{"notes":[{"id":"lightweight/lsnet","category":"lightweight","legacyPath":"","legacySlug":"","legacyPaths":[],"title":"LSNet","subtitle":"看得广，盯得紧","author":"Z. Yuan","paperDate":"2025-03-29","publishedOn":"2026-03-09","paperTitle":"LSNet: See Large, Focus Small","paperUrl":"https://arxiv.org/abs/2503.23135","bibliography":{"title":"LSNet: See Large, Focus Small","authors":["Ao Wang","Hui Chen","Zijia Lin","Jungong Han","Guiguang Ding"],"url":"https://arxiv.org/abs/2503.23135","pdfUrl":"https://arxiv.org/pdf/2503.23135v1","codeUrl":"https://github.com/jameslahm/lsnet","preprint":{"id":"2503.23135","submittedOn":"2025-03-29","url":"https://arxiv.org/abs/2503.23135"},"publication":{"venue":"CVPR","year":2025,"kind":"conference","url":"https://openaccess.thecvf.com/content/CVPR2025/html/Wang_LSNet_See_Large_Focus_Small_CVPR_2025_paper.html"},"citation":"@misc{lightweightlsnet2025,\n  title = {{LSNet: See Large, Focus Small}},\n  author = {Ao Wang and Hui Chen and Zijia Lin and Jungong Han and Guiguang Ding},\n  year = {2025},\n  eprint = {2503.23135},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2503.23135}\n}\n","citationPath":"/content/papers/citations/lightweight/lsnet.d31e2f81ec32.bib"},"thumbnail":{"src":"/content/papers/assets/6d3e348c9f8512697f7230b2ac3fde0efdd4fd84198dbb519d931acd26bcbb59.png","width":1728,"height":948},"locale":"zh-hans","path":"/content/papers/zh-hans/lightweight/lsnet.77ddeb86e3b5.md","headings":[{"id":"定义问题","text":"定义问题","depth":2},{"id":"感知和聚合","text":"感知和聚合","depth":3},{"id":"解决问题","text":"解决问题","depth":2},{"id":"ls-卷积","text":"LS 卷积","depth":3},{"id":"组成网络","text":"组成网络","depth":3},{"id":"讨论","text":"讨论","depth":2},{"id":"imagenet","text":"ImageNet","depth":3},{"id":"消融","text":"消融","depth":3},{"id":"部署","text":"部署","depth":3},{"id":"结论","text":"结论","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"reparameterization/repvit","relation":"referenced","topic":""},{"id":"object-detection/yolov10","relation":"referenced","topic":""},{"id":"classic-cnns/overlock","relation":"referenced","topic":""}]},{"id":"reparameterization/repvit","category":"reparameterization","legacyPath":"reparameterization/2307-repvit","legacySlug":"2307-repvit","legacyPaths":[],"title":"RepViT","subtitle":"参考 ViT 的观点","author":"Z. Yuan","paperDate":"2023-07-18","publishedOn":"2024-07-19","paperTitle":"RepViT: Revisiting Mobile CNN From ViT Perspective","paperUrl":"https://arxiv.org/abs/2307.09283","bibliography":{"title":"RepViT: Revisiting Mobile CNN From ViT Perspective","authors":["Ao Wang","Hui Chen","Zijia Lin","Jungong Han","Guiguang