{
  "id": 576668,
  "title": "Choice of training segments",
  "url": "/competitions/birdclef-2025/discussion/576668",
  "author_name": "",
  "post_date": "2025-05-06T14:57:29.874745Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I'm curious to know what is working best for choosing training segments, once you have excluded voice segments. For example:</p>\n<ol>\n<li>Use first and last segment in each recording.</li>\n<li>Use all non-overlapping segments in each recording.</li>\n<li>Overlap segments for under-represented classes.</li>\n<li>Train classifier, then use it to remove \"bad\" segments.</li>\n<li>Manually review and remove \"bad\" segments.</li>\n<li>Find matching segments in train_soundscapes and add to training.</li>\n</ol>",
  "messages": [
    {
      "id": "3195034",
      "postDate": "05/06/2025 14:57:29",
      "content": "<p>I'm curious to know what is working best for choosing training segments, once you have excluded voice segments. For example:</p>\n<ol>\n<li>Use first and last segment in each recording.</li>\n<li>Use all non-overlapping segments in each recording.</li>\n<li>Overlap segments for under-represented classes.</li>\n<li>Train classifier, then use it to remove \"bad\" segments.</li>\n<li>Manually review and remove \"bad\" segments.</li>\n<li>Find matching segments in train_soundscapes and add to training.</li>\n</ol>",
      "rawMarkdown": "I'm curious to know what is working best for choosing training segments, once you have excluded voice segments. For example:\n\n1. Use first and last segment in each recording.\n2. Use all non-overlapping segments in each recording.\n3. Overlap segments for under-represented classes.\n4. Train classifier, then use it to remove \"bad\" segments.\n5. Manually review and remove \"bad\" segments.\n6. Find matching segments in train_soundscapes and add to training.",
      "votes": null
    },
    {
      "id": "3195103",
      "postDate": "05/06/2025 16:27:22",
      "content": "<p>Just taking the middle worked for me the best, but when you look at the results of the classification report it kinda sucks </p>",
      "rawMarkdown": "Just taking the middle worked for me the best, but when you look at the results of the classification report it kinda sucks",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3195103,
      "author_name": "maxmelichov",
      "author_url": "",
      "post_date": "05/06/2025 16:27:22",
      "content": "<p>Just taking the middle worked for me the best, but when you look at the results of the classification report it kinda sucks </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3195034": "I'm curious to know what is working best for choosing training segments, once you have excluded voice segments. For example:\n\n1. Use first and last segment in each recording.\n2. Use all non-overlapping segments in each recording.\n3. Overlap segments for under-represented classes.\n4. Train classifier, then use it to remove \"bad\" segments.\n5. Manually review and remove \"bad\" segments.\n6. Find matching segments in train_soundscapes and add to training.",
    "3195103": "Just taking the middle worked for me the best, but when you look at the results of the classification report it kinda sucks"
  },
  "source": "meta"
}