Skip to content

D
DeepSpeech

PaddlePaddle / DeepSpeech
12 个月 前同步成功

通知 205
Star 8425
Fork 1598
D
DeepSpeech

体验新版 GitCode，发现更多精彩内容 >>

提交 15beee6e 编写于 作者: 小湉湉's avatar 小湉湉
浏览文件
操作

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

上级 9ccef7fa 980944da
展开全部 隐藏空白更改
内联 并排
Showing with 6823 addition and 227 deletion
.mergify.yml
浏览文件 @ 15beee6e
......@@ -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/
......
README.md
浏览文件 @ 15beee6e
......@@ -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)
......@@ -75,12 +68,15 @@ 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.
```shell
A tiny **ASR** DeepSpeech2 model training on toy set of LibriSpeech:
```bash
cd examples/tiny/s0/
# source the environment
source path.sh
......@@ -90,28 +86,50 @@ 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).
- 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
```
For **TTS**, try pretrained FastSpeech2 + Parallel WaveGAN on CSMSC:
```bash
cd examples/csmsc/tts3
# download the pretrained models and unaip them
wget https://paddlespeech.bj.bcebos.com/Parakeet/pwg_baker_ckpt_0.4.zip
unzip pwg_baker_ckpt_0.4.zip
wget https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_ckpt_0.4.zip
unzip fastspeech2_nosil_baker_ckpt_0.4.zip
# source the environment
source path.sh
# run end-to-end synthesize
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/synthesize_e2e.py \
--fastspeech2-config=fastspeech2_nosil_baker_ckpt_0.4/default.yaml \
--fastspeech2-checkpoint=fastspeech2_nosil_baker_ckpt_0.4/snapshot_iter_76000.pdz \
--fastspeech2-stat=fastspeech2_nosil_baker_ckpt_0.4/speech_stats.npy \
--pwg-config=pwg_baker_ckpt_0.4/pwg_default.yaml \
--pwg-checkpoint=pwg_baker_ckpt_0.4/pwg_snapshot_iter_400000.pdz \
--pwg-stat=pwg_baker_ckpt_0.4/pwg_stats.npy \
--text=${BIN_DIR}/../sentences.txt \
--output-dir=exp/default/test_e2e \
--inference-dir=exp/default/inference \
--device="gpu" \
--phones-dict=fastspeech2_nosil_baker_ckpt_0.4/phone_id_map.txt
```
If you want to try more functions like training and tuning, please see [ASR getting started](docs/source/asr/getting_started.md) and [TTS Basic Use](/docs/source/tts/basic_usage.md).
## 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 +143,7 @@ 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 +218,7 @@ 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 +226,41 @@ 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 +268,26 @@ 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 +295,14 @@ 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>
......
deepspeech/exps/deepspeech2/model.py
浏览文件 @ 15beee6e
......@@ -21,11 +21,6 @@ from typing import Optional
import jsonlines
import numpy as np
import paddle
from paddle import distributed as dist
from paddle import inference
from paddle.io import DataLoader
from yacs.config import CfgNode
from deepspeech.frontend.featurizer.text_featurizer import TextFeaturizer
from deepspeech.io.collator import SpeechCollator
from deepspeech.io.dataset import ManifestDataset
......@@ -44,6 +39,10 @@ from deepspeech.utils import mp_tools
from deepspeech.utils.log import Autolog
from deepspeech.utils.log import Log
from deepspeech.utils.utility import UpdateConfig
from paddle import distributed as dist
from paddle import inference
from paddle.io import DataLoader
from yacs.config import CfgNode
logger = Log(__name__).getlog()
......@@ -153,8 +152,12 @@ class DeepSpeech2Trainer(Trainer):
def setup_model(self):
config = self.config.clone()
with UpdateConfig(config):
config.model.feat_size = self.train_loader.collate_fn.feature_size
config.model.dict_size = self.train_loader.collate_fn.vocab_size
if self.train:
config.model.feat_size = self.train_loader.collate_fn.feature_size
config.model.dict_size = self.train_loader.collate_fn.vocab_size
else:
config.model.feat_size = self.test_loader.collate_fn.feature_size
config.model.dict_size = self.test_loader.collate_fn.vocab_size
if self.args.model_type == 'offline':
model = DeepSpeech2Model.from_config(config.model)
......@@ -167,6 +170,11 @@ 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 +188,76 @@ 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):
......@@ -401,6 +411,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
class DeepSpeech2ExportTester(DeepSpeech2Tester):
def __init__(self, config, args):
super().__init__(config, args)
self.apply_static = True
def compute_result_transcripts(self, audio, audio_len, vocab_list, cfg):
if self.args.model_type == "online":
......
deepspeech/exps/lm/transformer/__init__.py 0 → 100644
浏览文件 @ 15beee6e
# 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.
deepspeech/exps/lm/transformer/bin/cacu_perplexity.py 0 → 100644
浏览文件 @ 15beee6e
# 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 sys
import configargparse
def get_parser():
"""Get default arguments."""
parser = configargparse.ArgumentParser(
description="The parser for caculating the perplexity of transformer language model ",
config_file_parser_class=configargparse.YAMLConfigFileParser,
formatter_class=configargparse.ArgumentDefaultsHelpFormatter, )
parser.add_argument(
"--rnnlm", type=str, default=None, help="RNNLM model file to read")
parser.add_argument(
"--rnnlm-conf",
type=str,
default=None,
help="RNNLM model config file to read")
parser.add_argument(
"--vocab_path",
type=str,
default=None,
help="vocab path to for token2id")
parser.add_argument(
"--bpeprefix",
type=str,
default=None,
help="The path of bpeprefix for loading")
parser.add_argument(
"--text_path",
type=str,
default=None,
help="The path of text file for testing ")
parser.add_argument(
"--ngpu",
type=int,
default=0,
help="The number of gpu to use, 0 for using cpu instead")
parser.add_argument(
"--dtype",
choices=("float16", "float32", "float64"),
default="float32",
help="Float precision (only available in --api v2)", )
parser.add_argument(
"--output_dir",
type=str,
default=".",
help="The output directory to store the sentence PPL")
return parser
def main(args):
parser = get_parser()
args = parser.parse_args(args)
from deepspeech.exps.lm.transformer.lm_cacu_perplexity import run_get_perplexity
run_get_perplexity(args)
if __name__ == "__main__":
main(sys.argv[1:])
deepspeech/exps/lm/transformer/lm_cacu_perplexity.py 0 → 100644
浏览文件 @ 15beee6e
# 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.
# Caculating the PPL of LM model
import os
import numpy as np
import paddle
from paddle.io import DataLoader
from yacs.config import CfgNode
from deepspeech.io.collator import TextCollatorSpm
from deepspeech.io.dataset import TextDataset
from deepspeech.models.lm_interface import dynamic_import_lm
from deepspeech.utils.log import Log
logger = Log(__name__).getlog()
def get_config(config_path):
confs = CfgNode(new_allowed=True)
confs.merge_from_file(config_path)
return confs
def load_trained_lm(args):
lm_config = get_config(args.rnnlm_conf)
lm_model_module = lm_config.model_module
lm_class = dynamic_import_lm(lm_model_module)
lm = lm_class(**lm_config.model)
model_dict = paddle.load(args.rnnlm)
lm.set_state_dict(model_dict)
return lm, lm_config
def write_dict_into_file(ppl_dict, name):
with open(name, "w") as f:
for key in ppl_dict.keys():
f.write(key + " " + ppl_dict[key] + "\n")
return
def cacu_perplexity(
lm_model,
lm_config,
args,
log_base=None, ):
unit_type = lm_config.data.unit_type
batch_size = lm_config.decoding.batch_size
num_workers = lm_config.decoding.num_workers
text_file_path = args.text_path
total_nll = 0.0
total_ntokens = 0
ppl_dict = {}
len_dict = {}
text_dataset = TextDataset.from_file(text_file_path)
collate_fn_text = TextCollatorSpm(
unit_type=unit_type,
vocab_filepath=args.vocab_path,
spm_model_prefix=args.bpeprefix)
train_loader = DataLoader(
text_dataset,
batch_size=batch_size,
collate_fn=collate_fn_text,
num_workers=num_workers)
logger.info("start caculating PPL......")
for i, (keys, ys_input_pad, ys_output_pad,
y_lens) in enumerate(train_loader()):
ys_input_pad = paddle.to_tensor(ys_input_pad)
ys_output_pad = paddle.to_tensor(ys_output_pad)
_, unused_logp, unused_count, nll, nll_count = lm_model.forward(
ys_input_pad, ys_output_pad)
nll = nll.numpy()
nll_count = nll_count.numpy()
for key, _nll, ntoken in zip(keys, nll, nll_count):
if log_base is None:
utt_ppl = np.exp(_nll / ntoken)
else:
utt_ppl = log_base**(_nll / ntoken / np.log(log_base))
# Write PPL of each utts for debugging or analysis
ppl_dict[key] = str(utt_ppl)
len_dict[key] = str(ntoken)
total_nll += nll.sum()
total_ntokens += nll_count.sum()
logger.info("Current total nll: " + str(total_nll))
logger.info("Current total tokens: " + str(total_ntokens))
write_dict_into_file(ppl_dict, os.path.join(args.output_dir, "uttPPL"))
write_dict_into_file(len_dict, os.path.join(args.output_dir, "uttLEN"))
if log_base is None:
ppl = np.exp(total_nll / total_ntokens)
else:
ppl = log_base**(total_nll / total_ntokens / np.log(log_base))
if log_base is None:
log_base = np.e
else:
log_base = log_base
return ppl, log_base
def run_get_perplexity(args):
if args.ngpu > 1:
raise NotImplementedError("only single GPU decoding is supported")
if args.ngpu == 1:
device = "gpu:0"
else:
device = "cpu"
paddle.set_device(device)
dtype = getattr(paddle, args.dtype)
logger.info(f"Decoding device={device}, dtype={dtype}")
lm_model, lm_config = load_trained_lm(args)
lm_model.to(device=device, dtype=dtype)
lm_model.eval()
PPL, log_base = cacu_perplexity(lm_model, lm_config, args, None)
logger.info("Final PPL: " + str(PPL))
logger.info("The log base is:" + str("%.2f" % log_base))
deepspeech/exps/u2/model.py
浏览文件 @ 15beee6e
......@@ -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
......
deepspeech/exps/u2_kaldi/model.py
浏览文件 @ 15beee6e
......@@ -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
......
deepspeech/exps/u2_st/model.py
浏览文件 @ 15beee6e
......@@ -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
......
deepspeech/frontend/featurizer/text_featurizer.py
浏览文件 @ 15beee6e
......@@ -53,7 +53,7 @@ class TextFeaturizer():
self.maskctc = maskctc
if vocab_filepath:
self.vocab_dict, self._id2token, self.vocab_list, self.unk_id, self.eos_id = self._load_vocabulary_from_file(
self.vocab_dict, self._id2token, self.vocab_list, self.unk_id, self.eos_id, self.blank_id = self._load_vocabulary_from_file(
vocab_filepath, maskctc)
self.vocab_size = len(self.vocab_list)
......@@ -227,4 +227,4 @@ class TextFeaturizer():
logger.info(f"SOS id: {sos_id}")
logger.info(f"SPACE id: {space_id}")
logger.info(f"MASKCTC id: {maskctc_id}")
return token2id, id2token, vocab_list, unk_id, eos_id
return token2id, id2token, vocab_list, unk_id, eos_id, blank_id
deepspeech/io/collator.py
浏览文件 @ 15beee6e
......@@ -19,6 +19,7 @@ from yacs.config import CfgNode
from deepspeech.frontend.augmentor.augmentation import AugmentationPipeline
from deepspeech.frontend.featurizer.speech_featurizer import SpeechFeaturizer
from deepspeech.frontend.featurizer.text_featurizer import TextFeaturizer
from deepspeech.frontend.normalizer import FeatureNormalizer
from deepspeech.frontend.speech import SpeechSegment
from deepspeech.frontend.utility import IGNORE_ID
......@@ -33,7 +34,7 @@ logger = Log(__name__).getlog()
def _tokenids(text, keep_transcription_text):
# for training text is token ids
# for training text is token ids
tokens = text # token ids
if keep_transcription_text:
......@@ -45,6 +46,43 @@ def _tokenids(text, keep_transcription_text):
return tokens
class TextCollatorSpm():
def __init__(self, unit_type, vocab_filepath, spm_model_prefix):
assert (vocab_filepath is not None)
self.text_featurizer = TextFeaturizer(
unit_type=unit_type,
vocab_filepath=vocab_filepath,
spm_model_prefix=spm_model_prefix)
self.eos_id = self.text_featurizer.eos_id
self.blank_id = self.text_featurizer.blank_id
def __call__(self, batch):
"""
return type [List, np.array [B, T], np.array [B, T], np.array[B]]
"""
keys = []
texts = []
texts_input = []
texts_output = []
text_lens = []
for idx, item in enumerate(batch):
key = item.split(" ")[0].strip()
text = " ".join(item.split(" ")[1:])
keys.append(key)
token_ids = self.text_featurizer.featurize(text)
texts_input.append(
np.array([self.eos_id] + token_ids).astype(np.int64))
texts_output.append(
np.array(token_ids + [self.eos_id]).astype(np.int64))
text_lens.append(len(token_ids) + 1)
ys_input_pad = pad_list(texts_input, self.blank_id).astype(np.int64)
ys_output_pad = pad_list(texts_output, self.blank_id).astype(np.int64)
y_lens = np.array(text_lens).astype(np.int64)
return keys, ys_input_pad, ys_output_pad, y_lens
class SpeechCollatorBase():
def __init__(
self,
......
deepspeech/io/dataset.py
浏览文件 @ 15beee6e
......@@ -24,6 +24,25 @@ __all__ = ["ManifestDataset", "TransformDataset"]
logger = Log(__name__).getlog()
class TextDataset(Dataset):
@classmethod
def from_file(cls, file_path):
dataset = cls(file_path)
return dataset
def __init__(self, file_path):
self._manifest = []
with open(file_path) as f:
for line in f:
self._manifest.append(line.strip())
def __len__(self):
return len(self._manifest)
def __getitem__(self, idx):
return self._manifest[idx]
class ManifestDataset(Dataset):
@classmethod
def params(cls, config: Optional[CfgNode]=None) -> CfgNode:
......
