ML Practice 8_1

CNN(Convolution Neural Network)

• Neural network operations can also be applied to two-dimensional arrays by CNN.
• Neuron in CNN is called filter or kernel.

from tensorflow import keras
keras.layers.Conv2D(10, kernel_size=(3,3), activation="relu")

<keras.layers.convolutional.Conv2D at 0x7effd27dea10>

Padding & Stride

Padding : Filling the border of the input array with virtual elements

• To prevent the loss of the original features of the image even if you resize the array,
• Same padding : Padding to zero around the input to make the input and feature map the same size
• Valid padding : Convolution only in a pure input array without padding

Stride : Size of the filter moving over the input layer (default 1)

keras.layers.Conv2D(10, kernel_size=(3,3), activation='relu', padding='same', strides=1)

<keras.layers.convolutional.Conv2D at 0x7effceb4fb10>

Pooling

• Reducing the size of the feature map while maintaining the original features of the image
• Max pooling, Average pooling, etc

keras.layers.MaxPooling2D(2, strides=2, padding="valid")

<keras.layers.pooling.MaxPooling2D at 0x7effce850fd0>

keras.layers.AveragePooling2D(2, strides=2, padding="valid")

<keras.layers.pooling.AveragePooling2D at 0x7effcea305d0>

Overall process in CNN

1. Input Image Data
2. CNN Layer

• kernel_size, padding, stride
• activation function
• Calculate each feature map

3. Pooling Layer

• Maxpooling / Averagepooling
• final feature map

4. Repeat the above process
5. Fully Connected Layer
6. Calculate classification predictions (Softmax)

Ref.) 혼자 공부하는 머신러닝+딥러닝 (박해선, 한빛미디어)