Hello,
\nif you are familiar with computer vision you probably noticed that the baseline is called \"unet\" -- a classical model for image-to-image problems. The host considers interesting to solve the problem considering not only local information (as an MLP or XGBOOST will do). However, given the current dataset I do not think that is possible to reproduce the 2D grids, am I wrong/correct?
We get the flat values (556 values), but we cannot recover the map unless we know their position in the original low-resolution map. Thus, after spending 5 minutes reading about the data (and knowing the paper), I cannot find a way to solve the problem in 2D with the current train/test dataset. Unless it is save to assume that every 21,600 rows correspond to a map.
\nIf you know how to obtain the 2D input map for the test images, I would I appreciate your help/advice.
", "messages": [ { "id": 2775898, "postDate": "2024-04-25T21:38:59.133Z", "content": "Hello,
\nif you are familiar with computer vision you probably noticed that the baseline is called \"unet\" -- a classical model for image-to-image problems. The host considers interesting to solve the problem considering not only local information (as an MLP or XGBOOST will do). However, given the current dataset I do not think that is possible to reproduce the 2D grids, am I wrong/correct?
We get the flat values (556 values), but we cannot recover the map unless we know their position in the original low-resolution map. Thus, after spending 5 minutes reading about the data (and knowing the paper), I cannot find a way to solve the problem in 2D with the current train/test dataset. Unless it is save to assume that every 21,600 rows correspond to a map.
\nIf you know how to obtain the 2D input map for the test images, I would I appreciate your help/advice.
", "rawMarkdown": "Hello,\nif you are familiar with computer vision you probably noticed that the baseline is called \"unet\" -- a classical model for image-to-image problems. The host considers interesting to solve the problem considering not only local information (as an MLP or XGBOOST will do). However, given the current dataset I do not think that is possible to reproduce the 2D grids, am I wrong/correct?\n\nWe get the flat values (556 values), but we cannot recover the map unless we know their position in the original low-resolution map. Thus, after spending 5 minutes reading about the data (and knowing the paper), I cannot find a way to solve the problem in 2D with the current train/test dataset. Unless it is save to assume that every 21,600 rows correspond to a map.\n\nIf you know how to obtain the 2D input map for the test images, I would I appreciate your help/advice.", "votes": 7 }, { "id": 2777625, "postDate": "2024-04-26T18:32:32.230Z", "content": "Hello, this competition is pursuing a column-local 1D->1D regression approach. The U-net uses different variables as different 1D channels, each with 60 levels. Just be sure not to overweight the scalars when pursuing something similar.
", "rawMarkdown": "Hello, this competition is pursuing a column-local 1D->1D regression approach. The U-net uses different variables as different 1D channels, each with 60 levels. Just be sure not to overweight the scalars when pursuing something similar.", "votes": 5 }, { "id": 2775916, "postDate": "2024-04-25T22:00:24.053Z", "content": "Yes, we cannot reproduce the 2D grids as we don’t have access to the coordinates:
\nhttps://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/495258
Maybe they used a 1D U-Net architecture for variables with 60 associated levels. In this scenario, the spatial dimension of the input for the 1D U-Net would be 60, with 9 channels representing the 9 input variables each with 60 levels.
\nThis is just a guess of course.
thank you @fyangch
", "rawMarkdown": "thank you @fyangch ", "votes": 1 }, { "id": 2776675, "postDate": "2024-04-26T09:20:46.560Z", "content": "I think it is 1d unet too. Although if they confirm it is best.
\nAnother possibility is they turned the 1d data to 2d by reshaping e.g. [60*9] >> [60,9]/[30,27] etc. and then fed it to unet (probably with some resent/effnet/vit backbone) but I think it is less likely.
@jerrylin96 another related question.
\nIs it possible to reproduce the results in your paper (Table comparing CNN, MLP, etc)? Is the kaggle test set the same as in the paper or we would need to download the data and prepare the test set as in your paper?
In other words, is it possible to compare the best solution from this challenge with all your other solutions (besides Unet-1d)?
", "rawMarkdown": "@jerrylin96 another related question.\nIs it possible to reproduce the results in your paper (Table comparing CNN, MLP, etc)? Is the kaggle test set the same as in the paper or we would need to download the data and prepare the test set as in your paper? \n\nIn other words, is it possible to compare the best solution from this challenge with all your other solutions (besides Unet-1d)?\n" } ], "comments": [ { "id": 2777625, "author_name": "Jerry Lin", "author_url": "", "post_date": "2024-04-26T18:32:32.230000", "content": "Hello, this competition is pursuing a column-local 1D->1D regression approach. The U-net uses different variables as different 1D channels, each with 60 levels. Just be sure not to overweight the scalars when pursuing something similar.
", "votes": 5, "replies": [] }, { "id": 2775916, "author_name": "Felix Yang", "author_url": "", "post_date": "2024-04-25T22:00:24.053000", "content": "Yes, we cannot reproduce the 2D grids as we don’t have access to the coordinates:
\nhttps://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/495258
Maybe they used a 1D U-Net architecture for variables with 60 associated levels. In this scenario, the spatial dimension of the input for the 1D U-Net would be 60, with 9 channels representing the 9 input variables each with 60 levels.
\nThis is just a guess of course.
thank you @fyangch
", "votes": 1, "replies": [] }, { "id": 2776675, "author_name": "greySnow", "author_url": "", "post_date": "2024-04-26T09:20:46.560000", "content": "I think it is 1d unet too. Although if they confirm it is best.
\nAnother possibility is they turned the 1d data to 2d by reshaping e.g. [60*9] >> [60,9]/[30,27] etc. and then fed it to unet (probably with some resent/effnet/vit backbone) but I think it is less likely.
@jerrylin96 another related question.
\nIs it possible to reproduce the results in your paper (Table comparing CNN, MLP, etc)? Is the kaggle test set the same as in the paper or we would need to download the data and prepare the test set as in your paper?
In other words, is it possible to compare the best solution from this challenge with all your other solutions (besides Unet-1d)?
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2775898": "Hello,\nif you are familiar with computer vision you probably noticed that the baseline is called \"unet\" -- a classical model for image-to-image problems. The host considers interesting to solve the problem considering not only local information (as an MLP or XGBOOST will do). However, given the current dataset I do not think that is possible to reproduce the 2D grids, am I wrong/correct?\n\nWe get the flat values (556 values), but we cannot recover the map unless we know their position in the original low-resolution map. Thus, after spending 5 minutes reading about the data (and knowing the paper), I cannot find a way to solve the problem in 2D with the current train/test dataset. Unless it is save to assume that every 21,600 rows correspond to a map.\n\nIf you know how to obtain the 2D input map for the test images, I would I appreciate your help/advice.", "2777625": "Hello, this competition is pursuing a column-local 1D->1D regression approach. The U-net uses different variables as different 1D channels, each with 60 levels. Just be sure not to overweight the scalars when pursuing something similar.", "2775916": "Yes, we cannot reproduce the 2D grids as we don’t have access to the coordinates:\nhttps://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/495258\n\nMaybe they used a 1D U-Net architecture for variables with 60 associated levels. In this scenario, the spatial dimension of the input for the 1D U-Net would be 60, with 9 channels representing the 9 input variables each with 60 levels.\nThis is just a guess of course.\n", "2785505": "@jerrylin96 another related question.\nIs it possible to reproduce the results in your paper (Table comparing CNN, MLP, etc)? Is the kaggle test set the same as in the paper or we would need to download the data and prepare the test set as in your paper? \n\nIn other words, is it possible to compare the best solution from this challenge with all your other solutions (besides Unet-1d)?\n" } }