deepspeech/models/lm/transformer.py
浏览文件 @ 15beee6e
......@@ -111,6 +111,7 @@ class TransformerLM(nn.Layer, LMInterface, BatchScorerInterface):
in perplexity: p(t)^{-n} = exp(-log p(t) / n)
"""
batch_size = x.size(0)
xm = x != 0
xlen = xm.sum(axis=1)
if self.embed_drop is not None:
......@@ -121,11 +122,13 @@ class TransformerLM(nn.Layer, LMInterface, BatchScorerInterface):
y = self.decoder(h)
loss = F.cross_entropy(
y.view(-1, y.shape[-1]), t.view(-1), reduction="none")
mask = xm.to(dtype=loss.dtype)
mask = xm.to(loss.dtype)
logp = loss * mask.view(-1)
nll = logp.view(batch_size, -1).sum(-1)
nll_count = mask.sum(-1)
logp = logp.sum()
count = mask.sum()
return logp / count, logp, count
return logp / count, logp, count, nll, nll_count
# beam search API (see ScorerInterface)
def score(self, y: paddle.Tensor, state: Any,
......
deepspeech/training/trainer.py
浏览文件 @ 15beee6e
......@@ -18,9 +18,6 @@ from contextlib import contextmanager
from pathlib import Path
import paddle
from paddle import distributed as dist
from tensorboardX import SummaryWriter
from deepspeech.training.reporter import ObsScope
from deepspeech.training.reporter import report
from deepspeech.training.timer import Timer
......@@ -31,6 +28,8 @@ from deepspeech.utils.log import Log
from deepspeech.utils.utility import all_version
from deepspeech.utils.utility import seed_all
from deepspeech.utils.utility import UpdateConfig
from paddle import distributed as dist
from tensorboardX import SummaryWriter
__all__ = ["Trainer"]
......@@ -134,6 +133,10 @@ 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 +251,7 @@ 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 +324,7 @@ class Trainer():
"""
try:
with Timer("Training Done: {}"):
self.train()
self.do_train()
except KeyboardInterrupt:
exit(-1)
finally:
......@@ -344,8 +347,12 @@ class Trainer():
try:
with Timer("Test/Decode Done: {}"):
with self.eval():
self.restore()
self.test()
if hasattr(self,
"apply_static") and self.apply_static is True:
self.test()
else:
self.restore()
self.test()
except KeyboardInterrupt:
exit(-1)
......@@ -377,6 +384,8 @@ class Trainer():
elif self.args.checkpoint_path:
output_dir = Path(
self.args.checkpoint_path).expanduser().parent.parent
elif self.args.export_path:
output_dir = Path(self.args.export_path).expanduser().parent.parent
self.output_dir = output_dir
self.output_dir.mkdir(parents=True, exist_ok=True)
......@@ -432,7 +441,7 @@ 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"])
......
examples/aishell/s0/local/test.sh
浏览文件 @ 15beee6e
......@@ -13,7 +13,7 @@ ckpt_prefix=$2
model_type=$3
# download language model
bash local/download_lm_ch.sh
bash local/download_lm_ch.sh > /dev/null 2>&1
if [ $? -ne 0 ]; then
exit 1
fi
......
examples/aishell/s0/local/test_export.sh
浏览文件 @ 15beee6e
......@@ -13,7 +13,7 @@ jit_model_export_path=$2
model_type=$3
# download language model
bash local/download_lm_ch.sh
bash local/download_lm_ch.sh > /dev/null 2>&1
if [ $? -ne 0 ]; then
exit 1
fi
......
examples/csmsc/tts2/README.md
浏览文件 @ 15beee6e
......@@ -19,7 +19,7 @@ Run the command below to
4. synthesize wavs.
- synthesize waveform from `metadata.jsonl`.
- synthesize waveform from text file.
6. inference using static model.
5. inference using static model.
```bash
./run.sh
```
......
examples/csmsc/tts3/README.md
浏览文件 @ 15beee6e
......@@ -19,6 +19,7 @@ Run the command below to
4. synthesize wavs.
- synthesize waveform from `metadata.jsonl`.
- synthesize waveform from text file.
5. inference using static model.
```bash
./run.sh
```
......@@ -189,6 +190,13 @@ optional arguments:
5. `--output-dir` is the directory to save synthesized audio files.
6. `--device is` the type of device to run synthesis, 'cpu' and 'gpu' are supported. 'gpu' is recommended for faster synthesis.
### Inference
After Synthesize, we will get static models of fastspeech2 and pwgan in `${train_output_path}/inference`.
`./local/inference.sh` calls `${BIN_DIR}/inference.py`, which provides a paddle static model inference example for fastspeech2 + pwgan synthesize.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/inference.sh ${train_output_path}
```
## Pretrained Model
Pretrained FastSpeech2 model with no silence in the edge of audios. [fastspeech2_nosil_baker_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_ckpt_0.4.zip)
......@@ -215,6 +223,7 @@ python3 ${BIN_DIR}/synthesize_e2e.py \
--pwg-stat=pwg_baker_ckpt_0.4/pwg_stats.npy \
--text=${BIN_DIR}/../sentences.txt \
--output-dir=exp/default/test_e2e \
--inference-dir=exp/default/inference \
--device="gpu" \
--phones-dict=fastspeech2_nosil_baker_ckpt_0.4/phone_id_map.txt
```
examples/csmsc/tts3/local/inference.sh 0 → 100755
浏览文件 @ 15beee6e
#!/bin/bash
train_output_path=$1
python3 ${BIN_DIR}/inference.py \
--inference-dir=${train_output_path}/inference \
--text=${BIN_DIR}/../sentences.txt \
--output-dir=${train_output_path}/pd_infer_out \
--phones-dict=dump/phone_id_map.txt
examples/csmsc/tts3/local/synthesize_e2e.sh
浏览文件 @ 15beee6e
......@@ -15,5 +15,6 @@ python3 ${BIN_DIR}/synthesize_e2e.py \
--pwg-stat=pwg_baker_ckpt_0.4/pwg_stats.npy \
--text=${BIN_DIR}/../sentences.txt \
--output-dir=${train_output_path}/test_e2e \
--inference-dir=${train_output_path}/inference \
--device="gpu" \
--phones-dict=dump/phone_id_map.txt
examples/csmsc/tts3/run.sh
浏览文件 @ 15beee6e
......@@ -35,3 +35,8 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# synthesize_e2e, vocoder is pwgan
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# inference with static model
CUDA_VISIBLE_DEVICES=${gpus} ./local/inference.sh ${train_output_path} || exit -1
fi
examples/csmsc/voc3/README.md 0 → 100644
浏览文件 @ 15beee6e
# Multi Band MelGAN with CSMSC
This example contains code used to train a [Multi Band MelGAN](https://arxiv.org/abs/2005.05106) model with [Chinese Standard Mandarin Speech Copus](https://www.data-baker.com/open_source.html).
## Dataset
### Download and Extract the datasaet
Download CSMSC from the [official website](https://www.data-baker.com/data/index/source) and extract it to `~/datasets`. Then the dataset is in directory `~/datasets/BZNSYP`.
### Get MFA results for silence trim
We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) results to cut silence in the edge of audio.
You can download from here [baker_alignment_tone.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/BZNSYP/with_tone/baker_alignment_tone.tar.gz), or train your own MFA model reference to [use_mfa example](https://github.com/PaddlePaddle/Parakeet/tree/develop/examples/use_mfa) of our repo.
## Get Started
Assume the path to the dataset is `~/datasets/BZNSYP`.
Assume the path to the MFA result of CSMSC is `./baker_alignment_tone`.
Run the command below to
1. **source path**.
2. preprocess the dataset,
3. train the model.
4. synthesize wavs.
- synthesize waveform from `metadata.jsonl`.
```bash
./run.sh
```
### Preprocess the dataset
```bash
./local/preprocess.sh ${conf_path}
```
When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
```text
dump
├── dev
│ ├── norm
│ └── raw
├── test
│ ├── norm
│ └── raw
└── train
├── norm
├── raw
└── feats_stats.npy
```
The dataset is split into 3 parts, namely `train`, `dev` and `test`, each of which contains a `norm` and `raw` subfolder. The `raw` folder contains log magnitude of mel spectrogram of each utterances, while the norm folder contains normalized spectrogram. The statistics used to normalize the spectrogram is computed from the training set, which is located in `dump/train/feats_stats.npy`.
Also there is a `metadata.jsonl` in each subfolder. It is a table-like file which contains id and paths to spectrogam of each utterance.
### Train the model
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
```
`./local/train.sh` calls `${BIN_DIR}/train.py`.
Here's the complete help message.
```text
usage: train.py [-h] [--config CONFIG] [--train-metadata TRAIN_METADATA]
[--dev-metadata DEV_METADATA] [--output-dir OUTPUT_DIR]
[--device DEVICE] [--nprocs NPROCS] [--verbose VERBOSE]
[--batch-size BATCH_SIZE] [--max-iter MAX_ITER]
[--run-benchmark RUN_BENCHMARK]
[--profiler_options PROFILER_OPTIONS]
Train a ParallelWaveGAN model.
optional arguments:
-h, --help show this help message and exit
--config CONFIG config file to overwrite default config.
--train-metadata TRAIN_METADATA
training data.
--dev-metadata DEV_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--device DEVICE device type to use.
--nprocs NPROCS number of processes.
--verbose VERBOSE verbose.
benchmark:
arguments related to benchmark.
--batch-size BATCH_SIZE
batch size.
--max-iter MAX_ITER train max steps.
--run-benchmark RUN_BENCHMARK
runing benchmark or not, if True, use the --batch-size
and --max-iter.
--profiler_options PROFILER_OPTIONS
The option of profiler, which should be in format
"key1=value1;key2=value2;key3=value3".
```
1. `--config` is a config file in yaml format to overwrite the default config, which can be found at `conf/default.yaml`.
2. `--train-metadata` and `--dev-metadata` should be the metadata file in the normalized subfolder of `train` and `dev` in the `dump` folder.
3. `--output-dir` is the directory to save the results of the experiment. Checkpoints are save in `checkpoints/` inside this directory.
4. `--device` is the type of the device to run the experiment, 'cpu' or 'gpu' are supported.
5. `--nprocs` is the number of processes to run in parallel, note that nprocs > 1 is only supported when `--device` is 'gpu'.
### Synthesize
`./local/synthesize.sh` calls `${BIN_DIR}/synthesize.py`, which can synthesize waveform from `metadata.jsonl`.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}
```
```text
usage: synthesize.py [-h] [--config CONFIG] [--checkpoint CHECKPOINT]
[--test-metadata TEST_METADATA] [--output-dir OUTPUT_DIR]
[--device DEVICE] [--verbose VERBOSE]
Synthesize with parallel wavegan.
optional arguments:
-h, --help show this help message and exit
--config CONFIG parallel wavegan config file.
--checkpoint CHECKPOINT
snapshot to load.
--test-metadata TEST_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--device DEVICE device to run.
--verbose VERBOSE verbose.
```
1. `--config` parallel wavegan config file. You should use the same config with which the model is trained.
2. `--checkpoint` is the checkpoint to load. Pick one of the checkpoints from `checkpoints` inside the training output directory.
3. `--test-metadata` is the metadata of the test dataset. Use the `metadata.jsonl` in the `dev/norm` subfolder from the processed directory.
4. `--output-dir` is the directory to save the synthesized audio files.
5. `--device` is the type of device to run synthesis, 'cpu' and 'gpu' are supported.
## Pretrained Models
examples/csmsc/voc3/conf/default.yaml 0 → 100644
浏览文件 @ 15beee6e
# This is the hyperparameter configuration file for MelGAN.
# Please make sure this is adjusted for the CSMSC dataset. If you want to
# apply to the other dataset, you might need to carefully change some parameters.
# This configuration requires ~ 8GB memory and will finish within 7 days on Titan V.
# This configuration is based on full-band MelGAN but the hop size and sampling
# rate is different from the paper (16kHz vs 24kHz). The number of iteraions
# is not shown in the paper so currently we train 1M iterations (not sure enough
# to converge). The optimizer setting is based on @dathudeptrai advice.
# https://github.com/kan-bayashi/ParallelWaveGAN/issues/143#issuecomment-632539906
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 24000 # Sampling rate.
n_fft: 2048 # FFT size. (in samples)
n_shift: 300 # Hop size. (in samples)
win_length: 1200 # Window length. (in samples)
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
n_mels: 80 # Number of mel basis.
fmin: 80 # Minimum freq in mel basis calculation. (Hz)
fmax: 7600 # Maximum frequency in mel basis calculation. (Hz)
###########################################################
# GENERATOR NETWORK ARCHITECTURE SETTING #
###########################################################
generator_params:
in_channels: 80 # Number of input channels.
out_channels: 4 # Number of output channels.
kernel_size: 7 # Kernel size of initial and final conv layers.
channels: 384 # Initial number of channels for conv layers.
upsample_scales: [5, 5, 3] # List of Upsampling scales.
stack_kernel_size: 3 # Kernel size of dilated conv layers in residual stack.
stacks: 4 # Number of stacks in a single residual stack module.
use_weight_norm: True # Whether to use weight normalization.
use_causal_conv: False # Whether to use causal convolution.
use_final_nonlinear_activation: True
###########################################################
# DISCRIMINATOR NETWORK ARCHITECTURE SETTING #
###########################################################
discriminator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
scales: 3 # Number of multi-scales.
downsample_pooling: "AvgPool1D" # Pooling type for the input downsampling.
downsample_pooling_params: # Parameters of the above pooling function.
kernel_size: 4
stride: 2
padding: 1
exclusive: True
kernel_sizes: [5, 3] # List of kernel size.
channels: 16 # Number of channels of the initial conv layer.
max_downsample_channels: 512 # Maximum number of channels of downsampling layers.
downsample_scales: [4, 4, 4] # List of downsampling scales.
nonlinear_activation: "LeakyReLU" # Nonlinear activation function.
nonlinear_activation_params: # Parameters of nonlinear activation function.
negative_slope: 0.2
use_weight_norm: True # Whether to use weight norm.
###########################################################
# STFT LOSS SETTING #
###########################################################
use_stft_loss: true
stft_loss_params:
fft_sizes: [1024, 2048, 512] # List of FFT size for STFT-based loss.
hop_sizes: [120, 240, 50] # List of hop size for STFT-based loss
win_lengths: [600, 1200, 240] # List of window length for STFT-based loss.
window: "hann" # Window function for STFT-based loss
use_subband_stft_loss: true
subband_stft_loss_params:
fft_sizes: [384, 683, 171] # List of FFT size for STFT-based loss.
hop_sizes: [30, 60, 10] # List of hop size for STFT-based loss
win_lengths: [150, 300, 60] # List of window length for STFT-based loss.
window: "hann" # Window function for STFT-based loss
###########################################################
# ADVERSARIAL LOSS SETTING #
###########################################################
use_feat_match_loss: false # Whether to use feature matching loss.
lambda_adv: 2.5 # Loss balancing coefficient for adversarial loss.
