{
  "id": 534695,
  "title": "[Discussion] How to train/pretrain YOLO? ",
  "url": "/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/534695",
  "author_name": "Arindam Roy",
  "post_date": "2024-09-17T20:53:45.848000",
  "votes": 7,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hey! </p>\n<p>I have never worked with YOLO before, so looking for some clarity here. I wanted to share my approaches to pretraining and training;  and wanted to know this communities thoughts as well: </p>\n<ol>\n<li><p>Pretrain on the <a href=\"https://www.kaggle.com/datasets/brendanartley/lumbar-coordinate-pretraining-dataset\" target=\"_blank\">Lumbar Coordinate Dataset</a> and use these weights to train on available train data from comp. This performs a little worse than raw training for me </p></li>\n<li><p>Again, train YOLO on the above mentioned data and freeze its layers apart from the last (classification head; model22.*). Reason being, we train a model to first detect all levels of the spine and then trained it to classify the conditions from our dataset. Surprisingly, this is very close to the best performing model for me. </p></li>\n</ol>\n<p>Would love to hear your inputs! </p>",
  "messages": [
    {
      "id": 2991886,
      "postDate": "2024-09-17T20:53:45.847Z",
      "content": "<p>Hey! </p>\n<p>I have never worked with YOLO before, so looking for some clarity here. I wanted to share my approaches to pretraining and training;  and wanted to know this communities thoughts as well: </p>\n<ol>\n<li><p>Pretrain on the <a href=\"https://www.kaggle.com/datasets/brendanartley/lumbar-coordinate-pretraining-dataset\" target=\"_blank\">Lumbar Coordinate Dataset</a> and use these weights to train on available train data from comp. This performs a little worse than raw training for me </p></li>\n<li><p>Again, train YOLO on the above mentioned data and freeze its layers apart from the last (classification head; model22.*). Reason being, we train a model to first detect all levels of the spine and then trained it to classify the conditions from our dataset. Surprisingly, this is very close to the best performing model for me. </p></li>\n</ol>\n<p>Would love to hear your inputs! </p>",
      "rawMarkdown": "Hey! \n\nI have never worked with YOLO before, so looking for some clarity here. I wanted to share my approaches to pretraining and training;  and wanted to know this communities thoughts as well: \n\n1. Pretrain on the [Lumbar Coordinate Dataset](https://www.kaggle.com/datasets/brendanartley/lumbar-coordinate-pretraining-dataset) and use these weights to train on available train data from comp. This performs a little worse than raw training for me \n\n2. Again, train YOLO on the above mentioned data and freeze its layers apart from the last (classification head; model22.*). Reason being, we train a model to first detect all levels of the spine and then trained it to classify the conditions from our dataset. Surprisingly, this is very close to the best performing model for me. \n\nWould love to hear your inputs! ",
      "votes": 7
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2991886": "Hey! \n\nI have never worked with YOLO before, so looking for some clarity here. I wanted to share my approaches to pretraining and training;  and wanted to know this communities thoughts as well: \n\n1. Pretrain on the [Lumbar Coordinate Dataset](https://www.kaggle.com/datasets/brendanartley/lumbar-coordinate-pretraining-dataset) and use these weights to train on available train data from comp. This performs a little worse than raw training for me \n\n2. Again, train YOLO on the above mentioned data and freeze its layers apart from the last (classification head; model22.*). Reason being, we train a model to first detect all levels of the spine and then trained it to classify the conditions from our dataset. Surprisingly, this is very close to the best performing model for me. \n\nWould love to hear your inputs! "
  }
}