{
  "id": 234122,
  "title": "Question about piecing together submission mask",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/234122",
  "author_name": "",
  "post_date": "2021-04-22T18:50:10.039711100Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I just join this competition. Based on all the posts I viewed, it seems that people use cropped images to train instead of the entire complete image. Does this mean the testing prediction also require one to put together the entire prediction mask from pieces?</p>",
  "messages": [
    {
      "id": "1281268",
      "postDate": "04/22/2021 18:50:10",
      "content": "<p>I just join this competition. Based on all the posts I viewed, it seems that people use cropped images to train instead of the entire complete image. Does this mean the testing prediction also require one to put together the entire prediction mask from pieces?</p>",
      "rawMarkdown": "I just join this competition. Based on all the posts I viewed, it seems that people use cropped images to train instead of the entire complete image. Does this mean the testing prediction also require one to put together the entire prediction mask from pieces?",
      "votes": null
    },
    {
      "id": "1281283",
      "postDate": "04/22/2021 19:01:22",
      "content": "<p>That's the way that everyone, as far as I know, is doing it, yep.</p>\n<p>There are lots of excellent training and submission kernels outlying efficient ways of doing it.</p>",
      "rawMarkdown": "That's the way that everyone, as far as I know, is doing it, yep.\n\nThere are lots of excellent training and submission kernels outlying efficient ways of doing it.",
      "votes": null
    },
    {
      "id": "1281898",
      "postDate": "04/23/2021 12:12:41",
      "content": "<p>Yup that is the only way we can do it. Tiles of either 256, 512 or 1024. <br>\nIn that way you don't have to store the entire dataset i.e. image of the kidney in memory and it will give memory out errors. It is quite a huge dataset to be working on the entire image at once.<br>\nSo, perform prediction of tiles and then piece it back together.</p>",
      "rawMarkdown": "Yup that is the only way we can do it. Tiles of either 256, 512 or 1024. \nIn that way you don't have to store the entire dataset i.e. image of the kidney in memory and it will give memory out errors. It is quite a huge dataset to be working on the entire image at once.\nSo, perform prediction of tiles and then piece it back together.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1281283,
      "author_name": "jamesphoward",
      "author_url": "",
      "post_date": "04/22/2021 19:01:22",
      "content": "<p>That's the way that everyone, as far as I know, is doing it, yep.</p>\n<p>There are lots of excellent training and submission kernels outlying efficient ways of doing it.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1281898,
      "author_name": "ckanth090",
      "author_url": "",
      "post_date": "04/23/2021 12:12:41",
      "content": "<p>Yup that is the only way we can do it. Tiles of either 256, 512 or 1024. <br>\nIn that way you don't have to store the entire dataset i.e. image of the kidney in memory and it will give memory out errors. It is quite a huge dataset to be working on the entire image at once.<br>\nSo, perform prediction of tiles and then piece it back together.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1281268": "I just join this competition. Based on all the posts I viewed, it seems that people use cropped images to train instead of the entire complete image. Does this mean the testing prediction also require one to put together the entire prediction mask from pieces?",
    "1281283": "That's the way that everyone, as far as I know, is doing it, yep.\n\nThere are lots of excellent training and submission kernels outlying efficient ways of doing it.",
    "1281898": "Yup that is the only way we can do it. Tiles of either 256, 512 or 1024. \nIn that way you don't have to store the entire dataset i.e. image of the kidney in memory and it will give memory out errors. It is quite a huge dataset to be working on the entire image at once.\nSo, perform prediction of tiles and then piece it back together."
  },
  "source": "meta"
}