{
  "id": 213827,
  "title": "How to get 0.907？",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/213827",
  "author_name": "DarknessZX",
  "post_date": "2021-01-24T12:58:09.179000",
  "votes": 6,
  "comment_count": 32,
  "views": 0,
  "content": "<p>Now my score is 0.904，the normal thricks i had used in my notebook，just like tta，label smooth，cutout，mixup ，model essemble and so on.But i do not have a good idea now.<br>\nI think there must be some problem in noise label，but how to deal it？<br>\nI find two loss function in my last discussion，but there are some problem when i tyr to use them.<br>\nAnybody have some better ways to improve a higher score？</p>",
  "messages": [
    {
      "id": 1167724,
      "postDate": "2021-01-24T12:58:09.180Z",
      "content": "<p>Now my score is 0.904，the normal thricks i had used in my notebook，just like tta，label smooth，cutout，mixup ，model essemble and so on.But i do not have a good idea now.<br>\nI think there must be some problem in noise label，but how to deal it？<br>\nI find two loss function in my last discussion，but there are some problem when i tyr to use them.<br>\nAnybody have some better ways to improve a higher score？</p>",
      "rawMarkdown": "Now my score is 0.904，the normal thricks i had used in my notebook，just like tta，label smooth，cutout，mixup ，model essemble and so on.But i do not have a good idea now.\nI think there must be some problem in noise label，but how to deal it？\nI find two loss function in my last discussion，but there are some problem when i tyr to use them.\nAnybody have some better ways to improve a higher score？",
      "votes": 6
    },
    {
      "id": 1167814,
      "postDate": "2021-01-24T13:55:40.537Z",
      "content": "<p>Is your ultimate goal solely overfitting the public LB  ?</p>",
      "rawMarkdown": "Is your ultimate goal solely overfitting the public LB  ?",
      "votes": 1,
      "replies": [
        {
          "id": 1167871,
          "postDate": "2021-01-24T14:38:31.737Z",
          "content": "<p>of course not，i have not a robust way to improve my score now，so i think whether i can do something to data，just like clean the noisy data.Now i realize it is not a good approach.Thank you for reminding me.I need a robust way to improve my score，do you have some ideas about it？Thank you very much!!!!!</p>",
          "rawMarkdown": "of course not，i have not a robust way to improve my score now，so i think whether i can do something to data，just like clean the noisy data.Now i realize it is not a good approach.Thank you for reminding me.I need a robust way to improve my score，do you have some ideas about it？Thank you very much!!!!!",
          "votes": 2
        }
      ]
    },
    {
      "id": 1168473,
      "postDate": "2021-01-25T01:01:54.880Z",
      "content": "<p>Borden your mind and use your inner eye to overfit PB!</p>",
      "rawMarkdown": "Borden your mind and use your inner eye to overfit PB!",
      "votes": -1,
      "replies": [
        {
          "id": 1168642,
          "postDate": "2021-01-25T04:49:47.787Z",
          "content": "<p>aha，thank you very much</p>",
          "rawMarkdown": "aha，thank you very much"
        },
        {
          "id": 1168644,
          "postDate": "2021-01-25T04:50:25.127Z",
          "content": "<p>i can‘t agree more</p>",
          "rawMarkdown": "i can‘t agree more"
        },
        {
          "id": 1168771,
          "postDate": "2021-01-25T06:49:53.187Z",
          "content": "<p>In my experience, This kind of competition often end with pretty big shake up. And at the time when the bug of private LB exposed to the public, I only have one submission, scored 0.892 in PB, rank around 600+, but scored 0.897 in private LB and rank 170+. <br>\nSo It is better to trust your CV.</p>",
          "rawMarkdown": "In my experience, This kind of competition often end with pretty big shake up. And at the time when the bug of private LB exposed to the public, I only have one submission, scored 0.892 in PB, rank around 600+, but scored 0.897 in private LB and rank 170+. \nSo It is better to trust your CV.",
          "votes": 5
        },
        {
          "id": 1168787,
          "postDate": "2021-01-25T07:12:40.727Z",
          "content": "<p>Oh，This is I first time attends this kind of competition，thank you remind me！ I'll take note of the problem</p>",
          "rawMarkdown": "Oh，This is I first time attends this kind of competition，thank you remind me！ I'll take note of the problem"
        },
        {
          "id": 1168907,
          "postDate": "2021-01-25T08:41:58.723Z",
          "content": "<p><a href=\"https://www.kaggle.com/steamedsheep\" target=\"_blank\">@steamedsheep</a> Hi sheep😄, I wonder that such a big change is beacuse you have cleaned the data?</p>",
          "rawMarkdown": "@steamedsheep Hi sheep😄, I wonder that such a big change is beacuse you have cleaned the data?"
