{
  "id": 354686,
  "title": "How did you train SegFormer?",
  "url": "/competitions/hubmap-organ-segmentation/discussion/354686",
  "author_name": "",
  "post_date": "2022-09-23T10:11:48.390373200Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>SegFormer was one of the most popular models in this competition.<br>\nI also trained the SegFormer B2 model following <a href=\"https://www.kaggle.com/datasets/hengck23/hubmap-discuss-00\" target=\"_blank\">hengck23's code</a>, but failed to reproduce the score reported in <code>train-log-segformer-mit-b2</code>. <br>\nI'm keen to know what I could do to succeed. How did you train SegFormer?</p>\n<p>Just for your information, here is what I did:</p>\n<ul>\n<li><a href=\"https://www.kaggle.com/code/shionhonda/hubhpa-train-sf-mitb2\" target=\"_blank\">training</a></li>\n<li><a href=\"https://www.kaggle.com/code/shionhonda/hubhpa-infer-sf-mitb2\" target=\"_blank\">inference</a></li>\n</ul>",
  "messages": [
    {
      "id": "1951904",
      "postDate": "09/23/2022 10:11:48",
      "content": "<p>SegFormer was one of the most popular models in this competition.<br>\nI also trained the SegFormer B2 model following <a href=\"https://www.kaggle.com/datasets/hengck23/hubmap-discuss-00\" target=\"_blank\">hengck23's code</a>, but failed to reproduce the score reported in <code>train-log-segformer-mit-b2</code>. <br>\nI'm keen to know what I could do to succeed. How did you train SegFormer?</p>\n<p>Just for your information, here is what I did:</p>\n<ul>\n<li><a href=\"https://www.kaggle.com/code/shionhonda/hubhpa-train-sf-mitb2\" target=\"_blank\">training</a></li>\n<li><a href=\"https://www.kaggle.com/code/shionhonda/hubhpa-infer-sf-mitb2\" target=\"_blank\">inference</a></li>\n</ul>",
      "rawMarkdown": "SegFormer was one of the most popular models in this competition.\nI also trained the SegFormer B2 model following [hengck23's code](https://www.kaggle.com/datasets/hengck23/hubmap-discuss-00), but failed to reproduce the score reported in `train-log-segformer-mit-b2`. \nI'm keen to know what I could do to succeed. How did you train SegFormer?\n\nJust for your information, here is what I did:\n- [training](https://www.kaggle.com/code/shionhonda/hubhpa-train-sf-mitb2)\n- [inference](https://www.kaggle.com/code/shionhonda/hubhpa-infer-sf-mitb2)",
      "votes": null
    },
    {
      "id": "1951918",
      "postDate": "09/23/2022 10:20:39",
      "content": "<blockquote>\n  <p>Training of SegFormer is not very stable, which is sensitive to random seeds. <br>\n  from  <a href=\"https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer\" target=\"_blank\">https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer</a></p>\n</blockquote>",
      "rawMarkdown": "> Training of SegFormer is not very stable, which is sensitive to random seeds. \nfrom  https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer",
      "votes": null
    },
    {
      "id": "1951920",
      "postDate": "09/23/2022 10:22:08",
      "content": "<p>One possible issue may be the different data augmentation approaches, but I don't how much will that influence the results.</p>",
      "rawMarkdown": "One possible issue may be the different data augmentation approaches, but I don't how much will that influence the results.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1951918,
      "author_name": "befunny",
      "author_url": "",
      "post_date": "09/23/2022 10:20:39",
      "content": "<blockquote>\n  <p>Training of SegFormer is not very stable, which is sensitive to random seeds. <br>\n  from  <a href=\"https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer\" target=\"_blank\">https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer</a></p>\n</blockquote>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1951920,
      "author_name": "kuohsintu",
      "author_url": "",
      "post_date": "09/23/2022 10:22:08",
      "content": "<p>One possible issue may be the different data augmentation approaches, but I don't how much will that influence the results.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1951904": "SegFormer was one of the most popular models in this competition.\nI also trained the SegFormer B2 model following [hengck23's code](https://www.kaggle.com/datasets/hengck23/hubmap-discuss-00), but failed to reproduce the score reported in `train-log-segformer-mit-b2`. \nI'm keen to know what I could do to succeed. How did you train SegFormer?\n\nJust for your information, here is what I did:\n- [training](https://www.kaggle.com/code/shionhonda/hubhpa-train-sf-mitb2)\n- [inference](https://www.kaggle.com/code/shionhonda/hubhpa-infer-sf-mitb2)",
    "1951918": "> Training of SegFormer is not very stable, which is sensitive to random seeds. \nfrom  https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer",
    "1951920": "One possible issue may be the different data augmentation approaches, but I don't how much will that influence the results."
  },
  "source": "meta"
}