{
  "id": 426583,
  "title": "Only single model, single fold and dilation...",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/426583",
  "author_name": "",
  "post_date": "2023-07-24T09:24:54.524281600Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>For me only single model, single fold model (Hybrid Task Cascade, mmdet3) and dilation works..</p>\n<p>Not working for me:</p>\n<ul>\n<li>pseudo labels  (minus 3% LB)</li>\n<li>ensembling with all folds  (minus 5-7% LB)</li>\n<li>postprocessing (except dilation)  </li>\n<li>any thresholding at all  </li>\n<li>any complicated TTA (except simple flip and resize)</li>\n<li>training with glomerulus and filtering at inference stage</li>\n</ul>\n<p>What is working for you except pure one-fold model ?  </p>",
  "messages": [
    {
      "id": "2356579",
      "postDate": "07/24/2023 09:24:54",
      "content": "<p>For me only single model, single fold model (Hybrid Task Cascade, mmdet3) and dilation works..</p>\n<p>Not working for me:</p>\n<ul>\n<li>pseudo labels  (minus 3% LB)</li>\n<li>ensembling with all folds  (minus 5-7% LB)</li>\n<li>postprocessing (except dilation)  </li>\n<li>any thresholding at all  </li>\n<li>any complicated TTA (except simple flip and resize)</li>\n<li>training with glomerulus and filtering at inference stage</li>\n</ul>\n<p>What is working for you except pure one-fold model ?  </p>",
      "rawMarkdown": "For me only single model, single fold model (Hybrid Task Cascade, mmdet3) and dilation works..\n\nNot working for me:\n- pseudo labels  (minus 3% LB)\n- ensembling with all folds  (minus 5-7% LB)\n- postprocessing (except dilation)  \n- any thresholding at all  \n- any complicated TTA (except simple flip and resize)\n- training with glomerulus and filtering at inference stage\n\nWhat is working for you except pure one-fold model ?",
      "votes": null
    },
    {
      "id": "2356890",
      "postDate": "07/24/2023 12:21:56",
      "content": "<p>Ensemble depends on setting can boost a bit (Adding more models decrease LB score),  TTA does not work for me</p>",
      "rawMarkdown": "Ensemble depends on setting can boost a bit (Adding more models decrease LB score),  TTA does not work for me",
      "votes": null
    },
    {
      "id": "2356914",
      "postDate": "07/24/2023 12:34:44",
      "content": "<p>I'm pretty much in agreement with you. I guess how to spilt fold or do right ensembling may be the key (a hyperparameter game……) JUST guess!😂</p>",
      "rawMarkdown": "I'm pretty much in agreement with you. I guess how to spilt fold or do right ensembling may be the key (a hyperparameter game......) JUST guess!😂",
      "votes": null
    },
    {
      "id": "2357072",
      "postDate": "07/24/2023 14:24:02",
      "content": "<p>Ensemble is tricky here, because it depends a lot on the model and the setting. I saw it work well on some models but the boost on mmdet3 is small (about +0.01).<br>\nAlso very disappointed with TTA, it didn't work on any model I tried.</p>",
      "rawMarkdown": "Ensemble is tricky here, because it depends a lot on the model and the setting. I saw it work well on some models but the boost on mmdet3 is small (about +0.01).\nAlso very disappointed with TTA, it didn't work on any model I tried.",
      "votes": null
    },
    {
      "id": "2357164",
      "postDate": "07/24/2023 15:16:59",
      "content": "<p>Model ensembles work for me in some situations, when there is not a big difference between the scores of two single-folded models, the model ensemble improves by about 0.002<br>\nSimple TTA also works, but not by much, maybe 0.001 or so</p>",
      "rawMarkdown": "Model ensembles work for me in some situations, when there is not a big difference between the scores of two single-folded models, the model ensemble improves by about 0.002\nSimple TTA also works, but not by much, maybe 0.001 or so",
      "votes": null
    },
    {
      "id": "2357192",
      "postDate": "07/24/2023 15:30:42",
      "content": "<p>In my case <code>CopyPaste</code> augmentation increased my CV and LB for a single fold model (just a little though).</p>",
      "rawMarkdown": "In my case `CopyPaste` augmentation increased my CV and LB for a single fold model (just a little though).",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2356890,
      "author_name": "ptran1203",
      "author_url": "",
      "post_date": "07/24/2023 12:21:56",
      "content": "<p>Ensemble depends on setting can boost a bit (Adding more models decrease LB score),  TTA does not work for me</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2356914,
      "author_name": "wqx20000115",
      "author_url": "",
      "post_date": "07/24/2023 12:34:44",
      "content": "<p>I'm pretty much in agreement with you. I guess how to spilt fold or do right ensembling may be the key (a hyperparameter game……) JUST guess!😂</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2357072,
      "author_name": "nazariio",
      "author_url": "",
      "post_date": "07/24/2023 14:24:02",
      "content": "<p>Ensemble is tricky here, because it depends a lot on the model and the setting. I saw it work well on some models but the boost on mmdet3 is small (about +0.01).<br>\nAlso very disappointed with TTA, it didn't work on any model I tried.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2357164,
      "author_name": "bent1e",
      "author_url": "",
      "post_date": "07/24/2023 15:16:59",
      "content": "<p>Model ensembles work for me in some situations, when there is not a big difference between the scores of two single-folded models, the model ensemble improves by about 0.002<br>\nSimple TTA also works, but not by much, maybe 0.001 or so</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2357192,
      "author_name": "atamazian",
      "author_url": "",
      "post_date": "07/24/2023 15:30:42",
      "content": "<p>In my case <code>CopyPaste</code> augmentation increased my CV and LB for a single fold model (just a little though).</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2356579": "For me only single model, single fold model (Hybrid Task Cascade, mmdet3) and dilation works..\n\nNot working for me:\n- pseudo labels  (minus 3% LB)\n- ensembling with all folds  (minus 5-7% LB)\n- postprocessing (except dilation)  \n- any thresholding at all  \n- any complicated TTA (except simple flip and resize)\n- training with glomerulus and filtering at inference stage\n\nWhat is working for you except pure one-fold model ?",
    "2356890": "Ensemble depends on setting can boost a bit (Adding more models decrease LB score),  TTA does not work for me",
    "2356914": "I'm pretty much in agreement with you. I guess how to spilt fold or do right ensembling may be the key (a hyperparameter game......) JUST guess!😂",
    "2357072": "Ensemble is tricky here, because it depends a lot on the model and the setting. I saw it work well on some models but the boost on mmdet3 is small (about +0.01).\nAlso very disappointed with TTA, it didn't work on any model I tried.",
    "2357164": "Model ensembles work for me in some situations, when there is not a big difference between the scores of two single-folded models, the model ensemble improves by about 0.002\nSimple TTA also works, but not by much, maybe 0.001 or so",
    "2357192": "In my case `CopyPaste` augmentation increased my CV and LB for a single fold model (just a little though)."
  },
  "source": "meta"
}