Some say simple model is good enough, some say deeper models bring improvement.
\n\nSome say no/little augmentation is good enough, some say this is all about augmentation.
\n\nI think I need a drink...
", "messages": [ { "id": 759874, "postDate": "2020-02-29T15:09:11.610Z", "content": "Some say simple model is good enough, some say deeper models bring improvement.
\n\nSome say no/little augmentation is good enough, some say this is all about augmentation.
\n\nI think I need a drink...
", "rawMarkdown": "Some say simple model is good enough, some say deeper models bring improvement.\n\nSome say no/little augmentation is good enough, some say this is all about augmentation.\n\nI think I need a drink...", "votes": 18 }, { "id": 760074, "postDate": "2020-02-29T20:12:06.770Z", "content": "All Roads Lead to Rome
\n\nIt seems like they are many ways to get to top score.. I will be super curios to read the solutions once competition is over =)
", "rawMarkdown": "**All Roads Lead to Rome**\n\nIt seems like they are many ways to get to top score.. I will be super curios to read the solutions once competition is over =) ", "votes": 14, "replies": [ { "id": 760579, "postDate": "2020-03-01T13:42:40.953Z", "content": "\n", "rawMarkdown": "> All Roads Lead to Rome\n💜 ", "votes": 1 } ] }, { "id": 766232, "postDate": "2020-03-07T22:03:43.933Z", "content": "All Roads Lead to Rome\n 💜
\n
It looks like everyone just got stuck in his own local minimum in the hyper-parameter space) and if one tries to move in the direction that works for another local minimum, the score drops. Right now it is just about whose local minimum is better... Quite bumpy hyper-parameter space. Quite interesting competition that so many different approaches are working.
", "rawMarkdown": "It looks like everyone just got stuck in his own local minimum in the hyper-parameter space) and if one tries to move in the direction that works for another local minimum, the score drops. Right now it is just about whose local minimum is better... Quite bumpy hyper-parameter space. Quite interesting competition that so many different approaches are working. ", "votes": 11 }, { "id": 759945, "postDate": "2020-02-29T16:37:54.020Z", "content": "blend all of them, that is what we are doing, yielding improvements. Dont drown in all the recommendations because for some folds it works and for other folds it doesnt
", "rawMarkdown": "blend all of them, that is what we are doing, yielding improvements. Dont drown in all the recommendations because for some folds it works and for other folds it doesnt", "votes": 3, "replies": [ { "id": 760389, "postDate": "2020-03-01T08:08:49.643Z", "content": "\"You can use all the recommendations some of the fold and some of the recommendations all the fold, but you cannot use all the recommendations all the fold.\" ?
", "rawMarkdown": "\"You can use all the recommendations some of the fold and some of the recommendations all the fold, but you cannot use all the recommendations all the fold.\" ?", "votes": 1 } ] }, { "id": 759942, "postDate": "2020-02-29T16:32:03.937Z", "content": "Cheers!
", "rawMarkdown": "Cheers!", "votes": 3 }, { "id": 760256, "postDate": "2020-03-01T03:30:28.927Z", "content": "all solutions are \"correct\" and \"wrong\" at some time but not all time
", "rawMarkdown": "all solutions are \"correct\" and \"wrong\" at some time but not all time", "votes": 4, "replies": [ { "id": 770093, "postDate": "2020-03-12T14:58:36.857Z", "content": "heisenberg uncertainty principle :P
", "rawMarkdown": "heisenberg uncertainty principle :P ", "votes": 1 } ] }, { "id": 765409, "postDate": "2020-03-06T15:34:54.033Z", "content": "元気づける
", "rawMarkdown": "元気づける", "votes": 2 }, { "id": 760577, "postDate": "2020-03-01T13:41:14.970Z", "content": "I think compute power is literally the key to win every kaggle competitions! (More experiments, monster models, faster training, bigger batch size,...)\nI own a single 2070rtx 😅
", "rawMarkdown": "I think compute power is literally the key to win every kaggle competitions! (More experiments, monster models, faster training, bigger batch size,...)\nI own a single 2070rtx 😅 ", "votes": 2, "replies": [ { "id": 761287, "postDate": "2020-03-02T11:07:05.137Z", "content": "Definitely not every competition, I would even argue that most competitions can be won without crazy hardware. That said, it plays a huge role in this one, probably the most I have seen it yet (out of the competitions I participated in).
