{
  "id": 664970,
  "title": "about training strategy",
  "url": "/competitions/csiro-biomass/discussion/664970",
  "author_name": "",
  "post_date": "2025-12-29T09:47:26.971449800Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>did u guys used all data in training, first find the best cv strategy and then use all data to train, has anyone tried it?</p>",
  "messages": [
    {
      "id": "3383062",
      "postDate": "12/29/2025 09:47:26",
      "content": "<p>did u guys used all data in training, first find the best cv strategy and then use all data to train, has anyone tried it?</p>",
      "rawMarkdown": "did u guys used all data in training, first find the best cv strategy and then use all data to train, has anyone tried it?",
      "votes": null
    },
    {
      "id": "3383243",
      "postDate": "12/29/2025 17:16:07",
      "content": "<p>Have tried all versions of that :)</p>\n<p>With 357 images only - the real fun of this competition is figuring out a whole lot of things that I never would have needed to think about if 5000 images had been supplied.</p>",
      "rawMarkdown": "Have tried all versions of that :)\n\nWith 357 images only - the real fun of this competition is figuring out a whole lot of things that I never would have needed to think about if 5000 images had been supplied.",
      "votes": null
    },
    {
      "id": "3383272",
      "postDate": "12/29/2025 18:18:56",
      "content": "<p>\"Great point! I think most of us are currently 'sacrificing' that 20% of data just to keep our models honest. Once the CV is stable, the jump to all data training is a leap of faith. I've tried it once, and while the score improved slightly, the fear of private LB shake up is real!</p>\n<p>Curious to hear if others are finding that the extra data significantly changes  final biomass regression.\"</p>",
      "rawMarkdown": "\"Great point! I think most of us are currently 'sacrificing' that 20% of data just to keep our models honest. Once the CV is stable, the jump to all data training is a leap of faith. I've tried it once, and while the score improved slightly, the fear of private LB shake up is real!\n\nCurious to hear if others are finding that the extra data significantly changes  final biomass regression.\"",
      "votes": null
    },
    {
      "id": "3383274",
      "postDate": "12/29/2025 18:22:47",
      "content": "<p>For sure I will have one of my two selected submissions using a fairly safe (low lr and few epochs) to make a final tune and take the leap.</p>\n<p>With 357 images my prediction would be a huge LB shake up for most unless your a LB probing expert.</p>",
      "rawMarkdown": "For sure I will have one of my two selected submissions using a fairly safe (low lr and few epochs) to make a final tune and take the leap.\n\nWith 357 images my prediction would be a huge LB shake up for most unless your a LB probing expert.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3383243,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "12/29/2025 17:16:07",
      "content": "<p>Have tried all versions of that :)</p>\n<p>With 357 images only - the real fun of this competition is figuring out a whole lot of things that I never would have needed to think about if 5000 images had been supplied.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3383272,
      "author_name": "mugishapacifique",
      "author_url": "",
      "post_date": "12/29/2025 18:18:56",
      "content": "<p>\"Great point! I think most of us are currently 'sacrificing' that 20% of data just to keep our models honest. Once the CV is stable, the jump to all data training is a leap of faith. I've tried it once, and while the score improved slightly, the fear of private LB shake up is real!</p>\n<p>Curious to hear if others are finding that the extra data significantly changes  final biomass regression.\"</p>",
      "votes": null,
      "replies": [
        {
          "id": 3383274,
          "author_name": "pcjimmmy",
          "author_url": "",
          "post_date": "12/29/2025 18:22:47",
          "content": "<p>For sure I will have one of my two selected submissions using a fairly safe (low lr and few epochs) to make a final tune and take the leap.</p>\n<p>With 357 images my prediction would be a huge LB shake up for most unless your a LB probing expert.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3383062": "did u guys used all data in training, first find the best cv strategy and then use all data to train, has anyone tried it?",
    "3383243": "Have tried all versions of that :)\n\nWith 357 images only - the real fun of this competition is figuring out a whole lot of things that I never would have needed to think about if 5000 images had been supplied.",
    "3383272": "\"Great point! I think most of us are currently 'sacrificing' that 20% of data just to keep our models honest. Once the CV is stable, the jump to all data training is a leap of faith. I've tried it once, and while the score improved slightly, the fear of private LB shake up is real!\n\nCurious to hear if others are finding that the extra data significantly changes  final biomass regression.\"",
    "3383274": "For sure I will have one of my two selected submissions using a fairly safe (low lr and few epochs) to make a final tune and take the leap.\n\nWith 357 images my prediction would be a huge LB shake up for most unless your a LB probing expert."
  },
  "source": "meta"
}