Every day we see LB positions change like volatile stock market prices even since High Img resize kernel got published and lot of tricks and experiments shared in public discussion . Same old story continues to get repeated bringing disappointment to the participants.
\nIts high time kaggle should have a check on publicizing the notebooks that scores more than certain percentage of min Gold range score at any point of time.
Tips, tricks float freely , decimating every competition ,making it rather a horse run than fair performing platform .
", "messages": [ { "id": "1654127", "postDate": "01/18/2022 07:28:52", "content": "Every day we see LB positions change like volatile stock market prices even since High Img resize kernel got published and lot of tricks and experiments shared in public discussion . Same old story continues to get repeated bringing disappointment to the participants.
\nIts high time kaggle should have a check on publicizing the notebooks that scores more than certain percentage of min Gold range score at any point of time.
Tips, tricks float freely , decimating every competition ,making it rather a horse run than fair performing platform .
", "rawMarkdown": "Every day we see LB positions change like volatile stock market prices even since High Img resize kernel got published and lot of tricks and experiments shared in public discussion . Same old story continues to get repeated bringing disappointment to the participants. \nIts high time kaggle should have a check on publicizing the notebooks that scores more than certain percentage of min Gold range score at any point of time. \n\nTips, tricks float freely , decimating every competition ,making it rather a horse run than fair performing platform .", "votes": null }, { "id": "1654140", "postDate": "01/18/2022 07:42:54", "content": "I do not know who voted down your topic … but this is good question. It could be GPU war now …. but …. size is NOT ALL you need (you probably need higher resolution and specific training -> eg. batch size more then 1 to allow yolo train better model, better data cv, better training at all (different training strategy)). NOT ALL models will work with enlarged images. I am 100% sure …. that scaling imgs to enormous size is …. wasting time and …. (silent mode). Size shoud be specific to training… etc. etc. Topic about \"size is all you need\" is misleading but give people hope to jump … and many jump but not win in my opinion. This is only one factor which can improve your score but not a solution. As I said in different topic - sharing such information is not entirely responsible because many people who do not have the required computing power may now feel discouraged. Otherwise, they would look for solutions, and now it seems to them that the only solution is 10xGPU and huge pictures, which could not be true.
\nI am sure that this is important but could lead to some problem in private LB score … Could … but I'm not sure - investigation in progress (I am almost sure). Keep going on your best solution. Enjoy progress (or not - this is our case … we looking for improvement). There is a way to jump over 0.67 …. We are such case. We at best have Colab V100 with 16GB RAM. Could we jump higher, above 0.7? I think yes … still looking for solution. If not … it would be great lessons learned.
", "rawMarkdown": "I do not know who voted down your topic ... but this is good question. It **could be** GPU war now .... but .... size is **NOT ALL you need** (you probably need higher resolution and specific training -> eg. batch size more then 1 to allow yolo train better model, better data cv, better training at all (different training strategy)). NOT ALL models will work with enlarged images. I am 100% sure .... that scaling imgs to enormous size is .... wasting time and .... (silent mode). Size shoud be specific to training... etc. etc. Topic about \"size is all you need\" is misleading but give people hope to jump ... and many jump but not win in my opinion. This is only one factor which can improve your score but not a solution. As I said in different topic - sharing such information is not entirely responsible because many people who do not have the required computing power may now feel discouraged. Otherwise, they would look for solutions, and now it seems to them that the only solution is 10xGPU and huge pictures, which could not be true.\n\n I am sure that this is important but could lead to some problem in private LB score ... Could ... but I'm not sure - investigation in progress (I am almost sure). ***Keep going on your best solution. Enjoy progress (or not - this is our case ... we looking for improvement)***. There is a way to jump over 0.67 .... We are such case. We at best have Colab V100 with 16GB RAM. Could we jump higher, above 0.7? I think yes ... still looking for solution. If not ... it would be great lessons learned.", "votes": null }, { "id": "1654262", "postDate": "01/18/2022 11:07:51", "content": "@remekkinas may be my concern was more on sharing lot of tricks is what make others working hard to find them vulnerable towards slipping and make them again slog for days to find more things. My latest method not currently tested may have some thing to offer ,lets see.
