各位同学好，今天和大家分享一下如何使用 Tensorflow 构建 MobileNetV3 轻量化网络模型。
MobileNetV3 做了如下改动（1）更新了V2中的逆转残差结构；（2）使用NAS搜索参数的技术；（3）重新设计耗时层结构。MobileNetV3相比V2版本，在图像分类任务上，准确率上升了3.2%，延误降低了20%
MobileNetV3 是对MobileNetV2的改进，建议大家在学习之前，先了解 MobileNetV1、V2。
MobileNetV1：https://blog.csdn.net/dgvv4/article/details/123415708
MobileNetV2：https://blog.csdn.net/dgvv4/article/details/123417739
1. 网络核心模块介绍
1.1 MobileNetV1 深度可分离卷积
MobileNetV1 中主要使用了深度可分离卷积模块，大大减少了参数量和计算量。
普通卷积是一个卷积核处理所有的通道，输入特征图有多少个通道，卷积核就有几个通道，一个卷积核生成一张特征图。
深度可分离卷积 可理解为 深度卷积 + 逐点卷积
深度卷积只处理长宽方向的空间信息；逐点卷积只处理跨通道方向的信息。能大大减少参数量，提高计算效率
深度卷积： 一个卷积核只处理一个通道，即每个卷积核只处理自己对应的通道。输入特征图有多少个通道就有多少个卷积核。将每个卷积核处理后的特征图堆叠在一起。输入和输出特征图的通道数相同。
由于只处理长宽方向的信息会导致丢失跨通道信息，为了将跨通道的信息补充回来，需要进行逐点卷积。
扫描二维码关注公众号，回复：
13759682 查看本文章
逐点卷积： 是使用1x1卷积对跨通道维度处理，有多少个1x1卷积核就会生成多少个特征图。
1.2 MobileNetV2 逆转残差结构
MobileNetV2 使用了逆转残差模块。输入图像，先使用1x1卷积提升通道数；然后在高维空间下使用深度卷积；再使用1x1卷积下降通道数，降维时采用线性激活函数(y=x)。当步长等于1且输入和输出特征图的shape相同时，使用残差连接输入和输出；当步长=2（下采样阶段）直接输出降维后的特征图。
对比 ResNet 的残差结构。输入图像，先使用1x1卷积下降通道数；然后在低维空间下使用标准卷积，再使用1x1卷积上升通道数，激活函数都是ReLU函数。当步长等于1且输入和输出特征图的shape相同时，使用残差连接输入和输出；当步长=2（下采样阶段）直接输出降维后的特征图。
1.3 MobileNetV3 改进逆转残差结构
主要有以下改进:（1）添加SE注意力机制；（2）使用新的激活函数
1.3.1 SE注意力机制
（1）先将特征图进行全局平均池化，特征图有多少个通道，那么池化结果（一维向量）就有多少个元素，[h, w, c]==>[None, c]。
（2）然后经过两个全连接层得到输出向量。第一个全连接层的输出通道数等于原输入特征图的通道数的1/4；第二个全连接层的输出通道数等于原输入特征图的通道数。即先降维后升维。
（3）全连接层的输出向量可理解为，向量的每个元素是对每张特征图进行分析得出的权重关系。比较重要的特征图就会赋予更大的权重，即该特征图对应的向量元素的值较大。反之，不太重要的特征图对应的权重值较小。
（4）第一个全连接层使用ReLU激活函数，第二个全连接层使用 hard_sigmoid 激活函数
（5）经过两个全连接层得到一个由channel个元素组成的向量，每个元素是针对每个通道的权重，将权重和原特征图的对应相乘，得到新的特征图数据
以下图为例，特征图经过两个全连接层之后，比较重要的特征图对应的向量元素的值就较大。将得到的权重和对应特征图中的所有元素相乘，得到新的输出特征图
1.3.2 使用不同的激活函数
swish激活函数公式为：，尽管提高了网络精度，但是它的计算、求导复杂，对量化过程不友好，尤其对移动端设备的计算。
h_sigmoid激活函数公式为：，ReLU6激活函数公式为：
h_swish激活函数公式为：，替换之后网络的推理速度加快，对量化过程比较友好
1.3.3 总体流程
图像输入，先通过1x1卷积上升通道数；然后在高维空间下使用深度卷积；再经过SE注意力机制优化特征图数据；最后经过1x1卷积下降通道数（使用线性激活函数）。当步长等于1且输入和输出特征图的shape相同时，使用残差连接输入和输出；当步长=2（下采样阶段）直接输出降维后的特征图。
1.4 重新设计耗时层结构
（1）减少第一个卷积层的卷积核个数。将卷积核个数从32个降低到16个之后，准确率和降低之前是一样的。减少卷积核个数可以减少计算量，节省2ms时间
（2）简化最后的输出层。删除多余的卷积层，在准确率上没有变化，节省了7ms执行时间，这7ms占据了整个推理过程的11%的执行时间。明显提升计算速度。
2. 代码复现
2.1 网络结构图
网络模型结构如图所示。exp size 代表1*1卷积上升的通道数；#out 代表1*1卷积下降的通道数，即输出特征图数量；SE 代表是否使用注意力机制；NL 代表使用哪种激活函数；s 代表步长；bneck 代表逆残差结构；NBN 代表不使用批标准化。
2.2 搭建核心模块
（1）激活函数选择
根据上面的公式，定义hard_sigmoid激活函数和hard_swish激活函数。
#（1）激活函数：h-sigmoid
def h_sigmoid(input_tensor):
x = layers.Activation('hard_sigmoid')(input_tensor)
return x
#（2）激活函数：h-swish
def h_swish(input_tensor):
x = input_tensor * h_sigmoid(input_tensor)
return x（2）SE注意力机制
SE注意力机制由 全局平均池化 + 全连接层降维 + 全连接层升维 + 对应权重相乘 组成。为了减少参数量和计算量，全连接层由1*1普通卷积层代替
#（3）SE注意力机制
def se_block(input_tensor):
squeeze = input_tensor.shape[-1]/4 # 第一个全连接通道数下降1/4
excitation = input_tensor.shape[-1] # 第二个全连接通道数上升至原来
# 全局平均池化[b,h,w,c]==>[b,c]
x = layers.GlobalAveragePooling2D()(input_tensor)
# 添加宽度和高度的维度信息，因为下面要使用卷积层代替全连接层
x = layers.Reshape(target_shape=(1, 1, x.shape[-1]))(x) #[b,c]==>[b,1,1,c]
# 第一个全连接层下降通道数，用 1*1卷积层 代替，减少参数量
x = layers.Conv2D(filters=squeeze, # 通道数下降原来的1/4
kernel_size=(1,1), # 1*1卷积融合通道信息
strides=1, # 步长=1
padding='same')(x) # 卷积过程中特征图size不变
x = layers.ReLU()(x) # relu激活
# 第二个全连接层上升通道数，也用1*1卷积层代替
x = layers.Conv2D(filters=excitation, # 通道数上升至原始的特征图数量
kernel_size=(1,1),
strides=1,
padding='same')(x)
x = h_sigmoid(x) # hard_sigmoid激活函数
# 将输入特征图的每个通道和SE得到的针对每个通道的权重相乘
output = layers.Multiply()([input_tensor, x])
return output（3）标准卷积块
一个标准卷积块是由 普通卷积 + 批标准化 + 激活函数 组成
#（4）标准卷积块
def conv_block(input_tensor, filters, kernel_size, stride, activation):
# 判断使用什么类型的激活函数
if activation == 'RE':
act = layers.ReLU() # relu激活
elif activation == 'HS':
act = h_swish # hardswish激活
# 普通卷积
x = layers.Conv2D(filters, kernel_size, strides=stride, padding='same', use_bias=False)(input_tensor)
# BN层
x = layers.BatchNormalization()(x)
# 激活
x = act(x)
return x（4）逆残差模块
相比于MobileNetV2的逆残差模块，添加了注意力机制，使用不同的激活函数
#（5）逆转残差模块bneck
def bneck(x, expansion, filters, kernel_size, stride, se, activation):
"""
filters代表bottleneck模块输出特征图的通道数个数
se是bool类型, se=True 就使用注意力机制, 反之则不使用
activation表示使用什么类型的激活函数'RE'和'HS'
"""
# 残差边
residual = x
# 判断使用什么类型的激活函数
if activation == 'RE':
act = layers.ReLU() # relu激活
elif activation == 'HS':
act = h_swish # hardswish激活
# ① 1*1卷积上升通道数
if expansion != filters: # 第一个bneck模块不需要上升通道数，
x = layers.Conv2D(filters = expansion, # 上升的通道数
kernel_size = (1,1), # 1*1卷积融合通道信息
strides = 1, # 只处理通道方向的信息
padding = 'same', # 卷积过程中size不变，
use_bias = False)(x) # 有BN层就不使用偏置
x = layers.BatchNormalization()(x) # 批标准化
x = act(x) # 激活函数
# ② 深度卷积
x = layers.DepthwiseConv2D(kernel_size = kernel_size, # 卷积核size
strides = stride, # 是否进行下采样
padding = 'same', # 步长=2，特征图长宽减半
use_bias = False)(x) # 有BN层就不要偏置
x = layers.BatchNormalization()(x) # 批标准化
x = act(x) # 激活函数
# ③ 是否使用注意力机制
if se == True:
x = se_block(x)
# ④ 1*1卷积下降通道数
x = layers.Conv2D(filters = filters, # 输出特征图个数
kernel_size = (1,1), # 1*1卷积融合通道信息
strides = 1,
padding = 'same',
use_bias = False)(x)
x = layers.BatchNormalization()(x)
# 使用的是线性激活函数y=x
# ④ 如果深度卷积的步长=1并且输入和输出的shape相同，就叠加残差边
if stride == 1 and residual.shape==x.shape:
x = layers.Add()([residual, x])
return x # 如果步长=2，直接返回1*1卷积下降通道数后的结果2.3 完整代码展示
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import Model, layers
#（1）激活函数：h-sigmoid
