{
  "id": 475232,
  "title": "How did you guys get such a good public Score ?",
  "url": "/competitions/blood-vessel-segmentation/discussion/475232",
  "author_name": "JEANMPIA",
  "post_date": "2024-02-07T15:07:45.893000",
  "votes": 6,
  "comment_count": 5,
  "views": 0,
  "content": "<p>For the last month or so of the comp, I was stuck at 0.8~ (700/1200) for all of my submission. Most of them ended up scoring above .550 on the private LB.</p>\n<ul>\n<li><p>How did you get such an high scoring public LB ? I tried over-fitting the public LB with threshold and low regularization protocol to get a sense of my real position but got to .816 at max. </p></li>\n<li><p>Also, can you share some experiments that scored really high on the public LB and scored terrible on the private one ? what did you do different ? what did you learn from it ?</p></li>\n</ul>\n<p>I thought I would learn more about regularisation this comp but I now feel like I had a lucky protocol that worked well on private (I did focus on that tho, but seeing other's solution I feel like everything can work). What if they swapped the public and private sets ? Would I have shaken down ?</p>",
  "messages": [
    {
      "id": 2641581,
      "postDate": "2024-02-07T15:07:45.893Z",
      "content": "<p>For the last month or so of the comp, I was stuck at 0.8~ (700/1200) for all of my submission. Most of them ended up scoring above .550 on the private LB.</p>\n<ul>\n<li><p>How did you get such an high scoring public LB ? I tried over-fitting the public LB with threshold and low regularization protocol to get a sense of my real position but got to .816 at max. </p></li>\n<li><p>Also, can you share some experiments that scored really high on the public LB and scored terrible on the private one ? what did you do different ? what did you learn from it ?</p></li>\n</ul>\n<p>I thought I would learn more about regularisation this comp but I now feel like I had a lucky protocol that worked well on private (I did focus on that tho, but seeing other's solution I feel like everything can work). What if they swapped the public and private sets ? Would I have shaken down ?</p>",
      "rawMarkdown": "For the last month or so of the comp, I was stuck at 0.8~ (700/1200) for all of my submission. Most of them ended up scoring above .550 on the private LB.\n\n- How did you get such an high scoring public LB ? I tried over-fitting the public LB with threshold and low regularization protocol to get a sense of my real position but got to .816 at max. \n\n- Also, can you share some experiments that scored really high on the public LB and scored terrible on the private one ? what did you do different ? what did you learn from it ?\n\nI thought I would learn more about regularisation this comp but I now feel like I had a lucky protocol that worked well on private (I did focus on that tho, but seeing other's solution I feel like everything can work). What if they swapped the public and private sets ? Would I have shaken down ?",
      "votes": 6
    },
    {
      "id": 2641623,
      "postDate": "2024-02-07T15:20:31.537Z",
      "content": "<p>I had a convnext model with flip, size, rotation, and brightness/contrast augmentations. I had 4 of these with slightly different structure ensembled together. This was how I did it and got to 0.873</p>",
      "rawMarkdown": "I had a convnext model with flip, size, rotation, and brightness/contrast augmentations. I had 4 of these with slightly different structure ensembled together. This was how I did it and got to 0.873",
      "votes": 1,
      "replies": [
        {
          "id": 2641809,
          "postDate": "2024-02-07T17:18:57.077Z",
          "content": "<p>Was ensemble key in the boost of score ? I didn't try ensemble at all this comp.<br>\nWas this Tiling or full image size ? </p>",
          "rawMarkdown": "Was ensemble key in the boost of score ? I didn't try ensemble at all this comp.\nWas this Tiling or full image size ? ",
          "votes": 1,
          "replies": [
            {
              "id": 2641862,
              "postDate": "2024-02-07T18:00:50.207Z",
              "rawMarkdown": "",
              "isDeleted": true
            }
          ]
        }
      ]
    },
    {
      "id": 2641593,
      "postDate": "2024-02-07T15:11:36.200Z",
      "content": "<p>wait for their github code and rerun their experiments.</p>",
      "rawMarkdown": "wait for their github code and rerun their experiments.",
      "votes": -3,
      "replies": [
        {
          "id": 2641811,
          "postDate": "2024-02-07T17:19:15.957Z",
          "content": "<p>I can't learn this from looking at their final code:</p>\n<blockquote>\n  <p>Also, can you share some experiments that scored really high on the public LB and scored terrible on the private one ? what did you do different ? what did you learn from it ?</p>\n</blockquote>",
          "rawMarkdown": "I can't learn this from looking at their final code:\n\n> Also, can you share some experiments that scored really high on the public LB and scored terrible on the private one ? what did you do different ? what did you learn from it ?"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2641623,
      "author_name": "Cody_Null",
      "author_url": "",
      "post_date": "2024-02-07T15:20:31.537000",
      "content": "<p>I had a convnext model with flip, size, rotation, and brightness/contrast augmentations. I had 4 of these with slightly different structure ensembled together. This was how I did it and got to 0.873</p>",
      "votes": 1,
      "replies": [
        {
          "id": 2641809,
          "author_name": "JEANMPIA",
          "author_url": "",
          "post_date": "2024-02-07T17:18:57.077000",
          "content": "<p>Was ensemble key in the boost of score ? I didn't try ensemble at all this comp.<br>\nWas this Tiling or full image size ? </p>",
          "votes": 1,
          "replies": [
            {
              "id": 2641862,
              "author_name": "",
              "author_url": "",
              "post_date": "2024-02-07T18:00:50.207000",
              "content": "",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2641593,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2024-02-07T15:11:36.200000",
      "content": "<p>wait for their github code and rerun their experiments.</p>",
      "votes": -3,
      "replies": [
        {
          "id": 2641811,
          "author_name": "JEANMPIA",
          "author_url": "",
          "post_date": "2024-02-07T17:19:15.957000",
          "content": "<p>I can't learn this from looking at their final code:</p>\n<blockquote>\n  <p>Also, can you share some experiments that scored really high on the public LB and scored terrible on the private one ? what did you do different ? what did you learn from it ?</p>\n</blockquote>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2641581": "For the last month or so of the comp, I was stuck at 0.8~ (700/1200) for all of my submission. Most of them ended up scoring above .550 on the private LB.\n\n- How did you get such an high scoring public LB ? I tried over-fitting the public LB with threshold and low regularization protocol to get a sense of my real position but got to .816 at max. \n\n- Also, can you share some experiments that scored really high on the public LB and scored terrible on the private one ? what did you do different ? what did you learn from it ?\n\nI thought I would learn more about regularisation this comp but I now feel like I had a lucky protocol that worked well on private (I did focus on that tho, but seeing other's solution I feel like everything can work). What if they swapped the public and private sets ? Would I have shaken down ?",
    "2641623": "I had a convnext model with flip, size, rotation, and brightness/contrast augmentations. I had 4 of these with slightly different structure ensembled together. This was how I did it and got to 0.873",
    "2641593": "wait for their github code and rerun their experiments."
  }
}