\nI feel apologetic for you sliping from 1st place to now nearing out of gold range.
\nMay be final best solution is mix of lot other things
Slipping from #1 is not problem for me. This is just competition …. 😃
", "rawMarkdown": "Slipping from #1 is not problem for me. This is just competition .... 😃", "votes": null }, { "id": "1654625", "postDate": "01/18/2022 17:12:15", "content": "I'm curious of what a horse run is in this context. :)
", "rawMarkdown": "I'm curious of what a horse run is in this context. :)", "votes": null }, { "id": "1654695", "postDate": "01/18/2022 18:04:44", "content": "every is now gone crazy after img size.. showing their GPU mights to power their scores
", "rawMarkdown": "every is now gone crazy after img size.. showing their GPU mights to power their scores", "votes": null }, { "id": "1654701", "postDate": "01/18/2022 18:07:00", "content": "ds good.. hope you regain your position soon. i was planning to ask you for a team up if you were solo. our combination would have given resonating results :)
\nATB! keep moving Up..
", "rawMarkdown": "ds good.. hope you regain your position soon. i was planning to ask you for a team up if you were solo. our combination would have given resonating results :)\n\nATB! keep moving Up..", "votes": null }, { "id": "1654837", "postDate": "01/18/2022 21:21:33", "content": "High resolution images help increase LB.
\nBut that is a trade-off with inference time.
I think it can be a disadvantage when combining ensembles of multiple models and time-consuming post-processing.
", "rawMarkdown": "High resolution images help increase LB.\nBut that is a trade-off with inference time.\n\nI think it can be a disadvantage when combining ensembles of multiple models and time-consuming post-processing.", "votes": null }, { "id": "1654915", "postDate": "01/19/2022 00:15:04", "content": "Agree. The hard limit is inference resolution, because this must be performed on kaggle notebook. Inference time per image is roughly proportionate to the square of the resolution.
\nHaving said this, I am having trouble training higher resolutions with 16 Gb GPU. I think multiple GPUs won't help, as individual images must be trained on a single GPU. I think the top will be populated by those with access to 32Gb+ GPUs
", "rawMarkdown": "Agree. The hard limit is inference resolution, because this must be performed on kaggle notebook. Inference time per image is roughly proportionate to the square of the resolution.\n\nHaving said this, I am having trouble training higher resolutions with 16 Gb GPU. I think multiple GPUs won't help, as individual images must be trained on a single GPU. I think the top will be populated by those with access to 32Gb+ GPUs", "votes": null }, { "id": "1656270", "postDate": "01/19/2022 08:15:07", "content": "usually top rankers combine different engineering tricks to pop up,.
\nIt is not only pure model.
This is a platform for sharing ideas, and you're saying sharing ideas is bad?
\nA profound beauty of Kaggle is that it makes AI accessible to all.
\nCompetition medals is only one way to progress. The other way is to share notebooks and ideas in discussion. That's just Kaggle, man.
", "rawMarkdown": "This is a platform for sharing ideas, and you're saying sharing ideas is bad? \n\nA profound beauty of Kaggle is that it makes AI accessible to all.\n\nCompetition medals is only one way to progress. The other way is to share notebooks and ideas in discussion. That's just Kaggle, man.", "votes": null }, { "id": "1656543", "postDate": "01/19/2022 12:24:14", "content": "As I can see TOP10 use different solutions …. from light to …. complex one …
", "rawMarkdown": "As I can see TOP10 use different solutions .... from light to .... complex one ...", "votes": null }, { "id": "1656558", "postDate": "01/19/2022 12:41:50", "content": "yes @alexchwong I have trained large yolo models but all giving timeout on image sizes above 6000. So not exactly having 64 GB GPU is helping here.
