[Python] 딥러닝 ResNet 구현하기

ResNet 구현하기

ResNet에 처음 소개된 Residual Connection은 모델 내의 지름길을 새로 만든다고도 하여 Skip Connection이라고도 불리며,

레이어 개수가 매우 많은 경우에 발생할 수 있는 기울기 소실(Vanishing Gradient) 문제를 해결하고자 등장

 

 

import tensorflow as tf
from tensorflow.keras import layers, Model, Sequential

class ResidualBlock(Model):
    def __init__(self, num_kernels, kernel_size):
        super(ResidualBlock, self).__init__()

        # TODO: [지시사항 1번] 2개의 Conv2D Layer를 추가
        self.conv1 = layers.Conv2D(num_kernels, kernel_size=kernel_size, padding="same", activation="relu")
        self.conv2 = layers.Conv2D(num_kernels, kernel_size=kernel_size, padding="same", activation="relu")
        
        self.relu = layers.Activation("relu")
        
        # TODO: [지시사항 1번] Add Layer를 추가
        # Add(x, y)
        self.add = layers.Add()

    def call(self, input_tensor):
        x = self.conv1(input_tensor)
        x = self.conv2(x)

        x = self.add([x, input_tensor]) # 두 값을 더하는 과정
        x = self.relu(x)
        
        return x
        
def build_resnet(input_shape, num_classes):
    model = Sequential()
    
    model.add(layers.Conv2D(64, kernel_size=(3, 3), padding="same", activation="relu", input_shape=input_shape))
    model.add(layers.MaxPool2D(2))
    
    model.add(ResidualBlock(64, (3, 3)))
    model.add(ResidualBlock(64, (3, 3)))
    model.add(ResidualBlock(64, (3, 3)))
    
    model.add(layers.GlobalAveragePooling2D())
    model.add(layers.Dense(num_classes, activation="softmax"))
    
    return model
    
def main():
    input_shape = (32, 32, 3)
    num_classes = 10

    model = build_resnet(input_shape, num_classes)
    model.summary()

if __name__=="__main__":
    main()