:--------- | :---------- | :---------
[Ds2 Online Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds_online.5rnn.debug.tar.gz) | Aishell Dataset | Char-based | 345 MB | 2 Conv + 5 LSTM layers with only forward direction | 0.0824 |-| 151 h
[Ds2 Offline Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds2.offline.cer6p65.release.tar.gz)| Aishell Dataset | Char-based | 306 MB | 2 Conv + 3 bidirectional GRU layers| 0.065 |-| 151 h
[Conformer Online Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.chunk.release.tar.gz) | Aishell Dataset | Char-based | 283 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention + CTC | 0.0594 |-| 151 h
[Conformer Offline Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.release.tar.gz) | Aishell Dataset | Char-based | 284 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention | 0.0547 |-| 151 h
[Conformer Librispeech Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/conformer.release.tar.gz) | Librispeech Dataset | Word-based | 287 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention |-| 0.0325 | 960 h
[Transformer Librispeech Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/transformer.release.tar.gz) | Librispeech Dataset | Word-based | 195 MB | Encoder:Transformer, Decoder:Transformer, Decoding method: Attention |-| 0.0544 | 960 h
## Acoustic Model Transformed from paddle 1.8
Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech
:-------------:| :------------:| :-----: | -----: | :----------------- | :---------- | :---------- | :---------
[Ds2 Offline Aishell model](https://deepspeech.bj.bcebos.com/mandarin_models/aishell_model_v1.8_to_v2.x.tar.gz)|Aishell Dataset| Char-based| 234 MB| 2 Conv + 3 bidirectional GRU layers| 0.0804 |-| 151 h|
[Ds2 Offline Librispeech model](https://deepspeech.bj.bcebos.com/eng_models/librispeech_v1.8_to_v2.x.tar.gz)|Librispeech Dataset| Word-based| 307 MB| 2 Conv + 3 bidirectional sharing weight RNN layers |-| 0.0685| 960 h|
[Ds2 Offline Baidu en8k model](https://deepspeech.bj.bcebos.com/eng_models/baidu_en8k_v1.8_to_v2.x.tar.gz)|Baidu Internal English Dataset| Word-based| 273 MB| 2 Conv + 3 bidirectional GRU layers |-| 0.0541 | 8628 h|
## Language Model Released
Language Model | Training Data | Token-based | Size | Descriptions
:-------------:| :------------:| :-----: | -----: | :-----------------
[English LM](https://deepspeech.bj.bcebos.com/en_lm/common_crawl_00.prune01111.trie.klm) | [CommonCrawl(en.00)](http://web-language-models.s3-website-us-east-1.amazonaws.com/ngrams/en/deduped/en.00.deduped.xz) | Word-based | 8.3 GB | Pruned with 0 1 1 1 1; <br/> About 1.85 billion n-grams; <br/> 'trie' binary with '-a 22 -q 8 -b 8'
[Mandarin LM Small](https://deepspeech.bj.bcebos.com/zh_lm/zh_giga.no_cna_cmn.prune01244.klm) | Baidu Internal Corpus | Char-based | 2.8 GB | Pruned with 0 1 2 4 4; <br/> About 0.13 billion n-grams; <br/> 'probing' binary with default settings
[Mandarin LM Large](https://deepspeech.bj.bcebos.com/zh_lm/zhidao_giga.klm) | Baidu Internal Corpus | Char-based | 70.4 GB | No Pruning; <br/> About 3.7 billion n-grams; <br/> 'probing' binary with default settings

7
docs/source/conf.py
Unescape View File

@ -23,6 +23,8 @@
import recommonmark.parser
import sphinx_rtd_theme
autodoc_mock_imports = ["soundfile", "librosa"]
# -- Project information -----------------------------------------------------
project = 'paddle speech'
@ -46,10 +48,10 @@ pygments_style = 'sphinx'
extensions = [
'sphinx.ext.autodoc',
'sphinx.ext.viewcode',
'sphinx_rtd_theme',
"sphinx_rtd_theme",
'sphinx.ext.mathjax',
'sphinx.ext.autosummary',
'numpydoc',
'sphinx.ext.autosummary',
'myst_parser',
]
@ -76,6 +78,7 @@ smartquotes = False
# relative to this directory. They are copied after the builtin static files,
# so a file named "default.css" will overwrite the builtin "default.css".
html_static_path = ['_static']
html_logo = '../images/paddle.png'
# -- Extension configuration -------------------------------------------------
# numpydoc_show_class_members = False

39
docs/source/index.rst
Unescape View File

@ -10,34 +10,44 @@ Contents
.. toctree::
:maxdepth: 1
:caption: Introduction
asr/deepspeech_architecture
introduction
.. toctree::
:maxdepth: 1
:caption: Getting_started
asr/install
asr/getting_started
:caption: Quick Start
install
asr/quick_start
tts/quick_start
.. toctree::
:maxdepth: 1
:caption: More Information
:caption: Speech-To-Text
asr/models_introduction
asr/data_preparation
asr/augmentation
asr/feature_list
asr/ngram_lm
asr/ngram_lm
.. toctree::
:maxdepth: 1
:caption: Released_model
:caption: Text-To-Speech
asr/released_model
tts/basic_usage
tts/advanced_usage
tts/zh_text_frontend
tts/models_introduction
tts/gan_vocoder
tts/demo
tts/demo_2
.. toctree::
:maxdepth: 1
:caption: Released Models
released_model
.. toctree::
:maxdepth: 1
@ -45,3 +55,8 @@ Contents
asr/reference

10
docs/source/asr/install.md → docs/source/install.md
Unescape View File

@ -8,7 +8,7 @@ To avoid the trouble of environment setup, [running in Docker container](#runnin
## Setup (Important)
- Make sure these libraries or tools installed: `pkg-config`, `flac`, `ogg`, `vorbis`, `boost`, `sox, and `swig`, e.g. installing them via `apt-get`:
- Make sure these libraries or tools installed: `pkg-config`, `flac`, `ogg`, `vorbis`, `boost`, `sox`, and `swig`, e.g. installing them via `apt-get`:
```bash
sudo apt-get install -y sox pkg-config libflac-dev libogg-dev libvorbis-dev libboost-dev swig python3-dev
@ -44,6 +44,14 @@ bash setup.sh
source tools/venv/bin/activate
```
## Simple Setup
```python
git clone https://github.com/PaddlePaddle/DeepSpeech.git
cd DeepSpeech
pip install -e .
```
## Running in Docker Container (optional)
Docker is an open source tool to build, ship, and run distributed applications in an isolated environment. A Docker image for this project has been provided in [hub.docker.com](https://hub.docker.com) with all the dependencies installed. This Docker image requires the support of NVIDIA GPU, so please make sure its availiability and the [nvidia-docker](https://github.com/NVIDIA/nvidia-docker) has been installed.

33
docs/source/introduction.md
Unescape View File

@ -0,0 +1,33 @@
# PaddleSpeech
## What is PaddleSpeech?
PaddleSpeech is an open-source toolkit on PaddlePaddle platform for two critical tasks in Speech - Speech-To-Text (Automatic Speech Recognition, ASR) and Text-To-Speech Synthesis (TTS), with modules involving state-of-art and influential models.
## What can PaddleSpeech do?
### Speech-To-Text
(An introduce of ASR in PaddleSpeech is needed here!)
### Text-To-Speech
TTS mainly consists of components below:
- Implementation of models and commonly used neural network layers.
- Dataset abstraction and common data preprocessing pipelines.
- Ready-to-run experiments.
PaddleSpeech TTS provides you with a complete TTS pipeline, including:
- Text FrontEnd
- Rule based Chinese frontend.
- Acoustic Models
- FastSpeech2
- SpeedySpeech
- TransformerTTS
- Tacotron2
- Vocoders
- Multi Band MelGAN
- Parallel WaveGAN
- WaveFlow
- Voice Cloning
- Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis
- GE2E
Text-To-Speech helps you to train TTS models with simple commands.

55
docs/source/released_model.md
Unescape View File

@ -0,0 +1,55 @@
# Released Models
## Speech-To-Text Models
### Acoustic Model Released in paddle 2.X
Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech
:-------------:| :------------:| :-----: | -----: | :----------------- |:--------- | :---------- | :---------
[Ds2 Online Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds_online.5rnn.debug.tar.gz) | Aishell Dataset | Char-based | 345 MB | 2 Conv + 5 LSTM layers with only forward direction | 0.0824 |-| 151 h
[Ds2 Offline Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds2.offline.cer6p65.release.tar.gz)| Aishell Dataset | Char-based | 306 MB | 2 Conv + 3 bidirectional GRU layers| 0.065 |-| 151 h
[Conformer Online Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.chunk.release.tar.gz) | Aishell Dataset | Char-based | 283 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention + CTC | 0.0594 |-| 151 h
[Conformer Offline Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.release.tar.gz) | Aishell Dataset | Char-based | 284 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention | 0.0547 |-| 151 h
[Conformer Librispeech Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/conformer.release.tar.gz) | Librispeech Dataset | Word-based | 287 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention |-| 0.0325 | 960 h
[Transformer Librispeech Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/transformer.release.tar.gz) | Librispeech Dataset | Word-based | 195 MB | Encoder:Transformer, Decoder:Transformer, Decoding method: Attention |-| 0.0544 | 960 h
### Acoustic Model Transformed from paddle 1.8
Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech
:-------------:| :------------:| :-----: | -----: | :----------------- | :---------- | :---------- | :---------
[Ds2 Offline Aishell model](https://deepspeech.bj.bcebos.com/mandarin_models/aishell_model_v1.8_to_v2.x.tar.gz)|Aishell Dataset| Char-based| 234 MB| 2 Conv + 3 bidirectional GRU layers| 0.0804 |-| 151 h|
[Ds2 Offline Librispeech model](https://deepspeech.bj.bcebos.com/eng_models/librispeech_v1.8_to_v2.x.tar.gz)|Librispeech Dataset| Word-based| 307 MB| 2 Conv + 3 bidirectional sharing weight RNN layers |-| 0.0685| 960 h|
[Ds2 Offline Baidu en8k model](https://deepspeech.bj.bcebos.com/eng_models/baidu_en8k_v1.8_to_v2.x.tar.gz)|Baidu Internal English Dataset| Word-based| 273 MB| 2 Conv + 3 bidirectional GRU layers |-| 0.0541 | 8628 h|
### Language Model Released
Language Model | Training Data | Token-based | Size | Descriptions
:-------------:| :------------:| :-----: | -----: | :-----------------
[English LM](https://deepspeech.bj.bcebos.com/en_lm/common_crawl_00.prune01111.trie.klm) | [CommonCrawl(en.00)](http://web-language-models.s3-website-us-east-1.amazonaws.com/ngrams/en/deduped/en.00.deduped.xz) | Word-based | 8.3 GB | Pruned with 0 1 1 1 1; <br/> About 1.85 billion n-grams; <br/> 'trie' binary with '-a 22 -q 8 -b 8'
[Mandarin LM Small](https://deepspeech.bj.bcebos.com/zh_lm/zh_giga.no_cna_cmn.prune01244.klm) | Baidu Internal Corpus | Char-based | 2.8 GB | Pruned with 0 1 2 4 4; <br/> About 0.13 billion n-grams; <br/> 'probing' binary with default settings
[Mandarin LM Large](https://deepspeech.bj.bcebos.com/zh_lm/zhidao_giga.klm) | Baidu Internal Corpus | Char-based | 70.4 GB | No Pruning; <br/> About 3.7 billion n-grams; <br/> 'probing' binary with default settings
## Text-To-Speech Models
### Acoustic Models
Model Type | Dataset| Example Link | Pretrained Models
:-------------:| :------------:| :-----: | :-----
Tacotron2|LJSpeech|[tacotron2-vctk](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts0)|[tacotron2_ljspeech_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/tacotron2_ljspeech_ckpt_0.3.zip)
TransformerTTS| LJSpeech| [transformer-ljspeech](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts1)|[transformer_tts_ljspeech_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/transformer_tts_ljspeech_ckpt_0.4.zip)
SpeedySpeech| CSMSC | [speedyspeech-csmsc](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts2) |[speedyspeech_nosil_baker_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/speedyspeech_nosil_baker_ckpt_0.5.zip)
FastSpeech2| CSMSC |[fastspeech2-csmsc](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts3)|[fastspeech2_nosil_baker_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_ckpt_0.4.zip)
FastSpeech2| AISHELL-3 |[fastspeech2-aishell3](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/aishell3/tts3)|[fastspeech2_nosil_aishell3_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_aishell3_ckpt_0.4.zip)
FastSpeech2| LJSpeech |[fastspeech2-ljspeech](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts3)|[fastspeech2_nosil_ljspeech_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_ljspeech_ckpt_0.5.zip)
FastSpeech2| VCTK |[fastspeech2-csmsc](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/vctk/tts3)|[fastspeech2_nosil_vctk_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_vctk_ckpt_0.5.zip)
### Vocoders
Model Type | Dataset| Example Link | Pretrained Models
:-------------:| :------------:| :-----: | :-----
WaveFlow| LJSpeech |[waveflow-ljspeech](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/voc0)|[waveflow_ljspeech_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/waveflow_ljspeech_ckpt_0.3.zip)
Parallel WaveGAN| CSMSC |[PWGAN-csmsc](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/voc1)|[pwg_baker_ckpt_0.4.zip.](https://paddlespeech.bj.bcebos.com/Parakeet/pwg_baker_ckpt_0.4.zip)
Parallel WaveGAN| LJSpeech |[PWGAN-ljspeech](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/voc1)|[pwg_ljspeech_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/pwg_ljspeech_ckpt_0.5.zip)
Parallel WaveGAN| VCTK |[PWGAN-vctk](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/vctk/voc1)|[pwg_vctk_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/pwg_vctk_ckpt_0.5.zip)
### Voice Cloning
Model Type | Dataset| Example Link | Pretrained Models
:-------------:| :------------:| :-----: | :-----
GE2E| AISHELL-3, etc. |[ge2e](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/other/ge2e)|[ge2e_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/ge2e_ckpt_0.3.zip)
GE2E + Tactron2| AISHELL-3 |[ge2e-tactron2-aishell3](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/aishell3/vc0)|[tacotron2_aishell3_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/tacotron2_aishell3_ckpt_0.3.zip)

