![\"\"](\"https://media2.giphy.com/media/st83jeYy9L6Bq/giphy.gif\")
Hello, Kaggle community 🖖
\n\nActively solving the competition last month, having spent a lot of time, at the moment we can't exceed 0.982 on the leaderboard.\nWe can no longer calculate how many models we have trained, how many augmentations and what variations we have tried. Different learning rates, scheduler tactics, losses, variations of cutmix / mixup - only few things given us much of a boost from score with simple NN. But 0.99 very far from us 😦.
\n\nIf you know the key to a good solution, or if we missed something, we'd appreciate your advice, thank you 🙏 .
", "messages": [ { "id": "772090", "postDate": "03/15/2020 02:02:19", "content": "Hello, Kaggle community 🖖
\n\nActively solving the competition last month, having spent a lot of time, at the moment we can't exceed 0.982 on the leaderboard.\nWe can no longer calculate how many models we have trained, how many augmentations and what variations we have tried. Different learning rates, scheduler tactics, losses, variations of cutmix / mixup - only few things given us much of a boost from score with simple NN. But 0.99 very far from us 😦.
\n\nIf you know the key to a good solution, or if we missed something, we'd appreciate your advice, thank you 🙏 .
", "rawMarkdown": "Hello, Kaggle community 🖖 \n\nActively solving the competition last month, having spent a lot of time, at the moment we can't exceed 0.982 on the leaderboard.\nWe can no longer calculate how many models we have trained, how many augmentations and what variations we have tried. Different learning rates, scheduler tactics, losses, variations of cutmix / mixup - only few things given us much of a boost from score with simple NN. But 0.99 very far from us 😦.\n\nIf you know the key to a good solution, or if we missed something, we'd appreciate your advice, thank you 🙏 .", "votes": null }, { "id": "772152", "postDate": "03/15/2020 05:14:43", "content": "
Same...I have absolutely NO idea what the top teams are doing...I guess that's what makes kaggle competitions so interesting...like intensive and extended puzzle-solving
", "rawMarkdown": "Same...I have absolutely NO idea what the top teams are doing...I guess that's what makes kaggle competitions so interesting...like intensive and extended puzzle-solving", "votes": null }, { "id": "772355", "postDate": "03/15/2020 11:55:49", "content": "We also have no idea of what we are doing :D
", "rawMarkdown": "We also have no idea of what we are doing :D", "votes": null }, { "id": "772446", "postDate": "03/15/2020 14:00:44", "content": "Same here.\nI have no idea what to do next.\nThis competition is puzzling for me while top kagglers are at top of the leader board.\nReally looking forward to watch top solutions.
", "rawMarkdown": "Same here.\nI have no idea what to do next.\nThis competition is puzzling for me while top kagglers are at top of the leader board.\nReally looking forward to watch top solutions.", "votes": null }, { "id": "772510", "postDate": "03/15/2020 15:20:19", "content": "I can't get past the 0.98 barrier, even with mixups on TPUs. I know there is a secret sauce invlolving cutmix and mixup mentioned in discussions. Can't get it working yet
", "rawMarkdown": "I can't get past the 0.98 barrier, even with mixups on TPUs. I know there is a secret sauce invlolving cutmix and mixup mentioned in discussions. Can't get it working yet", "votes": null }, { "id": "772531", "postDate": "03/15/2020 15:39:46", "content": "I like to know that too. Be patient. they will all share in a day from now
", "rawMarkdown": "I like to know that too. Be patient. they will all share in a day from now", "votes": null }, { "id": "772538", "postDate": "03/15/2020 15:45:37", "content": "Well, this is my first actively trying Kaggle competition and its about to end. I have learned so many things and touched so many coding stuff that I never did before such as various types of augmentation methods. However, I spent a lot of time with the cropped images that caused my low LB scores. I should consider it early. There are many sharing of improving the scores, so many ways; I would very much love to hear the solution after the competition is over. 🙂
", "rawMarkdown": "Well, this is my first actively trying Kaggle competition and its about to end. I have learned so many things and touched so many coding stuff that I never did before such as various types of augmentation methods. However, I spent a lot of time with the cropped images that caused my low LB scores. I should consider it early. There are many sharing of improving the scores, so many ways; I would very much love to hear the solution after the competition is over. 🙂", "votes": null }, { "id": "772548", "postDate": "03/15/2020 15:56:09", "content": "@ipythonx Exactly same feelings!
