![\"Excalibur:](\"https://i.pinimg.com/originals/e8/4d/22/e84d22648b032e0b6da27f7094f16119.jpg\")
I just wanted to say that this has been one of the most enjoyable competitions I've been involved in on Kaggle. I would like to thank all those who have been extremely helpful by producing great datasets and those (often the same people) who have been suggesting approaches and ideas.
\n\nWe have barely started and I've climbed up that leaderboard and slipped and climbed and slipped.
\n\nI still have a lot to try and experiment with. I have Colabs working well for TPU (either TFRecords or JPEG), but I have to control my urge to see how I'm going by burning through my available Submissions to quickly.\n(yeah yeah... trust your CV lol)
\n\nRemember my day job is stacking potatoes neatly so this is a huge outlet for me.
\n\nthanks everyone (and Kaggle)
", "messages": [ { "id": "874189", "postDate": "06/04/2020 17:41:25", "content": "
I just wanted to say that this has been one of the most enjoyable competitions I've been involved in on Kaggle. I would like to thank all those who have been extremely helpful by producing great datasets and those (often the same people) who have been suggesting approaches and ideas.
\n\nWe have barely started and I've climbed up that leaderboard and slipped and climbed and slipped.
\n\nI still have a lot to try and experiment with. I have Colabs working well for TPU (either TFRecords or JPEG), but I have to control my urge to see how I'm going by burning through my available Submissions to quickly.\n(yeah yeah... trust your CV lol)
\n\nRemember my day job is stacking potatoes neatly so this is a huge outlet for me.
\n\nthanks everyone (and Kaggle)
", "rawMarkdown": "\n\nI just wanted to say that this has been one of the **most enjoyable competitions** I've been involved in on Kaggle. I would like to thank all those who have been extremely helpful by producing great datasets and those (often the same people) who have been suggesting approaches and ideas.\n\nWe have barely started and I've climbed up that leaderboard and slipped and climbed and slipped. \n\nI still have a lot to try and experiment with. I have Colabs working well for TPU (either TFRecords or JPEG), but I have to control my urge to see how I'm going by burning through my available Submissions to quickly.\n(yeah yeah... trust your CV lol)\n\nRemember my day job is stacking potatoes neatly so this is a huge outlet for me.\n\nthanks everyone (and Kaggle)", "votes": null }, { "id": "874231", "postDate": "06/04/2020 18:14:12", "content": "Bruce, you are not alone! I find this enjoyable as well - there is so many ways to approach this problem! \nGood luck!
", "rawMarkdown": "Bruce, you are not alone! I find this enjoyable as well - there is so many ways to approach this problem! \nGood luck!", "votes": null }, { "id": "974168", "postDate": "08/17/2020 19:05:06", "content": "And how do I feel now… with 5 hrs to go…
\nNot at all impressed and depressed by the whole experience.
\nThis is probably the first time I've actually seriously tried to win a competition that gives out medals. Sure I've done ok in some of the learning, playground, kudos type comps but I thought I'd be able to get a bronze medal at least.
\nI also tried to team up but found that my \"use them or loose them\" approach to submissions messed that up. I did get 2 other requests to merge but of course, I had the same issue.
\nI'd climb up 6 places and fall 26 when I went to work … then repeat … then repeat …
\nBut now, like most of us, I ended up on the wrong side of the last-minute high LB public submission and just got washed away in the flood of copy/paste that went on.
\nMy wife has been pretty tolerant of me spending so much time on this competition as she understands my need for an intellectual escape from stacking potatoes (yes, I really do that) but after this, she'll be pretty sceptical of the whole endeavour.
\nHopefully, my approach will see me through to a medal on the Private LB and yes, I have seen people rise hundreds of places in other competitions so I guess it's not over yet.
\nI'd still like to thank Chris, AgentAuers and the many others who have given freely and openly to help beginners and provide insight, even the \"grumpy\" ones ;) . That aspect has been a total success in this competition.
\nI have learnt a great deal, more than in any other of my 19 competitions, so I should just be thankful for that and move on. Motivation to do more may well depend on the result tonight though.
