{
  "id": 417264,
  "title": "75th solution",
  "url": "/competitions/vesuvius-challenge-ink-detection/discussion/417264",
  "author_name": "Séraphin Lampion",
  "post_date": "2023-06-15T00:57:44.838000",
  "votes": 5,
  "comment_count": 0,
  "views": 0,
  "content": "<p>public 0.68 - private 0.58<br>\n5 fold, val Fragment one<br>\n12 slices (23-35)<br>\n3DResnet34<br>\nBCE loss soft (0.2)<br>\nmask erosion<br>\nno pre-trained</p>\n<p>I tried a lot of different model 2d / 3D,  segformer,3defficientnet etc.</p>\n<p>to be fair i just learned two days ago that i should apply individual treshold instead of one of the different ensemble model before averaging so ensembling was alway bad lol (F1 usually th 0.45 and F2 0.8 and i put 0.5 for both ahah), I could have pushed higher with good ensemble  !! </p>\n<p>Very interesting i learned A LOT !<br>\nbest no picked: 0.67 public,0.6 private  with same setting except 22-34 slices</p>\n<p>A bit irrelevant solution but thats the occasion to thanks a lot all for this wonderful first competition !!</p>",
  "messages": [
    {
      "id": 2302924,
      "postDate": "2023-06-15T00:57:44.840Z",
      "content": "<p>public 0.68 - private 0.58<br>\n5 fold, val Fragment one<br>\n12 slices (23-35)<br>\n3DResnet34<br>\nBCE loss soft (0.2)<br>\nmask erosion<br>\nno pre-trained</p>\n<p>I tried a lot of different model 2d / 3D,  segformer,3defficientnet etc.</p>\n<p>to be fair i just learned two days ago that i should apply individual treshold instead of one of the different ensemble model before averaging so ensembling was alway bad lol (F1 usually th 0.45 and F2 0.8 and i put 0.5 for both ahah), I could have pushed higher with good ensemble  !! </p>\n<p>Very interesting i learned A LOT !<br>\nbest no picked: 0.67 public,0.6 private  with same setting except 22-34 slices</p>\n<p>A bit irrelevant solution but thats the occasion to thanks a lot all for this wonderful first competition !!</p>",
      "rawMarkdown": "public 0.68 - private 0.58\n5 fold, val Fragment one\n12 slices (23-35)\n3DResnet34\nBCE loss soft (0.2)\nmask erosion\nno pre-trained\n\nI tried a lot of different model 2d / 3D,  segformer,3defficientnet etc.\n\nto be fair i just learned two days ago that i should apply individual treshold instead of one of the different ensemble model before averaging so ensembling was alway bad lol (F1 usually th 0.45 and F2 0.8 and i put 0.5 for both ahah), I could have pushed higher with good ensemble  !! \n\nVery interesting i learned A LOT !\nbest no picked: 0.67 public,0.6 private  with same setting except 22-34 slices\n\nA bit irrelevant solution but thats the occasion to thanks a lot all for this wonderful first competition !!",
      "votes": 5
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2302924": "public 0.68 - private 0.58\n5 fold, val Fragment one\n12 slices (23-35)\n3DResnet34\nBCE loss soft (0.2)\nmask erosion\nno pre-trained\n\nI tried a lot of different model 2d / 3D,  segformer,3defficientnet etc.\n\nto be fair i just learned two days ago that i should apply individual treshold instead of one of the different ensemble model before averaging so ensembling was alway bad lol (F1 usually th 0.45 and F2 0.8 and i put 0.5 for both ahah), I could have pushed higher with good ensemble  !! \n\nVery interesting i learned A LOT !\nbest no picked: 0.67 public,0.6 private  with same setting except 22-34 slices\n\nA bit irrelevant solution but thats the occasion to thanks a lot all for this wonderful first competition !!"
  }
}