{
  "id": 230864,
  "title": "Anyone successfully using custom loss?",
  "url": "/competitions/indoor-location-navigation/discussion/230864",
  "author_name": "higepon",
  "post_date": "2021-04-06T00:04:39.397000",
  "votes": 11,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi there,<br>\nI have been looking into raw predictions of my model (raw = w/o the popular post processing) using my notebook <a href=\"https://www.kaggle.com/higepon/visualize-submissions-with-post-processing\" target=\"_blank\">Visualize submissions with post processing</a>.<br>\nI clearly see some bad predictions there, which are either out of hallway or building. Because we have floor map I wonder if we can give models some penalties as loss for the bad predictions.</p>\n<p>Did anyone successfully make the idea work?</p>\n<p>I quickly researched and wondered something similar to epsilon insensitive loss might work. But I'm not very sure. </p>\n<p>I'd appreciate your thoughts here.</p>",
  "messages": [
    {
      "id": 1264159,
      "postDate": "2021-04-06T00:04:39.397Z",
      "content": "<p>Hi there,<br>\nI have been looking into raw predictions of my model (raw = w/o the popular post processing) using my notebook <a href=\"https://www.kaggle.com/higepon/visualize-submissions-with-post-processing\" target=\"_blank\">Visualize submissions with post processing</a>.<br>\nI clearly see some bad predictions there, which are either out of hallway or building. Because we have floor map I wonder if we can give models some penalties as loss for the bad predictions.</p>\n<p>Did anyone successfully make the idea work?</p>\n<p>I quickly researched and wondered something similar to epsilon insensitive loss might work. But I'm not very sure. </p>\n<p>I'd appreciate your thoughts here.</p>",
      "rawMarkdown": "Hi there,\nI have been looking into raw predictions of my model (raw = w/o the popular post processing) using my notebook [Visualize submissions with post processing](https://www.kaggle.com/higepon/visualize-submissions-with-post-processing).\nI clearly see some bad predictions there, which are either out of hallway or building. Because we have floor map I wonder if we can give models some penalties as loss for the bad predictions.\n\nDid anyone successfully make the idea work?\n\nI quickly researched and wondered something similar to epsilon insensitive loss might work. But I'm not very sure. \n\nI'd appreciate your thoughts here.",
      "votes": 11
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1264159": "Hi there,\nI have been looking into raw predictions of my model (raw = w/o the popular post processing) using my notebook [Visualize submissions with post processing](https://www.kaggle.com/higepon/visualize-submissions-with-post-processing).\nI clearly see some bad predictions there, which are either out of hallway or building. Because we have floor map I wonder if we can give models some penalties as loss for the bad predictions.\n\nDid anyone successfully make the idea work?\n\nI quickly researched and wondered something similar to epsilon insensitive loss might work. But I'm not very sure. \n\nI'd appreciate your thoughts here."
  }
}