{
  "id": 506197,
  "title": "Is SpecAugment still useful this year?",
  "url": "/competitions/birdclef-2024/discussion/506197",
  "author_name": "Olly Powell",
  "post_date": "2024-05-20T21:51:09.447000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>One curiosity for me this year is that I can't seem to get Spec Augment to do anything useful.   It slows down my training as expected, but otherwise doesn't seem to help the score, or sends it backwards.  Is anybody else having much luck with SpecAugment?  Or just time or frequency masking?  Would you like to share your training parameters?  Or theories as to why it would/wouldn't be useful given this years training metric?</p>",
  "messages": [
    {
      "id": 2826369,
      "postDate": "2024-05-20T21:51:09.447Z",
      "content": "<p>One curiosity for me this year is that I can't seem to get Spec Augment to do anything useful.   It slows down my training as expected, but otherwise doesn't seem to help the score, or sends it backwards.  Is anybody else having much luck with SpecAugment?  Or just time or frequency masking?  Would you like to share your training parameters?  Or theories as to why it would/wouldn't be useful given this years training metric?</p>",
      "rawMarkdown": "One curiosity for me this year is that I can't seem to get Spec Augment to do anything useful.   It slows down my training as expected, but otherwise doesn't seem to help the score, or sends it backwards.  Is anybody else having much luck with SpecAugment?  Or just time or frequency masking?  Would you like to share your training parameters?  Or theories as to why it would/wouldn't be useful given this years training metric?\n\n",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2826369": "One curiosity for me this year is that I can't seem to get Spec Augment to do anything useful.   It slows down my training as expected, but otherwise doesn't seem to help the score, or sends it backwards.  Is anybody else having much luck with SpecAugment?  Or just time or frequency masking?  Would you like to share your training parameters?  Or theories as to why it would/wouldn't be useful given this years training metric?\n\n"
  }
}