###########################################################
# DATA LOADER SETTING #
###########################################################
batch_size: 64 # Batch size.
batch_max_steps: 16200 # Length of each audio in batch. Make sure dividable by hop_size.
num_workers: 2 # Number of workers in DataLoader.
###########################################################
# OPTIMIZER & SCHEDULER SETTING #
###########################################################
generator_optimizer_params:
epsilon: 1.0e-7 # Generator's epsilon.
weight_decay: 0.0 # Generator's weight decay coefficient.
generator_grad_norm: -1 # Generator's gradient norm.
generator_scheduler_params:
learning_rate: 1.0e-3 # Generator's learning rate.
gamma: 0.5 # Generator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 100000
- 200000
- 300000
- 400000
- 500000
- 600000
discriminator_optimizer_params:
epsilon: 1.0e-7 # Discriminator's epsilon.
weight_decay: 0.0 # Discriminator's weight decay coefficient.
discriminator_grad_norm: -1 # Discriminator's gradient norm.
discriminator_scheduler_params:
learning_rate: 1.0e-3 # Discriminator's learning rate.
gamma: 0.5 # Discriminator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 100000
- 200000
- 300000
- 400000
- 500000
- 600000
###########################################################
# INTERVAL SETTING #
###########################################################
discriminator_train_start_steps: 200000 # Number of steps to start to train discriminator.
train_max_steps: 1000000 # Number of training steps.
save_interval_steps: 5000 # Interval steps to save checkpoint.
eval_interval_steps: 1000 # Interval steps to evaluate the network.
###########################################################
# OTHER SETTING #
###########################################################
num_snapshots: 10 # max number of snapshots to keep while training
seed: 42 # random seed for paddle, random, and np.random
\ No newline at end of file
examples/csmsc/voc3/local/preprocess.sh 0 → 100755
浏览文件 @ 15beee6e
#!/bin/bash
stage=0
stop_stage=100
config_path=$1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# get durations from MFA's result
echo "Generate durations.txt from MFA results ..."
python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
--inputdir=./baker_alignment_tone \
--output=durations.txt \
--config=${config_path}
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# extract features
echo "Extract features ..."
python3 ${BIN_DIR}/../preprocess.py \
--rootdir=~/datasets/BZNSYP/ \
--dataset=baker \
--dumpdir=dump \
--dur-file=durations.txt \
--config=${config_path} \
--cut-sil=True \
--num-cpu=20
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# get features' stats(mean and std)
echo "Get features' stats ..."
python3 ${MAIN_ROOT}/utils/compute_statistics.py \
--metadata=dump/train/raw/metadata.jsonl \
--field-name="feats"
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# normalize, dev and test should use train's stats
echo "Normalize ..."
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/train/raw/metadata.jsonl \
--dumpdir=dump/train/norm \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/dev/raw/metadata.jsonl \
--dumpdir=dump/dev/norm \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/test/raw/metadata.jsonl \
--dumpdir=dump/test/norm \
--stats=dump/train/feats_stats.npy
fi
examples/csmsc/voc3/local/synthesize.sh 0 → 100755
浏览文件 @ 15beee6e
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/synthesize.py \
--config=${config_path} \
--checkpoint=${train_output_path}/checkpoints/${ckpt_name} \
--test-metadata=dump/test/norm/metadata.jsonl \
--output-dir=${train_output_path}/test
examples/csmsc/voc3/local/train.sh 0 → 100755
浏览文件 @ 15beee6e
#!/bin/bash
config_path=$1
train_output_path=$2
FLAGS_cudnn_exhaustive_search=true \
FLAGS_conv_workspace_size_limit=4000 \
python ${BIN_DIR}/train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=${config_path} \
--output-dir=${train_output_path} \
--nprocs=1
examples/csmsc/voc3/path.sh 0 → 100755
浏览文件 @ 15beee6e
#!/bin/bash
export MAIN_ROOT=`realpath ${PWD}/../../../`
export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
export LC_ALL=C
export PYTHONDONTWRITEBYTECODE=1
# Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
MODEL=multi_band_melgan
export BIN_DIR=${MAIN_ROOT}/parakeet/exps/gan_vocoder/${MODEL}
\ No newline at end of file
examples/csmsc/voc3/run.sh 0 → 100755
浏览文件 @ 15beee6e
#!/bin/bash
set -e
source path.sh
gpus=0,1
stage=0
stop_stage=100
conf_path=conf/default.yaml
train_output_path=exp/default
ckpt_name=snapshot_iter_50000.pdz
# with the following command, you can choice the stage range you want to run
# such as `./run.sh --stage 0 --stop-stage 0`
# this can not be mixed use with `$1`, `$2` ...
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# prepare data
./local/preprocess.sh ${conf_path} || exit -1
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# train model, all `ckpt` under `train_output_path/checkpoints/` dir
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# synthesize
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi
examples/librispeech/s0/local/test.sh
浏览文件 @ 15beee6e
......@@ -13,7 +13,7 @@ ckpt_prefix=$2
model_type=$3
# download language model
bash local/download_lm_en.sh
bash local/download_lm_en.sh > /dev/null 2>&1
if [ $? -ne 0 ]; then
exit 1
fi
......
examples/librispeech/s2/conf/lm/transformer.yaml
浏览文件 @ 15beee6e
model_module: transformer
data:
unit_type: spm
model:
n_vocab: 5002
pos_enc: null
......@@ -11,3 +15,7 @@ model:
emb_dropout_rate: 0.0
att_dropout_rate: 0.0
tie_weights: False
decoding:
batch_size: 30
num_workers: 2
examples/librispeech/s2/local/cacu_perplexity.sh 0 → 100755
浏览文件 @ 15beee6e
#!/bin/bash
set -e
stage=-1
stop_stage=100
expdir=exp
datadir=data
ngpu=0
# lm params
rnnlm_config_path=conf/lm/transformer.yaml
lmexpdir=exp/lm/transformer
lang_model=transformerLM.pdparams
#data path
test_set=${datadir}/test_clean/text
test_set_lower=${datadir}/test_clean/text_lower
train_set=train_960
# bpemode (unigram or bpe)
nbpe=5000
bpemode=unigram
bpeprefix=${datadir}/lang_char/${train_set}_${bpemode}${nbpe}
bpemodel=${bpeprefix}.model
vocabfile=${bpeprefix}_units.txt
vocabfile_lower=${bpeprefix}_units_lower.txt
output_dir=${expdir}/lm/transformer/perplexity
mkdir -p ${output_dir}
# Transform the data upper case to lower
if [ -f ${vocabfile} ]; then
tr A-Z a-z < ${vocabfile} > ${vocabfile_lower}
fi
if [ -f ${test_set} ]; then
tr A-Z a-z < ${test_set} > ${test_set_lower}
fi
python ${LM_BIN_DIR}/cacu_perplexity.py \
--rnnlm ${lmexpdir}/${lang_model} \
--rnnlm-conf ${rnnlm_config_path} \
--vocab_path ${vocabfile_lower} \
--bpeprefix ${bpeprefix} \
--text_path ${test_set_lower} \
--output_dir ${output_dir} \
--ngpu ${ngpu}
examples/librispeech/s2/run.sh
浏览文件 @ 15beee6e
......@@ -51,3 +51,7 @@ if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
# export ckpt avg_n
CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
fi
if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
CUDA_VISIBLE_DEVICES= ./local/cacu_perplexity.sh || exit -1
fi
examples/other/1xt2x/README.md
浏览文件 @ 15beee6e
......@@ -2,10 +2,18 @@
Convert Deepspeech 1.8 released model to 2.x.
## Model
## Model source directory
* Deepspeech2x
## Exp
* baidu_en8k
## Expriment directory
* aishell
* librispeech
* baidu_en8k
# The released model
Acoustic Model | Training Data | Hours of Speech | Token-based | CER | WER
:-------------:| :------------:| :---------------: | :---------: | :---: | :----:
Ds2 Offline Aishell 1xt2x model| Aishell Dataset | 151 h | Char-based | 0.080447 |
Ds2 Offline Librispeech 1xt2x model | Librispeech Dataset | 960 h | Word-based | | 0.068548
Ds2 Offline Baidu en8k 1x2x model | Baidu Internal English Dataset | 8628 h |Word-based | | 0.054112
examples/other/1xt2x/aishell/local/test.sh
浏览文件 @ 15beee6e
......@@ -13,7 +13,7 @@ ckpt_prefix=$2
model_type=$3
# download language model
bash local/download_lm_ch.sh
bash local/download_lm_ch.sh > /dev/null 2>&1
if [ $? -ne 0 ]; then
exit 1
fi
......
examples/other/1xt2x/baidu_en8k/local/test.sh
浏览文件 @ 15beee6e
......@@ -13,7 +13,7 @@ ckpt_prefix=$2
model_type=$3
# download language model
bash local/download_lm_en.sh
bash local/download_lm_en.sh > /dev/null 2>&1
if [ $? -ne 0 ]; then
exit 1
fi
......
examples/other/1xt2x/librispeech/local/test.sh
浏览文件 @ 15beee6e
......@@ -13,7 +13,7 @@ ckpt_prefix=$2
model_type=$3
# download language model
bash local/download_lm_en.sh
bash local/download_lm_en.sh > /dev/null 2>&1
if [ $? -ne 0 ]; then
exit 1
fi
......
examples/tiny/s0/conf/deepspeech2.yaml
浏览文件 @ 15beee6e
......@@ -45,7 +45,7 @@ model:
ctc_grad_norm_type: null
training:
n_epoch: 10
n_epoch: 5
accum_grad: 1
lr: 1e-5
lr_decay: 0.8
......
examples/tiny/s0/conf/deepspeech2_online.yaml
浏览文件 @ 15beee6e
......@@ -47,7 +47,7 @@ model:
ctc_grad_norm_type: null
training:
n_epoch: 10
n_epoch: 5
accum_grad: 1
lr: 1e-5
lr_decay: 1.0
......
examples/tiny/s0/local/test.sh
浏览文件 @ 15beee6e
......@@ -13,7 +13,7 @@ ckpt_prefix=$2
model_type=$3
# download language model
bash local/download_lm_en.sh
bash local/download_lm_en.sh > /dev/null 2>&1
if [ $? -ne 0 ]; then
exit 1
fi
......
examples/tiny/s1/conf/chunk_confermer.yaml
浏览文件 @ 15beee6e
......@@ -83,7 +83,7 @@ model:
training:
n_epoch: 20
n_epoch: 5
accum_grad: 1
global_grad_clip: 5.0
optim: adam
......
examples/tiny/s1/conf/chunk_transformer.yaml
浏览文件 @ 15beee6e
......@@ -76,7 +76,7 @@ model:
training:
n_epoch: 20
n_epoch: 5
accum_grad: 1
global_grad_clip: 5.0
optim: adam
......
examples/tiny/s1/conf/conformer.yaml
浏览文件 @ 15beee6e
......@@ -79,7 +79,7 @@ model:
training:
n_epoch: 20
n_epoch: 5
accum_grad: 4
global_grad_clip: 5.0
optim: adam
......
examples/tiny/s1/conf/transformer.yaml
浏览文件 @ 15beee6e
......@@ -73,7 +73,7 @@ model:
training:
n_epoch: 21
n_epoch: 5
accum_grad: 1
global_grad_clip: 5.0
optim: adam
......
parakeet/data/batch.py
浏览文件 @ 15beee6e
......@@ -53,8 +53,8 @@ def batch_text_id(minibatch, pad_id=0, dtype=np.int64):
peek_example = minibatch[0]
assert len(peek_example.shape) == 1, "text example is an 1D tensor"
lengths = [example.shape[0] for example in minibatch
] # assume (channel, n_samples) or (n_samples, )
lengths = [example.shape[0] for example in
minibatch] # assume (channel, n_samples) or (n_samples, )
max_len = np.max(lengths)
batch = []
......
parakeet/datasets/vocoder_batch_fn.py
浏览文件 @ 15beee6e
......@@ -107,8 +107,13 @@ class Clip(object):
features, this process will be needed.
"""
if len(x) < c.shape[1] * self.hop_size:
x = np.pad(x, (0, c.shape[1] * self.hop_size - len(x)), mode="edge")
if len(x) < c.shape[0] * self.hop_size:
x = np.pad(x, (0, c.shape[0] * self.hop_size - len(x)), mode="edge")
elif len(x) > c.shape[0] * self.hop_size:
print(
f"wave length: ({len(x)}), mel length: ({c.shape[0]}), hop size: ({self.hop_size })"
)
x = x[:c.shape[1] * self.hop_size]
# check the legnth is valid
assert len(x) == c.shape[
......
parakeet/exps/fastspeech2/inference.py 0 → 100644
浏览文件 @ 15beee6e
# 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 os
from pathlib import Path
import soundfile as sf
from paddle import inference
from parakeet.frontend.zh_frontend import Frontend
def main():
parser = argparse.ArgumentParser(
description="Paddle Infernce with speedyspeech & parallel wavegan.")
parser.add_argument(
"--inference-dir", type=str, help="dir to save inference models")
parser.add_argument(
"--text",
type=str,
help="text to synthesize, a 'utt_id sentence' pair per line")
parser.add_argument("--output-dir", type=str, help="output dir")
parser.add_argument(
"--enable-auto-log", action="store_true", help="use auto log")
parser.add_argument(
"--phones-dict",
type=str,
default="phones.txt",
help="phone vocabulary file.")
args, _ = parser.parse_known_args()
frontend = Frontend(phone_vocab_path=args.phones_dict)
print("frontend done!")
