jameslahm/lsnet_s_distill

LSNet: See Large, Focus Small

Paper: https://arxiv.org/abs/2503.23135

Code: https://github.com/jameslahm/lsnet

Usage

import timm
import torch
from PIL import Image
import requests
from timm.data import resolve_data_config, create_transform

# Load the model
model = timm.create_model(
    'hf_hub:jameslahm/lsnet_b',
    pretrained=True
)
model.eval()

# Load and transform image
# Example using a URL:
url = 'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
img = Image.open(requests.get(url, stream=True).raw)

config = resolve_data_config({}, model=model)
transform = create_transform(**config)
input_tensor = transform(img).unsqueeze(0) # transform and add batch dimension

# Make prediction
with torch.no_grad():
    output = model(input_tensor)
probabilities = torch.nn.functional.softmax(output[0], dim=0)

# Get top 5 predictions
top5_prob, top5_catid = torch.topk(probabilities, 5)
# Assuming you have imagenet labels list 'imagenet_labels'
# for i in range(top5_prob.size(0)):
#     print(imagenet_labels[top5_catid[i]], top5_prob[i].item())

Citation

If our code or models help your work, please cite our paper:

@misc{wang2025lsnetlargefocussmall,
      title={LSNet: See Large, Focus Small}, 
      author={Ao Wang and Hui Chen and Zijia Lin and Jungong Han and Guiguang Ding},
      year={2025},
      eprint={2503.23135},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.23135}, 
}

LSNet: See Large, Focus Small

Official PyTorch implementation of LSNet. CVPR 2025.

Models are trained on ImageNet-1K and the throughput is tested on a Nvidia RTX3090.

LSNet: See Large, Focus Small.
Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding

We introduce LSNet, a new family of lightweight vision models inspired by dynamic heteroscale capability of the human visual system, i.e., "See Large, Focus Small". LSNet achieves state-of-the-art performance and efficiency trade-offs across various vision tasks.

Abstract

Vision network designs, including Convolutional Neural Networks and Vision Transformers, have significantly advanced the field of computer vision. Yet, their complex computations pose challenges for practical deployments, particularly in real-time applications. To tackle this issue, researchers have explored various lightweight and efficient network designs. However, existing lightweight models predominantly leverage self-attention mechanisms and convolutions for token mixing. This dependence brings limitations in effectiveness and efficiency in the perception and aggregation processes of lightweight networks, hindering the balance between performance and efficiency under limited computational budgets. In this paper, we draw inspiration from the dynamic heteroscale vision ability inherent in the efficient human vision system and propose a "See Large, Focus Small" strategy for lightweight vision network design. We introduce LS (Large-Small) convolution, which combines large-kernel perception and small-kernel aggregation. It can efficiently capture a wide range of perceptual information and achieve precise feature aggregation for dynamic and complex visual representations, thus enabling proficient processing of visual information. Based on LS convolution, we present LSNet, a new family of lightweight models. Extensive experiments demonstrate that LSNet achieves superior performance and efficiency over existing lightweight networks in various vision tasks.

Classification on ImageNet-1K

Models

• * denotes the results with distillation.
• The throughput is tested on a Nvidia RTX3090 using speed.py.

Model	Top-1	Params	FLOPs	Throughput	Ckpt	Log
LSNet-T	74.9 / 76.1*	11.4M	0.3G	14708	T / T*	T / T*
LSNet-S	77.8 / 79.0*	16.1M	0.5G	9023	S / S*	S / S*
LSNet-B	80.3 / 81.6*	23.2M	1.3G	3996	B / B*	B / B*

ImageNet

Prerequisites

conda virtual environment is recommended.

conda create -n lsnet python=3.8
pip install -r requirements.txt

Data preparation

Download and extract ImageNet train and val images from http://image-net.org/. The training and validation data are expected to be in the train folder and val folder respectively:

|-- /path/to/imagenet/
    |-- train
    |-- val

Training

To train LSNet-T on an 8-GPU machine:

python -m torch.distributed.launch --nproc_per_node=8 --master_port 12345 --use_env main.py --model lsnet_t --data-path ~/imagenet --dist-eval
# For training with distillation, please add `--distillation-type hard`
# For LSNet-B, please add `--weight-decay 0.05`

Testing

python main.py --eval --model lsnet_t --resume ./pretrain/lsnet_t.pth --data-path ~/imagenet

Models can also be automatically downloaded from 🤗 like below.

import timm

model = timm.create_model(
    f'hf_hub:jameslahm/lsnet_{t/t_distill/s/s_distill/b/b_distill}',
    pretrained=True
)

Downstream Tasks

Object Detection and Instance Segmentation
Semantic Segmentation
Robustness Evaluation

Acknowledgement

Classification (ImageNet) code base is partly built with EfficientViT, LeViT, PoolFormer and EfficientFormer.

The detection and segmentation pipeline is from MMCV (MMDetection and MMSegmentation).

Thanks for the great implementations!