--- toc: true title: RepVGG tags: ['temp'] --- # RepVGG - [RepVGG: Making Vgg-style ConvNets Great Again](https://arxiv.org/abs/2101.03697) - stack of $3\times3$ [Conv](Conv.md) and [Relu](Relu.md) during inference time - training-time model has a multi-branch topology - decoupling of the training-time and inference-time architecture is realized by a structural re-parameterization technique - 5 stages and conducts down-[Sampling](Sampling) via stride-2 convolution at the beginning of a stage - identity and 1 \times 1 branches, but only for training - [ImageNet](ImageNet.md) - higher accuracy and show favorable accuracy-speed trade-off compared to the state-of-the-art models like [EfficientNet](EfficientNet.md) and [RegNet](RegNet.md) -