RepVGG: Making VGG-style ConvNets Great Again

Installation

conda create -n PyTorch python=3.8
conda activate PyTorch
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch-lts
pip install opencv-python==4.5.5.64
pip install pyyaml
pip install timm
pip install tqdm

Note

• The default training configuration is for RepVGG-A0
• The test results including accuracy, params and FLOP are obtained by using fused model

Parameters and FLOPS

Number of parameters: 8.309M
Time per operator type:
        13.9458 ms.    93.5989%. Conv
       0.606509 ms.    4.07065%. Relu
       0.332151 ms.    2.22927%. FC
      0.0136558 ms.  0.0916524%. AveragePool
     0.00142615 ms. 0.00957176%. Flatten
        14.8996 ms in Total
FLOP per operator type:
        2.72034 GFLOP.    99.9059%. Conv
       0.002561 GFLOP.   0.094054%. FC
              0 GFLOP.          0%. Relu
         2.7229 GFLOP in Total
Feature Memory Read per operator type:
        35.6399 MB.    74.3363%. Conv
        7.17517 MB.    14.9657%. Relu
        5.12912 MB.    10.6981%. FC
        47.9442 MB in Total
Feature Memory Written per operator type:
        7.17517 MB.    49.9861%. Conv
        7.17517 MB.    49.9861%. Relu
          0.004 MB.  0.0278661%. FC
        14.3543 MB in Total
Parameter Memory per operator type:
        28.0956 MB.    84.5753%. Conv
          5.124 MB.    15.4247%. FC
              0 MB.          0%. Relu
        33.2196 MB in Total

Train

• Configure your IMAGENET dataset path in main.py for training
• Run bash main.sh $ --train for training, $ is number of GPUs

Test

• Configure your IMAGENET path in main.py for testing
• Run python main.py --test for testing

Results

Version	Epochs	Top-1 Acc	Top-5 Acc	Params (M)	FLOP (G)	Download
RepVGG-A0	120	-	-	8.309	1.362	-
RepVGG-A0*	120	72.4	90.5	8.309	1.362	model
RepVGG-A1*	120	74.5	91.8	12.790	2.364	model
RepVGG-A2*	120	76.5	93.0	25.500	5.117	model
RepVGG-B0*	120	75.1	92.4	14.339	3.058	model
RepVGG-B1*	120	78.3	94.1	51.829	11.816	model
RepVGG-B2*	120	78.8	94.4	80.315	18.377	model

• * means that weights are ported from original repo, see reference

Reference

• https://github.com/DingXiaoH/RepVGG