I'm not sure to post this,
\nI may be wrong, or don't know,
There are a lot of great contributions in this compilation, which I really appreciate.
\nBut as a beginner, this is what I observed on LB, there are 3 categories of notebooks:
| Category I | \nCategory II | \nCategory III | \n
|---|---|---|
| - Gives approach and method to implement solutions. | \n- Improve existing approach. | \n- Simply copy the notebook from category 2, run more epoch, change the size, and Just make the model public. | \n
| - Helps new people to get on Leaderboard | \n- Improve LB scores, insights. | \n- No Insights, Nothing to do with improvement. | \n
| - Ex: Yolov5 notebook by @awsaf49, YoloR YoloX notebook @remekkinas, Other model notebooks, etc. | \n- Ex: YoloX + Tracking by @parapapapam, @steamedsheep Yolov5 High-Resolution image, Lot Of EDA notebook, Balance Fold Notebooks, etc. | \n- 🤐 | \n
Many people, spend a lot of time doing experiments and manage to get in the top 100-150 with their own solution, But Category III Notebook without any insights, Make LB worst.
", "messages": [ { "id": 1651373, "postDate": "2022-01-15T17:53:40.983Z", "content": "I'm not sure to post this,
\nI may be wrong, or don't know,
There are a lot of great contributions in this compilation, which I really appreciate.
\nBut as a beginner, this is what I observed on LB, there are 3 categories of notebooks:
| Category I | \nCategory II | \nCategory III | \n
|---|---|---|
| - Gives approach and method to implement solutions. | \n- Improve existing approach. | \n- Simply copy the notebook from category 2, run more epoch, change the size, and Just make the model public. | \n
| - Helps new people to get on Leaderboard | \n- Improve LB scores, insights. | \n- No Insights, Nothing to do with improvement. | \n
| - Ex: Yolov5 notebook by @awsaf49, YoloR YoloX notebook @remekkinas, Other model notebooks, etc. | \n- Ex: YoloX + Tracking by @parapapapam, @steamedsheep Yolov5 High-Resolution image, Lot Of EDA notebook, Balance Fold Notebooks, etc. | \n- 🤐 | \n
Many people, spend a lot of time doing experiments and manage to get in the top 100-150 with their own solution, But Category III Notebook without any insights, Make LB worst.
", "rawMarkdown": "I'm not sure to post this,\nI may be wrong, or don't know,\n\nThere are a lot of great contributions in this compilation, which I really appreciate. \nBut as a beginner, this is what I observed on LB, there are 3 categories of notebooks:\n| Category I | Category II | Category III\n| --- | --- |\n| - Gives approach and method to implement solutions. | - Improve existing approach. | - Simply copy the notebook from category 2, run more epoch, change the size, and **Just make the model public**.\n| - Helps new people to get on Leaderboard | - Improve LB scores, insights. | - No Insights, Nothing to do with improvement. \n| - Ex: Yolov5 notebook by @awsaf49, YoloR YoloX notebook @remekkinas, Other model notebooks, etc. | - Ex: YoloX + Tracking by @parapapapam, @steamedsheep Yolov5 High-Resolution image, Lot Of EDA notebook, Balance Fold Notebooks, etc. | - 🤐\n\nMany people, spend a lot of time doing experiments and manage to get in the top 100-150 with their own solution, But Category III Notebook without any insights, Make LB worst.\n", "votes": 16 }, { "id": 1651918, "postDate": "2022-01-16T07:27:23.240Z", "content": "Kind of thinking to give up this competition. Spend one and half month to beat the 0.539 LB solution. But it comes another 0.613 LB solution. Then I should start from the beginning again and have no idea my previous pipeline will work or not.
", "rawMarkdown": "Kind of thinking to give up this competition. Spend one and half month to beat the 0.539 LB solution. But it comes another 0.613 LB solution. Then I should start from the beginning again and have no idea my previous pipeline will work or not.", "votes": 9, "replies": [ { "id": 1652116, "postDate": "2022-01-16T11:00:46.107Z", "content": "I encourage you to stay in competition and learn. I am sure … it will be rewarded in the future. You can't control others having 8xGPU or people who share just after 2 months of competition… but you can make progress and improve your solution and enjoy the progress.
\nI was thinking a little bit about this situation. It is hard :) unfortunately. I understand people who shares solution and brilliant tricks … but …
\nFrom my point of view sharing solutions at the beggining of competition is really more then welcome. This is why we decided to share YoloX, YoloR … (we have not shared Yolo5 because it was published) - to enable people to easy and fast start (we give people rod … not fish).
\nSharing key parts of solution after 2 months of hard team work is just …. not ok (here we have prized competition contrary to …. many competition where goal is to share from start to end eg. TPS where you can't earn money and score points). This is my personal point of view. Sorry for that. It is only partially ok for person who shares. Why partially? He/she certainly got a lot of points … but usually other will use this solution better (this is my experience). Second … this gave people fish …. not …. net/rod so this won't improve people skills of solving problem (even they are really happy about \"progress\" they made …. unfortunately only on LB). Last point it ruins work (hours of work) people who truly were look for solution.
\nBut … I understand. This is competition. No regulations when to share, when not. But we have to think about our move consequences …
", "rawMarkdown": "I encourage you to stay in competition and learn. I am sure ... it will be rewarded in the future. You can't control others having 8xGPU or people who share just after 2 months of competition... but you can make progress and improve your solution and enjoy the progress. \n\nI was thinking a little bit about this situation. It is hard :) unfortunately. I understand people who shares solution and brilliant tricks ... but ...\n\nFrom my point of view sharing solutions at the beggining of competition is really more then welcome. This is why we decided to share YoloX, YoloR ... (we have not shared Yolo5 because it was published) - to enable people to easy and fast start (we give people rod ... not fish).\n\nSharing key parts of solution after 2 months of hard team work is just .... not ok (here we have prized competition contrary to .... many competition where goal is to share from start to end eg. TPS where you can't earn money and score points). This is my personal point of view. Sorry for that. It is only partially ok for person who shares. Why partially? He/she certainly got a lot of points ... but usually other will use this solution better (this is my experience). Second ... this gave people fish .... not .... net/rod so this won't improve people skills of solving problem (even they are really happy about \"progress\" they made .... unfortunately only on LB). Last point it ruins work (hours of work) people who truly were look for solution.\n\nBut ... I understand. This is competition. No regulations when to share, when not. But we have to think about our move consequences ... ", "votes": 18 }, { "id": 1652122, "postDate": "2022-01-16T11:10:47.950Z", "content": "I have never been a fan of this spoon feeding on Kaggle. I also highly doubt that it will lead to much learning of participants. My personal experience is that you learn the most, when you start with a baseline, and then try to improve that baseline by yourself.
\nSo all the sharing you guys did @remekkinas is very valuable, i.e. sharing baseline scripts that get you starting, but that are no free-lunch to high LB scores.
\nThis trend of submitting high scoring LB kernels, is something I also cannot appreciate.
\nIt is really a pain and demotivating for many kagglers, who struggled hard to go over let's say 0.5 on LB, and suddenly they are rank 500 instead of 50-100.
\nSo to summarize, my personal opinion is that sharing baseline scripts is valuable during competition and a lot of detailed sharing can and should happen after competition ends. You have so many past object detection competitions, where you can read up on all the solutions and details and you will have the same details available (hopefully) here after the competition ends.
