RepLKNet (CVPR 2022)

This is the official MegEngine implementation of RepLKNet, from the following CVPR-2022 paper:

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

The paper is released on arXiv: https://arxiv.org/abs/2203.06717.

Official PyTorch implementation

The PyTorch code, training script and all of the pretrained PyTorch models, ImageNet-1K models, and Cityscapes/ADE20K/COCO models have been released at https://github.com/DingXiaoH/RepLKNet-pytorch.

Other implementations

framework	link
Tensorflow	https://github.com/shkarupa-alex/tfreplknet
...

More implementations are welcomed.

Catalog

Results and Pre-trained Models

ImageNet-1K Models

name	resolution	acc	#params	FLOPs	download
RepLKNet-31B	224x224	83.58	79M	15.3G	0de394

ImageNet-22K Models

name	resolution	acc	#params	FLOPs	22K model	1K model

MegData-73M Models

name	resolution	acc@1	#params	FLOPs	MegData-73M model	1K model

Installation of MegEngine

pip3 install megengine -f https://megengine.org.cn/whl/mge.html --user

For more details, please check the HomePage.

Installation of BaseCls

BaseCls is an image classification framework built upon MegEngine. We ultilize BaseCls for ImageNet pretraining and finetuning.

pip3 install basecls --user

Training and evaluation are configured through file. All default configurations are listed here.

Evaluation

./main_imagenet_test.py -f configs/replknet31_base_224_pt1k.py -w [weights] batch_size 64 data.val_path /path/to/imagenet/val

Training

./main_imagenet_train.py -f configs/replknet31_base_224_pt1k.py data.train_path /path/to/imagenet/train data.val_path /path/to/imagenet/val

Benchmark large depth-wise kernels

We can compare the kernel speed of MegEngine against PyTorch. A minimum version of megengine 1.8.2 is required for optimized large depth-wise convolutions.

./main_benchmark.py

License

This project is released under the MIT license. Please see the LICENSE file for more information.

sicker2020/RepLKNet