\nthanks all
", "rawMarkdown": "And how do I feel now... with 5 hrs to go...\n\nNot at all impressed and depressed by the whole experience. \n\nThis is probably the first time I've actually seriously tried to win a competition that gives out medals. Sure I've done ok in some of the learning, playground, kudos type comps but I thought I'd be able to get a bronze medal at least.\n\nI also tried to team up but found that my \"use them or loose them\" approach to submissions messed that up. I did get 2 other requests to merge but of course, I had the same issue.\n\nI'd climb up 6 places and fall 26 when I went to work ... then repeat ... then repeat ...\n\nBut now, like most of us, I ended up on the wrong side of the last-minute high LB public submission and just got washed away in the flood of copy/paste that went on.\n\nMy wife has been pretty tolerant of me spending so much time on this competition as she understands my need for an intellectual escape from stacking potatoes (yes, I really do that) but after this, she'll be pretty sceptical of the whole endeavour.\n\nHopefully, my approach will see me through to a medal on the Private LB and yes, I have seen people rise hundreds of places in other competitions so I guess it's not over yet.\n\nI'd still like to thank Chris, AgentAuers and the many others who have given freely and openly to help beginners and provide insight, even the \"grumpy\" ones ;) . That aspect has been a total success in this competition. \n\nI have learnt a great deal, more than in any other of my 19 competitions, so I should just be thankful for that and move on. Motivation to do more may well depend on the result tonight though.\n\nthanks all", "votes": null }, { "id": "974240", "postDate": "08/17/2020 20:23:57", "content": "\n\nI thought I'd be able to get a bronze medal at least
\n
Don't judge yourself too quickly.
\nI bet you will climb in private part (bronze for sure) and with so much hard work I would say silver+ (top 100).
\nGood luck to you.
", "rawMarkdown": "> I thought I'd be able to get a bronze medal at least\n\nDon't judge yourself too quickly.\n\nI bet you will climb in private part (bronze for sure) and with so much hard work I would say silver+ (top 100).\n\nGood luck to you.", "votes": null }, { "id": "974281", "postDate": "08/17/2020 21:10:05", "content": "Thinking about if I would bet to rank above you - I don't dare :)
\n2 of my 3 selected submissions are well below your best result.
my public best managed about rank 1100 - private best has been more cv oriented - but huge shakeup I want to understand a bit….
", "rawMarkdown": "my public best managed about rank 1100 - private best has been more cv oriented - but huge shakeup I want to understand a bit....", "votes": null }, { "id": "974515", "postDate": "08/18/2020 01:05:35", "content": "My best “if only” score was 0.9456 putting me about 22nd
", "rawMarkdown": "My best “if only” score was 0.9456 putting me about 22nd", "votes": null }, { "id": "974524", "postDate": "08/18/2020 01:11:40", "content": "ok - it's over - and it's not easy to understand…
\nmy best would have been 0.9429 - about rank 70 - but no special reason to choose this one. Funny fact - this is Chris' score. Hard to understand why Chris did not score better. He did everything right.
If someone can explain the connection between private and public scores - pls do…
\n… too much metadata seems to harm - which is very much unexpected… Chris sees this different…
", "rawMarkdown": "If someone can explain the connection between private and public scores - pls do...\n\n... too much metadata seems to harm - which is very much unexpected... Chris sees this different...", "votes": null }, { "id": "975320", "postDate": "08/18/2020 08:42:22", "content": "Hello Bruce, as raddar already said, you are not alone.
\nI crashed down to place 1000, … and I really do not know why the shakeup did hit me so hard.
\nMy strategy was to ensemble as much models as possible without any selection based on public LB scores.
\nThe best submission scored private 0.9283 / public 0.9610 and consists of more than 100 models.
At some point in the competition (about 4 weeks ago) I was only one place behind chris deotte :-) and this without any dangerous selection of models based on LB scores.
\nNext time I will define fixed folds at the very beginning of the competition, and use cv scores for the whole pipeline, from modelselection to final ensembling and postprocessing.
\nThank you all! I really learned a lot in this competition from you all :-)
", "rawMarkdown": "Hello Bruce, as raddar already said, you are not alone.\nI crashed down to place 1000, ... and I really do not know why the shakeup did hit me so hard.\nMy strategy was to ensemble as much models as possible without any selection based on public LB scores. \nThe best submission scored private 0.9283 / public 0.9610 and consists of more than 100 models. \n\nAt some point in the competition (about 4 weeks ago) I was only one place behind chris deotte :-) and this without any dangerous selection of models based on LB scores.\n\nNext time I will define fixed folds at the very beginning of the competition, and use cv scores for the whole pipeline, from modelselection to final ensembling and postprocessing.\n\nThank you all! I really learned a lot in this competition from you all :-)", "votes": null }, { "id": "975549", "postDate": "08/18/2020 10:54:07", "content": "Hello AgentAuers - it really hit you hard (at least you can remember the rank easily) … I used your kernel a lot until Chris came up with his notebook - thx for that. I find the role of metadata very confusing. Had big lb ups with ensembling best image result and a good tabular submission with weights 0.7/0.3
\nBut my last experiments with ranked submissions even indicated less tabular data seems better (2nd private rank wrote, that he did not use tab data at all):
\nA: ranked 3 B555 ensembled (no good lb results - chosen by oof cv value around 0.92)
\nB: ranked best LB result (ensembled 3 B555 best lb result, best was 0.957)
\nC: ranked tab submission
\nResults:
\nweights 0.2 0.5 0.3: priv: 0.9095 public: 0.9346
\nweights 0.2 0.5 0.1: priv: 0.9324 public: 0.9541
\nweights 0.1 0.6 0.1: priv: 0.9301 public: 0.9543
\nweights 0.1 0.6 0.06: priv: 0.9336 public: 0.9568
\nthis confuses me…
\nlittle correction: ensemble A used some metadata for ensembling - calculated with the 'LGB/MLP/RIDGE STACKING Effnet OOF with META' notebook
", "rawMarkdown": "Hello AgentAuers - it really hit you hard (at least you can remember the rank easily) ... I used your kernel a lot until Chris came up with his notebook - thx for that. I find the role of metadata very confusing. Had big lb ups with ensembling best image result and a good tabular submission with weights 0.7/0.3\n\nBut my last experiments with ranked submissions even indicated less tabular data seems better (2nd private rank wrote, that he did not use tab data at all):\n\nA: ranked 3 B555 ensembled (no good lb results - chosen by oof cv value around 0.92)\n\nB: ranked best LB result (ensembled 3 B555 best lb result, best was 0.957)\n\nC: ranked tab submission\n\nResults: \n\nweights 0.2 0.5 0.3: priv: 0.9095 public: 0.9346\n\nweights 0.2 0.5 0.1: priv: 0.9324 public: 0.9541\n\nweights 0.1 0.6 0.1: priv: 0.9301 public: 0.9543\n\nweights 0.1 0.6 0.06: priv: 0.9336 public: 0.9568\n\nthis confuses me...\n\nlittle correction: ensemble A used some metadata for ensembling - calculated with the 'LGB/MLP/RIDGE STACKING Effnet OOF with META' notebook", "votes": null }, { "id": "975558", "postDate": "08/18/2020 10:57:28", "content": "\n\nMy best “if only” score was 0.9456 putting me about 22nd
\n
a promise…
", "rawMarkdown": ">My best “if only” score was 0.9456 putting me about 22nd\n\na promise...", "votes": null }, { "id": "981981", "postDate": "08/22/2020 23:17:00", "content": "As I said - all be good)))). Not silver (continue what you are doing and next silver or gold for sure), but congratulations with bronze medal.
", "rawMarkdown": "As I said - all be good)))). Not silver (continue what you are doing and next silver or gold for sure), but congratulations with bronze medal.", "votes": null }, { "id": "981994", "postDate": "08/22/2020 23:50:32", "content": "@agentauers Don't feel bad, I will always remember you as the MVP of this comp. Your notebook was the beginning of all strong models. And you invented aggressive TTA! Everyone benefited from your helpful contributions, creative models, and insights. My triple stratified notebook is just your notebook with the addition of KFold. All my offline models use your augmentation settings. I just varied the read size, crop size, external data, upsample, and EfficientNet backbone (and then maximized CV). I'm not sure why my private is higher than your private.
", "rawMarkdown": "agentauers Don't feel bad, I will always remember you as the MVP of this comp. Your notebook was the beginning of all strong models. And you invented aggressive TTA! Everyone benefited from your helpful contributions, creative models, and insights. My triple stratified notebook is just your notebook with the addition of KFold. All my offline models use your augmentation settings. I just varied the read size, crop size, external data, upsample, and EfficientNet backbone (and then maximized CV). I'm not sure why my private is higher than your private.", "votes": null } ], "comments": [ { "id": 974168, "author_name": "mutantspore", "author_url": "", "post_date": "08/17/2020 19:05:06", "content": "And how do I feel now… with 5 hrs to go…
\nNot at all impressed and depressed by the whole experience.
\nThis is probably the first time I've actually seriously tried to win a competition that gives out medals. Sure I've done ok in some of the learning, playground, kudos type comps but I thought I'd be able to get a bronze medal at least.
\nI also tried to team up but found that my \"use them or loose them\" approach to submissions messed that up. I did get 2 other requests to merge but of course, I had the same issue.
\nI'd climb up 6 places and fall 26 when I went to work … then repeat … then repeat …
\nBut now, like most of us, I ended up on the wrong side of the last-minute high LB public submission and just got washed away in the flood of copy/paste that went on.
\nMy wife has been pretty tolerant of me spending so much time on this competition as she understands my need for an intellectual escape from stacking potatoes (yes, I really do that) but after this, she'll be pretty sceptical of the whole endeavour.
\nHopefully, my approach will see me through to a medal on the Private LB and yes, I have seen people rise hundreds of places in other competitions so I guess it's not over yet.
\nI'd still like to thank Chris, AgentAuers and the many others who have given freely and openly to help beginners and provide insight, even the \"grumpy\" ones ;) . That aspect has been a total success in this competition.
\nI have learnt a great deal, more than in any other of my 19 competitions, so I should just be thankful for that and move on. Motivation to do more may well depend on the result tonight though.
\nthanks all
", "votes": null, "replies": [ { "id": 974240, "author_name": "kyakovlev", "author_url": "", "post_date": "08/17/2020 20:23:57", "content": "\n\nI thought I'd be able to get a bronze medal at least
\n
Don't judge yourself too quickly.
\nI bet you will climb in private part (bronze for sure) and with so much hard work I would say silver+ (top 100).
\nGood luck to you.
", "votes": null, "replies": [] }, { "id": 974281, "author_name": "romanweilguny", "author_url": "", "post_date": "08/17/2020 21:10:05", "content": "Thinking about if I would bet to rank above you - I don't dare :)
\n2 of my 3 selected submissions are well below your best result.
my public best managed about rank 1100 - private best has been more cv oriented - but huge shakeup I want to understand a bit….
", "votes": null, "replies": [] }, { "id": 974515, "author_name": "mutantspore", "author_url": "", "post_date": "08/18/2020 01:05:35", "content": "My best “if only” score was 0.9456 putting me about 22nd
", "votes": null, "replies": [] }, { "id": 974524, "author_name": "romanweilguny", "author_url": "", "post_date": "08/18/2020 01:11:40", "content": "ok - it's over - and it's not easy to understand…
\nmy best would have been 0.9429 - about rank 70 - but no special reason to choose this one. Funny fact - this is Chris' score. Hard to understand why Chris did not score better. He did everything right.
If someone can explain the connection between private and public scores - pls do…
\n… too much metadata seems to harm - which is very much unexpected… Chris sees this different…
", "votes": null, "replies": [] }, { "id": 975558, "author_name": "romanweilguny", "author_url": "", "post_date": "08/18/2020 10:57:28", "content": "\n\nMy best “if only” score was 0.9456 putting me about 22nd
\n
a promise…
", "votes": null, "replies": [] } ] }, { "id": 975320, "author_name": "agentauers", "author_url": "", "post_date": "08/18/2020 08:42:22", "content": "Hello Bruce, as raddar already said, you are not alone.
\nI crashed down to place 1000, … and I really do not know why the shakeup did hit me so hard.
\nMy strategy was to ensemble as much models as possible without any selection based on public LB scores.
\nThe best submission scored private 0.9283 / public 0.9610 and consists of more than 100 models.
At some point in the competition (about 4 weeks ago) I was only one place behind chris deotte :-) and this without any dangerous selection of models based on LB scores.
