{ "id": 671984, "title": "Notebook Time Limit Exceeded Error", "url": "/competitions/vesuvius-challenge-surface-detection/discussion/671984", "author_name": "", "post_date": "2026-02-05T09:56:53.391661Z", "votes": -2, "comment_count": 1, "views": 0, "content": "

Hi everyone,\nI am facing a Time Limit Exceeded (TLE) error on submission and trying to figure out if my inference strategy is fundamentally too slow or if I have a specific implementation bug.\nMy notebook runs for around 1-1.5 hrs on GPU P100. But after submitting it, I get a notebook Time limit exceeded error after being run for 10 hrs.\nMy Approach:\nModel: Standard 2D U-Net.\nInput: I am using a \"2.5D\" approach, taking 5 slices as input (current slice +/- 2 neighbors).\nResolution: During inference, I resize the input stack to 256x256 before feeding it to the model, then resize the prediction back to the original volume size.\nMy Inference Loop: Currently, I am iterating through the z-axis of the volume one slice at a time.\nFor every z, I load the 5 necessary slices from the .tif file.\nI resize them, normalize, and convert to tensor.\nI run the model (Batch size = 1).\nI resize the output mask and save it to a numpy array.\nThe Issue: This runs locally on sample data, but times out on the hidden test set.\nQuestions:

\n
    \n
  1. Is processing the volume slice-by-slice (Batch size 1) generally too slow for this competition's hidden test set size?
    2. \n
  2. Do I need to load the entire volume into RAM at once instead of reading from disk for every slice?\nWhat I feel is that inside my inference loop, I am opening a big TIFF file inside the loop for every single slice and that might be causing a huge overhead.\nfor z in range(pad, d - pad):\nwith Image.open(p) as img:
    \n stack = []\n for k in range(z-pad, z+pad+1):\n img.seek(k)\n stack.append(np.array(img))\nThanks for any suggestions.
    3. \n
", "messages": [ { "id": "3402136", "postDate": "02/05/2026 09:56:53", "content": "

Hi everyone,\nI am facing a Time Limit Exceeded (TLE) error on submission and trying to figure out if my inference strategy is fundamentally too slow or if I have a specific implementation bug.\nMy notebook runs for around 1-1.5 hrs on GPU P100. But after submitting it, I get a notebook Time limit exceeded error after being run for 10 hrs.\nMy Approach:\nModel: Standard 2D U-Net.\nInput: I am using a \"2.5D\" approach, taking 5 slices as input (current slice +/- 2 neighbors).\nResolution: During inference, I resize the input stack to 256x256 before feeding it to the model, then resize the prediction back to the original volume size.\nMy Inference Loop: Currently, I am iterating through the z-axis of the volume one slice at a time.\nFor every z, I load the 5 necessary slices from the .tif file.\nI resize them, normalize, and convert to tensor.\nI run the model (Batch size = 1).\nI resize the output mask and save it to a numpy array.\nThe Issue: This runs locally on sample data, but times out on the hidden test set.\nQuestions:

\n
    \n
  1. Is processing the volume slice-by-slice (Batch size 1) generally too slow for this competition's hidden test set size?
    2. \n
  2. Do I need to load the entire volume into RAM at once instead of reading from disk for every slice?\nWhat I feel is that inside my inference loop, I am opening a big TIFF file inside the loop for every single slice and that might be causing a huge overhead.\nfor z in range(pad, d - pad):\nwith Image.open(p) as img:
    \n stack = []\n for k in range(z-pad, z+pad+1):\n img.seek(k)\n stack.append(np.array(img))\nThanks for any suggestions.
    3. \n
", "rawMarkdown": "Hi everyone,\n\nI am facing a Time Limit Exceeded (TLE) error on submission and trying to figure out if my inference strategy is fundamentally too slow or if I have a specific implementation bug.\n\nMy notebook runs for around 1-1.5 hrs on GPU P100. But after submitting it, I get a notebook Time limit exceeded error after being run for 10 hrs.\n\nMy Approach:\n\nModel: Standard 2D U-Net.\n\nInput: I am using a \"2.5D\" approach, taking 5 slices as input (current slice +/- 2 neighbors).\n\nResolution: During inference, I resize the input stack to 256x256 before feeding it to the model, then resize the prediction back to the original volume size.\n\nMy Inference Loop: Currently, I am iterating through the z-axis of the volume one slice at a time.\n\nFor every z, I load the 5 necessary slices from the .tif file.\n\nI resize them, normalize, and convert to tensor.\n\nI run the model (Batch size = 1).\n\nI resize the output mask and save it to a numpy array.\n\nThe Issue: This runs locally on sample data, but times out on the hidden test set.\n\nQuestions:\n\n1. Is processing the volume slice-by-slice (Batch size 1) generally too slow for this competition's hidden test set size?\n\n2. Do I need to load the entire volume into RAM at once instead of reading from disk for every slice?\n\n\nWhat I feel is that inside my inference loop, I am opening a big TIFF file inside the loop for every single slice and that might be causing a huge overhead.\n\nfor z in range(pad, d - pad):\n with Image.open(p) as img: \n stack = []\n for k in range(z-pad, z+pad+1):\n img.seek(k)\n stack.append(np.array(img))\n\n\nThanks for any suggestions.", "votes": null }, { "id": "3402156", "postDate": "02/05/2026 11:08:25", "content": "

My notebook runs for around 1-1.5 hrs on GPU P100\nis this training or inference after submitting for competetion?

", "rawMarkdown": "My notebook runs for around 1-1.5 hrs on GPU P100\nis this training or inference after submitting for competetion?", "votes": null } ], "comments": [ { "id": 3402156, "author_name": "arjunashokbhandary", "author_url": "", "post_date": "02/05/2026 11:08:25", "content": "

My notebook runs for around 1-1.5 hrs on GPU P100\nis this training or inference after submitting for competetion?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3402136": "Hi everyone,\n\nI am facing a Time Limit Exceeded (TLE) error on submission and trying to figure out if my inference strategy is fundamentally too slow or if I have a specific implementation bug.\n\nMy notebook runs for around 1-1.5 hrs on GPU P100. But after submitting it, I get a notebook Time limit exceeded error after being run for 10 hrs.\n\nMy Approach:\n\nModel: Standard 2D U-Net.\n\nInput: I am using a \"2.5D\" approach, taking 5 slices as input (current slice +/- 2 neighbors).\n\nResolution: During inference, I resize the input stack to 256x256 before feeding it to the model, then resize the prediction back to the original volume size.\n\nMy Inference Loop: Currently, I am iterating through the z-axis of the volume one slice at a time.\n\nFor every z, I load the 5 necessary slices from the .tif file.\n\nI resize them, normalize, and convert to tensor.\n\nI run the model (Batch size = 1).\n\nI resize the output mask and save it to a numpy array.\n\nThe Issue: This runs locally on sample data, but times out on the hidden test set.\n\nQuestions:\n\n1. Is processing the volume slice-by-slice (Batch size 1) generally too slow for this competition's hidden test set size?\n\n2. Do I need to load the entire volume into RAM at once instead of reading from disk for every slice?\n\n\nWhat I feel is that inside my inference loop, I am opening a big TIFF file inside the loop for every single slice and that might be causing a huge overhead.\n\nfor z in range(pad, d - pad):\n with Image.open(p) as img: \n stack = []\n for k in range(z-pad, z+pad+1):\n img.seek(k)\n stack.append(np.array(img))\n\n\nThanks for any suggestions.", "3402156": "My notebook runs for around 1-1.5 hrs on GPU P100\nis this training or inference after submitting for competetion?" }, "source": "meta" }