Skip to content
master
Switch branches/tags
Go to file
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
kin
 
 
 
 
 
 
 
 

Naver AI Hackathon 2018

Team kozistr - Member : 김형찬(kozistr)


tl;dr

I participated in [Naver A.I Hackathon 2018] and ranked 4th/13th(over 200 teams total) as an individual participant (Team : kozistr)

And also i uploaded summary docs with the codes.

Final LeaderBoard

네이버 지식iN 질문 유사도 예측 (결선)

kin_leaderboard

네이버 영화 평점 예측 (결선)

movie_leaderboard

Result

Stage Mission Metric Score Rank Code
phase 1 kin acc
phase 1 movie-review mse
phase 2 kin acc
phase 2 movie-review mse
final kin acc 0.8115 4th code
final movie-review mse 0.0310 13th code

전처리를 하나도 하지 않고 기본 도커만 사용해서 시도 한 모델들의 main.py만 업로드 합니다!

_codes 폴더에 default 로 주어진 전처리 파일들 업로드!

Models

Soon~

Summary!

Here's summary docs! Summary

Hyper-Parameters

네이버 지식iN 질문 유사도 예측

Name Value Note
Epochs 100 70 ~ 80 에서 converge
Learning Rate 1e-3 exponential decay (rate 0.95)
Batch Size 64/128 본선에서는 128
DropOut Rate 0.7 0.7 is the best
Embeddings 384 384 ~ 400 good
CNN kernel size 10, 9, 7, 5, 3 10 이하에서 찾음
CNN filter size 256 256 ~ good
FC Unit 1024 512 ~ 1024 good
Optimizer Adam Adam, Momentum, SGD ~
...

네이버 영화 평점 예측

Name Value Note
Epochs 30 20 ~ 30 에서 converge
Learning Rate 2e-4 lr 에 엄청나게 민감
Batch Size 128 128 ~
DropOut Rate 0.6 0.6 is the best
Embeddings 128 128 ~ 256 good
CNN kernel size 3, 5, 7 10 이하에서 찾음
CNN filter size 256 256 ~ good
FC Unit 512 ~ 512 good
Optimizer Adam Adam, SGD ~
...

Author

HyeongChan Kim (@kozistr, kozistr@gmail.com)

Releases

No releases published

Packages

No packages published

Languages