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

unify api to 1.6 version and fix some problems

Browse Source
pull/369/head
lfchener 5 years ago
parent d74f4ff3f5
commit d89c3a48a7
14 changed files with 184 additions and 110 deletions
Show Stats Download Patch File Download Diff File

4
README.md
Unescape View File

@ -25,7 +25,7 @@ To avoid the trouble of environment setup, [running in Docker container](#runnin
### Prerequisites
- Python 2.7 only supported
- PaddlePaddle the latest version (please refer to the [Installation Guide](https://www.paddlepaddle.org.cn/documentation/docs/en/1.5/beginners_guide/install/index_en.html))
- PaddlePaddle 1.6 version (Coming soon ...)
### Setup
- Make sure these libraries or tools installed: `pkg-config`, `flac`, `ogg`, `vorbis`, `boost` and `swig`, e.g. installing them via `apt-get`:
@ -183,7 +183,7 @@ python tools/build_vocab.py --help
```
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python train.py \
--init_from_pretrain_model CHECKPOINT_PATH_TO_RESUME_FROM
--init_from_pretrained_model CHECKPOINT_PATH_TO_RESUME_FROM
```
For more help on arguments:

22
README_cn.md
Unescape View File

@ -1,6 +1,6 @@
# 语音识别: DeepSpeech2
*语音识别: DeepSpeech2*是一个采用[PaddlePaddle](https://github.com/PaddlePaddle/Paddle)平台的端到端自动语音识别（ASR）引擎的开源项目，具体原理参考这篇论文[Baidu's Deep Speech 2 paper](http://proceedings.mlr.press/v48/amodei16.pdf)。
*DeepSpeech2*是一个采用[PaddlePaddle](https://github.com/PaddlePaddle/Paddle)平台的端到端自动语音识别（ASR）引擎的开源项目，具体原理参考这篇论文[Baidu's Deep Speech 2 paper](http://proceedings.mlr.press/v48/amodei16.pdf)。
我们的愿景是为语音识别在工业应用和学术研究上，提供易于使用、高效和可扩展的工具，包括训练，推理，测试模块，以及 demo 部署。同时，我们还将发布一些预训练好的英语和普通话模型。
## 目录
@ -8,7 +8,7 @@
- [开始](#开始)
- [数据准备](#数据准备)
- [训练模型](#训练模型)
- [数据增强管道](#数据增强管道)
- [数据增强流水线](#数据增强流水线)
- [推断和评价](#推断和评价)
- [在 Docker 容器上运行](#在Docker容器上运行)
- [超参数调整](#超参数调整)
@ -23,7 +23,7 @@
### 前提
- 只支持Python 2.7
- PaddlePaddle最新版本(请参考[安装指南](https://www.paddlepaddle.org.cn/start))
- PaddlePaddle 1.6 版本（即将发布）
### 安装
- 请确保以下库或工具已安装完毕：`pkg-config`, `flac`, `ogg`, `vorbis`, `boost` 和 `swig`, 如可以通过`apt-get`安装：
@ -57,11 +57,11 @@ sh setup.sh
`./examples`里的一些 shell 脚本将帮助我们在一些公开数据集(比如：[LibriSpeech](http://www.openslr.org/12/), [Aishell](http://www.openslr.org/33)) 进行快速尝试，包括了数据准备，模型训练，案例推断和模型评价。阅读这些例子将帮助你理解如何使用你的数据集训练模型。
`./examples`目录中的一些脚本配置使用了8个GPU。如果你没有8个可用的GPU，请修改`CUDA_VISIBLE_DEVICES`。如果你没有可用的GPU，请设置`--use_gpu`为False，这样程序会用CPU代替GPU。另外如果发生内存不足的问题，减小`--batch_size`即可。
