deepray 0.1.2

DeePray（深度祈祷）：一个新的模块化、可伸缩、可配置、易于使用和扩展的基础设施，用于基于深度学习的推荐。在

简介

DeePray库为[深度学习推荐]提供最先进的算法。 DeePray基于最新的[TensorFlow 2]构建[(https://tensorflow.org/)]采用模块化结构设计 以表格的形式来发现问题和答案。在

DeePray的主要目标：

• 简单易用，新手可以通过深入学习快速掌握
• 具有良好的网络规模数据性能
• 易于扩展，模块化架构让你像玩乐高一样建立你的神经网络！在

我们开始吧！请参阅https://deepray.readthedocs.io/en/latest/的正式文件。在

安装

使用PyPI安装DeePray:

{{5}使用下面的cdray}安装命令执行{a5}：

pip install deepray

从Github安装DeePray源：

首先，使用git克隆DeePray存储库：

^{pr2}$

然后，cd到deepray文件夹，并通过执行以下命令安装库：

cd deepray
pip install .

教程

人口普查成人数据集

数据准备

在表格数据中，指定numeric作为继续特性，CATEGORY用于分类功能，VARIABLE用于可变长度的功能，显然，LABEL用于标签列。然后将它们处理为TFRecord格式，以便在大规模数据集中获得良好的性能。在

importpandasaspdfromsklearn.model_selectionimporttrain_test_splitfromsklearn.preprocessingimportMinMaxScaler,LabelEncoderfromdeepray.utils.converterimportCSV2TFRecord# http://archive.ics.uci.edu/ml/datasets/Adulttrain_data='DeePray/examples/census/data/raw_data/adult_data.csv'df=pd.read_csv(train_data)df['income_label']=(df["income_bracket"].apply(lambdax:">50K"inx)).astype(int)df.pop('income_bracket')NUMERICAL_FEATURES=['age','fnlwgt','hours_per_week','capital_gain','capital_loss','education_num']CATEGORY_FEATURES=[colforcolindf.columnsifcol!=LABELandcolnotinNUMERICAL_FEATURES]LABEL=['income_label']forfeatinCATEGORY_FEATURES:lbe=LabelEncoder()df[feat]=lbe.fit_transform(df[feat])# Feature normilizationmms=MinMaxScaler(feature_range=(0,1))df[NUMERICAL_FEATURES]=mms.fit_transform(df[NUMERICAL_FEATURES])prebatch=1# flags.prebatchconverter=CSV2TFRecord(LABEL,NUMERICAL_FEATURES,CATEGORY_FEATURES,VARIABLE_FEATURES=[],gzip=False)converter.write_feature_map(df,'./data/feature_map.csv')train_df,valid_df=train_test_split(df,test_size=0.2)converter(train_df,out_file='./data/train.tfrecord',prebatch=prebatch)converter(valid_df,out_file='./data/valid.tfrecord',prebatch=prebatch)

您将得到一个这样的特征映射文件：

9,workclass,CATEGORICAL
16,education,CATEGORICAL
7,marital_status,CATEGORICAL
15,occupation,CATEGORICAL
6,relationship,CATEGORICAL
5,race,CATEGORICAL
2,gender,CATEGORICAL
42,native_country,CATEGORICAL
1,hours_per_week,NUMERICAL
1,capital_gain,NUMERICAL
1,age,NUMERICAL
1,fnlwgt,NUMERICAL
1,capital_loss,NUMERICAL
1,education_num,NUMERICAL
2,income_label,LABEL

然后分别创建两个txt文件train和{}，记录列车集TFRecords和有效的set TFRecords文件路径。在

选择您的模式、培训和评估

"""build and train model"""importsysfromabslimportapp,flagsimportdeeprayasdpfromdeepray.base.trainerimporttrainfromdeepray.model.build_modelimportBuildModelFLAGS=flags.FLAGSdefmain(flags=None):FLAGS(flags,known_only=True)flags=FLAGSmodel=BuildModel(flags)history=train(model)print(history)argv=[sys.argv[0],'--model=lr','--train_data=/Users/vincent/Projects/DeePray/examples/census/data/train','--valid_data=/Users/vincent/Projects/DeePray/examples/census/data/valid','--feature_map=/Users/vincent/Projects/DeePray/examples/census/data/feature_map.csv','--learning_rate=0.01','--epochs=10','--batch_size=64',]main(flags=argv)