def h_sigmoid(input_tensor):
x = layers.Activation('hard_sigmoid')(input_tensor)
return x
#（2）激活函数：h-swish
def h_swish(input_tensor):
x = input_tensor * h_sigmoid(input_tensor)
return x
#（3）SE注意力机制
def se_block(input_tensor):
squeeze = input_tensor.shape[-1]/4 # 第一个全连接通道数下降1/4
excitation = input_tensor.shape[-1] # 第二个全连接通道数上升至原来
# 全局平均池化[b,h,w,c]==>[b,c]
x = layers.GlobalAveragePooling2D()(input_tensor)
# 添加宽度和高度的维度信息，因为下面要使用卷积层代替全连接层
x = layers.Reshape(target_shape=(1, 1, x.shape[-1]))(x) #[b,c]==>[b,1,1,c]
# 第一个全连接层下降通道数，用 1*1卷积层 代替，减少参数量
x = layers.Conv2D(filters=squeeze, # 通道数下降原来的1/4
kernel_size=(1,1), # 1*1卷积融合通道信息
strides=1, # 步长=1
padding='same')(x) # 卷积过程中特征图size不变
x = layers.ReLU()(x) # relu激活
# 第二个全连接层上升通道数，也用1*1卷积层代替
x = layers.Conv2D(filters=excitation, # 通道数上升至原始的特征图数量
kernel_size=(1,1),
strides=1,
padding='same')(x)
x = h_sigmoid(x) # hard_sigmoid激活函数
# 将输入特征图的每个通道和SE得到的针对每个通道的权重相乘
output = layers.Multiply()([input_tensor, x])
return output
#（4）标准卷积块
def conv_block(input_tensor, filters, kernel_size, stride, activation):
# 判断使用什么类型的激活函数
if activation == 'RE':
act = layers.ReLU() # relu激活
elif activation == 'HS':
act = h_swish # hardswish激活
# 普通卷积
x = layers.Conv2D(filters, kernel_size, strides=stride, padding='same', use_bias=False)(input_tensor)
# BN层
x = layers.BatchNormalization()(x)
# 激活
x = act(x)
return x
#（5）逆转残差模块bneck
def bneck(x, expansion, filters, kernel_size, stride, se, activation):
"""
filters代表bottleneck模块输出特征图的通道数个数
se是bool类型, se=True 就使用注意力机制, 反之则不使用
activation表示使用什么类型的激活函数'RE'和'HS'
"""
# 残差边
residual = x
# 判断使用什么类型的激活函数
if activation == 'RE':
act = layers.ReLU() # relu激活
elif activation == 'HS':
act = h_swish # hardswish激活
# ① 1*1卷积上升通道数
if expansion != filters: # 第一个bneck模块不需要上升通道数，
x = layers.Conv2D(filters = expansion, # 上升的通道数
kernel_size = (1,1), # 1*1卷积融合通道信息
strides = 1, # 只处理通道方向的信息
padding = 'same', # 卷积过程中size不变，
use_bias = False)(x) # 有BN层就不使用偏置
x = layers.BatchNormalization()(x) # 批标准化
x = act(x) # 激活函数
# ② 深度卷积
x = layers.DepthwiseConv2D(kernel_size = kernel_size, # 卷积核size
strides = stride, # 是否进行下采样
padding = 'same', # 步长=2，特征图长宽减半
use_bias = False)(x) # 有BN层就不要偏置
x = layers.BatchNormalization()(x) # 批标准化
x = act(x) # 激活函数
# ③ 是否使用注意力机制
if se == True:
x = se_block(x)
# ④ 1*1卷积下降通道数
x = layers.Conv2D(filters = filters, # 输出特征图个数
kernel_size = (1,1), # 1*1卷积融合通道信息
strides = 1,
padding = 'same',
use_bias = False)(x)
x = layers.BatchNormalization()(x)
# 使用的是线性激活函数y=x
# ④ 如果深度卷积的步长=1并且输入和输出的shape相同，就叠加残差边
if stride == 1 and residual.shape==x.shape:
x = layers.Add()([residual, x])
return x # 如果步长=2，直接返回1*1卷积下降通道数后的结果
#（6）主干网络
def mobilenet(input_shape, classes): # 输入图像shape，分类数
# 构造输入
inputs = keras.Input(shape=input_shape)
# [224,224,3] ==> [112,112,16]
x = conv_block(inputs, filters=16, kernel_size=(3,3), stride=2, activation='HS')
# [112,112,16] ==> [112,112,16]
x = bneck(x, expansion=16, filters=16, kernel_size=(3,3), stride=1, se=False, activation='RE')
# [112,112,16] ==> [56,56,24]
x = bneck(x, expansion=64, filters=24, kernel_size=(3,3), stride=2, se=False, activation='RE')
# [56,56,24] ==> [56,56,24]
x = bneck(x, expansion=72, filters=24, kernel_size=(3,3), stride=1, se=False, activation='RE')
# [56,56,24] ==> [28,28,40]
x = bneck(x, expansion=72, filters=40, kernel_size=(5,5), stride=2, se=True, activation='RE')
# [28,28,40] ==> [28,28,40]
x = bneck(x, expansion=120, filters=40, kernel_size=(5,5), stride=1, se=True, activation='RE')
# [28,28,40] ==> [28,28,40]
x = bneck(x, expansion=120, filters=40, kernel_size=(5,5), stride=1, se=True, activation='RE')
# [28,28,40] ==> [14,14,80]
x = bneck(x, expansion=240, filters=80, kernel_size=(3,3), stride=2, se=False, activation='HS')
# [14,14,80] ==> [14,14,80]
x = bneck(x, expansion=200, filters=80, kernel_size=(3,3), stride=1, se=False, activation='HS')
# [14,14,80] ==> [14,14,80]
x = bneck(x, expansion=184, filters=80, kernel_size=(3,3), stride=1, se=False, activation='HS')
# [14,14,80] ==> [14,14,80]
x = bneck(x, expansion=184, filters=80, kernel_size=(3,3), stride=1, se=False, activation='HS')
# [14,14,80] ==> [14,14,112]
x = bneck(x, expansion=480, filters=112, kernel_size=(3,3), stride=1, se=True, activation='HS')
# [14,14,112] ==> [14,14,112]
x = bneck(x, expansion=672, filters=112, kernel_size=(3,3), stride=1, se=True, activation='HS')
# [14,14,112] ==> [7,7,160]
x = bneck(x, expansion=672, filters=160, kernel_size=(5,5), stride=2, se=True, activation='HS')
# [7,7,160] ==> [7,7,160]
x = bneck(x, expansion=960, filters=160, kernel_size=(5,5), stride=1, se=True, activation='HS')
# [7,7,160] ==> [7,7,160]
x = bneck(x, expansion=960, filters=160, kernel_size=(5,5), stride=1, se=True, activation='HS')
# [7,7,160] ==> [7,7,960]
x = conv_block(x, filters=960, kernel_size=(1,1), stride=1, activation='HS')
# [7,7,960] ==> [None,960]
x = layers.MaxPooling2D(pool_size=(7,7))(x)
# [None,960] ==> [1,1,960]
x = layers.Reshape(target_shape=(1,1,x.shape[-1]))(x)
# [1,1,960] ==> [1,1,1280]
x = layers.Conv2D(filters=1280, kernel_size=(1,1), strides=1, padding='same')(x)
x = h_swish(x)
# [1,1,960] ==> [1,1,classes]
x = layers.Conv2D(filters=classes, kernel_size=(1,1), strides=1, padding='same')(x)
