![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15255201%2F2de6dec1eff1f4d8cedd49690bb4c192%2FScreenshot%202026-03-14%20181950.png?generation=1773483617375516&alt=media\")
\n160/4000After reading through the top-ranked solution, I came across the 4,000-epoch log file shared by @pgeiger .\nThe upward trend in the later stages of training caught my attention, so I decided to try running it for 4,000 epochs.
\n| Patch/Epoch | \nMean Validation Dice | \nFinal Score | \nSurf Dice | \nVOI Score | \nTopo Score | \n
|---|---|---|---|---|---|
| 192/1000 | \n0.6083 | \n0.6395 | \n0.8744 | \n0.5673 | \n0.4496 | \n
| 160/4000 | \n0.6173 | \n0.6446 | \n0.8867 | \n0.5678 | \n0.4516 | \n
Note: 192/1000 differs from the final CV in our 9th solution because we made minor adjustments.
\n\n160/4000
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15255201%2F33ae42bbecad89cf34e8a170cd2b9e2f%2FScreenshot%202026-03-14%20182033.png?generation=1773483650704097&alt=media\")
\n192/1000
After reading through the top-ranked solution, I came across the 4,000-epoch log file shared by @pgeiger .\nThe upward trend in the later stages of training caught my attention, so I decided to try running it for 4,000 epochs.
\n| Patch/Epoch | \nMean Validation Dice | \nFinal Score | \nSurf Dice | \nVOI Score | \nTopo Score | \n
|---|---|---|---|---|---|
| 192/1000 | \n0.6083 | \n0.6395 | \n0.8744 | \n0.5673 | \n0.4496 | \n
| 160/4000 | \n0.6173 | \n0.6446 | \n0.8867 | \n0.5678 | \n0.4516 | \n
Note: 192/1000 differs from the final CV in our 9th solution because we made minor adjustments.
\n\n160/4000
\n192/1000
Nice result @ggayoayogg!
\nCould you share whether the validation losses increased during this process? I tried continuing training from one of mine checkpoint and noticed that the Dice loss continued to go down while CE loss started to rise. However, I'm not using nnU-Net training pipeline, so the behavior might be different.
\nI also wonder if this improvement is not related to this problem:\n https://www.kaggle.com/competitions/vesuvius-challenge-surface-detection/discussion/679224
\n\n\n78 out of 110 test samples (≈71%) come from scroll IDs also present in the training data, meaning a large portion of the test set shares scroll IDs with the training set. This suggests that models may benefit significantly from in-distribution generalization rather than purely out-of-distribution generalization.
\n
Meaning that by training more epochs we could simply be overfitting to some scrolls IDs and this would not be useful to improve result in others out-of-distribution scrolls.
", "rawMarkdown": "Nice result @ggayoayogg!\n\nCould you share whether the validation losses increased during this process? I tried continuing training from one of mine checkpoint and noticed that the Dice loss continued to go down while CE loss started to rise. However, I'm not using nnU-Net training pipeline, so the behavior might be different.\n\nI also wonder if this improvement is not related to this problem:\n https://www.kaggle.com/competitions/vesuvius-challenge-surface-detection/discussion/679224\n>78 out of 110 test samples (≈71%) come from scroll IDs also present in the training data, meaning a large portion of the test set shares scroll IDs with the training set. This suggests that models may benefit significantly from in-distribution generalization rather than purely out-of-distribution generalization.\n \n\nMeaning that by training more epochs we could simply be overfitting to some scrolls IDs and this would not be useful to improve result in others out-of-distribution scrolls.", "votes": null }, { "id": "3420999", "postDate": "03/14/2026 11:59:29", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15255201%2F0f941ad5b5f00a73ccfc9423f045dba3%2Fprogress.png?generation=1773489345120790&alt=media\")
I feel validation losses hasn't changed.\nI did not monitor the individual curves for DC_SkelREC_and_CE_loss; I only monitored the combined curve.
