Hi all , I am afraid that is not related to the competition much but its my gratitude towards the community and all the Kaggle greats from whom I learn everyday.
\nPeople always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world . While this statement might be partly true , I strongly disagree , Kaggle allows you to compete against the best , learn from the best , and make you push your limits and this in turn prepares you for all the challenges in the real world . While Kaggle competitions might not simulate real world challenges every time , it teaches a lot of lessons , it makes you more confident , makes you think on your feet .
\nI have been working extra hard from past one month on this competition , learning techniques like spatial attention , Knowledge transfer models ,etc and still I am struggling to cross 0.970 and a team of greats (the elite club) comes in and reaches 2nd position with just 10 submissions , it just makes you wonder how tough it is to compete in a Kaggle competition
\nI feel very lucky that I get the chance to even compete with such great people . Everything which I have learned has been mostly through Kaggle and I hope I keep learning . It's a beautiful stage for every young data scientists like me . No matter how much challenges I face Kaggle pushes me to keep toiling.
\nThanks for taking out the time to read this and all the best for this competition
", "messages": [ { "id": "1233773", "postDate": "03/10/2021 16:44:55", "content": "Hi all , I am afraid that is not related to the competition much but its my gratitude towards the community and all the Kaggle greats from whom I learn everyday.
\nPeople always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world . While this statement might be partly true , I strongly disagree , Kaggle allows you to compete against the best , learn from the best , and make you push your limits and this in turn prepares you for all the challenges in the real world . While Kaggle competitions might not simulate real world challenges every time , it teaches a lot of lessons , it makes you more confident , makes you think on your feet .
\nI have been working extra hard from past one month on this competition , learning techniques like spatial attention , Knowledge transfer models ,etc and still I am struggling to cross 0.970 and a team of greats (the elite club) comes in and reaches 2nd position with just 10 submissions , it just makes you wonder how tough it is to compete in a Kaggle competition
\nI feel very lucky that I get the chance to even compete with such great people . Everything which I have learned has been mostly through Kaggle and I hope I keep learning . It's a beautiful stage for every young data scientists like me . No matter how much challenges I face Kaggle pushes me to keep toiling.
\nThanks for taking out the time to read this and all the best for this competition
", "rawMarkdown": "Hi all , I am afraid that is not related to the competition much but its my gratitude towards the community and all the Kaggle greats from whom I learn everyday.\n\nPeople always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world . While this statement might be partly true , I strongly disagree , Kaggle allows you to compete against the best , learn from the best , and make you push your limits and this in turn prepares you for all the challenges in the real world . While Kaggle competitions might not simulate real world challenges every time , it teaches a lot of lessons , it makes you more confident , makes you think on your feet .\n\nI have been working extra hard from past one month on this competition , learning techniques like spatial attention , Knowledge transfer models ,etc and still I am struggling to cross 0.970 and a team of greats (the elite club) comes in and reaches 2nd position with just 10 submissions , it just makes you wonder how tough it is to compete in a Kaggle competition\n\nI feel very lucky that I get the chance to even compete with such great people . Everything which I have learned has been mostly through Kaggle and I hope I keep learning . It's a beautiful stage for every young data scientists like me . No matter how much challenges I face Kaggle pushes me to keep toiling.\n\nThanks for taking out the time to read this and all the best for this competition", "votes": null }, { "id": "1233839", "postDate": "03/10/2021 18:01:11", "content": "I was just wondering why I couldn't find @haqishen on leaderboard in a computer vision competition and here they (team of elites) are 2nd on LB with just 10 subs.
", "rawMarkdown": "I was just wondering why I couldn't find @haqishen on leaderboard in a computer vision competition and here they (team of elites) are 2nd on LB with just 10 subs.", "votes": null }, { "id": "1234176", "postDate": "03/11/2021 03:13:12", "content": "Don't be discouraged guys.
\nWe also worked very hard on the past month, tried bunch of ideas, ran hundreds of experiments.
\nBut as we find that our CV socre is very well correlated to public LB in less than 8 submissions a month ago, we decided NOT to submit until the end of this competition.
