[Paper Review] RepVGG: Making VGG-style ConvNets Great Again

Xiaohan Ding, Xiangyu Zhang+4 more|arXiv (Cornell University)|Jan 11, 2021

Advanced Neural Network ApplicationsComputer Science43 references43 citations

TL;DR

RepVGG trains with a multi-branch, training-time architecture and then converts to a plain 3x3 conv body for inference via structural re-parameterization, achieving strong accuracy with fast inference on GPUs.

ABSTRACT

We present a simple but powerful architecture of convolutional neural network, which has a VGG-like inference-time body composed of nothing but a stack of 3x3 convolution and ReLU, while the training-time model has a multi-branch topology. Such decoupling of the training-time and inference-time architecture is realized by a structural re-parameterization technique so that the model is named RepVGG. On ImageNet, RepVGG reaches over 80% top-1 accuracy, which is the first time for a plain model, to the best of our knowledge. On NVIDIA 1080Ti GPU, RepVGG models run 83% faster than ResNet-50 or 101% faster than ResNet-101 with higher accuracy and show favorable accuracy-speed trade-off compared to the state-of-the-art models like EfficientNet and RegNet. The code and trained models are available at https://github.com/megvii-model/RepVGG.

Motivation & Objective

• Motivate the design of a simple, fast, memory-efficient CNN that rivals multi-branch architectures.
• Decouple training-time complexity from inference-time simplicity via structural re-parameterization.
• Demonstrate performance and speed benefits on ImageNet and semantic segmentation tasks.

Proposed method

• Propose RepVGG with a training-time multi-branch block (including identity and 1x1 branches) and a plain inference-time body of stacked 3x3 convolutions plus ReLU.
• Use structural re-parameterization to convert the training-time multi-branch block into an equivalent single 3x3 conv kernel for deployment (BN folding and kernel aggregation).
• Derive equations to merge BN and preceding convs, and to sum/pad 1x1 and identity contributions into a final 3x3 kernel (with matching strides and paddings).
• Instantiate architecture in a VGG-like plain topology with 5 stages, using 3x3 convs and controlled down-sampling, global average pooling, and a final classifier head.
• Explore variants with and without branches, ablate BN placement, and compare against ResNet, EfficientNet, and RegNet baselines.

Experimental results

Research questions

• RQ1 Can a training-time multi-branch design be re-parameterized into a fast, plain inference-time architecture without sacrificing accuracy?
• RQ2 How does RepVGG’s speed-accuracy trade-off compare to state-of-the-art models (ResNet, EfficientNet, RegNet) on ImageNet?
• RQ3 Is a VGG-style plain body with 3x3 convolutions competitive for ImageNet and transferable to semantic segmentation?
• RQ4 What is the impact of including or removing identity and 1x1 branches, BN placement, and ReLU in training-time structures on final performance?
• RQ5 Does structural re-parameterization enable practical gains in memory usage and deployment simplicity?

Key findings

• RepVGG achieves over 80% top-1 accuracy on ImageNet with plain inference-time architecture, a first for a plain model in their report.
• On NVIDIA 1080Ti, RepVGG models run faster than ResNet-50 and ResNet-101 with higher accuracy, and show favorable speed-accuracy compared to EfficientNet and RegNet.
• Architectural variants with interleaved groupwise convolutions (g2/g4) yield substantial speedups with modest accuracy loss (e.g., RepVGG-B1g4 vs ResNet-101).
• Structural re-parameterization (training-time multi-branch to inference-time plain) is validated as essential: removing branches or BN placement degrades accuracy, while full re-parameterization yields best results.
• In ablations, RepVGG’s training-time BN and branches provide advantages beyond parameter count, outperforming DiracNet and trivial re-parameterizations.
• RepVGG-B2/B3 variants demonstrate strong ImageNet performance (e.g., up to 80.52% top-1) with competitive FLOPs and lower parameter counts compared to several baselines.