ConvNeXt with Multi-Scale Depthwise Convolution

⚠️ This is an improved version of the original ConvNeXt architecture
We have enhanced the original ConvNeXt by replacing the standard 7×7 depthwise convolution with a multi-scale depthwise convolution module. This improvement captures features at multiple receptive field scales simultaneously.

Our Improvement: Multi-Scale Depthwise Convolution

We replace the single-scale 7×7 depthwise convolution in ConvNeXt blocks with a multi-scale module that processes features at three different scales:

• Local Path (3×3): Captures fine-grained details (25% of channels)
• Standard Path (7×7): Preserves original ConvNeXt receptive field (50% of channels)
• Global Path (7×7 dilated): Captures broader context with effective ~19×19 receptive field (25% of channels)

Architecture Diagram

Figure: Multi-scale depthwise convolution module with 1:2:1 channel split ratio

Training Results

Figure: Training accuracy comparison between original ConvNeXt and our multi-scale variant

---

A ConvNet for the 2020s

Official PyTorch implementation of ConvNeXt, from the following paper:

A ConvNet for the 2020s. CVPR 2022.
Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell and Saining Xie
Facebook AI Research, UC Berkeley
[arXiv][video]

---

We propose ConvNeXt, a pure ConvNet model constructed entirely from standard ConvNet modules. ConvNeXt is accurate, efficient, scalable and very simple in design.

Catalog

• ImageNet-1K Training Code
• ImageNet-22K Pre-training Code
• ImageNet-1K Fine-tuning Code
• Downstream Transfer (Detection, Segmentation) Code
• Image Classification [Colab] and Web Demo
• Fine-tune on CIFAR with Weights & Biases logging [Colab]

Results and Pre-trained Models

ImageNet-1K trained models

name	resolution	acc@1	#params	FLOPs	model
ConvNeXt-T	224x224	82.1	28M	4.5G	model
ConvNeXt-S	224x224	83.1	50M	8.7G	model
ConvNeXt-B	224x224	83.8	89M	15.4G	model
ConvNeXt-B	384x384	85.1	89M	45.0G	model
ConvNeXt-L	224x224	84.3	198M	34.4G	model
ConvNeXt-L	384x384	85.5	198M	101.0G	model

ImageNet-22K trained models

name	resolution	acc@1	#params	FLOPs	22k model	1k model
ConvNeXt-T	224x224	82.9	29M	4.5G	model	model
ConvNeXt-T	384x384	84.1	29M	13.1G	-	model
ConvNeXt-S	224x224	84.6	50M	8.7G	model	model
ConvNeXt-S	384x384	85.8	50M	25.5G	-	model
ConvNeXt-B	224x224	85.8	89M	15.4G	model	model
ConvNeXt-B	384x384	86.8	89M	47.0G	-	model
ConvNeXt-L	224x224	86.6	198M	34.4G	model	model
ConvNeXt-L	384x384	87.5	198M	101.0G	-	model
ConvNeXt-XL	224x224	87.0	350M	60.9G	model	model
ConvNeXt-XL	384x384	87.8	350M	179.0G	-	model

ImageNet-1K trained models (isotropic)

name	resolution	acc@1	#params	FLOPs	model
ConvNeXt-S	224x224	78.7	22M	4.3G	model
ConvNeXt-B	224x224	82.0	87M	16.9G	model
ConvNeXt-L	224x224	82.6	306M	59.7G	model

Installation

Please check INSTALL.md for installation instructions.

Evaluation

We give an example evaluation command for a ImageNet-22K pre-trained, then ImageNet-1K fine-tuned ConvNeXt-B:

Single-GPU

python main.py --model convnext_base --eval true \
--resume https://dl.fbaipublicfiles.com/convnext/convnext_base_22k_1k_224.pth \
--input_size 224 --drop_path 0.2 \
--data_path /path/to/imagenet-1k

Multi-GPU

python -m torch.distributed.launch --nproc_per_node=8 main.py \
--model convnext_base --eval true \
--resume https://dl.fbaipublicfiles.com/convnext/convnext_base_22k_1k_224.pth \
--input_size 224 --drop_path 0.2 \
--data_path /path/to/imagenet-1k

This should give

* Acc@1 85.820 Acc@5 97.868 loss 0.563

• For evaluating other model variants, change --model, --resume, --input_size accordingly. You can get the url to pre-trained models from the tables above.
• Setting model-specific --drop_path is not strictly required in evaluation, as the DropPath module in timm behaves the same during evaluation; but it is required in training. See TRAINING.md or our paper for the values used for different models.

Training

See TRAINING.md for training and fine-tuning instructions.

Acknowledgement

This repository is built using the timm library, DeiT and BEiT repositories.

License

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

Citation

If you find this repository helpful, please consider citing:

@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},
}