        },
        {
          "id": 1168992,
          "postDate": "2021-01-25T09:24:38.263Z",
          "content": "<p>That's my first submission, and haven't clean the data</p>",
          "rawMarkdown": "That's my first submission, and haven't clean the data",
          "votes": 1
        },
        {
          "id": 1169192,
          "postDate": "2021-01-25T11:28:32.607Z",
          "content": "<p>Thanks for rely:D  But in addition to being clean or noisy ,distribution differences, what can be the difference between LB and PB to make such a huge shake😔 . Or you balanced the class with few samples? that also can be the reason of the huge shake up</p>",
          "rawMarkdown": "Thanks for rely:D  But in addition to being clean or noisy ,distribution differences, what can be the difference between LB and PB to make such a huge shake😔 . Or you balanced the class with few samples? that also can be the reason of the huge shake up"
        },
        {
          "id": 1170653,
          "postDate": "2021-01-26T10:59:50.283Z",
          "content": "<p><a href=\"https://www.kaggle.com/steamedsheep\" target=\"_blank\">@steamedsheep</a> the difference is huge for your first submission.</p>",
          "rawMarkdown": "@steamedsheep the difference is huge for your first submission."
        }
      ]
    },
    {
      "id": 1170378,
      "postDate": "2021-01-26T07:13:45.500Z",
      "content": "<p>Hi,do you want to team up?Now, I can get 0.903 in single model.Please email me <a>bcwang1993@qq.com</a></p>",
      "rawMarkdown": "Hi,do you want to team up?Now, I can get 0.903 in single model.Please email me bcwang1993@qq.com",
      "replies": [
        {
          "id": 1170800,
          "postDate": "2021-01-26T12:55:33.643Z",
          "content": "<p>Hello,do u need teamates ? </p>",
          "rawMarkdown": "Hello,do u need teamates ? "
        },
        {
          "id": 1170829,
          "postDate": "2021-01-26T13:20:10.993Z",
          "content": "<p>sorry，I already have a team appointment with someone else. Can I add a QQ friend with you? We can get together next time😭</p>",
          "rawMarkdown": "sorry，I already have a team appointment with someone else. Can I add a QQ friend with you? We can get together next time😭"
        },
        {
          "id": 1170851,
          "postDate": "2021-01-26T13:46:06.123Z",
          "content": "<p><a href=\"https://www.kaggle.com/zekunn\" target=\"_blank\">@zekunn</a> yes,I am looking up a teamates,are you want to team up, email me <a>bcwang1993@qq.com</a></p>",
          "rawMarkdown": "@zekunn yes,I am looking up a teamates,are you want to team up, email me bcwang1993@qq.com"
        },
        {
          "id": 1170852,
          "postDate": "2021-01-26T13:46:58.240Z",
          "content": "<p><a href=\"https://www.kaggle.com/darknesszx\" target=\"_blank\">@darknesszx</a> ,you can send a email to me ,I will tell you my wechat.</p>",
          "rawMarkdown": "@darknesszx ,you can send a email to me ,I will tell you my wechat."
        }
      ]
    },
    {
      "id": 1169680,
      "postDate": "2021-01-25T17:06:38.860Z",
      "content": "<p>You are already in a high rank. :)</p>\n<p><a href=\"https://www.kaggle.com/darknesszx\" target=\"_blank\">@darknesszx</a> </p>",
      "rawMarkdown": "You are already in a high rank. :)\n\n@darknesszx ",
      "replies": [
        {
          "id": 1169763,
          "postDate": "2021-01-25T18:15:21.837Z",
          "content": "<p>Haha, there is still a big gap with you.I've seen some of your work on noise label in Notebook and it's great!!I'll try to catch up with you，Ha ha!!</p>",
          "rawMarkdown": "Haha, there is still a big gap with you.I've seen some of your work on noise label in Notebook and it's great!!I'll try to catch up with you，Ha ha!!"
        }
      ]
    },
    {
      "id": 1168326,
      "postDate": "2021-01-24T22:10:35.253Z",
      "content": "<p>What is your single model score?</p>",
      "rawMarkdown": "What is your single model score?",
      "replies": [
        {
          "id": 1168638,
          "postDate": "2021-01-25T04:49:10.333Z",
          "content": "<p>efficientnetb4，0.901</p>",
          "rawMarkdown": "efficientnetb4，0.901",
          "votes": 1
        },
        {
          "id": 1170125,
          "postDate": "2021-01-26T02:13:00.397Z",
          "content": "<p>Did you use TTA for inference?</p>",
          "rawMarkdown": "Did you use TTA for inference?"
        },
        {
          "id": 1172778,
          "postDate": "2021-01-27T14:47:13.127Z",
          "content": "<p>Wow, only TTA make you an increase of 0.03 ? My single model without TTA is doing 0.903 but haven't seen any improvement with TTA 😲</p>",
          "rawMarkdown": "Wow, only TTA make you an increase of 0.03 ? My single model without TTA is doing 0.903 but haven't seen any improvement with TTA 😲"
        },
        {
          "id": 1174835,
          "postDate": "2021-01-28T18:37:17.863Z",
          "content": "<p>which model are you using, none of my models are reaching above lb 8.8</p>",
          "rawMarkdown": "which model are you using, none of my models are reaching above lb 8.8"
        }
      ]
    },
    {
      "id": 1167794,
      "postDate": "2021-01-24T13:35:58.943Z",
      "content": "<p>Trust your CV and have robust approach on dealing label noise</p>",
      "rawMarkdown": "Trust your CV and have robust approach on dealing label noise",
      "replies": [
        {
          "id": 1167802,
          "postDate": "2021-01-24T13:49:42.897Z",
          "content": "<p>Is manual cleaning of noise data necessary？</p>",
          "rawMarkdown": "Is manual cleaning of noise data necessary？"
        },
        {
          "id": 1167803,
          "postDate": "2021-01-24T13:50:56.823Z",
          "content": "<p>Thank you very much！i will try it！</p>",
          "rawMarkdown": "Thank you very much！i will try it！"
        },
        {
          "id": 1167842,
          "postDate": "2021-01-24T14:06:52.593Z",
          "content": "<p>No, I recommend to not perform any manual label clean-up. Private lb also consist of noisy labels if you test your submission/s using different configuration like image size, augmentation etc, it generates different submission score. Otherwise, we can achieve a very high score in classification with a minimum of 0.95+, like other previous classification competition. </p>",
          "rawMarkdown": "No, I recommend to not perform any manual label clean-up. Private lb also consist of noisy labels if you test your submission/s using different configuration like image size, augmentation etc, it generates different submission score. Otherwise, we can achieve a very high score in classification with a minimum of 0.95+, like other previous classification competition. "
        },
        {
          "id": 1167875,
          "postDate": "2021-01-24T14:40:00.117Z",
          "content": "<p>Thank you very much,i will try to find some robust ways to improve my score!!</p>",
          "rawMarkdown": "Thank you very much,i will try to find some robust ways to improve my score!!"