", "rawMarkdown": "\n\nI feel validation losses hasn't changed.\nI did not monitor the individual curves for DC_SkelREC_and_CE_loss; I only monitored the combined curve.", "votes": null }, { "id": "3421045", "postDate": "03/14/2026 13:55:43", "content": "Thanks for sharing @ggayoayogg!\nOur loss curves are very different. Maybe because of optmizer and weight decay of nnUnet vs ours. I still have to explore this. For comparison:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2221915%2F2132a90d0c480ec1f408ad5c8f36bb76%2Fcurves.png?generation=1773496420505700&alt=media\")
https://github.com/ultralytics/ultralytics/blob/main/ultralytics/optim/muon.py
", "rawMarkdown": "https://github.com/ultralytics/ultralytics/blob/main/ultralytics/optim/muon.py", "votes": null } ], "comments": [ { "id": 3420984, "author_name": "sersasj", "author_url": "", "post_date": "03/14/2026 11:12:50", "content": "Nice result @ggayoayogg!
\nCould you share whether the validation losses increased during this process? I tried continuing training from one of mine checkpoint and noticed that the Dice loss continued to go down while CE loss started to rise. However, I'm not using nnU-Net training pipeline, so the behavior might be different.
\nI also wonder if this improvement is not related to this problem:\n https://www.kaggle.com/competitions/vesuvius-challenge-surface-detection/discussion/679224
\n\n\n78 out of 110 test samples (≈71%) come from scroll IDs also present in the training data, meaning a large portion of the test set shares scroll IDs with the training set. This suggests that models may benefit significantly from in-distribution generalization rather than purely out-of-distribution generalization.
\n
Meaning that by training more epochs we could simply be overfitting to some scrolls IDs and this would not be useful to improve result in others out-of-distribution scrolls.
", "votes": null, "replies": [ { "id": 3420999, "author_name": "ggayoayogg", "author_url": "", "post_date": "03/14/2026 11:59:29", "content": "
I feel validation losses hasn't changed.\nI did not monitor the individual curves for DC_SkelREC_and_CE_loss; I only monitored the combined curve.
", "votes": null, "replies": [ { "id": 3421045, "author_name": "sersasj", "author_url": "", "post_date": "03/14/2026 13:55:43", "content": "Thanks for sharing @ggayoayogg!\nOur loss curves are very different. Maybe because of optmizer and weight decay of nnUnet vs ours. I still have to explore this. For comparison:
\n
https://github.com/ultralytics/ultralytics/blob/main/ultralytics/optim/muon.py
", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "3420964": "After reading through the top-ranked solution, I came across the 4,000-epoch log file shared by @pgeiger .\nThe upward trend in the later stages of training caught my attention, so I decided to try running it for 4,000 epochs.\n\n| Patch/Epoch | Mean Validation Dice | Final Score | Surf Dice | VOI Score | Topo Score |\n|---|---|---|---|---|---|\n| **192/1000** | 0.6083 | 0.6395 | 0.8744 | 0.5673 | 0.4496 |\n| **160/4000** | 0.6173 | 0.6446 | 0.8867 | 0.5678 | 0.4516 |\n\nNote: 192/1000 differs from the final CV in our 9th solution because we made minor adjustments.\n\n\n160/4000\n\n\n192/1000\n\n#By the way, if you use ReLU during training 160, the GPU memory usage is less than 16 GB, so it fits on a T4.", "3420984": "Nice result @ggayoayogg!\n\nCould you share whether the validation losses increased during this process? I tried continuing training from one of mine checkpoint and noticed that the Dice loss continued to go down while CE loss started to rise. However, I'm not using nnU-Net training pipeline, so the behavior might be different.\n\nI also wonder if this improvement is not related to this problem:\n https://www.kaggle.com/competitions/vesuvius-challenge-surface-detection/discussion/679224\n>78 out of 110 test samples (≈71%) come from scroll IDs also present in the training data, meaning a large portion of the test set shares scroll IDs with the training set. This suggests that models may benefit significantly from in-distribution generalization rather than purely out-of-distribution generalization.\n \n\nMeaning that by training more epochs we could simply be overfitting to some scrolls IDs and this would not be useful to improve result in others out-of-distribution scrolls.", "3420999": "\n\nI feel validation losses hasn't changed.\nI did not monitor the individual curves for DC_SkelREC_and_CE_loss; I only monitored the combined curve.", "3421045": "Thanks for sharing @ggayoayogg!\nOur loss curves are very different. Maybe because of optmizer and weight decay of nnUnet vs ours. I still have to explore this. For comparison:\n\n", "3421057": "https://github.com/ultralytics/ultralytics/blob/main/ultralytics/optim/muon.py" }, "source": "meta" }