{
  "id": 412258,
  "title": "Introduction to SegFormer",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/412258",
  "author_name": "",
  "post_date": "2023-05-23T01:39:35.344793500Z",
  "votes": 7,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Since this is a image segmentation competition, it can be useful to understand advanced segmentation models. The model used by the <a href=\"https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/356201\" target=\"_blank\">winner of the previous</a> HuBMAP + HPA - Hacking the Human Body competition was a SegFormer, so here is some information about this model.</p>\n<h1>Overview of Model</h1>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2F3217f32058c82d8840284c040ae7e5b0%2FSegFormerImg.png?generation=1684803209937286&amp;alt=media\" alt=\"\"><br>\n<a href=\"https://arxiv.org/pdf/2105.15203.pdf\" target=\"_blank\">image and info from research paper</a></p>\n<p>SegFormer is a Transformer framework for semantic segmentation. They have redesigned the encoder and decoder with the following novelties:</p>\n<ul>\n<li>A novel positional-encoding-free and hierarchical Transformer encoder.</li>\n<li>A lightweight All-MLP decoder design that yields a powerful representation without complex and<br>\ncomputationally demanding modules.</li>\n<li>state-of-the-art results in terms of efficiency, accuracy and robustness in three publicly available semantic segmentation datasets</li>\n</ul>\n<p>There is a variety of Mix Transformer encoders (MiT) varients. They range from small to large following MiT-B0 to MiT-B5.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2F164f1b980259338277501ad807b41e71%2Fseg_former_varients.png?generation=1684805854995215&amp;alt=media\" alt=\"\"></p>\n<h1>References and Links</h1>\n<p><a href=\"https://arxiv.org/pdf/2105.15203.pdf\" target=\"_blank\">SegFormer Research Paper</a><br>\n<a href=\"https://huggingface.co/docs/transformers/model_doc/segformer\" target=\"_blank\">SegFormer from HuggingFace</a><br>\n<a href=\"https://huggingface.co/blog/fine-tune-segformer\" target=\"_blank\">Fine Tune SegFormer HuggingFace</a><br>\n<a href=\"https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer\" target=\"_blank\">mmsegmentation has a SegFormer implementation</a><br>\n<a href=\"https://github.com/NVlabs/SegFormer\" target=\"_blank\">NVLabs SegFormer implementation</a></p>",
  "messages": [
    {
      "id": "2270081",
      "postDate": "05/23/2023 01:39:35",
      "content": "<p>Since this is a image segmentation competition, it can be useful to understand advanced segmentation models. The model used by the <a href=\"https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/356201\" target=\"_blank\">winner of the previous</a> HuBMAP + HPA - Hacking the Human Body competition was a SegFormer, so here is some information about this model.</p>\n<h1>Overview of Model</h1>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2F3217f32058c82d8840284c040ae7e5b0%2FSegFormerImg.png?generation=1684803209937286&amp;alt=media\" alt=\"\"><br>\n<a href=\"https://arxiv.org/pdf/2105.15203.pdf\" target=\"_blank\">image and info from research paper</a></p>\n<p>SegFormer is a Transformer framework for semantic segmentation. They have redesigned the encoder and decoder with the following novelties:</p>\n<ul>\n<li>A novel positional-encoding-free and hierarchical Transformer encoder.</li>\n<li>A lightweight All-MLP decoder design that yields a powerful representation without complex and<br>\ncomputationally demanding modules.</li>\n<li>state-of-the-art results in terms of efficiency, accuracy and robustness in three publicly available semantic segmentation datasets</li>\n</ul>\n<p>There is a variety of Mix Transformer encoders (MiT) varients. They range from small to large following MiT-B0 to MiT-B5.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2F164f1b980259338277501ad807b41e71%2Fseg_former_varients.png?generation=1684805854995215&amp;alt=media\" alt=\"\"></p>\n<h1>References and Links</h1>\n<p><a href=\"https://arxiv.org/pdf/2105.15203.pdf\" target=\"_blank\">SegFormer Research Paper</a><br>\n<a href=\"https://huggingface.co/docs/transformers/model_doc/segformer\" target=\"_blank\">SegFormer from HuggingFace</a><br>\n<a href=\"https://huggingface.co/blog/fine-tune-segformer\" target=\"_blank\">Fine Tune SegFormer HuggingFace</a><br>\n<a href=\"https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer\" target=\"_blank\">mmsegmentation has a SegFormer implementation</a><br>\n<a href=\"https://github.com/NVlabs/SegFormer\" target=\"_blank\">NVLabs SegFormer implementation</a></p>",