fastspeech2_config = inference.Config(
str(Path(args.inference_dir) / "fastspeech2.pdmodel"),
str(Path(args.inference_dir) / "fastspeech2.pdiparams"))
fastspeech2_config.enable_use_gpu(50, 0)
# This line must be commented, if not, it will OOM
# fastspeech2_config.enable_memory_optim()
fastspeech2_predictor = inference.create_predictor(fastspeech2_config)
pwg_config = inference.Config(
str(Path(args.inference_dir) / "pwg.pdmodel"),
str(Path(args.inference_dir) / "pwg.pdiparams"))
pwg_config.enable_use_gpu(100, 0)
pwg_config.enable_memory_optim()
pwg_predictor = inference.create_predictor(pwg_config)
if args.enable_auto_log:
import auto_log
os.makedirs("output", exist_ok=True)
pid = os.getpid()
logger = auto_log.AutoLogger(
model_name="fastspeech2",
model_precision='float32',
batch_size=1,
data_shape="dynamic",
save_path="./output/auto_log.log",
inference_config=fastspeech2_config,
pids=pid,
process_name=None,
gpu_ids=0,
time_keys=['preprocess_time', 'inference_time', 'postprocess_time'],
warmup=0)
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
sentences = []
with open(args.text, 'rt') as f:
for line in f:
utt_id, sentence = line.strip().split()
sentences.append((utt_id, sentence))
for utt_id, sentence in sentences:
if args.enable_auto_log:
logger.times.start()
input_ids = frontend.get_input_ids(sentence, merge_sentences=True)
phone_ids = input_ids["phone_ids"]
phones = phone_ids[0].numpy()
if args.enable_auto_log:
logger.times.stamp()
input_names = fastspeech2_predictor.get_input_names()
phones_handle = fastspeech2_predictor.get_input_handle(input_names[0])
phones_handle.reshape(phones.shape)
phones_handle.copy_from_cpu(phones)
fastspeech2_predictor.run()
output_names = fastspeech2_predictor.get_output_names()
output_handle = fastspeech2_predictor.get_output_handle(output_names[0])
output_data = output_handle.copy_to_cpu()
input_names = pwg_predictor.get_input_names()
mel_handle = pwg_predictor.get_input_handle(input_names[0])
mel_handle.reshape(output_data.shape)
mel_handle.copy_from_cpu(output_data)
pwg_predictor.run()
output_names = pwg_predictor.get_output_names()
output_handle = pwg_predictor.get_output_handle(output_names[0])
wav = output_data = output_handle.copy_to_cpu()
if args.enable_auto_log:
logger.times.stamp()
sf.write(output_dir / (utt_id + ".wav"), wav, samplerate=24000)
if args.enable_auto_log:
logger.times.end(stamp=True)
print(f"{utt_id} done!")
if args.enable_auto_log:
logger.report()
if __name__ == "__main__":
main()
parakeet/exps/fastspeech2/synthesize_e2e.py
浏览文件 @ 15beee6e
......@@ -13,12 +13,15 @@
# limitations under the License.
import argparse
import logging
import os
from pathlib import Path
import numpy as np
import paddle
import soundfile as sf
import yaml
from paddle import jit
from paddle.static import InputSpec
from yacs.config import CfgNode
from parakeet.frontend.zh_frontend import Frontend
......@@ -74,7 +77,21 @@ def evaluate(args, fastspeech2_config, pwg_config):
pwg_normalizer = ZScore(mu, std)
fastspeech2_inference = FastSpeech2Inference(fastspeech2_normalizer, model)
fastspeech2_inference.eval()
fastspeech2_inference = jit.to_static(
fastspeech2_inference, input_spec=[InputSpec([-1], dtype=paddle.int64)])
paddle.jit.save(fastspeech2_inference,
os.path.join(args.inference_dir, "fastspeech2"))
fastspeech2_inference = paddle.jit.load(
os.path.join(args.inference_dir, "fastspeech2"))
pwg_inference = PWGInference(pwg_normalizer, vocoder)
pwg_inference.eval()
pwg_inference = jit.to_static(
pwg_inference, input_spec=[
InputSpec([-1, 80], dtype=paddle.float32),
])
paddle.jit.save(pwg_inference, os.path.join(args.inference_dir, "pwg"))
pwg_inference = paddle.jit.load(os.path.join(args.inference_dir, "pwg"))
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
......@@ -135,6 +152,8 @@ def main():
type=str,
help="text to synthesize, a 'utt_id sentence' pair per line.")
parser.add_argument("--output-dir", type=str, help="output dir.")
parser.add_argument(
"--inference-dir", type=str, help="dir to save inference models")
parser.add_argument(
"--device", type=str, default="gpu", help="device type to use.")
parser.add_argument("--verbose", type=int, default=1, help="verbose.")
......
parakeet/exps/fastspeech2/train.py
浏览文件 @ 15beee6e
......@@ -25,7 +25,6 @@ from paddle import DataParallel
from paddle import distributed as dist
from paddle.io import DataLoader
from paddle.io import DistributedBatchSampler
from visualdl import LogWriter
from yacs.config import CfgNode
from parakeet.datasets.am_batch_fn import fastspeech2_multi_spk_batch_fn
......@@ -160,8 +159,7 @@ def train_sp(args, config):
if dist.get_rank() == 0:
trainer.extend(evaluator, trigger=(1, "epoch"))
writer = LogWriter(str(output_dir))
trainer.extend(VisualDL(writer), trigger=(1, "iteration"))
trainer.extend(VisualDL(output_dir), trigger=(1, "iteration"))
trainer.extend(
Snapshot(max_size=config.num_snapshots), trigger=(1, 'epoch'))
# print(trainer.extensions)
......
parakeet/exps/gan_vocoder/multi_band_melgan/__init__.py 0 → 100644
浏览文件 @ 15beee6e
# 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.
parakeet/exps/gan_vocoder/multi_band_melgan/synthesize.py 0 → 100644
浏览文件 @ 15beee6e
# 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 os
from pathlib import Path
import jsonlines
import numpy as np
import paddle
import soundfile as sf
import yaml
from paddle import distributed as dist
from timer import timer
from yacs.config import CfgNode
from parakeet.datasets.data_table import DataTable
from parakeet.models.melgan import MelGANGenerator
def main():
parser = argparse.ArgumentParser(
description="Synthesize with parallel wavegan.")
parser.add_argument(
"--config", type=str, help="parallel wavegan config file.")
parser.add_argument("--checkpoint", type=str, help="snapshot to load.")
parser.add_argument("--test-metadata", type=str, help="dev data.")
parser.add_argument("--output-dir", type=str, help="output dir.")
parser.add_argument(
"--device", type=str, default="gpu", help="device to run.")
parser.add_argument("--verbose", type=int, default=1, help="verbose.")
args = parser.parse_args()
with open(args.config) as f:
config = CfgNode(yaml.safe_load(f))
print("========Args========")
print(yaml.safe_dump(vars(args)))
print("========Config========")
print(config)
print(
f"master see the word size: {dist.get_world_size()}, from pid: {os.getpid()}"
)
paddle.set_device(args.device)
generator = MelGANGenerator(**config["generator_params"])
state_dict = paddle.load(args.checkpoint)
generator.set_state_dict(state_dict["generator_params"])
generator.remove_weight_norm()
generator.eval()
with jsonlines.open(args.test_metadata, 'r') as reader:
metadata = list(reader)
test_dataset = DataTable(
metadata,
fields=['utt_id', 'feats'],
converters={
'utt_id': None,
'feats': np.load,
})
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
N = 0
T = 0
for example in test_dataset:
utt_id = example['utt_id']
mel = example['feats']
mel = paddle.to_tensor(mel) # (T, C)
with timer() as t:
with paddle.no_grad():
wav = generator.inference(c=mel)
wav = wav.numpy()
N += wav.size
T += t.elapse
speed = wav.size / t.elapse
rtf = config.fs / speed
print(
f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
)
sf.write(str(output_dir / (utt_id + ".wav")), wav, samplerate=config.fs)
print(f"generation speed: {N / T}Hz, RTF: {config.fs / (N / T) }")
if __name__ == "__main__":
main()
parakeet/exps/gan_vocoder/multi_band_melgan/train.py 0 → 100644
浏览文件 @ 15beee6e
# 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 logging
import os
import shutil
from pathlib import Path
import jsonlines
import numpy as np
import paddle
import yaml
from paddle import DataParallel
from paddle import distributed as dist
from paddle import nn
from paddle.io import DataLoader
from paddle.io import DistributedBatchSampler
from paddle.optimizer import Adam
from paddle.optimizer.lr import MultiStepDecay
from yacs.config import CfgNode
from parakeet.datasets.data_table import DataTable
from parakeet.datasets.vocoder_batch_fn import Clip
from parakeet.models.melgan import MBMelGANEvaluator
from parakeet.models.melgan import MBMelGANUpdater
from parakeet.models.melgan import MelGANGenerator
from parakeet.models.melgan import MelGANMultiScaleDiscriminator
from parakeet.modules.adversarial_loss import DiscriminatorAdversarialLoss
from parakeet.modules.adversarial_loss import GeneratorAdversarialLoss
from parakeet.modules.pqmf import PQMF
from parakeet.modules.stft_loss import MultiResolutionSTFTLoss
from parakeet.training.extensions.snapshot import Snapshot
from parakeet.training.extensions.visualizer import VisualDL
from parakeet.training.seeding import seed_everything
from parakeet.training.trainer import Trainer
def train_sp(args, config):
# decides device type and whether to run in parallel
# setup running environment correctly
world_size = paddle.distributed.get_world_size()
if not paddle.is_compiled_with_cuda():
paddle.set_device("cpu")
else:
paddle.set_device("gpu")
if world_size > 1:
paddle.distributed.init_parallel_env()
# set the random seed, it is a must for multiprocess training
seed_everything(config.seed)
print(
f"rank: {dist.get_rank()}, pid: {os.getpid()}, parent_pid: {os.getppid()}",
)
# dataloader has been too verbose
logging.getLogger("DataLoader").disabled = True
# construct dataset for training and validation
with jsonlines.open(args.train_metadata, 'r') as reader:
train_metadata = list(reader)
train_dataset = DataTable(
data=train_metadata,
fields=["wave", "feats"],
converters={
"wave": np.load,
"feats": np.load,
}, )
with jsonlines.open(args.dev_metadata, 'r') as reader:
dev_metadata = list(reader)
dev_dataset = DataTable(
data=dev_metadata,
fields=["wave", "feats"],
converters={
"wave": np.load,
"feats": np.load,
}, )
# collate function and dataloader
train_sampler = DistributedBatchSampler(
train_dataset,
batch_size=config.batch_size,
shuffle=True,
drop_last=True)
dev_sampler = DistributedBatchSampler(
dev_dataset,
batch_size=config.batch_size,
shuffle=False,
drop_last=False)
print("samplers done!")
if "aux_context_window" in config.generator_params:
aux_context_window = config.generator_params.aux_context_window
else:
aux_context_window = 0
train_batch_fn = Clip(
batch_max_steps=config.batch_max_steps,
hop_size=config.n_shift,
aux_context_window=aux_context_window)
train_dataloader = DataLoader(
train_dataset,
batch_sampler=train_sampler,
collate_fn=train_batch_fn,
num_workers=config.num_workers)
dev_dataloader = DataLoader(
dev_dataset,
batch_sampler=dev_sampler,
collate_fn=train_batch_fn,
num_workers=config.num_workers)
print("dataloaders done!")
generator = MelGANGenerator(**config["generator_params"])
discriminator = MelGANMultiScaleDiscriminator(
**config["discriminator_params"])
if world_size > 1:
generator = DataParallel(generator)
discriminator = DataParallel(discriminator)
print("models done!")
criterion_stft = MultiResolutionSTFTLoss(**config["stft_loss_params"])
criterion_sub_stft = MultiResolutionSTFTLoss(
**config["subband_stft_loss_params"])
criterion_gen_adv = GeneratorAdversarialLoss()
criterion_dis_adv = DiscriminatorAdversarialLoss()
# define special module for subband processing
criterion_pqmf = PQMF(subbands=config["generator_params"]["out_channels"])
print("criterions done!")
lr_schedule_g = MultiStepDecay(**config["generator_scheduler_params"])
# Compared to multi_band_melgan.v1 config, Adam optimizer without gradient norm is used
generator_grad_norm = config["generator_grad_norm"]
gradient_clip_g = nn.ClipGradByGlobalNorm(
generator_grad_norm) if generator_grad_norm > 0 else None
print("gradient_clip_g:", gradient_clip_g)
optimizer_g = Adam(
learning_rate=lr_schedule_g,
grad_clip=gradient_clip_g,
parameters=generator.parameters(),
**config["generator_optimizer_params"])
lr_schedule_d = MultiStepDecay(**config["discriminator_scheduler_params"])
discriminator_grad_norm = config["discriminator_grad_norm"]
gradient_clip_d = nn.ClipGradByGlobalNorm(
discriminator_grad_norm) if discriminator_grad_norm > 0 else None
print("gradient_clip_d:", gradient_clip_d)
optimizer_d = Adam(
learning_rate=lr_schedule_d,
grad_clip=gradient_clip_d,
parameters=discriminator.parameters(),
**config["discriminator_optimizer_params"])
print("optimizers done!")
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
if dist.get_rank() == 0:
config_name = args.config.split("/")[-1]
# copy conf to output_dir
shutil.copyfile(args.config, output_dir / config_name)
updater = MBMelGANUpdater(
models={
"generator": generator,
"discriminator": discriminator,
},
optimizers={
"generator": optimizer_g,
"discriminator": optimizer_d,
},
criterions={
"stft": criterion_stft,
"sub_stft": criterion_sub_stft,
"gen_adv": criterion_gen_adv,
"dis_adv": criterion_dis_adv,
"pqmf": criterion_pqmf
},
schedulers={
"generator": lr_schedule_g,
"discriminator": lr_schedule_d,
},
dataloader=train_dataloader,
discriminator_train_start_steps=config.discriminator_train_start_steps,
lambda_adv=config.lambda_adv,
output_dir=output_dir)
evaluator = MBMelGANEvaluator(
models={
"generator": generator,
"discriminator": discriminator,
},
criterions={
"stft": criterion_stft,
"sub_stft": criterion_sub_stft,
"gen_adv": criterion_gen_adv,
"dis_adv": criterion_dis_adv,
"pqmf": criterion_pqmf
},
dataloader=dev_dataloader,
lambda_adv=config.lambda_adv,
output_dir=output_dir)
trainer = Trainer(
updater,
stop_trigger=(config.train_max_steps, "iteration"),
out=output_dir)
if dist.get_rank() == 0:
trainer.extend(
evaluator, trigger=(config.eval_interval_steps, 'iteration'))
trainer.extend(VisualDL(output_dir), trigger=(1, 'iteration'))
trainer.extend(
Snapshot(max_size=config.num_snapshots),
trigger=(config.save_interval_steps, 'iteration'))
print("Trainer Done!")
trainer.run()
def main():
# parse args and config and redirect to train_sp
parser = argparse.ArgumentParser(
description="Train a Multi-Band MelGAN model.")
parser.add_argument(
"--config", type=str, help="config file to overwrite default config.")
parser.add_argument("--train-metadata", type=str, help="training data.")
parser.add_argument("--dev-metadata", type=str, help="dev data.")
parser.add_argument("--output-dir", type=str, help="output dir.")
parser.add_argument(
"--device", type=str, default="gpu", help="device type to use.")
