{
  "id": 237127,
  "title": "Imbalanced classes",
  "url": "/competitions/birdclef-2021/discussion/237127",
  "author_name": "",
  "post_date": "2021-05-07T09:23:48.551290Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I wonder how you've tackled the fact that in the short audio directory some classes are much more represented than others ? I went for class weighting in the loss function, but maybe that's not the best way to do it ?</p>",
  "messages": [
    {
      "id": "1296467",
      "postDate": "05/07/2021 09:23:48",
      "content": "<p>I wonder how you've tackled the fact that in the short audio directory some classes are much more represented than others ? I went for class weighting in the loss function, but maybe that's not the best way to do it ?</p>",
      "rawMarkdown": "I wonder how you've tackled the fact that in the short audio directory some classes are much more represented than others ? I went for class weighting in the loss function, but maybe that's not the best way to do it ?",
      "votes": null
    },
    {
      "id": "1296517",
      "postDate": "05/07/2021 10:04:36",
      "content": "<p>That's one of the challenges here indeed.  Another one is the weak labels, i.e. we don't know when bird calls happen  in the short train audio clips.</p>\n<p>Dealing well with these two will be key for good performance in this competition.</p>",
      "rawMarkdown": "That's one of the challenges here indeed.  Another one is the weak labels, i.e. we don't know when bird calls happen  in the short train audio clips.\n\nDealing well with these two will be key for good performance in this competition.",
      "votes": null
    },
    {
      "id": "1296549",
      "postDate": "05/07/2021 10:30:01",
      "content": "<p>I think the weak labels problem can be solved by ensemble methods, or at least it seems that in past competitions best submissions solved it like that.</p>",
      "rawMarkdown": "I think the weak labels problem can be solved by ensemble methods, or at least it seems that in past competitions best submissions solved it like that.",
      "votes": null
    },
    {
      "id": "1299167",
      "postDate": "05/09/2021 14:14:38",
      "content": "<p>Yeah! Weak labels is one of the major problem indeed. I saw few notebooks using random image from  5 or 6 images of one audio assuming there might be birdcall in that random image. </p>",
      "rawMarkdown": "Yeah! Weak labels is one of the major problem indeed. I saw few notebooks using random image from  5 or 6 images of one audio assuming there might be birdcall in that random image.",
      "votes": null
    },
    {
      "id": "1302529",
      "postDate": "05/11/2021 15:29:55",
      "content": "<p>Tried class weights and pos_weight of BCEWithLogitsLoss, neither works for me 😂</p>",
      "rawMarkdown": "Tried class weights and pos_weight of BCEWithLogitsLoss, neither works for me 😂",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1296517,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "05/07/2021 10:04:36",
      "content": "<p>That's one of the challenges here indeed.  Another one is the weak labels, i.e. we don't know when bird calls happen  in the short train audio clips.</p>\n<p>Dealing well with these two will be key for good performance in this competition.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1296549,
          "author_name": "josephamigo",
          "author_url": "",
          "post_date": "05/07/2021 10:30:01",
          "content": "<p>I think the weak labels problem can be solved by ensemble methods, or at least it seems that in past competitions best submissions solved it like that.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1299167,
          "author_name": "sarpal465",
          "author_url": "",
          "post_date": "05/09/2021 14:14:38",
          "content": "<p>Yeah! Weak labels is one of the major problem indeed. I saw few notebooks using random image from  5 or 6 images of one audio assuming there might be birdcall in that random image. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1302529,
      "author_name": "superchenhao",
      "author_url": "",
      "post_date": "05/11/2021 15:29:55",
      "content": "<p>Tried class weights and pos_weight of BCEWithLogitsLoss, neither works for me 😂</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1296467": "I wonder how you've tackled the fact that in the short audio directory some classes are much more represented than others ? I went for class weighting in the loss function, but maybe that's not the best way to do it ?",
    "1296517": "That's one of the challenges here indeed.  Another one is the weak labels, i.e. we don't know when bird calls happen  in the short train audio clips.\n\nDealing well with these two will be key for good performance in this competition.",
    "1296549": "I think the weak labels problem can be solved by ensemble methods, or at least it seems that in past competitions best submissions solved it like that.",
    "1299167": "Yeah! Weak labels is one of the major problem indeed. I saw few notebooks using random image from  5 or 6 images of one audio assuming there might be birdcall in that random image.",
    "1302529": "Tried class weights and pos_weight of BCEWithLogitsLoss, neither works for me 😂"
  },
  "source": "meta"
}