{
  "id": 343117,
  "title": "Prediction on Test Image",
  "url": "/competitions/hubmap-organ-segmentation/discussion/343117",
  "author_name": "Quan",
  "post_date": "2022-08-10T04:21:20.451000",
  "votes": 10,
  "comment_count": 20,
  "views": 0,
  "content": "<p>I am interested to see what our models predict on that one test image provided by the host. I will go first<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3436753%2F1d06bcee8e411b5efc8e5be2bb372349%2F__results___11_0.png?generation=1660105256147569&amp;alt=media\" alt=\"\"> </p>",
  "messages": [
    {
      "id": 1892362,
      "postDate": "2022-08-10T04:21:20.453Z",
      "content": "<p>I am interested to see what our models predict on that one test image provided by the host. I will go first<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3436753%2F1d06bcee8e411b5efc8e5be2bb372349%2F__results___11_0.png?generation=1660105256147569&amp;alt=media\" alt=\"\"> </p>",
      "rawMarkdown": "I am interested to see what our models predict on that one test image provided by the host. I will go first![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3436753%2F1d06bcee8e411b5efc8e5be2bb372349%2F__results___11_0.png?generation=1660105256147569&alt=media) ",
      "votes": 10
    },
    {
      "id": 1900744,
      "postDate": "2022-08-16T08:16:24.123Z",
      "content": "<p>if you really really want to know what is ground truth, here, is what you can do.<br>\nyou probably needs about 10 submission slots</p>\n<p><img src=\"https://i.ibb.co/FqH8SLJ/Selection-064.png\" alt=\"https://i.ibb.co/FqH8SLJ/Selection-064.png\"></p>\n<ul>\n<li>sort all objects according to score  </li>\n<li>candidates = { null }</li>\n<li>add most probable object and submit. did score improve?  <br>\nyes: candidates = candidates + new object  <br>\nno:  candidates = candidates  </li>\n<li>repeat with next object  </li>\n</ul>",
      "rawMarkdown": "if you really really want to know what is ground truth, here, is what you can do.\nyou probably needs about 10 submission slots\n\n![https://i.ibb.co/FqH8SLJ/Selection-064.png](https://i.ibb.co/FqH8SLJ/Selection-064.png)\n\n- sort all objects according to score  \n- candidates = { null }\n- add most probable object and submit. did score improve?  \nyes: candidates = candidates + new object  \nno:  candidates = candidates  \n- repeat with next object  \n",
      "votes": 3
    },
    {
      "id": 1893797,
      "postDate": "2022-08-11T04:53:55.243Z",
      "content": "<p>my efficientnet4 + deeplabv3 , lb 58 case<br>\n<img src=\"https://i.imgur.com/7EPLSG8.png\" alt=\"\"></p>",
      "rawMarkdown": "my efficientnet4 + deeplabv3 , lb 58 case\n![](https://i.imgur.com/7EPLSG8.png)",
      "votes": 1
    },
    {
      "id": 1892576,
      "postDate": "2022-08-10T07:11:39.873Z",
      "content": "<p>For shit and giggles here are my results… [work in progress] :)</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2386017%2F5f2572546578c55a28e78fce668b06f4%2F10078%20(1).jpeg?generation=1660115448759146&amp;alt=media\"></p>",
      "rawMarkdown": "For shit and giggles here are my results... [work in progress] :)\n\n<img width=\"300\" src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2386017%2F5f2572546578c55a28e78fce668b06f4%2F10078%20(1).jpeg?generation=1660115448759146&alt=media\"></img>",
      "votes": 1,
      "replies": [
        {
          "id": 1892592,
          "postDate": "2022-08-10T07:17:53.863Z",
          "content": "<p>I think our models agree on these parts<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3436753%2Fce49b7b470297b98dd06ae9276baa3e9%2Finbox_2386017_5f2572546578c55a28e78fce668b06f4_10078%20(1).jpeg?generation=1660115867989448&amp;alt=media\" alt=\"\"></p>",
          "rawMarkdown": "I think our models agree on these parts![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3436753%2Fce49b7b470297b98dd06ae9276baa3e9%2Finbox_2386017_5f2572546578c55a28e78fce668b06f4_10078%20(1).jpeg?generation=1660115867989448&alt=media)"
        },
        {
          "id": 1892604,
          "postDate": "2022-08-10T07:25:35.210Z",
          "content": "<p>It seems that what matters here are the density of the blue dots in the image, but my model is not so sharp picking those up.<br>\nYour seems much better on picking those dots up. 🤔</p>\n<p>Oh btw when you upload image you can use HTML <code>&lt;img width=\"300\" src=\"IMG.jpeg\"&gt;&lt;/img&gt;</code> </p>",
          "rawMarkdown": "It seems that what matters here are the density of the blue dots in the image, but my model is not so sharp picking those up.\nYour seems much better on picking those dots up. 🤔\n\nOh btw when you upload image you can use HTML `<img width=\"300\" src=\"IMG.jpeg\"></img>` "
        },
        {
          "id": 1892607,
          "postDate": "2022-08-10T07:27:14.853Z",
          "content": "<p>More like structured clusters of blue dots</p>",
          "rawMarkdown": "More like structured clusters of blue dots"
        },
        {
          "id": 1892968,
          "postDate": "2022-08-10T13:11:52.713Z",
          "content": "<p><img src=\"https://i.ibb.co/d5DSFzD/Selection-095.png\" alt=\"https://i.ibb.co/d5DSFzD/Selection-095.png\"></p>\n<p>open up in gimp and extract color component CYMK yellow</p>",
          "rawMarkdown": "![https://i.ibb.co/d5DSFzD/Selection-095.png](https://i.ibb.co/d5DSFzD/Selection-095.png)\n\nopen up in gimp and extract color component CYMK yellow"
        },
        {
          "id": 1893364,
          "postDate": "2022-08-10T18:32:27.120Z",
          "content": "<p>That's beautiful image, you can really see the dot clusters. You get a like from me 👍</p>",
          "rawMarkdown": "That's beautiful image, you can really see the dot clusters. You get a like from me 👍"
        },
        {
          "id": 1893820,
          "postDate": "2022-08-11T05:23:05.103Z",
          "content": "<p><img src=\"https://i.ibb.co/drbNG6y/Selection-060.png\" alt=\"https://i.ibb.co/drbNG6y/Selection-060.png\"></p>",
          "rawMarkdown": "![https://i.ibb.co/drbNG6y/Selection-060.png](https://i.ibb.co/drbNG6y/Selection-060.png)",
          "votes": 1
        },
        {
          "id": 1893905,
          "postDate": "2022-08-11T06:40:30.120Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1893915,
          "postDate": "2022-08-11T06:48:49.350Z",
          "content": "<p>I'm doing tilling with 50% overlaying and then averaging mask values. But I might consider go with joust resizing original image  on 1024 since I don't get submission points boost if I'm tilling.</p>\n<p><a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> Do you do your stuff all without tilling? I don't know if its good idea to go over 1/2 scaling factor, and my GPU is not so big as I would like.</p>",
          "rawMarkdown": "I'm doing tilling with 50% overlaying and then averaging mask values. But I might consider go with joust resizing original image  on 1024 since I don't get submission points boost if I'm tilling.\n\n@hengck23 Do you do your stuff all without tilling? I don't know if its good idea to go over 1/2 scaling factor, and my GPU is not so big as I would like."
        },
        {
          "id": 1893931,
          "postDate": "2022-08-11T07:04:40.650Z",
          "content": "<p>\"and my GPU is not so big as I would like.\"</p>\n<p>pretrain in tiles, finetune in full in the last step.<br>\nuse  gradient checking method to save GPU memory</p>",
          "rawMarkdown": "\"and my GPU is not so big as I would like.\"\n\npretrain in tiles, finetune in full in the last step.\nuse  gradient checking method to save GPU memory",
          "votes": 1
        }
      ]
    },
    {
      "id": 1907785,
      "postDate": "2022-08-21T04:38:56.097Z",
      "content": "<p>BTW, this image is a spleen sample, isn't it? </p>",
      "rawMarkdown": "BTW, this image is a spleen sample, isn't it? "
    },
    {
      "id": 1892553,
      "postDate": "2022-08-10T06:45:43.773Z",
      "content": "<p>while prediction is \"correct\" but i am not sure if small white pulp region would be annotated as ground truth.<br>\nwe will know later when the no. of decimal place for lb score is increase to 3.</p>",
      "rawMarkdown": "while prediction is \"correct\" but i am not sure if small white pulp region would be annotated as ground truth.\nwe will know later when the no. of decimal place for lb score is increase to 3.",
      "replies": [
        {
          "id": 1892595,
          "postDate": "2022-08-10T07:19:33.623Z",
          "content": "<p>I was also thinking about submitting only this test image to get dice score but then 2 decimal places are probably not enough</p>",
          "rawMarkdown": "I was also thinking about submitting only this test image to get dice score but then 2 decimal places are probably not enough"
        },
        {
          "id": 1892609,
          "postDate": "2022-08-10T07:31:48.027Z",
          "content": "<p>But how do you know if the testing image is really in the hidden dataset?<br>\nI would bet that its not.</p>",
          "rawMarkdown": "But how do you know if the testing image is really in the hidden dataset?\nI would bet that its not.\n"
        },
        {
          "id": 1892970,
          "postDate": "2022-08-10T13:14:30.057Z",
          "content": "<p>you can do a probe (check for filename, image size, some pixel value, etc, tiff tag)</p>",
          "rawMarkdown": "you can do a probe (check for filename, image size, some pixel value, etc, tiff tag)"
        },
        {
          "id": 1893369,
          "postDate": "2022-08-10T18:38:02.403Z",
          "content": "<p>I have checked if they are doing client side rounding of submission score, but unfortunately they doing rounding server side :(<br>\nHere is response from the server.</p>\n<pre><code>{\n  \"submissions\": [\n    {\n      \"id\": :),\n      \"name\": \"hubmap-organ-segmentation\",\n      \"details\": \"Notebook hubmap-organ-segmentation | baseline\",\n      \"teamName\": \"Uros Jarc\",\n      \"publicScore\": \"0.38\",\n      \"status\": \"COMPLETE\",\n      \"version\": 103,\n      \"latestVersion\": 103,\n      \"syncRerunState\": {\n        \"id\": \"PUBLISHED\",\n        \"name\": \"RERUN_STATE_NAME_PUBLISHED\",\n        \"sourceSubmissionStatus\": \"COMPLETE\"\n      }\n    },\n</code></pre>",
          "rawMarkdown": "I have checked if they are doing client side rounding of submission score, but unfortunately they doing rounding server side :(\nHere is response from the server.\n```\n{\n  \"submissions\": [\n    {\n      \"id\": :),\n      \"name\": \"hubmap-organ-segmentation\",\n      \"details\": \"Notebook hubmap-organ-segmentation | baseline\",\n      \"teamName\": \"Uros Jarc\",\n      \"publicScore\": \"0.38\",\n      \"status\": \"COMPLETE\",\n      \"version\": 103,\n      \"latestVersion\": 103,\n      \"syncRerunState\": {\n        \"id\": \"PUBLISHED\",\n        \"name\": \"RERUN_STATE_NAME_PUBLISHED\",\n        \"sourceSubmissionStatus\": \"COMPLETE\"\n      }\n    },\n```"
        },
        {
          "id": 1893824,
          "postDate": "2022-08-11T05:27:19.317Z",
          "content": "<p>if you want to check if \"10078.tiff\" is in LB test set, you should do something like</p>\n<pre><code>your sibmission code\n\ntest_df = pd.read_csv('test.csv')\nif test_df's id does not contain  \"10078\", raise an exception error\n</code></pre>",
          "rawMarkdown": "if you want to check if \"10078.tiff\" is in LB test set, you should do something like\n\n```\nyour sibmission code\n\ntest_df = pd.read_csv('test.csv')\nif test_df's id does not contain  \"10078\", raise an exception error\n\n\n```"
        },
        {
          "id": 1893926,
          "postDate": "2022-08-11T06:58:09.933Z",
          "content": "<p>Yes I could do that, but I don't want to waste 1 submission joust for getting an error :P</p>",
          "rawMarkdown": "Yes I could do that, but I don't want to waste 1 submission joust for getting an error :P"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1900744,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2022-08-16T08:16:24.123000",
      "content": "<p>if you really really want to know what is ground truth, here, is what you can do.<br>\nyou probably needs about 10 submission slots</p>\n<p><img src=\"https://i.ibb.co/FqH8SLJ/Selection-064.png\" alt=\"https://i.ibb.co/FqH8SLJ/Selection-064.png\"></p>\n<ul>\n<li>sort all objects according to score  </li>\n<li>candidates = { null }</li>\n<li>add most probable object and submit. did score improve?  <br>\nyes: candidates = candidates + new object  <br>\nno:  candidates = candidates  </li>\n<li>repeat with next object  </li>\n</ul>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 1893797,
      "author_name": "kaggler",
      "author_url": "",
      "post_date": "2022-08-11T04:53:55.243000",
      "content": "<p>my efficientnet4 + deeplabv3 , lb 58 case<br>\n<img src=\"https://i.imgur.com/7EPLSG8.png\" alt=\"\"></p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1892576,
      "author_name": "Uros Jarc",
      "author_url": "",
      "post_date": "2022-08-10T07:11:39.873000",
      "content": "<p>For shit and giggles here are my results… [work in progress] :)</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2386017%2F5f2572546578c55a28e78fce668b06f4%2F10078%20(1).jpeg?generation=1660115448759146&amp;alt=media\"></p>",
      "votes": 1,
      "replies": [
        {
          "id": 1892592,
          "author_name": "Quan",
          "author_url": "",
          "post_date": "2022-08-10T07:17:53.863000",
          "content": "<p>I think our models agree on these parts<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3436753%2Fce49b7b470297b98dd06ae9276baa3e9%2Finbox_2386017_5f2572546578c55a28e78fce668b06f4_10078%20(1).jpeg?generation=1660115867989448&amp;alt=media\" alt=\"\"></p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1892604,
          "author_name": "Uros Jarc",
          "author_url": "",
          "post_date": "2022-08-10T07:25:35.210000",
          "content": "<p>It seems that what matters here are the density of the blue dots in the image, but my model is not so sharp picking those up.<br>\nYour seems much better on picking those dots up. 🤔</p>\n<p>Oh btw when you upload image you can use HTML <code>&lt;img width=\"300\" src=\"IMG.jpeg\"&gt;&lt;/img&gt;</code> </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1892607,
          "author_name": "Quan",
          "author_url": "",
          "post_date": "2022-08-10T07:27:14.853000",
          "content": "<p>More like structured clusters of blue dots</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1892968,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2022-08-10T13:11:52.713000",
          "content": "<p><img src=\"https://i.ibb.co/d5DSFzD/Selection-095.png\" alt=\"https://i.ibb.co/d5DSFzD/Selection-095.png\"></p>\n<p>open up in gimp and extract color component CYMK yellow</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1893364,
          "author_name": "Uros Jarc",
          "author_url": "",
          "post_date": "2022-08-10T18:32:27.120000",
          "content": "<p>That's beautiful image, you can really see the dot clusters. You get a like from me 👍</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1893820,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2022-08-11T05:23:05.103000",
          "content": "<p><img src=\"https://i.ibb.co/drbNG6y/Selection-060.png\" alt=\"https://i.ibb.co/drbNG6y/Selection-060.png\"></p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1893905,
          "author_name": "",
          "author_url": "",
          "post_date": "2022-08-11T06:40:30.120000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1893915,
          "author_name": "Uros Jarc",
          "author_url": "",
          "post_date": "2022-08-11T06:48:49.350000",
          "content": "<p>I'm doing tilling with 50% overlaying and then averaging mask values. But I might consider go with joust resizing original image  on 1024 since I don't get submission points boost if I'm tilling.</p>\n<p><a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> Do you do your stuff all without tilling? I don't know if its good idea to go over 1/2 scaling factor, and my GPU is not so big as I would like.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1893931,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2022-08-11T07:04:40.650000",
          "content": "<p>\"and my GPU is not so big as I would like.\"</p>\n<p>pretrain in tiles, finetune in full in the last step.<br>\nuse  gradient checking method to save GPU memory</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1907785,
      "author_name": "Roni Heka",
      "author_url": "",
      "post_date": "2022-08-21T04:38:56.097000",
      "content": "<p>BTW, this image is a spleen sample, isn't it? </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1892553,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2022-08-10T06:45:43.773000",
      "content": "<p>while prediction is \"correct\" but i am not sure if small white pulp region would be annotated as ground truth.<br>\nwe will know later when the no. of decimal place for lb score is increase to 3.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1892595,
          "author_name": "Quan",
          "author_url": "",
          "post_date": "2022-08-10T07:19:33.623000",
          "content": "<p>I was also thinking about submitting only this test image to get dice score but then 2 decimal places are probably not enough</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1892609,
          "author_name": "Uros Jarc",
          "author_url": "",
          "post_date": "2022-08-10T07:31:48.027000",
          "content": "<p>But how do you know if the testing image is really in the hidden dataset?<br>\nI would bet that its not.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1892970,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2022-08-10T13:14:30.057000",
          "content": "<p>you can do a probe (check for filename, image size, some pixel value, etc, tiff tag)</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1893369,
          "author_name": "Uros Jarc",
          "author_url": "",
          "post_date": "2022-08-10T18:38:02.403000",
          "content": "<p>I have checked if they are doing client side rounding of submission score, but unfortunately they doing rounding server side :(<br>\nHere is response from the server.</p>\n<pre><code>{\n  \"submissions\": [\n    {\n      \"id\": :),\n      \"name\": \"hubmap-organ-segmentation\",\n      \"details\": \"Notebook hubmap-organ-segmentation | baseline\",\n      \"teamName\": \"Uros Jarc\",\n      \"publicScore\": \"0.38\",\n      \"status\": \"COMPLETE\",\n      \"version\": 103,\n      \"latestVersion\": 103,\n      \"syncRerunState\": {\n        \"id\": \"PUBLISHED\",\n        \"name\": \"RERUN_STATE_NAME_PUBLISHED\",\n        \"sourceSubmissionStatus\": \"COMPLETE\"\n      }\n    },\n</code></pre>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1893824,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2022-08-11T05:27:19.317000",
          "content": "<p>if you want to check if \"10078.tiff\" is in LB test set, you should do something like</p>\n<pre><code>your sibmission code\n\ntest_df = pd.read_csv('test.csv')\nif test_df's id does not contain  \"10078\", raise an exception error\n</code></pre>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1893926,
          "author_name": "Uros Jarc",
          "author_url": "",
          "post_date": "2022-08-11T06:58:09.933000",
          "content": "<p>Yes I could do that, but I don't want to waste 1 submission joust for getting an error :P</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1892362": "I am interested to see what our models predict on that one test image provided by the host. I will go first![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3436753%2F1d06bcee8e411b5efc8e5be2bb372349%2F__results___11_0.png?generation=1660105256147569&alt=media) ",
    "1900744": "if you really really want to know what is ground truth, here, is what you can do.\nyou probably needs about 10 submission slots\n\n![https://i.ibb.co/FqH8SLJ/Selection-064.png](https://i.ibb.co/FqH8SLJ/Selection-064.png)\n\n- sort all objects according to score  \n- candidates = { null }\n- add most probable object and submit. did score improve?  \nyes: candidates = candidates + new object  \nno:  candidates = candidates  \n- repeat with next object  \n",
    "1893797": "my efficientnet4 + deeplabv3 , lb 58 case\n![](https://i.imgur.com/7EPLSG8.png)",
    "1892576": "For shit and giggles here are my results... [work in progress] :)\n\n<img width=\"300\" src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2386017%2F5f2572546578c55a28e78fce668b06f4%2F10078%20(1).jpeg?generation=1660115448759146&alt=media\"></img>",
    "1907785": "BTW, this image is a spleen sample, isn't it? ",
    "1892553": "while prediction is \"correct\" but i am not sure if small white pulp region would be annotated as ground truth.\nwe will know later when the no. of decimal place for lb score is increase to 3."
  }
}