{
  "id": 227638,
  "title": "Training on a whole image?",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/227638",
  "author_name": "",
  "post_date": "2021-03-21T15:03:00.557721600Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p><strong>I know this isn't advised</strong> and probably hard to do in practice (due to RAM limitations) but I was wondering if anyone has been able to load few whole <strong>.tiff</strong> images instead of extracting tiles? </p>\n<p>For those that have more experience working with <a href=\"https://fr.wikipedia.org/wiki/Tagged_Image_File_Format\" target=\"_blank\">.tiff</a> images, is that something that could make sense? What are other ways to deal with very large images?</p>\n<p>Thanks for any insights. 🤓</p>",
  "messages": [
    {
      "id": "1247240",
      "postDate": "03/21/2021 15:03:00",
      "content": "<p><strong>I know this isn't advised</strong> and probably hard to do in practice (due to RAM limitations) but I was wondering if anyone has been able to load few whole <strong>.tiff</strong> images instead of extracting tiles? </p>\n<p>For those that have more experience working with <a href=\"https://fr.wikipedia.org/wiki/Tagged_Image_File_Format\" target=\"_blank\">.tiff</a> images, is that something that could make sense? What are other ways to deal with very large images?</p>\n<p>Thanks for any insights. 🤓</p>",
      "rawMarkdown": "**I know this isn't advised** and probably hard to do in practice (due to RAM limitations) but I was wondering if anyone has been able to load few whole **.tiff** images instead of extracting tiles? \n\nFor those that have more experience working with [.tiff](https://fr.wikipedia.org/wiki/Tagged_Image_File_Format) images, is that something that could make sense? What are other ways to deal with very large images?\n\nThanks for any insights. 🤓",
      "votes": null
    },
    {
      "id": "1247960",
      "postDate": "03/22/2021 08:14:41",
      "content": "<p>Most segmentations models are based on convolutional layers. And the way they work is indifferent to whether the images are given in one piece or in tiles. As long as the tiles have sufficient overlap in training, the network will learn the same patterns. See animation below:</p>\n<p>Cost aside, using multi-GPU setups will be harder because then you have to parallelize your model.<br>\n<img src=\"https://miro.medium.com/max/1000/1*GcI7G-JLAQiEoCON7xFbhg.gif\" alt=\"\"><br>\nConvoluting a 5x5x1 image with a 3x3x1 kernel to get a 3x3x1 convolved feature<br>\n<a href=\"https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-networks-the-eli5-way-3bd2b1164a53\" target=\"_blank\">Source</a></p>",
      "rawMarkdown": "Most segmentations models are based on convolutional layers. And the way they work is indifferent to whether the images are given in one piece or in tiles. As long as the tiles have sufficient overlap in training, the network will learn the same patterns. See animation below:\n\nCost aside, using multi-GPU setups will be harder because then you have to parallelize your model.\n![](https://miro.medium.com/max/1000/1*GcI7G-JLAQiEoCON7xFbhg.gif)\nConvoluting a 5x5x1 image with a 3x3x1 kernel to get a 3x3x1 convolved feature\n[Source](https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-networks-the-eli5-way-3bd2b1164a53)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1247960,
      "author_name": "erikdali",
      "author_url": "",
      "post_date": "03/22/2021 08:14:41",
      "content": "<p>Most segmentations models are based on convolutional layers. And the way they work is indifferent to whether the images are given in one piece or in tiles. As long as the tiles have sufficient overlap in training, the network will learn the same patterns. See animation below:</p>\n<p>Cost aside, using multi-GPU setups will be harder because then you have to parallelize your model.<br>\n<img src=\"https://miro.medium.com/max/1000/1*GcI7G-JLAQiEoCON7xFbhg.gif\" alt=\"\"><br>\nConvoluting a 5x5x1 image with a 3x3x1 kernel to get a 3x3x1 convolved feature<br>\n<a href=\"https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-networks-the-eli5-way-3bd2b1164a53\" target=\"_blank\">Source</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1247240": "**I know this isn't advised** and probably hard to do in practice (due to RAM limitations) but I was wondering if anyone has been able to load few whole **.tiff** images instead of extracting tiles? \n\nFor those that have more experience working with [.tiff](https://fr.wikipedia.org/wiki/Tagged_Image_File_Format) images, is that something that could make sense? What are other ways to deal with very large images?\n\nThanks for any insights. 🤓",
    "1247960": "Most segmentations models are based on convolutional layers. And the way they work is indifferent to whether the images are given in one piece or in tiles. As long as the tiles have sufficient overlap in training, the network will learn the same patterns. See animation below:\n\nCost aside, using multi-GPU setups will be harder because then you have to parallelize your model.\n![](https://miro.medium.com/max/1000/1*GcI7G-JLAQiEoCON7xFbhg.gif)\nConvoluting a 5x5x1 image with a 3x3x1 kernel to get a 3x3x1 convolved feature\n[Source](https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-networks-the-eli5-way-3bd2b1164a53)"
  },
  "source": "meta"
}