{
  "id": 154669,
  "title": "LB speculation",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/154669",
  "author_name": "Miroslav Valan",
  "post_date": "2020-05-29T10:06:12.376000",
  "votes": 3,
  "comment_count": 6,
  "views": 0,
  "content": "<p>let's predict the final winner LB and medal zones.</p>\n\n<p>0.92+ is doable with a single B0 trained on 75:25 split\nthe final (public) score could get to 0.96 and gold would be 0.95.</p>\n\n<p><strong>PS1</strong>: this is just a speculation - i.e i have no evidence yet and I just assume that larger models, better training, ensemble, metadata and external data will each give a significant boost. </p>\n\n<p><strong>PS2</strong>: if you have a track record in CV comps which includes minimum team gold or solo top2% and you are looking for a team to hunt gold reach out.</p>",
  "messages": [
    {
      "id": 866408,
      "postDate": "2020-05-29T11:20:34.207Z",
      "content": "<p>you can check the baseline performance of past ISIC competitions:\n<a href=\"https://challenge2019.isic-archive.com/leaderboard.html\">https://challenge2019.isic-archive.com/leaderboard.html</a></p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F113660%2Fd6c27cd9b953f74302e5d725b836eaee%2FSelection_040.png?generation=1590751229686289&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "you can check the baseline performance of past ISIC competitions:\nhttps://challenge2019.isic-archive.com/leaderboard.html\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F113660%2Fd6c27cd9b953f74302e5d725b836eaee%2FSelection_040.png?generation=1590751229686289&amp;alt=media)\n\n",
      "votes": 5,
      "replies": [
        {
          "id": 866423,
          "postDate": "2020-05-29T11:36:50.573Z",
          "content": "<p>thanks <a href=\"/hengck23\">@hengck23</a>  this is worth reading. </p>\n\n<p>the winner got 0.928 on melanoma so we might beat that in less than 48h :) (we also have more data).</p>\n\n<p>However, spending &lt;5 minutes on reading the winning solution I have a feeling that many kagglers should beat their score.</p>",
          "rawMarkdown": "thanks @hengck23  this is worth reading. \n\nthe winner got 0.928 on melanoma so we might beat that in less than 48h :) (we also have more data).\n\nHowever, spending &lt;5 minutes on reading the winning solution I have a feeling that many kagglers should beat their score.",
          "votes": 2
        }
      ]
    },
    {
      "id": 866345,
      "postDate": "2020-05-29T10:06:12.377Z",
      "content": "<p>let's predict the final winner LB and medal zones.</p>\n\n<p>0.92+ is doable with a single B0 trained on 75:25 split\nthe final (public) score could get to 0.96 and gold would be 0.95.</p>\n\n<p><strong>PS1</strong>: this is just a speculation - i.e i have no evidence yet and I just assume that larger models, better training, ensemble, metadata and external data will each give a significant boost. </p>\n\n<p><strong>PS2</strong>: if you have a track record in CV comps which includes minimum team gold or solo top2% and you are looking for a team to hunt gold reach out.</p>",
      "rawMarkdown": "let's predict the final winner LB and medal zones.\n\n0.92+ is doable with a single B0 trained on 75:25 split\nthe final (public) score could get to 0.96 and gold would be 0.95.\n\n**PS1**: this is just a speculation - i.e i have no evidence yet and I just assume that larger models, better training, ensemble, metadata and external data will each give a significant boost. \n\n**PS2**: if you have a track record in CV comps which includes minimum team gold or solo top2% and you are looking for a team to hunt gold reach out.",
      "votes": 3
    },
    {
      "id": 883012,
      "postDate": "2020-06-12T09:45:18.063Z",
      "content": "<p>2 weeks into the competition and my speculation still holds :)</p>",
      "rawMarkdown": "2 weeks into the competition and my speculation still holds :)",
      "votes": 1,
      "replies": [
        {
          "id": 883135,
          "postDate": "2020-06-12T11:53:20.573Z",
          "content": "<p>Well the amount by which everyone is climbing the LB and dropping , and climbing, regularly, I think it's going to be a complete chaos at the end.\nPS: This is just a speculation :)</p>",
          "rawMarkdown": "Well the amount by which everyone is climbing the LB and dropping , and climbing, regularly, I think it's going to be a complete chaos at the end.\nPS: This is just a speculation :)",
          "votes": 2
        },
        {
          "id": 883140,
          "postDate": "2020-06-12T12:02:11.490Z",
          "content": "<p>yes it will. the metric is super sensitive... and many of the top submissions are likely already overfitted to the LB. \nAt least I will not slide down from the first place :D</p>",
          "rawMarkdown": "yes it will. the metric is super sensitive... and many of the top submissions are likely already overfitted to the LB. \nAt least I will not slide down from the first place :D",
          "votes": 2
        },
        {
          "id": 883146,
          "postDate": "2020-06-12T12:09:42.500Z",
          "rawMarkdown": "",
          "votes": 1,
          "isDeleted": true
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 866408,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2020-05-29T11:20:34.207000",
      "content": "<p>you can check the baseline performance of past ISIC competitions:\n<a href=\"https://challenge2019.isic-archive.com/leaderboard.html\">https://challenge2019.isic-archive.com/leaderboard.html</a></p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F113660%2Fd6c27cd9b953f74302e5d725b836eaee%2FSelection_040.png?generation=1590751229686289&amp;alt=media\" alt=\"\"></p>",
      "votes": 5,
      "replies": [
        {
          "id": 866423,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2020-05-29T11:36:50.573000",
          "content": "<p>thanks <a href=\"/hengck23\">@hengck23</a>  this is worth reading. </p>\n\n<p>the winner got 0.928 on melanoma so we might beat that in less than 48h :) (we also have more data).</p>\n\n<p>However, spending &lt;5 minutes on reading the winning solution I have a feeling that many kagglers should beat their score.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 883012,
      "author_name": "Miroslav Valan",
      "author_url": "",
      "post_date": "2020-06-12T09:45:18.063000",
      "content": "<p>2 weeks into the competition and my speculation still holds :)</p>",
      "votes": 1,
      "replies": [
        {
          "id": 883135,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-06-12T11:53:20.573000",
          "content": "<p>Well the amount by which everyone is climbing the LB and dropping , and climbing, regularly, I think it's going to be a complete chaos at the end.\nPS: This is just a speculation :)</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 883140,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2020-06-12T12:02:11.490000",
          "content": "<p>yes it will. the metric is super sensitive... and many of the top submissions are likely already overfitted to the LB. \nAt least I will not slide down from the first place :D</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 883146,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-06-12T12:09:42.500000",
          "content": "",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "866408": "you can check the baseline performance of past ISIC competitions:\nhttps://challenge2019.isic-archive.com/leaderboard.html\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F113660%2Fd6c27cd9b953f74302e5d725b836eaee%2FSelection_040.png?generation=1590751229686289&amp;alt=media)\n\n",
    "866345": "let's predict the final winner LB and medal zones.\n\n0.92+ is doable with a single B0 trained on 75:25 split\nthe final (public) score could get to 0.96 and gold would be 0.95.\n\n**PS1**: this is just a speculation - i.e i have no evidence yet and I just assume that larger models, better training, ensemble, metadata and external data will each give a significant boost. \n\n**PS2**: if you have a track record in CV comps which includes minimum team gold or solo top2% and you are looking for a team to hunt gold reach out.",
    "883012": "2 weeks into the competition and my speculation still holds :)"
  }
}