", "rawMarkdown": "yes @alexchwong I have trained large yolo models but all giving timeout on image sizes above 6000. So not exactly having 64 GB GPU is helping here.", "votes": null }, { "id": "1656564", "postDate": "01/19/2022 12:49:55", "content": "And …. very very slow … we come to the conclution that \"resize is NOT all you need\" …. 😂😂😂
", "rawMarkdown": "And .... very very slow ... we come to the conclution that \"resize is NOT all you need\" .... 😂😂😂", "votes": null }, { "id": "1656572", "postDate": "01/19/2022 12:59:11", "content": "Exactly @remekkinas 😂. Once done with the experiments I will be posting comparison of CV and LB of all the large and small models. Hope that may clear the doubt of \" GPU War \"…
", "rawMarkdown": "Exactly @remekkinas 😂. Once done with the experiments I will be posting comparison of CV and LB of all the large and small models. Hope that may clear the doubt of \" GPU War \"...", "votes": null }, { "id": "1656587", "postDate": "01/19/2022 13:14:48", "content": "As I said … manipulations with image size can help (this is only part of solution) but this is not a final solution (all you need is just clickbait and lead people to depressing path - I need 10xGPU, I need 9h, train on 3600 and infer on 12000 or … train on 12000 and infer on 6542 …). The easiest way is to check LB position people who talk about this and check TOP10 LB submission time.
", "rawMarkdown": "As I said … manipulations with image size can help (this is only part of solution) but this is not a final solution (**all you need** is just clickbait and lead people to depressing path - I need 10xGPU, I need 9h, train on 3600 and infer on 12000 or … train on 12000 and infer on 6542 …). The easiest way is to check LB position people who talk about this and check TOP10 LB submission time.", "votes": null }, { "id": "1677408", "postDate": "02/05/2022 17:34:21", "content": "It looks now for me that … having more GPU is game changer unfortunately…
", "rawMarkdown": "It looks now for me that … having more GPU is game changer unfortunately…", "votes": null }, { "id": "1677813", "postDate": "02/06/2022 03:05:18", "content": "@remekkinas why all of the sudden you have this opinion ?
", "rawMarkdown": "remekkinas why all of the sudden you have this opinion ?", "votes": null }, { "id": "1677906", "postDate": "02/06/2022 05:59:58", "content": "I am waiting for competition end and TOP solutions description. I think (but this is only my assumption) that training with higher batch size (and higher res - but not extremely high) could change everything. I am not able to prove this because have only 16GB GPU so waiting for final days. As I can see with limited memory we could easily reach about 0.67-0.7 but then … it was end.
", "rawMarkdown": "I am waiting for competition end and TOP solutions description. I think (but this is only my assumption) that training with higher batch size (and higher res - but not extremely high) could change everything. I am not able to prove this because have only 16GB GPU so waiting for final days. As I can see with limited memory we could easily reach about 0.67-0.7 but then … it was end.", "votes": null }, { "id": "1677912", "postDate": "02/06/2022 06:09:12", "content": "Higher batch size of 8 with img size of 3200 is not making an impact. m6 and l6 are larger models which cannot go upto batch size of 8 and imgsz 3200 on 64 GB GPU. Plus larger models aren't making any difference because they need more data.