        }
      ]
    },
    {
      "id": 1169329,
      "postDate": "2021-01-25T12:45:17.577Z",
      "rawMarkdown": "",
      "votes": 2,
      "isDeleted": true,
      "replies": [
        {
          "id": 1169546,
          "postDate": "2021-01-25T14:52:36.080Z",
          "content": "<p>谢谢提醒！！！哈哈哈，我尝试一下看看！👍</p>",
          "rawMarkdown": "谢谢提醒！！！哈哈哈，我尝试一下看看！👍"
        },
        {
          "id": 1170127,
          "postDate": "2021-01-26T02:16:28.910Z",
          "content": "<p><a href=\"https://www.kaggle.com/jianxinhu\" target=\"_blank\">@jianxinhu</a> 那个说的是PANDA那个比赛</p>",
          "rawMarkdown": "@jianxinhu 那个说的是PANDA那个比赛"
        },
        {
          "id": 1170135,
          "postDate": "2021-01-26T02:29:19.043Z",
          "content": "<p>The host keeps silent all the time. Nobody knows the secret of test images. But the class distribution of first 31% test images is similar to train images.</p>",
          "rawMarkdown": "The host keeps silent all the time. Nobody knows the secret of test images. But the class distribution of first 31% test images is similar to train images.",
          "votes": 4
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1167814,
      "author_name": "Serigne ",
      "author_url": "",
      "post_date": "2021-01-24T13:55:40.537000",
      "content": "<p>Is your ultimate goal solely overfitting the public LB  ?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1167871,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-24T14:38:31.737000",
          "content": "<p>of course not，i have not a robust way to improve my score now，so i think whether i can do something to data，just like clean the noisy data.Now i realize it is not a good approach.Thank you for reminding me.I need a robust way to improve my score，do you have some ideas about it？Thank you very much!!!!!</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 1168473,
      "author_name": "sheep",
      "author_url": "",
      "post_date": "2021-01-25T01:01:54.880000",
      "content": "<p>Borden your mind and use your inner eye to overfit PB!</p>",
      "votes": -1,
      "replies": [
        {
          "id": 1168642,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-25T04:49:47.787000",
          "content": "<p>aha，thank you very much</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1168644,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-25T04:50:25.127000",
          "content": "<p>i can‘t agree more</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1168771,
          "author_name": "sheep",
          "author_url": "",
          "post_date": "2021-01-25T06:49:53.187000",
          "content": "<p>In my experience, This kind of competition often end with pretty big shake up. And at the time when the bug of private LB exposed to the public, I only have one submission, scored 0.892 in PB, rank around 600+, but scored 0.897 in private LB and rank 170+. <br>\nSo It is better to trust your CV.</p>",
          "votes": 5,
          "replies": []
        },
        {
          "id": 1168787,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-25T07:12:40.727000",
          "content": "<p>Oh，This is I first time attends this kind of competition，thank you remind me！ I'll take note of the problem</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1168907,
          "author_name": "Benlei Cui",
          "author_url": "",
          "post_date": "2021-01-25T08:41:58.723000",
          "content": "<p><a href=\"https://www.kaggle.com/steamedsheep\" target=\"_blank\">@steamedsheep</a> Hi sheep😄, I wonder that such a big change is beacuse you have cleaned the data?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1168992,
          "author_name": "sheep",
          "author_url": "",
          "post_date": "2021-01-25T09:24:38.263000",
          "content": "<p>That's my first submission, and haven't clean the data</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1169192,
          "author_name": "Benlei Cui",
          "author_url": "",
          "post_date": "2021-01-25T11:28:32.607000",
          "content": "<p>Thanks for rely:D  But in addition to being clean or noisy ,distribution differences, what can be the difference between LB and PB to make such a huge shake😔 . Or you balanced the class with few samples? that also can be the reason of the huge shake up</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1170653,
          "author_name": "ilovepotatoes",
          "author_url": "",
          "post_date": "2021-01-26T10:59:50.283000",
          "content": "<p><a href=\"https://www.kaggle.com/steamedsheep\" target=\"_blank\">@steamedsheep</a> the difference is huge for your first submission.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1170378,
      "author_name": "Bcw93",
      "author_url": "",
      "post_date": "2021-01-26T07:13:45.500000",
      "content": "<p>Hi,do you want to team up?Now, I can get 0.903 in single model.Please email me <a>bcwang1993@qq.com</a></p>",
      "votes": 0,
      "replies": [
        {
          "id": 1170800,
          "author_name": "Zekun",
          "author_url": "",
          "post_date": "2021-01-26T12:55:33.643000",
          "content": "<p>Hello,do u need teamates ? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1170829,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-26T13:20:10.993000",
          "content": "<p>sorry，I already have a team appointment with someone else. Can I add a QQ friend with you? We can get together next time😭</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1170851,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2021-01-26T13:46:06.123000",
