Sima Face-Recognition

Sima is an Open-source Pytorch face Recognition project. It`s based on ResNet architecture and TripletLoss(This loss has been developed by google team).
We use CASIA-WebFace and Sima Dataset for training and LFW dataset for validation.

You can download pre-trained model(ResNet18) from here. This model get 78% Accuracy on LFW dataset in verification mode and with 11.000.000 patameters.

My goal is to train ResNet34 and get better Accuracy on the LFW dataset. After that, I want to make a transfer learning for the Iranian Face dataset(Sima Dataset).

Deployment

Explain included files:

main.py : Start training loop.
loss.py : Implemetation of triplet loss function.
resnet.py : Implemetation of ResNet 18 & 34 architecture.
utils.py : Some function.

Explain Folders:

IranianFace-Dataset : You can use nootbooks that are in this folder for making a dataset based on instagram id.
LFW_Pairs: Making .npy list file for lfw pairs path

TODO:

Train on ResNet34 Arch.
Transfer Learning on Sima Dataset for get better accuracy on sima dataset
Using DenseNet

Authors

Reza Tanakizadeh

License

MIT

Feedback