`./examples`目录中的一些脚本配置使用了 8 个 GPU。如果你没有 8 个可用的 GPU，请修改环境变量`CUDA_VISIBLE_DEVICES`。如果你没有可用的 GPU，请设置`--use_gpu`为 False，这样程序会用 CPU 代替 GPU。另外如果发生内存不足的问题，减小`--batch_size`即可。
让我们先看看[LibriSpeech dataset](http://www.openslr.org/12/)小样本集的例子。
- 进到目录
- 进入目录
```bash
cd examples/tiny
@ -105,21 +105,21 @@ sh setup.sh
sh run_test_golden.sh
```
更多细节会在接下来的章节中阐述。祝你在*语音识别: DeepSpeech2*ASR引擎学习中过得愉快！
更多细节会在接下来的章节中阐述。祝你在*DeepSpeech2*ASR引擎学习中过得愉快！
## 数据准备
### 生成Manifest
*语音识别: DeepSpeech2*接受文本**manifest**文件作为数据接口。manifest文件包含了一系列语音数据，其中每一行代表一个[JSON](http://www.json.org/)格式的音频元数据（比如文件路径，描述，时长）。具体格式如下：
*DeepSpeech2*接受文本**manifest**文件作为数据接口。manifest 文件包含了一系列语音数据，其中每一行代表一个[JSON](http://www.json.org/)格式的音频元数据（比如文件路径，描述，时长）。具体格式如下：
```
{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0001.flac", "duration": 3.275, "text": "stuff it into you his belly counselled him"}
{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0007.flac", "duration": 4.275, "text": "a cold lucid indifference reigned in his soul"}
```
如果你要使用自定义数据，你只需要按照以上格式生成自己的manifest文件即可。给定manifest文件，训练，推断以及其他所有模块都能够访问到音频数据以及对应的时长和标签数据。
如果你要使用自定义数据，你只需要按照以上格式生成自己的 manifest 文件即可。给定 manifest 文件，训练、推断以及其它所有模块都能够访问到音频数据以及对应的时长和标签数据。
关于如何生成 manifest 文件，请参考`data/librispeech/librispeech.py`。该脚本将会下载 LibriSpeech 数据集并生成 manifest 文件。
@ -181,7 +181,7 @@ python tools/build_vocab.py --help
```
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python train.py \
--init_from_pretrain_model CHECKPOINT_PATH_TO_RESUME_FROM
--init_from_pretrained_model CHECKPOINT_PATH_TO_RESUME_FROM
```
获得更多帮助：
@ -192,7 +192,7 @@ python train.py --help
或参考 `example/librispeech/run_train.sh`.
## 数据增强管道
## 数据增强流水线
数据增强是用来提升深度学习性能的非常有效的技术。我们通过在原始音频中添加小的随机扰动（标签不变转换）获得新音频来增强我们的语音数据。你不必自己合成，因为数据增强已经嵌入到数据生成器中并且能够即时完成，在训练模型的每个epoch中随机合成音频。
@ -319,7 +319,7 @@ python test.py --help
`tools/tune.py`会进行2维网格查找超参数$\alpha$和$\beta$。你必须提供$\alpha$和$\beta$的范围，以及尝试的次数。
- 带GPU版的调整：
- GPU 版的调整：
```bash
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \

2
data_utils/data.py
Unescape View File

@ -57,7 +57,7 @@ class DataGenerator(object):
converting to index sequence.
:type keep_transcription_text: bool
:param place: The place to run the program.
:type place: CPU or GPU
:type place: CPUPlace or CUDAPlace
:param is_training: If set to True, generate text data for training,
otherwise, generate text data for infer.
