retinaface_model.py

import tensorflow as tf
import gdown
from pathlib import Path
import os

tf_version = int(tf.__version__.split(".")[0])

if tf_version == 1:
    from keras.models import Model
    from keras.layers import Input, BatchNormalization, ZeroPadding2D, Conv2D, ReLU, MaxPool2D, Add, UpSampling2D, concatenate, Softmax

else:
    from tensorflow.keras.models import Model
    from tensorflow.keras.layers import Input, BatchNormalization, ZeroPadding2D, Conv2D, ReLU, MaxPool2D, Add, UpSampling2D, concatenate, Softmax

def load_weights(model):

    exact_file = "retinaface.h5"

    model.load_weights(exact_file)

    return model

def build_model():

    data = Input(dtype=tf.float32, shape=(None, None, 3), name='data')

    bn_data = BatchNormalization(epsilon=1.9999999494757503e-05, name='bn_data', trainable=False)(data)

    conv0_pad = ZeroPadding2D(padding=tuple([3, 3]))(bn_data)

    conv0 = Conv2D(filters = 64, kernel_size = (7, 7), name = 'conv0', strides = [2, 2], padding = 'VALID', use_bias = False)(conv0_pad)

    bn0 = BatchNormalization(epsilon=1.9999999494757503e-05, name='bn0', trainable=False)(conv0)

    relu0 = ReLU(name='relu0')(bn0)

    pooling0_pad = ZeroPadding2D(padding=tuple([1, 1]))(relu0)

    pooling0 = MaxPool2D((3, 3), (2, 2), padding='VALID', name='pooling0')(pooling0_pad)

    stage1_unit1_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit1_bn1', trainable=False)(pooling0)

    stage1_unit1_relu1 = ReLU(name='stage1_unit1_relu1')(stage1_unit1_bn1)

    stage1_unit1_conv1 = Conv2D(filters = 64, kernel_size = (1, 1), name = 'stage1_unit1_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit1_relu1)

    stage1_unit1_sc = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage1_unit1_sc', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit1_relu1)

    stage1_unit1_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit1_bn2', trainable=False)(stage1_unit1_conv1)

    stage1_unit1_relu2 = ReLU(name='stage1_unit1_relu2')(stage1_unit1_bn2)

    stage1_unit1_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage1_unit1_relu2)

    stage1_unit1_conv2 = Conv2D(filters = 64, kernel_size = (3, 3), name = 'stage1_unit1_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit1_conv2_pad)

    stage1_unit1_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit1_bn3', trainable=False)(stage1_unit1_conv2)

    stage1_unit1_relu3 = ReLU(name='stage1_unit1_relu3')(stage1_unit1_bn3)

    stage1_unit1_conv3 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage1_unit1_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit1_relu3)

    plus0_v1 = Add()([stage1_unit1_conv3 , stage1_unit1_sc])

    stage1_unit2_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit2_bn1', trainable=False)(plus0_v1)

    stage1_unit2_relu1 = ReLU(name='stage1_unit2_relu1')(stage1_unit2_bn1)

    stage1_unit2_conv1 = Conv2D(filters = 64, kernel_size = (1, 1), name = 'stage1_unit2_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit2_relu1)

    stage1_unit2_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit2_bn2', trainable=False)(stage1_unit2_conv1)

    stage1_unit2_relu2 = ReLU(name='stage1_unit2_relu2')(stage1_unit2_bn2)

    stage1_unit2_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage1_unit2_relu2)

    stage1_unit2_conv2 = Conv2D(filters = 64, kernel_size = (3, 3), name = 'stage1_unit2_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit2_conv2_pad)

    stage1_unit2_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit2_bn3', trainable=False)(stage1_unit2_conv2)

    stage1_unit2_relu3 = ReLU(name='stage1_unit2_relu3')(stage1_unit2_bn3)

    stage1_unit2_conv3 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage1_unit2_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit2_relu3)

    plus1_v2 = Add()([stage1_unit2_conv3 , plus0_v1])

    stage1_unit3_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit3_bn1', trainable=False)(plus1_v2)

    stage1_unit3_relu1 = ReLU(name='stage1_unit3_relu1')(stage1_unit3_bn1)

    stage1_unit3_conv1 = Conv2D(filters = 64, kernel_size = (1, 1), name = 'stage1_unit3_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit3_relu1)

    stage1_unit3_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit3_bn2', trainable=False)(stage1_unit3_conv1)

