{
  "id": 502162,
  "title": "Any tried any Domain Adaption techniques?",
  "url": "/competitions/birdclef-2024/discussion/502162",
  "author_name": "",
  "post_date": "2024-05-12T11:35:22.103646600Z",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Pretty much the title. <br>\nI've been trying to do some type of Augs with unlabelled soundscapes to introduce some kind of unseen distribution to NN. </p>\n<p>Still deciding if I should investigate more into Domain Adoption or just wasting my time.</p>",
  "messages": [
    {
      "id": "2808739",
      "postDate": "05/12/2024 11:35:22",
      "content": "<p>Pretty much the title. <br>\nI've been trying to do some type of Augs with unlabelled soundscapes to introduce some kind of unseen distribution to NN. </p>\n<p>Still deciding if I should investigate more into Domain Adoption or just wasting my time.</p>",
      "rawMarkdown": "Pretty much the title. \nI've been trying to do some type of Augs with unlabelled soundscapes to introduce some kind of unseen distribution to NN. \n\nStill deciding if I should investigate more into Domain Adoption or just wasting my time.",
      "votes": null
    },
    {
      "id": "2809385",
      "postDate": "05/12/2024 17:33:18",
      "content": "<p>Hi, Salman!</p>\n<p>We are really interested in seeing a diversity of techniques! Notice that there's still a lot of head-room in the leader board, which suggests that there may be some real value in broader exploration. (and, as you've noticed, there's a bunch of unlabeled data just begging for innovative uses…)</p>\n<p>We also have a prize for best write-ups for the BirdCLEF conference. Past winners of this prize have included quality write-ups of interesting approaches which didn't end up placing high in the final leaderboard, [<a href=\"https://arxiv.org/abs/2306.16760](like\" target=\"_blank\">https://arxiv.org/abs/2306.16760](like</a> this one) from last year.</p>",
      "rawMarkdown": "Hi, Salman!\n\nWe are really interested in seeing a diversity of techniques! Notice that there's still a lot of head-room in the leader board, which suggests that there may be some real value in broader exploration. (and, as you've noticed, there's a bunch of unlabeled data just begging for innovative uses...)\n\nWe also have a prize for best write-ups for the BirdCLEF conference. Past winners of this prize have included quality write-ups of interesting approaches which didn't end up placing high in the final leaderboard, [https://arxiv.org/abs/2306.16760](like this one) from last year.",
      "votes": null
    },
    {
      "id": "2811939",
      "postDate": "05/14/2024 02:38:34",
      "content": "<p>I'm no expert, but I feel that the unlabelled soundscape will be the key point for this year (we never had this much in previous series).<br>\nI have not had the time to pretrain on them, but I think it's a good direction. I casually asked the owner of this repo from NTT lab  <a href=\"https://github.com/nttcslab/m2d\" target=\"_blank\">https://github.com/nttcslab/m2d</a>  , and his opinion is that it's worth trying (he even provided a 32khz checkpoint).<br>\nThe strict 2hr inference time on CPU is another problem to solve though. <br>\nAnyway, good luck !! </p>",
      "rawMarkdown": "I'm no expert, but I feel that the unlabelled soundscape will be the key point for this year (we never had this much in previous series).\nI have not had the time to pretrain on them, but I think it's a good direction. I casually asked the owner of this repo from NTT lab  https://github.com/nttcslab/m2d  , and his opinion is that it's worth trying (he even provided a 32khz checkpoint).\nThe strict 2hr inference time on CPU is another problem to solve though. \nAnyway, good luck !!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2809385,
      "author_name": "tomdenton",
      "author_url": "",
      "post_date": "05/12/2024 17:33:18",
      "content": "<p>Hi, Salman!</p>\n<p>We are really interested in seeing a diversity of techniques! Notice that there's still a lot of head-room in the leader board, which suggests that there may be some real value in broader exploration. (and, as you've noticed, there's a bunch of unlabeled data just begging for innovative uses…)</p>\n<p>We also have a prize for best write-ups for the BirdCLEF conference. Past winners of this prize have included quality write-ups of interesting approaches which didn't end up placing high in the final leaderboard, [<a href=\"https://arxiv.org/abs/2306.16760](like\" target=\"_blank\">https://arxiv.org/abs/2306.16760](like</a> this one) from last year.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2811939,
      "author_name": "nyleve",
      "author_url": "",
      "post_date": "05/14/2024 02:38:34",
      "content": "<p>I'm no expert, but I feel that the unlabelled soundscape will be the key point for this year (we never had this much in previous series).<br>\nI have not had the time to pretrain on them, but I think it's a good direction. I casually asked the owner of this repo from NTT lab  <a href=\"https://github.com/nttcslab/m2d\" target=\"_blank\">https://github.com/nttcslab/m2d</a>  , and his opinion is that it's worth trying (he even provided a 32khz checkpoint).<br>\nThe strict 2hr inference time on CPU is another problem to solve though. <br>\nAnyway, good luck !! </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2808739": "Pretty much the title. \nI've been trying to do some type of Augs with unlabelled soundscapes to introduce some kind of unseen distribution to NN. \n\nStill deciding if I should investigate more into Domain Adoption or just wasting my time.",
    "2809385": "Hi, Salman!\n\nWe are really interested in seeing a diversity of techniques! Notice that there's still a lot of head-room in the leader board, which suggests that there may be some real value in broader exploration. (and, as you've noticed, there's a bunch of unlabeled data just begging for innovative uses...)\n\nWe also have a prize for best write-ups for the BirdCLEF conference. Past winners of this prize have included quality write-ups of interesting approaches which didn't end up placing high in the final leaderboard, [https://arxiv.org/abs/2306.16760](like this one) from last year.",
    "2811939": "I'm no expert, but I feel that the unlabelled soundscape will be the key point for this year (we never had this much in previous series).\nI have not had the time to pretrain on them, but I think it's a good direction. I casually asked the owner of this repo from NTT lab  https://github.com/nttcslab/m2d  , and his opinion is that it's worth trying (he even provided a 32khz checkpoint).\nThe strict 2hr inference time on CPU is another problem to solve though. \nAnyway, good luck !!"
  },
  "source": "meta"
}