Kakao Arena - Product Classification

Code for '쇼핑몰 상품 카테고리 분류'

• Team: baseIine

Public Leaderboard(2019/01/07)

Features

• Fully dockerized environment
• Input Pipeline
  • Tokenize product metadata with Okt POS Tagger
  • Use TFRecord
• 5 classifiers with 2-layer MLP
  • one for concatenated label of b,m,s,d
  • 4 classifiers for each category
• Adversarial Training

Results

• The metric 'score' is calculated by the equation as follows:
  • score=(1.0 * b_acc + 1.2 * m_acc + 1.3 * s_acc + 1.4 * d_acc)
• The model Final was used to report our final results on dev, test
• Download trained weights here

Model	Dev score	Test score(TBD)	File Size
Intermediate	1.07799	-	966MB
Ensemble	1.080755	-	5*966MB
*Final	1.077696	-	966MB

Requirements

• Docker
• python >=2.7
  • Tensorflow >=1.12
  • Keras
  • Othres: h5py, tqdm, easydict
• Enough storage space at least 400GB

Reproduce results

Setup

1. Download datasets from kakao arena

Run a docker

$ bash build.sh
$ bash run.sh

[Note] Edit DATA_PATH from run.sh

For example,

ls $DATA_PATH
|- dev.chunk.01
|- test.chunk.01
|- test.chunk.02
|- train.chunk.01
|- train.chunk.02
|- train.chunk.03
|- train.chunk.04
|- train.chunk.05
|- train.chunk.06
|- train.chunk.07
|- train.chunk.08
`- train.chunk.09

Option1: Use pretrained weights

1. Download weights Dropbox Link
2. Copy weights to /data/output/interim, /data/output/final

$ bash scripts/eval.sh 0 interim 70 # for validation
$ bash scripts/inference.sh 0 interim 70 dev # for submission
$ bash scripts/inference.sh 0 interim 70 test # for submission

$ bash scripts/inference.sh 0 final 12 dev # for submission
$ bash scripts/inference.sh 0 final 12 test # for submission

Option2: Train a model from scratch

$ bash reproduce.sh

Reference

1. Baseline code
2. KoNLPy: Korean natural language processing in Python
3. Adversarial Training

License

© Taekmin Kim, 2019. Licensed under the MIT License.