{
  "id": 351954,
  "title": "Looking for an insight why a model trained with higher resolution performs worse than one with lower?",
  "url": "/competitions/hubmap-organ-segmentation/discussion/351954",
  "author_name": "",
  "post_date": "2022-09-12T13:24:30.760099Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hey guys. I'm running a CoaT model and it seemed to work pretty well until I tried to train it with 1024res images. Compared to the model trained with 786res images, models trained with 1024res images drastically performed worse, especially on evaluating Hubmap dataset in Public Leaderboard. I wonder why and how did you guys deal with this problem. Please give me your insights.</p>\n<p>PS same fold, augmentation, lr.</p>",
  "messages": [
    {
      "id": "1936028",
      "postDate": "09/12/2022 13:24:30",
      "content": "<p>Hey guys. I'm running a CoaT model and it seemed to work pretty well until I tried to train it with 1024res images. Compared to the model trained with 786res images, models trained with 1024res images drastically performed worse, especially on evaluating Hubmap dataset in Public Leaderboard. I wonder why and how did you guys deal with this problem. Please give me your insights.</p>\n<p>PS same fold, augmentation, lr.</p>",
      "rawMarkdown": "Hey guys. I'm running a CoaT model and it seemed to work pretty well until I tried to train it with 1024res images. Compared to the model trained with 786res images, models trained with 1024res images drastically performed worse, especially on evaluating Hubmap dataset in Public Leaderboard. I wonder why and how did you guys deal with this problem. Please give me your insights.\n\nPS same fold, augmentation, lr.",
      "votes": null
    },
    {
      "id": "1936091",
      "postDate": "09/12/2022 14:09:55",
      "content": "<blockquote>\n  <p>The HuBMAP images range in size from 4500x4500 down to 160x160 pixels</p>\n</blockquote>\n<p>This is from the data tab. Larger resolution means there will be some upsample artifacts on smaller images, but HPA images don't have them.</p>",
      "rawMarkdown": "> The HuBMAP images range in size from 4500x4500 down to 160x160 pixels\n\nThis is from the data tab. Larger resolution means there will be some upsample artifacts on smaller images, but HPA images don't have them.",
      "votes": null
    },
    {
      "id": "1936122",
      "postDate": "09/12/2022 14:22:08",
      "content": "<p>Thank you for the answer, <a href=\"https://www.kaggle.com/gunesevitan\" target=\"_blank\">@gunesevitan</a>. I'm going to google how to deal with the artifacts then. By the way, can you suggest related readings if any?</p>",
      "rawMarkdown": "Thank you for the answer, @gunesevitan. I'm going to google how to deal with the artifacts then. By the way, can you suggest related readings if any?",
      "votes": null
    },
    {
      "id": "1936237",
      "postDate": "09/12/2022 15:44:17",
      "content": "<p>I find transformer models &gt; 1280 resolution do worse, also.</p>",
      "rawMarkdown": "I find transformer models > 1280 resolution do worse, also.",
      "votes": null
    },
    {
      "id": "1936253",
      "postDate": "09/12/2022 16:01:22",
      "content": "<p>Thank you very much for the tip!</p>",
      "rawMarkdown": "Thank you very much for the tip!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1936091,
      "author_name": "gunesevitan",
      "author_url": "",
      "post_date": "09/12/2022 14:09:55",
      "content": "<blockquote>\n  <p>The HuBMAP images range in size from 4500x4500 down to 160x160 pixels</p>\n</blockquote>\n<p>This is from the data tab. Larger resolution means there will be some upsample artifacts on smaller images, but HPA images don't have them.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1936122,
          "author_name": "cheulkay",
          "author_url": "",
          "post_date": "09/12/2022 14:22:08",
          "content": "<p>Thank you for the answer, <a href=\"https://www.kaggle.com/gunesevitan\" target=\"_blank\">@gunesevitan</a>. I'm going to google how to deal with the artifacts then. By the way, can you suggest related readings if any?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1936237,
      "author_name": "jamesphoward",
      "author_url": "",
      "post_date": "09/12/2022 15:44:17",
      "content": "<p>I find transformer models &gt; 1280 resolution do worse, also.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1936253,
          "author_name": "cheulkay",
          "author_url": "",
          "post_date": "09/12/2022 16:01:22",
          "content": "<p>Thank you very much for the tip!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1936028": "Hey guys. I'm running a CoaT model and it seemed to work pretty well until I tried to train it with 1024res images. Compared to the model trained with 786res images, models trained with 1024res images drastically performed worse, especially on evaluating Hubmap dataset in Public Leaderboard. I wonder why and how did you guys deal with this problem. Please give me your insights.\n\nPS same fold, augmentation, lr.",
    "1936091": "> The HuBMAP images range in size from 4500x4500 down to 160x160 pixels\n\nThis is from the data tab. Larger resolution means there will be some upsample artifacts on smaller images, but HPA images don't have them.",
    "1936122": "Thank you for the answer, @gunesevitan. I'm going to google how to deal with the artifacts then. By the way, can you suggest related readings if any?",
    "1936237": "I find transformer models > 1280 resolution do worse, also.",
    "1936253": "Thank you very much for the tip!"
  },
  "source": "meta"
}