Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

add text classification models #24

Merged
merged 33 commits into from
May 17, 2017
Merged
Show file tree
Hide file tree
Changes from 29 commits
Commits
Show all changes
33 commits
Select commit Hold shift + click to select a range
100796f
add 'text_classification_cnn.py' and 'text_classification_dnn.py'
JiayiFeng May 2, 2017
dfc100a
Merge pull request #1 from Canpio/dev
JiayiFeng May 2, 2017
a651dc4
remove 'import sqlite3'
JiayiFeng May 2, 2017
31ef20f
Merge branch 'develop' of https://github.com/PaddlePaddle/models into…
JiayiFeng May 3, 2017
046d0a1
update readme file
JiayiFeng May 4, 2017
868d4cb
update readme file
JiayiFeng May 4, 2017
dd827a2
update README.md and add dnn_net.png
JiayiFeng May 4, 2017
20c2ec9
update README.md
JiayiFeng May 4, 2017
77abc3c
change dnn_net.png
JiayiFeng May 4, 2017
0619165
update dnn_net.png
JiayiFeng May 4, 2017
da0eb4e
add .gitignore
JiayiFeng May 4, 2017
5a4ad99
update README.md
JiayiFeng May 5, 2017
0a39b52
update README.md
JiayiFeng May 5, 2017
1b01f6a
update README.md
JiayiFeng May 5, 2017
78531d8
update README.md
JiayiFeng May 5, 2017
55dc1a3
add cnn_net.png and update README.md
JiayiFeng May 5, 2017
ac67d5a
add commits in cnn net
JiayiFeng May 5, 2017
947e70d
add comments in cnn net
JiayiFeng May 5, 2017
763615e
finish README.md
JiayiFeng May 9, 2017
304df71
finish README.md
JiayiFeng May 9, 2017
225979c
fix problems finded in review
JiayiFeng May 9, 2017
3a712e0
change README.md
JiayiFeng May 9, 2017
f18b572
add and test auc evaluator
JiayiFeng May 10, 2017
ae6a883
Merge pull request #2 from Canpio/dev
JiayiFeng May 10, 2017
18a4bd9
remove copyright statement
JiayiFeng May 10, 2017
344a9a4
add section of 'self-define data reader' into README.md
JiayiFeng May 12, 2017
5a59fcd
update README.md
JiayiFeng May 12, 2017
55a88ad
update README.md and fine-tune network hyper-params to avoid over-fit…
JiayiFeng May 12, 2017
b0071ff
update README.md
JiayiFeng May 15, 2017
631ffdd
remove .gitignore
JiayiFeng May 16, 2017
e6cfcf1
update README.md and renew dnn_net.png
JiayiFeng May 16, 2017
f46dcb1
read->red
JiayiFeng May 17, 2017
6ff3565
read->red
JiayiFeng May 17, 2017
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions .gitignore
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
.DS_Store
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这个文件暂时先删掉吧。ignore mac 的这个文件比较奇怪。

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这个文件还没删掉?

1 change: 1 addition & 0 deletions text_classification/.gitignore
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
temp/
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这个在目前上传的脚本中,并没有用到 temp 文件夹吧?如果是这样,请先从版本库中删除。

229 changes: 228 additions & 1 deletion text_classification/README.md
Original file line number Diff line number Diff line change
@@ -1 +1,228 @@
TBD
# 文本分类
文本分类是机器学习中的一项常见任务,主要目的是根据一条文本的内容,判断该文本所属的类别。在本例子中,我们利用有标注的语料库训练二分类DNN和CNN模型,完成对输入文本的分类任务。

DNN与CNN模型之间最大的区别在于:CNN是一种序列模型,能够提取一个局部区域之内的特征,能够处理变长的序列输入。而DNN大多使用基本的全连接结构,不是一种序列模型,只能接受固定维度的特征向量作为输入。举例来说,情感分类是一项常见的文本分类任务,在情感分类中,我们希望训练一个模型来判断句子中表现出的情感是正向还是负向。例如,"The apple is not bad",其中的"not bad"是决定这个句子情感的关键。对于DNN模型来说,只能知道句子中有一个"not"和一个"bad",并且在输入时“not”和“bad”之间的顺序关系已经丢失,网络不再有机会学习到序列之间蕴含的特征;而CNN模型接受文本序列作为输入,保留了"not bad"之间的顺序信息。因此,在大多数文本分类任务上,CNN模型的表现要好于DNN。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  1. 对比区别时使用列表更为清晰
  2. 既然开头是DNN与CNN,所以先讲DNN,后讲CNN。顺序要统一。

建议修改如下:

DNN与CNN模型之间最大的区别在于:

