{
  "id": 243206,
  "title": "大家都用啥模型呀，来讨论讨论呗",
  "url": "/competitions/classify-leaves/discussion/243206",
  "author_name": "Neko Kiku",
  "post_date": "2021-06-01T15:23:24.197000",
  "votes": 4,
  "comment_count": 15,
  "views": 0,
  "content": "<p>我自己用的res34，调了cos的学习率，加了一大堆也不知道有没有用的tricks，到96的成绩了</p>",
  "messages": [
    {
      "id": 2420137,
      "postDate": "2023-09-02T12:39:18.333Z",
      "content": "<p>细说tricks(不是)</p>",
      "rawMarkdown": "细说tricks(不是)",
      "votes": 1
    },
    {
      "id": 1331612,
      "postDate": "2021-06-01T15:23:24.197Z",
      "content": "<p>我自己用的res34，调了cos的学习率，加了一大堆也不知道有没有用的tricks，到96的成绩了</p>",
      "rawMarkdown": "我自己用的res34，调了cos的学习率，加了一大堆也不知道有没有用的tricks，到96的成绩了",
      "votes": 4
    },
    {
      "id": 2219941,
      "postDate": "2023-04-13T02:39:59.913Z",
      "content": "<p>我也是用了你的baseline，模型用的是densenet161，结果好像在0.9附近</p>",
      "rawMarkdown": "我也是用了你的baseline，模型用的是densenet161，结果好像在0.9附近"
    },
    {
      "id": 1334251,
      "postDate": "2021-06-03T11:07:03.307Z",
      "content": "<p>顺便我用的resnet50</p>",
      "rawMarkdown": "顺便我用的resnet50"
    },
    {
      "id": 1334248,
      "postDate": "2021-06-03T11:05:24.910Z",
      "content": "<p>有没有人试试transformer，我有点好奇，但我电脑不行跑不起来。</p>",
      "rawMarkdown": "有没有人试试transformer，我有点好奇，但我电脑不行跑不起来。",
      "replies": [
        {
          "id": 1334261,
          "postDate": "2021-06-03T11:14:28.397Z",
          "content": "<p>我刚下好了ViT的预训练模型，但是目前显卡被其他任务占满了，今天晚上估计能试试哈哈哈</p>",
          "rawMarkdown": "我刚下好了ViT的预训练模型，但是目前显卡被其他任务占满了，今天晚上估计能试试哈哈哈"
        },
        {
          "id": 1335031,
          "postDate": "2021-06-04T01:12:36.370Z",
          "content": "<p>Vit参数量太大了，过拟合很严重</p>",
          "rawMarkdown": "Vit参数量太大了，过拟合很严重"
        },
        {
          "id": 1335035,
          "postDate": "2021-06-04T01:19:29.447Z",
          "content": "<p>搞了个电脑俺也试一试。</p>",
          "rawMarkdown": "搞了个电脑俺也试一试。"
        }
      ]
    },
    {
      "id": 1332969,
      "postDate": "2021-06-02T12:09:44.700Z",
      "content": "<p>巧了，我也是用的res34</p>",
      "rawMarkdown": "巧了，我也是用的res34"
    },
    {
      "id": 1332158,
      "postDate": "2021-06-02T00:24:45.400Z",
      "content": "<p>我用resnet50跑，要么就用resnet的变形resnext，因为数据集比较小，没必要使用大模型，太费时间了，有条件的可以试一下efficientNet和Incepetion。最求快速的可以使用mobilenetV3,我感觉数据增强的效果挺一般的，不过还是有提升</p>",
      "rawMarkdown": "我用resnet50跑，要么就用resnet的变形resnext，因为数据集比较小，没必要使用大模型，太费时间了，有条件的可以试一下efficientNet和Incepetion。最求快速的可以使用mobilenetV3,我感觉数据增强的效果挺一般的，不过还是有提升"
    },
    {
      "id": 1332152,
      "postDate": "2021-06-02T00:17:47.997Z",
      "content": "<p>用了你的baseline，改成了resnet50，数据增强还不如不加，加了效果更差，然后现在是0.85+</p>",
      "rawMarkdown": "用了你的baseline，改成了resnet50，数据增强还不如不加，加了效果更差，然后现在是0.85+",
      "replies": [
        {
          "id": 1332215,
          "postDate": "2021-06-02T01:37:25.403Z",
          "content": "<p>不至于吧，图像增强还是有点用处，你试试加个学习率衰减，应该能更高，我最开始没加也是只有80+，加了就90+了</p>",
          "rawMarkdown": "不至于吧，图像增强还是有点用处，你试试加个学习率衰减，应该能更高，我最开始没加也是只有80+，加了就90+了"
        },
        {
          "id": 1332302,
          "postDate": "2021-06-02T03:25:22.730Z",
          "content": "<p>有什么图像增强的技巧吗，加了很多，效果不太明显</p>",
          "rawMarkdown": "有什么图像增强的技巧吗，加了很多，效果不太明显"
        },
        {
          "id": 1332375,
          "postDate": "2021-06-02T05:17:47.193Z",
          "content": "<p>改了。现在0.92+，但是加数据增强还是没有太大变化</p>",
          "rawMarkdown": "改了。现在0.92+，但是加数据增强还是没有太大变化"
        },
        {
          "id": 1332688,
          "postDate": "2021-06-02T08:42:06.153Z",
          "content": "<p>也不用太纠结数据增强，可以试下其他方法hhh</p>",
          "rawMarkdown": "也不用太纠结数据增强，可以试下其他方法hhh"
        }
      ]
    },
    {
      "id": 3237628,
      "postDate": "2025-07-01T07:46:23.937Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 2420137,
      "author_name": "Atri2333",
      "author_url": "",
      "post_date": "2023-09-02T12:39:18.333000",
      "content": "<p>细说tricks(不是)</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 2219941,
      "author_name": "我运气超好",
      "author_url": "",
      "post_date": "2023-04-13T02:39:59.913000",
      "content": "<p>我也是用了你的baseline，模型用的是densenet161，结果好像在0.9附近</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1334251,
      "author_name": "Tianfu Wang (王天夫)",
      "author_url": "",
      "post_date": "2021-06-03T11:07:03.307000",
      "content": "<p>顺便我用的resnet50</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1334248,
      "author_name": "Tianfu Wang (王天夫)",
      "author_url": "",
      "post_date": "2021-06-03T11:05:24.910000",
      "content": "<p>有没有人试试transformer，我有点好奇，但我电脑不行跑不起来。</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1334261,
          "author_name": "Neko Kiku",
          "author_url": "",
          "post_date": "2021-06-03T11:14:28.397000",
          "content": "<p>我刚下好了ViT的预训练模型，但是目前显卡被其他任务占满了，今天晚上估计能试试哈哈哈</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1335031,
          "author_name": "Neko Kiku",
          "author_url": "",
          "post_date": "2021-06-04T01:12:36.370000",
          "content": "<p>Vit参数量太大了，过拟合很严重</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1335035,
          "author_name": "Tianfu Wang (王天夫)",
          "author_url": "",
          "post_date": "2021-06-04T01:19:29.447000",
          "content": "<p>搞了个电脑俺也试一试。</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1332969,
      "author_name": "miugod",
      "author_url": "",
      "post_date": "2021-06-02T12:09:44.700000",
      "content": "<p>巧了，我也是用的res34</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1332158,
      "author_name": "迪迦奥特曼本曼",
      "author_url": "",
      "post_date": "2021-06-02T00:24:45.400000",
      "content": "<p>我用resnet50跑，要么就用resnet的变形resnext，因为数据集比较小，没必要使用大模型，太费时间了，有条件的可以试一下efficientNet和Incepetion。最求快速的可以使用mobilenetV3,我感觉数据增强的效果挺一般的，不过还是有提升</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1332152,
      "author_name": "Wangrongsheng",
      "author_url": "",
      "post_date": "2021-06-02T00:17:47.997000",
      "content": "<p>用了你的baseline，改成了resnet50，数据增强还不如不加，加了效果更差，然后现在是0.85+</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1332215,
          "author_name": "Neko Kiku",
          "author_url": "",
          "post_date": "2021-06-02T01:37:25.403000",
          "content": "<p>不至于吧，图像增强还是有点用处，你试试加个学习率衰减，应该能更高，我最开始没加也是只有80+，加了就90+了</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1332302,
          "author_name": "Wangrongsheng",
          "author_url": "",
          "post_date": "2021-06-02T03:25:22.730000",
          "content": "<p>有什么图像增强的技巧吗，加了很多，效果不太明显</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1332375,
          "author_name": "Wangrongsheng",
          "author_url": "",
          "post_date": "2021-06-02T05:17:47.193000",
          "content": "<p>改了。现在0.92+，但是加数据增强还是没有太大变化</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1332688,
          "author_name": "Neko Kiku",
          "author_url": "",
          "post_date": "2021-06-02T08:42:06.153000",
          "content": "<p>也不用太纠结数据增强，可以试下其他方法hhh</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3237628,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-07-01T07:46:23.937000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2420137": "细说tricks(不是)",
    "1331612": "我自己用的res34，调了cos的学习率，加了一大堆也不知道有没有用的tricks，到96的成绩了",
    "2219941": "我也是用了你的baseline，模型用的是densenet161，结果好像在0.9附近",
    "1334251": "顺便我用的resnet50",
    "1334248": "有没有人试试transformer，我有点好奇，但我电脑不行跑不起来。",
    "1332969": "巧了，我也是用的res34",
    "1332158": "我用resnet50跑，要么就用resnet的变形resnext，因为数据集比较小，没必要使用大模型，太费时间了，有条件的可以试一下efficientNet和Incepetion。最求快速的可以使用mobilenetV3,我感觉数据增强的效果挺一般的，不过还是有提升",
    "1332152": "用了你的baseline，改成了resnet50，数据增强还不如不加，加了效果更差，然后现在是0.85+",
    "3237628": ""
  }
}