{
  "id": 671613,
  "title": "Running more than 200 epochs?",
  "url": "/competitions/vesuvius-challenge-surface-detection/discussion/671613",
  "author_name": "",
  "post_date": "2026-02-03T00:14:03.885707200Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Can I know What's the training time per epoch, like iterations per second? I can see some discussion on running 1000 epoch default on nnUNet? If so, what's the training time? </p>\n<p>MonAI and TransUNet training time if running and how many epochs ?</p>",
  "messages": [
    {
      "id": "3401113",
      "postDate": "02/03/2026 00:14:03",
      "content": "<p>Can I know What's the training time per epoch, like iterations per second? I can see some discussion on running 1000 epoch default on nnUNet? If so, what's the training time? </p>\n<p>MonAI and TransUNet training time if running and how many epochs ?</p>",
      "rawMarkdown": "Can I know What's the training time per epoch, like iterations per second? I can see some discussion on running 1000 epoch default on nnUNet? If so, what's the training time? \n\nMonAI and TransUNet training time if running and how many epochs ?",
      "votes": null
    },
    {
      "id": "3401225",
      "postDate": "02/03/2026 07:52:26",
      "content": "<p>at least 100 epoch to see results (e.g. compare different models for selection) \nstop  at 200 if overfitting. but if you good augmentation you can go on to 800+.</p>\n<p>200 takes 7hr on m local machine(Ada rtx 6000)</p>\n<p>there are many 3d segmentation network. the ones in public code may be be the most efficient.\nchoose a mid-size network, tranformer encoder generally work the better. i recommend unetr++ and variants for speed</p>",
      "rawMarkdown": "at least 100 epoch to see results (e.g. compare different models for selection) \nstop  at 200 if overfitting. but if you good augmentation you can go on to 800+.\n\n200 takes 7hr on m local machine(Ada rtx 6000)\n\nthere are many 3d segmentation network. the ones in public code may be be the most efficient.\nchoose a mid-size network, tranformer encoder generally work the better. i recommend unetr++ and variants for speed",
      "votes": null
    },
    {
      "id": "3401253",
      "postDate": "02/03/2026 09:49:26",
      "content": "<p>thanks. let me try that. </p>",
      "rawMarkdown": "thanks. let me try that.",
      "votes": null
    },
    {
      "id": "3401255",
      "postDate": "02/03/2026 09:50:40",
      "content": "<p>Thanks for the training hour info. I ended using Hausdorff for loss calculation, it slowed training down. And I got curious. </p>\n<p>1.44it/s, loss=0.1775, dice=0.5262 .. augmentation with elastic deform takes my pass iteration per second lower and so slow. 200 epochs is like 20hrs for me. lol</p>",
      "rawMarkdown": "Thanks for the training hour info. I ended using Hausdorff for loss calculation, it slowed training down. And I got curious. \n\n1.44it/s, loss=0.1775, dice=0.5262 .. augmentation with elastic deform takes my pass iteration per second lower and so slow. 200 epochs is like 20hrs for me. lol",
      "votes": null
    },
    {
      "id": "3401449",
      "postDate": "02/03/2026 18:00:47",
      "content": "<p>Can I ask tips/guidance if using any of these:\nValidation Dice based on: Dice, clDice or skeleton recall, or precision and recall. </p>\n<p>These directly didn't relate to the competition metric score, Combination of: clDice, Surface Dice and VOI score neither correlated with final submission score. </p>\n<p>Whats the benchmark metric you look for in your validation score? for example noted increase Precision/Recall correlated to increase in submission score.</p>",
      "rawMarkdown": "Can I ask tips/guidance if using any of these:\nValidation Dice based on: Dice, clDice or skeleton recall, or precision and recall. \n\nThese directly didn't relate to the competition metric score, Combination of: clDice, Surface Dice and VOI score neither correlated with final submission score. \n\nWhats the benchmark metric you look for in your validation score? for example noted increase Precision/Recall correlated to increase in submission score.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3401225,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "02/03/2026 07:52:26",
      "content": "<p>at least 100 epoch to see results (e.g. compare different models for selection) \nstop  at 200 if overfitting. but if you good augmentation you can go on to 800+.</p>\n<p>200 takes 7hr on m local machine(Ada rtx 6000)</p>\n<p>there are many 3d segmentation network. the ones in public code may be be the most efficient.\nchoose a mid-size network, tranformer encoder generally work the better. i recommend unetr++ and variants for speed</p>",
      "votes": null,
      "replies": [
        {
          "id": 3401253,
          "author_name": "rajeshthevar",
          "author_url": "",
          "post_date": "02/03/2026 09:49:26",
          "content": "<p>thanks. let me try that. </p>",
          "votes": null,
          "replies": [
            {
              "id": 3401255,
              "author_name": "rajeshthevar",
              "author_url": "",
              "post_date": "02/03/2026 09:50:40",
              "content": "<p>Thanks for the training hour info. I ended using Hausdorff for loss calculation, it slowed training down. And I got curious. </p>\n<p>1.44it/s, loss=0.1775, dice=0.5262 .. augmentation with elastic deform takes my pass iteration per second lower and so slow. 200 epochs is like 20hrs for me. lol</p>",
              "votes": null,
              "replies": []
            }
          ]
        },
        {
          "id": 3401449,
          "author_name": "rajeshthevar",
          "author_url": "",
          "post_date": "02/03/2026 18:00:47",
          "content": "<p>Can I ask tips/guidance if using any of these:\nValidation Dice based on: Dice, clDice or skeleton recall, or precision and recall. </p>\n<p>These directly didn't relate to the competition metric score, Combination of: clDice, Surface Dice and VOI score neither correlated with final submission score. </p>\n<p>Whats the benchmark metric you look for in your validation score? for example noted increase Precision/Recall correlated to increase in submission score.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3401113": "Can I know What's the training time per epoch, like iterations per second? I can see some discussion on running 1000 epoch default on nnUNet? If so, what's the training time? \n\nMonAI and TransUNet training time if running and how many epochs ?",
    "3401225": "at least 100 epoch to see results (e.g. compare different models for selection) \nstop  at 200 if overfitting. but if you good augmentation you can go on to 800+.\n\n200 takes 7hr on m local machine(Ada rtx 6000)\n\nthere are many 3d segmentation network. the ones in public code may be be the most efficient.\nchoose a mid-size network, tranformer encoder generally work the better. i recommend unetr++ and variants for speed",
    "3401253": "thanks. let me try that.",
    "3401255": "Thanks for the training hour info. I ended using Hausdorff for loss calculation, it slowed training down. And I got curious. \n\n1.44it/s, loss=0.1775, dice=0.5262 .. augmentation with elastic deform takes my pass iteration per second lower and so slow. 200 epochs is like 20hrs for me. lol",
    "3401449": "Can I ask tips/guidance if using any of these:\nValidation Dice based on: Dice, clDice or skeleton recall, or precision and recall. \n\nThese directly didn't relate to the competition metric score, Combination of: clDice, Surface Dice and VOI score neither correlated with final submission score. \n\nWhats the benchmark metric you look for in your validation score? for example noted increase Precision/Recall correlated to increase in submission score."
  },
  "source": "meta"
}