4th Chalearn Face Anti-spoofing Workshop and Challenge@CVPR2023 —— Team SeaRecluse

Step

Install dependencies:

pip install -r requirements.txt

Data preprocessing:

If you have full content data data, it should look like this in the folder
--orig_data
    -dev *
    -test *
    -train *
    -dev.txt *
    -dev_label.txt *
    -test.txt *
    -train_label.txt *
    -data_arrange.py

Divide the dataset

# We use all the train data and 2/3 of the val data as the final training set, and the remaining val data as the validation dataset.

python data_arrange.py

Start training

# We have usd a pre-trained model based on ImageNet provided by timm. Please ensure that your network is accessible to download this model.
# Put the enhanced data into the folder "./orig_data" for training.
# It is recommended that the model training is not less than 300 epoch, default is 500 epoch.
# The default batch size is 64, which requires at least 24G GPU memory for training.

python main.py

Test your model

# When the training is over, you can copy the checkpoint.pth.tar in reulst to the root directory and use it for testing.

python test.py

other

The Flops requirement of this competition model is less than 100G, but this is still a huge number. If the organizer has sufficient training resources and can use a larger backbone, we believe that better results will be achieved.