{
  "id": 94114,
  "title": "1st place solution",
  "url": "/competitions/cassava-disease/writeups/gary-1st-place-solution",
  "author_name": "",
  "post_date": "2019-06-13T09:00:17.350Z",
  "votes": 64,
  "comment_count": 7,
  "views": 0,
  "content": "<h2>Train on train images</h2>\n\nThat is my best score on private leaderboard\n\n<ul>\n<li>model: 5 fold se_resnext101 </li>\n<li>data augumentaion: RandomCrop, VFlip, HFilp, RandomRotate</li>\n<li>3 * TTA</li>\n<li>cv: I don't remember it too clearly, but it should be over 0.925, LB(public): 0.92384, LB(private): 0.93507</li>\n</ul>\n\n<h2>Train on extra images as pseudo labels</h2>\n\nI use some models(such as resnet50, se_resnext50 and se_resnext101) that can get the best LB on leaderboard to ensemble for extraimgas. And then I did some experiments, and found that the CV could be higher if the samples with a filtering probability greater than 95 were used as training set for training.\n\n<ul>\n<li>config：same as above</li>\n<li>cv: around 0.935, LB(public): 0.92119(not work),  LB(private): 0.93771</li>\n</ul>\n\n<h2>ensemble score</h2>\n\n<ul>\n<li><p>ensemble se101 witn only trian images and se101 with extra images.\n&gt; LB(public): 0.920, LB(private): 0.93904</p></li>\n<li><p>ensemble se50 witn only trian images and se50 with extra images.\n&gt; LB(public): 0.923, LB(private): 0.93551</p></li>\n</ul>\n\n<h2>advices</h2>\n\n<ol>\n<li>Do not trust the LB, trust your CV.</li>\n<li>You can ensemble many many models，but it is not necessarily.\n...</li>\n</ol>",
  "messages": [
    {
      "id": "541296",
      "postDate": "06/02/2019 07:41:55",
      "content": "<h2>Train on train images</h2>\n\nThat is my best score on private leaderboard\n\n<ul>\n<li>model: 5 fold se_resnext101 </li>\n<li>data augumentaion: RandomCrop, VFlip, HFilp, RandomRotate</li>\n<li>3 * TTA</li>\n<li>cv: I don't remember it too clearly, but it should be over 0.925, LB(public): 0.92384, LB(private): 0.93507</li>\n</ul>\n\n<h2>Train on extra images as pseudo labels</h2>\n\nI use some models(such as resnet50, se_resnext50 and se_resnext101) that can get the best LB on leaderboard to ensemble for extraimgas. And then I did some experiments, and found that the CV could be higher if the samples with a filtering probability greater than 95 were used as training set for training.\n\n<ul>\n<li>config：same as above</li>\n<li>cv: around 0.935, LB(public): 0.92119(not work),  LB(private): 0.93771</li>\n</ul>\n\n<h2>ensemble score</h2>\n\n<ul>\n<li><p>ensemble se101 witn only trian images and se101 with extra images.\n&gt; LB(public): 0.920, LB(private): 0.93904</p></li>\n<li><p>ensemble se50 witn only trian images and se50 with extra images.\n&gt; LB(public): 0.923, LB(private): 0.93551</p></li>\n</ul>\n\n<h2>advices</h2>\n\n<ol>\n<li>Do not trust the LB, trust your CV.</li>\n<li>You can ensemble many many models，but it is not necessarily.\n...</li>\n</ol>",
      "rawMarkdown": "## Train on train images\n#### That is my best score on private leaderboard\n- model: 5 fold se_resnext101 \n- data augumentaion: RandomCrop, VFlip, HFilp, RandomRotate\n- 3 * TTA\n- cv: I don't remember it too clearly, but it should be over 0.925, LB(public): 0.92384, LB(private): 0.93507\n\n## Train on extra images as pseudo labels \n#### I use some models(such as resnet50, se_resnext50 and se_resnext101) that can get the best LB on leaderboard to ensemble for extraimgas. And then I did some experiments, and found that the CV could be higher if the samples with a filtering probability greater than 95 were used as training set for training.\n- config：same as above\n- cv: around 0.935, LB(public): 0.92119(not work),  LB(private): 0.93771\n\n## ensemble score\n- ensemble se101 witn only trian images and se101 with extra images.\n&gt; LB(public): 0.920, LB(private): 0.93904\n\n- ensemble se50 witn only trian images and se50 with extra images.\n&gt; LB(public): 0.923, LB(private): 0.93551\n\n## advices\n1. Do not trust the LB, trust your CV.\n2. You can ensemble many many models，but it is not necessarily.\n...",
      "votes": null
    },
    {
      "id": "541311",
      "postDate": "06/02/2019 08:07:58",
      "content": "<p>Good job!</p>",
      "rawMarkdown": "Good job!",
      "votes": null
    },
    {
      "id": "541428",
      "postDate": "06/02/2019 12:39:18",
      "content": "<p>The unbalanced train labels bother me. I am curious if you do something to figure it. thankyou</p>",
      "rawMarkdown": "The unbalanced train labels bother me. I am curious if you do something to figure it. thankyou",
      "votes": null
    },
    {
      "id": "542045",
      "postDate": "06/03/2019 11:30:00",
      "content": "<p>Good job! Thank you.</p>",
      "rawMarkdown": "Good job! Thank you.",
      "votes": null
    },
    {
      "id": "542063",
