![\"imge\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F9366732%2Fb67d38a569aa0527859d6b9e15abf511%2Fddd.png?generation=1772305736041024&alt=media\")
First of all,
\nThanks to the organizers for this fun and important competition, and congratulations to all the winners.
\nThis competition taught me a lot about 3D segmentation, long-term training stability, checkpoint management, and how Public and Private leaderboards can diverge during extended training.
Special thanks to @Innat for sharing the TFRecord dataset and the baseline training code.
\nThat baseline helped me start quickly and focus on improving training stability and long-run optimization.
I trained a 3D TransUNet model for volumetric segmentation and continuously resumed training from saved checkpoints. Instead of restarting training multiple times, I extended the same model progressively up to 1325 epochs.
\nThe main idea was:
\nI used TransUNet with a SEResNeXt50 encoder.
\n TransUNet(\n input_shape=(256,256,256,1),\n encoder_name=\"seresnext50\",\n classifier_activation=None,\n num_classes=3,\n dropout=0.1,\n )\nI trained in multiple phases and resumed from saved checkpoints:
\n| Epoch Range | \nLB Score | \nPB Score | \n
|---|---|---|
| 0 → 237 | \n0.546 | \n0.560 | \n
| 237 → 431 | \n0.553 | \n0.578 | \n
| 431 → 703 | \n0.549 | \n0.582 | \n
| 703 → 940 | \n0.569 | \n0.571 | \n
| 940 → 1325 | \n0.536 | \n0.586 | \n
Training code:
\nhttps://www.kaggle.com/code/engadamalmohammedi/vcsd3d25-train
Thank you again to the organizers and the community for this great experience.
", "messages": [ { "id": "3415366", "postDate": "02/28/2026 19:10:23", "content": "First of all,
\nThanks to the organizers for this fun and important competition, and congratulations to all the winners.
\nThis competition taught me a lot about 3D segmentation, long-term training stability, checkpoint management, and how Public and Private leaderboards can diverge during extended training.
Special thanks to @Innat for sharing the TFRecord dataset and the baseline training code.
\nThat baseline helped me start quickly and focus on improving training stability and long-run optimization.
I trained a 3D TransUNet model for volumetric segmentation and continuously resumed training from saved checkpoints. Instead of restarting training multiple times, I extended the same model progressively up to 1325 epochs.
\nThe main idea was:
\nI used TransUNet with a SEResNeXt50 encoder.
\n TransUNet(\n input_shape=(256,256,256,1),\n encoder_name=\"seresnext50\",\n classifier_activation=None,\n num_classes=3,\n dropout=0.1,\n )\nI trained in multiple phases and resumed from saved checkpoints:
\n| Epoch Range | \nLB Score | \nPB Score | \n
|---|---|---|
| 0 → 237 | \n0.546 | \n0.560 | \n
| 237 → 431 | \n0.553 | \n0.578 | \n
| 431 → 703 | \n0.549 | \n0.582 | \n
| 703 → 940 | \n0.569 | \n0.571 | \n
| 940 → 1325 | \n0.536 | \n0.586 | \n
Training code:
\nhttps://www.kaggle.com/code/engadamalmohammedi/vcsd3d25-train
Thank you again to the organizers and the community for this great experience.
", "rawMarkdown": "First of all,\n\nThanks to the organizers for this fun and important competition, and congratulations to all the winners. \nThis competition taught me a lot about 3D segmentation, long-term training stability, checkpoint management, and how Public and Private leaderboards can diverge during extended training.\n\n---\n\n## Acknowledgment\n\nSpecial thanks to **@Innat** for sharing the TFRecord dataset and the baseline training code. \nThat baseline helped me start quickly and focus on improving training stability and long-run optimization.\n\n---\n\n## Overview of My Approach\n\nI trained a 3D TransUNet model for volumetric segmentation and continuously resumed training from saved checkpoints. Instead of restarting training multiple times, I extended the same model progressively up to 1325 epochs.\n\nThe main idea was:\n\n- Train for a long duration \n- Save checkpoints regularly \n- Resume training from the last best state \n- Monitor LB and PB trends \n\n---\n\n## Model Architecture\n\nI used TransUNet with a SEResNeXt50 encoder.\n```python\n TransUNet(\n input_shape=(256,256,256,1),\n encoder_name=\"seresnext50\",\n classifier_activation=None,\n num_classes=3,\n dropout=0.1,\n )\n```\n\n### Configuration Highlights\n\n- Input size: (256, 256, 256) \n- Batch size: 2 \n- Number of classes: 3 \n- Dropout: 0.1 \n- Weight decay: 1e-5 \n- Optimizer: AdamW \n- Learning rate schedule: Warmup + Cosine decay \n- Loss:\n - Cross Entropy \n - clDice (for topology preservation) \n- Ignored class: 2 \n- Target class for clDice: 1 \n\n---\n\n## Training Strategy – Long Continuous Resume\n\nI trained in multiple phases and resumed from saved checkpoints:\n\n| Epoch Range | LB Score | PB Score |\n|-------------|----------|----------|\n| 0 → 237 | 0.546 | 0.560 |\n| 237 → 431 | 0.553 | 0.578 |\n| 431 → 703 | 0.549 | 0.582 |\n| 703 → 940 | 0.569 | 0.571 |\n| 940 → 1325 | 0.536 | 0.586 |\n\n\n---\n\n## Code\n\n[Training code](https://www.kaggle.com/code/engadamalmohammedi/vcsd3d25-train): \nhttps://www.kaggle.com/code/engadamalmohammedi/vcsd3d25-train \n\n[Inference code](https://www.kaggle.com/code/engadamalmohammedi/vs3dd-mixf16mgpu): \n---\nThank you again to the organizers and the community for this great experience.", "votes": null }, { "id": "3415411", "postDate": "02/28/2026 21:48:25", "content": "thanks for the writeup/code and good work!
