i am collecting questions that you might have on the results e.g. model training, reasons of failure, what is working etc …
\ni will use your feedback to create a solution (and answer your questions using experimental results) and particpate the solution prize.
\ni need your vote later if you think my solution writeup gives good answers and insights.
\nthanks!
to start with, this is what i have in mind:
\n1) https://www.kaggle.com/competitions/google-research-identify-contrails-reduce-global-warming/discussion/430479#2382728
\n@tascj0 mentions about truth mask alignment issue. is this the reasons why pesudo label fails? (without correct augmentation, we cannot learn good model and cannot generate good predictions. erros propagates)
i am collecting questions that you might have on the results e.g. model training, reasons of failure, what is working etc …
\ni will use your feedback to create a solution (and answer your questions using experimental results) and particpate the solution prize.
\ni need your vote later if you think my solution writeup gives good answers and insights.
\nthanks!
to start with, this is what i have in mind:
\n1) https://www.kaggle.com/competitions/google-research-identify-contrails-reduce-global-warming/discussion/430479#2382728
\n@tascj0 mentions about truth mask alignment issue. is this the reasons why pesudo label fails? (without correct augmentation, we cannot learn good model and cannot generate good predictions. erros propagates)
I would like to know how one would go about making use of available temporal information in general. We tried 3d unets, 2.5d approach, and some sequence models, but couldn't get them to work, while some top teams did.
\nHow should one be thinking about incorporating temporal information into their models?
", "rawMarkdown": "I would like to know how one would go about making use of available temporal information in general. We tried 3d unets, 2.5d approach, and some sequence models, but couldn't get them to work, while some top teams did.\n\nHow should one be thinking about incorporating temporal information into their models?", "votes": 1 }, { "id": 2383685, "postDate": "2023-08-10T14:11:54.450Z", "content": "The 0.5-pixel shift might be a good one to look into? 1st place solution
", "rawMarkdown": "The 0.5-pixel shift might be a good one to look into? [1st place solution](https://www.kaggle.com/competitions/google-research-identify-contrails-reduce-global-warming/discussion/430618)", "votes": 1 }, { "id": 2383066, "postDate": "2023-08-10T06:56:22.323Z", "content": "Great @hengck23 that you opened up this channel!
\nI would also love to know more about why specific solutions worked! If you have enough time I would really enjoy to hear your thoughts on how the validation strategy has influenced the top solutions.
\nIn the past weeks we found a potential source of leakage when building cross validation folds as there are many duplicates but also images with very high similarity:
\nhttps://www.kaggle.com/code/allunia/potential-leakage-between-folds-and-test?scriptVersionId=139486256
\nhttps://www.kaggle.com/code/allunia/building-balanced-cv-sets
A lot of people in the discussion forum also mentioned that CV and LB correlation was difficult. We have spend some effort on building CV sets that avoid this leakage. Do you think that one could improve top solutions even further by using leak-free folds or by trying to find out which kind of images can be found in test?
\nThank you! :-)
", "rawMarkdown": "Great @hengck23 that you opened up this channel!\n\nI would also love to know more about why specific solutions worked! If you have enough time I would really enjoy to hear your thoughts on how the validation strategy has influenced the top solutions. \n\nIn the past weeks we found a potential source of leakage when building cross validation folds as there are many duplicates but also images with very high similarity:\n\nhttps://www.kaggle.com/code/allunia/potential-leakage-between-folds-and-test?scriptVersionId=139486256\nhttps://www.kaggle.com/code/allunia/building-balanced-cv-sets\n\nA lot of people in the discussion forum also mentioned that CV and LB correlation was difficult. We have spend some effort on building CV sets that avoid this leakage. Do you think that one could improve top solutions even further by using leak-free folds or by trying to find out which kind of images can be found in test?\n\nThank you! :-)\n", "votes": 2 }, { "id": 2382967, "postDate": "2023-08-10T06:02:09.263Z", "content": "Thanks @hengck23 , this is the 1st competition in that I even can't understand the winners' solutions.
", "rawMarkdown": "Thanks @hengck23 , this is the 1st competition in that I even can't understand the winners' solutions. " }, { "id": 2382744, "postDate": "2023-08-10T02:21:55.737Z", "content": "using shiftscalerotate can improve the cv.
", "rawMarkdown": "using shiftscalerotate can improve the cv." }, { "id": 2382738, "postDate": "2023-08-10T02:05:49.163Z", "content": "note that it is possible to write solution that includes additional experiments and post submission:
\nhttps://www.kaggle.com/discussions/general/427114#2362386
I would like to know how one would go about making use of available temporal information in general. We tried 3d unets, 2.5d approach, and some sequence models, but couldn't get them to work, while some top teams did.
\nHow should one be thinking about incorporating temporal information into their models?
", "votes": 1, "replies": [] }, { "id": 2383685, "author_name": "Bartley", "author_url": "", "post_date": "2023-08-10T14:11:54.450000", "content": "The 0.5-pixel shift might be a good one to look into? 1st place solution
", "votes": 1, "replies": [] }, { "id": 2383066, "author_name": "Laura Fink", "author_url": "", "post_date": "2023-08-10T06:56:22.323000", "content": "Great @hengck23 that you opened up this channel!
\nI would also love to know more about why specific solutions worked! If you have enough time I would really enjoy to hear your thoughts on how the validation strategy has influenced the top solutions.
\nIn the past weeks we found a potential source of leakage when building cross validation folds as there are many duplicates but also images with very high similarity:
\nhttps://www.kaggle.com/code/allunia/potential-leakage-between-folds-and-test?scriptVersionId=139486256
\nhttps://www.kaggle.com/code/allunia/building-balanced-cv-sets
A lot of people in the discussion forum also mentioned that CV and LB correlation was difficult. We have spend some effort on building CV sets that avoid this leakage. Do you think that one could improve top solutions even further by using leak-free folds or by trying to find out which kind of images can be found in test?
\nThank you! :-)
", "votes": 2, "replies": [] }, { "id": 2382967, "author_name": "william.wu", "author_url": "", "post_date": "2023-08-10T06:02:09.263000", "content": "Thanks @hengck23 , this is the 1st competition in that I even can't understand the winners' solutions.
", "votes": 0, "replies": [] }, { "id": 2382744, "author_name": "pky", "author_url": "", "post_date": "2023-08-10T02:21:55.737000", "content": "using shiftscalerotate can improve the cv.
", "votes": 0, "replies": [] }, { "id": 2382738, "author_name": "hengck23", "author_url": "", "post_date": "2023-08-10T02:05:49.163000", "content": "note that it is possible to write solution that includes additional experiments and post submission:
\nhttps://www.kaggle.com/discussions/general/427114#2362386