{
  "id": 239261,
  "title": "Mixup: On MelSpec or Raw Wave?",
  "url": "/competitions/birdclef-2021/discussion/239261",
  "author_name": "",
  "post_date": "2021-05-15T14:22:41.517664400Z",
  "votes": 9,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Is it more appropriate to apply mixup on melspec or raw wave? I always thought it should be raw wave, so we are simulating the layering better, but as I dived into past competitions, many winning solutions applied mixup directly on melspec. What's your opinion?</p>\n<p>Also, regarding labels for mixup, since it's an audio competition, and we are trying to simulate multiple birds as we layer different files together, should the labels be 1 for all species present in mix up clips or still multiply the mixup ratio as we do in regular image mixups? </p>",
  "messages": [
    {
      "id": "1308886",
      "postDate": "05/15/2021 14:22:41",
      "content": "<p>Is it more appropriate to apply mixup on melspec or raw wave? I always thought it should be raw wave, so we are simulating the layering better, but as I dived into past competitions, many winning solutions applied mixup directly on melspec. What's your opinion?</p>\n<p>Also, regarding labels for mixup, since it's an audio competition, and we are trying to simulate multiple birds as we layer different files together, should the labels be 1 for all species present in mix up clips or still multiply the mixup ratio as we do in regular image mixups? </p>",
      "rawMarkdown": "Is it more appropriate to apply mixup on melspec or raw wave? I always thought it should be raw wave, so we are simulating the layering better, but as I dived into past competitions, many winning solutions applied mixup directly on melspec. What's your opinion?\n\nAlso, regarding labels for mixup, since it's an audio competition, and we are trying to simulate multiple birds as we layer different files together, should the labels be 1 for all species present in mix up clips or still multiply the mixup ratio as we do in regular image mixups?",
      "votes": null
    },
    {
      "id": "1309207",
      "postDate": "05/15/2021 18:35:10",
      "content": "<p>try both and see what works best</p>",
      "rawMarkdown": "try both and see what works best",
      "votes": null
    },
    {
      "id": "1314552",
      "postDate": "05/19/2021 08:23:32",
      "content": "<p>The advantage of doing it on melspec is the one can do it after first saving pre-processed 1ch melspec as torch/npy files, speeding up the training.</p>\n<p>My current evaluation setup with melspec is 2 different mixup, cutmix, specaugmenter, random_batch_cutout, randomcrop, random_frequencies_adjustment, random_contrast_adjustment.</p>",
      "rawMarkdown": "The advantage of doing it on melspec is the one can do it after first saving pre-processed 1ch melspec as torch/npy files, speeding up the training.\n\nMy current evaluation setup with melspec is 2 different mixup, cutmix, specaugmenter, random_batch_cutout, randomcrop, random_frequencies_adjustment, random_contrast_adjustment.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1309207,
      "author_name": "philippsinger",
      "author_url": "",
      "post_date": "05/15/2021 18:35:10",
      "content": "<p>try both and see what works best</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1314552,
      "author_name": "kirderf",
      "author_url": "",
      "post_date": "05/19/2021 08:23:32",
      "content": "<p>The advantage of doing it on melspec is the one can do it after first saving pre-processed 1ch melspec as torch/npy files, speeding up the training.</p>\n<p>My current evaluation setup with melspec is 2 different mixup, cutmix, specaugmenter, random_batch_cutout, randomcrop, random_frequencies_adjustment, random_contrast_adjustment.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1308886": "Is it more appropriate to apply mixup on melspec or raw wave? I always thought it should be raw wave, so we are simulating the layering better, but as I dived into past competitions, many winning solutions applied mixup directly on melspec. What's your opinion?\n\nAlso, regarding labels for mixup, since it's an audio competition, and we are trying to simulate multiple birds as we layer different files together, should the labels be 1 for all species present in mix up clips or still multiply the mixup ratio as we do in regular image mixups?",
    "1309207": "try both and see what works best",
    "1314552": "The advantage of doing it on melspec is the one can do it after first saving pre-processed 1ch melspec as torch/npy files, speeding up the training.\n\nMy current evaluation setup with melspec is 2 different mixup, cutmix, specaugmenter, random_batch_cutout, randomcrop, random_frequencies_adjustment, random_contrast_adjustment."
  },
  "source": "meta"
}