{
  "id": 244714,
  "title": "第五名思路，代码已公布。",
  "url": "/competitions/classify-leaves/discussion/244714",
  "author_name": "",
  "post_date": "2021-06-08T01:34:09.133016500Z",
  "votes": 2,
  "comment_count": 13,
  "views": 0,
  "content": "<ol>\n<li>数据增强：<br>\nresize 320, HorizontalFlip, VerticalFlip, Rotate, RandomBrightnesContrasr, ShiftScaleRotate, Normalize</li>\n<li>模型：<br>\nseresnext50和resnet50</li>\n<li>优化器：<br>\nAdamW CosineAnnealingLR</li>\n<li>其他：<br>\n5折交叉验证， 最终结果为五折准确率最高平均， 两个网络各自平均后再做平均。<br>\nloss就是CrossEntropy。</li>\n</ol>\n<p>总体来说太朴素了…….</p>\n<p>代码在 <a href=\"https://www.kaggle.com/wangtf96/0-98909-5th-place-solution\" target=\"_blank\">https://www.kaggle.com/wangtf96/0-98909-5th-place-solution</a></p>",
  "messages": [
    {
      "id": "1340465",
      "postDate": "06/08/2021 01:34:09",
      "content": "<ol>\n<li>数据增强：<br>\nresize 320, HorizontalFlip, VerticalFlip, Rotate, RandomBrightnesContrasr, ShiftScaleRotate, Normalize</li>\n<li>模型：<br>\nseresnext50和resnet50</li>\n<li>优化器：<br>\nAdamW CosineAnnealingLR</li>\n<li>其他：<br>\n5折交叉验证， 最终结果为五折准确率最高平均， 两个网络各自平均后再做平均。<br>\nloss就是CrossEntropy。</li>\n</ol>\n<p>总体来说太朴素了…….</p>\n<p>代码在 <a href=\"https://www.kaggle.com/wangtf96/0-98909-5th-place-solution\" target=\"_blank\">https://www.kaggle.com/wangtf96/0-98909-5th-place-solution</a></p>",
      "rawMarkdown": "1. 数据增强：\nresize 320, HorizontalFlip, VerticalFlip, Rotate, RandomBrightnesContrasr, ShiftScaleRotate, Normalize\n2. 模型：\nseresnext50和resnet50\n3. 优化器：\nAdamW CosineAnnealingLR\n4. 其他：\n5折交叉验证， 最终结果为五折准确率最高平均， 两个网络各自平均后再做平均。\nloss就是CrossEntropy。\n\n总体来说太朴素了.......\n\n代码在 https://www.kaggle.com/wangtf96/0-98909-5th-place-solution",
      "votes": null
    },
    {
      "id": "1340646",
      "postDate": "06/08/2021 07:02:42",
      "content": "<p>同resnet50，adam 就往死里练</p>",
      "rawMarkdown": "同resnet50，adam 就往死里练",
      "votes": null
    },
    {
      "id": "1340651",
      "postDate": "06/08/2021 07:05:30",
      "content": "<p>ResNet50, SGD 精度调参</p>",
      "rawMarkdown": "ResNet50, SGD 精度调参",
      "votes": null
    },
    {
      "id": "1341119",
      "postDate": "06/08/2021 12:46:39",
      "content": "<p>我也试了mixup，确实没用，我也是res50，Adam，咋就调不上去呢，死活上不了98。。。</p>",
      "rawMarkdown": "我也试了mixup，确实没用，我也是res50，Adam，咋就调不上去呢，死活上不了98。。。",
      "votes": null
    },
    {
      "id": "1341124",
      "postDate": "06/08/2021 12:49:37",
      "content": "<p>对了，efficientnet b2 b3效果也很好，能到97</p>",
      "rawMarkdown": "对了，efficientnet b2 b3效果也很好，能到97",
      "votes": null
    },
    {
      "id": "1341171",
      "postDate": "06/08/2021 13:30:39",
      "content": "<p>SGD我试了试，学习率太难调了，就懒得调了。</p>",
      "rawMarkdown": "SGD我试了试，学习率太难调了，就懒得调了。",
      "votes": null
    },
    {
      "id": "1341173",
      "postDate": "06/08/2021 13:33:48",
      "content": "<p>我也试了，蛮好的，但是还是resnet更好，也不知道为什么。</p>",
      "rawMarkdown": "我也试了，蛮好的，但是还是resnet更好，也不知道为什么。",
      "votes": null
    },
    {
      "id": "1341179",
      "postDate": "06/08/2021 13:35:27",
      "content": "<p>我用的是resnet50的改进版本，不完全是resnet50.</p>",
      "rawMarkdown": "我用的是resnet50的改进版本，不完全是resnet50.",
      "votes": null
    },
    {
      "id": "1341804",
      "postDate": "06/09/2021 02:56:08",
      "content": "<p>你res50,能到多少</p>",
      "rawMarkdown": "你res50,能到多少",
      "votes": null
    },
    {
      "id": "1341807",
      "postDate": "06/09/2021 02:59:55",
      "content": "<p>哈哈哈，你没发现楼主是现在的第一吗</p>",
      "rawMarkdown": "哈哈哈，你没发现楼主是现在的第一吗",
      "votes": null
    },
    {
      "id": "1342038",
      "postDate": "06/09/2021 06:42:35",
      "content": "<p>普通的resnet50，最高97以上，接近98.我用的是resnet50的改进版。</p>",
      "rawMarkdown": "普通的resnet50，最高97以上，接近98.我用的是resnet50的改进版。",
      "votes": null
    },
    {
      "id": "1343357",
      "postDate": "06/10/2021 07:06:50",
      "content": "<p>Almost the same setting but with randaugment with mixup and cutmix：<br>\nCV 0.980 LB 0.986 (with TTA)</p>",
      "rawMarkdown": "Almost the same setting but with randaugment with mixup and cutmix：\nCV 0.980 LB 0.986 (with TTA)",
      "votes": null
    },
    {
      "id": "1343378",
      "postDate": "06/10/2021 07:17:47",
      "content": "<p>这个数据每张图都有好大一片空白，为什么cutmix这样的增强会有用呢？</p>",
      "rawMarkdown": "这个数据每张图都有好大一片空白，为什么cutmix这样的增强会有用呢？",
      "votes": null
    },
    {
