https://github.com/haven-jeon/ko_en_neural_machine_translation
Korean English NMT(Neural Machine Translation) with Gluon
https://github.com/haven-jeon/ko_en_neural_machine_translation
gluon mxnet nmt-model seq2seq-attn
Last synced: 2 months ago
JSON representation
Korean English NMT(Neural Machine Translation) with Gluon
- Host: GitHub
- URL: https://github.com/haven-jeon/ko_en_neural_machine_translation
- Owner: haven-jeon
- Created: 2018-01-18T10:54:48.000Z (over 7 years ago)
- Default Branch: master
- Last Pushed: 2018-02-28T14:38:19.000Z (over 7 years ago)
- Last Synced: 2025-05-17T14:11:37.848Z (5 months ago)
- Topics: gluon, mxnet, nmt-model, seq2seq-attn
- Language: Jupyter Notebook
- Homepage:
- Size: 31.9 MB
- Stars: 60
- Watchers: 7
- Forks: 20
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
- Awesome-MXNet - ko_en_NMT(Gluon:star:)
README
### Neural Korean to English Machine Translater with Gluon
2018 마이크로소프트웨어 기사를 통해 오신분은 [이곳](maso_ver/)을 참고하세요.
#### 특징
- Seq2Seq with Attention
- Gluon 기반 [Hybridize](https://mxnet.incubator.apache.org/tutorials/gluon/hybrid.html)
- 공개 학습셋 (from https://github.com/jungyeul/korean-parallel-corpora) 기반 학습
- multi gpu 학습 옵션 제공
- Variable Sentence Length 학습
#### 요구사항
- Python 3.5 이상
- MXNet 1.0 이상
- tqdm, konlpy, stemming, gensim
#### 구성
- 아키텍처
- seq2seq
- Attention
- bi-directional GRU encoder
- 영한 교차 임베딩 학습
#### TODO
- Beam Search 구현
#### How to
- 임베딩 학습
```
python main.py --embedding
```
- 번역 모형 학습
```
CUDA_VISIBLE_DEVICES='4,5,6,7' python main.py --train --gpu-count 4 --batch-size 100
```
- 번역 모형 테스트
```
python main.py --test --init-model "models/trained_model.params"
kor > 오바마는 대통령이다.
오바마는 대통령이다.
eng > obama is a presid .
kor > 좋은 아침.
좋은 아침.
eng > good morn
kor > 북한은 핵무기를 가지고 있다.
북한은 핵무기를 가지고 있다.
eng > north korea has a nuclear weapon .
kor >
```
- models/trained_model.params
1. rmsprop(batch size : 100, lr : 0.01) : 7 epocs
1. sgd (batch size : 100, lr : 0.01) : 33 epocs