# [1,1,classes] ==> [None,classes]
logits = layers.Flatten()(x)
# 构造模型
model = Model(inputs, logits)
return model
#（7）接收网络模型
if __name__ == '__main__':
model = mobilenet(input_shape=[224,224,3], classes=1000)
model.summary() # 查看网络架构2.4 查看网络模型结构
通过函数model.summary()查看网络总体框架，约五百万参数量
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 224, 224, 3) 0
__________________________________________________________________________________________________
conv2d (Conv2D) (None, 112, 112, 16) 432 input_1[0][0]
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 112, 112, 16) 64 conv2d[0][0]
__________________________________________________________________________________________________
activation (Activation) (None, 112, 112, 16) 0 batch_normalization[0][0]
__________________________________________________________________________________________________
tf.math.multiply (TFOpLambda) (None, 112, 112, 16) 0 batch_normalization[0][0]
activation[0][0]
__________________________________________________________________________________________________
depthwise_conv2d (DepthwiseConv (None, 112, 112, 16) 144 tf.math.multiply[0][0]
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 112, 112, 16) 64 depthwise_conv2d[0][0]
__________________________________________________________________________________________________
re_lu (ReLU) (None, 112, 112, 16) 0 batch_normalization_1[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 112, 112, 16) 256 re_lu[0][0]
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 112, 112, 16) 64 conv2d_1[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 112, 112, 64) 1024 batch_normalization_2[0][0]
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 112, 112, 64) 256 conv2d_2[0][0]
__________________________________________________________________________________________________
re_lu_1 (ReLU) multiple 0 batch_normalization_3[0][0]
batch_normalization_4[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_1 (DepthwiseCo (None, 56, 56, 64) 576 re_lu_1[0][0]
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 56, 56, 64) 256 depthwise_conv2d_1[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 56, 56, 24) 1536 re_lu_1[1][0]
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 56, 56, 24) 96 conv2d_3[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 56, 56, 72) 1728 batch_normalization_5[0][0]
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 56, 56, 72) 288 conv2d_4[0][0]
__________________________________________________________________________________________________
re_lu_2 (ReLU) (None, 56, 56, 72) 0 batch_normalization_6[0][0]
batch_normalization_7[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_2 (DepthwiseCo (None, 56, 56, 72) 648 re_lu_2[0][0]
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 56, 56, 72) 288 depthwise_conv2d_2[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 56, 56, 24) 1728 re_lu_2[1][0]
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 56, 56, 24) 96 conv2d_5[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 56, 56, 72) 1728 batch_normalization_8[0][0]
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 56, 56, 72) 288 conv2d_6[0][0]
__________________________________________________________________________________________________
re_lu_3 (ReLU) multiple 0 batch_normalization_9[0][0]
batch_normalization_10[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_3 (DepthwiseCo (None, 28, 28, 72) 1800 re_lu_3[0][0]
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 28, 28, 72) 288 depthwise_conv2d_3[0][0]
__________________________________________________________________________________________________
global_average_pooling2d (Globa (None, 72) 0 re_lu_3[1][0]
__________________________________________________________________________________________________
reshape (Reshape) (None, 1, 1, 72) 0 global_average_pooling2d[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 1, 1, 18) 1314 reshape[0][0]
__________________________________________________________________________________________________
re_lu_4 (ReLU) (None, 1, 1, 18) 0 conv2d_7[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 1, 1, 72) 1368 re_lu_4[0][0]
__________________________________________________________________________________________________
activation_1 (Activation) (None, 1, 1, 72) 0 conv2d_8[0][0]
__________________________________________________________________________________________________
multiply (Multiply) (None, 28, 28, 72) 0 re_lu_3[1][0]
activation_1[0][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 28, 28, 40) 2880 multiply[0][0]
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 28, 28, 40) 160 conv2d_9[0][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 28, 28, 120) 4800 batch_normalization_11[0][0]
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 28, 28, 120) 480 conv2d_10[0][0]
__________________________________________________________________________________________________
re_lu_5 (ReLU) (None, 28, 28, 120) 0 batch_normalization_12[0][0]
batch_normalization_13[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_4 (DepthwiseCo (None, 28, 28, 120) 3000 re_lu_5[0][0]
