{"notes":[{"id":"lightweight/lsnet","category":"lightweight","legacyPath":"","legacySlug":"","legacyPaths":[],"title":"LSNet","subtitle":"看得廣，盯得緊","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-hant","path":"/content/papers/zh-hant/lightweight/lsnet.45424d2cbe89.md","headings":[{"id":"定義問題","text":"定義問題","depth":2},{"id":"感知和聚合","text":"感知和聚合","depth":3},{"id":"解決問題","text":"解決問題","depth":2},{"id":"ls-卷積","text":"LS 卷積","depth":3},{"id":"組成網路","text":"組成網路","depth":3},{"id":"討論","text":"討論","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":"結論","text":"結論","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":"參考 ViT 的觀點","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-hant","path":"/content/papers/zh-hant/reparameterization/repvit.f82730f3b964.md","headings":[{"id":"定義問題","text":"定義問題","depth":2},{"id":"解決問題","text":"解決問題","depth":2},{"id":"重新設計架構","text":"重新設計架構","depth":3},{"id":"重新參數化結構","text":"重新參數化結構","depth":3},{"id":"降低膨脹比並增加寬度","text":"降低膨脹比並增加寬度","depth":3},{"id":"stem-結構調整","text":"Stem 結構調整","depth":3},{"id":"降採樣結構調整","text":"降採樣結構調整","depth":3},{"id":"簡單分類器","text":"簡單分類器","depth":3},{"id":"階段比例調整","text":"階段比例調整","depth":3},{"id":"內核大小的選擇","text":"內核大小的選擇","depth":3},{"id":"sqeeze-and-excitation-的抉擇","text":"Sqeeze-and-Excitation 的抉擇","depth":3},{"id":"討論","text":"討論","depth":2},{"id":"imagenet-1k-上的性能","text":"ImageNet-1K 上的性能","depth":3},{"id":"消融實驗---重新參數化的結構","text":"消融實驗 - 重新參數化的結構","depth":3},{"id":"消融實驗---se-層的影響","text":"消融實驗 - SE 層的影響","depth":3},{"id":"結論","text":"結論","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"reparameterization/fastvit","relation":"comparison","topic":"面向部署的快速視覺骨架"},{"id":"reparameterization/mobileone","relation":"comparison","topic":"面向部署的快速視覺骨架"}]},{"id":"object-detection/yolov10","category":"object-detection","legacyPath":"object-detection/2405-yolov10","legacySlug":"2405-yolov10","legacyPaths":[],"title":"YOLOv10","subtitle":"NMS，再見","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-hant","path":"/content/papers/zh-hant/object-detection/yolov10.f7c8787f1a99.md","headings":[{"id":"定義問題","text":"定義問題","depth":2},{"id":"解決問題","text":"解決問題","depth":2},{"id":"雙重標註架構","text":"雙重標註架構","depth":3},{"id":"效率與精度優化","text":"效率與精度優化","depth":3},{"id":"精度導向強化","text":"精度導向強化","depth":3},{"id":"實作細節","text":"實作細節","depth":3},{"id":"討論","text":"討論","depth":2},{"id":"與-sota-方法比較","text":"與 