{
  "id": 428977,
  "title": "20th Place Solution: yolov8m single model",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/428977",
  "author_name": "tsobolev",
  "post_date": "2023-08-03T15:07:07.980000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Thanks to HuBMAP for hosting the exciting competition, and congrats to all winners!</p>\n<ul>\n<li><p>Summary </p>\n<p>Yolov8-m 1024x1024 trained on ds1 and ds2 with dilation iterations = 2 for ds2 and then finetuned only on ds1</p></li>\n<li><p>Training pipeline</p>\n<p>Dataset extended with different ElasticTransforms(torchvision) which is to slow to do realtime augmentation <br>\nTrain model with hsv,rotation,flip,mosaic,scale augmentation on ds1 and ds2<br>\nFine-tune model with less augumentation on dataset1 only<br>\nFine-tune model using the best parameters on full dataset1 with no val split (+0.01 on private and public lb)</p></li>\n<li><p>Post-precessing</p>\n<p>Remove small masks<br>\n Remove masks that highly overlaps glomerulus regions<br>\n Subtract glomerulus regions from predicted mask</p></li>\n<li><p>Notes</p>\n<p>Fine-tuning with relabeled dataset2 does not work for me (even with a small part)<br>\n Fine-tuning with dataset2 does not work for me (even with a small part)<br>\n TTA hvflip (picking a mask with the highest score from the intersected ones) lowering score just a bit - not used<br>\n Fine-tuning on ds1 increases all scores: val,public,private<br>\n Perhaps the same result or even better can be achieved with the same model only on ds1 with better parameters tuning ;)</p></li>\n</ul>",
  "messages": [
    {
      "id": 2372234,
      "postDate": "2023-08-03T15:07:07.980Z",
      "content": "<p>Thanks to HuBMAP for hosting the exciting competition, and congrats to all winners!</p>\n<ul>\n<li><p>Summary </p>\n<p>Yolov8-m 1024x1024 trained on ds1 and ds2 with dilation iterations = 2 for ds2 and then finetuned only on ds1</p></li>\n<li><p>Training pipeline</p>\n<p>Dataset extended with different ElasticTransforms(torchvision) which is to slow to do realtime augmentation <br>\nTrain model with hsv,rotation,flip,mosaic,scale augmentation on ds1 and ds2<br>\nFine-tune model with less augumentation on dataset1 only<br>\nFine-tune model using the best parameters on full dataset1 with no val split (+0.01 on private and public lb)</p></li>\n<li><p>Post-precessing</p>\n<p>Remove small masks<br>\n Remove masks that highly overlaps glomerulus regions<br>\n Subtract glomerulus regions from predicted mask</p></li>\n<li><p>Notes</p>\n<p>Fine-tuning with relabeled dataset2 does not work for me (even with a small part)<br>\n Fine-tuning with dataset2 does not work for me (even with a small part)<br>\n TTA hvflip (picking a mask with the highest score from the intersected ones) lowering score just a bit - not used<br>\n Fine-tuning on ds1 increases all scores: val,public,private<br>\n Perhaps the same result or even better can be achieved with the same model only on ds1 with better parameters tuning ;)</p></li>\n</ul>",
      "rawMarkdown": "Thanks to HuBMAP for hosting the exciting competition, and congrats to all winners!\n\n* Summary \n\n\tYolov8-m 1024x1024 trained on ds1 and ds2 with dilation iterations = 2 for ds2 and then finetuned only on ds1\n\n* Training pipeline\n\n\tDataset extended with different ElasticTransforms(torchvision) which is to slow to do realtime augmentation \n\tTrain model with hsv,rotation,flip,mosaic,scale augmentation on ds1 and ds2\n\tFine-tune model with less augumentation on dataset1 only\n\tFine-tune model using the best parameters on full dataset1 with no val split (+0.01 on private and public lb)\n\n* Post-precessing\n\n \tRemove small masks\n \tRemove masks that highly overlaps glomerulus regions\n \tSubtract glomerulus regions from predicted mask\n\n* Notes\n\n \tFine-tuning with relabeled dataset2 does not work for me (even with a small part)\n \tFine-tuning with dataset2 does not work for me (even with a small part)\n \tTTA hvflip (picking a mask with the highest score from the intersected ones) lowering score just a bit - not used\n \tFine-tuning on ds1 increases all scores: val,public,private\n \tPerhaps the same result or even better can be achieved with the same model only on ds1 with better parameters tuning ;)",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2372234": "Thanks to HuBMAP for hosting the exciting competition, and congrats to all winners!\n\n* Summary \n\n\tYolov8-m 1024x1024 trained on ds1 and ds2 with dilation iterations = 2 for ds2 and then finetuned only on ds1\n\n* Training pipeline\n\n\tDataset extended with different ElasticTransforms(torchvision) which is to slow to do realtime augmentation \n\tTrain model with hsv,rotation,flip,mosaic,scale augmentation on ds1 and ds2\n\tFine-tune model with less augumentation on dataset1 only\n\tFine-tune model using the best parameters on full dataset1 with no val split (+0.01 on private and public lb)\n\n* Post-precessing\n\n \tRemove small masks\n \tRemove masks that highly overlaps glomerulus regions\n \tSubtract glomerulus regions from predicted mask\n\n* Notes\n\n \tFine-tuning with relabeled dataset2 does not work for me (even with a small part)\n \tFine-tuning with dataset2 does not work for me (even with a small part)\n \tTTA hvflip (picking a mask with the highest score from the intersected ones) lowering score just a bit - not used\n \tFine-tuning on ds1 increases all scores: val,public,private\n \tPerhaps the same result or even better can be achieved with the same model only on ds1 with better parameters tuning ;)"
  }
}