53
docs/source/tts/advanced_usage.md
Unescape View File

@ -1,6 +1,5 @@
# Advanced Usage
This sections covers how to extend parakeet by implementing your own models and experiments. Guidelines on implementation are also elaborated.
This sections covers how to extend TTS by implementing your own models and experiments. Guidelines on implementation are also elaborated.
For the general deep learning experiment, there are several parts to deal with:
1. Preprocess the data according to the needs of the model, and iterate the dataset by batch.
@ -8,7 +7,7 @@ For the general deep learning experiment, there are several parts to deal with:
3. Write out the training process (generally including forward / backward calculation, parameter update, log recording, visualization, periodic evaluation, etc.).
5. Configure and run the experiment.
## Parakeet's Model Components
## PaddleSpeech TTS's Model Components
In order to balance the reusability and function of models, we divide models into several types according to its characteristics.
For the commonly used modules that can be used as part of other larger models, we try to implement them as simple and universal as possible, because they will be reused. Modules with trainable parameters are generally implemented as subclasses of `paddle.nn.Layer`. Modules without trainable parameters can be directly implemented as a function, and its input and output are `paddle.Tensor`.
@ -68,11 +67,11 @@ There are two common ways to define a model which consists of several modules.
```
When a model is a complicated and made up of several components, each of which has a separate functionality, and can be replaced by other components with the same functionality, we prefer to define it in this way.
In the directory structure of Parakeet, modules with high reusability are placed in `parakeet.modules`, but models for specific tasks are placed in `parakeet.models`. When developing a new model, developers need to consider the feasibility of splitting the modules, and the degree of generality of the modules, and place them in appropriate directories.
In the directory structure of PaddleSpeech TTS, modules with high reusability are placed in `parakeet.modules`, but models for specific tasks are placed in `parakeet.models`. When developing a new model, developers need to consider the feasibility of splitting the modules, and the degree of generality of the modules, and place them in appropriate directories.
## Parakeet's Data Components
## PaddleSpeech TTS's Data Components
Another critical componnet for a deep learning project is data.
Parakeet uses the following methods for training data:
PaddleSpeech TTS uses the following methods for training data:
1. Preprocess the data.
2. Load the preprocessed data for training.
@ -154,7 +153,7 @@ def _convert(self, meta_datum: Dict[str, Any]) -> Dict[str, Any]:
return example
```
## Parakeet's Training Components
## PaddleSpeech TTS's Training Components
A typical training process includes the following processes:
1. Iterate the dataset.
2. Process batch data.
@ -164,7 +163,7 @@ A typical training process includes the following processes:
6. Write logs, visualize, and in some cases save necessary intermediate results.
7. Save the state of the model and optimizer.
Here, we mainly introduce the training related components of Parakeet and why we designed it like this.
Here, we mainly introduce the training related components of TTS in Pa and why we designed it like this.
### Global Repoter
When training and modifying Deep Learning models，logging is often needed, and it has even become the key to model debugging and modifying. We usually use various visualization tools，such as , `visualdl` in `paddle`, `tensorboard` in `tensorflow` and `vidsom`, `wnb` ,etc. Besides, `logging` and `print` are usuaally used for different purpose.
@ -245,7 +244,7 @@ def test_reporter_scope():
In this way, when we write modular components, we can directly call `report`. The caller will decide where to report as long as it's ready for `OBSERVATION`, then it opens a `scope` and calls the component within this `scope`.
The `Trainer` in Parakeet report the information in this way.
The `Trainer` in PaddleSpeech TTS report the information in this way.
```python
while True:
self.observation = {}
@ -269,7 +268,7 @@ We made an abstraction for these intermediate processes, that is, `Updater`, whi
### Visualizer
Because we choose observation as the communication mode, we can simply write the things in observation into `visualizer`.
## Parakeet's Configuration Components
## PaddleSpeech TTS's Configuration Components
Deep learning experiments often have many options to configure. These configurations can be roughly divided into several categories.
1. Data source and data processing mode configuration.
2. Save path configuration of experimental results.
@ -293,28 +292,26 @@ The following is the basic `ArgumentParser`:
3. `--output-dir` is the dir to save the training results.（if there are checkpoints in `checkpoints/` of `--output-dir` , it's defalut to reload the newest checkpoint to train)
4. `--device` and `--nprocs` determine operation modes，`--device` specifies the type of running device, whether to run on `cpu` or `gpu`. `--nprocs` refers to the number of training processes. If `nprocs` > 1, it means that multi process parallel training is used. (Note: currently only GPU multi card multi process training is supported.)
Developers can refer to the examples in `Parakeet/examples` to write the default configuration file when adding new experiments.
Developers can refer to the examples in `examples` to write the default configuration file when adding new experiments.
## Parakeet's Experiment template
## PaddleSpeech TTS's Experiment template
The experimental codes in Parakeet are generally organized as follows:
The experimental codes in PaddleSpeech TTS are generally organized as follows:
```text
├── conf
│ └── default.yaml (defalut config)
├── README.md (help information)
├── batch_fn.py (organize metadata into batch)
├── config.py (code to read default config)
├── *_updater.py (Updater of a specific model)
├── preprocess.py (data preprocessing code)
├── preprocess.sh (script to call data preprocessing.py)
├── synthesis.py (synthesis from metadata)
├── synthesis.sh (script to call synthesis.py)
├── synthesis_e2e.py (synthesis from raw text)
├── synthesis_e2e.sh (script to call synthesis_e2e.py)
├── train.py (train code)
└── run.sh (script to call train.py)
.
├── README.md (help information)
├── conf
│ └── default.yaml (defalut config)
├── local
│ ├── preprocess.sh (script to call data preprocessing.py)
│ ├── synthesize.sh (script to call synthesis.py)
│ ├── synthesize_e2e.sh (script to call synthesis_e2e.py)
│ └──train.sh (script to call train.py)
├── path.sh (script include paths to be sourced)
└── run.sh (script to call scripts in local)
```
The `*.py` files called by above `*.sh` are located `${BIN_DIR}/`
We add a named argument. `--output-dir` to each training script to specify the output directory. The directory structure is as follows, It's best for developers to follow this specification:
```text
@ -330,4 +327,4 @@ exp/default/
└── test/ (output dir of synthesis results)
```
You can view the examples we provide in `Parakeet/examples`. These experiments are provided to users as examples which can be run directly. Users are welcome to add new models and experiments and contribute code to Parakeet.
You can view the examples we provide in `examples`. These experiments are provided to users as examples which can be run directly. Users are welcome to add new models and experiments and contribute code to PaddleSpeech.

115
docs/source/tts/basic_usage.md
Unescape View File

@ -1,115 +0,0 @@
# Basic Usage
This section shows how to use pretrained models provided by parakeet and make inference with them.
Pretrained models in v0.4 are provided in a archive. Extract it to get a folder like this:
```
checkpoint_name/
├──default.yaml
├──snapshot_iter_76000.pdz
├──speech_stats.npy
└──phone_id_map.txt
```
`default.yaml` stores the config used to train the model.
`snapshot_iter_N.pdz` is the chechpoint file, where `N` is the steps it has been trained.
`*_stats.npy` is the stats file of feature if it has been normalized before training.
`phone_id_map.txt` is the map of phonemes to phoneme_ids.
The example code below shows how to use the models for prediction.
## Acoustic Models (text to spectrogram)
The code below show how to use a `FastSpeech2` model. After loading the pretrained model, use it and normalizer object to construct a prediction object，then use fastspeech2_inferencet(phone_ids) to generate spectrograms, which can be further used to synthesize raw audio with a vocoder.
```python
from pathlib import Path
import numpy as np
import paddle
import yaml
from yacs.config import CfgNode
from parakeet.models.fastspeech2 import FastSpeech2
from parakeet.models.fastspeech2 import FastSpeech2Inference
from parakeet.modules.normalizer import ZScore
# Parakeet/examples/fastspeech2/baker/frontend.py
from frontend import Frontend
# load the pretrained model
checkpoint_dir = Path("fastspeech2_nosil_baker_ckpt_0.4")
with open(checkpoint_dir / "phone_id_map.txt", "r") as f:
phn_id = [line.strip().split() for line in f.readlines()]
vocab_size = len(phn_id)
with open(checkpoint_dir / "default.yaml") as f:
fastspeech2_config = CfgNode(yaml.safe_load(f))
odim = fastspeech2_config.n_mels
model = FastSpeech2(
idim=vocab_size, odim=odim, **fastspeech2_config["model"])
model.set_state_dict(
paddle.load(args.fastspeech2_checkpoint)["main_params"])
model.eval()
# load stats file
stat = np.load(checkpoint_dir / "speech_stats.npy")
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
fastspeech2_normalizer = ZScore(mu, std)
# construct a prediction object
fastspeech2_inference = FastSpeech2Inference(fastspeech2_normalizer, model)
# load Chinese Frontend
frontend = Frontend(checkpoint_dir / "phone_id_map.txt")
# text to spectrogram
sentence = "你好吗？"
input_ids = frontend.get_input_ids(sentence, merge_sentences=True)
phone_ids = input_ids["phone_ids"]
flags = 0
# The output of Chinese text frontend is segmented
for part_phone_ids in phone_ids:
with paddle.no_grad():
temp_mel = fastspeech2_inference(part_phone_ids)
if flags == 0:
mel = temp_mel
flags = 1
else:
mel = paddle.concat([mel, temp_mel])
```
## Vocoder (spectrogram to wave)
The code below show how to use a ` Parallel WaveGAN` model. Like the example above, after loading the pretrained model, use it and normalizer object to construct a prediction object，then use pwg_inference(mel) to generate raw audio (in wav format).
```python
from pathlib import Path
import numpy as np
import paddle
import soundfile as sf
import yaml
from yacs.config import CfgNode
from parakeet.models.parallel_wavegan import PWGGenerator
from parakeet.models.parallel_wavegan import PWGInference
from parakeet.modules.normalizer import ZScore
# load the pretrained model
checkpoint_dir = Path("parallel_wavegan_baker_ckpt_0.4")
with open(checkpoint_dir / "pwg_default.yaml") as f:
pwg_config = CfgNode(yaml.safe_load(f))
vocoder = PWGGenerator(**pwg_config["generator_params"])
vocoder.set_state_dict(paddle.load(args.pwg_params))
vocoder.remove_weight_norm()
vocoder.eval()
# load stats file
stat = np.load(checkpoint_dir / "pwg_stats.npy")
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
pwg_normalizer = ZScore(mu, std)
# construct a prediction object
pwg_inference = PWGInference(pwg_normalizer, vocoder)
# spectrogram to wave
wav = pwg_inference(mel)
sf.write(
audio_path,
wav.numpy(),
samplerate=fastspeech2_config.fs)
```

4
docs/source/tts/demo.rst
Unescape View File

@ -11,7 +11,7 @@ The main processes of TTS include:
When training ``Tacotron2``、``TransformerTTS`` and ``WaveFlow``, we use English single speaker TTS dataset `LJSpeech <https://keithito.com/LJ-Speech-Dataset/>`_ by default. However, when training ``SpeedySpeech``, ``FastSpeech2`` and ``ParallelWaveGAN``, we use Chinese single speaker dataset `CSMSC <https://test.data-baker.com/data/index/source/>`_ by default.
In the future, ``Parakeet`` will mainly use Chinese TTS datasets for default examples.
In the future, ``PaddleSpeech TTS`` will mainly use Chinese TTS datasets for default examples.
Here, we will display three types of audio samples:
@ -441,7 +441,7 @@ Audio samples generated by a TTS system. Text is first transformed into spectrog
Chinese TTS with/without text frontend
--------------------------------------
We provide a complete Chinese text frontend module in ``Parakeet``. ``Text Normalization`` and ``G2P`` are the most important modules in text frontend, We assume that the texts are normalized already, and mainly compare ``G2P`` module here.
We provide a complete Chinese text frontend module in ``PaddleSpeech TTS``. ``Text Normalization`` and ``G2P`` are the most important modules in text frontend, We assume that the texts are normalized already, and mainly compare ``G2P`` module here.
We use ``FastSpeech2`` + ``ParallelWaveGAN`` here.

7
docs/source/tts/demo_2.rst
Unescape View File

@ -0,0 +1,7 @@
Audio Sample (PaddleSpeech TTS VS Espnet TTS)
==================
This is an audio demo page to contrast PaddleSpeech TTS and Espnet TTS, We use their respective modules (Text Frontend, Acoustic model and Vocoder) here.
We use Espnet's released models here.
FastSpeech2 + Parallel WaveGAN in CSMSC

9
docs/source/tts/gan_vocoder.md
Unescape View File

@ -0,0 +1,9 @@
# GAN Vocoders
This is a brief introduction of GAN Vocoders, we mainly introduce the losses of different vocoders here.
Model | Generator Loss |Discriminator Loss
:-------------:| :------------:| :-----
Parallel Wave GAN| adversial loss <br> Feature Matching | Multi-Scale Discriminator |
Mel GAN |adversial loss <br> Multi-resolution STFT loss | adversial loss|
Multi-Band Mel GAN | adversial loss <br> full band Multi-resolution STFT loss <br> sub band Multi-resolution STFT loss |Multi-Scale Discriminator|
HiFi GAN |adversial loss <br> Feature Matching <br> Mel-Spectrogram Loss | Multi-Scale Discriminator <br> Multi-Period Discriminato |

45
docs/source/tts/index.rst
Unescape View File

@ -1,45 +0,0 @@
.. parakeet documentation master file, created by
sphinx-quickstart on Fri Sep 10 14:22:24 2021.
You can adapt this file completely to your liking, but it should at least
contain the root `toctree` directive.
Parakeet
====================================
``parakeet`` is a deep learning based text-to-speech toolkit built upon ``paddlepaddle`` framework. It aims to provide a flexible, efficient and state-of-the-art text-to-speech toolkit for the open-source community. It includes many influential TTS models proposed by `Baidu Research <http://research.baidu.com>`_ and other research groups.
``parakeet`` mainly consists of components below.
#. Implementation of models and commonly used neural network layers.
#. Dataset abstraction and common data preprocessing pipelines.
#. Ready-to-run experiments.
.. toctree::
:maxdepth: 1
:caption: Introduction
introduction
.. toctree::
:maxdepth: 1
:caption: Getting started
install
basic_usage
advanced_usage
cn_text_frontend
released_models
.. toctree::
:maxdepth: 1
:caption: Demos
demo
Indices and tables
==================
* :ref:`genindex`
* :ref:`modindex`
* :ref:`search`

47
docs/source/tts/install.md
Unescape View File

@ -1,47 +0,0 @@
# Installation
## Install PaddlePaddle
Parakeet requires PaddlePaddle as its backend. Note that 2.1.2 or newer versions of paddle is required.
Since paddlepaddle has multiple packages depending on the device (cpu or gpu) and the dependency libraries, it is recommended to install a proper package of paddlepaddle with respect to the device and dependency library versons via `pip`.
Installing paddlepaddle with conda or build paddlepaddle from source is also supported. Please refer to [PaddlePaddle installation](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/install/pip/linux-pip.html) for more details.
Example instruction to install paddlepaddle via pip is listed below.
### PaddlePaddle with GPU
```python
# PaddlePaddle for CUDA10.1
python -m pip install paddlepaddle-gpu==2.1.2.post101 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
# PaddlePaddle for CUDA10.2
python -m pip install paddlepaddle-gpu -i https://mirror.baidu.com/pypi/simple
# PaddlePaddle for CUDA11.0
python -m pip install paddlepaddle-gpu==2.1.2.post110 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
# PaddlePaddle for CUDA11.2
python -m pip install paddlepaddle-gpu==2.1.2.post112 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
```
### PaddlePaddle with CPU
```python
python -m pip install paddlepaddle==2.1.2 -i https://mirror.baidu.com/pypi/simple
```
## Install libsndfile
Experimemts in parakeet often involve audio and spectrum processing, thus `librosa` and `soundfile` are required. `soundfile` requires a extra C library `libsndfile`, which is not always handled by pip.
For Windows and Mac users, `libsndfile` is also installed when installing `soundfile` via pip, but for Linux users, installing `libsndfile` via system package manager is required. Example commands for popular distributions are listed below.
```bash
# ubuntu, debian
sudo apt-get install libsndfile1
# centos, fedora
sudo yum install libsndfile
# openSUSE
sudo zypper in libsndfile
```
For any problem with installtion of soundfile, please refer to [SoundFile](https://pypi.org/project/SoundFile/).
## Install Parakeet
There are two ways to install parakeet according to the purpose of using it.
1. If you want to run experiments provided by parakeet or add new models and experiments, it is recommended to clone the project from github (Parakeet), and install it in editable mode.
```python
git clone https://github.com/PaddlePaddle/Parakeet
cd Parakeet
pip install -e .
```

27
docs/source/tts/introduction.md
Unescape View File

@ -1,27 +0,0 @@
# Parakeet - PAddle PARAllel text-to-speech toolKIT
## What is Parakeet?
Parakeet is a deep learning based text-to-speech toolkit built upon paddlepaddle framework. It aims to provide a flexible, efficient and state-of-the-art text-to-speech toolkit for the open-source community. It includes many influential TTS models proposed by Baidu Research and other research groups.
## What can Parakeet do?
Parakeet mainly consists of components below:
- Implementation of models and commonly used neural network layers.
- Dataset abstraction and common data preprocessing pipelines.
- Ready-to-run experiments.
Parakeet provides you with a complete TTS pipeline, including:
- Text FrontEnd
- Rule based Chinese frontend.
- Acoustic Models
- FastSpeech2
- SpeedySpeech
- TransformerTTS
- Tacotron2
- Vocoders
- Parallel WaveGAN
- WaveFlow
- Voice Cloning
- Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis
- GE2E
Parakeet helps you to train TTS models with simple commands.

42
docs/source/tts/released_models.md → docs/source/tts/models_introduction.md
Unescape View File

@ -1,12 +1,12 @@
# Released Models
TTS system mainly includes three modules: `text frontend`, `Acoustic model` and `Vocoder`. We introduce a rule based Chinese text frontend in [cn_text_frontend.md](./cn_text_frontend.md). Here, we will introduce acoustic models and vocoders, which are trainable models.
# Models introduction
TTS system mainly includes three modules: `Text Frontend`, `Acoustic model` and `Vocoder`. We introduce a rule based Chinese text frontend in [cn_text_frontend.md](./cn_text_frontend.md). Here, we will introduce acoustic models and vocoders, which are trainable models.
The main processes of TTS include:
1. Convert the original text into characters/phonemes, through `text frontend` module.
2. Convert characters/phonemes into acoustic features , such as linear spectrogram, mel spectrogram, LPC features, etc. through `Acoustic models`.
3. Convert acoustic features into waveforms through `Vocoders`.
A simple text frontend module can be implemented by rules. Acoustic models and vocoders need to be trained. The models provided by Parakeet are acoustic models and vocoders.
A simple text frontend module can be implemented by rules. Acoustic models and vocoders need to be trained. The models provided by PaddleSpeech TTS are acoustic models and vocoders.
## Acoustic Models
### Modeling Objectives of Acoustic Models
@ -27,14 +27,14 @@ At present, there are two mainstream acoustic model structures.
- Acoustic decoder (N Frames - > N Frames).
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/frame_level_am.png" width=500 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/frame_level_am.png" width=500 /> <br>
</div>
- Sequence to sequence acoustic model:
- M Tokens - > N Frames.
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/seq2seq_am.png" width=500 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/seq2seq_am.png" width=500 /> <br>
</div>
### Tacotron2
@ -54,7 +54,7 @@ At present, there are two mainstream acoustic model structures.
- CBHG postprocess.
- Vocoder: Griffin-Lim.
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/tacotron.png" width=700 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/tacotron.png" width=700 /> <br>
</div>
**Advantage of Tacotron:**
@ -89,10 +89,10 @@ At present, there are two mainstream acoustic model structures.
- The alignment matrix of previous time is considered at the step `t` of decoder.
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/tacotron2.png" width=500 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/tacotron2.png" width=500 /> <br>
</div>
You can find Parakeet's tacotron2 example at `Parakeet/examples/tacotron2`.
You can find PaddleSpeech TTS's tacotron2 with LJSpeech dataset example at [examples/ljspeech/tts0](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts0).
### TransformerTTS
**Disadvantages of the Tacotrons:**
@ -118,7 +118,7 @@ Transformer TTS is a combination of Tacotron2 and Transformer.
- Positional Encoding.
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/transformer.png" width=500 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/transformer.png" width=500 /> <br>
</div>
#### Transformer TTS
@ -138,7 +138,7 @@ Transformer TTS is a seq2seq acoustic model based on Transformer and Tacotron2.
- Uniform scale position encoding may have a negative impact on input or output sequences.
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/transformer_tts.png" width=500 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/transformer_tts.png" width=500 /> <br>
</div>
**Disadvantages of Transformer TTS:**
@ -146,7 +146,7 @@ Transformer TTS is a seq2seq acoustic model based on Transformer and Tacotron2.
- The ability to perceive local information is weak, and local information is more related to pronunciation.
- Stability is worse than Tacotron2.
You can find Parakeet's Transformer TTS example at `Parakeet/examples/transformer_tts`.
You can find PaddleSpeech TTS's Transformer TTS with LJSpeech dataset example at [examples/ljspeech/tts1](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts1).
### FastSpeech2
@ -184,14 +184,14 @@ Instead of using the encoder-attention-decoder based architecture as adopted by
• Can be generated in parallel (decoding time is less affected by sequence length)
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/fastspeech.png" width=800 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/fastspeech.png" width=800 /> <br>
</div>
#### FastPitch
[FastPitch](https://arxiv.org/abs/2006.06873) follows FastSpeech. A single pitch value is predicted for every temporal location, which improves the overall quality of synthesized speech.
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/fastpitch.png" width=500 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/fastpitch.png" width=500 /> <br>
</div>
#### FastSpeech2
@ -209,10 +209,10 @@ Instead of using the encoder-attention-decoder based architecture as adopted by
FastSpeech2 is similar to FastPitch but introduces more variation information of speech.
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/fastspeech2.png" width=800 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/fastspeech2.png" width=800 /> <br>
</div>
You can find Parakeet's FastSpeech2/FastPitch example at `Parakeet/examples/fastspeech2`, We use token-averaged pitch and energy values introduced in FastPitch rather than frame level ones in FastSpeech2.
You can find PaddleSpeech TTS's FastSpeech2/FastPitch with CSMSC dataset example at [examples/csmsc/tts3](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts3), We use token-averaged pitch and energy values introduced in FastPitch rather than frame level ones in FastSpeech2.
### SpeedySpeech
[SpeedySpeech](https://arxiv.org/abs/2008.03802) simplify the teacher-student architecture of FastSpeech and provide a fast and stable training procedure.
@ -223,10 +223,10 @@ You can find Parakeet's FastSpeech2/FastPitch example at `Parakeet/examples/fast
- Describe a simple data augmentation technique that can be used early in the training to make the teacher network robust to sequential error propagation.
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/speedyspeech.png" width=500 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/speedyspeech.png" width=500 /> <br>
</div>
You can find Parakeet's SpeedySpeech example at `Parakeet/examples/speedyspeech/baker`.
You can find PaddleSpeech TTS's SpeedySpeech with CSMSC dataset example at [examples/csmsc/tts2](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts2).
## Vocoders
In speech synthesis, the main task of the vocoder is to convert the spectral parameters predicted by the acoustic model into the final speech waveform.
@ -276,7 +276,7 @@ Here, we introduce a Flow-based vocoder WaveFlow and a GAN-based vocoder Paralle
- It is a small-footprint flow-based model for raw audio. It has only 5.9M parameters, which is 15x smalller than WaveGlow (87.9M).
- It is directly trained with maximum likelihood without probability density distillation and auxiliary losses as used in [Parallel WaveNet](https://arxiv.org/abs/1711.10433) and [ClariNet](https://openreview.net/pdf?id=HklY120cYm), which simplifies the training pipeline and reduces the cost of development.
You can find Parakeet's WaveFlow example at `Parakeet/examples/waveflow`.
You can find PaddleSpeech TTS's WaveFlow with LJSpeech dataset example at [examples/ljspeech/voc0](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/voc0).
### Parallel WaveGAN
[Parallel WaveGAN](https://arxiv.org/abs/1910.11480) trains a non-autoregressive WaveNet variant as a generator in a GAN based training method.
@ -286,10 +286,10 @@ You can find Parakeet's WaveFlow example at `Parakeet/examples/waveflow`.
- Use non-causal convolution instead of causal convolution.
- The input is random Gaussian white noise.
- The model is non-autoregressive both in training and prediction, which is fast
- Multi-resolution STFT loss.
- Multi-resolution STFT loss.
<div align="left">
<img src="https://paddlespeech.bj.bcebos.com/Parakeet/docs/images/pwg.png" width=600 /> <br>
<img src="https://raw.githubusercontent.com/PaddlePaddle/DeepSpeech/develop/docs/images/pwg.png" width=600 /> <br>
</div>
You can find Parakeet's Parallel WaveGAN example at `Parakeet/examples/parallelwave_gan/baker`.
You can find PaddleSpeech TTS's Parallel WaveGAN with CSMSC example at [examples/csmsc/voc1](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/voc1).

193
docs/source/tts/quick_start.md
Unescape View File

@ -0,0 +1,193 @@
# Quick Start of Text-To-Speech
The examples in PaddleSpeech are mainly classified by datasets, the TTS datasets we mainly used are:
* CSMCS (Mandarin single speaker)
* AISHELL3 (Mandarin multiple speaker)
* LJSpeech (English single speaker)
* VCTK (English multiple speaker)
The models in PaddleSpeech TTS have the following mapping relationship:
* tts0 - Tactron2
* tts1 - TransformerTTS
* tts2 - SpeedySpeech
* tts3 - FastSpeech2
* voc0 - WaveFlow
* voc1 - Parallel WaveGAN
* voc2 - MelGAN
* voc3 - MultiBand MelGAN
* vc0 - Tactron2 Voice Clone with GE2E
## Quick Start
Let's take a FastSpeech2 + Parallel WaveGAN with CSMSC dataset for instance. (./examples/csmsc/)(https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc)
### Train Parallel WaveGAN with CSMSC
- Go to directory
```bash
cd examples/csmsc/voc1
```
- Source env
```bash
source path.sh
```
**Must do this before you start to do anything.**
Set `MAIN_ROOT` as project dir. Using `parallelwave_gan` model as `MODEL`.
- Main entrypoint
```bash
bash run.sh
```
This is just a demo, please make sure source data have been prepared well and every `step` works well before next `step`.
### Train FastSpeech2 with CSMSC
- Go to directory
```bash
cd examples/csmsc/tts3
```
- Source env
```bash
source path.sh
```
**Must do this before you start to do anything.**
Set `MAIN_ROOT` as project dir. Using `fastspeech2` model as `MODEL`.
- Main entrypoint
```bash
bash run.sh
```
This is just a demo, please make sure source data have been prepared well and every `step` works well before next `step`.
The steps in `run.sh` mainly include:
- source path.
- preprocess the dataset,
- train the model.
- synthesize waveform from metadata.jsonl.
- synthesize waveform from text file. (in acoustic models)
- inference using static model. (optional)
For more details , you can see `README.md` in examples.
## Pipeline of TTS
This section shows how to use pretrained models provided by TTS and make inference with them.
Pretrained models in TTS are provided in a archive. Extract it to get a folder like this:
**Acoustic Models:**
```text
checkpoint_name
├── default.yaml
├── snapshot_iter_*.pdz
├── speech_stats.npy
├── phone_id_map.txt
├── spk_id_map.txt (optimal)
└── tone_id_map.txt (optimal)
```
**Vocoders:**
```text
checkpoint_name
├── default.yaml
├── snapshot_iter_*.pdz
└── stats.npy
```
- `default.yaml` stores the config used to train the model.
- `snapshot_iter_*.pdz` is the chechpoint file, where `*` is the steps it has been trained.
- `*_stats.npy` is the stats file of feature if it has been normalized before training.
- `phone_id_map.txt` is the map of phonemes to phoneme_ids.
- `tone_id_map.txt` is the map of tones to tones_ids, when you split tones and phones before training acoustic models. (for example in our csmsc/speedyspeech example)
- `spk_id_map.txt` is the map of spkeaker to spk_ids in multi-spk acoustic models. (for example in our aishell3/fastspeech2 example)
The example code below shows how to use the models for prediction.
### Acoustic Models (text to spectrogram)
The code below show how to use a `FastSpeech2` model. After loading the pretrained model, use it and normalizer object to construct a prediction object，then use `fastspeech2_inferencet(phone_ids)` to generate spectrograms, which can be further used to synthesize raw audio with a vocoder.
```python
from pathlib import Path
import numpy as np
import paddle
import yaml
from yacs.config import CfgNode
from parakeet.models.fastspeech2 import FastSpeech2
from parakeet.models.fastspeech2 import FastSpeech2Inference
from parakeet.modules.normalizer import ZScore
# examples/fastspeech2/baker/frontend.py
from frontend import Frontend
# load the pretrained model
checkpoint_dir = Path("fastspeech2_nosil_baker_ckpt_0.4")
with open(checkpoint_dir / "phone_id_map.txt", "r") as f:
phn_id = [line.strip().split() for line in f.readlines()]
vocab_size = len(phn_id)
with open(checkpoint_dir / "default.yaml") as f:
fastspeech2_config = CfgNode(yaml.safe_load(f))
odim = fastspeech2_config.n_mels
model = FastSpeech2(
idim=vocab_size, odim=odim, **fastspeech2_config["model"])
model.set_state_dict(
paddle.load(args.fastspeech2_checkpoint)["main_params"])
model.eval()
# load stats file
stat = np.load(checkpoint_dir / "speech_stats.npy")
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
fastspeech2_normalizer = ZScore(mu, std)
# construct a prediction object
fastspeech2_inference = FastSpeech2Inference(fastspeech2_normalizer, model)
# load Chinese Frontend
frontend = Frontend(checkpoint_dir / "phone_id_map.txt")
# text to spectrogram
sentence = "你好吗？"
input_ids = frontend.get_input_ids(sentence, merge_sentences=True)
phone_ids = input_ids["phone_ids"]
flags = 0
# The output of Chinese text frontend is segmented
for part_phone_ids in phone_ids:
with paddle.no_grad():
temp_mel = fastspeech2_inference(part_phone_ids)
if flags == 0:
mel = temp_mel
flags = 1
else:
mel = paddle.concat([mel, temp_mel])
```
### Vocoder (spectrogram to wave)
The code below show how to use a ` Parallel WaveGAN` model. Like the example above, after loading the pretrained model, use it and normalizer object to construct a prediction object，then use `pwg_inference(mel)` to generate raw audio (in wav format).
```python
from pathlib import Path
import numpy as np
import paddle
import soundfile as sf
import yaml
from yacs.config import CfgNode
from parakeet.models.parallel_wavegan import PWGGenerator
from parakeet.models.parallel_wavegan import PWGInference
from parakeet.modules.normalizer import ZScore
# load the pretrained model
checkpoint_dir = Path("parallel_wavegan_baker_ckpt_0.4")
with open(checkpoint_dir / "pwg_default.yaml") as f:
pwg_config = CfgNode(yaml.safe_load(f))
vocoder = PWGGenerator(**pwg_config["generator_params"])
vocoder.set_state_dict(paddle.load(args.pwg_params))
vocoder.remove_weight_norm()
vocoder.eval()
# load stats file
stat = np.load(checkpoint_dir / "pwg_stats.npy")
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
pwg_normalizer = ZScore(mu, std)
# construct a prediction object
pwg_inference = PWGInference(pwg_normalizer, vocoder)
# spectrogram to wave
wav = pwg_inference(mel)
sf.write(
audio_path,
wav.numpy(),
samplerate=fastspeech2_config.fs)
```

2
docs/source/tts/cn_text_frontend.md → docs/source/tts/zh_text_frontend.md
Unescape View File

@ -1,5 +1,5 @@
# Chinese Rule Based Text Frontend
TTS system mainly includes three modules: `text frontend`, `Acoustic model` and `Vocoder`. We provide a complete Chinese text frontend module in Parakeet, see exapmle in `Parakeet/examples/text_frontend/`.
A TTS system mainly includes three modules: `Text Frontend`, `Acoustic model` and `Vocoder`. We provide a complete Chinese text frontend module in PaddleSpeech TTS, see exapmle in [examples/other/text_frontend/](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/other/text_frontend).
A text frontend module mainly includes:
- Text Segmentation

14
examples/other/text_frontend/README.md
Unescape View File

@ -20,5 +20,19 @@ Run the command below to get the results of test.
./run.sh
```
The `avg WER` of g2p is: 0.027495061517943988
```text
,--------------------------------------------------------------------.
| | # Snt # Wrd | Corr Sub Del Ins Err S.Err |
|--------+-----------------+-----------------------------------------|
| Sum/Avg| 9996 299181 | 97.3 2.7 0.0 0.0 2.7 52.5 |
`--------------------------------------------------------------------'
```
The `avg CER` of text normalization is: 0.006388318503308237
```text
,-----------------------------------------------------------------.
| | # Snt # Wrd | Corr Sub Del Ins Err S.Err |
|--------+--------------+-----------------------------------------|
| Sum/Avg| 125 2254 | 99.4 0.1 0.5 0.1 0.7 3.2 |
`-----------------------------------------------------------------'
```

