The following codes are the solutions (3st place, private score: 0.98274) for the dacon competition.
git clone https://github.com/GNOEYHEAT/LowResolution_ImgClf.git
cd LowResolution_ImgClf
conda create -n bird_cls python=3.10
conda activate bird_cls
conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -r requirements.txt
- Tested on NVIDIA RTX 3090, A100
- The dataset can be downloaded from the dacon link.
- Place the downloaded data inside
lowResolution_ImgClf
directory,
LowResolution_ImgClf
├── data
│ ├── train
│ │ ├── TRAIN_00000.jpg
│ │ ├── ...
│ ├── test
│ │ ├── TEST_00000.jpg
│ │ ├── ...
│ ├── upscale_train
│ │ ├── TRAIN_00000.jpg
│ │ ├── ...
│ ├── train.csv
│ ├── test.csv
│ └── sample_submission.csv
or modify train_csv_path
and test_csv_path
in configs/base.yaml
. It is recommended to input the absolute paths for train_csv_path
and test_csv_path
.
- If you want to train the model and generate the submission file all at once, execute the script below.
python main.py
- If you only want to train, execute the script below.
python scripts/train.py
- If you want to generate the submission file using prediction files, execute the script below.
python scripts/submission.py