", "rawMarkdown": "Don't be discouraged guys.\n\nWe also worked very hard on the past month, tried bunch of ideas, ran hundreds of experiments.\n\nBut as we find that our CV socre is very well correlated to public LB in less than 8 submissions a month ago, we decided NOT to submit until the end of this competition.", "votes": null }, { "id": "1234442", "postDate": "03/11/2021 08:51:46", "content": "\" it just makes you wonder how tough it is to compete in a Kaggle competition\"
\n\"turn prepares you for all the challenges in the real world\"
\n\" While Kaggle competitions might not simulate real world challenges every time β¦\"
let's relates to the real world!
\n\"We also worked very hard on the past month, tried bunch of ideas, ran hundreds of experiments.\"
\nAs a computer vision / deep learning algorithm developer in some commercial company, this is the plan that i need to submit to my customer, i.e. funding party, to bid for a project:
\n[expected performance of model: 95%]\n[expected milestone :]\na. 07-jun : 80% (min viable model ) -- first payment xx%\nb. 07-jul : 90% (first delivery) -- ....\n....\n\ne. 07-sep : ....\n\n\n[resource management]\nexpected number of models trained per day: 5 \ntotal number of models trained in milestone.a = 50 (8xGPU, expected utilisation >90%)\ntotal number of models trained in milestone.b = 50 (8xGPU, expected utilisation >90%)\n...\ntotal number of models trained in milestone.e = 50 (8xGPU, expected utilisation >90%)\n\n\n[pricing and speedup]\n... we can speedup results by increasing the number of experiments for 5 to 10. but this requires us to increase our gpu resource from 8 to 16 and this incur cost of $$ xxx to the prroject... \n\n... we think we can improve performance if we have 10x more data at the cost of $$yyy ...\n\n... this project is special because it requires re-design of any existing model in the market. we need xx data scientists for a period of xx months, incurring cost of $$ xxx\ntake home message:
\nwhen you are working in Kaggle or commerical project, you are not only deciding on how to develop algorithms but also how to schedule for GPU resources and training experiments.
no one can be sure if a modification will make an algorithm works. But he must be about to make the assessment that : \" by making at least x number of modifications and experiments, one can be sure that there will be improvement of at least zz% with probability 95% chance of success\"
\n\"People always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world\"
\nin some commercial applications +0.00001 means better price of your AI product. e.g. face recognition, prediction of airplane fuel usage
\n** YES! DATA SCIENCE is sometimes VERY TOUGH WORK !!! **
", "rawMarkdown": "\" it just makes you wonder how tough it is to compete in a Kaggle competition\"\n\"turn prepares you for all the challenges in the real world\"\n\" While Kaggle competitions might not simulate real world challenges every time ...\"\n\nlet's relates to the real world!\n\n---\n\"We also worked very hard on the past month, tried bunch of ideas, ran hundreds of experiments.\"\n\n---\n\nAs a computer vision / deep learning algorithm developer in some commercial company, this is the plan that i need to submit to my customer, i.e. funding party, to bid for a project:\n\n```\n[expected performance of model: 95%]\n[expected milestone :]\na. 07-jun : 80% (min viable model ) -- first payment xx%\nb. 07-jul : 90% (first delivery) -- ....\n....\n\ne. 07-sep : ....\n\n\n[resource management]\nexpected number of models trained per day: 5 \ntotal number of models trained in milestone.a = 50 (8xGPU, expected utilisation >90%)\ntotal number of models trained in milestone.b = 50 (8xGPU, expected utilisation >90%)\n...\ntotal number of models trained in milestone.e = 50 (8xGPU, expected utilisation >90%)\n\n\n[pricing and speedup]\n... we can speedup results by increasing the number of experiments for 5 to 10. but this requires us to increase our gpu resource from 8 to 16 and this incur cost of $$ xxx to the prroject... \n\n... we think we can improve performance if we have 10x more data at the cost of $$yyy ...