:type is_training: bool

2
deploy/demo_server.py
Unescape View File

@ -162,7 +162,7 @@ def start_server():
num_rnn_layers=args.num_rnn_layers,
rnn_layer_size=args.rnn_layer_size,
use_gru=args.use_gru,
init_from_pretrain_model=args.model_path,
init_from_pretrained_model=args.model_path,
place=place,
share_rnn_weights=args.share_rnn_weights)

2
examples/aishell/run_train.sh
Unescape View File

@ -3,7 +3,7 @@
cd ../.. > /dev/null
# train model
# if you wish to resume from an exists model, uncomment --init_from_pretrain_model
# if you wish to resume from an exists model, uncomment --init_from_pretrained_model
export FLAGS_sync_nccl_allreduce=0
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -u train.py \

2
examples/librispeech/run_train.sh
Unescape View File

@ -3,7 +3,7 @@
cd ../.. > /dev/null
# train model
# if you wish to resume from an exists model, uncomment --init_from_pretrain_model
# if you wish to resume from an exists model, uncomment --init_from_pretrained_model
export FLAGS_sync_nccl_allreduce=0
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \

2
examples/tiny/run_train.sh
Unescape View File

@ -3,7 +3,7 @@
cd ../.. > /dev/null
# train model
# if you wish to resume from an exists model, uncomment --init_from_pretrain_model
# if you wish to resume from an exists model, uncomment --init_from_pretrained_model
export FLAGS_sync_nccl_allreduce=0
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -u train.py \

9
infer.py
Unescape View File

@ -12,6 +12,7 @@ import functools
import paddle.fluid as fluid
from data_utils.data import DataGenerator
from model_utils.model import DeepSpeech2Model
from model_utils.model_check import check_cuda, check_version
from utils.error_rate import wer, cer
from utils.utility import add_arguments, print_arguments
@ -66,6 +67,12 @@ args = parser.parse_args()
def infer():
"""Inference for DeepSpeech2."""
# check if set use_gpu=True in paddlepaddle cpu version
check_cuda(args.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
@ -94,7 +101,7 @@ def infer():
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrain_model=args.model_path)
init_from_pretrained_model=args.model_path)
# decoders only accept string encoded in utf-8
vocab_list = [chars.encode("utf-8") for chars in data_generator.vocab_list]

76
model_utils/model.py
Unescape View File

@ -44,7 +44,7 @@ class DeepSpeech2Model(object):
for GRU, weight sharing is not supported.
:type share_rnn_weights: bool
:param place: Program running place.
:type place: CPU or GPU
:type place: CPUPlace or CUDAPlace
:param init_from_pretrained_model: Pretrained model path. If None, will train
from stratch.
:type init_from_pretrained_model: string|None
@ -60,7 +60,7 @@ class DeepSpeech2Model(object):
use_gru=False,
share_rnn_weights=True,
place=fluid.CPUPlace(),
init_from_pretrain_model=None,
init_from_pretrained_model=None,
output_model_dir=None):
self._vocab_size = vocab_size
self._num_conv_layers = num_conv_layers
@ -69,7 +69,7 @@ class DeepSpeech2Model(object):
self._use_gru = use_gru
self._share_rnn_weights = share_rnn_weights
self._place = place
self._init_from_pretrain_model = init_from_pretrain_model
self._init_from_pretrained_model = init_from_pretrained_model
self._output_model_dir = output_model_dir
self._ext_scorer = None
self.logger = logging.getLogger("")
@ -90,13 +90,14 @@ class DeepSpeech2Model(object):
if not is_infer:
input_fields = {
'names': ['audio_data', 'text_data', 'seq_len_data', 'masks'],
'shapes': [[-1, 161, 161], [-1, 1], [-1, 1], [-1, 32, 81, 1]],
'shapes':
[[None, 161, None], [None, 1], [None, 1], [None, 32, 81, None]],
'dtypes': ['float32', 'int32', 'int64', 'float32'],
'lod_levels': [0, 1, 0, 0]
}
inputs = [
