{
  "id": 231868,
  "title": "How to ensemble segmentation models",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/231868",
  "author_name": "",
  "post_date": "2021-04-10T19:09:29.273042400Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I was wondering how is everyone ensembling their different models together? Even when doing cross validation (5 folds for example), how are you putting the predictions together? For example when doing object detection there are some techniques used such as weighted boxes fusion, etc…</p>",
  "messages": [
    {
      "id": "1269678",
      "postDate": "04/10/2021 19:09:29",
      "content": "<p>I was wondering how is everyone ensembling their different models together? Even when doing cross validation (5 folds for example), how are you putting the predictions together? For example when doing object detection there are some techniques used such as weighted boxes fusion, etc…</p>",
      "rawMarkdown": "I was wondering how is everyone ensembling their different models together? Even when doing cross validation (5 folds for example), how are you putting the predictions together? For example when doing object detection there are some techniques used such as weighted boxes fusion, etc...",
      "votes": null
    },
    {
      "id": "1269758",
      "postDate": "04/10/2021 21:57:12",
      "content": "<p><a href=\"https://www.kaggle.com/mostafaibrahim17\" target=\"_blank\">@mostafaibrahim17</a> each model makes a prediction for a tile. for each pixel a probability is predicted. You can average those probabilities. Then pick a threshold … everything below the threshold is not part of the mask … everything above is part of the mask.</p>\n<p>A bit simplified but that's basically what it comes down to. Take a look at some inference kernels. There are plenty of examples. Goodluck!</p>",
      "rawMarkdown": "mostafaibrahim17 each model makes a prediction for a tile. for each pixel a probability is predicted. You can average those probabilities. Then pick a threshold ... everything below the threshold is not part of the mask ... everything above is part of the mask.\n\nA bit simplified but that's basically what it comes down to. Take a look at some inference kernels. There are plenty of examples. Goodluck!",
      "votes": null
    },
    {
      "id": "1271487",
      "postDate": "04/12/2021 16:53:21",
      "content": "<p>Well that actually make a lot of sense, thank you very much!</p>",
      "rawMarkdown": "Well that actually make a lot of sense, thank you very much!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1269758,
      "author_name": "rsmits",
      "author_url": "",
      "post_date": "04/10/2021 21:57:12",
      "content": "<p><a href=\"https://www.kaggle.com/mostafaibrahim17\" target=\"_blank\">@mostafaibrahim17</a> each model makes a prediction for a tile. for each pixel a probability is predicted. You can average those probabilities. Then pick a threshold … everything below the threshold is not part of the mask … everything above is part of the mask.</p>\n<p>A bit simplified but that's basically what it comes down to. Take a look at some inference kernels. There are plenty of examples. Goodluck!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1271487,
          "author_name": "victorasso",
          "author_url": "",
          "post_date": "04/12/2021 16:53:21",
          "content": "<p>Well that actually make a lot of sense, thank you very much!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1269678": "I was wondering how is everyone ensembling their different models together? Even when doing cross validation (5 folds for example), how are you putting the predictions together? For example when doing object detection there are some techniques used such as weighted boxes fusion, etc...",
    "1269758": "mostafaibrahim17 each model makes a prediction for a tile. for each pixel a probability is predicted. You can average those probabilities. Then pick a threshold ... everything below the threshold is not part of the mask ... everything above is part of the mask.\n\nA bit simplified but that's basically what it comes down to. Take a look at some inference kernels. There are plenty of examples. Goodluck!",
    "1271487": "Well that actually make a lot of sense, thank you very much!"
  },
  "source": "meta"
}