First of all, thanks to the organizers of the competition. My team used the competition data for our Machine Learning course project, and had a good time with it.
\nWe have a clean training code for Unet model and SwinTransformer model here: Ori-Replication/kaggle-human-vasculature (github.com)
\nIt have a detailed README and you can easily train your own model following the guide.
\nIt also have inference, score counting, and visulize script.
The top model in our submission (not the one we finally selected) is the one we trained with 3D-rotated data.
\nWe rotated the data on 2 or 3 axis with about 15 degrees, and used all the data for training.
\nThe rotate was implemented on kidney_1_dense, kidney_1_voi, and kidney_3_sparse.
It score 0.492 on private. However, on public it only scored 0.771 so I lose hope on it and didn't seleted it. It is about 0.04 higher than the previous baseline on private, but 0.074 lower than the previous baseline on public. Since I was busy these days and didn't implemented a local cross validation. So do not trust the public LB on kaggle. The baseline have a huge overfitting in this competition.
\nThe solution finally selcted only changed the threshold. Because we bet that the positive samples should be more than the public one. The Threshold was 0.152 in selected notebook.
", "messages": [ { "id": 2640488, "postDate": "2024-02-07T00:43:40.800Z", "content": "First of all, thanks to the organizers of the competition. My team used the competition data for our Machine Learning course project, and had a good time with it.
\nWe have a clean training code for Unet model and SwinTransformer model here: Ori-Replication/kaggle-human-vasculature (github.com)
\nIt have a detailed README and you can easily train your own model following the guide.
\nIt also have inference, score counting, and visulize script.
The top model in our submission (not the one we finally selected) is the one we trained with 3D-rotated data.
\nWe rotated the data on 2 or 3 axis with about 15 degrees, and used all the data for training.
\nThe rotate was implemented on kidney_1_dense, kidney_1_voi, and kidney_3_sparse.
It score 0.492 on private. However, on public it only scored 0.771 so I lose hope on it and didn't seleted it. It is about 0.04 higher than the previous baseline on private, but 0.074 lower than the previous baseline on public. Since I was busy these days and didn't implemented a local cross validation. So do not trust the public LB on kaggle. The baseline have a huge overfitting in this competition.
\nThe solution finally selcted only changed the threshold. Because we bet that the positive samples should be more than the public one. The Threshold was 0.152 in selected notebook.
", "rawMarkdown": "## TOP 17% Solution: 3D rotate augmentation works\nFirst of all, thanks to the organizers of the competition. My team used the competition data for our Machine Learning course project, and had a good time with it. \n\n## Code\nWe have a clean training code for Unet model and SwinTransformer model here: [Ori-Replication/kaggle-human-vasculature (github.com)](https://github.com/Ori-Replication/kaggle-human-vasculature/tree/master)\nIt have a detailed README and you can easily train your own model following the guide.\nIt also have inference, score counting, and visulize script.\n\n## Solution\nThe top model in our submission (not the one we finally selected) is the one we trained with 3D-rotated data.\n\nWe rotated the data on 2 or 3 axis with about 15 degrees, and used all the data for training.\nThe rotate was implemented on kidney_1_dense, kidney_1_voi, and kidney_3_sparse.\n\nIt score 0.492 on private. However, on public it only scored 0.771 so I lose hope on it and didn't seleted it. It is about 0.04 higher than the previous baseline on private, but 0.074 lower than the previous baseline on public. Since I was busy these days and didn't implemented a local cross validation. So do not trust the public LB on kaggle. The baseline have a huge overfitting in this competition.\n\nThe solution finally selcted only changed the threshold. Because we bet that the positive samples should be more than the public one. The Threshold was 0.152 in selected notebook.\n\n", "votes": 1 } ], "comments": [], "raw_markdown_by_id": { "2640488": "## TOP 17% Solution: 3D rotate augmentation works\nFirst of all, thanks to the organizers of the competition. My team used the competition data for our Machine Learning course project, and had a good time with it. \n\n## Code\nWe have a clean training code for Unet model and SwinTransformer model here: [Ori-Replication/kaggle-human-vasculature (github.com)](https://github.com/Ori-Replication/kaggle-human-vasculature/tree/master)\nIt have a detailed README and you can easily train your own model following the guide.\nIt also have inference, score counting, and visulize script.\n\n## Solution\nThe top model in our submission (not the one we finally selected) is the one we trained with 3D-rotated data.\n\nWe rotated the data on 2 or 3 axis with about 15 degrees, and used all the data for training.\nThe rotate was implemented on kidney_1_dense, kidney_1_voi, and kidney_3_sparse.\n\nIt score 0.492 on private. However, on public it only scored 0.771 so I lose hope on it and didn't seleted it. It is about 0.04 higher than the previous baseline on private, but 0.074 lower than the previous baseline on public. Since I was busy these days and didn't implemented a local cross validation. So do not trust the public LB on kaggle. The baseline have a huge overfitting in this competition.\n\nThe solution finally selcted only changed the threshold. Because we bet that the positive samples should be more than the public one. The Threshold was 0.152 in selected notebook.\n\n" } }