      "id": "1343392",
      "postDate": "06/10/2021 07:23:03",
      "content": "<p>我也试了mixup，发现没什么用，所以应该是cutmix起作用了</p>",
      "rawMarkdown": "我也试了mixup，发现没什么用，所以应该是cutmix起作用了",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1340646,
      "author_name": "litianfu1997",
      "author_url": "",
      "post_date": "06/08/2021 07:02:42",
      "content": "<p>同resnet50，adam 就往死里练</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1340651,
      "author_name": "yyosong",
      "author_url": "",
      "post_date": "06/08/2021 07:05:30",
      "content": "<p>ResNet50, SGD 精度调参</p>",
      "votes": null,
      "replies": [
        {
          "id": 1341171,
          "author_name": "wangtf96",
          "author_url": "",
          "post_date": "06/08/2021 13:30:39",
          "content": "<p>SGD我试了试，学习率太难调了，就懒得调了。</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1341119,
      "author_name": "nekokiku",
      "author_url": "",
      "post_date": "06/08/2021 12:46:39",
      "content": "<p>我也试了mixup，确实没用，我也是res50，Adam，咋就调不上去呢，死活上不了98。。。</p>",
      "votes": null,
      "replies": [
        {
          "id": 1341179,
          "author_name": "wangtf96",
          "author_url": "",
          "post_date": "06/08/2021 13:35:27",
          "content": "<p>我用的是resnet50的改进版本，不完全是resnet50.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1341124,
      "author_name": "nekokiku",
      "author_url": "",
      "post_date": "06/08/2021 12:49:37",
      "content": "<p>对了，efficientnet b2 b3效果也很好，能到97</p>",
      "votes": null,
      "replies": [
        {
          "id": 1341173,
          "author_name": "wangtf96",
          "author_url": "",
          "post_date": "06/08/2021 13:33:48",
          "content": "<p>我也试了，蛮好的，但是还是resnet更好，也不知道为什么。</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1341804,
      "author_name": "binzhongzhu1",
      "author_url": "",
      "post_date": "06/09/2021 02:56:08",
      "content": "<p>你res50,能到多少</p>",
      "votes": null,
      "replies": [
        {
          "id": 1341807,
          "author_name": "nekokiku",
          "author_url": "",
          "post_date": "06/09/2021 02:59:55",
          "content": "<p>哈哈哈，你没发现楼主是现在的第一吗</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1342038,
          "author_name": "wangtf96",
          "author_url": "",
          "post_date": "06/09/2021 06:42:35",
          "content": "<p>普通的resnet50，最高97以上，接近98.我用的是resnet50的改进版。</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1343357,
      "author_name": "seefun",
      "author_url": "",
      "post_date": "06/10/2021 07:06:50",
      "content": "<p>Almost the same setting but with randaugment with mixup and cutmix：<br>\nCV 0.980 LB 0.986 (with TTA)</p>",
      "votes": null,
      "replies": [
        {
          "id": 1343378,
          "author_name": "wangtf96",
          "author_url": "",
          "post_date": "06/10/2021 07:17:47",
          "content": "<p>这个数据每张图都有好大一片空白，为什么cutmix这样的增强会有用呢？</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1343392,
          "author_name": "nekokiku",
          "author_url": "",
          "post_date": "06/10/2021 07:23:03",
          "content": "<p>我也试了mixup，发现没什么用，所以应该是cutmix起作用了</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1340465": "1. 数据增强：\nresize 320, HorizontalFlip, VerticalFlip, Rotate, RandomBrightnesContrasr, ShiftScaleRotate, Normalize\n2. 模型：\nseresnext50和resnet50\n3. 优化器：\nAdamW CosineAnnealingLR\n4. 其他：\n5折交叉验证， 最终结果为五折准确率最高平均， 两个网络各自平均后再做平均。\nloss就是CrossEntropy。\n\n总体来说太朴素了.......\n\n代码在 https://www.kaggle.com/wangtf96/0-98909-5th-place-solution",
    "1340646": "同resnet50，adam 就往死里练",
    "1340651": "ResNet50, SGD 精度调参",
    "1341119": "我也试了mixup，确实没用，我也是res50，Adam，咋就调不上去呢，死活上不了98。。。",
    "1341124": "对了，efficientnet b2 b3效果也很好，能到97",
    "1341171": "SGD我试了试，学习率太难调了，就懒得调了。",
    "1341173": "我也试了，蛮好的，但是还是resnet更好，也不知道为什么。",
    "1341179": "我用的是resnet50的改进版本，不完全是resnet50.",
    "1341804": "你res50,能到多少",
    "1341807": "哈哈哈，你没发现楼主是现在的第一吗",
    "1342038": "普通的resnet50，最高97以上，接近98.我用的是resnet50的改进版。",
    "1343357": "Almost the same setting but with randaugment with mixup and cutmix：\nCV 0.980 LB 0.986 (with TTA)",
    "1343378": "这个数据每张图都有好大一片空白，为什么cutmix这样的增强会有用呢？",
    "1343392": "我也试了mixup，发现没什么用，所以应该是cutmix起作用了"
  },
  "source": "meta"
}