ensemble = 2-fold x 2-model
\n[2-model]
\n[2-fold]
\ntraining code and solution presentation (slide and video):
\nhttps://github.com/hengck23/solution-vesuvius-challenge-ink-detection
the public submission notebook (clean version) is at:
\nhttps://www.kaggle.com/code/hengck23/9th-place-final-ensemble-v1?scriptVersionId=136499959
more solution discussion :
\nhttps://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/407972
it is smiliar to @tanakar 11th place solution solution:
\nhttps://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/417281
ensemble = 2-fold x 2-model
\n[2-model]
\n[2-fold]
\ntraining code and solution presentation (slide and video):
\nhttps://github.com/hengck23/solution-vesuvius-challenge-ink-detection
the public submission notebook (clean version) is at:
\nhttps://www.kaggle.com/code/hengck23/9th-place-final-ensemble-v1?scriptVersionId=136499959
more solution discussion :
\nhttps://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/407972
it is smiliar to @tanakar 11th place solution solution:
\nhttps://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/417281
determine the correct threshold is important in this competition.
\nwe know test fragment (private and public) is smiliar to train fragment 3 from the data description page.
hence i:
\ncheck if optimal thresolds of the two fragements are the same. … which is obvious not the same.
\nthe best CV of whole of fragment 3 is about CV 0.66.
\nSo if you have higher CV for the smaller 10% segment, you have lower CV for the larger 90% segment.
This means that you should not choose the highest public LB threshold.
", "rawMarkdown": "determine the correct threshold is important in this competition.\nwe know test fragment (private and public) is smiliar to train fragment 3 from the data description page.\n\nhence i:\n- make a model with train = fragment1,2. validation = fragment 3\n- split fragment 3 into two (10%/90% area, to simulate public/private data set)\n\ncheck if optimal thresolds of the two fragements are the same. ... which is obvious not the same.\n\n---\n\nthe best CV of whole of fragment 3 is about CV 0.66.\nSo if you have higher CV for the smaller 10% segment, you have lower CV for the larger 90% segment.\n\nThis means that you should not choose the highest public LB threshold.", "votes": null }, { "id": "2303714", "postDate": "06/15/2023 12:29:52", "content": "@hengck23 congratulations for solo gold! I have learned a lot from you. Your discussion topics in many many competitions helped a lot to start and to develop our solutions. Congratulations!!
", "rawMarkdown": "hengck23 congratulations for solo gold! I have learned a lot from you. Your discussion topics in many many competitions helped a lot to start and to develop our solutions. Congratulations!!", "votes": null }, { "id": "2303729", "postDate": "06/15/2023 12:41:45", "content": "ggs, I wonder whats was your baseline 0.71 private score ?
", "rawMarkdown": "ggs, I wonder whats was your baseline 0.71 private score ?", "votes": null }, { "id": "2303752", "postDate": "06/15/2023 13:02:38", "content": "Congratulations @hengck23 on solo gold 🎉 You certainly deserve it !
\nI have learned a lot from all your discussion posts and code shared in many competitions. Thank you 🙏
I performed the exact same simulation and obtained the same conclusion. I agree it!
", "rawMarkdown": "I performed the exact same simulation and obtained the same conclusion. I agree it!", "votes": null }, { "id": "2303814", "postDate": "06/15/2023 13:40:47", "content": "Congratulations @hengck23
\nThank you very much for sharing the discussion topics . I learned a lot from it.
@hengck23 Congrats for solo-gold! And becoming the competition grand master 🎉 You absolutely deserve it !!
", "rawMarkdown": "hengck23 Congrats for solo-gold! And becoming the competition grand master 🎉 You absolutely deserve it !!", "votes": null }, { "id": "2303816", "postDate": "06/15/2023 13:43:12", "content": "@remekkinas
\nThanks for the post.
Good luck and all the best to your work!
\nhttps://www.kaggle.com/competitions/predict-student-performance-from-game-play/leaderboard
Extremely well-deserved solo gold! Congrats for becoming competitions GM, and many thanks for your valuable insights in this comp and in many others over the years. Your posts are always a must-read. Keep up the great work!
