{
  "id": 216044,
  "title": "Bi-Tempered Logistic Loss Vs Categorical Cross Entropy with Label Smoothening",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/216044",
  "author_name": "",
  "post_date": "2021-02-01T09:19:57.460683600Z",
  "votes": null,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I have this confusion is the usage of Loss Function Bi-Tempered Logistic Loss or using Categorical Cross Entropy with Label Smoothening are simillar approach.</p>\n<p>Can anyone guide me on this, or direct me to some reading material.</p>\n<p>Regards</p>",
  "messages": [
    {
      "id": "1180472",
      "postDate": "02/01/2021 09:19:57",
      "content": "<p>I have this confusion is the usage of Loss Function Bi-Tempered Logistic Loss or using Categorical Cross Entropy with Label Smoothening are simillar approach.</p>\n<p>Can anyone guide me on this, or direct me to some reading material.</p>\n<p>Regards</p>",
      "rawMarkdown": "I have this confusion is the usage of Loss Function Bi-Tempered Logistic Loss or using Categorical Cross Entropy with Label Smoothening are simillar approach.\n\nCan anyone guide me on this, or direct me to some reading material.\n\nRegards",
      "votes": null
    },
    {
      "id": "1180516",
      "postDate": "02/01/2021 10:00:31",
      "content": "<p>Actually you can use label smoothing with Bi-Tempered Logistic Loss too and there are discussions and notebooks about it.<br>\nIf you type in \"tempered\" either into discussion search or notebooks search you will find tons of them magically.</p>\n<p>But if you want some links directly, here you go:<br>\n<a href=\"https://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html\" target=\"_blank\">https://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html</a></p>",
      "rawMarkdown": "Actually you can use label smoothing with Bi-Tempered Logistic Loss too and there are discussions and notebooks about it.\nIf you type in \"tempered\" either into discussion search or notebooks search you will find tons of them magically.\n\n\nBut if you want some links directly, here you go:\nhttps://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html",
      "votes": null
    },
    {
      "id": "1183220",
      "postDate": "02/02/2021 20:29:54",
      "content": "<p>For me <code>bi_tempered_loss</code> with label_smoothing of <code>0.2</code> worked the best.</p>",
      "rawMarkdown": "For me `bi_tempered_loss` with label_smoothing of `0.2` worked the best.",
      "votes": null
    },
    {
      "id": "1185486",
      "postDate": "02/04/2021 07:12:06",
      "content": "<p>Hi, what values of t1 and t2 in bitempered loss are giving good results?</p>",
      "rawMarkdown": "Hi, what values of t1 and t2 in bitempered loss are giving good results?",
      "votes": null
    },
    {
      "id": "1186683",
      "postDate": "02/05/2021 01:47:22",
      "content": "<p>i have landed on 0.6 and 1.6 or 1.8, and label smoothing of 0.2.<br>\nthose were in a notebook i copied some weeks ago, and they are the best combo i have found after some brief experimentation.</p>",
      "rawMarkdown": "i have landed on 0.6 and 1.6 or 1.8, and label smoothing of 0.2.\nthose were in a notebook i copied some weeks ago, and they are the best combo i have found after some brief experimentation.",
      "votes": null
    },
    {
      "id": "1187052",
      "postDate": "02/05/2021 07:37:35",
      "content": "<p>Thanks, will try these values :)</p>",
      "rawMarkdown": "Thanks, will try these values :)",
      "votes": null
    },
    {
      "id": "1187930",
      "postDate": "02/05/2021 19:04:06",
      "content": "<p>There is a Demo for Bi-Tempered Loss where you can set some values and choose the noise, maybe it can help to find a direction:</p>\n<p><a href=\"https://google.github.io/bi-tempered-loss/\" target=\"_blank\">https://google.github.io/bi-tempered-loss/</a></p>",
      "rawMarkdown": "There is a Demo for Bi-Tempered Loss where you can set some values and choose the noise, maybe it can help to find a direction:\n\nhttps://google.github.io/bi-tempered-loss/",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1180516,
      "author_name": "killimi",
      "author_url": "",
      "post_date": "02/01/2021 10:00:31",
      "content": "<p>Actually you can use label smoothing with Bi-Tempered Logistic Loss too and there are discussions and notebooks about it.<br>\nIf you type in \"tempered\" either into discussion search or notebooks search you will find tons of them magically.</p>\n<p>But if you want some links directly, here you go:<br>\n<a href=\"https://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html\" target=\"_blank\">https://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1183220,
      "author_name": "aliabdin1",
      "author_url": "",
      "post_date": "02/02/2021 20:29:54",
      "content": "<p>For me <code>bi_tempered_loss</code> with label_smoothing of <code>0.2</code> worked the best.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1185486,
      "author_name": "sakshamaggarwal",
      "author_url": "",
      "post_date": "02/04/2021 07:12:06",
      "content": "<p>Hi, what values of t1 and t2 in bitempered loss are giving good results?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1186683,
          "author_name": "daveccampbell",
          "author_url": "",
          "post_date": "02/05/2021 01:47:22",
          "content": "<p>i have landed on 0.6 and 1.6 or 1.8, and label smoothing of 0.2.<br>\nthose were in a notebook i copied some weeks ago, and they are the best combo i have found after some brief experimentation.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1187052,
          "author_name": "sakshamaggarwal",
          "author_url": "",
          "post_date": "02/05/2021 07:37:35",
          "content": "<p>Thanks, will try these values :)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1187930,
          "author_name": "aliabdin1",
          "author_url": "",
          "post_date": "02/05/2021 19:04:06",
          "content": "<p>There is a Demo for Bi-Tempered Loss where you can set some values and choose the noise, maybe it can help to find a direction:</p>\n<p><a href=\"https://google.github.io/bi-tempered-loss/\" target=\"_blank\">https://google.github.io/bi-tempered-loss/</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1180472": "I have this confusion is the usage of Loss Function Bi-Tempered Logistic Loss or using Categorical Cross Entropy with Label Smoothening are simillar approach.\n\nCan anyone guide me on this, or direct me to some reading material.\n\nRegards",
    "1180516": "Actually you can use label smoothing with Bi-Tempered Logistic Loss too and there are discussions and notebooks about it.\nIf you type in \"tempered\" either into discussion search or notebooks search you will find tons of them magically.\n\n\nBut if you want some links directly, here you go:\nhttps://ai.googleblog.com/2019/08/bi-tempered-logistic-loss-for-training.html",
    "1183220": "For me `bi_tempered_loss` with label_smoothing of `0.2` worked the best.",
    "1185486": "Hi, what values of t1 and t2 in bitempered loss are giving good results?",
    "1186683": "i have landed on 0.6 and 1.6 or 1.8, and label smoothing of 0.2.\nthose were in a notebook i copied some weeks ago, and they are the best combo i have found after some brief experimentation.",
    "1187052": "Thanks, will try these values :)",
    "1187930": "There is a Demo for Bi-Tempered Loss where you can set some values and choose the noise, maybe it can help to find a direction:\n\nhttps://google.github.io/bi-tempered-loss/"
  },
  "source": "meta"
}