{
  "id": 114611,
  "title": "Optimisation of Threshold and Min Size doesn't work",
  "url": "/competitions/understanding_cloud_organization/discussion/114611",
  "author_name": "",
  "post_date": "2019-10-27T23:47:05.588129200Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi,</p>\n\n<p>I am using efficienet as backbone and is trying to optimize threshold and min size by calculating the dice score of different combinations of these two parameters with validation set. However, it doesn't seem to work. dice is just growing when threshold and min size increases. Did any one encounter a similar issue?</p>",
  "messages": [
    {
      "id": "659577",
      "postDate": "10/27/2019 23:47:05",
      "content": "<p>Hi,</p>\n\n<p>I am using efficienet as backbone and is trying to optimize threshold and min size by calculating the dice score of different combinations of these two parameters with validation set. However, it doesn't seem to work. dice is just growing when threshold and min size increases. Did any one encounter a similar issue?</p>",
      "rawMarkdown": "Hi,\n\nI am using efficienet as backbone and is trying to optimize threshold and min size by calculating the dice score of different combinations of these two parameters with validation set. However, it doesn't seem to work. dice is just growing when threshold and min size increases. Did any one encounter a similar issue?",
      "votes": null
    },
    {
      "id": "659582",
      "postDate": "10/28/2019 00:06:45",
      "content": "<p>Keep tuning the pose-processing might not be a good idea, since this just like fitting to the LB. </p>",
      "rawMarkdown": "Keep tuning the pose-processing might not be a good idea, since this just like fitting to the LB.",
      "votes": null
    },
    {
      "id": "660214",
      "postDate": "10/28/2019 22:01:38",
      "content": "<p>Remover helps by detecting FP. If it is not working, you need to figure out:</p>\n\n<ol>\n<li><p>What is the precision and recall of your remover？</p></li>\n<li><p>How many masks modified by the remover are FP/TP. </p></li>\n</ol>",
      "rawMarkdown": "Remover helps by detecting FP. If it is not working, you need to figure out:\n\n1. What is the precision and recall of your remover？\n\n2. How many masks modified by the remover are FP/TP.",
      "votes": null
    },
    {
      "id": "660380",
      "postDate": "10/29/2019 03:56:40",
      "content": "<p>By remover, do you mean the classifier? I'm not confused about the classifier but the threshold and minsize when converting unet prediction to mask and submission.</p>",
      "rawMarkdown": "By remover, do you mean the classifier? I'm not confused about the classifier but the threshold and minsize when converting unet prediction to mask and submission.",
      "votes": null
    },
    {
      "id": "660401",
      "postDate": "10/29/2019 04:32:10",
      "content": "<p>Remover is a general idea. </p>\n\n<p>Classifier is a trained remover, while threshold &amp; minsize are rule-based remover. </p>",
      "rawMarkdown": "Remover is a general idea. \n\nClassifier is a trained remover, while threshold &amp; minsize are rule-based remover.",
      "votes": null
    },
    {
      "id": "665153",
      "postDate": "11/04/2019 17:20:39",
      "content": "<p>So should I use things like PR or ROC to find the best thershold instead of dice?</p>",
      "rawMarkdown": "So should I use things like PR or ROC to find the best thershold instead of dice?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 659582,
      "author_name": "xiejialun",
      "author_url": "",
      "post_date": "10/28/2019 00:06:45",
      "content": "<p>Keep tuning the pose-processing might not be a good idea, since this just like fitting to the LB. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 660214,
      "author_name": "naivelamb",
      "author_url": "",
      "post_date": "10/28/2019 22:01:38",
      "content": "<p>Remover helps by detecting FP. If it is not working, you need to figure out:</p>\n\n<ol>\n<li><p>What is the precision and recall of your remover？</p></li>\n<li><p>How many masks modified by the remover are FP/TP. </p></li>\n</ol>",
      "votes": null,
      "replies": [
        {
          "id": 660380,
          "author_name": "tonychenxyz",
          "author_url": "",
          "post_date": "10/29/2019 03:56:40",
          "content": "<p>By remover, do you mean the classifier? I'm not confused about the classifier but the threshold and minsize when converting unet prediction to mask and submission.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 660401,
          "author_name": "naivelamb",
          "author_url": "",
          "post_date": "10/29/2019 04:32:10",
          "content": "<p>Remover is a general idea. </p>\n\n<p>Classifier is a trained remover, while threshold &amp; minsize are rule-based remover. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 665153,
          "author_name": "tonychenxyz",
          "author_url": "",
          "post_date": "11/04/2019 17:20:39",
          "content": "<p>So should I use things like PR or ROC to find the best thershold instead of dice?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "659577": "Hi,\n\nI am using efficienet as backbone and is trying to optimize threshold and min size by calculating the dice score of different combinations of these two parameters with validation set. However, it doesn't seem to work. dice is just growing when threshold and min size increases. Did any one encounter a similar issue?",
    "659582": "Keep tuning the pose-processing might not be a good idea, since this just like fitting to the LB.",
    "660214": "Remover helps by detecting FP. If it is not working, you need to figure out:\n\n1. What is the precision and recall of your remover？\n\n2. How many masks modified by the remover are FP/TP.",
    "660380": "By remover, do you mean the classifier? I'm not confused about the classifier but the threshold and minsize when converting unet prediction to mask and submission.",
    "660401": "Remover is a general idea. \n\nClassifier is a trained remover, while threshold &amp; minsize are rule-based remover.",
    "665153": "So should I use things like PR or ROC to find the best thershold instead of dice?"
  },
  "source": "meta"
}