      "rawMarkdown": "Since this is a image segmentation competition, it can be useful to understand advanced segmentation models. The model used by the [winner of the previous](https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/356201) HuBMAP + HPA - Hacking the Human Body competition was a SegFormer, so here is some information about this model.\n\n#Overview of Model\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2F3217f32058c82d8840284c040ae7e5b0%2FSegFormerImg.png?generation=1684803209937286&alt=media)\n[image and info from research paper](https://arxiv.org/pdf/2105.15203.pdf)\n\nSegFormer is a Transformer framework for semantic segmentation. They have redesigned the encoder and decoder with the following novelties:\n- A novel positional-encoding-free and hierarchical Transformer encoder.\n- A lightweight All-MLP decoder design that yields a powerful representation without complex and\ncomputationally demanding modules.\n- state-of-the-art results in terms of efficiency, accuracy and robustness in three publicly available semantic segmentation datasets\n\nThere is a variety of Mix Transformer encoders (MiT) varients. They range from small to large following MiT-B0 to MiT-B5.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2F164f1b980259338277501ad807b41e71%2Fseg_former_varients.png?generation=1684805854995215&alt=media)\n\n# References and Links\n[SegFormer Research Paper](https://arxiv.org/pdf/2105.15203.pdf)\n[SegFormer from HuggingFace](https://huggingface.co/docs/transformers/model_doc/segformer)\n[Fine Tune SegFormer HuggingFace](https://huggingface.co/blog/fine-tune-segformer)\n[mmsegmentation has a SegFormer implementation](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer)\n[NVLabs SegFormer implementation](https://github.com/NVlabs/SegFormer)",
      "votes": null
    },
    {
      "id": "2315754",
      "postDate": "06/24/2023 11:43:33",
      "content": "<p>This is completely new to me. Thanks for sharing! </p>",
      "rawMarkdown": "This is completely new to me. Thanks for sharing!",
      "votes": null
    },
    {
      "id": "2317292",
      "postDate": "06/25/2023 15:20:26",
      "content": "<p>According to host, segformer is not allowed in the competition<br>\n<a href=\"https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/413834#2281425\" target=\"_blank\">https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/413834#2281425</a></p>",
      "rawMarkdown": "According to host, segformer is not allowed in the competition\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/413834#2281425",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2315754,
      "author_name": "samuelcortinhas",
      "author_url": "",
      "post_date": "06/24/2023 11:43:33",
      "content": "<p>This is completely new to me. Thanks for sharing! </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2317292,
      "author_name": "ptran1203",
      "author_url": "",
      "post_date": "06/25/2023 15:20:26",
      "content": "<p>According to host, segformer is not allowed in the competition<br>\n<a href=\"https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/413834#2281425\" target=\"_blank\">https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/413834#2281425</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2270081": "Since this is a image segmentation competition, it can be useful to understand advanced segmentation models. The model used by the [winner of the previous](https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/356201) HuBMAP + HPA - Hacking the Human Body competition was a SegFormer, so here is some information about this model.\n\n#Overview of Model\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2F3217f32058c82d8840284c040ae7e5b0%2FSegFormerImg.png?generation=1684803209937286&alt=media)\n[image and info from research paper](https://arxiv.org/pdf/2105.15203.pdf)\n\nSegFormer is a Transformer framework for semantic segmentation. They have redesigned the encoder and decoder with the following novelties:\n- A novel positional-encoding-free and hierarchical Transformer encoder.\n- A lightweight All-MLP decoder design that yields a powerful representation without complex and\ncomputationally demanding modules.\n- state-of-the-art results in terms of efficiency, accuracy and robustness in three publicly available semantic segmentation datasets\n\nThere is a variety of Mix Transformer encoders (MiT) varients. They range from small to large following MiT-B0 to MiT-B5.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2F164f1b980259338277501ad807b41e71%2Fseg_former_varients.png?generation=1684805854995215&alt=media)\n\n# References and Links\n[SegFormer Research Paper](https://arxiv.org/pdf/2105.15203.pdf)\n[SegFormer from HuggingFace](https://huggingface.co/docs/transformers/model_doc/segformer)\n[Fine Tune SegFormer HuggingFace](https://huggingface.co/blog/fine-tune-segformer)\n[mmsegmentation has a SegFormer implementation](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/segformer)\n[NVLabs SegFormer implementation](https://github.com/NVlabs/SegFormer)",
    "2315754": "This is completely new to me. Thanks for sharing!",
    "2317292": "According to host, segformer is not allowed in the competition\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/413834#2281425"
  },
  "source": "meta"
}