7
text_processing/.gitignore vendored
Unescape View File

@ -0,0 +1,7 @@
data
glove
.pyc
checkpoints
epoch
__pycache__
glove.840B.300d.zip

25
text_processing/README.md
Unescape View File

@ -0,0 +1,25 @@
# PaddleSpeechTask
A speech library to deal with a series of related front-end and back-end tasks
## 环境
- python==3.6.13
- paddle==2.1.1
## 中/英文文本加标点任务 punctuation restoration：
### 数据集: data
- 中文数据来源：data/chinese
1.iwlst2012zh
2.平凡的世界
- 英文数据来源: data/english
1.iwlst2012en
- iwlst2012数据获取过程见data/README.md
### 模型：speechtask/punctuation_restoration/model
1.BLSTM模型
2.BertLinear模型
3.BertBLSTM模型

35
text_processing/examples/punctuation_restoration/chinese/README.md
Unescape View File

@ -0,0 +1,35 @@
# 中文实验例程
## 测试数据：
- IWLST2012中文：test2012
## 运行代码
- 运行 `run.sh 0 0 conf/train_conf/bertBLSTM_zh.yaml 1 conf/data_conf/chinese.yaml `
## 实验结果：
- BertLinear
- 实验配置：conf/train_conf/bertLinear_zh.yaml
- 测试结果
| | COMMA | PERIOD | QUESTION | OVERALL |
|-----------|-----------|-----------|-----------|--------- |
|Precision | 0.425665 | 0.335190 | 0.698113 | 0.486323 |
|Recall | 0.511278 | 0.572108 | 0.787234 | 0.623540 |
|F1 | 0.464560 | 0.422717 | 0.740000 | 0.542426 |
- BertBLSTM
- 实验配置：conf/train_conf/bertBLSTM_zh.yaml
- 测试结果 avg_1
| | COMMA | PERIOD | QUESTION | OVERALL |
|-----------|-----------|-----------|-----------|--------- |
|Precision | 0.469484 | 0.550604 | 0.801887 | 0.607325 |
|Recall | 0.580271 | 0.592408 | 0.817308 | 0.663329 |
|F1 | 0.519031 | 0.570741 | 0.809524 | 0.633099 |
- BertBLSTM/avg_1测试标贝合成数据
| | COMMA | PERIOD | QUESTION | OVERALL |
|-----------|-----------|-----------|-----------|--------- |
|Precision | 0.217192 | 0.196339 | 0.820717 | 0.411416 |
|Recall | 0.205922 | 0.892531 | 0.416162 | 0.504872 |
|F1 | 0.211407 | 0.321873 | 0.552279 | 0.361853 |

34
text_processing/examples/punctuation_restoration/chinese/conf/blstm.yaml
Unescape View File

@ -0,0 +1,34 @@
data:
language: chinese
raw_path: /data4/mahaoxin/PaddleSpeechTask/data/chinese/PFDSJ #path to raw dataset
raw_train_file: train
raw_dev_file: dev
raw_test_file: test
vocab_file: vocab
punc_file: punc_vocab
save_path: data/PFDSJ #path to save dataset
seq_len: 100
batch_size: 10
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 0
model_type: blstm
model_params:
vocab_size: 3751
embedding_size: 200
hidden_size: 100
num_layers: 3
num_class: 5
init_scale: 0.1
training:
n_epoch: 32
lr: !!float 1e-4
lr_decay: 1.0
weight_decay: !!float 1e-06
global_grad_clip: 5.0
log_interval: 10

7
text_processing/examples/punctuation_restoration/chinese/conf/data_conf/chinese.yaml
Unescape View File

@ -0,0 +1,7 @@
type: chinese
raw_path: /data4/mahaoxin/PaddleSpeechTask/data/chinese/iwslt2012_zh #path to raw dataset
raw_train_file: iwslt2012_train_zh
raw_dev_file: iwslt2010_dev_zh
raw_test_file: biaobei_asr
punc_file: punc_vocab
save_path: data/iwslt2012_zh #path to save dataset

49
text_processing/examples/punctuation_restoration/chinese/conf/train_conf/bertBLSTM_zh.yaml
Unescape View File

@ -0,0 +1,49 @@
data:
dataset_type: Bert
train_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/data/iwslt2012_zh/train
dev_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/data/iwslt2012_zh/dev
test_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/data/iwslt2012_zh/test2012_revise
data_params:
pretrained_token: bert-base-chinese
punc_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/data/iwslt2012_zh/punc_vocab
seq_len: 100
batch_size: 64
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 0
checkpoint:
kbest_n: 5
latest_n: 10
metric_type: F1
model_type: BertBLSTM
model_params:
pretrained_token: bert-base-chinese
output_size: 4
dropout: 0.0
bert_size: 768
blstm_size: 128
num_blstm_layers: 2
init_scale: 0.1
# model_type: BertChLinear
# model_params: bert-base-chinese
# pretrained_token:
# output_size: 4
# dropout: 0.0
# bert_size: 768
training:
n_epoch: 100
lr: !!float 1e-5
lr_decay: 1.0
weight_decay: !!float 1e-06
global_grad_clip: 5.0
log_interval: 10
log_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/log/bertBLSTM_zh0812.log
testing:
log_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/log/test_bertBLSTM_zh0812.log

42
text_processing/examples/punctuation_restoration/chinese/conf/train_conf/bertLinear_zh.yaml
Unescape View File

@ -0,0 +1,42 @@
data:
dataset_type: Bert
train_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/data/iwslt2012_zh/train
dev_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/data/iwslt2012_zh/dev
test_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/data/iwslt2012_zh/test2012
data_params:
pretrained_token: bert-base-chinese
punc_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/data/iwslt2012_zh/punc_vocab
seq_len: 100
batch_size: 32
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 0
checkpoint:
kbest_n: 10
latest_n: 10
metric_type: F1
model_type: BertLinear
model_params:
pretrained_token: bert-base-uncased
output_size: 4
dropout: 0.2
bert_size: 768
hiddensize: 1568
training:
n_epoch: 50
lr: !!float 1e-5
lr_decay: 1.0
weight_decay: !!float 1e-06
global_grad_clip: 5.0
log_interval: 10
log_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/log/train_linear0812.log
testing:
log_interval: 10
log_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/chinese/log/test_linear0812.log

23
text_processing/examples/punctuation_restoration/chinese/local/avg.sh
Unescape View File

@ -0,0 +1,23 @@
#! /usr/bin/env bash
if [ $# != 2 ]; then
echo "usage: ${0} ckpt_dir avg_num"
exit -1
fi
ckpt_dir=${1}
average_num=${2}
decode_checkpoint=${ckpt_dir}/avg_${average_num}.pdparams
python3 -u ${BIN_DIR}/avg_model.py \
--dst_model ${decode_checkpoint} \
--ckpt_dir ${ckpt_dir} \
--num ${average_num} \
--val_best
if [ $? -ne 0 ]; then
echo "Failed in avg ckpt!"
exit 1
fi
exit 0

19
text_processing/examples/punctuation_restoration/chinese/local/data.sh
Unescape View File

@ -0,0 +1,19 @@
#!/bin/bash
if [ $# != 1 ];then
echo "usage: ${0} data_pre_conf"
echo $1
exit -1
fi
data_pre_conf=$1
python3 -u ${BIN_DIR}/pre_data.py \
--config ${data_pre_conf}
if [ $? -ne 0 ]; then
echo "Failed in training!"
exit 1
fi
exit 0

32
text_processing/examples/punctuation_restoration/chinese/local/test.sh
Unescape View File

@ -0,0 +1,32 @@
#!/bin/bash
if [ $# != 2 ];then
echo "usage: ${0} config_path ckpt_path_prefix"
exit -1
fi
ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..."
device=gpu
if [ ${ngpu} == 0 ];then
device=cpu
fi
config_path=$1
ckpt_prefix=$2
python3 -u ${BIN_DIR}/test.py \
--device ${device} \
--nproc 1 \
--config ${config_path} \
--result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix}
if [ $? -ne 0 ]; then
echo "Failed in evaluation!"
exit 1
fi
exit 0

32
text_processing/examples/punctuation_restoration/chinese/local/train.sh
Unescape View File

@ -0,0 +1,32 @@
#!/bin/bash
if [ $# != 2 ];then
echo "usage: CUDA_VISIBLE_DEVICES=0 ${0} config_path ckpt_name"
exit -1
fi
ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..."
config_path=$1
ckpt_name=$2
device=gpu
if [ ${ngpu} == 0 ];then
device=cpu
fi
mkdir -p exp
python3 -u ${BIN_DIR}/train.py \
--device ${device} \
--nproc ${ngpu} \
--config ${config_path} \
--output exp/${ckpt_name}
if [ $? -ne 0 ]; then
echo "Failed in training!"
exit 1
fi
exit 0

13
text_processing/examples/punctuation_restoration/chinese/path.sh
Unescape View File

@ -0,0 +1,13 @@
export MAIN_ROOT=${PWD}/../../../
export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
export LC_ALL=C
# Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib/
export BIN_DIR=${MAIN_ROOT}/speechtask/punctuation_restoration/bin

47
text_processing/examples/punctuation_restoration/chinese/run.sh
Unescape View File

@ -0,0 +1,47 @@
#!/bin/bash
set -e
source path.sh
## stage, gpu, data_pre_config, train_config, avg_num
if [ $# -lt 4 ]; then
echo "usage: bash ./run.sh stage gpu train_config avg_num data_config"
echo "eg: bash ./run.sh 0 0 train_config 1 data_config "
exit -1
fi
stage=$1
stop_stage=100
gpus=$2
conf_path=$3
avg_num=$4
avg_ckpt=avg_${avg_num}
ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')
echo "checkpoint name ${ckpt}"
if [ $stage -le 0 ]; then
if [ $# -eq 5 ]; then
data_pre_conf=$5
# prepare data
bash ./local/data.sh ${data_pre_conf} || exit -1
else
echo "data_pre_conf is not exist!"
exit -1
fi
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# train model, all `ckpt` under `exp` dir
CUDA_VISIBLE_DEVICES=${gpus} bash ./local/train.sh ${conf_path} ${ckpt}
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# avg n best model
bash ./local/avg.sh exp/${ckpt}/checkpoints ${avg_num}
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n
CUDA_VISIBLE_DEVICES=${gpus} bash ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi

23
text_processing/examples/punctuation_restoration/english/README.md
Unescape View File

@ -0,0 +1,23 @@
# 英文实验例程
## 测试数据：
- IWLST2012英文：test2011
## 运行代码
- 运行 `run.sh 0 0 conf/train_conf/bertBLSTM_base_en.yaml 1 conf/data_conf/english.yaml `
## 相关论文实验结果：
> * Nagy, Attila, Bence Bial, and Judit Ács. "Automatic punctuation restoration with BERT models." arXiv preprint arXiv:2101.07343 (2021)*
>
## 实验结果：
- BertBLSTM
- 实验配置：conf/train_conf/bertLinear_en.yaml
- 测试结果：exp/bertLinear_enRe/checkpoints/3.pdparams
| | COMMA | PERIOD | QUESTION | OVERALL |
|-----------|-----------|-----------|-----------|--------- |
|Precision |0.667910 |0.715778 |0.822222 |0.735304 |
|Recall |0.755274 |0.868188 |0.804348 |0.809270 |
|F1 |0.708911 |0.784651 |0.813187 |0.768916 |

7
text_processing/examples/punctuation_restoration/english/conf/data_conf/english.yaml
Unescape View File

@ -0,0 +1,7 @@
type: english
raw_path: /data4/mahaoxin/PaddleSpeechTask/data/english/iwslt2012_en #path to raw dataset
raw_train_file: iwslt2012_train_en
raw_dev_file: iwslt2010_dev_en
raw_test_file: iwslt2011_test_en
punc_file: punc_vocab
save_path: data/iwslt2012_en #path to save dataset

47
text_processing/examples/punctuation_restoration/english/conf/train_conf/bertBLSTM_base_en.yaml
Unescape View File

@ -0,0 +1,47 @@
data:
dataset_type: Bert
train_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/english/data/iwslt2012_en/train
dev_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/english/data/iwslt2012_en/dev
test_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/english/data/iwslt2012_en/test2011
data_params:
pretrained_token: bert-base-uncased #english
punc_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/english/data/iwslt2012_en/punc_vocab
seq_len: 50
batch_size: 32
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 0
checkpoint:
kbest_n: 10
latest_n: 10
model_type: BertBLSTM
model_params:
pretrained_token: bert-base-uncased
output_size: 4
dropout: 0.0
bert_size: 768
blstm_size: 128
num_blstm_layers: 2
init_scale: 0.2
# model_type: BertChLinear
# model_params:
# pretrained_token: bert-large-uncased
# output_size: 4
# dropout: 0.0
# bert_size: 768
training:
n_epoch: 100
lr: !!float 1e-5
lr_decay: 1.0
weight_decay: !!float 1e-06
global_grad_clip: 5.0
log_interval: 10
log_path: log/bertBLSTM_base0812.log
testing:
log_path: log/testbertBLSTM_base0812.log

