{
  "id": 302029,
  "title": "Train from scratch?",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/302029",
  "author_name": "",
  "post_date": "2022-01-20T15:02:12.085597200Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Did you guys train a model from scratch or start with a pretrained model on public datasets like coco? Another question, is it necessary to retrain again after adjusting  very few hyper-params like lr? Appreciate guys.</p>",
  "messages": [
    {
      "id": "1657897",
      "postDate": "01/20/2022 15:02:12",
      "content": "<p>Did you guys train a model from scratch or start with a pretrained model on public datasets like coco? Another question, is it necessary to retrain again after adjusting  very few hyper-params like lr? Appreciate guys.</p>",
      "rawMarkdown": "Did you guys train a model from scratch or start with a pretrained model on public datasets like coco? Another question, is it necessary to retrain again after adjusting  very few hyper-params like lr? Appreciate guys.",
      "votes": null
    },
    {
      "id": "1657936",
      "postDate": "01/20/2022 15:41:06",
      "content": "<p>For me, fine tuning a pretrained model performed better than simply training from scratch.</p>",
      "rawMarkdown": "For me, fine tuning a pretrained model performed better than simply training from scratch.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1657936,
      "author_name": "dauriel",
      "author_url": "",
      "post_date": "01/20/2022 15:41:06",
      "content": "<p>For me, fine tuning a pretrained model performed better than simply training from scratch.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1657897": "Did you guys train a model from scratch or start with a pretrained model on public datasets like coco? Another question, is it necessary to retrain again after adjusting  very few hyper-params like lr? Appreciate guys.",
    "1657936": "For me, fine tuning a pretrained model performed better than simply training from scratch."
  },
  "source": "meta"
}