{
  "id": 420180,
  "title": "Augmentation Analysis",
  "url": "/competitions/google-research-identify-contrails-reduce-global-warming/discussion/420180",
  "author_name": "Balaji Selvaraj",
  "post_date": "2023-06-29T14:36:35.792000",
  "votes": 14,
  "comment_count": 11,
  "views": 0,
  "content": "<p>From my experiments so far, Randomresizedcrop alone gives improvement.</p>\n<p>Did anyone get find any interesting augmentation scheme ?</p>\n<p>I will add my findings in this section</p>\n<p>All the best👍</p>",
  "messages": [
    {
      "id": 2322868,
      "postDate": "2023-06-29T14:36:35.793Z",
      "content": "<p>From my experiments so far, Randomresizedcrop alone gives improvement.</p>\n<p>Did anyone get find any interesting augmentation scheme ?</p>\n<p>I will add my findings in this section</p>\n<p>All the best👍</p>",
      "rawMarkdown": "From my experiments so far, Randomresizedcrop alone gives improvement.\n\nDid anyone get find any interesting augmentation scheme ?\n\nI will add my findings in this section\n\nAll the best👍",
      "votes": 13
    },
    {
      "id": 2323301,
      "postDate": "2023-06-29T20:22:31.157Z",
      "content": "<p>In my case using augmentation helps a lot, if I don't use them model overfitts quickly. </p>\n<p>Have only tried flips and rotation augmentations.</p>",
      "rawMarkdown": "In my case using augmentation helps a lot, if I don't use them model overfitts quickly. \n\nHave only tried flips and rotation augmentations.",
      "votes": 5,
      "replies": [
        {
          "id": 2323803,
          "postDate": "2023-06-30T07:05:05.430Z",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/ragnar123\" target=\"_blank\">@ragnar123</a> </p>\n<p>That's good.. Are you using segmentation_models pytorch pipeline ?</p>",
          "rawMarkdown": "Hi @ragnar123 \n\nThat's good.. Are you using segmentation_models pytorch pipeline ?",
          "votes": 1,
          "replies": [
            {
              "id": 2324343,
              "postDate": "2023-06-30T14:43:57.613Z",
              "content": "<p>Yes, common pytorch pipeline and albumentations for augmentations.</p>",
              "rawMarkdown": "Yes, common pytorch pipeline and albumentations for augmentations.",
              "votes": 1
            },
            {
              "id": 2330476,
              "postDate": "2023-07-05T03:38:54.923Z",
              "content": "<p>Will flips and rotation not mess up the labels , just thinking 🤔 </p>",
              "rawMarkdown": "Will flips and rotation not mess up the labels , just thinking 🤔 "
            },
            {
              "id": 2330540,
              "postDate": "2023-07-05T04:27:26.947Z",
              "content": "<p>Albumentations offer flips and rotation augmentations that change the mask along with the image.<br>\nOr you can write your own</p>",
              "rawMarkdown": "Albumentations offer flips and rotation augmentations that change the mask along with the image.\nOr you can write your own",
              "votes": 2
            },
            {
              "id": 2330708,
              "postDate": "2023-07-05T06:43:10.660Z",
              "content": "<p>Yup was doing it without that for not but thanks for the pointer makes sense</p>",
              "rawMarkdown": "Yup was doing it without that for not but thanks for the pointer makes sense"
            }
          ]
        }
      ]
    },
    {
      "id": 2323525,
      "postDate": "2023-06-30T02:57:01.063Z",
      "content": "<p>Not helping at all here. :(</p>",
      "rawMarkdown": "Not helping at all here. :(",
      "votes": 1,
      "replies": [
        {
          "id": 2324422,
          "postDate": "2023-06-30T15:37:21.887Z",
          "content": "<p>give it more epochs ;)</p>",
          "rawMarkdown": "give it more epochs ;)",
          "votes": 1,
          "replies": [
            {
              "id": 2324583,
              "postDate": "2023-06-30T17:48:13.983Z",
              "content": "<p>Thanks. I will give it a try.</p>",
              "rawMarkdown": "Thanks. I will give it a try."
