![\"https://i.ibb.co/6r67fFK/Selection-836.png\"](\"https://i.ibb.co/6r67fFK/Selection-836.png\")
Is this competition now all about ensembling public submissions or having Resources mines that can easily train many many models .Barring few many could be doing this only.
\nI think host should seriously look into the way competition heading and should consider limiting no of models that can be used for ensembling .
\nSimilarly kaggle should also put restrictions for Gold medal zone
Practically is it possible to deploy so many models and then ensembling their predictions ?.
", "messages": [ { "id": "1508736", "postDate": "09/10/2021 14:13:43", "content": "Is this competition now all about ensembling public submissions or having Resources mines that can easily train many many models .Barring few many could be doing this only.
\nI think host should seriously look into the way competition heading and should consider limiting no of models that can be used for ensembling .
\nSimilarly kaggle should also put restrictions for Gold medal zone
Practically is it possible to deploy so many models and then ensembling their predictions ?.
", "rawMarkdown": "Is this competition now all about ensembling public submissions or having Resources mines that can easily train many many models .Barring few many could be doing this only. \n\nI think host should seriously look into the way competition heading and should consider limiting no of models that can be used for ensembling .\nSimilarly kaggle should also put restrictions for Gold medal zone\n \nPractically is it possible to deploy so many models and then ensembling their predictions ?.", "votes": null }, { "id": "1508864", "postDate": "09/10/2021 16:32:39", "content": "Are you saying \"If I have a lot of resources, I can win easily\"?
", "rawMarkdown": "Are you saying \"If I have a lot of resources, I can win easily\"?", "votes": null }, { "id": "1508958", "postDate": "09/10/2021 18:01:15", "content": "i dint meant that what i meant was when resources in hand one can do experiments that one have in mind easily ,there are pretty much ideas in the various discussions that one can easily try now if resources in hand, do flip flop.
\nDont take me wrong..
", "rawMarkdown": "i dint meant that what i meant was when resources in hand one can do experiments that one have in mind easily ,there are pretty much ideas in the various discussions that one can easily try now if resources in hand, do flip flop.\n\nDont take me wrong..", "votes": null }, { "id": "1508973", "postDate": "09/10/2021 18:19:35", "content": "
the hard truth:
\n\"in competition, it is not about how good you are. it is about how much better than others\"
\nyou can overcome resource limitations with better algorithm or better data.
\nyou can overcome data limitations with better algorithm or better hardware resources.
\nyou can overcome algorithm limitations with better data or better hardware resources.
\"i meant was when resources in hand one can do experiments that one have in mind easily ,there are pretty much ideas \"
\nthat is why investment in data science and AI includes:
\nI just told my PhD friend:
\n\"don't just of think of mathematical formulas …\" As a modern PhD, you should think of how to get more GPU to do write a better research paper. e.g. can you access the government supercomputers? can you get sponsorship from google, nvidia? your research proposal should include how you get the computation resources for your work
", "rawMarkdown": "\"i meant was when resources in hand one can do experiments that one have in mind easily ,there are pretty much ideas \"\n\nthat is why investment in data science and AI includes:\n- spend time to do read papers\n- spend money to buy GPUs, etc\n\nI just told my PhD friend:\n\n\"don't just of think of mathematical formulas ...\" As a modern PhD, you should think of how to get more GPU to do write a better research paper. e.g. can you access the government supercomputers? can you get sponsorship from google, nvidia? your research proposal should include how you get the computation resources for your work", "votes": null }, { "id": "1509063", "postDate": "09/10/2021 19:32:04", "content": "We have many single models that would be in top 10 alone. And we're not the best team. Ensembling helps. but not that much.
\nIMHO, best team makes a difference using better signal processing (based on forum answers).
", "rawMarkdown": "We have many single models that would be in top 10 alone. And we're not the best team. Ensembling helps. but not that much.\n\nIMHO, best team makes a difference using better signal processing (based on forum answers).", "votes": null }, { "id": "1509064", "postDate": "09/10/2021 19:33:07", "content": "I see you got lots of downvotes, I don't get why. It seems people are getting nervous, you included.