39
text_processing/examples/punctuation_restoration/english/conf/train_conf/bertLinear_en.yaml
Unescape View File

@ -0,0 +1,39 @@
data:
dataset_type: Bert
train_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/english/data/iwslt2012_en/train
dev_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/english/data/iwslt2012_en/dev
test_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/english/data/iwslt2012_en/test2011
data_params:
pretrained_token: bert-base-uncased #english
punc_path: /data4/mahaoxin/PaddleSpeechTask/examples/punctuation_restoration/english/data/iwslt2012_en/punc_vocab
seq_len: 100
batch_size: 32
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 0
checkpoint:
kbest_n: 10
latest_n: 10
model_type: BertLinear
model_params:
pretrained_token: bert-base-uncased
output_size: 4
dropout: 0.2
bert_size: 768
hiddensize: 1568
training:
n_epoch: 20
lr: !!float 1e-5
lr_decay: 1.0
weight_decay: !!float 1e-06
global_grad_clip: 3.0
log_interval: 10
log_path: log/train_linear0820.log
testing:
log_path: log/test2011_linear0820.log

23
text_processing/examples/punctuation_restoration/english/local/avg.sh
Unescape View File

@ -0,0 +1,23 @@
#! /usr/bin/env bash
if [ $# != 2 ]; then
echo "usage: ${0} ckpt_dir avg_num"
exit -1
fi
ckpt_dir=${1}
average_num=${2}
decode_checkpoint=${ckpt_dir}/avg_${average_num}.pdparams
python3 -u ${BIN_DIR}/avg_model.py \
--dst_model ${decode_checkpoint} \
--ckpt_dir ${ckpt_dir} \
--num ${average_num} \
--val_best
if [ $? -ne 0 ]; then
echo "Failed in avg ckpt!"
exit 1
fi
exit 0

18
text_processing/examples/punctuation_restoration/english/local/data.sh
Unescape View File

@ -0,0 +1,18 @@
#!/bin/bash
if [ $# != 1 ];then
echo "usage: ${0} config_path"
exit -1
fi
config_path=$1
python3 -u ${BIN_DIR}/pre_data.py \
--config ${config_path}
if [ $? -ne 0 ]; then
echo "Failed in training!"
exit 1
fi
exit 0

32
text_processing/examples/punctuation_restoration/english/local/test.sh
Unescape View File

@ -0,0 +1,32 @@
#!/bin/bash
if [ $# != 2 ];then
echo "usage: ${0} config_path ckpt_path_prefix"
exit -1
fi
ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..."
device=gpu
if [ ${ngpu} == 0 ];then
device=cpu
fi
config_path=$1
ckpt_prefix=$2
python3 -u ${BIN_DIR}/test.py \
--device ${device} \
--nproc 1 \
--config ${config_path} \
--result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix}
if [ $? -ne 0 ]; then
echo "Failed in evaluation!"
exit 1
fi
exit 0

32
text_processing/examples/punctuation_restoration/english/local/train.sh
Unescape View File

@ -0,0 +1,32 @@
#!/bin/bash
if [ $# != 2 ];then
echo "usage: CUDA_VISIBLE_DEVICES=0 ${0} config_path ckpt_name"
exit -1
fi
ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..."
config_path=$1
ckpt_name=$2
device=gpu
if [ ${ngpu} == 0 ];then
device=cpu
fi
mkdir -p exp
python3 -u ${BIN_DIR}/train.py \
--device ${device} \
--nproc ${ngpu} \
--config ${config_path} \
--output exp/${ckpt_name}
if [ $? -ne 0 ]; then
echo "Failed in training!"
exit 1
fi
exit 0

13
text_processing/examples/punctuation_restoration/english/path.sh
Unescape View File

@ -0,0 +1,13 @@
export MAIN_ROOT=${PWD}/../../../
export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
export LC_ALL=C
# Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib/
export BIN_DIR=${MAIN_ROOT}/speechtask/punctuation_restoration/bin

47
text_processing/examples/punctuation_restoration/english/run.sh
Unescape View File

@ -0,0 +1,47 @@
#!/bin/bash
set -e
source path.sh
## stage, gpu, data_pre_config, train_config, avg_num
if [ $# -lt 4 ]; then
echo "usage: bash ./run.sh stage gpu train_config avg_num data_config"
echo "eg: bash ./run.sh 0 0 train_config 1 data_config "
exit -1
fi
stage=$1
stop_stage=100
gpus=$2
conf_path=$3
avg_num=$4
avg_ckpt=avg_${avg_num}
ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')
echo "checkpoint name ${ckpt}"
if [ $stage -le 0 ]; then
if [ $# -eq 5 ]; then
data_pre_conf=$5
# prepare data
bash ./local/data.sh ${data_pre_conf} || exit -1
else
echo "data_pre_conf is not exist!"
exit -1
fi
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# train model, all `ckpt` under `exp` dir
CUDA_VISIBLE_DEVICES=${gpus} bash ./local/train.sh ${conf_path} ${ckpt}
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# avg n best model
bash ./local/avg.sh exp/${ckpt}/checkpoints ${avg_num}
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n
CUDA_VISIBLE_DEVICES=${gpus} bash ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi

6
text_processing/requirements.txt
Unescape View File

@ -0,0 +1,6 @@
numpy
pyyaml
tensorboardX
tqdm
ujson
yacs

112
text_processing/speechtask/punctuation_restoration/bin/avg_model.py
Unescape View File

@ -0,0 +1,112 @@
#!/usr/bin/env python3
# Copyright (c) 2021 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 glob
import json
import os
import numpy as np
import paddle
def main(args):
paddle.set_device('cpu')
val_scores = []
beat_val_scores = []
selected_epochs = []
if args.val_best:
jsons = glob.glob(f'{args.ckpt_dir}/[!train]*.json')
for y in jsons:
with open(y, 'r') as f:
dic_json = json.load(f)
loss = dic_json['F1']
epoch = dic_json['epoch']
if epoch >= args.min_epoch and epoch <= args.max_epoch:
val_scores.append((epoch, loss))
val_scores = np.array(val_scores)
sort_idx = np.argsort(val_scores[:, 1])
sorted_val_scores = val_scores[sort_idx]
path_list = [
args.ckpt_dir + '/{}.pdparams'.format(int(epoch))
for epoch in sorted_val_scores[:args.num, 0]
]
beat_val_scores = sorted_val_scores[:args.num, 1]
selected_epochs = sorted_val_scores[:args.num, 0].astype(np.int64)
print("best val scores = " + str(beat_val_scores))
print("selected epochs = " + str(selected_epochs))
else:
path_list = glob.glob(f'{args.ckpt_dir}/[!avg][!final]*.pdparams')
path_list = sorted(path_list, key=os.path.getmtime)
path_list = path_list[-args.num:]
print(path_list)
avg = None
num = args.num
assert num == len(path_list)
for path in path_list:
print(f'Processing {path}')
states = paddle.load(path)
if avg is None:
avg = states
else:
for k in avg.keys():
avg[k] += states[k]
# average
for k in avg.keys():
if avg[k] is not None:
avg[k] /= num
paddle.save(avg, args.dst_model)
print(f'Saving to {args.dst_model}')
meta_path = os.path.splitext(args.dst_model)[0] + '.avg.json'
with open(meta_path, 'w') as f:
data = json.dumps({
"avg_ckpt": args.dst_model,
"ckpt": path_list,
"epoch": selected_epochs.tolist(),
"val_loss": beat_val_scores.tolist(),
})
f.write(data + "\n")
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='average model')
parser.add_argument('--dst_model', required=True, help='averaged model')
parser.add_argument(
'--ckpt_dir', required=True, help='ckpt model dir for average')
parser.add_argument(
'--val_best', action="store_true", help='averaged model')
parser.add_argument(
'--num', default=5, type=int, help='nums for averaged model')
parser.add_argument(
'--min_epoch',
default=0,
type=int,
help='min epoch used for averaging model')
parser.add_argument(
'--max_epoch',
default=65536, # Big enough
type=int,
help='max epoch used for averaging model')
args = parser.parse_args()
print(args)
main(args)

48
text_processing/speechtask/punctuation_restoration/bin/pre_data.py
Unescape View File

@ -0,0 +1,48 @@
# Copyright (c) 2021 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.
"""Data preparation for punctuation_restoration task."""
import yaml
from speechtask.punctuation_restoration.utils.default_parser import default_argument_parser
from speechtask.punctuation_restoration.utils.punct_pre import process_chinese_pure_senetence
from speechtask.punctuation_restoration.utils.punct_pre import process_english_pure_senetence
from speechtask.punctuation_restoration.utils.utility import print_arguments
# create dataset from raw data files
def main(config, args):
print("Start preparing data from raw data.")
if (config['type'] == 'chinese'):
process_chinese_pure_senetence(config)
elif (config['type'] == 'english'):
print('english!!!!')
process_english_pure_senetence(config)
else:
print('Error: Type should be chinese or english!!!!')
raise ValueError('Type should be chinese or english')
print("Finish preparing data.")
if __name__ == "__main__":
parser = default_argument_parser()
args = parser.parse_args()
print_arguments(args, globals())
# https://yaml.org/type/float.html
with open(args.config, "r") as f:
config = yaml.load(f, Loader=yaml.FullLoader)
# config.freeze()
print(config)
main(config, args)

45
text_processing/speechtask/punctuation_restoration/bin/test.py
Unescape View File

@ -0,0 +1,45 @@
# Copyright (c) 2021 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.
"""Evaluation for model."""
import yaml
from speechtask.punctuation_restoration.training.trainer import Tester
from speechtask.punctuation_restoration.utils.default_parser import default_argument_parser
from speechtask.punctuation_restoration.utils.utility import print_arguments
def main_sp(config, args):
exp = Tester(config, args)
exp.setup()
exp.run_test()
def main(config, args):
main_sp(config, args)
if __name__ == "__main__":
parser = default_argument_parser()
args = parser.parse_args()
print_arguments(args, globals())
# https://yaml.org/type/float.html
with open(args.config, "r") as f:
config = yaml.load(f, Loader=yaml.FullLoader)
print(config)
if args.dump_config:
with open(args.dump_config, 'w') as f:
print(config, file=f)
main(config, args)

49
text_processing/speechtask/punctuation_restoration/bin/train.py
Unescape View File

@ -0,0 +1,49 @@
# Copyright (c) 2021 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.
"""Trainer for punctuation_restoration task."""
import yaml
from paddle import distributed as dist
from speechtask.punctuation_restoration.training.trainer import Trainer
from speechtask.punctuation_restoration.utils.default_parser import default_argument_parser
from speechtask.punctuation_restoration.utils.utility import print_arguments
def main_sp(config, args):
exp = Trainer(config, args)
exp.setup()
exp.run()
def main(config, args):
if args.device == "gpu" and args.nprocs > 1:
dist.spawn(main_sp, args=(config, args), nprocs=args.nprocs)
else:
main_sp(config, args)
if __name__ == "__main__":
parser = default_argument_parser()
args = parser.parse_args()
print_arguments(args, globals())
# https://yaml.org/type/float.html
with open(args.config, "r") as f:
config = yaml.load(f, Loader=yaml.FullLoader)
print(config)
if args.dump_config:
with open(args.dump_config, 'w') as f:
print(config, file=f)
main(config, args)

13
text_processing/speechtask/punctuation_restoration/io/__init__.py
Unescape View File

@ -0,0 +1,13 @@
# Copyright (c) 2021 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.

64
text_processing/speechtask/punctuation_restoration/io/collator.py
Unescape View File

@ -0,0 +1,64 @@
# Copyright (c) 2021 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
__all__ = ["TextCollator"]
class TextCollator():
def __init__(self, padding_value):
self.padding_value = padding_value
def __call__(self, batch):
"""batch examples
Args:
batch ([List]): batch is (text, punctuation)
text (List[int] ) shape (batch, L)
punctuation (List[int] or str): shape (batch, L)
Returns:
tuple(text, punctuation): batched data.
text : (B, Lmax)
punctuation : (B, Lmax)
"""
texts = []
punctuations = []
for text, punctuation in batch:
texts.append(text)
punctuations.append(punctuation)
#[B, T, D]
x_pad = self.pad_sequence(texts).astype(np.int64)
# print(x_pad.shape)
# pad_list(audios, 0.0).astype(np.float32)
# ilens = np.array(audio_lens).astype(np.int64)
y_pad = self.pad_sequence(punctuations).astype(np.int64)
# print(y_pad.shape)
# olens = np.array(text_lens).astype(np.int64)
return x_pad, y_pad
def pad_sequence(self, sequences):
# assuming trailing dimensions and type of all the Tensors
# in sequences are same and fetching those from sequences[0]
max_len = max([len(s) for s in sequences])
out_dims = (len(sequences), max_len)
out_tensor = np.full(out_dims,
self.padding_value) #, dtype=sequences[0].dtype)
for i, tensor in enumerate(sequences):
length = len(tensor)
# use index notation to prevent duplicate references to the tensor
out_tensor[i, :length] = tensor
return out_tensor

55
text_processing/speechtask/punctuation_restoration/io/common.py
Unescape View File

@ -0,0 +1,55 @@
# Copyright (c) 2021 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 codecs
import re
import unicodedata
import ujson
PAD = "<PAD>"
UNK = "<UNK>"
NUM = "<NUM>"
END = "</S>"
SPACE = "_SPACE"
def write_json(filename, dataset):
with codecs.open(filename, mode="w", encoding="utf-8") as f:
ujson.dump(dataset, f)
def word_convert(word, keep_number=True, lowercase=True):
if not keep_number:
if is_digit(word):
word = NUM
if lowercase:
word = word.lower()
return word
def is_digit(word):
try:
float(word)
return True
except ValueError:
pass
try:
unicodedata.numeric(word)
return True
except (TypeError, ValueError):
pass
result = re.compile(r'^[-+]?[0-9]+,[0-9]+$').match(word)
if result:
return True
return False

310
text_processing/speechtask/punctuation_restoration/io/dataset.py
Unescape View File

@ -0,0 +1,310 @@
# Copyright (c) 2021 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 random
import numpy as np
import paddle
from paddle.io import Dataset
from paddlenlp.transformers import BertTokenizer
# from speechtask.punctuation_restoration.utils.punct_prepro import load_dataset
__all__ = ["PuncDataset", "PuncDatasetFromBertTokenizer"]
class PuncDataset(Dataset):
"""Representing a Dataset
superclass
----------
data.Dataset :
Dataset is a abstract class, representing the real data.
"""
def __init__(self, train_path, vocab_path, punc_path, seq_len=100):
# 检查文件是否存在
print(train_path)
print(vocab_path)
assert os.path.exists(train_path), "train文件不存在"
assert os.path.exists(vocab_path), "词典文件不存在"
assert os.path.exists(punc_path), "标点文件不存在"
self.seq_len = seq_len
self.word2id = self.load_vocab(
vocab_path, extra_word_list=['<UNK>', '<END>'])
self.id2word = {v: k for k, v in self.word2id.items()}
self.punc2id = self.load_vocab(punc_path, extra_word_list=[" "])
self.id2punc = {k: v for (v, k) in self.punc2id.items()}
tmp_seqs = open(train_path, encoding='utf-8').readlines()
self.txt_seqs = [i for seq in tmp_seqs for i in seq.split()]
# print(self.txt_seqs[:10])
# with open('./txt_seq', 'w', encoding='utf-8') as w:
# print(self.txt_seqs, file=w)
self.preprocess(self.txt_seqs)
print('---punc-')
print(self.punc2id)
def __len__(self):
"""return the sentence nums in .txt
"""
return self.in_len
def __getitem__(self, index):
"""返回指定索引的张量对 (输入文本id的序列 , 其对应的标点id序列)
Parameters
----------
index : int 索引
"""
return self.input_data[index], self.label[index]
def load_vocab(self, vocab_path, extra_word_list=[], encoding='utf-8'):
n = len(extra_word_list)
with open(vocab_path, encoding='utf-8') as vf:
vocab = {word.strip(): i + n for i, word in enumerate(vf)}
for i, word in enumerate(extra_word_list):
vocab[word] = i
return vocab
def preprocess(self, txt_seqs: list):
"""将文本转为单词和应预测标点的id pair
Parameters
----------
txt : 文本
文本每个单词跟随一个空格，符号也跟一个空格
"""
input_data = []
label = []
input_r = []
label_r = []
# txt_seqs is a list like: ['char', 'char', 'char', '*，*', 'char', ......]
count = 0
length = len(txt_seqs)
for token in txt_seqs:
count += 1
if count == length:
break
if token in self.punc2id:
continue
punc = txt_seqs[count]
if punc not in self.punc2id:
# print('标点{}：'.format(count), self.punc2id[" "])
label.append(self.punc2id[" "])
input_data.append(
self.word2id.get(token, self.word2id["<UNK>"]))
input_r.append(token)
label_r.append(' ')
else:
# print('标点{}：'.format(count), self.punc2id[punc])
label.append(self.punc2id[punc])
input_data.append(
self.word2id.get(token, self.word2id["<UNK>"]))
input_r.append(token)
label_r.append(punc)
if len(input_data) != len(label):
assert 'error: length input_data != label'
# code below is for using 100 as a hidden size
print(len(input_data))
self.in_len = len(input_data) // self.seq_len
len_tmp = self.in_len * self.seq_len
input_data = input_data[:len_tmp]
label = label[:len_tmp]
self.input_data = paddle.to_tensor(
np.array(input_data, dtype='int64').reshape(-1, self.seq_len))
self.label = paddle.to_tensor(
np.array(label, dtype='int64').reshape(-1, self.seq_len))
# unk_token='[UNK]'
# sep_token='[SEP]'
# pad_token='[PAD]'
# cls_token='[CLS]'
# mask_token='[MASK]'
class PuncDatasetFromBertTokenizer(Dataset):
"""Representing a Dataset
superclass
----------
data.Dataset :
Dataset is a abstract class, representing the real data.
"""
def __init__(self,
train_path,
is_eval,
pretrained_token,
punc_path,
seq_len=100):
# 检查文件是否存在
print(train_path)
self.tokenizer = BertTokenizer.from_pretrained(
pretrained_token, do_lower_case=True)
self.paddingID = self.tokenizer.pad_token_id
assert os.path.exists(train_path), "train文件不存在"
assert os.path.exists(punc_path), "标点文件不存在"
self.seq_len = seq_len
self.punc2id = self.load_vocab(punc_path, extra_word_list=[" "])
self.id2punc = {k: v for (v, k) in self.punc2id.items()}
tmp_seqs = open(train_path, encoding='utf-8').readlines()
self.txt_seqs = [i for seq in tmp_seqs for i in seq.split()]
# print(self.txt_seqs[:10])
# with open('./txt_seq', 'w', encoding='utf-8') as w:
# print(self.txt_seqs, file=w)
if (is_eval):
self.preprocess(self.txt_seqs)
else:
self.preprocess_shift(self.txt_seqs)
print("data len: %d" % (len(self.input_data)))
print('---punc-')
print(self.punc2id)
def __len__(self):
"""return the sentence nums in .txt
"""
return self.in_len
def __getitem__(self, index):
"""返回指定索引的张量对 (输入文本id的序列 , 其对应的标点id序列)
Parameters
----------
index : int 索引
"""
return self.input_data[index], self.label[index]
def load_vocab(self, vocab_path, extra_word_list=[], encoding='utf-8'):
n = len(extra_word_list)
with open(vocab_path, encoding='utf-8') as vf:
vocab = {word.strip(): i + n for i, word in enumerate(vf)}
for i, word in enumerate(extra_word_list):
vocab[word] = i
return vocab
def preprocess(self, txt_seqs: list):
"""将文本转为单词和应预测标点的id pair
Parameters
----------
txt : 文本
文本每个单词跟随一个空格，符号也跟一个空格
"""
input_data = []
label = []
# txt_seqs is a list like: ['char', 'char', 'char', '*，*', 'char', ......]
count = 0
for i in range(len(txt_seqs) - 1):
word = txt_seqs[i]
punc = txt_seqs[i + 1]
if word in self.punc2id:
continue
token = self.tokenizer(word)
x = token["input_ids"][1:-1]
input_data.extend(x)
for i in range(len(x) - 1):
label.append(self.punc2id[" "])
if punc not in self.punc2id:
# print('标点{}：'.format(count), self.punc2id[" "])
label.append(self.punc2id[" "])
else:
label.append(self.punc2id[punc])
if len(input_data) != len(label):
assert 'error: length input_data != label'
# code below is for using 100 as a hidden size
# print(len(input_data[0]))
# print(len(label))
self.in_len = len(input_data) // self.seq_len
len_tmp = self.in_len * self.seq_len
input_data = input_data[:len_tmp]
label = label[:len_tmp]
# # print(input_data)
# print(type(input_data))
# tmp=np.array(input_data)
# print('--~~~~~~~~~~~~~')
# print(type(tmp))
# print(tmp.shape)
self.input_data = paddle.to_tensor(
np.array(input_data, dtype='int64').reshape(
-1, self.seq_len)) #, dtype='int64'
self.label = paddle.to_tensor(
np.array(label, dtype='int64').reshape(
-1, self.seq_len)) #, dtype='int64'
def preprocess_shift(self, txt_seqs: list):
"""将文本转为单词和应预测标点的id pair
Parameters
----------
txt : 文本
文本每个单词跟随一个空格，符号也跟一个空格
"""
input_data = []
label = []
# txt_seqs is a list like: ['char', 'char', 'char', '*，*', 'char', ......]
count = 0
for i in range(len(txt_seqs) - 1):
word = txt_seqs[i]
punc = txt_seqs[i + 1]
if word in self.punc2id:
continue
token = self.tokenizer(word)
x = token["input_ids"][1:-1]
input_data.extend(x)
for i in range(len(x) - 1):
label.append(self.punc2id[" "])
if punc not in self.punc2id:
# print('标点{}：'.format(count), self.punc2id[" "])
label.append(self.punc2id[" "])
else:
label.append(self.punc2id[punc])
if len(input_data) != len(label):
assert 'error: length input_data != label'
# print(len(input_data[0]))
# print(len(label))
start = 0
processed_data = []
processed_label = []
while (start < len(input_data) - self.seq_len):
# end=start+self.seq_len
end = random.randint(start + self.seq_len // 2,
start + self.seq_len)
processed_data.append(input_data[start:end])
processed_label.append(label[start:end])
start = start + random.randint(1, self.seq_len // 2)
self.in_len = len(processed_data)
# # print(input_data)
# print(type(input_data))
# tmp=np.array(input_data)
# print('--~~~~~~~~~~~~~')
# print(type(tmp))
# print(tmp.shape)
self.input_data = processed_data
#paddle.to_tensor(np.array(processed_data, dtype='int64')) #, dtype='int64'
self.label = processed_label
#paddle.to_tensor(np.array(processed_label, dtype='int64')) #, dtype='int64'
if __name__ == '__main__':
dataset = PuncDataset()

