![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9117256%2Ff4943226494a8c3812830fb923901bd4%2Foutput.png?generation=1737582820260504&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9117256%2Fdf05c3111a46cb47b660debb87e074b4%2Fnewplot1.png?generation=1737582919064444&alt=media\")
I tried training various 3d Unet based models but found that the DICE score for individual classes and traversky loss almost stayed same after many training epochs, with DICE score very close to 0 for all classes.
\nThen I tried to visualize the tomograms of patchsize (128) to see what is consumed by models, I found that there is no clear structure in the tomograms which the models can learn effectively…
\nThere are 2 awkward cases :
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9117256%2F20a550c1767c1a3269c1f888c92e2eb0%2Fnewplot.png?generation=1737582951366234&alt=media\")
Notice ~ in title of 3rd image.
Observation :
\nI have experimented with tomo-algo : denoised and scale : 0, without any preprocessing.
Here I attach my training report for detailed review. Pls help me understand the scenario and how can I tackle it…
\nI see people claiming to easily achive LB ~ 0.707 …. where am I getting wrong??
I tried training various 3d Unet based models but found that the DICE score for individual classes and traversky loss almost stayed same after many training epochs, with DICE score very close to 0 for all classes.
\nThen I tried to visualize the tomograms of patchsize (128) to see what is consumed by models, I found that there is no clear structure in the tomograms which the models can learn effectively…
\nThere are 2 awkward cases :
\n
Notice ~ in title of 3rd image.
Observation :
\nI have experimented with tomo-algo : denoised and scale : 0, without any preprocessing.
Here I attach my training report for detailed review. Pls help me understand the scenario and how can I tackle it…
\nI see people claiming to easily achive LB ~ 0.707 …. where am I getting wrong??
This notebook is a great starting point for 3D Unet. You might start there, and then try some of the higher scoring notebooks as well:
\nhttps://www.kaggle.com/code/fnands/baseline-unet-train-submit
I would take a look through the notebooks on the code tab.
", "rawMarkdown": "This notebook is a great starting point for 3D Unet. You might start there, and then try some of the higher scoring notebooks as well:\n[https://www.kaggle.com/code/fnands/baseline-unet-train-submit](https://www.kaggle.com/code/fnands/baseline-unet-train-submit)\n\nI would take a look through the notebooks on the code tab.", "votes": null }, { "id": "3109268", "postDate": "01/28/2025 17:00:54", "content": "Also keep in mind that the LB score of 0.707 is not necessarily easy but there is a public notebook that scored 0.707 that others have copied and submitted.
", "rawMarkdown": "Also keep in mind that the LB score of 0.707 is not necessarily easy but there is a public notebook that scored 0.707 that others have copied and submitted.", "votes": null } ], "comments": [ { "id": 3103165, "author_name": "davidlist", "author_url": "", "post_date": "01/23/2025 05:51:06", "content": "This notebook is a great starting point for 3D Unet. You might start there, and then try some of the higher scoring notebooks as well:
\nhttps://www.kaggle.com/code/fnands/baseline-unet-train-submit
I would take a look through the notebooks on the code tab.
", "votes": null, "replies": [] }, { "id": 3109268, "author_name": "connorjd", "author_url": "", "post_date": "01/28/2025 17:00:54", "content": "Also keep in mind that the LB score of 0.707 is not necessarily easy but there is a public notebook that scored 0.707 that others have copied and submitted.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3103006": "I tried training various 3d Unet based models but found that the DICE score for individual classes and traversky loss almost stayed same after many training epochs, with DICE score very close to 0 for all classes.\n\nThen I tried to visualize the tomograms of patchsize (128) to see what is consumed by models, I found that there is no clear structure in the tomograms which the models can learn effectively...\n\nThere are 2 awkward cases : \n- When I plot pixel histogram , there is very small range of pixel values with very high frequency. \n- When I plot the tomogram in 3d, allowing sampling of voxels from pixels within that high frequency range. \n\n\n\nNotice ```~``` in title of 3rd image.\n\nObservation : \n- The sampled pixels cover entire tomogram excluding the region surrounding the sphere (in red) (as segmentation ground truth for training)... the sizes of sphere correspond to various classes.\n\nI have experimented with tomo-algo : ```denoised``` and scale : ```0```, without any preprocessing.\n\nHere I attach my [training report](https://api.wandb.ai/links/shri_krishna/420dlwdc) for detailed review. Pls help me understand the scenario and how can I tackle it...\n\nI see people claiming to easily achive LB ~ ```0.707``` .... where am I getting wrong??", "3103165": "This notebook is a great starting point for 3D Unet. You might start there, and then try some of the higher scoring notebooks as well:\n[https://www.kaggle.com/code/fnands/baseline-unet-train-submit](https://www.kaggle.com/code/fnands/baseline-unet-train-submit)\n\nI would take a look through the notebooks on the code tab.", "3109268": "Also keep in mind that the LB score of 0.707 is not necessarily easy but there is a public notebook that scored 0.707 that others have copied and submitted." }, "source": "meta" }