parser.add_argument(
"--nprocs", type=int, default=1, help="number of processes.")
parser.add_argument("--verbose", type=int, default=1, help="verbose.")
args = parser.parse_args()
if args.device == "cpu" and args.nprocs > 1:
raise RuntimeError("Multiprocess training on CPU is not supported.")
with open(args.config, 'rt') as f:
config = CfgNode(yaml.safe_load(f))
print("========Args========")
print(yaml.safe_dump(vars(args)))
print("========Config========")
print(config)
print(
f"master see the word size: {dist.get_world_size()}, from pid: {os.getpid()}"
)
# dispatch
if args.nprocs > 1:
dist.spawn(train_sp, (args, config), nprocs=args.nprocs)
else:
train_sp(args, config)
if __name__ == "__main__":
main()
parakeet/exps/gan_vocoder/parallelwave_gan/synthesize.py
浏览文件 @ 15beee6e
......@@ -86,8 +86,9 @@ def main():
N += wav.size
T += t.elapse
speed = wav.size / t.elapse
rtf = config.fs / speed
print(
f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {config.fs / speed}."
f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
)
sf.write(str(output_dir / (utt_id + ".wav")), wav, samplerate=config.fs)
print(f"generation speed: {N / T}Hz, RTF: {config.fs / (N / T) }")
......
parakeet/exps/gan_vocoder/parallelwave_gan/train.py
浏览文件 @ 15beee6e
......@@ -28,7 +28,6 @@ from paddle.io import DataLoader
from paddle.io import DistributedBatchSampler
from paddle.optimizer import Adam # No RAdaom
from paddle.optimizer.lr import StepDecay
from visualdl import LogWriter
from yacs.config import CfgNode
from parakeet.datasets.data_table import DataTable
......@@ -193,8 +192,7 @@ def train_sp(args, config):
if dist.get_rank() == 0:
trainer.extend(
evaluator, trigger=(config.eval_interval_steps, 'iteration'))
writer = LogWriter(str(trainer.out))
trainer.extend(VisualDL(writer), trigger=(1, 'iteration'))
trainer.extend(VisualDL(output_dir), trigger=(1, 'iteration'))
trainer.extend(
Snapshot(max_size=config.num_snapshots),
trigger=(config.save_interval_steps, 'iteration'))
......
parakeet/exps/speedyspeech/inference.py
浏览文件 @ 15beee6e
......@@ -96,8 +96,8 @@ def main():
input_ids = frontend.get_input_ids(
sentence, merge_sentences=True, get_tone_ids=True)
phone_ids = input_ids["phone_ids"]
tone_ids = input_ids["tone_ids"]
phone_ids = input_ids["phone_ids"].numpy()
tone_ids = input_ids["tone_ids"].numpy()
phones = phone_ids[0]
tones = tone_ids[0]
......
parakeet/exps/speedyspeech/train.py
浏览文件 @ 15beee6e
......@@ -25,7 +25,6 @@ from paddle import DataParallel
from paddle import distributed as dist
from paddle.io import DataLoader
from paddle.io import DistributedBatchSampler
from visualdl import LogWriter
from yacs.config import CfgNode
from parakeet.datasets.am_batch_fn import speedyspeech_batch_fn
......@@ -153,8 +152,7 @@ def train_sp(args, config):
if dist.get_rank() == 0:
trainer.extend(evaluator, trigger=(1, "epoch"))
writer = LogWriter(str(output_dir))
trainer.extend(VisualDL(writer), trigger=(1, "iteration"))
trainer.extend(VisualDL(output_dir), trigger=(1, "iteration"))
trainer.extend(
Snapshot(max_size=config.num_snapshots), trigger=(1, 'epoch'))
trainer.run()
......
parakeet/exps/tacotron2/ljspeech.py
浏览文件 @ 15beee6e
......@@ -67,19 +67,16 @@ class LJSpeechCollector(object):
# Sort by text_len in descending order
texts = [
i
for i, _ in sorted(
i for i, _ in sorted(
zip(texts, text_lens), key=lambda x: x[1], reverse=True)
]
mels = [
i
for i, _ in sorted(
i for i, _ in sorted(
zip(mels, text_lens), key=lambda x: x[1], reverse=True)
]
mel_lens = [
i
for i, _ in sorted(
i for i, _ in sorted(
zip(mel_lens, text_lens), key=lambda x: x[1], reverse=True)
]
......
parakeet/exps/transformer_tts/train.py
浏览文件 @ 15beee6e
......@@ -25,7 +25,6 @@ from paddle import DataParallel
from paddle import distributed as dist
from paddle.io import DataLoader
from paddle.io import DistributedBatchSampler
from visualdl import LogWriter
from yacs.config import CfgNode
from parakeet.datasets.am_batch_fn import transformer_single_spk_batch_fn
......@@ -148,8 +147,7 @@ def train_sp(args, config):
if dist.get_rank() == 0:
trainer.extend(evaluator, trigger=(1, "epoch"))
writer = LogWriter(str(output_dir))
trainer.extend(VisualDL(writer), trigger=(1, "iteration"))
trainer.extend(VisualDL(output_dir), trigger=(1, "iteration"))
trainer.extend(
Snapshot(max_size=config.num_snapshots), trigger=(1, 'epoch'))
# print(trainer.extensions)
......
parakeet/models/fastspeech2/fastspeech2.py
浏览文件 @ 15beee6e
......@@ -341,6 +341,7 @@ class FastSpeech2(nn.Layer):
Tensor
speech_lengths, modified if reduction_factor > 1
"""
# input of embedding must be int64
xs = paddle.cast(text, 'int64')
ilens = paddle.cast(text_lengths, 'int64')
......@@ -388,7 +389,6 @@ class FastSpeech2(nn.Layer):
tone_id=None) -> Sequence[paddle.Tensor]:
# forward encoder
x_masks = self._source_mask(ilens)
# (B, Tmax, adim)
hs, _ = self.encoder(xs, x_masks)
......@@ -405,7 +405,6 @@ class FastSpeech2(nn.Layer):
if tone_id is not None:
tone_embs = self.tone_embedding_table(tone_id)
hs = self._integrate_with_tone_embed(hs, tone_embs)
# forward duration predictor and variance predictors
d_masks = make_pad_mask(ilens)
......@@ -437,6 +436,7 @@ class FastSpeech2(nn.Layer):
e_embs = self.energy_embed(e_outs.transpose((0, 2, 1))).transpose(
(0, 2, 1))
hs = hs + e_embs + p_embs
# (B, Lmax, adim)
hs = self.length_regulator(hs, d_outs, alpha)
else:
......@@ -447,6 +447,7 @@ class FastSpeech2(nn.Layer):
e_embs = self.energy_embed(es.transpose((0, 2, 1))).transpose(
(0, 2, 1))
hs = hs + e_embs + p_embs
# (B, Lmax, adim)
hs = self.length_regulator(hs, ds)
......@@ -461,9 +462,11 @@ class FastSpeech2(nn.Layer):
else:
h_masks = None
# (B, Lmax, adim)
zs, _ = self.decoder(hs, h_masks)
# (B, Lmax, odim)
before_outs = self.feat_out(zs).reshape((zs.shape[0], -1, self.odim))
before_outs = self.feat_out(zs).reshape(
(paddle.shape(zs)[0], -1, self.odim))
# postnet -> (B, Lmax//r * r, odim)
if self.postnet is None:
......@@ -526,8 +529,8 @@ class FastSpeech2(nn.Layer):
d = paddle.cast(durations, 'int64')
p, e = pitch, energy
# setup batch axis
ilens = paddle.to_tensor(
[x.shape[0]], dtype=paddle.int64, place=x.place)
ilens = paddle.shape(x)[0]
xs, ys = x.unsqueeze(0), None
if y is not None:
......
parakeet/models/melgan/__init__.py 0 → 100644
浏览文件 @ 15beee6e
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .melgan import *
from .multi_band_melgan_updater import *
parakeet/models/melgan/melgan.py 0 → 100644
浏览文件 @ 15beee6e
# 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.
"""MelGAN Modules."""
from typing import Any
from typing import Dict
from typing import List
import numpy as np
import paddle
from paddle import nn
from parakeet.modules.causal_conv import CausalConv1D
from parakeet.modules.causal_conv import CausalConv1DTranspose
from parakeet.modules.nets_utils import initialize
from parakeet.modules.pqmf import PQMF
from parakeet.modules.residual_stack import ResidualStack
class MelGANGenerator(nn.Layer):
"""MelGAN generator module."""
def __init__(
self,
in_channels: int=80,
out_channels: int=1,
kernel_size: int=7,
channels: int=512,
bias: bool=True,
upsample_scales: List[int]=[8, 8, 2, 2],
stack_kernel_size: int=3,
stacks: int=3,
nonlinear_activation: str="LeakyReLU",
nonlinear_activation_params: Dict[str, Any]={"negative_slope": 0.2},
pad: str="Pad1D",
pad_params: Dict[str, Any]={"mode": "reflect"},
use_final_nonlinear_activation: bool=True,
use_weight_norm: bool=True,
use_causal_conv: bool=False,
init_type: str="xavier_uniform", ):
"""Initialize MelGANGenerator module.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels,
the number of sub-band is out_channels in multi-band melgan.
kernel_size : int
Kernel size of initial and final conv layer.
channels : int
Initial number of channels for conv layer.
bias : bool
Whether to add bias parameter in convolution layers.
upsample_scales : List[int]
List of upsampling scales.
stack_kernel_size : int
Kernel size of dilated conv layers in residual stack.
stacks : int
Number of stacks in a single residual stack.
nonlinear_activation : Optional[str], optional
Non linear activation in upsample network, by default None
nonlinear_activation_params : Dict[str, Any], optional
Parameters passed to the linear activation in the upsample network,
by default {}
pad : str
Padding function module name before dilated convolution layer.
pad_params : dict
Hyperparameters for padding function.
use_final_nonlinear_activation : paddle.nn.Layer
Activation function for the final layer.
use_weight_norm : bool
Whether to use weight norm.
If set to true, it will be applied to all of the conv layers.
use_causal_conv : bool
Whether to use causal convolution.
"""
super().__init__()
# check hyper parameters is valid
assert channels >= np.prod(upsample_scales)
assert channels % (2**len(upsample_scales)) == 0
if not use_causal_conv:
assert (kernel_size - 1
) % 2 == 0, "Not support even number kernel size."
# initialize parameters
initialize(self, init_type)
layers = []
if not use_causal_conv:
layers += [
getattr(paddle.nn, pad)((kernel_size - 1) // 2, **pad_params),
nn.Conv1D(in_channels, channels, kernel_size, bias_attr=bias),
]
else:
layers += [
CausalConv1D(
in_channels,
channels,
kernel_size,
bias=bias,
pad=pad,
pad_params=pad_params, ),
]
for i, upsample_scale in enumerate(upsample_scales):
# add upsampling layer
layers += [
getattr(nn, nonlinear_activation)(**nonlinear_activation_params)
]
if not use_causal_conv:
layers += [
nn.Conv1DTranspose(
channels // (2**i),
channels // (2**(i + 1)),
upsample_scale * 2,
stride=upsample_scale,
padding=upsample_scale // 2 + upsample_scale % 2,
output_padding=upsample_scale % 2,
bias_attr=bias, )
]
else:
layers += [
CausalConv1DTranspose(
channels // (2**i),
channels // (2**(i + 1)),
upsample_scale * 2,
stride=upsample_scale,
bias=bias, )
]
# add residual stack
for j in range(stacks):
layers += [
ResidualStack(
kernel_size=stack_kernel_size,
channels=channels // (2**(i + 1)),
dilation=stack_kernel_size**j,
bias=bias,
nonlinear_activation=nonlinear_activation,
nonlinear_activation_params=nonlinear_activation_params,
pad=pad,
pad_params=pad_params,
use_causal_conv=use_causal_conv, )
]
# add final layer
layers += [
getattr(nn, nonlinear_activation)(**nonlinear_activation_params)
]
if not use_causal_conv:
layers += [
getattr(nn, pad)((kernel_size - 1) // 2, **pad_params),
nn.Conv1D(
channels // (2**(i + 1)),
out_channels,
kernel_size,
bias_attr=bias),
]
else:
layers += [
CausalConv1D(
channels // (2**(i + 1)),
out_channels,
kernel_size,
bias=bias,
pad=pad,
pad_params=pad_params, ),
]
if use_final_nonlinear_activation:
layers += [nn.Tanh()]
# define the model as a single function
self.melgan = nn.Sequential(*layers)
nn.initializer.set_global_initializer(None)
# apply weight norm
if use_weight_norm:
self.apply_weight_norm()
# reset parameters
self.reset_parameters()
# initialize pqmf for multi-band melgan inference
if out_channels > 1:
self.pqmf = PQMF(subbands=out_channels)
else:
self.pqmf = None
def forward(self, c):
"""Calculate forward propagation.
Parameters
----------
c : Tensor
Input tensor (B, in_channels, T).
Returns
----------
Tensor
Output tensor (B, out_channels, T ** prod(upsample_scales)).
"""
out = self.melgan(c)
return out
def apply_weight_norm(self):
"""Recursively apply weight normalization to all the Convolution layers
in the sublayers.
"""
def _apply_weight_norm(layer):
if isinstance(layer, (nn.Conv1D, nn.Conv2D, nn.Conv1DTranspose)):
nn.utils.weight_norm(layer)
self.apply(_apply_weight_norm)
def remove_weight_norm(self):
"""Recursively remove weight normalization from all the Convolution
layers in the sublayers.
"""
def _remove_weight_norm(layer):
try:
nn.utils.remove_weight_norm(layer)
except ValueError:
pass
self.apply(_remove_weight_norm)
def reset_parameters(self):
"""Reset parameters.
This initialization follows official implementation manner.
https://github.com/descriptinc/melgan-neurips/blob/master/mel2wav/modules.py
"""
# 定义参数为float的正态分布。
dist = paddle.distribution.Normal(loc=0.0, scale=0.02)
def _reset_parameters(m):
if isinstance(m, nn.Conv1D) or isinstance(m, nn.Conv1DTranspose):
w = dist.sample(m.weight.shape)
m.weight.set_value(w)
self.apply(_reset_parameters)
def inference(self, c):
"""Perform inference.
Parameters
----------
c : Union[Tensor, ndarray]
Input tensor (T, in_channels).
Returns
----------
Tensor
Output tensor (out_channels*T ** prod(upsample_scales), 1).