\n\n... this project is special because it requires re-design of any existing model in the market. we need xx data scientists for a period of xx months, incurring cost of $$ xxx\n\n```\n\n\ntake home message:\nwhen you are working in Kaggle or commerical project, you are not only deciding on how to develop algorithms but also how to schedule for GPU resources and training experiments.\n\nno one can be sure if a modification will make an algorithm works. But he must be about to make the assessment that : \" by making at least x number of modifications and experiments, one can be sure that there will be improvement of at least zz% with probability 95% chance of success\"\n\n---\n\n\"People always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world\"\n\nin some commercial applications +0.00001 means better price of your AI product. e.g. face recognition, prediction of airplane fuel usage\n\n** YES! DATA SCIENCE is sometimes VERY TOUGH WORK !!! **", "votes": null }, { "id": "1234637", "postDate": "03/11/2021 12:37:59", "content": "@hengck23 very well summarized
\nDATA SCIENCE is sometimes VERY TOUGH WORK
Isn't this the reason that makes it so beautiful . @haqishen not getting discouraged afterall we all will learn from the solution you publish after the competition
", "rawMarkdown": "hengck23 very well summarized \n\n` DATA SCIENCE is sometimes VERY TOUGH WORK`\n\nIsn't this the reason that makes it so beautiful . @haqishen not getting discouraged afterall we all will learn from the solution you publish after the competition", "votes": null }, { "id": "1234849", "postDate": "03/11/2021 16:07:56", "content": "\n\nPeople always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001.
\n
Yes i feel the same, i used to wonder whats the point doing 0.001 improvements. If you feel the same you're right. I have made more than 100 submissions on my last competition and it sucks, it was wrong. It was a waste of time. Kaggle is place of experiments and learning not to fight for 0.001 percent. You don't know much come you to kaggle learn from others, if you know something you experiment it on kaggle that makes more sense and not training the same thing for 0.001 improvement. I guess the medal thing is what keeps us hooked. π
", "rawMarkdown": "> People always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001.\n\n\nYes i feel the same, i used to wonder whats the point doing 0.001 improvements. If you feel the same you're right. I have made more than 100 submissions on my last competition and it sucks, it was wrong. It was a waste of time. Kaggle is place of experiments and learning not to fight for 0.001 percent. You don't know much come you to kaggle learn from others, if you know something you experiment it on kaggle that makes more sense and not training the same thing for 0.001 improvement. I guess the medal thing is what keeps us hooked. π", "votes": null }, { "id": "1234869", "postDate": "03/11/2021 16:29:44", "content": "Also dott moving into solo gold position!!! Unreal. π
", "rawMarkdown": "Also dott moving into solo gold position!!! Unreal. π", "votes": null }, { "id": "1235332", "postDate": "03/12/2021 04:10:57", "content": "I also want to express my opinion.
\nPeople always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world.
This is true to us, competitors, but it is maybe not true to the organizer. What we consider expensive today might not be expensive tomorrow. GPU was for the top companies decades ago, but now we have GPU (and even TPU) freely available to everyone. That being said, in the future, computational cost will decrease, and it will not remain a problem.
\nMoreover, 0.001-0.00001 might not be important to us, but it is important to the organizer. The number is small, but it might have large impact. Letβs say you want to use hard labels, then there is a very big difference between 0.499999 and 0.5: the decision to do or not to do it. This margin has significant important in medicine or financial trading. A patient with false positive only results in additional medical treatment, but with a false negative they can die.
\nIf you are hear to learn, then score is not a big thing, but if you are here to compete, then you must adapt to compete with the best around the world.
", "rawMarkdown": "I also want to express my opinion.\n`People always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world.`\n\nThis is true to us, competitors, but it is maybe not true to the organizer. What we consider expensive today might not be expensive tomorrow. GPU was for the top companies decades ago, but now we have GPU (and even TPU) freely available to everyone. That being said, in the future, computational cost will decrease, and it will not remain a problem.\n\nMoreover, 0.001-0.00001 might not be important to us, but it is important to the organizer. The number is small, but it might have large impact. Letβs say you want to use hard labels, then there is a very big difference between 0.499999 and 0.5: the decision to do or not to do it. This margin has significant important in medicine or financial trading. A patient with false positive only results in additional medical treatment, but with a false negative they can die.\n\nIf you are hear to learn, then score is not a big thing, but if you are here to compete, then you must adapt to compete with the best around the world.", "votes": null }, { "id": "1235440", "postDate": "03/12/2021 06:19:37", "content": "Hahaha True
", "rawMarkdown": "Hahaha True", "votes": null }, { "id": "1235441", "postDate": "03/12/2021 06:21:30", "content": "Well said and these are some of the reasons I strongly disagree with people criticizing kaggle competitions
", "rawMarkdown": "Well said and these are some of the reasons I strongly disagree with people criticizing kaggle competitions", "votes": null } ], "comments": [ { "id": 1233839, "author_name": "rsinda", "author_url": "", "post_date": "03/10/2021 18:01:11", "content": "I was just wondering why I couldn't find @haqishen on leaderboard in a computer vision competition and here they (team of elites) are 2nd on LB with just 10 subs.
", "votes": null, "replies": [] }, { "id": 1234176, "author_name": "haqishen", "author_url": "", "post_date": "03/11/2021 03:13:12", "content": "Don't be discouraged guys.
\nWe also worked very hard on the past month, tried bunch of ideas, ran hundreds of experiments.
\nBut as we find that our CV socre is very well correlated to public LB in less than 8 submissions a month ago, we decided NOT to submit until the end of this competition.
", "votes": null, "replies": [ { "id": 1234442, "author_name": "hengck23", "author_url": "", "post_date": "03/11/2021 08:51:46", "content": "\" it just makes you wonder how tough it is to compete in a Kaggle competition\"
\n\"turn prepares you for all the challenges in the real world\"
\n\" While Kaggle competitions might not simulate real world challenges every time β¦\"
let's relates to the real world!
\n\"We also worked very hard on the past month, tried bunch of ideas, ran hundreds of experiments.\"
\nAs a computer vision / deep learning algorithm developer in some commercial company, this is the plan that i need to submit to my customer, i.e. funding party, to bid for a project:
\n[expected performance of model: 95%]\n[expected milestone :]\na. 07-jun : 80% (min viable model ) -- first payment xx%\nb. 07-jul : 90% (first delivery) -- ....\n....\n\ne. 07-sep : ....\n\n\n[resource management]\nexpected number of models trained per day: 5 \ntotal number of models trained in milestone.a = 50 (8xGPU, expected utilisation >90%)\ntotal number of models trained in milestone.b = 50 (8xGPU, expected utilisation >90%)\n...\ntotal number of models trained in milestone.e = 50 (8xGPU, expected utilisation >90%)\n\n\n[pricing and speedup]\n... we can speedup results by increasing the number of experiments for 5 to 10. but this requires us to increase our gpu resource from 8 to 16 and this incur cost of $$ xxx to the prroject... \n\n... we think we can improve performance if we have 10x more data at the cost of $$yyy ...\n\n... this project is special because it requires re-design of any existing model in the market. we need xx data scientists for a period of xx months, incurring cost of $$ xxx\ntake home message:
\nwhen you are working in Kaggle or commerical project, you are not only deciding on how to develop algorithms but also how to schedule for GPU resources and training experiments.
no one can be sure if a modification will make an algorithm works. But he must be about to make the assessment that : \" by making at least x number of modifications and experiments, one can be sure that there will be improvement of at least zz% with probability 95% chance of success\"
\n\"People always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world\"
\nin some commercial applications +0.00001 means better price of your AI product. e.g. face recognition, prediction of airplane fuel usage
\n** YES! DATA SCIENCE is sometimes VERY TOUGH WORK !!! **
", "votes": null, "replies": [] }, { "id": 1234637, "author_name": "tanulsingh077", "author_url": "", "post_date": "03/11/2021 12:37:59", "content": "@hengck23 very well summarized
\nDATA SCIENCE is sometimes VERY TOUGH WORK
Isn't this the reason that makes it so beautiful . @haqishen not getting discouraged afterall we all will learn from the solution you publish after the competition
", "votes": null, "replies": [] } ] }, { "id": 1234849, "author_name": "anku5hk", "author_url": "", "post_date": "03/11/2021 16:07:56", "content": "\n\nPeople always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001.