", "rawMarkdown": "I have never been a fan of this spoon feeding on Kaggle. I also highly doubt that it will lead to much learning of participants. My personal experience is that you learn the most, when you start with a baseline, and then try to improve that baseline by yourself.\n\nSo all the sharing you guys did @remekkinas is very valuable, i.e. sharing baseline scripts that get you starting, but that are no free-lunch to high LB scores.\n\nThis trend of submitting high scoring LB kernels, is something I also cannot appreciate. \n\nIt is really a pain and demotivating for many kagglers, who struggled hard to go over let's say 0.5 on LB, and suddenly they are rank 500 instead of 50-100.\n\nSo to summarize, my personal opinion is that sharing baseline scripts is valuable during competition and a lot of detailed sharing can and should happen after competition ends. You have so many past object detection competitions, where you can read up on all the solutions and details and you will have the same details available (hopefully) here after the competition ends.\n", "votes": 24 }, { "id": 1652125, "postDate": "2022-01-16T11:13:34.940Z", "content": "@philippsinger to be well understood … I am cheering for you because I can see the work you have done … absolutely fantastic (I can see the number of submits). I keep my fingers crossed and I can't wait to publish your solution after the end of the competition. 👍💪💪🤛
", "rawMarkdown": "@philippsinger to be well understood ... I am cheering for you because I can see the work you have done ... absolutely fantastic (I can see the number of submits). I keep my fingers crossed and I can't wait to publish your solution after the end of the competition. 👍💪💪🤛", "votes": 3 }, { "id": 1652346, "postDate": "2022-01-16T14:26:04.053Z", "content": "@remekkinas Thanks for your encouragement. C'est la vie. We can't tell others not to share at such a moment. I can only all in myself and keep focus on the good things and positive side.
\nThis is my first Object Detection competition. So far, I really learned a lot from your notebooks, all the SOTA papers and open source libs.
", "rawMarkdown": "@remekkinas Thanks for your encouragement. C'est la vie. We can't tell others not to share at such a moment. I can only all in myself and keep focus on the good things and positive side.\n\nThis is my first Object Detection competition. So far, I really learned a lot from your notebooks, all the SOTA papers and open source libs.", "votes": 2 } ] }, { "id": 1685877, "postDate": "2022-02-11T15:40:38.357Z", "content": "Feeling so defeated when kernel that outperformed all my experiments in 2 months got published only in few days before ending.
", "rawMarkdown": "Feeling so defeated when kernel that outperformed all my experiments in 2 months got published only in few days before ending. ", "votes": 7 }, { "id": 1651398, "postDate": "2022-01-15T18:12:42.943Z", "content": "Thank you! 🙏
\nBTW: yoloR - today I published update - 0.521 :) butI am sure you can go higher.
", "rawMarkdown": "Thank you! 🙏 \n\nBTW: yoloR - today I published update - 0.521 :) butI am sure you can go higher.", "votes": 4, "replies": [ { "id": 1651786, "postDate": "2022-01-16T04:48:43.657Z", "content": "Sure, I'll try 👍
", "rawMarkdown": "Sure, I'll try 👍" } ] }, { "id": 1682549, "postDate": "2022-02-09T08:37:57.813Z", "content": "It's frustrating day by day in Kaggle now because of these scenarios. I and my team member have joined this competition from the beginning of the competition and submitted already 150+. The current LB scenario and high LB notebook sharing are totally frustrating to me, and happening in every Kaggle competition.
", "rawMarkdown": "It's frustrating day by day in Kaggle now because of these scenarios. I and my team member have joined this competition from the beginning of the competition and submitted already 150+. The current LB scenario and high LB notebook sharing are totally frustrating to me, and happening in every Kaggle competition. ", "votes": 1, "replies": [ { "id": 1682571, "postDate": "2022-02-09T08:51:00.227Z", "content": "Yes, this could be frustrating. We have made …. 372 submits 😂😜 looking for solution. But … I am happy even we dropped from gold …. Why?
\nNow it is the most important part - not to overfitt and choose the best solution - in this competition is kind of lottery since I do not trust local f2 score (it is hard to evaluate due to unknown test distribution and starfish characteristic)
We contributed a lot …. As a summary our team managed to publish ….
and finally we have learned a lot …. So we feel we good …. This is not only about winning in competition 😁
", "rawMarkdown": "Yes, this could be frustrating. We have made .... 372 submits 😂😜 looking for solution. But ... I am happy even we dropped from gold .... Why? \n1. Now it is the most important part - not to overfitt and choose the best solution - in this competition is kind of lottery since I do not trust local f2 score (it is hard to evaluate due to unknown test distribution and starfish characteristic)\n\n2. We contributed a lot .... As a summary our team managed to publish .... \n* 5 gold notebooks (YoloX/YoloX, YoloR/YoloR, SAHI)\n* 3 silver notebooks (GAN Training, GAN generator, Nueral Style Transfer)\n\nand finally we have learned a lot .... So we feel we good .... This is not only about winning in competition 😁 ", "votes": 7 }, { "id": 1682581, "postDate": "2022-02-09T08:58:17.140Z", "content": "I may miss my first bronze medal…because of those kinds of notebooks, trying my best, fingers crossed!
