/AICUP_2020_Mango

Repo for the competition(AICUP 2020 Mango Defect Classification). Private Leaderboard(3th / 222)

Primary LanguagePython

AICUP_2020_愛文芒果不良品分類競賽

Team Name : 一定達辰
Team Member : Dian-Shan Chen, Sian-Yao Huang
Public leaderboard / Private leaderboard : 0.7214973(3th/222) / 0.722768(3th/222)

This repo is the entire codebase for the competition AICUP_2020_愛文芒果不良品分類競賽. The backbone of this repo is SGNAS-A, an architecture searched by SGNAS[1].

Requirements

pip3 install -r requirements.txt

Data prepare

mkdir data
mv C2_TrainDev.rar data && cd data
unrar x FileName.rar

Pretrained weight of SGNAS-A

  • Download pretrained weight from Weight
mkdir pretrained
mv SGNAS_A_best.pth.tar pretrained

Train

python3 train.py --cfg config_file/mango_config.yml --title [EXPERIMENT_TITLE]

Inference

  1. Push Test_UploadSheet.csv file in the root of this repo
mv Test_UploadSheet.csv ./
python3 test.py --cfg config_file/mango_config.yml --tta [TTA_NUMBER] --output-file [OUTPUT_FILE_PATH]

Reference

[1] Sian-Yao Huang, Wei-Ta Chu. Searching by Generating: Flexible and Efficient One-Shot NAS with Architecture Generator.