{
  "id": 145155,
  "title": "Resized PNG",
  "url": "/competitions/prostate-cancer-grade-assessment/discussion/145155",
  "author_name": "xhlulu",
  "post_date": "2020-04-22T04:37:25.206000",
  "votes": 9,
  "comment_count": 7,
  "views": 0,
  "content": "<p>I think as a first step it's useful to have the images as resized PNGs, so I made a notebook about it. I also made a dataset in case people wanted to use it on TPUs: <a href=\"https://www.kaggle.com/xhlulu/panda-resized-train-data-512x512\">https://www.kaggle.com/xhlulu/panda-resized-train-data-512x512</a></p>",
  "messages": [
    {
      "id": 816056,
      "postDate": "2020-04-22T04:37:25.207Z",
      "content": "<p>I think as a first step it's useful to have the images as resized PNGs, so I made a notebook about it. I also made a dataset in case people wanted to use it on TPUs: <a href=\"https://www.kaggle.com/xhlulu/panda-resized-train-data-512x512\">https://www.kaggle.com/xhlulu/panda-resized-train-data-512x512</a></p>",
      "rawMarkdown": "I think as a first step it's useful to have the images as resized PNGs, so I made a notebook about it. I also made a dataset in case people wanted to use it on TPUs: https://www.kaggle.com/xhlulu/panda-resized-train-data-512x512\n\n",
      "votes": 9
    },
    {
      "id": 828647,
      "postDate": "2020-05-01T07:48:03.037Z",
      "content": "<p>Why your 1024x1024 dataset overall size less than 512x512?</p>\n\n<p>1024x1024 - 862M\n512x512.    - 1.02G</p>",
      "rawMarkdown": "Why your 1024x1024 dataset overall size less than 512x512?\n\n1024x1024 - 862M\n512x512.    - 1.02G",
      "replies": [
        {
          "id": 828700,
          "postDate": "2020-05-01T08:18:44.730Z",
          "content": "<p>png vs. jpeg maybe ? </p>",
          "rawMarkdown": "png vs. jpeg maybe ? ",
          "votes": 1
        }
      ]
    },
    {
      "id": 816533,
      "postDate": "2020-04-22T12:00:50.630Z",
      "content": "<p>Hi <a href=\"/xhlulu\">@xhlulu</a> !</p>\n\n<p>Thank you for putting this dataset together. May I ask what exactly was the strategy for having very large images become 512x512 png files ? Did you crop ? Downsample ? If you cropped, how did you choose which part to crop ?</p>\n\n<p>Thanks a lot.</p>",
      "rawMarkdown": "Hi @xhlulu !\n\nThank you for putting this dataset together. May I ask what exactly was the strategy for having very large images become 512x512 png files ? Did you crop ? Downsample ? If you cropped, how did you choose which part to crop ?\n\nThanks a lot.",
      "replies": [
        {
          "id": 816697,
          "postDate": "2020-04-22T14:20:11.843Z",
          "content": "<p>I resized it without cropping. It's not ideal but it's an easy start.</p>",
          "rawMarkdown": "I resized it without cropping. It's not ideal but it's an easy start.",
          "votes": 1
        },
        {
          "id": 816726,
          "postDate": "2020-04-22T14:35:49.687Z",
          "content": "<p>Thanks for your answer :)</p>\n\n<p>Okay, then maybe there is a catch. The original images might not be the same level of zoom, depending on whether they come from karolinska or radboud. I quickly analyzed your pictures in this kernel:</p>\n\n<p><a href=\"https://www.kaggle.com/bdubreu/differences-between-radboud-and-karolinska-data\">https://www.kaggle.com/bdubreu/differences-between-radboud-and-karolinska-data</a></p>\n\n<p>what I find is that a simple resnet50 gets 100% accuracy at predicting whether a pic comes from radboud or karolinska. I have a few ideas why it might be the case. I list them at the end of kernel. Any other thoughts ? </p>",
          "rawMarkdown": "Thanks for your answer :)\n\nOkay, then maybe there is a catch. The original images might not be the same level of zoom, depending on whether they come from karolinska or radboud. I quickly analyzed your pictures in this kernel:\n\nhttps://www.kaggle.com/bdubreu/differences-between-radboud-and-karolinska-data\n\nwhat I find is that a simple resnet50 gets 100% accuracy at predicting whether a pic comes from radboud or karolinska. I have a few ideas why it might be the case. I list them at the end of kernel. Any other thoughts ? "
        },
        {
          "id": 816732,
          "postDate": "2020-04-22T14:40:52.930Z",
          "content": "<p>I think that's definitely something bound to happen, whether you run it on the entire picture or on a downsized picture. Instead, I'd recommend breaking down the original images into sections of 512x512 pixels, and individually see if you can differentiate them (predict source, or predict the isup score).</p>",
          "rawMarkdown": "I think that's definitely something bound to happen, whether you run it on the entire picture or on a downsized picture. Instead, I'd recommend breaking down the original images into sections of 512x512 pixels, and individually see if you can differentiate them (predict source, or predict the isup score).",
          "votes": 1
        }
      ]
    },
    {
      "id": 816058,
      "postDate": "2020-04-22T04:37:30.960Z",
      "content": "<p>Dear organizers, please let me know if theres any problem with that. Let me know if you wish to be added as data owners.</p>",
      "rawMarkdown": "Dear organizers, please let me know if theres any problem with that. Let me know if you wish to be added as data owners."
    }
  ],
  "comments": [
    {
      "id": 828647,
      "author_name": "Aimoldin Anuar [dsmlkz]",
      "author_url": "",
      "post_date": "2020-05-01T07:48:03.037000",
      "content": "<p>Why your 1024x1024 dataset overall size less than 512x512?</p>\n\n<p>1024x1024 - 862M\n512x512.    - 1.02G</p>",
      "votes": 0,
      "replies": [
        {
          "id": 828700,
          "author_name": "Benjamin Dubreu",
          "author_url": "",
          "post_date": "2020-05-01T08:18:44.730000",
          "content": "<p>png vs. jpeg maybe ? </p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 816533,
      "author_name": "Benjamin Dubreu",
      "author_url": "",
      "post_date": "2020-04-22T12:00:50.630000",
      "content": "<p>Hi <a href=\"/xhlulu\">@xhlulu</a> !</p>\n\n<p>Thank you for putting this dataset together. May I ask what exactly was the strategy for having very large images become 512x512 png files ? Did you crop ? Downsample ? If you cropped, how did you choose which part to crop ?</p>\n\n<p>Thanks a lot.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 816697,
          "author_name": "xhlulu",
          "author_url": "",
          "post_date": "2020-04-22T14:20:11.843000",
          "content": "<p>I resized it without cropping. It's not ideal but it's an easy start.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 816726,
          "author_name": "Benjamin Dubreu",
          "author_url": "",
          "post_date": "2020-04-22T14:35:49.687000",
          "content": "<p>Thanks for your answer :)</p>\n\n<p>Okay, then maybe there is a catch. The original images might not be the same level of zoom, depending on whether they come from karolinska or radboud. I quickly analyzed your pictures in this kernel:</p>\n\n<p><a href=\"https://www.kaggle.com/bdubreu/differences-between-radboud-and-karolinska-data\">https://www.kaggle.com/bdubreu/differences-between-radboud-and-karolinska-data</a></p>\n\n<p>what I find is that a simple resnet50 gets 100% accuracy at predicting whether a pic comes from radboud or karolinska. I have a few ideas why it might be the case. I list them at the end of kernel. Any other thoughts ? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 816732,
          "author_name": "xhlulu",
          "author_url": "",
          "post_date": "2020-04-22T14:40:52.930000",
          "content": "<p>I think that's definitely something bound to happen, whether you run it on the entire picture or on a downsized picture. Instead, I'd recommend breaking down the original images into sections of 512x512 pixels, and individually see if you can differentiate them (predict source, or predict the isup score).</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 816058,
      "author_name": "xhlulu",
      "author_url": "",
      "post_date": "2020-04-22T04:37:30.960000",
      "content": "<p>Dear organizers, please let me know if theres any problem with that. Let me know if you wish to be added as data owners.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "816056": "I think as a first step it's useful to have the images as resized PNGs, so I made a notebook about it. I also made a dataset in case people wanted to use it on TPUs: https://www.kaggle.com/xhlulu/panda-resized-train-data-512x512\n\n",
    "828647": "Why your 1024x1024 dataset overall size less than 512x512?\n\n1024x1024 - 862M\n512x512.    - 1.02G",
    "816533": "Hi @xhlulu !\n\nThank you for putting this dataset together. May I ask what exactly was the strategy for having very large images become 512x512 png files ? Did you crop ? Downsample ? If you cropped, how did you choose which part to crop ?\n\nThanks a lot.",
    "816058": "Dear organizers, please let me know if theres any problem with that. Let me know if you wish to be added as data owners."
  }
}