{
  "id": 124045,
  "title": "Tuning Focal loss?",
  "url": "/competitions/pku-autonomous-driving/discussion/124045",
  "author_name": "",
  "post_date": "2020-01-01T15:50:42.910691300Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>what are the loss functions you tried? which one worked best for you?\nare you tuning focal loss? i mean are you using default apha and beta value for focal loss or modified one?</p>\n\n<p>i tried focal loss with no modification but for me focal loss didn't do well in validation,it keeps fluctuating and shows inconsistency.\nnote : i used 8% data for validation</p>",
  "messages": [
    {
      "id": "707863",
      "postDate": "01/01/2020 15:50:42",
      "content": "<p>what are the loss functions you tried? which one worked best for you?\nare you tuning focal loss? i mean are you using default apha and beta value for focal loss or modified one?</p>\n\n<p>i tried focal loss with no modification but for me focal loss didn't do well in validation,it keeps fluctuating and shows inconsistency.\nnote : i used 8% data for validation</p>",
      "rawMarkdown": "what are the loss functions you tried? which one worked best for you?\nare you tuning focal loss? i mean are you using default apha and beta value for focal loss or modified one?\n\ni tried focal loss with no modification but for me focal loss didn't do well in validation,it keeps fluctuating and shows inconsistency.\nnote : i used 8% data for validation",
      "votes": null
    },
    {
      "id": "708285",
      "postDate": "01/02/2020 06:37:14",
      "content": "<p>The default value works well. I tried to increase gamma from 2 to 4 and this fits positive objects better but causes ignorance to the negative objects, which is consistent with the description of the focal loss paper. Under this circumstance, the default threshold could be set higher to filter unwanted negative objects. However, no obvious improvements were achieved.</p>",
      "rawMarkdown": "The default value works well. I tried to increase gamma from 2 to 4 and this fits positive objects better but causes ignorance to the negative objects, which is consistent with the description of the focal loss paper. Under this circumstance, the default threshold could be set higher to filter unwanted negative objects. However, no obvious improvements were achieved.",
      "votes": null
    },
    {
      "id": "708310",
      "postDate": "01/02/2020 06:59:48",
      "content": "<p>thank you a lot mate for sharing your findings <a href=\"/syoya1997\">@syoya1997</a> </p>",
      "rawMarkdown": "thank you a lot mate for sharing your findings @syoya1997",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 708285,
      "author_name": "syoya1997",
      "author_url": "",
      "post_date": "01/02/2020 06:37:14",
      "content": "<p>The default value works well. I tried to increase gamma from 2 to 4 and this fits positive objects better but causes ignorance to the negative objects, which is consistent with the description of the focal loss paper. Under this circumstance, the default threshold could be set higher to filter unwanted negative objects. However, no obvious improvements were achieved.</p>",
      "votes": null,
      "replies": [
        {
          "id": 708310,
          "author_name": "mobassir",
          "author_url": "",
          "post_date": "01/02/2020 06:59:48",
          "content": "<p>thank you a lot mate for sharing your findings <a href=\"/syoya1997\">@syoya1997</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "707863": "what are the loss functions you tried? which one worked best for you?\nare you tuning focal loss? i mean are you using default apha and beta value for focal loss or modified one?\n\ni tried focal loss with no modification but for me focal loss didn't do well in validation,it keeps fluctuating and shows inconsistency.\nnote : i used 8% data for validation",
    "708285": "The default value works well. I tried to increase gamma from 2 to 4 and this fits positive objects better but causes ignorance to the negative objects, which is consistent with the description of the focal loss paper. Under this circumstance, the default threshold could be set higher to filter unwanted negative objects. However, no obvious improvements were achieved.",
    "708310": "thank you a lot mate for sharing your findings @syoya1997"
  },
  "source": "meta"
}