Ding"],"url":"https://arxiv.org/abs/2307.09283","pdfUrl":"https://arxiv.org/pdf/2307.09283","codeUrl":null,"preprint":{"id":"2307.09283","submittedOn":"2023-07-18","url":"https://arxiv.org/abs/2307.09283"},"publication":{"venue":"CVPR","year":2024,"kind":"conference","url":"https://arxiv.org/abs/2307.09283"},"citation":"@misc{reparameterizationrepvit2023,\n  title = {{RepViT: Revisiting Mobile CNN From ViT Perspective}},\n  author = {Ao Wang and Hui Chen and Zijia Lin and Jungong Han and Guiguang Ding},\n  year = {2023},\n  eprint = {2307.09283},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2307.09283}\n}\n","citationPath":"/content/papers/citations/reparameterization/repvit.583fd659fe6d.bib"},"thumbnail":{"src":"/content/papers/assets/f501c7d5eb209527b5401ba7607cd34496d2dda89aeb1848aa9927e0d0f713eb.jpg","width":1224,"height":704},"locale":"zh-hans","path":"/content/papers/zh-hans/reparameterization/repvit.4afbc49c4d19.md","headings":[{"id":"定义问题","text":"定义问题","depth":2},{"id":"解决问题","text":"解决问题","depth":2},{"id":"重新设计架构","text":"重新设计架构","depth":3},{"id":"重新参数化结构","text":"重新参数化结构","depth":3},{"id":"降低膨胀比并增加宽度","text":"降低膨胀比并增加宽度","depth":3},{"id":"stem-结构调整","text":"Stem 结构调整","depth":3},{"id":"降采样结构调整","text":"降采样结构调整","depth":3},{"id":"简单分类器","text":"简单分类器","depth":3},{"id":"阶段比例调整","text":"阶段比例调整","depth":3},{"id":"卷积核大小的选择","text":"卷积核大小的选择","depth":3},{"id":"squeeze-and-excitation-的抉择","text":"Squeeze-and-Excitation 的抉择","depth":3},{"id":"讨论","text":"讨论","depth":2},{"id":"imagenet-1k-上的性能","text":"ImageNet-1K 上的性能","depth":3},{"id":"消融实验---重新参数化的结构","text":"消融实验 - 重新参数化的结构","depth":3},{"id":"消融实验---se-层的影响","text":"消融实验 - SE 层的影响","depth":3},{"id":"结论","text":"结论","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"reparameterization/fastvit","relation":"comparison","topic":"面向部署的快速视觉骨干"},{"id":"reparameterization/mobileone","relation":"comparison","topic":"面向部署的快速视觉骨干"}]},{"id":"object-detection/yolov10","category":"object-detection","legacyPath":"object-detection/2405-yolov10","legacySlug":"2405-yolov10","legacyPaths":[],"title":"YOLOv10","subtitle":"NMS，再见","author":"Z. Yuan","paperDate":"2024-05-23","publishedOn":"2025-07-19","paperTitle":"YOLOv10: Real-Time End-to-End Object Detection","paperUrl":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/c34ddd05eb089991f06f3c5dc36836e0-Abstract-Conference.html","bibliography":{"title":"YOLOv10: Real-Time End-to-End Object Detection","authors":["Ao Wang","Hui Chen","Lihao Liu","Kai Chen","Zijia Lin","Jungong Han","Guiguang