convnext

ConvNeXt

PaddlePaddle reimplementation of facebookresearch's repository for the ConvneXt model that was released with the paper A ConvNet for the 2020s.

Requirements

To enjoy some new features, PaddlePaddle 2.4 is required. For more installation tutorials refer to installation.md

How to Train

# Note: Set the following environment variables 
# and then need to run the script on each node.
#export PADDLE_NNODES=4
#export PADDLE_MASTER="xxx.xxx.xxx.xxx:12538"
#export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

python -m paddle.distributed.launch \
    --nnodes=$PADDLE_NNODES \
    --master=$PADDLE_MASTER \
    --devices=$CUDA_VISIBLE_DEVICES \
    plsc-train \
    -c ./configs/ConvNeXt_base_224_in1k_4n32c_dp_fp32.yaml

How to Evaluation

# [Optional] Download checkpoint
mkdir -p pretrained/
wget -O ./pretrained/ConvNeXt_base_224_in1k_dp_fp32.pdparams https://plsc.bj.bcebos.com/models/convnext/v2.5/ConvNeXt_base_224_in1k_dp_fp32.pdparams

export PADDLE_NNODES=1
export PADDLE_MASTER="127.0.0.1:12538"
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

python -m paddle.distributed.launch \
  --nnodes=$PADDLE_NNODES \
  --master=$PADDLE_MASTER \
  --devices=$CUDA_VISIBLE_DEVICES \
  plsc-eval \
  -c ./configs/ConvNeXt_base_224_in1k_1n8c_dp_fp32.yaml \
  -o Global.pretrained_model=pretrained/ConvNeXt_base_224_in1k_dp_fp32 \
  -o Global.finetune=False

Other Configurations

We provide more directly runnable configurations, see ConvNeXt Configurations.

Models

Model	DType	Phase	Dataset	Configs	GPUs	Img/sec	Top1 Acc	Pre-trained checkpoint	Log
convnext_base	FP32	pretrain	ImageNet2012	config	A100*N4C32	7800	0.838	download	log

Citations

@Article{liu2022convnet,
  author  = {Zhuang Liu and Hanzi Mao and Chao-Yuan Wu and Christoph Feichtenhofer and Trevor Darrell and Saining Xie},
  title   = {A ConvNet for the 2020s},
  journal = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year    = {2022},
}