{
  "id": 492973,
  "title": "Choice of Loss Function",
  "url": "/competitions/birdclef-2024/discussion/492973",
  "author_name": "",
  "post_date": "2024-04-11T16:07:24.588955300Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>In previous years, I have seen multilabel losses used. However, this year it seems like multiclass is being favored more. </p>\n<p>What is everyone using this year and why?</p>\n<p>I'm currently using torch<code>nn.CrossEntropy</code> which is multiclass, but I used Focal in previous years. I don't really have reasoning other than that's what I saw others using. It seems like you could formulate the problem either way.</p>",
  "messages": [
    {
      "id": "2747038",
      "postDate": "04/11/2024 16:07:24",
      "content": "<p>In previous years, I have seen multilabel losses used. However, this year it seems like multiclass is being favored more. </p>\n<p>What is everyone using this year and why?</p>\n<p>I'm currently using torch<code>nn.CrossEntropy</code> which is multiclass, but I used Focal in previous years. I don't really have reasoning other than that's what I saw others using. It seems like you could formulate the problem either way.</p>",
      "rawMarkdown": "In previous years, I have seen multilabel losses used. However, this year it seems like multiclass is being favored more. \n\nWhat is everyone using this year and why?\n\nI'm currently using torch`nn.CrossEntropy` which is multiclass, but I used Focal in previous years. I don't really have reasoning other than that's what I saw others using. It seems like you could formulate the problem either way.",
      "votes": null
    },
    {
      "id": "2747059",
      "postDate": "04/11/2024 16:34:59",
      "content": "<p>Sometimes different loss function is more favored by the evaluation metrics.  I do not know which one is best to use </p>",
      "rawMarkdown": "Sometimes different loss function is more favored by the evaluation metrics.  I do not know which one is best to use",
      "votes": null
    },
    {
      "id": "2747096",
      "postDate": "04/11/2024 17:15:21",
      "content": "<p>What I like about BCE in contrast to CE is that it's a sigmoid layer and not a softmax, this means you can train your model on nocall audios, and you don't have to create an extra nocall class, just make a all 0s label.</p>",
      "rawMarkdown": "What I like about BCE in contrast to CE is that it's a sigmoid layer and not a softmax, this means you can train your model on nocall audios, and you don't have to create an extra nocall class, just make a all 0s label.",
      "votes": null
    },
    {
      "id": "2747261",
      "postDate": "04/11/2024 18:55:52",
      "content": "<p>I'm favoring focal losses, as I suspect they may better disambiguate near-target species. But I have no concrete results from this year's competition to support that and will very likely try others. </p>",
      "rawMarkdown": "I'm favoring focal losses, as I suspect they may better disambiguate near-target species. But I have no concrete results from this year's competition to support that and will very likely try others.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2747059,
      "author_name": "nyleve",
      "author_url": "",
      "post_date": "04/11/2024 16:34:59",
      "content": "<p>Sometimes different loss function is more favored by the evaluation metrics.  I do not know which one is best to use </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2747096,
      "author_name": "janmpia",
      "author_url": "",
      "post_date": "04/11/2024 17:15:21",
      "content": "<p>What I like about BCE in contrast to CE is that it's a sigmoid layer and not a softmax, this means you can train your model on nocall audios, and you don't have to create an extra nocall class, just make a all 0s label.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2747261,
      "author_name": "lobrien",
      "author_url": "",
      "post_date": "04/11/2024 18:55:52",
      "content": "<p>I'm favoring focal losses, as I suspect they may better disambiguate near-target species. But I have no concrete results from this year's competition to support that and will very likely try others. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2747038": "In previous years, I have seen multilabel losses used. However, this year it seems like multiclass is being favored more. \n\nWhat is everyone using this year and why?\n\nI'm currently using torch`nn.CrossEntropy` which is multiclass, but I used Focal in previous years. I don't really have reasoning other than that's what I saw others using. It seems like you could formulate the problem either way.",
    "2747059": "Sometimes different loss function is more favored by the evaluation metrics.  I do not know which one is best to use",
    "2747096": "What I like about BCE in contrast to CE is that it's a sigmoid layer and not a softmax, this means you can train your model on nocall audios, and you don't have to create an extra nocall class, just make a all 0s label.",
    "2747261": "I'm favoring focal losses, as I suspect they may better disambiguate near-target species. But I have no concrete results from this year's competition to support that and will very likely try others."
  },
  "source": "meta"
}