https://www.arxiv.org/pdf/2506.15346
\ncheck this out. it was released a couple days ago
\nthe math, model, code and everything else looks super advanced.
\nmaybe someone could implement and train it in this last 1 week.
Your efficiency is incredible, I even struggle just to read through the public notebooks.๐
", "rawMarkdown": "Your efficiency is incredible, I even struggle just to read through the public notebooks.๐\n", "votes": 5, "replies": [ { "id": 3230648, "postDate": "2025-06-23T09:31:55.227Z", "content": "Hehe
\nthnx mate ;)
https://www.arxiv.org/pdf/2506.15346
\ncheck this out. it was released a couple days ago
\nthe math, model, code and everything else looks super advanced.
\nmaybe someone could implement and train it in this last 1 week.
The numbers do not make sense. I did not read carefully, but I'll throw a wild guess that they normalized the VELs by 1000. Then, if we multiply their numbers by 1000, it makes sense, and the numbers are good but falling short of top solutions here.
", "rawMarkdown": "The numbers do not make sense. I did not read carefully, but I'll throw a wild guess that they normalized the VELs by 1000. Then, if we multiply their numbers by 1000, it makes sense, and the numbers are good but falling short of top solutions here.", "votes": 1, "replies": [ { "id": 3230052, "postDate": "2025-06-22T12:58:31.590Z", "content": "yes at first glance they dont make sense.
\ni went through the code a little bit. they did log transforms, min max, resizing and hell a lot of other things.
\nits a diffusion model based approach so i thought if someone could make use of it.
Usually when I read such paper, I read about relative performance. Eg if diffusion is better than just cnn. Kaggle results are much better here because of different data split and size.
\nThe paper are splitting public Openfwi for train and validation. Here we are using all Openfwi. For train ( and maybe other kagglers data etc) and additional hidden test data.
\nAfter said that, I still think kagglers solution are better engineered, eg better network with better backbone tweaks and infinite training loop. Most paper architecture are usually simple cnn and training limited by loss gain efficiency. But I am interested to know if we tweak the diffusion model, will it perform better than cnn. That will take a lot of work. In short donโt expect copy a solution and apply as it is to kaggle, it will not work without engineering and tweaking.
", "rawMarkdown": "Usually when I read such paper, I read about relative performance. Eg if diffusion is better than just cnn. Kaggle results are much better here because of different data split and size.\n\n\nThe paper are splitting public Openfwi for train and validation. Here we are using all Openfwi. For train ( and maybe other kagglers data etc) and additional hidden test data.\n\nAfter said that, I still think kagglers solution are better engineered, eg better network with better backbone tweaks and infinite training loop. Most paper architecture are usually simple cnn and training limited by loss gain efficiency. But I am interested to know if we tweak the diffusion model, will it perform better than cnn. That will take a lot of work. In short donโt expect copy a solution and apply as it is to kaggle, it will not work without engineering and tweaking.", "votes": 2, "replies": [ { "id": 3230541, "postDate": "2025-06-23T07:18:42.223Z", "rawMarkdown": "", "isDeleted": true } ] } ] } ], "comments": [ { "id": 3230632, "author_name": "Gold", "author_url": "", "post_date": "2025-06-23T09:18:07.670000", "content": "Your efficiency is incredible, I even struggle just to read through the public notebooks.๐
", "votes": 5, "replies": [ { "id": 3230648, "author_name": "", "author_url": "", "post_date": "2025-06-23T09:31:55.227000", "content": "Hehe
\nthnx mate ;)
The numbers do not make sense. I did not read carefully, but I'll throw a wild guess that they normalized the VELs by 1000. Then, if we multiply their numbers by 1000, it makes sense, and the numbers are good but falling short of top solutions here.
", "votes": 1, "replies": [ { "id": 3230052, "author_name": "", "author_url": "", "post_date": "2025-06-22T12:58:31.590000", "content": "yes at first glance they dont make sense.
\ni went through the code a little bit. they did log transforms, min max, resizing and hell a lot of other things.
\nits a diffusion model based approach so i thought if someone could make use of it.
Usually when I read such paper, I read about relative performance. Eg if diffusion is better than just cnn. Kaggle results are much better here because of different data split and size.
\nThe paper are splitting public Openfwi for train and validation. Here we are using all Openfwi. For train ( and maybe other kagglers data etc) and additional hidden test data.
\nAfter said that, I still think kagglers solution are better engineered, eg better network with better backbone tweaks and infinite training loop. Most paper architecture are usually simple cnn and training limited by loss gain efficiency. But I am interested to know if we tweak the diffusion model, will it perform better than cnn. That will take a lot of work. In short donโt expect copy a solution and apply as it is to kaggle, it will not work without engineering and tweaking.
", "votes": 2, "replies": [ { "id": 3230541, "author_name": "", "author_url": "", "post_date": "2025-06-23T07:18:42.223000", "content": "", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3230632": "Your efficiency is incredible, I even struggle just to read through the public notebooks.๐\n", "3230034": "https://www.arxiv.org/pdf/2506.15346\n\ncheck this out. it was released a couple days ago\nthe math, model, code and everything else looks super advanced. \nmaybe someone could implement and train it in this last 1 week. ", "3230047": "The numbers do not make sense. I did not read carefully, but I'll throw a wild guess that they normalized the VELs by 1000. Then, if we multiply their numbers by 1000, it makes sense, and the numbers are good but falling short of top solutions here." } }