{
  "id": 569828,
  "title": "Question about Competition",
  "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/569828",
  "author_name": "",
  "post_date": "2025-03-24T11:05:45.561946500Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello to everyone.<br>\nI am a newcomer to the ML field. In my opinion, everyone uses YOLO to predict bacterial flagellar motor activity. However, </p>\n<ul>\n<li>what techniques can lead to achieving the best LB score?</li>\n<li>how can I build the best loss function?</li>\n<li>What about training not only the z plane, but also x, and y?</li>\n<li>In my training, if I use YOLO is not enough memory for saving images. How can I resolve this issue?</li>\n</ul>",
  "messages": [
    {
      "id": "3158243",
      "postDate": "03/24/2025 11:05:45",
      "content": "<p>Hello to everyone.<br>\nI am a newcomer to the ML field. In my opinion, everyone uses YOLO to predict bacterial flagellar motor activity. However, </p>\n<ul>\n<li>what techniques can lead to achieving the best LB score?</li>\n<li>how can I build the best loss function?</li>\n<li>What about training not only the z plane, but also x, and y?</li>\n<li>In my training, if I use YOLO is not enough memory for saving images. How can I resolve this issue?</li>\n</ul>",
      "rawMarkdown": "Hello to everyone.\nI am a newcomer to the ML field. In my opinion, everyone uses YOLO to predict bacterial flagellar motor activity. However, \n- what techniques can lead to achieving the best LB score?\n- how can I build the best loss function?\n- What about training not only the z plane, but also x, and y?\n- In my training, if I use YOLO is not enough memory for saving images. How can I resolve this issue?",
      "votes": null
    },
    {
      "id": "3161927",
      "postDate": "03/28/2025 15:41:15",
      "content": "<p>The best thing to do is experiment.  The science to it is trying different things and seeing what works and what doesn't work.  Start with a simple solution that loads the data, trains the model, evaluations it, and builds some metrics and visualizations on the results.  From there, slowly improve.  </p>\n<ol>\n<li>Techniques vary and you should try everything and see what works, as well as read current literature on state of the art solutions</li>\n<li>You don't need to build a loss function at first, look at popular ones such as Dice Loss for segmentation or build in loss functions in YOLO.</li>\n<li>Try it, see what works</li>\n<li>Don't save the images.  Keep everything in memory instead of storage and only load the data in batches and free up the space once you are complete.</li>\n</ol>\n<p>Overall, start looking at the existing code notebooks for this competition and find a starting point.  If you are new this is the best way to learn :)</p>",
      "rawMarkdown": "The best thing to do is experiment.  The science to it is trying different things and seeing what works and what doesn't work.  Start with a simple solution that loads the data, trains the model, evaluations it, and builds some metrics and visualizations on the results.  From there, slowly improve.  \n\n1. Techniques vary and you should try everything and see what works, as well as read current literature on state of the art solutions\n2. You don't need to build a loss function at first, look at popular ones such as Dice Loss for segmentation or build in loss functions in YOLO.\n3. Try it, see what works\n4. Don't save the images.  Keep everything in memory instead of storage and only load the data in batches and free up the space once you are complete.\n\nOverall, start looking at the existing code notebooks for this competition and find a starting point.  If you are new this is the best way to learn :)",
      "votes": null
    },
    {
      "id": "3161979",
      "postDate": "03/28/2025 17:06:40",
      "content": "<p>I tried to train model in the X and Y planes but it doesn't work.</p>",
      "rawMarkdown": "I tried to train model in the X and Y planes but it doesn't work.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3161927,
      "author_name": "connorjd",
      "author_url": "",
      "post_date": "03/28/2025 15:41:15",
      "content": "<p>The best thing to do is experiment.  The science to it is trying different things and seeing what works and what doesn't work.  Start with a simple solution that loads the data, trains the model, evaluations it, and builds some metrics and visualizations on the results.  From there, slowly improve.  </p>\n<ol>\n<li>Techniques vary and you should try everything and see what works, as well as read current literature on state of the art solutions</li>\n<li>You don't need to build a loss function at first, look at popular ones such as Dice Loss for segmentation or build in loss functions in YOLO.</li>\n<li>Try it, see what works</li>\n<li>Don't save the images.  Keep everything in memory instead of storage and only load the data in batches and free up the space once you are complete.</li>\n</ol>\n<p>Overall, start looking at the existing code notebooks for this competition and find a starting point.  If you are new this is the best way to learn :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3161979,
      "author_name": "fautei",
      "author_url": "",
      "post_date": "03/28/2025 17:06:40",
      "content": "<p>I tried to train model in the X and Y planes but it doesn't work.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3158243": "Hello to everyone.\nI am a newcomer to the ML field. In my opinion, everyone uses YOLO to predict bacterial flagellar motor activity. However, \n- what techniques can lead to achieving the best LB score?\n- how can I build the best loss function?\n- What about training not only the z plane, but also x, and y?\n- In my training, if I use YOLO is not enough memory for saving images. How can I resolve this issue?",
    "3161927": "The best thing to do is experiment.  The science to it is trying different things and seeing what works and what doesn't work.  Start with a simple solution that loads the data, trains the model, evaluations it, and builds some metrics and visualizations on the results.  From there, slowly improve.  \n\n1. Techniques vary and you should try everything and see what works, as well as read current literature on state of the art solutions\n2. You don't need to build a loss function at first, look at popular ones such as Dice Loss for segmentation or build in loss functions in YOLO.\n3. Try it, see what works\n4. Don't save the images.  Keep everything in memory instead of storage and only load the data in batches and free up the space once you are complete.\n\nOverall, start looking at the existing code notebooks for this competition and find a starting point.  If you are new this is the best way to learn :)",
    "3161979": "I tried to train model in the X and Y planes but it doesn't work."
  },
  "source": "meta"
}