{
  "id": 599253,
  "title": "Regarding Submission Time",
  "url": "/competitions/rsna-intracranial-aneurysm-detection/discussion/599253",
  "author_name": "Skelp",
  "post_date": "2025-08-15T06:04:40.317000",
  "votes": 0,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hello everyone,<br>\nI am amidst training a model for the competition.<br>\nOne thing I was confused about, especially considering I've not participated in a Code Competition with inference API usage before, is how the 12 hour submission cap is evaluated.</p>\n<p>Specifically, <strong>does disk loading times count towards the total submission cap</strong>?<br>\nAt least for me, I've noticed that simply loading the DICOM series off of disk takes a couple of seconds for some larger series. Assuming the test set is similarily \"untouched, raw\", this would mean a <strong>considerable</strong> chunk of the submission time is spent on something none of the participants have power over and is practically a constant.<br>\nI assume kaggle themselves are the responsible party for the inference API and submission time evaluation, but I am not absolutely sure on that.</p>\n<p>Knowing this would help figure out how much or how little one can focus on aspects like pre-processing and model-size.<br>\nObviously, the smaller and faster the better in this regard, but getting the most out of the time we get is in the spirit of setting time restrictions in the first place.</p>\n<p>Hope to get some insight on this, I appreciate every bit of info from you guys!<br>\nThanks, and good luck to everybody.</p>",
  "messages": [
    {
      "id": 3270066,
      "postDate": "2025-08-15T16:03:56.417Z",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/skelpdev\" target=\"_blank\">@skelpdev</a>,</p>\n<p>Generally, anything that takes place inside your <code>predict</code> function will count towards the submission limit. This includes loading the data from disk. We have, however, increased the submission time limit by three hours to 12 hours from the usual 9 hours to compensate for the size of the dataset.</p>",
      "rawMarkdown": "Hi @skelpdev,\n\nGenerally, anything that takes place inside your `predict` function will count towards the submission limit. This includes loading the data from disk. We have, however, increased the submission time limit by three hours to 12 hours from the usual 9 hours to compensate for the size of the dataset.",
      "votes": 1
    },
    {
      "id": 3269833,
      "postDate": "2025-08-15T06:04:40.317Z",
      "content": "<p>Hello everyone,<br>\nI am amidst training a model for the competition.<br>\nOne thing I was confused about, especially considering I've not participated in a Code Competition with inference API usage before, is how the 12 hour submission cap is evaluated.</p>\n<p>Specifically, <strong>does disk loading times count towards the total submission cap</strong>?<br>\nAt least for me, I've noticed that simply loading the DICOM series off of disk takes a couple of seconds for some larger series. Assuming the test set is similarily \"untouched, raw\", this would mean a <strong>considerable</strong> chunk of the submission time is spent on something none of the participants have power over and is practically a constant.<br>\nI assume kaggle themselves are the responsible party for the inference API and submission time evaluation, but I am not absolutely sure on that.</p>\n<p>Knowing this would help figure out how much or how little one can focus on aspects like pre-processing and model-size.<br>\nObviously, the smaller and faster the better in this regard, but getting the most out of the time we get is in the spirit of setting time restrictions in the first place.</p>\n<p>Hope to get some insight on this, I appreciate every bit of info from you guys!<br>\nThanks, and good luck to everybody.</p>",
      "rawMarkdown": "Hello everyone,\nI am amidst training a model for the competition.\nOne thing I was confused about, especially considering I've not participated in a Code Competition with inference API usage before, is how the 12 hour submission cap is evaluated.\n\nSpecifically, **does disk loading times count towards the total submission cap**?\nAt least for me, I've noticed that simply loading the DICOM series off of disk takes a couple of seconds for some larger series. Assuming the test set is similarily \"untouched, raw\", this would mean a **considerable** chunk of the submission time is spent on something none of the participants have power over and is practically a constant.\nI assume kaggle themselves are the responsible party for the inference API and submission time evaluation, but I am not absolutely sure on that.\n\nKnowing this would help figure out how much or how little one can focus on aspects like pre-processing and model-size.\nObviously, the smaller and faster the better in this regard, but getting the most out of the time we get is in the spirit of setting time restrictions in the first place.\n\nHope to get some insight on this, I appreciate every bit of info from you guys!\nThanks, and good luck to everybody."
    }
  ],
  "comments": [
    {
      "id": 3270066,
      "author_name": "Ryan Holbrook",
      "author_url": "",
      "post_date": "2025-08-15T16:03:56.417000",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/skelpdev\" target=\"_blank\">@skelpdev</a>,</p>\n<p>Generally, anything that takes place inside your <code>predict</code> function will count towards the submission limit. This includes loading the data from disk. We have, however, increased the submission time limit by three hours to 12 hours from the usual 9 hours to compensate for the size of the dataset.</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3270066": "Hi @skelpdev,\n\nGenerally, anything that takes place inside your `predict` function will count towards the submission limit. This includes loading the data from disk. We have, however, increased the submission time limit by three hours to 12 hours from the usual 9 hours to compensate for the size of the dataset.",
    "3269833": "Hello everyone,\nI am amidst training a model for the competition.\nOne thing I was confused about, especially considering I've not participated in a Code Competition with inference API usage before, is how the 12 hour submission cap is evaluated.\n\nSpecifically, **does disk loading times count towards the total submission cap**?\nAt least for me, I've noticed that simply loading the DICOM series off of disk takes a couple of seconds for some larger series. Assuming the test set is similarily \"untouched, raw\", this would mean a **considerable** chunk of the submission time is spent on something none of the participants have power over and is practically a constant.\nI assume kaggle themselves are the responsible party for the inference API and submission time evaluation, but I am not absolutely sure on that.\n\nKnowing this would help figure out how much or how little one can focus on aspects like pre-processing and model-size.\nObviously, the smaller and faster the better in this regard, but getting the most out of the time we get is in the spirit of setting time restrictions in the first place.\n\nHope to get some insight on this, I appreciate every bit of info from you guys!\nThanks, and good luck to everybody."
  }
}