移动端facenet的人脸识别问题。

使用Keras实现MobileFaceNet模型或使用TensorFlow Lite来压缩模型大小并优化性能。

以Keras实现MobileFaceNet模型为例：

import tensorflow as tf
from tensorflow.keras import models, layers

class ConvBlock(layers.Layer):
    def __init__(self, out_channels, strides=1, **kwargs):
        super(ConvBlock, self).__init__(**kwargs)
        
        self.conv = layers.Conv2D(out_channels, kernel_size=3, strides=strides, padding="same")
        self.bn = layers.BatchNormalization()
        self.relu = layers.ReLU()
        
    def call(self, inputs):
        x = self.conv(inputs)
        x = self.bn(x)
        x = self.relu(x)
        return x
    
class MobileFaceNet(models.Model):
    def __init__(self, embeddings_size=128, **kwargs):
        super(MobileFaceNet, self).__init__(**kwargs)
        
        self.conv1 = ConvBlock(64, strides=2)
        self.conv2_dw = ConvBlock(64, strides=1)
        self.conv3_dw = ConvBlock(128, strides=2)
        self.conv4_dw = ConvBlock(128, strides=1)
        self.conv5_dw = ConvBlock(128, strides=1)
        self.conv6_dw = ConvBlock(128, strides=1)
        self.conv7_dw = ConvBlock(256, strides=2)
        self.conv8_dw = ConvBlock(256, strides=1)
        self.conv9_dw = ConvBlock(256, strides=1)
        self.conv10_dw = ConvBlock(256, strides=1)
        self.conv11_dw = ConvBlock(256, strides=1)
        self.conv12_dw = ConvBlock(256, strides=1)
        self.conv13_dw = ConvBlock(512, strides=2)
        self.conv14_dw = ConvBlock(512, strides=1)
        self.conv15_pw = ConvBlock(128, strides=1)
        self.conv16_pw = ConvBlock(512, strides=2)
        self.conv17_pw = ConvBlock(512, strides=1)
        self.conv18_pw = ConvBlock(512, strides=1)
        self.conv19