{
  "id": 497530,
  "title": "Train Sequence Length vs Model Performance",
  "url": "/competitions/birdclef-2024/discussion/497530",
  "author_name": "",
  "post_date": "2024-04-25T00:18:48.260919200Z",
  "votes": 3,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Is anyone getting improved performance from training with longer sequences?</p>\n<p>I would think that longer sequences would increase the chances of capturing in the label bird in the audio segment. However, I have been trying longer sequences and it does not seem to improve LB or CV. Is anyone else seeing this same behavior? </p>",
  "messages": [
    {
      "id": "2773881",
      "postDate": "04/25/2024 00:18:48",
      "content": "<p>Is anyone getting improved performance from training with longer sequences?</p>\n<p>I would think that longer sequences would increase the chances of capturing in the label bird in the audio segment. However, I have been trying longer sequences and it does not seem to improve LB or CV. Is anyone else seeing this same behavior? </p>",
      "rawMarkdown": "Is anyone getting improved performance from training with longer sequences?\n\nI would think that longer sequences would increase the chances of capturing in the label bird in the audio segment. However, I have been trying longer sequences and it does not seem to improve LB or CV. Is anyone else seeing this same behavior?",
      "votes": null
    },
    {
      "id": "2777129",
      "postDate": "04/26/2024 14:32:26",
      "content": "<p>I have a hypothesis. During the annotation process I have my doubts that the curators listened to the entire audio recording. I imagine the process is along the lines of listen to this recording and label the bird you hear. If the bird can be heard in the first couple seconds of the recording then it gets the label.</p>",
      "rawMarkdown": "I have a hypothesis. During the annotation process I have my doubts that the curators listened to the entire audio recording. I imagine the process is along the lines of listen to this recording and label the bird you hear. If the bird can be heard in the first couple seconds of the recording then it gets the label.",
      "votes": null
    },
    {
      "id": "2777281",
      "postDate": "04/26/2024 15:30:46",
      "content": "<p>You can see the assumption made here by our grandmaster:<br>\n<a href=\"https://www.kaggle.com/competitions/birdsong-recognition/discussion/183219\" target=\"_blank\">https://www.kaggle.com/competitions/birdsong-recognition/discussion/183219</a><br>\nsection \"Training data clips\"</p>",
      "rawMarkdown": "You can see the assumption made here by our grandmaster:\nhttps://www.kaggle.com/competitions/birdsong-recognition/discussion/183219\nsection \"Training data clips\"",
      "votes": null
    },
    {
      "id": "2777409",
      "postDate": "04/26/2024 16:25:23",
      "content": "<p>Thanks for the citation!</p>",
      "rawMarkdown": "Thanks for the citation!",
      "votes": null
    },
    {
      "id": "2777606",
      "postDate": "04/26/2024 18:24:15",
      "content": "<p>oh nice. thanks</p>",
      "rawMarkdown": "oh nice. thanks",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2777129,
      "author_name": "willrice",
      "author_url": "",
      "post_date": "04/26/2024 14:32:26",
      "content": "<p>I have a hypothesis. During the annotation process I have my doubts that the curators listened to the entire audio recording. I imagine the process is along the lines of listen to this recording and label the bird you hear. If the bird can be heard in the first couple seconds of the recording then it gets the label.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2777281,
          "author_name": "nyleve",
          "author_url": "",
          "post_date": "04/26/2024 15:30:46",
          "content": "<p>You can see the assumption made here by our grandmaster:<br>\n<a href=\"https://www.kaggle.com/competitions/birdsong-recognition/discussion/183219\" target=\"_blank\">https://www.kaggle.com/competitions/birdsong-recognition/discussion/183219</a><br>\nsection \"Training data clips\"</p>",
          "votes": null,
          "replies": [
            {
              "id": 2777409,
              "author_name": "cpmpml",
              "author_url": "",
              "post_date": "04/26/2024 16:25:23",
              "content": "<p>Thanks for the citation!</p>",
              "votes": null,
              "replies": []
            },
            {
              "id": 2777606,
              "author_name": "willrice",
              "author_url": "",
              "post_date": "04/26/2024 18:24:15",
              "content": "<p>oh nice. thanks</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2773881": "Is anyone getting improved performance from training with longer sequences?\n\nI would think that longer sequences would increase the chances of capturing in the label bird in the audio segment. However, I have been trying longer sequences and it does not seem to improve LB or CV. Is anyone else seeing this same behavior?",
    "2777129": "I have a hypothesis. During the annotation process I have my doubts that the curators listened to the entire audio recording. I imagine the process is along the lines of listen to this recording and label the bird you hear. If the bird can be heard in the first couple seconds of the recording then it gets the label.",
    "2777281": "You can see the assumption made here by our grandmaster:\nhttps://www.kaggle.com/competitions/birdsong-recognition/discussion/183219\nsection \"Training data clips\"",
    "2777409": "Thanks for the citation!",
    "2777606": "oh nice. thanks"
  },
  "source": "meta"
}