I don't mean to be vulgar but I just passed through the ridiculous bubble - 377 places with one submission. Now I can start having fun again.\n I don't mean to be vulgar but I just passed through the ridiculous bubble - 377 places with one submission. Now I can start having fun again.\nExactly. When our highest scoring subs are lower than that bubble zone, we were feaked out! Maybe people should not post these high scoring kernels. But anyways, it made us work so hard that actually gave us a gigantic boost! Its a bad thing and also maybe a good thing.
", "rawMarkdown": "Exactly. When our highest scoring subs are lower than that bubble zone, we were feaked out! Maybe people should not post these high scoring kernels. But anyways, it made us work so hard that actually gave us a gigantic boost! Its a bad thing and also maybe a good thing.", "votes": null }, { "id": "742136", "postDate": "02/11/2020 05:02:38", "content": "I am still technically below this stupid resnet kernel without bells and whistles... So humiliating :)
", "rawMarkdown": "I am still technically below this stupid resnet kernel without bells and whistles... So humiliating :)", "votes": null }, { "id": "742142", "postDate": "02/11/2020 05:09:04", "content": "I share the feeling with you. I literally started comparing my submission.csv to that kernel's submission.csv!
", "rawMarkdown": "I share the feeling with you. I literally started comparing my submission.csv to that kernel's submission.csv!", "votes": null }, { "id": "742146", "postDate": "02/11/2020 05:13:52", "content": "On the bright side, this kernel original author optimized it and added bells and whistles and it scored less... It is like that sometimes.
", "rawMarkdown": "On the bright side, this kernel original author optimized it and added bells and whistles and it scored less... It is like that sometimes.", "votes": null }, { "id": "742149", "postDate": "02/11/2020 05:17:08", "content": "Yeah. But still, just by copy-paste and will get you to top 10% is not really fair. No offense. I do understand public notebooks won't get you a medal BUT probably made some of the kagglers pretty desperate(especially there is no way to improve it).
", "rawMarkdown": "Yeah. But still, just by copy-paste and will get you to top 10% is not really fair. No offense. I do understand public notebooks won't get you a medal BUT probably made some of the kagglers pretty desperate(especially there is no way to improve it).", "votes": null }, { "id": "742154", "postDate": "02/11/2020 05:26:23", "content": "Meh, wait till the heavy guys start getting 0.1s... It's all part of the learning process. You need the drive.
", "rawMarkdown": "Meh, wait till the heavy guys start getting 0.1s... It's all part of the learning process. You need the drive.", "votes": null }, { "id": "742560", "postDate": "02/11/2020 11:30:07", "content": "Now imagine working extra hard and landing just in the middle :D ... I'm so frustrated that I'm going back to the drawing board now ... it looks like hand clustering the train set ... the \"bubble\" seems to be when you have a validation set leak in the train/val split :(
", "rawMarkdown": "Now imagine working extra hard and landing just in the middle :D ... I'm so frustrated that I'm going back to the drawing board now ... it looks like hand clustering the train set ... the \"bubble\" seems to be when you have a validation set leak in the train/val split :(", "votes": null }, { "id": "742620", "postDate": "02/11/2020 12:23:45", "content": "Congrats man!
", "rawMarkdown": "Congrats man!", "votes": null }, { "id": "743071", "postDate": "02/11/2020 19:00:14", "content": "Just scored a 0.426 LB (Before I was a public kernel forking sinner :D )! I was failing to reproduce the model from @humananalog public kernel for weeks; but with more data, augmentation, tuning batch size, I could reproduce it and get even a better score than the public kernel. My kernel is just an improvement over that kernel. My focus was entirely on reducing overfitting, rather than improving the score. Now, I have a baseline and will try to improve with different approaches.
\n\nAlso, if it helps: train BCE: 0.1551, val BCE: 0.3817 (Val folders 41-50).
", "rawMarkdown": "Just scored a 0.426 LB (Before I was a public kernel forking sinner :D )! I was failing to reproduce the model from @humananalog public kernel for weeks; but with more data, augmentation, tuning batch size, I could reproduce it and get even a better score than the public kernel. My kernel is just an improvement over that kernel. My focus was entirely on reducing overfitting, rather than improving the score. Now, I have a baseline and will try to improve with different approaches.\n\nAlso, if it helps: train BCE: 0.1551, val BCE: 0.3817 (Val folders 41-50).", "votes": null }, { "id": "743199", "postDate": "02/11/2020 21:57:14", "content": "@Zenify Your story is how the public kernels are supposed to work. After getting a head start and understanding the kernel and its issues, you make a significant improvement. In this case of this public kernel, because the trained model was just presented and was not easily reproducible, the process broke down. Congratulations on your successes.\nBTW I agree with you that the overfitting is the biggest hurdle.
", "rawMarkdown": "Zenify Your story is how the public kernels are supposed to work. After getting a head start and understanding the kernel and its issues, you make a significant improvement. In this case of this public kernel, because the trained model was just presented and was not easily reproducible, the process broke down. Congratulations on your successes.\nBTW I agree with you that the overfitting is the biggest hurdle.", "votes": null }, { "id": "743206", "postDate": "02/11/2020 22:06:39", "content": "@petewills, Thanks! Indeed it feels amazing :)))
", "rawMarkdown": "petewills, Thanks! Indeed it feels amazing :)))", "votes": null }, { "id": "743268", "postDate": "02/12/2020 00:29:32", "content": "actually, he also supplied a training kernel to reproduce the results.
", "rawMarkdown": "actually, he also supplied a training kernel to reproduce the results.", "votes": null }, { "id": "743269", "postDate": "02/12/2020 00:30:41", "content": "Its not reproducible because it's not using his training data. And will give a lot worse results.
", "rawMarkdown": "Its not reproducible because it's not using his training data. And will give a lot worse results.", "votes": null }, { "id": "743292", "postDate": "02/12/2020 00:58:04", "content": "@moshel, as @unkownhihi said, it's not. I have followed the kernel but results were far worse. I had to make lot of changes to the training kernel. But, the inference kernel is still the same as the original, just I am using a very high number of frames per video during inference.
", "rawMarkdown": "moshel, as @unkownhihi said, it's not. I have followed the kernel but results were far worse. I had to make lot of changes to the training kernel. But, the inference kernel is still the same as the original, just I am using a very high number of frames per video during inference.", "votes": null }, { "id": "745412", "postDate": "02/13/2020 19:45:24", "content": "To be fair I believe the original intent of both of the kernels is to provide a template for error free pipeline for training and inference ... data wrangling is what makes or breaks a model (plus tuning of course) ... the trained model he shared was trained on his dataset only for couple of epochs to demo the inference pipeline ... everyone needs to grind extra harder now I suppose ;)
", "rawMarkdown": "To be fair I believe the original intent of both of the kernels is to provide a template for error free pipeline for training and inference ... data wrangling is what makes or breaks a model (plus tuning of course) ... the trained model he shared was trained on his dataset only for couple of epochs to demo the inference pipeline ... everyone needs to grind extra harder now I suppose ;)", "votes": null }, { "id": "747200", "postDate": "02/16/2020 05:21:37", "content": "I have to admit his kernel did gave us extra effort to improve our LB, which resulted in our current best score and we couldn't have done it if there wasn't that kernel.
", "rawMarkdown": "I have to admit his kernel did gave us extra effort to improve our LB, which resulted in our current best score and we couldn't have done it if there wasn't that kernel.", "votes": null }, { "id": "755825", "postDate": "02/25/2020 06:35:34", "content": "The bubble doesn't die 🤕 I'm almost burning out now ... I have to say this competition plus the bubble pushed me to start reading papers/ SOTA again to rise higher than this floating bubble 💪
", "rawMarkdown": "The bubble doesn't die 🤕 I'm almost burning out now ... I have to say this competition plus the bubble pushed me to start reading papers/ SOTA again to rise higher than this floating bubble 💪", "votes": null }, { "id": "756682", "postDate": "02/26/2020 00:18:14", "content": "There was a new bubble that sprouted from the old when someone submitted two magic models, ensembled, much like the generation of inflationary universes. I am working my way through the new one now.
\n\nPeople need to think before they do this - it is a flaw in the kaggle model.
", "rawMarkdown": "There was a new bubble that sprouted from the old when someone submitted two magic models, ensembled, much like the generation of inflationary universes. I am working my way through the new one now.\n\nPeople need to think before they do this - it is a flaw in the kaggle model.", "votes": null } ], "comments": [ { "id": 742131, "author_name": "unkownhihi", "author_url": "", "post_date": "02/11/2020 04:57:53", "content": "Exactly. When our highest scoring subs are lower than that bubble zone, we were feaked out! Maybe people should not post these high scoring kernels. But anyways, it made us work so hard that actually gave us a gigantic boost! Its a bad thing and also maybe a good thing.
", "votes": null, "replies": [ { "id": 742136, "author_name": "moshel", "author_url": "", "post_date": "02/11/2020 05:02:38", "content": "I am still technically below this stupid resnet kernel without bells and whistles... So humiliating :)
", "votes": null, "replies": [] }, { "id": 742142, "author_name": "unkownhihi", "author_url": "", "post_date": "02/11/2020 05:09:04", "content": "I share the feeling with you. I literally started comparing my submission.csv to that kernel's submission.csv!
", "votes": null, "replies": [] }, { "id": 742146, "author_name": "moshel", "author_url": "", "post_date": "02/11/2020 05:13:52", "content": "On the bright side, this kernel original author optimized it and added bells and whistles and it scored less... It is like that sometimes.
", "votes": null, "replies": [] }, { "id": 742149, "author_name": "unkownhihi", "author_url": "", "post_date": "02/11/2020 05:17:08", "content": "Yeah. But still, just by copy-paste and will get you to top 10% is not really fair. No offense. I do understand public notebooks won't get you a medal BUT probably made some of the kagglers pretty desperate(especially there is no way to improve it).
", "votes": null, "replies": [] }, { "id": 742154, "author_name": "moshel", "author_url": "", "post_date": "02/11/2020 05:26:23", "content": "Meh, wait till the heavy guys start getting 0.1s... It's all part of the learning process. You need the drive.
", "votes": null, "replies": [] }, { "id": 742560, "author_name": "ma7moud", "author_url": "", "post_date": "02/11/2020 11:30:07", "content": "Now imagine working extra hard and landing just in the middle :D ... I'm so frustrated that I'm going back to the drawing board now ... it looks like hand clustering the train set ... the \"bubble\" seems to be when you have a validation set leak in the train/val split :(
", "votes": null, "replies": [] }, { "id": 743071, "author_name": "debanga", "author_url": "", "post_date": "02/11/2020 19:00:14", "content": "Just scored a 0.426 LB (Before I was a public kernel forking sinner :D )! I was failing to reproduce the model from @humananalog public kernel for weeks; but with more data, augmentation, tuning batch size, I could reproduce it and get even a better score than the public kernel. My kernel is just an improvement over that kernel. My focus was entirely on reducing overfitting, rather than improving the score. Now, I have a baseline and will try to improve with different approaches.
\n\nAlso, if it helps: train BCE: 0.1551, val BCE: 0.3817 (Val folders 41-50).
", "votes": null, "replies": [] }, { "id": 743199, "author_name": "petewills", "author_url": "", "post_date": "02/11/2020 21:57:14", "content": "@Zenify Your story is how the public kernels are supposed to work. After getting a head start and understanding the kernel and its issues, you make a significant improvement. In this case of this public kernel, because the trained model was just presented and was not easily reproducible, the process broke down. Congratulations on your successes.\nBTW I agree with you that the overfitting is the biggest hurdle.
", "votes": null, "replies": [] }, { "id": 743206, "author_name": "debanga", "author_url": "", "post_date": "02/11/2020 22:06:39", "content": "@petewills, Thanks! Indeed it feels amazing :)))
", "votes": null, "replies": [] }, { "id": 743268, "author_name": "moshel", "author_url": "", "post_date": "02/12/2020 00:29:32", "content": "actually, he also supplied a training kernel to reproduce the results.
", "votes": null, "replies": [] }, { "id": 743269, "author_name": "unkownhihi", "author_url": "", "post_date": "02/12/2020 00:30:41", "content": "Its not reproducible because it's not using his training data. And will give a lot worse results.
", "votes": null, "replies": [] }, { "id": 743292, "author_name": "debanga", "author_url": "", "post_date": "02/12/2020 00:58:04", "content": "@moshel, as @unkownhihi said, it's not. I have followed the kernel but results were far worse. I had to make lot of changes to the training kernel. But, the inference kernel is still the same as the original, just I am using a very high number of frames per video during inference.
", "votes": null, "replies": [] }, { "id": 745412, "author_name": "ma7moud", "author_url": "", "post_date": "02/13/2020 19:45:24", "content": "To be fair I believe the original intent of both of the kernels is to provide a template for error free pipeline for training and inference ... data wrangling is what makes or breaks a model (plus tuning of course) ... the trained model he shared was trained on his dataset only for couple of epochs to demo the inference pipeline ... everyone needs to grind extra harder now I suppose ;)
", "votes": null, "replies": [] }, { "id": 747200, "author_name": "unkownhihi", "author_url": "", "post_date": "02/16/2020 05:21:37", "content": "I have to admit his kernel did gave us extra effort to improve our LB, which resulted in our current best score and we couldn't have done it if there wasn't that kernel.
", "votes": null, "replies": [] } ] }, { "id": 742620, "author_name": "akashnandi", "author_url": "", "post_date": "02/11/2020 12:23:45", "content": "Congrats man!
", "votes": null, "replies": [] }, { "id": 755825, "author_name": "ma7moud", "author_url": "", "post_date": "02/25/2020 06:35:34", "content": "The bubble doesn't die 🤕 I'm almost burning out now ... I have to say this competition plus the bubble pushed me to start reading papers/ SOTA again to rise higher than this floating bubble 💪
", "votes": null, "replies": [ { "id": 756682, "author_name": "petewills", "author_url": "", "post_date": "02/26/2020 00:18:14", "content": "There was a new bubble that sprouted from the old when someone submitted two magic models, ensembled, much like the generation of inflationary universes. I am working my way through the new one now.
\n\nPeople need to think before they do this - it is a flaw in the kaggle model.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "742107": "I don't mean to be vulgar but I just passed through the ridiculous bubble - 377 places with one submission. Now I can start having fun again.", "742131": "Exactly. When our highest scoring subs are lower than that bubble zone, we were feaked out! Maybe people should not post these high scoring kernels. But anyways, it made us work so hard that actually gave us a gigantic boost! Its a bad thing and also maybe a good thing.", "742136": "I am still technically below this stupid resnet kernel without bells and whistles... So humiliating :)", "742142": "I share the feeling with you. I literally started comparing my submission.csv to that kernel's submission.csv!", "742146": "On the bright side, this kernel original author optimized it and added bells and whistles and it scored less... It is like that sometimes.", "742149": "Yeah. But still, just by copy-paste and will get you to top 10% is not really fair. No offense. I do understand public notebooks won't get you a medal BUT probably made some of the kagglers pretty desperate(especially there is no way to improve it).", "742154": "Meh, wait till the heavy guys start getting 0.1s... It's all part of the learning process. You need the drive.", "742560": "Now imagine working extra hard and landing just in the middle :D ... I'm so frustrated that I'm going back to the drawing board now ... it looks like hand clustering the train set ... the \"bubble\" seems to be when you have a validation set leak in the train/val split :(", "742620": "Congrats man!", "743071": "Just scored a 0.426 LB (Before I was a public kernel forking sinner :D )! I was failing to reproduce the model from @humananalog public kernel for weeks; but with more data, augmentation, tuning batch size, I could reproduce it and get even a better score than the public kernel. My kernel is just an improvement over that kernel. My focus was entirely on reducing overfitting, rather than improving the score. Now, I have a baseline and will try to improve with different approaches.\n\nAlso, if it helps: train BCE: 0.1551, val BCE: 0.3817 (Val folders 41-50).", "743199": "Zenify Your story is how the public kernels are supposed to work. After getting a head start and understanding the kernel and its issues, you make a significant improvement. In this case of this public kernel, because the trained model was just presented and was not easily reproducible, the process broke down. Congratulations on your successes.\nBTW I agree with you that the overfitting is the biggest hurdle.", "743206": "petewills, Thanks! Indeed it feels amazing :)))", "743268": "actually, he also supplied a training kernel to reproduce the results.", "743269": "Its not reproducible because it's not using his training data. And will give a lot worse results.", "743292": "moshel, as @unkownhihi said, it's not. I have followed the kernel but results were far worse. I had to make lot of changes to the training kernel. But, the inference kernel is still the same as the original, just I am using a very high number of frames per video during inference.", "745412": "To be fair I believe the original intent of both of the kernels is to provide a template for error free pipeline for training and inference ... data wrangling is what makes or breaks a model (plus tuning of course) ... the trained model he shared was trained on his dataset only for couple of epochs to demo the inference pipeline ... everyone needs to grind extra harder now I suppose ;)", "747200": "I have to admit his kernel did gave us extra effort to improve our LB, which resulted in our current best score and we couldn't have done it if there wasn't that kernel.", "755825": "The bubble doesn't die 🤕 I'm almost burning out now ... I have to say this competition plus the bubble pushed me to start reading papers/ SOTA again to rise higher than this floating bubble 💪", "756682": "There was a new bubble that sprouted from the old when someone submitted two magic models, ensembled, much like the generation of inflationary universes. I am working my way through the new one now.\n\nPeople need to think before they do this - it is a flaw in the kaggle model." }, "source": "meta" }