", "rawMarkdown": "Higher batch size of 8 with img size of 3200 is not making an impact. m6 and l6 are larger models which cannot go upto batch size of 8 and imgsz 3200 on 64 GB GPU. Plus larger models aren't making any difference because they need more data.", "votes": null } ], "comments": [ { "id": 1654140, "author_name": "remekkinas", "author_url": "", "post_date": "01/18/2022 07:42:54", "content": "I do not know who voted down your topic … but this is good question. It could be GPU war now …. but …. size is NOT ALL you need (you probably need higher resolution and specific training -> eg. batch size more then 1 to allow yolo train better model, better data cv, better training at all (different training strategy)). NOT ALL models will work with enlarged images. I am 100% sure …. that scaling imgs to enormous size is …. wasting time and …. (silent mode). Size shoud be specific to training… etc. etc. Topic about \"size is all you need\" is misleading but give people hope to jump … and many jump but not win in my opinion. This is only one factor which can improve your score but not a solution. As I said in different topic - sharing such information is not entirely responsible because many people who do not have the required computing power may now feel discouraged. Otherwise, they would look for solutions, and now it seems to them that the only solution is 10xGPU and huge pictures, which could not be true.
\nI am sure that this is important but could lead to some problem in private LB score … Could … but I'm not sure - investigation in progress (I am almost sure). Keep going on your best solution. Enjoy progress (or not - this is our case … we looking for improvement). There is a way to jump over 0.67 …. We are such case. We at best have Colab V100 with 16GB RAM. Could we jump higher, above 0.7? I think yes … still looking for solution. If not … it would be great lessons learned.
", "votes": null, "replies": [ { "id": 1654262, "author_name": "jaideepvalani", "author_url": "", "post_date": "01/18/2022 11:07:51", "content": "@remekkinas may be my concern was more on sharing lot of tricks is what make others working hard to find them vulnerable towards slipping and make them again slog for days to find more things. My latest method not currently tested may have some thing to offer ,lets see.
\nI feel apologetic for you sliping from 1st place to now nearing out of gold range.
\nMay be final best solution is mix of lot other things
Slipping from #1 is not problem for me. This is just competition …. 😃
", "votes": null, "replies": [] }, { "id": 1654701, "author_name": "jaideepvalani", "author_url": "", "post_date": "01/18/2022 18:07:00", "content": "ds good.. hope you regain your position soon. i was planning to ask you for a team up if you were solo. our combination would have given resonating results :)
\nATB! keep moving Up..
", "votes": null, "replies": [] }, { "id": 1656270, "author_name": "dragonzhang", "author_url": "", "post_date": "01/19/2022 08:15:07", "content": "usually top rankers combine different engineering tricks to pop up,.
\nIt is not only pure model.
As I can see TOP10 use different solutions …. from light to …. complex one …
", "votes": null, "replies": [] }, { "id": 1677408, "author_name": "remekkinas", "author_url": "", "post_date": "02/05/2022 17:34:21", "content": "It looks now for me that … having more GPU is game changer unfortunately…
", "votes": null, "replies": [] }, { "id": 1677813, "author_name": "nyleve", "author_url": "", "post_date": "02/06/2022 03:05:18", "content": "@remekkinas why all of the sudden you have this opinion ?
", "votes": null, "replies": [] }, { "id": 1677906, "author_name": "remekkinas", "author_url": "", "post_date": "02/06/2022 05:59:58", "content": "I am waiting for competition end and TOP solutions description. I think (but this is only my assumption) that training with higher batch size (and higher res - but not extremely high) could change everything. I am not able to prove this because have only 16GB GPU so waiting for final days. As I can see with limited memory we could easily reach about 0.67-0.7 but then … it was end.
", "votes": null, "replies": [] }, { "id": 1677912, "author_name": "sanchitvj", "author_url": "", "post_date": "02/06/2022 06:09:12", "content": "Higher batch size of 8 with img size of 3200 is not making an impact. m6 and l6 are larger models which cannot go upto batch size of 8 and imgsz 3200 on 64 GB GPU. Plus larger models aren't making any difference because they need more data.
