A Keras implementation of MobileFaceNet from https://arxiv.org/abs/1804.07573. It is the 1st time I uploaded my own work. If the license or citation is wrong, please inform me and I will change it ASAP. The Keras model for inference is ~14 Mb, as well as the TensorFlow model in optimized protobuf format, so I might not follow all the details in the original work (should be around 5 Mb).
(1) Use the celebrity & msra datasets from the Trillion Pairs dataset: http://trillionpairs.deepglint.com/data.
(2) For each identity folder:
a. n > 350
Randomly pick 350 pics from the origin data set
b. 200 < n <= 350
Keep all the pics
c. 90 < n <= 200
Keep all the pics & Apply Opening to them (Double the data)
d. 30 < n <= 90
Keep all the pics, Apply Opening to them, Add Gaussian noise to them & Add Salt & Pepper noise to them (Four times the data)
e. n <= 30
Drop the folder
(3) Crop & resize:
a. Use MTCNN to locate the face (which has already been aligned) & Crop it out
b. Resize the image to 112 x 112
(4) PS: The OpenCV ops in Python run really slow and all the operations above have a quite similar code implementation in C++. Thus it is recommended to use C++ for prerocessing the image dataset.
At first I did not finish the data preprocessing step and use parts of the processed data for training with 512 as the mini batch size. Since I only bought 1 Nvidia 1080 Ti GPU myself and its memory is not enough when the dataset becomes larger (there is one intermediate layer whose size is related to the number of labels, while now I have 26928 labels for 6151666 pics, which makes that layer the shape (26928, 26928) in float32 and needs more than 21 Gb space, meaning that even 2 2080 Ti (12 Gb each) is not enough yet). So for now the training has been paused and I am waiting for upgrading my hardware when I get money.
(1) Original paper: https://arxiv.org/abs/1804.07573
(2) The idea of the implementation of model structure is from MobileNet v2: https://github.com/xiaochus/MobileNetV2
(3) The implementation of ArcFace loss (InsightFace loss) is from: https://github.com/ewrfcas/Machine-Learning-Toolbox/blob/127d6e5d336614d1efb21e78865501435cdb7b8b/loss_function/ArcFace_loss.py