为了探究影响模型运行时间的变量，之前运用了参数量做标准
参数量在TF框架下还是很容易计算的
TF框架下运用
tf.keras.models.Model().summary()
就能一键生成包含模型的layers，output，parameters的报告
为了探究其他标准用于反映模型的运行时间，我们在网上找到了三个标准：
参数量（parameters）
浮点运算次数（FLOPs）
内存访问次数（MAC）
这次我们就来探究FLOPs对模型latency的影响
一上来先踩坑
TF2.X取消了Profiler接口对于FLOPs的统计
即使通过网上给的TF1.X接口再用TF2.X compat.v1接口依然不能成功返回FLOPs的值
于是只能写个程序硬算
FLOPs本质上是模型中的乘法和加法运算，
模型里面的layers有：
Input layer
Zero Padding layer
Conv2D
BatchNormalization
Activation
Depthwise conv2D
Dense
其中因为浮点运算次数很少而可以忽略不记的layers：
Input，
zero padding,
BatchNormalization,
Activation
剩下需要计算的layers就是：
Conv2D,
Depthwise Conv2D,
Dense
Conv2D：
F L O P s = C i n ∗ K ∗ K ∗ H ∗ W ∗ C o u t FLOPs=Cin*K*K*H*W*Cout FLOPs=Cin∗K∗K∗H∗W∗Cout
Cin是输入的channel
K*K是kernel size
H*W是输出size
Cout是输出channel
Depthwise Conv2D：
F L O P s = C i n ∗ H ∗ W ∗ K ∗ K / S / S FLOPs=Cin*H*W*K*K/S/S FLOPs=Cin∗H∗W∗K∗K/S/S
Cin是输入的channel
K*K是kernel size
H*W是输入size
S*S是strides
Dense：
F L O P s = 2 ∗ C i n ∗ C o u t FLOPs=2*Cin*Cout FLOPs=2∗Cin∗Cout
Cin是输入的channel
Cout是输出channel
基本层定义好了之后就按照模型的结构将运算FLOPs的结构搭起来
每一层输出的数据跟模型的也是一样的
#for conv2D layers in Depthwise Res Block
#padding=same stride=(1,1)
#request input:
# channel out(filter),
# kernel size(k[2]),
# input(cin[3])
#
#output:
# Flops of this layer(conv2dDRB_FLPOs)
# output(cout[3])
def conv2DDRB(filter,kernel,cin):
H=cin[0]
W=cin[1]
conv2dDRB_FLOPs=cin[2]*kernel[0]*kernel[1]*filter*H*W
cout=[H,W,filter]
return conv2dDRB_FLOPs,cout
#for DepthwiseConv2D layers in Depthwise Res Block
#request input:
# kernel size(kernal[2]),
# stride(stride[2]),
# input(cin[3])
#output:
# FLOPs of this layer(DepConv2dDRB_FLOPs)
# output(out[3])
def DepthwiseConv2DDRB(kernel,stride,cin):
out=[0,0,0]
if stride[0]==1 :
DepConv2dDRB_FLOPs=cin[0]*cin[1]*cin[2]*kernel[0]*kernel[1]
out=cin
else:
DepConv2dDRB_FLOPs=cin[0]*cin[1]*cin[2]*kernel[0]*kernel[1]/stride[0]/stride[1]
out[0]=cin[0]/stride[0]
out[1]=cin[1]/stride[1]
out[2]=cin[2]
return DepConv2dDRB_FLOPs, out
def DRB(cin, filter,kernel,stride,t):
exp_channel=cin[2]*t
alpha=filter
block_counter=0
conv_flop,cout=conv2DDRB(exp_channel,(1,1),cin)
block_counter+=conv_flop
dep_flop,cout=DepthwiseConv2DDRB(kernel,stride,cout)
block_counter+=dep_flop
conv_flop,cout=conv2DDRB(alpha,(1,1),cout)
block_counter+=conv_flop
print('This Block FLOPs:',block_counter,'Output:',cout)
return block_counter,cout
def conv(filter,kernel,stride,cin):
H=cin[0]/stride[0]
W=cin[1]/stride[1]
conv_FLOPs=cin[2]*kernel[0]*kernel[1]*filter*H*W
cout=[H,W,filter]
return conv_FLOPs, cout
def GolbalAvgPool(cin):
features=cin[2]
print('Output:',features)
return features
def den(classes,fin):
flop=2*fin*classes
print('This layer FLOPs:',flop,'Output:',classes)
return flop, classes
def MNV2(classes):
total_count=0
img_input=(224,224,3)
conv_flop,cout=conv(32,(3,3),(2,2),img_input)
total_count+=conv_flop
print('This layer FLOPs:',conv_flop,'Output:',cout)
Dep_flop,cout=DepthwiseConv2DDRB((3,3),(1,1),cout)
total_count+=Dep_flop
print('This layer FLOPs:', Dep_flop, 'Output:', cout)
conv_flop,cout=conv(16,(1,1),(1,1),cout)
total_count += conv_flop
print('This layer FLOPs:', conv_flop, 'Output:', cout)
DRB_flops,cout=DRB(cout,24,(3,3),(2,2),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,24,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,32,(3,3),(2,2),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,32,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,32,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,64,(3,3),(2,2),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,64,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,64,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,64,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,96,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,96,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,96,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,160,(3,3),(2,2),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,160,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,160,(3,3),(1,1),6)
total_count+=DRB_flops
DRB_flops,cout=DRB(cout,320,(3,3),(1,1),6)
total_count+=DRB_flops
conv_flop,cout=conv(1280,(1,1),(1,1),cout)
total_count+=conv_flop
cout=GolbalAvgPool(cout)
Den_flops,classes=den(classes,cout)
total_count+=Den_flops
print('TOTAL FLOPs:',total_count,'Output:',classes)
return total_count,classes
def main():
MNV2(5)
if __name__=='__main__':
main()