"""
if not isinstance(c, paddle.Tensor):
c = paddle.to_tensor(c, dtype="float32")
# pseudo batch
c = c.transpose([1, 0]).unsqueeze(0)
# (B, out_channels, T ** prod(upsample_scales)
out = self.melgan(c)
if self.pqmf is not None:
# (B, 1, out_channels * T ** prod(upsample_scales)
out = self.pqmf.synthesis(out)
out = out.squeeze(0).transpose([1, 0])
return out
class MelGANDiscriminator(nn.Layer):
"""MelGAN discriminator module."""
def __init__(
self,
in_channels: int=1,
out_channels: int=1,
kernel_sizes: List[int]=[5, 3],
channels: int=16,
max_downsample_channels: int=1024,
bias: bool=True,
downsample_scales: List[int]=[4, 4, 4, 4],
nonlinear_activation: str="LeakyReLU",
nonlinear_activation_params: Dict[str, Any]={"negative_slope": 0.2},
pad: str="Pad1D",
pad_params: Dict[str, Any]={"mode": "reflect"}, ):
"""Initilize MelGAN discriminator module.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_sizes : List[int]
List of two kernel sizes. The prod will be used for the first conv layer,
and the first and the second kernel sizes will be used for the last two layers.
For example if kernel_sizes = [5, 3], the first layer kernel size will be 5 * 3 = 15,
the last two layers' kernel size will be 5 and 3, respectively.
channels : int
Initial number of channels for conv layer.
max_downsample_channels : int
Maximum number of channels for downsampling layers.
bias : bool
Whether to add bias parameter in convolution layers.
downsample_scales : List[int]
List of downsampling scales.
nonlinear_activation : str
Activation function module name.
nonlinear_activation_params : dict
Hyperparameters for activation function.
pad : str
Padding function module name before dilated convolution layer.
pad_params : dict
Hyperparameters for padding function.
"""
super().__init__()
self.layers = nn.LayerList()
# check kernel size is valid
assert len(kernel_sizes) == 2
assert kernel_sizes[0] % 2 == 1
assert kernel_sizes[1] % 2 == 1
# add first layer
self.layers.append(
nn.Sequential(
getattr(nn, pad)((np.prod(kernel_sizes) - 1) // 2, **
pad_params),
nn.Conv1D(
in_channels,
channels,
int(np.prod(kernel_sizes)),
bias_attr=bias),
getattr(nn, nonlinear_activation)(
**nonlinear_activation_params), ))
# add downsample layers
in_chs = channels
for downsample_scale in downsample_scales:
out_chs = min(in_chs * downsample_scale, max_downsample_channels)
self.layers.append(
nn.Sequential(
nn.Conv1D(
in_chs,
out_chs,
kernel_size=downsample_scale * 10 + 1,
stride=downsample_scale,
padding=downsample_scale * 5,
groups=in_chs // 4,
bias_attr=bias, ),
getattr(nn, nonlinear_activation)(
**nonlinear_activation_params), ))
in_chs = out_chs
# add final layers
out_chs = min(in_chs * 2, max_downsample_channels)
self.layers.append(
nn.Sequential(
nn.Conv1D(
in_chs,
out_chs,
kernel_sizes[0],
padding=(kernel_sizes[0] - 1) // 2,
bias_attr=bias, ),
getattr(nn, nonlinear_activation)(
**nonlinear_activation_params), ))
self.layers.append(
nn.Conv1D(
out_chs,
out_channels,
kernel_sizes[1],
padding=(kernel_sizes[1] - 1) // 2,
bias_attr=bias, ), )
def forward(self, x):
"""Calculate forward propagation.
Parameters
----------
x : Tensor
Input noise signal (B, 1, T).
Returns
----------
List
List of output tensors of each layer (for feat_match_loss).
"""
outs = []
for f in self.layers:
x = f(x)
outs += [x]
return outs
class MelGANMultiScaleDiscriminator(nn.Layer):
"""MelGAN multi-scale discriminator module."""
def __init__(
self,
in_channels: int=1,
out_channels: int=1,
scales: int=3,
downsample_pooling: str="AvgPool1D",
# follow the official implementation setting
downsample_pooling_params: Dict[str, Any]={
"kernel_size": 4,
"stride": 2,
"padding": 1,
"exclusive": True,
},
kernel_sizes: List[int]=[5, 3],
channels: int=16,
max_downsample_channels: int=1024,
bias: bool=True,
downsample_scales: List[int]=[4, 4, 4, 4],
nonlinear_activation: str="LeakyReLU",
nonlinear_activation_params: Dict[str, Any]={"negative_slope": 0.2},
pad: str="Pad1D",
pad_params: Dict[str, Any]={"mode": "reflect"},
use_weight_norm: bool=True,
init_type: str="xavier_uniform", ):
"""Initilize MelGAN multi-scale discriminator module.
Parameters
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
scales : int
Number of multi-scales.
downsample_pooling : str
Pooling module name for downsampling of the inputs.
downsample_pooling_params : dict
Parameters for the above pooling module.
kernel_sizes : List[int]
List of two kernel sizes. The sum will be used for the first conv layer,
and the first and the second kernel sizes will be used for the last two layers.
channels : int
Initial number of channels for conv layer.
max_downsample_channels : int
Maximum number of channels for downsampling layers.
bias : bool
Whether to add bias parameter in convolution layers.
downsample_scales : List[int]
List of downsampling scales.
nonlinear_activation : str
Activation function module name.
nonlinear_activation_params : dict
Hyperparameters for activation function.
pad : str
Padding function module name before dilated convolution layer.
pad_params : dict
Hyperparameters for padding function.
use_causal_conv : bool
Whether to use causal convolution.
"""
super().__init__()
# initialize parameters
initialize(self, init_type)
self.discriminators = nn.LayerList()
# add discriminators
for _ in range(scales):
self.discriminators.append(
MelGANDiscriminator(
in_channels=in_channels,
out_channels=out_channels,
kernel_sizes=kernel_sizes,
channels=channels,
max_downsample_channels=max_downsample_channels,
bias=bias,
downsample_scales=downsample_scales,
nonlinear_activation=nonlinear_activation,
nonlinear_activation_params=nonlinear_activation_params,
pad=pad,
pad_params=pad_params, ))
self.pooling = getattr(nn, downsample_pooling)(
**downsample_pooling_params)
nn.initializer.set_global_initializer(None)
# apply weight norm
if use_weight_norm:
self.apply_weight_norm()
# reset parameters
self.reset_parameters()
def forward(self, x):
"""Calculate forward propagation.
Parameters
----------
x : Tensor
Input noise signal (B, 1, T).
Returns
----------
List
List of list of each discriminator outputs, which consists of each layer output tensors.
"""
outs = []
for f in self.discriminators:
outs += [f(x)]
x = self.pooling(x)
return outs
def apply_weight_norm(self):
"""Recursively apply weight normalization to all the Convolution layers
in the sublayers.
"""
def _apply_weight_norm(layer):
if isinstance(layer, (nn.Conv1D, nn.Conv2D, nn.Conv1DTranspose)):
nn.utils.weight_norm(layer)
self.apply(_apply_weight_norm)
def remove_weight_norm(self):
"""Recursively remove weight normalization from all the Convolution
layers in the sublayers.
"""
def _remove_weight_norm(layer):
try:
nn.utils.remove_weight_norm(layer)
except ValueError:
pass
self.apply(_remove_weight_norm)
def reset_parameters(self):
"""Reset parameters.
This initialization follows official implementation manner.
https://github.com/descriptinc/melgan-neurips/blob/master/mel2wav/modules.py
"""
# 定义参数为float的正态分布。
dist = paddle.distribution.Normal(loc=0.0, scale=0.02)
def _reset_parameters(m):
if isinstance(m, nn.Conv1D) or isinstance(m, nn.Conv1DTranspose):
w = dist.sample(m.weight.shape)
m.weight.set_value(w)
self.apply(_reset_parameters)
parakeet/models/melgan/multi_band_melgan_updater.py 0 → 100644
浏览文件 @ 15beee6e
# 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
from typing import Dict
import paddle
from paddle import distributed as dist
from paddle.io import DataLoader
from paddle.nn import Layer
from paddle.optimizer import Optimizer
from paddle.optimizer.lr import LRScheduler
from parakeet.training.extensions.evaluator import StandardEvaluator
from parakeet.training.reporter import report
from parakeet.training.updaters.standard_updater import StandardUpdater
from parakeet.training.updaters.standard_updater import UpdaterState
logging.basicConfig(
format='%(asctime)s [%(levelname)s] [%(filename)s:%(lineno)d] %(message)s',
datefmt='[%Y-%m-%d %H:%M:%S]')
logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
class MBMelGANUpdater(StandardUpdater):
def __init__(self,
models: Dict[str, Layer],
optimizers: Dict[str, Optimizer],
criterions: Dict[str, Layer],
schedulers: Dict[str, LRScheduler],
dataloader: DataLoader,
discriminator_train_start_steps: int,
lambda_adv: float,
output_dir=None):
self.models = models
self.generator: Layer = models['generator']
self.discriminator: Layer = models['discriminator']
self.optimizers = optimizers
self.optimizer_g: Optimizer = optimizers['generator']
self.optimizer_d: Optimizer = optimizers['discriminator']
self.criterions = criterions
self.criterion_stft = criterions['stft']
self.criterion_sub_stft = criterions['sub_stft']
self.criterion_pqmf = criterions['pqmf']
self.criterion_gen_adv = criterions["gen_adv"]
self.criterion_dis_adv = criterions["dis_adv"]
self.schedulers = schedulers
self.scheduler_g = schedulers['generator']
self.scheduler_d = schedulers['discriminator']
self.dataloader = dataloader
self.discriminator_train_start_steps = discriminator_train_start_steps
self.lambda_adv = lambda_adv
self.state = UpdaterState(iteration=0, epoch=0)
self.train_iterator = iter(self.dataloader)
log_file = output_dir / 'worker_{}.log'.format(dist.get_rank())
self.filehandler = logging.FileHandler(str(log_file))
logger.addHandler(self.filehandler)
self.logger = logger
self.msg = ""
def update_core(self, batch):
self.msg = "Rank: {}, ".format(dist.get_rank())
losses_dict = {}
# parse batch
wav, mel = batch
# Generator
# (B, out_channels, T ** prod(upsample_scales)
wav_ = self.generator(mel)
wav_mb_ = wav_
# (B, 1, out_channels*T ** prod(upsample_scales)
wav_ = self.criterion_pqmf.synthesis(wav_mb_)
# initialize
gen_loss = 0.0
# full band Multi-resolution stft loss
sc_loss, mag_loss = self.criterion_stft(wav_, wav)
# for balancing with subband stft loss
# Eq.(9) in paper
gen_loss += 0.5 * (sc_loss + mag_loss)
report("train/spectral_convergence_loss", float(sc_loss))
report("train/log_stft_magnitude_loss", float(mag_loss))
losses_dict["spectral_convergence_loss"] = float(sc_loss)
losses_dict["log_stft_magnitude_loss"] = float(mag_loss)
# sub band Multi-resolution stft loss
# (B, subbands, T // subbands)
wav_mb = self.criterion_pqmf.analysis(wav)
sub_sc_loss, sub_mag_loss = self.criterion_sub_stft(wav_mb_, wav_mb)
# Eq.(9) in paper
gen_loss += 0.5 * (sub_sc_loss + sub_mag_loss)
report("train/sub_spectral_convergence_loss", float(sub_sc_loss))
report("train/sub_log_stft_magnitude_loss", float(sub_mag_loss))
losses_dict["sub_spectral_convergence_loss"] = float(sub_sc_loss)
losses_dict["sub_log_stft_magnitude_loss"] = float(sub_mag_loss)
## Adversarial loss
if self.state.iteration > self.discriminator_train_start_steps:
p_ = self.discriminator(wav_)
adv_loss = self.criterion_gen_adv(p_)
report("train/adversarial_loss", float(adv_loss))
losses_dict["adversarial_loss"] = float(adv_loss)
gen_loss += self.lambda_adv * adv_loss
report("train/generator_loss", float(gen_loss))
losses_dict["generator_loss"] = float(gen_loss)
self.optimizer_g.clear_grad()
gen_loss.backward()
self.optimizer_g.step()
self.scheduler_g.step()
# Disctiminator
if self.state.iteration > self.discriminator_train_start_steps:
# re-compute wav_ which leads better quality
with paddle.no_grad():
wav_ = self.generator(mel)
wav_ = self.criterion_pqmf.synthesis(wav_)
p = self.discriminator(wav)
p_ = self.discriminator(wav_.detach())
real_loss, fake_loss = self.criterion_dis_adv(p_, p)
dis_loss = real_loss + fake_loss
report("train/real_loss", float(real_loss))
report("train/fake_loss", float(fake_loss))
report("train/discriminator_loss", float(dis_loss))
losses_dict["real_loss"] = float(real_loss)
losses_dict["fake_loss"] = float(fake_loss)
losses_dict["discriminator_loss"] = float(dis_loss)
self.optimizer_d.clear_grad()
dis_loss.backward()
self.optimizer_d.step()
self.scheduler_d.step()
self.msg += ', '.join('{}: {:>.6f}'.format(k, v)
for k, v in losses_dict.items())
class MBMelGANEvaluator(StandardEvaluator):
def __init__(self,
models,
criterions,
dataloader,
lambda_adv,
output_dir=None):
self.models = models
self.generator = models['generator']
self.discriminator = models['discriminator']
self.criterions = criterions
self.criterion_stft = criterions['stft']
self.criterion_sub_stft = criterions['sub_stft']
self.criterion_pqmf = criterions['pqmf']
self.criterion_gen_adv = criterions["gen_adv"]
self.criterion_dis_adv = criterions["dis_adv"]
self.dataloader = dataloader
self.lambda_adv = lambda_adv
log_file = output_dir / 'worker_{}.log'.format(dist.get_rank())
self.filehandler = logging.FileHandler(str(log_file))
logger.addHandler(self.filehandler)
self.logger = logger
self.msg = ""
def evaluate_core(self, batch):
# logging.debug("Evaluate: ")
self.msg = "Evaluate: "
losses_dict = {}
wav, mel = batch
# Generator
# (B, out_channels, T ** prod(upsample_scales)
wav_ = self.generator(mel)
wav_mb_ = wav_
# (B, 1, out_channels*T ** prod(upsample_scales)
wav_ = self.criterion_pqmf.synthesis(wav_mb_)
## Adversarial loss
p_ = self.discriminator(wav_)
adv_loss = self.criterion_gen_adv(p_)
report("eval/adversarial_loss", float(adv_loss))
losses_dict["adversarial_loss"] = float(adv_loss)
gen_loss = self.lambda_adv * adv_loss
# Multi-resolution stft loss
sc_loss, mag_loss = self.criterion_stft(wav_, wav)
# Eq.(9) in paper
gen_loss += 0.5 * (sc_loss + mag_loss)
report("eval/spectral_convergence_loss", float(sc_loss))
report("eval/log_stft_magnitude_loss", float(mag_loss))
losses_dict["spectral_convergence_loss"] = float(sc_loss)
losses_dict["log_stft_magnitude_loss"] = float(mag_loss)
# sub band Multi-resolution stft loss
# (B, subbands, T // subbands)
wav_mb = self.criterion_pqmf.analysis(wav)
sub_sc_loss, sub_mag_loss = self.criterion_sub_stft(wav_mb_, wav_mb)
# Eq.(9) in paper
gen_loss += 0.5 * (sub_sc_loss + sub_mag_loss)
report("eval/sub_spectral_convergence_loss", float(sub_sc_loss))
report("eval/sub_log_stft_magnitude_loss", float(sub_mag_loss))
losses_dict["sub_spectral_convergence_loss"] = float(sub_sc_loss)
losses_dict["sub_log_stft_magnitude_loss"] = float(sub_mag_loss)
report("eval/generator_loss", float(gen_loss))
losses_dict["generator_loss"] = float(gen_loss)
# Disctiminator
p = self.discriminator(wav)
real_loss, fake_loss = self.criterion_dis_adv(p_, p)
dis_loss = real_loss + fake_loss
report("eval/real_loss", float(real_loss))
report("eval/fake_loss", float(fake_loss))
report("eval/discriminator_loss", float(dis_loss))
losses_dict["real_loss"] = float(real_loss)
losses_dict["fake_loss"] = float(fake_loss)
losses_dict["discriminator_loss"] = float(dis_loss)
self.msg += ', '.join('{}: {:>.6f}'.format(k, v)
for k, v in losses_dict.items())
self.logger.info(self.msg)
parakeet/models/parallel_wavegan/parallel_wavegan.py
浏览文件 @ 15beee6e
......@@ -498,7 +498,6 @@ class PWGGenerator(nn.Layer):
def inference(self, c=None):
"""Waveform generation. This function is used for single instance
inference.