Ding"],"url":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/c34ddd05eb089991f06f3c5dc36836e0-Abstract-Conference.html","pdfUrl":"https://proceedings.neurips.cc/paper_files/paper/2024/file/c34ddd05eb089991f06f3c5dc36836e0-Paper-Conference.pdf","codeUrl":"https://github.com/THU-MIG/yolov10","preprint":{"id":"2405.14458","submittedOn":"2024-05-23","url":"https://arxiv.org/abs/2405.14458"},"publication":{"venue":"NeurIPS","year":2024,"kind":"conference","url":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/c34ddd05eb089991f06f3c5dc36836e0-Abstract-Conference.html"},"citation":"@misc{objectdetectionyolov102024,\n  title = {{YOLOv10: Real-Time End-to-End Object Detection}},\n  author = {Ao Wang and Hui Chen and Lihao Liu and Kai Chen and Zijia Lin and Jungong Han and Guiguang Ding},\n  year = {2024},\n  eprint = {2405.14458},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2405.14458}\n}\n","citationPath":"/content/papers/citations/object-detection/yolov10.f624ec9e20b6.bib"},"thumbnail":{"src":"/content/papers/assets/43843817d1e6667933eb36b7b588db3ca318fe74e00d2d47095426923f3a5b44.jpg","width":1614,"height":536},"locale":"zh-hans","path":"/content/papers/zh-hans/object-detection/yolov10.ba77eced71b7.md","headings":[{"id":"定义问题","text":"定义问题","depth":2},{"id":"解决问题","text":"解决问题","depth":2},{"id":"双重标注架构","text":"双重标注架构","depth":3},{"id":"效率与精度优化","text":"效率与精度优化","depth":3},{"id":"精度导向强化","text":"精度导向强化","depth":3},{"id":"实现细节","text":"实现细节","depth":3},{"id":"讨论","text":"讨论","depth":2},{"id":"与-sota-方法比较","text":"与 SOTA 方法比较","depth":3},{"id":"消融实验","text":"消融实验","depth":3},{"id":"nms-free-真的能完全取代传统做法吗","text":"NMS-free 真的能完全取代传统做法吗？","depth":3},{"id":"为什么精简不会拖垮性能","text":"为什么精简不会拖垮性能？","depth":3},{"id":"精度提升延迟却几乎不变","text":"精度提升，延迟却几乎不变？","depth":3},{"id":"结论","text":"结论","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"object-detection/yolov11","relation":"comparison","topic":"实时检测的近期架构设计"},{"id":"object-detection/yolo-tiny","relation":"comparison","topic":"实时检测的近期架构设计"}]},{"id":"classic-cnns/overlock","category":"classic-cnns","legacyPath":"","legacySlug":"","legacyPaths":[],"title":"OverLoCK","subtitle":"先看大局","author":"Z. Yuan","paperDate":"2025-02-27","publishedOn":"2026-02-24","paperTitle":"OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels","paperUrl":"https://arxiv.org/abs/2502.20087","bibliography":{"title":"OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels","authors":["Meng Lou","Yizhou Yu"],"url":"https://arxiv.org/abs/2502.20087","pdfUrl":"https://arxiv.org/pdf/2502.20087v3","codeUrl":"https://github.com/LMMMEng/OverLoCK","preprint":{"id":"2502.20087","submittedOn":"2025-02-27","url":"https://arxiv.org/abs/2502.20087"},"publication":{"venue":"CVPR","year":2025,"kind":"conference","url":"https://openaccess.thecvf.com/content/CVPR2025/html/Lou_OverLoCK_An_Overview-first-Look-Closely-next_ConvNet_with_Context-Mixing_Dynamic_Kernels_CVPR_2025_paper.html"},"citation":"@misc{classiccnnsoverlock2025,\n  title = {{OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels}},\n  