Model in "face_detection_front.tflite" is inconsistent with the model in the paper "BlazeFace"

[liqiang311]

Hi

I studied this paper BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs, which led me to download face_detection_front.tflite.

But their model is completely different. I read tflite using tensorflow's tf.lite.Interpreter, but I didn't find the double BlazeBlock in the paper.

In addition, the paper mentions the use of 5x5 DepthWise to replace 3x3 Depthwise, but the tflite model uses 3x3 Depthwise.

Why is this?

Thank you.

[mgyong]

@liqiang311 See answer from bazarevsky in #21