Skip to content

M
models

PaddlePaddle / models
大约 1 年 前同步成功

通知 222
Star 6828
Fork 2962
M
models
提交 555e0899 编写于 作者: C caoying03
浏览文件
操作

rewrite nmt without attention.

上级 06f272ab
隐藏空白更改
内联 并排
Showing with 294 addition and 30 deletion
nmt_without_attention/README.md
浏览文件 @ 555e0899
# 神经网络机器翻译模型
## 背景介绍
- PaddleBook中[机器翻译](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.cn.md)的相关章节中，已介绍了带注意力机制（Attention Mechanism）的 Encoder-Decoder 结构，本例则介绍的是不带注意力机制的 Encoder-Decoder 结构。关于注意力机制，读者可进一步参考 PaddleBook 和参考文献\[[3](#参考文献)]。
机器翻译利用计算机将源语言转换成目标语言的同义表达，是自然语言处理中重要的研究方向，有着广泛的应用需求，其实现方式也经历了不断地演化。传统机器翻译方法主要基于规则或统计模型，需要人为地指定翻译规则或设计语言特征，效果依赖于人对源语言与目标语言的理解程度。近些年来，深度学习的提出与迅速发展使得特征的自动学习成为可能。深度学习首先在图像识别和语音识别中取得成功，进而在机器翻译等自然语言处理领域中掀起了研究热潮。机器翻译中的深度学习模型直接学习源语言到目标语言的映射，大为减少了学习过程中人的介入，同时显著地提高了翻译质量。本例介绍在PaddlePaddle中如何利用循环神经网络（Recurrent Neural Network, RNN）构建一个端到端（End-to-End）的神经网络机器翻译（Neural Machine Translation, NMT）模型。
## 模型概览
......@@ -84,7 +86,6 @@ encoded_vector = paddle.networks.bidirectional_gru(
### 无注意力机制的解码器
PaddleBook中[机器翻译](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.cn.md)的相关章节中，已介绍了带注意力机制（Attention Mechanism）的 Encoder-Decoder 结构，本例则介绍的是不带注意力机制的 Encoder-Decoder 结构。关于注意力机制，读者可进一步参考 PaddleBook 和参考文献\[[3](#参考文献)]。
对于流行的RNN单元，PaddlePaddle 已有很好的实现均可直接调用。如果希望在 RNN 每一个时间步实现某些自定义操作，可使用 PaddlePaddle 中的`recurrent_layer_group`。首先，自定义单步逻辑函数，再利用函数 `recurrent_group()` 循环调用单步逻辑函数处理整个序列。本例中的无注意力机制的解码器便是使用`recurrent_layer_group`来实现，其中，单步逻辑函数`gru_decoder_without_attention()`相关代码如下：
......@@ -198,7 +199,7 @@ else:
**a) 由网络定义，解析网络结构，初始化模型参数**
```
```python
# initialize model
cost = seq2seq_net(source_dict_dim, target_dict_dim)
parameters = paddle.parameters.create(cost)
......@@ -206,7 +207,7 @@ parameters = paddle.parameters.create(cost)
**b) 设定训练过程中的优化策略、定义训练数据读取 `reader`**
```
```python
# define optimize method and trainer
optimizer = paddle.optimizer.RMSProp(
learning_rate=1e-3,
......@@ -223,7 +224,7 @@ wmt14_reader = paddle.batch(
**c) 定义事件句柄，打印训练中间结果、保存模型快照**
```
```python
# define event_handler callback
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
......@@ -242,7 +243,7 @@ def event_handler(event):
**d) 开始训练**
```
```python
# start to train
trainer.train(
reader=wmt14_reader, event_handler=event_handler, num_passes=2)
......@@ -250,13 +251,13 @@ trainer.train(
启动模型训练的十分简单，只需在命令行窗口中执行
```
python nmt_without_attention_v2.py --train
```bash
python train.py
```
输出样例为
```
```text
Pass 0, Batch 0, Cost 267.674663, {'classification_error_evaluator': 1.0}
.........