", "rawMarkdown": "Definitely not every competition, I would even argue that most competitions can be won without crazy hardware. That said, it plays a huge role in this one, probably the most I have seen it yet (out of the competitions I participated in).", "votes": 6 } ] }, { "id": 760227, "postDate": "2020-03-01T02:14:54.833Z", "content": "Welcome to world of Computer Vision :)
", "rawMarkdown": "Welcome to world of Computer Vision :)", "votes": 2 }, { "id": 759949, "postDate": "2020-02-29T16:42:54.283Z", "content": "All the tricks to get 98+ are publicly shared, keep trying and you'll find the right combination!
", "rawMarkdown": "All the tricks to get 98+ are publicly shared, keep trying and you'll find the right combination!", "votes": 3 }, { "id": 764323, "postDate": "2020-03-05T10:47:04.757Z", "content": "Hahaha well, it sounds like you are frustrated enough!\nIn the meanwhile, I would like to say -I think it is all about the experience that speaks in the form of methods or ideas you are implementing. We can't predict the solution without experimenting, maybe that is why data scientists are highly paid!.
", "rawMarkdown": "Hahaha well, it sounds like you are frustrated enough!\nIn the meanwhile, I would like to say -I think it is all about the experience that speaks in the form of methods or ideas you are implementing. We can't predict the solution without experimenting, maybe that is why data scientists are highly paid!.\n", "votes": 1 }, { "id": 764169, "postDate": "2020-03-05T07:42:54.930Z", "content": "Well yeah, each person has different opinions, Everyone's solution and implementation of this competition are different.\nFor me, deeper model, little augmentations are good.
", "rawMarkdown": "Well yeah, each person has different opinions, Everyone's solution and implementation of this competition are different.\nFor me, deeper model, little augmentations are good.", "votes": 1 }, { "id": 762769, "postDate": "2020-03-03T19:44:44.413Z", "content": "Can feel it too 🤕
\n\n10 epochs => 0.9634\n100 epochs => 0.9635
\n\non single model, nomather what augmentation.
", "rawMarkdown": "Can feel it too 🤕 \n\n10 epochs => 0.9634\n100 epochs => 0.9635\n\non single model, nomather what augmentation.", "votes": 1 }, { "id": 761193, "postDate": "2020-03-02T08:34:48.937Z", "content": "I am also confused. Stucked for 2 weeks...
", "rawMarkdown": "I am also confused. Stucked for 2 weeks...", "votes": 1 }, { "id": 761052, "postDate": "2020-03-02T04:36:28.793Z", "content": "Everyone's solution and implementation of this competition are different. I'm afraid there's any universal way to improve your score steadily but just keep trying.
", "rawMarkdown": "Everyone's solution and implementation of this competition are different. I'm afraid there's any universal way to improve your score steadily but just keep trying.", "votes": 1 }, { "id": 760041, "postDate": "2020-02-29T19:08:45.627Z", "content": "True😂
", "rawMarkdown": "True😂 ", "votes": 1 }, { "id": 761746, "postDate": "2020-03-02T22:20:20.263Z", "content": "It depends on how you define good \"enough\". A simple model without augmentation is good enough to get you a decent score. Not good enough to get you a gold medal.
", "rawMarkdown": "It depends on how you define good \"enough\". A simple model without augmentation is good enough to get you a decent score. Not good enough to get you a gold medal. ", "votes": 2 }, { "id": 760333, "postDate": "2020-03-01T06:17:44.477Z", "content": "Because this is a competition, excluding people like @hengck23's useful experimentation posts, the rest I take as something to try for myself time and resources permitting. These shared comments are without knowing their exact architecture, hyperparameters, and other constraints they have use. So it is very hard to know. Just my 2 cents.