    stage1_unit3_relu2 = ReLU(name='stage1_unit3_relu2')(stage1_unit3_bn2)

    stage1_unit3_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage1_unit3_relu2)

    stage1_unit3_conv2 = Conv2D(filters = 64, kernel_size = (3, 3), name = 'stage1_unit3_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit3_conv2_pad)

    stage1_unit3_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit3_bn3', trainable=False)(stage1_unit3_conv2)

    stage1_unit3_relu3 = ReLU(name='stage1_unit3_relu3')(stage1_unit3_bn3)

    stage1_unit3_conv3 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage1_unit3_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit3_relu3)

    plus2 = Add()([stage1_unit3_conv3 , plus1_v2])

    stage2_unit1_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit1_bn1', trainable=False)(plus2)

    stage2_unit1_relu1 = ReLU(name='stage2_unit1_relu1')(stage2_unit1_bn1)

    stage2_unit1_conv1 = Conv2D(filters = 128, kernel_size = (1, 1), name = 'stage2_unit1_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit1_relu1)

    stage2_unit1_sc = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit1_sc', strides = [2, 2], padding = 'VALID', use_bias = False)(stage2_unit1_relu1)

    stage2_unit1_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit1_bn2', trainable=False)(stage2_unit1_conv1)

    stage2_unit1_relu2 = ReLU(name='stage2_unit1_relu2')(stage2_unit1_bn2)

    stage2_unit1_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage2_unit1_relu2)

    stage2_unit1_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'stage2_unit1_conv2', strides = [2, 2], padding = 'VALID', use_bias = False)(stage2_unit1_conv2_pad)

    stage2_unit1_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit1_bn3', trainable=False)(stage2_unit1_conv2)

    stage2_unit1_relu3 = ReLU(name='stage2_unit1_relu3')(stage2_unit1_bn3)

    stage2_unit1_conv3 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit1_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit1_relu3)

    plus3 = Add()([stage2_unit1_conv3 , stage2_unit1_sc])

    stage2_unit2_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit2_bn1', trainable=False)(plus3)

    stage2_unit2_relu1 = ReLU(name='stage2_unit2_relu1')(stage2_unit2_bn1)

    stage2_unit2_conv1 = Conv2D(filters = 128, kernel_size = (1, 1), name = 'stage2_unit2_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit2_relu1)

    stage2_unit2_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit2_bn2', trainable=False)(stage2_unit2_conv1)

    stage2_unit2_relu2 = ReLU(name='stage2_unit2_relu2')(stage2_unit2_bn2)

    stage2_unit2_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage2_unit2_relu2)

    stage2_unit2_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'stage2_unit2_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit2_conv2_pad)

    stage2_unit2_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit2_bn3', trainable=False)(stage2_unit2_conv2)

    stage2_unit2_relu3 = ReLU(name='stage2_unit2_relu3')(stage2_unit2_bn3)

    stage2_unit2_conv3 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit2_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit2_relu3)

    plus4 = Add()([stage2_unit2_conv3 , plus3])

    stage2_unit3_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit3_bn1', trainable=False)(plus4)

    stage2_unit3_relu1 = ReLU(name='stage2_unit3_relu1')(stage2_unit3_bn1)

    stage2_unit3_conv1 = Conv2D(filters = 128, kernel_size = (1, 1), name = 'stage2_unit3_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit3_relu1)

    stage2_unit3_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit3_bn2', trainable=False)(stage2_unit3_conv1)

    stage2_unit3_relu2 = ReLU(name='stage2_unit3_relu2')(stage2_unit3_bn2)

    stage2_unit3_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage2_unit3_relu2)

    stage2_unit3_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'stage2_unit3_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit3_conv2_pad)

    stage2_unit3_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit3_bn3', trainable=False)(stage2_unit3_conv2)

    stage2_unit3_relu3 = ReLU(name='stage2_unit3_relu3')(stage2_unit3_bn3)

    stage2_unit3_conv3 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit3_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit3_relu3)

    plus5 = Add()([stage2_unit3_conv3 , plus4])

    stage2_unit4_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit4_bn1', trainable=False)(plus5)

    stage2_unit4_relu1 = ReLU(name='stage2_unit4_relu1')(stage2_unit4_bn1)

    stage2_unit4_conv1 = Conv2D(filters = 128, kernel_size = (1, 1), name = 'stage2_unit4_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit4_relu1)

    stage2_unit4_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit4_bn2', trainable=False)(stage2_unit4_conv1)

