{
  "id": 580560,
  "title": "Custom Architectures From Scratch",
  "url": "/competitions/stanford-rna-3d-folding/discussion/580560",
  "author_name": "",
  "post_date": "2025-05-24T23:37:53.452986600Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Has anyone tried experimenting with training custom model architectures from scratch? What has worked/not worked? From what I've seen so far, this type of problem greatly benefits from incorporating pre-trained models in some way, but would love to hear others thoughts. </p>",
  "messages": [
    {
      "id": "3208921",
      "postDate": "05/24/2025 23:37:53",
      "content": "<p>Has anyone tried experimenting with training custom model architectures from scratch? What has worked/not worked? From what I've seen so far, this type of problem greatly benefits from incorporating pre-trained models in some way, but would love to hear others thoughts. </p>",
      "rawMarkdown": "Has anyone tried experimenting with training custom model architectures from scratch? What has worked/not worked? From what I've seen so far, this type of problem greatly benefits from incorporating pre-trained models in some way, but would love to hear others thoughts.",
      "votes": null
    },
    {
      "id": "3209885",
      "postDate": "05/26/2025 12:59:15",
      "content": "<p>I think it's too tough to train a custom model. pre-trained model might be easier and low required in computability</p>",
      "rawMarkdown": "I think it's too tough to train a custom model. pre-trained model might be easier and low required in computability",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3209885,
      "author_name": "nigmatrahim4",
      "author_url": "",
      "post_date": "05/26/2025 12:59:15",
      "content": "<p>I think it's too tough to train a custom model. pre-trained model might be easier and low required in computability</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3208921": "Has anyone tried experimenting with training custom model architectures from scratch? What has worked/not worked? From what I've seen so far, this type of problem greatly benefits from incorporating pre-trained models in some way, but would love to hear others thoughts.",
    "3209885": "I think it's too tough to train a custom model. pre-trained model might be easier and low required in computability"
  },
  "source": "meta"
}