", "votes": null, "replies": [] } ] }, { "id": 1654625, "author_name": "davidbroberts", "author_url": "", "post_date": "01/18/2022 17:12:15", "content": "I'm curious of what a horse run is in this context. :)
", "votes": null, "replies": [ { "id": 1654695, "author_name": "jaideepvalani", "author_url": "", "post_date": "01/18/2022 18:04:44", "content": "every is now gone crazy after img size.. showing their GPU mights to power their scores
", "votes": null, "replies": [] } ] }, { "id": 1654837, "author_name": "syurenuko", "author_url": "", "post_date": "01/18/2022 21:21:33", "content": "High resolution images help increase LB.
\nBut that is a trade-off with inference time.
I think it can be a disadvantage when combining ensembles of multiple models and time-consuming post-processing.
", "votes": null, "replies": [ { "id": 1654915, "author_name": "alexchwong", "author_url": "", "post_date": "01/19/2022 00:15:04", "content": "Agree. The hard limit is inference resolution, because this must be performed on kaggle notebook. Inference time per image is roughly proportionate to the square of the resolution.
\nHaving said this, I am having trouble training higher resolutions with 16 Gb GPU. I think multiple GPUs won't help, as individual images must be trained on a single GPU. I think the top will be populated by those with access to 32Gb+ GPUs
", "votes": null, "replies": [] }, { "id": 1656558, "author_name": "sanchitvj", "author_url": "", "post_date": "01/19/2022 12:41:50", "content": "yes @alexchwong I have trained large yolo models but all giving timeout on image sizes above 6000. So not exactly having 64 GB GPU is helping here.
", "votes": null, "replies": [] }, { "id": 1656564, "author_name": "remekkinas", "author_url": "", "post_date": "01/19/2022 12:49:55", "content": "And …. very very slow … we come to the conclution that \"resize is NOT all you need\" …. 😂😂😂
", "votes": null, "replies": [] }, { "id": 1656572, "author_name": "sanchitvj", "author_url": "", "post_date": "01/19/2022 12:59:11", "content": "Exactly @remekkinas 😂. Once done with the experiments I will be posting comparison of CV and LB of all the large and small models. Hope that may clear the doubt of \" GPU War \"…
", "votes": null, "replies": [] }, { "id": 1656587, "author_name": "remekkinas", "author_url": "", "post_date": "01/19/2022 13:14:48", "content": "As I said … manipulations with image size can help (this is only part of solution) but this is not a final solution (all you need is just clickbait and lead people to depressing path - I need 10xGPU, I need 9h, train on 3600 and infer on 12000 or … train on 12000 and infer on 6542 …). The easiest way is to check LB position people who talk about this and check TOP10 LB submission time.
", "votes": null, "replies": [] } ] }, { "id": 1656307, "author_name": "kaggleqrdl", "author_url": "", "post_date": "01/19/2022 08:48:21", "content": "This is a platform for sharing ideas, and you're saying sharing ideas is bad?
\nA profound beauty of Kaggle is that it makes AI accessible to all.
