lgcn

Linkage-based Face Clustering via GCN

Main modification:

• refactory the training and testing with Runner from mmcv, which is modularized for easy extension.
• replace torch.Tensor with np.ndarray in dataloader, which makes the dataloader usable in frameworks in addition to PyTorch.
• evaluate lgcn under the same setting as dsgcn.

Test

Download the pretrained models in the model zoo.

Test

# Testing takes about 3 hours on 1 TitanX.
sh scripts/lgcn/test_lgcn_ms1m.sh

Train

We use the training parameters with best performance in our experiments as the default config.

# Training takes about 27 hours on 1 TitanX.
sh scripts/lgcn/train_lgcn_ms1m.sh

If there is better training config, you are welcome to report to us.

Reference

• Paper: https://arxiv.org/abs/1903.11306
• Code: https://github.com/Zhongdao/gcn_clustering