SOTA 方法比較","depth":3},{"id":"消融實驗","text":"消融實驗","depth":3},{"id":"nms-free-真的能完全取代傳統做法嗎","text":"NMS-free 真的能完全取代傳統做法嗎？","depth":3},{"id":"為什麼精簡不會拖垮效能","text":"為什麼精簡不會拖垮效能？","depth":3},{"id":"精度提升延遲卻幾乎不變","text":"精度提升，延遲卻幾乎不變？","depth":3},{"id":"結論","text":"結論","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"object-detection/yolov11","relation":"comparison","topic":"即時偵測的近期架構設計"},{"id":"object-detection/yolo-tiny","relation":"comparison","topic":"即時偵測的近期架構設計"}]},{"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-hant","path":"/content/papers/zh-hant/classic-cnns/overlock.f654e8b388c7.md","headings":[{"id":"定義問題","text":"定義問題","depth":2},{"id":"中間層缺一個方向","text":"中間層缺一個方向","depth":3},{"id":"卷積核看不遠","text":"卷積核看不遠","depth":3},{"id":"解決問題","text":"解決問題","depth":2},{"id":"拆成三個子網路","text":"拆成三個子網路","depth":3},{"id":"contmix把全圖塞進卷積核","text":"ContMix：把全圖塞進卷積核","depth":3},{"id":"四個尺寸","text":"四個尺寸","depth":3},{"id":"討論","text":"討論","depth":2},{"id":"imagenet","text":"ImageNet","depth":3},{"id":"偵測與分割","text":"偵測與分割","depth":3},{"id":"一步一步加上去","text":"一步一步加上去","depth":3},{"id":"和其他動態運算子比","text":"和其他動態運算子比","depth":3},{"id":"實際用起來","text":"實際用起來","depth":3},{"id":"結論","text":"結論","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"classic-cnns/internimage","relation":"referenced","topic":""},{"id":"classic-cnns/dcnv4","relation":"referenced","topic":""}]}],"text":"輕量網路的空間混合，通常只有兩個選擇。自注意力看得廣，可是貴；小卷積核便宜，可是看得窄，權重訓練完還是固定的。\n\n這篇來自清華大學，第一作者 Ao Wang 也是 RepViT 和 YOLOv10 的第一作者。他們的靈感來自視網膜上的兩種感光細胞。桿細胞分布在周邊，看得廣但模糊；錐細胞集中在中央，範圍小卻最清楚。論文標題 See Large, Focus Small 說的就是這種分工。\n\n\n> [!TIP]\n> \n> 同一位作者的另外兩篇工作：\n> \n> - **[[23.07] RepViT：參考 ViT 的觀點](/blog/papers/reparameterization/repvit)**\n> - **[[24.05] YOLOv10：NMS，再見](/blog/papers/object-detection/yolov10)**\n\n\n## 定義問題\n\n### 感知和聚合\n\n作者把 token mixing，也就是空間方向的資訊交換，拆成兩步。\n\n- **感知（perception）**：決定這個位置和周圍哪些位置有關、關係多強，也就是算出聚合權重。\n- **聚合（aggregation）**：用這些權重，把周圍的特徵加權加總。\n\n[![原論文 Figure 1：(a) 自注意力，左邊是貓狗照片的格子，一個深藍色格子是查詢，橘色箭頭代表感知，經過 softmax 得到一個大範圍的權重矩陣，綠色是有用的權重、紅色是多餘的權重，藍色箭頭代表聚合，範圍和感知一樣大。(b) 卷積，感知和聚合都只在一個 3×3 的小範圍裡，權重是固定的卷積核。(c) 人眼，視網膜分成周邊的桿細胞和中央的錐細胞，左邊橘色箭頭從大範圍的模糊影像進入，標示 See Large、Perception；右邊藍色箭頭聚焦到清晰的局部，標示 Focus Small、Aggregation。(d) 折線圖，橫軸是離中央窩的偏心角度，藍線是桿細胞密度，在周邊很高、在中央降到零；紅線是錐細胞密度，只在中央附近出現一個很高的尖峰。](/content/papers/assets/85567a02dcfb2f88ff1be569a4d077a422f781df88b73b63effe301953d2985b.png)](/content/papers/assets/85567a02dcfb2f88ff1be569a4d077a422f781df88b73b63effe301953d2985b.png)\n\n*原論文 Figure 1（Wang 等人），取自 [**PDF 第 2 頁**](https://arxiv.org/pdf/2503.23135v1#page=2)，CC BY 4.0。