", "rawMarkdown": "Extremely well-deserved solo gold! Congrats for becoming competitions GM, and many thanks for your valuable insights in this comp and in many others over the years. Your posts are always a must-read. Keep up the great work!", "votes": null }, { "id": "2304260", "postDate": "06/15/2023 21:15:11", "content": "Congratulations on becoming grandmaster @hengck23 you deserve it! You are grandmaster in sharing too!
", "rawMarkdown": "Congratulations on becoming grandmaster @hengck23 you deserve it! You are grandmaster in sharing too!", "votes": null }, { "id": "2304351", "postDate": "06/16/2023 00:10:58", "content": "Congrats on the solo gold and becoming competition GM. You are certainly one of the most deserving of that title in the kaggle community.
", "rawMarkdown": "Congrats on the solo gold and becoming competition GM. You are certainly one of the most deserving of that title in the kaggle community.", "votes": null }, { "id": "2304360", "postDate": "06/16/2023 00:14:35", "content": "Congratulations competition GM 🎉🎉🎉
", "rawMarkdown": "Congratulations competition GM 🎉🎉🎉", "votes": null }, { "id": "2304371", "postDate": "06/16/2023 00:26:10", "content": "@hengck23 Congratulation! and thanks a lot for sharing your work.
\nI(& other people) learned a lot from your notebook and discussion.
If you don't mind, can you share your training code?
\nBecause I tried to reproduce your result(this one),
\nbut I can't just because of my poor experience..
Congrats, we were studying your notebooks trying to understand why they scored as well as they did and learned a lot from what you shared.
", "rawMarkdown": "Congrats, we were studying your notebooks trying to understand why they scored as well as they did and learned a lot from what you shared.", "votes": null }, { "id": "2304779", "postDate": "06/16/2023 08:27:51", "content": "\n\nwe know test fragment (private and public) is smiliar to train fragment 3 from the data description page.
\n
I can't seem to find this, where exactly did you see this?
\nAlso, if indeed the test fragment is similar to fragment 3 in the training data, wouldn't it make sense to use valid_id = 3 instead of valid_id = 1 for training your models:
\n\n", "rawMarkdown": "> we know test fragment (private and public) is smiliar to train fragment 3 from the data description page.\n\nI can't seem to find this, where exactly did you see this?\n\n------------\n\nAlso, if indeed the test fragment is similar to fragment 3 in the training data, wouldn't it make sense to use valid_id = 3 instead of valid_id = 1 for training your models:\n\n> [2-fold]\n> validation fragment-1 and fragment-2aa", "votes": null }, { "id": "2304957", "postDate": "06/16/2023 10:50:33", "content": "[2-fold]
\n
\n validation fragment-1 and fragment-2aa
Congrats on solo gold and becoming Competitions GM Heng!
", "rawMarkdown": "Congrats on solo gold and becoming Competitions GM Heng!", "votes": null }, { "id": "2310181", "postDate": "06/20/2023 07:11:03", "content": "Thank you for all the notebooks and thoughtful information you shared in this competition. They were very beneficial. Congratulations on GM -- well deserved.
", "rawMarkdown": "Thank you for all the notebooks and thoughtful information you shared in this competition. They were very beneficial. Congratulations on GM -- well deserved.", "votes": null }, { "id": "2353895", "postDate": "07/22/2023 05:37:43", "content": "@hengck23
\nthanks for sharing training code, I'll learn more :)
determine the correct threshold is important in this competition.
\nwe know test fragment (private and public) is smiliar to train fragment 3 from the data description page.
hence i:
\ncheck if optimal thresolds of the two fragements are the same. … which is obvious not the same.
\nthe best CV of whole of fragment 3 is about CV 0.66.
\nSo if you have higher CV for the smaller 10% segment, you have lower CV for the larger 90% segment.
This means that you should not choose the highest public LB threshold.
", "votes": null, "replies": [ { "id": 2303808, "author_name": "chumajin", "author_url": "", "post_date": "06/15/2023 13:37:38", "content": "I performed the exact same simulation and obtained the same conclusion. I agree it!