Pass 0, Batch 10, Cost 172.892294, {'classification_error_evaluator': 0.953895092010498}
......@@ -274,7 +275,7 @@ Pass 0, Batch 40, Cost 168.170543, {'classification_error_evaluator': 0.83481836
**a) 加载测试样本**
```
```python
# load data samples for generation
gen_creator = paddle.dataset.wmt14.gen(source_dict_dim)
gen_data = []
......@@ -284,7 +285,7 @@ for item in gen_creator():
**b) 初始化模型，执行`infer()`为每个输入样本生成`beam search`的翻译结果**
```
```python
beam_gen = seq2seq_net(source_dict_dim, target_dict_dim, True)
with gzip.open(init_models_path) as f:
parameters = paddle.parameters.Parameters.from_tar(f)
......@@ -298,7 +299,7 @@ beam_result = paddle.infer(
**c) 加载源语言和目标语言词典，将`id`序列表示的句子转化成原语言并输出结果**
```
```python
# get the dictionary
src_dict, trg_dict = paddle.dataset.wmt14.get_dict(source_dict_dim)
......@@ -323,19 +324,19 @@ for i in xrange(len(gen_data)):
模型测试的执行与模型训练类似，只需执行
```
python nmt_without_attention_v2.py --generate
```bash
python generate.py
```
则自动为测试数据生成了对应的翻译结果。
设置beam search的宽度为3，输入某个法文句子
```
```text
src: <s> Elles connaissent leur entreprise mieux que personne . <e>
```
其对应的英文翻译结果为
```
```text
prob = -3.754819: They know their business better than anyone . <e>
prob = -4.445528: They know their businesses better than anyone . <e>
prob = -5.026885: They know their business better than anybody . <e>
......
nmt_without_attention/generate.py 0 → 100644
浏览文件 @ 555e0899
#!/usr/bin/env python
import os
from network_conf import *
def infer_a_batch(inferer, test_batch, beam_size, src_dict, trg_dict):
beam_result = inferer.infer(input=test_batch, field=["prob", "id"])
# the delimited element of generated sequences is -1,
# the first element of each generated sequence is the sequence length
seq_list, seq = [], []
for w in beam_result[1]:
if w != -1:
seq.append(w)
else:
seq_list.append(" ".join([trg_dict.get(w) for w in seq[1:]]))
seq = []
prob = beam_result[0]
for i, sample in enumerate(test_batch):
print("src:", " ".join([src_dict.get(w) for w in sample[0]]), "\n")
for j in xrange(beam_size):
print("prob = %f:" % (prob[i][j]), seq_list[i * beam_size + j])
print("\n")
def generate(source_dict_dim, target_dict_dim, model_path, batch_size):
"""
Generating function for NMT
:param source_dict_dim: size of source dictionary
:type source_dict_dim: int
:param target_dict_dim: size of target dictionary
:type target_dict_dim: int
:param model_path: path for inital model
:type model_path: string
"""
assert os.path.exists(model_path), "trained model does not exist."
# step 1: prepare dictionary
src_dict, trg_dict = paddle.dataset.wmt14.get_dict(source_dict_dim)
beam_size = 5
# step 2: load the trained model
paddle.init(use_gpu=True, trainer_count=1)
with gzip.open(model_path) as f:
parameters = paddle.parameters.Parameters.from_tar(f)
beam_gen = seq2seq_net(
source_dict_dim,
target_dict_dim,
beam_size=beam_size,
max_length=100,
is_generating=True)
