{
  "id": 201750,
  "title": "Training at different resolutions",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/201750",
  "author_name": "",
  "post_date": "2020-12-06T15:23:46.632540500Z",
  "votes": 3,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Has anyone one experienced doing training at 1024 tiles (no reduction).I have used 256 and got 0.83x on CV. Can 1024 resolution improve the results?</p>",
  "messages": [
    {
      "id": "1104072",
      "postDate": "12/06/2020 15:23:46",
      "content": "<p>Has anyone one experienced doing training at 1024 tiles (no reduction).I have used 256 and got 0.83x on CV. Can 1024 resolution improve the results?</p>",
      "rawMarkdown": "Has anyone one experienced doing training at 1024 tiles (no reduction).I have used 256 and got 0.83x on CV. Can 1024 resolution improve the results?",
      "votes": null
    },
    {
      "id": "1113597",
      "postDate": "12/15/2020 15:10:57",
      "content": "<p>Similar at 64 resolution: 0.835</p>",
      "rawMarkdown": "Similar at 64 resolution: 0.835",
      "votes": null
    },
    {
      "id": "1116274",
      "postDate": "12/17/2020 02:58:28",
      "content": "<p>you should do this test:</p>\n<ol>\n<li>truth= full resolution, predict = truth at full resolution. measure local dice score, which is equal to 1.00</li>\n<li>set predict = truth at full 0.5 resolution. then upscale your prediction to the original resolution. <br>\nmeasure local  dice score</li>\n<li>set predict = truth at full 0.25 resolution .. repeat</li>\n<li>set predict = truth at full 0.125 resolution .. repeat</li>\n</ol>\n<p>this shows the max theoretical gain if you increase resolution from 512 to say 1024 …<br>\nthen you have to think about efficiency and efficacy … or the priority of methods</p>",
      "rawMarkdown": "you should do this test:\n1. truth= full resolution, predict = truth at full resolution. measure local dice score, which is equal to 1.00\n2. set predict = truth at full 0.5 resolution. then upscale your prediction to the original resolution. \nmeasure local  dice score\n3. set predict = truth at full 0.25 resolution .. repeat\n4. set predict = truth at full 0.125 resolution .. repeat\n\n\nthis shows the max theoretical gain if you increase resolution from 512 to say 1024 ...\nthen you have to think about efficiency and efficacy ... or the priority of methods",
      "votes": null
    },
    {
      "id": "1116602",
      "postDate": "12/17/2020 10:07:28",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a>, in your 2nd bullet point, you mean feeding 0.5 resolution images for inference ==&gt; x2 to predicted masks ==&gt; measure dice loss w.r.t full resolution groud-truth masks? Am I getting what you're saying here?</p>\n<p>Thank you!</p>",
      "rawMarkdown": "Hi @hengck23, in your 2nd bullet point, you mean feeding 0.5 resolution images for inference ==> x2 to predicted masks ==> measure dice loss w.r.t full resolution groud-truth masks? Am I getting what you're saying here?\n\nThank you!",
      "votes": null
    },
    {
      "id": "1117069",
      "postDate": "12/17/2020 17:45:36",
      "content": "<p>assume you want to train a model at 600x600 patch. what is the best you can get? it is the ground truth. hence best prediction is ground truth at 600x600, which is created by downsizing the original ground truth.</p>\n<p>but to make a submission, you need to upsize your best prediction to the submission size (the original size).</p>\n<p>hence you downsample and upsample to get the best theoretical results.</p>",
      "rawMarkdown": "assume you want to train a model at 600x600 patch. what is the best you can get? it is the ground truth. hence best prediction is ground truth at 600x600, which is created by downsizing the original ground truth.\n\nbut to make a submission, you need to upsize your best prediction to the submission size (the original size).\n\nhence you downsample and upsample to get the best theoretical results.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1113597,
      "author_name": "andrasferenczi",
      "author_url": "",
      "post_date": "12/15/2020 15:10:57",
      "content": "<p>Similar at 64 resolution: 0.835</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1116274,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "12/17/2020 02:58:28",
      "content": "<p>you should do this test:</p>\n<ol>\n<li>truth= full resolution, predict = truth at full resolution. measure local dice score, which is equal to 1.00</li>\n<li>set predict = truth at full 0.5 resolution. then upscale your prediction to the original resolution. <br>\nmeasure local  dice score</li>\n<li>set predict = truth at full 0.25 resolution .. repeat</li>\n<li>set predict = truth at full 0.125 resolution .. repeat</li>\n</ol>\n<p>this shows the max theoretical gain if you increase resolution from 512 to say 1024 …<br>\nthen you have to think about efficiency and efficacy … or the priority of methods</p>",
      "votes": null,
      "replies": [
        {
          "id": 1116602,
          "author_name": "fuckvenkatraman",
          "author_url": "",
          "post_date": "12/17/2020 10:07:28",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a>, in your 2nd bullet point, you mean feeding 0.5 resolution images for inference ==&gt; x2 to predicted masks ==&gt; measure dice loss w.r.t full resolution groud-truth masks? Am I getting what you're saying here?</p>\n<p>Thank you!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1117069,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "12/17/2020 17:45:36",
          "content": "<p>assume you want to train a model at 600x600 patch. what is the best you can get? it is the ground truth. hence best prediction is ground truth at 600x600, which is created by downsizing the original ground truth.</p>\n<p>but to make a submission, you need to upsize your best prediction to the submission size (the original size).</p>\n<p>hence you downsample and upsample to get the best theoretical results.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1104072": "Has anyone one experienced doing training at 1024 tiles (no reduction).I have used 256 and got 0.83x on CV. Can 1024 resolution improve the results?",
    "1113597": "Similar at 64 resolution: 0.835",
    "1116274": "you should do this test:\n1. truth= full resolution, predict = truth at full resolution. measure local dice score, which is equal to 1.00\n2. set predict = truth at full 0.5 resolution. then upscale your prediction to the original resolution. \nmeasure local  dice score\n3. set predict = truth at full 0.25 resolution .. repeat\n4. set predict = truth at full 0.125 resolution .. repeat\n\n\nthis shows the max theoretical gain if you increase resolution from 512 to say 1024 ...\nthen you have to think about efficiency and efficacy ... or the priority of methods",
    "1116602": "Hi @hengck23, in your 2nd bullet point, you mean feeding 0.5 resolution images for inference ==> x2 to predicted masks ==> measure dice loss w.r.t full resolution groud-truth masks? Am I getting what you're saying here?\n\nThank you!",
    "1117069": "assume you want to train a model at 600x600 patch. what is the best you can get? it is the ground truth. hence best prediction is ground truth at 600x600, which is created by downsizing the original ground truth.\n\nbut to make a submission, you need to upsize your best prediction to the submission size (the original size).\n\nhence you downsample and upsample to get the best theoretical results."
  },
  "source": "meta"
}