      "postDate": "06/03/2019 12:01:25",
      "content": "<p>Thanks Gary</p>",
      "rawMarkdown": "Thanks Gary",
      "votes": null
    },
    {
      "id": "550660",
      "postDate": "06/11/2019 22:16:18",
      "content": "<p>Hello, can you please post your email address here? For some reason Kaggle isn't allowing me to contact you. I wanted to send you instructions for presenting your work, prizes etc.</p>",
      "rawMarkdown": "Hello, can you please post your email address here? For some reason Kaggle isn't allowing me to contact you. I wanted to send you instructions for presenting your work, prizes etc.",
      "votes": null
    },
    {
      "id": "551145",
      "postDate": "06/12/2019 11:43:36",
      "content": "<p>Sure</p>",
      "rawMarkdown": "Sure",
      "votes": null
    },
    {
      "id": "1132396",
      "postDate": "12/30/2020 11:33:57",
      "content": "<p>The unbalanced train labels bother me. I am curious if you do something to figure it. thankyou</p>",
      "rawMarkdown": "The unbalanced train labels bother me. I am curious if you do something to figure it. thankyou",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1132396,
      "author_name": "clwwlc",
      "author_url": "",
      "post_date": "12/30/2020 11:33:57",
      "content": "<p>The unbalanced train labels bother me. I am curious if you do something to figure it. thankyou</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 541311,
      "author_name": "mamontenokk",
      "author_url": "",
      "post_date": "06/02/2019 08:07:58",
      "content": "<p>Good job!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 541428,
      "author_name": "erichen188",
      "author_url": "",
      "post_date": "06/02/2019 12:39:18",
      "content": "<p>The unbalanced train labels bother me. I am curious if you do something to figure it. thankyou</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 542045,
      "author_name": "luxinfeng",
      "author_url": "",
      "post_date": "06/03/2019 11:30:00",
      "content": "<p>Good job! Thank you.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 542063,
      "author_name": "seefun",
      "author_url": "",
      "post_date": "06/03/2019 12:01:25",
      "content": "<p>Thanks Gary</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 550660,
      "author_name": "tgebru123",
      "author_url": "",
      "post_date": "06/11/2019 22:16:18",
      "content": "<p>Hello, can you please post your email address here? For some reason Kaggle isn't allowing me to contact you. I wanted to send you instructions for presenting your work, prizes etc.</p>",
      "votes": null,
      "replies": [
        {
          "id": 551145,
          "author_name": "garybios",
          "author_url": "",
          "post_date": "06/12/2019 11:43:36",
          "content": "<p>Sure</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "541296": "## Train on train images\n#### That is my best score on private leaderboard\n- model: 5 fold se_resnext101 \n- data augumentaion: RandomCrop, VFlip, HFilp, RandomRotate\n- 3 * TTA\n- cv: I don't remember it too clearly, but it should be over 0.925, LB(public): 0.92384, LB(private): 0.93507\n\n## Train on extra images as pseudo labels \n#### I use some models(such as resnet50, se_resnext50 and se_resnext101) that can get the best LB on leaderboard to ensemble for extraimgas. And then I did some experiments, and found that the CV could be higher if the samples with a filtering probability greater than 95 were used as training set for training.\n- config：same as above\n- cv: around 0.935, LB(public): 0.92119(not work),  LB(private): 0.93771\n\n## ensemble score\n- ensemble se101 witn only trian images and se101 with extra images.\n&gt; LB(public): 0.920, LB(private): 0.93904\n\n- ensemble se50 witn only trian images and se50 with extra images.\n&gt; LB(public): 0.923, LB(private): 0.93551\n\n## advices\n1. Do not trust the LB, trust your CV.\n2. You can ensemble many many models，but it is not necessarily.\n...",
    "541311": "Good job!",
    "541428": "The unbalanced train labels bother me. I am curious if you do something to figure it. thankyou",
    "542045": "Good job! Thank you.",
    "542063": "Thanks Gary",
    "550660": "Hello, can you please post your email address here? For some reason Kaggle isn't allowing me to contact you. I wanted to send you instructions for presenting your work, prizes etc.",
    "551145": "Sure",
    "1132396": "The unbalanced train labels bother me. I am curious if you do something to figure it. thankyou"
  },
  "source": "existing_discussion",
  "complete": true,
  "competition_id": 13277,
  "competition_slug": "cassava-disease",
  "team_id": 3118163,
  "solution_url": "/competitions/cassava-disease/writeups/gary-1st-place-solution"
}