    stage2_unit4_relu2 = ReLU(name='stage2_unit4_relu2')(stage2_unit4_bn2)

    stage2_unit4_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage2_unit4_relu2)

    stage2_unit4_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'stage2_unit4_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit4_conv2_pad)

    stage2_unit4_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit4_bn3', trainable=False)(stage2_unit4_conv2)

    stage2_unit4_relu3 = ReLU(name='stage2_unit4_relu3')(stage2_unit4_bn3)

    stage2_unit4_conv3 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit4_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit4_relu3)

    plus6 = Add()([stage2_unit4_conv3 , plus5])

    stage3_unit1_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit1_bn1', trainable=False)(plus6)

    stage3_unit1_relu1 = ReLU(name='stage3_unit1_relu1')(stage3_unit1_bn1)

    stage3_unit1_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit1_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit1_relu1)

    stage3_unit1_sc = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit1_sc', strides = [2, 2], padding = 'VALID', use_bias = False)(stage3_unit1_relu1)

    stage3_unit1_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit1_bn2', trainable=False)(stage3_unit1_conv1)

    stage3_unit1_relu2 = ReLU(name='stage3_unit1_relu2')(stage3_unit1_bn2)

    stage3_unit1_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit1_relu2)

    stage3_unit1_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit1_conv2', strides = [2, 2], padding = 'VALID', use_bias = False)(stage3_unit1_conv2_pad)

    ssh_m1_red_conv = Conv2D(filters = 256, kernel_size = (1, 1), name = 'ssh_m1_red_conv', strides = [1, 1], padding = 'VALID', use_bias = True)(stage3_unit1_relu2)

    stage3_unit1_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit1_bn3', trainable=False)(stage3_unit1_conv2)

    ssh_m1_red_conv_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_red_conv_bn', trainable=False)(ssh_m1_red_conv)

    stage3_unit1_relu3 = ReLU(name='stage3_unit1_relu3')(stage3_unit1_bn3)

    ssh_m1_red_conv_relu = ReLU(name='ssh_m1_red_conv_relu')(ssh_m1_red_conv_bn)

    stage3_unit1_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit1_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit1_relu3)

    plus7 = Add()([stage3_unit1_conv3 , stage3_unit1_sc])

    stage3_unit2_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit2_bn1', trainable=False)(plus7)

    stage3_unit2_relu1 = ReLU(name='stage3_unit2_relu1')(stage3_unit2_bn1)

    stage3_unit2_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit2_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit2_relu1)

    stage3_unit2_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit2_bn2', trainable=False)(stage3_unit2_conv1)

    stage3_unit2_relu2 = ReLU(name='stage3_unit2_relu2')(stage3_unit2_bn2)

    stage3_unit2_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit2_relu2)

    stage3_unit2_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit2_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit2_conv2_pad)

    stage3_unit2_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit2_bn3', trainable=False)(stage3_unit2_conv2)

    stage3_unit2_relu3 = ReLU(name='stage3_unit2_relu3')(stage3_unit2_bn3)

    stage3_unit2_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit2_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit2_relu3)

    plus8 = Add()([stage3_unit2_conv3 , plus7])

    stage3_unit3_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit3_bn1', trainable=False)(plus8)

    stage3_unit3_relu1 = ReLU(name='stage3_unit3_relu1')(stage3_unit3_bn1)

    stage3_unit3_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit3_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit3_relu1)

    stage3_unit3_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit3_bn2', trainable=False)(stage3_unit3_conv1)

    stage3_unit3_relu2 = ReLU(name='stage3_unit3_relu2')(stage3_unit3_bn2)

    stage3_unit3_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit3_relu2)

    stage3_unit3_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit3_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit3_conv2_pad)

    stage3_unit3_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit3_bn3', trainable=False)(stage3_unit3_conv2)

    stage3_unit3_relu3 = ReLU(name='stage3_unit3_relu3')(stage3_unit3_bn3)

    stage3_unit3_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit3_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit3_relu3)

    plus9 = Add()([stage3_unit3_conv3 , plus8])

    stage3_unit4_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit4_bn1', trainable=False)(plus9)

    stage3_unit4_relu1 = ReLU(name='stage3_unit4_relu1')(stage3_unit4_bn1)

    stage3_unit4_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit4_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit4_relu1)

    stage3_unit4_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit4_bn2', trainable=False)(stage3_unit4_conv1)