", "votes": null, "replies": [] }, { "id": 2304779, "author_name": "lucasvw", "author_url": "", "post_date": "06/16/2023 08:27:51", "content": "\n\nwe know test fragment (private and public) is smiliar to train fragment 3 from the data description page.
\n
I can't seem to find this, where exactly did you see this?
\nAlso, if indeed the test fragment is similar to fragment 3 in the training data, wouldn't it make sense to use valid_id = 3 instead of valid_id = 1 for training your models:
\n\n", "votes": null, "replies": [] } ] }, { "id": 2303714, "author_name": "remekkinas", "author_url": "", "post_date": "06/15/2023 12:29:52", "content": "[2-fold]
\n
\n validation fragment-1 and fragment-2aa
@hengck23 congratulations for solo gold! I have learned a lot from you. Your discussion topics in many many competitions helped a lot to start and to develop our solutions. Congratulations!!
", "votes": null, "replies": [ { "id": 2303816, "author_name": "hengck23", "author_url": "", "post_date": "06/15/2023 13:43:12", "content": "@remekkinas
\nThanks for the post.
Good luck and all the best to your work!
\nhttps://www.kaggle.com/competitions/predict-student-performance-from-game-play/leaderboard
ggs, I wonder whats was your baseline 0.71 private score ?
", "votes": null, "replies": [] }, { "id": 2303752, "author_name": "rashmibanthia", "author_url": "", "post_date": "06/15/2023 13:02:38", "content": "Congratulations @hengck23 on solo gold 🎉 You certainly deserve it !
\nI have learned a lot from all your discussion posts and code shared in many competitions. Thank you 🙏
Congratulations @hengck23
\nThank you very much for sharing the discussion topics . I learned a lot from it.
@hengck23 Congrats for solo-gold! And becoming the competition grand master 🎉 You absolutely deserve it !!
", "votes": null, "replies": [] }, { "id": 2304041, "author_name": "headsortails", "author_url": "", "post_date": "06/15/2023 16:27:49", "content": "Extremely well-deserved solo gold! Congrats for becoming competitions GM, and many thanks for your valuable insights in this comp and in many others over the years. Your posts are always a must-read. Keep up the great work!
", "votes": null, "replies": [] }, { "id": 2304260, "author_name": "alejopaullier", "author_url": "", "post_date": "06/15/2023 21:15:11", "content": "Congratulations on becoming grandmaster @hengck23 you deserve it! You are grandmaster in sharing too!
", "votes": null, "replies": [] }, { "id": 2304351, "author_name": "nymfree", "author_url": "", "post_date": "06/16/2023 00:10:58", "content": "Congrats on the solo gold and becoming competition GM. You are certainly one of the most deserving of that title in the kaggle community.
", "votes": null, "replies": [] }, { "id": 2304360, "author_name": "robsonsan", "author_url": "", "post_date": "06/16/2023 00:14:35", "content": "Congratulations competition GM 🎉🎉🎉
", "votes": null, "replies": [] }, { "id": 2304371, "author_name": "jumdpwl", "author_url": "", "post_date": "06/16/2023 00:26:10", "content": "@hengck23 Congratulation! and thanks a lot for sharing your work.
\nI(& other people) learned a lot from your notebook and discussion.
If you don't mind, can you share your training code?
\nBecause I tried to reproduce your result(this one),
\nbut I can't just because of my poor experience..
@hengck23
\nthanks for sharing training code, I'll learn more :)
Congrats, we were studying your notebooks trying to understand why they scored as well as they did and learned a lot from what you shared.
", "votes": null, "replies": [] }, { "id": 2304957, "author_name": "returnofsputnik", "author_url": "", "post_date": "06/16/2023 10:50:33", "content": "Congrats on solo gold and becoming Competitions GM Heng!