", "rawMarkdown": "Because this is a competition, excluding people like @hengck23's useful experimentation posts, the rest I take as something to try for myself time and resources permitting. These shared comments are without knowing their exact architecture, hyperparameters, and other constraints they have use. So it is very hard to know. Just my 2 cents. ", "votes": 2 }, { "id": 759894, "postDate": "2020-02-29T15:36:13.440Z", "content": "yeah They all say differently. 😂 \nfor me, deeper model, little augmentations are good
", "rawMarkdown": "yeah They all say differently. 😂 \nfor me, deeper model, little augmentations are good", "votes": 2 }, { "id": 770286, "postDate": "2020-03-12T19:23:54.903Z", "rawMarkdown": "" }, { "id": 761380, "postDate": "2020-03-02T12:58:12.640Z", "content": "I am also getting lost. My case for a single model CV: 0.99 and LB: 0.9568 and for 3 separate models it has CV: 0.99 but LB 0.92. What's wrong with me trying to figure it out. Can you please suggest anything like a single model vs multiple models? How people are getting higher LB? Using separate 3 models or single.
", "rawMarkdown": "I am also getting lost. My case for a single model CV: 0.99 and LB: 0.9568 and for 3 separate models it has CV: 0.99 but LB 0.92. What's wrong with me trying to figure it out. Can you please suggest anything like a single model vs multiple models? How people are getting higher LB? Using separate 3 models or single. ", "replies": [ { "id": 761426, "postDate": "2020-03-02T13:59:18.737Z", "content": "Do you use the trained data as a validation set?\nI have made such a mistake.
", "rawMarkdown": "Do you use the trained data as a validation set?\nI have made such a mistake." } ] }, { "id": 763882, "postDate": "2020-03-04T23:54:01.753Z", "rawMarkdown": "", "votes": -1, "isDeleted": true } ], "comments": [ { "id": 760074, "author_name": "DrHB", "author_url": "", "post_date": "2020-02-29T20:12:06.770000", "content": "All Roads Lead to Rome
\n\nIt seems like they are many ways to get to top score.. I will be super curios to read the solutions once competition is over =)
", "votes": 14, "replies": [ { "id": 760579, "author_name": "Innat", "author_url": "", "post_date": "2020-03-01T13:42:40.953000", "content": "\n", "votes": 1, "replies": [] } ] }, { "id": 766232, "author_name": "Iafoss", "author_url": "", "post_date": "2020-03-07T22:03:43.933000", "content": "All Roads Lead to Rome\n 💜
\n
It looks like everyone just got stuck in his own local minimum in the hyper-parameter space) and if one tries to move in the direction that works for another local minimum, the score drops. Right now it is just about whose local minimum is better... Quite bumpy hyper-parameter space. Quite interesting competition that so many different approaches are working.
", "votes": 11, "replies": [] }, { "id": 759945, "author_name": "CoreyJamesLevinson", "author_url": "", "post_date": "2020-02-29T16:37:54.020000", "content": "blend all of them, that is what we are doing, yielding improvements. Dont drown in all the recommendations because for some folds it works and for other folds it doesnt
", "votes": 3, "replies": [ { "id": 760389, "author_name": "hengck23", "author_url": "", "post_date": "2020-03-01T08:08:49.643000", "content": "\"You can use all the recommendations some of the fold and some of the recommendations all the fold, but you cannot use all the recommendations all the fold.\" ?
", "votes": 1, "replies": [] } ] }, { "id": 759942, "author_name": "Andrei Dukhounik", "author_url": "", "post_date": "2020-02-29T16:32:03.937000", "content": "Cheers!