fluid.layers.data(
fluid.data(
name=input_fields['names'][i],
shape=input_fields['shapes'][i],
dtype=input_fields['dtypes'][i],
@ -104,7 +105,7 @@ class DeepSpeech2Model(object):
for i in range(len(input_fields['names']))
]
reader = fluid.io.PyReader(
reader = fluid.io.DataLoader.from_generator(
feed_list=inputs,
capacity=64,
iterable=False,
@ -112,16 +113,19 @@ class DeepSpeech2Model(object):
(audio_data, text_data, seq_len_data, masks) = inputs
else:
audio_data = fluid.layers.data(
audio_data = fluid.data(
name='audio_data',
shape=[-1, 161, 161],
shape=[None, 161, None],
dtype='float32',
lod_level=0)
seq_len_data = fluid.layers.data(
name='seq_len_data', shape=[-1, 1], dtype='int64', lod_level=0)
masks = fluid.layers.data(
seq_len_data = fluid.data(
name='seq_len_data',
shape=[None, 1],
dtype='int64',
lod_level=0)
masks = fluid.data(
name='masks',
shape=[-1, 32, 81, 1],
shape=[None, 32, 81, None],
dtype='float32',
lod_level=0)
text_data = None
@ -141,26 +145,26 @@ class DeepSpeech2Model(object):
share_rnn_weights=self._share_rnn_weights)
return reader, log_probs, loss
def init_from_pretrain_model(self, exe, program):
def init_from_pretrained_model(self, exe, program):
'''Init params from pretrain model. '''
assert isinstance(self._init_from_pretrain_model, str)
assert isinstance(self._init_from_pretrained_model, str)
if not os.path.exists(self._init_from_pretrain_model):
print(self._init_from_pretrain_model)
if not os.path.exists(self._init_from_pretrained_model):
print(self._init_from_pretrained_model)
raise Warning("The pretrained params do not exist.")
return False
fluid.io.load_params(
exe,
self._init_from_pretrain_model,
self._init_from_pretrained_model,
main_program=program,
filename="params.pdparams")
print("finish initing model from pretrained params from %s" %
(self._init_from_pretrain_model))
(self._init_from_pretrained_model))
pre_epoch = 0
dir_name = self._init_from_pretrain_model.split('_')
dir_name = self._init_from_pretrained_model.split('_')
if len(dir_name) >= 2 and dir_name[-2].endswith('epoch') and dir_name[
-1].isdigit():
pre_epoch = int(dir_name[-1])
@ -186,7 +190,7 @@ class DeepSpeech2Model(object):
return True
def test(self, exe, dev_batch_reader, test_program, test_pyreader,
def test(self, exe, dev_batch_reader, test_program, test_reader,
fetch_list):
'''Test the model.
@ -196,14 +200,14 @@ class DeepSpeech2Model(object):
:type dev_batch_reader: read generator
:param test_program: The program of test.
:type test_program: Program
:param test_pyreader: Pyreader of test.
:type test_pyreader: Pyreader
:param test_reader: Reader of test.
:type test_reader: Reader
:param fetch_list: Fetch list.
:type fetch_list: list
:return: An output unnormalized log probability.
:rtype: array
'''
test_pyreader.start()
test_reader.start()
epoch_loss = []
while True:
try:
@ -214,7 +218,7 @@ class DeepSpeech2Model(object):
epoch_loss.extend(np.array(each_loss[0]))
except fluid.core.EOFException:
test_pyreader.reset()
test_reader.reset()
break
return np.mean(np.array(epoch_loss))
@ -274,7 +278,7 @@ class DeepSpeech2Model(object):
startup_prog = fluid.Program()
with fluid.program_guard(train_program, startup_prog):
with fluid.unique_name.guard():
train_pyreader, log_probs, ctc_loss = self.create_network()
train_reader, log_probs, ctc_loss = self.create_network()
# prepare optimizer
optimizer = fluid.optimizer.AdamOptimizer(
learning_rate=fluid.layers.exponential_decay(
@ -290,7 +294,7 @@ class DeepSpeech2Model(object):
test_prog = fluid.Program()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
test_pyreader, _, ctc_loss = self.create_network()
test_reader, _, ctc_loss = self.create_network()
test_prog = test_prog.clone(for_test=True)
