All Projects → nlpjoe → Ccf Bdci Automotive Field Asc 2018

nlpjoe / Ccf Bdci Automotive Field Asc 2018

CCF-BDCI 2018年汽车行业用户观点主题及情感识别挑战赛 第6名解决方案

Projects that are alternatives of or similar to Ccf Bdci Automotive Field Asc 2018

Benchmarks
Comparison tools
Stars: ✭ 139 (-1.42%)
Mutual labels:  jupyter-notebook
Data Structure And Algorithm Using Python
In computer science, a data structure is a data organization, management, and storage format that enables efficient access and modification. More precisely, a data structure is a collection of data values, the relationships among them, and the functions or operations that can be applied to the data. (Using Python 3)
Stars: ✭ 141 (+0%)
Mutual labels:  jupyter-notebook
Book
This book serves as an introduction to a whole new way of thinking systematically about geographic data, using geographical analysis and computation to unlock new insights hidden within data.
Stars: ✭ 141 (+0%)
Mutual labels:  jupyter-notebook
Python For Algorithms Data Structures And Interviews
Files for Udemy Course on Algorithms and Data Structures
Stars: ✭ 1,917 (+1259.57%)
Mutual labels:  jupyter-notebook
Lmkor
Pretrained Language Models for Korean
Stars: ✭ 140 (-0.71%)
Mutual labels:  jupyter-notebook
Hpo2nas Tutorial Cvpr Eccv2020
Hands-on Tutorial on Automated Deep Learning
Stars: ✭ 141 (+0%)
Mutual labels:  jupyter-notebook
Interactive machine learning
IPython widgets, interactive plots, interactive machine learning
Stars: ✭ 140 (-0.71%)
Mutual labels:  jupyter-notebook
Agem
Official implementation of the Averaged Gradient Episodic Memory (A-GEM) in Tensorflow
Stars: ✭ 141 (+0%)
Mutual labels:  jupyter-notebook
Python Youtube Code
Stars: ✭ 139 (-1.42%)
Mutual labels:  jupyter-notebook
Swift4tf
A set of notebooks explaining swift for tensorflow optimized to run in Google Collaboratory.
Stars: ✭ 141 (+0%)
Mutual labels:  jupyter-notebook
Nlpaug
Data augmentation for NLP
Stars: ✭ 2,761 (+1858.16%)
Mutual labels:  jupyter-notebook
Scipy Tutorial 2015
Geospatial data tutorial
Stars: ✭ 140 (-0.71%)
Mutual labels:  jupyter-notebook
Vaal
Variational Adversarial Active Learning (ICCV 2019)
Stars: ✭ 140 (-0.71%)
Mutual labels:  jupyter-notebook
Practical Machine Learning With Python
Master the essential skills needed to recognize and solve complex real-world problems with Machine Learning and Deep Learning by leveraging the highly popular Python Machine Learning Eco-system.
Stars: ✭ 1,868 (+1224.82%)
Mutual labels:  jupyter-notebook
Stanford cs231n 2019
Solutions and comments to assignments for 2019 Stanford's course on convolutional neural networks
Stars: ✭ 141 (+0%)
Mutual labels:  jupyter-notebook
Imageprocessing
MicaSense RedEdge and Altum image processing tutorials
Stars: ✭ 139 (-1.42%)
Mutual labels:  jupyter-notebook
Data science blogs
A repository to keep track of all the code that I end up writing for my blog posts.
Stars: ✭ 139 (-1.42%)
Mutual labels:  jupyter-notebook
Py4chemoinformatics
Python for chemoinformatics
Stars: ✭ 140 (-0.71%)
Mutual labels:  jupyter-notebook
Quantum Native Dojo
量子コンピュータ初学者のための自習教材
Stars: ✭ 141 (+0%)
Mutual labels:  jupyter-notebook
19 udacity dsa
Data Structures & Algorithms Nanodegree Program from Udacity
Stars: ✭ 140 (-0.71%)
Mutual labels:  jupyter-notebook

CCF-BDCI2018 汽车领域ASC挑战赛

以前没接触过ASC、TSC领域,最开始纠结这是单分类还是多分类问题,走了一些弯路。最终我们回到ASC赛道上,根据直觉,我们设计了一个基于memory的lstm-attention模型,复赛B榜线上在0.69左右,融合最终得分0.70,单模型结构图如下:

后面时间比较紧张,复现今年ASC论文的代码效果都不好,最终排名6/1701,思路如同代码所写,很简单。

原始数据可在比赛数据处下载,由于这次我们问题建模方式比较多,数据预处理代码也比较多,所以我会上传一份处理好的数据(包括处理好的Bert特征和百度百科词向量)放在百度云盘,下载后请解压放在data/目录下。

ELMo哈工大基于pytorch的pretrain版本和我用tf pretrain训练集的效果都不好,但是我也保留了tf pretrain版本代码。

Bert我们没有弄finetune,直接抽取的特征,效果和百度百科词向量相当。

若有任何想法可以提issue或者pull request,也可以微信与我直接讨论。希望大家一起学习进步。

一、环境

环境/库 版本
Ubuntu 16.04.5 LTS
python 3.6
jupyter notebook 4.2.3
tensorflow-gpu 1.9.1
numpy 1.14.1
pandas 0.23.0
matplotlib 2.2.2
tqdm 4.24.0

这里最重要的就是我们用的Cudnn版本的lstm,所以需要tensorflow版本大于1.4.0,相应的cuda版本不能用8.0,需要9.0及以上。

二、数据预处理

都写在jupyter里了,运行src/preprocess/EDA.ipynb生成各种文件,可用看看思路,但是建议直接下载云盘处理好的结果。

三、深度模型训练

数据预处理好即可用直接train模型,单GPU运行,模型请参考src/config.py自选,参数名含义请参考src/train_predict.py

python train_predict.py --gpu 7 --model aspv0 --feature word --epoch 20 --bs 128 --oe

四、模型融合输出

python stacking.py --gpu 1 --data_type 3

这里是stackingpesudo label一起做了,请修改代码自选是否用伪标签。

这里数据集比较合适,伪标签有一定提分作用。

五、提交结果

修改src/pack_sub_dt2.py里对应stacking生成的pre_path概率结果路径,运行

python python pack_sub_dt2.py

生成提交结果。

Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].