inferer = paddle.inference.Inference(
output_layer=beam_gen, parameters=parameters)
# step 3: iterating over the testing dataset
test_batch = []
for idx, item in enumerate(paddle.dataset.wmt14.gen(source_dict_dim)()):
test_batch.append([item[0]])
if len(test_batch) == batch_size:
infer_a_batch(inferer, test_batch, beam_size, src_dict, trg_dict)
test_batch = []
if len(test_batch):
infer_a_batch(inferer, test_batch, beam_size, src_dict, trg_dict)
test_batch = []
if __name__ == "__main__":
generate(
source_dict_dim=3000,
target_dict_dim=3000,
batch_size=5,
model_path="models/nmt_without_att_params_batch_00001.tar.gz")
nmt_without_attention/index.html
浏览文件 @ 555e0899
......@@ -43,6 +43,8 @@
# 神经网络机器翻译模型
## 背景介绍
- PaddleBook中[机器翻译](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.cn.md)的相关章节中，已介绍了带注意力机制（Attention Mechanism）的 Encoder-Decoder 结构，本例则介绍的是不带注意力机制的 Encoder-Decoder 结构。关于注意力机制，读者可进一步参考 PaddleBook 和参考文献\[[3](#参考文献)]。
机器翻译利用计算机将源语言转换成目标语言的同义表达，是自然语言处理中重要的研究方向，有着广泛的应用需求，其实现方式也经历了不断地演化。传统机器翻译方法主要基于规则或统计模型，需要人为地指定翻译规则或设计语言特征，效果依赖于人对源语言与目标语言的理解程度。近些年来，深度学习的提出与迅速发展使得特征的自动学习成为可能。深度学习首先在图像识别和语音识别中取得成功，进而在机器翻译等自然语言处理领域中掀起了研究热潮。机器翻译中的深度学习模型直接学习源语言到目标语言的映射，大为减少了学习过程中人的介入，同时显著地提高了翻译质量。本例介绍在PaddlePaddle中如何利用循环神经网络（Recurrent Neural Network, RNN）构建一个端到端（End-to-End）的神经网络机器翻译（Neural Machine Translation, NMT）模型。
## 模型概览
......@@ -126,7 +128,6 @@ encoded_vector = paddle.networks.bidirectional_gru(
### 无注意力机制的解码器
PaddleBook中[机器翻译](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.cn.md)的相关章节中，已介绍了带注意力机制（Attention Mechanism）的 Encoder-Decoder 结构，本例则介绍的是不带注意力机制的 Encoder-Decoder 结构。关于注意力机制，读者可进一步参考 PaddleBook 和参考文献\[[3](#参考文献)]。
对于流行的RNN单元，PaddlePaddle 已有很好的实现均可直接调用。如果希望在 RNN 每一个时间步实现某些自定义操作，可使用 PaddlePaddle 中的`recurrent_layer_group`。首先，自定义单步逻辑函数，再利用函数 `recurrent_group()` 循环调用单步逻辑函数处理整个序列。本例中的无注意力机制的解码器便是使用`recurrent_layer_group`来实现，其中，单步逻辑函数`gru_decoder_without_attention()`相关代码如下：
......@@ -240,7 +241,7 @@ else:
**a) 由网络定义，解析网络结构，初始化模型参数**
```
```python
# initialize model
cost = seq2seq_net(source_dict_dim, target_dict_dim)
parameters = paddle.parameters.create(cost)
......@@ -248,7 +249,7 @@ parameters = paddle.parameters.create(cost)
**b) 设定训练过程中的优化策略、定义训练数据读取 `reader`**
```
```python
# define optimize method and trainer
optimizer = paddle.optimizer.RMSProp(
learning_rate=1e-3,
......@@ -265,7 +266,7 @@ wmt14_reader = paddle.batch(
**c) 定义事件句柄，打印训练中间结果、保存模型快照**
```
```python
# define event_handler callback
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
......@@ -284,7 +285,7 @@ def event_handler(event):
**d) 开始训练**
```
```python
# start to train
trainer.train(
reader=wmt14_reader, event_handler=event_handler, num_passes=2)
......@@ -292,13 +293,13 @@ trainer.train(
启动模型训练的十分简单，只需在命令行窗口中执行
```
python nmt_without_attention_v2.py --train
```bash
python train.py
```
输出样例为
```
```text
Pass 0, Batch 0, Cost 267.674663, {'classification_error_evaluator': 1.0}
.........