", "rawMarkdown": "ipythonx Exactly same feelings!", "votes": null }, { "id": "772553", "postDate": "03/15/2020 16:01:23", "content": "There are so many discussion threads about cut mix/mixup but it never helped me to cross even 0.97. I've tried many model architectures with it but the highest I could score is 0.9692. There must be something which I'm missing out!
", "rawMarkdown": "There are so many discussion threads about cut mix/mixup but it never helped me to cross even 0.97. I've tried many model architectures with it but the highest I could score is 0.9692. There must be something which I'm missing out!", "votes": null }, { "id": "772554", "postDate": "03/15/2020 16:06:17", "content": "Same! it's like struggling every submission, my improvement is so little or none at all, while the LB scores take off! Cannot wait to see the high score solutions!
", "rawMarkdown": "Same! it's like struggling every submission, my improvement is so little or none at all, while the LB scores take off! Cannot wait to see the high score solutions!", "votes": null }, { "id": "772594", "postDate": "03/15/2020 16:59:13", "content": "No idea what the secret is in this comp but I suspect one aspect that a lot of people haven't tried is adding extra data to the training set. I was only ever able to get minor improvements from doing it but considering there's almost 800 k extra images available, there should be some potential.
", "rawMarkdown": "No idea what the secret is in this comp but I suspect one aspect that a lot of people haven't tried is adding extra data to the training set. I was only ever able to get minor improvements from doing it but considering there's almost 800 k extra images available, there should be some potential.", "votes": null }, { "id": "772671", "postDate": "03/15/2020 18:43:02", "content": "I believe that top teams changed targets. Graphemes consist of simple graphemes (!). Today I have tried convert all targets into 105 simple graphemes. Now I can do training on almost any graphemes (ekush, matrivasha and fonts). Now my validation score is reached 0.989, I am waiting for my model - tomorrow will do submission :)
\n\nWish everyone good luck on Private LB
", "rawMarkdown": "I believe that top teams changed targets. Graphemes consist of simple graphemes (!). Today I have tried convert all targets into 105 simple graphemes. Now I can do training on almost any graphemes (ekush, matrivasha and fonts). Now my validation score is reached 0.989, I am waiting for my model - tomorrow will do submission :)\n\nWish everyone good luck on Private LB", "votes": null }, { "id": "772724", "postDate": "03/15/2020 20:22:33", "content": "I cannot wait to see the things you are doing that you dont know... if you decide to share after the competition of course :D
", "rawMarkdown": "I cannot wait to see the things you are doing that you dont know... if you decide to share after the competition of course :D", "votes": null }, { "id": "772732", "postDate": "03/15/2020 20:38:54", "content": "What data you are talking about?
", "rawMarkdown": "What data you are talking about?", "votes": null }, { "id": "774920", "postDate": "03/16/2020 04:00:03", "content": "We will share (as usual). But I'm sure I will learn a lot from top teams.
", "rawMarkdown": "We will share (as usual). But I'm sure I will learn a lot from top teams.", "votes": null }, { "id": "774938", "postDate": "03/16/2020 04:44:13", "content": "agree. \nthe secret should be the optimization of the the unseen samples.
", "rawMarkdown": "agree. \nthe secret should be the optimization of the the unseen samples.", "votes": null }, { "id": "774970", "postDate": "03/16/2020 06:01:08", "content": "Nice
", "rawMarkdown": "Nice", "votes": null }, { "id": "775016", "postDate": "03/16/2020 07:39:42", "content": "\"Should be\" or \"is\" hahaha....but anyways it's no use only couple of hours left....
", "rawMarkdown": "\"Should be\" or \"is\" hahaha....but anyways it's no use only couple of hours left....", "votes": null }, { "id": "775037", "postDate": "03/16/2020 07:58:49", "content": "Ekush, BanglaLekha-Isolated, CMATERdb. I have created files labeled according to the competition for these datasets, I can upload them after it ends I suppose
", "rawMarkdown": "Ekush, BanglaLekha-Isolated, CMATERdb. I have created files labeled according to the competition for these datasets, I can upload them after it ends I suppose", "votes": null }, { "id": "775042", "postDate": "03/16/2020 08:15:28", "content": "ok, is 😄
I feel like its a combintation of big image size, big network and a combination of cutmix/mixup/griddrop and ohem loss.
