Face Recognition

Models

1. SphereFace with A-Softmax
   • Review paper: link

Environment

• PyTorch 0.3.1
• Python 2.7.13
• Option1: TensorboardX (Use for visualization)

Training

In this project, there are two different parallel method have been implemented.

1. Data Parallel
   A good choice for this way when your class number is less than 300000(Maybe?).
   Start training with this cmd:

   python training.py --input_paths=/your/data/path/10000_caffe_format.lst --working_root=/your/path/sphereface_pytorch --max_epoch=100 --img_size=112 --feature_dim=512 --label_num=10000 --process_num=15 --learning_rate=0.1 --model=SphereFaceNet --model_params='{"lamb_iter":0,"lamb_base":1000,"lamb_gamma":0.0001,"lamb_power":1,"lamb_min":10, "layer_type": "20layer"}' --try=0 --gpu_device=0,1,2,3 --batch_size=128 --parallel_mode='DataParallel'
   

2. Model Parallel
   We can not fit full model into one GPU when class number is huge in last margin fc layer.
   So? The model parallel is a good idea!!!
   Start training with this cmd:

   python training.py --input_paths=/your/data/path/1000000_caffe_format.lst --working_root=/your/path/sphereface_pytorch --max_epoch=100 --img_size=112 --feature_dim=512 --label_num=1000000 --process_num=15 --learning_rate=0.1 --model=SphereFaceNet --model_params='{"lamb_iter":0,"lamb_base":1000,"lamb_gamma":0.00001,"lamb_power":1,"lamb_min":10, "layer_type": "20layer"}' --try=0 --gpu_device=0,1,2,3 --batch_size=128 --parallel_mode='ModelParallel'
   

Testing & Benchmark

TODO

References

Pytorch
sphereface
sphereface