74
text_processing/speechtask/punctuation_restoration/model/BertBLSTM.py
Unescape View File

@ -0,0 +1,74 @@
# Copyright (c) 2021 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
import paddle.nn as nn
import paddle.nn.initializer as I
from paddlenlp.transformers import BertForTokenClassification
class BertBLSTMPunc(nn.Layer):
def __init__(self,
pretrained_token="bert-large-uncased",
output_size=4,
dropout=0.0,
bert_size=768,
blstm_size=128,
num_blstm_layers=2,
init_scale=0.1):
super(BertBLSTMPunc, self).__init__()
self.output_size = output_size
self.bert = BertForTokenClassification.from_pretrained(
pretrained_token, num_classes=bert_size)
# self.bert_vocab_size = vocab_size
# self.bn = nn.BatchNorm1d(segment_size*self.bert_vocab_size)
# self.fc = nn.Linear(segment_size*self.bert_vocab_size, output_size)
self.lstm = nn.LSTM(
input_size=bert_size,
hidden_size=blstm_size,
num_layers=num_blstm_layers,
direction="bidirect",
weight_ih_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)),
weight_hh_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)))
# NOTE dense*2 使用bert中间层 dense hidden_state self.bert_size
self.dropout = nn.Dropout(dropout)
self.fc = nn.Linear(blstm_size * 2, output_size)
self.softmax = nn.Softmax()
def forward(self, x):
# print('input :', x.shape)
x = self.bert(x) #[0]
# print('after bert :', x.shape)
y, (_, _) = self.lstm(x)
# print('after lstm :', y.shape)
y = self.fc(self.dropout(y))
y = paddle.reshape(y, shape=[-1, self.output_size])
# print('after fc :', y.shape)
logit = self.softmax(y)
# print('after softmax :', logit.shape)
return y, logit
if __name__ == '__main__':
print('start model')
model = BertBLSTMPunc()
x = paddle.randint(low=0, high=40, shape=[2, 5])
print(x)
y, logit = model(x)

63
text_processing/speechtask/punctuation_restoration/model/BertLinear.py
Unescape View File

@ -0,0 +1,63 @@
# Copyright (c) 2021 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
import paddle.nn as nn
from paddlenlp.transformers import BertForTokenClassification
class BertLinearPunc(nn.Layer):
def __init__(self,
pretrained_token="bert-base-uncased",
output_size=4,
dropout=0.2,
bert_size=768,
hiddensize=1568):
super(BertLinearPunc, self).__init__()
self.output_size = output_size
self.bert = BertForTokenClassification.from_pretrained(
pretrained_token, num_classes=bert_size)
# self.bert_vocab_size = vocab_size
# self.bn = nn.BatchNorm1d(segment_size*self.bert_vocab_size)
# self.fc = nn.Linear(segment_size*self.bert_vocab_size, output_size)
# NOTE dense*2 使用bert中间层 dense hidden_state self.bert_size
self.dropout1 = nn.Dropout(dropout)
self.fc1 = nn.Linear(bert_size, hiddensize)
self.dropout2 = nn.Dropout(dropout)
self.relu = nn.ReLU()
self.fc2 = nn.Linear(hiddensize, output_size)
self.softmax = nn.Softmax()
def forward(self, x):
# print('input :', x.shape)
x = self.bert(x) #[0]
# print('after bert :', x.shape)
x = self.fc1(self.dropout1(x))
x = self.fc2(self.relu(self.dropout2(x)))
x = paddle.reshape(x, shape=[-1, self.output_size])
# print('after fc :', x.shape)
logit = self.softmax(x)
# print('after softmax :', logit.shape)
return x, logit
if __name__ == '__main__':
print('start model')
model = BertLinearPunc()
x = paddle.randint(low=0, high=40, shape=[2, 5])
print(x)
y, logit = model(x)

89
text_processing/speechtask/punctuation_restoration/model/blstm.py
Unescape View File

@ -0,0 +1,89 @@
# Copyright (c) 2021 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
import paddle.nn as nn
import paddle.nn.initializer as I
class BiLSTM(nn.Layer):
"""LSTM for Punctuation Restoration
"""
def __init__(self,
vocab_size,
embedding_size,
hidden_size,
num_layers,
num_class,
init_scale=0.1):
super(BiLSTM, self).__init__()
# hyper parameters
self.vocab_size = vocab_size
self.embedding_size = embedding_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.num_class = num_class
# 网络中的层
self.embedding = nn.Embedding(
vocab_size,
embedding_size,
weight_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)))
# print(hidden_size)
# print(embedding_size)
self.lstm = nn.LSTM(
input_size=embedding_size,
hidden_size=hidden_size,
num_layers=num_layers,
direction="bidirect",
weight_ih_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)),
weight_hh_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)))
# Here is a one direction LSTM. If bidirection LSTM, (hidden_size*2(,))
self.fc = nn.Linear(
in_features=hidden_size * 2,
out_features=num_class,
weight_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)),
bias_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)))
# self.fc = nn.Linear(hidden_size, num_class)
self.softmax = nn.Softmax()
def forward(self, input):
"""The forward process of Net
Parameters
----------
inputs : tensor
Training data, batch first
"""
# Inherit the knowledge of context
# hidden = self.init_hidden(inputs.size(0))
# print('input_size',inputs.size())
embedding = self.embedding(input)
# print('embedding_size', embedding.size())
# packed = pack_sequence(embedding, inputs_lengths, batch_first=True)
# embedding本身是同样长度的，用这个函数主要是为了用pack
# *****************************************************************************
y, (_, _) = self.lstm(embedding)
# print(y.size())
y = self.fc(y)
y = paddle.reshape(y, shape=[-1, self.num_class])
logit = self.softmax(y)
return y, logit

85
text_processing/speechtask/punctuation_restoration/model/lstm.py
Unescape View File

@ -0,0 +1,85 @@
# Copyright (c) 2021 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
import paddle.nn as nn
import paddle.nn.initializer as I
class RnnLm(nn.Layer):
def __init__(self,
vocab_size,
punc_size,
hidden_size,
num_layers=1,
init_scale=0.1,
dropout=0.0):
super(RnnLm, self).__init__()
self.hidden_size = hidden_size
self.num_layers = num_layers
self.init_scale = init_scale
self.punc_size = punc_size
self.embedder = nn.Embedding(
vocab_size,
hidden_size,
weight_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)))
self.lstm = nn.LSTM(
input_size=hidden_size,
hidden_size=hidden_size,
num_layers=num_layers,
dropout=dropout,
weight_ih_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)),
weight_hh_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)))
self.fc = nn.Linear(
hidden_size,
punc_size,
weight_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)),
bias_attr=paddle.ParamAttr(initializer=I.Uniform(
low=-init_scale, high=init_scale)))
self.dropout = nn.Dropout(p=dropout)
self.softmax = nn.Softmax()
def forward(self, inputs):
x = inputs
x_emb = self.embedder(x)
x_emb = self.dropout(x_emb)
y, (_, _) = self.lstm(x_emb)
y = self.dropout(y)
y = self.fc(y)
y = paddle.reshape(y, shape=[-1, self.punc_size])
logit = self.softmax(y)
return y, logit
class CrossEntropyLossForLm(nn.Layer):
def __init__(self):
super(CrossEntropyLossForLm, self).__init__()
def forward(self, y, label):
label = paddle.unsqueeze(label, axis=2)
loss = paddle.nn.functional.cross_entropy(
input=y, label=label, reduction='none')
loss = paddle.squeeze(loss, axis=[2])
loss = paddle.mean(loss, axis=[0])
loss = paddle.sum(loss)
return loss

13
text_processing/speechtask/punctuation_restoration/modules/__init__.py
Unescape View File

@ -0,0 +1,13 @@
# Copyright (c) 2021 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.

141
text_processing/speechtask/punctuation_restoration/modules/activation.py
Unescape View File

@ -0,0 +1,141 @@
# Copyright (c) 2021 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 collections import OrderedDict
import paddle
from paddle import nn
__all__ = ["get_activation", "brelu", "LinearGLUBlock", "ConvGLUBlock"]
def brelu(x, t_min=0.0, t_max=24.0, name=None):
# paddle.to_tensor is dygraph_only can not work under JIT
t_min = paddle.full(shape=[1], fill_value=t_min, dtype='float32')
t_max = paddle.full(shape=[1], fill_value=t_max, dtype='float32')
return x.maximum(t_min).minimum(t_max)
class LinearGLUBlock(nn.Layer):
"""A linear Gated Linear Units (GLU) block."""
def __init__(self, idim: int):
""" GLU.
Args:
idim (int): input and output dimension
"""
super().__init__()
self.fc = nn.Linear(idim, idim * 2)
def forward(self, xs):
return glu(self.fc(xs), dim=-1)
class ConvGLUBlock(nn.Layer):
def __init__(self, kernel_size, in_ch, out_ch, bottlececk_dim=0,
dropout=0.):
"""A convolutional Gated Linear Units (GLU) block.
Args:
kernel_size (int): kernel size
in_ch (int): number of input channels
out_ch (int): number of output channels
bottlececk_dim (int): dimension of the bottleneck layers for computational efficiency. Defaults to 0.
dropout (float): dropout probability. Defaults to 0..
"""
super().__init__()
self.conv_residual = None
if in_ch != out_ch:
self.conv_residual = nn.utils.weight_norm(
nn.Conv2D(
in_channels=in_ch, out_channels=out_ch, kernel_size=(1, 1)),
name='weight',
dim=0)
self.dropout_residual = nn.Dropout(p=dropout)
self.pad_left = ConstantPad2d((0, 0, kernel_size - 1, 0), 0)
layers = OrderedDict()
if bottlececk_dim == 0:
layers['conv'] = nn.utils.weight_norm(
nn.Conv2D(
in_channels=in_ch,
out_channels=out_ch * 2,
kernel_size=(kernel_size, 1)),
name='weight',
dim=0)
# TODO(hirofumi0810): padding?
layers['dropout'] = nn.Dropout(p=dropout)
layers['glu'] = GLU()
elif bottlececk_dim > 0:
layers['conv_in'] = nn.utils.weight_norm(
nn.Conv2D(
in_channels=in_ch,
out_channels=bottlececk_dim,
kernel_size=(1, 1)),
name='weight',
dim=0)
layers['dropout_in'] = nn.Dropout(p=dropout)
layers['conv_bottleneck'] = nn.utils.weight_norm(
nn.Conv2D(
in_channels=bottlececk_dim,
out_channels=bottlececk_dim,
kernel_size=(kernel_size, 1)),
name='weight',
dim=0)
layers['dropout'] = nn.Dropout(p=dropout)
layers['glu'] = GLU()
layers['conv_out'] = nn.utils.weight_norm(
nn.Conv2D(
in_channels=bottlececk_dim,
out_channels=out_ch * 2,
kernel_size=(1, 1)),
name='weight',
dim=0)
layers['dropout_out'] = nn.Dropout(p=dropout)
self.layers = nn.Sequential(layers)
def forward(self, xs):
"""Forward pass.
Args:
xs (FloatTensor): `[B, in_ch, T, feat_dim]`
Returns:
out (FloatTensor): `[B, out_ch, T, feat_dim]`
"""
residual = xs
if self.conv_residual is not None:
residual = self.dropout_residual(self.conv_residual(residual))
xs = self.pad_left(xs) # `[B, embed_dim, T+kernel-1, 1]`
xs = self.layers(xs) # `[B, out_ch * 2, T ,1]`
xs = xs + residual
return xs
def get_activation(act):
"""Return activation function."""
# Lazy load to avoid unused import
activation_funcs = {
"hardtanh": paddle.nn.Hardtanh,
"tanh": paddle.nn.Tanh,
"relu": paddle.nn.ReLU,
"selu": paddle.nn.SELU,
"swish": paddle.nn.Swish,
"gelu": paddle.nn.GELU,
"brelu": brelu,
}
return activation_funcs[act]()

229
text_processing/speechtask/punctuation_restoration/modules/attention.py
Unescape View File

@ -0,0 +1,229 @@
# Copyright (c) 2021 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.
"""Multi-Head Attention layer definition."""
import math
from typing import Optional
from typing import Tuple
import paddle
from paddle import nn
from paddle.nn import initializer as I
__all__ = ["MultiHeadedAttention", "RelPositionMultiHeadedAttention"]
# Relative Positional Encodings
# https://www.jianshu.com/p/c0608efcc26f
# https://zhuanlan.zhihu.com/p/344604604
class MultiHeadedAttention(nn.Layer):
"""Multi-Head Attention layer."""
def __init__(self, n_head: int, n_feat: int, dropout_rate: float):
"""Construct an MultiHeadedAttention object.
Args:
n_head (int): The number of heads.
n_feat (int): The number of features.
dropout_rate (float): Dropout rate.
"""
super().__init__()
assert n_feat % n_head == 0
# We assume d_v always equals d_k
self.d_k = n_feat // n_head
self.h = n_head
self.linear_q = nn.Linear(n_feat, n_feat)
self.linear_k = nn.Linear(n_feat, n_feat)
self.linear_v = nn.Linear(n_feat, n_feat)
self.linear_out = nn.Linear(n_feat, n_feat)
self.dropout = nn.Dropout(p=dropout_rate)
def forward_qkv(self,
query: paddle.Tensor,
key: paddle.Tensor,
value: paddle.Tensor
) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
"""Transform query, key and value.
Args:
query (paddle.Tensor): Query tensor (#batch, time1, size).
key (paddle.Tensor): Key tensor (#batch, time2, size).
value (paddle.Tensor): Value tensor (#batch, time2, size).
Returns:
paddle.Tensor: Transformed query tensor, size
(#batch, n_head, time1, d_k).
paddle.Tensor: Transformed key tensor, size
(#batch, n_head, time2, d_k).
paddle.Tensor: Transformed value tensor, size
(#batch, n_head, time2, d_k).
"""
n_batch = query.size(0)
q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
q = q.transpose([0, 2, 1, 3]) # (batch, head, time1, d_k)
k = k.transpose([0, 2, 1, 3]) # (batch, head, time2, d_k)
v = v.transpose([0, 2, 1, 3]) # (batch, head, time2, d_k)
return q, k, v
def forward_attention(self,
value: paddle.Tensor,
scores: paddle.Tensor,
mask: Optional[paddle.Tensor]) -> paddle.Tensor:
"""Compute attention context vector.
Args:
value (paddle.Tensor): Transformed value, size
(#batch, n_head, time2, d_k).
scores (paddle.Tensor): Attention score, size
(#batch, n_head, time1, time2).
mask (paddle.Tensor): Mask, size (#batch, 1, time2) or
(#batch, time1, time2).
Returns:
paddle.Tensor: Transformed value weighted
by the attention score, (#batch, time1, d_model).
"""
n_batch = value.size(0)
if mask is not None:
mask = mask.unsqueeze(1).eq(0) # (batch, 1, *, time2)
scores = scores.masked_fill(mask, -float('inf'))
attn = paddle.softmax(
scores, axis=-1).masked_fill(mask,
0.0) # (batch, head, time1, time2)
else:
attn = paddle.softmax(
scores, axis=-1) # (batch, head, time1, time2)
p_attn = self.dropout(attn)
x = paddle.matmul(p_attn, value) # (batch, head, time1, d_k)
x = x.transpose([0, 2, 1, 3]).contiguous().view(
n_batch, -1, self.h * self.d_k) # (batch, time1, d_model)
return self.linear_out(x) # (batch, time1, d_model)
def forward(self,
query: paddle.Tensor,
key: paddle.Tensor,
value: paddle.Tensor,
mask: Optional[paddle.Tensor]) -> paddle.Tensor:
"""Compute scaled dot product attention.
Args:
query (torch.Tensor): Query tensor (#batch, time1, size).
key (torch.Tensor): Key tensor (#batch, time2, size).
value (torch.Tensor): Value tensor (#batch, time2, size).
mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
(#batch, time1, time2).
Returns:
torch.Tensor: Output tensor (#batch, time1, d_model).
"""
q, k, v = self.forward_qkv(query, key, value)
scores = paddle.matmul(q,
k.transpose([0, 1, 3, 2])) / math.sqrt(self.d_k)
return self.forward_attention(v, scores, mask)
class RelPositionMultiHeadedAttention(MultiHeadedAttention):
"""Multi-Head Attention layer with relative position encoding."""
def __init__(self, n_head, n_feat, dropout_rate):
"""Construct an RelPositionMultiHeadedAttention object.
Paper: https://arxiv.org/abs/1901.02860
Args:
n_head (int): The number of heads.
n_feat (int): The number of features.
dropout_rate (float): Dropout rate.
"""
super().__init__(n_head, n_feat, dropout_rate)
# linear transformation for positional encoding
self.linear_pos = nn.Linear(n_feat, n_feat, bias_attr=False)
# these two learnable bias are used in matrix c and matrix d
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
#self.pos_bias_u = nn.Parameter(torch.Tensor(self.h, self.d_k))
#self.pos_bias_v = nn.Parameter(torch.Tensor(self.h, self.d_k))
#torch.nn.init.xavier_uniform_(self.pos_bias_u)
#torch.nn.init.xavier_uniform_(self.pos_bias_v)
pos_bias_u = self.create_parameter(
[self.h, self.d_k], default_initializer=I.XavierUniform())
self.add_parameter('pos_bias_u', pos_bias_u)
pos_bias_v = self.create_parameter(
(self.h, self.d_k), default_initializer=I.XavierUniform())
self.add_parameter('pos_bias_v', pos_bias_v)
def rel_shift(self, x, zero_triu: bool=False):
"""Compute relative positinal encoding.
Args:
x (paddle.Tensor): Input tensor (batch, head, time1, time1).
zero_triu (bool): If true, return the lower triangular part of
the matrix.
Returns:
paddle.Tensor: Output tensor. (batch, head, time1, time1)
"""
zero_pad = paddle.zeros(
(x.size(0), x.size(1), x.size(2), 1), dtype=x.dtype)
x_padded = paddle.cat([zero_pad, x], dim=-1)
x_padded = x_padded.view(x.size(0), x.size(1), x.size(3) + 1, x.size(2))
x = x_padded[:, :, 1:].view_as(x) # [B, H, T1, T1]
if zero_triu:
ones = paddle.ones((x.size(2), x.size(3)))
x = x * paddle.tril(ones, x.size(3) - x.size(2))[None, None, :, :]
return x
def forward(self,
query: paddle.Tensor,
key: paddle.Tensor,
value: paddle.Tensor,
pos_emb: paddle.Tensor,
mask: Optional[paddle.Tensor]):
"""Compute 'Scaled Dot Product Attention' with rel. positional encoding.
Args:
query (paddle.Tensor): Query tensor (#batch, time1, size).
key (paddle.Tensor): Key tensor (#batch, time2, size).
value (paddle.Tensor): Value tensor (#batch, time2, size).
pos_emb (paddle.Tensor): Positional embedding tensor
(#batch, time1, size).
mask (paddle.Tensor): Mask tensor (#batch, 1, time2) or
(#batch, time1, time2).
Returns:
paddle.Tensor: Output tensor (#batch, time1, d_model).
"""
q, k, v = self.forward_qkv(query, key, value)
q = q.transpose([0, 2, 1, 3]) # (batch, time1, head, d_k)
n_batch_pos = pos_emb.size(0)
p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k)
p = p.transpose([0, 2, 1, 3]) # (batch, head, time1, d_k)
# (batch, head, time1, d_k)
q_with_bias_u = (q + self.pos_bias_u).transpose([0, 2, 1, 3])
# (batch, head, time1, d_k)
q_with_bias_v = (q + self.pos_bias_v).transpose([0, 2, 1, 3])
# compute attention score
# first compute matrix a and matrix c
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
# (batch, head, time1, time2)
matrix_ac = paddle.matmul(q_with_bias_u, k.transpose([0, 1, 3, 2]))
# compute matrix b and matrix d
# (batch, head, time1, time2)
matrix_bd = paddle.matmul(q_with_bias_v, p.transpose([0, 1, 3, 2]))
# Remove rel_shift since it is useless in speech recognition,
# and it requires special attention for streaming.
# matrix_bd = self.rel_shift(matrix_bd)
scores = (matrix_ac + matrix_bd) / math.sqrt(
self.d_k) # (batch, head, time1, time2)
return self.forward_attention(v, scores, mask)

