Hi all, this is my first time participating in a segmentation competition. I started by training SegResNet on a single A4000 GPU. It took me ~3 days to train for 200 epochs, and I achieved a leaderboard score of 0.475. If possible, could you share what GPU you are using and how long it takes to train one fold? I feel that my code is not optimized yet.
", "messages": [ { "id": "3389681", "postDate": "01/11/2026 17:23:05", "content": "Hi all, this is my first time participating in a segmentation competition. I started by training SegResNet on a single A4000 GPU. It took me ~3 days to train for 200 epochs, and I achieved a leaderboard score of 0.475. If possible, could you share what GPU you are using and how long it takes to train one fold? I feel that my code is not optimized yet.
", "rawMarkdown": "Hi all, this is my first time participating in a segmentation competition. I started by training SegResNet on a single A4000 GPU. It took me ~3 days to train for 200 epochs, and I achieved a leaderboard score of 0.475. If possible, could you share what GPU you are using and how long it takes to train one fold? I feel that my code is not optimized yet.", "votes": null }, { "id": "3389721", "postDate": "01/11/2026 18:57:44", "content": "There are public notebooks which train UNets a lot faster than what your code does, refer to them. 3 days for 200 epochs is too much.
", "rawMarkdown": "There are public notebooks which train UNets a lot faster than what your code does, refer to them. 3 days for 200 epochs is too much.", "votes": null }, { "id": "3389870", "postDate": "01/12/2026 05:32:47", "content": "Thanks. I found some comments in the discussion where someone shared that they trained nnUNet on a 4070TiS in only about 40,000 seconds, which is very impressive.
", "rawMarkdown": "Thanks. I found some comments in the discussion where someone shared that they trained nnUNet on a 4070TiS in only about 40,000 seconds, which is very impressive.", "votes": null }, { "id": "3389920", "postDate": "01/12/2026 07:38:34", "content": "An AI code agent can help you with profiling to see where the bottlenecks are
", "rawMarkdown": "An AI code agent can help you with profiling to see where the bottlenecks are", "votes": null }, { "id": "3390289", "postDate": "01/12/2026 23:08:52", "content": "FYI. My LB.548 score was using TransUnet, with a Google Colab A100 (80GB) GPU, taking 121s per epoch.\nTherefore, if were to train for 200 epochs, it would take about 6.7 hours.
\nEpoch 3\n195/195 ━━━━━━━━━━━━━━━━━━━━ 121s \nEpoch 4\n195/195 ━━━━━━━━━━━━━━━━━━━━ 121s \nEpoch 5\n195/195 ━━━━━━━━━━━━━━━━━━━━ 121s \n\n...etc\nThanks. If possible, could you share the total time required for one fold? I have ~200 units on google colab, it seem worth trying transUnet
", "rawMarkdown": "Thanks. If possible, could you share the total time required for one fold? I have ~200 units on google colab, it seem worth trying transUnet", "votes": null }, { "id": "3391184", "postDate": "01/14/2026 14:53:50", "content": "model size, ROI, batches per epoch, all this also deeply influences time per epoch.
", "rawMarkdown": "model size, ROI, batches per epoch, all this also deeply influences time per epoch.", "votes": null }, { "id": "3391357", "postDate": "01/14/2026 19:27:54", "content": "You will have to wait a bit for TPUs on Kaggle but it is a good option for TransUNet. Look for Manas' notebook.
", "rawMarkdown": "You will have to wait a bit for TPUs on Kaggle but it is a good option for TransUNet. Look for Manas' notebook.", "votes": null }, { "id": "3391492", "postDate": "01/15/2026 04:17:28", "content": "@rob1080ti Thanks, but I’m often 50+ in the queue 😭, nnUNet seem to be appropriate, only take 300s per epoch
", "rawMarkdown": "rob1080ti Thanks, but I’m often 50+ in the queue 😭, nnUNet seem to be appropriate, only take 300s per epoch", "votes": null }, { "id": "3391883", "postDate": "01/15/2026 21:36:26", "content": "You could simply do a full commit or run all with file persistence but you’d need to keep your computer connected or you’ll lose your place in the queue
", "rawMarkdown": "You could simply do a full commit or run all with file persistence but you’d need to keep your computer connected or you’ll lose your place in the queue", "votes": null } ], "comments": [ { "id": 3389721, "author_name": "choudharymanas", "author_url": "", "post_date": "01/11/2026 18:57:44", "content": "There are public notebooks which train UNets a lot faster than what your code does, refer to them. 3 days for 200 epochs is too much.