\nthe drop in lb seems suspicious ....\n940 → 1325 0.536 (lb)\nCongrats. I am confused, you say \"PB 0.586\", what does this refer to? Your private LB is 0.572 and your public LB is 0.569.
", "rawMarkdown": "Congrats. I am confused, you say \"PB 0.586\", what does this refer to? Your private LB is 0.572 and your public LB is 0.569.", "votes": null }, { "id": "3415434", "postDate": "02/28/2026 23:04:01", "content": "yes i was not select that private LB 0.586 as shown in the image.\ni was select other with public LB is 0.569 and private LB is 0.572 😟
", "rawMarkdown": "yes i was not select that private LB 0.586 as shown in the image.\ni was select other with public LB is 0.569 and private LB is 0.572 😟", "votes": null }, { "id": "3415435", "postDate": "02/28/2026 23:10:05", "content": "Ah, sorry for your drop on private LB. Did you compute local CV on all your models? Did your \"PB 0.586\" achieve a good local CV score?
", "rawMarkdown": "Ah, sorry for your drop on private LB. Did you compute local CV on all your models? Did your \"PB 0.586\" achieve a good local CV score?", "votes": null }, { "id": "3415436", "postDate": "02/28/2026 23:11:10", "content": "That's drop prevent me from choice that submission.
\ni think if i would continue training the private LB will increased.
", "rawMarkdown": "That's drop prevent me from choice that submission.\n\ni think if i would continue training the private LB will increased.", "votes": null }, { "id": "3415438", "postDate": "02/28/2026 23:20:33", "content": "Because only 6 samples used for validate, i can't decide which the best.\nall that models between 0.51-0.526 score.
", "rawMarkdown": "Because only 6 samples used for validate, i can't decide which the best.\nall that models between 0.51-0.526 score.", "votes": null }, { "id": "3415443", "postDate": "02/28/2026 23:26:43", "content": "Yes I agree. I trained many model using only 6 volumes to validate Dice metric. Then I retrained my best models with 75% train data and validated with 25% data on the official metric to determine which models are best. I observed that there was a big difference. So it was hard to select final models if only using validation on 6 volumes.
", "rawMarkdown": "Yes I agree. I trained many model using only 6 volumes to validate Dice metric. Then I retrained my best models with 75% train data and validated with 25% data on the official metric to determine which models are best. I observed that there was a big difference. So it was hard to select final models if only using validation on 6 volumes.", "votes": null } ], "comments": [ { "id": 3415411, "author_name": "hengck23", "author_url": "", "post_date": "02/28/2026 21:48:25", "content": "thanks for the writeup/code and good work!
\nthe drop in lb seems suspicious ....\n940 → 1325 0.536 (lb)\nThat's drop prevent me from choice that submission.
\ni think if i would continue training the private LB will increased.
", "votes": null, "replies": [] } ] }, { "id": 3415422, "author_name": "cdeotte", "author_url": "", "post_date": "02/28/2026 22:01:52", "content": "Congrats. I am confused, you say \"PB 0.586\", what does this refer to? Your private LB is 0.572 and your public LB is 0.569.
", "votes": null, "replies": [ { "id": 3415434, "author_name": "engadamalmohammedi", "author_url": "", "post_date": "02/28/2026 23:04:01", "content": "yes i was not select that private LB 0.586 as shown in the image.\ni was select other with public LB is 0.569 and private LB is 0.572 😟
", "votes": null, "replies": [ { "id": 3415435, "author_name": "cdeotte", "author_url": "", "post_date": "02/28/2026 23:10:05", "content": "Ah, sorry for your drop on private LB. Did you compute local CV on all your models? Did your \"PB 0.586\" achieve a good local CV score?
