YoloV1网络尺寸及FlOPS计算

以Conv1为例:

input:448x448x3

conv ksize:7x7x3(64),strides=2,padding=3

output dimension:(448 + 2 x 3 - 7) / 2 + 1 = 224.5(向下取整)=>>224x224x64

params:7 x 7 x 3 x 64 + 64(biases) =  9.472K

FLOPS = (7 x 7 x 3 + 7 x 7 x 3 - 1 + 1) x 224 x 224 x 64 = 944.111M

其余各层:

Name

Filters

Output Dimension

Params

Flops

Conv1

7x7x3,64,strides=2,padding=3

 224x224x64

9.472K

 944.111M

Max pool1

2x2,strides=2

112x112x64

0

 

Conv2

3x3x64,192,strides=1,padding=1

 112x112x192                      

110.784k

2774.5M

Max pool2

2x2,strides=2

56x56x192

 

 

Conv3

1x1x192,128,strides=1,padding=0

56x56x128

24.704K

154.140M

Conv4

3x3x128,256,strides=1,padding=1

56x56x256

295.168K

1849.688M

Conv5

1x1x256,256,strides=1,padding=0

56x56x256

65.792K

411.041M

Conv6

3x3x256,512,strides=1,padding=1

56x56x512

1.180M

7398.752M

Max pool3

2x2,strides=2

28x28x512

 

 

Conv7

1x1x512,256,strides=1,padding=0

28x28x256

131.328K

205.550M

Conv8

3x3x256,512,strides=1,padding=1

28x28x512

1.180M

1849.688M

Conv9

1x1x512,256,strides=1,padding=0

28x28x256

131.328K

205.550M

Conv10

3x3x256,512,strides=1,padding=1

28x28x512

1.180M

1849.688M

Conv11

1x1x512,256,strides=1,padding=0

28x28x256

131.328K

205.550M

Conv12

3x3x256,512,strides=1,padding=1

28x28x512

1.180M

1849.688M

Conv13

1x1x512,256,strides=1,padding=0

28x28x256

131.328K

205.520M

Conv14

3x3x256,512,strides=1,padding=1

28x28x512

1.180M

1849.688M

Conv15

1x1x512,512,strides=1,padding=0

28x28x512

262.656K

411.041M

Conv16

3x3x512,1024,strides=1,padding=1

28x28x1024

4.720M

7398.752M

Max pool4

2x2,strides=2

14x14x1024

 

 

Conv17

1x1x1024,512,strides=1,padding=0

14x14x512

524.800K

205.520M

Conv18

3x3x512,1024,strides=1,padding=1

14x14x1024

4.720M

1849.688M

Conv19

1x1x1024,512,strides=1,padding=0

14x14x512

524.800K

205.520M

Conv20

3x3x512,1024,strides=1,padding=1

14x14x1024

4.720M

1849.688M

Conv21

3x3x1024,1024,strides=1,padding=1

14x14x1024

9.438M

3699.376M

Conv22

3x3x1024,1024,strides=2,padding=1

7x7x1024

9.438M

924.844M

Conv23

3x3x1024,1024,strides=1,padding=1

7x7x1024

9.438M

924.844M

Conv24

3x3x1024,1024,strides=1,padding=1

7x7x1024

9.438M

924.844M

Full con1

 

4096

205.524M

411.041M

Full con2

 

7x7x30

6.022M

12.042M

Total

 

 

271.701M

40.570B

 

 

 

你可能感兴趣的:(深度学习)