            }
          ]
        },
        {
          "id": 2324626,
          "postDate": "2023-06-30T18:26:26.203Z",
          "content": "<p>Similar to what <a href=\"https://www.kaggle.com/ragnar123\" target=\"_blank\">@ragnar123</a> said.</p>\n<p>Randomresizedcrop did not increase my CV, but the epoch-to-epoch validation dice increase is more stable.</p>",
          "rawMarkdown": "Similar to what @ragnar123 said.\n\nRandomresizedcrop did not increase my CV, but the epoch-to-epoch validation dice increase is more stable.",
          "votes": 2
        }
      ]
    },
    {
      "id": 2324448,
      "postDate": "2023-06-30T15:55:17.817Z",
      "content": "<p>I've tried using augmentations but none have worked so far (even when increasing number of epochs from 20 to 30). I haven't tried using larger models yet though.</p>",
      "rawMarkdown": "I've tried using augmentations but none have worked so far (even when increasing number of epochs from 20 to 30). I haven't tried using larger models yet though."
    }
  ],
  "comments": [
    {
      "id": 2323301,
      "author_name": "Martin Kovacevic Buvinic",
      "author_url": "",
      "post_date": "2023-06-29T20:22:31.157000",
      "content": "<p>In my case using augmentation helps a lot, if I don't use them model overfitts quickly. </p>\n<p>Have only tried flips and rotation augmentations.</p>",
      "votes": 5,
      "replies": [
        {
          "id": 2323803,
          "author_name": "Balaji Selvaraj",
          "author_url": "",
          "post_date": "2023-06-30T07:05:05.430000",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/ragnar123\" target=\"_blank\">@ragnar123</a> </p>\n<p>That's good.. Are you using segmentation_models pytorch pipeline ?</p>",
          "votes": 1,
          "replies": [
            {
              "id": 2324343,
              "author_name": "Martin Kovacevic Buvinic",
              "author_url": "",
              "post_date": "2023-06-30T14:43:57.613000",
              "content": "<p>Yes, common pytorch pipeline and albumentations for augmentations.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 2330476,
              "author_name": "Gaurav Rawat",
              "author_url": "",
              "post_date": "2023-07-05T03:38:54.923000",
              "content": "<p>Will flips and rotation not mess up the labels , just thinking 🤔 </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2330540,
              "author_name": "Balaji Selvaraj",
              "author_url": "",
              "post_date": "2023-07-05T04:27:26.947000",
              "content": "<p>Albumentations offer flips and rotation augmentations that change the mask along with the image.<br>\nOr you can write your own</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 2330708,
              "author_name": "Gaurav Rawat",
              "author_url": "",
              "post_date": "2023-07-05T06:43:10.660000",
              "content": "<p>Yup was doing it without that for not but thanks for the pointer makes sense</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2323525,
      "author_name": "june",
      "author_url": "",
      "post_date": "2023-06-30T02:57:01.063000",
      "content": "<p>Not helping at all here. :(</p>",
      "votes": 1,
      "replies": [
        {
          "id": 2324422,
          "author_name": "JEANMPIA",
          "author_url": "",
          "post_date": "2023-06-30T15:37:21.887000",
          "content": "<p>give it more epochs ;)</p>",
          "votes": 1,
          "replies": [
            {
              "id": 2324583,
              "author_name": "june",
              "author_url": "",
              "post_date": "2023-06-30T17:48:13.983000",
              "content": "<p>Thanks. I will give it a try.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 2324626,
          "author_name": "Bartley",
          "author_url": "",
          "post_date": "2023-06-30T18:26:26.203000",
          "content": "<p>Similar to what <a href=\"https://www.kaggle.com/ragnar123\" target=\"_blank\">@ragnar123</a> said.</p>\n<p>Randomresizedcrop did not increase my CV, but the epoch-to-epoch validation dice increase is more stable.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 2324448,
      "author_name": "Ari",
      "author_url": "",
      "post_date": "2023-06-30T15:55:17.817000",
      "content": "<p>I've tried using augmentations but none have worked so far (even when increasing number of epochs from 20 to 30). I haven't tried using larger models yet though.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2322868": "From my experiments so far, Randomresizedcrop alone gives improvement.\n\nDid anyone get find any interesting augmentation scheme ?\n\nI will add my findings in this section\n\nAll the best👍",
    "2323301": "In my case using augmentation helps a lot, if I don't use them model overfitts quickly. \n\nHave only tried flips and rotation augmentations.",
    "2323525": "Not helping at all here. :(",
    "2324448": "I've tried using augmentations but none have worked so far (even when increasing number of epochs from 20 to 30). I haven't tried using larger models yet though."
  }
}