  • DNN不属于序列模型,大多使用基本的全连接结构,只能接受固定维度的特征向量作为输入。
  • CNN属于序列模型,能够提取一个局部区域之内的特征,从而处理变长的序列输入。

举例来说,情感分类是一项常见的文本分类任务。在情感分类中,我们希望训练一个模型来判断句子中表现出的情感是正向还是负向。例如,"The apple is not bad",其中"not bad"是决定这个句子情感的关键。

  • 对于DNN模型来说,只能知道句子中有一个"not"和一个"bad",但两者之间的顺序关系在输入时已经丢失,网络不再有机会学习序列之间的顺序信息。
  • CNN模型接受文本序列作为输入,保留了"not bad"之间的顺序信息。因此,在大多数文本分类任务上,CNN模型的表现要好于DNN。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done


## 实验数据
本例子的实验在IMDB数据集上进行([数据集下载](http://ai.stanford.edu/%7Eamaas/data/sentiment/aclImdb_v1.tar.gz))。IMDB数据集包含了来自IMDB(互联网电影数据库)网站的5万条电影影评,并被标注为正面/负面两种评价。数据集被划分为train和test两部分,各2.5万条数据,正负样本的比例基本为1:1。样本直接以英文原文的形式表示。
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

数据集下载这里可以删去,放在第一次引入 IMDB 时加上。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

本例子的实验在IMDB数据集上进行。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done


## DNN模型

**DNN的模型结构入下图所示:**
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

“DNN的模型结构如下图所示:”,有个错别字


<p align="center">
<img src="images/dnn_net.png" width = "90%" align="center"/><br/>
图1. DNN文本分类模型
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这张图能显示出边框,请重新上传全白底的图片
image

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done

</p>

**可以看到,模型主要分为如下几个部分:**

- **词向量层**:IMDB的样本由原始的英文单词组成,为了更好地表示不同的词之间语义上的关系,首先将英文单词转化为固定维度的向量。训练完成后,词与词语义上的相似程度将可以用它们的词向量之间的距离来表示,语义上越相似,距离越近。关于词向量的更多信息请参考PaddleBook中的[词向量](https://github.com/PaddlePaddle/book/tree/develop/04.word2vec)一节。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  1. 不同的词-》不同词
  2. 将可以(去掉将)

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done


- **最大池化层**:最大池化在时间序列上进行,池化过程消除了不同语料样本在单词数量多少上的差异,并提炼出词向量中每一下标位置上的最大值。经过池化后,词向量层输出的向量的序列被转化为一条固定维度的向量。例如,假设最大池化前向量的序列为`[[2,3,5],[7,3,6],[1,4,0]]`,则最大池化的结果为:`[7,4,6]`。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

输出的向量的序列:“的”太多了

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done


- **全连接隐层**:经过最大池化后的向量被送入两个连续的隐层,隐层之间为全连接结构。


- **输出层**:输出层的神经元数量和样本的类别数一致,例如在二分类问题中,输出层会有2个神经元。通过Softmax激活函数,输出结果是一个归一化的概率分布,和为1,因此第i个神经元的输出就可以认为是样本属于第i类的预测概率。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

。因此第$i$个神经元的输出就可以认为是样本属于第$i$类的预测概率。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done


**通过PaddlePaddle实现该DNN结构的代码如下:**

```python
import paddle.v2 as paddle

def fc_net(dict_dim, class_dim=2, emb_dim=28):
"""
dnn network definition

:param dict_dim: size of word dictionary
:type input_dim: int
:params class_dim: number of instance class
:type class_dim: int
:params emb_dim: embedding vector dimension
:type emb_dim: int
"""

# input layers
data = paddle.layer.data("word",
paddle.data_type.integer_value_sequence(dict_dim))
lbl = paddle.layer.data("label", paddle.data_type.integer_value(class_dim))

# embedding layer
emb = paddle.layer.embedding(input=data, size=emb_dim)
# max pooling
seq_pool = paddle.layer.pooling(
input=emb, pooling_type=paddle.pooling.Max())

# two hidden layers
hd_layer_size = [28, 8]
hd_layer_init_std = [1.0 / math.sqrt(s) for s in hd_layer_size]
hd1 = paddle.layer.fc(
input=seq_pool,
size=hd_layer_size[0],
act=paddle.activation.Tanh(),
param_attr=paddle.attr.Param(initial_std=hd_layer_init_std[0]))
hd2 = paddle.layer.fc(
input=hd1,
size=hd_layer_size[1],
act=paddle.activation.Tanh(),
param_attr=paddle.attr.Param(initial_std=hd_layer_init_std[1]))

# output layer
output = paddle.layer.fc(
input=hd2,
size=class_dim,
act=paddle.activation.Softmax(),
param_attr=paddle.attr.Param(initial_std=1.0 / math.sqrt(class_dim)))

cost = paddle.layer.classification_cost(input=output, label=lbl)

return cost, output, lbl