圖說未收錄。*\n\n用這兩步來看，常見的兩種 token mixer 各有問題。\n\n自注意力的感知和聚合範圍一樣大。查詢和所有位置算相似度、做 softmax，再對所有位置加權加總。範圍越大，計算量越重；而且常常把不少權重分給背景這類沒什麼資訊的地方，就是圖 (a) 裡紅色的格子。\n\n卷積的感知和聚合也一樣大，只是都很小。聚合權重就是卷積核，只看相對位置，對每個位置都一樣，訓練完就固定了。輕量網路的卷積核通常很小，看的範圍有限，也沒辦法依內容調整。\n\n兩者都是「看多大就加多大」。作者想要的是把兩個範圍拆開，感知看得廣，聚合只在小範圍裡做。\n\n## 解決問題\n\n### LS 卷積\n\n[![原論文 Figure 3：(a) LS 卷積。上半部橘色虛線框是 Large Kernel Perception：H×W×C 的輸入先經過 1×1 卷積降到 C/2，再經過 K_L×K_L 的 depthwise 卷積，再經過兩個 1×1 卷積，輸出 H×W×D，重塑成 H×W×G×K_s×K_s。下半部藍色虛線框是 Small Kernel Aggregation：每個位置拿到自己的 K_s×K_s 卷積核，同一組的 C/G 個通道共用，對 3×3 的鄰域做卷積。(b) LSNet 架構，由下往上是 Stem、三段 LS Block 與 Downsample、最後一段 MSA Block、Pooling、FC；右側展開 Stem 是三個步長 2 的 3×3 卷積，Downsample 是步長 2 的 depthwise 3×3 加 1×1；LS Block 由 DW3×3、SE、FFN、LS Convolution、FFN 組成；MSA Block 把 LS Convolution 換成 MHSA。](/content/papers/assets/6d3e348c9f8512697f7230b2ac3fde0efdd4fd84198dbb519d931acd26bcbb59.png)](/content/papers/assets/6d3e348c9f8512697f7230b2ac3fde0efdd4fd84198dbb519d931acd26bcbb59.png)\n\n*原論文 Figure 3（Wang 等人），取自 [**PDF 第 5 頁**](https://arxiv.org/pdf/2503.23135v1#page=5)，CC BY 4.0。圖說未收錄。*\n\nLS（Large-Small）卷積就是這個想法的實作，分成兩段：\n\n1. **大卷積核感知**：輸入先用 1×1 卷積把通道降到一半，再過一個 $K_L\\times K_L$ 的 depthwise 卷積，看一大片鄰域，最後用 1×1 卷積產生每個位置的權重。論文預設 $K_L=7$。\n2. **小卷積核聚合**：把每個位置的權重整理成 $G$ 個 $K_S\\times K_S$ 的卷積核，對這個位置周圍的小鄰域做加權加總。論文預設 $K_S=3$，每組 8 個通道共用一個卷積核。\n\n所以聚合權重是動態的，每個位置不一樣，而且是看過 7×7 的範圍才決定的；真正被加總的，卻只有 3×3 的鄰居。我們把它寫出來，用梯度看中心那個輸出受哪些輸入影響：\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 組、7×7 負責看、3×3 負責聚合\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)))                               # 大卷積核感知：每個位置、每一組各產生一個 3×3 卷積核\n    if detach_weights:\n        w = w.detach()                                # 只留下「拿來加總的值」那條路\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)      # 小卷積核聚合：同一組的通道共用卷積核\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())} 個位置，範圍 {int(rows.max() - rows.min() + 1)}×{int(cols.max() - cols.min() + 1)}\"\n\nprint(\"中心輸出受哪些輸入影響：\", footprint(False))\nprint(\"其中真正被加總的值：\", footprint(True))\n```\n\n```text\n中心輸出受哪些輸入影響： 49 個位置，範圍 7×7\n其中真正被加總的值： 9 個位置，範圍 3×3\n```\n\n49 個位置參與決定權重，只有 9 個位置的值被加起來，這就是 See Large, Focus Small。實際的模組還有正規化和啟動函數，這裡省略了。\n\n計算量也因此壓得很低。大範圍的部分只用 depthwise 卷積，小範圍的部分只加 9 個值，整體和解析度呈線性關係，不像自注意力那樣隨範圍平方成長。