", "rawMarkdown": "I feel like its a combintation of big image size, big network and a combination of cutmix/mixup/griddrop and ohem loss.", "votes": null }, { "id": "775094", "postDate": "03/16/2020 09:46:12", "content": "", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 772152, "author_name": "kupchanski", "author_url": "", "post_date": "03/15/2020 05:14:43", "content": "
Same...I have absolutely NO idea what the top teams are doing...I guess that's what makes kaggle competitions so interesting...like intensive and extended puzzle-solving
", "votes": null, "replies": [ { "id": 772355, "author_name": "cpmpml", "author_url": "", "post_date": "03/15/2020 11:55:49", "content": "We also have no idea of what we are doing :D
", "votes": null, "replies": [] }, { "id": 772724, "author_name": "yl1202", "author_url": "", "post_date": "03/15/2020 20:22:33", "content": "I cannot wait to see the things you are doing that you dont know... if you decide to share after the competition of course :D
", "votes": null, "replies": [] }, { "id": 774920, "author_name": "cpmpml", "author_url": "", "post_date": "03/16/2020 04:00:03", "content": "We will share (as usual). But I'm sure I will learn a lot from top teams.
", "votes": null, "replies": [] }, { "id": 774938, "author_name": "garybios", "author_url": "", "post_date": "03/16/2020 04:44:13", "content": "agree. \nthe secret should be the optimization of the the unseen samples.
", "votes": null, "replies": [] }, { "id": 775016, "author_name": "roguekk007", "author_url": "", "post_date": "03/16/2020 07:39:42", "content": "\"Should be\" or \"is\" hahaha....but anyways it's no use only couple of hours left....
", "votes": null, "replies": [] }, { "id": 775042, "author_name": "garybios", "author_url": "", "post_date": "03/16/2020 08:15:28", "content": "ok, is 😄
Same here.\nI have no idea what to do next.\nThis competition is puzzling for me while top kagglers are at top of the leader board.\nReally looking forward to watch top solutions.
", "votes": null, "replies": [] }, { "id": 772510, "author_name": "kurianbenoy", "author_url": "", "post_date": "03/15/2020 15:20:19", "content": "I can't get past the 0.98 barrier, even with mixups on TPUs. I know there is a secret sauce invlolving cutmix and mixup mentioned in discussions. Can't get it working yet
", "votes": null, "replies": [] }, { "id": 772531, "author_name": "mhviraf", "author_url": "", "post_date": "03/15/2020 15:39:46", "content": "I like to know that too. Be patient. they will all share in a day from now
", "votes": null, "replies": [] }, { "id": 772538, "author_name": "ipythonx", "author_url": "", "post_date": "03/15/2020 15:45:37", "content": "Well, this is my first actively trying Kaggle competition and its about to end. I have learned so many things and touched so many coding stuff that I never did before such as various types of augmentation methods. However, I spent a lot of time with the cropped images that caused my low LB scores. I should consider it early. There are many sharing of improving the scores, so many ways; I would very much love to hear the solution after the competition is over. 🙂
", "votes": null, "replies": [ { "id": 772548, "author_name": "kaushal2896", "author_url": "", "post_date": "03/15/2020 15:56:09", "content": "@ipythonx Exactly same feelings!
", "votes": null, "replies": [] } ] }, { "id": 772553, "author_name": "kaushal2896", "author_url": "", "post_date": "03/15/2020 16:01:23", "content": "There are so many discussion threads about cut mix/mixup but it never helped me to cross even 0.97. I've tried many model architectures with it but the highest I could score is 0.9692. There must be something which I'm missing out!
", "votes": null, "replies": [] }, { "id": 772554, "author_name": "yuanlin08", "author_url": "", "post_date": "03/15/2020 16:06:17", "content": "Same! it's like struggling every submission, my improvement is so little or none at all, while the LB scores take off! Cannot wait to see the high score solutions!
", "votes": null, "replies": [] }, { "id": 772594, "author_name": "maxjon", "author_url": "", "post_date": "03/15/2020 16:59:13", "content": "No idea what the secret is in this comp but I suspect one aspect that a lot of people haven't tried is adding extra data to the training set. I was only ever able to get minor improvements from doing it but considering there's almost 800 k extra images available, there should be some potential.
", "votes": null, "replies": [ { "id": 772732, "author_name": "dukhovnik", "author_url": "", "post_date": "03/15/2020 20:38:54", "content": "What data you are talking about?