author = {Meng Lou and Yizhou Yu},\n  year = {2025},\n  eprint = {2502.20087},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2502.20087}\n}\n","citationPath":"/content/papers/citations/classic-cnns/overlock.a1eecba8875b.bib"},"thumbnail":{"src":"/content/papers/assets/99458c439c8f5458671d09ec6d68f52f31a20702b6a0bdba99e2fe33dbf86ea7.png","width":2112,"height":681},"locale":"zh-hans","path":"/content/papers/zh-hans/classic-cnns/overlock.c982182e0497.md","headings":[{"id":"定义问题","text":"定义问题","depth":2},{"id":"中间层缺一个方向","text":"中间层缺一个方向","depth":3},{"id":"卷积核看不远","text":"卷积核看不远","depth":3},{"id":"解决问题","text":"解决问题","depth":2},{"id":"拆成三个子网络","text":"拆成三个子网络","depth":3},{"id":"contmix把全图塞进卷积核","text":"ContMix：把全图塞进卷积核","depth":3},{"id":"四个尺寸","text":"四个尺寸","depth":3},{"id":"讨论","text":"讨论","depth":2},{"id":"imagenet","text":"ImageNet","depth":3},{"id":"检测与分割","text":"检测与分割","depth":3},{"id":"一步一步加上去","text":"一步一步加上去","depth":3},{"id":"和其他动态算子比","text":"和其他动态算子比","depth":3},{"id":"实际用起来","text":"实际用起来","depth":3},{"id":"结论","text":"结论","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"classic-cnns/internimage","relation":"referenced","topic":""},{"id":"classic-cnns/dcnv4","relation":"referenced","topic":""}]}],"text":"轻量网络的空间混合，通常只有两个选择。自注意力看得广，可是贵；小卷积核便宜，可是看得窄，权重训练完还是固定的。\n\n这篇来自清华大学，第一作者 Ao Wang 也是 RepViT 和 YOLOv10 的第一作者。他们的灵感来自视网膜上的两种感光细胞。杆细胞分布在周边，看得广但模糊；锥细胞集中在中央，范围小却最清楚。论文标题 See Large, Focus Small 说的就是这种分工。\n\n\n> [!TIP]\n> \n> 同一位作者的另外两篇工作：\n> \n> - **[[23.07] RepViT：参考 ViT 的观点](/blog/papers/reparameterization/repvit)**\n> - **[[24.05] YOLOv10：NMS，再见](/blog/papers/object-detection/yolov10)**\n\n\n## 定义问题\n\n### 感知和聚合\n\n作者把 token mixing，也就是空间方向的信息交换，拆成两步。\n\n- **感知（perception）**：决定这个位置和周围哪些位置有关、关系多强，也就是算出聚合权重。\n- **聚合（aggregation）**：用这些权重，把周围的特征加权加总。\n\n[![原论文 Figure 1：(a) 自注意力，左边是猫狗照片的格子，一个深蓝色格子是查询，橘色箭头代表感知，经过 softmax 得到一个大范围的权重矩阵，绿色是有用的权重、红色是多余的权重，蓝色箭头代表聚合，范围和感知一样大。(b) 卷积，感知和聚合都只在一个 3×3 的小范围里，权重是固定的卷积核。(c) 人眼，视网膜分成周边的杆细胞和中央的锥细胞，左边橘色箭头从大范围的模糊图像进入，标示 See Large、Perception；右边蓝色箭头聚焦到清晰的局部，标示 Focus Small、Aggregation。(d) 折线图，横轴是离中央窝的偏心角度，蓝线是杆细胞密度，在周边很高、在中央降到零；红线是锥细胞密度，只在中央附近出现一个很高的尖峰。](/content/papers/assets/85567a02dcfb2f88ff1be569a4d077a422f781df88b73b63effe301953d2985b.png)](/content/papers/assets/85567a02dcfb2f88ff1be569a4d077a422f781df88b73b63effe301953d2985b.png)\n\n*原论文 Figure 1（Wang 等人），取自 [**PDF 第 2 页**](https://arxiv.org/pdf/2503.23135v1#page=2)，CC BY 4.0。原图注未收录。*\n\n用这两步来看，常见的两种 token mixer 各有问题。\n\n自注意力的感知和聚合范围一样大。查询和所有位置算相似度、做 softmax，再对所有位置加权加总。范围越大，计算量越重；而且常常把不少权重分给背景这类没什么信息的地方，就是图 (a) 里红色的格子。\n\n卷积的感知和聚合也一样大，只是都很小。聚合权重就是卷积核，只看相对位置，对每个位置都一样，训练完就固定了。轻量网络的卷积核通常很小，看的范围有限，也没办法依内容调整。\n\n两者都是「看多大就加多大」。作者想要的是把两个范围拆开，感知看得广，聚合只在小范围里做。