\nCompetition medals is only one way to progress. The other way is to share notebooks and ideas in discussion. That's just Kaggle, man.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1654127": "Every day we see LB positions change like volatile stock market prices even since High Img resize kernel got published and lot of tricks and experiments shared in public discussion . Same old story continues to get repeated bringing disappointment to the participants. \nIts high time kaggle should have a check on publicizing the notebooks that scores more than certain percentage of min Gold range score at any point of time. \n\nTips, tricks float freely , decimating every competition ,making it rather a horse run than fair performing platform .", "1654140": "I do not know who voted down your topic ... but this is good question. It **could be** GPU war now .... but .... size is **NOT ALL you need** (you probably need higher resolution and specific training -> eg. batch size more then 1 to allow yolo train better model, better data cv, better training at all (different training strategy)). NOT ALL models will work with enlarged images. I am 100% sure .... that scaling imgs to enormous size is .... wasting time and .... (silent mode). Size shoud be specific to training... etc. etc. Topic about \"size is all you need\" is misleading but give people hope to jump ... and many jump but not win in my opinion. This is only one factor which can improve your score but not a solution. As I said in different topic - sharing such information is not entirely responsible because many people who do not have the required computing power may now feel discouraged. Otherwise, they would look for solutions, and now it seems to them that the only solution is 10xGPU and huge pictures, which could not be true.\n\n I am sure that this is important but could lead to some problem in private LB score ... Could ... but I'm not sure - investigation in progress (I am almost sure). ***Keep going on your best solution. Enjoy progress (or not - this is our case ... we looking for improvement)***. There is a way to jump over 0.67 .... We are such case. We at best have Colab V100 with 16GB RAM. Could we jump higher, above 0.7? I think yes ... still looking for solution. If not ... it would be great lessons learned.", "1654262": "remekkinas may be my concern was more on sharing lot of tricks is what make others working hard to find them vulnerable towards slipping and make them again slog for days to find more things. My latest method not currently tested may have some thing to offer ,lets see. \nI feel apologetic for you sliping from 1st place to now nearing out of gold range. \nMay be final best solution is mix of lot other things", "1654505": "Slipping from #1 is not problem for me. This is just competition .... 😃", "1654625": "I'm curious of what a horse run is in this context. :)", "1654695": "every is now gone crazy after img size.. showing their GPU mights to power their scores", "1654701": "ds good.. hope you regain your position soon. i was planning to ask you for a team up if you were solo. our combination would have given resonating results :)\n\nATB! keep moving Up..", "1654837": "High resolution images help increase LB.\nBut that is a trade-off with inference time.\n\nI think it can be a disadvantage when combining ensembles of multiple models and time-consuming post-processing.", "1654915": "Agree. The hard limit is inference resolution, because this must be performed on kaggle notebook. Inference time per image is roughly proportionate to the square of the resolution.\n\nHaving said this, I am having trouble training higher resolutions with 16 Gb GPU. I think multiple GPUs won't help, as individual images must be trained on a single GPU. I think the top will be populated by those with access to 32Gb+ GPUs", "1656270": "usually top rankers combine different engineering tricks to pop up,. \nIt is not only pure model.", "1656307": "This is a platform for sharing ideas, and you're saying sharing ideas is bad? \n\nA profound beauty of Kaggle is that it makes AI accessible to all.\n\nCompetition medals is only one way to progress. The other way is to share notebooks and ideas in discussion. That's just Kaggle, man.", "1656543": "As I can see TOP10 use different solutions .... from light to .... complex one ...", "1656558": "yes @alexchwong I have trained large yolo models but all giving timeout on image sizes above 6000. So not exactly having 64 GB GPU is helping here.", "1656564": "And .... very very slow ... we come to the conclution that \"resize is NOT all you need\" .... 😂😂😂", "1656572": "Exactly @remekkinas 😂. Once done with the experiments I will be posting comparison of CV and LB of all the large and small models. Hope that may clear the doubt of \" GPU War \"...", "1656587": "As I said … manipulations with image size can help (this is only part of solution) but this is not a final solution (**all you need** is just clickbait and lead people to depressing path - I need 10xGPU, I need 9h, train on 3600 and infer on 12000 or … train on 12000 and infer on 6542 …). The easiest way is to check LB position people who talk about this and check TOP10 LB submission time.", "1677408": "It looks now for me that … having more GPU is game changer unfortunately…", "1677813": "remekkinas why all of the sudden you have this opinion ?", "1677906": "I am waiting for competition end and TOP solutions description. I think (but this is only my assumption) that training with higher batch size (and higher res - but not extremely high) could change everything. I am not able to prove this because have only 16GB GPU so waiting for final days. As I can see with limited memory we could easily reach about 0.67-0.7 but then … it was end.", "1677912": "Higher batch size of 8 with img size of 3200 is not making an impact. m6 and l6 are larger models which cannot go upto batch size of 8 and imgsz 3200 on 64 GB GPU. Plus larger models aren't making any difference because they need more data." }, "source": "meta" }