{
  "id": 231214,
  "title": "Ways to integrate models from CV",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/231214",
  "author_name": "",
  "post_date": "2021-04-07T13:40:45.639557600Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Currently I integrated 5 models trained from 5-fold CV, by doing union (OR) calculation for 5 generated masks by 5 models. But my LB is always good enough (~0.91), even worse than single model.</p>\n<p>So I am wondering whether my interation strategy is not good. I suppose there should be better ways to integrate prediction results from CV models.</p>\n<p>BTW, I used the resized images (4x) so I need to convert the probability to binary values for better interpolation. I also suspect that the problem happens in this step.</p>",
  "messages": [
    {
      "id": "1266107",
      "postDate": "04/07/2021 13:40:45",
      "content": "<p>Currently I integrated 5 models trained from 5-fold CV, by doing union (OR) calculation for 5 generated masks by 5 models. But my LB is always good enough (~0.91), even worse than single model.</p>\n<p>So I am wondering whether my interation strategy is not good. I suppose there should be better ways to integrate prediction results from CV models.</p>\n<p>BTW, I used the resized images (4x) so I need to convert the probability to binary values for better interpolation. I also suspect that the problem happens in this step.</p>",
      "rawMarkdown": "Currently I integrated 5 models trained from 5-fold CV, by doing union (OR) calculation for 5 generated masks by 5 models. But my LB is always good enough (~0.91), even worse than single model.\n\nSo I am wondering whether my interation strategy is not good. I suppose there should be better ways to integrate prediction results from CV models.\n\nBTW, I used the resized images (4x) so I need to convert the probability to binary values for better interpolation. I also suspect that the problem happens in this step.",
      "votes": null
    },
    {
      "id": "1271511",
      "postDate": "04/12/2021 17:09:56",
      "content": "<p>I'm about to start integrating more than 1 model, i havn't implemented yet but for what i read around the discussions, you will get a predicted probability for each pixel on the image, than you can average the probabilities and strictly determine a threshold to cut lower probs out.<br>\nI believe that if you convert all masks to binary before averaging them, you will lose at least some part of the predictions to this. Am i wrong?</p>\n<p>That is in theory what i will do, i let you know once i actually manage to do it.</p>",
      "rawMarkdown": "I'm about to start integrating more than 1 model, i havn't implemented yet but for what i read around the discussions, you will get a predicted probability for each pixel on the image, than you can average the probabilities and strictly determine a threshold to cut lower probs out.\nI believe that if you convert all masks to binary before averaging them, you will lose at least some part of the predictions to this. Am i wrong?\n\nThat is in theory what i will do, i let you know once i actually manage to do it.",
      "votes": null
    },
    {
      "id": "1274194",
      "postDate": "04/15/2021 05:13:30",
      "content": "<p>OK, that's great</p>",
      "rawMarkdown": "OK, that's great",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1271511,
      "author_name": "victorasso",
      "author_url": "",
      "post_date": "04/12/2021 17:09:56",
      "content": "<p>I'm about to start integrating more than 1 model, i havn't implemented yet but for what i read around the discussions, you will get a predicted probability for each pixel on the image, than you can average the probabilities and strictly determine a threshold to cut lower probs out.<br>\nI believe that if you convert all masks to binary before averaging them, you will lose at least some part of the predictions to this. Am i wrong?</p>\n<p>That is in theory what i will do, i let you know once i actually manage to do it.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1274194,
          "author_name": "lijiaqi96",
          "author_url": "",
          "post_date": "04/15/2021 05:13:30",
          "content": "<p>OK, that's great</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1266107": "Currently I integrated 5 models trained from 5-fold CV, by doing union (OR) calculation for 5 generated masks by 5 models. But my LB is always good enough (~0.91), even worse than single model.\n\nSo I am wondering whether my interation strategy is not good. I suppose there should be better ways to integrate prediction results from CV models.\n\nBTW, I used the resized images (4x) so I need to convert the probability to binary values for better interpolation. I also suspect that the problem happens in this step.",
    "1271511": "I'm about to start integrating more than 1 model, i havn't implemented yet but for what i read around the discussions, you will get a predicted probability for each pixel on the image, than you can average the probabilities and strictly determine a threshold to cut lower probs out.\nI believe that if you convert all masks to binary before averaging them, you will lose at least some part of the predictions to this. Am i wrong?\n\nThat is in theory what i will do, i let you know once i actually manage to do it.",
    "1274194": "OK, that's great"
  },
  "source": "meta"
}