366
text_processing/speechtask/punctuation_restoration/modules/crf.py
Unescape View File

@ -0,0 +1,366 @@
# Copyright (c) 2021 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
__all__ = ['CRF']
class CRF(nn.Layer):
"""
Linear-chain Conditional Random Field (CRF).
Args:
nb_labels (int): number of labels in your tagset, including special symbols.
bos_tag_id (int): integer representing the beginning of sentence symbol in
your tagset.
eos_tag_id (int): integer representing the end of sentence symbol in your tagset.
pad_tag_id (int, optional): integer representing the pad symbol in your tagset.
If None, the model will treat the PAD as a normal tag. Otherwise, the model
will apply constraints for PAD transitions.
batch_first (bool): Whether the first dimension represents the batch dimension.
"""
def __init__(self,
nb_labels: int,
bos_tag_id: int,
eos_tag_id: int,
pad_tag_id: int=None,
batch_first: bool=True):
super().__init__()
self.nb_labels = nb_labels
self.BOS_TAG_ID = bos_tag_id
self.EOS_TAG_ID = eos_tag_id
self.PAD_TAG_ID = pad_tag_id
self.batch_first = batch_first
# initialize transitions from a random uniform distribution between -0.1 and 0.1
self.transitions = self.create_parameter(
[self.nb_labels, self.nb_labels],
default_initializer=nn.initializer.Uniform(-0.1, 0.1))
self.init_weights()
def init_weights(self):
# enforce contraints (rows=from, columns=to) with a big negative number
# so exp(-10000) will tend to zero
# no transitions allowed to the beginning of sentence
self.transitions[:, self.BOS_TAG_ID] = -10000.0
# no transition alloed from the end of sentence
self.transitions[self.EOS_TAG_ID, :] = -10000.0
if self.PAD_TAG_ID is not None:
# no transitions from padding
self.transitions[self.PAD_TAG_ID, :] = -10000.0
# no transitions to padding
self.transitions[:, self.PAD_TAG_ID] = -10000.0
# except if the end of sentence is reached
# or we are already in a pad position
self.transitions[self.PAD_TAG_ID, self.EOS_TAG_ID] = 0.0
self.transitions[self.PAD_TAG_ID, self.PAD_TAG_ID] = 0.0
def forward(self,
emissions: paddle.Tensor,
tags: paddle.Tensor,
mask: paddle.Tensor=None) -> paddle.Tensor:
"""Compute the negative log-likelihood. See `log_likelihood` method."""
nll = -self.log_likelihood(emissions, tags, mask=mask)
return nll
def log_likelihood(self, emissions, tags, mask=None):
"""Compute the probability of a sequence of tags given a sequence of
emissions scores.
Args:
emissions (paddle.Tensor): Sequence of emissions for each label.
Shape of (batch_size, seq_len, nb_labels) if batch_first is True,
(seq_len, batch_size, nb_labels) otherwise.
tags (paddle.LongTensor): Sequence of labels.
Shape of (batch_size, seq_len) if batch_first is True,
(seq_len, batch_size) otherwise.
mask (paddle.FloatTensor, optional): Tensor representing valid positions.
If None, all positions are considered valid.
Shape of (batch_size, seq_len) if batch_first is True,
(seq_len, batch_size) otherwise.
Returns:
paddle.Tensor: sum of the log-likelihoods for each sequence in the batch.
Shape of ()
"""
# fix tensors order by setting batch as the first dimension
if not self.batch_first:
emissions = emissions.transpose(0, 1)
tags = tags.transpose(0, 1)
if mask is None:
mask = paddle.ones(emissions.shape[:2], dtype=paddle.float)
scores = self._compute_scores(emissions, tags, mask=mask)
partition = self._compute_log_partition(emissions, mask=mask)
return paddle.sum(scores - partition)
def decode(self, emissions, mask=None):
"""Find the most probable sequence of labels given the emissions using
the Viterbi algorithm.
Args:
emissions (paddle.Tensor): Sequence of emissions for each label.
Shape (batch_size, seq_len, nb_labels) if batch_first is True,
(seq_len, batch_size, nb_labels) otherwise.
mask (paddle.FloatTensor, optional): Tensor representing valid positions.
If None, all positions are considered valid.
Shape (batch_size, seq_len) if batch_first is True,
(seq_len, batch_size) otherwise.
Returns:
paddle.Tensor: the viterbi score for the for each batch.
Shape of (batch_size,)
list of lists: the best viterbi sequence of labels for each batch. [B, T]
"""
# fix tensors order by setting batch as the first dimension
if not self.batch_first:
emissions = emissions.transpose(0, 1)
tags = tags.transpose(0, 1)
if mask is None:
mask = paddle.ones(emissions.shape[:2], dtype=paddle.float)
scores, sequences = self._viterbi_decode(emissions, mask)
return scores, sequences
def _compute_scores(self, emissions, tags, mask):
"""Compute the scores for a given batch of emissions with their tags.
Args:
emissions (paddle.Tensor): (batch_size, seq_len, nb_labels)
tags (Paddle.LongTensor): (batch_size, seq_len)
mask (Paddle.FloatTensor): (batch_size, seq_len)
Returns:
paddle.Tensor: Scores for each batch.
Shape of (batch_size,)
"""
batch_size, seq_length = tags.shape
scores = paddle.zeros([batch_size])
# save first and last tags to be used later
first_tags = tags[:, 0]
last_valid_idx = mask.int().sum(1) - 1
# TODO(Hui Zhang): not support fancy index.
# last_tags = tags.gather(last_valid_idx.unsqueeze(1), axis=1).squeeze()
batch_idx = paddle.arange(batch_size, dtype=last_valid_idx.dtype)
gather_last_valid_idx = paddle.stack(
[batch_idx, last_valid_idx], axis=-1)
last_tags = tags.gather_nd(gather_last_valid_idx)
# add the transition from BOS to the first tags for each batch
# t_scores = self.transitions[self.BOS_TAG_ID, first_tags]
t_scores = self.transitions[self.BOS_TAG_ID].gather(first_tags)
# add the [unary] emission scores for the first tags for each batch
# for all batches, the first word, see the correspondent emissions
# for the first tags (which is a list of ids):
# emissions[:, 0, [tag_1, tag_2, ..., tag_nblabels]]
# e_scores = emissions[:, 0].gather(1, first_tags.unsqueeze(1)).squeeze()
gather_first_tags_idx = paddle.stack([batch_idx, first_tags], axis=-1)
e_scores = emissions[:, 0].gather_nd(gather_first_tags_idx)
# the scores for a word is just the sum of both scores
scores += e_scores + t_scores
# now lets do this for each remaining word
for i in range(1, seq_length):
# we could: iterate over batches, check if we reached a mask symbol
# and stop the iteration, but vecotrizing is faster due to gpu,
# so instead we perform an element-wise multiplication
is_valid = mask[:, i]
previous_tags = tags[:, i - 1]
current_tags = tags[:, i]
# calculate emission and transition scores as we did before
# e_scores = emissions[:, i].gather(1, current_tags.unsqueeze(1)).squeeze()
gather_current_tags_idx = paddle.stack(
[batch_idx, current_tags], axis=-1)
e_scores = emissions[:, i].gather_nd(gather_current_tags_idx)
# t_scores = self.transitions[previous_tags, current_tags]
gather_transitions_idx = paddle.stack(
[previous_tags, current_tags], axis=-1)
t_scores = self.transitions.gather_nd(gather_transitions_idx)
# apply the mask
e_scores = e_scores * is_valid
t_scores = t_scores * is_valid
scores += e_scores + t_scores
# add the transition from the end tag to the EOS tag for each batch
# scores += self.transitions[last_tags, self.EOS_TAG_ID]
scores += self.transitions.gather(last_tags)[:, self.EOS_TAG_ID]
return scores
def _compute_log_partition(self, emissions, mask):
"""Compute the partition function in log-space using the forward-algorithm.
Args:
emissions (paddle.Tensor): (batch_size, seq_len, nb_labels)
mask (Paddle.FloatTensor): (batch_size, seq_len)
Returns:
paddle.Tensor: the partition scores for each batch.
Shape of (batch_size,)
"""
batch_size, seq_length, nb_labels = emissions.shape
# in the first iteration, BOS will have all the scores
alphas = self.transitions[self.BOS_TAG_ID, :].unsqueeze(
0) + emissions[:, 0]
for i in range(1, seq_length):
# (bs, nb_labels) -> (bs, 1, nb_labels)
e_scores = emissions[:, i].unsqueeze(1)
# (nb_labels, nb_labels) -> (bs, nb_labels, nb_labels)
t_scores = self.transitions.unsqueeze(0)
# (bs, nb_labels) -> (bs, nb_labels, 1)
a_scores = alphas.unsqueeze(2)
scores = e_scores + t_scores + a_scores
new_alphas = paddle.logsumexp(scores, axis=1)
# set alphas if the mask is valid, otherwise keep the current values
is_valid = mask[:, i].unsqueeze(-1)
alphas = is_valid * new_alphas + (1 - is_valid) * alphas
# add the scores for the final transition
last_transition = self.transitions[:, self.EOS_TAG_ID]
end_scores = alphas + last_transition.unsqueeze(0)
# return a *log* of sums of exps
return paddle.logsumexp(end_scores, axis=1)
def _viterbi_decode(self, emissions, mask):
"""Compute the viterbi algorithm to find the most probable sequence of labels
given a sequence of emissions.
Args:
emissions (paddle.Tensor): (batch_size, seq_len, nb_labels)
mask (Paddle.FloatTensor): (batch_size, seq_len)
Returns:
paddle.Tensor: the viterbi score for the for each batch.
Shape of (batch_size,)
list of lists of ints: the best viterbi sequence of labels for each batch
"""
batch_size, seq_length, nb_labels = emissions.shape
# in the first iteration, BOS will have all the scores and then, the max
alphas = self.transitions[self.BOS_TAG_ID, :].unsqueeze(
0) + emissions[:, 0]
backpointers = []
for i in range(1, seq_length):
# (bs, nb_labels) -> (bs, 1, nb_labels)
e_scores = emissions[:, i].unsqueeze(1)
# (nb_labels, nb_labels) -> (bs, nb_labels, nb_labels)
t_scores = self.transitions.unsqueeze(0)
# (bs, nb_labels) -> (bs, nb_labels, 1)
a_scores = alphas.unsqueeze(2)
# combine current scores with previous alphas
scores = e_scores + t_scores + a_scores
# so far is exactly like the forward algorithm,
# but now, instead of calculating the logsumexp,
# we will find the highest score and the tag associated with it
# max_scores, max_score_tags = paddle.max(scores, axis=1)
max_scores = paddle.max(scores, axis=1)
max_score_tags = paddle.argmax(scores, axis=1)
# set alphas if the mask is valid, otherwise keep the current values
is_valid = mask[:, i].unsqueeze(-1)
alphas = is_valid * max_scores + (1 - is_valid) * alphas
# add the max_score_tags for our list of backpointers
# max_scores has shape (batch_size, nb_labels) so we transpose it to
# be compatible with our previous loopy version of viterbi
backpointers.append(max_score_tags.t())
# add the scores for the final transition
last_transition = self.transitions[:, self.EOS_TAG_ID]
end_scores = alphas + last_transition.unsqueeze(0)
# get the final most probable score and the final most probable tag
# max_final_scores, max_final_tags = paddle.max(end_scores, axis=1)
max_final_scores = paddle.max(end_scores, axis=1)
max_final_tags = paddle.argmax(end_scores, axis=1)
# find the best sequence of labels for each sample in the batch
best_sequences = []
emission_lengths = mask.int().sum(axis=1)
for i in range(batch_size):
# recover the original sentence length for the i-th sample in the batch
sample_length = emission_lengths[i].item()
# recover the max tag for the last timestep
sample_final_tag = max_final_tags[i].item()
# limit the backpointers until the last but one
# since the last corresponds to the sample_final_tag
sample_backpointers = backpointers[:sample_length - 1]
# follow the backpointers to build the sequence of labels
sample_path = self._find_best_path(i, sample_final_tag,
sample_backpointers)
# add this path to the list of best sequences
best_sequences.append(sample_path)
return max_final_scores, best_sequences
def _find_best_path(self, sample_id, best_tag, backpointers):
"""Auxiliary function to find the best path sequence for a specific sample.
Args:
sample_id (int): sample index in the range [0, batch_size)
best_tag (int): tag which maximizes the final score
backpointers (list of lists of tensors): list of pointers with
shape (seq_len_i-1, nb_labels, batch_size) where seq_len_i
represents the length of the ith sample in the batch
Returns:
list of ints: a list of tag indexes representing the bast path
"""
# add the final best_tag to our best path
best_path = [best_tag]
# traverse the backpointers in backwards
for backpointers_t in reversed(backpointers):
# recover the best_tag at this timestep
best_tag = backpointers_t[best_tag][sample_id].item()
# append to the beginning of the list so we don't need to reverse it later
best_path.insert(0, best_tag)
return best_path

13
text_processing/speechtask/punctuation_restoration/training/__init__.py
Unescape View File

@ -0,0 +1,13 @@
# Copyright (c) 2021 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.

98
text_processing/speechtask/punctuation_restoration/training/loss.py
Unescape View File

@ -0,0 +1,98 @@
# Copyright (c) 2021 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
import paddle.nn as nn
import paddle.nn.functional as F
class FocalLossHX(nn.Layer):
def __init__(self, gamma=0, size_average=True):
super(FocalLoss, self).__init__()
self.gamma = gamma
self.size_average = size_average
def forward(self, input, target):
# print('input')
# print(input.shape)
# print(target.shape)
if input.dim() > 2:
input = paddle.reshape(
input,
shape=[input.size(0), input.size(1), -1]) # N,C,H,W => N,C,H*W
input = input.transpose(1, 2) # N,C,H*W => N,H*W,C
input = paddle.reshape(
input, shape=[-1, input.size(2)]) # N,H*W,C => N*H*W,C
target = paddle.reshape(target, shape=[-1])
logpt = F.log_softmax(input)
# print('logpt')
# print(logpt.shape)
# print(logpt)
# get true class column from each row
all_rows = paddle.arange(len(input))
# print(target)
log_pt = logpt.numpy()[all_rows.numpy(), target.numpy()]
pt = paddle.to_tensor(log_pt, dtype='float64').exp()
ce = F.cross_entropy(input, target, reduction='none')
# print('ce')
# print(ce.shape)
loss = (1 - pt)**self.gamma * ce
# print('ce:%f'%ce.mean())
# print('fl:%f'%loss.mean())
if self.size_average:
return loss.mean()
else:
return loss.sum()
class FocalLoss(nn.Layer):
"""
Focal Loss.
Code referenced from:
https://github.com/clcarwin/focal_loss_pytorch/blob/master/focalloss.py
Args:
gamma (float): the coefficient of Focal Loss.
ignore_index (int64): Specifies a target value that is ignored
and does not contribute to the input gradient. Default ``255``.
"""
def __init__(self, gamma=2.0):
super(FocalLoss, self).__init__()
self.gamma = gamma
def forward(self, logit, label):
#####logit = F.softmax(logit)
# logit = paddle.reshape(
# logit, [logit.shape[0], logit.shape[1], -1]) # N,C,H,W => N,C,H*W
# logit = paddle.transpose(logit, [0, 2, 1]) # N,C,H*W => N,H*W,C
# logit = paddle.reshape(logit,
# [-1, logit.shape[2]]) # N,H*W,C => N*H*W,C
label = paddle.reshape(label, [-1, 1])
range_ = paddle.arange(0, label.shape[0])
range_ = paddle.unsqueeze(range_, axis=-1)
label = paddle.cast(label, dtype='int64')
label = paddle.concat([range_, label], axis=-1)
logpt = F.log_softmax(logit)
logpt = paddle.gather_nd(logpt, label)
pt = paddle.exp(logpt.detach())
loss = -1 * (1 - pt)**self.gamma * logpt
loss = paddle.mean(loss)
# print(loss)
# print(logpt)
return loss

651
text_processing/speechtask/punctuation_restoration/training/trainer.py
Unescape View File