", "votes": null, "replies": [ { "id": 3415438, "author_name": "engadamalmohammedi", "author_url": "", "post_date": "02/28/2026 23:20:33", "content": "Because only 6 samples used for validate, i can't decide which the best.\nall that models between 0.51-0.526 score.
", "votes": null, "replies": [ { "id": 3415443, "author_name": "cdeotte", "author_url": "", "post_date": "02/28/2026 23:26:43", "content": "Yes I agree. I trained many model using only 6 volumes to validate Dice metric. Then I retrained my best models with 75% train data and validated with 25% data on the official metric to determine which models are best. I observed that there was a big difference. So it was hard to select final models if only using validation on 6 volumes.
", "votes": null, "replies": [] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "3415366": "First of all,\n\nThanks to the organizers for this fun and important competition, and congratulations to all the winners. \nThis competition taught me a lot about 3D segmentation, long-term training stability, checkpoint management, and how Public and Private leaderboards can diverge during extended training.\n\n---\n\n## Acknowledgment\n\nSpecial thanks to **@Innat** for sharing the TFRecord dataset and the baseline training code. \nThat baseline helped me start quickly and focus on improving training stability and long-run optimization.\n\n---\n\n## Overview of My Approach\n\nI trained a 3D TransUNet model for volumetric segmentation and continuously resumed training from saved checkpoints. Instead of restarting training multiple times, I extended the same model progressively up to 1325 epochs.\n\nThe main idea was:\n\n- Train for a long duration \n- Save checkpoints regularly \n- Resume training from the last best state \n- Monitor LB and PB trends \n\n---\n\n## Model Architecture\n\nI used TransUNet with a SEResNeXt50 encoder.\n```python\n TransUNet(\n input_shape=(256,256,256,1),\n encoder_name=\"seresnext50\",\n classifier_activation=None,\n num_classes=3,\n dropout=0.1,\n )\n```\n\n### Configuration Highlights\n\n- Input size: (256, 256, 256) \n- Batch size: 2 \n- Number of classes: 3 \n- Dropout: 0.1 \n- Weight decay: 1e-5 \n- Optimizer: AdamW \n- Learning rate schedule: Warmup + Cosine decay \n- Loss:\n - Cross Entropy \n - clDice (for topology preservation) \n- Ignored class: 2 \n- Target class for clDice: 1 \n\n---\n\n## Training Strategy – Long Continuous Resume\n\nI trained in multiple phases and resumed from saved checkpoints:\n\n| Epoch Range | LB Score | PB Score |\n|-------------|----------|----------|\n| 0 → 237 | 0.546 | 0.560 |\n| 237 → 431 | 0.553 | 0.578 |\n| 431 → 703 | 0.549 | 0.582 |\n| 703 → 940 | 0.569 | 0.571 |\n| 940 → 1325 | 0.536 | 0.586 |\n\n\n---\n\n## Code\n\n[Training code](https://www.kaggle.com/code/engadamalmohammedi/vcsd3d25-train): \nhttps://www.kaggle.com/code/engadamalmohammedi/vcsd3d25-train \n\n[Inference code](https://www.kaggle.com/code/engadamalmohammedi/vs3dd-mixf16mgpu): \n---\nThank you again to the organizers and the community for this great experience.", "3415411": "thanks for the writeup/code and good work!\n\n```\nthe drop in lb seems suspicious ....\n940 → 1325\t0.536 (lb)\n\n```", "3415422": "Congrats. I am confused, you say \"PB 0.586\", what does this refer to? Your private LB is 0.572 and your public LB is 0.569.", "3415434": "yes i was not select that private LB 0.586 as shown in the image.\ni was select other with public LB is 0.569 and private LB is 0.572 😟", "3415435": "Ah, sorry for your drop on private LB. Did you compute local CV on all your models? Did your \"PB 0.586\" achieve a good local CV score?", "3415436": "That's drop prevent me from choice that submission.\n\ni think if i would continue training the private LB will increased.", "3415438": "Because only 6 samples used for validate, i can't decide which the best.\nall that models between 0.51-0.526 score.", "3415443": "Yes I agree. I trained many model using only 6 volumes to validate Dice metric. Then I retrained my best models with 75% train data and validated with 25% data on the official metric to determine which models are best. I observed that there was a big difference. So it was hard to select final models if only using validation on 6 volumes." }, "source": "meta" }