{
  "id": 209485,
  "title": "High IoU and Dice Loss",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/209485",
  "author_name": "",
  "post_date": "2021-01-07T17:41:09.345600300Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi Everyone,</p>\n<p>While I was training with a new set of data and applied normalization to my PyTorch pipeline, my IoU and Dice loss reach values greater than 23.3 and -0.54 respectively.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1586379%2F619961385d525a986daabec89dcff425%2FScreenshot%202021-01-07%20at%2011.07.36%20PM.png?generation=1610041100406039&amp;alt=media\" alt=\"\"></p>\n<p>Even if I don't normalize the images I get the same problem. Also, the images look very weird after I apply these normalizations. Can anyone please help me with this?</p>\n<p>My normalization code is as follows:<br>\n<code>albu.Normalize(mean=(0.65806392, 0.4906465, 0.69688281), std=(0.15952521, 0.24545997, 0.13793028), max_pixel_value=255.0, always_apply=True)</code></p>",
  "messages": [
    {
      "id": "1142974",
      "postDate": "01/07/2021 17:41:09",
      "content": "<p>Hi Everyone,</p>\n<p>While I was training with a new set of data and applied normalization to my PyTorch pipeline, my IoU and Dice loss reach values greater than 23.3 and -0.54 respectively.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1586379%2F619961385d525a986daabec89dcff425%2FScreenshot%202021-01-07%20at%2011.07.36%20PM.png?generation=1610041100406039&amp;alt=media\" alt=\"\"></p>\n<p>Even if I don't normalize the images I get the same problem. Also, the images look very weird after I apply these normalizations. Can anyone please help me with this?</p>\n<p>My normalization code is as follows:<br>\n<code>albu.Normalize(mean=(0.65806392, 0.4906465, 0.69688281), std=(0.15952521, 0.24545997, 0.13793028), max_pixel_value=255.0, always_apply=True)</code></p>",
      "rawMarkdown": "Hi Everyone,\n\nWhile I was training with a new set of data and applied normalization to my PyTorch pipeline, my IoU and Dice loss reach values greater than 23.3 and -0.54 respectively.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1586379%2F619961385d525a986daabec89dcff425%2FScreenshot%202021-01-07%20at%2011.07.36%20PM.png?generation=1610041100406039&alt=media)\n\nEven if I don't normalize the images I get the same problem. Also, the images look very weird after I apply these normalizations. Can anyone please help me with this?\n\nMy normalization code is as follows:\n`albu.Normalize(mean=(0.65806392, 0.4906465, 0.69688281), std=(0.15952521, 0.24545997, 0.13793028), max_pixel_value=255.0, always_apply=True)`",
      "votes": null
    },
    {
      "id": "1143047",
      "postDate": "01/07/2021 18:25:20",
      "content": "<p><a href=\"https://www.kaggle.com/ckanth090\" target=\"_blank\">@ckanth090</a> <br>\nThis is not problem of normalization.<br>\nThis seems to be a problem with the output of the model.</p>\n<p>Maybe it is because you have not applied the sigmoid function to the output of the model.<br>\nPlease check these points.</p>\n<ul>\n<li>Output of your model (logits or after sigmoid value?)</li>\n<li>Argument types required by your loss function<ul>\n<li>logits? or after sigmoid value?</li></ul></li>\n<li>Argument types required by your iou_score function<ul>\n<li>logits? or binary(0 or 1)? or float(0 ~ 1)?</li></ul></li>\n</ul>",
      "rawMarkdown": "ckanth090 \nThis is not problem of normalization.\nThis seems to be a problem with the output of the model.\n\nMaybe it is because you have not applied the sigmoid function to the output of the model.\nPlease check these points.\n\n- Output of your model (logits or after sigmoid value?)\n- Argument types required by your loss function\n    - logits? or after sigmoid value?\n- Argument types required by your iou_score function\n    - logits? or binary(0 or 1)? or float(0 ~ 1)?",
      "votes": null
    },
    {
      "id": "1145610",
      "postDate": "01/09/2021 08:55:12",
      "content": "<p>hey <a href=\"https://www.kaggle.com/yukkyo\" target=\"_blank\">@yukkyo</a> i really want to talk to you, can i send a message on Linkedin?</p>",
      "rawMarkdown": "hey @yukkyo i really want to talk to you, can i send a message on Linkedin?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1143047,
      "author_name": "yukkyo",
      "author_url": "",
      "post_date": "01/07/2021 18:25:20",
      "content": "<p><a href=\"https://www.kaggle.com/ckanth090\" target=\"_blank\">@ckanth090</a> <br>\nThis is not problem of normalization.<br>\nThis seems to be a problem with the output of the model.</p>\n<p>Maybe it is because you have not applied the sigmoid function to the output of the model.<br>\nPlease check these points.</p>\n<ul>\n<li>Output of your model (logits or after sigmoid value?)</li>\n<li>Argument types required by your loss function<ul>\n<li>logits? or after sigmoid value?</li></ul></li>\n<li>Argument types required by your iou_score function<ul>\n<li>logits? or binary(0 or 1)? or float(0 ~ 1)?</li></ul></li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 1145610,
          "author_name": "biswajitghosh145",
          "author_url": "",
          "post_date": "01/09/2021 08:55:12",
          "content": "<p>hey <a href=\"https://www.kaggle.com/yukkyo\" target=\"_blank\">@yukkyo</a> i really want to talk to you, can i send a message on Linkedin?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1142974": "Hi Everyone,\n\nWhile I was training with a new set of data and applied normalization to my PyTorch pipeline, my IoU and Dice loss reach values greater than 23.3 and -0.54 respectively.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1586379%2F619961385d525a986daabec89dcff425%2FScreenshot%202021-01-07%20at%2011.07.36%20PM.png?generation=1610041100406039&alt=media)\n\nEven if I don't normalize the images I get the same problem. Also, the images look very weird after I apply these normalizations. Can anyone please help me with this?\n\nMy normalization code is as follows:\n`albu.Normalize(mean=(0.65806392, 0.4906465, 0.69688281), std=(0.15952521, 0.24545997, 0.13793028), max_pixel_value=255.0, always_apply=True)`",
    "1143047": "ckanth090 \nThis is not problem of normalization.\nThis seems to be a problem with the output of the model.\n\nMaybe it is because you have not applied the sigmoid function to the output of the model.\nPlease check these points.\n\n- Output of your model (logits or after sigmoid value?)\n- Argument types required by your loss function\n    - logits? or after sigmoid value?\n- Argument types required by your iou_score function\n    - logits? or binary(0 or 1)? or float(0 ~ 1)?",
    "1145610": "hey @yukkyo i really want to talk to you, can i send a message on Linkedin?"
  },
  "source": "meta"
}