{
  "id": 71648,
  "title": "How about your thresholds?",
  "url": "/competitions/human-protein-atlas-image-classification/discussion/71648",
  "author_name": "",
  "post_date": "2018-11-15T11:01:34.926645Z",
  "votes": 9,
  "comment_count": 7,
  "views": 0,
  "content": "<p>I'm using thresholds: 0.5 for class [0], 0.2 for classes [8,9,10,15,20,27] and 0.4 for the rest. It's better than fine-tuning thresholds obtained from my validation set. How about yours?</p>",
  "messages": [
    {
      "id": "421741",
      "postDate": "11/15/2018 11:01:34",
      "content": "<p>I'm using thresholds: 0.5 for class [0], 0.2 for classes [8,9,10,15,20,27] and 0.4 for the rest. It's better than fine-tuning thresholds obtained from my validation set. How about yours?</p>",
      "rawMarkdown": "I'm using thresholds: 0.5 for class [0], 0.2 for classes [8,9,10,15,20,27] and 0.4 for the rest. It's better than fine-tuning thresholds obtained from my validation set. How about yours?",
      "votes": null
    },
    {
      "id": "421978",
      "postDate": "11/15/2018 16:19:41",
      "content": "<p>I fine tune the thresholds first, then I reduce the thresholds and try again on images that had no predictions.</p>",
      "rawMarkdown": "I fine tune the thresholds first, then I reduce the thresholds and try again on images that had no predictions.",
      "votes": null
    },
    {
      "id": "422237",
      "postDate": "11/16/2018 00:07:58",
      "content": "<p>Mau: your method seems reasonable.</p>",
      "rawMarkdown": "Mau: your method seems reasonable.",
      "votes": null
    },
    {
      "id": "422325",
      "postDate": "11/16/2018 04:19:52",
      "content": "<p>Oh. That is interesting, maybe I will try it</p>",
      "rawMarkdown": "Oh. That is interesting, maybe I will try it",
      "votes": null
    },
    {
      "id": "422461",
      "postDate": "11/16/2018 08:54:11",
      "content": "<p>hello,  how do you set these thresholds,  what is the logistic reason? </p>",
      "rawMarkdown": "hello,  how do you set these thresholds,  what is the logistic reason?",
      "votes": null
    },
    {
      "id": "422773",
      "postDate": "11/16/2018 19:06:05",
      "content": "<p>As a single threshold ~0.4 is the best for unclear reason. Class 0 is the most popular so it makes sense to unbias it a bit by increasing it, same for the other classes which are minorities. </p>",
      "rawMarkdown": "As a single threshold ~0.4 is the best for unclear reason. Class 0 is the most popular so it makes sense to unbias it a bit by increasing it, same for the other classes which are minorities.",
      "votes": null
    },
    {
      "id": "422775",
      "postDate": "11/16/2018 19:09:21",
      "content": "<p>@Brian Did you finetune your thresholds on a single fold or with CV + averaging? Does it correlate well with public LB result?</p>",
      "rawMarkdown": "Brian Did you finetune your thresholds on a single fold or with CV + averaging? Does it correlate well with public LB result?",
      "votes": null
    },
    {
      "id": "422854",
      "postDate": "11/16/2018 22:19:02",
      "content": "<p>For the thresholds, I used the entire training set. During training I am only doing a single fold, 20% validation split.</p>",
      "rawMarkdown": "For the thresholds, I used the entire training set. During training I am only doing a single fold, 20% validation split.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 421978,
      "author_name": "ldm314",
      "author_url": "",
      "post_date": "11/15/2018 16:19:41",
      "content": "<p>I fine tune the thresholds first, then I reduce the thresholds and try again on images that had no predictions.</p>",
      "votes": null,
      "replies": [
        {
          "id": 422325,
          "author_name": "maudung164",
          "author_url": "",
          "post_date": "11/16/2018 04:19:52",
          "content": "<p>Oh. That is interesting, maybe I will try it</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 422775,
          "author_name": "suicaokhoailang",
          "author_url": "",
          "post_date": "11/16/2018 19:09:21",
          "content": "<p>@Brian Did you finetune your thresholds on a single fold or with CV + averaging? Does it correlate well with public LB result?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 422854,
          "author_name": "ldm314",
          "author_url": "",
          "post_date": "11/16/2018 22:19:02",
          "content": "<p>For the thresholds, I used the entire training set. During training I am only doing a single fold, 20% validation split.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 422237,
      "author_name": "petewills",
      "author_url": "",
      "post_date": "11/16/2018 00:07:58",
      "content": "<p>Mau: your method seems reasonable.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 422461,
      "author_name": "wchzhhxj",
      "author_url": "",
      "post_date": "11/16/2018 08:54:11",
      "content": "<p>hello,  how do you set these thresholds,  what is the logistic reason? </p>",
      "votes": null,
      "replies": [
        {
          "id": 422773,
          "author_name": "suicaokhoailang",
          "author_url": "",
          "post_date": "11/16/2018 19:06:05",
          "content": "<p>As a single threshold ~0.4 is the best for unclear reason. Class 0 is the most popular so it makes sense to unbias it a bit by increasing it, same for the other classes which are minorities. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "421741": "I'm using thresholds: 0.5 for class [0], 0.2 for classes [8,9,10,15,20,27] and 0.4 for the rest. It's better than fine-tuning thresholds obtained from my validation set. How about yours?",
    "421978": "I fine tune the thresholds first, then I reduce the thresholds and try again on images that had no predictions.",
    "422237": "Mau: your method seems reasonable.",
    "422325": "Oh. That is interesting, maybe I will try it",
    "422461": "hello,  how do you set these thresholds,  what is the logistic reason?",
    "422773": "As a single threshold ~0.4 is the best for unclear reason. Class 0 is the most popular so it makes sense to unbias it a bit by increasing it, same for the other classes which are minorities.",
    "422775": "Brian Did you finetune your thresholds on a single fold or with CV + averaging? Does it correlate well with public LB result?",
    "422854": "For the thresholds, I used the entire training set. During training I am only doing a single fold, 20% validation split."
  },
  "source": "meta"
}