\nThis is just a Kaggle competition, no need to get emotional about it.
", "rawMarkdown": "I see you got lots of downvotes, I don't get why. It seems people are getting nervous, you included.\n\nThis is just a Kaggle competition, no need to get emotional about it.", "votes": null }, { "id": "1509087", "postDate": "09/10/2021 20:10:20", "content": "If \"adding more models to the ensemble\" would necessary mean \"improving score\", you would be right. It's not that easy and I guess that many top teams have single models currently on top of the LB.
\nHowever, there were lots of movements on the LB in the past days and after 2+ months of competition, it's quite natural to get emotional about losing places. Just have a look at that great post from CPMP and try to project yourself in the future, past the current competition. Its not all about LB climbing but also about gaining knowledge (so that you'll be on top of the private LB in future competitions.)
", "rawMarkdown": "If \"adding more models to the ensemble\" would necessary mean \"improving score\", you would be right. It's not that easy and I guess that many top teams have single models currently on top of the LB.\n\nHowever, there were lots of movements on the LB in the past days and after 2+ months of competition, it's quite natural to get emotional about losing places. Just have a look at [that great post from CPMP](https://www.kaggle.com/c/ieee-fraud-detection/discussion/110848) and try to project yourself in the future, past the current competition. Its not all about LB climbing but also about gaining knowledge (so that you'll be on top of the private LB in future competitions.)", "votes": null }, { "id": "1509113", "postDate": "09/10/2021 20:54:21", "content": "Thanks for linking to that post of mine. It is indeed quite relevant.
", "rawMarkdown": "Thanks for linking to that post of mine. It is indeed quite relevant.", "votes": null }, { "id": "1509137", "postDate": "09/10/2021 22:14:11", "content": "I think that solid preprocessing can still be a game-changer.
\nAlso, creativity in the modelling stage might provide good results.
At this competition, my only computational resource is Kaggle TPU, so I partially understand your embarrassment.
\nThe amount of data and the nature of the problem are demanding resources.
But we still have chances if we rigorously research the problem.
\nSo my advice is don't give up and continue trying!
I dont care about down votes .. one other post on ethical dead line too got down votes . My intention is not to undermine hard efforts of any one .. I put this post after seeing rapid changes in lb positions , a week or so ago you might have an understanding 88 .1 is hard to beat but now things are different .
", "rawMarkdown": "I dont care about down votes .. one other post on ethical dead line too got down votes . My intention is not to undermine hard efforts of any one .. I put this post after seeing rapid changes in lb positions , a week or so ago you might have an understanding 88 .1 is hard to beat but now things are different .", "votes": null }, { "id": "1509421", "postDate": "09/11/2021 08:59:05", "content": "\n\na week or so ago you might have an understanding 88 .1 is hard to beat but now things are different
\n
@jaideepvalani 0.881 is easy to beat now? Thanks to what? I'd like to know because it was hard for us and I welcome ways to easily improve our score.
\nAnd also, if 0.881 is easy to beat then show us.
", "rawMarkdown": "> a week or so ago you might have an understanding 88 .1 is hard to beat but now things are different\n\n@jaideepvalani 0.881 is easy to beat now? Thanks to what? I'd like to know because it was hard for us and I welcome ways to easily improve our score.\n\nAnd also, if 0.881 is easy to beat then show us.", "votes": null }, { "id": "1509460", "postDate": "09/11/2021 10:26:19", "content": "@cpmpml i dint mean its easy .. what i meant was by looking at current trend of changes happening in LB positions score seems achievable for good no of people (very difficult with single model ) not for me as of now as my single model is at just 8771 lol:) . It would be interesting to see how many score of this range remain stand still on Private LB..
\nI project 88.1 to 88.3 to be likely gold range score on a public LB atleast.
I dont enjoy medals or ranking if my efforts are at least in getting same..or i dint learn any thing new from competition..
", "rawMarkdown": "I dont enjoy medals or ranking if my efforts are at least in getting same..or i dint learn any thing new from competition..", "votes": null }, { "id": "1509468", "postDate": "09/11/2021 10:35:24", "content": "Yes now its time to look for different source of scoring :)..