```
该DNN模型默认对输入的语料进行二分类(`class_dim=2`),embedding的词向量维度默认为28(`emd_dim=28`),两个隐层均使用Tanh激活函数(`act=paddle.activation.Tanh()`)。

需要注意的是,该模型的输入数据为整数序列,而不是原始的英文单词序列。事实上,为了处理方便我们一般会事先将单词根据词频顺序进行id化,即将单词用整数替代, 也就是单词在字典中的序号。这一步一般在DNN模型之外完成。

## CNN模型

**CNN的模型结构如下图所示:**

<p align="center">
<img src="images/cnn_net.png" width = "90%" align="center"/><br/>
图2. CNN文本分类模型
</p>

**可以看到,模型主要分为如下几个部分:**

- **词向量层**:与DNN中词向量层的作用一样,将英文单词转化为固定维度的向量,利用向量之间的距离来表示词之间的语义相关程度。如图2中所示,将得到的词向量定义为行向量,再将语料中所有的单词产生的行向量拼接在一起组成矩阵。假设词向量维度为5,语料“The cat sat on the read mat”包含7个单词,那么得到的矩阵维度为7*5。关于词向量的更多信息请参考PaddleBook中的[词向量](https://github.com/PaddlePaddle/book/tree/develop/04.word2vec)一节。

- **卷积层**: 文本分类中的卷积在时间序列上进行,即卷积核的宽度和词向量层产出的矩阵一致,卷积验证矩阵的高度方向进行。卷积后得到的结果被称为“特征图”(feature map)。假设卷积核的高度为h,矩阵的高度为N,卷积的步长为1,则得到的特征图为一个高度为N+1-h的向量。可以同时使用多个不同高度的卷积核,得到多个特征图。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  1. 卷积验证矩阵的高度方向进行:这句话不是很明白,验证矩阵是什么?
  2. $h$, $N$, $N+1-h$

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

此处为手误,“验证”应该为“沿着”,done


- **最大池化层**: 对卷积得到的各个特征图分别进行最大池化操作。由于特征图本身已经是向量,因此这里的最大池化实际上就是简单地选出各个向量中的最大元素。各个最大元素又被并置在一起,组成新的向量,显然,该向量的维度等于特征图的数量,也就是卷积核的数量。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  1. 并置-》拼接?
  2. DNN的最大池化举了例子,这里是否也可举个例子。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

确实是“拼接”表达得更加精确,done


- **全连接与输出层**:将最大池化的结果通过全连接层输出,与DNN模型一样,最后输出层的神经元个数与样本的类别数量一致,且输出之和为1。

**通过PaddlePaddle实现该CNN结构的代码如下:**

```python
import paddle.v2 as paddle

def convolution_net(dict_dim, class_dim=2, emb_dim=28, hid_dim=128):
"""
cnn network definition

:param dict_dim: size of word dictionary
:type input_dim: int
:params class_dim: number of instance class
:type class_dim: int
:params emb_dim: embedding vector dimension
:type emb_dim: int
:params hid_dim: number of same size convolution kernels
:type hid_dim: int
"""

# input layers
data = paddle.layer.data("word",
paddle.data_type.integer_value_sequence(dict_dim))
lbl = paddle.layer.data("label", paddle.data_type.integer_value(2))

#embedding layer
emb = paddle.layer.embedding(input=data, size=emb_dim)

# convolution layers with max pooling
conv_3 = paddle.networks.sequence_conv_pool(
input=emb, context_len=3, hidden_size=hid_dim)
conv_4 = paddle.networks.sequence_conv_pool(
input=emb, context_len=4, hidden_size=hid_dim)

# fc and output layer
output = paddle.layer.fc(
input=[conv_3, conv_4], size=class_dim, act=paddle.activation.Softmax())

cost = paddle.layer.classification_cost(input=output, label=lbl)

return cost, output, lbl
```

该CNN网络的输入数据类型和前面介绍过的DNN一致。`paddle.networks.sequence_conv_pool`为PaddlePaddle中已经封装好的带有池化的文本序列卷积模块,该模块的`context_len`参数用于指定卷积核在同一时间覆盖的文本长度,也即图2中的卷积核的高度;`hidden_size`用于指定该类型的卷积核的数量。可以看到,上述代码定义的结构中使用了128个大小为3的卷积核和128个大小为4的卷积核,这些卷积的结果经过最大池化和结果并置后产生一个256维的向量,向量经过一个全连接层输出最终预测结果。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  1. 也即-》即
  2. 并置-》拼接?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done