Pass 0, Batch 10, Cost 172.892294, {'classification_error_evaluator': 0.953895092010498}
......@@ -316,7 +317,7 @@ Pass 0, Batch 40, Cost 168.170543, {'classification_error_evaluator': 0.83481836
**a) 加载测试样本**
```
```python
# load data samples for generation
gen_creator = paddle.dataset.wmt14.gen(source_dict_dim)
gen_data = []
......@@ -326,7 +327,7 @@ for item in gen_creator():
**b) 初始化模型，执行`infer()`为每个输入样本生成`beam search`的翻译结果**
```
```python
beam_gen = seq2seq_net(source_dict_dim, target_dict_dim, True)
with gzip.open(init_models_path) as f:
parameters = paddle.parameters.Parameters.from_tar(f)
......@@ -340,7 +341,7 @@ beam_result = paddle.infer(
**c) 加载源语言和目标语言词典，将`id`序列表示的句子转化成原语言并输出结果**
```
```python
# get the dictionary
src_dict, trg_dict = paddle.dataset.wmt14.get_dict(source_dict_dim)
......@@ -365,19 +366,19 @@ for i in xrange(len(gen_data)):
模型测试的执行与模型训练类似，只需执行
```
python nmt_without_attention_v2.py --generate
```bash
python generate.py
```
则自动为测试数据生成了对应的翻译结果。
设置beam search的宽度为3，输入某个法文句子
```
```text
src: <s> Elles connaissent leur entreprise mieux que personne . <e>
```
其对应的英文翻译结果为
```
```text
prob = -3.754819: They know their business better than anyone . <e>
prob = -4.445528: They know their businesses better than anyone . <e>
prob = -5.026885: They know their business better than anybody . <e>
......
nmt_without_attention/nmt_without_attention.py → nmt_without_attention/network_conf.py
浏览文件 @ 555e0899
#!/usr/bin/env python
import paddle.v2 as paddle
import sys
import gzip
import paddle.v2 as paddle
### Parameters
word_vector_dim = 620
latent_chain_dim = 1000
beam_size = 5
max_length = 50
def seq2seq_net(source_dict_dim, target_dict_dim, generating=False):
'''
def seq2seq_net(source_dict_dim,
target_dict_dim,
word_vector_dim=620,
rnn_hidden_size=1000,
beam_size=1,
max_length=50,
is_generating=False):
"""
Define the network structure of NMT, including encoder and decoder.
:param source_dict_dim: size of source dictionary
:type source_dict_dim : int
:param target_dict_dim: size of target dictionary
:type target_dict_dim: int
'''
:param word_vector_dim: size of source language word embedding
:type word_vector_dim: int
:param rnn_hidden_size: size of hidden state of encoder and decoder RNN
:type rnn_hidden_size: int
:param beam_size: expansion width in each step when generating
:type beam_size: int
:param max_length: max iteration number in generation
:type max_length: int
:param generating: whether to generate sequence or to train
:type generating: bool
"""
decoder_size = encoder_size = rnn_hidden_size
decoder_size = encoder_size = latent_chain_dim
#### Encoder
src_word_id = paddle.layer.data(
name='source_language_word',
name="source_language_word",
type=paddle.data_type.integer_value_sequence(source_dict_dim))
src_embedding = paddle.layer.embedding(
input=src_word_id, size=word_vector_dim)
# use bidirectional_gru
# use bidirectional_gru as the encoder
encoded_vector = paddle.networks.bidirectional_gru(
input=src_embedding,
size=encoder_size,
......@@ -41,79 +49,52 @@ def seq2seq_net(source_dict_dim, target_dict_dim, generating=False):
return_seq=True)
#### Decoder
encoder_last = paddle.layer.last_seq(input=encoded_vector)
encoder_last_projected = paddle.layer.mixed(
size=decoder_size,
act=paddle.activation.Tanh(),
input=paddle.layer.full_matrix_projection(input=encoder_last))
encoder_last_projected = paddle.layer.fc(
size=decoder_size, act=paddle.activation.Tanh(), input=encoder_last)
# gru step
def gru_decoder_without_attention(enc_vec, current_word):
'''
"""
Step function for gru decoder
:param enc_vec: encoded vector of source language
:type enc_vec: layer object
:param current_word: current input of decoder
:type current_word: layer object
'''
"""
decoder_mem = paddle.layer.memory(
name='gru_decoder',
name="gru_decoder",
size=decoder_size,
boot_layer=encoder_last_projected)