\n
Yes i feel the same, i used to wonder whats the point doing 0.001 improvements. If you feel the same you're right. I have made more than 100 submissions on my last competition and it sucks, it was wrong. It was a waste of time. Kaggle is place of experiments and learning not to fight for 0.001 percent. You don't know much come you to kaggle learn from others, if you know something you experiment it on kaggle that makes more sense and not training the same thing for 0.001 improvement. I guess the medal thing is what keeps us hooked. π
", "votes": null, "replies": [] }, { "id": 1234869, "author_name": "anjum48", "author_url": "", "post_date": "03/11/2021 16:29:44", "content": "Also dott moving into solo gold position!!! Unreal. π
", "votes": null, "replies": [ { "id": 1235440, "author_name": "tanulsingh077", "author_url": "", "post_date": "03/12/2021 06:19:37", "content": "Hahaha True
", "votes": null, "replies": [] } ] }, { "id": 1235332, "author_name": "aeryss", "author_url": "", "post_date": "03/12/2021 04:10:57", "content": "I also want to express my opinion.
\nPeople always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world.
This is true to us, competitors, but it is maybe not true to the organizer. What we consider expensive today might not be expensive tomorrow. GPU was for the top companies decades ago, but now we have GPU (and even TPU) freely available to everyone. That being said, in the future, computational cost will decrease, and it will not remain a problem.
\nMoreover, 0.001-0.00001 might not be important to us, but it is important to the organizer. The number is small, but it might have large impact. Letβs say you want to use hard labels, then there is a very big difference between 0.499999 and 0.5: the decision to do or not to do it. This margin has significant important in medicine or financial trading. A patient with false positive only results in additional medical treatment, but with a false negative they can die.
\nIf you are hear to learn, then score is not a big thing, but if you are here to compete, then you must adapt to compete with the best around the world.
", "votes": null, "replies": [ { "id": 1235441, "author_name": "tanulsingh077", "author_url": "", "post_date": "03/12/2021 06:21:30", "content": "Well said and these are some of the reasons I strongly disagree with people criticizing kaggle competitions
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1233773": "Hi all , I am afraid that is not related to the competition much but its my gratitude towards the community and all the Kaggle greats from whom I learn everyday.\n\nPeople always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world . While this statement might be partly true , I strongly disagree , Kaggle allows you to compete against the best , learn from the best , and make you push your limits and this in turn prepares you for all the challenges in the real world . While Kaggle competitions might not simulate real world challenges every time , it teaches a lot of lessons , it makes you more confident , makes you think on your feet .\n\nI have been working extra hard from past one month on this competition , learning techniques like spatial attention , Knowledge transfer models ,etc and still I am struggling to cross 0.970 and a team of greats (the elite club) comes in and reaches 2nd position with just 10 submissions , it just makes you wonder how tough it is to compete in a Kaggle competition\n\nI feel very lucky that I get the chance to even compete with such great people . Everything which I have learned has been mostly through Kaggle and I hope I keep learning . It's a beautiful stage for every young data scientists like me . No matter how much challenges I face Kaggle pushes me to keep toiling.\n\nThanks for taking out the time to read this and all the best for this competition", "1233839": "I was just wondering why I couldn't find @haqishen on leaderboard in a computer vision competition and here they (team of elites) are 2nd on LB with just 10 subs.", "1234176": "Don't be discouraged guys.\n\nWe also worked very hard on the past month, tried bunch of ideas, ran hundreds of experiments.\n\nBut as we find that our CV socre is very well correlated to public LB in less than 8 submissions a month ago, we decided NOT to submit until the end of this competition.", "1234442": "\" it just makes you wonder how tough it is to compete in a Kaggle competition\"\n\"turn prepares you for all the challenges in the real world\"\n\" While Kaggle competitions might not simulate real world challenges every time ...