MixNet: Mixed Depthwise Convolut

google新论文，继续探索网络结构。基于混合使用多size的卷积核可以提升精度和效率这一观察，提出一种新的混合可分离卷积(MDConv）。即：在可分离卷积基础上考虑不同的卷积核。有点类似ASPP，但没有使用dilated convolution。再借助AutoML，搜索出一组效果惊艳但网络。

which significantly outperform previous models including MobileNetV2 [19] (ImageNet top-1 accuracy +4.2%), ShuffleNetV2 [15] (+3.5%), MnasNet [25] (+1.3%), ProxylessNAS [2] (+2.2%), and FBNet [26] (+2.0%). In particular, our MixNet-L achieves a new state-of-the-art 78.9% ImageNet top-1 accuracy under typical mobile settings (<600M FLOPS)

network.png perform.png

• paper
• github
• 解读