- This repository is the pytorch implement of the paper: MobileFaceNets: Efficient CNNs for Accurate Real-Time Face Verification on Mobile Devices and I almost follow the implement details of the paper.
- I train the model on CASIA-WebFace dataset, and evaluate on LFW dataset.
- Python 3.5
- pytorch 0.4+
- GPU menory
- All images of dataset are preprocessed following the SphereFace and you can download the aligned images at Align-CASIA-WebFace@BaiduDrive and Align-LFW@BaiduDrive.
-
Change the CAISIA_DATA_DIR and LFW_DATA_DAR in
config.pyto your data path. -
Train the mobilefacenet model.
Note: The default settings set the batch size of 512, use 2 gpus and train the model on 70 epochs. You can change the settings in
config.pypython3 train.py
-
Test the model on LFW.
Note: I have tested
lfw_eval.pyon the caffe model at SphereFace, it gets the same result.python3 lfw_eval.py --resume --feature_save_dir--resume:path of saved model--feature_save_dir:path to save the extracted features (must be .mat file)
- You can just run the
lfw_eval.pyto get the result, the accuracy on LFW like this:
| Fold | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | AVE(ours) | Paper(112x96) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ACC | 99.00 | 99.00 | 99.00 | 98.67 | 99.33 | 99.67 | 99.17 | 99.50 | 100.00 | 99.67 | 99.30 | 99.18 |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent