From 61st to 532nd place, trying to trust our CV, still such a terrible shakeup.
\n\n@abhishek classifies this competition as a lottery 🤐
", "messages": [ { "id": 940420, "postDate": "2020-07-23T00:06:53.950Z", "content": "From 61st to 532nd place, trying to trust our CV, still such a terrible shakeup.
\n\n@abhishek classifies this competition as a lottery 🤐
", "rawMarkdown": "From 61st to 532nd place, trying to trust our CV, still such a terrible shakeup. \n\n@abhishek classifies this competition as a lottery 🤐 ", "votes": 22 }, { "id": 940455, "postDate": "2020-07-23T00:39:07.567Z", "content": "Very rough, but not totally unexpected. I stopped working on the competition a few weeks before the end since it was becoming apparent that it was going to be a lottery for private LB.
", "rawMarkdown": "Very rough, but not totally unexpected. I stopped working on the competition a few weeks before the end since it was becoming apparent that it was going to be a lottery for private LB.", "votes": 14, "replies": [ { "id": 941241, "postDate": "2020-07-23T06:32:37.633Z", "content": "I still need to learn that.
", "rawMarkdown": "I still need to learn that.", "votes": 5 } ] }, { "id": 940446, "postDate": "2020-07-23T00:30:46.930Z", "content": "Same thing happened to me, from 76th to 507th somehow. For me the CV was greater than the public LB by roughly 0.005, to be honest I don't know how I could have predicted such an awful final standing
", "rawMarkdown": "Same thing happened to me, from 76th to 507th somehow. For me the CV was greater than the public LB by roughly 0.005, to be honest I don't know how I could have predicted such an awful final standing", "votes": 1 }, { "id": 940435, "postDate": "2020-07-23T00:21:51.023Z", "content": "We too experiences the same and fell from 41 to 536. Shakeup is really terrible.
", "rawMarkdown": "We too experiences the same and fell from 41 to 536. Shakeup is really terrible.", "votes": 1 }, { "id": 940679, "postDate": "2020-07-23T04:46:28.353Z", "content": "Its bloodbath . Either data split was too wrong and model get biased to images available for LB.\nIf this is the case then all to those who organized this. But not good as spent a lot of time and effort .\nWe did not select one model for final ranking as has low score of 87 only, got the score of 92.7 in PB which is a top 20 score. However BECAUSE we did not select this low score we are at 536 , a fall from 41.\nHad we been select it , we would be in top 20. I think this is not the way to waste time of people .\ntoo bad .
", "rawMarkdown": "Its bloodbath . Either data split was too wrong and model get biased to images available for LB.\nIf this is the case then all to those who organized this. But not good as spent a lot of time and effort .\nWe did not select one model for final ranking as has low score of 87 only, got the score of 92.7 in PB which is a top 20 score. However BECAUSE we did not select this low score we are at 536 , a fall from 41.\nHad we been select it , we would be in top 20. I think this is not the way to waste time of people .\ntoo bad .", "replies": [ { "id": 940882, "postDate": "2020-07-23T06:02:37.203Z", "content": "Welcome to Kaggle!
\nOne of the best thing about Kaggle is precisely the fact that you cannot choose your best submission \"a posteriori\". This is not \"a waste of time for people\" but an effective strategy to reward competitors who can:
\n- built models which generalize.
\n- understand the models they built.
I know that it hurts to be penalized by the private lb, but this aspect it's really crucial to Kaggle and you'd better embrace it if you want to enjoy the competitive side here. Otherwise you can always focus on what you have learnt, which it's almost always worth the time spent.
\n\nOn a side note, if everyone could choose his/her best submission you would probably not be in the top 20 anyway ;)
", "rawMarkdown": "Welcome to Kaggle! \nOne of the best thing about Kaggle is precisely the fact that you cannot choose your best submission \"a posteriori\". This is not \"a waste of time for people\" but an effective strategy to reward competitors who can: \n- built models which generalize. \n- understand the models they built. \n\nI know that it hurts to be penalized by the private lb, but this aspect it's really crucial to Kaggle and you'd better embrace it if you want to enjoy the competitive side here. Otherwise you can always focus on what you have learnt, which it's almost always worth the time spent.\n\nOn a side note, if everyone could choose his/her best submission you would probably not be in the top 20 anyway ;)", "votes": 8 }, { "id": 941344, "postDate": "2020-07-23T07:15:52.400Z", "content": "For sure , a lot of learning , but my brain is little near to boiling point at the moment.\nAfter cooling , could be in better position to think more about this comp and what not do in next.
", "rawMarkdown": "For sure , a lot of learning , but my brain is little near to boiling point at the moment.\nAfter cooling , could be in better position to think more about this comp and what not do in next.", "votes": 1 } ] }, { "id": 940453, "postDate": "2020-07-23T00:38:56.287Z", "content": "Well, I feel more for the organizers... I am not sure they will be happy giving money to the third guy, who probably barely tweeked a public kernel... and I don't know how they will sort out what to use out of this competition (which is the goal in the end... here we have no way of knowing which kernel would perform best in a real setup, and so good luck sorting out which preprocessing/deep learning trick would actually be useful in the future...).
\n\nI look forward to see the approaches that were stable over time/ and across private and public, such as lafoss', save the prostate, NS pathology and Vanda. Maybe also the bears, and perhaps ChienYiChi, because I believe Arnaud Roussel has an interesting approach to the problem... I hope their solution can provide the organizers with some insight about what works best...
", "rawMarkdown": "Well, I feel more for the organizers... I am not sure they will be happy giving money to the third guy, who probably barely tweeked a public kernel... and I don't know how they will sort out what to use out of this competition (which is the goal in the end... here we have no way of knowing which kernel would perform best in a real setup, and so good luck sorting out which preprocessing/deep learning trick would actually be useful in the future...).\n\nI look forward to see the approaches that were stable over time/ and across private and public, such as lafoss', save the prostate, NS pathology and Vanda. Maybe also the bears, and perhaps ChienYiChi, because I believe Arnaud Roussel has an interesting approach to the problem... I hope their solution can provide the organizers with some insight about what works best...", "votes": -2, "replies": [ { "id": 940457, "postDate": "2020-07-23T00:42:25.270Z", "content": "Our public lb of 0.921 corresponds to a private of 0.927. Although when we made a mistake during inference, that same run gave public lb of 0.921 and a private of 0.93! This really shows you how random this competition is
", "rawMarkdown": "Our public lb of 0.921 corresponds to a private of 0.927. Although when we made a mistake during inference, that same run gave public lb of 0.921 and a private of 0.93! This really shows you how random this competition is", "votes": 1 } ] }, { "id": 940439, "postDate": "2020-07-23T00:25:48.973Z", "content": "It's unfortunate... The test set is way too small
", "rawMarkdown": "It's unfortunate... The test set is way too small" }, { "id": 940661, "postDate": "2020-07-23T04:30:52.897Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 940455, "author_name": "Ian Pan", "author_url": "", "post_date": "2020-07-23T00:39:07.567000", "content": "Very rough, but not totally unexpected. I stopped working on the competition a few weeks before the end since it was becoming apparent that it was going to be a lottery for private LB.
", "votes": 14, "replies": [ { "id": 941241, "author_name": "Psi", "author_url": "", "post_date": "2020-07-23T06:32:37.633000", "content": "I still need to learn that.
", "votes": 5, "replies": [] } ] }, { "id": 940446, "author_name": "Pasquale", "author_url": "", "post_date": "2020-07-23T00:30:46.930000", "content": "Same thing happened to me, from 76th to 507th somehow. For me the CV was greater than the public LB by roughly 0.005, to be honest I don't know how I could have predicted such an awful final standing
", "votes": 1, "replies": [] }, { "id": 940435, "author_name": "Karan", "author_url": "", "post_date": "2020-07-23T00:21:51.023000", "content": "We too experiences the same and fell from 41 to 536. Shakeup is really terrible.
", "votes": 1, "replies": [] }, { "id": 940679, "author_name": "Rajnish Chauhan", "author_url": "", "post_date": "2020-07-23T04:46:28.353000", "content": "Its bloodbath . Either data split was too wrong and model get biased to images available for LB.\nIf this is the case then all to those who organized this. But not good as spent a lot of time and effort .\nWe did not select one model for final ranking as has low score of 87 only, got the score of 92.7 in PB which is a top 20 score. However BECAUSE we did not select this low score we are at 536 , a fall from 41.\nHad we been select it , we would be in top 20. I think this is not the way to waste time of people .\ntoo bad .
", "votes": 0, "replies": [ { "id": 940882, "author_name": "bluetrain", "author_url": "", "post_date": "2020-07-23T06:02:37.203000", "content": "Welcome to Kaggle!
\nOne of the best thing about Kaggle is precisely the fact that you cannot choose your best submission \"a posteriori\". This is not \"a waste of time for people\" but an effective strategy to reward competitors who can:
\n- built models which generalize.
\n- understand the models they built.
I know that it hurts to be penalized by the private lb, but this aspect it's really crucial to Kaggle and you'd better embrace it if you want to enjoy the competitive side here. Otherwise you can always focus on what you have learnt, which it's almost always worth the time spent.
\n\nOn a side note, if everyone could choose his/her best submission you would probably not be in the top 20 anyway ;)
", "votes": 8, "replies": [] }, { "id": 941344, "author_name": "Rajnish Chauhan", "author_url": "", "post_date": "2020-07-23T07:15:52.400000", "content": "For sure , a lot of learning , but my brain is little near to boiling point at the moment.\nAfter cooling , could be in better position to think more about this comp and what not do in next.
", "votes": 1, "replies": [] } ] }, { "id": 940453, "author_name": "Benjamin Dubreu", "author_url": "", "post_date": "2020-07-23T00:38:56.287000", "content": "Well, I feel more for the organizers... I am not sure they will be happy giving money to the third guy, who probably barely tweeked a public kernel... and I don't know how they will sort out what to use out of this competition (which is the goal in the end... here we have no way of knowing which kernel would perform best in a real setup, and so good luck sorting out which preprocessing/deep learning trick would actually be useful in the future...).
\n\nI look forward to see the approaches that were stable over time/ and across private and public, such as lafoss', save the prostate, NS pathology and Vanda. Maybe also the bears, and perhaps ChienYiChi, because I believe Arnaud Roussel has an interesting approach to the problem... I hope their solution can provide the organizers with some insight about what works best...
", "votes": -2, "replies": [ { "id": 940457, "author_name": "Shujun", "author_url": "", "post_date": "2020-07-23T00:42:25.270000", "content": "Our public lb of 0.921 corresponds to a private of 0.927. Although when we made a mistake during inference, that same run gave public lb of 0.921 and a private of 0.93! This really shows you how random this competition is
", "votes": 1, "replies": [] } ] }, { "id": 940439, "author_name": "Shujun", "author_url": "", "post_date": "2020-07-23T00:25:48.973000", "content": "It's unfortunate... The test set is way too small
", "votes": 0, "replies": [] }, { "id": 940661, "author_name": "", "author_url": "", "post_date": "2020-07-23T04:30:52.897000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "940420": "From 61st to 532nd place, trying to trust our CV, still such a terrible shakeup. \n\n@abhishek classifies this competition as a lottery 🤐 ", "940455": "Very rough, but not totally unexpected. I stopped working on the competition a few weeks before the end since it was becoming apparent that it was going to be a lottery for private LB.", "940446": "Same thing happened to me, from 76th to 507th somehow. For me the CV was greater than the public LB by roughly 0.005, to be honest I don't know how I could have predicted such an awful final standing", "940435": "We too experiences the same and fell from 41 to 536. Shakeup is really terrible.", "940679": "Its bloodbath . Either data split was too wrong and model get biased to images available for LB.\nIf this is the case then all to those who organized this. But not good as spent a lot of time and effort .\nWe did not select one model for final ranking as has low score of 87 only, got the score of 92.7 in PB which is a top 20 score. However BECAUSE we did not select this low score we are at 536 , a fall from 41.\nHad we been select it , we would be in top 20. I think this is not the way to waste time of people .\ntoo bad .", "940453": "Well, I feel more for the organizers... I am not sure they will be happy giving money to the third guy, who probably barely tweeked a public kernel... and I don't know how they will sort out what to use out of this competition (which is the goal in the end... here we have no way of knowing which kernel would perform best in a real setup, and so good luck sorting out which preprocessing/deep learning trick would actually be useful in the future...).\n\nI look forward to see the approaches that were stable over time/ and across private and public, such as lafoss', save the prostate, NS pathology and Vanda. Maybe also the bears, and perhaps ChienYiChi, because I believe Arnaud Roussel has an interesting approach to the problem... I hope their solution can provide the organizers with some insight about what works best...", "940439": "It's unfortunate... The test set is way too small", "940661": "" } }