\nlot of things i discovered few weeks ago are now known to more set of people its okay its part of competition i dont have any ill feeling about this .. Will work more to explore more things. Sincerely DSP had always gone above my head right since beginning :)
As I wrote above, you can be in top 10 at this point with a single model that can be trained without too much compute resources.
", "rawMarkdown": "As I wrote above, you can be in top 10 at this point with a single model that can be trained without too much compute resources.", "votes": null }, { "id": "1509732", "postDate": "09/11/2021 16:15:58", "content": "I think you can use easily free TPU on colab which a very good hardware. I also use TPU on colab, so I have 16 free TPU.
\nI addition to that, if you have 10€ you can suscribe one month on colab pro for better GPU and TPU which is not very expensive.
\ngl hf for the end of the competition, enjoy there are more and more angry post this week about the competition
", "rawMarkdown": "I think you can use easily free TPU on colab which a very good hardware. I also use TPU on colab, so I have 16 free TPU. \n\nI addition to that, if you have 10€ you can suscribe one month on colab pro for better GPU and TPU which is not very expensive.\n\ngl hf for the end of the competition, enjoy there are more and more angry post this week about the competition", "votes": null }, { "id": "1510467", "postDate": "09/12/2021 12:51:39", "content": "Still a little bit early for this to be an issue. Besides, ensembling notebooks aren't very useful \"in general\" (sometimes they end up giving good final LB scores) and in the long run, it is better to try things for yourself and learn from what others share and what you read.
\nI hope this helps give an optimistic spin on things. 😄
", "rawMarkdown": "Still a little bit early for this to be an issue. Besides, ensembling notebooks aren't very useful \"in general\" (sometimes they end up giving good final LB scores) and in the long run, it is better to try things for yourself and learn from what others share and what you read. \n\nI hope this helps give an optimistic spin on things. 😄", "votes": null }, { "id": "1559702", "postDate": "10/27/2021 07:07:45", "content": "Hey All,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris
Are you saying \"If I have a lot of resources, I can win easily\"?
", "votes": null, "replies": [ { "id": 1508958, "author_name": "jaideepvalani", "author_url": "", "post_date": "09/10/2021 18:01:15", "content": "i dint meant that what i meant was when resources in hand one can do experiments that one have in mind easily ,there are pretty much ideas in the various discussions that one can easily try now if resources in hand, do flip flop.
\nDont take me wrong..
", "votes": null, "replies": [] }, { "id": 1508978, "author_name": "hengck23", "author_url": "", "post_date": "09/10/2021 18:22:32", "content": "\"i meant was when resources in hand one can do experiments that one have in mind easily ,there are pretty much ideas \"
\nthat is why investment in data science and AI includes:
\nI just told my PhD friend:
\n\"don't just of think of mathematical formulas …\" As a modern PhD, you should think of how to get more GPU to do write a better research paper. e.g. can you access the government supercomputers? can you get sponsorship from google, nvidia? your research proposal should include how you get the computation resources for your work
", "votes": null, "replies": [] } ] }, { "id": 1508973, "author_name": "hengck23", "author_url": "", "post_date": "09/10/2021 18:19:35", "content": "
the hard truth:
\n\"in competition, it is not about how good you are. it is about how much better than others\"
\nyou can overcome resource limitations with better algorithm or better data.
\nyou can overcome data limitations with better algorithm or better hardware resources.
\nyou can overcome algorithm limitations with better data or better hardware resources.
Yes now its time to look for different source of scoring :)..
\nlot of things i discovered few weeks ago are now known to more set of people its okay its part of competition i dont have any ill feeling about this .. Will work more to explore more things. Sincerely DSP had always gone above my head right since beginning :)
We have many single models that would be in top 10 alone. And we're not the best team. Ensembling helps. but not that much.
\nIMHO, best team makes a difference using better signal processing (based on forum answers).
", "votes": null, "replies": [ { "id": 1509064, "author_name": "cpmpml", "author_url": "", "post_date": "09/10/2021 19:33:07", "content": "I see you got lots of downvotes, I don't get why. It seems people are getting nervous, you included.
\nThis is just a Kaggle competition, no need to get emotional about it.