    stage3_unit4_relu2 = ReLU(name='stage3_unit4_relu2')(stage3_unit4_bn2)

    stage3_unit4_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit4_relu2)

    stage3_unit4_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit4_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit4_conv2_pad)

    stage3_unit4_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit4_bn3', trainable=False)(stage3_unit4_conv2)

    stage3_unit4_relu3 = ReLU(name='stage3_unit4_relu3')(stage3_unit4_bn3)

    stage3_unit4_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit4_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit4_relu3)

    plus10 = Add()([stage3_unit4_conv3 , plus9])

    stage3_unit5_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit5_bn1', trainable=False)(plus10)

    stage3_unit5_relu1 = ReLU(name='stage3_unit5_relu1')(stage3_unit5_bn1)

    stage3_unit5_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit5_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit5_relu1)

    stage3_unit5_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit5_bn2', trainable=False)(stage3_unit5_conv1)

    stage3_unit5_relu2 = ReLU(name='stage3_unit5_relu2')(stage3_unit5_bn2)

    stage3_unit5_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit5_relu2)

    stage3_unit5_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit5_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit5_conv2_pad)

    stage3_unit5_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit5_bn3', trainable=False)(stage3_unit5_conv2)

    stage3_unit5_relu3 = ReLU(name='stage3_unit5_relu3')(stage3_unit5_bn3)

    stage3_unit5_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit5_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit5_relu3)

    plus11 = Add()([stage3_unit5_conv3 , plus10])

    stage3_unit6_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit6_bn1', trainable=False)(plus11)

    stage3_unit6_relu1 = ReLU(name='stage3_unit6_relu1')(stage3_unit6_bn1)

    stage3_unit6_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit6_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit6_relu1)

    stage3_unit6_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit6_bn2', trainable=False)(stage3_unit6_conv1)

    stage3_unit6_relu2 = ReLU(name='stage3_unit6_relu2')(stage3_unit6_bn2)

    stage3_unit6_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit6_relu2)

    stage3_unit6_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit6_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit6_conv2_pad)

    stage3_unit6_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit6_bn3', trainable=False)(stage3_unit6_conv2)

    stage3_unit6_relu3 = ReLU(name='stage3_unit6_relu3')(stage3_unit6_bn3)

    stage3_unit6_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit6_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit6_relu3)

    plus12 = Add()([stage3_unit6_conv3 , plus11])

    stage4_unit1_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit1_bn1', trainable=False)(plus12)

    stage4_unit1_relu1 = ReLU(name='stage4_unit1_relu1')(stage4_unit1_bn1)

    stage4_unit1_conv1 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage4_unit1_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit1_relu1)

    stage4_unit1_sc = Conv2D(filters = 2048, kernel_size = (1, 1), name = 'stage4_unit1_sc', strides = [2, 2], padding = 'VALID', use_bias = False)(stage4_unit1_relu1)

    stage4_unit1_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit1_bn2', trainable=False)(stage4_unit1_conv1)

    stage4_unit1_relu2 = ReLU(name='stage4_unit1_relu2')(stage4_unit1_bn2)

    stage4_unit1_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage4_unit1_relu2)

    stage4_unit1_conv2 = Conv2D(filters = 512, kernel_size = (3, 3), name = 'stage4_unit1_conv2', strides = [2, 2], padding = 'VALID', use_bias = False)(stage4_unit1_conv2_pad)

    ssh_c2_lateral = Conv2D(filters = 256, kernel_size = (1, 1), name = 'ssh_c2_lateral', strides = [1, 1], padding = 'VALID', use_bias = True)(stage4_unit1_relu2)

    stage4_unit1_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit1_bn3', trainable=False)(stage4_unit1_conv2)

    ssh_c2_lateral_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_c2_lateral_bn', trainable=False)(ssh_c2_lateral)

    stage4_unit1_relu3 = ReLU(name='stage4_unit1_relu3')(stage4_unit1_bn3)

    ssh_c2_lateral_relu = ReLU(name='ssh_c2_lateral_relu')(ssh_c2_lateral_bn)

    stage4_unit1_conv3 = Conv2D(filters = 2048, kernel_size = (1, 1), name = 'stage4_unit1_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit1_relu3)

    plus13 = Add()([stage4_unit1_conv3 , stage4_unit1_sc])

    stage4_unit2_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit2_bn1', trainable=False)(plus13)

    stage4_unit2_relu1 = ReLU(name='stage4_unit2_relu1')(stage4_unit2_bn1)