", "votes": null, "replies": [ { "id": 3389870, "author_name": "nguyncdngs", "author_url": "", "post_date": "01/12/2026 05:32:47", "content": "Thanks. I found some comments in the discussion where someone shared that they trained nnUNet on a 4070TiS in only about 40,000 seconds, which is very impressive.
", "votes": null, "replies": [] } ] }, { "id": 3389920, "author_name": "wzyfromhust", "author_url": "", "post_date": "01/12/2026 07:38:34", "content": "An AI code agent can help you with profiling to see where the bottlenecks are
", "votes": null, "replies": [] }, { "id": 3390289, "author_name": "hideyukizushi", "author_url": "", "post_date": "01/12/2026 23:08:52", "content": "FYI. My LB.548 score was using TransUnet, with a Google Colab A100 (80GB) GPU, taking 121s per epoch.\nTherefore, if were to train for 200 epochs, it would take about 6.7 hours.
\nEpoch 3\n195/195 ━━━━━━━━━━━━━━━━━━━━ 121s \nEpoch 4\n195/195 ━━━━━━━━━━━━━━━━━━━━ 121s \nEpoch 5\n195/195 ━━━━━━━━━━━━━━━━━━━━ 121s \n\n...etc\nThanks. If possible, could you share the total time required for one fold? I have ~200 units on google colab, it seem worth trying transUnet
", "votes": null, "replies": [ { "id": 3391357, "author_name": "rob1080ti", "author_url": "", "post_date": "01/14/2026 19:27:54", "content": "You will have to wait a bit for TPUs on Kaggle but it is a good option for TransUNet. Look for Manas' notebook.
", "votes": null, "replies": [ { "id": 3391492, "author_name": "nguyncdngs", "author_url": "", "post_date": "01/15/2026 04:17:28", "content": "@rob1080ti Thanks, but I’m often 50+ in the queue 😭, nnUNet seem to be appropriate, only take 300s per epoch
", "votes": null, "replies": [ { "id": 3391883, "author_name": "rob1080ti", "author_url": "", "post_date": "01/15/2026 21:36:26", "content": "You could simply do a full commit or run all with file persistence but you’d need to keep your computer connected or you’ll lose your place in the queue
", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 3391184, "author_name": "arjunashokbhandary", "author_url": "", "post_date": "01/14/2026 14:53:50", "content": "model size, ROI, batches per epoch, all this also deeply influences time per epoch.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3389681": "Hi all, this is my first time participating in a segmentation competition. I started by training SegResNet on a single A4000 GPU. It took me ~3 days to train for 200 epochs, and I achieved a leaderboard score of 0.475. If possible, could you share what GPU you are using and how long it takes to train one fold? I feel that my code is not optimized yet.", "3389721": "There are public notebooks which train UNets a lot faster than what your code does, refer to them. 3 days for 200 epochs is too much.", "3389870": "Thanks. I found some comments in the discussion where someone shared that they trained nnUNet on a 4070TiS in only about 40,000 seconds, which is very impressive.", "3389920": "An AI code agent can help you with profiling to see where the bottlenecks are", "3390289": "FYI. My LB.548 score was using TransUnet, with a **Google Colab A100 (80GB) GPU, taking 121s per epoch.**\nTherefore, if were to train for 200 epochs, it would take about 6.7 hours.\n\n```\nEpoch 3\n195/195 ━━━━━━━━━━━━━━━━━━━━ 121s \nEpoch 4\n195/195 ━━━━━━━━━━━━━━━━━━━━ 121s \nEpoch 5\n195/195 ━━━━━━━━━━━━━━━━━━━━ 121s \n\n...etc\n```", "3390432": "Thanks. If possible, could you share the total time required for one fold? I have ~200 units on google colab, it seem worth trying transUnet", "3391184": "model size, ROI, batches per epoch, all this also deeply influences time per epoch.", "3391357": "You will have to wait a bit for TPUs on Kaggle but it is a good option for TransUNet. Look for Manas' notebook.", "3391492": "rob1080ti Thanks, but I’m often 50+ in the queue 😭, nnUNet seem to be appropriate, only take 300s per epoch", "3391883": "You could simply do a full commit or run all with file persistence but you’d need to keep your computer connected or you’ll lose your place in the queue" }, "source": "meta" }