\"\n\nlet's relates to the real world!\n\n---\n\"We also worked very hard on the past month, tried bunch of ideas, ran hundreds of experiments.\"\n\n---\n\nAs a computer vision / deep learning algorithm developer in some commercial company, this is the plan that i need to submit to my customer, i.e. funding party, to bid for a project:\n\n```\n[expected performance of model: 95%]\n[expected milestone :]\na. 07-jun : 80% (min viable model ) -- first payment xx%\nb. 07-jul : 90% (first delivery) -- ....\n....\n\ne. 07-sep : ....\n\n\n[resource management]\nexpected number of models trained per day: 5 \ntotal number of models trained in milestone.a = 50 (8xGPU, expected utilisation >90%)\ntotal number of models trained in milestone.b = 50 (8xGPU, expected utilisation >90%)\n...\ntotal number of models trained in milestone.e = 50 (8xGPU, expected utilisation >90%)\n\n\n[pricing and speedup]\n... we can speedup results by increasing the number of experiments for 5 to 10. but this requires us to increase our gpu resource from 8 to 16 and this incur cost of $$ xxx to the prroject... \n\n... we think we can improve performance if we have 10x more data at the cost of $$yyy ...\n\n... this project is special because it requires re-design of any existing model in the market. we need xx data scientists for a period of xx months, incurring cost of $$ xxx\n\n```\n\n\ntake home message:\nwhen you are working in Kaggle or commerical project, you are not only deciding on how to develop algorithms but also how to schedule for GPU resources and training experiments.\n\nno one can be sure if a modification will make an algorithm works. But he must be about to make the assessment that : \" by making at least x number of modifications and experiments, one can be sure that there will be improvement of at least zz% with probability 95% chance of success\"\n\n---\n\n\"People always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world\"\n\nin some commercial applications +0.00001 means better price of your AI product. e.g. face recognition, prediction of airplane fuel usage\n\n** YES! DATA SCIENCE is sometimes VERY TOUGH WORK !!! **", "1234637": "hengck23 very well summarized \n\n` DATA SCIENCE is sometimes VERY TOUGH WORK`\n\nIsn't this the reason that makes it so beautiful . @haqishen not getting discouraged afterall we all will learn from the solution you publish after the competition", "1234849": "> People always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001.\n\n\nYes i feel the same, i used to wonder whats the point doing 0.001 improvements. If you feel the same you're right. I have made more than 100 submissions on my last competition and it sucks, it was wrong. It was a waste of time. Kaggle is place of experiments and learning not to fight for 0.001 percent. You don't know much come you to kaggle learn from others, if you know something you experiment it on kaggle that makes more sense and not training the same thing for 0.001 improvement. I guess the medal thing is what keeps us hooked. π", "1234869": "Also dott moving into solo gold position!!! Unreal. π", "1235332": "I also want to express my opinion.\n`People always criticize Kaggle to be a platform where people fight for accuracies in the range 0.001-0.00001 and it remains far away from the real world.`\n\nThis is true to us, competitors, but it is maybe not true to the organizer. What we consider expensive today might not be expensive tomorrow. GPU was for the top companies decades ago, but now we have GPU (and even TPU) freely available to everyone. That being said, in the future, computational cost will decrease, and it will not remain a problem.\n\nMoreover, 0.001-0.00001 might not be important to us, but it is important to the organizer. The number is small, but it might have large impact. Letβs say you want to use hard labels, then there is a very big difference between 0.499999 and 0.5: the decision to do or not to do it. This margin has significant important in medicine or financial trading. A patient with false positive only results in additional medical treatment, but with a false negative they can die.\n\nIf you are hear to learn, then score is not a big thing, but if you are here to compete, then you must adapt to compete with the best around the world.", "1235440": "Hahaha True", "1235441": "Well said and these are some of the reasons I strongly disagree with people criticizing kaggle competitions" }, "source": "meta" }