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 28, 28, 120) 480 depthwise_conv2d_4[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_1 (Glo (None, 120) 0 re_lu_5[1][0]
__________________________________________________________________________________________________
reshape_1 (Reshape) (None, 1, 1, 120) 0 global_average_pooling2d_1[0][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 1, 1, 30) 3630 reshape_1[0][0]
__________________________________________________________________________________________________
re_lu_6 (ReLU) (None, 1, 1, 30) 0 conv2d_11[0][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 1, 1, 120) 3720 re_lu_6[0][0]
__________________________________________________________________________________________________
activation_2 (Activation) (None, 1, 1, 120) 0 conv2d_12[0][0]
__________________________________________________________________________________________________
multiply_1 (Multiply) (None, 28, 28, 120) 0 re_lu_5[1][0]
activation_2[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 28, 28, 40) 4800 multiply_1[0][0]
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 28, 28, 40) 160 conv2d_13[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 28, 28, 120) 4800 batch_normalization_14[0][0]
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 28, 28, 120) 480 conv2d_14[0][0]
__________________________________________________________________________________________________
re_lu_7 (ReLU) (None, 28, 28, 120) 0 batch_normalization_15[0][0]
batch_normalization_16[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_5 (DepthwiseCo (None, 28, 28, 120) 3000 re_lu_7[0][0]
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 28, 28, 120) 480 depthwise_conv2d_5[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_2 (Glo (None, 120) 0 re_lu_7[1][0]
__________________________________________________________________________________________________
reshape_2 (Reshape) (None, 1, 1, 120) 0 global_average_pooling2d_2[0][0]
__________________________________________________________________________________________________
conv2d_15 (Conv2D) (None, 1, 1, 30) 3630 reshape_2[0][0]
__________________________________________________________________________________________________
re_lu_8 (ReLU) (None, 1, 1, 30) 0 conv2d_15[0][0]
__________________________________________________________________________________________________
conv2d_16 (Conv2D) (None, 1, 1, 120) 3720 re_lu_8[0][0]
__________________________________________________________________________________________________
activation_3 (Activation) (None, 1, 1, 120) 0 conv2d_16[0][0]
__________________________________________________________________________________________________
multiply_2 (Multiply) (None, 28, 28, 120) 0 re_lu_7[1][0]
activation_3[0][0]
__________________________________________________________________________________________________
conv2d_17 (Conv2D) (None, 28, 28, 40) 4800 multiply_2[0][0]
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 28, 28, 40) 160 conv2d_17[0][0]
__________________________________________________________________________________________________
conv2d_18 (Conv2D) (None, 28, 28, 240) 9600 batch_normalization_17[0][0]
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 28, 28, 240) 960 conv2d_18[0][0]
__________________________________________________________________________________________________
activation_4 (Activation) (None, 28, 28, 240) 0 batch_normalization_18[0][0]
__________________________________________________________________________________________________
tf.math.multiply_1 (TFOpLambda) (None, 28, 28, 240) 0 batch_normalization_18[0][0]
activation_4[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_6 (DepthwiseCo (None, 14, 14, 240) 2160 tf.math.multiply_1[0][0]
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 14, 14, 240) 960 depthwise_conv2d_6[0][0]
__________________________________________________________________________________________________
activation_5 (Activation) (None, 14, 14, 240) 0 batch_normalization_19[0][0]
__________________________________________________________________________________________________
tf.math.multiply_2 (TFOpLambda) (None, 14, 14, 240) 0 batch_normalization_19[0][0]
activation_5[0][0]
__________________________________________________________________________________________________
conv2d_19 (Conv2D) (None, 14, 14, 80) 19200 tf.math.multiply_2[0][0]
__________________________________________________________________________________________________
batch_normalization_20 (BatchNo (None, 14, 14, 80) 320 conv2d_19[0][0]
__________________________________________________________________________________________________
conv2d_20 (Conv2D) (None, 14, 14, 200) 16000 batch_normalization_20[0][0]
__________________________________________________________________________________________________
batch_normalization_21 (BatchNo (None, 14, 14, 200) 800 conv2d_20[0][0]
__________________________________________________________________________________________________
activation_6 (Activation) (None, 14, 14, 200) 0 batch_normalization_21[0][0]
__________________________________________________________________________________________________
tf.math.multiply_3 (TFOpLambda) (None, 14, 14, 200) 0 batch_normalization_21[0][0]
