UniConvNet: Expanding Effective Receptive Field while Maintaining Asymptotically Gaussian Distribution for ConvNets of Any Scale

Official PyTorch implementation of UniConvNet, from the following paper:

UniConvNet: Expanding Effective Receptive Field while Maintaining Asymptotically Gaussian Distribution for ConvNets of Any Scale.
ICCV 2025.
Yuhao Wang, Wei Xi
Xi'an Jiaotong University
[arXiv]

---

We propose UniConvNet, a pure ConvNet model constructed entirely from standard ConvNet modules. UniConvNet performs well on both lightweight and large-scale models.

Catalog

• ImageNet-1K Training Code
• ImageNet-22K Pre-training Code
• ImageNet-1K Fine-tuning Code
• Downstream Transfer (Detection, Segmentation) Code (Coming soon ...)

Results and Pre-trained Models

ImageNet-1K trained models

name	resolution	acc@1	#params	FLOPs	model(hugging face)	model(baidu)
UniConvNet-A	224x224	77.0	3.4M	0.589G	model	model
UniConvNet-P0	224x224	79.1	5.2M	0.932G	model	model
UniConvNet-P1	224x224	79.6	6.1M	0.895G	model	model
UniConvNet-P2	224x224	80.5	7.6M	1.25G	model	model
UniConvNet-N0	224x224	81.6	10.2M	1.65G	model	model
UniConvNet-N1	224x224	82.2	13.1M	1.88G	model	model
UniConvNet-N2	224x224	82.7	15.0M	2.47G	model	model
UniConvNet-N3	224x224	83.2	19.7M	3.37G	model	model
UniConvNet-T	224x224	84.2	30.3M	5.1G	model	model
UniConvNet-T	384x384	85.4	30.3M	15.0G	model	model
UniConvNet-S	224x224	84.5	50.0M	8.48G	model	model
UniConvNet-S	384x384	85.7	50.0M	24.9G	model	model
UniConvNet-B	224x224	85.0	97.6M	15.9G	model	model
UniConvNet-B	384x384	85.9	97.6M	46.6G	model	model

ImageNet-22K trained models

name	resolution	acc@1	#params	FLOPs	22k model (hugging face)	22k model (baidu)	1k model (hugging face)	22k model (baidu)
ConvNeXt-L	384x384	88.2	201.8M	100.1G	model	model	model	model
ConvNeXt-XL	384x384	88.4	226.7M	115.2G	model	model	model	model

Installation

Please check INSTALL.md for installation instructions.

Evaluation

We give an example evaluation command for a ImageNet-1K pre-trained UniConvNet-A:

Single-GPU

python main.py --model UniConvNet_A --eval true \
--resume https://huggingface.co/ai-modelwithcode/UniConvNet/resolve/main/uniconvnet_a_1k_224.pth \
--input_size 224 --drop_path 0.05 \
--data_path /path/to/imagenet-1k

Multi-GPU

python -m torch.distributed.launch --nproc_per_node=8 main.py \
--model UniConvNet_A --eval true \
--resume https://huggingface.co/ai-modelwithcode/UniConvNet/resolve/main/uniconvnet_a_1k_224.pth \
--input_size 224 --drop_path 0.05 \
--data_path /path/to/imagenet-1k

This should give

* Acc@1 77.030 Acc@5 93.364 loss 0.983

• 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, ConvNeXt and InternImage 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:

@InProceedings{Wang_2025_ICCV,
    author    = {Wang, Yuhao and Xi, Wei},
    title     = {UniConvNet: Expanding Effective Receptive Field while Maintaining Asymptotically Gaussian Distribution for ConvNets of Any Scale},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {20922-20933}
}