{
  "id": 64205,
  "title": "loss function",
  "url": "/competitions/airbus-ship-detection/discussion/64205",
  "author_name": "",
  "post_date": "2018-08-26T19:29:32.735123200Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I am just curious what loss function people use in the training process?\nI am using binary cross entropy as the loss function, but I feel it is not a good one. \nThanks!</p>",
  "messages": [
    {
      "id": "376078",
      "postDate": "08/26/2018 19:29:32",
      "content": "<p>I am just curious what loss function people use in the training process?\nI am using binary cross entropy as the loss function, but I feel it is not a good one. \nThanks!</p>",
      "rawMarkdown": "I am just curious what loss function people use in the training process?\nI am using binary cross entropy as the loss function, but I feel it is not a good one. \nThanks!",
      "votes": null
    },
    {
      "id": "376194",
      "postDate": "08/27/2018 01:28:41",
      "content": "<p>BCE should be fine. If you want a fancier loss function, check out this interview with the winners from the Carvana segmentation challenge. They basically used combinations of BCE and dice score.</p>\n\n<p><a href=\"http://blog.kaggle.com/2017/12/22/carvana-image-masking-first-place-interview/\">http://blog.kaggle.com/2017/12/22/carvana-image-masking-first-place-interview/</a></p>",
      "rawMarkdown": "BCE should be fine. If you want a fancier loss function, check out this interview with the winners from the Carvana segmentation challenge. They basically used combinations of BCE and dice score.\n\nhttp://blog.kaggle.com/2017/12/22/carvana-image-masking-first-place-interview/",
      "votes": null
    },
    {
      "id": "376780",
      "postDate": "08/28/2018 03:48:56",
      "content": "<p>Thanks!</p>",
      "rawMarkdown": "Thanks!",
      "votes": null
    },
    {
      "id": "380407",
      "postDate": "09/02/2018 15:19:34",
      "content": "<p>I used focal loss - log(soft dice loss), check my kernel <a href=\"https://www.kaggle.com/iafoss/unet34-dice-0-87/notebook\">https://www.kaggle.com/iafoss/unet34-dice-0-87/notebook</a> for more details.</p>",
      "rawMarkdown": "I used focal loss - log(soft dice loss), check my kernel https://www.kaggle.com/iafoss/unet34-dice-0-87/notebook for more details.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 376194,
      "author_name": "towardsentropy",
      "author_url": "",
      "post_date": "08/27/2018 01:28:41",
      "content": "<p>BCE should be fine. If you want a fancier loss function, check out this interview with the winners from the Carvana segmentation challenge. They basically used combinations of BCE and dice score.</p>\n\n<p><a href=\"http://blog.kaggle.com/2017/12/22/carvana-image-masking-first-place-interview/\">http://blog.kaggle.com/2017/12/22/carvana-image-masking-first-place-interview/</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 376780,
          "author_name": "zhengrui315",
          "author_url": "",
          "post_date": "08/28/2018 03:48:56",
          "content": "<p>Thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 380407,
      "author_name": "iafoss",
      "author_url": "",
      "post_date": "09/02/2018 15:19:34",
      "content": "<p>I used focal loss - log(soft dice loss), check my kernel <a href=\"https://www.kaggle.com/iafoss/unet34-dice-0-87/notebook\">https://www.kaggle.com/iafoss/unet34-dice-0-87/notebook</a> for more details.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "376078": "I am just curious what loss function people use in the training process?\nI am using binary cross entropy as the loss function, but I feel it is not a good one. \nThanks!",
    "376194": "BCE should be fine. If you want a fancier loss function, check out this interview with the winners from the Carvana segmentation challenge. They basically used combinations of BCE and dice score.\n\nhttp://blog.kaggle.com/2017/12/22/carvana-image-masking-first-place-interview/",
    "376780": "Thanks!",
    "380407": "I used focal loss - log(soft dice loss), check my kernel https://www.kaggle.com/iafoss/unet34-dice-0-87/notebook for more details."
  },
  "source": "meta"
}