/LightFaceNet

This tensorflow project is based on MobileFaceNet, which use SqueezeNet and ShuffleNet as stem CNN to improve speed.

Primary LanguagePython

MobileFaceNet_Tensorflow

Tensorflow implementation for MobileFaceNet which is modified from MobileFaceNet_TF

Requirements

  • tensorflow >= r1.2 (support cuda 8.0, original needs tensorflow >= r1.5 and cuda 9.0)
  • opencv-python
  • python 3.x ( if you want to use python 2.x, somewhere in load_data function need to change, see details in comment)
  • mxnet
  • anaconda (recommend)

Construction

├── MobileFaceNet
│   ├── arch
│       ├── img
│       ├── txt
│   ├── datasets
│       ├── faces_ms1m_112x112
│       ├── tfrecords
│   ├── losses
│   ├── nets
│   ├── output
│       ├── ckpt
│       ├── ckpt_best
│       ├── logs
│       ├── summary
│   ├── utils

Datasets

  1. choose one of The following links to download dataset which is provide by insightface. (Special Recommend MS1M)
  1. move dataset to ${MobileFaceNet_TF_ROOT}/datasets.
  2. run ${MobileFaceNet_TF_ROOT}/utils/data_process.py.

Training

train_nets.py --max_epoch=10
              --train_batch_size=128
              --model_type=0  # 0-mobilefacenet 1-tinymobilefacenet 2-squeezefacenet 3-shufflefacenet 4-shufflefacenetV2
              --gpu=0

Inference

MobileFaceNet

python inference.py --pretrained_model='./output/ckpt_best/mobilefacenet_best_ckpt'
                    --model_type=0
                    --gpu=0

Performance

Original result

size LFW(%) Val@1e-3(%) inference@MSM8976(ms)
5.7M 99.25+ 96.8+ 260-

My training results

Models LFW Cfp_FF Cfp_FP Agedb_30 inference@i7-7700 16G 240G (fps)
MobileFaceNet(Bad training) 0.983+-0.008 0.980+-0.005 0.827+-0.019 0.878+-0.023 27
Tiny_MobileFaceNet 0.981+-0.008 0.984+-0.006 0.835+-0.019 0.882+-0.023 50
SqueezeFaceNet 0.972+-0.008 0.962+-0.006 0.785+-0.019 0.837+-0.023 83
ShuffleFaceNet v1 0.962+-0.008 0.941+-0.006 0.763+-0.019 0.747+-0.023 33
ShuffleFaceNet v2 0.980+-0.008 0.971+-0.006 0.820+-0.019 0.823+-0.023 50

References

  1. facenet
  2. InsightFace mxnet
  3. InsightFace_TF
  4. MobileFaceNets: Efficient CNNs for Accurate Real-Time Face Verification on Mobile Devices
  5. CosFace: Large Margin Cosine Loss for Deep Face Recognition
  6. InsightFace : Additive Angular Margin Loss for Deep Face Recognition
  7. tensorflow-triplet-loss
  8. MobileFaceNet_TF
  9. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size
  10. ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices