{
  "id": 280017,
  "title": "255th place solution - Using 3D Resnet50 model",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/writeups/aryaman-sharma-255th-place-solution-using-3d-resne",
  "author_name": "",
  "post_date": "2021-10-20T02:24:56.458270300Z",
  "votes": 9,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Although my model didn't generalise i'd still like to summarise my approach, this being my first kaggle competition.</p>\n<h1>Data preprocessing steps</h1>\n<ul>\n<li>Load T2w Images</li>\n<li>Removed Blank slices </li>\n<li>Crop slices </li>\n<li>SIZ [1] depth 64</li>\n<li>Normalise</li>\n<li>Zero centre</li>\n</ul>\n<h1>Training</h1>\n<ul>\n<li>Trained a Resnet50 3D with a LR scheduler (factor 10, tolerance 5) Initial LR=1e-2, batch size = 16</li>\n<li>Single fold cross validation<br>\nGenerally the results appeared like:<br>\n<a href=\"https://postimg.cc/QHzXL2c8\" target=\"_blank\">Screenshot-from-2021-10-20-12-20-00.png</a><br>\nChose best 5 models from each training cycle and created a final ensemble for submission</li>\n</ul>\n<p>[1] <a href=\"https://link.springer.com/chapter/10.1007%2F978-3-030-59354-4_15\" target=\"_blank\">https://link.springer.com/chapter/10.1007%2F978-3-030-59354-4_15</a></p>",
  "messages": [
    {
      "id": "1550767",
      "postDate": "10/20/2021 02:24:56",
      "content": "<p>Although my model didn't generalise i'd still like to summarise my approach, this being my first kaggle competition.</p>\n<h1>Data preprocessing steps</h1>\n<ul>\n<li>Load T2w Images</li>\n<li>Removed Blank slices </li>\n<li>Crop slices </li>\n<li>SIZ [1] depth 64</li>\n<li>Normalise</li>\n<li>Zero centre</li>\n</ul>\n<h1>Training</h1>\n<ul>\n<li>Trained a Resnet50 3D with a LR scheduler (factor 10, tolerance 5) Initial LR=1e-2, batch size = 16</li>\n<li>Single fold cross validation<br>\nGenerally the results appeared like:<br>\n<a href=\"https://postimg.cc/QHzXL2c8\" target=\"_blank\">Screenshot-from-2021-10-20-12-20-00.png</a><br>\nChose best 5 models from each training cycle and created a final ensemble for submission</li>\n</ul>\n<p>[1] <a href=\"https://link.springer.com/chapter/10.1007%2F978-3-030-59354-4_15\" target=\"_blank\">https://link.springer.com/chapter/10.1007%2F978-3-030-59354-4_15</a></p>",
      "rawMarkdown": "Although my model didn't generalise i'd still like to summarise my approach, this being my first kaggle competition.\n\n# Data preprocessing steps\n- Load T2w Images\n- Removed Blank slices \n- Crop slices \n- SIZ [1] depth 64\n- Normalise\n- Zero centre\n\n# Training\n- Trained a Resnet50 3D with a LR scheduler (factor 10, tolerance 5) Initial LR=1e-2, batch size = 16\n- Single fold cross validation\nGenerally the results appeared like:\n[Screenshot-from-2021-10-20-12-20-00.png](https://postimg.cc/QHzXL2c8)\nChose best 5 models from each training cycle and created a final ensemble for submission\n\n\n[1] https://link.springer.com/chapter/10.1007%2F978-3-030-59354-4_15",
      "votes": null
    },
    {
      "id": "1551074",
      "postDate": "10/20/2021 09:09:47",
      "content": "<p>Best of luck with your future competitions, </p>",
      "rawMarkdown": "Best of luck with your future competitions,",
      "votes": null
    },
    {
      "id": "1551172",
      "postDate": "10/20/2021 11:32:04",
      "content": "<p>Cheers mate, you too</p>",
      "rawMarkdown": "Cheers mate, you too",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1551074,
      "author_name": "mrinath",
      "author_url": "",
      "post_date": "10/20/2021 09:09:47",
      "content": "<p>Best of luck with your future competitions, </p>",
      "votes": null,
      "replies": [
        {
          "id": 1551172,
          "author_name": "aryamansharma47",
          "author_url": "",
          "post_date": "10/20/2021 11:32:04",
          "content": "<p>Cheers mate, you too</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1550767": "Although my model didn't generalise i'd still like to summarise my approach, this being my first kaggle competition.\n\n# Data preprocessing steps\n- Load T2w Images\n- Removed Blank slices \n- Crop slices \n- SIZ [1] depth 64\n- Normalise\n- Zero centre\n\n# Training\n- Trained a Resnet50 3D with a LR scheduler (factor 10, tolerance 5) Initial LR=1e-2, batch size = 16\n- Single fold cross validation\nGenerally the results appeared like:\n[Screenshot-from-2021-10-20-12-20-00.png](https://postimg.cc/QHzXL2c8)\nChose best 5 models from each training cycle and created a final ensemble for submission\n\n\n[1] https://link.springer.com/chapter/10.1007%2F978-3-030-59354-4_15",
    "1551074": "Best of luck with your future competitions,",
    "1551172": "Cheers mate, you too"
  },
  "source": "meta"
}