## 自定义数据
本样例中的代码通过`Paddle.dataset.imdb.train`接口使用了PaddlePaddle自带的样例数据,在第一次运行代码时,PaddlePaddle会自动下载并缓存所需的数据。如果希望使用自己的数据进行训练,需要自行编写数据读取接口。

编写数据读取接口的关键在于实现一个Python生成器,生成器负责从原始输入文本中解析出一条训练样本,并组合成适当的数据形式传送给网络中的data layer。例如在本样例中,data layer需要的数据类型为`paddle.data_type.integer_value_sequence`,本质上是一个Python list。因此我们的生成器需要完成:1. 从文件中读取数据, 2. 转换成适当形式的Python list,这两件事情。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

因此我们的生成器需要完成:从文件中读取数据,以及转换成适当形式的Python list,这两件事情。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done


假设原始数据的格式为:

```
PaddlePaddle is good 1
What a terrible weather 0
```
每一行为一条样本,样本包括了原始语料和标签,语料内部单词以空格分隔,语料和标签之间用`\t`分隔。对以上格式的数据,可以使用如下自定义的数据读取接口为PaddlePaddle返回训练数据:

```python
def encode_word(word, word_dict):
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

在代码中为每一个参数添加注释。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done

"""
map word to id

:param word: the word to be mapped
:type word: str
:param word_dict: word dictionary
:type word_dict: Python dict
"""

if word_dict.has_key(word):
return word_dict[word]
else:
return word_dict['<unk>']

def data_reader(file_name, word_dict):
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

请按照python 的docstring 为每一个参数添加注释。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done

"""
Reader interface for training data

:param file_name: data file name
:type file_name: str
:param word_dict: word dictionary
:type word_dict: Python dict
"""

def reader():
with open(file_name, "r") as f:
for line in f:
ins, label = line.strip('\n').split('\t')
ins_data = [int(encode_word(w, word_dict)) for w in ins.split(' ')]
yield ins_data, int(label)
return reader
```

`word_dict`是字典,用来将原始的单词字符串转化为在字典中的序号。可以用`data_reader`替换原先代码中的`Paddle.dataset.imdb.train`接口用以提供自定义的训练数据。

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

我有一个建议,PaddleBook 是目前官方唯一可以找到正式托管 V2 API 用法的“教程”,但是PaddleBook读数据直接 import paddle.dataset.×,隐层了读数据的过程。

普通用户如果需要替换自己的数据,还是会比较难以运行这些例子。我们可以简单地增加一节:使用用户数据训练。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

好的,将在接下来添加这一部分

## 运行与输出

本部分以上文介绍的DNN网络为例,介绍如何利用样例中的`text_classification_dnn.py`脚本进行DNN网络的训练和对新样本的预测。

`text_classification_dnn.py`中的代码分为四部分:

- **fc_net函数**:定义dnn网络结构,上文已经有说明。

- **train\_dnn\_model函数**:模型训练函数。定义优化方式、训练输出等内容,并组织训练流程。每完成一个pass的训练,程序都会将当前的模型参数保存在硬盘上,文件名为:`dnn_params_pass***.tar.gz`,其中`***`表示pass的id,从0开始计数。本函数接受一个整数类型的参数,表示训练pass的总轮数。

- **dnn_infer函数**:载入已有模型并对新样本进行预测。函数开始运行后会从当前路径下寻找并读取指定名称的参数文件,加载其中的模型参数,并对test数据集中的样本进行预测。

- **main函数**:主函数

要运行本样例,直接在`text_classification_dnn.py`所在路径下执行`python ./text_classification_dnn.py`即可,样例会自动依次执行数据集下载、数据读取、模型训练和保存、模型读取、新样本预测等步骤。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

一般运行命令是python text_classification_dnn.py

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done


预测的输出形式为:

```
[ 0.99892634 0.00107362] 0
[ 0.00107638 0.9989236 ] 1
[ 0.98185927 0.01814074] 0
[ 0.31667888 0.68332112] 1
[ 0.98853314 0.01146684] 0
```

每一行表示一条样本的预测结果。前两列表示该样本属于0、1这两个类别的预测概率,最后一列表示样本的实际label。
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这里加上一些说明, CNN 模型的训练运行 ×,预测运行×。和DNN是一样的。

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done,感谢曹总的细致review


在运行CNN的模型的`text_classification_cnn.py`脚本中,网络模型定义在`convolution_net`函数中,模型训练函数名为`train_cnn_model`,预测函数名为`cnn_infer`。其他的用法和`text_classification_dnn.py`是一致的。
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

