{
  "id": 669686,
  "title": "Sharing Training Methods",
  "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/669686",
  "author_name": "Nick Pellegrin",
  "post_date": "2026-01-23T18:01:14.584000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Now that the competition is over for submissions, does anyone care to share how they trained their models and what score you achieved with it?</p>\n<p>I was able to get a somewhat decent score (0.325) using dinov2 + small segmentation head, and here's how I trained it:</p>\n<p>A two-stage training where both stages used a CosineAnnealingLR:</p>\n<ol>\n<li>Keep dinov2 backbone frozen and train only the segmentation head using Adamw and vanilla BCEWithLogitsLoss.</li>\n<li>Unfreeze the backbone, and again train with Adamw and vanilla BCEWithLogitsLoss while keeping the learning rate for the backbone very small (5e-7)</li>\n</ol>\n<p>As far as post-processing, I employed many of the common approaches you can find in almost any public notebook, but I'm curious if anyone found a large jump in LB from different post-processing techniques.</p>\n<p>Did anyone get good results using different loss functions (BCE + DICE)?</p>",
  "messages": [
    {
      "id": 3395823,
      "postDate": "2026-01-23T18:01:14.583Z",
      "content": "<p>Now that the competition is over for submissions, does anyone care to share how they trained their models and what score you achieved with it?</p>\n<p>I was able to get a somewhat decent score (0.325) using dinov2 + small segmentation head, and here's how I trained it:</p>\n<p>A two-stage training where both stages used a CosineAnnealingLR:</p>\n<ol>\n<li>Keep dinov2 backbone frozen and train only the segmentation head using Adamw and vanilla BCEWithLogitsLoss.</li>\n<li>Unfreeze the backbone, and again train with Adamw and vanilla BCEWithLogitsLoss while keeping the learning rate for the backbone very small (5e-7)</li>\n</ol>\n<p>As far as post-processing, I employed many of the common approaches you can find in almost any public notebook, but I'm curious if anyone found a large jump in LB from different post-processing techniques.</p>\n<p>Did anyone get good results using different loss functions (BCE + DICE)?</p>",
      "rawMarkdown": "Now that the competition is over for submissions, does anyone care to share how they trained their models and what score you achieved with it?\n\nI was able to get a somewhat decent score (0.325) using dinov2 + small segmentation head, and here's how I trained it:\n\nA two-stage training where both stages used a CosineAnnealingLR:\n1. Keep dinov2 backbone frozen and train only the segmentation head using Adamw and vanilla BCEWithLogitsLoss.\n2. Unfreeze the backbone, and again train with Adamw and vanilla BCEWithLogitsLoss while keeping the learning rate for the backbone very small (5e-7)\n\nAs far as post-processing, I employed many of the common approaches you can find in almost any public notebook, but I'm curious if anyone found a large jump in LB from different post-processing techniques.\n\nDid anyone get good results using different loss functions (BCE + DICE)?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3395823": "Now that the competition is over for submissions, does anyone care to share how they trained their models and what score you achieved with it?\n\nI was able to get a somewhat decent score (0.325) using dinov2 + small segmentation head, and here's how I trained it:\n\nA two-stage training where both stages used a CosineAnnealingLR:\n1. Keep dinov2 backbone frozen and train only the segmentation head using Adamw and vanilla BCEWithLogitsLoss.\n2. Unfreeze the backbone, and again train with Adamw and vanilla BCEWithLogitsLoss while keeping the learning rate for the backbone very small (5e-7)\n\nAs far as post-processing, I employed many of the common approaches you can find in almost any public notebook, but I'm curious if anyone found a large jump in LB from different post-processing techniques.\n\nDid anyone get good results using different loss functions (BCE + DICE)?"
  }
}