cuongnm5 / zac-2022

Zalo AI Challenge 2022 - Liveness Detection - 1st place solution

1. INSTALLATION

1.1 Phần cứng

• 2 x NVIDIA TESLA A100 40G GPU
• Ubuntu 18.04.5 LTS
• CUDA 11.4
• Python 3.6.9

1.2 Cài đặt môi trường:

conda env create -f environment.yml
conda activate zac2022

2. DATASET

2.1 Chuẩn bị dữ liệu huấn luyện

mkdir raw
mkdir train_video&&cd train_video
wget https://dl-challenge.zalo.ai/liveness-detection/train.zip

unzip train.zip
rm -rf train.zip
mv ./train/videos/*.mp4 .
mv ./train/label.csv ../raw
rm -rf train

2.2 Chuẩn bị dữ liệu test

cd ..
mkdir data&&cd data
wget https://dl-challenge.zalo.ai/liveness-detection/private_test.zip

unzip private_test.zip
rm -rf private_test.zip
mv ./private_test/videos/*.mp4 .
rm -rf private_test

2.3 Extract frames từ dữ liệu huấn luyện

cd ..
bash prepare_data.sh

2.4 Tạo sample test

python create_sample_sub.py

3. SOLUTION SUMMARY

4. TRAINING

4.1 Tải pretrained ConvNeXt

wget https://dl.fbaipublicfiles.com/convnext/convnext_xlarge_22k_1k_224_ema.pth -P weights
wget https://dl.fbaipublicfiles.com/convnext/convnext_large_22k_1k_224.pth -P weights

4.2 Huấn luyện

bash scripts/train_convnext_xlarge_fold0.sh
bash scripts/train_convnext_xlarge_fold1.sh
bash scripts/train_convnext_xlarge_fold2.sh
bash scripts/train_convnext_xlarge_fold3.sh
bash scripts/train_convnext_xlarge_fold4.sh


bash scripts/train_convnext_large_fold0.sh
bash scripts/train_convnext_large_fold1.sh
bash scripts/train_convnext_large_fold2.sh

5. INFERENCE

bash predict.sh

Có thể sử dụng checkpoints nhóm mình đã train bằng cách download từ google drive:

bash download_weights.sh
bash predict.sh

Output là file submission.csv sẽ được lưu vào folder result. Sử dụng file này để nộp kết quả lên hệ thống.

6. RESULTS

6.1 Comparision of ConvNeXt with different models

6.2 Ablation study

6.3 Final result

7. REFERENCES

• Zalo AI Challenge 2022

• timm

• ConvNeXt