{
  "id": 566583,
  "title": "Clarification on RNA Lengths",
  "url": "/competitions/stanford-rna-3d-folding/discussion/566583",
  "author_name": "",
  "post_date": "2025-03-06T05:34:57.706115900Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I want some details on the maximum length of the RNA sequence that can be present tin the testing set. I need this data because the number of parameters scale pretty steeply with the increase with the increase in input length. So it would be easier on the compute time, if I could stay as close to the optimum as possible. Any Ideas?</p>",
  "messages": [
    {
      "id": "3142117",
      "postDate": "03/06/2025 05:34:57",
      "content": "<p>I want some details on the maximum length of the RNA sequence that can be present tin the testing set. I need this data because the number of parameters scale pretty steeply with the increase with the increase in input length. So it would be easier on the compute time, if I could stay as close to the optimum as possible. Any Ideas?</p>",
      "rawMarkdown": "I want some details on the maximum length of the RNA sequence that can be present tin the testing set. I need this data because the number of parameters scale pretty steeply with the increase with the increase in input length. So it would be easier on the compute time, if I could stay as close to the optimum as possible. Any Ideas?",
      "votes": null
    },
    {
      "id": "3142781",
      "postDate": "03/06/2025 16:58:12",
      "content": "<p>If you use a sequence-to-sequence model to extract embeddings and then a structure prediction regressor, your model size shouldn't depend on the sequence length. The activation/gradient/optimizer weights depend on the sequence length. Not the model :)</p>",
      "rawMarkdown": "If you use a sequence-to-sequence model to extract embeddings and then a structure prediction regressor, your model size shouldn't depend on the sequence length. The activation/gradient/optimizer weights depend on the sequence length. Not the model :)",
      "votes": null
    },
    {
      "id": "3158760",
      "postDate": "03/24/2025 21:12:17",
      "content": "<p>Although the model size doesn't depend on sequence length, the memory used during inference does, so I think it is a valid question. For example one could be worried about Out of Memory issues. <a href=\"https://www.kaggle.com/shujun717\" target=\"_blank\">@shujun717</a> could you clarify what will be the maximum sequence length found in the private test set?</p>",
      "rawMarkdown": "Although the model size doesn't depend on sequence length, the memory used during inference does, so I think it is a valid question. For example one could be worried about Out of Memory issues. @shujun717 could you clarify what will be the maximum sequence length found in the private test set?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3142781,
      "author_name": "louisstefanuto",
      "author_url": "",
      "post_date": "03/06/2025 16:58:12",
      "content": "<p>If you use a sequence-to-sequence model to extract embeddings and then a structure prediction regressor, your model size shouldn't depend on the sequence length. The activation/gradient/optimizer weights depend on the sequence length. Not the model :)</p>",
      "votes": null,
      "replies": [
        {
          "id": 3158760,
          "author_name": "aleixgim",
          "author_url": "",
          "post_date": "03/24/2025 21:12:17",
          "content": "<p>Although the model size doesn't depend on sequence length, the memory used during inference does, so I think it is a valid question. For example one could be worried about Out of Memory issues. <a href=\"https://www.kaggle.com/shujun717\" target=\"_blank\">@shujun717</a> could you clarify what will be the maximum sequence length found in the private test set?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3142117": "I want some details on the maximum length of the RNA sequence that can be present tin the testing set. I need this data because the number of parameters scale pretty steeply with the increase with the increase in input length. So it would be easier on the compute time, if I could stay as close to the optimum as possible. Any Ideas?",
    "3142781": "If you use a sequence-to-sequence model to extract embeddings and then a structure prediction regressor, your model size shouldn't depend on the sequence length. The activation/gradient/optimizer weights depend on the sequence length. Not the model :)",
    "3158760": "Although the model size doesn't depend on sequence length, the memory used during inference does, so I think it is a valid question. For example one could be worried about Out of Memory issues. @shujun717 could you clarify what will be the maximum sequence length found in the private test set?"
  },
  "source": "meta"
}