We used 500,000 Asian amounts.
A total of 7 emotions were used: happy, angry, anxious, peaceful, pain, and sad.
We used YOLOv8-face as a preprocessing method to cut out only the face to save manpower and learning time.
Repvgg, VIT (Vision Transformer), and YOLO were used as 2-stage models, and YOLO was used as 1-stage model.
Streamlit adds various new features.
GPU server : 4GPU A-100 (AWS) OS : Linux Language : Python
- DoYeon Kim - Team Leader
- HyunJun Kang
- Jongseong Kim
- Chaewook Lee
- HaNeul Pyeon
Experiment Report | Presentation Materials |
---|---|
결과보고서.pdf | 감정AI발표자료.pdf |
결과보고서.docx |
git clone https://github.com/JamieSKinard/est_wassup_03.git
cd est_wassup_03
pip install -r requirements.txt
pip install -e .
or
# using conda
conda env create -f env.yaml
# check default path
python preprocess/preprocess.py --data-dir {your_data_path}
Create a cropped photo after face recognition with the yolo8n-face model
python main.py --data-dir {your_data_path}
### Check defult
### Change model -> choice(defalut = RepVGG, VIT)
python main.py -mn VIT --data-dir {your_data_path} -mp {your_model} -mn {Reppvgg or VIT}
History and model aved in Models folder/{your_choice_model}
Preprocessing is possible with box_labeling_yolov8.ipynb in the folder called preprocess. and Go to the file named YOLO.ipynb and Just Shift + F5
python eval.py --data-dir {your_test_folder_path} -mp {your_model_path} -mn {Repvgg, VIT}
We used f1, R2, Precision, and recall as metrics.
YOLO(1Stage) | YOLO(2Stage) | ReppVgg | VIT | |
---|---|---|---|---|
val_loss | 0.233 | 0.533 | 1.470 | 1.251 |
train_acc | 78.3% | 72.2% | ||
val_acc | 73.9% | 68.7% | 62.5% |