{
  "topic": {
    "id": 736678,
    "title": "Maybe worth checking the Efficiency LB?",
    "authorName": "k256.dev",
    "commentCount": 4,
    "votes": 13,
    "postDate": "2026-08-22T01:27:30.292000"
  },
  "comments": [
    {
      "id": 3515576,
      "authorName": "Scott Willis",
      "votes": 1,
      "postDate": "2026-08-22T02:45:51.867000",
      "content": "<p>So I'm surprisingly #1 on the efficiency LB right now.  No ensemble and I'm not too sure how much more I can squeeze out of it, but it's been really fun seeing how far I can push it without hurting efficiency.</p>"
    },
    {
      "id": 3515583,
      "authorName": "k256.dev",
      "votes": 0,
      "postDate": "2026-08-22T04:03:07.060000",
      "content": "<p>Honestly, I've come to think the Efficiency LB is the real public leaderboard. Let's both watch out for overfitting to public — good luck to us both!</p>"
    },
    {
      "id": 3515569,
      "authorName": "Chikuwabu",
      "votes": 1,
      "postDate": "2026-08-22T02:16:44.817000",
      "content": "<p>I agree. </p>\n<p>One thing to keep in mind is that the top two Public LB submissions seem to be used for the Efficiency LB as well, so we should be careful not to accidentally get a high score through ensembling.</p>"
    },
    {
      "id": 3516470,
      "authorName": "Berat Kirbiyik",
      "votes": 0,
      "postDate": "2026-08-24T13:24:15.703000",
      "content": "<p>Thanks for opening this up — the Efficiency LB really is the more informative board.</p>\n<p>I profiled our inference pipeline and hit a wall I can't get past. Breakdown for ~1300 studies:</p>\n<p>header pass          0.5 min\nslice ordering      12.0 min   &lt;- the wall\ndecode + resize      4.1 min\ncontainer + import   5.0 min\nGPU (5-fold)        10.0 min</p>\n<p>The ordering pass reads the geometry tags of every slice in every selected series (~186 files per study) to sort by the projection onto the slice normal, then keeps only 9. Things I tried that didn't work:</p>\n<p>Threads: saturates at 64. Going to 128/256/512 changes nothing, so it's not latency-bound.\nReading only every Nth slice and interpolating: slice agreement with the full ordering drops to ~39%, no better than chance for the ones not sampled.\nInstanceNumber instead of the projection: 94.5% agreement but zero speedup — the cost is opening the file, not parsing the tag.</p>\n<p>You mentioned ~20 min total with preprocessing plus a single 5-fold model. Are you doing something fundamentally different at the ordering step, or skipping it? Genuinely curious whether the geometric sort is worth its cost at all</p>"
    }
  ],
  "index": {
    "id": "736678",
    "title": "Maybe worth checking the Efficiency LB?",
    "authorName": "",
    "commentCount": "4",
    "votes": "13",
    "postDate": "2026-08-22 01:27:30.292000"
  },
  "competition": "rsna-knee-abnormality-detection"
}