    stage4_unit2_conv1 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage4_unit2_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit2_relu1)

    stage4_unit2_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit2_bn2', trainable=False)(stage4_unit2_conv1)

    stage4_unit2_relu2 = ReLU(name='stage4_unit2_relu2')(stage4_unit2_bn2)

    stage4_unit2_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage4_unit2_relu2)

    stage4_unit2_conv2 = Conv2D(filters = 512, kernel_size = (3, 3), name = 'stage4_unit2_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit2_conv2_pad)

    stage4_unit2_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit2_bn3', trainable=False)(stage4_unit2_conv2)

    stage4_unit2_relu3 = ReLU(name='stage4_unit2_relu3')(stage4_unit2_bn3)

    stage4_unit2_conv3 = Conv2D(filters = 2048, kernel_size = (1, 1), name = 'stage4_unit2_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit2_relu3)

    plus14 = Add()([stage4_unit2_conv3 , plus13])

    stage4_unit3_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit3_bn1', trainable=False)(plus14)

    stage4_unit3_relu1 = ReLU(name='stage4_unit3_relu1')(stage4_unit3_bn1)

    stage4_unit3_conv1 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage4_unit3_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit3_relu1)

    stage4_unit3_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit3_bn2', trainable=False)(stage4_unit3_conv1)

    stage4_unit3_relu2 = ReLU(name='stage4_unit3_relu2')(stage4_unit3_bn2)

    stage4_unit3_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage4_unit3_relu2)

    stage4_unit3_conv2 = Conv2D(filters = 512, kernel_size = (3, 3), name = 'stage4_unit3_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit3_conv2_pad)

    stage4_unit3_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit3_bn3', trainable=False)(stage4_unit3_conv2)

    stage4_unit3_relu3 = ReLU(name='stage4_unit3_relu3')(stage4_unit3_bn3)

    stage4_unit3_conv3 = Conv2D(filters = 2048, kernel_size = (1, 1), name = 'stage4_unit3_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit3_relu3)

    plus15 = Add()([stage4_unit3_conv3 , plus14])

    bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='bn1', trainable=False)(plus15)

    relu1 = ReLU(name='relu1')(bn1)

    ssh_c3_lateral = Conv2D(filters = 256, kernel_size = (1, 1), name = 'ssh_c3_lateral', strides = [1, 1], padding = 'VALID', use_bias = True)(relu1)

    ssh_c3_lateral_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_c3_lateral_bn', trainable=False)(ssh_c3_lateral)

    ssh_c3_lateral_relu = ReLU(name='ssh_c3_lateral_relu')(ssh_c3_lateral_bn)

    ssh_m3_det_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c3_lateral_relu)

    ssh_m3_det_conv1 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_m3_det_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_conv1_pad)

    ssh_m3_det_context_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c3_lateral_relu)

    ssh_m3_det_context_conv1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m3_det_context_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_context_conv1_pad)

    ssh_c3_up = UpSampling2D(size=(2, 2), interpolation="nearest", name="ssh_c3_up")(ssh_c3_lateral_relu)

    ssh_m3_det_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_conv1_bn', trainable=False)(ssh_m3_det_conv1)

    ssh_m3_det_context_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_context_conv1_bn', trainable=False)(ssh_m3_det_context_conv1)

    x1_shape = tf.shape(ssh_c3_up)
    x2_shape = tf.shape(ssh_c2_lateral_relu)
    offsets = [0, (x1_shape[1] - x2_shape[1]) // 2, (x1_shape[2] - x2_shape[2]) // 2, 0]
    size = [-1, x2_shape[1], x2_shape[2], -1]
    crop0 = tf.slice(ssh_c3_up, offsets, size, "crop0")

    ssh_m3_det_context_conv1_relu = ReLU(name='ssh_m3_det_context_conv1_relu')(ssh_m3_det_context_conv1_bn)

    plus0_v2 = Add()([ssh_c2_lateral_relu , crop0])

    ssh_m3_det_context_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m3_det_context_conv1_relu)

    ssh_m3_det_context_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m3_det_context_conv2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_context_conv2_pad)

    ssh_m3_det_context_conv3_1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m3_det_context_conv1_relu)

    ssh_m3_det_context_conv3_1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m3_det_context_conv3_1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_context_conv3_1_pad)

    ssh_c2_aggr_pad = ZeroPadding2D(padding=tuple([1, 1]))(plus0_v2)

