{
  "id": 577914,
  "title": "YOLO training: fine-tuning all weights or freezing the last ones?",
  "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/577914",
  "author_name": "",
  "post_date": "2025-05-07T20:34:23.146255900Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hello Kaggle community,</p>\n<p>I'm new here and learning YOLO. For transfer learning with a pre-trained YOLO model on a custom dataset, what's the general approach for this competition:</p>\n<p>Fine-tuning all weights?<br>\nOr freezing some layers and training only the last ones?<br>\nJust looking for general practices or tips. Thanks!</p>",
  "messages": [
    {
      "id": "3197124",
      "postDate": "05/07/2025 20:34:23",
      "content": "<p>Hello Kaggle community,</p>\n<p>I'm new here and learning YOLO. For transfer learning with a pre-trained YOLO model on a custom dataset, what's the general approach for this competition:</p>\n<p>Fine-tuning all weights?<br>\nOr freezing some layers and training only the last ones?<br>\nJust looking for general practices or tips. Thanks!</p>",
      "rawMarkdown": "Hello Kaggle community,\n\nI'm new here and learning YOLO. For transfer learning with a pre-trained YOLO model on a custom dataset, what's the general approach for this competition:\n\nFine-tuning all weights?\nOr freezing some layers and training only the last ones?\nJust looking for general practices or tips. Thanks!",
      "votes": null
    },
    {
      "id": "3197189",
      "postDate": "05/07/2025 21:34:46",
      "content": "<p>Haven't tested this extensively, so take what I am about to say with a grain of salt.</p>\n<p>Not freezing any layers generally worked better in my hands, regardless of the order of operations. When I froze the first 8-10 layers (backbone sizes are different), the model was converging very slowly and generally didn't come close to mAP values of the model without any freezing. When I trained without freezing until early stopping, followed by freezing the backbone and fine-tuning with very low learning rates, the models still had slightly lower mAP values.</p>",
      "rawMarkdown": "Haven't tested this extensively, so take what I am about to say with a grain of salt.\n\nNot freezing any layers generally worked better in my hands, regardless of the order of operations. When I froze the first 8-10 layers (backbone sizes are different), the model was converging very slowly and generally didn't come close to mAP values of the model without any freezing. When I trained without freezing until early stopping, followed by freezing the backbone and fine-tuning with very low learning rates, the models still had slightly lower mAP values.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3197189,
      "author_name": "tilii7",
      "author_url": "",
      "post_date": "05/07/2025 21:34:46",
      "content": "<p>Haven't tested this extensively, so take what I am about to say with a grain of salt.</p>\n<p>Not freezing any layers generally worked better in my hands, regardless of the order of operations. When I froze the first 8-10 layers (backbone sizes are different), the model was converging very slowly and generally didn't come close to mAP values of the model without any freezing. When I trained without freezing until early stopping, followed by freezing the backbone and fine-tuning with very low learning rates, the models still had slightly lower mAP values.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3197124": "Hello Kaggle community,\n\nI'm new here and learning YOLO. For transfer learning with a pre-trained YOLO model on a custom dataset, what's the general approach for this competition:\n\nFine-tuning all weights?\nOr freezing some layers and training only the last ones?\nJust looking for general practices or tips. Thanks!",
    "3197189": "Haven't tested this extensively, so take what I am about to say with a grain of salt.\n\nNot freezing any layers generally worked better in my hands, regardless of the order of operations. When I froze the first 8-10 layers (backbone sizes are different), the model was converging very slowly and generally didn't come close to mAP values of the model without any freezing. When I trained without freezing until early stopping, followed by freezing the backbone and fine-tuning with very low learning rates, the models still had slightly lower mAP values."
  },
  "source": "meta"
}