@ -299,8 +303,8 @@ class DeepSpeech2Model(object):
# init from some pretrain models, to better solve the current task
pre_epoch = 0
if self._init_from_pretrain_model:
pre_epoch = self.init_from_pretrain_model(exe, train_program)
if self._init_from_pretrained_model:
pre_epoch = self.init_from_pretrained_model(exe, train_program)
build_strategy = compiler.BuildStrategy()
exec_strategy = fluid.ExecutionStrategy()
@ -312,12 +316,12 @@ class DeepSpeech2Model(object):
build_strategy=build_strategy,
exec_strategy=exec_strategy)
train_pyreader.decorate_batch_generator(train_batch_reader)
test_pyreader.decorate_batch_generator(dev_batch_reader)
train_reader.set_batch_generator(train_batch_reader)
test_reader.set_batch_generator(dev_batch_reader)
# run train
for epoch_id in range(num_epoch):
train_pyreader.start()
train_reader.start()
epoch_loss = []
time_begin = time.time()
batch_id = 0
@ -346,7 +350,7 @@ class DeepSpeech2Model(object):
batch_id = batch_id + 1
except fluid.core.EOFException:
train_pyreader.reset()
train_reader.reset()
break
time_end = time.time()
used_time = time_end - time_begin
@ -359,7 +363,7 @@ class DeepSpeech2Model(object):
exe,
dev_batch_reader=dev_batch_reader,
test_program=test_prog,
test_pyreader=test_pyreader,
test_reader=test_reader,
fetch_list=[ctc_loss])
print(
"--------Time: %f sec, epoch: %d, train loss: %f, test loss: %f"
@ -402,10 +406,10 @@ class DeepSpeech2Model(object):
exe = fluid.Executor(self._place)
exe.run(startup_prog)
# init param from pretrain_model
if not self._init_from_pretrain_model:
# init param from pretrained_model
if not self._init_from_pretrained_model:
exit("No pretrain model file path!")
self.init_from_pretrain_model(exe, infer_program)
self.init_from_pretrained_model(exe, infer_program)
infer_results = []
time_begin = time.time()

49
model_utils/model_check.py
Unescape View File

@ -0,0 +1,49 @@
# Copyright (c) 2019 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 paddle
import paddle.fluid as fluid
def check_cuda(use_cuda, err = \
"\nYou can not set use_cuda = True in the model because you are using paddlepaddle-cpu.\n \
Please: 1. Install paddlepaddle-gpu to run your models on GPU or 2. Set use_cuda = False to run models on CPU.\n"
):
"""
Log error and exit when set use_gpu=true in paddlepaddle
cpu version.
"""
try:
if use_cuda == True and fluid.is_compiled_with_cuda() == False:
print(err)
sys.exit(1)
except Exception as e:
pass
def check_version():
"""
Log error and exit when the installed version of paddlepaddle is
not satisfied.
"""
err = "PaddlePaddle version 1.6 or higher is required, " \
"or a suitable develop version is satisfied as well. \n" \
"Please make sure the version is good with your code." \
try:
fluid.require_version('1.6.0')
except Exception as e:
print(err)
sys.exit(1)

10
model_utils/network.py
Unescape View File

@ -63,7 +63,7 @@ def simple_rnn(input, size, param_attr=None, bias_attr=None, is_reverse=False):
'''A simple rnn layer.
:param input: input layer.
:type input: Variable
:param size:Number of RNN cells.
:param size: Dimension of RNN cells.
:type size: int
:param param_attr: Parameter properties of hidden layer weights that
can be learned
@ -112,7 +112,7 @@ def bidirectional_simple_rnn_bn_layer(name, input, size, share_weights):
:type name: string
:param input: Input layer.
:type input: Variable
:param size: Number of RNN cells.
:param size: Dimension of RNN cells.
:type size: int
:param share_weights: Whether to share input-hidden weights between
forward and backward directional RNNs.
@ -206,7 +206,7 @@ def bidirectional_gru_bn_layer(name, input, size, act):
:type name: string
:param input: Input layer.
:type input: Variable
:param size: Number of GRU cells.