", "votes": null, "replies": [] }, { "id": 2310181, "author_name": "socated", "author_url": "", "post_date": "06/20/2023 07:11:03", "content": "Thank you for all the notebooks and thoughtful information you shared in this competition. They were very beneficial. Congratulations on GM -- well deserved.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2303652": "ensemble = 2-fold x 2-model\n\n[2-model]\n- resnet34d-unet (crop=256. depth=32 slices)\n- pvtv2-b3-daformer (crop=384. depth=16 slices)\n\n[2-fold]\n- validation fragment-1 and fragment-2aa \n\n---\n\ntraining code and solution presentation (slide and video):\nhttps://github.com/hengck23/solution-vesuvius-challenge-ink-detection\n\n\nthe public submission notebook (clean version) is at:\nhttps://www.kaggle.com/code/hengck23/9th-place-final-ensemble-v1?scriptVersionId=136499959\n\n---\n\nmore solution discussion : \nhttps://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/407972\n\nit is smiliar to @tanakar 11th place solution solution:\nhttps://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/417281", "2303700": "determine the correct threshold is important in this competition.\nwe know test fragment (private and public) is smiliar to train fragment 3 from the data description page.\n\nhence i:\n- make a model with train = fragment1,2. validation = fragment 3\n- split fragment 3 into two (10%/90% area, to simulate public/private data set)\n\ncheck if optimal thresolds of the two fragements are the same. ... which is obvious not the same.\n\n---\n\nthe best CV of whole of fragment 3 is about CV 0.66.\nSo if you have higher CV for the smaller 10% segment, you have lower CV for the larger 90% segment.\n\nThis means that you should not choose the highest public LB threshold.", "2303714": "hengck23 congratulations for solo gold! I have learned a lot from you. Your discussion topics in many many competitions helped a lot to start and to develop our solutions. Congratulations!!", "2303729": "ggs, I wonder whats was your baseline 0.71 private score ?", "2303752": "Congratulations @hengck23 on solo gold 🎉 You certainly deserve it ! \nI have learned a lot from all your discussion posts and code shared in many competitions. Thank you 🙏", "2303808": "I performed the exact same simulation and obtained the same conclusion. I agree it!", "2303814": "Congratulations @hengck23 \nThank you very much for sharing the discussion topics . I learned a lot from it.", "2303815": "hengck23 Congrats for solo-gold! And becoming the competition grand master 🎉 You absolutely deserve it !!", "2303816": "remekkinas \nThanks for the post.\n\nGood luck and all the best to your work!\nhttps://www.kaggle.com/competitions/predict-student-performance-from-game-play/leaderboard", "2304041": "Extremely well-deserved solo gold! Congrats for becoming competitions GM, and many thanks for your valuable insights in this comp and in many others over the years. Your posts are always a must-read. Keep up the great work!", "2304260": "Congratulations on becoming grandmaster @hengck23 you deserve it! You are grandmaster in sharing too!", "2304351": "Congrats on the solo gold and becoming competition GM. You are certainly one of the most deserving of that title in the kaggle community.", "2304360": "Congratulations competition GM 🎉🎉🎉", "2304371": "hengck23 Congratulation! and thanks a lot for sharing your work.\nI(& other people) learned a lot from your notebook and discussion.\n \nIf you don't mind, can you share your training code? \nBecause I tried to reproduce your result([this one](https://www.kaggle.com/code/hengck23/lb0-68-one-fold-stacked-unet)),\nbut I can't just because of my poor experience..", "2304377": "Congrats, we were studying your notebooks trying to understand why they scored as well as they did and learned a lot from what you shared.", "2304779": "> we know test fragment (private and public) is smiliar to train fragment 3 from the data description page.\n\nI can't seem to find this, where exactly did you see this?\n\n------------\n\nAlso, if indeed the test fragment is similar to fragment 3 in the training data, wouldn't it make sense to use valid_id = 3 instead of valid_id = 1 for training your models:\n\n> [2-fold]\n> validation fragment-1 and fragment-2aa", "2304957": "Congrats on solo gold and becoming Competitions GM Heng!", "2310181": "Thank you for all the notebooks and thoughtful information you shared in this competition. They were very beneficial. Congratulations on GM -- well deserved.", "2353895": "hengck23 \nthanks for sharing training code, I'll learn more :)" }, "source": "meta" }