RepVGG

RepVGG: Making VGG-style ConvNets Great Again

简介

RepVGG is a VGG-style convolutional architecture. It has the following advantages:

1. The model has a VGG-like plain (a.k.a. feed-forward) topology 1 without any branches. I.e., every layer takes the output of its only preceding layer as input and feeds the output into its only following layer.

2. The model’s body uses only 3 × 3 conv and ReLU.

3. The concrete architecture (including the specific depth and layer widths) is instantiated with no automatic search, manual refinement, compound scaling, nor other heavy designs.

摘要

Show the paper's 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.

How to use

The checkpoints provided are all training-time models. Use the reparameterize tool or switch_to_deploy interface to switch them to more efficient inference-time architecture, which not only has fewer parameters but also less calculations.

Predict imageUse the modelTrain/Test Command

Use classifier.backbone.switch_to_deploy() interface to switch the RepVGG models into inference mode.

>>> import torch
>>> from mmcls.apis import init_model, inference_model
>>>
>>> model = init_model('configs/repvgg/repvgg-A0_8xb32_in1k.py', 'https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth')
>>> results = inference_model(model, 'demo/demo.JPEG')
>>> print( (results['pred_class'], results['pred_score']) )
('sea snake' 0.8338906168937683)
>>>
>>> # switch to deploy mode
>>> model.backbone.switch_to_deploy()
>>> results = inference_model(model, 'demo/demo.JPEG')
>>> print( (results['pred_class'], results['pred_score']) )
('sea snake', 0.7883061170578003)

For more configurable parameters, please refer to the API.

How to use the reparameterisation tool(click to show)

Use provided tool to reparameterize the given model and save the checkpoint:

python tools/convert_models/reparameterize_model.py ${CFG_PATH} ${SRC_CKPT_PATH} ${TARGET_CKPT_PATH}

${CFG_PATH} is the config file path, ${SRC_CKPT_PATH} is the source chenpoint file path, ${TARGET_CKPT_PATH} is the target deploy weight file path.

For example:

# download the weight
wget https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth
# reparameterize unfused weight to fused weight
python ./tools/convert_models/reparameterize_model.py configs/repvgg/repvgg-a0_8xb32_in1k.py repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth repvgg-A0_deploy.pth

To use reparameterized weights, the config file must switch to the deploy config files as the deploy_A0 example or add --cfg-options model.backbone.deploy=True in command.

For example of using the reparameterized weights above:

python ./tools/test.py ./configs/repvgg/repvgg-A0_deploy_in1k.py  repvgg-A0_deploy.pth

You can get other deploy configs by modifying the A0_deploy example:

# in repvgg-A0_deploy_in1k.py
_base_ = '../repvgg-A0_8xb32_in1k.py'  # basic A0 config

model = dict(backbone=dict(deploy=True))  # switch model into deploy mode

or add --cfg-options model.backbone.deploy=True in command as following：

python tools/test.py configs/repvgg/repvgg-A0_8xb32_in1k.py repvgg_A0_deploy.pth --cfg-options model.backbone.deploy=True

结果和模型

ImageNet-1k

模型	预训练	Params(M) （train|deploy)	Flops(G) （train|deploy)	Top-1 (%)	Top-5 (%)	配置文件	下载
repvgg-A0_8xb32_in1k	从头训练	9.11 | 8.31	1.53 | 1.36	72.37	90.56	config	model | log
repvgg-A1_8xb32_in1k	从头训练	14.09 | 12.79	2.65 | 2.37	74.47	91.85	config	model | log
repvgg-A2_8xb32_in1k	从头训练	28.21 | 25.5	5.72 | 5.12	76.49	93.09	config	model | log
repvgg-B0_8xb32_in1k	从头训练	15.82 | 14.34	3.43 | 3.06	75.27	92.21	config	model | log
repvgg-B1_8xb32_in1k	从头训练	57.42 | 51.83	13.20 | 11.81	78.19	94.04	config	model | log
repvgg-B1g2_8xb32_in1k	从头训练	45.78 | 41.36	9.86 | 8.80	77.87	93.99	config	model | log
repvgg-B1g4_8xb32_in1k	从头训练	39.97 | 36.13	8.19 | 7.30	77.81	93.77	config	model | log
repvgg-B2_8xb32_in1k	从头训练	89.02 | 80.32	20.5 | 18.4	78.58	94.23	config	model | log
repvgg-B2g4_8xb32_in1k	从头训练	61.76 | 55.78	12.7 | 11.3	79.44	94.72	config	model | log
repvgg-B3_8xb32_in1k	从头训练	123.09 | 110.96	29.2 | 26.2	80.58	95.33	config	model | log
repvgg-B3g4_8xb32_in1k	从头训练	83.83 | 75.63	18.0 | 16.1	80.26	95.15	config	model | log
repvgg-D2se_3rdparty_in1k*	从头训练	133.33 | 120.39	36.6 | 32.8	81.81	95.94	config	model

Models with * are converted from the official repo. The config files of these models are only for inference. We don’t ensure these config files’ training accuracy and welcome you to contribute your reproduction results.

引用

@inproceedings{ding2021repvgg,
  title={Repvgg: Making vgg-style convnets great again},
  author={Ding, Xiaohan and Zhang, Xiangyu and Ma, Ningning and Han, Jungong and Ding, Guiguang and Sun, Jian},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={13733--13742},
  year={2021}
}