EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

[chullhwan-song]

https://arxiv.org/abs/1905.11946

[chullhwan-song]

• 좋은리뷰가 있어 소개
• EfficientNet： Rethinking Model Scaling for Convolutional Neural Networks 리뷰

[chullhwan-song]

EfficientNet-B0 구조

1 - Stem    - Conv3x3|BN|Swish

2 - Blocks  - MBConv1, k3x3 
            - MBConv6, k3x3 repeated 2 times
            - MBConv6, k5x5 repeated 2 times
            - MBConv6, k3x3 repeated 3 times
            - MBConv6, k5x5 repeated 3 times
            - MBConv6, k5x5 repeated 4 times
            - MBConv6, k3x3
                            totally 16 blocks

3 - Head    - Conv1x1|BN|Swish 
            - Pooling
            - Dropout
            - FC
where

Swish(x) = x * sigmoid(x)
and MBConvX stands for mobile inverted bottleneck convolution, X - denotes expansion ratio:

MBConv1 : 
  -> DepthwiseConv|BN|Swish -> SqueezeExcitation -> Conv|BN

MBConv6 : 
  -> Conv|BN|Swish -> DepthwiseConv|BN|Swish -> SqueezeExcitation -> Conv|BN

MBConv6+IdentitySkip : 
  -.-> Conv|BN|Swish -> DepthwiseConv|BN|Swish -> SqueezeExcitation -> Conv|BN-(+)->
   \___________________________________________________________________________/