context = paddle.layer.last_seq(input=enc_vec)
decoder_inputs = paddle.layer.mixed(
size=decoder_size * 3,
input=[
paddle.layer.full_matrix_projection(input=context),
paddle.layer.full_matrix_projection(input=current_word)
])
decoder_inputs = paddle.layer.fc(
size=decoder_size * 3, input=[context, current_word])
gru_step = paddle.layer.gru_step(
name='gru_decoder',
name="gru_decoder",
act=paddle.activation.Tanh(),
gate_act=paddle.activation.Sigmoid(),
input=decoder_inputs,
output_mem=decoder_mem,
size=decoder_size)
out = paddle.layer.mixed(
out = paddle.layer.fc(
size=target_dict_dim,
bias_attr=True,
act=paddle.activation.Softmax(),
input=paddle.layer.full_matrix_projection(input=gru_step))
input=gru_step)
return out
decoder_group_name = "decoder_group"
group_input1 = paddle.layer.StaticInput(input=encoded_vector, is_seq=True)
group_input1 = paddle.layer.StaticInput(input=encoded_vector)
group_inputs = [group_input1]
if not generating:
trg_embedding = paddle.layer.embedding(
input=paddle.layer.data(
name='target_language_word',
type=paddle.data_type.integer_value_sequence(target_dict_dim)),
size=word_vector_dim,
param_attr=paddle.attr.ParamAttr(name='_target_language_embedding'))
group_inputs.append(trg_embedding)
decoder = paddle.layer.recurrent_group(
name=decoder_group_name,
step=gru_decoder_without_attention,
input=group_inputs)
lbl = paddle.layer.data(
name='target_language_next_word',
type=paddle.data_type.integer_value_sequence(target_dict_dim))
cost = paddle.layer.classification_cost(input=decoder, label=lbl)
return cost
else:
decoder_group_name = "decoder_group"
if is_generating:
trg_embedding = paddle.layer.GeneratedInput(
size=target_dict_dim,
embedding_name='_target_language_embedding',
embedding_name="_target_language_embedding",
embedding_size=word_vector_dim)
group_inputs.append(trg_embedding)
......@@ -127,137 +108,23 @@ def seq2seq_net(source_dict_dim, target_dict_dim, generating=False):
max_length=max_length)
return beam_gen
def train(source_dict_dim, target_dict_dim):
'''
Training function for NMT
:param source_dict_dim: size of source dictionary
:type source_dict_dim: int
:param target_dict_dim: size of target dictionary
:type target_dict_dim: int
'''
# initialize model
cost = seq2seq_net(source_dict_dim, target_dict_dim)
parameters = paddle.parameters.create(cost)
# define optimize method and trainer
optimizer = paddle.optimizer.RMSProp(
learning_rate=1e-3,
gradient_clipping_threshold=10.0,
regularization=paddle.optimizer.L2Regularization(rate=8e-4))
trainer = paddle.trainer.SGD(
cost=cost, parameters=parameters, update_equation=optimizer)
# define data reader
wmt14_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.wmt14.train(source_dict_dim), buf_size=8192),
batch_size=55)
# define event_handler callback
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0 and event.batch_id > 0:
with gzip.open('models/nmt_without_att_params_batch_%d.tar.gz' %
event.batch_id, 'w') as f:
parameters.to_tar(f)
if event.batch_id % 10 == 0:
print "\nPass %d, Batch %d, Cost %f, %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics)
else:
sys.stdout.write('.')
sys.stdout.flush()
# start to train
trainer.train(
reader=wmt14_reader, event_handler=event_handler, num_passes=2)
def generate(source_dict_dim, target_dict_dim, init_models_path):
'''
Generating function for NMT
:param source_dict_dim: size of source dictionary
:type source_dict_dim: int
:param target_dict_dim: size of target dictionary
:type target_dict_dim: int
:param init_models_path: path for inital model
:type init_models_path: string
'''
# load data samples for generation
gen_creator = paddle.dataset.wmt14.gen(source_dict_dim)
gen_data = []
for item in gen_creator():
gen_data.append((item[0], ))
beam_gen = seq2seq_net(source_dict_dim, target_dict_dim, True)
with gzip.open(init_models_path) as f:
parameters = paddle.parameters.Parameters.from_tar(f)
# prob is the prediction probabilities, and id is the prediction word.
beam_result = paddle.infer(
output_layer=beam_gen,
parameters=parameters,
input=gen_data,
field=['prob', 'id'])
# get the dictionary
src_dict, trg_dict = paddle.dataset.wmt14.get_dict(source_dict_dim)
# the delimited element of generated sequences is -1,
# the first element of each generated sequence is the sequence length
seq_list, seq = [], []
for w in beam_result[1]:
if w != -1:
seq.append(w)
else:
seq_list.append(' '.join([trg_dict.get(w) for w in seq[1:]]))
seq = []
prob = beam_result[0]
for i in xrange(len(gen_data)):
print "\n*******************************************************\n"
print "src:", ' '.join([src_dict.get(w) for w in gen_data[i][0]]), "\n"
for j in xrange(beam_size):
print "prob = %f:" % (prob[i][j]), seq_list[i * beam_size + j]
def usage_helper():
print "Please specify training/generating phase!"
print "Usage: python nmt_without_attention_v2.py --train/generate"
exit(1)
def main():
if not (len(sys.argv) == 2):
usage_helper()
if sys.argv[1] == '--train':
generating = False
elif sys.argv[1] == '--generate':
generating = True
else:
usage_helper()
# initialize paddle
paddle.init(use_gpu=False, trainer_count=1)
source_language_dict_dim = 30000
target_language_dict_dim = 30000
trg_embedding = paddle.layer.embedding(
input=paddle.layer.data(
name="target_language_word",
type=paddle.data_type.integer_value_sequence(target_dict_dim)),
size=word_vector_dim,
param_attr=paddle.attr.ParamAttr(name="_target_language_embedding"))
group_inputs.append(trg_embedding)
if generating:
# modify this path to speicify a trained model.
init_models_path = 'models/nmt_without_att_params_batch_1800.tar.gz'
if not os.path.exists(init_models_path):
print "trained model cannot be found."
exit(1)
generate(source_language_dict_dim, target_language_dict_dim,
init_models_path)
else:
if not os.path.exists('./models'):
os.system('mkdir ./models')
train(source_language_dict_dim, target_language_dict_dim)
decoder = paddle.layer.recurrent_group(
name=decoder_group_name,
step=gru_decoder_without_attention,
input=group_inputs)
lbl = paddle.layer.data(
name="target_language_next_word",
type=paddle.data_type.integer_value_sequence(target_dict_dim))
cost = paddle.layer.classification_cost(input=decoder, label=lbl)
if __name__ == '__main__':
main()
return cost
nmt_without_attention/train.py 0 → 100644
浏览文件 @ 555e0899
#!/usr/bin/env python
from network_conf import *
def train(source_dict_dim, target_dict_dim):
'''
Training function for NMT
:param source_dict_dim: size of source dictionary
:type source_dict_dim: int
:param target_dict_dim: size of target dictionary
:type target_dict_dim: int
'''
# initialize model
paddle.init(use_gpu=False, trainer_count=1)
cost = seq2seq_net(source_dict_dim, target_dict_dim)
parameters = paddle.parameters.create(cost)
# define optimize method and trainer
optimizer = paddle.optimizer.RMSProp(
learning_rate=1e-3,
gradient_clipping_threshold=10.0,
regularization=paddle.optimizer.L2Regularization(rate=8e-4))
trainer = paddle.trainer.SGD(
cost=cost, parameters=parameters, update_equation=optimizer)
# define data reader
wmt14_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.wmt14.train(source_dict_dim), buf_size=8192),
batch_size=8)
# define event_handler callback
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if not event.batch_id % 500 and event.batch_id:
with gzip.open("models/nmt_without_att_params_batch_%05d.tar.gz"
% event.batch_id, "w") as f:
parameters.to_tar(f)
if event.batch_id and not event.batch_id % 10:
print("\nPass %d, Batch %d, Cost %f, %s" %
(event.pass_id, event.batch_id, event.cost,
event.metrics))
else:
sys.stdout.write('.')
sys.stdout.flush()
# start to train
trainer.train(
reader=wmt14_reader, event_handler=event_handler, num_passes=2)
if __name__ == '__main__':
train(source_dict_dim=3000, target_dict_dim=3000)
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