KernelDNA

KernelDNA is a plug-and-play convolution module that enhances model capacity through adaptive kernel specialization while maintaining hardware-friendly efficiency.

Key Innovations

✅ Parameter-Efficient Design

• Replaces dense convolutional layers with derived "child" kernels from a shared "parent" kernel
• Avoids linear parameter growth typical in dynamic convolutions

✅ Hardware-Optimized Inference

• Decouples adaptation into:
  • Input-dependent dynamic routing
  • Pre-trained static modulation
• Preserves standard convolution's native computational efficiency

✅ Enhanced Representation

• Achieves input-adaptive kernel adjustments without structural changes
• Outperforms existing dynamic convolutions in accuracy-efficiency trade-off

Performance Highlights

• State-of-the-art results on image classification & dense prediction tasks
• Maintains >90% of baseline throughput while improving accuracy
• Compatible with pre-trained CNNs via adapter-based fine-tuning

Installation

git clone https://github.com/haiduo/KernelDNA.git
cd KernelDNA
pip install -r requirements

Training on Classification Task

KernelDNA for ResNet18

python main.py --data /path/to/imagenet-1k --ckpt_dir /path/to/save_ckpt --log_dir /path/to/save_log --arch resnet18 --customize --epochs 90 --warmup-epochs 0 --lr 0.1 --wd 1e-4 --batch-size 256

KernelDNA for ResNet50

python main.py --data /path/to/imagenet-1k --ckpt_dir /path/to/save_ckpt --log_dir /path/to/save_log --arch resnet50 --customize --epochs 300 --warmup-epochs 20 --lr 0.1 --wd 1e-4 --batch-size 256

KernelDNA for MobileNetV2-1x

python main.py --data /path/to/imagenet-1k --ckpt_dir /path/to/save_ckpt --log_dir /path/to/save_log --arch mobilenet_v2 --customize --epochs 150 --warmup-epochs 0 --lr 0.05 --wd 4e-5 --batch-size 256

KernelDNA for MobileNetV2-0.5x

python main.py --data /path/to/imagenet-1k --ckpt_dir /path/to/save_ckpt --log_dir /path/to/save_log --arch mobilenet_v2_1d2 --customize --epochs 150 --warmup-epochs 0 --lr 0.05 --wd 4e-5 --batch-size 256

KernelDNA for ConvNeXt-Tiny

python main.py --data /path/to/imagenet-1k --ckpt_dir /path/to/save_ckpt --log_dir /path/to/save_log --arch convNeXt-tiny --customize --epochs 150 --warmup-epochs 0 --lr 0.05 --wd 4e-5 --batch-size 256

Training on Dense Prediction Task

KernelDNA for ResNet18

cd detection_and_segmentation
bash mmdetection/tools/dist_train.sh

Evaluation

KernelDNA ResNet18

python main.py --data /path/to/imagenet-1k --evaluate --resume /path/to/ckpt --log_dir /path/to/save_log --arch resnet18 --customize 

KernelDNA ResNet50

python main.py --data /path/to/imagenet-1k --evaluate --resume /path/to/ckpt --log_dir /path/to/save_log --arch resnet50 --customize 

KernelDNA MobileNetV2-1x

python main.py --data /path/to/imagenet-1k --evaluate --resume /path/to/ckpt --log_dir /path/to/save_log --arch mobilenet_v2 --customize 

KernelDNA MobileNetV2-0.5x

python main.py --data /path/to/imagenet-1k --evaluate --resume /path/to/ckpt --log_dir /path/to/save_log --arch mobilenet+_v2_1d2 --customize 

KernelDNA ConvNeXt-Tiny

python main.py --data /path/to/imagenet-1k --evaluate --resume /path/to/ckpt --log_dir /path/to/save_log --arch convNeXt-tiny --customize 

Throughput & Latency on GPU

KernelDNA ResNet18

python metrics.py --params --latency --gpu --iter 1000 --batch-size 128 --arch resnet18 --customize 

KernelDNA ResNet50

python metrics.py --params --latency --gpu --iter 1000 --batch-size 128 --arch resnet50 --customize 

KernelDNA MobileNetV2-1x

python metrics.py --params --latency --gpu --iter 1000 --batch-size 128 --arch mobilenet_v2 --customize 

KernelDNA MobileNetV2-0.5x

python metrics.py --params --latency --gpu --iter 1000 --batch-size 128 --arch mobilenet+_v2_1d2 --customize 

Results

For technical details and full experimental results, please check the paper of KernelDNA.

Reference

@misc{huang2025kerneldnadynamickernelsharing,
      title={KernelDNA: Dynamic Kernel Sharing via Decoupled Naive Adapters}, 
      author={Haiduo Huang and Yadong Zhang and Pengju Ren},
      year={2025},
      eprint={2503.23379},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.23379}, 
}