/xinye_competition

Primary LanguagePython

xinye_competition

xinye_competition
├── data
│   ├── train
│   │   ├── a_images
│   │   ├── a_annotations.json
│   │   ├── b_images
│   │   ├── b_annotations.json
│   ├── test
│   │   ├── a_images
│   │   ├── a_annotations.json
│   │   ├── b_images
│   │   ├── b_annotations.json
│   ├── cropped_train
│   ├── cropped_test
│   ├── retail
│   │   ├── train
│   │   ├── val
├── model
├── model_files
├── submit
├── tools
├── utils

train and test folders are original data.

cropped_train/test and retail folders are generated by tools/prepare_data.py.

cropped_train for regressor training.

cropped_test for regressor testing.

retail for detector training.

TODO List

Regressor

  • more feature dims(cancel the feature concat)
  • more augmentation
  • add auxiliary class loss
  • CGD
  • EMA
  • TTA
  • CoAtNet
  • VOLO

Detector

  • train retail as single class
  • more dataset
  • WBF
  • more powerful detector
cd xinye_competition

Prepare data

python tools/prepare_data.py --regressor-data --detector-data --generate-pair

Train regressor

python tools/train_regressor.py --epochs 150 --input_size 224 --batch_size 200 --backbone swin_transformer --swin_type small --loss_head Circleloss --use_cgd
python tools/train_regressor.py --epochs 150 --input_size 224 --batch_size 100 --backbone swin_transformer --swin_type large --loss_head Circleloss --use_cgd
python tools/train_regressor.py --epochs 150 --input_size 224 --batch_size 100 --backbone swin_transformer --swin_type large --loss_head Circleloss --use_cgd

Train detector

python tools/train_detector.py

Inference

python inference.py