@ -0,0 +1,651 @@
# Copyright (c) 2021 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 logging
import time
from collections import defaultdict
from pathlib import Path
import numpy as np
import paddle
import paddle.nn as nn
import pandas as pd
from paddle import distributed as dist
from paddle.io import DataLoader
from sklearn.metrics import classification_report
from sklearn.metrics import f1_score
from sklearn.metrics import precision_recall_fscore_support
from speechtask.punctuation_restoration.io.dataset import PuncDataset
from speechtask.punctuation_restoration.io.dataset import PuncDatasetFromBertTokenizer
from speechtask.punctuation_restoration.model.BertBLSTM import BertBLSTMPunc
from speechtask.punctuation_restoration.model.BertLinear import BertLinearPunc
from speechtask.punctuation_restoration.model.blstm import BiLSTM
from speechtask.punctuation_restoration.model.lstm import RnnLm
from speechtask.punctuation_restoration.utils import layer_tools
from speechtask.punctuation_restoration.utils import mp_tools
from speechtask.punctuation_restoration.utils.checkpoint import Checkpoint
from tensorboardX import SummaryWriter
__all__ = ["Trainer", "Tester"]
DefinedClassifier = {
"lstm": RnnLm,
"blstm": BiLSTM,
"BertLinear": BertLinearPunc,
"BertBLSTM": BertBLSTMPunc
}
DefinedLoss = {
"ce": nn.CrossEntropyLoss,
}
DefinedDataset = {
'PuncCh': PuncDataset,
'Bert': PuncDatasetFromBertTokenizer,
}
class Trainer():
"""
An experiment template in order to structure the training code and take
care of saving, loading, logging, visualization stuffs. It"s intended to
be flexible and simple.
So it only handles output directory (create directory for the output,
create a checkpoint directory, dump the config in use and create
visualizer and logger) in a standard way without enforcing any
input-output protocols to the model and dataloader. It leaves the main
part for the user to implement their own (setup the model, criterion,
optimizer, define a training step, define a validation function and
customize all the text and visual logs).
It does not save too much boilerplate code. The users still have to write
the forward/backward/update mannually, but they are free to add
non-standard behaviors if needed.
We have some conventions to follow.
1. Experiment should have ``model``, ``optimizer``, ``train_loader`` and
``valid_loader``, ``config`` and ``args`` attributes.
2. The config should have a ``training`` field, which has
``valid_interval``, ``save_interval`` and ``max_iteration`` keys. It is
used as the trigger to invoke validation, checkpointing and stop of the
experiment.
3. There are four methods, namely ``train_batch``, ``valid``,
``setup_model`` and ``setup_dataloader`` that should be implemented.
Feel free to add/overwrite other methods and standalone functions if you
need.
Parameters
----------
config: yacs.config.CfgNode
The configuration used for the experiment.
args: argparse.Namespace
The parsed command line arguments.
Examples
--------
>>> def main_sp(config, args):
>>> exp = Trainer(config, args)
>>> exp.setup()
>>> exp.run()
>>>
>>> config = get_cfg_defaults()
>>> parser = default_argument_parser()
>>> args = parser.parse_args()
>>> if args.config:
>>> config.merge_from_file(args.config)
>>> if args.opts:
>>> config.merge_from_list(args.opts)
>>> config.freeze()
>>>
>>> if args.nprocs > 1 and args.device == "gpu":
>>> dist.spawn(main_sp, args=(config, args), nprocs=args.nprocs)
>>> else:
>>> main_sp(config, args)
"""
def __init__(self, config, args):
self.config = config
self.args = args
self.optimizer = None
self.visualizer = None
self.output_dir = None
self.checkpoint_dir = None
self.iteration = 0
self.epoch = 0
def setup(self):
"""Setup the experiment.
"""
self.setup_logger()
paddle.set_device(self.args.device)
if self.parallel:
self.init_parallel()
self.setup_output_dir()
self.dump_config()
self.setup_visualizer()
self.setup_checkpointer()
self.setup_model()
self.setup_dataloader()
self.iteration = 0
self.epoch = 0
@property
def parallel(self):
"""A flag indicating whether the experiment should run with
multiprocessing.
"""
return self.args.device == "gpu" and self.args.nprocs > 1
def init_parallel(self):
"""Init environment for multiprocess training.
"""
dist.init_parallel_env()
@mp_tools.rank_zero_only
def save(self, tag=None, infos: dict=None):
"""Save checkpoint (model parameters and optimizer states).
Args:
tag (int or str, optional): None for step, else using tag, e.g epoch. Defaults to None.
infos (dict, optional): meta data to save. Defaults to None.
"""
infos = infos if infos else dict()
infos.update({
"step": self.iteration,
"epoch": self.epoch,
"lr": self.optimizer.get_lr()
})
self.checkpointer.add_checkpoint(self.checkpoint_dir, self.iteration
if tag is None else tag, self.model,
self.optimizer, infos)
def resume_or_scratch(self):
"""Resume from latest checkpoint at checkpoints in the output
directory or load a specified checkpoint.
If ``args.checkpoint_path`` is not None, load the checkpoint, else
resume training.
"""
scratch = None
infos = self.checkpointer.load_parameters(
self.model,
self.optimizer,
checkpoint_dir=self.checkpoint_dir,
checkpoint_path=self.args.checkpoint_path)
if infos:
# restore from ckpt
self.iteration = infos["step"]
self.epoch = infos["epoch"]
scratch = False
else:
self.iteration = 0
self.epoch = 0
scratch = True
return scratch
def new_epoch(self):
"""Reset the train loader seed and increment `epoch`.
"""
self.epoch += 1
if self.parallel:
self.train_loader.batch_sampler.set_epoch(self.epoch)
def train(self):
"""The training process control by epoch."""
from_scratch = self.resume_or_scratch()
if from_scratch:
# save init model, i.e. 0 epoch
self.save(tag="init")
self.lr_scheduler.step(self.iteration)
if self.parallel:
self.train_loader.batch_sampler.set_epoch(self.epoch)
self.logger.info(
f"Train Total Examples: {len(self.train_loader.dataset)}")
self.punc_list = []
for i in range(len(self.train_loader.dataset.id2punc)):
self.punc_list.append(self.train_loader.dataset.id2punc[i])
while self.epoch < self.config["training"]["n_epoch"]:
self.model.train()
self.total_label_train = []
self.total_predict_train = []
try:
data_start_time = time.time()
for batch_index, batch in enumerate(self.train_loader):
dataload_time = time.time() - data_start_time
msg = "Train: Rank: {}, ".format(dist.get_rank())
msg += "epoch: {}, ".format(self.epoch)
msg += "step: {}, ".format(self.iteration)
msg += "batch : {}/{}, ".format(batch_index + 1,
len(self.train_loader))
msg += "lr: {:>.8f}, ".format(self.lr_scheduler())
msg += "data time: {:>.3f}s, ".format(dataload_time)
self.train_batch(batch_index, batch, msg)
data_start_time = time.time()
t = classification_report(
self.total_label_train,
self.total_predict_train,
target_names=self.punc_list)
self.logger.info(t)
except Exception as e:
self.logger.error(e)
raise e
total_loss, F1_score = self.valid()
self.logger.info("Epoch {} Val info val_loss {}, F1_score {}".
format(self.epoch, total_loss, F1_score))
if self.visualizer:
self.visualizer.add_scalars("epoch", {
"total_loss": total_loss,
"lr": self.lr_scheduler()
}, self.epoch)
self.save(
tag=self.epoch, infos={"val_loss": total_loss,
"F1": F1_score})
# step lr every epoch
self.lr_scheduler.step()
self.new_epoch()
def run(self):
"""The routine of the experiment after setup. This method is intended
to be used by the user.
"""
try:
self.train()
except KeyboardInterrupt:
self.save()
exit(-1)
finally:
self.destory()
self.logger.info("Training Done.")
def setup_output_dir(self):
"""Create a directory used for output.
"""
# output dir
output_dir = Path(self.args.output).expanduser()
output_dir.mkdir(parents=True, exist_ok=True)
self.output_dir = output_dir
def setup_checkpointer(self):
"""Create a directory used to save checkpoints into.
It is "checkpoints" inside the output directory.
"""
# checkpoint dir
self.checkpointer = Checkpoint(self.logger,
self.config["checkpoint"]["kbest_n"],
self.config["checkpoint"]["latest_n"])
checkpoint_dir = self.output_dir / "checkpoints"
checkpoint_dir.mkdir(exist_ok=True)
self.checkpoint_dir = checkpoint_dir
def setup_logger(self):
LOG_FORMAT = "%(asctime)s - %(pathname)s[line:%(lineno)d] - %(levelname)s: %(message)s"
format_str = logging.Formatter(
'%(asctime)s - %(pathname)s[line:%(lineno)d] - %(levelname)s: %(message)s'
)
logging.basicConfig(
filename=self.config["training"]["log_path"],
level=logging.INFO,
format=LOG_FORMAT)
self.logger = logging.getLogger(__name__)
# self.logger = logging.getLogger(self.config["training"]["log_path"].strip().split('/')[-1].split('.')[0])
self.logger.setLevel(logging.INFO) #设置日志级别
sh = logging.StreamHandler() #往屏幕上输出
sh.setFormatter(format_str) #设置屏幕上显示的格式
self.logger.addHandler(sh) #把对象加到logger里
self.logger.info('info')
print("setup logger!!!")
@mp_tools.rank_zero_only
def destory(self):
"""Close visualizer to avoid hanging after training"""
# https://github.com/pytorch/fairseq/issues/2357
if self.visualizer:
self.visualizer.close()
@mp_tools.rank_zero_only
def setup_visualizer(self):
"""Initialize a visualizer to log the experiment.
The visual log is saved in the output directory.
Notes
------
Only the main process has a visualizer with it. Use multiple
visualizers in multiprocess to write to a same log file may cause
unexpected behaviors.
"""
# visualizer
visualizer = SummaryWriter(logdir=str(self.output_dir))
self.visualizer = visualizer
@mp_tools.rank_zero_only
def dump_config(self):
"""Save the configuration used for this experiment.
It is saved in to ``config.yaml`` in the output directory at the
beginning of the experiment.
"""
with open(self.output_dir / "config.yaml", "wt") as f:
print(self.config, file=f)
def train_batch(self, batch_index, batch_data, msg):
start = time.time()
input, label = batch_data
label = paddle.reshape(label, shape=[-1])
y, logit = self.model(input)
pred = paddle.argmax(logit, axis=1)
self.total_label_train.extend(label.numpy().tolist())
self.total_predict_train.extend(pred.numpy().tolist())
# self.total_predict.append(logit.numpy().tolist())
# print('--after model----')
# # print(label.shape)
# # print(pred.shape)
# # print('--!!!!!!!!!!!!!----')
# print("self.total_label")
# print(self.total_label)
# print("self.total_predict")
# print(self.total_predict)
loss = self.crit(y, label)
loss.backward()
layer_tools.print_grads(self.model, print_func=None)
self.optimizer.step()
self.optimizer.clear_grad()
iteration_time = time.time() - start
losses_np = {
"train_loss": float(loss),
}
msg += "train time: {:>.3f}s, ".format(iteration_time)
msg += "batch size: {}, ".format(self.config["data"]["batch_size"])
msg += ", ".join("{}: {:>.6f}".format(k, v)
for k, v in losses_np.items())
self.logger.info(msg)
# print(msg)
if dist.get_rank() == 0 and self.visualizer:
for k, v in losses_np.items():
self.visualizer.add_scalar("train/{}".format(k), v,
self.iteration)
self.iteration += 1
@paddle.no_grad()
def valid(self):
self.logger.info(
f"Valid Total Examples: {len(self.valid_loader.dataset)}")
self.model.eval()
valid_losses = defaultdict(list)
num_seen_utts = 1
total_loss = 0.0
valid_total_label = []
valid_total_predict = []
for i, batch in enumerate(self.valid_loader):
input, label = batch
label = paddle.reshape(label, shape=[-1])
y, logit = self.model(input)
pred = paddle.argmax(logit, axis=1)
valid_total_label.extend(label.numpy().tolist())
valid_total_predict.extend(pred.numpy().tolist())
loss = self.crit(y, label)
if paddle.isfinite(loss):
num_utts = batch[1].shape[0]
num_seen_utts += num_utts
total_loss += float(loss) * num_utts
valid_losses["val_loss"].append(float(loss))
if (i + 1) % self.config["training"]["log_interval"] == 0:
valid_dump = {k: np.mean(v) for k, v in valid_losses.items()}
valid_dump["val_history_loss"] = total_loss / num_seen_utts
# logging
msg = f"Valid: Rank: {dist.get_rank()}, "
msg += "epoch: {}, ".format(self.epoch)
msg += "step: {}, ".format(self.iteration)
msg += "batch : {}/{}, ".format(i + 1, len(self.valid_loader))
msg += ", ".join("{}: {:>.6f}".format(k, v)
for k, v in valid_dump.items())
self.logger.info(msg)
# print(msg)
self.logger.info("Rank {} Val info val_loss {}".format(
dist.get_rank(), total_loss / num_seen_utts))
# print("Rank {} Val info val_loss {} acc: {}".format(
# dist.get_rank(), total_loss / num_seen_utts, acc))
F1_score = f1_score(
valid_total_label, valid_total_predict, average="macro")
return total_loss / num_seen_utts, F1_score
def setup_model(self):
config = self.config
model = DefinedClassifier[self.config["model_type"]](
**self.config["model_params"])
self.crit = DefinedLoss[self.config["loss_type"]](**self.config[
"loss"]) if "loss_type" in self.config else DefinedLoss["ce"]()
if self.parallel:
model = paddle.DataParallel(model)
self.logger.info(f"{model}")
layer_tools.print_params(model, self.logger.info)
lr_scheduler = paddle.optimizer.lr.ExponentialDecay(
learning_rate=config["training"]["lr"],
gamma=config["training"]["lr_decay"],
verbose=True)
optimizer = paddle.optimizer.Adam(
learning_rate=lr_scheduler,
parameters=model.parameters(),
weight_decay=paddle.regularizer.L2Decay(
config["training"]["weight_decay"]))
self.model = model
self.optimizer = optimizer
self.lr_scheduler = lr_scheduler
self.logger.info("Setup model/criterion/optimizer/lr_scheduler!")
def setup_dataloader(self):
print("setup_dataloader!!!")
config = self.config["data"].copy()
print(config["batch_size"])
train_dataset = DefinedDataset[config["dataset_type"]](
train_path=config["train_path"], **config["data_params"])
dev_dataset = DefinedDataset[config["dataset_type"]](
train_path=config["dev_path"], **config["data_params"])
# train_dataset = config["dataset_type"](os.path.join(config["save_path"], "train"),
# os.path.join(config["save_path"], config["vocab_file"]),
# os.path.join(config["save_path"], config["punc_file"]),
# config["seq_len"])
# dev_dataset = PuncDataset(os.path.join(config["save_path"], "dev"),
# os.path.join(config["save_path"], config["vocab_file"]),
# os.path.join(config["save_path"], config["punc_file"]),
# config["seq_len"])
# if self.parallel:
# batch_sampler = SortagradDistributedBatchSampler(
# train_dataset,
# batch_size=config["batch_size"],
# num_replicas=None,
# rank=None,
# shuffle=True,
# drop_last=True,
# sortagrad=config["sortagrad"],
# shuffle_method=config["shuffle_method"])
# else:
# batch_sampler = SortagradBatchSampler(
# train_dataset,
# shuffle=True,
# batch_size=config["batch_size"],
# drop_last=True,
# sortagrad=config["sortagrad"],
# shuffle_method=config["shuffle_method"])
self.train_loader = DataLoader(
train_dataset,
num_workers=config["num_workers"],
batch_size=config["batch_size"])
self.valid_loader = DataLoader(
dev_dataset,
batch_size=config["batch_size"],
shuffle=False,
drop_last=False,
num_workers=config["num_workers"])
self.logger.info("Setup train/valid Dataloader!")
class Tester(Trainer):
def __init__(self, config, args):
super().__init__(config, args)
@mp_tools.rank_zero_only
@paddle.no_grad()
def test(self):
self.logger.info(
f"Test Total Examples: {len(self.test_loader.dataset)}")
self.punc_list = []
for i in range(len(self.test_loader.dataset.id2punc)):
self.punc_list.append(self.test_loader.dataset.id2punc[i])
self.model.eval()
test_total_label = []
test_total_predict = []
with open(self.args.result_file, 'w') as fout:
for i, batch in enumerate(self.test_loader):
input, label = batch
label = paddle.reshape(label, shape=[-1])
y, logit = self.model(input)
pred = paddle.argmax(logit, axis=1)
test_total_label.extend(label.numpy().tolist())
test_total_predict.extend(pred.numpy().tolist())
# print(type(logit))
# logging
msg = "Test: "
msg += "epoch: {}, ".format(self.epoch)
msg += "step: {}, ".format(self.iteration)
self.logger.info(msg)
# print(msg)
t = classification_report(
test_total_label, test_total_predict, target_names=self.punc_list)
print(t)
t2 = self.evaluation(test_total_label, test_total_predict)
print(t2)
def evaluation(self, y_pred, y_test):
precision, recall, f1, _ = precision_recall_fscore_support(
y_test, y_pred, average=None, labels=[1, 2, 3])
overall = precision_recall_fscore_support(
y_test, y_pred, average='macro', labels=[1, 2, 3])
result = pd.DataFrame(
np.array([precision, recall, f1]),
columns=list(['O', 'COMMA', 'PERIOD', 'QUESTION'])[1:],
index=['Precision', 'Recall', 'F1'])
result['OVERALL'] = overall[:3]
return result
def run_test(self):
self.resume_or_scratch()
try:
self.test()
except KeyboardInterrupt:
exit(-1)
def setup(self):
"""Setup the experiment.
"""
paddle.set_device(self.args.device)
self.setup_logger()
self.setup_output_dir()
self.setup_checkpointer()
self.setup_dataloader()
self.setup_model()
self.iteration = 0
self.epoch = 0
def setup_model(self):
config = self.config
model = DefinedClassifier[self.config["model_type"]](
**self.config["model_params"])
self.model = model
self.logger.info("Setup model!")
def setup_dataloader(self):
config = self.config["data"].copy()
test_dataset = DefinedDataset[config["dataset_type"]](
train_path=config["test_path"], **config["data_params"])
self.test_loader = DataLoader(
test_dataset,
batch_size=config["batch_size"],
shuffle=False,
drop_last=False)
self.logger.info("Setup test Dataloader!")
def setup_output_dir(self):
"""Create a directory used for output.
"""
# output dir
if self.args.output:
output_dir = Path(self.args.output).expanduser()
output_dir.mkdir(parents=True, exist_ok=True)
else:
output_dir = Path(
self.args.checkpoint_path).expanduser().parent.parent
output_dir.mkdir(parents=True, exist_ok=True)
self.output_dir = output_dir
def setup_logger(self):
LOG_FORMAT = "%(asctime)s - %(pathname)s[line:%(lineno)d] - %(levelname)s: %(message)s"
format_str = logging.Formatter(
'%(asctime)s - %(pathname)s[line:%(lineno)d] - %(levelname)s: %(message)s'
)
logging.basicConfig(
filename=self.config["testing"]["log_path"],
level=logging.INFO,
format=LOG_FORMAT)
self.logger = logging.getLogger(__name__)
# self.logger = logging.getLogger(self.config["training"]["log_path"].strip().split('/')[-1].split('.')[0])
self.logger.setLevel(logging.INFO) #设置日志级别
sh = logging.StreamHandler() #往屏幕上输出
sh.setFormatter(format_str) #设置屏幕上显示的格式
self.logger.addHandler(sh) #把对象加到logger里
self.logger.info('info')
print("setup test logger!!!")

13
text_processing/speechtask/punctuation_restoration/utils/__init__.py
Unescape View File

@ -0,0 +1,13 @@
# Copyright (c) 2021 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.

