ArcFace in PyTorch, for the MXNET-weary

What it does This repo ports the ArcFace/InsightFace face-recognition pipeline to PyTorch 0.4.1. It bundles the backbone architectures—IR-SE50 and MobileFacenet—with pretrained weights that hit competitive accuracy on standard benchmarks like LFW and CFP-FP. You can enroll identities by dropping images into folders, average their embeddings, and run verification against a live camera or video feed.

The interesting bit The author includes a converter that turns the original MXNET record files into plain image folders, which lowers the barrier if you are migrating from the reference implementation. It is essentially a compatibility layer with batteries included.

Key highlights

• Pretrained IR-SE50 and MobileFacenet models posted to BaiduNetDisk and OneDrive
• Benchmarked on seven standard face-recognition datasets (LFW, CFP-FF, CFP-FP, AgeDB-30, calfw, cplfw, vgg2_fp)
• Supports facebank-style enrollment: drop images into folders and the model averages embeddings per identity
• Camera and video inference modes included
• Training code provided for both backbones

Caveats

• Locked to PyTorch 0.4.1, which is several major releases behind current stable versions
• The README’s clone command accidentally points to a different MTCNN repository, suggesting the docs have some copy-paste drift
• Author notes they lack compute to train the deeper 100-layer and 151-layer variants

Verdict Useful for researchers or hobbyists who want a working PyTorch face-recognition baseline without wrestling with MXNET. Skip it if you need modern PyTorch features or the latest large-scale backbones.