Scribd
Upload
9 views

AlexNet

Uploaded by

nericjoel2003
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
9 views

AlexNet

Uploaded by

nericjoel2003
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 20

20IS712

Deep Learning
300 (3)
Dr. Lekshmi R. R.
Asst. Prof.
Department of Electrical & Electronics
Engineering
Amrita School of Engineering 1
AlexNet
Neural Network
• VGG – by Alex and team
• Solves many overfitting related problem
• Looks similar to LeNet
• There are 60 million parameters
• Architecture includes:
– 5 Convolution layers
– 3 Fully connected layer
• includes output layer
• More filers per layer
– Hence called as Deep Neural Network
• Includes dropout
Architecture
Convolution

• Convolution layers ReLu 𝑃𝑖𝑥𝑒𝑙, 𝑛 = 227,

• Input: 227x227x3 𝐹𝑖𝑙𝑡𝑒𝑟 𝑠𝑖𝑧𝑒, 𝑓 = 11


Stride 𝑠𝑖𝑧𝑒, 𝑠 = 4
• Filter: 96 P𝑎𝑑𝑑𝑖𝑛𝑔, 𝑝 = 0
– Size: 11x11 𝑛 + 2𝑝 − 𝑓 227 + 2 × 0 − 11
+1 +1
• Stride: 4 𝑠 4

= 55
Convolution

• Convolution layers ReLu 𝑃𝑖𝑥𝑒𝑙, 𝑛 = 227,

• Input: 227x227x3 𝐹𝑖𝑙𝑡𝑒𝑟 𝑠𝑖𝑧𝑒, 𝑓 = 11


Stride 𝑠𝑖𝑧𝑒, 𝑠 = 4
• Filter: 96 P𝑎𝑑𝑑𝑖𝑛𝑔, 𝑝 = 0
– Size: 11x11 𝑛 + 2𝑝 − 𝑓 227 + 2 × 0 − 11
+1 +1
• Stride: 4 𝑠 4

= 55
Max pooling

• Input: 56x56x96
𝑛 + 2𝑝 − 𝑓
• Filter size: 3x3 +1
𝑠
• Stride: 2
55 + 2 × 0 − 3
+1
2

= 27
Max pooling

• Input: 56x56x96
𝑛 + 2𝑝 − 𝑓
• Filter size: 3x3 +1
𝑠
• Stride: 2
55 + 2 × 0 − 3
+1
2

= 27
Convolution layer

• Convolution layers ReLu 𝑃𝑖𝑥𝑒𝑙, 𝑛 = 27,

• Input: 27x27x96 𝐹𝑖𝑙𝑡𝑒𝑟 𝑠𝑖𝑧𝑒, 𝑓 =5


Stride 𝑠𝑖𝑧𝑒, 𝑠 = 1
• Padding: 2 P𝑎𝑑𝑑𝑖𝑛𝑔, 𝑝 = 2
• Filter: 256 𝑛 + 2𝑝 − 𝑓 27 + 2 × 2 − 5
+1 +1
– Size: 5x5 𝑠 1

• Stride: 1 = 27
Max pooling

• Input:27x27x256
𝑛 + 2𝑝 − 𝑓
• Size: 3x3 𝑠
+1

• Stride: 2 27 + 2 × 0 − 3
+1
2
• Output:13x13x256
=13
Convolution layer

• Convolution layers ReLu 𝑃𝑖𝑥𝑒𝑙, 𝑛 = 27,

• Input: 13x13x256 𝐹𝑖𝑙𝑡𝑒𝑟 𝑠𝑖𝑧𝑒, 𝑓 =5


Stride 𝑠𝑖𝑧𝑒, 𝑠 = 1
• Padding: 1 P𝑎𝑑𝑑𝑖𝑛𝑔, 𝑝 = 2
• Filter: 384 𝑛 + 2𝑝 − 𝑓 13 + 2 × 1 − 3
+1 +1
– Size: 3x3 𝑠 1

• Stride: 1 = 13
Convolution layer

• Convolution layers ReLu 𝑃𝑖𝑥𝑒𝑙, 𝑛 = 27,

• Input: 13x13x384 𝐹𝑖𝑙𝑡𝑒𝑟 𝑠𝑖𝑧𝑒, 𝑓 =5


Stride 𝑠𝑖𝑧𝑒, 𝑠 = 1
• Padding: 1 P𝑎𝑑𝑑𝑖𝑛𝑔, 𝑝 = 2
• Filter: 384 𝑛 + 2𝑝 − 𝑓 13 + 2 × 1 − 3
+1 +1
– Size: 3x3 𝑠 1

• Stride: 1 = 13
Convolution layer

• Convolution layers ReLu 𝑃𝑖𝑥𝑒𝑙, 𝑛 = 27,

• Input: 13x13x384 𝐹𝑖𝑙𝑡𝑒𝑟 𝑠𝑖𝑧𝑒, 𝑓 =5


Stride 𝑠𝑖𝑧𝑒, 𝑠 = 1
• Padding: 1 P𝑎𝑑𝑑𝑖𝑛𝑔, 𝑝 = 2
• Filter: 256 𝑛 + 2𝑝 − 𝑓 13 + 2 × 1 − 3
+1 +1
– Size: 3x3 𝑠 1

• Stride: 1 = 13
Max pooling

• Input:13x13x256
𝑛 + 2𝑝 − 𝑓
• Size: 3x3 𝑠
+1

• Stride: 2 13 + 2 × 0 − 3
+1
2
• Output:6x6x256
=6
Fully connected

• Input:6x6x256 (flatten)
• Neurons: 4096
Fully connected

• Input:4096
• Neurons: 4096
Fully connected (Output)

• Input:4096
• Activation: Softmax
• Neurons: 1000
Activity
Image Operation
Output
227x227x Conv 11x11
3 4 stride 55x55x96
96 kernels

Input Operation Output


55x55x96 Max pool 3x3 27x27x96
2 stride

Input Operation Output


27x27x96 Conv 5x5 27x27x256
2 padding
256 kernels

Input Operation Output


27x27x256 Max pool 3x3 13x13x256
2 stride
Activity
Input Operation
Output
13x13x256 Conv 3x3
1 padding 13x13x384
384 kernels
Input Operation
13x13x384 Conv 3x3 Output
1 padding 13x13x384
384 kernels

Input Operation
Output
13x13x384 Conv 3x3
1 padding 13x13x256
256 kernels
Input Operation
Output
13x13x256 Max pool 3x3
2 stride 6x6x256
Thank you

You might also like