{
  "id": 419709,
  "title": "What things make your results better?",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/419709",
  "author_name": "Roberto",
  "post_date": "2023-06-27T07:58:31.969000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi,</p>\n<p>I am relatively new in this world of image segmentation. Sorry for my direct question, but I do not know what can I do more to improve my results. This is my notebook (it has a lot of parts from code of other people here): <a href=\"https://www.kaggle.com/code/robertsun2/hubmap-detectron-training\" target=\"_blank\">https://www.kaggle.com/code/robertsun2/hubmap-detectron-training</a>.</p>\n<p>I have tried the UNet arquitecture in another notebook. Here I am trying Detectron2. I trained my model for almost 8 hours. I tried to improve it by changing the parameters. I did data augmentation (from 1633 images to almost 7000) by rotating and flipping images and masks.</p>\n<p>But, the results are really bad. I have seen other people using a big pretrained model (.pth file, I do not know where they trained it, but they load it as a dataset) and they get a 0.25 score (I get a 0 score).</p>\n<p>Anyone can give me a clue? What really makes the difference between a good model and a bad model? More training time? I do not know why my model cannot learn the patters well. Please, I would appreciate so much your help!</p>\n<p>Thank you</p>",
  "messages": [
    {
      "id": 2319573,
      "postDate": "2023-06-27T07:58:31.970Z",
      "content": "<p>Hi,</p>\n<p>I am relatively new in this world of image segmentation. Sorry for my direct question, but I do not know what can I do more to improve my results. This is my notebook (it has a lot of parts from code of other people here): <a href=\"https://www.kaggle.com/code/robertsun2/hubmap-detectron-training\" target=\"_blank\">https://www.kaggle.com/code/robertsun2/hubmap-detectron-training</a>.</p>\n<p>I have tried the UNet arquitecture in another notebook. Here I am trying Detectron2. I trained my model for almost 8 hours. I tried to improve it by changing the parameters. I did data augmentation (from 1633 images to almost 7000) by rotating and flipping images and masks.</p>\n<p>But, the results are really bad. I have seen other people using a big pretrained model (.pth file, I do not know where they trained it, but they load it as a dataset) and they get a 0.25 score (I get a 0 score).</p>\n<p>Anyone can give me a clue? What really makes the difference between a good model and a bad model? More training time? I do not know why my model cannot learn the patters well. Please, I would appreciate so much your help!</p>\n<p>Thank you</p>",
      "rawMarkdown": "Hi,\n\nI am relatively new in this world of image segmentation. Sorry for my direct question, but I do not know what can I do more to improve my results. This is my notebook (it has a lot of parts from code of other people here): https://www.kaggle.com/code/robertsun2/hubmap-detectron-training.\n\nI have tried the UNet arquitecture in another notebook. Here I am trying Detectron2. I trained my model for almost 8 hours. I tried to improve it by changing the parameters. I did data augmentation (from 1633 images to almost 7000) by rotating and flipping images and masks.\n\nBut, the results are really bad. I have seen other people using a big pretrained model (.pth file, I do not know where they trained it, but they load it as a dataset) and they get a 0.25 score (I get a 0 score).\n\nAnyone can give me a clue? What really makes the difference between a good model and a bad model? More training time? I do not know why my model cannot learn the patters well. Please, I would appreciate so much your help!\n\nThank you",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2319573": "Hi,\n\nI am relatively new in this world of image segmentation. Sorry for my direct question, but I do not know what can I do more to improve my results. This is my notebook (it has a lot of parts from code of other people here): https://www.kaggle.com/code/robertsun2/hubmap-detectron-training.\n\nI have tried the UNet arquitecture in another notebook. Here I am trying Detectron2. I trained my model for almost 8 hours. I tried to improve it by changing the parameters. I did data augmentation (from 1633 images to almost 7000) by rotating and flipping images and masks.\n\nBut, the results are really bad. I have seen other people using a big pretrained model (.pth file, I do not know where they trained it, but they load it as a dataset) and they get a 0.25 score (I get a 0 score).\n\nAnyone can give me a clue? What really makes the difference between a good model and a bad model? More training time? I do not know why my model cannot learn the patters well. Please, I would appreciate so much your help!\n\nThank you"
  }
}