{
  "id": 580864,
  "title": "Better data preprocessing when using different model aside Yolo e.g. 3D U-Net model",
  "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/580864",
  "author_name": "",
  "post_date": "2025-05-26T22:45:21.791863800Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>While running my model (3D U-Net) on this task, i have been noticing this that my val_f1 score value remains stagnant ( at val_f1=0.463 ) even  my model val_loss and  val_reg_error score values decreasing gradually as expected.                                                                                                                               For example:                                                                                                         <br>\n Epoch 2: 243/243 [12:08&lt;00:00,  0.33it/s, v_num=0, train_loss=nan.0, val_loss=0.580, val_reg_error=99.50, val_f1=0.463]</p>",
  "messages": [
    {
      "id": "3210222",
      "postDate": "05/26/2025 22:45:21",
      "content": "<p>While running my model (3D U-Net) on this task, i have been noticing this that my val_f1 score value remains stagnant ( at val_f1=0.463 ) even  my model val_loss and  val_reg_error score values decreasing gradually as expected.                                                                                                                               For example:                                                                                                         <br>\n Epoch 2: 243/243 [12:08&lt;00:00,  0.33it/s, v_num=0, train_loss=nan.0, val_loss=0.580, val_reg_error=99.50, val_f1=0.463]</p>",
      "rawMarkdown": "While running my model (3D U-Net) on this task, i have been noticing this that my val_f1 score value remains stagnant ( at val_f1=0.463 ) even  my model val_loss and  val_reg_error score values decreasing gradually as expected.                                                                                                                               For example:                                                                                                         \n Epoch 2: 243/243 [12:08<00:00,  0.33it/s, v_num=0, train_loss=nan.0, val_loss=0.580, val_reg_error=99.50, val_f1=0.463]",
      "votes": null
    },
    {
      "id": "3210684",
      "postDate": "05/27/2025 15:18:14",
      "content": "<p>There are things that will lower your loss but won't improve the fbeta score. For example less detections inside a given tomogram will usually lead to a lower classification loss, but won't change the score if you still have the one detection that matters. Some parameters like ball size and loss weights will better align the two.</p>",
      "rawMarkdown": "There are things that will lower your loss but won't improve the fbeta score. For example less detections inside a given tomogram will usually lead to a lower classification loss, but won't change the score if you still have the one detection that matters. Some parameters like ball size and loss weights will better align the two.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3210684,
      "author_name": "tennogh",
      "author_url": "",
      "post_date": "05/27/2025 15:18:14",
      "content": "<p>There are things that will lower your loss but won't improve the fbeta score. For example less detections inside a given tomogram will usually lead to a lower classification loss, but won't change the score if you still have the one detection that matters. Some parameters like ball size and loss weights will better align the two.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3210222": "While running my model (3D U-Net) on this task, i have been noticing this that my val_f1 score value remains stagnant ( at val_f1=0.463 ) even  my model val_loss and  val_reg_error score values decreasing gradually as expected.                                                                                                                               For example:                                                                                                         \n Epoch 2: 243/243 [12:08<00:00,  0.33it/s, v_num=0, train_loss=nan.0, val_loss=0.580, val_reg_error=99.50, val_f1=0.463]",
    "3210684": "There are things that will lower your loss but won't improve the fbeta score. For example less detections inside a given tomogram will usually lead to a lower classification loss, but won't change the score if you still have the one detection that matters. Some parameters like ball size and loss weights will better align the two."
  },
  "source": "meta"
}