{
  "source_file": "0930",
  "source_platform": "研究机构",
  "author": "Dive into Deep Learning",
  "title_listed": "Modern Convolutional Neural Networks（章节）",
  "keywords": "ResNet,GoogleNet,DenseNet",
  "domain": "图像处理",
  "content_type": "教程/书籍",
  "is_paid_flag": "否",
  "url": "https://www.d2l.ai/chapter_convolutional-modern/index.html",
  "publish_date": "持续更新",
  "corpus_stage": "图像处理",
  "notes": "现代 CNN 架构",
  "id": "84742f1e9591",
  "collected_at": "2026-10-02T16:07:38",
  "title": "8. Modern Convolutional Neural Networks — Dive into Deep Learning 1.0.3 documentation",
  "body_markdown": "Table Of Contents\n8.\nModern Convolutional Neural Networks\n¶\nNow that we understand the basics of wiring together CNNs, let’s take a\ntour of modern CNN architectures. This tour is, by necessity,\nincomplete, thanks to the plethora of exciting new designs being added.\nTheir importance derives from the fact that not only can they be used\ndirectly for vision tasks, but they also serve as basic feature\ngenerators for more advanced tasks such as tracking\n(\nZhang\net al.\n, 2021\n)\n, segmentation\n(\nLong\net al.\n, 2015\n)\n, object detection\n(\nRedmon and Farhadi, 2018\n)\n, or style transformation\n(\nGatys\net al.\n, 2016\n)\n. In this chapter, most sections\ncorrespond to a significant CNN architecture that was at some point (or\ncurrently) the base model upon which many research projects and deployed\nsystems were built. Each of these networks was briefly a dominant\narchitecture and many were winners or runners-up in the\nImageNet\ncompetition\nwhich has\nserved as a barometer of progress on supervised learning in computer\nvision since 2010. It is only recently that Transformers have begun to\ndisplace CNNs, starting with\nDosovitskiy\net al.\n(\n2021\n)\nand followed by the Swin\nTransformer\n(\nLiu\net al.\n, 2021\n)\n. We will cover this development later\nin\nSection 11\n.\nWhile the idea of\ndeep\nneural networks is quite simple (stack together\na bunch of layers), performance can vary wildly across architectures and\nhyperparameter choices. The neural networks described in this chapter\nare the product of intuition, a few mathematical insights, and a lot of\ntrial and error. We present these models in chronological order, partly\nto convey a sense of the history so that you can form your own\nintuitions about where the field is heading and perhaps develop your own\narchitectures. For instance, batch normalization and residual\nconnections described in this chapter have offered two popular ideas for\ntraining and designing deep models, both of which have since also been\napplied to architectures beyond computer vision.\nWe begin our tour of modern CNNs with AlexNet\n(\nKrizhevsky\net al.\n, 2012\n)\n, the first large-scale\nnetwork deployed to beat conventional computer vision methods on a\nlarge-scale vision challenge; the VGG network\n(\nSimonyan and Zisserman, 2014\n)\n, which makes use of a number of\nrepeating blocks of elements; the network in network (NiN) that\nconvolves whole neural networks patch-wise over inputs\n(\nLin\net al.\n, 2013\n)\n; GoogLeNet that uses networks with\nmulti-branch convolutions\n(\nSzegedy\net al.\n, 2015\n)\n; the\nresidual network (ResNet)\n(\nHe\net al.\n, 2016\n)\n, which remains\none of the most popular off-the-shelf architectures in computer vision;\nResNeXt blocks\n(\nXie\net al.\n, 2017\n)\nfor sparser\nconnections; and DenseNet\n(\nHuang\net al.\n, 2017\n)\nfor\na generalization of the residual architecture. Over time many special\noptimizations for efficient networks have been developed, such as\ncoordinate shifts (ShiftNet)\n(\nWu\net al.\n, 2018\n)\n. This culminated in\nthe automatic search for efficient architectures such as MobileNet v3\n(\nHoward\net al.\n, 2019\n)\n. It also includes the\nsemi-automatic design exploration of\nRadosavovic\net al.\n(\n2020\n)\nthat led to the\nRegNetX/Y which we will discuss later in this chapter. The work is\ninstructive insofar as it offers a path for marrying brute force\ncomputation with the ingenuity of an experimenter in the search for\nefficient design spaces. Of note is also the work of\nLiu\net al.\n(\n2022\n)\nas it shows that training techniques (e.g.,\noptimizers, data augmentation, and regularization) play a pivotal role\nin improving accuracy. It also shows that long-held assumptions, such as\nthe size of a convolution window, may need to be revisited, given the\nincrease in computation and data. We will cover this and many more\nquestions in due course throughout this chapter.