PyTorch implementation of Seesaw-shuffleFaceNet
Differences from official implementation are minimal:
- Li-ArcFace loss
- Zero γ
- standard SE Block (no bias, no prelu)
- ReLU after first two convolutions
- Hard versions for Swish and Sigmoid
Pretrained model (TorchScript)
Validation metrics can be computed by running python3 eval.py <path to model> <path to a folder containing InsightFace bin files>:
cfp_ff: 99.56%
lfw: 99.62%
agedb_30: 96.12%
vgg2_fp: 93.24%
cfp_fp: 94.49%
cplfw: 89.52%
For training code see foamliu/InsightFace-v3 and similar projects.
Tested with nightly builds of PyTorch (1.4.0.dev20191216).