{
  "id": 317236,
  "title": "Problem with data augmentation",
  "url": "/competitions/happy-whale-and-dolphin/discussion/317236",
  "author_name": "",
  "post_date": "2022-04-06T04:01:10.075050500Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I've tried some off-line data augmentations to expand the training set.When I tested with resnet18 backbone,it improve with 5% on my local CV.But when I changed to bigger backbone, however, I got lower score than the model which I didn't use off-line augmentations.Did i miss something?😳</p>",
  "messages": [
    {
      "id": "1746739",
      "postDate": "04/06/2022 04:01:10",
      "content": "<p>I've tried some off-line data augmentations to expand the training set.When I tested with resnet18 backbone,it improve with 5% on my local CV.But when I changed to bigger backbone, however, I got lower score than the model which I didn't use off-line augmentations.Did i miss something?😳</p>",
      "rawMarkdown": "I've tried some off-line data augmentations to expand the training set.When I tested with resnet18 backbone,it improve with 5% on my local CV.But when I changed to bigger backbone, however, I got lower score than the model which I didn't use off-line augmentations.Did i miss something?😳",
      "votes": null
    },
    {
      "id": "1746880",
      "postDate": "04/06/2022 07:31:17",
      "content": "<blockquote>\n  <p>Did i miss something</p>\n</blockquote>\n<p>Not necessary, bigger model have more power to learn (and more prone to overfitting). Bigger models sometimes need stronger augmentation, also there may be more causes, bad CV as example.</p>",
      "rawMarkdown": "> Did i miss something\n\nNot necessary, bigger model have more power to learn (and more prone to overfitting). Bigger models sometimes need stronger augmentation, also there may be more causes, bad CV as example.",
      "votes": null
    },
    {
      "id": "1746902",
      "postDate": "04/06/2022 07:44:23",
      "content": "<p>thanks for your comment. I didn't make changes to  the valid fold. I will check if the model is overfitted.</p>",
      "rawMarkdown": "thanks for your comment. I didn't make changes to  the valid fold. I will check if the model is overfitted.",
      "votes": null
    },
    {
      "id": "1746908",
      "postDate": "04/06/2022 07:48:44",
      "content": "<p>I didn't mean you have some changes in valid fold, I mean the source valid folds may not represent the real situation :) The validation in this competition is so non-obvious.</p>",
      "rawMarkdown": "I didn't mean you have some changes in valid fold, I mean the source valid folds may not represent the real situation :) The validation in this competition is so non-obvious.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1746880,
      "author_name": "kwentar",
      "author_url": "",
      "post_date": "04/06/2022 07:31:17",
      "content": "<blockquote>\n  <p>Did i miss something</p>\n</blockquote>\n<p>Not necessary, bigger model have more power to learn (and more prone to overfitting). Bigger models sometimes need stronger augmentation, also there may be more causes, bad CV as example.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1746902,
          "author_name": "jimmysmith1009",
          "author_url": "",
          "post_date": "04/06/2022 07:44:23",
          "content": "<p>thanks for your comment. I didn't make changes to  the valid fold. I will check if the model is overfitted.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1746908,
          "author_name": "kwentar",
          "author_url": "",
          "post_date": "04/06/2022 07:48:44",
          "content": "<p>I didn't mean you have some changes in valid fold, I mean the source valid folds may not represent the real situation :) The validation in this competition is so non-obvious.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1746739": "I've tried some off-line data augmentations to expand the training set.When I tested with resnet18 backbone,it improve with 5% on my local CV.But when I changed to bigger backbone, however, I got lower score than the model which I didn't use off-line augmentations.Did i miss something?😳",
    "1746880": "> Did i miss something\n\nNot necessary, bigger model have more power to learn (and more prone to overfitting). Bigger models sometimes need stronger augmentation, also there may be more causes, bad CV as example.",
    "1746902": "thanks for your comment. I didn't make changes to  the valid fold. I will check if the model is overfitted.",
    "1746908": "I didn't mean you have some changes in valid fold, I mean the source valid folds may not represent the real situation :) The validation in this competition is so non-obvious."
  },
  "source": "meta"
}