A ConvNet for the 2020s, arxiv
PaddlePaddle training/validation code and pretrained models for the model released in CVPR2022: ConvNeXt.
The official PyTorch implementation is here.
This implementation is developed by PaddleViT.
- Update (2022-05-16): Code is released and ported weights are uploaded.
Model | Acc@1 | Acc@5 | #Params | FLOPs | Image Size | Crop_pct | Interpolation | Link |
---|---|---|---|---|---|---|---|---|
ConvNeXt-T(22kft1k) | 82.91 | 96.62 | 28.6M | 4.5G | 224 | 0.9 | bicubic | google/baidu |
ConvNeXt-S(22kft1k) | 84.57 | 97.39 | 50.2M | 8.8G | 224 | 0.9 | bicubic | google/baidu |
ConvNeXt-B(22kft1k) | 85.82 | 97.87 | 88.6M | 15.5G | 224 | 0.9 | bicubic | google/baidu |
ConvNeXt-L(22kft1k) | 86.64 | 98.03 | 197.7M | 34.6G | 224 | 0.9 | bicubic | google/baidu |
ConvNeXt-XL(22kft1k) | 87.01 | 98.20 | 350.2M | 61.3G | 224 | 0.9 | bicubic | google/baidu |
*The results are evaluated on ImageNet2012 validation set.
ImageNet2012 dataset is used in the following file structure:
│imagenet/
├──train_list.txt
├──val_list.txt
├──train/
│ ├── n01440764
│ │ ├── n01440764_10026.JPEG
│ │ ├── n01440764_10027.JPEG
│ │ ├── ......
│ ├── ......
├──val/
│ ├── n01440764
│ │ ├── ILSVRC2012_val_00000293.JPEG
│ │ ├── ILSVRC2012_val_00002138.JPEG
│ │ ├── ......
│ ├── ......
train_list.txt
: list of relative paths and labels of training images. You can download it from: google/baiduval_list.txt
: list of relative paths and labels of validation images. You can download it from: google/baidu
To use the model with pretrained weights, download the .pdparam
weight file and change related file paths in the following python scripts. The model config files are located in ./configs/
.
For example, assume weight file is downloaded in ./convnext_tiny.pdparams
, to use the convnext_tiny
model in python:
from config import get_config
from convnext import build_convnext as build_model
# config files in ./configs/
config = get_config('./configs/convnext_tiny.yaml')
# build model
model = build_model(config)
# load pretrained weights
model_state_dict = paddle.load('./convnext_tiny.pdparams')
model.set_state_dict(model_state_dict)
To evaluate model performance on ImageNet2012, run the following script using command line:
sh run_eval_multi.sh
or
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python main_multi_gpu.py \
-cfg='./configs/convnext_tiny.yaml' \
-dataset='imagenet2012' \
-batch_size=256 \
-data_path='/dataset/imagenet' \
-eval \
-pretrained='./convnext_tiny.pdparams' \
-amp
Note: if you have only 1 GPU, change device number to
CUDA_VISIBLE_DEVICES=0
would run the evaluation on single GPU.
To train the model on ImageNet2012, run the following script using command line:
sh run_train_multi.sh
or
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python main_multi_gpu.py \
-cfg='./configs/convnext_tiny.yaml' \
-dataset='imagenet2012' \
-batch_size=256 \
-data_path='/dataset/imagenet' \
-amp
Note: it is highly recommanded to run the training using multiple GPUs / multi-node GPUs.
@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},
}