304
text_processing/speechtask/punctuation_restoration/utils/checkpoint.py
Unescape View File

@ -0,0 +1,304 @@
# Copyright (c) 2021 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 glob
import json
import os
import re
from pathlib import Path
from typing import Text
from typing import Union
import paddle
from paddle import distributed as dist
from paddle.optimizer import Optimizer
from speechtask.punctuation_restoration.utils import mp_tools
# from speechtask.punctuation_restoration.utils.log import Log
# logger = Log(__name__).getlog()
__all__ = ["Checkpoint"]
class Checkpoint():
def __init__(self,
logger,
kbest_n: int=5,
latest_n: int=1,
metric_type='val_loss'):
self.best_records: Mapping[Path, float] = {}
self.latest_records = []
self.kbest_n = kbest_n
self.latest_n = latest_n
self._save_all = (kbest_n == -1)
self.logger = logger
self.metric_type = metric_type
def add_checkpoint(self,
checkpoint_dir,
tag_or_iteration: Union[int, Text],
model: paddle.nn.Layer,
optimizer: Optimizer=None,
infos: dict=None):
"""Save checkpoint in best_n and latest_n.
Args:
checkpoint_dir (str): the directory where checkpoint is saved.
tag_or_iteration (int or str): the latest iteration(step or epoch) number or tag.
model (Layer): model to be checkpointed.
optimizer (Optimizer, optional): optimizer to be checkpointed.
infos (dict or None)): any info you want to save.
metric_type (str, optional): metric type. Defaults to 'val_loss'.
"""
metric_type = self.metric_type
if (metric_type not in infos.keys()):
self._save_parameters(checkpoint_dir, tag_or_iteration, model,
optimizer, infos)
return
#save best
if self._should_save_best(infos[metric_type]):
self._save_best_checkpoint_and_update(
infos[metric_type], checkpoint_dir, tag_or_iteration, model,
optimizer, infos)
#save latest
self._save_latest_checkpoint_and_update(
checkpoint_dir, tag_or_iteration, model, optimizer, infos)
if isinstance(tag_or_iteration, int):
self._save_checkpoint_record(checkpoint_dir, tag_or_iteration)
def load_parameters(self,
model,
optimizer=None,
checkpoint_dir=None,
checkpoint_path=None,
record_file="checkpoint_latest"):
"""Load a last model checkpoint from disk.
Args:
model (Layer): model to load parameters.
optimizer (Optimizer, optional): optimizer to load states if needed.
Defaults to None.
checkpoint_dir (str, optional): the directory where checkpoint is saved.
checkpoint_path (str, optional): if specified, load the checkpoint
stored in the checkpoint_path(prefix) and the argument 'checkpoint_dir' will
be ignored. Defaults to None.
record_file "checkpoint_latest" or "checkpoint_best"
Returns:
configs (dict): epoch or step, lr and other meta info should be saved.
"""
configs = {}
if checkpoint_path is not None:
pass
elif checkpoint_dir is not None and record_file is not None:
# load checkpint from record file
checkpoint_record = os.path.join(checkpoint_dir, record_file)
iteration = self._load_checkpoint_idx(checkpoint_record)
if iteration == -1:
return configs
checkpoint_path = os.path.join(checkpoint_dir,
"{}".format(iteration))
else:
raise ValueError(
"At least one of 'checkpoint_path' or 'checkpoint_dir' should be specified!"
)
rank = dist.get_rank()
params_path = checkpoint_path + ".pdparams"
model_dict = paddle.load(params_path)
model.set_state_dict(model_dict)
self.logger.info(
"Rank {}: loaded model from {}".format(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)
self.logger.info("Rank {}: loaded optimizer state from {}".format(
rank, optimizer_path))
info_path = re.sub('.pdparams$', '.json', params_path)
if os.path.exists(info_path):
with open(info_path, 'r') as fin:
configs = json.load(fin)
return configs
def load_latest_parameters(self,
model,
optimizer=None,
checkpoint_dir=None,
checkpoint_path=None):
"""Load a last model checkpoint from disk.
Args:
model (Layer): model to load parameters.
optimizer (Optimizer, optional): optimizer to load states if needed.
Defaults to None.
checkpoint_dir (str, optional): the directory where checkpoint is saved.
checkpoint_path (str, optional): if specified, load the checkpoint
stored in the checkpoint_path(prefix) and the argument 'checkpoint_dir' will
be ignored. Defaults to None.
Returns:
configs (dict): epoch or step, lr and other meta info should be saved.
"""
return self.load_parameters(model, optimizer, checkpoint_dir,
checkpoint_path, "checkpoint_latest")
def load_best_parameters(self,
model,
optimizer=None,
checkpoint_dir=None,
checkpoint_path=None):
"""Load a last model checkpoint from disk.
Args:
model (Layer): model to load parameters.
optimizer (Optimizer, optional): optimizer to load states if needed.
Defaults to None.
checkpoint_dir (str, optional): the directory where checkpoint is saved.
checkpoint_path (str, optional): if specified, load the checkpoint
stored in the checkpoint_path(prefix) and the argument 'checkpoint_dir' will
be ignored. Defaults to None.
Returns:
configs (dict): epoch or step, lr and other meta info should be saved.
"""
return self.load_parameters(model, optimizer, checkpoint_dir,
checkpoint_path, "checkpoint_best")
def _should_save_best(self, metric: float) -> bool:
if not self._best_full():
return True
# already full
worst_record_path = max(self.best_records, key=self.best_records.get)
# worst_record_path = max(self.best_records.iteritems(), key=operator.itemgetter(1))[0]
worst_metric = self.best_records[worst_record_path]
return metric < worst_metric
def _best_full(self):
return (not self._save_all) and len(self.best_records) == self.kbest_n
def _latest_full(self):
return len(self.latest_records) == self.latest_n
def _save_best_checkpoint_and_update(self, metric, checkpoint_dir,
tag_or_iteration, model, optimizer,
infos):
# remove the worst
if self._best_full():
worst_record_path = max(self.best_records,
key=self.best_records.get)
self.best_records.pop(worst_record_path)
if (worst_record_path not in self.latest_records):
self.logger.info(
"remove the worst checkpoint: {}".format(worst_record_path))
self._del_checkpoint(checkpoint_dir, worst_record_path)
# add the new one
self._save_parameters(checkpoint_dir, tag_or_iteration, model,
optimizer, infos)
self.best_records[tag_or_iteration] = metric
def _save_latest_checkpoint_and_update(
self, checkpoint_dir, tag_or_iteration, model, optimizer, infos):
# remove the old
if self._latest_full():
to_del_fn = self.latest_records.pop(0)
if (to_del_fn not in self.best_records.keys()):
self.logger.info(
"remove the latest checkpoint: {}".format(to_del_fn))
self._del_checkpoint(checkpoint_dir, to_del_fn)
self.latest_records.append(tag_or_iteration)
self._save_parameters(checkpoint_dir, tag_or_iteration, model,
optimizer, infos)
def _del_checkpoint(self, checkpoint_dir, tag_or_iteration):
checkpoint_path = os.path.join(checkpoint_dir,
"{}".format(tag_or_iteration))
for filename in glob.glob(checkpoint_path + ".*"):
os.remove(filename)
self.logger.info("delete file: {}".format(filename))
def _load_checkpoint_idx(self, checkpoint_record: str) -> int:
"""Get the iteration number corresponding to the latest saved checkpoint.
Args:
checkpoint_path (str): the saved path of checkpoint.
Returns:
int: the latest iteration number. -1 for no checkpoint to load.
"""
if not os.path.isfile(checkpoint_record):
return -1
# Fetch the latest checkpoint index.
with open(checkpoint_record, "rt") as handle:
latest_checkpoint = handle.readlines()[-1].strip()
iteration = int(latest_checkpoint.split(":")[-1])
return iteration
def _save_checkpoint_record(self, checkpoint_dir: str, iteration: int):
"""Save the iteration number of the latest model to be checkpoint record.
Args:
checkpoint_dir (str): the directory where checkpoint is saved.
iteration (int): the latest iteration number.
Returns:
None
"""
checkpoint_record_latest = os.path.join(checkpoint_dir,
"checkpoint_latest")
checkpoint_record_best = os.path.join(checkpoint_dir, "checkpoint_best")
with open(checkpoint_record_best, "w") as handle:
for i in self.best_records.keys():
handle.write("model_checkpoint_path:{}\n".format(i))
with open(checkpoint_record_latest, "w") as handle:
for i in self.latest_records:
handle.write("model_checkpoint_path:{}\n".format(i))
@mp_tools.rank_zero_only
def _save_parameters(self,
checkpoint_dir: str,
tag_or_iteration: Union[int, str],
model: paddle.nn.Layer,
optimizer: Optimizer=None,
infos: dict=None):
"""Checkpoint the latest trained model parameters.
Args:
checkpoint_dir (str): the directory where checkpoint is saved.
tag_or_iteration (int or str): the latest iteration(step or epoch) number.
model (Layer): model to be checkpointed.
optimizer (Optimizer, optional): optimizer to be checkpointed.
Defaults to None.
infos (dict or None): any info you want to save.
Returns:
None
"""
checkpoint_path = os.path.join(checkpoint_dir,
"{}".format(tag_or_iteration))
model_dict = model.state_dict()
params_path = checkpoint_path + ".pdparams"
paddle.save(model_dict, params_path)
self.logger.info("Saved model to {}".format(params_path))
if optimizer:
opt_dict = optimizer.state_dict()
optimizer_path = checkpoint_path + ".pdopt"
paddle.save(opt_dict, optimizer_path)
self.logger.info(
"Saved optimzier state to {}".format(optimizer_path))
info_path = re.sub('.pdparams$', '.json', params_path)
infos = {} if infos is None else infos
with open(info_path, 'w') as fout:
data = json.dumps(infos)
fout.write(data)

74
text_processing/speechtask/punctuation_restoration/utils/default_parser.py
Unescape View File

@ -0,0 +1,74 @@
# Copyright (c) 2021 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.
This is used in examples with parakeet. And it is intended to be used by
other experiments with parakeet. It requires a minimal set of command line
arguments to start a training script.
The ``--config`` and ``--opts`` are used for overwrite the deault
configuration.
The ``--data`` and ``--output`` specifies the data path and output path.
Resuming training from existing progress at the output directory is the
intended default behavior.
The ``--checkpoint_path`` specifies the checkpoint to load from.
The ``--device`` and ``--nprocs`` specifies how to run the training.
See Also
--------
parakeet.training.experiment
Returns
-------
argparse.ArgumentParser
the parser
"""
parser = argparse.ArgumentParser()
# yapf: disable
# data and output
parser.add_argument("--config", metavar="FILE", help="path of the config file to overwrite to default config with.")
parser.add_argument("--dump-config", metavar="FILE", help="dump config to yaml file.")
# parser.add_argument("--data", metavar="DATA_DIR", help="path to the datatset.")
parser.add_argument("--output", metavar="OUTPUT_DIR", help="path to save checkpoint and logs.")
# load from saved checkpoint
parser.add_argument("--checkpoint_path", type=str, help="path of the checkpoint to load")
# save jit model to
parser.add_argument("--export_path", type=str, help="path of the jit model to save")
# save asr result to
parser.add_argument("--result_file", type=str, help="path of save the asr result")
# running
parser.add_argument("--device", type=str, default='gpu', choices=["cpu", "gpu"],
help="device type to use, cpu and gpu are supported.")
parser.add_argument("--nprocs", type=int, default=1, help="number of parallel processes to use.")
# 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")
parser.add_argument("--opts", type=str, default=[], nargs='+',
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs")
# yapd: enable
return parser

88
text_processing/speechtask/punctuation_restoration/utils/layer_tools.py
Unescape View File

@ -0,0 +1,88 @@
# Copyright (c) 2021 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 import nn
__all__ = [
"summary", "gradient_norm", "freeze", "unfreeze", "print_grads",
"print_params"
]
def summary(layer: nn.Layer, print_func=print):
if print_func is None:
return
num_params = num_elements = 0
for name, param in layer.state_dict().items():
if print_func:
print_func(
"{} | {} | {}".format(name, param.shape, np.prod(param.shape)))
num_elements += np.prod(param.shape)
num_params += 1
if print_func:
num_elements = num_elements / 1024**2
print_func(
f"Total parameters: {num_params}, {num_elements:.2f}M elements.")
def print_grads(model, print_func=print):
if print_func is None:
return
for n, p in model.named_parameters():
msg = f"param grad: {n}: shape: {p.shape} grad: {p.grad}"
print_func(msg)
def print_params(model, print_func=print):
if print_func is None:
return
total = 0.0
num_params = 0.0
for n, p in model.named_parameters():
msg = f"{n} | {p.shape} | {np.prod(p.shape)} | {not p.stop_gradient}"
total += np.prod(p.shape)
num_params += 1
if print_func:
print_func(msg)
if print_func:
total = total / 1024**2
print_func(f"Total parameters: {num_params}, {total:.2f}M elements.")
def gradient_norm(layer: nn.Layer):
grad_norm_dict = {}
for name, param in layer.state_dict().items():
if param.trainable:
grad = param.gradient() # return numpy.ndarray
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:
nn.utils.remove_weight_norm(layer)
except ValueError as e:
# 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

30
text_processing/speechtask/punctuation_restoration/utils/mp_tools.py
Unescape View File

@ -0,0 +1,30 @@
# Copyright (c) 2021 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 functools import wraps
from paddle import distributed as dist
__all__ = ["rank_zero_only"]
def rank_zero_only(func):
@wraps(func)
def wrapper(*args, **kwargs):
rank = dist.get_rank()
if rank != 0:
return
result = func(*args, **kwargs)
return result
return wrapper

163
text_processing/speechtask/punctuation_restoration/utils/punct_pre.py
Unescape View File

@ -0,0 +1,163 @@
# Copyright (c) 2021 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 shutil
CHINESE_PUNCTUATION_MAPPING = {
'O': '',
'，': "，",
'。': '。',
'？': '？',
}
def process_one_file_chinese(raw_path, save_path):
f = open(raw_path, 'r', encoding='utf-8')
save_file = open(save_path, 'w', encoding='utf-8')
for line in f.readlines():
line = line.strip().replace(' ', '').replace(' ', '')
for i in line:
save_file.write(i + ' ')
save_file.write('\n')
save_file.close()
def process_chinese_pure_senetence(config):
####need raw_path, raw_train_file, raw_dev_file, raw_test_file, punc_file, save_path
assert os.path.exists(
os.path.join(config["raw_path"], config[
"raw_train_file"])), "train file doesn't exist."
assert os.path.exists(
os.path.join(config["raw_path"], config[
"raw_dev_file"])), "dev file doesn't exist."
assert os.path.exists(
os.path.join(config["raw_path"], config[
"raw_test_file"])), "test file doesn't exist."
assert os.path.exists(
os.path.join(config["raw_path"], config[
"punc_file"])), "punc file doesn't exist."
train_file = os.path.join(config["raw_path"], config["raw_train_file"])
dev_file = os.path.join(config["raw_path"], config["raw_dev_file"])
test_file = os.path.join(config["raw_path"], config["raw_test_file"])
if not os.path.exists(config["save_path"]):
os.makedirs(config["save_path"])
shutil.copy(
os.path.join(config["raw_path"], config["punc_file"]),
os.path.join(config["save_path"], config["punc_file"]))
process_one_file_chinese(train_file,
os.path.join(config["save_path"], "train"))
process_one_file_chinese(dev_file, os.path.join(config["save_path"], "dev"))
process_one_file_chinese(test_file,
os.path.join(config["save_path"], "test"))
def process_one_chinese_pair(raw_path, save_path):
f = open(raw_path, 'r', encoding='utf-8')
save_file = open(save_path, 'w', encoding='utf-8')
for line in f.readlines():
if (len(line.strip().split()) == 2):
word, punc = line.strip().split()
save_file.write(word + ' ' + CHINESE_PUNCTUATION_MAPPING[punc])
if (punc == "。"):
save_file.write("\n")
else:
save_file.write(" ")
save_file.close()
def process_chinese_pair(config):
### need raw_path, raw_train_file, raw_dev_file, raw_test_file, punc_file, save_path
assert os.path.exists(
os.path.join(config["raw_path"], config[
"raw_train_file"])), "train file doesn't exist."
assert os.path.exists(
os.path.join(config["raw_path"], config[
"raw_dev_file"])), "dev file doesn't exist."
assert os.path.exists(
os.path.join(config["raw_path"], config[
"raw_test_file"])), "test file doesn't exist."
assert os.path.exists(
os.path.join(config["raw_path"], config[
"punc_file"])), "punc file doesn't exist."
train_file = os.path.join(config["raw_path"], config["raw_train_file"])
dev_file = os.path.join(config["raw_path"], config["raw_dev_file"])
test_file = os.path.join(config["raw_path"], config["raw_test_file"])
process_one_chinese_pair(train_file,
os.path.join(config["save_path"], "train"))
process_one_chinese_pair(dev_file, os.path.join(config["save_path"], "dev"))
process_one_chinese_pair(test_file,
os.path.join(config["save_path"], "test"))
shutil.copy(
os.path.join(config["raw_path"], config["punc_file"]),
os.path.join(config["save_path"], config["punc_file"]))
english_punc = [',', '.', '?']
ignore_english_punc = ['\"', '/']
def process_one_file_english(raw_path, save_path):
f = open(raw_path, 'r', encoding='utf-8')
save_file = open(save_path, 'w', encoding='utf-8')
for line in f.readlines():
for i in ignore_english_punc:
line = line.replace(i, '')
for i in english_punc:
line = line.replace(i, ' ' + i)
wordlist = line.strip().split(' ')
# print(type(wordlist))
# print(wordlist)
for i in wordlist:
save_file.write(i + ' ')
save_file.write('\n')
save_file.close()
def process_english_pure_senetence(config):
####need raw_path, raw_train_file, raw_dev_file, raw_test_file, punc_file, save_path
assert os.path.exists(
os.path.join(config["raw_path"], config[
"raw_train_file"])), "train file doesn't exist."
assert os.path.exists(
os.path.join(config["raw_path"], config[
"raw_dev_file"])), "dev file doesn't exist."
assert os.path.exists(
os.path.join(config["raw_path"], config[
"raw_test_file"])), "test file doesn't exist."
assert os.path.exists(
os.path.join(config["raw_path"], config[
"punc_file"])), "punc file doesn't exist."
train_file = os.path.join(config["raw_path"], config["raw_train_file"])
dev_file = os.path.join(config["raw_path"], config["raw_dev_file"])
test_file = os.path.join(config["raw_path"], config["raw_test_file"])
if not os.path.exists(config["save_path"]):
os.makedirs(config["save_path"])
shutil.copy(
os.path.join(config["raw_path"], config["punc_file"]),
os.path.join(config["save_path"], config["punc_file"]))
process_one_file_english(train_file,
os.path.join(config["save_path"], "train"))
process_one_file_english(dev_file, os.path.join(config["save_path"], "dev"))
process_one_file_english(test_file,
os.path.join(config["save_path"], "test"))

81
text_processing/speechtask/punctuation_restoration/utils/utility.py
Unescape View File

@ -0,0 +1,81 @@
# Copyright (c) 2021 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.
"""Contains common utility functions."""
import distutils.util
import math
import os
from typing import List
__all__ = ['print_arguments', 'add_arguments', "log_add"]
def print_arguments(args, info=None):
"""Print argparse's arguments.
Usage:
.. code-block:: python
parser = argparse.ArgumentParser()
parser.add_argument("name", default="Jonh", type=str, help="User name.")
args = parser.parse_args()
print_arguments(args)
:param args: Input argparse.Namespace for printing.
:type args: argparse.Namespace
"""
filename = ""
if info:
filename = info["__file__"]
filename = os.path.basename(filename)
print(f"----------- {filename} Configuration Arguments -----------")
for arg, value in sorted(vars(args).items()):
print("%s: %s" % (arg, value))
print("-----------------------------------------------------------")
def add_arguments(argname, type, default, help, argparser, **kwargs):
"""Add argparse's argument.
Usage:
.. code-block:: python
parser = argparse.ArgumentParser()
add_argument("name", str, "Jonh", "User name.", parser)
args = parser.parse_args()
"""
type = distutils.util.strtobool if type == bool else type
argparser.add_argument(
"--" + argname,
default=default,
type=type,
help=help + ' Default: %(default)s.',
**kwargs)
def log_add(args: List[int]) -> float:
"""Stable log add
Args:
args (List[int]): log scores
Returns:
float: sum of log scores
"""
if all(a == -float('inf') for a in args):
return -float('inf')
a_max = max(args)
lsp = math.log(sum(math.exp(a - a_max) for a in args))
return a_max + lsp
Write Preview
Loading…
Cancel
Save