FanChiMao / Competition-2022-Pytorch-Orchid_Classification

AICUP 2022 orchid classification

Competition-2022-Pytorch-Orchid_Classification

TEAM_482: Chi-Mao Fan, Yu-Chen Su, Wei-Hsiang Liao

• 尋找花中君子－蘭花種類辨識及分類

Installation

The model is built in PyTorch 1.8.0 and tested on Windows10 environment
(Python: 3.8, CUDA: 10.2, cudnn: 7.6).

For installing, follow these intructions

conda create -n pytorch python=3.8  
conda activate pytorch  
conda install pytorch=1.8 torchvision cudatoolkit=10.2 -c pytorch  
conda install -c conda-forge tensorboardx
pip install git+https://github.com/rwightman/pytorch-image-models.git
pip install -r requirements.txt

Dataset

You can refer the README.md to prepare the dataset.

Train each classifier

Set hyperparameters and revelent training path in train.yaml and simply run train.py.

Predict via each trained classifier

You can download our pretrained model from pretrained.
To predict the orchid images by single classifier, see predict.py and run:

python predict.py --model model_name --input_dir images_folder_path --result_dir save_csv_here --weights path_to_models

Ensemble strategy

We support the code of three different ensemble methods as following:

Ensemble figures...

Average ensemble	Traditional ensemble	Resnet ensemble

However, due to the time limitation, we only use Average ensemble method to improve our performance. Traditional ensemble and Resnet ensemble cost about 8 hours to test on 81710 images by our GTX GPU 1080Ti.😱

• Average ensemble
  Before predicting the results via average ensemble, please first check the parameters in predict_ensemble.yaml are all correctly set. And directly run:

  python predict_ensemble.py
  

• Traditional ensemble
  Train the ensemble mlp via running train_ensemble_mlp.py.

• Resnet ensemble
  Train the res-ensemble net via running faster_res_ensemble_train.py and predict the output label using faster_res_ensemble_test.py.

Final result

• Score (accuracy)

  • Public dataset: 90.00%

  • Private dataset: 78.03%

  • General final score¹: 81.63%

  • Specific orchids: 96.15%

    	Public set	Private set	General final score
    Best accuracy	0.900077	0.780395	0.816300277

• Official final leaderboard

  • Leaderboard: leaderboard pdf
  • Registration teams: 743
  • Participating teams: 275
  • Our (TEAM_482) final rank: 18-th²

Reference

• https://github.com/rwightman/pytorch-image-models.git

Contact us

• Chi-Mao Fan (leader): [e-mail]
• Yu-Chen Su: [e-mail]
• Wei-Hsiang Liao: [e-mail]

Footnotes

1. General final score = 0.3xPublic + 0.7xPrivate ↩

2. Winners depend on the baseline score where general final score > 0.79 ↩