去掉多余“的”字:

  1. 运行CNN模型
  2. 其他用法

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done

Binary file added text_classification/images/cnn_net.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added text_classification/images/dnn_net.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
146 changes: 146 additions & 0 deletions text_classification/text_classification_cnn.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,146 @@
import sys
import paddle.v2 as paddle
import gzip


def convolution_net(dict_dim, class_dim=2, emb_dim=28, hid_dim=128):
"""
cnn network definition

:param dict_dim: size of word dictionary
:type input_dim: int
:params class_dim: number of instance class
:type class_dim: int
:params emb_dim: embedding vector dimension
:type emb_dim: int
:params hid_dim: number of same size convolution kernels
:type hid_dim: int
"""

# input layers
data = paddle.layer.data("word",
paddle.data_type.integer_value_sequence(dict_dim))
lbl = paddle.layer.data("label", paddle.data_type.integer_value(2))

#embedding layer
emb = paddle.layer.embedding(input=data, size=emb_dim)

# convolution layers with max pooling
conv_3 = paddle.networks.sequence_conv_pool(
input=emb, context_len=3, hidden_size=hid_dim)
conv_4 = paddle.networks.sequence_conv_pool(
input=emb, context_len=4, hidden_size=hid_dim)

# fc and output layer
output = paddle.layer.fc(
input=[conv_3, conv_4], size=class_dim, act=paddle.activation.Softmax())

cost = paddle.layer.classification_cost(input=output, label=lbl)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you help with adding more than one evaluators to this configuration, for example, besides an evaluator to calculate the error rate, if it possible to add a precision-recall evaluator? I hope to test a configuration with more than one evaluators. Thanks for your work.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done, added auc evaluator


return cost, output, lbl


def train_cnn_model(num_pass):
"""
train cnn model

:params num_pass: train pass number
:type num_pass: int
"""

# load word dictionary
print 'load dictionary...'
word_dict = paddle.dataset.imdb.word_dict()

dict_dim = len(word_dict)
class_dim = 2
# define data reader
train_reader = paddle.batch(
paddle.reader.shuffle(
lambda: paddle.dataset.imdb.train(word_dict), buf_size=1000),
batch_size=100)
test_reader = paddle.batch(
lambda: paddle.dataset.imdb.test(word_dict), batch_size=100)

# network config
[cost, output, label] = convolution_net(dict_dim, class_dim=class_dim)
# create parameters
parameters = paddle.parameters.create(cost)
# create optimizer
adam_optimizer = paddle.optimizer.Adam(
learning_rate=1e-3,
regularization=paddle.optimizer.L2Regularization(rate=1e-3),
model_average=paddle.optimizer.ModelAverage(average_window=0.5))

# add auc evaluator
paddle.evaluator.auc(input=output, label=label)

# create trainer
trainer = paddle.trainer.SGD(
cost=cost, parameters=parameters, update_equation=adam_optimizer)

# Define end batch and end pass event handler
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 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()
if isinstance(event, paddle.event.EndPass):
result = trainer.test(reader=test_reader, feeding=feeding)
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics)
with gzip.open("cnn_params_pass" + str(event.pass_id) + ".tar.gz",
'w') as f:
parameters.to_tar(f)

# begin training network
feeding = {'word': 0, 'label': 1}
trainer.train(
reader=train_reader,
event_handler=event_handler,
feeding=feeding,
num_passes=num_pass)

print("Training finished.")


def cnn_infer(file_name):
"""
predict instance labels by cnn network

:params file_name: network parameter file
:type file_name: str
"""

print("Begin to predict...")

word_dict = paddle.dataset.imdb.word_dict()
dict_dim = len(word_dict)
class_dim = 2

[_, output, _] = convolution_net(dict_dim, class_dim=class_dim)
parameters = paddle.parameters.Parameters.from_tar(gzip.open(file_name))

infer_data = []
infer_data_label = []
for item in paddle.dataset.imdb.test(word_dict):
infer_data.append([item[0]])
infer_data_label.append(item[1])

predictions = paddle.infer(
output_layer=output,
parameters=parameters,
input=infer_data,
field=['value'])
for i, prob in enumerate(predictions):
print prob, infer_data_label[i]


if __name__ == "__main__":
paddle.init(use_gpu=False, trainer_count=1)
num_pass = 5
train_cnn_model(num_pass=num_pass)
param_file_name = "cnn_params_pass" + str(num_pass - 1) + ".tar.gz"
cnn_infer(file_name=param_file_name)
Loading