\n8.1. Deep Convolutional Neural Networks (AlexNet)\n8.1.1. Representation Learning\n8.1.2. AlexNet\n8.1.3. Training\n8.1.4. Discussion\n8.1.5. Exercises\n8.2. Networks Using Blocks (VGG)\n8.2.1. VGG Blocks\n8.2.2. VGG Network\n8.2.3. Training\n8.2.4. Summary\n8.2.5. Exercises\n8.3. Network in Network (NiN)\n8.3.1. NiN Blocks\n8.3.2. NiN Model\n8.3.3. Training\n8.3.4. Summary\n8.3.5. Exercises\n8.4. Multi-Branch Networks (GoogLeNet)\n8.4.1. Inception Blocks\n8.4.2. GoogLeNet Model\n8.4.3. Training\n8.4.4. Discussion\n8.4.5. Exercises\n8.5. Batch Normalization\n8.5.1. Training Deep Networks\n8.5.2. Batch Normalization Layers\n8.5.3. Implementation from Scratch\n8.5.4. LeNet with Batch Normalization\n8.5.5. Concise Implementation\n8.5.6. Discussion\n8.5.7. Exercises\n8.6. Residual Networks (ResNet) and ResNeXt\n8.6.1. Function Classes\n8.6.2. Residual Blocks\n8.6.3. ResNet Model\n8.6.4. Training\n8.6.5. ResNeXt\n8.6.6. Summary and Discussion\n8.6.7. Exercises\n8.7. Densely Connected Networks (DenseNet)\n8.7.1. From ResNet to DenseNet\n8.7.2. Dense Blocks\n8.7.3. Transition Layers\n8.7.4. DenseNet Model\n8.7.5. Training\n8.7.6. Summary and Discussion\n8.7.7. Exercises\n8.8. Designing Convolution Network Architectures\n8.8.1. The AnyNet Design Space\n8.8.2. Distributions and Parameters of Design Spaces\n8.8.3. RegNet\n8.8.4. Training\n8.8.5. Discussion\n8.8.6. Exercises\nPrevious\n7.6. Convolutional Neural Networks (LeNet)\nNext\n8.1. Deep Convolutional Neural Networks (AlexNet)",
  "metadata": {
    "title": "8. Modern Convolutional Neural Networks — Dive into Deep Learning 1.0.3 documentation",
    "description": null,
    "keywords": null,
    "author": null
  },
  "paywall": {
    "listed_paid": false,
    "detected": false,
    "marker": "",
    "public_only": false,
    "policy": "付费/会员内容只保留页面公开可见部分，不绕过付费墙"
  },
  "crawl": {
    "success": true,
    "status_code": 200,
    "error": ""
  },
  "code_blocks": [],
  "images": [
    {
      "status": "ok",
      "url": "https://www.d2l.ai/_static/logo-with-text.png",
      "bytes": 87175,
      "reused": false,
      "local_path": "attachments/84742f1e9591/images/body-509f124ba61493e64845.png",
      "src": "https://www.d2l.ai/_static/logo-with-text.png",
      "alt": "Dive into Deep Learning",
      "downloaded": true
    }
  ],
  "audio": [],
  "videos": [],
  "links": [
    {
      "href": "https://www.d2l.ai/index.html",
      "label": "",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_references/zreferences.html",
      "label": "Zhang et al. , 2021",
      "type": "a"
    },
    {
      "href": "https://www.image-net.org/challenges/LSVRC/",
      "label": "ImageNet\ncompetition",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_attention-mechanisms-and-transformers/index.html",
      "label": "Section 11",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_convolutional-modern/alexnet.html",
      "label": "8.1. Deep Convolutional Neural Networks (AlexNet)",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_convolutional-modern/vgg.html",
      "label": "8.2. Networks Using Blocks (VGG)",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_convolutional-modern/nin.html",
      "label": "8.3. Network in Network (NiN)",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_convolutional-modern/googlenet.html",
      "label": "8.4. Multi-Branch Networks (GoogLeNet)",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_convolutional-modern/batch-norm.html",
      "label": "8.5. Batch Normalization",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_convolutional-modern/resnet.html",
      "label": "8.6. Residual Networks (ResNet) and ResNeXt",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_convolutional-modern/densenet.html",
      "label": "8.7. Densely Connected Networks (DenseNet)",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_convolutional-modern/cnn-design.html",
      "label": "8.8. Designing Convolution Network Architectures",
      "type": "a"
    },
    {
      "href": "https://www.d2l.ai/chapter_convolutional-neural-networks/lenet.html",
      "label": "Previous 7.6. Convolutional Neural Networks (LeNet)",
      "type": "a"
    }
  ],
  "repair_version": 2,
  "repaired_at": "2026-10-14T10:53:45Z",
  "source_url": "https://www.d2l.ai/chapter_convolutional-modern/index.html",
  "papers": [],
  "actual_video_files": [],
  "body_chars": 5246,
  "expected_body_images": 1,
  "discovery_status": "ok",
  "delivery_status": "ok",
  "missing_reason": "",
  "raw_html": "attachments/84742f1e9591/pages/source.html",
  "local_html": "attachments/84742f1e9591/pages/article.html",
  "body_base_path": "attachments/84742f1e9591/pages/",
  "body_markdown_path": "attachments/84742f1e9591/pages/article.md"
}