MobileFaceNet_TF

Tensorflow implementation for MobileFaceNet.
This version is different from the original repo sirius-ai/MobileFaceNet_TF:

• Model checkpoint and re-train continuously is more easier. This is suitable for model training progress on Google Colaboratory.
• Bug fixes
• Convert to Tensorflow Lite format
• Provide examples on ipython notebooks

---

Pre-trained model

Pretrain is available at arch/pretrained_model/MobileFaceNet_TFLite/.

---

Dependencies

• tensorflow 1.x >= r1.5
• opencv-python 3.x
• python 3.x
• scipy
• sklearn
• numpy
• mxnet
• pickle

---

Prepare dataset

1. Choose one of the following links to download dataset which is provide by insightface. (Special Recommend MS1M-refine-v2)

• MS1M-refine-v2@BaiduDrive, MS1M-refine-v2@GoogleDrive
• Refined-MS1M@BaiduDrive, Refined-MS1M@GoogleDrive
• VGGFace2@BaiduDrive, VGGFace2@GoogleDrive
• Insightface Dataset Zoo

2. Move dataset to ${MobileFaceNet_TF_ROOT}/datasets.
3. Run ${MobileFaceNet_TF_ROOT}/utils/data_process.py.

Take a look at data_prepair.ipynb for more details.

---

Training

1. Refined super parameters by yourself special project.
2. Run script ${MobileFaceNet_TF_ROOT}/train_nets.py
3. Have a snapshot result at ${MobileFaceNet_TF_ROOT}/output.

Take a look at train.ipynb for more details.

---

Convert to Tensorflow Lite format

1. Choose your best model checkpoint.
2. Freeze graph and get the .pb model file using freeze_graph.py

python3 freeze_graph.py \
--pretrained_model ./arch/pretrained_model/MobileFaceNet_TFLite/checkpoints/895000_MobileFaceNet.ckpt   \
--output_file ./arch/pretrained_model/MobileFaceNet_TFLite/MobileFaceNet.pb

3. Convert to TFLite format
   For example:

tflite_convert \
--output_file ./arch/pretrained_model/MobileFaceNet_TFLite/MobileFaceNet.tflite  \
--graph_def_file ./arch/pretrained_model/MobileFaceNet_TFLite/MobileFaceNet.pb  \
--input_arrays "input" \
--input_shapes "1,112,112,3"  \
--output_arrays embeddings \
--output_format TFLITE

You may need to choose the suitable inference batch size. The previous example use batchsize = 1.

Take a look at test.ipynb for more details.

---

Performance

size	LFW(%)	Val@1e-3(%)	inference@GPU: Qualcomm Adreno TM 509(ms)
4.9MB	99.4+	98.3+	30-

---

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