sphereface_pytorch

To evaluate a trained model's performance on the LFW (Labeled Faces in the Wild) dataset, use the lfw_eval.py script. You must provide the path to your model checkpoint using the --model flag.

python lfw_eval.py --model model/sphere20a_20171020.pth

If your LFW dataset is in .tgz format, use the following commands to extract it and repackage it into a .zip format required by the evaluation script:

1. Extract the tarball: tar zxf lfw.tgz
2. Enter the directory: cd lfw
3. Zip the contents: zip -r ../lfw.zip *
4. Return to the parent directory: cd ..

tar zxf lfw.tgz; cd lfw; zip -r ../lfw.zip *; cd ..

To start the training process for SphereFace, execute the train.py script. The implementation is designed to be trained on the CASIA-Webface dataset.

python train.py

You can use pre-trained models to achieve high accuracy on the LFW dataset without training from scratch. The following model is available:

• Model Name: 20171020
• LFW Accuracy: 0.9922
• Training Dataset: CASIA-WebFace