{
  "id": 339034,
  "title": "Problems about mission, datasets, and models",
  "url": "/competitions/hubmap-organ-segmentation/discussion/339034",
  "author_name": "",
  "post_date": "2022-07-23T03:18:21.597593500Z",
  "votes": 7,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Mission: What's the mission of this comp? semantic or instance segmentation?<br>\nDatasets: May I ask what dataset u use, coco or else?<br>\nModels: Which model is recommended? According to my investigation, U-Net seems to be the most recommended?<br>\nand, Can I use such as mmdet, mmseg for this comp?<br>\nThanks to u responses!</p>",
  "messages": [
    {
      "id": "1867148",
      "postDate": "07/23/2022 03:18:21",
      "content": "<p>Mission: What's the mission of this comp? semantic or instance segmentation?<br>\nDatasets: May I ask what dataset u use, coco or else?<br>\nModels: Which model is recommended? According to my investigation, U-Net seems to be the most recommended?<br>\nand, Can I use such as mmdet, mmseg for this comp?<br>\nThanks to u responses!</p>",
      "rawMarkdown": "Mission: What's the mission of this comp? semantic or instance segmentation?\nDatasets: May I ask what dataset u use, coco or else?\nModels: Which model is recommended? According to my investigation, U-Net seems to be the most recommended?\nand, Can I use such as mmdet, mmseg for this comp?\nThanks to u responses!",
      "votes": null
    },
    {
      "id": "1870144",
      "postDate": "07/25/2022 10:16:14",
      "content": "<p>Hi! As for me:</p>\n<ul>\n<li>I'm approaching the problem as semantic segmentation (stick to the requirements)</li>\n<li>I converted the dataset to ADE20K format, 512x512 non-empty crops (simplicity always pays back)</li>\n<li>I'm using UPerNet as head and trying different backbones (I'm concentrating more on augmentations at the beginning)</li>\n<li>I'm using mmsegmentation, even though I modified the code heavily (integrated EfficientNet, Albumentations). The notebook installation is not straightforward so I'd not recommend using it.</li>\n</ul>",
      "rawMarkdown": "Hi! As for me:\n- I'm approaching the problem as semantic segmentation (stick to the requirements)\n- I converted the dataset to ADE20K format, 512x512 non-empty crops (simplicity always pays back)\n- I'm using UPerNet as head and trying different backbones (I'm concentrating more on augmentations at the beginning)\n- I'm using mmsegmentation, even though I modified the code heavily (integrated EfficientNet, Albumentations). The notebook installation is not straightforward so I'd not recommend using it.",
      "votes": null
    },
    {
      "id": "1870313",
      "postDate": "07/25/2022 13:16:48",
      "content": "<p>thanks!<br>\n :)</p>",
      "rawMarkdown": "thanks!\n :)",
      "votes": null
    },
    {
      "id": "1871002",
      "postDate": "07/26/2022 02:10:57",
      "content": "<p>maybe you should try small batchsize.</p>",
      "rawMarkdown": "maybe you should try small batchsize.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1870144,
      "author_name": "mawanda",
      "author_url": "",
      "post_date": "07/25/2022 10:16:14",
      "content": "<p>Hi! As for me:</p>\n<ul>\n<li>I'm approaching the problem as semantic segmentation (stick to the requirements)</li>\n<li>I converted the dataset to ADE20K format, 512x512 non-empty crops (simplicity always pays back)</li>\n<li>I'm using UPerNet as head and trying different backbones (I'm concentrating more on augmentations at the beginning)</li>\n<li>I'm using mmsegmentation, even though I modified the code heavily (integrated EfficientNet, Albumentations). The notebook installation is not straightforward so I'd not recommend using it.</li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 1870313,
          "author_name": "fortelean",
          "author_url": "",
          "post_date": "07/25/2022 13:16:48",
          "content": "<p>thanks!<br>\n :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1871002,
      "author_name": "befunny",
      "author_url": "",
      "post_date": "07/26/2022 02:10:57",
      "content": "<p>maybe you should try small batchsize.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1867148": "Mission: What's the mission of this comp? semantic or instance segmentation?\nDatasets: May I ask what dataset u use, coco or else?\nModels: Which model is recommended? According to my investigation, U-Net seems to be the most recommended?\nand, Can I use such as mmdet, mmseg for this comp?\nThanks to u responses!",
    "1870144": "Hi! As for me:\n- I'm approaching the problem as semantic segmentation (stick to the requirements)\n- I converted the dataset to ADE20K format, 512x512 non-empty crops (simplicity always pays back)\n- I'm using UPerNet as head and trying different backbones (I'm concentrating more on augmentations at the beginning)\n- I'm using mmsegmentation, even though I modified the code heavily (integrated EfficientNet, Albumentations). The notebook installation is not straightforward so I'd not recommend using it.",
    "1870313": "thanks!\n :)",
    "1871002": "maybe you should try small batchsize."
  },
  "source": "meta"
}