          "content": "<p><a href=\"https://www.kaggle.com/zekunn\" target=\"_blank\">@zekunn</a> yes,I am looking up a teamates,are you want to team up, email me <a>bcwang1993@qq.com</a></p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1170852,
          "author_name": "Bcw93",
          "author_url": "",
          "post_date": "2021-01-26T13:46:58.240000",
          "content": "<p><a href=\"https://www.kaggle.com/darknesszx\" target=\"_blank\">@darknesszx</a> ,you can send a email to me ,I will tell you my wechat.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1169680,
      "author_name": "Heroseo",
      "author_url": "",
      "post_date": "2021-01-25T17:06:38.860000",
      "content": "<p>You are already in a high rank. :)</p>\n<p><a href=\"https://www.kaggle.com/darknesszx\" target=\"_blank\">@darknesszx</a> </p>",
      "votes": 0,
      "replies": [
        {
          "id": 1169763,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-25T18:15:21.837000",
          "content": "<p>Haha, there is still a big gap with you.I've seen some of your work on noise label in Notebook and it's great!!I'll try to catch up with you，Ha ha!!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1168326,
      "author_name": "LeoF",
      "author_url": "",
      "post_date": "2021-01-24T22:10:35.253000",
      "content": "<p>What is your single model score?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1168638,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-25T04:49:10.333000",
          "content": "<p>efficientnetb4，0.901</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1170125,
          "author_name": "LeoF",
          "author_url": "",
          "post_date": "2021-01-26T02:13:00.397000",
          "content": "<p>Did you use TTA for inference?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1172778,
          "author_name": "Shiro",
          "author_url": "",
          "post_date": "2021-01-27T14:47:13.127000",
          "content": "<p>Wow, only TTA make you an increase of 0.03 ? My single model without TTA is doing 0.903 but haven't seen any improvement with TTA 😲</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1174835,
          "author_name": "MivHashBrown",
          "author_url": "",
          "post_date": "2021-01-28T18:37:17.863000",
          "content": "<p>which model are you using, none of my models are reaching above lb 8.8</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1167794,
      "author_name": "FGPC",
      "author_url": "",
      "post_date": "2021-01-24T13:35:58.943000",
      "content": "<p>Trust your CV and have robust approach on dealing label noise</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1167802,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-24T13:49:42.897000",
          "content": "<p>Is manual cleaning of noise data necessary？</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1167803,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-24T13:50:56.823000",
          "content": "<p>Thank you very much！i will try it！</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1167842,
          "author_name": "FGPC",
          "author_url": "",
          "post_date": "2021-01-24T14:06:52.593000",
          "content": "<p>No, I recommend to not perform any manual label clean-up. Private lb also consist of noisy labels if you test your submission/s using different configuration like image size, augmentation etc, it generates different submission score. Otherwise, we can achieve a very high score in classification with a minimum of 0.95+, like other previous classification competition. </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1167875,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-24T14:40:00.117000",
          "content": "<p>Thank you very much,i will try to find some robust ways to improve my score!!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1169329,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-01-25T12:45:17.577000",
      "content": "",
      "votes": 2,
      "replies": [
        {
          "id": 1169546,
          "author_name": "DarknessZX",
          "author_url": "",
          "post_date": "2021-01-25T14:52:36.080000",
          "content": "<p>谢谢提醒！！！哈哈哈，我尝试一下看看！👍</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1170127,
          "author_name": "Benlei Cui",
          "author_url": "",
          "post_date": "2021-01-26T02:16:28.910000",
          "content": "<p><a href=\"https://www.kaggle.com/jianxinhu\" target=\"_blank\">@jianxinhu</a> 那个说的是PANDA那个比赛</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1170135,
          "author_name": "LeoF",
          "author_url": "",
          "post_date": "2021-01-26T02:29:19.043000",
          "content": "<p>The host keeps silent all the time. Nobody knows the secret of test images. But the class distribution of first 31% test images is similar to train images.</p>",
          "votes": 4,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1167724": "Now my score is 0.904，the normal thricks i had used in my notebook，just like tta，label smooth，cutout，mixup ，model essemble and so on.But i do not have a good idea now.\nI think there must be some problem in noise label，but how to deal it？\nI find two loss function in my last discussion，but there are some problem when i tyr to use them.\nAnybody have some better ways to improve a higher score？",
    "1167814": "Is your ultimate goal solely overfitting the public LB  ?",
    "1168473": "Borden your mind and use your inner eye to overfit PB!",
    "1170378": "Hi,do you want to team up?Now, I can get 0.903 in single model.Please email me bcwang1993@qq.com",
    "1169680": "You are already in a high rank. :)\n\n@darknesszx ",
    "1168326": "What is your single model score?",
    "1167794": "Trust your CV and have robust approach on dealing label noise",
    "1169329": ""
  }
}