{
  "id": 573000,
  "title": "Data Augmentation",
  "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/573000",
  "author_name": "",
  "post_date": "2025-04-12T20:02:39.160455900Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>In YOLO training, I want to rotate images in four stages, with each stage being 90 degrees. Does this help in training?<br>\nWith this method, the training data will increase to about 20,000 images. Is this number of images enough?</p>",
  "messages": [
    {
      "id": "3177474",
      "postDate": "04/12/2025 20:02:39",
      "content": "<p>In YOLO training, I want to rotate images in four stages, with each stage being 90 degrees. Does this help in training?<br>\nWith this method, the training data will increase to about 20,000 images. Is this number of images enough?</p>",
      "rawMarkdown": "In YOLO training, I want to rotate images in four stages, with each stage being 90 degrees. Does this help in training?\nWith this method, the training data will increase to about 20,000 images. Is this number of images enough?",
      "votes": null
    },
    {
      "id": "3178695",
      "postDate": "04/14/2025 13:36:57",
      "content": "<p>If you are doing transfer learning then yes it is enough but I would consider other techniques than just rotating.  Consider things like random patch dropouts and spectral augmentations instead of only spatial.  There are a lot of things you can try.  The benefit to including 90 degree rotations is that it may help with TTA on your submissions.</p>",
      "rawMarkdown": "If you are doing transfer learning then yes it is enough but I would consider other techniques than just rotating.  Consider things like random patch dropouts and spectral augmentations instead of only spatial.  There are a lot of things you can try.  The benefit to including 90 degree rotations is that it may help with TTA on your submissions.",
      "votes": null
    },
    {
      "id": "3178700",
      "postDate": "04/14/2025 13:47:06",
      "content": "<p>I'll use spectral augmentations going forward, but I'm not sure how to apply patch dropout on the data.</p>",
      "rawMarkdown": "I'll use spectral augmentations going forward, but I'm not sure how to apply patch dropout on the data.",
      "votes": null
    },
    {
      "id": "3179094",
      "postDate": "04/15/2025 03:11:41",
      "content": "<p>The MONAI library has augmentations that are popular for biology problems and tomographs.  That is what I use for all my augmentations in competitions such as this and there is a 3d dropout augmentation in there that you can use.</p>",
      "rawMarkdown": "The MONAI library has augmentations that are popular for biology problems and tomographs.  That is what I use for all my augmentations in competitions such as this and there is a 3d dropout augmentation in there that you can use.",
      "votes": null
    },
    {
      "id": "3179614",
      "postDate": "04/15/2025 14:35:03",
      "content": "<p>I applied some augmentations, such as vertical flip and Gaussian blur, but they didn’t seem to improve the LB score. Maybe some bugs in my code.😅</p>",
      "rawMarkdown": "I applied some augmentations, such as vertical flip and Gaussian blur, but they didn’t seem to improve the LB score. Maybe some bugs in my code.😅",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3178695,
      "author_name": "connorjd",
      "author_url": "",
      "post_date": "04/14/2025 13:36:57",
      "content": "<p>If you are doing transfer learning then yes it is enough but I would consider other techniques than just rotating.  Consider things like random patch dropouts and spectral augmentations instead of only spatial.  There are a lot of things you can try.  The benefit to including 90 degree rotations is that it may help with TTA on your submissions.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3178700,
          "author_name": "ali78kabirzadeh",
          "author_url": "",
          "post_date": "04/14/2025 13:47:06",
          "content": "<p>I'll use spectral augmentations going forward, but I'm not sure how to apply patch dropout on the data.</p>",
          "votes": null,
          "replies": [
            {
              "id": 3179094,
              "author_name": "connorjd",
              "author_url": "",
              "post_date": "04/15/2025 03:11:41",
              "content": "<p>The MONAI library has augmentations that are popular for biology problems and tomographs.  That is what I use for all my augmentations in competitions such as this and there is a 3d dropout augmentation in there that you can use.</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3179614,
      "author_name": "seeingtimes",
      "author_url": "",
      "post_date": "04/15/2025 14:35:03",
      "content": "<p>I applied some augmentations, such as vertical flip and Gaussian blur, but they didn’t seem to improve the LB score. Maybe some bugs in my code.😅</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3177474": "In YOLO training, I want to rotate images in four stages, with each stage being 90 degrees. Does this help in training?\nWith this method, the training data will increase to about 20,000 images. Is this number of images enough?",
    "3178695": "If you are doing transfer learning then yes it is enough but I would consider other techniques than just rotating.  Consider things like random patch dropouts and spectral augmentations instead of only spatial.  There are a lot of things you can try.  The benefit to including 90 degree rotations is that it may help with TTA on your submissions.",
    "3178700": "I'll use spectral augmentations going forward, but I'm not sure how to apply patch dropout on the data.",
    "3179094": "The MONAI library has augmentations that are popular for biology problems and tomographs.  That is what I use for all my augmentations in competitions such as this and there is a 3d dropout augmentation in there that you can use.",
    "3179614": "I applied some augmentations, such as vertical flip and Gaussian blur, but they didn’t seem to improve the LB score. Maybe some bugs in my code.😅"
  },
  "source": "meta"
}