activation_6[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_7 (DepthwiseCo (None, 14, 14, 200) 1800 tf.math.multiply_3[0][0]
__________________________________________________________________________________________________
batch_normalization_22 (BatchNo (None, 14, 14, 200) 800 depthwise_conv2d_7[0][0]
__________________________________________________________________________________________________
activation_7 (Activation) (None, 14, 14, 200) 0 batch_normalization_22[0][0]
__________________________________________________________________________________________________
tf.math.multiply_4 (TFOpLambda) (None, 14, 14, 200) 0 batch_normalization_22[0][0]
activation_7[0][0]
__________________________________________________________________________________________________
conv2d_21 (Conv2D) (None, 14, 14, 80) 16000 tf.math.multiply_4[0][0]
__________________________________________________________________________________________________
batch_normalization_23 (BatchNo (None, 14, 14, 80) 320 conv2d_21[0][0]
__________________________________________________________________________________________________
conv2d_22 (Conv2D) (None, 14, 14, 184) 14720 batch_normalization_23[0][0]
__________________________________________________________________________________________________
batch_normalization_24 (BatchNo (None, 14, 14, 184) 736 conv2d_22[0][0]
__________________________________________________________________________________________________
activation_8 (Activation) (None, 14, 14, 184) 0 batch_normalization_24[0][0]
__________________________________________________________________________________________________
tf.math.multiply_5 (TFOpLambda) (None, 14, 14, 184) 0 batch_normalization_24[0][0]
activation_8[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_8 (DepthwiseCo (None, 14, 14, 184) 1656 tf.math.multiply_5[0][0]
__________________________________________________________________________________________________
batch_normalization_25 (BatchNo (None, 14, 14, 184) 736 depthwise_conv2d_8[0][0]
__________________________________________________________________________________________________
activation_9 (Activation) (None, 14, 14, 184) 0 batch_normalization_25[0][0]
__________________________________________________________________________________________________
tf.math.multiply_6 (TFOpLambda) (None, 14, 14, 184) 0 batch_normalization_25[0][0]
activation_9[0][0]
__________________________________________________________________________________________________
conv2d_23 (Conv2D) (None, 14, 14, 80) 14720 tf.math.multiply_6[0][0]
__________________________________________________________________________________________________
batch_normalization_26 (BatchNo (None, 14, 14, 80) 320 conv2d_23[0][0]
__________________________________________________________________________________________________
conv2d_24 (Conv2D) (None, 14, 14, 184) 14720 batch_normalization_26[0][0]
__________________________________________________________________________________________________
batch_normalization_27 (BatchNo (None, 14, 14, 184) 736 conv2d_24[0][0]
__________________________________________________________________________________________________
activation_10 (Activation) (None, 14, 14, 184) 0 batch_normalization_27[0][0]
__________________________________________________________________________________________________
tf.math.multiply_7 (TFOpLambda) (None, 14, 14, 184) 0 batch_normalization_27[0][0]
activation_10[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_9 (DepthwiseCo (None, 14, 14, 184) 1656 tf.math.multiply_7[0][0]
__________________________________________________________________________________________________
batch_normalization_28 (BatchNo (None, 14, 14, 184) 736 depthwise_conv2d_9[0][0]
__________________________________________________________________________________________________
activation_11 (Activation) (None, 14, 14, 184) 0 batch_normalization_28[0][0]
__________________________________________________________________________________________________
tf.math.multiply_8 (TFOpLambda) (None, 14, 14, 184) 0 batch_normalization_28[0][0]
activation_11[0][0]
__________________________________________________________________________________________________
conv2d_25 (Conv2D) (None, 14, 14, 80) 14720 tf.math.multiply_8[0][0]
__________________________________________________________________________________________________
batch_normalization_29 (BatchNo (None, 14, 14, 80) 320 conv2d_25[0][0]
__________________________________________________________________________________________________
conv2d_26 (Conv2D) (None, 14, 14, 480) 38400 batch_normalization_29[0][0]
__________________________________________________________________________________________________
batch_normalization_30 (BatchNo (None, 14, 14, 480) 1920 conv2d_26[0][0]
__________________________________________________________________________________________________
activation_12 (Activation) (None, 14, 14, 480) 0 batch_normalization_30[0][0]
__________________________________________________________________________________________________
tf.math.multiply_9 (TFOpLambda) (None, 14, 14, 480) 0 batch_normalization_30[0][0]