", "votes": null, "replies": [] }, { "id": 775037, "author_name": "maxjon", "author_url": "", "post_date": "03/16/2020 07:58:49", "content": "Ekush, BanglaLekha-Isolated, CMATERdb. I have created files labeled according to the competition for these datasets, I can upload them after it ends I suppose
", "votes": null, "replies": [] } ] }, { "id": 772671, "author_name": "shonenkov", "author_url": "", "post_date": "03/15/2020 18:43:02", "content": "I believe that top teams changed targets. Graphemes consist of simple graphemes (!). Today I have tried convert all targets into 105 simple graphemes. Now I can do training on almost any graphemes (ekush, matrivasha and fonts). Now my validation score is reached 0.989, I am waiting for my model - tomorrow will do submission :)
\n\nWish everyone good luck on Private LB
", "votes": null, "replies": [ { "id": 775094, "author_name": "seshurajup", "author_url": "", "post_date": "03/16/2020 09:46:12", "content": "", "votes": null, "replies": [] } ] }, { "id": 774970, "author_name": "mohsinusmani", "author_url": "", "post_date": "03/16/2020 06:01:08", "content": "Nice
", "votes": null, "replies": [] }, { "id": 775078, "author_name": "garyongguanjie", "author_url": "", "post_date": "03/16/2020 09:21:07", "content": "I feel like its a combintation of big image size, big network and a combination of cutmix/mixup/griddrop and ohem loss.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "772090": "Hello, Kaggle community 🖖 \n\nActively solving the competition last month, having spent a lot of time, at the moment we can't exceed 0.982 on the leaderboard.\nWe can no longer calculate how many models we have trained, how many augmentations and what variations we have tried. Different learning rates, scheduler tactics, losses, variations of cutmix / mixup - only few things given us much of a boost from score with simple NN. But 0.99 very far from us 😦.\n\nIf you know the key to a good solution, or if we missed something, we'd appreciate your advice, thank you 🙏 .", "772152": "", "772278": "Same...I have absolutely NO idea what the top teams are doing...I guess that's what makes kaggle competitions so interesting...like intensive and extended puzzle-solving", "772355": "We also have no idea of what we are doing :D", "772446": "Same here.\nI have no idea what to do next.\nThis competition is puzzling for me while top kagglers are at top of the leader board.\nReally looking forward to watch top solutions.", "772510": "I can't get past the 0.98 barrier, even with mixups on TPUs. I know there is a secret sauce invlolving cutmix and mixup mentioned in discussions. Can't get it working yet", "772531": "I like to know that too. Be patient. they will all share in a day from now", "772538": "Well, this is my first actively trying Kaggle competition and its about to end. I have learned so many things and touched so many coding stuff that I never did before such as various types of augmentation methods. However, I spent a lot of time with the cropped images that caused my low LB scores. I should consider it early. There are many sharing of improving the scores, so many ways; I would very much love to hear the solution after the competition is over. 🙂", "772548": "ipythonx Exactly same feelings!", "772553": "There are so many discussion threads about cut mix/mixup but it never helped me to cross even 0.97. I've tried many model architectures with it but the highest I could score is 0.9692. There must be something which I'm missing out!", "772554": "Same! it's like struggling every submission, my improvement is so little or none at all, while the LB scores take off! Cannot wait to see the high score solutions!", "772594": "No idea what the secret is in this comp but I suspect one aspect that a lot of people haven't tried is adding extra data to the training set. I was only ever able to get minor improvements from doing it but considering there's almost 800 k extra images available, there should be some potential.", "772671": "I believe that top teams changed targets. Graphemes consist of simple graphemes (!). Today I have tried convert all targets into 105 simple graphemes. Now I can do training on almost any graphemes (ekush, matrivasha and fonts). Now my validation score is reached 0.989, I am waiting for my model - tomorrow will do submission :)\n\nWish everyone good luck on Private LB", "772724": "I cannot wait to see the things you are doing that you dont know... if you decide to share after the competition of course :D", "772732": "What data you are talking about?", "774920": "We will share (as usual). But I'm sure I will learn a lot from top teams.", "774938": "agree. \nthe secret should be the optimization of the the unseen samples.", "774970": "Nice", "775016": "\"Should be\" or \"is\" hahaha....but anyways it's no use only couple of hours left....", "775037": "Ekush, BanglaLekha-Isolated, CMATERdb. I have created files labeled according to the competition for these datasets, I can upload them after it ends I suppose", "775042": "ok, `is` 😄", "775078": "I feel like its a combintation of big image size, big network and a combination of cutmix/mixup/griddrop and ohem loss.", "775094": "" }, "source": "meta" }