Skip to main content

Table 1 GoogleNet architecture modified with Batch Normalization

From: Going deeper in the automated identification of Herbarium specimens

Type

Patch size / Stride

Output size

Depth

Params

Ops

Convolution

7 ×7/2

112 ×112×64

1

2.7K

34M

Max pool

3 ×3/2

56 ×56×64

0

  

Batch norm

 

56 ×56×64

0

  

LRN

 

56 ×56×64

0

  

Convolution

3 ×3/1

56 ×56×192

2

112K

360M

Max pool

3 ×3/2

28 ×28×192

0

  

Batch norm

 

28 ×28×192

0

  

LRN

 

28 ×28×192

0

  

Inception (3a)

 

28 ×28×256

2

159K

128M

Inception (3b)

 

28 ×28×480

2

380K

304M

Max pool

3 ×3/2

14 ×14×480

0

  

Batch norm

 

14 ×14×480

0

  

Inception (4a)

 

14 ×14×512

2

364K

73M

Inception (4b)

 

14 ×14×512

2

437K

88M

inception (4c)

 

14 ×14×512

2

463K

100M

Inception (4d)

 

14 ×14×528

2

580K

119M

Inception (4e)

 

14 ×14×832

2

840K

170M

Max pool

3 ×3/2

7 ×7×832

0

  

Batch norm

 

7 ×7×832

0

  

Inception (5a)

 

7 ×7×832

2

1072K

54M

Inception (5b)

 

7 ×7×1024

2

1388K

71M

Avg pool

7 ×7/1

1 ×1×1024

0

  

Batch norm

 

1 ×1×1024

0

  

Linear

 

1 ×1×10000

1

1000K

1M

Softmax

 

1 ×1×10000

0

 Â