", "votes": null, "replies": [] }, { "id": 1509260, "author_name": "jaideepvalani", "author_url": "", "post_date": "09/11/2021 04:58:52", "content": "I dont care about down votes .. one other post on ethical dead line too got down votes . My intention is not to undermine hard efforts of any one .. I put this post after seeing rapid changes in lb positions , a week or so ago you might have an understanding 88 .1 is hard to beat but now things are different .
", "votes": null, "replies": [] }, { "id": 1509421, "author_name": "cpmpml", "author_url": "", "post_date": "09/11/2021 08:59:05", "content": "\n\na week or so ago you might have an understanding 88 .1 is hard to beat but now things are different
\n
@jaideepvalani 0.881 is easy to beat now? Thanks to what? I'd like to know because it was hard for us and I welcome ways to easily improve our score.
\nAnd also, if 0.881 is easy to beat then show us.
", "votes": null, "replies": [] }, { "id": 1509460, "author_name": "jaideepvalani", "author_url": "", "post_date": "09/11/2021 10:26:19", "content": "@cpmpml i dint mean its easy .. what i meant was by looking at current trend of changes happening in LB positions score seems achievable for good no of people (very difficult with single model ) not for me as of now as my single model is at just 8771 lol:) . It would be interesting to see how many score of this range remain stand still on Private LB..
\nI project 88.1 to 88.3 to be likely gold range score on a public LB atleast.
As I wrote above, you can be in top 10 at this point with a single model that can be trained without too much compute resources.
", "votes": null, "replies": [] } ] }, { "id": 1509087, "author_name": "fabiendaniel", "author_url": "", "post_date": "09/10/2021 20:10:20", "content": "If \"adding more models to the ensemble\" would necessary mean \"improving score\", you would be right. It's not that easy and I guess that many top teams have single models currently on top of the LB.
\nHowever, there were lots of movements on the LB in the past days and after 2+ months of competition, it's quite natural to get emotional about losing places. Just have a look at that great post from CPMP and try to project yourself in the future, past the current competition. Its not all about LB climbing but also about gaining knowledge (so that you'll be on top of the private LB in future competitions.)
", "votes": null, "replies": [ { "id": 1509113, "author_name": "cpmpml", "author_url": "", "post_date": "09/10/2021 20:54:21", "content": "Thanks for linking to that post of mine. It is indeed quite relevant.
", "votes": null, "replies": [] }, { "id": 1509463, "author_name": "jaideepvalani", "author_url": "", "post_date": "09/11/2021 10:31:16", "content": "I dont enjoy medals or ranking if my efforts are at least in getting same..or i dint learn any thing new from competition..
", "votes": null, "replies": [] } ] }, { "id": 1509137, "author_name": "rolandw0w", "author_url": "", "post_date": "09/10/2021 22:14:11", "content": "I think that solid preprocessing can still be a game-changer.
\nAlso, creativity in the modelling stage might provide good results.
At this competition, my only computational resource is Kaggle TPU, so I partially understand your embarrassment.
\nThe amount of data and the nature of the problem are demanding resources.
But we still have chances if we rigorously research the problem.
\nSo my advice is don't give up and continue trying!
I think you can use easily free TPU on colab which a very good hardware. I also use TPU on colab, so I have 16 free TPU.
\nI addition to that, if you have 10€ you can suscribe one month on colab pro for better GPU and TPU which is not very expensive.
\ngl hf for the end of the competition, enjoy there are more and more angry post this week about the competition
", "votes": null, "replies": [] }, { "id": 1510467, "author_name": "yassinealouini", "author_url": "", "post_date": "09/12/2021 12:51:39", "content": "Still a little bit early for this to be an issue. Besides, ensembling notebooks aren't very useful \"in general\" (sometimes they end up giving good final LB scores) and in the long run, it is better to try things for yourself and learn from what others share and what you read.
\nI hope this helps give an optimistic spin on things. 😄
", "votes": null, "replies": [] }, { "id": 1559702, "author_name": "zerafachris", "author_url": "", "post_date": "10/27/2021 07:07:45", "content": "Hey All,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris