Skip to content

B
book

PaddlePaddle / book

通知 16
Star 4
Fork 0
B
book
提交 3b5bd69a 编写于 作者: H helinwang 提交者： GitHub
浏览文件
操作

Merge pull request #125 from helinwang/word2vec_tutorial 

revise word2vec according to new code
上级 d3f760e3 9dfec6bc
隐藏空白更改
内联 并排
Showing with 209 addition and 342 deletion
word2vec/README.md
浏览文件 @ 3b5bd69a
......@@ -141,7 +141,7 @@ CBOW的好处是对上下文词语的分布在词向量上进行了平滑，去
## 数据准备
### 数据介绍与下载
### 数据介绍
本教程使用Penn Tree Bank (PTB)数据集。PTB数据集较小，训练速度快，应用于Mikolov的公开语言模型训练工具\[[2](#参考文献)\]中。其统计情况如下：
......@@ -165,109 +165,24 @@ CBOW的好处是对上下文词语的分布在词向量上进行了平滑，去
</table>
</p>
执行以下命令，可下载该数据集，并分别将训练数据和验证数据输入`train.list`和`test.list`文件中，供PaddlePaddle训练时使用。
```bash
./data/getdata.sh
```
### 提供数据给PaddlePaddle
1. 使用initializer函数进行dataprovider的初始化，包括字典的建立（build_dict函数中）和PaddlePaddle输入字段的格式定义。注意：这里N为n-gram模型中的`n`, 本章代码中，定义$N=5$, 表示在PaddlePaddle训练时，每条数据的前4个词用来预测第5个词。大家也可以根据自己的数据和需求自行调整N，但调整的同时要在模型配置文件中加入/减少相应输入字段。
```python
from paddle.trainer.PyDataProvider2 import *
import collections
import logging
import pdb
logging.basicConfig(
format='[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s', )
logger = logging.getLogger('paddle')
logger.setLevel(logging.INFO)
N = 5 # Ngram
cutoff = 50 # select words with frequency > cutoff to dictionary
def build_dict(ftrain, fdict):
sentences = []
with open(ftrain) as fin:
for line in fin:
line = ['<s>'] + line.strip().split() + ['<e>']
sentences += line
wordfreq = collections.Counter(sentences)
wordfreq = filter(lambda x: x[1] > cutoff, wordfreq.items())
dictionary = sorted(wordfreq, key = lambda x: (-x[1], x[0]))
words, _ = list(zip(*dictionary))
for word in words:
print >> fdict, word
word_idx = dict(zip(words, xrange(len(words))))
logger.info("Dictionary size=%s" %len(words))
return word_idx
def initializer(settings, srcText, dictfile, **xargs):
with open(dictfile, 'w') as fdict:
settings.dicts = build_dict(srcText, fdict)
input_types = []
for i in xrange(N):
input_types.append(integer_value(len(settings.dicts)))
settings.input_types = input_types
```
2. 使用process函数中将数据逐一提供给PaddlePaddle。具体来说，将每句话前面补上N-1个开始符号 `<s>`, 末尾补上一个结束符号 `<e>`，然后以N为窗口大小，从头到尾每次向右滑动窗口并生成一条数据。
```python
@provider(init_hook=initializer)
def process(settings, filename):
UNKID = settings.dicts['<unk>']
with open(filename) as fin:
for line in fin:
line = ['<s>']*(N-1) + line.strip().split() + ['<e>']
line = [settings.dicts.get(w, UNKID) for w in line]
for i in range(N, len(line) + 1):
yield line[i-N: i]
```
如"I have a dream" 一句提供了5条数据:
> `<s> <s> <s> <s> I` <br>
> `<s> <s> <s> I have` <br>
> `<s> <s> I have a` <br>
> `<s> I have a dream` <br>
> `I have a dream <e>` <br>
## 模型配置说明
### 数据定义
通过`define_py_data_sources2`函数从dataprovider中读入数据，其中args指定了训练文本(srcText)和词汇表(dictfile)。
```python
from paddle.trainer_config_helpers import *
import math
### 数据预处理
args = {'srcText': 'data/simple-examples/data/ptb.train.txt',
'dictfile': 'data/vocabulary.txt'}
define_py_data_sources2(
train_list="data/train.list",
test_list="data/test.list",
module="dataprovider",
obj="process",
args=args)
```
本章训练的是5-gram模型，表示在PaddlePaddle训练时，每条数据的前4个词用来预测第5个词。PaddlePaddle提供了对应PTB数据集的python包`paddle.dataset.imikolov`，自动做数据的下载与预处理，方便大家使用。
### 算法配置
预处理会把数据集中的每一句话前后加上开始符号`<s>`以及结束符号`<e>`。然后依据窗口大小（本教程中为5），从头到尾每次向右滑动窗口并生成一条数据。
在这里，我们指定了模型的训练参数, L2正则项系数、学习率和batch size。
如"I have a dream that one day" 一句提供了5条数据：
```python
settings(
batch_size=100, regularization=L2Regularization(8e-4), learning_rate=3e-3)
```text
<s> I have a dream
I have a dream that
have a dream that one
a dream that one day
dream that one day <e>
```
### 模型结构
## 编程实现
本配置的模型结构如下图所示：
......@@ -276,94 +191,132 @@ settings(
图5. 模型配置中的N-gram神经网络模型
</p>
1. 定义参数维度和和数据输入。
```python
dictsize = 1953 # 字典大小
embsize = 32 # 词向量维度
hiddensize = 256 # 隐层维度
firstword = data_layer(name = "firstw", size = dictsize)
secondword = data_layer(name = "secondw", size = dictsize)
thirdword = data_layer(name = "thirdw", size = dictsize)
fourthword = data_layer(name = "fourthw", size = dictsize)
nextword = data_layer(name = "fifthw", size = dictsize)
```
2. 将$w_t$之前的$n-1$个词 $w_{t-n+1},...w_{t-1}$，通过$|V|\times D$的矩阵映射到D维词向量（本例中取D=32）。
首先，加载所需要的包：
```python
import math
import paddle.v2 as paddle
```
然后，定义参数：
```python
embsize = 32 # 词向量维度
hiddensize = 256 # 隐层维度
N = 5 # 训练5-Gram
```
接着，定义网络结构：
- 将$w_t$之前的$n-1$个词 $w_{t-n+1},...w_{t-1}$，通过$|V|\times D$的矩阵映射到D维词向量（本例中取D=32）。
```python
def wordemb(inlayer):
wordemb = table_projection(
input = inlayer,
size = embsize,
param_attr=ParamAttr(name = "_proj",
initial_std=0.001, # 参数初始化标准差
l2_rate= 0,)) # 词向量不需要稀疏化，因此其l2_rate设为0
```python
def wordemb(inlayer):
wordemb = paddle.layer.table_projection(
input=inlayer,
size=embsize,
param_attr=paddle.attr.Param(
name="_proj",
initial_std=0.001,
learning_rate=1,
l2_rate=0, ))
return wordemb
```
Efirst = wordemb(firstword)
Esecond = wordemb(secondword)
Ethird = wordemb(thirdword)
Efourth = wordemb(fourthword)
```
3. 接着，将这n-1个词向量经过concat_layer连接成一个大向量作为历史文本特征。
```python
contextemb = concat_layer(input = [Efirst, Esecond, Ethird, Efourth])
```
4. 然后，将历史文本特征经过一个全连接得到文本隐层特征。
```python
hidden1 = fc_layer(
input = contextemb,
size = hiddensize,
act = SigmoidActivation(),
layer_attr = ExtraAttr(drop_rate=0.5),
bias_attr = ParamAttr(learning_rate = 2),
param_attr = ParamAttr(
initial_std = 1./math.sqrt(embsize*8),
learning_rate = 1))
```
5. 最后，将文本隐层特征，再经过一个全连接，映射成一个$|V|$维向量，同时通过softmax归一化得到这`|V|`个词的生成概率。
```python
# use context embedding to predict nextword
predictword = fc_layer(
input = hidden1,
size = dictsize,
bias_attr = ParamAttr(learning_rate = 2),
act = SoftmaxActivation())
```
6. 网络的损失函数为多分类交叉熵，可直接调用`classification_cost`函数。
```python
cost = classification_cost(
input = predictword,
label = nextword)
# network input and output
outputs(cost)
```
- 定义输入层接受的数据类型以及名字。
```python
def main():
paddle.init(use_gpu=False, trainer_count=1) # 初始化PaddlePaddle
word_dict = paddle.dataset.imikolov.build_dict()
dict_size = len(word_dict)
# 每个输入层都接受整形数据，这些数据的范围是[0, dict_size)
firstword = paddle.layer.data(
name="firstw", type=paddle.data_type.integer_value(dict_size))
secondword = paddle.layer.data(
name="secondw", type=paddle.data_type.integer_value(dict_size))
thirdword = paddle.layer.data(
name="thirdw", type=paddle.data_type.integer_value(dict_size))
fourthword = paddle.layer.data(
name="fourthw", type=paddle.data_type.integer_value(dict_size))
nextword = paddle.layer.data(
name="fifthw", type=paddle.data_type.integer_value(dict_size))
Efirst = wordemb(firstword)
Esecond = wordemb(secondword)
Ethird = wordemb(thirdword)
Efourth = wordemb(fourthword)
```
- 将这n-1个词向量经过concat_layer连接成一个大向量作为历史文本特征。
```python
contextemb = paddle.layer.concat(input=[Efirst, Esecond, Ethird, Efourth])
```
- 将历史文本特征经过一个全连接得到文本隐层特征。
```python
hidden1 = paddle.layer.fc(input=contextemb,
size=hiddensize,
act=paddle.activation.Sigmoid(),
layer_attr=paddle.attr.Extra(drop_rate=0.5),
bias_attr=paddle.attr.Param(learning_rate=2),
param_attr=paddle.attr.Param(
initial_std=1. / math.sqrt(embsize * 8),
learning_rate=1))
```
##训练模型
- 将文本隐层特征，再经过一个全连接，映射成一个$|V|$维向量，同时通过softmax归一化得到这`|V|`个词的生成概率。
```python
predictword = paddle.layer.fc(input=hidden1,
size=dict_size,
bias_attr=paddle.attr.Param(learning_rate=2),
act=paddle.activation.Softmax())
```
模型训练命令为`./train.sh`。脚本内容如下，其中指定了总共需要执行30个pass。
- 网络的损失函数为多分类交叉熵，可直接调用`classification_cost`函数。
```bash
paddle train \
--config ngram.py \
--use_gpu=1 \
--dot_period=100 \
--log_period=3000 \
--test_period=0 \
--save_dir=model \
--num_passes=30
```python
cost = paddle.layer.classification_cost(input=predictword, label=nextword)
```
然后，指定训练相关的参数：
- 训练方法（optimizer)： 代表训练过程在更新权重时采用动量优化器，本教程使用Adam优化器。
- 训练速度（learning_rate）： 迭代的速度，与网络的训练收敛速度有关系。
- 正则化（regularization）： 是防止网络过拟合的一种手段，此处采用L2正则化。
```python
parameters = paddle.parameters.create(cost)
adam_optimizer = paddle.optimizer.Adam(
learning_rate=3e-3,
regularization=paddle.optimizer.L2Regularization(8e-4))
trainer = paddle.trainer.SGD(cost, parameters, adam_optimizer)
```
下一步，我们开始训练过程。`paddle.dataset.imikolov.train()`和`paddle.dataset.imikolov.test()`分别做训练和测试数据集。这两个函数各自返回一个reader——PaddlePaddle中的reader是一个Python函数，每次调用的时候返回一个Python generator。
`paddle.batch`的输入是一个reader，输出是一个batched reader —— 在PaddlePaddle里，一个reader每次yield一条训练数据，而一个batched reader每次yield一个minbatch。
```python
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
result = trainer.test(
paddle.batch(
paddle.dataset.imikolov.test(word_dict, N), 32))
print "Pass %d, Batch %d, Cost %f, %s, Testing metrics %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics,
result.metrics)
trainer.train(
paddle.batch(paddle.dataset.imikolov.train(word_dict, N), 32),
num_passes=30,
event_handler=event_handler)
```
一个pass的训练日志如下所示：
训练过程是完全自动的，event_handler里打印的日志类似如下所示：
```text
.............................
......
word2vec/data/getdata.sh 已删除 100755 → 0
浏览文件 @ d3f760e3
#!/bin/bash
set -e
wget http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz
tar -zxf simple-examples.tgz
echo `pwd`/simple-examples/data/ptb.train.txt > train.list
echo `pwd`/simple-examples/data/ptb.valid.txt > test.list
word2vec/dataprovider.py 已删除 100644 → 0
浏览文件 @ d3f760e3
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer.PyDataProvider2 import *
import collections
import logging
import pdb
logging.basicConfig(
format='[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s', )
logger = logging.getLogger('paddle')
logger.setLevel(logging.INFO)
N = 5 # Ngram
cutoff = 50 # select words with frequency > cutoff to dictionary
def build_dict(ftrain, fdict):
sentences = []
with open(ftrain) as fin:
for line in fin:
line = ['<s>'] + line.strip().split() + ['<e>']
sentences += line
wordfreq = collections.Counter(sentences)
wordfreq = filter(lambda x: x[1] > cutoff, wordfreq.items())
dictionary = sorted(wordfreq, key=lambda x: (-x[1], x[0]))
words, _ = list(zip(*dictionary))
for word in words:
print >> fdict, word
word_idx = dict(zip(words, xrange(len(words))))
logger.info("Dictionary size=%s" % len(words))
return word_idx
def initializer(settings, srcText, dictfile, **xargs):
with open(dictfile, 'w') as fdict:
settings.dicts = build_dict(srcText, fdict)
input_types = []
for i in xrange(N):
input_types.append(integer_value(len(settings.dicts)))
settings.input_types = input_types
@provider(init_hook=initializer)
def process(settings, filename):
UNKID = settings.dicts['<unk>']
with open(filename) as fin:
for line in fin:
line = ['<s>'] * (N - 1) + line.strip().split() + ['<e>']
line = [settings.dicts.get(w, UNKID) for w in line]
for i in range(N, len(line) + 1):
yield line[i - N:i]
word2vec/ngram.py 已删除 100644 → 0
浏览文件 @ d3f760e3
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer_config_helpers import *
import math
#################### Data Configure ####################
args = {
'srcText': 'data/simple-examples/data/ptb.train.txt',
'dictfile': 'data/vocabulary.txt'
}
define_py_data_sources2(
train_list="data/train.list",
test_list="data/test.list",
module="dataprovider",
obj="process",
args=args)
settings(
batch_size=100, regularization=L2Regularization(8e-4), learning_rate=3e-3)
dictsize = 1953
embsize = 32
hiddensize = 256
firstword = data_layer(name="firstw", size=dictsize)
secondword = data_layer(name="secondw", size=dictsize)
thirdword = data_layer(name="thirdw", size=dictsize)
fourthword = data_layer(name="fourthw", size=dictsize)
nextword = data_layer(name="fifthw", size=dictsize)
# construct word embedding for each datalayer
def wordemb(inlayer):
wordemb = table_projection(
input=inlayer,
size=embsize,
param_attr=ParamAttr(
name="_proj",
initial_std=0.001,
learning_rate=1,
l2_rate=0, ))
return wordemb
Efirst = wordemb(firstword)
Esecond = wordemb(secondword)
Ethird = wordemb(thirdword)
Efourth = wordemb(fourthword)
# concatentate Ngram embeddings into context embedding
contextemb = concat_layer(input=[Efirst, Esecond, Ethird, Efourth])
