FPT-AI-DATA-COMPETITION

This repository clone from FPT-AI

Update:

• src/: contains all utils files for EDA.
• eda/: contains all files for EDA (NOTE: don't commit notebook file)
• download_data.sh: file bash for downloading the dataset.

1. Create virtual environment:

conda create -n fptai python=3.7
conda activate fptai

2. Clone this repository:

git clone https://github.com/DatacollectorVN/fpt-ai-data-competition.git

3. Install required packages:

pip install -r requirements.txt

4. Download the standard and additional data after processing:

Run to download raw dataset:

bash download_data.sh

Check annotation:

We use Streamlit to display and check annotations of image.

streamlit run eda/streamlit_annotations.py

Data pre-processing:

• Increase brightness

python eda/increase_brightness.py

• Enhence face of people

python eda/enhence_face.py

NOTE: Remember to change config correctly

Data augmentation:

• Mosaic | Flip | Rotate | Mixup

python src/{augmentation_name}_augmentation.py

NOTE: Change path of dataset and number images to generate

• Auto augmentation based on Yolov5 source code

python auto_augmentation.py

5. Baseline:

Val:

Public_test:

For more details DRIVE-CHUNG

6. Train:

• On Google Colab: (Note: Make a copy in drive)

• On server:

python train.py --batch-size 32 --device 0 --name <version_name> 

Note: Change the number of epochs to 70 in config/train_cfg.yaml

7. Evaluation:

python val.py --weights results/train/<version_name>/weights/best.pt  --task test --name <version_name> --batch-size 64 --device 0
                                                                             val
                                                                             train

• Results are saved at results/evaluate/<task>/<version_name>.

8. Prediction:

• Results are saved at <save_dir>.

python detect.py --weights results/train/<version_name>/weights/best.pt --source <path_to_folder> --dir <save_dir> --device 0

Note:

• <path_to_folder>: folder contain images to predict (Usually ./dataset/public_test)
• <save_dir>: path to save images predict

9. Result on leaderboard:

10. Result official:

In the final result, our team finished 15th out of 394 participating teams. We are very happy with this result and will try to do better in the upcoming competitions.

Try your best 🔥