\n\n\n> [!TIP]\n> \n> LS 卷積產生權重的方式，和同樣「先看大範圍、再決定小範圍權重」的 OverLoCK 可以對照著看：\n> \n> - **[[25.02] OverLoCK：先看大局](/blog/papers/classic-cnns/overlock)**\n\n\n### 組成網路\n\nLS 區塊把 LS 卷積當成空間混合，旁邊加上殘差，再多放一個 depthwise 卷積和 SE 層，最後用 FFN 混合通道。\n\nLSNet 有四個階段。前三個階段用 LS 區塊；最後一個階段解析度已經只剩 $H/64$，改用多頭自注意力的 MSA 區塊。所以 LSNet 嚴格說來不是純卷積網路，最後一段用了注意力。\n\n三個尺寸 T、S、B 的 FLOPs 分別是 0.3G、0.5G 和 1.3G，區塊都集中在後面的階段，因為前面解析度高、比較耗時。\n\n## 討論\n\n### ImageNet\n\n吞吐量在一張 RTX 3090 上量，批次設成記憶體放得下的最大二的次方：\n\n| 模型 | 參數量 | FLOPs | 吞吐量（張／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：摘自原論文 Table 1 的部分列，不含蒸餾版本。*\n\n同樣的 FLOPs，LSNet 的準確率和吞吐量都排在前面。LSNet-T 比 StarNet-S1 高 1.4 個百分點，吞吐量快了將近三倍；LSNet-B 比 RepViT-M1.1 高 0.9。用 RegNetY-16GF 當老師做蒸餾，三個尺寸還能再多 1.2 到 1.3 個百分點。\n\n有兩件事要一起看。\n\n第一是參數量。LSNet-T 只有 0.3G FLOPs，參數卻有 11.4M，是 StarNet-S1 的四倍；LSNet-B 有 23.2M。如果部署環境卡的是模型大小或記憶體，這張表的排名就要重新看。\n\n第二是測速的方式。論文只量了 RTX 3090 上的大批次吞吐量，沒有手機或 CPU 的延遲。StarNet 的優勢在 iPhone 上批次 1 的延遲，兩篇量的是不同的東西，這張表上的「快將近三倍」不能直接搬到手機上。\n\n### 消融\n\n這組實驗用 LSNet-T 只訓練 100 個 epoch，基準是 71.6%：\n\n| 改動 | top-1 |\n| --- | ---: |\n| LS 卷積（基準） | 71.6% |\n| 拿掉 LS 卷積 | 69.3% |\n| 拿掉大卷積核感知 | 70.5% |\n| 感知的卷積核 7 改成 3 | 70.9% |\n| 感知的卷積核 7 改成 9 | 71.5% |\n| 拿掉小卷積核聚合（大卷積核接一般小卷積） | 70.1% |\n| 聚合的卷積核 3 改成 1 | 69.6% |\n| 聚合的卷積核 3 改成 5 | 71.6% |\n\n*表：摘自原論文 Table 5 與 Table 7。*\n\n有兩列特別值得看。\n\n「拿掉小卷積核聚合」是把大卷積核感知的結果，直接接一個一般的靜態小卷積核，也就是「大核加小核」最直覺的組法，結果只有 70.1%，比 LS 卷積少 1.5。大範圍的資訊要拿來決定權重，比直接疊上去有用。\n\n聚合範圍從 3 加到 5，準確率完全沒變，FLOPs 卻從 0.31G 漲到 0.34G。感知範圍從 7 加到 9 也沒有更好。看 7×7、加 3×3，剛好。\n\n作者也把 LS 卷積換進其他網路。ResNet-50 的 3×3 卷積全部換掉，FLOPs 從 4.1G 降到 2.6G，準確率從 78.8% 升到 80.7%；DeiT-T 的自注意力換掉，FLOPs 從 1.3G 降到 0.9G，準確率從 72.2% 升到 73.0%。\n\n### 部署\n\nLS 卷積的小卷積核聚合不是一般的卷積。每個位置的卷積核都不一樣，PyTorch 內建的 `conv2d` 做不到，官方程式碼用 Triton 寫了專用的 GPU 運算子。GitHub 上有使用者[問為什麼不用內建的卷積](https://github.com/jameslahm/lsnet/issues/8)，也提到轉成 TensorRT 時相容性不好。要部署到手機或其他推論引擎，這一步得自己改寫，例如用 `unfold` 加逐元素相乘，就像上面那段示範程式。\n\n## 結論\n\nLSNet 把 token mixing 拆成感知和聚合兩步，再讓兩步用不同的範圍。7×7 的 depthwise 卷積負責看，決定每個位置的 3×3 權重，聚合只做 9 個值的加權加總。消融表說得很清楚，大範圍的資訊拿去決定權重，比直接把大卷積核和小卷積核疊起來有用。\n\n比起整個 LSNet，LS 卷積這個零件更值得帶走。它只用到 depthwise 卷積和逐位置的加權加總，換進 ResNet-50 之後，FLOPs 少了將近四成，準確率還多了 1.9 個百分點。\n","imageSizes":{"/content/papers/assets/85567a02dcfb2f88ff1be569a4d077a422f781df88b73b63effe301953d2985b.png":{"width":1743,"height":432},"/content/papers/assets/6d3e348c9f8512697f7230b2ac3fde0efdd4fd84198dbb519d931acd26bcbb59.png":{"width":1728,"height":948}}}