NAVER A.I Hackathon 2018
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
_codes
_images
_refs
kin
movie-review
LICENSE
README.md

README.md

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)