{
  "id": 211055,
  "title": "Training went well when excluding non glom tiles. Any help?",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/211055",
  "author_name": "",
  "post_date": "2021-01-13T12:42:59.988802300Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<ul>\n<li>Image size : 256 x 256</li>\n<li>model : efficient unet</li>\n<li>loss : focal loss</li>\n<li>data preprocess: only min max scale</li>\n</ul>\n<p>When using only glom tiles, it seemed that training the model went well.<br>\nBut when i used all tiles, the loss was very low (= 0.99) and dice coefficient was very low(&lt; 0.1).</p>\n<p>What did i miss? Could you give me any suggestions?</p>\n<p>Any help will be appreciated.</p>",
  "messages": [
    {
      "id": "1151595",
      "postDate": "01/13/2021 12:42:59",
      "content": "<ul>\n<li>Image size : 256 x 256</li>\n<li>model : efficient unet</li>\n<li>loss : focal loss</li>\n<li>data preprocess: only min max scale</li>\n</ul>\n<p>When using only glom tiles, it seemed that training the model went well.<br>\nBut when i used all tiles, the loss was very low (= 0.99) and dice coefficient was very low(&lt; 0.1).</p>\n<p>What did i miss? Could you give me any suggestions?</p>\n<p>Any help will be appreciated.</p>",
      "rawMarkdown": "Image size : 256 x 256\n- model : efficient unet\n- loss : focal loss\n- data preprocess: only min max scale\n\nWhen using only glom tiles, it seemed that training the model went well.\nBut when i used all tiles, the loss was very low (<= 0.050) and accuracy was very high(>= 0.99) and dice coefficient was very low(< 0.1).\n\nWhat did i miss? Could you give me any suggestions?\n\nAny help will be appreciated.",
      "votes": null
    },
    {
      "id": "1151598",
      "postDate": "01/13/2021 12:44:05",
      "content": "<p>The loss was very low even at start of the training.</p>",
      "rawMarkdown": "The loss was very low even at start of the training.",
      "votes": null
    },
    {
      "id": "1157456",
      "postDate": "01/17/2021 23:09:17",
      "content": "<p>doesn't look right, for sure. Cannot comment due to lack of info, except 256x256 is too low resolution, in my opinion, to achieve competitive results. Also, make sure training and test data distribution should be the same</p>",
      "rawMarkdown": "doesn't look right, for sure. Cannot comment due to lack of info, except 256x256 is too low resolution, in my opinion, to achieve competitive results. Also, make sure training and test data distribution should be the same",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1151598,
      "author_name": "daviek",
      "author_url": "",
      "post_date": "01/13/2021 12:44:05",
      "content": "<p>The loss was very low even at start of the training.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1157456,
      "author_name": "andrasferenczi",
      "author_url": "",
      "post_date": "01/17/2021 23:09:17",
      "content": "<p>doesn't look right, for sure. Cannot comment due to lack of info, except 256x256 is too low resolution, in my opinion, to achieve competitive results. Also, make sure training and test data distribution should be the same</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1151595": "Image size : 256 x 256\n- model : efficient unet\n- loss : focal loss\n- data preprocess: only min max scale\n\nWhen using only glom tiles, it seemed that training the model went well.\nBut when i used all tiles, the loss was very low (<= 0.050) and accuracy was very high(>= 0.99) and dice coefficient was very low(< 0.1).\n\nWhat did i miss? Could you give me any suggestions?\n\nAny help will be appreciated.",
    "1151598": "The loss was very low even at start of the training.",
    "1157456": "doesn't look right, for sure. Cannot comment due to lack of info, except 256x256 is too low resolution, in my opinion, to achieve competitive results. Also, make sure training and test data distribution should be the same"
  },
  "source": "meta"
}