{
  "id": 355065,
  "title": "What is your best score for single model and single fold?",
  "url": "/competitions/hubmap-organ-segmentation/discussion/355065",
  "author_name": "",
  "post_date": "2022-09-25T10:22:26.089916600Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>My best score  is 0.7568 for public LB, is 0.7580 for private LB.<br>\nThe model is PVT-v2_unet  forked from <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> notebook. The HPA images are normalized to Hubmap and are used stainaugmentor.</p>",
  "messages": [
    {
      "id": "1954535",
      "postDate": "09/25/2022 10:22:26",
      "content": "<p>My best score  is 0.7568 for public LB, is 0.7580 for private LB.<br>\nThe model is PVT-v2_unet  forked from <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> notebook. The HPA images are normalized to Hubmap and are used stainaugmentor.</p>",
      "rawMarkdown": "My best score  is 0.7568 for public LB, is 0.7580 for private LB.\nThe model is PVT-v2_unet  forked from @hengck23 notebook. The HPA images are normalized to Hubmap and are used stainaugmentor.",
      "votes": null
    },
    {
      "id": "1955390",
      "postDate": "09/26/2022 01:35:02",
      "content": "<p>Using pvtv2-b4 backbone with conv3x3 decoder (training for 768 * 768),  my best score is 0.78861 for public LB and 0.78165 for private LB. And I use pixel size adaptation (test image resize scale = (target pixel size/ source pixel size * training scale), if you training for 768, the  training scale is 768/3000) for test dataset.</p>",
      "rawMarkdown": "Using pvtv2-b4 backbone with conv3x3 decoder (training for 768 * 768),  my best score is 0.78861 for public LB and 0.78165 for private LB. And I use pixel size adaptation (test image resize scale = (target pixel size/ source pixel size * training scale), if you training for 768, the  training scale is 768/3000) for test dataset.",
      "votes": null
    },
    {
      "id": "1955829",
      "postDate": "09/26/2022 06:55:02",
      "content": "<p>Great methods, How do you process your training dataset?</p>",
      "rawMarkdown": "Great methods, How do you process your training dataset?",
      "votes": null
    },
    {
      "id": "1957485",
      "postDate": "09/27/2022 02:32:26",
      "content": "<p>I use 5 simple augmentations for training dataset:</p>\n<ul>\n<li>Random flip</li>\n<li>Rotate 90 degrees</li>\n<li>Random noise</li>\n<li>Random contrast</li>\n<li>Random HSV</li>\n</ul>\n<p>And all training images just resized to 768 * 768 (no tiling), also referenced from <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a>  :D</p>",
      "rawMarkdown": "I use 5 simple augmentations for training dataset:\n- Random flip\n- Rotate 90 degrees\n- Random noise\n- Random contrast\n- Random HSV\n\nAnd all training images just resized to 768 * 768 (no tiling), also referenced from @hengck23  :D",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1955390,
      "author_name": "chris666",
      "author_url": "",
      "post_date": "09/26/2022 01:35:02",
      "content": "<p>Using pvtv2-b4 backbone with conv3x3 decoder (training for 768 * 768),  my best score is 0.78861 for public LB and 0.78165 for private LB. And I use pixel size adaptation (test image resize scale = (target pixel size/ source pixel size * training scale), if you training for 768, the  training scale is 768/3000) for test dataset.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1955829,
          "author_name": "ynhuhu",
          "author_url": "",
          "post_date": "09/26/2022 06:55:02",
          "content": "<p>Great methods, How do you process your training dataset?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1957485,
          "author_name": "chris666",
          "author_url": "",
          "post_date": "09/27/2022 02:32:26",
          "content": "<p>I use 5 simple augmentations for training dataset:</p>\n<ul>\n<li>Random flip</li>\n<li>Rotate 90 degrees</li>\n<li>Random noise</li>\n<li>Random contrast</li>\n<li>Random HSV</li>\n</ul>\n<p>And all training images just resized to 768 * 768 (no tiling), also referenced from <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a>  :D</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1954535": "My best score  is 0.7568 for public LB, is 0.7580 for private LB.\nThe model is PVT-v2_unet  forked from @hengck23 notebook. The HPA images are normalized to Hubmap and are used stainaugmentor.",
    "1955390": "Using pvtv2-b4 backbone with conv3x3 decoder (training for 768 * 768),  my best score is 0.78861 for public LB and 0.78165 for private LB. And I use pixel size adaptation (test image resize scale = (target pixel size/ source pixel size * training scale), if you training for 768, the  training scale is 768/3000) for test dataset.",
    "1955829": "Great methods, How do you process your training dataset?",
    "1957485": "I use 5 simple augmentations for training dataset:\n- Random flip\n- Rotate 90 degrees\n- Random noise\n- Random contrast\n- Random HSV\n\nAnd all training images just resized to 768 * 768 (no tiling), also referenced from @hengck23  :D"
  },
  "source": "meta"
}