Parameters
----------
c : Tensor, optional
......@@ -506,12 +505,12 @@ class PWGGenerator(nn.Layer):
x : Tensor, optional
Shape (T, C_in), the noise waveform, by default None
If not provided, a sample is drawn from a gaussian distribution.
Returns
-------
Tensor
Shape (T, C_out), the generated waveform
"""
# when to static, can not input x, see https://github.com/PaddlePaddle/Parakeet/pull/132/files
x = paddle.randn(
[1, self.in_channels, paddle.shape(c)[0] * self.upsample_factor])
c = paddle.transpose(c, [1, 0]).unsqueeze(0) # pseudo batch
......
parakeet/modules/adversarial_loss.py 0 → 100644
浏览文件 @ 15beee6e
# 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.
"""Adversarial loss modules."""
import paddle
import paddle.nn.functional as F
from paddle import nn
class GeneratorAdversarialLoss(nn.Layer):
"""Generator adversarial loss module."""
def __init__(
self,
average_by_discriminators=True,
loss_type="mse", ):
"""Initialize GeneratorAversarialLoss module."""
super().__init__()
self.average_by_discriminators = average_by_discriminators
assert loss_type in ["mse", "hinge"], f"{loss_type} is not supported."
if loss_type == "mse":
self.criterion = self._mse_loss
else:
self.criterion = self._hinge_loss
def forward(self, outputs):
"""Calcualate generator adversarial loss.
Parameters
----------
outputs: Tensor or List
Discriminator outputs or list of discriminator outputs.
Returns
----------
Tensor
Generator adversarial loss value.
"""
if isinstance(outputs, (tuple, list)):
adv_loss = 0.0
for i, outputs_ in enumerate(outputs):
if isinstance(outputs_, (tuple, list)):
# case including feature maps
outputs_ = outputs_[-1]
adv_loss += self.criterion(outputs_)
if self.average_by_discriminators:
adv_loss /= i + 1
else:
adv_loss = self.criterion(outputs)
return adv_loss
def _mse_loss(self, x):
return F.mse_loss(x, paddle.ones_like(x))
def _hinge_loss(self, x):
return -x.mean()
class DiscriminatorAdversarialLoss(nn.Layer):
"""Discriminator adversarial loss module."""
def __init__(
self,
average_by_discriminators=True,
loss_type="mse", ):
"""Initialize DiscriminatorAversarialLoss module."""
super().__init__()
self.average_by_discriminators = average_by_discriminators
assert loss_type in ["mse"], f"{loss_type} is not supported."
if loss_type == "mse":
self.fake_criterion = self._mse_fake_loss
self.real_criterion = self._mse_real_loss
def forward(self, outputs_hat, outputs):
"""Calcualate discriminator adversarial loss.
Parameters
----------
outputs_hat : Tensor or list
Discriminator outputs or list of
discriminator outputs calculated from generator outputs.
outputs : Tensor or list
Discriminator outputs or list of
discriminator outputs calculated from groundtruth.
Returns
----------
Tensor
Discriminator real loss value.
Tensor
Discriminator fake loss value.
"""
if isinstance(outputs, (tuple, list)):
real_loss = 0.0
fake_loss = 0.0
for i, (outputs_hat_,
outputs_) in enumerate(zip(outputs_hat, outputs)):
if isinstance(outputs_hat_, (tuple, list)):
# case including feature maps
outputs_hat_ = outputs_hat_[-1]
outputs_ = outputs_[-1]
real_loss += self.real_criterion(outputs_)
fake_loss += self.fake_criterion(outputs_hat_)
if self.average_by_discriminators:
fake_loss /= i + 1
real_loss /= i + 1
else:
real_loss = self.real_criterion(outputs)
fake_loss = self.fake_criterion(outputs_hat)
return real_loss, fake_loss
def _mse_real_loss(self, x):
return F.mse_loss(x, paddle.ones_like(x))
def _mse_fake_loss(self, x):
return F.mse_loss(x, paddle.zeros_like(x))
parakeet/modules/causal_conv.py 0 → 100644
浏览文件 @ 15beee6e
# 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.
"""Causal convolusion layer modules."""
import paddle
class CausalConv1D(paddle.nn.Layer):
"""CausalConv1D module with customized initialization."""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
dilation=1,
bias=True,
pad="Pad1D",
pad_params={"value": 0.0}, ):
"""Initialize CausalConv1d module."""
super().__init__()
self.pad = getattr(paddle.nn, pad)((kernel_size - 1) * dilation,
**pad_params)
self.conv = paddle.nn.Conv1D(
in_channels,
out_channels,
kernel_size,
dilation=dilation,
bias_attr=bias)
def forward(self, x):
"""Calculate forward propagation.
Parameters
----------
x : Tensor
Input tensor (B, in_channels, T).
Returns
----------
Tensor
Output tensor (B, out_channels, T).
"""
return self.conv(self.pad(x))[:, :, :x.shape[2]]
class CausalConv1DTranspose(paddle.nn.Layer):
"""CausalConv1DTranspose module with customized initialization."""
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
bias=True):
"""Initialize CausalConvTranspose1d module."""
super().__init__()
self.deconv = paddle.nn.Conv1DTranspose(
in_channels, out_channels, kernel_size, stride, bias_attr=bias)
self.stride = stride
def forward(self, x):
"""Calculate forward propagation.
Parameters
----------
x : Tensor
Input tensor (B, in_channels, T_in).
Returns
----------
Tensor
Output tensor (B, out_channels, T_out).
"""
return self.deconv(x)[:, :, :-self.stride]
parakeet/modules/fastspeech2_predictor/length_regulator.py
浏览文件 @ 15beee6e
......@@ -12,7 +12,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Length regulator related modules."""
import numpy as np
import paddle
from paddle import nn
......@@ -49,11 +48,10 @@ class LengthRegulator(nn.Layer):
encodings: (B, T, C)
durations: (B, T)
"""
batch_size, t_enc = durations.shape
durations = durations.numpy()
slens = np.sum(durations, -1)
t_dec = np.max(slens)
M = np.zeros([batch_size, t_dec, t_enc])
batch_size, t_enc = paddle.shape(durations)
slens = durations.sum(-1)
t_dec = slens.max()
M = paddle.zeros([batch_size, t_dec, t_enc])
for i in range(batch_size):
k = 0
for j in range(t_enc):
......@@ -61,7 +59,6 @@ class LengthRegulator(nn.Layer):
if d >= 1:
M[i, k:k + d, j] = 1
k += d
M = paddle.to_tensor(M, dtype=encodings.dtype)
encodings = paddle.matmul(M, encodings)
return encodings
......@@ -82,6 +79,7 @@ class LengthRegulator(nn.Layer):
Tensor
replicated input tensor based on durations (B, T*, D).
"""
if alpha != 1.0:
assert alpha > 0
ds = paddle.round(ds.cast(dtype=paddle.float32) * alpha)
......
parakeet/modules/fastspeech2_transformer/attention.py
浏览文件 @ 15beee6e
......@@ -106,13 +106,11 @@ class MultiHeadedAttention(nn.Layer):
n_batch = value.shape[0]
softmax = paddle.nn.Softmax(axis=-1)
if mask is not None:
mask = mask.unsqueeze(1)
mask = paddle.logical_not(mask)
min_value = float(
numpy.finfo(
paddle.to_tensor(0, dtype=scores.dtype).numpy().dtype).min)
# assume scores.dtype==paddle.float32, we only use "float32" here
dtype = str(scores.dtype).split(".")[-1]
min_value = numpy.finfo(dtype).min
scores = masked_fill(scores, mask, min_value)
# (batch, head, time1, time2)
self.attn = softmax(scores)
......
parakeet/modules/fastspeech2_transformer/embedding.py
浏览文件 @ 15beee6e
......@@ -31,9 +31,16 @@ class PositionalEncoding(nn.Layer):
Maximum input length.
reverse : bool
Whether to reverse the input position.
type : str
dtype of param
"""
def __init__(self, d_model, dropout_rate, max_len=5000, reverse=False):
def __init__(self,
d_model,
dropout_rate,
max_len=5000,
dtype="float32",
reverse=False):
"""Construct an PositionalEncoding object."""
super(PositionalEncoding, self).__init__()
self.d_model = d_model
......@@ -41,20 +48,21 @@ class PositionalEncoding(nn.Layer):
self.xscale = math.sqrt(self.d_model)
self.dropout = nn.Dropout(p=dropout_rate)
self.pe = None
self.extend_pe(paddle.expand(paddle.to_tensor(0.0), (1, max_len)))
self.dtype = dtype
self.extend_pe(paddle.expand(paddle.zeros([1]), (1, max_len)))
def extend_pe(self, x):
"""Reset the positional encodings."""
pe = paddle.zeros([x.shape[1], self.d_model])
x_shape = paddle.shape(x)
pe = paddle.zeros([x_shape[1], self.d_model])
if self.reverse:
position = paddle.arange(
x.shape[1] - 1, -1, -1.0, dtype=paddle.float32).unsqueeze(1)
x_shape[1] - 1, -1, -1.0, dtype=self.dtype).unsqueeze(1)
else:
position = paddle.arange(
0, x.shape[1], dtype=paddle.float32).unsqueeze(1)
0, x_shape[1], dtype=self.dtype).unsqueeze(1)
div_term = paddle.exp(
paddle.arange(0, self.d_model, 2, dtype=paddle.float32) *
paddle.arange(0, self.d_model, 2, dtype=self.dtype) *
-(math.log(10000.0) / self.d_model))
pe[:, 0::2] = paddle.sin(position * div_term)
pe[:, 1::2] = paddle.cos(position * div_term)
......@@ -75,7 +83,8 @@ class PositionalEncoding(nn.Layer):
Encoded tensor (batch, time, `*`).
"""
self.extend_pe(x)
x = x * self.xscale + self.pe[:, :x.shape[1]]
T = paddle.shape(x)[1]
x = x * self.xscale + self.pe[:, :T]
return self.dropout(x)
......@@ -92,21 +101,26 @@ class ScaledPositionalEncoding(PositionalEncoding):
Dropout rate.
max_len : int
Maximum input length.
dtype : str
dtype of param
"""
def __init__(self, d_model, dropout_rate, max_len=5000):
def __init__(self, d_model, dropout_rate, max_len=5000, dtype="float32"):
"""Initialize class."""
super().__init__(
d_model=d_model, dropout_rate=dropout_rate, max_len=max_len)
x = paddle.ones([1], dtype="float32")
d_model=d_model,
dropout_rate=dropout_rate,
max_len=max_len,
dtype=dtype)
x = paddle.ones([1], dtype=self.dtype)
self.alpha = paddle.create_parameter(
shape=x.shape,
dtype=str(x.numpy().dtype),
dtype=self.dtype,
default_initializer=paddle.nn.initializer.Assign(x))
def reset_parameters(self):
"""Reset parameters."""
self.alpha = paddle.to_tensor(1.0)
self.alpha = paddle.ones([1])
def forward(self, x):
"""Add positional encoding.
......@@ -115,12 +129,12 @@ class ScaledPositionalEncoding(PositionalEncoding):
----------
x : paddle.Tensor
Input tensor (batch, time, `*`).
Returns
----------
paddle.Tensor
Encoded tensor (batch, time, `*`).
"""
self.extend_pe(x)
x = x + self.alpha * self.pe[:, :x.shape[1]]
T = paddle.shape(x)[1]
x = x + self.alpha * self.pe[:, :T]
return self.dropout(x)
parakeet/modules/fastspeech2_transformer/encoder.py
浏览文件 @ 15beee6e
......@@ -185,6 +185,7 @@ class Encoder(nn.Layer):
paddle.Tensor
Mask tensor (#batch, time).
"""
xs = self.embed(xs)
xs, masks = self.encoders(xs, masks)
if self.normalize_before:
......
parakeet/modules/layer_norm.py
浏览文件 @ 15beee6e
......@@ -44,6 +44,7 @@ class LayerNorm(paddle.nn.LayerNorm):
paddle.Tensor
Normalized tensor.