activation_12[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_10 (DepthwiseC (None, 14, 14, 480) 4320 tf.math.multiply_9[0][0]
__________________________________________________________________________________________________
batch_normalization_31 (BatchNo (None, 14, 14, 480) 1920 depthwise_conv2d_10[0][0]
__________________________________________________________________________________________________
activation_13 (Activation) (None, 14, 14, 480) 0 batch_normalization_31[0][0]
__________________________________________________________________________________________________
tf.math.multiply_10 (TFOpLambda (None, 14, 14, 480) 0 batch_normalization_31[0][0]
activation_13[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_3 (Glo (None, 480) 0 tf.math.multiply_10[0][0]
__________________________________________________________________________________________________
reshape_3 (Reshape) (None, 1, 1, 480) 0 global_average_pooling2d_3[0][0]
__________________________________________________________________________________________________
conv2d_27 (Conv2D) (None, 1, 1, 120) 57720 reshape_3[0][0]
__________________________________________________________________________________________________
re_lu_9 (ReLU) (None, 1, 1, 120) 0 conv2d_27[0][0]
__________________________________________________________________________________________________
conv2d_28 (Conv2D) (None, 1, 1, 480) 58080 re_lu_9[0][0]
__________________________________________________________________________________________________
activation_14 (Activation) (None, 1, 1, 480) 0 conv2d_28[0][0]
__________________________________________________________________________________________________
multiply_3 (Multiply) (None, 14, 14, 480) 0 tf.math.multiply_10[0][0]
activation_14[0][0]
__________________________________________________________________________________________________
conv2d_29 (Conv2D) (None, 14, 14, 112) 53760 multiply_3[0][0]
__________________________________________________________________________________________________
batch_normalization_32 (BatchNo (None, 14, 14, 112) 448 conv2d_29[0][0]
__________________________________________________________________________________________________
conv2d_30 (Conv2D) (None, 14, 14, 672) 75264 batch_normalization_32[0][0]
__________________________________________________________________________________________________
batch_normalization_33 (BatchNo (None, 14, 14, 672) 2688 conv2d_30[0][0]
__________________________________________________________________________________________________
activation_15 (Activation) (None, 14, 14, 672) 0 batch_normalization_33[0][0]
__________________________________________________________________________________________________
tf.math.multiply_11 (TFOpLambda (None, 14, 14, 672) 0 batch_normalization_33[0][0]
activation_15[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_11 (DepthwiseC (None, 14, 14, 672) 6048 tf.math.multiply_11[0][0]
__________________________________________________________________________________________________
batch_normalization_34 (BatchNo (None, 14, 14, 672) 2688 depthwise_conv2d_11[0][0]
__________________________________________________________________________________________________
activation_16 (Activation) (None, 14, 14, 672) 0 batch_normalization_34[0][0]
__________________________________________________________________________________________________
tf.math.multiply_12 (TFOpLambda (None, 14, 14, 672) 0 batch_normalization_34[0][0]
activation_16[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_4 (Glo (None, 672) 0 tf.math.multiply_12[0][0]
__________________________________________________________________________________________________
reshape_4 (Reshape) (None, 1, 1, 672) 0 global_average_pooling2d_4[0][0]
__________________________________________________________________________________________________
conv2d_31 (Conv2D) (None, 1, 1, 168) 113064 reshape_4[0][0]
__________________________________________________________________________________________________
re_lu_10 (ReLU) (None, 1, 1, 168) 0 conv2d_31[0][0]
__________________________________________________________________________________________________
conv2d_32 (Conv2D) (None, 1, 1, 672) 113568 re_lu_10[0][0]
__________________________________________________________________________________________________
activation_17 (Activation) (None, 1, 1, 672) 0 conv2d_32[0][0]
__________________________________________________________________________________________________
multiply_4 (Multiply) (None, 14, 14, 672) 0 tf.math.multiply_12[0][0]
activation_17[0][0]
__________________________________________________________________________________________________
conv2d_33 (Conv2D) (None, 14, 14, 112) 75264 multiply_4[0][0]
__________________________________________________________________________________________________
batch_normalization_35 (BatchNo (None, 14, 14, 112) 448 conv2d_33[0][0]
__________________________________________________________________________________________________
conv2d_34 (Conv2D) (None, 14, 14, 672) 75264 batch_normalization_35[0][0]