    ssh_c2_aggr = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_c2_aggr', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_c2_aggr_pad)

    ssh_m3_det_context_conv2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_context_conv2_bn', trainable=False)(ssh_m3_det_context_conv2)

    ssh_m3_det_context_conv3_1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_context_conv3_1_bn', trainable=False)(ssh_m3_det_context_conv3_1)

    ssh_c2_aggr_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_c2_aggr_bn', trainable=False)(ssh_c2_aggr)

    ssh_m3_det_context_conv3_1_relu = ReLU(name='ssh_m3_det_context_conv3_1_relu')(ssh_m3_det_context_conv3_1_bn)

    ssh_c2_aggr_relu = ReLU(name='ssh_c2_aggr_relu')(ssh_c2_aggr_bn)

    ssh_m3_det_context_conv3_2_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m3_det_context_conv3_1_relu)

    ssh_m3_det_context_conv3_2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m3_det_context_conv3_2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_context_conv3_2_pad)

    ssh_m2_det_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c2_aggr_relu)

    ssh_m2_det_conv1 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_m2_det_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_conv1_pad)

    ssh_m2_det_context_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c2_aggr_relu)

    ssh_m2_det_context_conv1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m2_det_context_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_context_conv1_pad)

    ssh_m2_red_up = UpSampling2D(size=(2, 2), interpolation="nearest", name="ssh_m2_red_up")(ssh_c2_aggr_relu)

    ssh_m3_det_context_conv3_2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_context_conv3_2_bn', trainable=False)(ssh_m3_det_context_conv3_2)

    ssh_m2_det_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_conv1_bn', trainable=False)(ssh_m2_det_conv1)

    ssh_m2_det_context_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_context_conv1_bn', trainable=False)(ssh_m2_det_context_conv1)

    x1_shape = tf.shape(ssh_m2_red_up)
    x2_shape = tf.shape(ssh_m1_red_conv_relu)
    offsets = [0, (x1_shape[1] - x2_shape[1]) // 2, (x1_shape[2] - x2_shape[2]) // 2, 0]
    size = [-1, x2_shape[1], x2_shape[2], -1]
    crop1 = tf.slice(ssh_m2_red_up, offsets, size, "crop1")

    ssh_m3_det_concat = concatenate([ssh_m3_det_conv1_bn, ssh_m3_det_context_conv2_bn, ssh_m3_det_context_conv3_2_bn], 3, name='ssh_m3_det_concat')

    ssh_m2_det_context_conv1_relu = ReLU(name='ssh_m2_det_context_conv1_relu')(ssh_m2_det_context_conv1_bn)

    plus1_v1 = Add()([ssh_m1_red_conv_relu , crop1])

    ssh_m3_det_concat_relu = ReLU(name='ssh_m3_det_concat_relu')(ssh_m3_det_concat)

    ssh_m2_det_context_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m2_det_context_conv1_relu)

    ssh_m2_det_context_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m2_det_context_conv2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_context_conv2_pad)

    ssh_m2_det_context_conv3_1_pad  = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m2_det_context_conv1_relu)

    ssh_m2_det_context_conv3_1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m2_det_context_conv3_1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_context_conv3_1_pad)

    ssh_c1_aggr_pad = ZeroPadding2D(padding=tuple([1, 1]))(plus1_v1)

    ssh_c1_aggr = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_c1_aggr', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_c1_aggr_pad)

    face_rpn_cls_score_stride32 = Conv2D(filters = 4, kernel_size = (1, 1), name = 'face_rpn_cls_score_stride32', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_concat_relu)

    inter_1 = concatenate([face_rpn_cls_score_stride32[:, :, :, 0], face_rpn_cls_score_stride32[:, :, :, 1]], axis=1)
    inter_2 = concatenate([face_rpn_cls_score_stride32[:, :, :, 2], face_rpn_cls_score_stride32[:, :, :, 3]], axis=1)
    final = tf.stack([inter_1, inter_2])
    face_rpn_cls_score_reshape_stride32 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_score_reshape_stride32")

    face_rpn_bbox_pred_stride32 = Conv2D(filters = 8, kernel_size = (1, 1), name = 'face_rpn_bbox_pred_stride32', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_concat_relu)

    face_rpn_landmark_pred_stride32 = Conv2D(filters = 20, kernel_size = (1, 1), name = 'face_rpn_landmark_pred_stride32', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_concat_relu)

    ssh_m2_det_context_conv2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_context_conv2_bn', trainable=False)(ssh_m2_det_context_conv2)

    ssh_m2_det_context_conv3_1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_context_conv3_1_bn', trainable=False)(ssh_m2_det_context_conv3_1)

    ssh_c1_aggr_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_c1_aggr_bn', trainable=False)(ssh_c1_aggr)

    ssh_m2_det_context_conv3_1_relu = ReLU(name='ssh_m2_det_context_conv3_1_relu')(ssh_m2_det_context_conv3_1_bn)

    ssh_c1_aggr_relu = ReLU(name='ssh_c1_aggr_relu')(ssh_c1_aggr_bn)

    face_rpn_cls_prob_stride32 = Softmax(name = 'face_rpn_cls_prob_stride32')(face_rpn_cls_score_reshape_stride32)

    input_shape = [tf.shape(face_rpn_cls_prob_stride32)[k] for k in range(4)]
    sz = tf.dtypes.cast(input_shape[1] / 2, dtype=tf.int32)
    inter_1 = face_rpn_cls_prob_stride32[:, 0:sz, :, 0]
    inter_2 = face_rpn_cls_prob_stride32[:, 0:sz, :, 1]
    inter_3 = face_rpn_cls_prob_stride32[:, sz:, :, 0]
    inter_4 = face_rpn_cls_prob_stride32[:, sz:, :, 1]
    final = tf.stack([inter_1, inter_3, inter_2, inter_4])
    face_rpn_cls_prob_reshape_stride32 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_prob_reshape_stride32")

    ssh_m2_det_context_conv3_2_pad  = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m2_det_context_conv3_1_relu)

    ssh_m2_det_context_conv3_2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m2_det_context_conv3_2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_context_conv3_2_pad)

    ssh_m1_det_conv1_pad            = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c1_aggr_relu)

    ssh_m1_det_conv1 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_m1_det_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_conv1_pad)

    ssh_m1_det_context_conv1_pad    = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c1_aggr_relu)

    ssh_m1_det_context_conv1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m1_det_context_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_context_conv1_pad)

    ssh_m2_det_context_conv3_2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_context_conv3_2_bn', trainable=False)(ssh_m2_det_context_conv3_2)

    ssh_m1_det_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_conv1_bn', trainable=False)(ssh_m1_det_conv1)

    ssh_m1_det_context_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_context_conv1_bn', trainable=False)(ssh_m1_det_context_conv1)

    ssh_m2_det_concat               = concatenate([ssh_m2_det_conv1_bn, ssh_m2_det_context_conv2_bn, ssh_m2_det_context_conv3_2_bn], 3, name='ssh_m2_det_concat')

    ssh_m1_det_context_conv1_relu = ReLU(name='ssh_m1_det_context_conv1_relu')(ssh_m1_det_context_conv1_bn)

    ssh_m2_det_concat_relu = ReLU(name='ssh_m2_det_concat_relu')(ssh_m2_det_concat)

    ssh_m1_det_context_conv2_pad    = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m1_det_context_conv1_relu)

    ssh_m1_det_context_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m1_det_context_conv2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_context_conv2_pad)

    ssh_m1_det_context_conv3_1_pad  = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m1_det_context_conv1_relu)

    ssh_m1_det_context_conv3_1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m1_det_context_conv3_1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_context_conv3_1_pad)

    face_rpn_cls_score_stride16 = Conv2D(filters = 4, kernel_size = (1, 1), name = 'face_rpn_cls_score_stride16', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_concat_relu)

    inter_1 = concatenate([face_rpn_cls_score_stride16[:, :, :, 0], face_rpn_cls_score_stride16[:, :, :, 1]], axis=1)
    inter_2 = concatenate([face_rpn_cls_score_stride16[:, :, :, 2], face_rpn_cls_score_stride16[:, :, :, 3]], axis=1)
    final = tf.stack([inter_1, inter_2])
    face_rpn_cls_score_reshape_stride16 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_score_reshape_stride16")

    face_rpn_bbox_pred_stride16 = Conv2D(filters = 8, kernel_size = (1, 1), name = 'face_rpn_bbox_pred_stride16', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_concat_relu)

    face_rpn_landmark_pred_stride16 = Conv2D(filters = 20, kernel_size = (1, 1), name = 'face_rpn_landmark_pred_stride16', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_concat_relu)

    ssh_m1_det_context_conv2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_context_conv2_bn', trainable=False)(ssh_m1_det_context_conv2)

    ssh_m1_det_context_conv3_1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_context_conv3_1_bn', trainable=False)(ssh_m1_det_context_conv3_1)

    ssh_m1_det_context_conv3_1_relu = ReLU(name='ssh_m1_det_context_conv3_1_relu')(ssh_m1_det_context_conv3_1_bn)

    face_rpn_cls_prob_stride16      = Softmax(name = 'face_rpn_cls_prob_stride16')(face_rpn_cls_score_reshape_stride16)

    input_shape = [tf.shape(face_rpn_cls_prob_stride16)[k] for k in range(4)]
    sz = tf.dtypes.cast(input_shape[1] / 2, dtype=tf.int32)
    inter_1 = face_rpn_cls_prob_stride16[:, 0:sz, :, 0]
    inter_2 = face_rpn_cls_prob_stride16[:, 0:sz, :, 1]
    inter_3 = face_rpn_cls_prob_stride16[:, sz:, :, 0]
    inter_4 = face_rpn_cls_prob_stride16[:, sz:, :, 1]
    final = tf.stack([inter_1, inter_3, inter_2, inter_4])
    face_rpn_cls_prob_reshape_stride16 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_prob_reshape_stride16")

    ssh_m1_det_context_conv3_2_pad  = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m1_det_context_conv3_1_relu)

    ssh_m1_det_context_conv3_2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m1_det_context_conv3_2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_context_conv3_2_pad)

    ssh_m1_det_context_conv3_2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_context_conv3_2_bn', trainable=False)(ssh_m1_det_context_conv3_2)

    ssh_m1_det_concat               = concatenate([ssh_m1_det_conv1_bn, ssh_m1_det_context_conv2_bn, ssh_m1_det_context_conv3_2_bn], 3, name='ssh_m1_det_concat')

    ssh_m1_det_concat_relu = ReLU(name='ssh_m1_det_concat_relu')(ssh_m1_det_concat)
    face_rpn_cls_score_stride8 = Conv2D(filters = 4, kernel_size = (1, 1), name = 'face_rpn_cls_score_stride8', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_concat_relu)

    inter_1 = concatenate([face_rpn_cls_score_stride8[:, :, :, 0], face_rpn_cls_score_stride8[:, :, :, 1]], axis=1)
    inter_2 = concatenate([face_rpn_cls_score_stride8[:, :, :, 2], face_rpn_cls_score_stride8[:, :, :, 3]], axis=1)
    final = tf.stack([inter_1, inter_2])
    face_rpn_cls_score_reshape_stride8 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_score_reshape_stride8")

    face_rpn_bbox_pred_stride8 = Conv2D(filters = 8, kernel_size = (1, 1), name = 'face_rpn_bbox_pred_stride8', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_concat_relu)

    face_rpn_landmark_pred_stride8 = Conv2D(filters = 20, kernel_size = (1, 1), name = 'face_rpn_landmark_pred_stride8', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_concat_relu)

    face_rpn_cls_prob_stride8       = Softmax(name = 'face_rpn_cls_prob_stride8')(face_rpn_cls_score_reshape_stride8)

    input_shape = [tf.shape(face_rpn_cls_prob_stride8)[k] for k in range(4)]
    sz = tf.dtypes.cast(input_shape[1] / 2, dtype=tf.int32)
    inter_1 = face_rpn_cls_prob_stride8[:, 0:sz, :, 0]
    inter_2 = face_rpn_cls_prob_stride8[:, 0:sz, :, 1]
    inter_3 = face_rpn_cls_prob_stride8[:, sz:, :, 0]
    inter_4 = face_rpn_cls_prob_stride8[:, sz:, :, 1]
    final = tf.stack([inter_1, inter_3, inter_2, inter_4])
    face_rpn_cls_prob_reshape_stride8 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_prob_reshape_stride8")

    model = Model(inputs=data,
                    outputs=[face_rpn_cls_prob_reshape_stride32,
                                                   face_rpn_bbox_pred_stride32,
                                                   face_rpn_landmark_pred_stride32,
                                                   face_rpn_cls_prob_reshape_stride16,
                                                   face_rpn_bbox_pred_stride16,
                                                   face_rpn_landmark_pred_stride16,
                                                   face_rpn_cls_prob_reshape_stride8,
                                                   face_rpn_bbox_pred_stride8,
                                                   face_rpn_landmark_pred_stride8
                                                   ])
    model = load_weights(model)

    return model