", "votes": 3, "replies": [] }, { "id": 760256, "author_name": "hengck23", "author_url": "", "post_date": "2020-03-01T03:30:28.927000", "content": "all solutions are \"correct\" and \"wrong\" at some time but not all time
", "votes": 4, "replies": [ { "id": 770093, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2020-03-12T14:58:36.857000", "content": "heisenberg uncertainty principle :P
", "votes": 1, "replies": [] } ] }, { "id": 765409, "author_name": "Fuaifeng", "author_url": "", "post_date": "2020-03-06T15:34:54.033000", "content": "元気づける
", "votes": 2, "replies": [] }, { "id": 760577, "author_name": "Quan", "author_url": "", "post_date": "2020-03-01T13:41:14.970000", "content": "I think compute power is literally the key to win every kaggle competitions! (More experiments, monster models, faster training, bigger batch size,...)\nI own a single 2070rtx 😅
", "votes": 2, "replies": [ { "id": 761287, "author_name": "Psi", "author_url": "", "post_date": "2020-03-02T11:07:05.137000", "content": "Definitely not every competition, I would even argue that most competitions can be won without crazy hardware. That said, it plays a huge role in this one, probably the most I have seen it yet (out of the competitions I participated in).
", "votes": 6, "replies": [] } ] }, { "id": 760227, "author_name": "Bibek", "author_url": "", "post_date": "2020-03-01T02:14:54.833000", "content": "Welcome to world of Computer Vision :)
", "votes": 2, "replies": [] }, { "id": 759949, "author_name": "GreatGameDota", "author_url": "", "post_date": "2020-02-29T16:42:54.283000", "content": "All the tricks to get 98+ are publicly shared, keep trying and you'll find the right combination!
", "votes": 3, "replies": [] }, { "id": 764323, "author_name": "Charchit Dhawan", "author_url": "", "post_date": "2020-03-05T10:47:04.757000", "content": "Hahaha well, it sounds like you are frustrated enough!\nIn the meanwhile, I would like to say -I think it is all about the experience that speaks in the form of methods or ideas you are implementing. We can't predict the solution without experimenting, maybe that is why data scientists are highly paid!.
", "votes": 1, "replies": [] }, { "id": 764169, "author_name": "Anusha Nidamanuri", "author_url": "", "post_date": "2020-03-05T07:42:54.930000", "content": "Well yeah, each person has different opinions, Everyone's solution and implementation of this competition are different.\nFor me, deeper model, little augmentations are good.
", "votes": 1, "replies": [] }, { "id": 762769, "author_name": "Jo Tom", "author_url": "", "post_date": "2020-03-03T19:44:44.413000", "content": "Can feel it too 🤕
\n\n10 epochs => 0.9634\n100 epochs => 0.9635
\n\non single model, nomather what augmentation.
", "votes": 1, "replies": [] }, { "id": 761193, "author_name": "yama", "author_url": "", "post_date": "2020-03-02T08:34:48.937000", "content": "I am also confused. Stucked for 2 weeks...
", "votes": 1, "replies": [] }, { "id": 761052, "author_name": "syoya", "author_url": "", "post_date": "2020-03-02T04:36:28.793000", "content": "Everyone's solution and implementation of this competition are different. I'm afraid there's any universal way to improve your score steadily but just keep trying.
", "votes": 1, "replies": [] }, { "id": 760041, "author_name": "Akshit Sharma", "author_url": "", "post_date": "2020-02-29T19:08:45.627000", "content": "True😂
", "votes": 1, "replies": [] }, { "id": 761746, "author_name": "Massoud Hosseinali", "author_url": "", "post_date": "2020-03-02T22:20:20.263000", "content": "It depends on how you define good \"enough\". A simple model without augmentation is good enough to get you a decent score. Not good enough to get you a gold medal.
", "votes": 2, "replies": [] }, { "id": 760333, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2020-03-01T06:17:44.477000", "content": "Because this is a competition, excluding people like @hengck23's useful experimentation posts, the rest I take as something to try for myself time and resources permitting. These shared comments are without knowing their exact architecture, hyperparameters, and other constraints they have use. So it is very hard to know. Just my 2 cents.