__________________________________________________________________________________________________
batch_normalization_36 (BatchNo (None, 14, 14, 672) 2688 conv2d_34[0][0]
__________________________________________________________________________________________________
activation_18 (Activation) (None, 14, 14, 672) 0 batch_normalization_36[0][0]
__________________________________________________________________________________________________
tf.math.multiply_13 (TFOpLambda (None, 14, 14, 672) 0 batch_normalization_36[0][0]
activation_18[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_12 (DepthwiseC (None, 7, 7, 672) 16800 tf.math.multiply_13[0][0]
__________________________________________________________________________________________________
batch_normalization_37 (BatchNo (None, 7, 7, 672) 2688 depthwise_conv2d_12[0][0]
__________________________________________________________________________________________________
activation_19 (Activation) (None, 7, 7, 672) 0 batch_normalization_37[0][0]
__________________________________________________________________________________________________
tf.math.multiply_14 (TFOpLambda (None, 7, 7, 672) 0 batch_normalization_37[0][0]
activation_19[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_5 (Glo (None, 672) 0 tf.math.multiply_14[0][0]
__________________________________________________________________________________________________
reshape_5 (Reshape) (None, 1, 1, 672) 0 global_average_pooling2d_5[0][0]
__________________________________________________________________________________________________
conv2d_35 (Conv2D) (None, 1, 1, 168) 113064 reshape_5[0][0]
__________________________________________________________________________________________________
re_lu_11 (ReLU) (None, 1, 1, 168) 0 conv2d_35[0][0]
__________________________________________________________________________________________________
conv2d_36 (Conv2D) (None, 1, 1, 672) 113568 re_lu_11[0][0]
__________________________________________________________________________________________________
activation_20 (Activation) (None, 1, 1, 672) 0 conv2d_36[0][0]
__________________________________________________________________________________________________
multiply_5 (Multiply) (None, 7, 7, 672) 0 tf.math.multiply_14[0][0]
activation_20[0][0]
__________________________________________________________________________________________________
conv2d_37 (Conv2D) (None, 7, 7, 160) 107520 multiply_5[0][0]
__________________________________________________________________________________________________
batch_normalization_38 (BatchNo (None, 7, 7, 160) 640 conv2d_37[0][0]
__________________________________________________________________________________________________
conv2d_38 (Conv2D) (None, 7, 7, 960) 153600 batch_normalization_38[0][0]
__________________________________________________________________________________________________
batch_normalization_39 (BatchNo (None, 7, 7, 960) 3840 conv2d_38[0][0]
__________________________________________________________________________________________________
activation_21 (Activation) (None, 7, 7, 960) 0 batch_normalization_39[0][0]
__________________________________________________________________________________________________
tf.math.multiply_15 (TFOpLambda (None, 7, 7, 960) 0 batch_normalization_39[0][0]
activation_21[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_13 (DepthwiseC (None, 7, 7, 960) 24000 tf.math.multiply_15[0][0]
__________________________________________________________________________________________________
batch_normalization_40 (BatchNo (None, 7, 7, 960) 3840 depthwise_conv2d_13[0][0]
__________________________________________________________________________________________________
activation_22 (Activation) (None, 7, 7, 960) 0 batch_normalization_40[0][0]
__________________________________________________________________________________________________
tf.math.multiply_16 (TFOpLambda (None, 7, 7, 960) 0 batch_normalization_40[0][0]
activation_22[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_6 (Glo (None, 960) 0 tf.math.multiply_16[0][0]
__________________________________________________________________________________________________
reshape_6 (Reshape) (None, 1, 1, 960) 0 global_average_pooling2d_6[0][0]
__________________________________________________________________________________________________
conv2d_39 (Conv2D) (None, 1, 1, 240) 230640 reshape_6[0][0]
__________________________________________________________________________________________________
re_lu_12 (ReLU) (None, 1, 1, 240) 0 conv2d_39[0][0]
__________________________________________________________________________________________________
conv2d_40 (Conv2D) (None, 1, 1, 960) 231360 re_lu_12[0][0]
__________________________________________________________________________________________________
activation_23 (Activation) (None, 1, 1, 960) 0 conv2d_40[0][0]
__________________________________________________________________________________________________
multiply_6 (Multiply) (None, 7, 7, 960) 0 tf.math.multiply_16[0][0]
activation_23[0][0]
__________________________________________________________________________________________________
conv2d_41 (Conv2D) (None, 7, 7, 160) 153600 multiply_6[0][0]
__________________________________________________________________________________________________
batch_normalization_41 (BatchNo (None, 7, 7, 160) 640 conv2d_41[0][0]
__________________________________________________________________________________________________
conv2d_42 (Conv2D) (None, 7, 7, 960) 153600 batch_normalization_41[0][0]
__________________________________________________________________________________________________
batch_normalization_42 (BatchNo (None, 7, 7, 960) 3840 conv2d_42[0][0]
__________________________________________________________________________________________________
activation_24 (Activation) (None, 7, 7, 960) 0 batch_normalization_42[0][0]
__________________________________________________________________________________________________
tf.math.multiply_17 (TFOpLambda (None, 7, 7, 960) 0 batch_normalization_42[0][0]
activation_24[0][0]
__________________________________________________________________________________________________
depthwise_conv2d_14 (DepthwiseC (None, 7, 7, 960) 24000 tf.math.multiply_17[0][0]
__________________________________________________________________________________________________
batch_normalization_43 (BatchNo (None, 7, 7, 960) 3840 depthwise_conv2d_14[0][0]
__________________________________________________________________________________________________
activation_25 (Activation) (None, 7, 7, 960) 0 batch_normalization_43[0][0]
__________________________________________________________________________________________________
tf.math.multiply_18 (TFOpLambda (None, 7, 7, 960) 0 batch_normalization_43[0][0]
activation_25[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_7 (Glo (None, 960) 0 tf.math.multiply_18[0][0]
__________________________________________________________________________________________________
reshape_7 (Reshape) (None, 1, 1, 960) 0 global_average_pooling2d_7[0][0]
__________________________________________________________________________________________________
conv2d_43 (Conv2D) (None, 1, 1, 240) 230640 reshape_7[0][0]
__________________________________________________________________________________________________
re_lu_13 (ReLU) (None, 1, 1, 240) 0 conv2d_43[0][0]
__________________________________________________________________________________________________
conv2d_44 (Conv2D) (None, 1, 1, 960) 231360 re_lu_13[0][0]
__________________________________________________________________________________________________
activation_26 (Activation) (None, 1, 1, 960) 0 conv2d_44[0][0]
__________________________________________________________________________________________________
multiply_7 (Multiply) (None, 7, 7, 960) 0 tf.math.multiply_18[0][0]
activation_26[0][0]
__________________________________________________________________________________________________
conv2d_45 (Conv2D) (None, 7, 7, 160) 153600 multiply_7[0][0]
__________________________________________________________________________________________________
batch_normalization_44 (BatchNo (None, 7, 7, 160) 640 conv2d_45[0][0]
__________________________________________________________________________________________________
conv2d_46 (Conv2D) (None, 7, 7, 960) 153600 batch_normalization_44[0][0]
__________________________________________________________________________________________________
batch_normalization_45 (BatchNo (None, 7, 7, 960) 3840 conv2d_46[0][0]
__________________________________________________________________________________________________
activation_27 (Activation) (None, 7, 7, 960) 0 batch_normalization_45[0][0]
__________________________________________________________________________________________________
tf.math.multiply_19 (TFOpLambda (None, 7, 7, 960) 0 batch_normalization_45[0][0]
activation_27[0][0]
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 1, 1, 960) 0 tf.math.multiply_19[0][0]
__________________________________________________________________________________________________
reshape_8 (Reshape) (None, 1, 1, 960) 0 max_pooling2d[0][0]
__________________________________________________________________________________________________
conv2d_47 (Conv2D) (None, 1, 1, 1280) 1230080 reshape_8[0][0]
__________________________________________________________________________________________________
activation_28 (Activation) (None, 1, 1, 1280) 0 conv2d_47[0][0]
__________________________________________________________________________________________________
tf.math.multiply_20 (TFOpLambda (None, 1, 1, 1280) 0 conv2d_47[0][0]
activation_28[0][0]
__________________________________________________________________________________________________
conv2d_48 (Conv2D) (None, 1, 1, 1000) 1281000 tf.math.multiply_20[0][0]
__________________________________________________________________________________________________
flatten (Flatten) (None, 1000) 0 conv2d_48[0][0]
==================================================================================================
Total params: 5,505,598
Trainable params: 5,481,198
Non-trainable params: 24,400
__________________________________________________________________________________________________