", "rawMarkdown": "I may miss my first bronze medal...because of those kinds of notebooks, trying my best, fingers crossed!", "votes": 1 } ] }, { "id": 1651778, "postDate": "2022-01-16T04:34:37.657Z", "content": "I think those notebooks are toxic, But we can do nothing. It is happening in every kaggle competition.
\nBTW, The LB in Jigsaw, Homework and this game was ruined in a week. 👀
", "rawMarkdown": "I think those notebooks are toxic, But we can do nothing. It is happening in every kaggle competition. \n\nBTW, The LB in Jigsaw, Homework and this game was ruined in a week. 👀", "votes": 2, "replies": [ { "id": 1651788, "postDate": "2022-01-16T04:50:58.270Z", "content": "Agree, I participated in Jigsaw competition, those notebooks are everywhere…
", "rawMarkdown": "Agree, I participated in Jigsaw competition, those notebooks are everywhere... ", "votes": 2 } ] }, { "id": 1651395, "postDate": "2022-01-15T18:10:20.917Z", "content": "Im also not sure what is happening on the LB, after the High-Resolution approach got public, within two days there is a drastic change in the LB. Dont know how people are getting that LB scores, even in that short amount of time🤔, by increasing the res of the img higher and higher, or something clicked for a new approach of models or everybody started taking the comp very very seriously.
", "rawMarkdown": "Im also not sure what is happening on the LB, after the High-Resolution approach got public, within two days there is a drastic change in the LB. Dont know how people are getting that LB scores, even in that short amount of time🤔, by increasing the res of the img higher and higher, or something clicked for a new approach of models or everybody started taking the comp very very seriously.", "votes": 2, "replies": [ { "id": 1651407, "postDate": "2022-01-15T18:18:42.597Z", "content": "Exactly., High res image give high LB scores, now people with good GPU just making their model public, and LB getting the worst😪
", "rawMarkdown": "Exactly., High res image give high LB scores, now people with good GPU just making their model public, and LB getting the worst😪" } ] }, { "id": 1651405, "postDate": "2022-01-15T18:16:39.860Z", "rawMarkdown": "", "votes": -8, "isDeleted": true } ], "comments": [ { "id": 1651918, "author_name": "DeepInvolution", "author_url": "", "post_date": "2022-01-16T07:27:23.240000", "content": "Kind of thinking to give up this competition. Spend one and half month to beat the 0.539 LB solution. But it comes another 0.613 LB solution. Then I should start from the beginning again and have no idea my previous pipeline will work or not.
", "votes": 9, "replies": [ { "id": 1652116, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-16T11:00:46.107000", "content": "I encourage you to stay in competition and learn. I am sure … it will be rewarded in the future. You can't control others having 8xGPU or people who share just after 2 months of competition… but you can make progress and improve your solution and enjoy the progress.
\nI was thinking a little bit about this situation. It is hard :) unfortunately. I understand people who shares solution and brilliant tricks … but …
\nFrom my point of view sharing solutions at the beggining of competition is really more then welcome. This is why we decided to share YoloX, YoloR … (we have not shared Yolo5 because it was published) - to enable people to easy and fast start (we give people rod … not fish).
\nSharing key parts of solution after 2 months of hard team work is just …. not ok (here we have prized competition contrary to …. many competition where goal is to share from start to end eg. TPS where you can't earn money and score points). This is my personal point of view. Sorry for that. It is only partially ok for person who shares. Why partially? He/she certainly got a lot of points … but usually other will use this solution better (this is my experience). Second … this gave people fish …. not …. net/rod so this won't improve people skills of solving problem (even they are really happy about \"progress\" they made …. unfortunately only on LB). Last point it ruins work (hours of work) people who truly were look for solution.
\nBut … I understand. This is competition. No regulations when to share, when not. But we have to think about our move consequences …
", "votes": 18, "replies": [] }, { "id": 1652122, "author_name": "Psi", "author_url": "", "post_date": "2022-01-16T11:10:47.950000", "content": "I have never been a fan of this spoon feeding on Kaggle. I also highly doubt that it will lead to much learning of participants. My personal experience is that you learn the most, when you start with a baseline, and then try to improve that baseline by yourself.
\nSo all the sharing you guys did @remekkinas is very valuable, i.e. sharing baseline scripts that get you starting, but that are no free-lunch to high LB scores.
\nThis trend of submitting high scoring LB kernels, is something I also cannot appreciate.
\nIt is really a pain and demotivating for many kagglers, who struggled hard to go over let's say 0.5 on LB, and suddenly they are rank 500 instead of 50-100.
\nSo to summarize, my personal opinion is that sharing baseline scripts is valuable during competition and a lot of detailed sharing can and should happen after competition ends. You have so many past object detection competitions, where you can read up on all the solutions and details and you will have the same details available (hopefully) here after the competition ends.
", "votes": 24, "replies": [] }, { "id": 1652125, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-16T11:13:34.940000", "content": "@philippsinger to be well understood … I am cheering for you because I can see the work you have done … absolutely fantastic (I can see the number of submits). I keep my fingers crossed and I can't wait to publish your solution after the end of the competition. 👍💪💪🤛
", "votes": 3, "replies": [] }, { "id": 1652346, "author_name": "DeepInvolution", "author_url": "", "post_date": "2022-01-16T14:26:04.053000", "content": "@remekkinas Thanks for your encouragement. C'est la vie. We can't tell others not to share at such a moment. I can only all in myself and keep focus on the good things and positive side.
\nThis is my first Object Detection competition. So far, I really learned a lot from your notebooks, all the SOTA papers and open source libs.
", "votes": 2, "replies": [] } ] }, { "id": 1685877, "author_name": "Bao Loc Pham", "author_url": "", "post_date": "2022-02-11T15:40:38.357000", "content": "Feeling so defeated when kernel that outperformed all my experiments in 2 months got published only in few days before ending.
", "votes": 7, "replies": [] }, { "id": 1651398, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-15T18:12:42.943000", "content": "Thank you! 🙏
\nBTW: yoloR - today I published update - 0.521 :) butI am sure you can go higher.
", "votes": 4, "replies": [ { "id": 1651786, "author_name": "Mahipal Singh", "author_url": "", "post_date": "2022-01-16T04:48:43.657000", "content": "Sure, I'll try 👍
", "votes": 0, "replies": [] } ] }, { "id": 1682549, "author_name": "Townim Faisal", "author_url": "", "post_date": "2022-02-09T08:37:57.813000", "content": "It's frustrating day by day in Kaggle now because of these scenarios. I and my team member have joined this competition from the beginning of the competition and submitted already 150+. The current LB scenario and high LB notebook sharing are totally frustrating to me, and happening in every Kaggle competition.
", "votes": 1, "replies": [ { "id": 1682571, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-02-09T08:51:00.227000", "content": "Yes, this could be frustrating. We have made …. 372 submits 😂😜 looking for solution. But … I am happy even we dropped from gold …. Why?
\nNow it is the most important part - not to overfitt and choose the best solution - in this competition is kind of lottery since I do not trust local f2 score (it is hard to evaluate due to unknown test distribution and starfish characteristic)
We contributed a lot …. As a summary our team managed to publish ….
and finally we have learned a lot …. So we feel we good …. This is not only about winning in competition 😁
", "votes": 7, "replies": [] }, { "id": 1682581, "author_name": "Mahipal Singh", "author_url": "", "post_date": "2022-02-09T08:58:17.140000", "content": "I may miss my first bronze medal…because of those kinds of notebooks, trying my best, fingers crossed!
", "votes": 1, "replies": [] } ] }, { "id": 1651778, "author_name": "ForcewithMe", "author_url": "", "post_date": "2022-01-16T04:34:37.657000", "content": "I think those notebooks are toxic, But we can do nothing. It is happening in every kaggle competition.
\nBTW, The LB in Jigsaw, Homework and this game was ruined in a week. 👀
", "votes": 2, "replies": [ { "id": 1651788, "author_name": "Mahipal Singh", "author_url": "", "post_date": "2022-01-16T04:50:58.270000", "content": "Agree, I participated in Jigsaw competition, those notebooks are everywhere…
", "votes": 2, "replies": [] } ] }, { "id": 1651395, "author_name": "somuSan", "author_url": "", "post_date": "2022-01-15T18:10:20.917000", "content": "Im also not sure what is happening on the LB, after the High-Resolution approach got public, within two days there is a drastic change in the LB. Dont know how people are getting that LB scores, even in that short amount of time🤔, by increasing the res of the img higher and higher, or something clicked for a new approach of models or everybody started taking the comp very very seriously.
", "votes": 2, "replies": [ { "id": 1651407, "author_name": "Mahipal Singh", "author_url": "", "post_date": "2022-01-15T18:18:42.597000", "content": "Exactly., High res image give high LB scores, now people with good GPU just making their model public, and LB getting the worst😪
", "votes": 0, "replies": [] } ] }, { "id": 1651405, "author_name": "", "author_url": "", "post_date": "2022-01-15T18:16:39.860000", "content": "", "votes": -8, "replies": [] } ], "raw_markdown_by_id": { "1651373": "I'm not sure to post this,\nI may be wrong, or don't know,\n\nThere are a lot of great contributions in this compilation, which I really appreciate. \nBut as a beginner, this is what I observed on LB, there are 3 categories of notebooks:\n| Category I | Category II | Category III\n| --- | --- |\n| - Gives approach and method to implement solutions. | - Improve existing approach. | - Simply copy the notebook from category 2, run more epoch, change the size, and **Just make the model public**.\n| - Helps new people to get on Leaderboard | - Improve LB scores, insights. | - No Insights, Nothing to do with improvement. \n| - Ex: Yolov5 notebook by @awsaf49, YoloR YoloX notebook @remekkinas, Other model notebooks, etc. | - Ex: YoloX + Tracking by @parapapapam, @steamedsheep Yolov5 High-Resolution image, Lot Of EDA notebook, Balance Fold Notebooks, etc. | - 🤐\n\nMany people, spend a lot of time doing experiments and manage to get in the top 100-150 with their own solution, But Category III Notebook without any insights, Make LB worst.\n", "1651918": "Kind of thinking to give up this competition. Spend one and half month to beat the 0.539 LB solution. But it comes another 0.613 LB solution. Then I should start from the beginning again and have no idea my previous pipeline will work or not.", "1685877": "Feeling so defeated when kernel that outperformed all my experiments in 2 months got published only in few days before ending. ", "1651398": "Thank you! 🙏 \n\nBTW: yoloR - today I published update - 0.521 :) butI am sure you can go higher.", "1682549": "It's frustrating day by day in Kaggle now because of these scenarios. I and my team member have joined this competition from the beginning of the competition and submitted already 150+. The current LB scenario and high LB notebook sharing are totally frustrating to me, and happening in every Kaggle competition. ", "1651778": "I think those notebooks are toxic, But we can do nothing. It is happening in every kaggle competition. \n\nBTW, The LB in Jigsaw, Homework and this game was ruined in a week. 👀", "1651395": "Im also not sure what is happening on the LB, after the High-Resolution approach got public, within two days there is a drastic change in the LB. Dont know how people are getting that LB scores, even in that short amount of time🤔, by increasing the res of the img higher and higher, or something clicked for a new approach of models or everybody started taking the comp very very seriously.", "1651405": "" } }