hidden1 = fc_layer(
input=contextemb,
size=hiddensize,
act=SigmoidActivation(),
layer_attr=ExtraAttr(drop_rate=0.5),
bias_attr=ParamAttr(learning_rate=2),
param_attr=ParamAttr(
initial_std=1. / math.sqrt(embsize * 8), learning_rate=1))
# use context embedding to predict nextword
predictword = fc_layer(
input=hidden1,
size=dictsize,
bias_attr=ParamAttr(learning_rate=2),
act=SoftmaxActivation())
cost = classification_cost(input=predictword, label=nextword)
# network input and output
outputs(cost)
word2vec/train.py 0 → 100644
浏览文件 @ 3b5bd69a
import math
import paddle.v2 as paddle
embsize = 32
hiddensize = 256
N = 5
def wordemb(inlayer):
wordemb = paddle.layer.table_projection(
input=inlayer,
size=embsize,
param_attr=paddle.attr.Param(
name="_proj",
initial_std=0.001,
learning_rate=1,
l2_rate=0, ))
return wordemb
def main():
paddle.init(use_gpu=False, trainer_count=1)
word_dict = paddle.dataset.imikolov.build_dict()
dict_size = len(word_dict)
firstword = paddle.layer.data(
name="firstw", type=paddle.data_type.integer_value(dict_size))
secondword = paddle.layer.data(
name="secondw", type=paddle.data_type.integer_value(dict_size))
thirdword = paddle.layer.data(
name="thirdw", type=paddle.data_type.integer_value(dict_size))
fourthword = paddle.layer.data(
name="fourthw", type=paddle.data_type.integer_value(dict_size))
nextword = paddle.layer.data(
name="fifthw", type=paddle.data_type.integer_value(dict_size))
Efirst = wordemb(firstword)
Esecond = wordemb(secondword)
Ethird = wordemb(thirdword)
Efourth = wordemb(fourthword)
contextemb = paddle.layer.concat(input=[Efirst, Esecond, Ethird, Efourth])
hidden1 = paddle.layer.fc(input=contextemb,
size=hiddensize,
act=paddle.activation.Sigmoid(),
layer_attr=paddle.attr.Extra(drop_rate=0.5),
bias_attr=paddle.attr.Param(learning_rate=2),
param_attr=paddle.attr.Param(
initial_std=1. / math.sqrt(embsize * 8),
learning_rate=1))
predictword = paddle.layer.fc(input=hidden1,
size=dict_size,
bias_attr=paddle.attr.Param(learning_rate=2),
act=paddle.activation.Softmax())
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
result = trainer.test(
paddle.batch(
paddle.dataset.imikolov.test(word_dict, N), 32))
print "Pass %d, Batch %d, Cost %f, %s, Testing metrics %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics,
result.metrics)
cost = paddle.layer.classification_cost(input=predictword, label=nextword)
parameters = paddle.parameters.create(cost)
adam_optimizer = paddle.optimizer.Adam(
learning_rate=3e-3,
regularization=paddle.optimizer.L2Regularization(8e-4))
trainer = paddle.trainer.SGD(cost, parameters, adam_optimizer)
trainer.train(
paddle.batch(paddle.dataset.imikolov.train(word_dict, N), 32),
num_passes=30,
event_handler=event_handler)
if __name__ == '__main__':
main()
word2vec/train.sh 已删除 100755 → 0
浏览文件 @ d3f760e3
#!/bin/bash
set -e
paddle train \
--config ngram.py \
--use_gpu=1 \
--dot_period=100 \
--log_period=3000 \
--test_period=0 \
--save_dir=model \
--num_passes=30
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