LiteFace - InsightFace with TensorFlow Lite

With LiteFace we convert the state-of-the-art face detection and recognition models InsightFace, from MXNet to TensorFlow Lite to be deployed and used in Android, iOS, embedded devices etc... for real-time face detection and recognition.
There is no other documented way of doing this.

Installation - Reference Important Notes Below

Clone and install everything we need:

git clone https://github.com/deepinsight/insightface.git
pip3 install mmdnn
cd insightface/python-package
pip3 install -e .

Download all InsightFace models, they should be in ~/.insightface/models/:

python3 << END
from insightface.app import FaceAnalysis; FaceAnalysis()
END

Convert ArcFace face recognition model to TensorFlow.

mmconvert -sf mxnet -in ~/.insightface/models/arcface_r100_v1/model-symbol.json -iw ~/.insightface/models/arcface_r100_v1/model-0000.params -df tensorflow -om tf_arcface_100_v1 --inputShape 3,112,112 --dump_tag SERVING

Convert ArcFace TensorFlow model to TFLite:

tflite_convert \
  --output_file=tf_arcface_100_v1/tf_arcface_100_v1.tflite \
  --saved_model_dir=tf_arcface_100_v1

Since RetinaFace detection model has multiple outputs, you must modify the MMDNN/conversion/tensorflow/saver.py script like so:

Now run MMDNN to convert RetinaFace:

mmconvert -sf mxnet -in ~/.insightface/models/retinaface_r50_v1/R50-symbol.json -iw ~/.insightface/models/retinaface_r50_v1/R50-0000.params -df tensorflow -om tf_retinaface_r50_v1 --inputShape 3,480,640 --dump_tag SERVING

Convert RetinaFace TensorFlow model to TFLite:

tflite_convert \
  --output_file=tf_retinaface_r50_v1/retinaface_r50_v1.tflite \
  --saved_model_dir=tf_retinaface_r50_v1

Important Notes

The appropriate MXnet and TensorFlow installations for your system are required (GPU/CPU, specific CUDA version).

---

If you run into missing package errors e.g you miss python-opencv then run pip3 install for that package.

---

If you run into:

from .cv2 import *
ImportError: libSM.so.6: cannot open shared object file: No such file or directory

Install these packages:

apt-get install libsm6 libxext6 libxrender-dev

---

If you run into:

ValueError: Object arrays cannot be loaded when allow_pickle=False

That's because your numpy version is too recent where allow_pickle=False by default. Downgrade:

pip3 install numpy==1.16.1

---