", "votes": 2, "replies": [] }, { "id": 759894, "author_name": "Jaebb", "author_url": "", "post_date": "2020-02-29T15:36:13.440000", "content": "yeah They all say differently. 😂 \nfor me, deeper model, little augmentations are good
", "votes": 2, "replies": [] }, { "id": 770286, "author_name": "timetraveller", "author_url": "", "post_date": "2020-03-12T19:23:54.903000", "content": "", "votes": 0, "replies": [] }, { "id": 761380, "author_name": "Shamrat Akbar", "author_url": "", "post_date": "2020-03-02T12:58:12.640000", "content": "I am also getting lost. My case for a single model CV: 0.99 and LB: 0.9568 and for 3 separate models it has CV: 0.99 but LB 0.92. What's wrong with me trying to figure it out. Can you please suggest anything like a single model vs multiple models? How people are getting higher LB? Using separate 3 models or single.
", "votes": 0, "replies": [ { "id": 761426, "author_name": "0chen", "author_url": "", "post_date": "2020-03-02T13:59:18.737000", "content": "Do you use the trained data as a validation set?\nI have made such a mistake.
", "votes": 0, "replies": [] } ] }, { "id": 763882, "author_name": "", "author_url": "", "post_date": "2020-03-04T23:54:01.753000", "content": "", "votes": -1, "replies": [] } ], "raw_markdown_by_id": { "759874": "Some say simple model is good enough, some say deeper models bring improvement.\n\nSome say no/little augmentation is good enough, some say this is all about augmentation.\n\nI think I need a drink...", "760074": "**All Roads Lead to Rome**\n\nIt seems like they are many ways to get to top score.. I will be super curios to read the solutions once competition is over =) ", "766232": "It looks like everyone just got stuck in his own local minimum in the hyper-parameter space) and if one tries to move in the direction that works for another local minimum, the score drops. Right now it is just about whose local minimum is better... Quite bumpy hyper-parameter space. Quite interesting competition that so many different approaches are working. ", "759945": "blend all of them, that is what we are doing, yielding improvements. Dont drown in all the recommendations because for some folds it works and for other folds it doesnt", "759942": "Cheers!", "760256": "all solutions are \"correct\" and \"wrong\" at some time but not all time", "765409": "元気づける", "760577": "I think compute power is literally the key to win every kaggle competitions! (More experiments, monster models, faster training, bigger batch size,...)\nI own a single 2070rtx 😅 ", "760227": "Welcome to world of Computer Vision :)", "759949": "All the tricks to get 98+ are publicly shared, keep trying and you'll find the right combination!", "764323": "Hahaha well, it sounds like you are frustrated enough!\nIn the meanwhile, I would like to say -I think it is all about the experience that speaks in the form of methods or ideas you are implementing. We can't predict the solution without experimenting, maybe that is why data scientists are highly paid!.\n", "764169": "Well yeah, each person has different opinions, Everyone's solution and implementation of this competition are different.\nFor me, deeper model, little augmentations are good.", "762769": "Can feel it too 🤕 \n\n10 epochs => 0.9634\n100 epochs => 0.9635\n\non single model, nomather what augmentation.", "761193": "I am also confused. Stucked for 2 weeks...", "761052": "Everyone's solution and implementation of this competition are different. I'm afraid there's any universal way to improve your score steadily but just keep trying.", "760041": "True😂 ", "761746": "It depends on how you define good \"enough\". A simple model without augmentation is good enough to get you a decent score. Not good enough to get you a gold medal. ", "760333": "Because this is a competition, excluding people like @hengck23's useful experimentation posts, the rest I take as something to try for myself time and resources permitting. These shared comments are without knowing their exact architecture, hyperparameters, and other constraints they have use. So it is very hard to know. Just my 2 cents. ", "759894": "yeah They all say differently. 😂 \nfor me, deeper model, little augmentations are good", "770286": "", "761380": "I am also getting lost. My case for a single model CV: 0.99 and LB: 0.9568 and for 3 separate models it has CV: 0.99 but LB 0.92. What's wrong with me trying to figure it out. Can you please suggest anything like a single model vs multiple models? How people are getting higher LB? Using separate 3 models or single. ", "763882": "" } }