\n\n## 解决问题\n\n### LS 卷积\n\n[![原论文 Figure 3：(a) LS 卷积。上半部橘色虚线框是 Large Kernel Perception：H×W×C 的输入先经过 1×1 卷积降到 C/2，再经过 K_L×K_L 的 depthwise 卷积，再经过两个 1×1 卷积，输出 H×W×D，重塑成 H×W×G×K_s×K_s。下半部蓝色虚线框是 Small Kernel Aggregation：每个位置拿到自己的 K_s×K_s 卷积核，同一组的 C/G 个通道共用，对 3×3 的邻域做卷积。(b) LSNet 架构，由下往上是 Stem、三段 LS Block 与 Downsample、最后一段 MSA Block、Pooling、FC；右侧展开 Stem 是三个步长 2 的 3×3 卷积，Downsample 是步长 2 的 depthwise 3×3 加 1×1；LS Block 由 DW3×3、SE、FFN、LS Convolution、FFN 组成；MSA Block 把 LS Convolution 换成 MHSA。](/content/papers/assets/6d3e348c9f8512697f7230b2ac3fde0efdd4fd84198dbb519d931acd26bcbb59.png)](/content/papers/assets/6d3e348c9f8512697f7230b2ac3fde0efdd4fd84198dbb519d931acd26bcbb59.png)\n\n*原论文 Figure 3（Wang 等人），取自 [**PDF 第 5 页**](https://arxiv.org/pdf/2503.23135v1#page=5)，CC BY 4.0。原图注未收录。*\n\nLS（Large-Small）卷积就是这个想法的实现，分成两段：\n\n1. **大卷积核感知**：输入先用 1×1 卷积把通道降到一半，再过一个 $K_L\\times K_L$ 的 depthwise 卷积，看一大片邻域，最后用 1×1 卷积产生每个位置的权重。论文默认 $K_L=7$。\n2. **小卷积核聚合**：把每个位置的权重整理成 $G$ 个 $K_S\\times K_S$ 的卷积核，对这个位置周围的小邻域做加权加总。论文默认 $K_S=3$，每组 8 个通道共用一个卷积核。\n\n所以聚合权重是动态的，每个位置不一样，而且是看过 7×7 的范围才决定的；真正被加总的，却只有 3×3 的邻居。我们把它写出来，用梯度看中心那个输出受哪些输入影响：\n\n```python\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\ntorch.manual_seed(0)\nC, H, W = 16, 15, 15\nG, KL, KS = 2, 7, 3                                   # 2 组、7×7 负责看、3×3 负责聚合\npw1 = nn.Conv2d(C, C // 2, 1)\ndw = nn.Conv2d(C // 2, C // 2, KL, padding=KL // 2, groups=C // 2)\npw2 = nn.Conv2d(C // 2, G * KS * KS, 1)\n\ndef ls_conv(x, detach_weights=False):\n    w = pw2(dw(pw1(x)))                               # 大卷积核感知：每个位置、每一组各产生一个 3×3 卷积核\n    if detach_weights:\n        w = w.detach()                                # 只留下「拿来加总的值」那条路\n    w = w.view(G, 1, KS * KS, H * W)\n    patches = F.unfold(x, KS, padding=KS // 2).view(G, C // G, KS * KS, H * W)\n    return (patches * w).sum(2).view(1, C, H, W)      # 小卷积核聚合：同一组的通道共用卷积核\n\ndef footprint(detach_weights):\n    x = torch.randn(1, C, H, W, requires_grad=True)\n    ls_conv(x, detach_weights)[0, :, H // 2, W // 2].sum().backward()\n    grid = x.grad[0].abs().sum(0) > 0\n    rows, cols = grid.nonzero().T\n    return f\"{int(grid.sum())} 个位置，范围 {int(rows.max() - rows.min() + 1)}×{int(cols.max() - cols.min() + 1)}\"\n\nprint(\"中心输出受哪些输入影响：\", footprint(False))\nprint(\"其中真正被加总的值：\", footprint(True))\n```\n\n```text\n中心输出受哪些输入影响： 49 个位置，范围 7×7\n其中真正被加总的值： 9 个位置，范围 3×3\n```\n\n49 个位置参与决定权重，只有 9 个位置的值被加起来，这就是 See Large, Focus Small。实际的模块还有归一化和激活函数，这里省略了。\n\n计算量也因此压得很低。大范围的部分只用 depthwise 卷积，小范围的部分只加 9 个值，整体和分辨率呈线性关系，不像自注意力那样随范围平方增长。\n\n\n> [!TIP]\n> \n> LS 卷积产生权重的方式，和同样「先看大范围、再决定小范围权重」的 OverLoCK 可以对照着看：\n> \n> - **[[25.02] OverLoCK：先看大局](/blog/papers/classic-cnns/overlock)**\n\n\n### 组成网络\n\nLS 区块把 LS 卷积当成空间混合，旁边加上残差，再多放一个 depthwise 卷积和 SE 层，最后用 FFN 混合通道。\n\nLSNet 有四个阶段。前三个阶段用 LS 区块；最后一个阶段分辨率已经只剩 $H/64$，改用多头自注意力的 MSA 区块。所以 LSNet 严格说来不是纯卷积网络，最后一段用了注意力。\n\n三个尺寸 T、S、B 的 FLOPs 分别是 0.3G、0.5G 和 1.3G，区块都集中在后面的阶段，因为前面分辨率高、比较耗时。\n\n## 讨论\n\n### ImageNet\n\n吞吐量在一张 RTX 3090 上量，批量设成内存放得下的最大二的次方：\n\n| 模型 | 参数量 | FLOPs | 吞吐量（张／秒） | top-1 |\n| --- | ---: | ---: | ---: | ---: |\n| FasterNet-T0 | 3.9M | 0.3G | 14467 | 71.9% |\n| EfficientViT-M3 | 6.9M | 0.3G | 14613 | 73.4% |\n| StarNet-S1 | 2.9M | 0.4G | 5034 | 73.5% |\n| LSNet-T | 11.4M | 0.3G | **14708** | **74.9%** |\n| UniRepLKNet-A | 4.4M | 0.6G | 3931 | 77.0% |\n| RepViT-M0.9 | 5.1M | 0.8G | 4817 | 77.4% |\n| SHViT-S3 | 14.2M | 0.6G | 8993 | 77.4% |\n| LSNet-S | 16.1M | 0.5G | **9023** | **77.8%** |\n| FastViT-T12 | 6.8M | 1.4G | 2586 | 79.1% |\n| RepViT-M1.1 | 8.2M | 1.3G | 3604 | 79.4% |\n| LSNet-B | 23.2M | 1.3G | **3996** | **80.3%** |\n\n*表：摘自原论文 Table 1 的部分行，不含蒸馏版本。*\n\n同样的 FLOPs，LSNet 的准确率和吞吐量都排在前面。LSNet-T 比 StarNet-S1 高 1.4 个百分点，吞吐量快了将近三倍；LSNet-B 比 RepViT-M1.1 高 0.9。用 RegNetY-16GF 当老师做蒸馏，三个尺寸还能再多 1.2 到 1.3 个百分点。\n\n有两件事要一起看。\n\n第一是参数量。LSNet-T 只有 0.3G FLOPs，参数却有 11.4M，是 StarNet-S1 的四倍；LSNet-B 有 23.2M。如果部署环境卡的是模型大小或内存，这张表的排名就要重新看。\n\n第二是测速的方式。论文只量了 RTX 3090 上的大批量吞吐量，没有手机或 CPU 的延迟。StarNet 的优势在 iPhone 上批量 1 的延迟，两篇量的是不同的东西，这张表上的「快将近三倍」不能直接搬到手机上。\n\n### 消融\n\n这组实验用 LSNet-T 只训练 100 个 epoch，基准是 71.6%：\n\n| 改动 | top-1 |\n| --- | ---: |\n| LS 卷积（基准） | 71.6% |\n| 拿掉 LS 卷积 | 69.3% |\n| 拿掉大卷积核感知 | 70.5% |\n| 感知的卷积核 7 改成 3 | 70.9% |\n| 感知的卷积核 7 改成 9 | 71.5% |\n| 拿掉小卷积核聚合（大卷积核接一般小卷积） | 70.1% |\n| 聚合的卷积核 3 改成 1 | 69.6% |\n| 聚合的卷积核 3 改成 5 | 71.6% |\n\n*表：摘自原论文 Table 5 与 Table 7。*\n\n有两行特别值得看。\n\n「拿掉小卷积核聚合」是把大卷积核感知的结果，直接接一个一般的静态小卷积核，也就是「大核加小核」最直觉的组法，结果只有 70.1%，比 LS 卷积少 1.5。大范围的信息要拿来决定权重，比直接叠上去有用。\n\n聚合范围从 3 加到 5，准确率完全没变，FLOPs 却从 0.31G 涨到 0.34G。感知范围从 7 加到 9 也没有更好。看 7×7、加 3×3，刚好。\n\n作者也把 LS 卷积换进其他网络。ResNet-50 的 3×3 卷积全部换掉，FLOPs 从 4.1G 降到 2.6G，准确率从 78.8% 升到 80.7%；DeiT-T 的自注意力换掉，FLOPs 从 1.3G 降到 0.9G，准确率从 72.2% 升到 73.0%。\n\n### 部署\n\nLS 卷积的小卷积核聚合不是一般的卷积。每个位置的卷积核都不一样，PyTorch 内置的 `conv2d` 做不到，官方代码用 Triton 写了专用的 GPU 算子。GitHub 上有用户[问为什么不用内置的卷积](https://github.com/jameslahm/lsnet/issues/8)，也提到转成 TensorRT 时兼容性不好。要部署到手机或其他推理引擎，这一步得自己改写，例如用 `unfold` 加逐元素相乘，就像上面那段示范程序。\n\n## 结论\n\nLSNet 把 token mixing 拆成感知和聚合两步，再让两步用不同的范围。7×7 的 depthwise 卷积负责看，决定每个位置的 3×3 权重，聚合只做 9 个值的加权加总。消融表说得很清楚，大范围的信息拿去决定权重，比直接把大卷积核和小卷积核叠起来有用。\n\n比起整个 LSNet，LS 卷积这个零件更值得带走。它只用到 depthwise 卷积和逐位置的加权加总，换进 ResNet-50 之后，FLOPs 少了将近四成，准确率还多了 1.9 个百分点。\n","imageSizes":{"/content/papers/assets/85567a02dcfb2f88ff1be569a4d077a422f781df88b73b63effe301953d2985b.png":{"width":1743,"height":432},"/content/papers/assets/6d3e348c9f8512697f7230b2ac3fde0efdd4fd84198dbb519d931acd26bcbb59.png":{"width":1728,"height":948}}}