Skip to content

Latest commit

 

History

History
 
 

replknet

RepLKNet

Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs

Abstract

We revisit large kernel design in modern convolutional neural networks (CNNs). Inspired by recent advances in vision transformers (ViTs), in this paper, we demonstrate that using a few large convolutional kernels instead of a stack of small kernels could be a more powerful paradigm. We suggested five guidelines, e.g., applying re-parameterized large depth-wise convolutions, to design efficient highperformance large-kernel CNNs. Following the guidelines, we propose RepLKNet, a pure CNN architecture whose kernel size is as large as 31×31, in contrast to commonly used 3×3. RepLKNet greatly closes the performance gap between CNNs and ViTs, e.g., achieving comparable or superior results than Swin Transformer on ImageNet and a few typical downstream tasks, with lower latency. RepLKNet also shows nice scalability to big data and large models, obtaining 87.8% top-1 accuracy on ImageNet and 56.0% mIoU on ADE20K, which is very competitive among the state-of-the-arts with similar model sizes. Our study further reveals that, in contrast to small-kernel CNNs, large kernel CNNs have much larger effective receptive fields and higher shape bias rather than texture bias.

How to use it?

Predict image

from mmpretrain import inference_model, get_model

model = get_model('replknet-31B_3rdparty_in1k', pretrained=True)
model.backbone.switch_to_deploy()
predict = inference_model(model, 'demo/bird.JPEG')
print(predict['pred_class'])
print(predict['pred_score'])

Use the model

import torch
from mmpretrain import get_model

model = get_model('replknet-31B_3rdparty_in1k', pretrained=True)
inputs = torch.rand(1, 3, 224, 224)
out = model(inputs)
print(type(out))
# To extract features.
feats = model.extract_feat(inputs)
print(type(feats))

Test Command

Prepare your dataset according to the docs.

Test:

python tools/test.py configs/replknet/replknet-31B_32xb64_in1k.py https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_3rdparty_in1k_20221118-fd08e268.pth

Reparameterization

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

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

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

To use reparameterized weights, the config file must switch to the deploy config files.

python tools/test.py ${deploy_cfg} ${deploy_checkpoint} --metrics accuracy

You can also use backbone.switch_to_deploy() to switch to the deploy mode in Python code. For example:

from mmpretrain.models import RepLKNet

backbone = RepLKNet(arch='31B')
backbone.switch_to_deploy()

Models and results

Image Classification on ImageNet-1k

Model Pretrain Params (M) Flops (G) Top-1 (%) Top-5 (%) Config Download
replknet-31B_3rdparty_in1k* From scratch 79.86 15.64 83.48 96.57 config model
replknet-31B_3rdparty_in1k-384px* From scratch 79.86 45.95 84.84 97.34 config model
replknet-31B_in21k-pre_3rdparty_in1k* ImageNet-21k 79.86 15.64 85.20 97.56 config model
replknet-31B_in21k-pre_3rdparty_in1k-384px* ImageNet-21k 79.86 45.95 85.99 97.75 config model
replknet-31L_in21k-pre_3rdparty_in1k-384px* ImageNet-21k 172.67 97.24 86.63 98.00 config model
replknet-XL_meg73m-pre_3rdparty_in1k-320px* MEG73M 335.44 129.57 87.57 98.39 config model

Models with * are converted from the official repo. The config files of these models are only for inference. We haven't reprodcue the training results.

Citation

@inproceedings{ding2022scaling,
  title={Scaling up your kernels to 31x31: Revisiting large kernel design in cnns},
  author={Ding, Xiaohan and Zhang, Xiangyu and Han, Jungong and Ding, Guiguang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={11963--11975},
  year={2022}
}