"""
if self.dim == -1:
return super(LayerNorm, self).forward(x)
else:
......@@ -54,9 +55,12 @@ class LayerNorm(paddle.nn.LayerNorm):
orig_perm = list(range(len_dim))
new_perm = orig_perm[:]
new_perm[self.dim], new_perm[len_dim -
1] = new_perm[len_dim -
1], new_perm[self.dim]
# Python style item change is not able when converting dygraph to static graph.
# new_perm[self.dim], new_perm[len_dim -1] = new_perm[len_dim -1], new_perm[self.dim]
# use C++ style item change here
temp = new_perm[self.dim]
new_perm[self.dim] = new_perm[len_dim - 1]
new_perm[len_dim - 1] = temp
return paddle.transpose(
super(LayerNorm, self).forward(paddle.transpose(x, new_perm)),
......
parakeet/modules/masked_fill.py
浏览文件 @ 15beee6e
......@@ -25,12 +25,24 @@ def is_broadcastable(shp1, shp2):
return True
# assume that len(shp1) == len(shp2)
def broadcast_shape(shp1, shp2):
result = []
for a, b in zip(shp1[::-1], shp2[::-1]):
result.append(max(a, b))
return result[::-1]
def masked_fill(xs: paddle.Tensor,
mask: paddle.Tensor,
value: Union[float, int]):
assert is_broadcastable(xs.shape, mask.shape) is True
bshape = paddle.broadcast_shape(xs.shape, mask.shape)
# comment following line for converting dygraph to static graph.
# assert is_broadcastable(xs.shape, mask.shape) is True
# bshape = paddle.broadcast_shape(xs.shape, mask.shape)
bshape = broadcast_shape(xs.shape, mask.shape)
mask.stop_gradient = True
mask = mask.broadcast_to(bshape)
trues = paddle.ones_like(xs) * value
mask = mask.cast(dtype=paddle.bool)
xs = paddle.where(mask, trues, xs)
......
parakeet/modules/nets_utils.py
浏览文件 @ 15beee6e
......@@ -56,7 +56,7 @@ def make_pad_mask(lengths, length_dim=-1):
Parameters
----------
lengths : LongTensor or List
lengths : LongTensor
Batch of lengths (B,).
Returns
......@@ -77,17 +77,11 @@ def make_pad_mask(lengths, length_dim=-1):
if length_dim == 0:
raise ValueError("length_dim cannot be 0: {}".format(length_dim))
if not isinstance(lengths, list):
lengths = lengths.tolist()
bs = int(len(lengths))
maxlen = int(max(lengths))
bs = paddle.shape(lengths)[0]
maxlen = lengths.max()
seq_range = paddle.arange(0, maxlen, dtype=paddle.int64)
seq_range_expand = seq_range.unsqueeze(0).expand([bs, maxlen])
seq_length_expand = paddle.to_tensor(
lengths, dtype=seq_range_expand.dtype).unsqueeze(-1)
seq_length_expand = lengths.unsqueeze(-1)
mask = seq_range_expand >= seq_length_expand
return mask
......
parakeet/modules/pqmf.py 0 → 100644
浏览文件 @ 15beee6e
# 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.
"""Pseudo QMF modules."""
import numpy as np
import paddle
import paddle.nn.functional as F
from scipy.signal import kaiser
def design_prototype_filter(taps=62, cutoff_ratio=0.142, beta=9.0):
"""Design prototype filter for PQMF.
This method is based on `A Kaiser window approach for the design of prototype
filters of cosine modulated filterbanks`_.
Parameters
----------
taps : int
The number of filter taps.
cutoff_ratio : float
Cut-off frequency ratio.
beta : float
Beta coefficient for kaiser window.
Returns
----------
ndarray
Impluse response of prototype filter (taps + 1,).
.. _`A Kaiser window approach for the design of prototype filters of cosine modulated filterbanks`:
https://ieeexplore.ieee.org/abstract/document/681427
"""
# check the arguments are valid
assert taps % 2 == 0, "The number of taps mush be even number."
assert 0.0 < cutoff_ratio < 1.0, "Cutoff ratio must be > 0.0 and < 1.0."
# make initial filter
omega_c = np.pi * cutoff_ratio
with np.errstate(invalid="ignore"):
h_i = np.sin(omega_c * (np.arange(taps + 1) - 0.5 * taps)) / (
np.pi * (np.arange(taps + 1) - 0.5 * taps))
h_i[taps //
2] = np.cos(0) * cutoff_ratio # fix nan due to indeterminate form
# apply kaiser window
w = kaiser(taps + 1, beta)
h = h_i * w
return h
class PQMF(paddle.nn.Layer):
"""PQMF module.
This module is based on `Near-perfect-reconstruction pseudo-QMF banks`_.
.. _`Near-perfect-reconstruction pseudo-QMF banks`:
https://ieeexplore.ieee.org/document/258122
"""
def __init__(self, subbands=4, taps=62, cutoff_ratio=0.142, beta=9.0):
"""Initilize PQMF module.
The cutoff_ratio and beta parameters are optimized for #subbands = 4.
See dicussion in https://github.com/kan-bayashi/ParallelWaveGAN/issues/195.
Parameters
----------
subbands : int
The number of subbands.
taps : int
The number of filter taps.
cutoff_ratio : float
Cut-off frequency ratio.
beta : float
Beta coefficient for kaiser window.
"""
super(PQMF, self).__init__()
# build analysis & synthesis filter coefficients
h_proto = design_prototype_filter(taps, cutoff_ratio, beta)
h_analysis = np.zeros((subbands, len(h_proto)))
h_synthesis = np.zeros((subbands, len(h_proto)))
for k in range(subbands):
h_analysis[k] = (
2 * h_proto * np.cos((2 * k + 1) * (np.pi / (2 * subbands)) * (
np.arange(taps + 1) - (taps / 2)) + (-1)**k * np.pi / 4))
h_synthesis[k] = (
2 * h_proto * np.cos((2 * k + 1) * (np.pi / (2 * subbands)) * (
np.arange(taps + 1) - (taps / 2)) - (-1)**k * np.pi / 4))
# convert to tensor
self.analysis_filter = paddle.to_tensor(
h_analysis, dtype="float32").unsqueeze(1)
self.synthesis_filter = paddle.to_tensor(
h_synthesis, dtype="float32").unsqueeze(0)
# filter for downsampling & upsampling
updown_filter = paddle.zeros(
(subbands, subbands, subbands), dtype="float32")
for k in range(subbands):
updown_filter[k, k, 0] = 1.0
self.updown_filter = updown_filter
self.subbands = subbands
# keep padding info
self.pad_fn = paddle.nn.Pad1D(taps // 2, mode='constant', value=0.0)
def analysis(self, x):
"""Analysis with PQMF.
Parameters
----------
x : Tensor
Input tensor (B, 1, T).
Returns
----------
Tensor
Output tensor (B, subbands, T // subbands).
"""
x = F.conv1d(self.pad_fn(x), self.analysis_filter)
return F.conv1d(x, self.updown_filter, stride=self.subbands)
def synthesis(self, x):
"""Synthesis with PQMF.
Parameters
----------
x : Tensor
Input tensor (B, subbands, T // subbands).
Returns
----------
Tensor
Output tensor (B, 1, T).
"""
x = F.conv1d_transpose(
x, self.updown_filter * self.subbands, stride=self.subbands)
return F.conv1d(self.pad_fn(x), self.synthesis_filter)
parakeet/modules/residual_stack.py 0 → 100644
浏览文件 @ 15beee6e
# 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.
"""Residual stack module in MelGAN."""
from typing import Any
from typing import Dict
from paddle import nn
from parakeet.modules.causal_conv import CausalConv1D
class ResidualStack(nn.Layer):
"""Residual stack module introduced in MelGAN."""
def __init__(
self,
kernel_size: int=3,
channels: int=32,
dilation: int=1,
bias: bool=True,
nonlinear_activation: str="LeakyReLU",
nonlinear_activation_params: Dict[str, Any]={"negative_slope": 0.2},
pad: str="Pad1D",
pad_params: Dict[str, Any]={"mode": "reflect"},
use_causal_conv: bool=False, ):
"""Initialize ResidualStack module.
Parameters
----------
kernel_size : int
Kernel size of dilation convolution layer.
channels : int
Number of channels of convolution layers.
dilation : int
Dilation factor.
bias : bool
Whether to add bias parameter in convolution layers.
nonlinear_activation : str
Activation function module name.
nonlinear_activation_params : Dict[str,Any]
Hyperparameters for activation function.
pad : str
Padding function module name before dilated convolution layer.
pad_params : Dict[str, Any]
Hyperparameters for padding function.
use_causal_conv : bool
Whether to use causal convolution.
"""
super().__init__()
# defile residual stack part
if not use_causal_conv:
assert (kernel_size - 1
) % 2 == 0, "Not support even number kernel size."
self.stack = nn.Sequential(
getattr(nn, nonlinear_activation)(
**nonlinear_activation_params),
getattr(nn, pad)((kernel_size - 1) // 2 * dilation,
**pad_params),
nn.Conv1D(
channels,
channels,
kernel_size,
dilation=dilation,
bias_attr=bias),
getattr(nn, nonlinear_activation)(
**nonlinear_activation_params),
nn.Conv1D(channels, channels, 1, bias_attr=bias), )
else:
self.stack = nn.Sequential(
getattr(nn, nonlinear_activation)(
**nonlinear_activation_params),
CausalConv1D(
channels,
channels,
kernel_size,
dilation=dilation,
bias=bias,
pad=pad,
pad_params=pad_params, ),
getattr(nn, nonlinear_activation)(
**nonlinear_activation_params),
nn.Conv1D(channels, channels, 1, bias_attr=bias), )
# defile extra layer for skip connection
self.skip_layer = nn.Conv1D(channels, channels, 1, bias_attr=bias)
def forward(self, c):
"""Calculate forward propagation.
Parameters
----------
c : Tensor
Input tensor (B, channels, T).
Returns
----------
Tensor
Output tensor (B, chennels, T).
"""
return self.stack(c) + self.skip_layer(c)
parakeet/modules/stft_loss.py
浏览文件 @ 15beee6e
......@@ -51,7 +51,7 @@ def stft(x,
# calculate window
window = signal.get_window(window, win_length, fftbins=True)
window = paddle.to_tensor(window)
x_stft = paddle.tensor.signal.stft(
x_stft = paddle.signal.stft(
x,
fft_size,
hop_length,
......
parakeet/training/extensions/visualizer.py
浏览文件 @ 15beee6e
......@@ -11,6 +11,8 @@
# 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 visualdl import LogWriter
from parakeet.training import extension
from parakeet.training.trainer import Trainer
......@@ -26,8 +28,8 @@ class VisualDL(extension.Extension):
default_name = 'visualdl'
priority = extension.PRIORITY_READER
def __init__(self, writer):
self.writer = writer
def __init__(self, logdir):
self.writer = LogWriter(str(logdir))
def __call__(self, trainer: Trainer):
for k, v in trainer.observation.items():
......
text_processing/.gitignore 0 → 100644
浏览文件 @ 15beee6e
data
glove
.pyc
checkpoints
epoch
__pycache__
glove.840B.300d.zip
text_processing/README.md 0 → 100644
浏览文件 @ 15beee6e
# 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
text_processing/examples/punctuation_restoration/chinese/README.md 0 → 100644
浏览文件 @ 15beee6e
# 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 |
text_processing/examples/punctuation_restoration/chinese/conf/blstm.yaml 0 → 100644
浏览文件 @ 15beee6e
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
text_processing/examples/punctuation_restoration/chinese/conf/data_conf/chinese.yaml 0 → 100644
浏览文件 @ 15beee6e
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
\ No newline at end of file
text_processing/examples/punctuation_restoration/chinese/conf/train_conf/bertBLSTM_zh.yaml 0 → 100644
浏览文件 @ 15beee6e
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
text_processing/examples/punctuation_restoration/chinese/conf/train_conf/bertLinear_zh.yaml 0 → 100644
浏览文件 @ 15beee6e
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
text_processing/examples/punctuation_restoration/chinese/local/avg.sh 0 → 100644
浏览文件 @ 15beee6e
#! /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
\ No newline at end of file
text_processing/examples/punctuation_restoration/chinese/local/data.sh 0 → 100644
浏览文件 @ 15beee6e
#!/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
text_processing/examples/punctuation_restoration/chinese/local/test.sh 0 → 100644
浏览文件 @ 15beee6e
#!/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
text_processing/examples/punctuation_restoration/chinese/local/train.sh 0 → 100644
浏览文件 @ 15beee6e
#!/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 [ $? -ne 0 ]; then
echo "Failed in training!"
exit 1
fi
exit 0
text_processing/examples/punctuation_restoration/chinese/path.sh 0 → 100644
浏览文件 @ 15beee6e
export MAIN_ROOT=${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
text_processing/examples/punctuation_restoration/chinese/run.sh 0 → 100644
浏览文件 @ 15beee6e
#!/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
text_processing/examples/punctuation_restoration/english/README.md 0 → 100644
浏览文件 @ 15beee6e
# 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 |
text_processing/examples/punctuation_restoration/english/conf/data_conf/english.yaml 0 → 100644
浏览文件 @ 15beee6e
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
\ No newline at end of file
text_processing/examples/punctuation_restoration/english/conf/train_conf/bertBLSTM_base_en.yaml 0 → 100644
浏览文件 @ 15beee6e
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
text_processing/examples/punctuation_restoration/english/conf/train_conf/bertLinear_en.yaml 0 → 100644
浏览文件 @ 15beee6e
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
text_processing/examples/punctuation_restoration/english/local/avg.sh 0 → 100644
浏览文件 @ 15beee6e
#! /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
\ No newline at end of file
text_processing/examples/punctuation_restoration/english/local/data.sh 0 → 100644
浏览文件 @ 15beee6e
#!/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
text_processing/examples/punctuation_restoration/english/local/test.sh 0 → 100644
浏览文件 @ 15beee6e
#!/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
text_processing/examples/punctuation_restoration/english/local/train.sh 0 → 100644
浏览文件 @ 15beee6e
#!/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 [ $? -ne 0 ]; then
echo "Failed in training!"
exit 1
fi
exit 0
text_processing/examples/punctuation_restoration/english/path.sh 0 → 100644
浏览文件 @ 15beee6e
export MAIN_ROOT=${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
text_processing/examples/punctuation_restoration/english/run.sh 0 → 100644
浏览文件 @ 15beee6e
#!/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
text_processing/requirements.txt 0 → 100644
浏览文件 @ 15beee6e
numpy
pyyaml
tensorboardX
tqdm
ujson
yacs
text_processing/speechtask/punctuation_restoration/bin/avg_model.py 0 → 100644
浏览文件 @ 15beee6e
#!/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)
tools/Makefile
浏览文件 @ 15beee6e
此差异已折叠。
tools/extras/README.md
浏览文件 @ 15beee6e
此差异已折叠。
tools/extras/install_soundfile.sh 0 → 100755
浏览文件 @ 15beee6e
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或