{
  "id": 466403,
  "title": "3D UNETR score is strangely low! Any help with debugging?",
  "url": "/competitions/blood-vessel-segmentation/discussion/466403",
  "author_name": "",
  "post_date": "2024-01-08T13:15:17.749161500Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I've trained a UNETR model on 3D subvolumes of size 64x256x256 using kidney_1_dense images. <a href=\"https://www.kaggle.com/code/tahseenislamsajon/sennet-hoa-3d-training-monai-pytorch\" target=\"_blank\">Here</a> is the public demo. After training for 40 epochs, and using kidney_3_dense as the validation data, here are the scores on the validation set: </p>\n<pre><code>Valid Dice:  | Valid Jaccard: \nValid Loss: \n</code></pre>\n<p>However, for inference, the score was as low as 0.05. <a href=\"https://www.kaggle.com/code/tahseenislamsajon/submission-example-for-2d-to-3d-unet\" target=\"_blank\">Here</a> is my inference code, based on <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a>'s example. I've tried adjusting the threshold and cc3d values, but to no avail. Shouldn't the score be at least somewhat close to the validation results? Not sure what i'm doing wrong. I would greatly appreciate any insights or recommendations on this issue. Thanks! </p>",
  "messages": [
    {
      "id": "2592258",
      "postDate": "01/08/2024 13:15:17",
      "content": "<p>I've trained a UNETR model on 3D subvolumes of size 64x256x256 using kidney_1_dense images. <a href=\"https://www.kaggle.com/code/tahseenislamsajon/sennet-hoa-3d-training-monai-pytorch\" target=\"_blank\">Here</a> is the public demo. After training for 40 epochs, and using kidney_3_dense as the validation data, here are the scores on the validation set: </p>\n<pre><code>Valid Dice:  | Valid Jaccard: \nValid Loss: \n</code></pre>\n<p>However, for inference, the score was as low as 0.05. <a href=\"https://www.kaggle.com/code/tahseenislamsajon/submission-example-for-2d-to-3d-unet\" target=\"_blank\">Here</a> is my inference code, based on <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a>'s example. I've tried adjusting the threshold and cc3d values, but to no avail. Shouldn't the score be at least somewhat close to the validation results? Not sure what i'm doing wrong. I would greatly appreciate any insights or recommendations on this issue. Thanks! </p>",
      "rawMarkdown": "I've trained a UNETR model on 3D subvolumes of size 64x256x256 using kidney_1_dense images. [Here](https://www.kaggle.com/code/tahseenislamsajon/sennet-hoa-3d-training-monai-pytorch) is the public demo. After training for 40 epochs, and using kidney_3_dense as the validation data, here are the scores on the validation set: \n```python\nValid Dice: 0.6275 | Valid Jaccard: 0.6233\nValid Loss: 0.37287195917073784\n```\n\nHowever, for inference, the score was as low as 0.05. [Here](https://www.kaggle.com/code/tahseenislamsajon/submission-example-for-2d-to-3d-unet) is my inference code, based on @hengck23's example. I've tried adjusting the threshold and cc3d values, but to no avail. Shouldn't the score be at least somewhat close to the validation results? Not sure what i'm doing wrong. I would greatly appreciate any insights or recommendations on this issue. Thanks!",
      "votes": null
    },
    {
      "id": "2621968",
      "postDate": "01/27/2024 07:00:37",
      "content": "<p>3D UNETR is not on this challenge?</p>",
      "rawMarkdown": "3D UNETR is not on this challenge?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2621968,
      "author_name": "junhanzangai",
      "author_url": "",
      "post_date": "01/27/2024 07:00:37",
      "content": "<p>3D UNETR is not on this challenge?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2592258": "I've trained a UNETR model on 3D subvolumes of size 64x256x256 using kidney_1_dense images. [Here](https://www.kaggle.com/code/tahseenislamsajon/sennet-hoa-3d-training-monai-pytorch) is the public demo. After training for 40 epochs, and using kidney_3_dense as the validation data, here are the scores on the validation set: \n```python\nValid Dice: 0.6275 | Valid Jaccard: 0.6233\nValid Loss: 0.37287195917073784\n```\n\nHowever, for inference, the score was as low as 0.05. [Here](https://www.kaggle.com/code/tahseenislamsajon/submission-example-for-2d-to-3d-unet) is my inference code, based on @hengck23's example. I've tried adjusting the threshold and cc3d values, but to no avail. Shouldn't the score be at least somewhat close to the validation results? Not sure what i'm doing wrong. I would greatly appreciate any insights or recommendations on this issue. Thanks!",
    "2621968": "3D UNETR is not on this challenge?"
  },
  "source": "meta"
}