型号列表

Titile	Booktitle	Resources
FM: Factorization Machines	ICDM'2010	[pdf][code]
FFM: Field-aware Factorization Machines for CTR Prediction	RecSys'2016	[pdf][code]
FNN: Deep Learning over Multi-field Categorical Data: A Case Study on User Response Prediction	ECIR'2016	[pdf][code]
PNN: Product-based Neural Networks for User Response Prediction	ICDM'2016	[pdf][code]
Wide&Deep: Wide & Deep Learning for Recommender Systems	DLRS'2016	[pdf][code]
AFM: Attentional Factorization Machines: Learning the Weight of Feature Interactions via Attention Networks	IJCAI'2017	[pdf][code]
NFM: Neural Factorization Machines for Sparse Predictive Analytics	SIGIR'2017	[pdf][code]
DeepFM: DeepFM: A Factorization-Machine based Neural Network for CTR Prediction[C]	IJCAI'2017	[pdf][code]
DCN: Deep & Cross Network for Ad Click Predictions	ADKDD'2017	[pdf][code]
xDeepFM: xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems	KDD'2018	[pdf][code]
DIN: DIN: Deep Interest Network for Click-Through Rate Prediction	KDD'2018	[pdf][code]
DIEN: DIEN: Deep Interest Evolution Network for Click-Through Rate Prediction	AAAI'2019	[pdf][code]
DSIN: Deep Session Interest Network for Click-Through Rate Prediction	IJCAI'2019	[pdf][code]
AutoInt: Automatic Feature Interaction Learning via Self-Attentive Neural Networks	CIKM'2019	[pdf][code]
FLEN: Leveraging Field for Scalable CTR Prediction	AAAI'2020	[pdf][code]
DFN: Deep Feedback Network for Recommendation	IJCAI'2020	[pdf][code]

如何使用DeePray构建自己的模型

从from deepray.model.model_ctr继承BaseCTRModel类，并实现您自己的build_network()方法！在

贡献

DeePray仍在开发中，请大家捐款！在

* Hailin Fu (`Hailin <https://github.com/fuhailin>`)
* Call for contributions!

深水区

引用

DeePray由Hailin设计、开发和支持。 如果您在研究中使用了本库的任何部分，请使用以下BibTex条目引用它

@misc{DeePray,
  author = {Hailin Fu},
  title = {DeePray: A new Modular, Scalable, Configurable, Easy-to-Use and Extend infrastructure for Deep Learning based Recommendation},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub Repository},
  howpublished = {\url{https://github.com/fuhailin/deepray}},
}

许可证

版权所有（c）版权所有©2020 The DeePray Authors。版权所有。在

根据Apach许可证授权。在

参考文献

https://github.com/shenweichen/DeepCTR

https://github.com/aimetrics/jarvis

https://github.com/shichence/AutoInt

联系人

如果您需要合作或有任何疑问，请关注我的微信公众号：

编号：【最新技术】

标签：

• 项目

欢迎加入QQ群-->： 979659372

推荐PyPI第三方库

• pyesp

  基于mircopython和nodemcu的esp8266平台编程/通信工具

• flask_markdown2

  向模板添加{%markdown%}标记的烧瓶扩展。

• python-actors

  巨蟒演员

• eptransition

  python中的energyplus文件转换。

• pholcidae

  pholcidae-微型python网络爬虫库

• affine6p-cstest

  用六个自由参数计算仿射变换参数。

• instaloader

  从Instagram下载图片（或视频）及其标题和其他元数据。

• wechat-web-auth

  微信公众平台网页授权获取用户信息python封装

• python-ecore

  ecore的python绑定

• ultraif

  Ultraify项目

• i3-output-selector

  用于更改i3中当前工作区的输出的gui工具。

• odoo10-addon-cms-delete-content-example

  显示CMS删除内容的基本内容

• GoldSaxTamilTranslate

  使python应用程序能够轻松地将这个包从英语翻译成泰米尔语，包括一个音译。

• pyharbor

  港口python sdk

• jolt

  任务执行人