Research map / Architectures and efficiency / CNN design
KernelWarehouse: Rethinking the Design of Dynamic Convolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KernelWarehouse, a generalized dynamic convolution that partitions kernels into cells, shares a warehouse of cells across same-stage layers, and uses a contrasting-driven attention function to enable n>100 while controlling parameter budget.
Problem
Dynamic convolution multiplies convolutional parameters by the kernel number n, so it is parameter-inefficient and prevents exploring large n (n>100).
Approach
Redefines kernels as small kernel cells obtained by partitioning static kernels, assembles a warehouse of n cells shared across same-stage layers with a convolutional parameter budget b=n/m_t, and replaces softmax attention with a contrasting-driven attention function (CAF) that allows negative and positive attentions with temperature-based one-hot initialization.
Key results
- KW (4x) on ResNet18 reaches 76.05% top-1 on ImageNet with advanced training, a 5.61% gain over baseline (vs. 4.75% for KW 2x and 4.33% for KW 1x).
- KW (1/4x) on ResNet18 uses 4.08M params (65.10% reduction) and still gains 2.29% top-1.
- KW (4x) on DeiT-Tiny reaches 76.51% top-1, a 4.38% gain over the 72.13% baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace normal convolution with KW (4×) | ResNet18 · ImageNet val (advanced training) | Top-1 Acc (%): 70.44 -> 76.05 | Helped |
| use KW (1/4×) instead of the baseline | ResNet18 · ImageNet val (advanced training) | Top-1 Acc (%): 70.44 -> 72.73 | Helped |
| replace normal convolution with KW (4×) | ResNet50 · ImageNet val (advanced training) | Top-1 Acc (%): 78.44 -> 81.05 | Helped |
| replace normal convolution with KW (1×) | ConvNeXt-Tiny · ImageNet val | Top-1 Acc (%): 82.07 -> 82.55 | Helped |
| replace normal convolution with KW (4×) | MobileNetV2 (1.0×) · ImageNet val | Top-1 Acc (%): 72.02 -> 75.92 | Helped |
| replace normal convolution with KW (4×) | DeiT-Tiny · ImageNet val | Top-1 Acc (%): 72.13 -> 76.51 | Helped |
| replace normal convolution with KW (4×) | Mask R-CNN with ResNet50 · MS-COCO 2017 val | Object detection AP: 39.6 -> 42.4 | Helped |
| add kernel partition (ablation) | KW (1×) on ResNet18 · ImageNet val | Top-1 Acc (%): 70.49 -> 74.77 | Helped |
| share the warehouse within each stage instead of within each layer (ablation) | KW (1×) on ResNet18 · ImageNet val | Top-1 Acc (%): 74.34 -> 74.77 | Helped |
| add warehouse sharing instead of no sharing (ablation) | KW (1×) on ResNet18 · ImageNet val | Top-1 Acc (%): 72.49 -> 74.77 | Helped |
| share across different-dimensioned rather than only same-dimensioned kernels (ablation) | KW (1×) on ResNet50 · ImageNet val | Top-1 Acc (%): 79.80 -> 80.38 | Helped |
| use CAF rather than Softmax attention (ablation) | KW (1×) on ResNet18 · ImageNet val | Top-1 Acc (%): 72.67 -> 74.77 | Helped |
Surprising findings
- Popular attention functions (softmax, sigmoid) lose effectiveness in the warehouse setting, and CAF gives slight accuracy drops when applied to DY-Conv and ODConv.
- KW (1/2x) on ResNet50 and ConvNeXt-Tiny KW (3/4x) reduce parameters while still improving accuracy.
Limitations
- Runtime speed is slower than vanilla DY-Conv on larger backbones such as ResNet50 (265.4 vs 322.7 FPS for KW 1x vs DY-Conv 4x).
- CAF gives slight accuracy drops when applied to DY-Conv and ODConv, so the design is tailored to KernelWarehouse.
- Attention initialization is not used during detection fine-tuning, and many hyperparameters (n, m per layer) are hand-designed per backbone.
Experimental setup
Datasets: ImageNet-1K MS-COCO 2017
Benchmarks: ImageNet-1K MS-COCO 2017
Metrics: top-1 accuracy top-5 accuracy AP AP50 AP75 APS APM APL params FPS
Baselines: SE CBAM ECA CGC WeightNet DCD CondConv DY-Conv ODConv
Builds on: CondConv DY-Conv ODConv
Closest papers by meaning
- Dynamic Convolution: Attention Over Convolution Kernels (CVPR 2020 · similarity 0.92)
- KernelDNA: Cross-Layer Kernel Sharing via Decoupled Neural Adapters (NeurIPS 2026 · similarity 0.91)
- Approximation and Learning with Deep Convolutional Models: a Kernel Perspective (ICLR 2022 · similarity 0.91)
- Omni-Dimensional Dynamic Convolution (ICLR 2022 · similarity 0.91)
- Frequency Dynamic Convolution for Dense Image Prediction (CVPR 2025 · similarity 0.87)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.