:param size: Dimension of GRU cells.
:type size: int
:param act: Activation type.
:type act: string
@ -317,7 +317,7 @@ def rnn_group(input, size, num_stacks, num_conv_layers, use_gru,
"""RNN group with stacked bidirectional simple RNN or GRU layers.
:param input: Input layer.
:type input: Variable
:param size: Number of RNN cells in each layer.
:param size: Dimension of RNN cells in each layer.
:type size: int
:param num_stacks: Number of stacked rnn layers.
:type num_stacks: int
@ -373,7 +373,7 @@ def deep_speech_v2_network(audio_data,
:type num_conv_layers: int
:param num_rnn_layers: Number of stacking RNN layers.
:type num_rnn_layers: int
:param rnn_size: RNN layer size (number of RNN cells).
:param rnn_size: RNN layer size (dimension of RNN cells).
:type rnn_size: int
:param use_gru: Use gru if set True. Use simple rnn if set False.
:type use_gru: bool

9
test.py
Unescape View File

@ -8,6 +8,7 @@ import functools
import paddle.fluid as fluid
from data_utils.data import DataGenerator
from model_utils.model import DeepSpeech2Model
from model_utils.model_check import check_cuda, check_version
from utils.error_rate import char_errors, word_errors
from utils.utility import add_arguments, print_arguments
@ -62,6 +63,12 @@ args = parser.parse_args()
def evaluate():
"""Evaluate on whole test data for DeepSpeech2."""
# check if set use_gpu=True in paddlepaddle cpu version
check_cuda(args.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
@ -89,7 +96,7 @@ def evaluate():
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrain_model=args.model_path)
init_from_pretrained_model=args.model_path)
# decoders only accept string encoded in utf-8
vocab_list = [chars.encode("utf-8") for chars in data_generator.vocab_list]

2
tools/tune.py
Unescape View File

@ -103,7 +103,7 @@ def tune():
rnn_layer_size=args.rnn_layer_size,
use_gru=args.use_gru,
place=place,
init_from_pretrain_model=args.model_path,
init_from_pretrained_model=args.model_path,
share_rnn_weights=args.share_rnn_weights)
# decoders only accept string encoded in utf-8

13
train.py
Unescape View File

@ -7,6 +7,7 @@ import argparse
import functools
import io
from model_utils.model import DeepSpeech2Model
from model_utils.model_check import check_cuda, check_version
from data_utils.data import DataGenerator
from utils.utility import add_arguments, print_arguments
@ -34,7 +35,7 @@ add_arg('use_gru', bool, False, "Use GRUs instead of simple RNNs.")
add_arg('is_local', bool, True, "Use pserver or not.")
add_arg('share_rnn_weights',bool, True, "Share input-hidden weights across "
"bi-directional RNNs. Not for GRU.")
add_arg('init_from_pretrain_model',str,
add_arg('init_from_pretrained_model',str,
None,
"If None, the training starts from scratch, "
"otherwise, it resumes from the pre-trained model.")
@ -71,6 +72,12 @@ args = parser.parse_args()
def train():
"""DeepSpeech2 training."""
# check if set use_gpu=True in paddlepaddle cpu version
check_cuda(args.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
@ -93,7 +100,7 @@ def train():
train_batch_reader = train_generator.batch_reader_creator(
manifest_path=args.train_manifest,
batch_size=args.batch_size,
sortagrad=args.use_sortagrad if args.init_from_pretrain_model is None else False,
sortagrad=args.use_sortagrad if args.init_from_pretrained_model is None else False,
shuffle_method=args.shuffle_method)
dev_batch_reader = dev_generator.batch_reader_creator(
manifest_path=args.dev_manifest,
@ -109,7 +116,7 @@ def train():
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrain_model=args.init_from_pretrain_model,
init_from_pretrained_model=args.init_from_pretrained_model,
output_model_dir=args.output_model_dir)
ds2_model.train(

Write Preview
Loading…
Cancel
Save