\nNext time I will define fixed folds at the very beginning of the competition, and use cv scores for the whole pipeline, from modelselection to final ensembling and postprocessing.
\nThank you all! I really learned a lot in this competition from you all :-)
", "votes": null, "replies": [ { "id": 975549, "author_name": "romanweilguny", "author_url": "", "post_date": "08/18/2020 10:54:07", "content": "Hello AgentAuers - it really hit you hard (at least you can remember the rank easily) … I used your kernel a lot until Chris came up with his notebook - thx for that. I find the role of metadata very confusing. Had big lb ups with ensembling best image result and a good tabular submission with weights 0.7/0.3
\nBut my last experiments with ranked submissions even indicated less tabular data seems better (2nd private rank wrote, that he did not use tab data at all):
\nA: ranked 3 B555 ensembled (no good lb results - chosen by oof cv value around 0.92)
\nB: ranked best LB result (ensembled 3 B555 best lb result, best was 0.957)
\nC: ranked tab submission
\nResults:
\nweights 0.2 0.5 0.3: priv: 0.9095 public: 0.9346
\nweights 0.2 0.5 0.1: priv: 0.9324 public: 0.9541
\nweights 0.1 0.6 0.1: priv: 0.9301 public: 0.9543
\nweights 0.1 0.6 0.06: priv: 0.9336 public: 0.9568
\nthis confuses me…
\nlittle correction: ensemble A used some metadata for ensembling - calculated with the 'LGB/MLP/RIDGE STACKING Effnet OOF with META' notebook
", "votes": null, "replies": [] }, { "id": 981994, "author_name": "cdeotte", "author_url": "", "post_date": "08/22/2020 23:50:32", "content": "@agentauers Don't feel bad, I will always remember you as the MVP of this comp. Your notebook was the beginning of all strong models. And you invented aggressive TTA! Everyone benefited from your helpful contributions, creative models, and insights. My triple stratified notebook is just your notebook with the addition of KFold. All my offline models use your augmentation settings. I just varied the read size, crop size, external data, upsample, and EfficientNet backbone (and then maximized CV). I'm not sure why my private is higher than your private.
", "votes": null, "replies": [] } ] }, { "id": 981981, "author_name": "kyakovlev", "author_url": "", "post_date": "08/22/2020 23:17:00", "content": "As I said - all be good)))). Not silver (continue what you are doing and next silver or gold for sure), but congratulations with bronze medal.
", "votes": null, "replies": [] }, { "id": 874231, "author_name": "raddar", "author_url": "", "post_date": "06/04/2020 18:14:12", "content": "Bruce, you are not alone! I find this enjoyable as well - there is so many ways to approach this problem! \nGood luck!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "874189": "\n\nI just wanted to say that this has been one of the **most enjoyable competitions** I've been involved in on Kaggle. I would like to thank all those who have been extremely helpful by producing great datasets and those (often the same people) who have been suggesting approaches and ideas.\n\nWe have barely started and I've climbed up that leaderboard and slipped and climbed and slipped. \n\nI still have a lot to try and experiment with. I have Colabs working well for TPU (either TFRecords or JPEG), but I have to control my urge to see how I'm going by burning through my available Submissions to quickly.\n(yeah yeah... trust your CV lol)\n\nRemember my day job is stacking potatoes neatly so this is a huge outlet for me.\n\nthanks everyone (and Kaggle)", "874231": "Bruce, you are not alone! I find this enjoyable as well - there is so many ways to approach this problem! \nGood luck!", "974168": "And how do I feel now... with 5 hrs to go...\n\nNot at all impressed and depressed by the whole experience. \n\nThis is probably the first time I've actually seriously tried to win a competition that gives out medals. Sure I've done ok in some of the learning, playground, kudos type comps but I thought I'd be able to get a bronze medal at least.\n\nI also tried to team up but found that my \"use them or loose them\" approach to submissions messed that up. I did get 2 other requests to merge but of course, I had the same issue.\n\nI'd climb up 6 places and fall 26 when I went to work ... then repeat ... then repeat ...\n\nBut now, like most of us, I ended up on the wrong side of the last-minute high LB public submission and just got washed away in the flood of copy/paste that went on.\n\nMy wife has been pretty tolerant of me spending so much time on this competition as she understands my need for an intellectual escape from stacking potatoes (yes, I really do that) but after this, she'll be pretty sceptical of the whole endeavour.\n\nHopefully, my approach will see me through to a medal on the Private LB and yes, I have seen people rise hundreds of places in other competitions so I guess it's not over yet.\n\nI'd still like to thank Chris, AgentAuers and the many others who have given freely and openly to help beginners and provide insight, even the \"grumpy\" ones ;) . That aspect has been a total success in this competition. \n\nI have learnt a great deal, more than in any other of my 19 competitions, so I should just be thankful for that and move on. Motivation to do more may well depend on the result tonight though.\n\nthanks all", "974240": "> I thought I'd be able to get a bronze medal at least\n\nDon't judge yourself too quickly.\n\nI bet you will climb in private part (bronze for sure) and with so much hard work I would say silver+ (top 100).\n\nGood luck to you.", "974281": "Thinking about if I would bet to rank above you - I don't dare :)\n2 of my 3 selected submissions are well below your best result.", "974462": "my public best managed about rank 1100 - private best has been more cv oriented - but huge shakeup I want to understand a bit....", "974515": "My best “if only” score was 0.9456 putting me about 22nd", "974524": "ok - it's over - and it's not easy to understand...\nmy best would have been 0.9429 - about rank 70 - but no special reason to choose this one. Funny fact - this is Chris' score. Hard to understand why Chris did not score better. He did everything right.", "974556": "If someone can explain the connection between private and public scores - pls do...\n\n... too much metadata seems to harm - which is very much unexpected... Chris sees this different...", "975320": "Hello Bruce, as raddar already said, you are not alone.\nI crashed down to place 1000, ... and I really do not know why the shakeup did hit me so hard.\nMy strategy was to ensemble as much models as possible without any selection based on public LB scores. \nThe best submission scored private 0.9283 / public 0.9610 and consists of more than 100 models. \n\nAt some point in the competition (about 4 weeks ago) I was only one place behind chris deotte :-) and this without any dangerous selection of models based on LB scores.\n\nNext time I will define fixed folds at the very beginning of the competition, and use cv scores for the whole pipeline, from modelselection to final ensembling and postprocessing.\n\nThank you all! I really learned a lot in this competition from you all :-)", "975549": "Hello AgentAuers - it really hit you hard (at least you can remember the rank easily) ... I used your kernel a lot until Chris came up with his notebook - thx for that. I find the role of metadata very confusing. Had big lb ups with ensembling best image result and a good tabular submission with weights 0.7/0.3\n\nBut my last experiments with ranked submissions even indicated less tabular data seems better (2nd private rank wrote, that he did not use tab data at all):\n\nA: ranked 3 B555 ensembled (no good lb results - chosen by oof cv value around 0.92)\n\nB: ranked best LB result (ensembled 3 B555 best lb result, best was 0.957)\n\nC: ranked tab submission\n\nResults: \n\nweights 0.2 0.5 0.3: priv: 0.9095 public: 0.9346\n\nweights 0.2 0.5 0.1: priv: 0.9324 public: 0.9541\n\nweights 0.1 0.6 0.1: priv: 0.9301 public: 0.9543\n\nweights 0.1 0.6 0.06: priv: 0.9336 public: 0.9568\n\nthis confuses me...\n\nlittle correction: ensemble A used some metadata for ensembling - calculated with the 'LGB/MLP/RIDGE STACKING Effnet OOF with META' notebook", "975558": ">My best “if only” score was 0.9456 putting me about 22nd\n\na promise...", "981981": "As I said - all be good)))). Not silver (continue what you are doing and next silver or gold for sure), but congratulations with bronze medal.", "981994": "agentauers Don't feel bad, I will always remember you as the MVP of this comp. Your notebook was the beginning of all strong models. And you invented aggressive TTA! Everyone benefited from your helpful contributions, creative models, and insights. My triple stratified notebook is just your notebook with the addition of KFold. All my offline models use your augmentation settings. I just varied the read size, crop size, external data, upsample, and EfficientNet backbone (and then maximized CV). I'm not sure why my private is higher than your private." }, "source": "meta" }